* [PATCH v1 2/3] CryptoPkg/OpensslLib: Commit the auto-generated assembly files for IA32
2022-09-21 20:25 [PATCH v1 0/3] CryptoPkg/OpensslLib: Add native instruction support for IA32 Christopher Zurcher
2022-09-21 20:25 ` [PATCH v1 1/3] " Christopher Zurcher
@ 2022-09-21 20:25 ` Christopher Zurcher
2022-09-21 20:25 ` [PATCH v1 3/3] CryptoPkg/OpensslLib: Update generated files for native X64 Christopher Zurcher
2022-09-22 0:53 ` [PATCH v1 0/3] CryptoPkg/OpensslLib: Add native instruction support for IA32 Yao, Jiewen
3 siblings, 0 replies; 8+ messages in thread
From: Christopher Zurcher @ 2022-09-21 20:25 UTC (permalink / raw)
To: devel; +Cc: Yi Li, Jiewen Yao, Jian J Wang, Xiaoyu Lu, Guomin Jiang
REF: https://bugzilla.tianocore.org/show_bug.cgi?id=3654
Adding the auto-generated assembly files for IA32 architectures.
Cc: Yi Li <yi1.li@intel.com>
Cc: Jiewen Yao <jiewen.yao@intel.com>
Cc: Jian J Wang <jian.j.wang@intel.com>
Cc: Xiaoyu Lu <xiaoyu1.lu@intel.com>
Cc: Guomin Jiang <guomin.jiang@intel.com>
Signed-off-by: Christopher Zurcher <christopher.zurcher@microsoft.com>
---
CryptoPkg/Library/OpensslLib/IA32/crypto/aes/aesni-x86.nasm | 3212 +++++++++++++++++++
CryptoPkg/Library/OpensslLib/IA32/crypto/aes/vpaes-x86.nasm | 651 ++++
CryptoPkg/Library/OpensslLib/IA32/crypto/modes/ghash-x86.nasm | 700 ++++
CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha1-586.nasm | 1394 ++++++++
CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha256-586.nasm | 3364 ++++++++++++++++++++
CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha512-586.nasm | 579 ++++
CryptoPkg/Library/OpensslLib/IA32/crypto/x86cpuid.nasm | 433 +++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/aesni-x86.S | 3247 +++++++++++++++++++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/vpaes-x86.S | 670 ++++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/modes/ghash-x86.S | 703 ++++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha1-586.S | 1389 ++++++++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha256-586.S | 3356 +++++++++++++++++++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha512-586.S | 574 ++++
CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/x86cpuid.S | 449 +++
14 files changed, 20721 insertions(+)
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/aes/aesni-x86.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/aes/aesni-x86.nasm
new file mode 100644
index 0000000000..e0b5058273
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/aes/aesni-x86.nasm
@@ -0,0 +1,3212 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/aes/asm/aesni-x86.pl
+;
+; Copyright 2009-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+;extern _OPENSSL_ia32cap_P
+global _aesni_encrypt
+align 16
+_aesni_encrypt:
+L$_aesni_encrypt_begin:
+ mov eax,DWORD [4+esp]
+ mov edx,DWORD [12+esp]
+ movups xmm2,[eax]
+ mov ecx,DWORD [240+edx]
+ mov eax,DWORD [8+esp]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$000enc1_loop_1:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$000enc1_loop_1
+db 102,15,56,221,209
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ movups [eax],xmm2
+ pxor xmm2,xmm2
+ ret
+global _aesni_decrypt
+align 16
+_aesni_decrypt:
+L$_aesni_decrypt_begin:
+ mov eax,DWORD [4+esp]
+ mov edx,DWORD [12+esp]
+ movups xmm2,[eax]
+ mov ecx,DWORD [240+edx]
+ mov eax,DWORD [8+esp]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$001dec1_loop_2:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$001dec1_loop_2
+db 102,15,56,223,209
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ movups [eax],xmm2
+ pxor xmm2,xmm2
+ ret
+align 16
+__aesni_encrypt2:
+ movups xmm0,[edx]
+ shl ecx,4
+ movups xmm1,[16+edx]
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ movups xmm0,[32+edx]
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+ add ecx,16
+L$002enc2_loop:
+db 102,15,56,220,209
+db 102,15,56,220,217
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,220,208
+db 102,15,56,220,216
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$002enc2_loop
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,221,208
+db 102,15,56,221,216
+ ret
+align 16
+__aesni_decrypt2:
+ movups xmm0,[edx]
+ shl ecx,4
+ movups xmm1,[16+edx]
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ movups xmm0,[32+edx]
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+ add ecx,16
+L$003dec2_loop:
+db 102,15,56,222,209
+db 102,15,56,222,217
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,222,208
+db 102,15,56,222,216
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$003dec2_loop
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,223,208
+db 102,15,56,223,216
+ ret
+align 16
+__aesni_encrypt3:
+ movups xmm0,[edx]
+ shl ecx,4
+ movups xmm1,[16+edx]
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+ movups xmm0,[32+edx]
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+ add ecx,16
+L$004enc3_loop:
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,220,208
+db 102,15,56,220,216
+db 102,15,56,220,224
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$004enc3_loop
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+db 102,15,56,221,208
+db 102,15,56,221,216
+db 102,15,56,221,224
+ ret
+align 16
+__aesni_decrypt3:
+ movups xmm0,[edx]
+ shl ecx,4
+ movups xmm1,[16+edx]
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+ movups xmm0,[32+edx]
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+ add ecx,16
+L$005dec3_loop:
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,222,208
+db 102,15,56,222,216
+db 102,15,56,222,224
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$005dec3_loop
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+db 102,15,56,223,208
+db 102,15,56,223,216
+db 102,15,56,223,224
+ ret
+align 16
+__aesni_encrypt4:
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ shl ecx,4
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+ pxor xmm5,xmm0
+ movups xmm0,[32+edx]
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+db 15,31,64,0
+ add ecx,16
+L$006enc4_loop:
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+db 102,15,56,220,233
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,220,208
+db 102,15,56,220,216
+db 102,15,56,220,224
+db 102,15,56,220,232
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$006enc4_loop
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+db 102,15,56,220,233
+db 102,15,56,221,208
+db 102,15,56,221,216
+db 102,15,56,221,224
+db 102,15,56,221,232
+ ret
+align 16
+__aesni_decrypt4:
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ shl ecx,4
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+ pxor xmm5,xmm0
+ movups xmm0,[32+edx]
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+db 15,31,64,0
+ add ecx,16
+L$007dec4_loop:
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+db 102,15,56,222,233
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,222,208
+db 102,15,56,222,216
+db 102,15,56,222,224
+db 102,15,56,222,232
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$007dec4_loop
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+db 102,15,56,222,233
+db 102,15,56,223,208
+db 102,15,56,223,216
+db 102,15,56,223,224
+db 102,15,56,223,232
+ ret
+align 16
+__aesni_encrypt6:
+ movups xmm0,[edx]
+ shl ecx,4
+ movups xmm1,[16+edx]
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+db 102,15,56,220,209
+ pxor xmm5,xmm0
+ pxor xmm6,xmm0
+db 102,15,56,220,217
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+db 102,15,56,220,225
+ pxor xmm7,xmm0
+ movups xmm0,[ecx*1+edx]
+ add ecx,16
+ jmp NEAR L$008_aesni_encrypt6_inner
+align 16
+L$009enc6_loop:
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+L$008_aesni_encrypt6_inner:
+db 102,15,56,220,233
+db 102,15,56,220,241
+db 102,15,56,220,249
+L$_aesni_encrypt6_enter:
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,220,208
+db 102,15,56,220,216
+db 102,15,56,220,224
+db 102,15,56,220,232
+db 102,15,56,220,240
+db 102,15,56,220,248
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$009enc6_loop
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+db 102,15,56,220,233
+db 102,15,56,220,241
+db 102,15,56,220,249
+db 102,15,56,221,208
+db 102,15,56,221,216
+db 102,15,56,221,224
+db 102,15,56,221,232
+db 102,15,56,221,240
+db 102,15,56,221,248
+ ret
+align 16
+__aesni_decrypt6:
+ movups xmm0,[edx]
+ shl ecx,4
+ movups xmm1,[16+edx]
+ xorps xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+db 102,15,56,222,209
+ pxor xmm5,xmm0
+ pxor xmm6,xmm0
+db 102,15,56,222,217
+ lea edx,[32+ecx*1+edx]
+ neg ecx
+db 102,15,56,222,225
+ pxor xmm7,xmm0
+ movups xmm0,[ecx*1+edx]
+ add ecx,16
+ jmp NEAR L$010_aesni_decrypt6_inner
+align 16
+L$011dec6_loop:
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+L$010_aesni_decrypt6_inner:
+db 102,15,56,222,233
+db 102,15,56,222,241
+db 102,15,56,222,249
+L$_aesni_decrypt6_enter:
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,222,208
+db 102,15,56,222,216
+db 102,15,56,222,224
+db 102,15,56,222,232
+db 102,15,56,222,240
+db 102,15,56,222,248
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$011dec6_loop
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+db 102,15,56,222,233
+db 102,15,56,222,241
+db 102,15,56,222,249
+db 102,15,56,223,208
+db 102,15,56,223,216
+db 102,15,56,223,224
+db 102,15,56,223,232
+db 102,15,56,223,240
+db 102,15,56,223,248
+ ret
+global _aesni_ecb_encrypt
+align 16
+_aesni_ecb_encrypt:
+L$_aesni_ecb_encrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ mov ebx,DWORD [36+esp]
+ and eax,-16
+ jz NEAR L$012ecb_ret
+ mov ecx,DWORD [240+edx]
+ test ebx,ebx
+ jz NEAR L$013ecb_decrypt
+ mov ebp,edx
+ mov ebx,ecx
+ cmp eax,96
+ jb NEAR L$014ecb_enc_tail
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ movdqu xmm7,[80+esi]
+ lea esi,[96+esi]
+ sub eax,96
+ jmp NEAR L$015ecb_enc_loop6_enter
+align 16
+L$016ecb_enc_loop6:
+ movups [edi],xmm2
+ movdqu xmm2,[esi]
+ movups [16+edi],xmm3
+ movdqu xmm3,[16+esi]
+ movups [32+edi],xmm4
+ movdqu xmm4,[32+esi]
+ movups [48+edi],xmm5
+ movdqu xmm5,[48+esi]
+ movups [64+edi],xmm6
+ movdqu xmm6,[64+esi]
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ movdqu xmm7,[80+esi]
+ lea esi,[96+esi]
+L$015ecb_enc_loop6_enter:
+ call __aesni_encrypt6
+ mov edx,ebp
+ mov ecx,ebx
+ sub eax,96
+ jnc NEAR L$016ecb_enc_loop6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ add eax,96
+ jz NEAR L$012ecb_ret
+L$014ecb_enc_tail:
+ movups xmm2,[esi]
+ cmp eax,32
+ jb NEAR L$017ecb_enc_one
+ movups xmm3,[16+esi]
+ je NEAR L$018ecb_enc_two
+ movups xmm4,[32+esi]
+ cmp eax,64
+ jb NEAR L$019ecb_enc_three
+ movups xmm5,[48+esi]
+ je NEAR L$020ecb_enc_four
+ movups xmm6,[64+esi]
+ xorps xmm7,xmm7
+ call __aesni_encrypt6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ jmp NEAR L$012ecb_ret
+align 16
+L$017ecb_enc_one:
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$021enc1_loop_3:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$021enc1_loop_3
+db 102,15,56,221,209
+ movups [edi],xmm2
+ jmp NEAR L$012ecb_ret
+align 16
+L$018ecb_enc_two:
+ call __aesni_encrypt2
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ jmp NEAR L$012ecb_ret
+align 16
+L$019ecb_enc_three:
+ call __aesni_encrypt3
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ jmp NEAR L$012ecb_ret
+align 16
+L$020ecb_enc_four:
+ call __aesni_encrypt4
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ jmp NEAR L$012ecb_ret
+align 16
+L$013ecb_decrypt:
+ mov ebp,edx
+ mov ebx,ecx
+ cmp eax,96
+ jb NEAR L$022ecb_dec_tail
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ movdqu xmm7,[80+esi]
+ lea esi,[96+esi]
+ sub eax,96
+ jmp NEAR L$023ecb_dec_loop6_enter
+align 16
+L$024ecb_dec_loop6:
+ movups [edi],xmm2
+ movdqu xmm2,[esi]
+ movups [16+edi],xmm3
+ movdqu xmm3,[16+esi]
+ movups [32+edi],xmm4
+ movdqu xmm4,[32+esi]
+ movups [48+edi],xmm5
+ movdqu xmm5,[48+esi]
+ movups [64+edi],xmm6
+ movdqu xmm6,[64+esi]
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ movdqu xmm7,[80+esi]
+ lea esi,[96+esi]
+L$023ecb_dec_loop6_enter:
+ call __aesni_decrypt6
+ mov edx,ebp
+ mov ecx,ebx
+ sub eax,96
+ jnc NEAR L$024ecb_dec_loop6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ add eax,96
+ jz NEAR L$012ecb_ret
+L$022ecb_dec_tail:
+ movups xmm2,[esi]
+ cmp eax,32
+ jb NEAR L$025ecb_dec_one
+ movups xmm3,[16+esi]
+ je NEAR L$026ecb_dec_two
+ movups xmm4,[32+esi]
+ cmp eax,64
+ jb NEAR L$027ecb_dec_three
+ movups xmm5,[48+esi]
+ je NEAR L$028ecb_dec_four
+ movups xmm6,[64+esi]
+ xorps xmm7,xmm7
+ call __aesni_decrypt6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ jmp NEAR L$012ecb_ret
+align 16
+L$025ecb_dec_one:
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$029dec1_loop_4:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$029dec1_loop_4
+db 102,15,56,223,209
+ movups [edi],xmm2
+ jmp NEAR L$012ecb_ret
+align 16
+L$026ecb_dec_two:
+ call __aesni_decrypt2
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ jmp NEAR L$012ecb_ret
+align 16
+L$027ecb_dec_three:
+ call __aesni_decrypt3
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ jmp NEAR L$012ecb_ret
+align 16
+L$028ecb_dec_four:
+ call __aesni_decrypt4
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+L$012ecb_ret:
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ pxor xmm5,xmm5
+ pxor xmm6,xmm6
+ pxor xmm7,xmm7
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_ccm64_encrypt_blocks
+align 16
+_aesni_ccm64_encrypt_blocks:
+L$_aesni_ccm64_encrypt_blocks_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ mov ebx,DWORD [36+esp]
+ mov ecx,DWORD [40+esp]
+ mov ebp,esp
+ sub esp,60
+ and esp,-16
+ mov DWORD [48+esp],ebp
+ movdqu xmm7,[ebx]
+ movdqu xmm3,[ecx]
+ mov ecx,DWORD [240+edx]
+ mov DWORD [esp],202182159
+ mov DWORD [4+esp],134810123
+ mov DWORD [8+esp],67438087
+ mov DWORD [12+esp],66051
+ mov ebx,1
+ xor ebp,ebp
+ mov DWORD [16+esp],ebx
+ mov DWORD [20+esp],ebp
+ mov DWORD [24+esp],ebp
+ mov DWORD [28+esp],ebp
+ shl ecx,4
+ mov ebx,16
+ lea ebp,[edx]
+ movdqa xmm5,[esp]
+ movdqa xmm2,xmm7
+ lea edx,[32+ecx*1+edx]
+ sub ebx,ecx
+db 102,15,56,0,253
+L$030ccm64_enc_outer:
+ movups xmm0,[ebp]
+ mov ecx,ebx
+ movups xmm6,[esi]
+ xorps xmm2,xmm0
+ movups xmm1,[16+ebp]
+ xorps xmm0,xmm6
+ xorps xmm3,xmm0
+ movups xmm0,[32+ebp]
+L$031ccm64_enc2_loop:
+db 102,15,56,220,209
+db 102,15,56,220,217
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,220,208
+db 102,15,56,220,216
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$031ccm64_enc2_loop
+db 102,15,56,220,209
+db 102,15,56,220,217
+ paddq xmm7,[16+esp]
+ dec eax
+db 102,15,56,221,208
+db 102,15,56,221,216
+ lea esi,[16+esi]
+ xorps xmm6,xmm2
+ movdqa xmm2,xmm7
+ movups [edi],xmm6
+db 102,15,56,0,213
+ lea edi,[16+edi]
+ jnz NEAR L$030ccm64_enc_outer
+ mov esp,DWORD [48+esp]
+ mov edi,DWORD [40+esp]
+ movups [edi],xmm3
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ pxor xmm5,xmm5
+ pxor xmm6,xmm6
+ pxor xmm7,xmm7
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_ccm64_decrypt_blocks
+align 16
+_aesni_ccm64_decrypt_blocks:
+L$_aesni_ccm64_decrypt_blocks_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ mov ebx,DWORD [36+esp]
+ mov ecx,DWORD [40+esp]
+ mov ebp,esp
+ sub esp,60
+ and esp,-16
+ mov DWORD [48+esp],ebp
+ movdqu xmm7,[ebx]
+ movdqu xmm3,[ecx]
+ mov ecx,DWORD [240+edx]
+ mov DWORD [esp],202182159
+ mov DWORD [4+esp],134810123
+ mov DWORD [8+esp],67438087
+ mov DWORD [12+esp],66051
+ mov ebx,1
+ xor ebp,ebp
+ mov DWORD [16+esp],ebx
+ mov DWORD [20+esp],ebp
+ mov DWORD [24+esp],ebp
+ mov DWORD [28+esp],ebp
+ movdqa xmm5,[esp]
+ movdqa xmm2,xmm7
+ mov ebp,edx
+ mov ebx,ecx
+db 102,15,56,0,253
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$032enc1_loop_5:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$032enc1_loop_5
+db 102,15,56,221,209
+ shl ebx,4
+ mov ecx,16
+ movups xmm6,[esi]
+ paddq xmm7,[16+esp]
+ lea esi,[16+esi]
+ sub ecx,ebx
+ lea edx,[32+ebx*1+ebp]
+ mov ebx,ecx
+ jmp NEAR L$033ccm64_dec_outer
+align 16
+L$033ccm64_dec_outer:
+ xorps xmm6,xmm2
+ movdqa xmm2,xmm7
+ movups [edi],xmm6
+ lea edi,[16+edi]
+db 102,15,56,0,213
+ sub eax,1
+ jz NEAR L$034ccm64_dec_break
+ movups xmm0,[ebp]
+ mov ecx,ebx
+ movups xmm1,[16+ebp]
+ xorps xmm6,xmm0
+ xorps xmm2,xmm0
+ xorps xmm3,xmm6
+ movups xmm0,[32+ebp]
+L$035ccm64_dec2_loop:
+db 102,15,56,220,209
+db 102,15,56,220,217
+ movups xmm1,[ecx*1+edx]
+ add ecx,32
+db 102,15,56,220,208
+db 102,15,56,220,216
+ movups xmm0,[ecx*1+edx-16]
+ jnz NEAR L$035ccm64_dec2_loop
+ movups xmm6,[esi]
+ paddq xmm7,[16+esp]
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,221,208
+db 102,15,56,221,216
+ lea esi,[16+esi]
+ jmp NEAR L$033ccm64_dec_outer
+align 16
+L$034ccm64_dec_break:
+ mov ecx,DWORD [240+ebp]
+ mov edx,ebp
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ xorps xmm6,xmm0
+ lea edx,[32+edx]
+ xorps xmm3,xmm6
+L$036enc1_loop_6:
+db 102,15,56,220,217
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$036enc1_loop_6
+db 102,15,56,221,217
+ mov esp,DWORD [48+esp]
+ mov edi,DWORD [40+esp]
+ movups [edi],xmm3
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ pxor xmm5,xmm5
+ pxor xmm6,xmm6
+ pxor xmm7,xmm7
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_ctr32_encrypt_blocks
+align 16
+_aesni_ctr32_encrypt_blocks:
+L$_aesni_ctr32_encrypt_blocks_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ mov ebx,DWORD [36+esp]
+ mov ebp,esp
+ sub esp,88
+ and esp,-16
+ mov DWORD [80+esp],ebp
+ cmp eax,1
+ je NEAR L$037ctr32_one_shortcut
+ movdqu xmm7,[ebx]
+ mov DWORD [esp],202182159
+ mov DWORD [4+esp],134810123
+ mov DWORD [8+esp],67438087
+ mov DWORD [12+esp],66051
+ mov ecx,6
+ xor ebp,ebp
+ mov DWORD [16+esp],ecx
+ mov DWORD [20+esp],ecx
+ mov DWORD [24+esp],ecx
+ mov DWORD [28+esp],ebp
+db 102,15,58,22,251,3
+db 102,15,58,34,253,3
+ mov ecx,DWORD [240+edx]
+ bswap ebx
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ movdqa xmm2,[esp]
+db 102,15,58,34,195,0
+ lea ebp,[3+ebx]
+db 102,15,58,34,205,0
+ inc ebx
+db 102,15,58,34,195,1
+ inc ebp
+db 102,15,58,34,205,1
+ inc ebx
+db 102,15,58,34,195,2
+ inc ebp
+db 102,15,58,34,205,2
+ movdqa [48+esp],xmm0
+db 102,15,56,0,194
+ movdqu xmm6,[edx]
+ movdqa [64+esp],xmm1
+db 102,15,56,0,202
+ pshufd xmm2,xmm0,192
+ pshufd xmm3,xmm0,128
+ cmp eax,6
+ jb NEAR L$038ctr32_tail
+ pxor xmm7,xmm6
+ shl ecx,4
+ mov ebx,16
+ movdqa [32+esp],xmm7
+ mov ebp,edx
+ sub ebx,ecx
+ lea edx,[32+ecx*1+edx]
+ sub eax,6
+ jmp NEAR L$039ctr32_loop6
+align 16
+L$039ctr32_loop6:
+ pshufd xmm4,xmm0,64
+ movdqa xmm0,[32+esp]
+ pshufd xmm5,xmm1,192
+ pxor xmm2,xmm0
+ pshufd xmm6,xmm1,128
+ pxor xmm3,xmm0
+ pshufd xmm7,xmm1,64
+ movups xmm1,[16+ebp]
+ pxor xmm4,xmm0
+ pxor xmm5,xmm0
+db 102,15,56,220,209
+ pxor xmm6,xmm0
+ pxor xmm7,xmm0
+db 102,15,56,220,217
+ movups xmm0,[32+ebp]
+ mov ecx,ebx
+db 102,15,56,220,225
+db 102,15,56,220,233
+db 102,15,56,220,241
+db 102,15,56,220,249
+ call L$_aesni_encrypt6_enter
+ movups xmm1,[esi]
+ movups xmm0,[16+esi]
+ xorps xmm2,xmm1
+ movups xmm1,[32+esi]
+ xorps xmm3,xmm0
+ movups [edi],xmm2
+ movdqa xmm0,[16+esp]
+ xorps xmm4,xmm1
+ movdqa xmm1,[64+esp]
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ paddd xmm1,xmm0
+ paddd xmm0,[48+esp]
+ movdqa xmm2,[esp]
+ movups xmm3,[48+esi]
+ movups xmm4,[64+esi]
+ xorps xmm5,xmm3
+ movups xmm3,[80+esi]
+ lea esi,[96+esi]
+ movdqa [48+esp],xmm0
+db 102,15,56,0,194
+ xorps xmm6,xmm4
+ movups [48+edi],xmm5
+ xorps xmm7,xmm3
+ movdqa [64+esp],xmm1
+db 102,15,56,0,202
+ movups [64+edi],xmm6
+ pshufd xmm2,xmm0,192
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ pshufd xmm3,xmm0,128
+ sub eax,6
+ jnc NEAR L$039ctr32_loop6
+ add eax,6
+ jz NEAR L$040ctr32_ret
+ movdqu xmm7,[ebp]
+ mov edx,ebp
+ pxor xmm7,[32+esp]
+ mov ecx,DWORD [240+ebp]
+L$038ctr32_tail:
+ por xmm2,xmm7
+ cmp eax,2
+ jb NEAR L$041ctr32_one
+ pshufd xmm4,xmm0,64
+ por xmm3,xmm7
+ je NEAR L$042ctr32_two
+ pshufd xmm5,xmm1,192
+ por xmm4,xmm7
+ cmp eax,4
+ jb NEAR L$043ctr32_three
+ pshufd xmm6,xmm1,128
+ por xmm5,xmm7
+ je NEAR L$044ctr32_four
+ por xmm6,xmm7
+ call __aesni_encrypt6
+ movups xmm1,[esi]
+ movups xmm0,[16+esi]
+ xorps xmm2,xmm1
+ movups xmm1,[32+esi]
+ xorps xmm3,xmm0
+ movups xmm0,[48+esi]
+ xorps xmm4,xmm1
+ movups xmm1,[64+esi]
+ xorps xmm5,xmm0
+ movups [edi],xmm2
+ xorps xmm6,xmm1
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ jmp NEAR L$040ctr32_ret
+align 16
+L$037ctr32_one_shortcut:
+ movups xmm2,[ebx]
+ mov ecx,DWORD [240+edx]
+L$041ctr32_one:
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$045enc1_loop_7:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$045enc1_loop_7
+db 102,15,56,221,209
+ movups xmm6,[esi]
+ xorps xmm6,xmm2
+ movups [edi],xmm6
+ jmp NEAR L$040ctr32_ret
+align 16
+L$042ctr32_two:
+ call __aesni_encrypt2
+ movups xmm5,[esi]
+ movups xmm6,[16+esi]
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ jmp NEAR L$040ctr32_ret
+align 16
+L$043ctr32_three:
+ call __aesni_encrypt3
+ movups xmm5,[esi]
+ movups xmm6,[16+esi]
+ xorps xmm2,xmm5
+ movups xmm7,[32+esi]
+ xorps xmm3,xmm6
+ movups [edi],xmm2
+ xorps xmm4,xmm7
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ jmp NEAR L$040ctr32_ret
+align 16
+L$044ctr32_four:
+ call __aesni_encrypt4
+ movups xmm6,[esi]
+ movups xmm7,[16+esi]
+ movups xmm1,[32+esi]
+ xorps xmm2,xmm6
+ movups xmm0,[48+esi]
+ xorps xmm3,xmm7
+ movups [edi],xmm2
+ xorps xmm4,xmm1
+ movups [16+edi],xmm3
+ xorps xmm5,xmm0
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+L$040ctr32_ret:
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ movdqa [32+esp],xmm0
+ pxor xmm5,xmm5
+ movdqa [48+esp],xmm0
+ pxor xmm6,xmm6
+ movdqa [64+esp],xmm0
+ pxor xmm7,xmm7
+ mov esp,DWORD [80+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_xts_encrypt
+align 16
+_aesni_xts_encrypt:
+L$_aesni_xts_encrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov edx,DWORD [36+esp]
+ mov esi,DWORD [40+esp]
+ mov ecx,DWORD [240+edx]
+ movups xmm2,[esi]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$046enc1_loop_8:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$046enc1_loop_8
+db 102,15,56,221,209
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ mov ebp,esp
+ sub esp,120
+ mov ecx,DWORD [240+edx]
+ and esp,-16
+ mov DWORD [96+esp],135
+ mov DWORD [100+esp],0
+ mov DWORD [104+esp],1
+ mov DWORD [108+esp],0
+ mov DWORD [112+esp],eax
+ mov DWORD [116+esp],ebp
+ movdqa xmm1,xmm2
+ pxor xmm0,xmm0
+ movdqa xmm3,[96+esp]
+ pcmpgtd xmm0,xmm1
+ and eax,-16
+ mov ebp,edx
+ mov ebx,ecx
+ sub eax,96
+ jc NEAR L$047xts_enc_short
+ shl ecx,4
+ mov ebx,16
+ sub ebx,ecx
+ lea edx,[32+ecx*1+edx]
+ jmp NEAR L$048xts_enc_loop6
+align 16
+L$048xts_enc_loop6:
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [16+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [32+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [48+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm7,xmm0,19
+ movdqa [64+esp],xmm1
+ paddq xmm1,xmm1
+ movups xmm0,[ebp]
+ pand xmm7,xmm3
+ movups xmm2,[esi]
+ pxor xmm7,xmm1
+ mov ecx,ebx
+ movdqu xmm3,[16+esi]
+ xorps xmm2,xmm0
+ movdqu xmm4,[32+esi]
+ pxor xmm3,xmm0
+ movdqu xmm5,[48+esi]
+ pxor xmm4,xmm0
+ movdqu xmm6,[64+esi]
+ pxor xmm5,xmm0
+ movdqu xmm1,[80+esi]
+ pxor xmm6,xmm0
+ lea esi,[96+esi]
+ pxor xmm2,[esp]
+ movdqa [80+esp],xmm7
+ pxor xmm7,xmm1
+ movups xmm1,[16+ebp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+db 102,15,56,220,209
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+db 102,15,56,220,217
+ pxor xmm7,xmm0
+ movups xmm0,[32+ebp]
+db 102,15,56,220,225
+db 102,15,56,220,233
+db 102,15,56,220,241
+db 102,15,56,220,249
+ call L$_aesni_encrypt6_enter
+ movdqa xmm1,[80+esp]
+ pxor xmm0,xmm0
+ xorps xmm2,[esp]
+ pcmpgtd xmm0,xmm1
+ xorps xmm3,[16+esp]
+ movups [edi],xmm2
+ xorps xmm4,[32+esp]
+ movups [16+edi],xmm3
+ xorps xmm5,[48+esp]
+ movups [32+edi],xmm4
+ xorps xmm6,[64+esp]
+ movups [48+edi],xmm5
+ xorps xmm7,xmm1
+ movups [64+edi],xmm6
+ pshufd xmm2,xmm0,19
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ movdqa xmm3,[96+esp]
+ pxor xmm0,xmm0
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ sub eax,96
+ jnc NEAR L$048xts_enc_loop6
+ mov ecx,DWORD [240+ebp]
+ mov edx,ebp
+ mov ebx,ecx
+L$047xts_enc_short:
+ add eax,96
+ jz NEAR L$049xts_enc_done6x
+ movdqa xmm5,xmm1
+ cmp eax,32
+ jb NEAR L$050xts_enc_one
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ je NEAR L$051xts_enc_two
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa xmm6,xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ cmp eax,64
+ jb NEAR L$052xts_enc_three
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa xmm7,xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ movdqa [esp],xmm5
+ movdqa [16+esp],xmm6
+ je NEAR L$053xts_enc_four
+ movdqa [32+esp],xmm7
+ pshufd xmm7,xmm0,19
+ movdqa [48+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm7,xmm3
+ pxor xmm7,xmm1
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ pxor xmm2,[esp]
+ movdqu xmm5,[48+esi]
+ pxor xmm3,[16+esp]
+ movdqu xmm6,[64+esi]
+ pxor xmm4,[32+esp]
+ lea esi,[80+esi]
+ pxor xmm5,[48+esp]
+ movdqa [64+esp],xmm7
+ pxor xmm6,xmm7
+ call __aesni_encrypt6
+ movaps xmm1,[64+esp]
+ xorps xmm2,[esp]
+ xorps xmm3,[16+esp]
+ xorps xmm4,[32+esp]
+ movups [edi],xmm2
+ xorps xmm5,[48+esp]
+ movups [16+edi],xmm3
+ xorps xmm6,xmm1
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ lea edi,[80+edi]
+ jmp NEAR L$054xts_enc_done
+align 16
+L$050xts_enc_one:
+ movups xmm2,[esi]
+ lea esi,[16+esi]
+ xorps xmm2,xmm5
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$055enc1_loop_9:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$055enc1_loop_9
+db 102,15,56,221,209
+ xorps xmm2,xmm5
+ movups [edi],xmm2
+ lea edi,[16+edi]
+ movdqa xmm1,xmm5
+ jmp NEAR L$054xts_enc_done
+align 16
+L$051xts_enc_two:
+ movaps xmm6,xmm1
+ movups xmm2,[esi]
+ movups xmm3,[16+esi]
+ lea esi,[32+esi]
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ call __aesni_encrypt2
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ lea edi,[32+edi]
+ movdqa xmm1,xmm6
+ jmp NEAR L$054xts_enc_done
+align 16
+L$052xts_enc_three:
+ movaps xmm7,xmm1
+ movups xmm2,[esi]
+ movups xmm3,[16+esi]
+ movups xmm4,[32+esi]
+ lea esi,[48+esi]
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ xorps xmm4,xmm7
+ call __aesni_encrypt3
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ xorps xmm4,xmm7
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ lea edi,[48+edi]
+ movdqa xmm1,xmm7
+ jmp NEAR L$054xts_enc_done
+align 16
+L$053xts_enc_four:
+ movaps xmm6,xmm1
+ movups xmm2,[esi]
+ movups xmm3,[16+esi]
+ movups xmm4,[32+esi]
+ xorps xmm2,[esp]
+ movups xmm5,[48+esi]
+ lea esi,[64+esi]
+ xorps xmm3,[16+esp]
+ xorps xmm4,xmm7
+ xorps xmm5,xmm6
+ call __aesni_encrypt4
+ xorps xmm2,[esp]
+ xorps xmm3,[16+esp]
+ xorps xmm4,xmm7
+ movups [edi],xmm2
+ xorps xmm5,xmm6
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ lea edi,[64+edi]
+ movdqa xmm1,xmm6
+ jmp NEAR L$054xts_enc_done
+align 16
+L$049xts_enc_done6x:
+ mov eax,DWORD [112+esp]
+ and eax,15
+ jz NEAR L$056xts_enc_ret
+ movdqa xmm5,xmm1
+ mov DWORD [112+esp],eax
+ jmp NEAR L$057xts_enc_steal
+align 16
+L$054xts_enc_done:
+ mov eax,DWORD [112+esp]
+ pxor xmm0,xmm0
+ and eax,15
+ jz NEAR L$056xts_enc_ret
+ pcmpgtd xmm0,xmm1
+ mov DWORD [112+esp],eax
+ pshufd xmm5,xmm0,19
+ paddq xmm1,xmm1
+ pand xmm5,[96+esp]
+ pxor xmm5,xmm1
+L$057xts_enc_steal:
+ movzx ecx,BYTE [esi]
+ movzx edx,BYTE [edi-16]
+ lea esi,[1+esi]
+ mov BYTE [edi-16],cl
+ mov BYTE [edi],dl
+ lea edi,[1+edi]
+ sub eax,1
+ jnz NEAR L$057xts_enc_steal
+ sub edi,DWORD [112+esp]
+ mov edx,ebp
+ mov ecx,ebx
+ movups xmm2,[edi-16]
+ xorps xmm2,xmm5
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$058enc1_loop_10:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$058enc1_loop_10
+db 102,15,56,221,209
+ xorps xmm2,xmm5
+ movups [edi-16],xmm2
+L$056xts_enc_ret:
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ movdqa [esp],xmm0
+ pxor xmm3,xmm3
+ movdqa [16+esp],xmm0
+ pxor xmm4,xmm4
+ movdqa [32+esp],xmm0
+ pxor xmm5,xmm5
+ movdqa [48+esp],xmm0
+ pxor xmm6,xmm6
+ movdqa [64+esp],xmm0
+ pxor xmm7,xmm7
+ movdqa [80+esp],xmm0
+ mov esp,DWORD [116+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_xts_decrypt
+align 16
+_aesni_xts_decrypt:
+L$_aesni_xts_decrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov edx,DWORD [36+esp]
+ mov esi,DWORD [40+esp]
+ mov ecx,DWORD [240+edx]
+ movups xmm2,[esi]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$059enc1_loop_11:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$059enc1_loop_11
+db 102,15,56,221,209
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ mov ebp,esp
+ sub esp,120
+ and esp,-16
+ xor ebx,ebx
+ test eax,15
+ setnz bl
+ shl ebx,4
+ sub eax,ebx
+ mov DWORD [96+esp],135
+ mov DWORD [100+esp],0
+ mov DWORD [104+esp],1
+ mov DWORD [108+esp],0
+ mov DWORD [112+esp],eax
+ mov DWORD [116+esp],ebp
+ mov ecx,DWORD [240+edx]
+ mov ebp,edx
+ mov ebx,ecx
+ movdqa xmm1,xmm2
+ pxor xmm0,xmm0
+ movdqa xmm3,[96+esp]
+ pcmpgtd xmm0,xmm1
+ and eax,-16
+ sub eax,96
+ jc NEAR L$060xts_dec_short
+ shl ecx,4
+ mov ebx,16
+ sub ebx,ecx
+ lea edx,[32+ecx*1+edx]
+ jmp NEAR L$061xts_dec_loop6
+align 16
+L$061xts_dec_loop6:
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [16+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [32+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa [48+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ pshufd xmm7,xmm0,19
+ movdqa [64+esp],xmm1
+ paddq xmm1,xmm1
+ movups xmm0,[ebp]
+ pand xmm7,xmm3
+ movups xmm2,[esi]
+ pxor xmm7,xmm1
+ mov ecx,ebx
+ movdqu xmm3,[16+esi]
+ xorps xmm2,xmm0
+ movdqu xmm4,[32+esi]
+ pxor xmm3,xmm0
+ movdqu xmm5,[48+esi]
+ pxor xmm4,xmm0
+ movdqu xmm6,[64+esi]
+ pxor xmm5,xmm0
+ movdqu xmm1,[80+esi]
+ pxor xmm6,xmm0
+ lea esi,[96+esi]
+ pxor xmm2,[esp]
+ movdqa [80+esp],xmm7
+ pxor xmm7,xmm1
+ movups xmm1,[16+ebp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+db 102,15,56,222,209
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+db 102,15,56,222,217
+ pxor xmm7,xmm0
+ movups xmm0,[32+ebp]
+db 102,15,56,222,225
+db 102,15,56,222,233
+db 102,15,56,222,241
+db 102,15,56,222,249
+ call L$_aesni_decrypt6_enter
+ movdqa xmm1,[80+esp]
+ pxor xmm0,xmm0
+ xorps xmm2,[esp]
+ pcmpgtd xmm0,xmm1
+ xorps xmm3,[16+esp]
+ movups [edi],xmm2
+ xorps xmm4,[32+esp]
+ movups [16+edi],xmm3
+ xorps xmm5,[48+esp]
+ movups [32+edi],xmm4
+ xorps xmm6,[64+esp]
+ movups [48+edi],xmm5
+ xorps xmm7,xmm1
+ movups [64+edi],xmm6
+ pshufd xmm2,xmm0,19
+ movups [80+edi],xmm7
+ lea edi,[96+edi]
+ movdqa xmm3,[96+esp]
+ pxor xmm0,xmm0
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ sub eax,96
+ jnc NEAR L$061xts_dec_loop6
+ mov ecx,DWORD [240+ebp]
+ mov edx,ebp
+ mov ebx,ecx
+L$060xts_dec_short:
+ add eax,96
+ jz NEAR L$062xts_dec_done6x
+ movdqa xmm5,xmm1
+ cmp eax,32
+ jb NEAR L$063xts_dec_one
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ je NEAR L$064xts_dec_two
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa xmm6,xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ cmp eax,64
+ jb NEAR L$065xts_dec_three
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa xmm7,xmm1
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+ movdqa [esp],xmm5
+ movdqa [16+esp],xmm6
+ je NEAR L$066xts_dec_four
+ movdqa [32+esp],xmm7
+ pshufd xmm7,xmm0,19
+ movdqa [48+esp],xmm1
+ paddq xmm1,xmm1
+ pand xmm7,xmm3
+ pxor xmm7,xmm1
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ pxor xmm2,[esp]
+ movdqu xmm5,[48+esi]
+ pxor xmm3,[16+esp]
+ movdqu xmm6,[64+esi]
+ pxor xmm4,[32+esp]
+ lea esi,[80+esi]
+ pxor xmm5,[48+esp]
+ movdqa [64+esp],xmm7
+ pxor xmm6,xmm7
+ call __aesni_decrypt6
+ movaps xmm1,[64+esp]
+ xorps xmm2,[esp]
+ xorps xmm3,[16+esp]
+ xorps xmm4,[32+esp]
+ movups [edi],xmm2
+ xorps xmm5,[48+esp]
+ movups [16+edi],xmm3
+ xorps xmm6,xmm1
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ movups [64+edi],xmm6
+ lea edi,[80+edi]
+ jmp NEAR L$067xts_dec_done
+align 16
+L$063xts_dec_one:
+ movups xmm2,[esi]
+ lea esi,[16+esi]
+ xorps xmm2,xmm5
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$068dec1_loop_12:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$068dec1_loop_12
+db 102,15,56,223,209
+ xorps xmm2,xmm5
+ movups [edi],xmm2
+ lea edi,[16+edi]
+ movdqa xmm1,xmm5
+ jmp NEAR L$067xts_dec_done
+align 16
+L$064xts_dec_two:
+ movaps xmm6,xmm1
+ movups xmm2,[esi]
+ movups xmm3,[16+esi]
+ lea esi,[32+esi]
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ call __aesni_decrypt2
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ lea edi,[32+edi]
+ movdqa xmm1,xmm6
+ jmp NEAR L$067xts_dec_done
+align 16
+L$065xts_dec_three:
+ movaps xmm7,xmm1
+ movups xmm2,[esi]
+ movups xmm3,[16+esi]
+ movups xmm4,[32+esi]
+ lea esi,[48+esi]
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ xorps xmm4,xmm7
+ call __aesni_decrypt3
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ xorps xmm4,xmm7
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ lea edi,[48+edi]
+ movdqa xmm1,xmm7
+ jmp NEAR L$067xts_dec_done
+align 16
+L$066xts_dec_four:
+ movaps xmm6,xmm1
+ movups xmm2,[esi]
+ movups xmm3,[16+esi]
+ movups xmm4,[32+esi]
+ xorps xmm2,[esp]
+ movups xmm5,[48+esi]
+ lea esi,[64+esi]
+ xorps xmm3,[16+esp]
+ xorps xmm4,xmm7
+ xorps xmm5,xmm6
+ call __aesni_decrypt4
+ xorps xmm2,[esp]
+ xorps xmm3,[16+esp]
+ xorps xmm4,xmm7
+ movups [edi],xmm2
+ xorps xmm5,xmm6
+ movups [16+edi],xmm3
+ movups [32+edi],xmm4
+ movups [48+edi],xmm5
+ lea edi,[64+edi]
+ movdqa xmm1,xmm6
+ jmp NEAR L$067xts_dec_done
+align 16
+L$062xts_dec_done6x:
+ mov eax,DWORD [112+esp]
+ and eax,15
+ jz NEAR L$069xts_dec_ret
+ mov DWORD [112+esp],eax
+ jmp NEAR L$070xts_dec_only_one_more
+align 16
+L$067xts_dec_done:
+ mov eax,DWORD [112+esp]
+ pxor xmm0,xmm0
+ and eax,15
+ jz NEAR L$069xts_dec_ret
+ pcmpgtd xmm0,xmm1
+ mov DWORD [112+esp],eax
+ pshufd xmm2,xmm0,19
+ pxor xmm0,xmm0
+ movdqa xmm3,[96+esp]
+ paddq xmm1,xmm1
+ pand xmm2,xmm3
+ pcmpgtd xmm0,xmm1
+ pxor xmm1,xmm2
+L$070xts_dec_only_one_more:
+ pshufd xmm5,xmm0,19
+ movdqa xmm6,xmm1
+ paddq xmm1,xmm1
+ pand xmm5,xmm3
+ pxor xmm5,xmm1
+ mov edx,ebp
+ mov ecx,ebx
+ movups xmm2,[esi]
+ xorps xmm2,xmm5
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$071dec1_loop_13:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$071dec1_loop_13
+db 102,15,56,223,209
+ xorps xmm2,xmm5
+ movups [edi],xmm2
+L$072xts_dec_steal:
+ movzx ecx,BYTE [16+esi]
+ movzx edx,BYTE [edi]
+ lea esi,[1+esi]
+ mov BYTE [edi],cl
+ mov BYTE [16+edi],dl
+ lea edi,[1+edi]
+ sub eax,1
+ jnz NEAR L$072xts_dec_steal
+ sub edi,DWORD [112+esp]
+ mov edx,ebp
+ mov ecx,ebx
+ movups xmm2,[edi]
+ xorps xmm2,xmm6
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$073dec1_loop_14:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$073dec1_loop_14
+db 102,15,56,223,209
+ xorps xmm2,xmm6
+ movups [edi],xmm2
+L$069xts_dec_ret:
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ movdqa [esp],xmm0
+ pxor xmm3,xmm3
+ movdqa [16+esp],xmm0
+ pxor xmm4,xmm4
+ movdqa [32+esp],xmm0
+ pxor xmm5,xmm5
+ movdqa [48+esp],xmm0
+ pxor xmm6,xmm6
+ movdqa [64+esp],xmm0
+ pxor xmm7,xmm7
+ movdqa [80+esp],xmm0
+ mov esp,DWORD [116+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_ocb_encrypt
+align 16
+_aesni_ocb_encrypt:
+L$_aesni_ocb_encrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov ecx,DWORD [40+esp]
+ mov ebx,DWORD [48+esp]
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ movdqu xmm0,[ecx]
+ mov ebp,DWORD [36+esp]
+ movdqu xmm1,[ebx]
+ mov ebx,DWORD [44+esp]
+ mov ecx,esp
+ sub esp,132
+ and esp,-16
+ sub edi,esi
+ shl eax,4
+ lea eax,[eax*1+esi-96]
+ mov DWORD [120+esp],edi
+ mov DWORD [124+esp],eax
+ mov DWORD [128+esp],ecx
+ mov ecx,DWORD [240+edx]
+ test ebp,1
+ jnz NEAR L$074odd
+ bsf eax,ebp
+ add ebp,1
+ shl eax,4
+ movdqu xmm7,[eax*1+ebx]
+ mov eax,edx
+ movdqu xmm2,[esi]
+ lea esi,[16+esi]
+ pxor xmm7,xmm0
+ pxor xmm1,xmm2
+ pxor xmm2,xmm7
+ movdqa xmm6,xmm1
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$075enc1_loop_15:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$075enc1_loop_15
+db 102,15,56,221,209
+ xorps xmm2,xmm7
+ movdqa xmm0,xmm7
+ movdqa xmm1,xmm6
+ movups [esi*1+edi-16],xmm2
+ mov ecx,DWORD [240+eax]
+ mov edx,eax
+ mov eax,DWORD [124+esp]
+L$074odd:
+ shl ecx,4
+ mov edi,16
+ sub edi,ecx
+ mov DWORD [112+esp],edx
+ lea edx,[32+ecx*1+edx]
+ mov DWORD [116+esp],edi
+ cmp esi,eax
+ ja NEAR L$076short
+ jmp NEAR L$077grandloop
+align 32
+L$077grandloop:
+ lea ecx,[1+ebp]
+ lea eax,[3+ebp]
+ lea edi,[5+ebp]
+ add ebp,6
+ bsf ecx,ecx
+ bsf eax,eax
+ bsf edi,edi
+ shl ecx,4
+ shl eax,4
+ shl edi,4
+ movdqu xmm2,[ebx]
+ movdqu xmm3,[ecx*1+ebx]
+ mov ecx,DWORD [116+esp]
+ movdqa xmm4,xmm2
+ movdqu xmm5,[eax*1+ebx]
+ movdqa xmm6,xmm2
+ movdqu xmm7,[edi*1+ebx]
+ pxor xmm2,xmm0
+ pxor xmm3,xmm2
+ movdqa [esp],xmm2
+ pxor xmm4,xmm3
+ movdqa [16+esp],xmm3
+ pxor xmm5,xmm4
+ movdqa [32+esp],xmm4
+ pxor xmm6,xmm5
+ movdqa [48+esp],xmm5
+ pxor xmm7,xmm6
+ movdqa [64+esp],xmm6
+ movdqa [80+esp],xmm7
+ movups xmm0,[ecx*1+edx-48]
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ movdqu xmm7,[80+esi]
+ lea esi,[96+esi]
+ pxor xmm1,xmm2
+ pxor xmm2,xmm0
+ pxor xmm1,xmm3
+ pxor xmm3,xmm0
+ pxor xmm1,xmm4
+ pxor xmm4,xmm0
+ pxor xmm1,xmm5
+ pxor xmm5,xmm0
+ pxor xmm1,xmm6
+ pxor xmm6,xmm0
+ pxor xmm1,xmm7
+ pxor xmm7,xmm0
+ movdqa [96+esp],xmm1
+ movups xmm1,[ecx*1+edx-32]
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+ pxor xmm7,[80+esp]
+ movups xmm0,[ecx*1+edx-16]
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+db 102,15,56,220,233
+db 102,15,56,220,241
+db 102,15,56,220,249
+ mov edi,DWORD [120+esp]
+ mov eax,DWORD [124+esp]
+ call L$_aesni_encrypt6_enter
+ movdqa xmm0,[80+esp]
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+ pxor xmm7,xmm0
+ movdqa xmm1,[96+esp]
+ movdqu [esi*1+edi-96],xmm2
+ movdqu [esi*1+edi-80],xmm3
+ movdqu [esi*1+edi-64],xmm4
+ movdqu [esi*1+edi-48],xmm5
+ movdqu [esi*1+edi-32],xmm6
+ movdqu [esi*1+edi-16],xmm7
+ cmp esi,eax
+ jb NEAR L$077grandloop
+L$076short:
+ add eax,96
+ sub eax,esi
+ jz NEAR L$078done
+ cmp eax,32
+ jb NEAR L$079one
+ je NEAR L$080two
+ cmp eax,64
+ jb NEAR L$081three
+ je NEAR L$082four
+ lea ecx,[1+ebp]
+ lea eax,[3+ebp]
+ bsf ecx,ecx
+ bsf eax,eax
+ shl ecx,4
+ shl eax,4
+ movdqu xmm2,[ebx]
+ movdqu xmm3,[ecx*1+ebx]
+ mov ecx,DWORD [116+esp]
+ movdqa xmm4,xmm2
+ movdqu xmm5,[eax*1+ebx]
+ movdqa xmm6,xmm2
+ pxor xmm2,xmm0
+ pxor xmm3,xmm2
+ movdqa [esp],xmm2
+ pxor xmm4,xmm3
+ movdqa [16+esp],xmm3
+ pxor xmm5,xmm4
+ movdqa [32+esp],xmm4
+ pxor xmm6,xmm5
+ movdqa [48+esp],xmm5
+ pxor xmm7,xmm6
+ movdqa [64+esp],xmm6
+ movups xmm0,[ecx*1+edx-48]
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ pxor xmm7,xmm7
+ pxor xmm1,xmm2
+ pxor xmm2,xmm0
+ pxor xmm1,xmm3
+ pxor xmm3,xmm0
+ pxor xmm1,xmm4
+ pxor xmm4,xmm0
+ pxor xmm1,xmm5
+ pxor xmm5,xmm0
+ pxor xmm1,xmm6
+ pxor xmm6,xmm0
+ movdqa [96+esp],xmm1
+ movups xmm1,[ecx*1+edx-32]
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+ movups xmm0,[ecx*1+edx-16]
+db 102,15,56,220,209
+db 102,15,56,220,217
+db 102,15,56,220,225
+db 102,15,56,220,233
+db 102,15,56,220,241
+db 102,15,56,220,249
+ mov edi,DWORD [120+esp]
+ call L$_aesni_encrypt6_enter
+ movdqa xmm0,[64+esp]
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,xmm0
+ movdqa xmm1,[96+esp]
+ movdqu [esi*1+edi],xmm2
+ movdqu [16+esi*1+edi],xmm3
+ movdqu [32+esi*1+edi],xmm4
+ movdqu [48+esi*1+edi],xmm5
+ movdqu [64+esi*1+edi],xmm6
+ jmp NEAR L$078done
+align 16
+L$079one:
+ movdqu xmm7,[ebx]
+ mov edx,DWORD [112+esp]
+ movdqu xmm2,[esi]
+ mov ecx,DWORD [240+edx]
+ pxor xmm7,xmm0
+ pxor xmm1,xmm2
+ pxor xmm2,xmm7
+ movdqa xmm6,xmm1
+ mov edi,DWORD [120+esp]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$083enc1_loop_16:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$083enc1_loop_16
+db 102,15,56,221,209
+ xorps xmm2,xmm7
+ movdqa xmm0,xmm7
+ movdqa xmm1,xmm6
+ movups [esi*1+edi],xmm2
+ jmp NEAR L$078done
+align 16
+L$080two:
+ lea ecx,[1+ebp]
+ mov edx,DWORD [112+esp]
+ bsf ecx,ecx
+ shl ecx,4
+ movdqu xmm6,[ebx]
+ movdqu xmm7,[ecx*1+ebx]
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ mov ecx,DWORD [240+edx]
+ pxor xmm6,xmm0
+ pxor xmm7,xmm6
+ pxor xmm1,xmm2
+ pxor xmm2,xmm6
+ pxor xmm1,xmm3
+ pxor xmm3,xmm7
+ movdqa xmm5,xmm1
+ mov edi,DWORD [120+esp]
+ call __aesni_encrypt2
+ xorps xmm2,xmm6
+ xorps xmm3,xmm7
+ movdqa xmm0,xmm7
+ movdqa xmm1,xmm5
+ movups [esi*1+edi],xmm2
+ movups [16+esi*1+edi],xmm3
+ jmp NEAR L$078done
+align 16
+L$081three:
+ lea ecx,[1+ebp]
+ mov edx,DWORD [112+esp]
+ bsf ecx,ecx
+ shl ecx,4
+ movdqu xmm5,[ebx]
+ movdqu xmm6,[ecx*1+ebx]
+ movdqa xmm7,xmm5
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ mov ecx,DWORD [240+edx]
+ pxor xmm5,xmm0
+ pxor xmm6,xmm5
+ pxor xmm7,xmm6
+ pxor xmm1,xmm2
+ pxor xmm2,xmm5
+ pxor xmm1,xmm3
+ pxor xmm3,xmm6
+ pxor xmm1,xmm4
+ pxor xmm4,xmm7
+ movdqa [96+esp],xmm1
+ mov edi,DWORD [120+esp]
+ call __aesni_encrypt3
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ xorps xmm4,xmm7
+ movdqa xmm0,xmm7
+ movdqa xmm1,[96+esp]
+ movups [esi*1+edi],xmm2
+ movups [16+esi*1+edi],xmm3
+ movups [32+esi*1+edi],xmm4
+ jmp NEAR L$078done
+align 16
+L$082four:
+ lea ecx,[1+ebp]
+ lea eax,[3+ebp]
+ bsf ecx,ecx
+ bsf eax,eax
+ mov edx,DWORD [112+esp]
+ shl ecx,4
+ shl eax,4
+ movdqu xmm4,[ebx]
+ movdqu xmm5,[ecx*1+ebx]
+ movdqa xmm6,xmm4
+ movdqu xmm7,[eax*1+ebx]
+ pxor xmm4,xmm0
+ movdqu xmm2,[esi]
+ pxor xmm5,xmm4
+ movdqu xmm3,[16+esi]
+ pxor xmm6,xmm5
+ movdqa [esp],xmm4
+ pxor xmm7,xmm6
+ movdqa [16+esp],xmm5
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ mov ecx,DWORD [240+edx]
+ pxor xmm1,xmm2
+ pxor xmm2,[esp]
+ pxor xmm1,xmm3
+ pxor xmm3,[16+esp]
+ pxor xmm1,xmm4
+ pxor xmm4,xmm6
+ pxor xmm1,xmm5
+ pxor xmm5,xmm7
+ movdqa [96+esp],xmm1
+ mov edi,DWORD [120+esp]
+ call __aesni_encrypt4
+ xorps xmm2,[esp]
+ xorps xmm3,[16+esp]
+ xorps xmm4,xmm6
+ movups [esi*1+edi],xmm2
+ xorps xmm5,xmm7
+ movups [16+esi*1+edi],xmm3
+ movdqa xmm0,xmm7
+ movups [32+esi*1+edi],xmm4
+ movdqa xmm1,[96+esp]
+ movups [48+esi*1+edi],xmm5
+L$078done:
+ mov edx,DWORD [128+esp]
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ movdqa [esp],xmm2
+ pxor xmm4,xmm4
+ movdqa [16+esp],xmm2
+ pxor xmm5,xmm5
+ movdqa [32+esp],xmm2
+ pxor xmm6,xmm6
+ movdqa [48+esp],xmm2
+ pxor xmm7,xmm7
+ movdqa [64+esp],xmm2
+ movdqa [80+esp],xmm2
+ movdqa [96+esp],xmm2
+ lea esp,[edx]
+ mov ecx,DWORD [40+esp]
+ mov ebx,DWORD [48+esp]
+ movdqu [ecx],xmm0
+ pxor xmm0,xmm0
+ movdqu [ebx],xmm1
+ pxor xmm1,xmm1
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_ocb_decrypt
+align 16
+_aesni_ocb_decrypt:
+L$_aesni_ocb_decrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov ecx,DWORD [40+esp]
+ mov ebx,DWORD [48+esp]
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ movdqu xmm0,[ecx]
+ mov ebp,DWORD [36+esp]
+ movdqu xmm1,[ebx]
+ mov ebx,DWORD [44+esp]
+ mov ecx,esp
+ sub esp,132
+ and esp,-16
+ sub edi,esi
+ shl eax,4
+ lea eax,[eax*1+esi-96]
+ mov DWORD [120+esp],edi
+ mov DWORD [124+esp],eax
+ mov DWORD [128+esp],ecx
+ mov ecx,DWORD [240+edx]
+ test ebp,1
+ jnz NEAR L$084odd
+ bsf eax,ebp
+ add ebp,1
+ shl eax,4
+ movdqu xmm7,[eax*1+ebx]
+ mov eax,edx
+ movdqu xmm2,[esi]
+ lea esi,[16+esi]
+ pxor xmm7,xmm0
+ pxor xmm2,xmm7
+ movdqa xmm6,xmm1
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$085dec1_loop_17:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$085dec1_loop_17
+db 102,15,56,223,209
+ xorps xmm2,xmm7
+ movaps xmm1,xmm6
+ movdqa xmm0,xmm7
+ xorps xmm1,xmm2
+ movups [esi*1+edi-16],xmm2
+ mov ecx,DWORD [240+eax]
+ mov edx,eax
+ mov eax,DWORD [124+esp]
+L$084odd:
+ shl ecx,4
+ mov edi,16
+ sub edi,ecx
+ mov DWORD [112+esp],edx
+ lea edx,[32+ecx*1+edx]
+ mov DWORD [116+esp],edi
+ cmp esi,eax
+ ja NEAR L$086short
+ jmp NEAR L$087grandloop
+align 32
+L$087grandloop:
+ lea ecx,[1+ebp]
+ lea eax,[3+ebp]
+ lea edi,[5+ebp]
+ add ebp,6
+ bsf ecx,ecx
+ bsf eax,eax
+ bsf edi,edi
+ shl ecx,4
+ shl eax,4
+ shl edi,4
+ movdqu xmm2,[ebx]
+ movdqu xmm3,[ecx*1+ebx]
+ mov ecx,DWORD [116+esp]
+ movdqa xmm4,xmm2
+ movdqu xmm5,[eax*1+ebx]
+ movdqa xmm6,xmm2
+ movdqu xmm7,[edi*1+ebx]
+ pxor xmm2,xmm0
+ pxor xmm3,xmm2
+ movdqa [esp],xmm2
+ pxor xmm4,xmm3
+ movdqa [16+esp],xmm3
+ pxor xmm5,xmm4
+ movdqa [32+esp],xmm4
+ pxor xmm6,xmm5
+ movdqa [48+esp],xmm5
+ pxor xmm7,xmm6
+ movdqa [64+esp],xmm6
+ movdqa [80+esp],xmm7
+ movups xmm0,[ecx*1+edx-48]
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ movdqu xmm7,[80+esi]
+ lea esi,[96+esi]
+ movdqa [96+esp],xmm1
+ pxor xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+ pxor xmm5,xmm0
+ pxor xmm6,xmm0
+ pxor xmm7,xmm0
+ movups xmm1,[ecx*1+edx-32]
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+ pxor xmm7,[80+esp]
+ movups xmm0,[ecx*1+edx-16]
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+db 102,15,56,222,233
+db 102,15,56,222,241
+db 102,15,56,222,249
+ mov edi,DWORD [120+esp]
+ mov eax,DWORD [124+esp]
+ call L$_aesni_decrypt6_enter
+ movdqa xmm0,[80+esp]
+ pxor xmm2,[esp]
+ movdqa xmm1,[96+esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+ pxor xmm7,xmm0
+ pxor xmm1,xmm2
+ movdqu [esi*1+edi-96],xmm2
+ pxor xmm1,xmm3
+ movdqu [esi*1+edi-80],xmm3
+ pxor xmm1,xmm4
+ movdqu [esi*1+edi-64],xmm4
+ pxor xmm1,xmm5
+ movdqu [esi*1+edi-48],xmm5
+ pxor xmm1,xmm6
+ movdqu [esi*1+edi-32],xmm6
+ pxor xmm1,xmm7
+ movdqu [esi*1+edi-16],xmm7
+ cmp esi,eax
+ jb NEAR L$087grandloop
+L$086short:
+ add eax,96
+ sub eax,esi
+ jz NEAR L$088done
+ cmp eax,32
+ jb NEAR L$089one
+ je NEAR L$090two
+ cmp eax,64
+ jb NEAR L$091three
+ je NEAR L$092four
+ lea ecx,[1+ebp]
+ lea eax,[3+ebp]
+ bsf ecx,ecx
+ bsf eax,eax
+ shl ecx,4
+ shl eax,4
+ movdqu xmm2,[ebx]
+ movdqu xmm3,[ecx*1+ebx]
+ mov ecx,DWORD [116+esp]
+ movdqa xmm4,xmm2
+ movdqu xmm5,[eax*1+ebx]
+ movdqa xmm6,xmm2
+ pxor xmm2,xmm0
+ pxor xmm3,xmm2
+ movdqa [esp],xmm2
+ pxor xmm4,xmm3
+ movdqa [16+esp],xmm3
+ pxor xmm5,xmm4
+ movdqa [32+esp],xmm4
+ pxor xmm6,xmm5
+ movdqa [48+esp],xmm5
+ pxor xmm7,xmm6
+ movdqa [64+esp],xmm6
+ movups xmm0,[ecx*1+edx-48]
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ pxor xmm7,xmm7
+ movdqa [96+esp],xmm1
+ pxor xmm2,xmm0
+ pxor xmm3,xmm0
+ pxor xmm4,xmm0
+ pxor xmm5,xmm0
+ pxor xmm6,xmm0
+ movups xmm1,[ecx*1+edx-32]
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,[64+esp]
+ movups xmm0,[ecx*1+edx-16]
+db 102,15,56,222,209
+db 102,15,56,222,217
+db 102,15,56,222,225
+db 102,15,56,222,233
+db 102,15,56,222,241
+db 102,15,56,222,249
+ mov edi,DWORD [120+esp]
+ call L$_aesni_decrypt6_enter
+ movdqa xmm0,[64+esp]
+ pxor xmm2,[esp]
+ movdqa xmm1,[96+esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,[32+esp]
+ pxor xmm5,[48+esp]
+ pxor xmm6,xmm0
+ pxor xmm1,xmm2
+ movdqu [esi*1+edi],xmm2
+ pxor xmm1,xmm3
+ movdqu [16+esi*1+edi],xmm3
+ pxor xmm1,xmm4
+ movdqu [32+esi*1+edi],xmm4
+ pxor xmm1,xmm5
+ movdqu [48+esi*1+edi],xmm5
+ pxor xmm1,xmm6
+ movdqu [64+esi*1+edi],xmm6
+ jmp NEAR L$088done
+align 16
+L$089one:
+ movdqu xmm7,[ebx]
+ mov edx,DWORD [112+esp]
+ movdqu xmm2,[esi]
+ mov ecx,DWORD [240+edx]
+ pxor xmm7,xmm0
+ pxor xmm2,xmm7
+ movdqa xmm6,xmm1
+ mov edi,DWORD [120+esp]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$093dec1_loop_18:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$093dec1_loop_18
+db 102,15,56,223,209
+ xorps xmm2,xmm7
+ movaps xmm1,xmm6
+ movdqa xmm0,xmm7
+ xorps xmm1,xmm2
+ movups [esi*1+edi],xmm2
+ jmp NEAR L$088done
+align 16
+L$090two:
+ lea ecx,[1+ebp]
+ mov edx,DWORD [112+esp]
+ bsf ecx,ecx
+ shl ecx,4
+ movdqu xmm6,[ebx]
+ movdqu xmm7,[ecx*1+ebx]
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ mov ecx,DWORD [240+edx]
+ movdqa xmm5,xmm1
+ pxor xmm6,xmm0
+ pxor xmm7,xmm6
+ pxor xmm2,xmm6
+ pxor xmm3,xmm7
+ mov edi,DWORD [120+esp]
+ call __aesni_decrypt2
+ xorps xmm2,xmm6
+ xorps xmm3,xmm7
+ movdqa xmm0,xmm7
+ xorps xmm5,xmm2
+ movups [esi*1+edi],xmm2
+ xorps xmm5,xmm3
+ movups [16+esi*1+edi],xmm3
+ movaps xmm1,xmm5
+ jmp NEAR L$088done
+align 16
+L$091three:
+ lea ecx,[1+ebp]
+ mov edx,DWORD [112+esp]
+ bsf ecx,ecx
+ shl ecx,4
+ movdqu xmm5,[ebx]
+ movdqu xmm6,[ecx*1+ebx]
+ movdqa xmm7,xmm5
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ mov ecx,DWORD [240+edx]
+ movdqa [96+esp],xmm1
+ pxor xmm5,xmm0
+ pxor xmm6,xmm5
+ pxor xmm7,xmm6
+ pxor xmm2,xmm5
+ pxor xmm3,xmm6
+ pxor xmm4,xmm7
+ mov edi,DWORD [120+esp]
+ call __aesni_decrypt3
+ movdqa xmm1,[96+esp]
+ xorps xmm2,xmm5
+ xorps xmm3,xmm6
+ xorps xmm4,xmm7
+ movups [esi*1+edi],xmm2
+ pxor xmm1,xmm2
+ movdqa xmm0,xmm7
+ movups [16+esi*1+edi],xmm3
+ pxor xmm1,xmm3
+ movups [32+esi*1+edi],xmm4
+ pxor xmm1,xmm4
+ jmp NEAR L$088done
+align 16
+L$092four:
+ lea ecx,[1+ebp]
+ lea eax,[3+ebp]
+ bsf ecx,ecx
+ bsf eax,eax
+ mov edx,DWORD [112+esp]
+ shl ecx,4
+ shl eax,4
+ movdqu xmm4,[ebx]
+ movdqu xmm5,[ecx*1+ebx]
+ movdqa xmm6,xmm4
+ movdqu xmm7,[eax*1+ebx]
+ pxor xmm4,xmm0
+ movdqu xmm2,[esi]
+ pxor xmm5,xmm4
+ movdqu xmm3,[16+esi]
+ pxor xmm6,xmm5
+ movdqa [esp],xmm4
+ pxor xmm7,xmm6
+ movdqa [16+esp],xmm5
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ mov ecx,DWORD [240+edx]
+ movdqa [96+esp],xmm1
+ pxor xmm2,[esp]
+ pxor xmm3,[16+esp]
+ pxor xmm4,xmm6
+ pxor xmm5,xmm7
+ mov edi,DWORD [120+esp]
+ call __aesni_decrypt4
+ movdqa xmm1,[96+esp]
+ xorps xmm2,[esp]
+ xorps xmm3,[16+esp]
+ xorps xmm4,xmm6
+ movups [esi*1+edi],xmm2
+ pxor xmm1,xmm2
+ xorps xmm5,xmm7
+ movups [16+esi*1+edi],xmm3
+ pxor xmm1,xmm3
+ movdqa xmm0,xmm7
+ movups [32+esi*1+edi],xmm4
+ pxor xmm1,xmm4
+ movups [48+esi*1+edi],xmm5
+ pxor xmm1,xmm5
+L$088done:
+ mov edx,DWORD [128+esp]
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ movdqa [esp],xmm2
+ pxor xmm4,xmm4
+ movdqa [16+esp],xmm2
+ pxor xmm5,xmm5
+ movdqa [32+esp],xmm2
+ pxor xmm6,xmm6
+ movdqa [48+esp],xmm2
+ pxor xmm7,xmm7
+ movdqa [64+esp],xmm2
+ movdqa [80+esp],xmm2
+ movdqa [96+esp],xmm2
+ lea esp,[edx]
+ mov ecx,DWORD [40+esp]
+ mov ebx,DWORD [48+esp]
+ movdqu [ecx],xmm0
+ pxor xmm0,xmm0
+ movdqu [ebx],xmm1
+ pxor xmm1,xmm1
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _aesni_cbc_encrypt
+align 16
+_aesni_cbc_encrypt:
+L$_aesni_cbc_encrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov ebx,esp
+ mov edi,DWORD [24+esp]
+ sub ebx,24
+ mov eax,DWORD [28+esp]
+ and ebx,-16
+ mov edx,DWORD [32+esp]
+ mov ebp,DWORD [36+esp]
+ test eax,eax
+ jz NEAR L$094cbc_abort
+ cmp DWORD [40+esp],0
+ xchg ebx,esp
+ movups xmm7,[ebp]
+ mov ecx,DWORD [240+edx]
+ mov ebp,edx
+ mov DWORD [16+esp],ebx
+ mov ebx,ecx
+ je NEAR L$095cbc_decrypt
+ movaps xmm2,xmm7
+ cmp eax,16
+ jb NEAR L$096cbc_enc_tail
+ sub eax,16
+ jmp NEAR L$097cbc_enc_loop
+align 16
+L$097cbc_enc_loop:
+ movups xmm7,[esi]
+ lea esi,[16+esi]
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ xorps xmm7,xmm0
+ lea edx,[32+edx]
+ xorps xmm2,xmm7
+L$098enc1_loop_19:
+db 102,15,56,220,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$098enc1_loop_19
+db 102,15,56,221,209
+ mov ecx,ebx
+ mov edx,ebp
+ movups [edi],xmm2
+ lea edi,[16+edi]
+ sub eax,16
+ jnc NEAR L$097cbc_enc_loop
+ add eax,16
+ jnz NEAR L$096cbc_enc_tail
+ movaps xmm7,xmm2
+ pxor xmm2,xmm2
+ jmp NEAR L$099cbc_ret
+L$096cbc_enc_tail:
+ mov ecx,eax
+dd 2767451785
+ mov ecx,16
+ sub ecx,eax
+ xor eax,eax
+dd 2868115081
+ lea edi,[edi-16]
+ mov ecx,ebx
+ mov esi,edi
+ mov edx,ebp
+ jmp NEAR L$097cbc_enc_loop
+align 16
+L$095cbc_decrypt:
+ cmp eax,80
+ jbe NEAR L$100cbc_dec_tail
+ movaps [esp],xmm7
+ sub eax,80
+ jmp NEAR L$101cbc_dec_loop6_enter
+align 16
+L$102cbc_dec_loop6:
+ movaps [esp],xmm0
+ movups [edi],xmm7
+ lea edi,[16+edi]
+L$101cbc_dec_loop6_enter:
+ movdqu xmm2,[esi]
+ movdqu xmm3,[16+esi]
+ movdqu xmm4,[32+esi]
+ movdqu xmm5,[48+esi]
+ movdqu xmm6,[64+esi]
+ movdqu xmm7,[80+esi]
+ call __aesni_decrypt6
+ movups xmm1,[esi]
+ movups xmm0,[16+esi]
+ xorps xmm2,[esp]
+ xorps xmm3,xmm1
+ movups xmm1,[32+esi]
+ xorps xmm4,xmm0
+ movups xmm0,[48+esi]
+ xorps xmm5,xmm1
+ movups xmm1,[64+esi]
+ xorps xmm6,xmm0
+ movups xmm0,[80+esi]
+ xorps xmm7,xmm1
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ lea esi,[96+esi]
+ movups [32+edi],xmm4
+ mov ecx,ebx
+ movups [48+edi],xmm5
+ mov edx,ebp
+ movups [64+edi],xmm6
+ lea edi,[80+edi]
+ sub eax,96
+ ja NEAR L$102cbc_dec_loop6
+ movaps xmm2,xmm7
+ movaps xmm7,xmm0
+ add eax,80
+ jle NEAR L$103cbc_dec_clear_tail_collected
+ movups [edi],xmm2
+ lea edi,[16+edi]
+L$100cbc_dec_tail:
+ movups xmm2,[esi]
+ movaps xmm6,xmm2
+ cmp eax,16
+ jbe NEAR L$104cbc_dec_one
+ movups xmm3,[16+esi]
+ movaps xmm5,xmm3
+ cmp eax,32
+ jbe NEAR L$105cbc_dec_two
+ movups xmm4,[32+esi]
+ cmp eax,48
+ jbe NEAR L$106cbc_dec_three
+ movups xmm5,[48+esi]
+ cmp eax,64
+ jbe NEAR L$107cbc_dec_four
+ movups xmm6,[64+esi]
+ movaps [esp],xmm7
+ movups xmm2,[esi]
+ xorps xmm7,xmm7
+ call __aesni_decrypt6
+ movups xmm1,[esi]
+ movups xmm0,[16+esi]
+ xorps xmm2,[esp]
+ xorps xmm3,xmm1
+ movups xmm1,[32+esi]
+ xorps xmm4,xmm0
+ movups xmm0,[48+esi]
+ xorps xmm5,xmm1
+ movups xmm7,[64+esi]
+ xorps xmm6,xmm0
+ movups [edi],xmm2
+ movups [16+edi],xmm3
+ pxor xmm3,xmm3
+ movups [32+edi],xmm4
+ pxor xmm4,xmm4
+ movups [48+edi],xmm5
+ pxor xmm5,xmm5
+ lea edi,[64+edi]
+ movaps xmm2,xmm6
+ pxor xmm6,xmm6
+ sub eax,80
+ jmp NEAR L$108cbc_dec_tail_collected
+align 16
+L$104cbc_dec_one:
+ movups xmm0,[edx]
+ movups xmm1,[16+edx]
+ lea edx,[32+edx]
+ xorps xmm2,xmm0
+L$109dec1_loop_20:
+db 102,15,56,222,209
+ dec ecx
+ movups xmm1,[edx]
+ lea edx,[16+edx]
+ jnz NEAR L$109dec1_loop_20
+db 102,15,56,223,209
+ xorps xmm2,xmm7
+ movaps xmm7,xmm6
+ sub eax,16
+ jmp NEAR L$108cbc_dec_tail_collected
+align 16
+L$105cbc_dec_two:
+ call __aesni_decrypt2
+ xorps xmm2,xmm7
+ xorps xmm3,xmm6
+ movups [edi],xmm2
+ movaps xmm2,xmm3
+ pxor xmm3,xmm3
+ lea edi,[16+edi]
+ movaps xmm7,xmm5
+ sub eax,32
+ jmp NEAR L$108cbc_dec_tail_collected
+align 16
+L$106cbc_dec_three:
+ call __aesni_decrypt3
+ xorps xmm2,xmm7
+ xorps xmm3,xmm6
+ xorps xmm4,xmm5
+ movups [edi],xmm2
+ movaps xmm2,xmm4
+ pxor xmm4,xmm4
+ movups [16+edi],xmm3
+ pxor xmm3,xmm3
+ lea edi,[32+edi]
+ movups xmm7,[32+esi]
+ sub eax,48
+ jmp NEAR L$108cbc_dec_tail_collected
+align 16
+L$107cbc_dec_four:
+ call __aesni_decrypt4
+ movups xmm1,[16+esi]
+ movups xmm0,[32+esi]
+ xorps xmm2,xmm7
+ movups xmm7,[48+esi]
+ xorps xmm3,xmm6
+ movups [edi],xmm2
+ xorps xmm4,xmm1
+ movups [16+edi],xmm3
+ pxor xmm3,xmm3
+ xorps xmm5,xmm0
+ movups [32+edi],xmm4
+ pxor xmm4,xmm4
+ lea edi,[48+edi]
+ movaps xmm2,xmm5
+ pxor xmm5,xmm5
+ sub eax,64
+ jmp NEAR L$108cbc_dec_tail_collected
+align 16
+L$103cbc_dec_clear_tail_collected:
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ pxor xmm5,xmm5
+ pxor xmm6,xmm6
+L$108cbc_dec_tail_collected:
+ and eax,15
+ jnz NEAR L$110cbc_dec_tail_partial
+ movups [edi],xmm2
+ pxor xmm0,xmm0
+ jmp NEAR L$099cbc_ret
+align 16
+L$110cbc_dec_tail_partial:
+ movaps [esp],xmm2
+ pxor xmm0,xmm0
+ mov ecx,16
+ mov esi,esp
+ sub ecx,eax
+dd 2767451785
+ movdqa [esp],xmm2
+L$099cbc_ret:
+ mov esp,DWORD [16+esp]
+ mov ebp,DWORD [36+esp]
+ pxor xmm2,xmm2
+ pxor xmm1,xmm1
+ movups [ebp],xmm7
+ pxor xmm7,xmm7
+L$094cbc_abort:
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 16
+__aesni_set_encrypt_key:
+ push ebp
+ push ebx
+ test eax,eax
+ jz NEAR L$111bad_pointer
+ test edx,edx
+ jz NEAR L$111bad_pointer
+ call L$112pic
+L$112pic:
+ pop ebx
+ lea ebx,[(L$key_const-L$112pic)+ebx]
+ lea ebp,[_OPENSSL_ia32cap_P]
+ movups xmm0,[eax]
+ xorps xmm4,xmm4
+ mov ebp,DWORD [4+ebp]
+ lea edx,[16+edx]
+ and ebp,268437504
+ cmp ecx,256
+ je NEAR L$11314rounds
+ cmp ecx,192
+ je NEAR L$11412rounds
+ cmp ecx,128
+ jne NEAR L$115bad_keybits
+align 16
+L$11610rounds:
+ cmp ebp,268435456
+ je NEAR L$11710rounds_alt
+ mov ecx,9
+ movups [edx-16],xmm0
+db 102,15,58,223,200,1
+ call L$118key_128_cold
+db 102,15,58,223,200,2
+ call L$119key_128
+db 102,15,58,223,200,4
+ call L$119key_128
+db 102,15,58,223,200,8
+ call L$119key_128
+db 102,15,58,223,200,16
+ call L$119key_128
+db 102,15,58,223,200,32
+ call L$119key_128
+db 102,15,58,223,200,64
+ call L$119key_128
+db 102,15,58,223,200,128
+ call L$119key_128
+db 102,15,58,223,200,27
+ call L$119key_128
+db 102,15,58,223,200,54
+ call L$119key_128
+ movups [edx],xmm0
+ mov DWORD [80+edx],ecx
+ jmp NEAR L$120good_key
+align 16
+L$119key_128:
+ movups [edx],xmm0
+ lea edx,[16+edx]
+L$118key_128_cold:
+ shufps xmm4,xmm0,16
+ xorps xmm0,xmm4
+ shufps xmm4,xmm0,140
+ xorps xmm0,xmm4
+ shufps xmm1,xmm1,255
+ xorps xmm0,xmm1
+ ret
+align 16
+L$11710rounds_alt:
+ movdqa xmm5,[ebx]
+ mov ecx,8
+ movdqa xmm4,[32+ebx]
+ movdqa xmm2,xmm0
+ movdqu [edx-16],xmm0
+L$121loop_key128:
+db 102,15,56,0,197
+db 102,15,56,221,196
+ pslld xmm4,1
+ lea edx,[16+edx]
+ movdqa xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm2,xmm3
+ pxor xmm0,xmm2
+ movdqu [edx-16],xmm0
+ movdqa xmm2,xmm0
+ dec ecx
+ jnz NEAR L$121loop_key128
+ movdqa xmm4,[48+ebx]
+db 102,15,56,0,197
+db 102,15,56,221,196
+ pslld xmm4,1
+ movdqa xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm2,xmm3
+ pxor xmm0,xmm2
+ movdqu [edx],xmm0
+ movdqa xmm2,xmm0
+db 102,15,56,0,197
+db 102,15,56,221,196
+ movdqa xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm3,xmm2
+ pslldq xmm2,4
+ pxor xmm2,xmm3
+ pxor xmm0,xmm2
+ movdqu [16+edx],xmm0
+ mov ecx,9
+ mov DWORD [96+edx],ecx
+ jmp NEAR L$120good_key
+align 16
+L$11412rounds:
+ movq xmm2,[16+eax]
+ cmp ebp,268435456
+ je NEAR L$12212rounds_alt
+ mov ecx,11
+ movups [edx-16],xmm0
+db 102,15,58,223,202,1
+ call L$123key_192a_cold
+db 102,15,58,223,202,2
+ call L$124key_192b
+db 102,15,58,223,202,4
+ call L$125key_192a
+db 102,15,58,223,202,8
+ call L$124key_192b
+db 102,15,58,223,202,16
+ call L$125key_192a
+db 102,15,58,223,202,32
+ call L$124key_192b
+db 102,15,58,223,202,64
+ call L$125key_192a
+db 102,15,58,223,202,128
+ call L$124key_192b
+ movups [edx],xmm0
+ mov DWORD [48+edx],ecx
+ jmp NEAR L$120good_key
+align 16
+L$125key_192a:
+ movups [edx],xmm0
+ lea edx,[16+edx]
+align 16
+L$123key_192a_cold:
+ movaps xmm5,xmm2
+L$126key_192b_warm:
+ shufps xmm4,xmm0,16
+ movdqa xmm3,xmm2
+ xorps xmm0,xmm4
+ shufps xmm4,xmm0,140
+ pslldq xmm3,4
+ xorps xmm0,xmm4
+ pshufd xmm1,xmm1,85
+ pxor xmm2,xmm3
+ pxor xmm0,xmm1
+ pshufd xmm3,xmm0,255
+ pxor xmm2,xmm3
+ ret
+align 16
+L$124key_192b:
+ movaps xmm3,xmm0
+ shufps xmm5,xmm0,68
+ movups [edx],xmm5
+ shufps xmm3,xmm2,78
+ movups [16+edx],xmm3
+ lea edx,[32+edx]
+ jmp NEAR L$126key_192b_warm
+align 16
+L$12212rounds_alt:
+ movdqa xmm5,[16+ebx]
+ movdqa xmm4,[32+ebx]
+ mov ecx,8
+ movdqu [edx-16],xmm0
+L$127loop_key192:
+ movq [edx],xmm2
+ movdqa xmm1,xmm2
+db 102,15,56,0,213
+db 102,15,56,221,212
+ pslld xmm4,1
+ lea edx,[24+edx]
+ movdqa xmm3,xmm0
+ pslldq xmm0,4
+ pxor xmm3,xmm0
+ pslldq xmm0,4
+ pxor xmm3,xmm0
+ pslldq xmm0,4
+ pxor xmm0,xmm3
+ pshufd xmm3,xmm0,255
+ pxor xmm3,xmm1
+ pslldq xmm1,4
+ pxor xmm3,xmm1
+ pxor xmm0,xmm2
+ pxor xmm2,xmm3
+ movdqu [edx-16],xmm0
+ dec ecx
+ jnz NEAR L$127loop_key192
+ mov ecx,11
+ mov DWORD [32+edx],ecx
+ jmp NEAR L$120good_key
+align 16
+L$11314rounds:
+ movups xmm2,[16+eax]
+ lea edx,[16+edx]
+ cmp ebp,268435456
+ je NEAR L$12814rounds_alt
+ mov ecx,13
+ movups [edx-32],xmm0
+ movups [edx-16],xmm2
+db 102,15,58,223,202,1
+ call L$129key_256a_cold
+db 102,15,58,223,200,1
+ call L$130key_256b
+db 102,15,58,223,202,2
+ call L$131key_256a
+db 102,15,58,223,200,2
+ call L$130key_256b
+db 102,15,58,223,202,4
+ call L$131key_256a
+db 102,15,58,223,200,4
+ call L$130key_256b
+db 102,15,58,223,202,8
+ call L$131key_256a
+db 102,15,58,223,200,8
+ call L$130key_256b
+db 102,15,58,223,202,16
+ call L$131key_256a
+db 102,15,58,223,200,16
+ call L$130key_256b
+db 102,15,58,223,202,32
+ call L$131key_256a
+db 102,15,58,223,200,32
+ call L$130key_256b
+db 102,15,58,223,202,64
+ call L$131key_256a
+ movups [edx],xmm0
+ mov DWORD [16+edx],ecx
+ xor eax,eax
+ jmp NEAR L$120good_key
+align 16
+L$131key_256a:
+ movups [edx],xmm2
+ lea edx,[16+edx]
+L$129key_256a_cold:
+ shufps xmm4,xmm0,16
+ xorps xmm0,xmm4
+ shufps xmm4,xmm0,140
+ xorps xmm0,xmm4
+ shufps xmm1,xmm1,255
+ xorps xmm0,xmm1
+ ret
+align 16
+L$130key_256b:
+ movups [edx],xmm0
+ lea edx,[16+edx]
+ shufps xmm4,xmm2,16
+ xorps xmm2,xmm4
+ shufps xmm4,xmm2,140
+ xorps xmm2,xmm4
+ shufps xmm1,xmm1,170
+ xorps xmm2,xmm1
+ ret
+align 16
+L$12814rounds_alt:
+ movdqa xmm5,[ebx]
+ movdqa xmm4,[32+ebx]
+ mov ecx,7
+ movdqu [edx-32],xmm0
+ movdqa xmm1,xmm2
+ movdqu [edx-16],xmm2
+L$132loop_key256:
+db 102,15,56,0,213
+db 102,15,56,221,212
+ movdqa xmm3,xmm0
+ pslldq xmm0,4
+ pxor xmm3,xmm0
+ pslldq xmm0,4
+ pxor xmm3,xmm0
+ pslldq xmm0,4
+ pxor xmm0,xmm3
+ pslld xmm4,1
+ pxor xmm0,xmm2
+ movdqu [edx],xmm0
+ dec ecx
+ jz NEAR L$133done_key256
+ pshufd xmm2,xmm0,255
+ pxor xmm3,xmm3
+db 102,15,56,221,211
+ movdqa xmm3,xmm1
+ pslldq xmm1,4
+ pxor xmm3,xmm1
+ pslldq xmm1,4
+ pxor xmm3,xmm1
+ pslldq xmm1,4
+ pxor xmm1,xmm3
+ pxor xmm2,xmm1
+ movdqu [16+edx],xmm2
+ lea edx,[32+edx]
+ movdqa xmm1,xmm2
+ jmp NEAR L$132loop_key256
+L$133done_key256:
+ mov ecx,13
+ mov DWORD [16+edx],ecx
+L$120good_key:
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ pxor xmm5,xmm5
+ xor eax,eax
+ pop ebx
+ pop ebp
+ ret
+align 4
+L$111bad_pointer:
+ mov eax,-1
+ pop ebx
+ pop ebp
+ ret
+align 4
+L$115bad_keybits:
+ pxor xmm0,xmm0
+ mov eax,-2
+ pop ebx
+ pop ebp
+ ret
+global _aesni_set_encrypt_key
+align 16
+_aesni_set_encrypt_key:
+L$_aesni_set_encrypt_key_begin:
+ mov eax,DWORD [4+esp]
+ mov ecx,DWORD [8+esp]
+ mov edx,DWORD [12+esp]
+ call __aesni_set_encrypt_key
+ ret
+global _aesni_set_decrypt_key
+align 16
+_aesni_set_decrypt_key:
+L$_aesni_set_decrypt_key_begin:
+ mov eax,DWORD [4+esp]
+ mov ecx,DWORD [8+esp]
+ mov edx,DWORD [12+esp]
+ call __aesni_set_encrypt_key
+ mov edx,DWORD [12+esp]
+ shl ecx,4
+ test eax,eax
+ jnz NEAR L$134dec_key_ret
+ lea eax,[16+ecx*1+edx]
+ movups xmm0,[edx]
+ movups xmm1,[eax]
+ movups [eax],xmm0
+ movups [edx],xmm1
+ lea edx,[16+edx]
+ lea eax,[eax-16]
+L$135dec_key_inverse:
+ movups xmm0,[edx]
+ movups xmm1,[eax]
+db 102,15,56,219,192
+db 102,15,56,219,201
+ lea edx,[16+edx]
+ lea eax,[eax-16]
+ movups [16+eax],xmm0
+ movups [edx-16],xmm1
+ cmp eax,edx
+ ja NEAR L$135dec_key_inverse
+ movups xmm0,[edx]
+db 102,15,56,219,192
+ movups [edx],xmm0
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ xor eax,eax
+L$134dec_key_ret:
+ ret
+align 64
+L$key_const:
+dd 202313229,202313229,202313229,202313229
+dd 67569157,67569157,67569157,67569157
+dd 1,1,1,1
+dd 27,27,27,27
+db 65,69,83,32,102,111,114,32,73,110,116,101,108,32,65,69
+db 83,45,78,73,44,32,67,82,89,80,84,79,71,65,77,83
+db 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
+db 115,108,46,111,114,103,62,0
+segment .bss
+common _OPENSSL_ia32cap_P 16
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/aes/vpaes-x86.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/aes/vpaes-x86.nasm
new file mode 100644
index 0000000000..cf2e13f620
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/aes/vpaes-x86.nasm
@@ -0,0 +1,651 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/aes/asm/vpaes-x86.pl
+;
+; Copyright 2011-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+align 64
+L$_vpaes_consts:
+dd 218628480,235210255,168496130,67568393
+dd 252381056,17041926,33884169,51187212
+dd 252645135,252645135,252645135,252645135
+dd 1512730624,3266504856,1377990664,3401244816
+dd 830229760,1275146365,2969422977,3447763452
+dd 3411033600,2979783055,338359620,2782886510
+dd 4209124096,907596821,221174255,1006095553
+dd 191964160,3799684038,3164090317,1589111125
+dd 182528256,1777043520,2877432650,3265356744
+dd 1874708224,3503451415,3305285752,363511674
+dd 1606117888,3487855781,1093350906,2384367825
+dd 197121,67569157,134941193,202313229
+dd 67569157,134941193,202313229,197121
+dd 134941193,202313229,197121,67569157
+dd 202313229,197121,67569157,134941193
+dd 33619971,100992007,168364043,235736079
+dd 235736079,33619971,100992007,168364043
+dd 168364043,235736079,33619971,100992007
+dd 100992007,168364043,235736079,33619971
+dd 50462976,117835012,185207048,252579084
+dd 252314880,51251460,117574920,184942860
+dd 184682752,252054788,50987272,118359308
+dd 118099200,185467140,251790600,50727180
+dd 2946363062,528716217,1300004225,1881839624
+dd 1532713819,1532713819,1532713819,1532713819
+dd 3602276352,4288629033,3737020424,4153884961
+dd 1354558464,32357713,2958822624,3775749553
+dd 1201988352,132424512,1572796698,503232858
+dd 2213177600,1597421020,4103937655,675398315
+dd 2749646592,4273543773,1511898873,121693092
+dd 3040248576,1103263732,2871565598,1608280554
+dd 2236667136,2588920351,482954393,64377734
+dd 3069987328,291237287,2117370568,3650299247
+dd 533321216,3573750986,2572112006,1401264716
+dd 1339849704,2721158661,548607111,3445553514
+dd 2128193280,3054596040,2183486460,1257083700
+dd 655635200,1165381986,3923443150,2344132524
+dd 190078720,256924420,290342170,357187870
+dd 1610966272,2263057382,4103205268,309794674
+dd 2592527872,2233205587,1335446729,3402964816
+dd 3973531904,3225098121,3002836325,1918774430
+dd 3870401024,2102906079,2284471353,4117666579
+dd 617007872,1021508343,366931923,691083277
+dd 2528395776,3491914898,2968704004,1613121270
+dd 3445188352,3247741094,844474987,4093578302
+dd 651481088,1190302358,1689581232,574775300
+dd 4289380608,206939853,2555985458,2489840491
+dd 2130264064,327674451,3566485037,3349835193
+dd 2470714624,316102159,3636825756,3393945945
+db 86,101,99,116,111,114,32,80,101,114,109,117,116,97,116,105
+db 111,110,32,65,69,83,32,102,111,114,32,120,56,54,47,83
+db 83,83,69,51,44,32,77,105,107,101,32,72,97,109,98,117
+db 114,103,32,40,83,116,97,110,102,111,114,100,32,85,110,105
+db 118,101,114,115,105,116,121,41,0
+align 64
+align 16
+__vpaes_preheat:
+ add ebp,DWORD [esp]
+ movdqa xmm7,[ebp-48]
+ movdqa xmm6,[ebp-16]
+ ret
+align 16
+__vpaes_encrypt_core:
+ mov ecx,16
+ mov eax,DWORD [240+edx]
+ movdqa xmm1,xmm6
+ movdqa xmm2,[ebp]
+ pandn xmm1,xmm0
+ pand xmm0,xmm6
+ movdqu xmm5,[edx]
+db 102,15,56,0,208
+ movdqa xmm0,[16+ebp]
+ pxor xmm2,xmm5
+ psrld xmm1,4
+ add edx,16
+db 102,15,56,0,193
+ lea ebx,[192+ebp]
+ pxor xmm0,xmm2
+ jmp NEAR L$000enc_entry
+align 16
+L$001enc_loop:
+ movdqa xmm4,[32+ebp]
+ movdqa xmm0,[48+ebp]
+db 102,15,56,0,226
+db 102,15,56,0,195
+ pxor xmm4,xmm5
+ movdqa xmm5,[64+ebp]
+ pxor xmm0,xmm4
+ movdqa xmm1,[ecx*1+ebx-64]
+db 102,15,56,0,234
+ movdqa xmm2,[80+ebp]
+ movdqa xmm4,[ecx*1+ebx]
+db 102,15,56,0,211
+ movdqa xmm3,xmm0
+ pxor xmm2,xmm5
+db 102,15,56,0,193
+ add edx,16
+ pxor xmm0,xmm2
+db 102,15,56,0,220
+ add ecx,16
+ pxor xmm3,xmm0
+db 102,15,56,0,193
+ and ecx,48
+ sub eax,1
+ pxor xmm0,xmm3
+L$000enc_entry:
+ movdqa xmm1,xmm6
+ movdqa xmm5,[ebp-32]
+ pandn xmm1,xmm0
+ psrld xmm1,4
+ pand xmm0,xmm6
+db 102,15,56,0,232
+ movdqa xmm3,xmm7
+ pxor xmm0,xmm1
+db 102,15,56,0,217
+ movdqa xmm4,xmm7
+ pxor xmm3,xmm5
+db 102,15,56,0,224
+ movdqa xmm2,xmm7
+ pxor xmm4,xmm5
+db 102,15,56,0,211
+ movdqa xmm3,xmm7
+ pxor xmm2,xmm0
+db 102,15,56,0,220
+ movdqu xmm5,[edx]
+ pxor xmm3,xmm1
+ jnz NEAR L$001enc_loop
+ movdqa xmm4,[96+ebp]
+ movdqa xmm0,[112+ebp]
+db 102,15,56,0,226
+ pxor xmm4,xmm5
+db 102,15,56,0,195
+ movdqa xmm1,[64+ecx*1+ebx]
+ pxor xmm0,xmm4
+db 102,15,56,0,193
+ ret
+align 16
+__vpaes_decrypt_core:
+ lea ebx,[608+ebp]
+ mov eax,DWORD [240+edx]
+ movdqa xmm1,xmm6
+ movdqa xmm2,[ebx-64]
+ pandn xmm1,xmm0
+ mov ecx,eax
+ psrld xmm1,4
+ movdqu xmm5,[edx]
+ shl ecx,4
+ pand xmm0,xmm6
+db 102,15,56,0,208
+ movdqa xmm0,[ebx-48]
+ xor ecx,48
+db 102,15,56,0,193
+ and ecx,48
+ pxor xmm2,xmm5
+ movdqa xmm5,[176+ebp]
+ pxor xmm0,xmm2
+ add edx,16
+ lea ecx,[ecx*1+ebx-352]
+ jmp NEAR L$002dec_entry
+align 16
+L$003dec_loop:
+ movdqa xmm4,[ebx-32]
+ movdqa xmm1,[ebx-16]
+db 102,15,56,0,226
+db 102,15,56,0,203
+ pxor xmm0,xmm4
+ movdqa xmm4,[ebx]
+ pxor xmm0,xmm1
+ movdqa xmm1,[16+ebx]
+db 102,15,56,0,226
+db 102,15,56,0,197
+db 102,15,56,0,203
+ pxor xmm0,xmm4
+ movdqa xmm4,[32+ebx]
+ pxor xmm0,xmm1
+ movdqa xmm1,[48+ebx]
+db 102,15,56,0,226
+db 102,15,56,0,197
+db 102,15,56,0,203
+ pxor xmm0,xmm4
+ movdqa xmm4,[64+ebx]
+ pxor xmm0,xmm1
+ movdqa xmm1,[80+ebx]
+db 102,15,56,0,226
+db 102,15,56,0,197
+db 102,15,56,0,203
+ pxor xmm0,xmm4
+ add edx,16
+db 102,15,58,15,237,12
+ pxor xmm0,xmm1
+ sub eax,1
+L$002dec_entry:
+ movdqa xmm1,xmm6
+ movdqa xmm2,[ebp-32]
+ pandn xmm1,xmm0
+ pand xmm0,xmm6
+ psrld xmm1,4
+db 102,15,56,0,208
+ movdqa xmm3,xmm7
+ pxor xmm0,xmm1
+db 102,15,56,0,217
+ movdqa xmm4,xmm7
+ pxor xmm3,xmm2
+db 102,15,56,0,224
+ pxor xmm4,xmm2
+ movdqa xmm2,xmm7
+db 102,15,56,0,211
+ movdqa xmm3,xmm7
+ pxor xmm2,xmm0
+db 102,15,56,0,220
+ movdqu xmm0,[edx]
+ pxor xmm3,xmm1
+ jnz NEAR L$003dec_loop
+ movdqa xmm4,[96+ebx]
+db 102,15,56,0,226
+ pxor xmm4,xmm0
+ movdqa xmm0,[112+ebx]
+ movdqa xmm2,[ecx]
+db 102,15,56,0,195
+ pxor xmm0,xmm4
+db 102,15,56,0,194
+ ret
+align 16
+__vpaes_schedule_core:
+ add ebp,DWORD [esp]
+ movdqu xmm0,[esi]
+ movdqa xmm2,[320+ebp]
+ movdqa xmm3,xmm0
+ lea ebx,[ebp]
+ movdqa [4+esp],xmm2
+ call __vpaes_schedule_transform
+ movdqa xmm7,xmm0
+ test edi,edi
+ jnz NEAR L$004schedule_am_decrypting
+ movdqu [edx],xmm0
+ jmp NEAR L$005schedule_go
+L$004schedule_am_decrypting:
+ movdqa xmm1,[256+ecx*1+ebp]
+db 102,15,56,0,217
+ movdqu [edx],xmm3
+ xor ecx,48
+L$005schedule_go:
+ cmp eax,192
+ ja NEAR L$006schedule_256
+ je NEAR L$007schedule_192
+L$008schedule_128:
+ mov eax,10
+L$009loop_schedule_128:
+ call __vpaes_schedule_round
+ dec eax
+ jz NEAR L$010schedule_mangle_last
+ call __vpaes_schedule_mangle
+ jmp NEAR L$009loop_schedule_128
+align 16
+L$007schedule_192:
+ movdqu xmm0,[8+esi]
+ call __vpaes_schedule_transform
+ movdqa xmm6,xmm0
+ pxor xmm4,xmm4
+ movhlps xmm6,xmm4
+ mov eax,4
+L$011loop_schedule_192:
+ call __vpaes_schedule_round
+db 102,15,58,15,198,8
+ call __vpaes_schedule_mangle
+ call __vpaes_schedule_192_smear
+ call __vpaes_schedule_mangle
+ call __vpaes_schedule_round
+ dec eax
+ jz NEAR L$010schedule_mangle_last
+ call __vpaes_schedule_mangle
+ call __vpaes_schedule_192_smear
+ jmp NEAR L$011loop_schedule_192
+align 16
+L$006schedule_256:
+ movdqu xmm0,[16+esi]
+ call __vpaes_schedule_transform
+ mov eax,7
+L$012loop_schedule_256:
+ call __vpaes_schedule_mangle
+ movdqa xmm6,xmm0
+ call __vpaes_schedule_round
+ dec eax
+ jz NEAR L$010schedule_mangle_last
+ call __vpaes_schedule_mangle
+ pshufd xmm0,xmm0,255
+ movdqa [20+esp],xmm7
+ movdqa xmm7,xmm6
+ call L$_vpaes_schedule_low_round
+ movdqa xmm7,[20+esp]
+ jmp NEAR L$012loop_schedule_256
+align 16
+L$010schedule_mangle_last:
+ lea ebx,[384+ebp]
+ test edi,edi
+ jnz NEAR L$013schedule_mangle_last_dec
+ movdqa xmm1,[256+ecx*1+ebp]
+db 102,15,56,0,193
+ lea ebx,[352+ebp]
+ add edx,32
+L$013schedule_mangle_last_dec:
+ add edx,-16
+ pxor xmm0,[336+ebp]
+ call __vpaes_schedule_transform
+ movdqu [edx],xmm0
+ pxor xmm0,xmm0
+ pxor xmm1,xmm1
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ pxor xmm4,xmm4
+ pxor xmm5,xmm5
+ pxor xmm6,xmm6
+ pxor xmm7,xmm7
+ ret
+align 16
+__vpaes_schedule_192_smear:
+ pshufd xmm1,xmm6,128
+ pshufd xmm0,xmm7,254
+ pxor xmm6,xmm1
+ pxor xmm1,xmm1
+ pxor xmm6,xmm0
+ movdqa xmm0,xmm6
+ movhlps xmm6,xmm1
+ ret
+align 16
+__vpaes_schedule_round:
+ movdqa xmm2,[8+esp]
+ pxor xmm1,xmm1
+db 102,15,58,15,202,15
+db 102,15,58,15,210,15
+ pxor xmm7,xmm1
+ pshufd xmm0,xmm0,255
+db 102,15,58,15,192,1
+ movdqa [8+esp],xmm2
+L$_vpaes_schedule_low_round:
+ movdqa xmm1,xmm7
+ pslldq xmm7,4
+ pxor xmm7,xmm1
+ movdqa xmm1,xmm7
+ pslldq xmm7,8
+ pxor xmm7,xmm1
+ pxor xmm7,[336+ebp]
+ movdqa xmm4,[ebp-16]
+ movdqa xmm5,[ebp-48]
+ movdqa xmm1,xmm4
+ pandn xmm1,xmm0
+ psrld xmm1,4
+ pand xmm0,xmm4
+ movdqa xmm2,[ebp-32]
+db 102,15,56,0,208
+ pxor xmm0,xmm1
+ movdqa xmm3,xmm5
+db 102,15,56,0,217
+ pxor xmm3,xmm2
+ movdqa xmm4,xmm5
+db 102,15,56,0,224
+ pxor xmm4,xmm2
+ movdqa xmm2,xmm5
+db 102,15,56,0,211
+ pxor xmm2,xmm0
+ movdqa xmm3,xmm5
+db 102,15,56,0,220
+ pxor xmm3,xmm1
+ movdqa xmm4,[32+ebp]
+db 102,15,56,0,226
+ movdqa xmm0,[48+ebp]
+db 102,15,56,0,195
+ pxor xmm0,xmm4
+ pxor xmm0,xmm7
+ movdqa xmm7,xmm0
+ ret
+align 16
+__vpaes_schedule_transform:
+ movdqa xmm2,[ebp-16]
+ movdqa xmm1,xmm2
+ pandn xmm1,xmm0
+ psrld xmm1,4
+ pand xmm0,xmm2
+ movdqa xmm2,[ebx]
+db 102,15,56,0,208
+ movdqa xmm0,[16+ebx]
+db 102,15,56,0,193
+ pxor xmm0,xmm2
+ ret
+align 16
+__vpaes_schedule_mangle:
+ movdqa xmm4,xmm0
+ movdqa xmm5,[128+ebp]
+ test edi,edi
+ jnz NEAR L$014schedule_mangle_dec
+ add edx,16
+ pxor xmm4,[336+ebp]
+db 102,15,56,0,229
+ movdqa xmm3,xmm4
+db 102,15,56,0,229
+ pxor xmm3,xmm4
+db 102,15,56,0,229
+ pxor xmm3,xmm4
+ jmp NEAR L$015schedule_mangle_both
+align 16
+L$014schedule_mangle_dec:
+ movdqa xmm2,[ebp-16]
+ lea esi,[416+ebp]
+ movdqa xmm1,xmm2
+ pandn xmm1,xmm4
+ psrld xmm1,4
+ pand xmm4,xmm2
+ movdqa xmm2,[esi]
+db 102,15,56,0,212
+ movdqa xmm3,[16+esi]
+db 102,15,56,0,217
+ pxor xmm3,xmm2
+db 102,15,56,0,221
+ movdqa xmm2,[32+esi]
+db 102,15,56,0,212
+ pxor xmm2,xmm3
+ movdqa xmm3,[48+esi]
+db 102,15,56,0,217
+ pxor xmm3,xmm2
+db 102,15,56,0,221
+ movdqa xmm2,[64+esi]
+db 102,15,56,0,212
+ pxor xmm2,xmm3
+ movdqa xmm3,[80+esi]
+db 102,15,56,0,217
+ pxor xmm3,xmm2
+db 102,15,56,0,221
+ movdqa xmm2,[96+esi]
+db 102,15,56,0,212
+ pxor xmm2,xmm3
+ movdqa xmm3,[112+esi]
+db 102,15,56,0,217
+ pxor xmm3,xmm2
+ add edx,-16
+L$015schedule_mangle_both:
+ movdqa xmm1,[256+ecx*1+ebp]
+db 102,15,56,0,217
+ add ecx,-16
+ and ecx,48
+ movdqu [edx],xmm3
+ ret
+global _vpaes_set_encrypt_key
+align 16
+_vpaes_set_encrypt_key:
+L$_vpaes_set_encrypt_key_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ lea ebx,[esp-56]
+ mov eax,DWORD [24+esp]
+ and ebx,-16
+ mov edx,DWORD [28+esp]
+ xchg ebx,esp
+ mov DWORD [48+esp],ebx
+ mov ebx,eax
+ shr ebx,5
+ add ebx,5
+ mov DWORD [240+edx],ebx
+ mov ecx,48
+ mov edi,0
+ lea ebp,[(L$_vpaes_consts+0x30-L$016pic_point)]
+ call __vpaes_schedule_core
+L$016pic_point:
+ mov esp,DWORD [48+esp]
+ xor eax,eax
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _vpaes_set_decrypt_key
+align 16
+_vpaes_set_decrypt_key:
+L$_vpaes_set_decrypt_key_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ lea ebx,[esp-56]
+ mov eax,DWORD [24+esp]
+ and ebx,-16
+ mov edx,DWORD [28+esp]
+ xchg ebx,esp
+ mov DWORD [48+esp],ebx
+ mov ebx,eax
+ shr ebx,5
+ add ebx,5
+ mov DWORD [240+edx],ebx
+ shl ebx,4
+ lea edx,[16+ebx*1+edx]
+ mov edi,1
+ mov ecx,eax
+ shr ecx,1
+ and ecx,32
+ xor ecx,32
+ lea ebp,[(L$_vpaes_consts+0x30-L$017pic_point)]
+ call __vpaes_schedule_core
+L$017pic_point:
+ mov esp,DWORD [48+esp]
+ xor eax,eax
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _vpaes_encrypt
+align 16
+_vpaes_encrypt:
+L$_vpaes_encrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ lea ebp,[(L$_vpaes_consts+0x30-L$018pic_point)]
+ call __vpaes_preheat
+L$018pic_point:
+ mov esi,DWORD [20+esp]
+ lea ebx,[esp-56]
+ mov edi,DWORD [24+esp]
+ and ebx,-16
+ mov edx,DWORD [28+esp]
+ xchg ebx,esp
+ mov DWORD [48+esp],ebx
+ movdqu xmm0,[esi]
+ call __vpaes_encrypt_core
+ movdqu [edi],xmm0
+ mov esp,DWORD [48+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _vpaes_decrypt
+align 16
+_vpaes_decrypt:
+L$_vpaes_decrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ lea ebp,[(L$_vpaes_consts+0x30-L$019pic_point)]
+ call __vpaes_preheat
+L$019pic_point:
+ mov esi,DWORD [20+esp]
+ lea ebx,[esp-56]
+ mov edi,DWORD [24+esp]
+ and ebx,-16
+ mov edx,DWORD [28+esp]
+ xchg ebx,esp
+ mov DWORD [48+esp],ebx
+ movdqu xmm0,[esi]
+ call __vpaes_decrypt_core
+ movdqu [edi],xmm0
+ mov esp,DWORD [48+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _vpaes_cbc_encrypt
+align 16
+_vpaes_cbc_encrypt:
+L$_vpaes_cbc_encrypt_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov edx,DWORD [32+esp]
+ sub eax,16
+ jc NEAR L$020cbc_abort
+ lea ebx,[esp-56]
+ mov ebp,DWORD [36+esp]
+ and ebx,-16
+ mov ecx,DWORD [40+esp]
+ xchg ebx,esp
+ movdqu xmm1,[ebp]
+ sub edi,esi
+ mov DWORD [48+esp],ebx
+ mov DWORD [esp],edi
+ mov DWORD [4+esp],edx
+ mov DWORD [8+esp],ebp
+ mov edi,eax
+ lea ebp,[(L$_vpaes_consts+0x30-L$021pic_point)]
+ call __vpaes_preheat
+L$021pic_point:
+ cmp ecx,0
+ je NEAR L$022cbc_dec_loop
+ jmp NEAR L$023cbc_enc_loop
+align 16
+L$023cbc_enc_loop:
+ movdqu xmm0,[esi]
+ pxor xmm0,xmm1
+ call __vpaes_encrypt_core
+ mov ebx,DWORD [esp]
+ mov edx,DWORD [4+esp]
+ movdqa xmm1,xmm0
+ movdqu [esi*1+ebx],xmm0
+ lea esi,[16+esi]
+ sub edi,16
+ jnc NEAR L$023cbc_enc_loop
+ jmp NEAR L$024cbc_done
+align 16
+L$022cbc_dec_loop:
+ movdqu xmm0,[esi]
+ movdqa [16+esp],xmm1
+ movdqa [32+esp],xmm0
+ call __vpaes_decrypt_core
+ mov ebx,DWORD [esp]
+ mov edx,DWORD [4+esp]
+ pxor xmm0,[16+esp]
+ movdqa xmm1,[32+esp]
+ movdqu [esi*1+ebx],xmm0
+ lea esi,[16+esi]
+ sub edi,16
+ jnc NEAR L$022cbc_dec_loop
+L$024cbc_done:
+ mov ebx,DWORD [8+esp]
+ mov esp,DWORD [48+esp]
+ movdqu [ebx],xmm1
+L$020cbc_abort:
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/modes/ghash-x86.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/modes/ghash-x86.nasm
new file mode 100644
index 0000000000..88c52f746c
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/modes/ghash-x86.nasm
@@ -0,0 +1,700 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/modes/asm/ghash-x86.pl
+;
+; Copyright 2010-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+global _gcm_gmult_4bit_x86
+align 16
+_gcm_gmult_4bit_x86:
+L$_gcm_gmult_4bit_x86_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ sub esp,84
+ mov edi,DWORD [104+esp]
+ mov esi,DWORD [108+esp]
+ mov ebp,DWORD [edi]
+ mov edx,DWORD [4+edi]
+ mov ecx,DWORD [8+edi]
+ mov ebx,DWORD [12+edi]
+ mov DWORD [16+esp],0
+ mov DWORD [20+esp],471859200
+ mov DWORD [24+esp],943718400
+ mov DWORD [28+esp],610271232
+ mov DWORD [32+esp],1887436800
+ mov DWORD [36+esp],1822425088
+ mov DWORD [40+esp],1220542464
+ mov DWORD [44+esp],1423966208
+ mov DWORD [48+esp],3774873600
+ mov DWORD [52+esp],4246732800
+ mov DWORD [56+esp],3644850176
+ mov DWORD [60+esp],3311403008
+ mov DWORD [64+esp],2441084928
+ mov DWORD [68+esp],2376073216
+ mov DWORD [72+esp],2847932416
+ mov DWORD [76+esp],3051356160
+ mov DWORD [esp],ebp
+ mov DWORD [4+esp],edx
+ mov DWORD [8+esp],ecx
+ mov DWORD [12+esp],ebx
+ shr ebx,20
+ and ebx,240
+ mov ebp,DWORD [4+ebx*1+esi]
+ mov edx,DWORD [ebx*1+esi]
+ mov ecx,DWORD [12+ebx*1+esi]
+ mov ebx,DWORD [8+ebx*1+esi]
+ xor eax,eax
+ mov edi,15
+ jmp NEAR L$000x86_loop
+align 16
+L$000x86_loop:
+ mov al,bl
+ shrd ebx,ecx,4
+ and al,15
+ shrd ecx,edx,4
+ shrd edx,ebp,4
+ shr ebp,4
+ xor ebp,DWORD [16+eax*4+esp]
+ mov al,BYTE [edi*1+esp]
+ and al,240
+ xor ebx,DWORD [8+eax*1+esi]
+ xor ecx,DWORD [12+eax*1+esi]
+ xor edx,DWORD [eax*1+esi]
+ xor ebp,DWORD [4+eax*1+esi]
+ dec edi
+ js NEAR L$001x86_break
+ mov al,bl
+ shrd ebx,ecx,4
+ and al,15
+ shrd ecx,edx,4
+ shrd edx,ebp,4
+ shr ebp,4
+ xor ebp,DWORD [16+eax*4+esp]
+ mov al,BYTE [edi*1+esp]
+ shl al,4
+ xor ebx,DWORD [8+eax*1+esi]
+ xor ecx,DWORD [12+eax*1+esi]
+ xor edx,DWORD [eax*1+esi]
+ xor ebp,DWORD [4+eax*1+esi]
+ jmp NEAR L$000x86_loop
+align 16
+L$001x86_break:
+ bswap ebx
+ bswap ecx
+ bswap edx
+ bswap ebp
+ mov edi,DWORD [104+esp]
+ mov DWORD [12+edi],ebx
+ mov DWORD [8+edi],ecx
+ mov DWORD [4+edi],edx
+ mov DWORD [edi],ebp
+ add esp,84
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _gcm_ghash_4bit_x86
+align 16
+_gcm_ghash_4bit_x86:
+L$_gcm_ghash_4bit_x86_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ sub esp,84
+ mov ebx,DWORD [104+esp]
+ mov esi,DWORD [108+esp]
+ mov edi,DWORD [112+esp]
+ mov ecx,DWORD [116+esp]
+ add ecx,edi
+ mov DWORD [116+esp],ecx
+ mov ebp,DWORD [ebx]
+ mov edx,DWORD [4+ebx]
+ mov ecx,DWORD [8+ebx]
+ mov ebx,DWORD [12+ebx]
+ mov DWORD [16+esp],0
+ mov DWORD [20+esp],471859200
+ mov DWORD [24+esp],943718400
+ mov DWORD [28+esp],610271232
+ mov DWORD [32+esp],1887436800
+ mov DWORD [36+esp],1822425088
+ mov DWORD [40+esp],1220542464
+ mov DWORD [44+esp],1423966208
+ mov DWORD [48+esp],3774873600
+ mov DWORD [52+esp],4246732800
+ mov DWORD [56+esp],3644850176
+ mov DWORD [60+esp],3311403008
+ mov DWORD [64+esp],2441084928
+ mov DWORD [68+esp],2376073216
+ mov DWORD [72+esp],2847932416
+ mov DWORD [76+esp],3051356160
+align 16
+L$002x86_outer_loop:
+ xor ebx,DWORD [12+edi]
+ xor ecx,DWORD [8+edi]
+ xor edx,DWORD [4+edi]
+ xor ebp,DWORD [edi]
+ mov DWORD [12+esp],ebx
+ mov DWORD [8+esp],ecx
+ mov DWORD [4+esp],edx
+ mov DWORD [esp],ebp
+ shr ebx,20
+ and ebx,240
+ mov ebp,DWORD [4+ebx*1+esi]
+ mov edx,DWORD [ebx*1+esi]
+ mov ecx,DWORD [12+ebx*1+esi]
+ mov ebx,DWORD [8+ebx*1+esi]
+ xor eax,eax
+ mov edi,15
+ jmp NEAR L$003x86_loop
+align 16
+L$003x86_loop:
+ mov al,bl
+ shrd ebx,ecx,4
+ and al,15
+ shrd ecx,edx,4
+ shrd edx,ebp,4
+ shr ebp,4
+ xor ebp,DWORD [16+eax*4+esp]
+ mov al,BYTE [edi*1+esp]
+ and al,240
+ xor ebx,DWORD [8+eax*1+esi]
+ xor ecx,DWORD [12+eax*1+esi]
+ xor edx,DWORD [eax*1+esi]
+ xor ebp,DWORD [4+eax*1+esi]
+ dec edi
+ js NEAR L$004x86_break
+ mov al,bl
+ shrd ebx,ecx,4
+ and al,15
+ shrd ecx,edx,4
+ shrd edx,ebp,4
+ shr ebp,4
+ xor ebp,DWORD [16+eax*4+esp]
+ mov al,BYTE [edi*1+esp]
+ shl al,4
+ xor ebx,DWORD [8+eax*1+esi]
+ xor ecx,DWORD [12+eax*1+esi]
+ xor edx,DWORD [eax*1+esi]
+ xor ebp,DWORD [4+eax*1+esi]
+ jmp NEAR L$003x86_loop
+align 16
+L$004x86_break:
+ bswap ebx
+ bswap ecx
+ bswap edx
+ bswap ebp
+ mov edi,DWORD [112+esp]
+ lea edi,[16+edi]
+ cmp edi,DWORD [116+esp]
+ mov DWORD [112+esp],edi
+ jb NEAR L$002x86_outer_loop
+ mov edi,DWORD [104+esp]
+ mov DWORD [12+edi],ebx
+ mov DWORD [8+edi],ecx
+ mov DWORD [4+edi],edx
+ mov DWORD [edi],ebp
+ add esp,84
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 16
+__mmx_gmult_4bit_inner:
+ xor ecx,ecx
+ mov edx,ebx
+ mov cl,dl
+ shl cl,4
+ and edx,240
+ movq mm0,[8+ecx*1+esi]
+ movq mm1,[ecx*1+esi]
+ movd ebp,mm0
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [14+edi]
+ psllq mm2,60
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [13+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [12+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [11+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [10+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [9+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [8+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [7+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [6+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [5+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [4+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [3+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [2+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [1+edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ mov cl,BYTE [edi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ mov edx,ecx
+ movd ebx,mm0
+ pxor mm0,mm2
+ shl cl,4
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+ecx*1+esi]
+ psllq mm2,60
+ and edx,240
+ pxor mm1,[ebp*8+eax]
+ and ebx,15
+ pxor mm1,[ecx*1+esi]
+ movd ebp,mm0
+ pxor mm0,mm2
+ psrlq mm0,4
+ movq mm2,mm1
+ psrlq mm1,4
+ pxor mm0,[8+edx*1+esi]
+ psllq mm2,60
+ pxor mm1,[ebx*8+eax]
+ and ebp,15
+ pxor mm1,[edx*1+esi]
+ movd ebx,mm0
+ pxor mm0,mm2
+ mov edi,DWORD [4+ebp*8+eax]
+ psrlq mm0,32
+ movd edx,mm1
+ psrlq mm1,32
+ movd ecx,mm0
+ movd ebp,mm1
+ shl edi,4
+ bswap ebx
+ bswap edx
+ bswap ecx
+ xor ebp,edi
+ bswap ebp
+ ret
+global _gcm_gmult_4bit_mmx
+align 16
+_gcm_gmult_4bit_mmx:
+L$_gcm_gmult_4bit_mmx_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov edi,DWORD [20+esp]
+ mov esi,DWORD [24+esp]
+ call L$005pic_point
+L$005pic_point:
+ pop eax
+ lea eax,[(L$rem_4bit-L$005pic_point)+eax]
+ movzx ebx,BYTE [15+edi]
+ call __mmx_gmult_4bit_inner
+ mov edi,DWORD [20+esp]
+ emms
+ mov DWORD [12+edi],ebx
+ mov DWORD [4+edi],edx
+ mov DWORD [8+edi],ecx
+ mov DWORD [edi],ebp
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _gcm_ghash_4bit_mmx
+align 16
+_gcm_ghash_4bit_mmx:
+L$_gcm_ghash_4bit_mmx_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov ebp,DWORD [20+esp]
+ mov esi,DWORD [24+esp]
+ mov edi,DWORD [28+esp]
+ mov ecx,DWORD [32+esp]
+ call L$006pic_point
+L$006pic_point:
+ pop eax
+ lea eax,[(L$rem_4bit-L$006pic_point)+eax]
+ add ecx,edi
+ mov DWORD [32+esp],ecx
+ sub esp,20
+ mov ebx,DWORD [12+ebp]
+ mov edx,DWORD [4+ebp]
+ mov ecx,DWORD [8+ebp]
+ mov ebp,DWORD [ebp]
+ jmp NEAR L$007mmx_outer_loop
+align 16
+L$007mmx_outer_loop:
+ xor ebx,DWORD [12+edi]
+ xor edx,DWORD [4+edi]
+ xor ecx,DWORD [8+edi]
+ xor ebp,DWORD [edi]
+ mov DWORD [48+esp],edi
+ mov DWORD [12+esp],ebx
+ mov DWORD [4+esp],edx
+ mov DWORD [8+esp],ecx
+ mov DWORD [esp],ebp
+ mov edi,esp
+ shr ebx,24
+ call __mmx_gmult_4bit_inner
+ mov edi,DWORD [48+esp]
+ lea edi,[16+edi]
+ cmp edi,DWORD [52+esp]
+ jb NEAR L$007mmx_outer_loop
+ mov edi,DWORD [40+esp]
+ emms
+ mov DWORD [12+edi],ebx
+ mov DWORD [4+edi],edx
+ mov DWORD [8+edi],ecx
+ mov DWORD [edi],ebp
+ add esp,20
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 64
+L$rem_4bit:
+dd 0,0,0,29491200,0,58982400,0,38141952
+dd 0,117964800,0,113901568,0,76283904,0,88997888
+dd 0,235929600,0,265420800,0,227803136,0,206962688
+dd 0,152567808,0,148504576,0,177995776,0,190709760
+db 71,72,65,83,72,32,102,111,114,32,120,56,54,44,32,67
+db 82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112
+db 112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62
+db 0
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha1-586.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha1-586.nasm
new file mode 100644
index 0000000000..bf2831e5ea
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha1-586.nasm
@@ -0,0 +1,1394 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/sha/asm/sha1-586.pl
+;
+; Copyright 1998-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+global _sha1_block_data_order
+align 16
+_sha1_block_data_order:
+L$_sha1_block_data_order_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov ebp,DWORD [20+esp]
+ mov esi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ sub esp,76
+ shl eax,6
+ add eax,esi
+ mov DWORD [104+esp],eax
+ mov edi,DWORD [16+ebp]
+ jmp NEAR L$000loop
+align 16
+L$000loop:
+ mov eax,DWORD [esi]
+ mov ebx,DWORD [4+esi]
+ mov ecx,DWORD [8+esi]
+ mov edx,DWORD [12+esi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ mov DWORD [esp],eax
+ mov DWORD [4+esp],ebx
+ mov DWORD [8+esp],ecx
+ mov DWORD [12+esp],edx
+ mov eax,DWORD [16+esi]
+ mov ebx,DWORD [20+esi]
+ mov ecx,DWORD [24+esi]
+ mov edx,DWORD [28+esi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ mov DWORD [16+esp],eax
+ mov DWORD [20+esp],ebx
+ mov DWORD [24+esp],ecx
+ mov DWORD [28+esp],edx
+ mov eax,DWORD [32+esi]
+ mov ebx,DWORD [36+esi]
+ mov ecx,DWORD [40+esi]
+ mov edx,DWORD [44+esi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ mov DWORD [32+esp],eax
+ mov DWORD [36+esp],ebx
+ mov DWORD [40+esp],ecx
+ mov DWORD [44+esp],edx
+ mov eax,DWORD [48+esi]
+ mov ebx,DWORD [52+esi]
+ mov ecx,DWORD [56+esi]
+ mov edx,DWORD [60+esi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ mov DWORD [48+esp],eax
+ mov DWORD [52+esp],ebx
+ mov DWORD [56+esp],ecx
+ mov DWORD [60+esp],edx
+ mov DWORD [100+esp],esi
+ mov eax,DWORD [ebp]
+ mov ebx,DWORD [4+ebp]
+ mov ecx,DWORD [8+ebp]
+ mov edx,DWORD [12+ebp]
+ ; 00_15 0
+ mov esi,ecx
+ mov ebp,eax
+ rol ebp,5
+ xor esi,edx
+ add ebp,edi
+ mov edi,DWORD [esp]
+ and esi,ebx
+ ror ebx,2
+ xor esi,edx
+ lea ebp,[1518500249+edi*1+ebp]
+ add ebp,esi
+ ; 00_15 1
+ mov edi,ebx
+ mov esi,ebp
+ rol ebp,5
+ xor edi,ecx
+ add ebp,edx
+ mov edx,DWORD [4+esp]
+ and edi,eax
+ ror eax,2
+ xor edi,ecx
+ lea ebp,[1518500249+edx*1+ebp]
+ add ebp,edi
+ ; 00_15 2
+ mov edx,eax
+ mov edi,ebp
+ rol ebp,5
+ xor edx,ebx
+ add ebp,ecx
+ mov ecx,DWORD [8+esp]
+ and edx,esi
+ ror esi,2
+ xor edx,ebx
+ lea ebp,[1518500249+ecx*1+ebp]
+ add ebp,edx
+ ; 00_15 3
+ mov ecx,esi
+ mov edx,ebp
+ rol ebp,5
+ xor ecx,eax
+ add ebp,ebx
+ mov ebx,DWORD [12+esp]
+ and ecx,edi
+ ror edi,2
+ xor ecx,eax
+ lea ebp,[1518500249+ebx*1+ebp]
+ add ebp,ecx
+ ; 00_15 4
+ mov ebx,edi
+ mov ecx,ebp
+ rol ebp,5
+ xor ebx,esi
+ add ebp,eax
+ mov eax,DWORD [16+esp]
+ and ebx,edx
+ ror edx,2
+ xor ebx,esi
+ lea ebp,[1518500249+eax*1+ebp]
+ add ebp,ebx
+ ; 00_15 5
+ mov eax,edx
+ mov ebx,ebp
+ rol ebp,5
+ xor eax,edi
+ add ebp,esi
+ mov esi,DWORD [20+esp]
+ and eax,ecx
+ ror ecx,2
+ xor eax,edi
+ lea ebp,[1518500249+esi*1+ebp]
+ add ebp,eax
+ ; 00_15 6
+ mov esi,ecx
+ mov eax,ebp
+ rol ebp,5
+ xor esi,edx
+ add ebp,edi
+ mov edi,DWORD [24+esp]
+ and esi,ebx
+ ror ebx,2
+ xor esi,edx
+ lea ebp,[1518500249+edi*1+ebp]
+ add ebp,esi
+ ; 00_15 7
+ mov edi,ebx
+ mov esi,ebp
+ rol ebp,5
+ xor edi,ecx
+ add ebp,edx
+ mov edx,DWORD [28+esp]
+ and edi,eax
+ ror eax,2
+ xor edi,ecx
+ lea ebp,[1518500249+edx*1+ebp]
+ add ebp,edi
+ ; 00_15 8
+ mov edx,eax
+ mov edi,ebp
+ rol ebp,5
+ xor edx,ebx
+ add ebp,ecx
+ mov ecx,DWORD [32+esp]
+ and edx,esi
+ ror esi,2
+ xor edx,ebx
+ lea ebp,[1518500249+ecx*1+ebp]
+ add ebp,edx
+ ; 00_15 9
+ mov ecx,esi
+ mov edx,ebp
+ rol ebp,5
+ xor ecx,eax
+ add ebp,ebx
+ mov ebx,DWORD [36+esp]
+ and ecx,edi
+ ror edi,2
+ xor ecx,eax
+ lea ebp,[1518500249+ebx*1+ebp]
+ add ebp,ecx
+ ; 00_15 10
+ mov ebx,edi
+ mov ecx,ebp
+ rol ebp,5
+ xor ebx,esi
+ add ebp,eax
+ mov eax,DWORD [40+esp]
+ and ebx,edx
+ ror edx,2
+ xor ebx,esi
+ lea ebp,[1518500249+eax*1+ebp]
+ add ebp,ebx
+ ; 00_15 11
+ mov eax,edx
+ mov ebx,ebp
+ rol ebp,5
+ xor eax,edi
+ add ebp,esi
+ mov esi,DWORD [44+esp]
+ and eax,ecx
+ ror ecx,2
+ xor eax,edi
+ lea ebp,[1518500249+esi*1+ebp]
+ add ebp,eax
+ ; 00_15 12
+ mov esi,ecx
+ mov eax,ebp
+ rol ebp,5
+ xor esi,edx
+ add ebp,edi
+ mov edi,DWORD [48+esp]
+ and esi,ebx
+ ror ebx,2
+ xor esi,edx
+ lea ebp,[1518500249+edi*1+ebp]
+ add ebp,esi
+ ; 00_15 13
+ mov edi,ebx
+ mov esi,ebp
+ rol ebp,5
+ xor edi,ecx
+ add ebp,edx
+ mov edx,DWORD [52+esp]
+ and edi,eax
+ ror eax,2
+ xor edi,ecx
+ lea ebp,[1518500249+edx*1+ebp]
+ add ebp,edi
+ ; 00_15 14
+ mov edx,eax
+ mov edi,ebp
+ rol ebp,5
+ xor edx,ebx
+ add ebp,ecx
+ mov ecx,DWORD [56+esp]
+ and edx,esi
+ ror esi,2
+ xor edx,ebx
+ lea ebp,[1518500249+ecx*1+ebp]
+ add ebp,edx
+ ; 00_15 15
+ mov ecx,esi
+ mov edx,ebp
+ rol ebp,5
+ xor ecx,eax
+ add ebp,ebx
+ mov ebx,DWORD [60+esp]
+ and ecx,edi
+ ror edi,2
+ xor ecx,eax
+ lea ebp,[1518500249+ebx*1+ebp]
+ mov ebx,DWORD [esp]
+ add ecx,ebp
+ ; 16_19 16
+ mov ebp,edi
+ xor ebx,DWORD [8+esp]
+ xor ebp,esi
+ xor ebx,DWORD [32+esp]
+ and ebp,edx
+ xor ebx,DWORD [52+esp]
+ rol ebx,1
+ xor ebp,esi
+ add eax,ebp
+ mov ebp,ecx
+ ror edx,2
+ mov DWORD [esp],ebx
+ rol ebp,5
+ lea ebx,[1518500249+eax*1+ebx]
+ mov eax,DWORD [4+esp]
+ add ebx,ebp
+ ; 16_19 17
+ mov ebp,edx
+ xor eax,DWORD [12+esp]
+ xor ebp,edi
+ xor eax,DWORD [36+esp]
+ and ebp,ecx
+ xor eax,DWORD [56+esp]
+ rol eax,1
+ xor ebp,edi
+ add esi,ebp
+ mov ebp,ebx
+ ror ecx,2
+ mov DWORD [4+esp],eax
+ rol ebp,5
+ lea eax,[1518500249+esi*1+eax]
+ mov esi,DWORD [8+esp]
+ add eax,ebp
+ ; 16_19 18
+ mov ebp,ecx
+ xor esi,DWORD [16+esp]
+ xor ebp,edx
+ xor esi,DWORD [40+esp]
+ and ebp,ebx
+ xor esi,DWORD [60+esp]
+ rol esi,1
+ xor ebp,edx
+ add edi,ebp
+ mov ebp,eax
+ ror ebx,2
+ mov DWORD [8+esp],esi
+ rol ebp,5
+ lea esi,[1518500249+edi*1+esi]
+ mov edi,DWORD [12+esp]
+ add esi,ebp
+ ; 16_19 19
+ mov ebp,ebx
+ xor edi,DWORD [20+esp]
+ xor ebp,ecx
+ xor edi,DWORD [44+esp]
+ and ebp,eax
+ xor edi,DWORD [esp]
+ rol edi,1
+ xor ebp,ecx
+ add edx,ebp
+ mov ebp,esi
+ ror eax,2
+ mov DWORD [12+esp],edi
+ rol ebp,5
+ lea edi,[1518500249+edx*1+edi]
+ mov edx,DWORD [16+esp]
+ add edi,ebp
+ ; 20_39 20
+ mov ebp,esi
+ xor edx,DWORD [24+esp]
+ xor ebp,eax
+ xor edx,DWORD [48+esp]
+ xor ebp,ebx
+ xor edx,DWORD [4+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [16+esp],edx
+ lea edx,[1859775393+ecx*1+edx]
+ mov ecx,DWORD [20+esp]
+ add edx,ebp
+ ; 20_39 21
+ mov ebp,edi
+ xor ecx,DWORD [28+esp]
+ xor ebp,esi
+ xor ecx,DWORD [52+esp]
+ xor ebp,eax
+ xor ecx,DWORD [8+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [20+esp],ecx
+ lea ecx,[1859775393+ebx*1+ecx]
+ mov ebx,DWORD [24+esp]
+ add ecx,ebp
+ ; 20_39 22
+ mov ebp,edx
+ xor ebx,DWORD [32+esp]
+ xor ebp,edi
+ xor ebx,DWORD [56+esp]
+ xor ebp,esi
+ xor ebx,DWORD [12+esp]
+ rol ebx,1
+ add eax,ebp
+ ror edx,2
+ mov ebp,ecx
+ rol ebp,5
+ mov DWORD [24+esp],ebx
+ lea ebx,[1859775393+eax*1+ebx]
+ mov eax,DWORD [28+esp]
+ add ebx,ebp
+ ; 20_39 23
+ mov ebp,ecx
+ xor eax,DWORD [36+esp]
+ xor ebp,edx
+ xor eax,DWORD [60+esp]
+ xor ebp,edi
+ xor eax,DWORD [16+esp]
+ rol eax,1
+ add esi,ebp
+ ror ecx,2
+ mov ebp,ebx
+ rol ebp,5
+ mov DWORD [28+esp],eax
+ lea eax,[1859775393+esi*1+eax]
+ mov esi,DWORD [32+esp]
+ add eax,ebp
+ ; 20_39 24
+ mov ebp,ebx
+ xor esi,DWORD [40+esp]
+ xor ebp,ecx
+ xor esi,DWORD [esp]
+ xor ebp,edx
+ xor esi,DWORD [20+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ mov DWORD [32+esp],esi
+ lea esi,[1859775393+edi*1+esi]
+ mov edi,DWORD [36+esp]
+ add esi,ebp
+ ; 20_39 25
+ mov ebp,eax
+ xor edi,DWORD [44+esp]
+ xor ebp,ebx
+ xor edi,DWORD [4+esp]
+ xor ebp,ecx
+ xor edi,DWORD [24+esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ mov DWORD [36+esp],edi
+ lea edi,[1859775393+edx*1+edi]
+ mov edx,DWORD [40+esp]
+ add edi,ebp
+ ; 20_39 26
+ mov ebp,esi
+ xor edx,DWORD [48+esp]
+ xor ebp,eax
+ xor edx,DWORD [8+esp]
+ xor ebp,ebx
+ xor edx,DWORD [28+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [40+esp],edx
+ lea edx,[1859775393+ecx*1+edx]
+ mov ecx,DWORD [44+esp]
+ add edx,ebp
+ ; 20_39 27
+ mov ebp,edi
+ xor ecx,DWORD [52+esp]
+ xor ebp,esi
+ xor ecx,DWORD [12+esp]
+ xor ebp,eax
+ xor ecx,DWORD [32+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [44+esp],ecx
+ lea ecx,[1859775393+ebx*1+ecx]
+ mov ebx,DWORD [48+esp]
+ add ecx,ebp
+ ; 20_39 28
+ mov ebp,edx
+ xor ebx,DWORD [56+esp]
+ xor ebp,edi
+ xor ebx,DWORD [16+esp]
+ xor ebp,esi
+ xor ebx,DWORD [36+esp]
+ rol ebx,1
+ add eax,ebp
+ ror edx,2
+ mov ebp,ecx
+ rol ebp,5
+ mov DWORD [48+esp],ebx
+ lea ebx,[1859775393+eax*1+ebx]
+ mov eax,DWORD [52+esp]
+ add ebx,ebp
+ ; 20_39 29
+ mov ebp,ecx
+ xor eax,DWORD [60+esp]
+ xor ebp,edx
+ xor eax,DWORD [20+esp]
+ xor ebp,edi
+ xor eax,DWORD [40+esp]
+ rol eax,1
+ add esi,ebp
+ ror ecx,2
+ mov ebp,ebx
+ rol ebp,5
+ mov DWORD [52+esp],eax
+ lea eax,[1859775393+esi*1+eax]
+ mov esi,DWORD [56+esp]
+ add eax,ebp
+ ; 20_39 30
+ mov ebp,ebx
+ xor esi,DWORD [esp]
+ xor ebp,ecx
+ xor esi,DWORD [24+esp]
+ xor ebp,edx
+ xor esi,DWORD [44+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ mov DWORD [56+esp],esi
+ lea esi,[1859775393+edi*1+esi]
+ mov edi,DWORD [60+esp]
+ add esi,ebp
+ ; 20_39 31
+ mov ebp,eax
+ xor edi,DWORD [4+esp]
+ xor ebp,ebx
+ xor edi,DWORD [28+esp]
+ xor ebp,ecx
+ xor edi,DWORD [48+esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ mov DWORD [60+esp],edi
+ lea edi,[1859775393+edx*1+edi]
+ mov edx,DWORD [esp]
+ add edi,ebp
+ ; 20_39 32
+ mov ebp,esi
+ xor edx,DWORD [8+esp]
+ xor ebp,eax
+ xor edx,DWORD [32+esp]
+ xor ebp,ebx
+ xor edx,DWORD [52+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [esp],edx
+ lea edx,[1859775393+ecx*1+edx]
+ mov ecx,DWORD [4+esp]
+ add edx,ebp
+ ; 20_39 33
+ mov ebp,edi
+ xor ecx,DWORD [12+esp]
+ xor ebp,esi
+ xor ecx,DWORD [36+esp]
+ xor ebp,eax
+ xor ecx,DWORD [56+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [4+esp],ecx
+ lea ecx,[1859775393+ebx*1+ecx]
+ mov ebx,DWORD [8+esp]
+ add ecx,ebp
+ ; 20_39 34
+ mov ebp,edx
+ xor ebx,DWORD [16+esp]
+ xor ebp,edi
+ xor ebx,DWORD [40+esp]
+ xor ebp,esi
+ xor ebx,DWORD [60+esp]
+ rol ebx,1
+ add eax,ebp
+ ror edx,2
+ mov ebp,ecx
+ rol ebp,5
+ mov DWORD [8+esp],ebx
+ lea ebx,[1859775393+eax*1+ebx]
+ mov eax,DWORD [12+esp]
+ add ebx,ebp
+ ; 20_39 35
+ mov ebp,ecx
+ xor eax,DWORD [20+esp]
+ xor ebp,edx
+ xor eax,DWORD [44+esp]
+ xor ebp,edi
+ xor eax,DWORD [esp]
+ rol eax,1
+ add esi,ebp
+ ror ecx,2
+ mov ebp,ebx
+ rol ebp,5
+ mov DWORD [12+esp],eax
+ lea eax,[1859775393+esi*1+eax]
+ mov esi,DWORD [16+esp]
+ add eax,ebp
+ ; 20_39 36
+ mov ebp,ebx
+ xor esi,DWORD [24+esp]
+ xor ebp,ecx
+ xor esi,DWORD [48+esp]
+ xor ebp,edx
+ xor esi,DWORD [4+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ mov DWORD [16+esp],esi
+ lea esi,[1859775393+edi*1+esi]
+ mov edi,DWORD [20+esp]
+ add esi,ebp
+ ; 20_39 37
+ mov ebp,eax
+ xor edi,DWORD [28+esp]
+ xor ebp,ebx
+ xor edi,DWORD [52+esp]
+ xor ebp,ecx
+ xor edi,DWORD [8+esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ mov DWORD [20+esp],edi
+ lea edi,[1859775393+edx*1+edi]
+ mov edx,DWORD [24+esp]
+ add edi,ebp
+ ; 20_39 38
+ mov ebp,esi
+ xor edx,DWORD [32+esp]
+ xor ebp,eax
+ xor edx,DWORD [56+esp]
+ xor ebp,ebx
+ xor edx,DWORD [12+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [24+esp],edx
+ lea edx,[1859775393+ecx*1+edx]
+ mov ecx,DWORD [28+esp]
+ add edx,ebp
+ ; 20_39 39
+ mov ebp,edi
+ xor ecx,DWORD [36+esp]
+ xor ebp,esi
+ xor ecx,DWORD [60+esp]
+ xor ebp,eax
+ xor ecx,DWORD [16+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [28+esp],ecx
+ lea ecx,[1859775393+ebx*1+ecx]
+ mov ebx,DWORD [32+esp]
+ add ecx,ebp
+ ; 40_59 40
+ mov ebp,edi
+ xor ebx,DWORD [40+esp]
+ xor ebp,esi
+ xor ebx,DWORD [esp]
+ and ebp,edx
+ xor ebx,DWORD [20+esp]
+ rol ebx,1
+ add ebp,eax
+ ror edx,2
+ mov eax,ecx
+ rol eax,5
+ mov DWORD [32+esp],ebx
+ lea ebx,[2400959708+ebp*1+ebx]
+ mov ebp,edi
+ add ebx,eax
+ and ebp,esi
+ mov eax,DWORD [36+esp]
+ add ebx,ebp
+ ; 40_59 41
+ mov ebp,edx
+ xor eax,DWORD [44+esp]
+ xor ebp,edi
+ xor eax,DWORD [4+esp]
+ and ebp,ecx
+ xor eax,DWORD [24+esp]
+ rol eax,1
+ add ebp,esi
+ ror ecx,2
+ mov esi,ebx
+ rol esi,5
+ mov DWORD [36+esp],eax
+ lea eax,[2400959708+ebp*1+eax]
+ mov ebp,edx
+ add eax,esi
+ and ebp,edi
+ mov esi,DWORD [40+esp]
+ add eax,ebp
+ ; 40_59 42
+ mov ebp,ecx
+ xor esi,DWORD [48+esp]
+ xor ebp,edx
+ xor esi,DWORD [8+esp]
+ and ebp,ebx
+ xor esi,DWORD [28+esp]
+ rol esi,1
+ add ebp,edi
+ ror ebx,2
+ mov edi,eax
+ rol edi,5
+ mov DWORD [40+esp],esi
+ lea esi,[2400959708+ebp*1+esi]
+ mov ebp,ecx
+ add esi,edi
+ and ebp,edx
+ mov edi,DWORD [44+esp]
+ add esi,ebp
+ ; 40_59 43
+ mov ebp,ebx
+ xor edi,DWORD [52+esp]
+ xor ebp,ecx
+ xor edi,DWORD [12+esp]
+ and ebp,eax
+ xor edi,DWORD [32+esp]
+ rol edi,1
+ add ebp,edx
+ ror eax,2
+ mov edx,esi
+ rol edx,5
+ mov DWORD [44+esp],edi
+ lea edi,[2400959708+ebp*1+edi]
+ mov ebp,ebx
+ add edi,edx
+ and ebp,ecx
+ mov edx,DWORD [48+esp]
+ add edi,ebp
+ ; 40_59 44
+ mov ebp,eax
+ xor edx,DWORD [56+esp]
+ xor ebp,ebx
+ xor edx,DWORD [16+esp]
+ and ebp,esi
+ xor edx,DWORD [36+esp]
+ rol edx,1
+ add ebp,ecx
+ ror esi,2
+ mov ecx,edi
+ rol ecx,5
+ mov DWORD [48+esp],edx
+ lea edx,[2400959708+ebp*1+edx]
+ mov ebp,eax
+ add edx,ecx
+ and ebp,ebx
+ mov ecx,DWORD [52+esp]
+ add edx,ebp
+ ; 40_59 45
+ mov ebp,esi
+ xor ecx,DWORD [60+esp]
+ xor ebp,eax
+ xor ecx,DWORD [20+esp]
+ and ebp,edi
+ xor ecx,DWORD [40+esp]
+ rol ecx,1
+ add ebp,ebx
+ ror edi,2
+ mov ebx,edx
+ rol ebx,5
+ mov DWORD [52+esp],ecx
+ lea ecx,[2400959708+ebp*1+ecx]
+ mov ebp,esi
+ add ecx,ebx
+ and ebp,eax
+ mov ebx,DWORD [56+esp]
+ add ecx,ebp
+ ; 40_59 46
+ mov ebp,edi
+ xor ebx,DWORD [esp]
+ xor ebp,esi
+ xor ebx,DWORD [24+esp]
+ and ebp,edx
+ xor ebx,DWORD [44+esp]
+ rol ebx,1
+ add ebp,eax
+ ror edx,2
+ mov eax,ecx
+ rol eax,5
+ mov DWORD [56+esp],ebx
+ lea ebx,[2400959708+ebp*1+ebx]
+ mov ebp,edi
+ add ebx,eax
+ and ebp,esi
+ mov eax,DWORD [60+esp]
+ add ebx,ebp
+ ; 40_59 47
+ mov ebp,edx
+ xor eax,DWORD [4+esp]
+ xor ebp,edi
+ xor eax,DWORD [28+esp]
+ and ebp,ecx
+ xor eax,DWORD [48+esp]
+ rol eax,1
+ add ebp,esi
+ ror ecx,2
+ mov esi,ebx
+ rol esi,5
+ mov DWORD [60+esp],eax
+ lea eax,[2400959708+ebp*1+eax]
+ mov ebp,edx
+ add eax,esi
+ and ebp,edi
+ mov esi,DWORD [esp]
+ add eax,ebp
+ ; 40_59 48
+ mov ebp,ecx
+ xor esi,DWORD [8+esp]
+ xor ebp,edx
+ xor esi,DWORD [32+esp]
+ and ebp,ebx
+ xor esi,DWORD [52+esp]
+ rol esi,1
+ add ebp,edi
+ ror ebx,2
+ mov edi,eax
+ rol edi,5
+ mov DWORD [esp],esi
+ lea esi,[2400959708+ebp*1+esi]
+ mov ebp,ecx
+ add esi,edi
+ and ebp,edx
+ mov edi,DWORD [4+esp]
+ add esi,ebp
+ ; 40_59 49
+ mov ebp,ebx
+ xor edi,DWORD [12+esp]
+ xor ebp,ecx
+ xor edi,DWORD [36+esp]
+ and ebp,eax
+ xor edi,DWORD [56+esp]
+ rol edi,1
+ add ebp,edx
+ ror eax,2
+ mov edx,esi
+ rol edx,5
+ mov DWORD [4+esp],edi
+ lea edi,[2400959708+ebp*1+edi]
+ mov ebp,ebx
+ add edi,edx
+ and ebp,ecx
+ mov edx,DWORD [8+esp]
+ add edi,ebp
+ ; 40_59 50
+ mov ebp,eax
+ xor edx,DWORD [16+esp]
+ xor ebp,ebx
+ xor edx,DWORD [40+esp]
+ and ebp,esi
+ xor edx,DWORD [60+esp]
+ rol edx,1
+ add ebp,ecx
+ ror esi,2
+ mov ecx,edi
+ rol ecx,5
+ mov DWORD [8+esp],edx
+ lea edx,[2400959708+ebp*1+edx]
+ mov ebp,eax
+ add edx,ecx
+ and ebp,ebx
+ mov ecx,DWORD [12+esp]
+ add edx,ebp
+ ; 40_59 51
+ mov ebp,esi
+ xor ecx,DWORD [20+esp]
+ xor ebp,eax
+ xor ecx,DWORD [44+esp]
+ and ebp,edi
+ xor ecx,DWORD [esp]
+ rol ecx,1
+ add ebp,ebx
+ ror edi,2
+ mov ebx,edx
+ rol ebx,5
+ mov DWORD [12+esp],ecx
+ lea ecx,[2400959708+ebp*1+ecx]
+ mov ebp,esi
+ add ecx,ebx
+ and ebp,eax
+ mov ebx,DWORD [16+esp]
+ add ecx,ebp
+ ; 40_59 52
+ mov ebp,edi
+ xor ebx,DWORD [24+esp]
+ xor ebp,esi
+ xor ebx,DWORD [48+esp]
+ and ebp,edx
+ xor ebx,DWORD [4+esp]
+ rol ebx,1
+ add ebp,eax
+ ror edx,2
+ mov eax,ecx
+ rol eax,5
+ mov DWORD [16+esp],ebx
+ lea ebx,[2400959708+ebp*1+ebx]
+ mov ebp,edi
+ add ebx,eax
+ and ebp,esi
+ mov eax,DWORD [20+esp]
+ add ebx,ebp
+ ; 40_59 53
+ mov ebp,edx
+ xor eax,DWORD [28+esp]
+ xor ebp,edi
+ xor eax,DWORD [52+esp]
+ and ebp,ecx
+ xor eax,DWORD [8+esp]
+ rol eax,1
+ add ebp,esi
+ ror ecx,2
+ mov esi,ebx
+ rol esi,5
+ mov DWORD [20+esp],eax
+ lea eax,[2400959708+ebp*1+eax]
+ mov ebp,edx
+ add eax,esi
+ and ebp,edi
+ mov esi,DWORD [24+esp]
+ add eax,ebp
+ ; 40_59 54
+ mov ebp,ecx
+ xor esi,DWORD [32+esp]
+ xor ebp,edx
+ xor esi,DWORD [56+esp]
+ and ebp,ebx
+ xor esi,DWORD [12+esp]
+ rol esi,1
+ add ebp,edi
+ ror ebx,2
+ mov edi,eax
+ rol edi,5
+ mov DWORD [24+esp],esi
+ lea esi,[2400959708+ebp*1+esi]
+ mov ebp,ecx
+ add esi,edi
+ and ebp,edx
+ mov edi,DWORD [28+esp]
+ add esi,ebp
+ ; 40_59 55
+ mov ebp,ebx
+ xor edi,DWORD [36+esp]
+ xor ebp,ecx
+ xor edi,DWORD [60+esp]
+ and ebp,eax
+ xor edi,DWORD [16+esp]
+ rol edi,1
+ add ebp,edx
+ ror eax,2
+ mov edx,esi
+ rol edx,5
+ mov DWORD [28+esp],edi
+ lea edi,[2400959708+ebp*1+edi]
+ mov ebp,ebx
+ add edi,edx
+ and ebp,ecx
+ mov edx,DWORD [32+esp]
+ add edi,ebp
+ ; 40_59 56
+ mov ebp,eax
+ xor edx,DWORD [40+esp]
+ xor ebp,ebx
+ xor edx,DWORD [esp]
+ and ebp,esi
+ xor edx,DWORD [20+esp]
+ rol edx,1
+ add ebp,ecx
+ ror esi,2
+ mov ecx,edi
+ rol ecx,5
+ mov DWORD [32+esp],edx
+ lea edx,[2400959708+ebp*1+edx]
+ mov ebp,eax
+ add edx,ecx
+ and ebp,ebx
+ mov ecx,DWORD [36+esp]
+ add edx,ebp
+ ; 40_59 57
+ mov ebp,esi
+ xor ecx,DWORD [44+esp]
+ xor ebp,eax
+ xor ecx,DWORD [4+esp]
+ and ebp,edi
+ xor ecx,DWORD [24+esp]
+ rol ecx,1
+ add ebp,ebx
+ ror edi,2
+ mov ebx,edx
+ rol ebx,5
+ mov DWORD [36+esp],ecx
+ lea ecx,[2400959708+ebp*1+ecx]
+ mov ebp,esi
+ add ecx,ebx
+ and ebp,eax
+ mov ebx,DWORD [40+esp]
+ add ecx,ebp
+ ; 40_59 58
+ mov ebp,edi
+ xor ebx,DWORD [48+esp]
+ xor ebp,esi
+ xor ebx,DWORD [8+esp]
+ and ebp,edx
+ xor ebx,DWORD [28+esp]
+ rol ebx,1
+ add ebp,eax
+ ror edx,2
+ mov eax,ecx
+ rol eax,5
+ mov DWORD [40+esp],ebx
+ lea ebx,[2400959708+ebp*1+ebx]
+ mov ebp,edi
+ add ebx,eax
+ and ebp,esi
+ mov eax,DWORD [44+esp]
+ add ebx,ebp
+ ; 40_59 59
+ mov ebp,edx
+ xor eax,DWORD [52+esp]
+ xor ebp,edi
+ xor eax,DWORD [12+esp]
+ and ebp,ecx
+ xor eax,DWORD [32+esp]
+ rol eax,1
+ add ebp,esi
+ ror ecx,2
+ mov esi,ebx
+ rol esi,5
+ mov DWORD [44+esp],eax
+ lea eax,[2400959708+ebp*1+eax]
+ mov ebp,edx
+ add eax,esi
+ and ebp,edi
+ mov esi,DWORD [48+esp]
+ add eax,ebp
+ ; 20_39 60
+ mov ebp,ebx
+ xor esi,DWORD [56+esp]
+ xor ebp,ecx
+ xor esi,DWORD [16+esp]
+ xor ebp,edx
+ xor esi,DWORD [36+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ mov DWORD [48+esp],esi
+ lea esi,[3395469782+edi*1+esi]
+ mov edi,DWORD [52+esp]
+ add esi,ebp
+ ; 20_39 61
+ mov ebp,eax
+ xor edi,DWORD [60+esp]
+ xor ebp,ebx
+ xor edi,DWORD [20+esp]
+ xor ebp,ecx
+ xor edi,DWORD [40+esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ mov DWORD [52+esp],edi
+ lea edi,[3395469782+edx*1+edi]
+ mov edx,DWORD [56+esp]
+ add edi,ebp
+ ; 20_39 62
+ mov ebp,esi
+ xor edx,DWORD [esp]
+ xor ebp,eax
+ xor edx,DWORD [24+esp]
+ xor ebp,ebx
+ xor edx,DWORD [44+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [56+esp],edx
+ lea edx,[3395469782+ecx*1+edx]
+ mov ecx,DWORD [60+esp]
+ add edx,ebp
+ ; 20_39 63
+ mov ebp,edi
+ xor ecx,DWORD [4+esp]
+ xor ebp,esi
+ xor ecx,DWORD [28+esp]
+ xor ebp,eax
+ xor ecx,DWORD [48+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [60+esp],ecx
+ lea ecx,[3395469782+ebx*1+ecx]
+ mov ebx,DWORD [esp]
+ add ecx,ebp
+ ; 20_39 64
+ mov ebp,edx
+ xor ebx,DWORD [8+esp]
+ xor ebp,edi
+ xor ebx,DWORD [32+esp]
+ xor ebp,esi
+ xor ebx,DWORD [52+esp]
+ rol ebx,1
+ add eax,ebp
+ ror edx,2
+ mov ebp,ecx
+ rol ebp,5
+ mov DWORD [esp],ebx
+ lea ebx,[3395469782+eax*1+ebx]
+ mov eax,DWORD [4+esp]
+ add ebx,ebp
+ ; 20_39 65
+ mov ebp,ecx
+ xor eax,DWORD [12+esp]
+ xor ebp,edx
+ xor eax,DWORD [36+esp]
+ xor ebp,edi
+ xor eax,DWORD [56+esp]
+ rol eax,1
+ add esi,ebp
+ ror ecx,2
+ mov ebp,ebx
+ rol ebp,5
+ mov DWORD [4+esp],eax
+ lea eax,[3395469782+esi*1+eax]
+ mov esi,DWORD [8+esp]
+ add eax,ebp
+ ; 20_39 66
+ mov ebp,ebx
+ xor esi,DWORD [16+esp]
+ xor ebp,ecx
+ xor esi,DWORD [40+esp]
+ xor ebp,edx
+ xor esi,DWORD [60+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ mov DWORD [8+esp],esi
+ lea esi,[3395469782+edi*1+esi]
+ mov edi,DWORD [12+esp]
+ add esi,ebp
+ ; 20_39 67
+ mov ebp,eax
+ xor edi,DWORD [20+esp]
+ xor ebp,ebx
+ xor edi,DWORD [44+esp]
+ xor ebp,ecx
+ xor edi,DWORD [esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ mov DWORD [12+esp],edi
+ lea edi,[3395469782+edx*1+edi]
+ mov edx,DWORD [16+esp]
+ add edi,ebp
+ ; 20_39 68
+ mov ebp,esi
+ xor edx,DWORD [24+esp]
+ xor ebp,eax
+ xor edx,DWORD [48+esp]
+ xor ebp,ebx
+ xor edx,DWORD [4+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [16+esp],edx
+ lea edx,[3395469782+ecx*1+edx]
+ mov ecx,DWORD [20+esp]
+ add edx,ebp
+ ; 20_39 69
+ mov ebp,edi
+ xor ecx,DWORD [28+esp]
+ xor ebp,esi
+ xor ecx,DWORD [52+esp]
+ xor ebp,eax
+ xor ecx,DWORD [8+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [20+esp],ecx
+ lea ecx,[3395469782+ebx*1+ecx]
+ mov ebx,DWORD [24+esp]
+ add ecx,ebp
+ ; 20_39 70
+ mov ebp,edx
+ xor ebx,DWORD [32+esp]
+ xor ebp,edi
+ xor ebx,DWORD [56+esp]
+ xor ebp,esi
+ xor ebx,DWORD [12+esp]
+ rol ebx,1
+ add eax,ebp
+ ror edx,2
+ mov ebp,ecx
+ rol ebp,5
+ mov DWORD [24+esp],ebx
+ lea ebx,[3395469782+eax*1+ebx]
+ mov eax,DWORD [28+esp]
+ add ebx,ebp
+ ; 20_39 71
+ mov ebp,ecx
+ xor eax,DWORD [36+esp]
+ xor ebp,edx
+ xor eax,DWORD [60+esp]
+ xor ebp,edi
+ xor eax,DWORD [16+esp]
+ rol eax,1
+ add esi,ebp
+ ror ecx,2
+ mov ebp,ebx
+ rol ebp,5
+ mov DWORD [28+esp],eax
+ lea eax,[3395469782+esi*1+eax]
+ mov esi,DWORD [32+esp]
+ add eax,ebp
+ ; 20_39 72
+ mov ebp,ebx
+ xor esi,DWORD [40+esp]
+ xor ebp,ecx
+ xor esi,DWORD [esp]
+ xor ebp,edx
+ xor esi,DWORD [20+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ mov DWORD [32+esp],esi
+ lea esi,[3395469782+edi*1+esi]
+ mov edi,DWORD [36+esp]
+ add esi,ebp
+ ; 20_39 73
+ mov ebp,eax
+ xor edi,DWORD [44+esp]
+ xor ebp,ebx
+ xor edi,DWORD [4+esp]
+ xor ebp,ecx
+ xor edi,DWORD [24+esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ mov DWORD [36+esp],edi
+ lea edi,[3395469782+edx*1+edi]
+ mov edx,DWORD [40+esp]
+ add edi,ebp
+ ; 20_39 74
+ mov ebp,esi
+ xor edx,DWORD [48+esp]
+ xor ebp,eax
+ xor edx,DWORD [8+esp]
+ xor ebp,ebx
+ xor edx,DWORD [28+esp]
+ rol edx,1
+ add ecx,ebp
+ ror esi,2
+ mov ebp,edi
+ rol ebp,5
+ mov DWORD [40+esp],edx
+ lea edx,[3395469782+ecx*1+edx]
+ mov ecx,DWORD [44+esp]
+ add edx,ebp
+ ; 20_39 75
+ mov ebp,edi
+ xor ecx,DWORD [52+esp]
+ xor ebp,esi
+ xor ecx,DWORD [12+esp]
+ xor ebp,eax
+ xor ecx,DWORD [32+esp]
+ rol ecx,1
+ add ebx,ebp
+ ror edi,2
+ mov ebp,edx
+ rol ebp,5
+ mov DWORD [44+esp],ecx
+ lea ecx,[3395469782+ebx*1+ecx]
+ mov ebx,DWORD [48+esp]
+ add ecx,ebp
+ ; 20_39 76
+ mov ebp,edx
+ xor ebx,DWORD [56+esp]
+ xor ebp,edi
+ xor ebx,DWORD [16+esp]
+ xor ebp,esi
+ xor ebx,DWORD [36+esp]
+ rol ebx,1
+ add eax,ebp
+ ror edx,2
+ mov ebp,ecx
+ rol ebp,5
+ mov DWORD [48+esp],ebx
+ lea ebx,[3395469782+eax*1+ebx]
+ mov eax,DWORD [52+esp]
+ add ebx,ebp
+ ; 20_39 77
+ mov ebp,ecx
+ xor eax,DWORD [60+esp]
+ xor ebp,edx
+ xor eax,DWORD [20+esp]
+ xor ebp,edi
+ xor eax,DWORD [40+esp]
+ rol eax,1
+ add esi,ebp
+ ror ecx,2
+ mov ebp,ebx
+ rol ebp,5
+ lea eax,[3395469782+esi*1+eax]
+ mov esi,DWORD [56+esp]
+ add eax,ebp
+ ; 20_39 78
+ mov ebp,ebx
+ xor esi,DWORD [esp]
+ xor ebp,ecx
+ xor esi,DWORD [24+esp]
+ xor ebp,edx
+ xor esi,DWORD [44+esp]
+ rol esi,1
+ add edi,ebp
+ ror ebx,2
+ mov ebp,eax
+ rol ebp,5
+ lea esi,[3395469782+edi*1+esi]
+ mov edi,DWORD [60+esp]
+ add esi,ebp
+ ; 20_39 79
+ mov ebp,eax
+ xor edi,DWORD [4+esp]
+ xor ebp,ebx
+ xor edi,DWORD [28+esp]
+ xor ebp,ecx
+ xor edi,DWORD [48+esp]
+ rol edi,1
+ add edx,ebp
+ ror eax,2
+ mov ebp,esi
+ rol ebp,5
+ lea edi,[3395469782+edx*1+edi]
+ add edi,ebp
+ mov ebp,DWORD [96+esp]
+ mov edx,DWORD [100+esp]
+ add edi,DWORD [ebp]
+ add esi,DWORD [4+ebp]
+ add eax,DWORD [8+ebp]
+ add ebx,DWORD [12+ebp]
+ add ecx,DWORD [16+ebp]
+ mov DWORD [ebp],edi
+ add edx,64
+ mov DWORD [4+ebp],esi
+ cmp edx,DWORD [104+esp]
+ mov DWORD [8+ebp],eax
+ mov edi,ecx
+ mov DWORD [12+ebp],ebx
+ mov esi,edx
+ mov DWORD [16+ebp],ecx
+ jb NEAR L$000loop
+ add esp,76
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+db 83,72,65,49,32,98,108,111,99,107,32,116,114,97,110,115
+db 102,111,114,109,32,102,111,114,32,120,56,54,44,32,67,82
+db 89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112
+db 114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha256-586.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha256-586.nasm
new file mode 100644
index 0000000000..4598dae75a
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha256-586.nasm
@@ -0,0 +1,3364 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/sha/asm/sha256-586.pl
+;
+; Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+;extern _OPENSSL_ia32cap_P
+global _sha256_block_data_order
+align 16
+_sha256_block_data_order:
+L$_sha256_block_data_order_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov ebx,esp
+ call L$000pic_point
+L$000pic_point:
+ pop ebp
+ lea ebp,[(L$001K256-L$000pic_point)+ebp]
+ sub esp,16
+ and esp,-64
+ shl eax,6
+ add eax,edi
+ mov DWORD [esp],esi
+ mov DWORD [4+esp],edi
+ mov DWORD [8+esp],eax
+ mov DWORD [12+esp],ebx
+ jmp NEAR L$002loop
+align 16
+L$002loop:
+ mov eax,DWORD [edi]
+ mov ebx,DWORD [4+edi]
+ mov ecx,DWORD [8+edi]
+ bswap eax
+ mov edx,DWORD [12+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ mov eax,DWORD [16+edi]
+ mov ebx,DWORD [20+edi]
+ mov ecx,DWORD [24+edi]
+ bswap eax
+ mov edx,DWORD [28+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ mov eax,DWORD [32+edi]
+ mov ebx,DWORD [36+edi]
+ mov ecx,DWORD [40+edi]
+ bswap eax
+ mov edx,DWORD [44+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ mov eax,DWORD [48+edi]
+ mov ebx,DWORD [52+edi]
+ mov ecx,DWORD [56+edi]
+ bswap eax
+ mov edx,DWORD [60+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ add edi,64
+ lea esp,[esp-36]
+ mov DWORD [104+esp],edi
+ mov eax,DWORD [esi]
+ mov ebx,DWORD [4+esi]
+ mov ecx,DWORD [8+esi]
+ mov edi,DWORD [12+esi]
+ mov DWORD [8+esp],ebx
+ xor ebx,ecx
+ mov DWORD [12+esp],ecx
+ mov DWORD [16+esp],edi
+ mov DWORD [esp],ebx
+ mov edx,DWORD [16+esi]
+ mov ebx,DWORD [20+esi]
+ mov ecx,DWORD [24+esi]
+ mov edi,DWORD [28+esi]
+ mov DWORD [24+esp],ebx
+ mov DWORD [28+esp],ecx
+ mov DWORD [32+esp],edi
+align 16
+L$00300_15:
+ mov ecx,edx
+ mov esi,DWORD [24+esp]
+ ror ecx,14
+ mov edi,DWORD [28+esp]
+ xor ecx,edx
+ xor esi,edi
+ mov ebx,DWORD [96+esp]
+ ror ecx,5
+ and esi,edx
+ mov DWORD [20+esp],edx
+ xor edx,ecx
+ add ebx,DWORD [32+esp]
+ xor esi,edi
+ ror edx,6
+ mov ecx,eax
+ add ebx,esi
+ ror ecx,9
+ add ebx,edx
+ mov edi,DWORD [8+esp]
+ xor ecx,eax
+ mov DWORD [4+esp],eax
+ lea esp,[esp-4]
+ ror ecx,11
+ mov esi,DWORD [ebp]
+ xor ecx,eax
+ mov edx,DWORD [20+esp]
+ xor eax,edi
+ ror ecx,2
+ add ebx,esi
+ mov DWORD [esp],eax
+ add edx,ebx
+ and eax,DWORD [4+esp]
+ add ebx,ecx
+ xor eax,edi
+ add ebp,4
+ add eax,ebx
+ cmp esi,3248222580
+ jne NEAR L$00300_15
+ mov ecx,DWORD [156+esp]
+ jmp NEAR L$00416_63
+align 16
+L$00416_63:
+ mov ebx,ecx
+ mov esi,DWORD [104+esp]
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [160+esp]
+ shr edi,10
+ add ebx,DWORD [124+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [24+esp]
+ ror ecx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor ecx,edx
+ xor esi,edi
+ mov DWORD [96+esp],ebx
+ ror ecx,5
+ and esi,edx
+ mov DWORD [20+esp],edx
+ xor edx,ecx
+ add ebx,DWORD [32+esp]
+ xor esi,edi
+ ror edx,6
+ mov ecx,eax
+ add ebx,esi
+ ror ecx,9
+ add ebx,edx
+ mov edi,DWORD [8+esp]
+ xor ecx,eax
+ mov DWORD [4+esp],eax
+ lea esp,[esp-4]
+ ror ecx,11
+ mov esi,DWORD [ebp]
+ xor ecx,eax
+ mov edx,DWORD [20+esp]
+ xor eax,edi
+ ror ecx,2
+ add ebx,esi
+ mov DWORD [esp],eax
+ add edx,ebx
+ and eax,DWORD [4+esp]
+ add ebx,ecx
+ xor eax,edi
+ mov ecx,DWORD [156+esp]
+ add ebp,4
+ add eax,ebx
+ cmp esi,3329325298
+ jne NEAR L$00416_63
+ mov esi,DWORD [356+esp]
+ mov ebx,DWORD [8+esp]
+ mov ecx,DWORD [16+esp]
+ add eax,DWORD [esi]
+ add ebx,DWORD [4+esi]
+ add edi,DWORD [8+esi]
+ add ecx,DWORD [12+esi]
+ mov DWORD [esi],eax
+ mov DWORD [4+esi],ebx
+ mov DWORD [8+esi],edi
+ mov DWORD [12+esi],ecx
+ mov eax,DWORD [24+esp]
+ mov ebx,DWORD [28+esp]
+ mov ecx,DWORD [32+esp]
+ mov edi,DWORD [360+esp]
+ add edx,DWORD [16+esi]
+ add eax,DWORD [20+esi]
+ add ebx,DWORD [24+esi]
+ add ecx,DWORD [28+esi]
+ mov DWORD [16+esi],edx
+ mov DWORD [20+esi],eax
+ mov DWORD [24+esi],ebx
+ mov DWORD [28+esi],ecx
+ lea esp,[356+esp]
+ sub ebp,256
+ cmp edi,DWORD [8+esp]
+ jb NEAR L$002loop
+ mov esp,DWORD [12+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 32
+L$005loop_shrd:
+ mov eax,DWORD [edi]
+ mov ebx,DWORD [4+edi]
+ mov ecx,DWORD [8+edi]
+ bswap eax
+ mov edx,DWORD [12+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ mov eax,DWORD [16+edi]
+ mov ebx,DWORD [20+edi]
+ mov ecx,DWORD [24+edi]
+ bswap eax
+ mov edx,DWORD [28+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ mov eax,DWORD [32+edi]
+ mov ebx,DWORD [36+edi]
+ mov ecx,DWORD [40+edi]
+ bswap eax
+ mov edx,DWORD [44+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ mov eax,DWORD [48+edi]
+ mov ebx,DWORD [52+edi]
+ mov ecx,DWORD [56+edi]
+ bswap eax
+ mov edx,DWORD [60+edi]
+ bswap ebx
+ push eax
+ bswap ecx
+ push ebx
+ bswap edx
+ push ecx
+ push edx
+ add edi,64
+ lea esp,[esp-36]
+ mov DWORD [104+esp],edi
+ mov eax,DWORD [esi]
+ mov ebx,DWORD [4+esi]
+ mov ecx,DWORD [8+esi]
+ mov edi,DWORD [12+esi]
+ mov DWORD [8+esp],ebx
+ xor ebx,ecx
+ mov DWORD [12+esp],ecx
+ mov DWORD [16+esp],edi
+ mov DWORD [esp],ebx
+ mov edx,DWORD [16+esi]
+ mov ebx,DWORD [20+esi]
+ mov ecx,DWORD [24+esi]
+ mov edi,DWORD [28+esi]
+ mov DWORD [24+esp],ebx
+ mov DWORD [28+esp],ecx
+ mov DWORD [32+esp],edi
+align 16
+L$00600_15_shrd:
+ mov ecx,edx
+ mov esi,DWORD [24+esp]
+ shrd ecx,ecx,14
+ mov edi,DWORD [28+esp]
+ xor ecx,edx
+ xor esi,edi
+ mov ebx,DWORD [96+esp]
+ shrd ecx,ecx,5
+ and esi,edx
+ mov DWORD [20+esp],edx
+ xor edx,ecx
+ add ebx,DWORD [32+esp]
+ xor esi,edi
+ shrd edx,edx,6
+ mov ecx,eax
+ add ebx,esi
+ shrd ecx,ecx,9
+ add ebx,edx
+ mov edi,DWORD [8+esp]
+ xor ecx,eax
+ mov DWORD [4+esp],eax
+ lea esp,[esp-4]
+ shrd ecx,ecx,11
+ mov esi,DWORD [ebp]
+ xor ecx,eax
+ mov edx,DWORD [20+esp]
+ xor eax,edi
+ shrd ecx,ecx,2
+ add ebx,esi
+ mov DWORD [esp],eax
+ add edx,ebx
+ and eax,DWORD [4+esp]
+ add ebx,ecx
+ xor eax,edi
+ add ebp,4
+ add eax,ebx
+ cmp esi,3248222580
+ jne NEAR L$00600_15_shrd
+ mov ecx,DWORD [156+esp]
+ jmp NEAR L$00716_63_shrd
+align 16
+L$00716_63_shrd:
+ mov ebx,ecx
+ mov esi,DWORD [104+esp]
+ shrd ecx,ecx,11
+ mov edi,esi
+ shrd esi,esi,2
+ xor ecx,ebx
+ shr ebx,3
+ shrd ecx,ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ shrd esi,esi,17
+ add ebx,DWORD [160+esp]
+ shr edi,10
+ add ebx,DWORD [124+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [24+esp]
+ shrd ecx,ecx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor ecx,edx
+ xor esi,edi
+ mov DWORD [96+esp],ebx
+ shrd ecx,ecx,5
+ and esi,edx
+ mov DWORD [20+esp],edx
+ xor edx,ecx
+ add ebx,DWORD [32+esp]
+ xor esi,edi
+ shrd edx,edx,6
+ mov ecx,eax
+ add ebx,esi
+ shrd ecx,ecx,9
+ add ebx,edx
+ mov edi,DWORD [8+esp]
+ xor ecx,eax
+ mov DWORD [4+esp],eax
+ lea esp,[esp-4]
+ shrd ecx,ecx,11
+ mov esi,DWORD [ebp]
+ xor ecx,eax
+ mov edx,DWORD [20+esp]
+ xor eax,edi
+ shrd ecx,ecx,2
+ add ebx,esi
+ mov DWORD [esp],eax
+ add edx,ebx
+ and eax,DWORD [4+esp]
+ add ebx,ecx
+ xor eax,edi
+ mov ecx,DWORD [156+esp]
+ add ebp,4
+ add eax,ebx
+ cmp esi,3329325298
+ jne NEAR L$00716_63_shrd
+ mov esi,DWORD [356+esp]
+ mov ebx,DWORD [8+esp]
+ mov ecx,DWORD [16+esp]
+ add eax,DWORD [esi]
+ add ebx,DWORD [4+esi]
+ add edi,DWORD [8+esi]
+ add ecx,DWORD [12+esi]
+ mov DWORD [esi],eax
+ mov DWORD [4+esi],ebx
+ mov DWORD [8+esi],edi
+ mov DWORD [12+esi],ecx
+ mov eax,DWORD [24+esp]
+ mov ebx,DWORD [28+esp]
+ mov ecx,DWORD [32+esp]
+ mov edi,DWORD [360+esp]
+ add edx,DWORD [16+esi]
+ add eax,DWORD [20+esi]
+ add ebx,DWORD [24+esi]
+ add ecx,DWORD [28+esi]
+ mov DWORD [16+esi],edx
+ mov DWORD [20+esi],eax
+ mov DWORD [24+esi],ebx
+ mov DWORD [28+esi],ecx
+ lea esp,[356+esp]
+ sub ebp,256
+ cmp edi,DWORD [8+esp]
+ jb NEAR L$005loop_shrd
+ mov esp,DWORD [12+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 64
+L$001K256:
+dd 1116352408,1899447441,3049323471,3921009573,961987163,1508970993,2453635748,2870763221,3624381080,310598401,607225278,1426881987,1925078388,2162078206,2614888103,3248222580,3835390401,4022224774,264347078,604807628,770255983,1249150122,1555081692,1996064986,2554220882,2821834349,2952996808,3210313671,3336571891,3584528711,113926993,338241895,666307205,773529912,1294757372,1396182291,1695183700,1986661051,2177026350,2456956037,2730485921,2820302411,3259730800,3345764771,3516065817,3600352804,4094571909,275423344,430227734,506948616,659060556,883997877,958139571,1322822218,1537002063,1747873779,1955562222,2024104815,2227730452,2361852424,2428436474,2756734187,3204031479,3329325298
+dd 66051,67438087,134810123,202182159
+db 83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97
+db 110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+db 67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+db 112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+db 62,0
+align 16
+L$008unrolled:
+ lea esp,[esp-96]
+ mov eax,DWORD [esi]
+ mov ebp,DWORD [4+esi]
+ mov ecx,DWORD [8+esi]
+ mov ebx,DWORD [12+esi]
+ mov DWORD [4+esp],ebp
+ xor ebp,ecx
+ mov DWORD [8+esp],ecx
+ mov DWORD [12+esp],ebx
+ mov edx,DWORD [16+esi]
+ mov ebx,DWORD [20+esi]
+ mov ecx,DWORD [24+esi]
+ mov esi,DWORD [28+esi]
+ mov DWORD [20+esp],ebx
+ mov DWORD [24+esp],ecx
+ mov DWORD [28+esp],esi
+ jmp NEAR L$009grand_loop
+align 16
+L$009grand_loop:
+ mov ebx,DWORD [edi]
+ mov ecx,DWORD [4+edi]
+ bswap ebx
+ mov esi,DWORD [8+edi]
+ bswap ecx
+ mov DWORD [32+esp],ebx
+ bswap esi
+ mov DWORD [36+esp],ecx
+ mov DWORD [40+esp],esi
+ mov ebx,DWORD [12+edi]
+ mov ecx,DWORD [16+edi]
+ bswap ebx
+ mov esi,DWORD [20+edi]
+ bswap ecx
+ mov DWORD [44+esp],ebx
+ bswap esi
+ mov DWORD [48+esp],ecx
+ mov DWORD [52+esp],esi
+ mov ebx,DWORD [24+edi]
+ mov ecx,DWORD [28+edi]
+ bswap ebx
+ mov esi,DWORD [32+edi]
+ bswap ecx
+ mov DWORD [56+esp],ebx
+ bswap esi
+ mov DWORD [60+esp],ecx
+ mov DWORD [64+esp],esi
+ mov ebx,DWORD [36+edi]
+ mov ecx,DWORD [40+edi]
+ bswap ebx
+ mov esi,DWORD [44+edi]
+ bswap ecx
+ mov DWORD [68+esp],ebx
+ bswap esi
+ mov DWORD [72+esp],ecx
+ mov DWORD [76+esp],esi
+ mov ebx,DWORD [48+edi]
+ mov ecx,DWORD [52+edi]
+ bswap ebx
+ mov esi,DWORD [56+edi]
+ bswap ecx
+ mov DWORD [80+esp],ebx
+ bswap esi
+ mov DWORD [84+esp],ecx
+ mov DWORD [88+esp],esi
+ mov ebx,DWORD [60+edi]
+ add edi,64
+ bswap ebx
+ mov DWORD [100+esp],edi
+ mov DWORD [92+esp],ebx
+ mov ecx,edx
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov ebx,DWORD [32+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1116352408+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov ebx,DWORD [36+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1899447441+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov ebx,DWORD [40+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3049323471+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov ebx,DWORD [44+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3921009573+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov ebx,DWORD [48+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[961987163+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov ebx,DWORD [52+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1508970993+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov ebx,DWORD [56+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2453635748+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov ebx,DWORD [60+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2870763221+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov ebx,DWORD [64+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3624381080+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov ebx,DWORD [68+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[310598401+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov ebx,DWORD [72+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[607225278+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov ebx,DWORD [76+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1426881987+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov ebx,DWORD [80+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1925078388+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov ebx,DWORD [84+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2162078206+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov ecx,edx
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov ebx,DWORD [88+esp]
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2614888103+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov esi,edx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov ebx,DWORD [92+esp]
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3248222580+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [36+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [88+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [32+esp]
+ shr edi,10
+ add ebx,DWORD [68+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov DWORD [32+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3835390401+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [40+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov ecx,DWORD [92+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [36+esp]
+ shr edi,10
+ add ebx,DWORD [72+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov DWORD [36+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[4022224774+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [44+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov esi,DWORD [32+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [40+esp]
+ shr edi,10
+ add ebx,DWORD [76+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov DWORD [40+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[264347078+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [48+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov ecx,DWORD [36+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [44+esp]
+ shr edi,10
+ add ebx,DWORD [80+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov DWORD [44+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[604807628+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [52+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov esi,DWORD [40+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [48+esp]
+ shr edi,10
+ add ebx,DWORD [84+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov DWORD [48+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[770255983+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [56+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov ecx,DWORD [44+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [52+esp]
+ shr edi,10
+ add ebx,DWORD [88+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov DWORD [52+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1249150122+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [60+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov esi,DWORD [48+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [56+esp]
+ shr edi,10
+ add ebx,DWORD [92+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov DWORD [56+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1555081692+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [64+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov ecx,DWORD [52+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [60+esp]
+ shr edi,10
+ add ebx,DWORD [32+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov DWORD [60+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1996064986+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [68+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [56+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [64+esp]
+ shr edi,10
+ add ebx,DWORD [36+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov DWORD [64+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2554220882+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [72+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov ecx,DWORD [60+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [68+esp]
+ shr edi,10
+ add ebx,DWORD [40+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov DWORD [68+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2821834349+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [76+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov esi,DWORD [64+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [72+esp]
+ shr edi,10
+ add ebx,DWORD [44+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov DWORD [72+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2952996808+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [80+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov ecx,DWORD [68+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [76+esp]
+ shr edi,10
+ add ebx,DWORD [48+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov DWORD [76+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3210313671+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [84+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov esi,DWORD [72+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [80+esp]
+ shr edi,10
+ add ebx,DWORD [52+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov DWORD [80+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3336571891+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [88+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov ecx,DWORD [76+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [84+esp]
+ shr edi,10
+ add ebx,DWORD [56+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov DWORD [84+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3584528711+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [92+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov esi,DWORD [80+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [88+esp]
+ shr edi,10
+ add ebx,DWORD [60+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov DWORD [88+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[113926993+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [32+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov ecx,DWORD [84+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [92+esp]
+ shr edi,10
+ add ebx,DWORD [64+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov DWORD [92+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[338241895+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [36+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [88+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [32+esp]
+ shr edi,10
+ add ebx,DWORD [68+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov DWORD [32+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[666307205+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [40+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov ecx,DWORD [92+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [36+esp]
+ shr edi,10
+ add ebx,DWORD [72+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov DWORD [36+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[773529912+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [44+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov esi,DWORD [32+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [40+esp]
+ shr edi,10
+ add ebx,DWORD [76+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov DWORD [40+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1294757372+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [48+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov ecx,DWORD [36+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [44+esp]
+ shr edi,10
+ add ebx,DWORD [80+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov DWORD [44+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1396182291+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [52+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov esi,DWORD [40+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [48+esp]
+ shr edi,10
+ add ebx,DWORD [84+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov DWORD [48+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1695183700+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [56+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov ecx,DWORD [44+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [52+esp]
+ shr edi,10
+ add ebx,DWORD [88+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov DWORD [52+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1986661051+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [60+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov esi,DWORD [48+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [56+esp]
+ shr edi,10
+ add ebx,DWORD [92+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov DWORD [56+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2177026350+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [64+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov ecx,DWORD [52+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [60+esp]
+ shr edi,10
+ add ebx,DWORD [32+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov DWORD [60+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2456956037+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [68+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [56+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [64+esp]
+ shr edi,10
+ add ebx,DWORD [36+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov DWORD [64+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2730485921+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [72+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov ecx,DWORD [60+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [68+esp]
+ shr edi,10
+ add ebx,DWORD [40+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov DWORD [68+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2820302411+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [76+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov esi,DWORD [64+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [72+esp]
+ shr edi,10
+ add ebx,DWORD [44+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov DWORD [72+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3259730800+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [80+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov ecx,DWORD [68+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [76+esp]
+ shr edi,10
+ add ebx,DWORD [48+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov DWORD [76+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3345764771+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [84+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov esi,DWORD [72+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [80+esp]
+ shr edi,10
+ add ebx,DWORD [52+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov DWORD [80+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3516065817+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [88+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov ecx,DWORD [76+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [84+esp]
+ shr edi,10
+ add ebx,DWORD [56+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov DWORD [84+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3600352804+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [92+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov esi,DWORD [80+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [88+esp]
+ shr edi,10
+ add ebx,DWORD [60+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov DWORD [88+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[4094571909+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [32+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov ecx,DWORD [84+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [92+esp]
+ shr edi,10
+ add ebx,DWORD [64+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov DWORD [92+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[275423344+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [36+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [88+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [32+esp]
+ shr edi,10
+ add ebx,DWORD [68+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov DWORD [32+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[430227734+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [40+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov ecx,DWORD [92+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [36+esp]
+ shr edi,10
+ add ebx,DWORD [72+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov DWORD [36+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[506948616+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [44+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov esi,DWORD [32+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [40+esp]
+ shr edi,10
+ add ebx,DWORD [76+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov DWORD [40+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[659060556+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [48+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov ecx,DWORD [36+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [44+esp]
+ shr edi,10
+ add ebx,DWORD [80+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov DWORD [44+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[883997877+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [52+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov esi,DWORD [40+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [48+esp]
+ shr edi,10
+ add ebx,DWORD [84+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov DWORD [48+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[958139571+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [56+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov ecx,DWORD [44+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [52+esp]
+ shr edi,10
+ add ebx,DWORD [88+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov DWORD [52+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1322822218+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [60+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov esi,DWORD [48+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [56+esp]
+ shr edi,10
+ add ebx,DWORD [92+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ mov DWORD [56+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1537002063+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [64+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov ecx,DWORD [52+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [60+esp]
+ shr edi,10
+ add ebx,DWORD [32+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ mov DWORD [60+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[1747873779+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [68+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [56+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [64+esp]
+ shr edi,10
+ add ebx,DWORD [36+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [20+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [24+esp]
+ xor edx,ecx
+ mov DWORD [64+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [16+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [28+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [4+esp]
+ xor ecx,eax
+ mov DWORD [esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[1955562222+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [72+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [12+esp]
+ add ebp,ecx
+ mov ecx,DWORD [60+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [68+esp]
+ shr edi,10
+ add ebx,DWORD [40+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [16+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [20+esp]
+ xor edx,esi
+ mov DWORD [68+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [12+esp],esi
+ xor edx,esi
+ add ebx,DWORD [24+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [esp]
+ xor esi,ebp
+ mov DWORD [28+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2024104815+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [76+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [8+esp]
+ add eax,esi
+ mov esi,DWORD [64+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [72+esp]
+ shr edi,10
+ add ebx,DWORD [44+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [12+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [16+esp]
+ xor edx,ecx
+ mov DWORD [72+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [8+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [20+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [28+esp]
+ xor ecx,eax
+ mov DWORD [24+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2227730452+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [80+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [4+esp]
+ add ebp,ecx
+ mov ecx,DWORD [68+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [76+esp]
+ shr edi,10
+ add ebx,DWORD [48+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [8+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [12+esp]
+ xor edx,esi
+ mov DWORD [76+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [4+esp],esi
+ xor edx,esi
+ add ebx,DWORD [16+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [24+esp]
+ xor esi,ebp
+ mov DWORD [20+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2361852424+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [84+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [esp]
+ add eax,esi
+ mov esi,DWORD [72+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [80+esp]
+ shr edi,10
+ add ebx,DWORD [52+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [4+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [8+esp]
+ xor edx,ecx
+ mov DWORD [80+esp],ebx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [12+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [20+esp]
+ xor ecx,eax
+ mov DWORD [16+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[2428436474+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [88+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [28+esp]
+ add ebp,ecx
+ mov ecx,DWORD [76+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [84+esp]
+ shr edi,10
+ add ebx,DWORD [56+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [4+esp]
+ xor edx,esi
+ mov DWORD [84+esp],ebx
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [28+esp],esi
+ xor edx,esi
+ add ebx,DWORD [8+esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [16+esp]
+ xor esi,ebp
+ mov DWORD [12+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[2756734187+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ mov ecx,DWORD [92+esp]
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [24+esp]
+ add eax,esi
+ mov esi,DWORD [80+esp]
+ mov ebx,ecx
+ ror ecx,11
+ mov edi,esi
+ ror esi,2
+ xor ecx,ebx
+ shr ebx,3
+ ror ecx,7
+ xor esi,edi
+ xor ebx,ecx
+ ror esi,17
+ add ebx,DWORD [88+esp]
+ shr edi,10
+ add ebx,DWORD [60+esp]
+ mov ecx,edx
+ xor edi,esi
+ mov esi,DWORD [28+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [esp]
+ xor edx,ecx
+ xor esi,edi
+ ror edx,5
+ and esi,ecx
+ mov DWORD [24+esp],ecx
+ xor edx,ecx
+ add ebx,DWORD [4+esp]
+ xor edi,esi
+ ror edx,6
+ mov ecx,eax
+ add ebx,edi
+ ror ecx,9
+ mov esi,eax
+ mov edi,DWORD [12+esp]
+ xor ecx,eax
+ mov DWORD [8+esp],eax
+ xor eax,edi
+ ror ecx,11
+ and ebp,eax
+ lea edx,[3204031479+edx*1+ebx]
+ xor ecx,esi
+ xor ebp,edi
+ mov esi,DWORD [32+esp]
+ ror ecx,2
+ add ebp,edx
+ add edx,DWORD [20+esp]
+ add ebp,ecx
+ mov ecx,DWORD [84+esp]
+ mov ebx,esi
+ ror esi,11
+ mov edi,ecx
+ ror ecx,2
+ xor esi,ebx
+ shr ebx,3
+ ror esi,7
+ xor ecx,edi
+ xor ebx,esi
+ ror ecx,17
+ add ebx,DWORD [92+esp]
+ shr edi,10
+ add ebx,DWORD [64+esp]
+ mov esi,edx
+ xor edi,ecx
+ mov ecx,DWORD [24+esp]
+ ror edx,14
+ add ebx,edi
+ mov edi,DWORD [28+esp]
+ xor edx,esi
+ xor ecx,edi
+ ror edx,5
+ and ecx,esi
+ mov DWORD [20+esp],esi
+ xor edx,esi
+ add ebx,DWORD [esp]
+ xor edi,ecx
+ ror edx,6
+ mov esi,ebp
+ add ebx,edi
+ ror esi,9
+ mov ecx,ebp
+ mov edi,DWORD [8+esp]
+ xor esi,ebp
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ ror esi,11
+ and eax,ebp
+ lea edx,[3329325298+edx*1+ebx]
+ xor esi,ecx
+ xor eax,edi
+ ror esi,2
+ add eax,edx
+ add edx,DWORD [16+esp]
+ add eax,esi
+ mov esi,DWORD [96+esp]
+ xor ebp,edi
+ mov ecx,DWORD [12+esp]
+ add eax,DWORD [esi]
+ add ebp,DWORD [4+esi]
+ add edi,DWORD [8+esi]
+ add ecx,DWORD [12+esi]
+ mov DWORD [esi],eax
+ mov DWORD [4+esi],ebp
+ mov DWORD [8+esi],edi
+ mov DWORD [12+esi],ecx
+ mov DWORD [4+esp],ebp
+ xor ebp,edi
+ mov DWORD [8+esp],edi
+ mov DWORD [12+esp],ecx
+ mov edi,DWORD [20+esp]
+ mov ebx,DWORD [24+esp]
+ mov ecx,DWORD [28+esp]
+ add edx,DWORD [16+esi]
+ add edi,DWORD [20+esi]
+ add ebx,DWORD [24+esi]
+ add ecx,DWORD [28+esi]
+ mov DWORD [16+esi],edx
+ mov DWORD [20+esi],edi
+ mov DWORD [24+esi],ebx
+ mov DWORD [28+esi],ecx
+ mov DWORD [20+esp],edi
+ mov edi,DWORD [100+esp]
+ mov DWORD [24+esp],ebx
+ mov DWORD [28+esp],ecx
+ cmp edi,DWORD [104+esp]
+ jb NEAR L$009grand_loop
+ mov esp,DWORD [108+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+segment .bss
+common _OPENSSL_ia32cap_P 16
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha512-586.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha512-586.nasm
new file mode 100644
index 0000000000..a168b4a799
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/sha/sha512-586.nasm
@@ -0,0 +1,579 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/sha/asm/sha512-586.pl
+;
+; Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+global _sha512_block_data_order
+align 16
+_sha512_block_data_order:
+L$_sha512_block_data_order_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov esi,DWORD [20+esp]
+ mov edi,DWORD [24+esp]
+ mov eax,DWORD [28+esp]
+ mov ebx,esp
+ call L$000pic_point
+L$000pic_point:
+ pop ebp
+ lea ebp,[(L$001K512-L$000pic_point)+ebp]
+ sub esp,16
+ and esp,-64
+ shl eax,7
+ add eax,edi
+ mov DWORD [esp],esi
+ mov DWORD [4+esp],edi
+ mov DWORD [8+esp],eax
+ mov DWORD [12+esp],ebx
+align 16
+L$002loop_x86:
+ mov eax,DWORD [edi]
+ mov ebx,DWORD [4+edi]
+ mov ecx,DWORD [8+edi]
+ mov edx,DWORD [12+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [16+edi]
+ mov ebx,DWORD [20+edi]
+ mov ecx,DWORD [24+edi]
+ mov edx,DWORD [28+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [32+edi]
+ mov ebx,DWORD [36+edi]
+ mov ecx,DWORD [40+edi]
+ mov edx,DWORD [44+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [48+edi]
+ mov ebx,DWORD [52+edi]
+ mov ecx,DWORD [56+edi]
+ mov edx,DWORD [60+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [64+edi]
+ mov ebx,DWORD [68+edi]
+ mov ecx,DWORD [72+edi]
+ mov edx,DWORD [76+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [80+edi]
+ mov ebx,DWORD [84+edi]
+ mov ecx,DWORD [88+edi]
+ mov edx,DWORD [92+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [96+edi]
+ mov ebx,DWORD [100+edi]
+ mov ecx,DWORD [104+edi]
+ mov edx,DWORD [108+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ mov eax,DWORD [112+edi]
+ mov ebx,DWORD [116+edi]
+ mov ecx,DWORD [120+edi]
+ mov edx,DWORD [124+edi]
+ bswap eax
+ bswap ebx
+ bswap ecx
+ bswap edx
+ push eax
+ push ebx
+ push ecx
+ push edx
+ add edi,128
+ sub esp,72
+ mov DWORD [204+esp],edi
+ lea edi,[8+esp]
+ mov ecx,16
+dd 2784229001
+align 16
+L$00300_15_x86:
+ mov ecx,DWORD [40+esp]
+ mov edx,DWORD [44+esp]
+ mov esi,ecx
+ shr ecx,9
+ mov edi,edx
+ shr edx,9
+ mov ebx,ecx
+ shl esi,14
+ mov eax,edx
+ shl edi,14
+ xor ebx,esi
+ shr ecx,5
+ xor eax,edi
+ shr edx,5
+ xor eax,ecx
+ shl esi,4
+ xor ebx,edx
+ shl edi,4
+ xor ebx,esi
+ shr ecx,4
+ xor eax,edi
+ shr edx,4
+ xor eax,ecx
+ shl esi,5
+ xor ebx,edx
+ shl edi,5
+ xor eax,esi
+ xor ebx,edi
+ mov ecx,DWORD [48+esp]
+ mov edx,DWORD [52+esp]
+ mov esi,DWORD [56+esp]
+ mov edi,DWORD [60+esp]
+ add eax,DWORD [64+esp]
+ adc ebx,DWORD [68+esp]
+ xor ecx,esi
+ xor edx,edi
+ and ecx,DWORD [40+esp]
+ and edx,DWORD [44+esp]
+ add eax,DWORD [192+esp]
+ adc ebx,DWORD [196+esp]
+ xor ecx,esi
+ xor edx,edi
+ mov esi,DWORD [ebp]
+ mov edi,DWORD [4+ebp]
+ add eax,ecx
+ adc ebx,edx
+ mov ecx,DWORD [32+esp]
+ mov edx,DWORD [36+esp]
+ add eax,esi
+ adc ebx,edi
+ mov DWORD [esp],eax
+ mov DWORD [4+esp],ebx
+ add eax,ecx
+ adc ebx,edx
+ mov ecx,DWORD [8+esp]
+ mov edx,DWORD [12+esp]
+ mov DWORD [32+esp],eax
+ mov DWORD [36+esp],ebx
+ mov esi,ecx
+ shr ecx,2
+ mov edi,edx
+ shr edx,2
+ mov ebx,ecx
+ shl esi,4
+ mov eax,edx
+ shl edi,4
+ xor ebx,esi
+ shr ecx,5
+ xor eax,edi
+ shr edx,5
+ xor ebx,ecx
+ shl esi,21
+ xor eax,edx
+ shl edi,21
+ xor eax,esi
+ shr ecx,21
+ xor ebx,edi
+ shr edx,21
+ xor eax,ecx
+ shl esi,5
+ xor ebx,edx
+ shl edi,5
+ xor eax,esi
+ xor ebx,edi
+ mov ecx,DWORD [8+esp]
+ mov edx,DWORD [12+esp]
+ mov esi,DWORD [16+esp]
+ mov edi,DWORD [20+esp]
+ add eax,DWORD [esp]
+ adc ebx,DWORD [4+esp]
+ or ecx,esi
+ or edx,edi
+ and ecx,DWORD [24+esp]
+ and edx,DWORD [28+esp]
+ and esi,DWORD [8+esp]
+ and edi,DWORD [12+esp]
+ or ecx,esi
+ or edx,edi
+ add eax,ecx
+ adc ebx,edx
+ mov DWORD [esp],eax
+ mov DWORD [4+esp],ebx
+ mov dl,BYTE [ebp]
+ sub esp,8
+ lea ebp,[8+ebp]
+ cmp dl,148
+ jne NEAR L$00300_15_x86
+align 16
+L$00416_79_x86:
+ mov ecx,DWORD [312+esp]
+ mov edx,DWORD [316+esp]
+ mov esi,ecx
+ shr ecx,1
+ mov edi,edx
+ shr edx,1
+ mov eax,ecx
+ shl esi,24
+ mov ebx,edx
+ shl edi,24
+ xor ebx,esi
+ shr ecx,6
+ xor eax,edi
+ shr edx,6
+ xor eax,ecx
+ shl esi,7
+ xor ebx,edx
+ shl edi,1
+ xor ebx,esi
+ shr ecx,1
+ xor eax,edi
+ shr edx,1
+ xor eax,ecx
+ shl edi,6
+ xor ebx,edx
+ xor eax,edi
+ mov DWORD [esp],eax
+ mov DWORD [4+esp],ebx
+ mov ecx,DWORD [208+esp]
+ mov edx,DWORD [212+esp]
+ mov esi,ecx
+ shr ecx,6
+ mov edi,edx
+ shr edx,6
+ mov eax,ecx
+ shl esi,3
+ mov ebx,edx
+ shl edi,3
+ xor eax,esi
+ shr ecx,13
+ xor ebx,edi
+ shr edx,13
+ xor eax,ecx
+ shl esi,10
+ xor ebx,edx
+ shl edi,10
+ xor ebx,esi
+ shr ecx,10
+ xor eax,edi
+ shr edx,10
+ xor ebx,ecx
+ shl edi,13
+ xor eax,edx
+ xor eax,edi
+ mov ecx,DWORD [320+esp]
+ mov edx,DWORD [324+esp]
+ add eax,DWORD [esp]
+ adc ebx,DWORD [4+esp]
+ mov esi,DWORD [248+esp]
+ mov edi,DWORD [252+esp]
+ add eax,ecx
+ adc ebx,edx
+ add eax,esi
+ adc ebx,edi
+ mov DWORD [192+esp],eax
+ mov DWORD [196+esp],ebx
+ mov ecx,DWORD [40+esp]
+ mov edx,DWORD [44+esp]
+ mov esi,ecx
+ shr ecx,9
+ mov edi,edx
+ shr edx,9
+ mov ebx,ecx
+ shl esi,14
+ mov eax,edx
+ shl edi,14
+ xor ebx,esi
+ shr ecx,5
+ xor eax,edi
+ shr edx,5
+ xor eax,ecx
+ shl esi,4
+ xor ebx,edx
+ shl edi,4
+ xor ebx,esi
+ shr ecx,4
+ xor eax,edi
+ shr edx,4
+ xor eax,ecx
+ shl esi,5
+ xor ebx,edx
+ shl edi,5
+ xor eax,esi
+ xor ebx,edi
+ mov ecx,DWORD [48+esp]
+ mov edx,DWORD [52+esp]
+ mov esi,DWORD [56+esp]
+ mov edi,DWORD [60+esp]
+ add eax,DWORD [64+esp]
+ adc ebx,DWORD [68+esp]
+ xor ecx,esi
+ xor edx,edi
+ and ecx,DWORD [40+esp]
+ and edx,DWORD [44+esp]
+ add eax,DWORD [192+esp]
+ adc ebx,DWORD [196+esp]
+ xor ecx,esi
+ xor edx,edi
+ mov esi,DWORD [ebp]
+ mov edi,DWORD [4+ebp]
+ add eax,ecx
+ adc ebx,edx
+ mov ecx,DWORD [32+esp]
+ mov edx,DWORD [36+esp]
+ add eax,esi
+ adc ebx,edi
+ mov DWORD [esp],eax
+ mov DWORD [4+esp],ebx
+ add eax,ecx
+ adc ebx,edx
+ mov ecx,DWORD [8+esp]
+ mov edx,DWORD [12+esp]
+ mov DWORD [32+esp],eax
+ mov DWORD [36+esp],ebx
+ mov esi,ecx
+ shr ecx,2
+ mov edi,edx
+ shr edx,2
+ mov ebx,ecx
+ shl esi,4
+ mov eax,edx
+ shl edi,4
+ xor ebx,esi
+ shr ecx,5
+ xor eax,edi
+ shr edx,5
+ xor ebx,ecx
+ shl esi,21
+ xor eax,edx
+ shl edi,21
+ xor eax,esi
+ shr ecx,21
+ xor ebx,edi
+ shr edx,21
+ xor eax,ecx
+ shl esi,5
+ xor ebx,edx
+ shl edi,5
+ xor eax,esi
+ xor ebx,edi
+ mov ecx,DWORD [8+esp]
+ mov edx,DWORD [12+esp]
+ mov esi,DWORD [16+esp]
+ mov edi,DWORD [20+esp]
+ add eax,DWORD [esp]
+ adc ebx,DWORD [4+esp]
+ or ecx,esi
+ or edx,edi
+ and ecx,DWORD [24+esp]
+ and edx,DWORD [28+esp]
+ and esi,DWORD [8+esp]
+ and edi,DWORD [12+esp]
+ or ecx,esi
+ or edx,edi
+ add eax,ecx
+ adc ebx,edx
+ mov DWORD [esp],eax
+ mov DWORD [4+esp],ebx
+ mov dl,BYTE [ebp]
+ sub esp,8
+ lea ebp,[8+ebp]
+ cmp dl,23
+ jne NEAR L$00416_79_x86
+ mov esi,DWORD [840+esp]
+ mov edi,DWORD [844+esp]
+ mov eax,DWORD [esi]
+ mov ebx,DWORD [4+esi]
+ mov ecx,DWORD [8+esi]
+ mov edx,DWORD [12+esi]
+ add eax,DWORD [8+esp]
+ adc ebx,DWORD [12+esp]
+ mov DWORD [esi],eax
+ mov DWORD [4+esi],ebx
+ add ecx,DWORD [16+esp]
+ adc edx,DWORD [20+esp]
+ mov DWORD [8+esi],ecx
+ mov DWORD [12+esi],edx
+ mov eax,DWORD [16+esi]
+ mov ebx,DWORD [20+esi]
+ mov ecx,DWORD [24+esi]
+ mov edx,DWORD [28+esi]
+ add eax,DWORD [24+esp]
+ adc ebx,DWORD [28+esp]
+ mov DWORD [16+esi],eax
+ mov DWORD [20+esi],ebx
+ add ecx,DWORD [32+esp]
+ adc edx,DWORD [36+esp]
+ mov DWORD [24+esi],ecx
+ mov DWORD [28+esi],edx
+ mov eax,DWORD [32+esi]
+ mov ebx,DWORD [36+esi]
+ mov ecx,DWORD [40+esi]
+ mov edx,DWORD [44+esi]
+ add eax,DWORD [40+esp]
+ adc ebx,DWORD [44+esp]
+ mov DWORD [32+esi],eax
+ mov DWORD [36+esi],ebx
+ add ecx,DWORD [48+esp]
+ adc edx,DWORD [52+esp]
+ mov DWORD [40+esi],ecx
+ mov DWORD [44+esi],edx
+ mov eax,DWORD [48+esi]
+ mov ebx,DWORD [52+esi]
+ mov ecx,DWORD [56+esi]
+ mov edx,DWORD [60+esi]
+ add eax,DWORD [56+esp]
+ adc ebx,DWORD [60+esp]
+ mov DWORD [48+esi],eax
+ mov DWORD [52+esi],ebx
+ add ecx,DWORD [64+esp]
+ adc edx,DWORD [68+esp]
+ mov DWORD [56+esi],ecx
+ mov DWORD [60+esi],edx
+ add esp,840
+ sub ebp,640
+ cmp edi,DWORD [8+esp]
+ jb NEAR L$002loop_x86
+ mov esp,DWORD [12+esp]
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 64
+L$001K512:
+dd 3609767458,1116352408
+dd 602891725,1899447441
+dd 3964484399,3049323471
+dd 2173295548,3921009573
+dd 4081628472,961987163
+dd 3053834265,1508970993
+dd 2937671579,2453635748
+dd 3664609560,2870763221
+dd 2734883394,3624381080
+dd 1164996542,310598401
+dd 1323610764,607225278
+dd 3590304994,1426881987
+dd 4068182383,1925078388
+dd 991336113,2162078206
+dd 633803317,2614888103
+dd 3479774868,3248222580
+dd 2666613458,3835390401
+dd 944711139,4022224774
+dd 2341262773,264347078
+dd 2007800933,604807628
+dd 1495990901,770255983
+dd 1856431235,1249150122
+dd 3175218132,1555081692
+dd 2198950837,1996064986
+dd 3999719339,2554220882
+dd 766784016,2821834349
+dd 2566594879,2952996808
+dd 3203337956,3210313671
+dd 1034457026,3336571891
+dd 2466948901,3584528711
+dd 3758326383,113926993
+dd 168717936,338241895
+dd 1188179964,666307205
+dd 1546045734,773529912
+dd 1522805485,1294757372
+dd 2643833823,1396182291
+dd 2343527390,1695183700
+dd 1014477480,1986661051
+dd 1206759142,2177026350
+dd 344077627,2456956037
+dd 1290863460,2730485921
+dd 3158454273,2820302411
+dd 3505952657,3259730800
+dd 106217008,3345764771
+dd 3606008344,3516065817
+dd 1432725776,3600352804
+dd 1467031594,4094571909
+dd 851169720,275423344
+dd 3100823752,430227734
+dd 1363258195,506948616
+dd 3750685593,659060556
+dd 3785050280,883997877
+dd 3318307427,958139571
+dd 3812723403,1322822218
+dd 2003034995,1537002063
+dd 3602036899,1747873779
+dd 1575990012,1955562222
+dd 1125592928,2024104815
+dd 2716904306,2227730452
+dd 442776044,2361852424
+dd 593698344,2428436474
+dd 3733110249,2756734187
+dd 2999351573,3204031479
+dd 3815920427,3329325298
+dd 3928383900,3391569614
+dd 566280711,3515267271
+dd 3454069534,3940187606
+dd 4000239992,4118630271
+dd 1914138554,116418474
+dd 2731055270,174292421
+dd 3203993006,289380356
+dd 320620315,460393269
+dd 587496836,685471733
+dd 1086792851,852142971
+dd 365543100,1017036298
+dd 2618297676,1126000580
+dd 3409855158,1288033470
+dd 4234509866,1501505948
+dd 987167468,1607167915
+dd 1246189591,1816402316
+dd 67438087,66051
+dd 202182159,134810123
+db 83,72,65,53,49,50,32,98,108,111,99,107,32,116,114,97
+db 110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+db 67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+db 112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+db 62,0
diff --git a/CryptoPkg/Library/OpensslLib/IA32/crypto/x86cpuid.nasm b/CryptoPkg/Library/OpensslLib/IA32/crypto/x86cpuid.nasm
new file mode 100644
index 0000000000..542e759ecd
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32/crypto/x86cpuid.nasm
@@ -0,0 +1,433 @@
+; WARNING: do not edit!
+; Generated from openssl/crypto/x86cpuid.pl
+;
+; Copyright 2004-2020 The OpenSSL Project Authors. All Rights Reserved.
+;
+; Licensed under the OpenSSL license (the "License"). You may not use
+; this file except in compliance with the License. You can obtain a copy
+; in the file LICENSE in the source distribution or at
+; https://www.openssl.org/source/license.html
+
+%ifidn __OUTPUT_FORMAT__,obj
+section code use32 class=code align=64
+%elifidn __OUTPUT_FORMAT__,win32
+$@feat.00 equ 1
+section .text code align=64
+%else
+section .text code
+%endif
+global _OPENSSL_ia32_cpuid
+align 16
+_OPENSSL_ia32_cpuid:
+L$_OPENSSL_ia32_cpuid_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ xor edx,edx
+ pushfd
+ pop eax
+ mov ecx,eax
+ xor eax,2097152
+ push eax
+ popfd
+ pushfd
+ pop eax
+ xor ecx,eax
+ xor eax,eax
+ mov esi,DWORD [20+esp]
+ mov DWORD [8+esi],eax
+ bt ecx,21
+ jnc NEAR L$000nocpuid
+ cpuid
+ mov edi,eax
+ xor eax,eax
+ cmp ebx,1970169159
+ setne al
+ mov ebp,eax
+ cmp edx,1231384169
+ setne al
+ or ebp,eax
+ cmp ecx,1818588270
+ setne al
+ or ebp,eax
+ jz NEAR L$001intel
+ cmp ebx,1752462657
+ setne al
+ mov esi,eax
+ cmp edx,1769238117
+ setne al
+ or esi,eax
+ cmp ecx,1145913699
+ setne al
+ or esi,eax
+ jnz NEAR L$001intel
+ mov eax,2147483648
+ cpuid
+ cmp eax,2147483649
+ jb NEAR L$001intel
+ mov esi,eax
+ mov eax,2147483649
+ cpuid
+ or ebp,ecx
+ and ebp,2049
+ cmp esi,2147483656
+ jb NEAR L$001intel
+ mov eax,2147483656
+ cpuid
+ movzx esi,cl
+ inc esi
+ mov eax,1
+ xor ecx,ecx
+ cpuid
+ bt edx,28
+ jnc NEAR L$002generic
+ shr ebx,16
+ and ebx,255
+ cmp ebx,esi
+ ja NEAR L$002generic
+ and edx,4026531839
+ jmp NEAR L$002generic
+L$001intel:
+ cmp edi,4
+ mov esi,-1
+ jb NEAR L$003nocacheinfo
+ mov eax,4
+ mov ecx,0
+ cpuid
+ mov esi,eax
+ shr esi,14
+ and esi,4095
+L$003nocacheinfo:
+ mov eax,1
+ xor ecx,ecx
+ cpuid
+ and edx,3220176895
+ cmp ebp,0
+ jne NEAR L$004notintel
+ or edx,1073741824
+ and ah,15
+ cmp ah,15
+ jne NEAR L$004notintel
+ or edx,1048576
+L$004notintel:
+ bt edx,28
+ jnc NEAR L$002generic
+ and edx,4026531839
+ cmp esi,0
+ je NEAR L$002generic
+ or edx,268435456
+ shr ebx,16
+ cmp bl,1
+ ja NEAR L$002generic
+ and edx,4026531839
+L$002generic:
+ and ebp,2048
+ and ecx,4294965247
+ mov esi,edx
+ or ebp,ecx
+ cmp edi,7
+ mov edi,DWORD [20+esp]
+ jb NEAR L$005no_extended_info
+ mov eax,7
+ xor ecx,ecx
+ cpuid
+ mov DWORD [8+edi],ebx
+L$005no_extended_info:
+ bt ebp,27
+ jnc NEAR L$006clear_avx
+ xor ecx,ecx
+db 15,1,208
+ and eax,6
+ cmp eax,6
+ je NEAR L$007done
+ cmp eax,2
+ je NEAR L$006clear_avx
+L$008clear_xmm:
+ and ebp,4261412861
+ and esi,4278190079
+L$006clear_avx:
+ and ebp,4026525695
+ and DWORD [8+edi],4294967263
+L$007done:
+ mov eax,esi
+ mov edx,ebp
+L$000nocpuid:
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+;extern _OPENSSL_ia32cap_P
+global _OPENSSL_rdtsc
+align 16
+_OPENSSL_rdtsc:
+L$_OPENSSL_rdtsc_begin:
+ xor eax,eax
+ xor edx,edx
+ lea ecx,[_OPENSSL_ia32cap_P]
+ bt DWORD [ecx],4
+ jnc NEAR L$009notsc
+ rdtsc
+L$009notsc:
+ ret
+global _OPENSSL_instrument_halt
+align 16
+_OPENSSL_instrument_halt:
+L$_OPENSSL_instrument_halt_begin:
+ lea ecx,[_OPENSSL_ia32cap_P]
+ bt DWORD [ecx],4
+ jnc NEAR L$010nohalt
+dd 2421723150
+ and eax,3
+ jnz NEAR L$010nohalt
+ pushfd
+ pop eax
+ bt eax,9
+ jnc NEAR L$010nohalt
+ rdtsc
+ push edx
+ push eax
+ hlt
+ rdtsc
+ sub eax,DWORD [esp]
+ sbb edx,DWORD [4+esp]
+ add esp,8
+ ret
+L$010nohalt:
+ xor eax,eax
+ xor edx,edx
+ ret
+global _OPENSSL_far_spin
+align 16
+_OPENSSL_far_spin:
+L$_OPENSSL_far_spin_begin:
+ pushfd
+ pop eax
+ bt eax,9
+ jnc NEAR L$011nospin
+ mov eax,DWORD [4+esp]
+ mov ecx,DWORD [8+esp]
+dd 2430111262
+ xor eax,eax
+ mov edx,DWORD [ecx]
+ jmp NEAR L$012spin
+align 16
+L$012spin:
+ inc eax
+ cmp edx,DWORD [ecx]
+ je NEAR L$012spin
+dd 529567888
+ ret
+L$011nospin:
+ xor eax,eax
+ xor edx,edx
+ ret
+global _OPENSSL_wipe_cpu
+align 16
+_OPENSSL_wipe_cpu:
+L$_OPENSSL_wipe_cpu_begin:
+ xor eax,eax
+ xor edx,edx
+ lea ecx,[_OPENSSL_ia32cap_P]
+ mov ecx,DWORD [ecx]
+ bt DWORD [ecx],1
+ jnc NEAR L$013no_x87
+dd 4007259865,4007259865,4007259865,4007259865,2430851995
+L$013no_x87:
+ lea eax,[4+esp]
+ ret
+global _OPENSSL_atomic_add
+align 16
+_OPENSSL_atomic_add:
+L$_OPENSSL_atomic_add_begin:
+ mov edx,DWORD [4+esp]
+ mov ecx,DWORD [8+esp]
+ push ebx
+ nop
+ mov eax,DWORD [edx]
+L$014spin:
+ lea ebx,[ecx*1+eax]
+ nop
+dd 447811568
+ jne NEAR L$014spin
+ mov eax,ebx
+ pop ebx
+ ret
+global _OPENSSL_cleanse
+align 16
+_OPENSSL_cleanse:
+L$_OPENSSL_cleanse_begin:
+ mov edx,DWORD [4+esp]
+ mov ecx,DWORD [8+esp]
+ xor eax,eax
+ cmp ecx,7
+ jae NEAR L$015lot
+ cmp ecx,0
+ je NEAR L$016ret
+L$017little:
+ mov BYTE [edx],al
+ sub ecx,1
+ lea edx,[1+edx]
+ jnz NEAR L$017little
+L$016ret:
+ ret
+align 16
+L$015lot:
+ test edx,3
+ jz NEAR L$018aligned
+ mov BYTE [edx],al
+ lea ecx,[ecx-1]
+ lea edx,[1+edx]
+ jmp NEAR L$015lot
+L$018aligned:
+ mov DWORD [edx],eax
+ lea ecx,[ecx-4]
+ test ecx,-4
+ lea edx,[4+edx]
+ jnz NEAR L$018aligned
+ cmp ecx,0
+ jne NEAR L$017little
+ ret
+global _CRYPTO_memcmp
+align 16
+_CRYPTO_memcmp:
+L$_CRYPTO_memcmp_begin:
+ push esi
+ push edi
+ mov esi,DWORD [12+esp]
+ mov edi,DWORD [16+esp]
+ mov ecx,DWORD [20+esp]
+ xor eax,eax
+ xor edx,edx
+ cmp ecx,0
+ je NEAR L$019no_data
+L$020loop:
+ mov dl,BYTE [esi]
+ lea esi,[1+esi]
+ xor dl,BYTE [edi]
+ lea edi,[1+edi]
+ or al,dl
+ dec ecx
+ jnz NEAR L$020loop
+ neg eax
+ shr eax,31
+L$019no_data:
+ pop edi
+ pop esi
+ ret
+global _OPENSSL_instrument_bus
+align 16
+_OPENSSL_instrument_bus:
+L$_OPENSSL_instrument_bus_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov eax,0
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _OPENSSL_instrument_bus2
+align 16
+_OPENSSL_instrument_bus2:
+L$_OPENSSL_instrument_bus2_begin:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ mov eax,0
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+global _OPENSSL_ia32_rdrand_bytes
+align 16
+_OPENSSL_ia32_rdrand_bytes:
+L$_OPENSSL_ia32_rdrand_bytes_begin:
+ push edi
+ push ebx
+ xor eax,eax
+ mov edi,DWORD [12+esp]
+ mov ebx,DWORD [16+esp]
+ cmp ebx,0
+ je NEAR L$021done
+ mov ecx,8
+L$022loop:
+db 15,199,242
+ jc NEAR L$023break
+ loop L$022loop
+ jmp NEAR L$021done
+align 16
+L$023break:
+ cmp ebx,4
+ jb NEAR L$024tail
+ mov DWORD [edi],edx
+ lea edi,[4+edi]
+ add eax,4
+ sub ebx,4
+ jz NEAR L$021done
+ mov ecx,8
+ jmp NEAR L$022loop
+align 16
+L$024tail:
+ mov BYTE [edi],dl
+ lea edi,[1+edi]
+ inc eax
+ shr edx,8
+ dec ebx
+ jnz NEAR L$024tail
+L$021done:
+ xor edx,edx
+ pop ebx
+ pop edi
+ ret
+global _OPENSSL_ia32_rdseed_bytes
+align 16
+_OPENSSL_ia32_rdseed_bytes:
+L$_OPENSSL_ia32_rdseed_bytes_begin:
+ push edi
+ push ebx
+ xor eax,eax
+ mov edi,DWORD [12+esp]
+ mov ebx,DWORD [16+esp]
+ cmp ebx,0
+ je NEAR L$025done
+ mov ecx,8
+L$026loop:
+db 15,199,250
+ jc NEAR L$027break
+ loop L$026loop
+ jmp NEAR L$025done
+align 16
+L$027break:
+ cmp ebx,4
+ jb NEAR L$028tail
+ mov DWORD [edi],edx
+ lea edi,[4+edi]
+ add eax,4
+ sub ebx,4
+ jz NEAR L$025done
+ mov ecx,8
+ jmp NEAR L$026loop
+align 16
+L$028tail:
+ mov BYTE [edi],dl
+ lea edi,[1+edi]
+ inc eax
+ shr edx,8
+ dec ebx
+ jnz NEAR L$028tail
+L$025done:
+ xor edx,edx
+ pop ebx
+ pop edi
+ ret
+segment .bss
+common _OPENSSL_ia32cap_P 16
+segment .CRT$XCU data align=4
+extern _OPENSSL_cpuid_setup
+dd _OPENSSL_cpuid_setup
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/aesni-x86.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/aesni-x86.S
new file mode 100644
index 0000000000..836e831b18
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/aesni-x86.S
@@ -0,0 +1,3247 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/aes/asm/aesni-x86.pl
+#
+# Copyright 2009-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.globl aesni_encrypt
+.type aesni_encrypt,@function
+.align 16
+aesni_encrypt:
+.L_aesni_encrypt_begin:
+ movl 4(%esp),%eax
+ movl 12(%esp),%edx
+ movups (%eax),%xmm2
+ movl 240(%edx),%ecx
+ movl 8(%esp),%eax
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L000enc1_loop_1:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L000enc1_loop_1
+.byte 102,15,56,221,209
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ movups %xmm2,(%eax)
+ pxor %xmm2,%xmm2
+ ret
+.size aesni_encrypt,.-.L_aesni_encrypt_begin
+.globl aesni_decrypt
+.type aesni_decrypt,@function
+.align 16
+aesni_decrypt:
+.L_aesni_decrypt_begin:
+ movl 4(%esp),%eax
+ movl 12(%esp),%edx
+ movups (%eax),%xmm2
+ movl 240(%edx),%ecx
+ movl 8(%esp),%eax
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L001dec1_loop_2:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L001dec1_loop_2
+.byte 102,15,56,223,209
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ movups %xmm2,(%eax)
+ pxor %xmm2,%xmm2
+ ret
+.size aesni_decrypt,.-.L_aesni_decrypt_begin
+.type _aesni_encrypt2,@function
+.align 16
+_aesni_encrypt2:
+ movups (%edx),%xmm0
+ shll $4,%ecx
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ movups 32(%edx),%xmm0
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+ addl $16,%ecx
+.L002enc2_loop:
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,220,208
+.byte 102,15,56,220,216
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L002enc2_loop
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,221,208
+.byte 102,15,56,221,216
+ ret
+.size _aesni_encrypt2,.-_aesni_encrypt2
+.type _aesni_decrypt2,@function
+.align 16
+_aesni_decrypt2:
+ movups (%edx),%xmm0
+ shll $4,%ecx
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ movups 32(%edx),%xmm0
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+ addl $16,%ecx
+.L003dec2_loop:
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,222,208
+.byte 102,15,56,222,216
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L003dec2_loop
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,223,208
+.byte 102,15,56,223,216
+ ret
+.size _aesni_decrypt2,.-_aesni_decrypt2
+.type _aesni_encrypt3,@function
+.align 16
+_aesni_encrypt3:
+ movups (%edx),%xmm0
+ shll $4,%ecx
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+ movups 32(%edx),%xmm0
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+ addl $16,%ecx
+.L004enc3_loop:
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,220,208
+.byte 102,15,56,220,216
+.byte 102,15,56,220,224
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L004enc3_loop
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.byte 102,15,56,221,208
+.byte 102,15,56,221,216
+.byte 102,15,56,221,224
+ ret
+.size _aesni_encrypt3,.-_aesni_encrypt3
+.type _aesni_decrypt3,@function
+.align 16
+_aesni_decrypt3:
+ movups (%edx),%xmm0
+ shll $4,%ecx
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+ movups 32(%edx),%xmm0
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+ addl $16,%ecx
+.L005dec3_loop:
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,222,208
+.byte 102,15,56,222,216
+.byte 102,15,56,222,224
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L005dec3_loop
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.byte 102,15,56,223,208
+.byte 102,15,56,223,216
+.byte 102,15,56,223,224
+ ret
+.size _aesni_decrypt3,.-_aesni_decrypt3
+.type _aesni_encrypt4,@function
+.align 16
+_aesni_encrypt4:
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ shll $4,%ecx
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+ pxor %xmm0,%xmm5
+ movups 32(%edx),%xmm0
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+.byte 15,31,64,0
+ addl $16,%ecx
+.L006enc4_loop:
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,220,208
+.byte 102,15,56,220,216
+.byte 102,15,56,220,224
+.byte 102,15,56,220,232
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L006enc4_loop
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+.byte 102,15,56,221,208
+.byte 102,15,56,221,216
+.byte 102,15,56,221,224
+.byte 102,15,56,221,232
+ ret
+.size _aesni_encrypt4,.-_aesni_encrypt4
+.type _aesni_decrypt4,@function
+.align 16
+_aesni_decrypt4:
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ shll $4,%ecx
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+ pxor %xmm0,%xmm5
+ movups 32(%edx),%xmm0
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+.byte 15,31,64,0
+ addl $16,%ecx
+.L007dec4_loop:
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.byte 102,15,56,222,233
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,222,208
+.byte 102,15,56,222,216
+.byte 102,15,56,222,224
+.byte 102,15,56,222,232
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L007dec4_loop
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.byte 102,15,56,222,233
+.byte 102,15,56,223,208
+.byte 102,15,56,223,216
+.byte 102,15,56,223,224
+.byte 102,15,56,223,232
+ ret
+.size _aesni_decrypt4,.-_aesni_decrypt4
+.type _aesni_encrypt6,@function
+.align 16
+_aesni_encrypt6:
+ movups (%edx),%xmm0
+ shll $4,%ecx
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+.byte 102,15,56,220,209
+ pxor %xmm0,%xmm5
+ pxor %xmm0,%xmm6
+.byte 102,15,56,220,217
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+.byte 102,15,56,220,225
+ pxor %xmm0,%xmm7
+ movups (%edx,%ecx,1),%xmm0
+ addl $16,%ecx
+ jmp .L008_aesni_encrypt6_inner
+.align 16
+.L009enc6_loop:
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.L008_aesni_encrypt6_inner:
+.byte 102,15,56,220,233
+.byte 102,15,56,220,241
+.byte 102,15,56,220,249
+.L_aesni_encrypt6_enter:
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,220,208
+.byte 102,15,56,220,216
+.byte 102,15,56,220,224
+.byte 102,15,56,220,232
+.byte 102,15,56,220,240
+.byte 102,15,56,220,248
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L009enc6_loop
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+.byte 102,15,56,220,241
+.byte 102,15,56,220,249
+.byte 102,15,56,221,208
+.byte 102,15,56,221,216
+.byte 102,15,56,221,224
+.byte 102,15,56,221,232
+.byte 102,15,56,221,240
+.byte 102,15,56,221,248
+ ret
+.size _aesni_encrypt6,.-_aesni_encrypt6
+.type _aesni_decrypt6,@function
+.align 16
+_aesni_decrypt6:
+ movups (%edx),%xmm0
+ shll $4,%ecx
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+.byte 102,15,56,222,209
+ pxor %xmm0,%xmm5
+ pxor %xmm0,%xmm6
+.byte 102,15,56,222,217
+ leal 32(%edx,%ecx,1),%edx
+ negl %ecx
+.byte 102,15,56,222,225
+ pxor %xmm0,%xmm7
+ movups (%edx,%ecx,1),%xmm0
+ addl $16,%ecx
+ jmp .L010_aesni_decrypt6_inner
+.align 16
+.L011dec6_loop:
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.L010_aesni_decrypt6_inner:
+.byte 102,15,56,222,233
+.byte 102,15,56,222,241
+.byte 102,15,56,222,249
+.L_aesni_decrypt6_enter:
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,222,208
+.byte 102,15,56,222,216
+.byte 102,15,56,222,224
+.byte 102,15,56,222,232
+.byte 102,15,56,222,240
+.byte 102,15,56,222,248
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L011dec6_loop
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.byte 102,15,56,222,233
+.byte 102,15,56,222,241
+.byte 102,15,56,222,249
+.byte 102,15,56,223,208
+.byte 102,15,56,223,216
+.byte 102,15,56,223,224
+.byte 102,15,56,223,232
+.byte 102,15,56,223,240
+.byte 102,15,56,223,248
+ ret
+.size _aesni_decrypt6,.-_aesni_decrypt6
+.globl aesni_ecb_encrypt
+.type aesni_ecb_encrypt,@function
+.align 16
+aesni_ecb_encrypt:
+.L_aesni_ecb_encrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movl 36(%esp),%ebx
+ andl $-16,%eax
+ jz .L012ecb_ret
+ movl 240(%edx),%ecx
+ testl %ebx,%ebx
+ jz .L013ecb_decrypt
+ movl %edx,%ebp
+ movl %ecx,%ebx
+ cmpl $96,%eax
+ jb .L014ecb_enc_tail
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ movdqu 80(%esi),%xmm7
+ leal 96(%esi),%esi
+ subl $96,%eax
+ jmp .L015ecb_enc_loop6_enter
+.align 16
+.L016ecb_enc_loop6:
+ movups %xmm2,(%edi)
+ movdqu (%esi),%xmm2
+ movups %xmm3,16(%edi)
+ movdqu 16(%esi),%xmm3
+ movups %xmm4,32(%edi)
+ movdqu 32(%esi),%xmm4
+ movups %xmm5,48(%edi)
+ movdqu 48(%esi),%xmm5
+ movups %xmm6,64(%edi)
+ movdqu 64(%esi),%xmm6
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ movdqu 80(%esi),%xmm7
+ leal 96(%esi),%esi
+.L015ecb_enc_loop6_enter:
+ call _aesni_encrypt6
+ movl %ebp,%edx
+ movl %ebx,%ecx
+ subl $96,%eax
+ jnc .L016ecb_enc_loop6
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ addl $96,%eax
+ jz .L012ecb_ret
+.L014ecb_enc_tail:
+ movups (%esi),%xmm2
+ cmpl $32,%eax
+ jb .L017ecb_enc_one
+ movups 16(%esi),%xmm3
+ je .L018ecb_enc_two
+ movups 32(%esi),%xmm4
+ cmpl $64,%eax
+ jb .L019ecb_enc_three
+ movups 48(%esi),%xmm5
+ je .L020ecb_enc_four
+ movups 64(%esi),%xmm6
+ xorps %xmm7,%xmm7
+ call _aesni_encrypt6
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L017ecb_enc_one:
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L021enc1_loop_3:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L021enc1_loop_3
+.byte 102,15,56,221,209
+ movups %xmm2,(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L018ecb_enc_two:
+ call _aesni_encrypt2
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L019ecb_enc_three:
+ call _aesni_encrypt3
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L020ecb_enc_four:
+ call _aesni_encrypt4
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L013ecb_decrypt:
+ movl %edx,%ebp
+ movl %ecx,%ebx
+ cmpl $96,%eax
+ jb .L022ecb_dec_tail
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ movdqu 80(%esi),%xmm7
+ leal 96(%esi),%esi
+ subl $96,%eax
+ jmp .L023ecb_dec_loop6_enter
+.align 16
+.L024ecb_dec_loop6:
+ movups %xmm2,(%edi)
+ movdqu (%esi),%xmm2
+ movups %xmm3,16(%edi)
+ movdqu 16(%esi),%xmm3
+ movups %xmm4,32(%edi)
+ movdqu 32(%esi),%xmm4
+ movups %xmm5,48(%edi)
+ movdqu 48(%esi),%xmm5
+ movups %xmm6,64(%edi)
+ movdqu 64(%esi),%xmm6
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ movdqu 80(%esi),%xmm7
+ leal 96(%esi),%esi
+.L023ecb_dec_loop6_enter:
+ call _aesni_decrypt6
+ movl %ebp,%edx
+ movl %ebx,%ecx
+ subl $96,%eax
+ jnc .L024ecb_dec_loop6
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ addl $96,%eax
+ jz .L012ecb_ret
+.L022ecb_dec_tail:
+ movups (%esi),%xmm2
+ cmpl $32,%eax
+ jb .L025ecb_dec_one
+ movups 16(%esi),%xmm3
+ je .L026ecb_dec_two
+ movups 32(%esi),%xmm4
+ cmpl $64,%eax
+ jb .L027ecb_dec_three
+ movups 48(%esi),%xmm5
+ je .L028ecb_dec_four
+ movups 64(%esi),%xmm6
+ xorps %xmm7,%xmm7
+ call _aesni_decrypt6
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L025ecb_dec_one:
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L029dec1_loop_4:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L029dec1_loop_4
+.byte 102,15,56,223,209
+ movups %xmm2,(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L026ecb_dec_two:
+ call _aesni_decrypt2
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L027ecb_dec_three:
+ call _aesni_decrypt3
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ jmp .L012ecb_ret
+.align 16
+.L028ecb_dec_four:
+ call _aesni_decrypt4
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+.L012ecb_ret:
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ pxor %xmm5,%xmm5
+ pxor %xmm6,%xmm6
+ pxor %xmm7,%xmm7
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_ecb_encrypt,.-.L_aesni_ecb_encrypt_begin
+.globl aesni_ccm64_encrypt_blocks
+.type aesni_ccm64_encrypt_blocks,@function
+.align 16
+aesni_ccm64_encrypt_blocks:
+.L_aesni_ccm64_encrypt_blocks_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movl 36(%esp),%ebx
+ movl 40(%esp),%ecx
+ movl %esp,%ebp
+ subl $60,%esp
+ andl $-16,%esp
+ movl %ebp,48(%esp)
+ movdqu (%ebx),%xmm7
+ movdqu (%ecx),%xmm3
+ movl 240(%edx),%ecx
+ movl $202182159,(%esp)
+ movl $134810123,4(%esp)
+ movl $67438087,8(%esp)
+ movl $66051,12(%esp)
+ movl $1,%ebx
+ xorl %ebp,%ebp
+ movl %ebx,16(%esp)
+ movl %ebp,20(%esp)
+ movl %ebp,24(%esp)
+ movl %ebp,28(%esp)
+ shll $4,%ecx
+ movl $16,%ebx
+ leal (%edx),%ebp
+ movdqa (%esp),%xmm5
+ movdqa %xmm7,%xmm2
+ leal 32(%edx,%ecx,1),%edx
+ subl %ecx,%ebx
+.byte 102,15,56,0,253
+.L030ccm64_enc_outer:
+ movups (%ebp),%xmm0
+ movl %ebx,%ecx
+ movups (%esi),%xmm6
+ xorps %xmm0,%xmm2
+ movups 16(%ebp),%xmm1
+ xorps %xmm6,%xmm0
+ xorps %xmm0,%xmm3
+ movups 32(%ebp),%xmm0
+.L031ccm64_enc2_loop:
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,220,208
+.byte 102,15,56,220,216
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L031ccm64_enc2_loop
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+ paddq 16(%esp),%xmm7
+ decl %eax
+.byte 102,15,56,221,208
+.byte 102,15,56,221,216
+ leal 16(%esi),%esi
+ xorps %xmm2,%xmm6
+ movdqa %xmm7,%xmm2
+ movups %xmm6,(%edi)
+.byte 102,15,56,0,213
+ leal 16(%edi),%edi
+ jnz .L030ccm64_enc_outer
+ movl 48(%esp),%esp
+ movl 40(%esp),%edi
+ movups %xmm3,(%edi)
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ pxor %xmm5,%xmm5
+ pxor %xmm6,%xmm6
+ pxor %xmm7,%xmm7
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_ccm64_encrypt_blocks,.-.L_aesni_ccm64_encrypt_blocks_begin
+.globl aesni_ccm64_decrypt_blocks
+.type aesni_ccm64_decrypt_blocks,@function
+.align 16
+aesni_ccm64_decrypt_blocks:
+.L_aesni_ccm64_decrypt_blocks_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movl 36(%esp),%ebx
+ movl 40(%esp),%ecx
+ movl %esp,%ebp
+ subl $60,%esp
+ andl $-16,%esp
+ movl %ebp,48(%esp)
+ movdqu (%ebx),%xmm7
+ movdqu (%ecx),%xmm3
+ movl 240(%edx),%ecx
+ movl $202182159,(%esp)
+ movl $134810123,4(%esp)
+ movl $67438087,8(%esp)
+ movl $66051,12(%esp)
+ movl $1,%ebx
+ xorl %ebp,%ebp
+ movl %ebx,16(%esp)
+ movl %ebp,20(%esp)
+ movl %ebp,24(%esp)
+ movl %ebp,28(%esp)
+ movdqa (%esp),%xmm5
+ movdqa %xmm7,%xmm2
+ movl %edx,%ebp
+ movl %ecx,%ebx
+.byte 102,15,56,0,253
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L032enc1_loop_5:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L032enc1_loop_5
+.byte 102,15,56,221,209
+ shll $4,%ebx
+ movl $16,%ecx
+ movups (%esi),%xmm6
+ paddq 16(%esp),%xmm7
+ leal 16(%esi),%esi
+ subl %ebx,%ecx
+ leal 32(%ebp,%ebx,1),%edx
+ movl %ecx,%ebx
+ jmp .L033ccm64_dec_outer
+.align 16
+.L033ccm64_dec_outer:
+ xorps %xmm2,%xmm6
+ movdqa %xmm7,%xmm2
+ movups %xmm6,(%edi)
+ leal 16(%edi),%edi
+.byte 102,15,56,0,213
+ subl $1,%eax
+ jz .L034ccm64_dec_break
+ movups (%ebp),%xmm0
+ movl %ebx,%ecx
+ movups 16(%ebp),%xmm1
+ xorps %xmm0,%xmm6
+ xorps %xmm0,%xmm2
+ xorps %xmm6,%xmm3
+ movups 32(%ebp),%xmm0
+.L035ccm64_dec2_loop:
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+ movups (%edx,%ecx,1),%xmm1
+ addl $32,%ecx
+.byte 102,15,56,220,208
+.byte 102,15,56,220,216
+ movups -16(%edx,%ecx,1),%xmm0
+ jnz .L035ccm64_dec2_loop
+ movups (%esi),%xmm6
+ paddq 16(%esp),%xmm7
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,221,208
+.byte 102,15,56,221,216
+ leal 16(%esi),%esi
+ jmp .L033ccm64_dec_outer
+.align 16
+.L034ccm64_dec_break:
+ movl 240(%ebp),%ecx
+ movl %ebp,%edx
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm6
+ leal 32(%edx),%edx
+ xorps %xmm6,%xmm3
+.L036enc1_loop_6:
+.byte 102,15,56,220,217
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L036enc1_loop_6
+.byte 102,15,56,221,217
+ movl 48(%esp),%esp
+ movl 40(%esp),%edi
+ movups %xmm3,(%edi)
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ pxor %xmm5,%xmm5
+ pxor %xmm6,%xmm6
+ pxor %xmm7,%xmm7
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_ccm64_decrypt_blocks,.-.L_aesni_ccm64_decrypt_blocks_begin
+.globl aesni_ctr32_encrypt_blocks
+.type aesni_ctr32_encrypt_blocks,@function
+.align 16
+aesni_ctr32_encrypt_blocks:
+.L_aesni_ctr32_encrypt_blocks_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movl 36(%esp),%ebx
+ movl %esp,%ebp
+ subl $88,%esp
+ andl $-16,%esp
+ movl %ebp,80(%esp)
+ cmpl $1,%eax
+ je .L037ctr32_one_shortcut
+ movdqu (%ebx),%xmm7
+ movl $202182159,(%esp)
+ movl $134810123,4(%esp)
+ movl $67438087,8(%esp)
+ movl $66051,12(%esp)
+ movl $6,%ecx
+ xorl %ebp,%ebp
+ movl %ecx,16(%esp)
+ movl %ecx,20(%esp)
+ movl %ecx,24(%esp)
+ movl %ebp,28(%esp)
+.byte 102,15,58,22,251,3
+.byte 102,15,58,34,253,3
+ movl 240(%edx),%ecx
+ bswap %ebx
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ movdqa (%esp),%xmm2
+.byte 102,15,58,34,195,0
+ leal 3(%ebx),%ebp
+.byte 102,15,58,34,205,0
+ incl %ebx
+.byte 102,15,58,34,195,1
+ incl %ebp
+.byte 102,15,58,34,205,1
+ incl %ebx
+.byte 102,15,58,34,195,2
+ incl %ebp
+.byte 102,15,58,34,205,2
+ movdqa %xmm0,48(%esp)
+.byte 102,15,56,0,194
+ movdqu (%edx),%xmm6
+ movdqa %xmm1,64(%esp)
+.byte 102,15,56,0,202
+ pshufd $192,%xmm0,%xmm2
+ pshufd $128,%xmm0,%xmm3
+ cmpl $6,%eax
+ jb .L038ctr32_tail
+ pxor %xmm6,%xmm7
+ shll $4,%ecx
+ movl $16,%ebx
+ movdqa %xmm7,32(%esp)
+ movl %edx,%ebp
+ subl %ecx,%ebx
+ leal 32(%edx,%ecx,1),%edx
+ subl $6,%eax
+ jmp .L039ctr32_loop6
+.align 16
+.L039ctr32_loop6:
+ pshufd $64,%xmm0,%xmm4
+ movdqa 32(%esp),%xmm0
+ pshufd $192,%xmm1,%xmm5
+ pxor %xmm0,%xmm2
+ pshufd $128,%xmm1,%xmm6
+ pxor %xmm0,%xmm3
+ pshufd $64,%xmm1,%xmm7
+ movups 16(%ebp),%xmm1
+ pxor %xmm0,%xmm4
+ pxor %xmm0,%xmm5
+.byte 102,15,56,220,209
+ pxor %xmm0,%xmm6
+ pxor %xmm0,%xmm7
+.byte 102,15,56,220,217
+ movups 32(%ebp),%xmm0
+ movl %ebx,%ecx
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+.byte 102,15,56,220,241
+.byte 102,15,56,220,249
+ call .L_aesni_encrypt6_enter
+ movups (%esi),%xmm1
+ movups 16(%esi),%xmm0
+ xorps %xmm1,%xmm2
+ movups 32(%esi),%xmm1
+ xorps %xmm0,%xmm3
+ movups %xmm2,(%edi)
+ movdqa 16(%esp),%xmm0
+ xorps %xmm1,%xmm4
+ movdqa 64(%esp),%xmm1
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ paddd %xmm0,%xmm1
+ paddd 48(%esp),%xmm0
+ movdqa (%esp),%xmm2
+ movups 48(%esi),%xmm3
+ movups 64(%esi),%xmm4
+ xorps %xmm3,%xmm5
+ movups 80(%esi),%xmm3
+ leal 96(%esi),%esi
+ movdqa %xmm0,48(%esp)
+.byte 102,15,56,0,194
+ xorps %xmm4,%xmm6
+ movups %xmm5,48(%edi)
+ xorps %xmm3,%xmm7
+ movdqa %xmm1,64(%esp)
+.byte 102,15,56,0,202
+ movups %xmm6,64(%edi)
+ pshufd $192,%xmm0,%xmm2
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ pshufd $128,%xmm0,%xmm3
+ subl $6,%eax
+ jnc .L039ctr32_loop6
+ addl $6,%eax
+ jz .L040ctr32_ret
+ movdqu (%ebp),%xmm7
+ movl %ebp,%edx
+ pxor 32(%esp),%xmm7
+ movl 240(%ebp),%ecx
+.L038ctr32_tail:
+ por %xmm7,%xmm2
+ cmpl $2,%eax
+ jb .L041ctr32_one
+ pshufd $64,%xmm0,%xmm4
+ por %xmm7,%xmm3
+ je .L042ctr32_two
+ pshufd $192,%xmm1,%xmm5
+ por %xmm7,%xmm4
+ cmpl $4,%eax
+ jb .L043ctr32_three
+ pshufd $128,%xmm1,%xmm6
+ por %xmm7,%xmm5
+ je .L044ctr32_four
+ por %xmm7,%xmm6
+ call _aesni_encrypt6
+ movups (%esi),%xmm1
+ movups 16(%esi),%xmm0
+ xorps %xmm1,%xmm2
+ movups 32(%esi),%xmm1
+ xorps %xmm0,%xmm3
+ movups 48(%esi),%xmm0
+ xorps %xmm1,%xmm4
+ movups 64(%esi),%xmm1
+ xorps %xmm0,%xmm5
+ movups %xmm2,(%edi)
+ xorps %xmm1,%xmm6
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ jmp .L040ctr32_ret
+.align 16
+.L037ctr32_one_shortcut:
+ movups (%ebx),%xmm2
+ movl 240(%edx),%ecx
+.L041ctr32_one:
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L045enc1_loop_7:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L045enc1_loop_7
+.byte 102,15,56,221,209
+ movups (%esi),%xmm6
+ xorps %xmm2,%xmm6
+ movups %xmm6,(%edi)
+ jmp .L040ctr32_ret
+.align 16
+.L042ctr32_two:
+ call _aesni_encrypt2
+ movups (%esi),%xmm5
+ movups 16(%esi),%xmm6
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ jmp .L040ctr32_ret
+.align 16
+.L043ctr32_three:
+ call _aesni_encrypt3
+ movups (%esi),%xmm5
+ movups 16(%esi),%xmm6
+ xorps %xmm5,%xmm2
+ movups 32(%esi),%xmm7
+ xorps %xmm6,%xmm3
+ movups %xmm2,(%edi)
+ xorps %xmm7,%xmm4
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ jmp .L040ctr32_ret
+.align 16
+.L044ctr32_four:
+ call _aesni_encrypt4
+ movups (%esi),%xmm6
+ movups 16(%esi),%xmm7
+ movups 32(%esi),%xmm1
+ xorps %xmm6,%xmm2
+ movups 48(%esi),%xmm0
+ xorps %xmm7,%xmm3
+ movups %xmm2,(%edi)
+ xorps %xmm1,%xmm4
+ movups %xmm3,16(%edi)
+ xorps %xmm0,%xmm5
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+.L040ctr32_ret:
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ movdqa %xmm0,32(%esp)
+ pxor %xmm5,%xmm5
+ movdqa %xmm0,48(%esp)
+ pxor %xmm6,%xmm6
+ movdqa %xmm0,64(%esp)
+ pxor %xmm7,%xmm7
+ movl 80(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_ctr32_encrypt_blocks,.-.L_aesni_ctr32_encrypt_blocks_begin
+.globl aesni_xts_encrypt
+.type aesni_xts_encrypt,@function
+.align 16
+aesni_xts_encrypt:
+.L_aesni_xts_encrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 36(%esp),%edx
+ movl 40(%esp),%esi
+ movl 240(%edx),%ecx
+ movups (%esi),%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L046enc1_loop_8:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L046enc1_loop_8
+.byte 102,15,56,221,209
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movl %esp,%ebp
+ subl $120,%esp
+ movl 240(%edx),%ecx
+ andl $-16,%esp
+ movl $135,96(%esp)
+ movl $0,100(%esp)
+ movl $1,104(%esp)
+ movl $0,108(%esp)
+ movl %eax,112(%esp)
+ movl %ebp,116(%esp)
+ movdqa %xmm2,%xmm1
+ pxor %xmm0,%xmm0
+ movdqa 96(%esp),%xmm3
+ pcmpgtd %xmm1,%xmm0
+ andl $-16,%eax
+ movl %edx,%ebp
+ movl %ecx,%ebx
+ subl $96,%eax
+ jc .L047xts_enc_short
+ shll $4,%ecx
+ movl $16,%ebx
+ subl %ecx,%ebx
+ leal 32(%edx,%ecx,1),%edx
+ jmp .L048xts_enc_loop6
+.align 16
+.L048xts_enc_loop6:
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,16(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,32(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,48(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm7
+ movdqa %xmm1,64(%esp)
+ paddq %xmm1,%xmm1
+ movups (%ebp),%xmm0
+ pand %xmm3,%xmm7
+ movups (%esi),%xmm2
+ pxor %xmm1,%xmm7
+ movl %ebx,%ecx
+ movdqu 16(%esi),%xmm3
+ xorps %xmm0,%xmm2
+ movdqu 32(%esi),%xmm4
+ pxor %xmm0,%xmm3
+ movdqu 48(%esi),%xmm5
+ pxor %xmm0,%xmm4
+ movdqu 64(%esi),%xmm6
+ pxor %xmm0,%xmm5
+ movdqu 80(%esi),%xmm1
+ pxor %xmm0,%xmm6
+ leal 96(%esi),%esi
+ pxor (%esp),%xmm2
+ movdqa %xmm7,80(%esp)
+ pxor %xmm1,%xmm7
+ movups 16(%ebp),%xmm1
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+.byte 102,15,56,220,209
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+.byte 102,15,56,220,217
+ pxor %xmm0,%xmm7
+ movups 32(%ebp),%xmm0
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+.byte 102,15,56,220,241
+.byte 102,15,56,220,249
+ call .L_aesni_encrypt6_enter
+ movdqa 80(%esp),%xmm1
+ pxor %xmm0,%xmm0
+ xorps (%esp),%xmm2
+ pcmpgtd %xmm1,%xmm0
+ xorps 16(%esp),%xmm3
+ movups %xmm2,(%edi)
+ xorps 32(%esp),%xmm4
+ movups %xmm3,16(%edi)
+ xorps 48(%esp),%xmm5
+ movups %xmm4,32(%edi)
+ xorps 64(%esp),%xmm6
+ movups %xmm5,48(%edi)
+ xorps %xmm1,%xmm7
+ movups %xmm6,64(%edi)
+ pshufd $19,%xmm0,%xmm2
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ movdqa 96(%esp),%xmm3
+ pxor %xmm0,%xmm0
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ subl $96,%eax
+ jnc .L048xts_enc_loop6
+ movl 240(%ebp),%ecx
+ movl %ebp,%edx
+ movl %ecx,%ebx
+.L047xts_enc_short:
+ addl $96,%eax
+ jz .L049xts_enc_done6x
+ movdqa %xmm1,%xmm5
+ cmpl $32,%eax
+ jb .L050xts_enc_one
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ je .L051xts_enc_two
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,%xmm6
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ cmpl $64,%eax
+ jb .L052xts_enc_three
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,%xmm7
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ movdqa %xmm5,(%esp)
+ movdqa %xmm6,16(%esp)
+ je .L053xts_enc_four
+ movdqa %xmm7,32(%esp)
+ pshufd $19,%xmm0,%xmm7
+ movdqa %xmm1,48(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm7
+ pxor %xmm1,%xmm7
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ pxor (%esp),%xmm2
+ movdqu 48(%esi),%xmm5
+ pxor 16(%esp),%xmm3
+ movdqu 64(%esi),%xmm6
+ pxor 32(%esp),%xmm4
+ leal 80(%esi),%esi
+ pxor 48(%esp),%xmm5
+ movdqa %xmm7,64(%esp)
+ pxor %xmm7,%xmm6
+ call _aesni_encrypt6
+ movaps 64(%esp),%xmm1
+ xorps (%esp),%xmm2
+ xorps 16(%esp),%xmm3
+ xorps 32(%esp),%xmm4
+ movups %xmm2,(%edi)
+ xorps 48(%esp),%xmm5
+ movups %xmm3,16(%edi)
+ xorps %xmm1,%xmm6
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ leal 80(%edi),%edi
+ jmp .L054xts_enc_done
+.align 16
+.L050xts_enc_one:
+ movups (%esi),%xmm2
+ leal 16(%esi),%esi
+ xorps %xmm5,%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L055enc1_loop_9:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L055enc1_loop_9
+.byte 102,15,56,221,209
+ xorps %xmm5,%xmm2
+ movups %xmm2,(%edi)
+ leal 16(%edi),%edi
+ movdqa %xmm5,%xmm1
+ jmp .L054xts_enc_done
+.align 16
+.L051xts_enc_two:
+ movaps %xmm1,%xmm6
+ movups (%esi),%xmm2
+ movups 16(%esi),%xmm3
+ leal 32(%esi),%esi
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ call _aesni_encrypt2
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ leal 32(%edi),%edi
+ movdqa %xmm6,%xmm1
+ jmp .L054xts_enc_done
+.align 16
+.L052xts_enc_three:
+ movaps %xmm1,%xmm7
+ movups (%esi),%xmm2
+ movups 16(%esi),%xmm3
+ movups 32(%esi),%xmm4
+ leal 48(%esi),%esi
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm7,%xmm4
+ call _aesni_encrypt3
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm7,%xmm4
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ leal 48(%edi),%edi
+ movdqa %xmm7,%xmm1
+ jmp .L054xts_enc_done
+.align 16
+.L053xts_enc_four:
+ movaps %xmm1,%xmm6
+ movups (%esi),%xmm2
+ movups 16(%esi),%xmm3
+ movups 32(%esi),%xmm4
+ xorps (%esp),%xmm2
+ movups 48(%esi),%xmm5
+ leal 64(%esi),%esi
+ xorps 16(%esp),%xmm3
+ xorps %xmm7,%xmm4
+ xorps %xmm6,%xmm5
+ call _aesni_encrypt4
+ xorps (%esp),%xmm2
+ xorps 16(%esp),%xmm3
+ xorps %xmm7,%xmm4
+ movups %xmm2,(%edi)
+ xorps %xmm6,%xmm5
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ leal 64(%edi),%edi
+ movdqa %xmm6,%xmm1
+ jmp .L054xts_enc_done
+.align 16
+.L049xts_enc_done6x:
+ movl 112(%esp),%eax
+ andl $15,%eax
+ jz .L056xts_enc_ret
+ movdqa %xmm1,%xmm5
+ movl %eax,112(%esp)
+ jmp .L057xts_enc_steal
+.align 16
+.L054xts_enc_done:
+ movl 112(%esp),%eax
+ pxor %xmm0,%xmm0
+ andl $15,%eax
+ jz .L056xts_enc_ret
+ pcmpgtd %xmm1,%xmm0
+ movl %eax,112(%esp)
+ pshufd $19,%xmm0,%xmm5
+ paddq %xmm1,%xmm1
+ pand 96(%esp),%xmm5
+ pxor %xmm1,%xmm5
+.L057xts_enc_steal:
+ movzbl (%esi),%ecx
+ movzbl -16(%edi),%edx
+ leal 1(%esi),%esi
+ movb %cl,-16(%edi)
+ movb %dl,(%edi)
+ leal 1(%edi),%edi
+ subl $1,%eax
+ jnz .L057xts_enc_steal
+ subl 112(%esp),%edi
+ movl %ebp,%edx
+ movl %ebx,%ecx
+ movups -16(%edi),%xmm2
+ xorps %xmm5,%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L058enc1_loop_10:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L058enc1_loop_10
+.byte 102,15,56,221,209
+ xorps %xmm5,%xmm2
+ movups %xmm2,-16(%edi)
+.L056xts_enc_ret:
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ movdqa %xmm0,(%esp)
+ pxor %xmm3,%xmm3
+ movdqa %xmm0,16(%esp)
+ pxor %xmm4,%xmm4
+ movdqa %xmm0,32(%esp)
+ pxor %xmm5,%xmm5
+ movdqa %xmm0,48(%esp)
+ pxor %xmm6,%xmm6
+ movdqa %xmm0,64(%esp)
+ pxor %xmm7,%xmm7
+ movdqa %xmm0,80(%esp)
+ movl 116(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_xts_encrypt,.-.L_aesni_xts_encrypt_begin
+.globl aesni_xts_decrypt
+.type aesni_xts_decrypt,@function
+.align 16
+aesni_xts_decrypt:
+.L_aesni_xts_decrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 36(%esp),%edx
+ movl 40(%esp),%esi
+ movl 240(%edx),%ecx
+ movups (%esi),%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L059enc1_loop_11:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L059enc1_loop_11
+.byte 102,15,56,221,209
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movl %esp,%ebp
+ subl $120,%esp
+ andl $-16,%esp
+ xorl %ebx,%ebx
+ testl $15,%eax
+ setnz %bl
+ shll $4,%ebx
+ subl %ebx,%eax
+ movl $135,96(%esp)
+ movl $0,100(%esp)
+ movl $1,104(%esp)
+ movl $0,108(%esp)
+ movl %eax,112(%esp)
+ movl %ebp,116(%esp)
+ movl 240(%edx),%ecx
+ movl %edx,%ebp
+ movl %ecx,%ebx
+ movdqa %xmm2,%xmm1
+ pxor %xmm0,%xmm0
+ movdqa 96(%esp),%xmm3
+ pcmpgtd %xmm1,%xmm0
+ andl $-16,%eax
+ subl $96,%eax
+ jc .L060xts_dec_short
+ shll $4,%ecx
+ movl $16,%ebx
+ subl %ecx,%ebx
+ leal 32(%edx,%ecx,1),%edx
+ jmp .L061xts_dec_loop6
+.align 16
+.L061xts_dec_loop6:
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,16(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,32(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,48(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ pshufd $19,%xmm0,%xmm7
+ movdqa %xmm1,64(%esp)
+ paddq %xmm1,%xmm1
+ movups (%ebp),%xmm0
+ pand %xmm3,%xmm7
+ movups (%esi),%xmm2
+ pxor %xmm1,%xmm7
+ movl %ebx,%ecx
+ movdqu 16(%esi),%xmm3
+ xorps %xmm0,%xmm2
+ movdqu 32(%esi),%xmm4
+ pxor %xmm0,%xmm3
+ movdqu 48(%esi),%xmm5
+ pxor %xmm0,%xmm4
+ movdqu 64(%esi),%xmm6
+ pxor %xmm0,%xmm5
+ movdqu 80(%esi),%xmm1
+ pxor %xmm0,%xmm6
+ leal 96(%esi),%esi
+ pxor (%esp),%xmm2
+ movdqa %xmm7,80(%esp)
+ pxor %xmm1,%xmm7
+ movups 16(%ebp),%xmm1
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+.byte 102,15,56,222,209
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+.byte 102,15,56,222,217
+ pxor %xmm0,%xmm7
+ movups 32(%ebp),%xmm0
+.byte 102,15,56,222,225
+.byte 102,15,56,222,233
+.byte 102,15,56,222,241
+.byte 102,15,56,222,249
+ call .L_aesni_decrypt6_enter
+ movdqa 80(%esp),%xmm1
+ pxor %xmm0,%xmm0
+ xorps (%esp),%xmm2
+ pcmpgtd %xmm1,%xmm0
+ xorps 16(%esp),%xmm3
+ movups %xmm2,(%edi)
+ xorps 32(%esp),%xmm4
+ movups %xmm3,16(%edi)
+ xorps 48(%esp),%xmm5
+ movups %xmm4,32(%edi)
+ xorps 64(%esp),%xmm6
+ movups %xmm5,48(%edi)
+ xorps %xmm1,%xmm7
+ movups %xmm6,64(%edi)
+ pshufd $19,%xmm0,%xmm2
+ movups %xmm7,80(%edi)
+ leal 96(%edi),%edi
+ movdqa 96(%esp),%xmm3
+ pxor %xmm0,%xmm0
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ subl $96,%eax
+ jnc .L061xts_dec_loop6
+ movl 240(%ebp),%ecx
+ movl %ebp,%edx
+ movl %ecx,%ebx
+.L060xts_dec_short:
+ addl $96,%eax
+ jz .L062xts_dec_done6x
+ movdqa %xmm1,%xmm5
+ cmpl $32,%eax
+ jb .L063xts_dec_one
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ je .L064xts_dec_two
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,%xmm6
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ cmpl $64,%eax
+ jb .L065xts_dec_three
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa %xmm1,%xmm7
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+ movdqa %xmm5,(%esp)
+ movdqa %xmm6,16(%esp)
+ je .L066xts_dec_four
+ movdqa %xmm7,32(%esp)
+ pshufd $19,%xmm0,%xmm7
+ movdqa %xmm1,48(%esp)
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm7
+ pxor %xmm1,%xmm7
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ pxor (%esp),%xmm2
+ movdqu 48(%esi),%xmm5
+ pxor 16(%esp),%xmm3
+ movdqu 64(%esi),%xmm6
+ pxor 32(%esp),%xmm4
+ leal 80(%esi),%esi
+ pxor 48(%esp),%xmm5
+ movdqa %xmm7,64(%esp)
+ pxor %xmm7,%xmm6
+ call _aesni_decrypt6
+ movaps 64(%esp),%xmm1
+ xorps (%esp),%xmm2
+ xorps 16(%esp),%xmm3
+ xorps 32(%esp),%xmm4
+ movups %xmm2,(%edi)
+ xorps 48(%esp),%xmm5
+ movups %xmm3,16(%edi)
+ xorps %xmm1,%xmm6
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ movups %xmm6,64(%edi)
+ leal 80(%edi),%edi
+ jmp .L067xts_dec_done
+.align 16
+.L063xts_dec_one:
+ movups (%esi),%xmm2
+ leal 16(%esi),%esi
+ xorps %xmm5,%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L068dec1_loop_12:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L068dec1_loop_12
+.byte 102,15,56,223,209
+ xorps %xmm5,%xmm2
+ movups %xmm2,(%edi)
+ leal 16(%edi),%edi
+ movdqa %xmm5,%xmm1
+ jmp .L067xts_dec_done
+.align 16
+.L064xts_dec_two:
+ movaps %xmm1,%xmm6
+ movups (%esi),%xmm2
+ movups 16(%esi),%xmm3
+ leal 32(%esi),%esi
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ call _aesni_decrypt2
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ leal 32(%edi),%edi
+ movdqa %xmm6,%xmm1
+ jmp .L067xts_dec_done
+.align 16
+.L065xts_dec_three:
+ movaps %xmm1,%xmm7
+ movups (%esi),%xmm2
+ movups 16(%esi),%xmm3
+ movups 32(%esi),%xmm4
+ leal 48(%esi),%esi
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm7,%xmm4
+ call _aesni_decrypt3
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm7,%xmm4
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ leal 48(%edi),%edi
+ movdqa %xmm7,%xmm1
+ jmp .L067xts_dec_done
+.align 16
+.L066xts_dec_four:
+ movaps %xmm1,%xmm6
+ movups (%esi),%xmm2
+ movups 16(%esi),%xmm3
+ movups 32(%esi),%xmm4
+ xorps (%esp),%xmm2
+ movups 48(%esi),%xmm5
+ leal 64(%esi),%esi
+ xorps 16(%esp),%xmm3
+ xorps %xmm7,%xmm4
+ xorps %xmm6,%xmm5
+ call _aesni_decrypt4
+ xorps (%esp),%xmm2
+ xorps 16(%esp),%xmm3
+ xorps %xmm7,%xmm4
+ movups %xmm2,(%edi)
+ xorps %xmm6,%xmm5
+ movups %xmm3,16(%edi)
+ movups %xmm4,32(%edi)
+ movups %xmm5,48(%edi)
+ leal 64(%edi),%edi
+ movdqa %xmm6,%xmm1
+ jmp .L067xts_dec_done
+.align 16
+.L062xts_dec_done6x:
+ movl 112(%esp),%eax
+ andl $15,%eax
+ jz .L069xts_dec_ret
+ movl %eax,112(%esp)
+ jmp .L070xts_dec_only_one_more
+.align 16
+.L067xts_dec_done:
+ movl 112(%esp),%eax
+ pxor %xmm0,%xmm0
+ andl $15,%eax
+ jz .L069xts_dec_ret
+ pcmpgtd %xmm1,%xmm0
+ movl %eax,112(%esp)
+ pshufd $19,%xmm0,%xmm2
+ pxor %xmm0,%xmm0
+ movdqa 96(%esp),%xmm3
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm2
+ pcmpgtd %xmm1,%xmm0
+ pxor %xmm2,%xmm1
+.L070xts_dec_only_one_more:
+ pshufd $19,%xmm0,%xmm5
+ movdqa %xmm1,%xmm6
+ paddq %xmm1,%xmm1
+ pand %xmm3,%xmm5
+ pxor %xmm1,%xmm5
+ movl %ebp,%edx
+ movl %ebx,%ecx
+ movups (%esi),%xmm2
+ xorps %xmm5,%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L071dec1_loop_13:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L071dec1_loop_13
+.byte 102,15,56,223,209
+ xorps %xmm5,%xmm2
+ movups %xmm2,(%edi)
+.L072xts_dec_steal:
+ movzbl 16(%esi),%ecx
+ movzbl (%edi),%edx
+ leal 1(%esi),%esi
+ movb %cl,(%edi)
+ movb %dl,16(%edi)
+ leal 1(%edi),%edi
+ subl $1,%eax
+ jnz .L072xts_dec_steal
+ subl 112(%esp),%edi
+ movl %ebp,%edx
+ movl %ebx,%ecx
+ movups (%edi),%xmm2
+ xorps %xmm6,%xmm2
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L073dec1_loop_14:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L073dec1_loop_14
+.byte 102,15,56,223,209
+ xorps %xmm6,%xmm2
+ movups %xmm2,(%edi)
+.L069xts_dec_ret:
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ movdqa %xmm0,(%esp)
+ pxor %xmm3,%xmm3
+ movdqa %xmm0,16(%esp)
+ pxor %xmm4,%xmm4
+ movdqa %xmm0,32(%esp)
+ pxor %xmm5,%xmm5
+ movdqa %xmm0,48(%esp)
+ pxor %xmm6,%xmm6
+ movdqa %xmm0,64(%esp)
+ pxor %xmm7,%xmm7
+ movdqa %xmm0,80(%esp)
+ movl 116(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_xts_decrypt,.-.L_aesni_xts_decrypt_begin
+.globl aesni_ocb_encrypt
+.type aesni_ocb_encrypt,@function
+.align 16
+aesni_ocb_encrypt:
+.L_aesni_ocb_encrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 40(%esp),%ecx
+ movl 48(%esp),%ebx
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movdqu (%ecx),%xmm0
+ movl 36(%esp),%ebp
+ movdqu (%ebx),%xmm1
+ movl 44(%esp),%ebx
+ movl %esp,%ecx
+ subl $132,%esp
+ andl $-16,%esp
+ subl %esi,%edi
+ shll $4,%eax
+ leal -96(%esi,%eax,1),%eax
+ movl %edi,120(%esp)
+ movl %eax,124(%esp)
+ movl %ecx,128(%esp)
+ movl 240(%edx),%ecx
+ testl $1,%ebp
+ jnz .L074odd
+ bsfl %ebp,%eax
+ addl $1,%ebp
+ shll $4,%eax
+ movdqu (%ebx,%eax,1),%xmm7
+ movl %edx,%eax
+ movdqu (%esi),%xmm2
+ leal 16(%esi),%esi
+ pxor %xmm0,%xmm7
+ pxor %xmm2,%xmm1
+ pxor %xmm7,%xmm2
+ movdqa %xmm1,%xmm6
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L075enc1_loop_15:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L075enc1_loop_15
+.byte 102,15,56,221,209
+ xorps %xmm7,%xmm2
+ movdqa %xmm7,%xmm0
+ movdqa %xmm6,%xmm1
+ movups %xmm2,-16(%edi,%esi,1)
+ movl 240(%eax),%ecx
+ movl %eax,%edx
+ movl 124(%esp),%eax
+.L074odd:
+ shll $4,%ecx
+ movl $16,%edi
+ subl %ecx,%edi
+ movl %edx,112(%esp)
+ leal 32(%edx,%ecx,1),%edx
+ movl %edi,116(%esp)
+ cmpl %eax,%esi
+ ja .L076short
+ jmp .L077grandloop
+.align 32
+.L077grandloop:
+ leal 1(%ebp),%ecx
+ leal 3(%ebp),%eax
+ leal 5(%ebp),%edi
+ addl $6,%ebp
+ bsfl %ecx,%ecx
+ bsfl %eax,%eax
+ bsfl %edi,%edi
+ shll $4,%ecx
+ shll $4,%eax
+ shll $4,%edi
+ movdqu (%ebx),%xmm2
+ movdqu (%ebx,%ecx,1),%xmm3
+ movl 116(%esp),%ecx
+ movdqa %xmm2,%xmm4
+ movdqu (%ebx,%eax,1),%xmm5
+ movdqa %xmm2,%xmm6
+ movdqu (%ebx,%edi,1),%xmm7
+ pxor %xmm0,%xmm2
+ pxor %xmm2,%xmm3
+ movdqa %xmm2,(%esp)
+ pxor %xmm3,%xmm4
+ movdqa %xmm3,16(%esp)
+ pxor %xmm4,%xmm5
+ movdqa %xmm4,32(%esp)
+ pxor %xmm5,%xmm6
+ movdqa %xmm5,48(%esp)
+ pxor %xmm6,%xmm7
+ movdqa %xmm6,64(%esp)
+ movdqa %xmm7,80(%esp)
+ movups -48(%edx,%ecx,1),%xmm0
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ movdqu 80(%esi),%xmm7
+ leal 96(%esi),%esi
+ pxor %xmm2,%xmm1
+ pxor %xmm0,%xmm2
+ pxor %xmm3,%xmm1
+ pxor %xmm0,%xmm3
+ pxor %xmm4,%xmm1
+ pxor %xmm0,%xmm4
+ pxor %xmm5,%xmm1
+ pxor %xmm0,%xmm5
+ pxor %xmm6,%xmm1
+ pxor %xmm0,%xmm6
+ pxor %xmm7,%xmm1
+ pxor %xmm0,%xmm7
+ movdqa %xmm1,96(%esp)
+ movups -32(%edx,%ecx,1),%xmm1
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+ pxor 80(%esp),%xmm7
+ movups -16(%edx,%ecx,1),%xmm0
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+.byte 102,15,56,220,241
+.byte 102,15,56,220,249
+ movl 120(%esp),%edi
+ movl 124(%esp),%eax
+ call .L_aesni_encrypt6_enter
+ movdqa 80(%esp),%xmm0
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+ pxor %xmm0,%xmm7
+ movdqa 96(%esp),%xmm1
+ movdqu %xmm2,-96(%edi,%esi,1)
+ movdqu %xmm3,-80(%edi,%esi,1)
+ movdqu %xmm4,-64(%edi,%esi,1)
+ movdqu %xmm5,-48(%edi,%esi,1)
+ movdqu %xmm6,-32(%edi,%esi,1)
+ movdqu %xmm7,-16(%edi,%esi,1)
+ cmpl %eax,%esi
+ jb .L077grandloop
+.L076short:
+ addl $96,%eax
+ subl %esi,%eax
+ jz .L078done
+ cmpl $32,%eax
+ jb .L079one
+ je .L080two
+ cmpl $64,%eax
+ jb .L081three
+ je .L082four
+ leal 1(%ebp),%ecx
+ leal 3(%ebp),%eax
+ bsfl %ecx,%ecx
+ bsfl %eax,%eax
+ shll $4,%ecx
+ shll $4,%eax
+ movdqu (%ebx),%xmm2
+ movdqu (%ebx,%ecx,1),%xmm3
+ movl 116(%esp),%ecx
+ movdqa %xmm2,%xmm4
+ movdqu (%ebx,%eax,1),%xmm5
+ movdqa %xmm2,%xmm6
+ pxor %xmm0,%xmm2
+ pxor %xmm2,%xmm3
+ movdqa %xmm2,(%esp)
+ pxor %xmm3,%xmm4
+ movdqa %xmm3,16(%esp)
+ pxor %xmm4,%xmm5
+ movdqa %xmm4,32(%esp)
+ pxor %xmm5,%xmm6
+ movdqa %xmm5,48(%esp)
+ pxor %xmm6,%xmm7
+ movdqa %xmm6,64(%esp)
+ movups -48(%edx,%ecx,1),%xmm0
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ pxor %xmm7,%xmm7
+ pxor %xmm2,%xmm1
+ pxor %xmm0,%xmm2
+ pxor %xmm3,%xmm1
+ pxor %xmm0,%xmm3
+ pxor %xmm4,%xmm1
+ pxor %xmm0,%xmm4
+ pxor %xmm5,%xmm1
+ pxor %xmm0,%xmm5
+ pxor %xmm6,%xmm1
+ pxor %xmm0,%xmm6
+ movdqa %xmm1,96(%esp)
+ movups -32(%edx,%ecx,1),%xmm1
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+ movups -16(%edx,%ecx,1),%xmm0
+.byte 102,15,56,220,209
+.byte 102,15,56,220,217
+.byte 102,15,56,220,225
+.byte 102,15,56,220,233
+.byte 102,15,56,220,241
+.byte 102,15,56,220,249
+ movl 120(%esp),%edi
+ call .L_aesni_encrypt6_enter
+ movdqa 64(%esp),%xmm0
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor %xmm0,%xmm6
+ movdqa 96(%esp),%xmm1
+ movdqu %xmm2,(%edi,%esi,1)
+ movdqu %xmm3,16(%edi,%esi,1)
+ movdqu %xmm4,32(%edi,%esi,1)
+ movdqu %xmm5,48(%edi,%esi,1)
+ movdqu %xmm6,64(%edi,%esi,1)
+ jmp .L078done
+.align 16
+.L079one:
+ movdqu (%ebx),%xmm7
+ movl 112(%esp),%edx
+ movdqu (%esi),%xmm2
+ movl 240(%edx),%ecx
+ pxor %xmm0,%xmm7
+ pxor %xmm2,%xmm1
+ pxor %xmm7,%xmm2
+ movdqa %xmm1,%xmm6
+ movl 120(%esp),%edi
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L083enc1_loop_16:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L083enc1_loop_16
+.byte 102,15,56,221,209
+ xorps %xmm7,%xmm2
+ movdqa %xmm7,%xmm0
+ movdqa %xmm6,%xmm1
+ movups %xmm2,(%edi,%esi,1)
+ jmp .L078done
+.align 16
+.L080two:
+ leal 1(%ebp),%ecx
+ movl 112(%esp),%edx
+ bsfl %ecx,%ecx
+ shll $4,%ecx
+ movdqu (%ebx),%xmm6
+ movdqu (%ebx,%ecx,1),%xmm7
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movl 240(%edx),%ecx
+ pxor %xmm0,%xmm6
+ pxor %xmm6,%xmm7
+ pxor %xmm2,%xmm1
+ pxor %xmm6,%xmm2
+ pxor %xmm3,%xmm1
+ pxor %xmm7,%xmm3
+ movdqa %xmm1,%xmm5
+ movl 120(%esp),%edi
+ call _aesni_encrypt2
+ xorps %xmm6,%xmm2
+ xorps %xmm7,%xmm3
+ movdqa %xmm7,%xmm0
+ movdqa %xmm5,%xmm1
+ movups %xmm2,(%edi,%esi,1)
+ movups %xmm3,16(%edi,%esi,1)
+ jmp .L078done
+.align 16
+.L081three:
+ leal 1(%ebp),%ecx
+ movl 112(%esp),%edx
+ bsfl %ecx,%ecx
+ shll $4,%ecx
+ movdqu (%ebx),%xmm5
+ movdqu (%ebx,%ecx,1),%xmm6
+ movdqa %xmm5,%xmm7
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movl 240(%edx),%ecx
+ pxor %xmm0,%xmm5
+ pxor %xmm5,%xmm6
+ pxor %xmm6,%xmm7
+ pxor %xmm2,%xmm1
+ pxor %xmm5,%xmm2
+ pxor %xmm3,%xmm1
+ pxor %xmm6,%xmm3
+ pxor %xmm4,%xmm1
+ pxor %xmm7,%xmm4
+ movdqa %xmm1,96(%esp)
+ movl 120(%esp),%edi
+ call _aesni_encrypt3
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm7,%xmm4
+ movdqa %xmm7,%xmm0
+ movdqa 96(%esp),%xmm1
+ movups %xmm2,(%edi,%esi,1)
+ movups %xmm3,16(%edi,%esi,1)
+ movups %xmm4,32(%edi,%esi,1)
+ jmp .L078done
+.align 16
+.L082four:
+ leal 1(%ebp),%ecx
+ leal 3(%ebp),%eax
+ bsfl %ecx,%ecx
+ bsfl %eax,%eax
+ movl 112(%esp),%edx
+ shll $4,%ecx
+ shll $4,%eax
+ movdqu (%ebx),%xmm4
+ movdqu (%ebx,%ecx,1),%xmm5
+ movdqa %xmm4,%xmm6
+ movdqu (%ebx,%eax,1),%xmm7
+ pxor %xmm0,%xmm4
+ movdqu (%esi),%xmm2
+ pxor %xmm4,%xmm5
+ movdqu 16(%esi),%xmm3
+ pxor %xmm5,%xmm6
+ movdqa %xmm4,(%esp)
+ pxor %xmm6,%xmm7
+ movdqa %xmm5,16(%esp)
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movl 240(%edx),%ecx
+ pxor %xmm2,%xmm1
+ pxor (%esp),%xmm2
+ pxor %xmm3,%xmm1
+ pxor 16(%esp),%xmm3
+ pxor %xmm4,%xmm1
+ pxor %xmm6,%xmm4
+ pxor %xmm5,%xmm1
+ pxor %xmm7,%xmm5
+ movdqa %xmm1,96(%esp)
+ movl 120(%esp),%edi
+ call _aesni_encrypt4
+ xorps (%esp),%xmm2
+ xorps 16(%esp),%xmm3
+ xorps %xmm6,%xmm4
+ movups %xmm2,(%edi,%esi,1)
+ xorps %xmm7,%xmm5
+ movups %xmm3,16(%edi,%esi,1)
+ movdqa %xmm7,%xmm0
+ movups %xmm4,32(%edi,%esi,1)
+ movdqa 96(%esp),%xmm1
+ movups %xmm5,48(%edi,%esi,1)
+.L078done:
+ movl 128(%esp),%edx
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ movdqa %xmm2,(%esp)
+ pxor %xmm4,%xmm4
+ movdqa %xmm2,16(%esp)
+ pxor %xmm5,%xmm5
+ movdqa %xmm2,32(%esp)
+ pxor %xmm6,%xmm6
+ movdqa %xmm2,48(%esp)
+ pxor %xmm7,%xmm7
+ movdqa %xmm2,64(%esp)
+ movdqa %xmm2,80(%esp)
+ movdqa %xmm2,96(%esp)
+ leal (%edx),%esp
+ movl 40(%esp),%ecx
+ movl 48(%esp),%ebx
+ movdqu %xmm0,(%ecx)
+ pxor %xmm0,%xmm0
+ movdqu %xmm1,(%ebx)
+ pxor %xmm1,%xmm1
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_ocb_encrypt,.-.L_aesni_ocb_encrypt_begin
+.globl aesni_ocb_decrypt
+.type aesni_ocb_decrypt,@function
+.align 16
+aesni_ocb_decrypt:
+.L_aesni_ocb_decrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 40(%esp),%ecx
+ movl 48(%esp),%ebx
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ movdqu (%ecx),%xmm0
+ movl 36(%esp),%ebp
+ movdqu (%ebx),%xmm1
+ movl 44(%esp),%ebx
+ movl %esp,%ecx
+ subl $132,%esp
+ andl $-16,%esp
+ subl %esi,%edi
+ shll $4,%eax
+ leal -96(%esi,%eax,1),%eax
+ movl %edi,120(%esp)
+ movl %eax,124(%esp)
+ movl %ecx,128(%esp)
+ movl 240(%edx),%ecx
+ testl $1,%ebp
+ jnz .L084odd
+ bsfl %ebp,%eax
+ addl $1,%ebp
+ shll $4,%eax
+ movdqu (%ebx,%eax,1),%xmm7
+ movl %edx,%eax
+ movdqu (%esi),%xmm2
+ leal 16(%esi),%esi
+ pxor %xmm0,%xmm7
+ pxor %xmm7,%xmm2
+ movdqa %xmm1,%xmm6
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L085dec1_loop_17:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L085dec1_loop_17
+.byte 102,15,56,223,209
+ xorps %xmm7,%xmm2
+ movaps %xmm6,%xmm1
+ movdqa %xmm7,%xmm0
+ xorps %xmm2,%xmm1
+ movups %xmm2,-16(%edi,%esi,1)
+ movl 240(%eax),%ecx
+ movl %eax,%edx
+ movl 124(%esp),%eax
+.L084odd:
+ shll $4,%ecx
+ movl $16,%edi
+ subl %ecx,%edi
+ movl %edx,112(%esp)
+ leal 32(%edx,%ecx,1),%edx
+ movl %edi,116(%esp)
+ cmpl %eax,%esi
+ ja .L086short
+ jmp .L087grandloop
+.align 32
+.L087grandloop:
+ leal 1(%ebp),%ecx
+ leal 3(%ebp),%eax
+ leal 5(%ebp),%edi
+ addl $6,%ebp
+ bsfl %ecx,%ecx
+ bsfl %eax,%eax
+ bsfl %edi,%edi
+ shll $4,%ecx
+ shll $4,%eax
+ shll $4,%edi
+ movdqu (%ebx),%xmm2
+ movdqu (%ebx,%ecx,1),%xmm3
+ movl 116(%esp),%ecx
+ movdqa %xmm2,%xmm4
+ movdqu (%ebx,%eax,1),%xmm5
+ movdqa %xmm2,%xmm6
+ movdqu (%ebx,%edi,1),%xmm7
+ pxor %xmm0,%xmm2
+ pxor %xmm2,%xmm3
+ movdqa %xmm2,(%esp)
+ pxor %xmm3,%xmm4
+ movdqa %xmm3,16(%esp)
+ pxor %xmm4,%xmm5
+ movdqa %xmm4,32(%esp)
+ pxor %xmm5,%xmm6
+ movdqa %xmm5,48(%esp)
+ pxor %xmm6,%xmm7
+ movdqa %xmm6,64(%esp)
+ movdqa %xmm7,80(%esp)
+ movups -48(%edx,%ecx,1),%xmm0
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ movdqu 80(%esi),%xmm7
+ leal 96(%esi),%esi
+ movdqa %xmm1,96(%esp)
+ pxor %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+ pxor %xmm0,%xmm5
+ pxor %xmm0,%xmm6
+ pxor %xmm0,%xmm7
+ movups -32(%edx,%ecx,1),%xmm1
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+ pxor 80(%esp),%xmm7
+ movups -16(%edx,%ecx,1),%xmm0
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.byte 102,15,56,222,233
+.byte 102,15,56,222,241
+.byte 102,15,56,222,249
+ movl 120(%esp),%edi
+ movl 124(%esp),%eax
+ call .L_aesni_decrypt6_enter
+ movdqa 80(%esp),%xmm0
+ pxor (%esp),%xmm2
+ movdqa 96(%esp),%xmm1
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+ pxor %xmm0,%xmm7
+ pxor %xmm2,%xmm1
+ movdqu %xmm2,-96(%edi,%esi,1)
+ pxor %xmm3,%xmm1
+ movdqu %xmm3,-80(%edi,%esi,1)
+ pxor %xmm4,%xmm1
+ movdqu %xmm4,-64(%edi,%esi,1)
+ pxor %xmm5,%xmm1
+ movdqu %xmm5,-48(%edi,%esi,1)
+ pxor %xmm6,%xmm1
+ movdqu %xmm6,-32(%edi,%esi,1)
+ pxor %xmm7,%xmm1
+ movdqu %xmm7,-16(%edi,%esi,1)
+ cmpl %eax,%esi
+ jb .L087grandloop
+.L086short:
+ addl $96,%eax
+ subl %esi,%eax
+ jz .L088done
+ cmpl $32,%eax
+ jb .L089one
+ je .L090two
+ cmpl $64,%eax
+ jb .L091three
+ je .L092four
+ leal 1(%ebp),%ecx
+ leal 3(%ebp),%eax
+ bsfl %ecx,%ecx
+ bsfl %eax,%eax
+ shll $4,%ecx
+ shll $4,%eax
+ movdqu (%ebx),%xmm2
+ movdqu (%ebx,%ecx,1),%xmm3
+ movl 116(%esp),%ecx
+ movdqa %xmm2,%xmm4
+ movdqu (%ebx,%eax,1),%xmm5
+ movdqa %xmm2,%xmm6
+ pxor %xmm0,%xmm2
+ pxor %xmm2,%xmm3
+ movdqa %xmm2,(%esp)
+ pxor %xmm3,%xmm4
+ movdqa %xmm3,16(%esp)
+ pxor %xmm4,%xmm5
+ movdqa %xmm4,32(%esp)
+ pxor %xmm5,%xmm6
+ movdqa %xmm5,48(%esp)
+ pxor %xmm6,%xmm7
+ movdqa %xmm6,64(%esp)
+ movups -48(%edx,%ecx,1),%xmm0
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ pxor %xmm7,%xmm7
+ movdqa %xmm1,96(%esp)
+ pxor %xmm0,%xmm2
+ pxor %xmm0,%xmm3
+ pxor %xmm0,%xmm4
+ pxor %xmm0,%xmm5
+ pxor %xmm0,%xmm6
+ movups -32(%edx,%ecx,1),%xmm1
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor 64(%esp),%xmm6
+ movups -16(%edx,%ecx,1),%xmm0
+.byte 102,15,56,222,209
+.byte 102,15,56,222,217
+.byte 102,15,56,222,225
+.byte 102,15,56,222,233
+.byte 102,15,56,222,241
+.byte 102,15,56,222,249
+ movl 120(%esp),%edi
+ call .L_aesni_decrypt6_enter
+ movdqa 64(%esp),%xmm0
+ pxor (%esp),%xmm2
+ movdqa 96(%esp),%xmm1
+ pxor 16(%esp),%xmm3
+ pxor 32(%esp),%xmm4
+ pxor 48(%esp),%xmm5
+ pxor %xmm0,%xmm6
+ pxor %xmm2,%xmm1
+ movdqu %xmm2,(%edi,%esi,1)
+ pxor %xmm3,%xmm1
+ movdqu %xmm3,16(%edi,%esi,1)
+ pxor %xmm4,%xmm1
+ movdqu %xmm4,32(%edi,%esi,1)
+ pxor %xmm5,%xmm1
+ movdqu %xmm5,48(%edi,%esi,1)
+ pxor %xmm6,%xmm1
+ movdqu %xmm6,64(%edi,%esi,1)
+ jmp .L088done
+.align 16
+.L089one:
+ movdqu (%ebx),%xmm7
+ movl 112(%esp),%edx
+ movdqu (%esi),%xmm2
+ movl 240(%edx),%ecx
+ pxor %xmm0,%xmm7
+ pxor %xmm7,%xmm2
+ movdqa %xmm1,%xmm6
+ movl 120(%esp),%edi
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L093dec1_loop_18:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L093dec1_loop_18
+.byte 102,15,56,223,209
+ xorps %xmm7,%xmm2
+ movaps %xmm6,%xmm1
+ movdqa %xmm7,%xmm0
+ xorps %xmm2,%xmm1
+ movups %xmm2,(%edi,%esi,1)
+ jmp .L088done
+.align 16
+.L090two:
+ leal 1(%ebp),%ecx
+ movl 112(%esp),%edx
+ bsfl %ecx,%ecx
+ shll $4,%ecx
+ movdqu (%ebx),%xmm6
+ movdqu (%ebx,%ecx,1),%xmm7
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movl 240(%edx),%ecx
+ movdqa %xmm1,%xmm5
+ pxor %xmm0,%xmm6
+ pxor %xmm6,%xmm7
+ pxor %xmm6,%xmm2
+ pxor %xmm7,%xmm3
+ movl 120(%esp),%edi
+ call _aesni_decrypt2
+ xorps %xmm6,%xmm2
+ xorps %xmm7,%xmm3
+ movdqa %xmm7,%xmm0
+ xorps %xmm2,%xmm5
+ movups %xmm2,(%edi,%esi,1)
+ xorps %xmm3,%xmm5
+ movups %xmm3,16(%edi,%esi,1)
+ movaps %xmm5,%xmm1
+ jmp .L088done
+.align 16
+.L091three:
+ leal 1(%ebp),%ecx
+ movl 112(%esp),%edx
+ bsfl %ecx,%ecx
+ shll $4,%ecx
+ movdqu (%ebx),%xmm5
+ movdqu (%ebx,%ecx,1),%xmm6
+ movdqa %xmm5,%xmm7
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movl 240(%edx),%ecx
+ movdqa %xmm1,96(%esp)
+ pxor %xmm0,%xmm5
+ pxor %xmm5,%xmm6
+ pxor %xmm6,%xmm7
+ pxor %xmm5,%xmm2
+ pxor %xmm6,%xmm3
+ pxor %xmm7,%xmm4
+ movl 120(%esp),%edi
+ call _aesni_decrypt3
+ movdqa 96(%esp),%xmm1
+ xorps %xmm5,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm7,%xmm4
+ movups %xmm2,(%edi,%esi,1)
+ pxor %xmm2,%xmm1
+ movdqa %xmm7,%xmm0
+ movups %xmm3,16(%edi,%esi,1)
+ pxor %xmm3,%xmm1
+ movups %xmm4,32(%edi,%esi,1)
+ pxor %xmm4,%xmm1
+ jmp .L088done
+.align 16
+.L092four:
+ leal 1(%ebp),%ecx
+ leal 3(%ebp),%eax
+ bsfl %ecx,%ecx
+ bsfl %eax,%eax
+ movl 112(%esp),%edx
+ shll $4,%ecx
+ shll $4,%eax
+ movdqu (%ebx),%xmm4
+ movdqu (%ebx,%ecx,1),%xmm5
+ movdqa %xmm4,%xmm6
+ movdqu (%ebx,%eax,1),%xmm7
+ pxor %xmm0,%xmm4
+ movdqu (%esi),%xmm2
+ pxor %xmm4,%xmm5
+ movdqu 16(%esi),%xmm3
+ pxor %xmm5,%xmm6
+ movdqa %xmm4,(%esp)
+ pxor %xmm6,%xmm7
+ movdqa %xmm5,16(%esp)
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movl 240(%edx),%ecx
+ movdqa %xmm1,96(%esp)
+ pxor (%esp),%xmm2
+ pxor 16(%esp),%xmm3
+ pxor %xmm6,%xmm4
+ pxor %xmm7,%xmm5
+ movl 120(%esp),%edi
+ call _aesni_decrypt4
+ movdqa 96(%esp),%xmm1
+ xorps (%esp),%xmm2
+ xorps 16(%esp),%xmm3
+ xorps %xmm6,%xmm4
+ movups %xmm2,(%edi,%esi,1)
+ pxor %xmm2,%xmm1
+ xorps %xmm7,%xmm5
+ movups %xmm3,16(%edi,%esi,1)
+ pxor %xmm3,%xmm1
+ movdqa %xmm7,%xmm0
+ movups %xmm4,32(%edi,%esi,1)
+ pxor %xmm4,%xmm1
+ movups %xmm5,48(%edi,%esi,1)
+ pxor %xmm5,%xmm1
+.L088done:
+ movl 128(%esp),%edx
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ movdqa %xmm2,(%esp)
+ pxor %xmm4,%xmm4
+ movdqa %xmm2,16(%esp)
+ pxor %xmm5,%xmm5
+ movdqa %xmm2,32(%esp)
+ pxor %xmm6,%xmm6
+ movdqa %xmm2,48(%esp)
+ pxor %xmm7,%xmm7
+ movdqa %xmm2,64(%esp)
+ movdqa %xmm2,80(%esp)
+ movdqa %xmm2,96(%esp)
+ leal (%edx),%esp
+ movl 40(%esp),%ecx
+ movl 48(%esp),%ebx
+ movdqu %xmm0,(%ecx)
+ pxor %xmm0,%xmm0
+ movdqu %xmm1,(%ebx)
+ pxor %xmm1,%xmm1
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_ocb_decrypt,.-.L_aesni_ocb_decrypt_begin
+.globl aesni_cbc_encrypt
+.type aesni_cbc_encrypt,@function
+.align 16
+aesni_cbc_encrypt:
+.L_aesni_cbc_encrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl %esp,%ebx
+ movl 24(%esp),%edi
+ subl $24,%ebx
+ movl 28(%esp),%eax
+ andl $-16,%ebx
+ movl 32(%esp),%edx
+ movl 36(%esp),%ebp
+ testl %eax,%eax
+ jz .L094cbc_abort
+ cmpl $0,40(%esp)
+ xchgl %esp,%ebx
+ movups (%ebp),%xmm7
+ movl 240(%edx),%ecx
+ movl %edx,%ebp
+ movl %ebx,16(%esp)
+ movl %ecx,%ebx
+ je .L095cbc_decrypt
+ movaps %xmm7,%xmm2
+ cmpl $16,%eax
+ jb .L096cbc_enc_tail
+ subl $16,%eax
+ jmp .L097cbc_enc_loop
+.align 16
+.L097cbc_enc_loop:
+ movups (%esi),%xmm7
+ leal 16(%esi),%esi
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ xorps %xmm0,%xmm7
+ leal 32(%edx),%edx
+ xorps %xmm7,%xmm2
+.L098enc1_loop_19:
+.byte 102,15,56,220,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L098enc1_loop_19
+.byte 102,15,56,221,209
+ movl %ebx,%ecx
+ movl %ebp,%edx
+ movups %xmm2,(%edi)
+ leal 16(%edi),%edi
+ subl $16,%eax
+ jnc .L097cbc_enc_loop
+ addl $16,%eax
+ jnz .L096cbc_enc_tail
+ movaps %xmm2,%xmm7
+ pxor %xmm2,%xmm2
+ jmp .L099cbc_ret
+.L096cbc_enc_tail:
+ movl %eax,%ecx
+.long 2767451785
+ movl $16,%ecx
+ subl %eax,%ecx
+ xorl %eax,%eax
+.long 2868115081
+ leal -16(%edi),%edi
+ movl %ebx,%ecx
+ movl %edi,%esi
+ movl %ebp,%edx
+ jmp .L097cbc_enc_loop
+.align 16
+.L095cbc_decrypt:
+ cmpl $80,%eax
+ jbe .L100cbc_dec_tail
+ movaps %xmm7,(%esp)
+ subl $80,%eax
+ jmp .L101cbc_dec_loop6_enter
+.align 16
+.L102cbc_dec_loop6:
+ movaps %xmm0,(%esp)
+ movups %xmm7,(%edi)
+ leal 16(%edi),%edi
+.L101cbc_dec_loop6_enter:
+ movdqu (%esi),%xmm2
+ movdqu 16(%esi),%xmm3
+ movdqu 32(%esi),%xmm4
+ movdqu 48(%esi),%xmm5
+ movdqu 64(%esi),%xmm6
+ movdqu 80(%esi),%xmm7
+ call _aesni_decrypt6
+ movups (%esi),%xmm1
+ movups 16(%esi),%xmm0
+ xorps (%esp),%xmm2
+ xorps %xmm1,%xmm3
+ movups 32(%esi),%xmm1
+ xorps %xmm0,%xmm4
+ movups 48(%esi),%xmm0
+ xorps %xmm1,%xmm5
+ movups 64(%esi),%xmm1
+ xorps %xmm0,%xmm6
+ movups 80(%esi),%xmm0
+ xorps %xmm1,%xmm7
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ leal 96(%esi),%esi
+ movups %xmm4,32(%edi)
+ movl %ebx,%ecx
+ movups %xmm5,48(%edi)
+ movl %ebp,%edx
+ movups %xmm6,64(%edi)
+ leal 80(%edi),%edi
+ subl $96,%eax
+ ja .L102cbc_dec_loop6
+ movaps %xmm7,%xmm2
+ movaps %xmm0,%xmm7
+ addl $80,%eax
+ jle .L103cbc_dec_clear_tail_collected
+ movups %xmm2,(%edi)
+ leal 16(%edi),%edi
+.L100cbc_dec_tail:
+ movups (%esi),%xmm2
+ movaps %xmm2,%xmm6
+ cmpl $16,%eax
+ jbe .L104cbc_dec_one
+ movups 16(%esi),%xmm3
+ movaps %xmm3,%xmm5
+ cmpl $32,%eax
+ jbe .L105cbc_dec_two
+ movups 32(%esi),%xmm4
+ cmpl $48,%eax
+ jbe .L106cbc_dec_three
+ movups 48(%esi),%xmm5
+ cmpl $64,%eax
+ jbe .L107cbc_dec_four
+ movups 64(%esi),%xmm6
+ movaps %xmm7,(%esp)
+ movups (%esi),%xmm2
+ xorps %xmm7,%xmm7
+ call _aesni_decrypt6
+ movups (%esi),%xmm1
+ movups 16(%esi),%xmm0
+ xorps (%esp),%xmm2
+ xorps %xmm1,%xmm3
+ movups 32(%esi),%xmm1
+ xorps %xmm0,%xmm4
+ movups 48(%esi),%xmm0
+ xorps %xmm1,%xmm5
+ movups 64(%esi),%xmm7
+ xorps %xmm0,%xmm6
+ movups %xmm2,(%edi)
+ movups %xmm3,16(%edi)
+ pxor %xmm3,%xmm3
+ movups %xmm4,32(%edi)
+ pxor %xmm4,%xmm4
+ movups %xmm5,48(%edi)
+ pxor %xmm5,%xmm5
+ leal 64(%edi),%edi
+ movaps %xmm6,%xmm2
+ pxor %xmm6,%xmm6
+ subl $80,%eax
+ jmp .L108cbc_dec_tail_collected
+.align 16
+.L104cbc_dec_one:
+ movups (%edx),%xmm0
+ movups 16(%edx),%xmm1
+ leal 32(%edx),%edx
+ xorps %xmm0,%xmm2
+.L109dec1_loop_20:
+.byte 102,15,56,222,209
+ decl %ecx
+ movups (%edx),%xmm1
+ leal 16(%edx),%edx
+ jnz .L109dec1_loop_20
+.byte 102,15,56,223,209
+ xorps %xmm7,%xmm2
+ movaps %xmm6,%xmm7
+ subl $16,%eax
+ jmp .L108cbc_dec_tail_collected
+.align 16
+.L105cbc_dec_two:
+ call _aesni_decrypt2
+ xorps %xmm7,%xmm2
+ xorps %xmm6,%xmm3
+ movups %xmm2,(%edi)
+ movaps %xmm3,%xmm2
+ pxor %xmm3,%xmm3
+ leal 16(%edi),%edi
+ movaps %xmm5,%xmm7
+ subl $32,%eax
+ jmp .L108cbc_dec_tail_collected
+.align 16
+.L106cbc_dec_three:
+ call _aesni_decrypt3
+ xorps %xmm7,%xmm2
+ xorps %xmm6,%xmm3
+ xorps %xmm5,%xmm4
+ movups %xmm2,(%edi)
+ movaps %xmm4,%xmm2
+ pxor %xmm4,%xmm4
+ movups %xmm3,16(%edi)
+ pxor %xmm3,%xmm3
+ leal 32(%edi),%edi
+ movups 32(%esi),%xmm7
+ subl $48,%eax
+ jmp .L108cbc_dec_tail_collected
+.align 16
+.L107cbc_dec_four:
+ call _aesni_decrypt4
+ movups 16(%esi),%xmm1
+ movups 32(%esi),%xmm0
+ xorps %xmm7,%xmm2
+ movups 48(%esi),%xmm7
+ xorps %xmm6,%xmm3
+ movups %xmm2,(%edi)
+ xorps %xmm1,%xmm4
+ movups %xmm3,16(%edi)
+ pxor %xmm3,%xmm3
+ xorps %xmm0,%xmm5
+ movups %xmm4,32(%edi)
+ pxor %xmm4,%xmm4
+ leal 48(%edi),%edi
+ movaps %xmm5,%xmm2
+ pxor %xmm5,%xmm5
+ subl $64,%eax
+ jmp .L108cbc_dec_tail_collected
+.align 16
+.L103cbc_dec_clear_tail_collected:
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ pxor %xmm5,%xmm5
+ pxor %xmm6,%xmm6
+.L108cbc_dec_tail_collected:
+ andl $15,%eax
+ jnz .L110cbc_dec_tail_partial
+ movups %xmm2,(%edi)
+ pxor %xmm0,%xmm0
+ jmp .L099cbc_ret
+.align 16
+.L110cbc_dec_tail_partial:
+ movaps %xmm2,(%esp)
+ pxor %xmm0,%xmm0
+ movl $16,%ecx
+ movl %esp,%esi
+ subl %eax,%ecx
+.long 2767451785
+ movdqa %xmm2,(%esp)
+.L099cbc_ret:
+ movl 16(%esp),%esp
+ movl 36(%esp),%ebp
+ pxor %xmm2,%xmm2
+ pxor %xmm1,%xmm1
+ movups %xmm7,(%ebp)
+ pxor %xmm7,%xmm7
+.L094cbc_abort:
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size aesni_cbc_encrypt,.-.L_aesni_cbc_encrypt_begin
+.type _aesni_set_encrypt_key,@function
+.align 16
+_aesni_set_encrypt_key:
+ pushl %ebp
+ pushl %ebx
+ testl %eax,%eax
+ jz .L111bad_pointer
+ testl %edx,%edx
+ jz .L111bad_pointer
+ call .L112pic
+.L112pic:
+ popl %ebx
+ leal .Lkey_const-.L112pic(%ebx),%ebx
+ leal OPENSSL_ia32cap_P,%ebp
+ movups (%eax),%xmm0
+ xorps %xmm4,%xmm4
+ movl 4(%ebp),%ebp
+ leal 16(%edx),%edx
+ andl $268437504,%ebp
+ cmpl $256,%ecx
+ je .L11314rounds
+ cmpl $192,%ecx
+ je .L11412rounds
+ cmpl $128,%ecx
+ jne .L115bad_keybits
+.align 16
+.L11610rounds:
+ cmpl $268435456,%ebp
+ je .L11710rounds_alt
+ movl $9,%ecx
+ movups %xmm0,-16(%edx)
+.byte 102,15,58,223,200,1
+ call .L118key_128_cold
+.byte 102,15,58,223,200,2
+ call .L119key_128
+.byte 102,15,58,223,200,4
+ call .L119key_128
+.byte 102,15,58,223,200,8
+ call .L119key_128
+.byte 102,15,58,223,200,16
+ call .L119key_128
+.byte 102,15,58,223,200,32
+ call .L119key_128
+.byte 102,15,58,223,200,64
+ call .L119key_128
+.byte 102,15,58,223,200,128
+ call .L119key_128
+.byte 102,15,58,223,200,27
+ call .L119key_128
+.byte 102,15,58,223,200,54
+ call .L119key_128
+ movups %xmm0,(%edx)
+ movl %ecx,80(%edx)
+ jmp .L120good_key
+.align 16
+.L119key_128:
+ movups %xmm0,(%edx)
+ leal 16(%edx),%edx
+.L118key_128_cold:
+ shufps $16,%xmm0,%xmm4
+ xorps %xmm4,%xmm0
+ shufps $140,%xmm0,%xmm4
+ xorps %xmm4,%xmm0
+ shufps $255,%xmm1,%xmm1
+ xorps %xmm1,%xmm0
+ ret
+.align 16
+.L11710rounds_alt:
+ movdqa (%ebx),%xmm5
+ movl $8,%ecx
+ movdqa 32(%ebx),%xmm4
+ movdqa %xmm0,%xmm2
+ movdqu %xmm0,-16(%edx)
+.L121loop_key128:
+.byte 102,15,56,0,197
+.byte 102,15,56,221,196
+ pslld $1,%xmm4
+ leal 16(%edx),%edx
+ movdqa %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm3,%xmm2
+ pxor %xmm2,%xmm0
+ movdqu %xmm0,-16(%edx)
+ movdqa %xmm0,%xmm2
+ decl %ecx
+ jnz .L121loop_key128
+ movdqa 48(%ebx),%xmm4
+.byte 102,15,56,0,197
+.byte 102,15,56,221,196
+ pslld $1,%xmm4
+ movdqa %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm3,%xmm2
+ pxor %xmm2,%xmm0
+ movdqu %xmm0,(%edx)
+ movdqa %xmm0,%xmm2
+.byte 102,15,56,0,197
+.byte 102,15,56,221,196
+ movdqa %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm2,%xmm3
+ pslldq $4,%xmm2
+ pxor %xmm3,%xmm2
+ pxor %xmm2,%xmm0
+ movdqu %xmm0,16(%edx)
+ movl $9,%ecx
+ movl %ecx,96(%edx)
+ jmp .L120good_key
+.align 16
+.L11412rounds:
+ movq 16(%eax),%xmm2
+ cmpl $268435456,%ebp
+ je .L12212rounds_alt
+ movl $11,%ecx
+ movups %xmm0,-16(%edx)
+.byte 102,15,58,223,202,1
+ call .L123key_192a_cold
+.byte 102,15,58,223,202,2
+ call .L124key_192b
+.byte 102,15,58,223,202,4
+ call .L125key_192a
+.byte 102,15,58,223,202,8
+ call .L124key_192b
+.byte 102,15,58,223,202,16
+ call .L125key_192a
+.byte 102,15,58,223,202,32
+ call .L124key_192b
+.byte 102,15,58,223,202,64
+ call .L125key_192a
+.byte 102,15,58,223,202,128
+ call .L124key_192b
+ movups %xmm0,(%edx)
+ movl %ecx,48(%edx)
+ jmp .L120good_key
+.align 16
+.L125key_192a:
+ movups %xmm0,(%edx)
+ leal 16(%edx),%edx
+.align 16
+.L123key_192a_cold:
+ movaps %xmm2,%xmm5
+.L126key_192b_warm:
+ shufps $16,%xmm0,%xmm4
+ movdqa %xmm2,%xmm3
+ xorps %xmm4,%xmm0
+ shufps $140,%xmm0,%xmm4
+ pslldq $4,%xmm3
+ xorps %xmm4,%xmm0
+ pshufd $85,%xmm1,%xmm1
+ pxor %xmm3,%xmm2
+ pxor %xmm1,%xmm0
+ pshufd $255,%xmm0,%xmm3
+ pxor %xmm3,%xmm2
+ ret
+.align 16
+.L124key_192b:
+ movaps %xmm0,%xmm3
+ shufps $68,%xmm0,%xmm5
+ movups %xmm5,(%edx)
+ shufps $78,%xmm2,%xmm3
+ movups %xmm3,16(%edx)
+ leal 32(%edx),%edx
+ jmp .L126key_192b_warm
+.align 16
+.L12212rounds_alt:
+ movdqa 16(%ebx),%xmm5
+ movdqa 32(%ebx),%xmm4
+ movl $8,%ecx
+ movdqu %xmm0,-16(%edx)
+.L127loop_key192:
+ movq %xmm2,(%edx)
+ movdqa %xmm2,%xmm1
+.byte 102,15,56,0,213
+.byte 102,15,56,221,212
+ pslld $1,%xmm4
+ leal 24(%edx),%edx
+ movdqa %xmm0,%xmm3
+ pslldq $4,%xmm0
+ pxor %xmm0,%xmm3
+ pslldq $4,%xmm0
+ pxor %xmm0,%xmm3
+ pslldq $4,%xmm0
+ pxor %xmm3,%xmm0
+ pshufd $255,%xmm0,%xmm3
+ pxor %xmm1,%xmm3
+ pslldq $4,%xmm1
+ pxor %xmm1,%xmm3
+ pxor %xmm2,%xmm0
+ pxor %xmm3,%xmm2
+ movdqu %xmm0,-16(%edx)
+ decl %ecx
+ jnz .L127loop_key192
+ movl $11,%ecx
+ movl %ecx,32(%edx)
+ jmp .L120good_key
+.align 16
+.L11314rounds:
+ movups 16(%eax),%xmm2
+ leal 16(%edx),%edx
+ cmpl $268435456,%ebp
+ je .L12814rounds_alt
+ movl $13,%ecx
+ movups %xmm0,-32(%edx)
+ movups %xmm2,-16(%edx)
+.byte 102,15,58,223,202,1
+ call .L129key_256a_cold
+.byte 102,15,58,223,200,1
+ call .L130key_256b
+.byte 102,15,58,223,202,2
+ call .L131key_256a
+.byte 102,15,58,223,200,2
+ call .L130key_256b
+.byte 102,15,58,223,202,4
+ call .L131key_256a
+.byte 102,15,58,223,200,4
+ call .L130key_256b
+.byte 102,15,58,223,202,8
+ call .L131key_256a
+.byte 102,15,58,223,200,8
+ call .L130key_256b
+.byte 102,15,58,223,202,16
+ call .L131key_256a
+.byte 102,15,58,223,200,16
+ call .L130key_256b
+.byte 102,15,58,223,202,32
+ call .L131key_256a
+.byte 102,15,58,223,200,32
+ call .L130key_256b
+.byte 102,15,58,223,202,64
+ call .L131key_256a
+ movups %xmm0,(%edx)
+ movl %ecx,16(%edx)
+ xorl %eax,%eax
+ jmp .L120good_key
+.align 16
+.L131key_256a:
+ movups %xmm2,(%edx)
+ leal 16(%edx),%edx
+.L129key_256a_cold:
+ shufps $16,%xmm0,%xmm4
+ xorps %xmm4,%xmm0
+ shufps $140,%xmm0,%xmm4
+ xorps %xmm4,%xmm0
+ shufps $255,%xmm1,%xmm1
+ xorps %xmm1,%xmm0
+ ret
+.align 16
+.L130key_256b:
+ movups %xmm0,(%edx)
+ leal 16(%edx),%edx
+ shufps $16,%xmm2,%xmm4
+ xorps %xmm4,%xmm2
+ shufps $140,%xmm2,%xmm4
+ xorps %xmm4,%xmm2
+ shufps $170,%xmm1,%xmm1
+ xorps %xmm1,%xmm2
+ ret
+.align 16
+.L12814rounds_alt:
+ movdqa (%ebx),%xmm5
+ movdqa 32(%ebx),%xmm4
+ movl $7,%ecx
+ movdqu %xmm0,-32(%edx)
+ movdqa %xmm2,%xmm1
+ movdqu %xmm2,-16(%edx)
+.L132loop_key256:
+.byte 102,15,56,0,213
+.byte 102,15,56,221,212
+ movdqa %xmm0,%xmm3
+ pslldq $4,%xmm0
+ pxor %xmm0,%xmm3
+ pslldq $4,%xmm0
+ pxor %xmm0,%xmm3
+ pslldq $4,%xmm0
+ pxor %xmm3,%xmm0
+ pslld $1,%xmm4
+ pxor %xmm2,%xmm0
+ movdqu %xmm0,(%edx)
+ decl %ecx
+ jz .L133done_key256
+ pshufd $255,%xmm0,%xmm2
+ pxor %xmm3,%xmm3
+.byte 102,15,56,221,211
+ movdqa %xmm1,%xmm3
+ pslldq $4,%xmm1
+ pxor %xmm1,%xmm3
+ pslldq $4,%xmm1
+ pxor %xmm1,%xmm3
+ pslldq $4,%xmm1
+ pxor %xmm3,%xmm1
+ pxor %xmm1,%xmm2
+ movdqu %xmm2,16(%edx)
+ leal 32(%edx),%edx
+ movdqa %xmm2,%xmm1
+ jmp .L132loop_key256
+.L133done_key256:
+ movl $13,%ecx
+ movl %ecx,16(%edx)
+.L120good_key:
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ pxor %xmm5,%xmm5
+ xorl %eax,%eax
+ popl %ebx
+ popl %ebp
+ ret
+.align 4
+.L111bad_pointer:
+ movl $-1,%eax
+ popl %ebx
+ popl %ebp
+ ret
+.align 4
+.L115bad_keybits:
+ pxor %xmm0,%xmm0
+ movl $-2,%eax
+ popl %ebx
+ popl %ebp
+ ret
+.size _aesni_set_encrypt_key,.-_aesni_set_encrypt_key
+.globl aesni_set_encrypt_key
+.type aesni_set_encrypt_key,@function
+.align 16
+aesni_set_encrypt_key:
+.L_aesni_set_encrypt_key_begin:
+ movl 4(%esp),%eax
+ movl 8(%esp),%ecx
+ movl 12(%esp),%edx
+ call _aesni_set_encrypt_key
+ ret
+.size aesni_set_encrypt_key,.-.L_aesni_set_encrypt_key_begin
+.globl aesni_set_decrypt_key
+.type aesni_set_decrypt_key,@function
+.align 16
+aesni_set_decrypt_key:
+.L_aesni_set_decrypt_key_begin:
+ movl 4(%esp),%eax
+ movl 8(%esp),%ecx
+ movl 12(%esp),%edx
+ call _aesni_set_encrypt_key
+ movl 12(%esp),%edx
+ shll $4,%ecx
+ testl %eax,%eax
+ jnz .L134dec_key_ret
+ leal 16(%edx,%ecx,1),%eax
+ movups (%edx),%xmm0
+ movups (%eax),%xmm1
+ movups %xmm0,(%eax)
+ movups %xmm1,(%edx)
+ leal 16(%edx),%edx
+ leal -16(%eax),%eax
+.L135dec_key_inverse:
+ movups (%edx),%xmm0
+ movups (%eax),%xmm1
+.byte 102,15,56,219,192
+.byte 102,15,56,219,201
+ leal 16(%edx),%edx
+ leal -16(%eax),%eax
+ movups %xmm0,16(%eax)
+ movups %xmm1,-16(%edx)
+ cmpl %edx,%eax
+ ja .L135dec_key_inverse
+ movups (%edx),%xmm0
+.byte 102,15,56,219,192
+ movups %xmm0,(%edx)
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ xorl %eax,%eax
+.L134dec_key_ret:
+ ret
+.size aesni_set_decrypt_key,.-.L_aesni_set_decrypt_key_begin
+.align 64
+.Lkey_const:
+.long 202313229,202313229,202313229,202313229
+.long 67569157,67569157,67569157,67569157
+.long 1,1,1,1
+.long 27,27,27,27
+.byte 65,69,83,32,102,111,114,32,73,110,116,101,108,32,65,69
+.byte 83,45,78,73,44,32,67,82,89,80,84,79,71,65,77,83
+.byte 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
+.byte 115,108,46,111,114,103,62,0
+.comm OPENSSL_ia32cap_P,16,4
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/vpaes-x86.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/vpaes-x86.S
new file mode 100644
index 0000000000..6f62534682
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/aes/vpaes-x86.S
@@ -0,0 +1,670 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/aes/asm/vpaes-x86.pl
+#
+# Copyright 2011-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.align 64
+.L_vpaes_consts:
+.long 218628480,235210255,168496130,67568393
+.long 252381056,17041926,33884169,51187212
+.long 252645135,252645135,252645135,252645135
+.long 1512730624,3266504856,1377990664,3401244816
+.long 830229760,1275146365,2969422977,3447763452
+.long 3411033600,2979783055,338359620,2782886510
+.long 4209124096,907596821,221174255,1006095553
+.long 191964160,3799684038,3164090317,1589111125
+.long 182528256,1777043520,2877432650,3265356744
+.long 1874708224,3503451415,3305285752,363511674
+.long 1606117888,3487855781,1093350906,2384367825
+.long 197121,67569157,134941193,202313229
+.long 67569157,134941193,202313229,197121
+.long 134941193,202313229,197121,67569157
+.long 202313229,197121,67569157,134941193
+.long 33619971,100992007,168364043,235736079
+.long 235736079,33619971,100992007,168364043
+.long 168364043,235736079,33619971,100992007
+.long 100992007,168364043,235736079,33619971
+.long 50462976,117835012,185207048,252579084
+.long 252314880,51251460,117574920,184942860
+.long 184682752,252054788,50987272,118359308
+.long 118099200,185467140,251790600,50727180
+.long 2946363062,528716217,1300004225,1881839624
+.long 1532713819,1532713819,1532713819,1532713819
+.long 3602276352,4288629033,3737020424,4153884961
+.long 1354558464,32357713,2958822624,3775749553
+.long 1201988352,132424512,1572796698,503232858
+.long 2213177600,1597421020,4103937655,675398315
+.long 2749646592,4273543773,1511898873,121693092
+.long 3040248576,1103263732,2871565598,1608280554
+.long 2236667136,2588920351,482954393,64377734
+.long 3069987328,291237287,2117370568,3650299247
+.long 533321216,3573750986,2572112006,1401264716
+.long 1339849704,2721158661,548607111,3445553514
+.long 2128193280,3054596040,2183486460,1257083700
+.long 655635200,1165381986,3923443150,2344132524
+.long 190078720,256924420,290342170,357187870
+.long 1610966272,2263057382,4103205268,309794674
+.long 2592527872,2233205587,1335446729,3402964816
+.long 3973531904,3225098121,3002836325,1918774430
+.long 3870401024,2102906079,2284471353,4117666579
+.long 617007872,1021508343,366931923,691083277
+.long 2528395776,3491914898,2968704004,1613121270
+.long 3445188352,3247741094,844474987,4093578302
+.long 651481088,1190302358,1689581232,574775300
+.long 4289380608,206939853,2555985458,2489840491
+.long 2130264064,327674451,3566485037,3349835193
+.long 2470714624,316102159,3636825756,3393945945
+.byte 86,101,99,116,111,114,32,80,101,114,109,117,116,97,116,105
+.byte 111,110,32,65,69,83,32,102,111,114,32,120,56,54,47,83
+.byte 83,83,69,51,44,32,77,105,107,101,32,72,97,109,98,117
+.byte 114,103,32,40,83,116,97,110,102,111,114,100,32,85,110,105
+.byte 118,101,114,115,105,116,121,41,0
+.align 64
+.type _vpaes_preheat,@function
+.align 16
+_vpaes_preheat:
+ addl (%esp),%ebp
+ movdqa -48(%ebp),%xmm7
+ movdqa -16(%ebp),%xmm6
+ ret
+.size _vpaes_preheat,.-_vpaes_preheat
+.type _vpaes_encrypt_core,@function
+.align 16
+_vpaes_encrypt_core:
+ movl $16,%ecx
+ movl 240(%edx),%eax
+ movdqa %xmm6,%xmm1
+ movdqa (%ebp),%xmm2
+ pandn %xmm0,%xmm1
+ pand %xmm6,%xmm0
+ movdqu (%edx),%xmm5
+.byte 102,15,56,0,208
+ movdqa 16(%ebp),%xmm0
+ pxor %xmm5,%xmm2
+ psrld $4,%xmm1
+ addl $16,%edx
+.byte 102,15,56,0,193
+ leal 192(%ebp),%ebx
+ pxor %xmm2,%xmm0
+ jmp .L000enc_entry
+.align 16
+.L001enc_loop:
+ movdqa 32(%ebp),%xmm4
+ movdqa 48(%ebp),%xmm0
+.byte 102,15,56,0,226
+.byte 102,15,56,0,195
+ pxor %xmm5,%xmm4
+ movdqa 64(%ebp),%xmm5
+ pxor %xmm4,%xmm0
+ movdqa -64(%ebx,%ecx,1),%xmm1
+.byte 102,15,56,0,234
+ movdqa 80(%ebp),%xmm2
+ movdqa (%ebx,%ecx,1),%xmm4
+.byte 102,15,56,0,211
+ movdqa %xmm0,%xmm3
+ pxor %xmm5,%xmm2
+.byte 102,15,56,0,193
+ addl $16,%edx
+ pxor %xmm2,%xmm0
+.byte 102,15,56,0,220
+ addl $16,%ecx
+ pxor %xmm0,%xmm3
+.byte 102,15,56,0,193
+ andl $48,%ecx
+ subl $1,%eax
+ pxor %xmm3,%xmm0
+.L000enc_entry:
+ movdqa %xmm6,%xmm1
+ movdqa -32(%ebp),%xmm5
+ pandn %xmm0,%xmm1
+ psrld $4,%xmm1
+ pand %xmm6,%xmm0
+.byte 102,15,56,0,232
+ movdqa %xmm7,%xmm3
+ pxor %xmm1,%xmm0
+.byte 102,15,56,0,217
+ movdqa %xmm7,%xmm4
+ pxor %xmm5,%xmm3
+.byte 102,15,56,0,224
+ movdqa %xmm7,%xmm2
+ pxor %xmm5,%xmm4
+.byte 102,15,56,0,211
+ movdqa %xmm7,%xmm3
+ pxor %xmm0,%xmm2
+.byte 102,15,56,0,220
+ movdqu (%edx),%xmm5
+ pxor %xmm1,%xmm3
+ jnz .L001enc_loop
+ movdqa 96(%ebp),%xmm4
+ movdqa 112(%ebp),%xmm0
+.byte 102,15,56,0,226
+ pxor %xmm5,%xmm4
+.byte 102,15,56,0,195
+ movdqa 64(%ebx,%ecx,1),%xmm1
+ pxor %xmm4,%xmm0
+.byte 102,15,56,0,193
+ ret
+.size _vpaes_encrypt_core,.-_vpaes_encrypt_core
+.type _vpaes_decrypt_core,@function
+.align 16
+_vpaes_decrypt_core:
+ leal 608(%ebp),%ebx
+ movl 240(%edx),%eax
+ movdqa %xmm6,%xmm1
+ movdqa -64(%ebx),%xmm2
+ pandn %xmm0,%xmm1
+ movl %eax,%ecx
+ psrld $4,%xmm1
+ movdqu (%edx),%xmm5
+ shll $4,%ecx
+ pand %xmm6,%xmm0
+.byte 102,15,56,0,208
+ movdqa -48(%ebx),%xmm0
+ xorl $48,%ecx
+.byte 102,15,56,0,193
+ andl $48,%ecx
+ pxor %xmm5,%xmm2
+ movdqa 176(%ebp),%xmm5
+ pxor %xmm2,%xmm0
+ addl $16,%edx
+ leal -352(%ebx,%ecx,1),%ecx
+ jmp .L002dec_entry
+.align 16
+.L003dec_loop:
+ movdqa -32(%ebx),%xmm4
+ movdqa -16(%ebx),%xmm1
+.byte 102,15,56,0,226
+.byte 102,15,56,0,203
+ pxor %xmm4,%xmm0
+ movdqa (%ebx),%xmm4
+ pxor %xmm1,%xmm0
+ movdqa 16(%ebx),%xmm1
+.byte 102,15,56,0,226
+.byte 102,15,56,0,197
+.byte 102,15,56,0,203
+ pxor %xmm4,%xmm0
+ movdqa 32(%ebx),%xmm4
+ pxor %xmm1,%xmm0
+ movdqa 48(%ebx),%xmm1
+.byte 102,15,56,0,226
+.byte 102,15,56,0,197
+.byte 102,15,56,0,203
+ pxor %xmm4,%xmm0
+ movdqa 64(%ebx),%xmm4
+ pxor %xmm1,%xmm0
+ movdqa 80(%ebx),%xmm1
+.byte 102,15,56,0,226
+.byte 102,15,56,0,197
+.byte 102,15,56,0,203
+ pxor %xmm4,%xmm0
+ addl $16,%edx
+.byte 102,15,58,15,237,12
+ pxor %xmm1,%xmm0
+ subl $1,%eax
+.L002dec_entry:
+ movdqa %xmm6,%xmm1
+ movdqa -32(%ebp),%xmm2
+ pandn %xmm0,%xmm1
+ pand %xmm6,%xmm0
+ psrld $4,%xmm1
+.byte 102,15,56,0,208
+ movdqa %xmm7,%xmm3
+ pxor %xmm1,%xmm0
+.byte 102,15,56,0,217
+ movdqa %xmm7,%xmm4
+ pxor %xmm2,%xmm3
+.byte 102,15,56,0,224
+ pxor %xmm2,%xmm4
+ movdqa %xmm7,%xmm2
+.byte 102,15,56,0,211
+ movdqa %xmm7,%xmm3
+ pxor %xmm0,%xmm2
+.byte 102,15,56,0,220
+ movdqu (%edx),%xmm0
+ pxor %xmm1,%xmm3
+ jnz .L003dec_loop
+ movdqa 96(%ebx),%xmm4
+.byte 102,15,56,0,226
+ pxor %xmm0,%xmm4
+ movdqa 112(%ebx),%xmm0
+ movdqa (%ecx),%xmm2
+.byte 102,15,56,0,195
+ pxor %xmm4,%xmm0
+.byte 102,15,56,0,194
+ ret
+.size _vpaes_decrypt_core,.-_vpaes_decrypt_core
+.type _vpaes_schedule_core,@function
+.align 16
+_vpaes_schedule_core:
+ addl (%esp),%ebp
+ movdqu (%esi),%xmm0
+ movdqa 320(%ebp),%xmm2
+ movdqa %xmm0,%xmm3
+ leal (%ebp),%ebx
+ movdqa %xmm2,4(%esp)
+ call _vpaes_schedule_transform
+ movdqa %xmm0,%xmm7
+ testl %edi,%edi
+ jnz .L004schedule_am_decrypting
+ movdqu %xmm0,(%edx)
+ jmp .L005schedule_go
+.L004schedule_am_decrypting:
+ movdqa 256(%ebp,%ecx,1),%xmm1
+.byte 102,15,56,0,217
+ movdqu %xmm3,(%edx)
+ xorl $48,%ecx
+.L005schedule_go:
+ cmpl $192,%eax
+ ja .L006schedule_256
+ je .L007schedule_192
+.L008schedule_128:
+ movl $10,%eax
+.L009loop_schedule_128:
+ call _vpaes_schedule_round
+ decl %eax
+ jz .L010schedule_mangle_last
+ call _vpaes_schedule_mangle
+ jmp .L009loop_schedule_128
+.align 16
+.L007schedule_192:
+ movdqu 8(%esi),%xmm0
+ call _vpaes_schedule_transform
+ movdqa %xmm0,%xmm6
+ pxor %xmm4,%xmm4
+ movhlps %xmm4,%xmm6
+ movl $4,%eax
+.L011loop_schedule_192:
+ call _vpaes_schedule_round
+.byte 102,15,58,15,198,8
+ call _vpaes_schedule_mangle
+ call _vpaes_schedule_192_smear
+ call _vpaes_schedule_mangle
+ call _vpaes_schedule_round
+ decl %eax
+ jz .L010schedule_mangle_last
+ call _vpaes_schedule_mangle
+ call _vpaes_schedule_192_smear
+ jmp .L011loop_schedule_192
+.align 16
+.L006schedule_256:
+ movdqu 16(%esi),%xmm0
+ call _vpaes_schedule_transform
+ movl $7,%eax
+.L012loop_schedule_256:
+ call _vpaes_schedule_mangle
+ movdqa %xmm0,%xmm6
+ call _vpaes_schedule_round
+ decl %eax
+ jz .L010schedule_mangle_last
+ call _vpaes_schedule_mangle
+ pshufd $255,%xmm0,%xmm0
+ movdqa %xmm7,20(%esp)
+ movdqa %xmm6,%xmm7
+ call .L_vpaes_schedule_low_round
+ movdqa 20(%esp),%xmm7
+ jmp .L012loop_schedule_256
+.align 16
+.L010schedule_mangle_last:
+ leal 384(%ebp),%ebx
+ testl %edi,%edi
+ jnz .L013schedule_mangle_last_dec
+ movdqa 256(%ebp,%ecx,1),%xmm1
+.byte 102,15,56,0,193
+ leal 352(%ebp),%ebx
+ addl $32,%edx
+.L013schedule_mangle_last_dec:
+ addl $-16,%edx
+ pxor 336(%ebp),%xmm0
+ call _vpaes_schedule_transform
+ movdqu %xmm0,(%edx)
+ pxor %xmm0,%xmm0
+ pxor %xmm1,%xmm1
+ pxor %xmm2,%xmm2
+ pxor %xmm3,%xmm3
+ pxor %xmm4,%xmm4
+ pxor %xmm5,%xmm5
+ pxor %xmm6,%xmm6
+ pxor %xmm7,%xmm7
+ ret
+.size _vpaes_schedule_core,.-_vpaes_schedule_core
+.type _vpaes_schedule_192_smear,@function
+.align 16
+_vpaes_schedule_192_smear:
+ pshufd $128,%xmm6,%xmm1
+ pshufd $254,%xmm7,%xmm0
+ pxor %xmm1,%xmm6
+ pxor %xmm1,%xmm1
+ pxor %xmm0,%xmm6
+ movdqa %xmm6,%xmm0
+ movhlps %xmm1,%xmm6
+ ret
+.size _vpaes_schedule_192_smear,.-_vpaes_schedule_192_smear
+.type _vpaes_schedule_round,@function
+.align 16
+_vpaes_schedule_round:
+ movdqa 8(%esp),%xmm2
+ pxor %xmm1,%xmm1
+.byte 102,15,58,15,202,15
+.byte 102,15,58,15,210,15
+ pxor %xmm1,%xmm7
+ pshufd $255,%xmm0,%xmm0
+.byte 102,15,58,15,192,1
+ movdqa %xmm2,8(%esp)
+.L_vpaes_schedule_low_round:
+ movdqa %xmm7,%xmm1
+ pslldq $4,%xmm7
+ pxor %xmm1,%xmm7
+ movdqa %xmm7,%xmm1
+ pslldq $8,%xmm7
+ pxor %xmm1,%xmm7
+ pxor 336(%ebp),%xmm7
+ movdqa -16(%ebp),%xmm4
+ movdqa -48(%ebp),%xmm5
+ movdqa %xmm4,%xmm1
+ pandn %xmm0,%xmm1
+ psrld $4,%xmm1
+ pand %xmm4,%xmm0
+ movdqa -32(%ebp),%xmm2
+.byte 102,15,56,0,208
+ pxor %xmm1,%xmm0
+ movdqa %xmm5,%xmm3
+.byte 102,15,56,0,217
+ pxor %xmm2,%xmm3
+ movdqa %xmm5,%xmm4
+.byte 102,15,56,0,224
+ pxor %xmm2,%xmm4
+ movdqa %xmm5,%xmm2
+.byte 102,15,56,0,211
+ pxor %xmm0,%xmm2
+ movdqa %xmm5,%xmm3
+.byte 102,15,56,0,220
+ pxor %xmm1,%xmm3
+ movdqa 32(%ebp),%xmm4
+.byte 102,15,56,0,226
+ movdqa 48(%ebp),%xmm0
+.byte 102,15,56,0,195
+ pxor %xmm4,%xmm0
+ pxor %xmm7,%xmm0
+ movdqa %xmm0,%xmm7
+ ret
+.size _vpaes_schedule_round,.-_vpaes_schedule_round
+.type _vpaes_schedule_transform,@function
+.align 16
+_vpaes_schedule_transform:
+ movdqa -16(%ebp),%xmm2
+ movdqa %xmm2,%xmm1
+ pandn %xmm0,%xmm1
+ psrld $4,%xmm1
+ pand %xmm2,%xmm0
+ movdqa (%ebx),%xmm2
+.byte 102,15,56,0,208
+ movdqa 16(%ebx),%xmm0
+.byte 102,15,56,0,193
+ pxor %xmm2,%xmm0
+ ret
+.size _vpaes_schedule_transform,.-_vpaes_schedule_transform
+.type _vpaes_schedule_mangle,@function
+.align 16
+_vpaes_schedule_mangle:
+ movdqa %xmm0,%xmm4
+ movdqa 128(%ebp),%xmm5
+ testl %edi,%edi
+ jnz .L014schedule_mangle_dec
+ addl $16,%edx
+ pxor 336(%ebp),%xmm4
+.byte 102,15,56,0,229
+ movdqa %xmm4,%xmm3
+.byte 102,15,56,0,229
+ pxor %xmm4,%xmm3
+.byte 102,15,56,0,229
+ pxor %xmm4,%xmm3
+ jmp .L015schedule_mangle_both
+.align 16
+.L014schedule_mangle_dec:
+ movdqa -16(%ebp),%xmm2
+ leal 416(%ebp),%esi
+ movdqa %xmm2,%xmm1
+ pandn %xmm4,%xmm1
+ psrld $4,%xmm1
+ pand %xmm2,%xmm4
+ movdqa (%esi),%xmm2
+.byte 102,15,56,0,212
+ movdqa 16(%esi),%xmm3
+.byte 102,15,56,0,217
+ pxor %xmm2,%xmm3
+.byte 102,15,56,0,221
+ movdqa 32(%esi),%xmm2
+.byte 102,15,56,0,212
+ pxor %xmm3,%xmm2
+ movdqa 48(%esi),%xmm3
+.byte 102,15,56,0,217
+ pxor %xmm2,%xmm3
+.byte 102,15,56,0,221
+ movdqa 64(%esi),%xmm2
+.byte 102,15,56,0,212
+ pxor %xmm3,%xmm2
+ movdqa 80(%esi),%xmm3
+.byte 102,15,56,0,217
+ pxor %xmm2,%xmm3
+.byte 102,15,56,0,221
+ movdqa 96(%esi),%xmm2
+.byte 102,15,56,0,212
+ pxor %xmm3,%xmm2
+ movdqa 112(%esi),%xmm3
+.byte 102,15,56,0,217
+ pxor %xmm2,%xmm3
+ addl $-16,%edx
+.L015schedule_mangle_both:
+ movdqa 256(%ebp,%ecx,1),%xmm1
+.byte 102,15,56,0,217
+ addl $-16,%ecx
+ andl $48,%ecx
+ movdqu %xmm3,(%edx)
+ ret
+.size _vpaes_schedule_mangle,.-_vpaes_schedule_mangle
+.globl vpaes_set_encrypt_key
+.type vpaes_set_encrypt_key,@function
+.align 16
+vpaes_set_encrypt_key:
+.L_vpaes_set_encrypt_key_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ leal -56(%esp),%ebx
+ movl 24(%esp),%eax
+ andl $-16,%ebx
+ movl 28(%esp),%edx
+ xchgl %esp,%ebx
+ movl %ebx,48(%esp)
+ movl %eax,%ebx
+ shrl $5,%ebx
+ addl $5,%ebx
+ movl %ebx,240(%edx)
+ movl $48,%ecx
+ movl $0,%edi
+ leal .L_vpaes_consts+0x30-.L016pic_point,%ebp
+ call _vpaes_schedule_core
+.L016pic_point:
+ movl 48(%esp),%esp
+ xorl %eax,%eax
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size vpaes_set_encrypt_key,.-.L_vpaes_set_encrypt_key_begin
+.globl vpaes_set_decrypt_key
+.type vpaes_set_decrypt_key,@function
+.align 16
+vpaes_set_decrypt_key:
+.L_vpaes_set_decrypt_key_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ leal -56(%esp),%ebx
+ movl 24(%esp),%eax
+ andl $-16,%ebx
+ movl 28(%esp),%edx
+ xchgl %esp,%ebx
+ movl %ebx,48(%esp)
+ movl %eax,%ebx
+ shrl $5,%ebx
+ addl $5,%ebx
+ movl %ebx,240(%edx)
+ shll $4,%ebx
+ leal 16(%edx,%ebx,1),%edx
+ movl $1,%edi
+ movl %eax,%ecx
+ shrl $1,%ecx
+ andl $32,%ecx
+ xorl $32,%ecx
+ leal .L_vpaes_consts+0x30-.L017pic_point,%ebp
+ call _vpaes_schedule_core
+.L017pic_point:
+ movl 48(%esp),%esp
+ xorl %eax,%eax
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size vpaes_set_decrypt_key,.-.L_vpaes_set_decrypt_key_begin
+.globl vpaes_encrypt
+.type vpaes_encrypt,@function
+.align 16
+vpaes_encrypt:
+.L_vpaes_encrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ leal .L_vpaes_consts+0x30-.L018pic_point,%ebp
+ call _vpaes_preheat
+.L018pic_point:
+ movl 20(%esp),%esi
+ leal -56(%esp),%ebx
+ movl 24(%esp),%edi
+ andl $-16,%ebx
+ movl 28(%esp),%edx
+ xchgl %esp,%ebx
+ movl %ebx,48(%esp)
+ movdqu (%esi),%xmm0
+ call _vpaes_encrypt_core
+ movdqu %xmm0,(%edi)
+ movl 48(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size vpaes_encrypt,.-.L_vpaes_encrypt_begin
+.globl vpaes_decrypt
+.type vpaes_decrypt,@function
+.align 16
+vpaes_decrypt:
+.L_vpaes_decrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ leal .L_vpaes_consts+0x30-.L019pic_point,%ebp
+ call _vpaes_preheat
+.L019pic_point:
+ movl 20(%esp),%esi
+ leal -56(%esp),%ebx
+ movl 24(%esp),%edi
+ andl $-16,%ebx
+ movl 28(%esp),%edx
+ xchgl %esp,%ebx
+ movl %ebx,48(%esp)
+ movdqu (%esi),%xmm0
+ call _vpaes_decrypt_core
+ movdqu %xmm0,(%edi)
+ movl 48(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size vpaes_decrypt,.-.L_vpaes_decrypt_begin
+.globl vpaes_cbc_encrypt
+.type vpaes_cbc_encrypt,@function
+.align 16
+vpaes_cbc_encrypt:
+.L_vpaes_cbc_encrypt_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl 32(%esp),%edx
+ subl $16,%eax
+ jc .L020cbc_abort
+ leal -56(%esp),%ebx
+ movl 36(%esp),%ebp
+ andl $-16,%ebx
+ movl 40(%esp),%ecx
+ xchgl %esp,%ebx
+ movdqu (%ebp),%xmm1
+ subl %esi,%edi
+ movl %ebx,48(%esp)
+ movl %edi,(%esp)
+ movl %edx,4(%esp)
+ movl %ebp,8(%esp)
+ movl %eax,%edi
+ leal .L_vpaes_consts+0x30-.L021pic_point,%ebp
+ call _vpaes_preheat
+.L021pic_point:
+ cmpl $0,%ecx
+ je .L022cbc_dec_loop
+ jmp .L023cbc_enc_loop
+.align 16
+.L023cbc_enc_loop:
+ movdqu (%esi),%xmm0
+ pxor %xmm1,%xmm0
+ call _vpaes_encrypt_core
+ movl (%esp),%ebx
+ movl 4(%esp),%edx
+ movdqa %xmm0,%xmm1
+ movdqu %xmm0,(%ebx,%esi,1)
+ leal 16(%esi),%esi
+ subl $16,%edi
+ jnc .L023cbc_enc_loop
+ jmp .L024cbc_done
+.align 16
+.L022cbc_dec_loop:
+ movdqu (%esi),%xmm0
+ movdqa %xmm1,16(%esp)
+ movdqa %xmm0,32(%esp)
+ call _vpaes_decrypt_core
+ movl (%esp),%ebx
+ movl 4(%esp),%edx
+ pxor 16(%esp),%xmm0
+ movdqa 32(%esp),%xmm1
+ movdqu %xmm0,(%ebx,%esi,1)
+ leal 16(%esi),%esi
+ subl $16,%edi
+ jnc .L022cbc_dec_loop
+.L024cbc_done:
+ movl 8(%esp),%ebx
+ movl 48(%esp),%esp
+ movdqu %xmm1,(%ebx)
+.L020cbc_abort:
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size vpaes_cbc_encrypt,.-.L_vpaes_cbc_encrypt_begin
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/modes/ghash-x86.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/modes/ghash-x86.S
new file mode 100644
index 0000000000..f52b445df2
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/modes/ghash-x86.S
@@ -0,0 +1,703 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/modes/asm/ghash-x86.pl
+#
+# Copyright 2010-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.globl gcm_gmult_4bit_x86
+.type gcm_gmult_4bit_x86,@function
+.align 16
+gcm_gmult_4bit_x86:
+.L_gcm_gmult_4bit_x86_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ subl $84,%esp
+ movl 104(%esp),%edi
+ movl 108(%esp),%esi
+ movl (%edi),%ebp
+ movl 4(%edi),%edx
+ movl 8(%edi),%ecx
+ movl 12(%edi),%ebx
+ movl $0,16(%esp)
+ movl $471859200,20(%esp)
+ movl $943718400,24(%esp)
+ movl $610271232,28(%esp)
+ movl $1887436800,32(%esp)
+ movl $1822425088,36(%esp)
+ movl $1220542464,40(%esp)
+ movl $1423966208,44(%esp)
+ movl $3774873600,48(%esp)
+ movl $4246732800,52(%esp)
+ movl $3644850176,56(%esp)
+ movl $3311403008,60(%esp)
+ movl $2441084928,64(%esp)
+ movl $2376073216,68(%esp)
+ movl $2847932416,72(%esp)
+ movl $3051356160,76(%esp)
+ movl %ebp,(%esp)
+ movl %edx,4(%esp)
+ movl %ecx,8(%esp)
+ movl %ebx,12(%esp)
+ shrl $20,%ebx
+ andl $240,%ebx
+ movl 4(%esi,%ebx,1),%ebp
+ movl (%esi,%ebx,1),%edx
+ movl 12(%esi,%ebx,1),%ecx
+ movl 8(%esi,%ebx,1),%ebx
+ xorl %eax,%eax
+ movl $15,%edi
+ jmp .L000x86_loop
+.align 16
+.L000x86_loop:
+ movb %bl,%al
+ shrdl $4,%ecx,%ebx
+ andb $15,%al
+ shrdl $4,%edx,%ecx
+ shrdl $4,%ebp,%edx
+ shrl $4,%ebp
+ xorl 16(%esp,%eax,4),%ebp
+ movb (%esp,%edi,1),%al
+ andb $240,%al
+ xorl 8(%esi,%eax,1),%ebx
+ xorl 12(%esi,%eax,1),%ecx
+ xorl (%esi,%eax,1),%edx
+ xorl 4(%esi,%eax,1),%ebp
+ decl %edi
+ js .L001x86_break
+ movb %bl,%al
+ shrdl $4,%ecx,%ebx
+ andb $15,%al
+ shrdl $4,%edx,%ecx
+ shrdl $4,%ebp,%edx
+ shrl $4,%ebp
+ xorl 16(%esp,%eax,4),%ebp
+ movb (%esp,%edi,1),%al
+ shlb $4,%al
+ xorl 8(%esi,%eax,1),%ebx
+ xorl 12(%esi,%eax,1),%ecx
+ xorl (%esi,%eax,1),%edx
+ xorl 4(%esi,%eax,1),%ebp
+ jmp .L000x86_loop
+.align 16
+.L001x86_break:
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ bswap %ebp
+ movl 104(%esp),%edi
+ movl %ebx,12(%edi)
+ movl %ecx,8(%edi)
+ movl %edx,4(%edi)
+ movl %ebp,(%edi)
+ addl $84,%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size gcm_gmult_4bit_x86,.-.L_gcm_gmult_4bit_x86_begin
+.globl gcm_ghash_4bit_x86
+.type gcm_ghash_4bit_x86,@function
+.align 16
+gcm_ghash_4bit_x86:
+.L_gcm_ghash_4bit_x86_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ subl $84,%esp
+ movl 104(%esp),%ebx
+ movl 108(%esp),%esi
+ movl 112(%esp),%edi
+ movl 116(%esp),%ecx
+ addl %edi,%ecx
+ movl %ecx,116(%esp)
+ movl (%ebx),%ebp
+ movl 4(%ebx),%edx
+ movl 8(%ebx),%ecx
+ movl 12(%ebx),%ebx
+ movl $0,16(%esp)
+ movl $471859200,20(%esp)
+ movl $943718400,24(%esp)
+ movl $610271232,28(%esp)
+ movl $1887436800,32(%esp)
+ movl $1822425088,36(%esp)
+ movl $1220542464,40(%esp)
+ movl $1423966208,44(%esp)
+ movl $3774873600,48(%esp)
+ movl $4246732800,52(%esp)
+ movl $3644850176,56(%esp)
+ movl $3311403008,60(%esp)
+ movl $2441084928,64(%esp)
+ movl $2376073216,68(%esp)
+ movl $2847932416,72(%esp)
+ movl $3051356160,76(%esp)
+.align 16
+.L002x86_outer_loop:
+ xorl 12(%edi),%ebx
+ xorl 8(%edi),%ecx
+ xorl 4(%edi),%edx
+ xorl (%edi),%ebp
+ movl %ebx,12(%esp)
+ movl %ecx,8(%esp)
+ movl %edx,4(%esp)
+ movl %ebp,(%esp)
+ shrl $20,%ebx
+ andl $240,%ebx
+ movl 4(%esi,%ebx,1),%ebp
+ movl (%esi,%ebx,1),%edx
+ movl 12(%esi,%ebx,1),%ecx
+ movl 8(%esi,%ebx,1),%ebx
+ xorl %eax,%eax
+ movl $15,%edi
+ jmp .L003x86_loop
+.align 16
+.L003x86_loop:
+ movb %bl,%al
+ shrdl $4,%ecx,%ebx
+ andb $15,%al
+ shrdl $4,%edx,%ecx
+ shrdl $4,%ebp,%edx
+ shrl $4,%ebp
+ xorl 16(%esp,%eax,4),%ebp
+ movb (%esp,%edi,1),%al
+ andb $240,%al
+ xorl 8(%esi,%eax,1),%ebx
+ xorl 12(%esi,%eax,1),%ecx
+ xorl (%esi,%eax,1),%edx
+ xorl 4(%esi,%eax,1),%ebp
+ decl %edi
+ js .L004x86_break
+ movb %bl,%al
+ shrdl $4,%ecx,%ebx
+ andb $15,%al
+ shrdl $4,%edx,%ecx
+ shrdl $4,%ebp,%edx
+ shrl $4,%ebp
+ xorl 16(%esp,%eax,4),%ebp
+ movb (%esp,%edi,1),%al
+ shlb $4,%al
+ xorl 8(%esi,%eax,1),%ebx
+ xorl 12(%esi,%eax,1),%ecx
+ xorl (%esi,%eax,1),%edx
+ xorl 4(%esi,%eax,1),%ebp
+ jmp .L003x86_loop
+.align 16
+.L004x86_break:
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ bswap %ebp
+ movl 112(%esp),%edi
+ leal 16(%edi),%edi
+ cmpl 116(%esp),%edi
+ movl %edi,112(%esp)
+ jb .L002x86_outer_loop
+ movl 104(%esp),%edi
+ movl %ebx,12(%edi)
+ movl %ecx,8(%edi)
+ movl %edx,4(%edi)
+ movl %ebp,(%edi)
+ addl $84,%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size gcm_ghash_4bit_x86,.-.L_gcm_ghash_4bit_x86_begin
+.type _mmx_gmult_4bit_inner,@function
+.align 16
+_mmx_gmult_4bit_inner:
+ xorl %ecx,%ecx
+ movl %ebx,%edx
+ movb %dl,%cl
+ shlb $4,%cl
+ andl $240,%edx
+ movq 8(%esi,%ecx,1),%mm0
+ movq (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 14(%edi),%cl
+ psllq $60,%mm2
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 13(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 12(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 11(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 10(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 9(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 8(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 7(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 6(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 5(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 4(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 3(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 2(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb 1(%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ movb (%edi),%cl
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movl %ecx,%edx
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ shlb $4,%cl
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%ecx,1),%mm0
+ psllq $60,%mm2
+ andl $240,%edx
+ pxor (%eax,%ebp,8),%mm1
+ andl $15,%ebx
+ pxor (%esi,%ecx,1),%mm1
+ movd %mm0,%ebp
+ pxor %mm2,%mm0
+ psrlq $4,%mm0
+ movq %mm1,%mm2
+ psrlq $4,%mm1
+ pxor 8(%esi,%edx,1),%mm0
+ psllq $60,%mm2
+ pxor (%eax,%ebx,8),%mm1
+ andl $15,%ebp
+ pxor (%esi,%edx,1),%mm1
+ movd %mm0,%ebx
+ pxor %mm2,%mm0
+ movl 4(%eax,%ebp,8),%edi
+ psrlq $32,%mm0
+ movd %mm1,%edx
+ psrlq $32,%mm1
+ movd %mm0,%ecx
+ movd %mm1,%ebp
+ shll $4,%edi
+ bswap %ebx
+ bswap %edx
+ bswap %ecx
+ xorl %edi,%ebp
+ bswap %ebp
+ ret
+.size _mmx_gmult_4bit_inner,.-_mmx_gmult_4bit_inner
+.globl gcm_gmult_4bit_mmx
+.type gcm_gmult_4bit_mmx,@function
+.align 16
+gcm_gmult_4bit_mmx:
+.L_gcm_gmult_4bit_mmx_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%edi
+ movl 24(%esp),%esi
+ call .L005pic_point
+.L005pic_point:
+ popl %eax
+ leal .Lrem_4bit-.L005pic_point(%eax),%eax
+ movzbl 15(%edi),%ebx
+ call _mmx_gmult_4bit_inner
+ movl 20(%esp),%edi
+ emms
+ movl %ebx,12(%edi)
+ movl %edx,4(%edi)
+ movl %ecx,8(%edi)
+ movl %ebp,(%edi)
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size gcm_gmult_4bit_mmx,.-.L_gcm_gmult_4bit_mmx_begin
+.globl gcm_ghash_4bit_mmx
+.type gcm_ghash_4bit_mmx,@function
+.align 16
+gcm_ghash_4bit_mmx:
+.L_gcm_ghash_4bit_mmx_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%ebp
+ movl 24(%esp),%esi
+ movl 28(%esp),%edi
+ movl 32(%esp),%ecx
+ call .L006pic_point
+.L006pic_point:
+ popl %eax
+ leal .Lrem_4bit-.L006pic_point(%eax),%eax
+ addl %edi,%ecx
+ movl %ecx,32(%esp)
+ subl $20,%esp
+ movl 12(%ebp),%ebx
+ movl 4(%ebp),%edx
+ movl 8(%ebp),%ecx
+ movl (%ebp),%ebp
+ jmp .L007mmx_outer_loop
+.align 16
+.L007mmx_outer_loop:
+ xorl 12(%edi),%ebx
+ xorl 4(%edi),%edx
+ xorl 8(%edi),%ecx
+ xorl (%edi),%ebp
+ movl %edi,48(%esp)
+ movl %ebx,12(%esp)
+ movl %edx,4(%esp)
+ movl %ecx,8(%esp)
+ movl %ebp,(%esp)
+ movl %esp,%edi
+ shrl $24,%ebx
+ call _mmx_gmult_4bit_inner
+ movl 48(%esp),%edi
+ leal 16(%edi),%edi
+ cmpl 52(%esp),%edi
+ jb .L007mmx_outer_loop
+ movl 40(%esp),%edi
+ emms
+ movl %ebx,12(%edi)
+ movl %edx,4(%edi)
+ movl %ecx,8(%edi)
+ movl %ebp,(%edi)
+ addl $20,%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size gcm_ghash_4bit_mmx,.-.L_gcm_ghash_4bit_mmx_begin
+.align 64
+.Lrem_4bit:
+.long 0,0,0,29491200,0,58982400,0,38141952
+.long 0,117964800,0,113901568,0,76283904,0,88997888
+.long 0,235929600,0,265420800,0,227803136,0,206962688
+.long 0,152567808,0,148504576,0,177995776,0,190709760
+.byte 71,72,65,83,72,32,102,111,114,32,120,56,54,44,32,67
+.byte 82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112
+.byte 112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62
+.byte 0
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha1-586.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha1-586.S
new file mode 100644
index 0000000000..0b22a3e553
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha1-586.S
@@ -0,0 +1,1389 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/sha/asm/sha1-586.pl
+#
+# Copyright 1998-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.globl sha1_block_data_order
+.type sha1_block_data_order,@function
+.align 16
+sha1_block_data_order:
+.L_sha1_block_data_order_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%ebp
+ movl 24(%esp),%esi
+ movl 28(%esp),%eax
+ subl $76,%esp
+ shll $6,%eax
+ addl %esi,%eax
+ movl %eax,104(%esp)
+ movl 16(%ebp),%edi
+ jmp .L000loop
+.align 16
+.L000loop:
+ movl (%esi),%eax
+ movl 4(%esi),%ebx
+ movl 8(%esi),%ecx
+ movl 12(%esi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ movl %eax,(%esp)
+ movl %ebx,4(%esp)
+ movl %ecx,8(%esp)
+ movl %edx,12(%esp)
+ movl 16(%esi),%eax
+ movl 20(%esi),%ebx
+ movl 24(%esi),%ecx
+ movl 28(%esi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ movl %eax,16(%esp)
+ movl %ebx,20(%esp)
+ movl %ecx,24(%esp)
+ movl %edx,28(%esp)
+ movl 32(%esi),%eax
+ movl 36(%esi),%ebx
+ movl 40(%esi),%ecx
+ movl 44(%esi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ movl %eax,32(%esp)
+ movl %ebx,36(%esp)
+ movl %ecx,40(%esp)
+ movl %edx,44(%esp)
+ movl 48(%esi),%eax
+ movl 52(%esi),%ebx
+ movl 56(%esi),%ecx
+ movl 60(%esi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ movl %eax,48(%esp)
+ movl %ebx,52(%esp)
+ movl %ecx,56(%esp)
+ movl %edx,60(%esp)
+ movl %esi,100(%esp)
+ movl (%ebp),%eax
+ movl 4(%ebp),%ebx
+ movl 8(%ebp),%ecx
+ movl 12(%ebp),%edx
+
+ movl %ecx,%esi
+ movl %eax,%ebp
+ roll $5,%ebp
+ xorl %edx,%esi
+ addl %edi,%ebp
+ movl (%esp),%edi
+ andl %ebx,%esi
+ rorl $2,%ebx
+ xorl %edx,%esi
+ leal 1518500249(%ebp,%edi,1),%ebp
+ addl %esi,%ebp
+
+ movl %ebx,%edi
+ movl %ebp,%esi
+ roll $5,%ebp
+ xorl %ecx,%edi
+ addl %edx,%ebp
+ movl 4(%esp),%edx
+ andl %eax,%edi
+ rorl $2,%eax
+ xorl %ecx,%edi
+ leal 1518500249(%ebp,%edx,1),%ebp
+ addl %edi,%ebp
+
+ movl %eax,%edx
+ movl %ebp,%edi
+ roll $5,%ebp
+ xorl %ebx,%edx
+ addl %ecx,%ebp
+ movl 8(%esp),%ecx
+ andl %esi,%edx
+ rorl $2,%esi
+ xorl %ebx,%edx
+ leal 1518500249(%ebp,%ecx,1),%ebp
+ addl %edx,%ebp
+
+ movl %esi,%ecx
+ movl %ebp,%edx
+ roll $5,%ebp
+ xorl %eax,%ecx
+ addl %ebx,%ebp
+ movl 12(%esp),%ebx
+ andl %edi,%ecx
+ rorl $2,%edi
+ xorl %eax,%ecx
+ leal 1518500249(%ebp,%ebx,1),%ebp
+ addl %ecx,%ebp
+
+ movl %edi,%ebx
+ movl %ebp,%ecx
+ roll $5,%ebp
+ xorl %esi,%ebx
+ addl %eax,%ebp
+ movl 16(%esp),%eax
+ andl %edx,%ebx
+ rorl $2,%edx
+ xorl %esi,%ebx
+ leal 1518500249(%ebp,%eax,1),%ebp
+ addl %ebx,%ebp
+
+ movl %edx,%eax
+ movl %ebp,%ebx
+ roll $5,%ebp
+ xorl %edi,%eax
+ addl %esi,%ebp
+ movl 20(%esp),%esi
+ andl %ecx,%eax
+ rorl $2,%ecx
+ xorl %edi,%eax
+ leal 1518500249(%ebp,%esi,1),%ebp
+ addl %eax,%ebp
+
+ movl %ecx,%esi
+ movl %ebp,%eax
+ roll $5,%ebp
+ xorl %edx,%esi
+ addl %edi,%ebp
+ movl 24(%esp),%edi
+ andl %ebx,%esi
+ rorl $2,%ebx
+ xorl %edx,%esi
+ leal 1518500249(%ebp,%edi,1),%ebp
+ addl %esi,%ebp
+
+ movl %ebx,%edi
+ movl %ebp,%esi
+ roll $5,%ebp
+ xorl %ecx,%edi
+ addl %edx,%ebp
+ movl 28(%esp),%edx
+ andl %eax,%edi
+ rorl $2,%eax
+ xorl %ecx,%edi
+ leal 1518500249(%ebp,%edx,1),%ebp
+ addl %edi,%ebp
+
+ movl %eax,%edx
+ movl %ebp,%edi
+ roll $5,%ebp
+ xorl %ebx,%edx
+ addl %ecx,%ebp
+ movl 32(%esp),%ecx
+ andl %esi,%edx
+ rorl $2,%esi
+ xorl %ebx,%edx
+ leal 1518500249(%ebp,%ecx,1),%ebp
+ addl %edx,%ebp
+
+ movl %esi,%ecx
+ movl %ebp,%edx
+ roll $5,%ebp
+ xorl %eax,%ecx
+ addl %ebx,%ebp
+ movl 36(%esp),%ebx
+ andl %edi,%ecx
+ rorl $2,%edi
+ xorl %eax,%ecx
+ leal 1518500249(%ebp,%ebx,1),%ebp
+ addl %ecx,%ebp
+
+ movl %edi,%ebx
+ movl %ebp,%ecx
+ roll $5,%ebp
+ xorl %esi,%ebx
+ addl %eax,%ebp
+ movl 40(%esp),%eax
+ andl %edx,%ebx
+ rorl $2,%edx
+ xorl %esi,%ebx
+ leal 1518500249(%ebp,%eax,1),%ebp
+ addl %ebx,%ebp
+
+ movl %edx,%eax
+ movl %ebp,%ebx
+ roll $5,%ebp
+ xorl %edi,%eax
+ addl %esi,%ebp
+ movl 44(%esp),%esi
+ andl %ecx,%eax
+ rorl $2,%ecx
+ xorl %edi,%eax
+ leal 1518500249(%ebp,%esi,1),%ebp
+ addl %eax,%ebp
+
+ movl %ecx,%esi
+ movl %ebp,%eax
+ roll $5,%ebp
+ xorl %edx,%esi
+ addl %edi,%ebp
+ movl 48(%esp),%edi
+ andl %ebx,%esi
+ rorl $2,%ebx
+ xorl %edx,%esi
+ leal 1518500249(%ebp,%edi,1),%ebp
+ addl %esi,%ebp
+
+ movl %ebx,%edi
+ movl %ebp,%esi
+ roll $5,%ebp
+ xorl %ecx,%edi
+ addl %edx,%ebp
+ movl 52(%esp),%edx
+ andl %eax,%edi
+ rorl $2,%eax
+ xorl %ecx,%edi
+ leal 1518500249(%ebp,%edx,1),%ebp
+ addl %edi,%ebp
+
+ movl %eax,%edx
+ movl %ebp,%edi
+ roll $5,%ebp
+ xorl %ebx,%edx
+ addl %ecx,%ebp
+ movl 56(%esp),%ecx
+ andl %esi,%edx
+ rorl $2,%esi
+ xorl %ebx,%edx
+ leal 1518500249(%ebp,%ecx,1),%ebp
+ addl %edx,%ebp
+
+ movl %esi,%ecx
+ movl %ebp,%edx
+ roll $5,%ebp
+ xorl %eax,%ecx
+ addl %ebx,%ebp
+ movl 60(%esp),%ebx
+ andl %edi,%ecx
+ rorl $2,%edi
+ xorl %eax,%ecx
+ leal 1518500249(%ebp,%ebx,1),%ebp
+ movl (%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edi,%ebp
+ xorl 8(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 32(%esp),%ebx
+ andl %edx,%ebp
+ xorl 52(%esp),%ebx
+ roll $1,%ebx
+ xorl %esi,%ebp
+ addl %ebp,%eax
+ movl %ecx,%ebp
+ rorl $2,%edx
+ movl %ebx,(%esp)
+ roll $5,%ebp
+ leal 1518500249(%ebx,%eax,1),%ebx
+ movl 4(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %edx,%ebp
+ xorl 12(%esp),%eax
+ xorl %edi,%ebp
+ xorl 36(%esp),%eax
+ andl %ecx,%ebp
+ xorl 56(%esp),%eax
+ roll $1,%eax
+ xorl %edi,%ebp
+ addl %ebp,%esi
+ movl %ebx,%ebp
+ rorl $2,%ecx
+ movl %eax,4(%esp)
+ roll $5,%ebp
+ leal 1518500249(%eax,%esi,1),%eax
+ movl 8(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ecx,%ebp
+ xorl 16(%esp),%esi
+ xorl %edx,%ebp
+ xorl 40(%esp),%esi
+ andl %ebx,%ebp
+ xorl 60(%esp),%esi
+ roll $1,%esi
+ xorl %edx,%ebp
+ addl %ebp,%edi
+ movl %eax,%ebp
+ rorl $2,%ebx
+ movl %esi,8(%esp)
+ roll $5,%ebp
+ leal 1518500249(%esi,%edi,1),%esi
+ movl 12(%esp),%edi
+ addl %ebp,%esi
+
+ movl %ebx,%ebp
+ xorl 20(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 44(%esp),%edi
+ andl %eax,%ebp
+ xorl (%esp),%edi
+ roll $1,%edi
+ xorl %ecx,%ebp
+ addl %ebp,%edx
+ movl %esi,%ebp
+ rorl $2,%eax
+ movl %edi,12(%esp)
+ roll $5,%ebp
+ leal 1518500249(%edi,%edx,1),%edi
+ movl 16(%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl 24(%esp),%edx
+ xorl %eax,%ebp
+ xorl 48(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 4(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,16(%esp)
+ leal 1859775393(%edx,%ecx,1),%edx
+ movl 20(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 28(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 52(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 8(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,20(%esp)
+ leal 1859775393(%ecx,%ebx,1),%ecx
+ movl 24(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edx,%ebp
+ xorl 32(%esp),%ebx
+ xorl %edi,%ebp
+ xorl 56(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 12(%esp),%ebx
+ roll $1,%ebx
+ addl %ebp,%eax
+ rorl $2,%edx
+ movl %ecx,%ebp
+ roll $5,%ebp
+ movl %ebx,24(%esp)
+ leal 1859775393(%ebx,%eax,1),%ebx
+ movl 28(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %ecx,%ebp
+ xorl 36(%esp),%eax
+ xorl %edx,%ebp
+ xorl 60(%esp),%eax
+ xorl %edi,%ebp
+ xorl 16(%esp),%eax
+ roll $1,%eax
+ addl %ebp,%esi
+ rorl $2,%ecx
+ movl %ebx,%ebp
+ roll $5,%ebp
+ movl %eax,28(%esp)
+ leal 1859775393(%eax,%esi,1),%eax
+ movl 32(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl 40(%esp),%esi
+ xorl %ecx,%ebp
+ xorl (%esp),%esi
+ xorl %edx,%ebp
+ xorl 20(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ movl %esi,32(%esp)
+ leal 1859775393(%esi,%edi,1),%esi
+ movl 36(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 44(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 4(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 24(%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ movl %edi,36(%esp)
+ leal 1859775393(%edi,%edx,1),%edi
+ movl 40(%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl 48(%esp),%edx
+ xorl %eax,%ebp
+ xorl 8(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 28(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,40(%esp)
+ leal 1859775393(%edx,%ecx,1),%edx
+ movl 44(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 52(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 12(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 32(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,44(%esp)
+ leal 1859775393(%ecx,%ebx,1),%ecx
+ movl 48(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edx,%ebp
+ xorl 56(%esp),%ebx
+ xorl %edi,%ebp
+ xorl 16(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 36(%esp),%ebx
+ roll $1,%ebx
+ addl %ebp,%eax
+ rorl $2,%edx
+ movl %ecx,%ebp
+ roll $5,%ebp
+ movl %ebx,48(%esp)
+ leal 1859775393(%ebx,%eax,1),%ebx
+ movl 52(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %ecx,%ebp
+ xorl 60(%esp),%eax
+ xorl %edx,%ebp
+ xorl 20(%esp),%eax
+ xorl %edi,%ebp
+ xorl 40(%esp),%eax
+ roll $1,%eax
+ addl %ebp,%esi
+ rorl $2,%ecx
+ movl %ebx,%ebp
+ roll $5,%ebp
+ movl %eax,52(%esp)
+ leal 1859775393(%eax,%esi,1),%eax
+ movl 56(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl (%esp),%esi
+ xorl %ecx,%ebp
+ xorl 24(%esp),%esi
+ xorl %edx,%ebp
+ xorl 44(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ movl %esi,56(%esp)
+ leal 1859775393(%esi,%edi,1),%esi
+ movl 60(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 4(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 28(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 48(%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ movl %edi,60(%esp)
+ leal 1859775393(%edi,%edx,1),%edi
+ movl (%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl 8(%esp),%edx
+ xorl %eax,%ebp
+ xorl 32(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 52(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,(%esp)
+ leal 1859775393(%edx,%ecx,1),%edx
+ movl 4(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 12(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 36(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 56(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,4(%esp)
+ leal 1859775393(%ecx,%ebx,1),%ecx
+ movl 8(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edx,%ebp
+ xorl 16(%esp),%ebx
+ xorl %edi,%ebp
+ xorl 40(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 60(%esp),%ebx
+ roll $1,%ebx
+ addl %ebp,%eax
+ rorl $2,%edx
+ movl %ecx,%ebp
+ roll $5,%ebp
+ movl %ebx,8(%esp)
+ leal 1859775393(%ebx,%eax,1),%ebx
+ movl 12(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %ecx,%ebp
+ xorl 20(%esp),%eax
+ xorl %edx,%ebp
+ xorl 44(%esp),%eax
+ xorl %edi,%ebp
+ xorl (%esp),%eax
+ roll $1,%eax
+ addl %ebp,%esi
+ rorl $2,%ecx
+ movl %ebx,%ebp
+ roll $5,%ebp
+ movl %eax,12(%esp)
+ leal 1859775393(%eax,%esi,1),%eax
+ movl 16(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl 24(%esp),%esi
+ xorl %ecx,%ebp
+ xorl 48(%esp),%esi
+ xorl %edx,%ebp
+ xorl 4(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ movl %esi,16(%esp)
+ leal 1859775393(%esi,%edi,1),%esi
+ movl 20(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 28(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 52(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 8(%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ movl %edi,20(%esp)
+ leal 1859775393(%edi,%edx,1),%edi
+ movl 24(%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl 32(%esp),%edx
+ xorl %eax,%ebp
+ xorl 56(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 12(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,24(%esp)
+ leal 1859775393(%edx,%ecx,1),%edx
+ movl 28(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 36(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 60(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 16(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,28(%esp)
+ leal 1859775393(%ecx,%ebx,1),%ecx
+ movl 32(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edi,%ebp
+ xorl 40(%esp),%ebx
+ xorl %esi,%ebp
+ xorl (%esp),%ebx
+ andl %edx,%ebp
+ xorl 20(%esp),%ebx
+ roll $1,%ebx
+ addl %eax,%ebp
+ rorl $2,%edx
+ movl %ecx,%eax
+ roll $5,%eax
+ movl %ebx,32(%esp)
+ leal 2400959708(%ebx,%ebp,1),%ebx
+ movl %edi,%ebp
+ addl %eax,%ebx
+ andl %esi,%ebp
+ movl 36(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %edx,%ebp
+ xorl 44(%esp),%eax
+ xorl %edi,%ebp
+ xorl 4(%esp),%eax
+ andl %ecx,%ebp
+ xorl 24(%esp),%eax
+ roll $1,%eax
+ addl %esi,%ebp
+ rorl $2,%ecx
+ movl %ebx,%esi
+ roll $5,%esi
+ movl %eax,36(%esp)
+ leal 2400959708(%eax,%ebp,1),%eax
+ movl %edx,%ebp
+ addl %esi,%eax
+ andl %edi,%ebp
+ movl 40(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ecx,%ebp
+ xorl 48(%esp),%esi
+ xorl %edx,%ebp
+ xorl 8(%esp),%esi
+ andl %ebx,%ebp
+ xorl 28(%esp),%esi
+ roll $1,%esi
+ addl %edi,%ebp
+ rorl $2,%ebx
+ movl %eax,%edi
+ roll $5,%edi
+ movl %esi,40(%esp)
+ leal 2400959708(%esi,%ebp,1),%esi
+ movl %ecx,%ebp
+ addl %edi,%esi
+ andl %edx,%ebp
+ movl 44(%esp),%edi
+ addl %ebp,%esi
+
+ movl %ebx,%ebp
+ xorl 52(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 12(%esp),%edi
+ andl %eax,%ebp
+ xorl 32(%esp),%edi
+ roll $1,%edi
+ addl %edx,%ebp
+ rorl $2,%eax
+ movl %esi,%edx
+ roll $5,%edx
+ movl %edi,44(%esp)
+ leal 2400959708(%edi,%ebp,1),%edi
+ movl %ebx,%ebp
+ addl %edx,%edi
+ andl %ecx,%ebp
+ movl 48(%esp),%edx
+ addl %ebp,%edi
+
+ movl %eax,%ebp
+ xorl 56(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 16(%esp),%edx
+ andl %esi,%ebp
+ xorl 36(%esp),%edx
+ roll $1,%edx
+ addl %ecx,%ebp
+ rorl $2,%esi
+ movl %edi,%ecx
+ roll $5,%ecx
+ movl %edx,48(%esp)
+ leal 2400959708(%edx,%ebp,1),%edx
+ movl %eax,%ebp
+ addl %ecx,%edx
+ andl %ebx,%ebp
+ movl 52(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %esi,%ebp
+ xorl 60(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 20(%esp),%ecx
+ andl %edi,%ebp
+ xorl 40(%esp),%ecx
+ roll $1,%ecx
+ addl %ebx,%ebp
+ rorl $2,%edi
+ movl %edx,%ebx
+ roll $5,%ebx
+ movl %ecx,52(%esp)
+ leal 2400959708(%ecx,%ebp,1),%ecx
+ movl %esi,%ebp
+ addl %ebx,%ecx
+ andl %eax,%ebp
+ movl 56(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edi,%ebp
+ xorl (%esp),%ebx
+ xorl %esi,%ebp
+ xorl 24(%esp),%ebx
+ andl %edx,%ebp
+ xorl 44(%esp),%ebx
+ roll $1,%ebx
+ addl %eax,%ebp
+ rorl $2,%edx
+ movl %ecx,%eax
+ roll $5,%eax
+ movl %ebx,56(%esp)
+ leal 2400959708(%ebx,%ebp,1),%ebx
+ movl %edi,%ebp
+ addl %eax,%ebx
+ andl %esi,%ebp
+ movl 60(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %edx,%ebp
+ xorl 4(%esp),%eax
+ xorl %edi,%ebp
+ xorl 28(%esp),%eax
+ andl %ecx,%ebp
+ xorl 48(%esp),%eax
+ roll $1,%eax
+ addl %esi,%ebp
+ rorl $2,%ecx
+ movl %ebx,%esi
+ roll $5,%esi
+ movl %eax,60(%esp)
+ leal 2400959708(%eax,%ebp,1),%eax
+ movl %edx,%ebp
+ addl %esi,%eax
+ andl %edi,%ebp
+ movl (%esp),%esi
+ addl %ebp,%eax
+
+ movl %ecx,%ebp
+ xorl 8(%esp),%esi
+ xorl %edx,%ebp
+ xorl 32(%esp),%esi
+ andl %ebx,%ebp
+ xorl 52(%esp),%esi
+ roll $1,%esi
+ addl %edi,%ebp
+ rorl $2,%ebx
+ movl %eax,%edi
+ roll $5,%edi
+ movl %esi,(%esp)
+ leal 2400959708(%esi,%ebp,1),%esi
+ movl %ecx,%ebp
+ addl %edi,%esi
+ andl %edx,%ebp
+ movl 4(%esp),%edi
+ addl %ebp,%esi
+
+ movl %ebx,%ebp
+ xorl 12(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 36(%esp),%edi
+ andl %eax,%ebp
+ xorl 56(%esp),%edi
+ roll $1,%edi
+ addl %edx,%ebp
+ rorl $2,%eax
+ movl %esi,%edx
+ roll $5,%edx
+ movl %edi,4(%esp)
+ leal 2400959708(%edi,%ebp,1),%edi
+ movl %ebx,%ebp
+ addl %edx,%edi
+ andl %ecx,%ebp
+ movl 8(%esp),%edx
+ addl %ebp,%edi
+
+ movl %eax,%ebp
+ xorl 16(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 40(%esp),%edx
+ andl %esi,%ebp
+ xorl 60(%esp),%edx
+ roll $1,%edx
+ addl %ecx,%ebp
+ rorl $2,%esi
+ movl %edi,%ecx
+ roll $5,%ecx
+ movl %edx,8(%esp)
+ leal 2400959708(%edx,%ebp,1),%edx
+ movl %eax,%ebp
+ addl %ecx,%edx
+ andl %ebx,%ebp
+ movl 12(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %esi,%ebp
+ xorl 20(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 44(%esp),%ecx
+ andl %edi,%ebp
+ xorl (%esp),%ecx
+ roll $1,%ecx
+ addl %ebx,%ebp
+ rorl $2,%edi
+ movl %edx,%ebx
+ roll $5,%ebx
+ movl %ecx,12(%esp)
+ leal 2400959708(%ecx,%ebp,1),%ecx
+ movl %esi,%ebp
+ addl %ebx,%ecx
+ andl %eax,%ebp
+ movl 16(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edi,%ebp
+ xorl 24(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 48(%esp),%ebx
+ andl %edx,%ebp
+ xorl 4(%esp),%ebx
+ roll $1,%ebx
+ addl %eax,%ebp
+ rorl $2,%edx
+ movl %ecx,%eax
+ roll $5,%eax
+ movl %ebx,16(%esp)
+ leal 2400959708(%ebx,%ebp,1),%ebx
+ movl %edi,%ebp
+ addl %eax,%ebx
+ andl %esi,%ebp
+ movl 20(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %edx,%ebp
+ xorl 28(%esp),%eax
+ xorl %edi,%ebp
+ xorl 52(%esp),%eax
+ andl %ecx,%ebp
+ xorl 8(%esp),%eax
+ roll $1,%eax
+ addl %esi,%ebp
+ rorl $2,%ecx
+ movl %ebx,%esi
+ roll $5,%esi
+ movl %eax,20(%esp)
+ leal 2400959708(%eax,%ebp,1),%eax
+ movl %edx,%ebp
+ addl %esi,%eax
+ andl %edi,%ebp
+ movl 24(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ecx,%ebp
+ xorl 32(%esp),%esi
+ xorl %edx,%ebp
+ xorl 56(%esp),%esi
+ andl %ebx,%ebp
+ xorl 12(%esp),%esi
+ roll $1,%esi
+ addl %edi,%ebp
+ rorl $2,%ebx
+ movl %eax,%edi
+ roll $5,%edi
+ movl %esi,24(%esp)
+ leal 2400959708(%esi,%ebp,1),%esi
+ movl %ecx,%ebp
+ addl %edi,%esi
+ andl %edx,%ebp
+ movl 28(%esp),%edi
+ addl %ebp,%esi
+
+ movl %ebx,%ebp
+ xorl 36(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 60(%esp),%edi
+ andl %eax,%ebp
+ xorl 16(%esp),%edi
+ roll $1,%edi
+ addl %edx,%ebp
+ rorl $2,%eax
+ movl %esi,%edx
+ roll $5,%edx
+ movl %edi,28(%esp)
+ leal 2400959708(%edi,%ebp,1),%edi
+ movl %ebx,%ebp
+ addl %edx,%edi
+ andl %ecx,%ebp
+ movl 32(%esp),%edx
+ addl %ebp,%edi
+
+ movl %eax,%ebp
+ xorl 40(%esp),%edx
+ xorl %ebx,%ebp
+ xorl (%esp),%edx
+ andl %esi,%ebp
+ xorl 20(%esp),%edx
+ roll $1,%edx
+ addl %ecx,%ebp
+ rorl $2,%esi
+ movl %edi,%ecx
+ roll $5,%ecx
+ movl %edx,32(%esp)
+ leal 2400959708(%edx,%ebp,1),%edx
+ movl %eax,%ebp
+ addl %ecx,%edx
+ andl %ebx,%ebp
+ movl 36(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %esi,%ebp
+ xorl 44(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 4(%esp),%ecx
+ andl %edi,%ebp
+ xorl 24(%esp),%ecx
+ roll $1,%ecx
+ addl %ebx,%ebp
+ rorl $2,%edi
+ movl %edx,%ebx
+ roll $5,%ebx
+ movl %ecx,36(%esp)
+ leal 2400959708(%ecx,%ebp,1),%ecx
+ movl %esi,%ebp
+ addl %ebx,%ecx
+ andl %eax,%ebp
+ movl 40(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edi,%ebp
+ xorl 48(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 8(%esp),%ebx
+ andl %edx,%ebp
+ xorl 28(%esp),%ebx
+ roll $1,%ebx
+ addl %eax,%ebp
+ rorl $2,%edx
+ movl %ecx,%eax
+ roll $5,%eax
+ movl %ebx,40(%esp)
+ leal 2400959708(%ebx,%ebp,1),%ebx
+ movl %edi,%ebp
+ addl %eax,%ebx
+ andl %esi,%ebp
+ movl 44(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %edx,%ebp
+ xorl 52(%esp),%eax
+ xorl %edi,%ebp
+ xorl 12(%esp),%eax
+ andl %ecx,%ebp
+ xorl 32(%esp),%eax
+ roll $1,%eax
+ addl %esi,%ebp
+ rorl $2,%ecx
+ movl %ebx,%esi
+ roll $5,%esi
+ movl %eax,44(%esp)
+ leal 2400959708(%eax,%ebp,1),%eax
+ movl %edx,%ebp
+ addl %esi,%eax
+ andl %edi,%ebp
+ movl 48(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl 56(%esp),%esi
+ xorl %ecx,%ebp
+ xorl 16(%esp),%esi
+ xorl %edx,%ebp
+ xorl 36(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ movl %esi,48(%esp)
+ leal 3395469782(%esi,%edi,1),%esi
+ movl 52(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 60(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 20(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 40(%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ movl %edi,52(%esp)
+ leal 3395469782(%edi,%edx,1),%edi
+ movl 56(%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl (%esp),%edx
+ xorl %eax,%ebp
+ xorl 24(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 44(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,56(%esp)
+ leal 3395469782(%edx,%ecx,1),%edx
+ movl 60(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 4(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 28(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 48(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,60(%esp)
+ leal 3395469782(%ecx,%ebx,1),%ecx
+ movl (%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edx,%ebp
+ xorl 8(%esp),%ebx
+ xorl %edi,%ebp
+ xorl 32(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 52(%esp),%ebx
+ roll $1,%ebx
+ addl %ebp,%eax
+ rorl $2,%edx
+ movl %ecx,%ebp
+ roll $5,%ebp
+ movl %ebx,(%esp)
+ leal 3395469782(%ebx,%eax,1),%ebx
+ movl 4(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %ecx,%ebp
+ xorl 12(%esp),%eax
+ xorl %edx,%ebp
+ xorl 36(%esp),%eax
+ xorl %edi,%ebp
+ xorl 56(%esp),%eax
+ roll $1,%eax
+ addl %ebp,%esi
+ rorl $2,%ecx
+ movl %ebx,%ebp
+ roll $5,%ebp
+ movl %eax,4(%esp)
+ leal 3395469782(%eax,%esi,1),%eax
+ movl 8(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl 16(%esp),%esi
+ xorl %ecx,%ebp
+ xorl 40(%esp),%esi
+ xorl %edx,%ebp
+ xorl 60(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ movl %esi,8(%esp)
+ leal 3395469782(%esi,%edi,1),%esi
+ movl 12(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 20(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 44(%esp),%edi
+ xorl %ecx,%ebp
+ xorl (%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ movl %edi,12(%esp)
+ leal 3395469782(%edi,%edx,1),%edi
+ movl 16(%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl 24(%esp),%edx
+ xorl %eax,%ebp
+ xorl 48(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 4(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,16(%esp)
+ leal 3395469782(%edx,%ecx,1),%edx
+ movl 20(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 28(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 52(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 8(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,20(%esp)
+ leal 3395469782(%ecx,%ebx,1),%ecx
+ movl 24(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edx,%ebp
+ xorl 32(%esp),%ebx
+ xorl %edi,%ebp
+ xorl 56(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 12(%esp),%ebx
+ roll $1,%ebx
+ addl %ebp,%eax
+ rorl $2,%edx
+ movl %ecx,%ebp
+ roll $5,%ebp
+ movl %ebx,24(%esp)
+ leal 3395469782(%ebx,%eax,1),%ebx
+ movl 28(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %ecx,%ebp
+ xorl 36(%esp),%eax
+ xorl %edx,%ebp
+ xorl 60(%esp),%eax
+ xorl %edi,%ebp
+ xorl 16(%esp),%eax
+ roll $1,%eax
+ addl %ebp,%esi
+ rorl $2,%ecx
+ movl %ebx,%ebp
+ roll $5,%ebp
+ movl %eax,28(%esp)
+ leal 3395469782(%eax,%esi,1),%eax
+ movl 32(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl 40(%esp),%esi
+ xorl %ecx,%ebp
+ xorl (%esp),%esi
+ xorl %edx,%ebp
+ xorl 20(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ movl %esi,32(%esp)
+ leal 3395469782(%esi,%edi,1),%esi
+ movl 36(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 44(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 4(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 24(%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ movl %edi,36(%esp)
+ leal 3395469782(%edi,%edx,1),%edi
+ movl 40(%esp),%edx
+ addl %ebp,%edi
+
+ movl %esi,%ebp
+ xorl 48(%esp),%edx
+ xorl %eax,%ebp
+ xorl 8(%esp),%edx
+ xorl %ebx,%ebp
+ xorl 28(%esp),%edx
+ roll $1,%edx
+ addl %ebp,%ecx
+ rorl $2,%esi
+ movl %edi,%ebp
+ roll $5,%ebp
+ movl %edx,40(%esp)
+ leal 3395469782(%edx,%ecx,1),%edx
+ movl 44(%esp),%ecx
+ addl %ebp,%edx
+
+ movl %edi,%ebp
+ xorl 52(%esp),%ecx
+ xorl %esi,%ebp
+ xorl 12(%esp),%ecx
+ xorl %eax,%ebp
+ xorl 32(%esp),%ecx
+ roll $1,%ecx
+ addl %ebp,%ebx
+ rorl $2,%edi
+ movl %edx,%ebp
+ roll $5,%ebp
+ movl %ecx,44(%esp)
+ leal 3395469782(%ecx,%ebx,1),%ecx
+ movl 48(%esp),%ebx
+ addl %ebp,%ecx
+
+ movl %edx,%ebp
+ xorl 56(%esp),%ebx
+ xorl %edi,%ebp
+ xorl 16(%esp),%ebx
+ xorl %esi,%ebp
+ xorl 36(%esp),%ebx
+ roll $1,%ebx
+ addl %ebp,%eax
+ rorl $2,%edx
+ movl %ecx,%ebp
+ roll $5,%ebp
+ movl %ebx,48(%esp)
+ leal 3395469782(%ebx,%eax,1),%ebx
+ movl 52(%esp),%eax
+ addl %ebp,%ebx
+
+ movl %ecx,%ebp
+ xorl 60(%esp),%eax
+ xorl %edx,%ebp
+ xorl 20(%esp),%eax
+ xorl %edi,%ebp
+ xorl 40(%esp),%eax
+ roll $1,%eax
+ addl %ebp,%esi
+ rorl $2,%ecx
+ movl %ebx,%ebp
+ roll $5,%ebp
+ leal 3395469782(%eax,%esi,1),%eax
+ movl 56(%esp),%esi
+ addl %ebp,%eax
+
+ movl %ebx,%ebp
+ xorl (%esp),%esi
+ xorl %ecx,%ebp
+ xorl 24(%esp),%esi
+ xorl %edx,%ebp
+ xorl 44(%esp),%esi
+ roll $1,%esi
+ addl %ebp,%edi
+ rorl $2,%ebx
+ movl %eax,%ebp
+ roll $5,%ebp
+ leal 3395469782(%esi,%edi,1),%esi
+ movl 60(%esp),%edi
+ addl %ebp,%esi
+
+ movl %eax,%ebp
+ xorl 4(%esp),%edi
+ xorl %ebx,%ebp
+ xorl 28(%esp),%edi
+ xorl %ecx,%ebp
+ xorl 48(%esp),%edi
+ roll $1,%edi
+ addl %ebp,%edx
+ rorl $2,%eax
+ movl %esi,%ebp
+ roll $5,%ebp
+ leal 3395469782(%edi,%edx,1),%edi
+ addl %ebp,%edi
+ movl 96(%esp),%ebp
+ movl 100(%esp),%edx
+ addl (%ebp),%edi
+ addl 4(%ebp),%esi
+ addl 8(%ebp),%eax
+ addl 12(%ebp),%ebx
+ addl 16(%ebp),%ecx
+ movl %edi,(%ebp)
+ addl $64,%edx
+ movl %esi,4(%ebp)
+ cmpl 104(%esp),%edx
+ movl %eax,8(%ebp)
+ movl %ecx,%edi
+ movl %ebx,12(%ebp)
+ movl %edx,%esi
+ movl %ecx,16(%ebp)
+ jb .L000loop
+ addl $76,%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size sha1_block_data_order,.-.L_sha1_block_data_order_begin
+.byte 83,72,65,49,32,98,108,111,99,107,32,116,114,97,110,115
+.byte 102,111,114,109,32,102,111,114,32,120,56,54,44,32,67,82
+.byte 89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112
+.byte 114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha256-586.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha256-586.S
new file mode 100644
index 0000000000..260dde7ad7
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha256-586.S
@@ -0,0 +1,3356 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/sha/asm/sha256-586.pl
+#
+# Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.globl sha256_block_data_order
+.type sha256_block_data_order,@function
+.align 16
+sha256_block_data_order:
+.L_sha256_block_data_order_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl %esp,%ebx
+ call .L000pic_point
+.L000pic_point:
+ popl %ebp
+ leal .L001K256-.L000pic_point(%ebp),%ebp
+ subl $16,%esp
+ andl $-64,%esp
+ shll $6,%eax
+ addl %edi,%eax
+ movl %esi,(%esp)
+ movl %edi,4(%esp)
+ movl %eax,8(%esp)
+ movl %ebx,12(%esp)
+ jmp .L002loop
+.align 16
+.L002loop:
+ movl (%edi),%eax
+ movl 4(%edi),%ebx
+ movl 8(%edi),%ecx
+ bswap %eax
+ movl 12(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ movl 16(%edi),%eax
+ movl 20(%edi),%ebx
+ movl 24(%edi),%ecx
+ bswap %eax
+ movl 28(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ movl 32(%edi),%eax
+ movl 36(%edi),%ebx
+ movl 40(%edi),%ecx
+ bswap %eax
+ movl 44(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ movl 48(%edi),%eax
+ movl 52(%edi),%ebx
+ movl 56(%edi),%ecx
+ bswap %eax
+ movl 60(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ addl $64,%edi
+ leal -36(%esp),%esp
+ movl %edi,104(%esp)
+ movl (%esi),%eax
+ movl 4(%esi),%ebx
+ movl 8(%esi),%ecx
+ movl 12(%esi),%edi
+ movl %ebx,8(%esp)
+ xorl %ecx,%ebx
+ movl %ecx,12(%esp)
+ movl %edi,16(%esp)
+ movl %ebx,(%esp)
+ movl 16(%esi),%edx
+ movl 20(%esi),%ebx
+ movl 24(%esi),%ecx
+ movl 28(%esi),%edi
+ movl %ebx,24(%esp)
+ movl %ecx,28(%esp)
+ movl %edi,32(%esp)
+.align 16
+.L00300_15:
+ movl %edx,%ecx
+ movl 24(%esp),%esi
+ rorl $14,%ecx
+ movl 28(%esp),%edi
+ xorl %edx,%ecx
+ xorl %edi,%esi
+ movl 96(%esp),%ebx
+ rorl $5,%ecx
+ andl %edx,%esi
+ movl %edx,20(%esp)
+ xorl %ecx,%edx
+ addl 32(%esp),%ebx
+ xorl %edi,%esi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %esi,%ebx
+ rorl $9,%ecx
+ addl %edx,%ebx
+ movl 8(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,4(%esp)
+ leal -4(%esp),%esp
+ rorl $11,%ecx
+ movl (%ebp),%esi
+ xorl %eax,%ecx
+ movl 20(%esp),%edx
+ xorl %edi,%eax
+ rorl $2,%ecx
+ addl %esi,%ebx
+ movl %eax,(%esp)
+ addl %ebx,%edx
+ andl 4(%esp),%eax
+ addl %ecx,%ebx
+ xorl %edi,%eax
+ addl $4,%ebp
+ addl %ebx,%eax
+ cmpl $3248222580,%esi
+ jne .L00300_15
+ movl 156(%esp),%ecx
+ jmp .L00416_63
+.align 16
+.L00416_63:
+ movl %ecx,%ebx
+ movl 104(%esp),%esi
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 160(%esp),%ebx
+ shrl $10,%edi
+ addl 124(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 24(%esp),%esi
+ rorl $14,%ecx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %edx,%ecx
+ xorl %edi,%esi
+ movl %ebx,96(%esp)
+ rorl $5,%ecx
+ andl %edx,%esi
+ movl %edx,20(%esp)
+ xorl %ecx,%edx
+ addl 32(%esp),%ebx
+ xorl %edi,%esi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %esi,%ebx
+ rorl $9,%ecx
+ addl %edx,%ebx
+ movl 8(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,4(%esp)
+ leal -4(%esp),%esp
+ rorl $11,%ecx
+ movl (%ebp),%esi
+ xorl %eax,%ecx
+ movl 20(%esp),%edx
+ xorl %edi,%eax
+ rorl $2,%ecx
+ addl %esi,%ebx
+ movl %eax,(%esp)
+ addl %ebx,%edx
+ andl 4(%esp),%eax
+ addl %ecx,%ebx
+ xorl %edi,%eax
+ movl 156(%esp),%ecx
+ addl $4,%ebp
+ addl %ebx,%eax
+ cmpl $3329325298,%esi
+ jne .L00416_63
+ movl 356(%esp),%esi
+ movl 8(%esp),%ebx
+ movl 16(%esp),%ecx
+ addl (%esi),%eax
+ addl 4(%esi),%ebx
+ addl 8(%esi),%edi
+ addl 12(%esi),%ecx
+ movl %eax,(%esi)
+ movl %ebx,4(%esi)
+ movl %edi,8(%esi)
+ movl %ecx,12(%esi)
+ movl 24(%esp),%eax
+ movl 28(%esp),%ebx
+ movl 32(%esp),%ecx
+ movl 360(%esp),%edi
+ addl 16(%esi),%edx
+ addl 20(%esi),%eax
+ addl 24(%esi),%ebx
+ addl 28(%esi),%ecx
+ movl %edx,16(%esi)
+ movl %eax,20(%esi)
+ movl %ebx,24(%esi)
+ movl %ecx,28(%esi)
+ leal 356(%esp),%esp
+ subl $256,%ebp
+ cmpl 8(%esp),%edi
+ jb .L002loop
+ movl 12(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.align 32
+.L005loop_shrd:
+ movl (%edi),%eax
+ movl 4(%edi),%ebx
+ movl 8(%edi),%ecx
+ bswap %eax
+ movl 12(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ movl 16(%edi),%eax
+ movl 20(%edi),%ebx
+ movl 24(%edi),%ecx
+ bswap %eax
+ movl 28(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ movl 32(%edi),%eax
+ movl 36(%edi),%ebx
+ movl 40(%edi),%ecx
+ bswap %eax
+ movl 44(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ movl 48(%edi),%eax
+ movl 52(%edi),%ebx
+ movl 56(%edi),%ecx
+ bswap %eax
+ movl 60(%edi),%edx
+ bswap %ebx
+ pushl %eax
+ bswap %ecx
+ pushl %ebx
+ bswap %edx
+ pushl %ecx
+ pushl %edx
+ addl $64,%edi
+ leal -36(%esp),%esp
+ movl %edi,104(%esp)
+ movl (%esi),%eax
+ movl 4(%esi),%ebx
+ movl 8(%esi),%ecx
+ movl 12(%esi),%edi
+ movl %ebx,8(%esp)
+ xorl %ecx,%ebx
+ movl %ecx,12(%esp)
+ movl %edi,16(%esp)
+ movl %ebx,(%esp)
+ movl 16(%esi),%edx
+ movl 20(%esi),%ebx
+ movl 24(%esi),%ecx
+ movl 28(%esi),%edi
+ movl %ebx,24(%esp)
+ movl %ecx,28(%esp)
+ movl %edi,32(%esp)
+.align 16
+.L00600_15_shrd:
+ movl %edx,%ecx
+ movl 24(%esp),%esi
+ shrdl $14,%ecx,%ecx
+ movl 28(%esp),%edi
+ xorl %edx,%ecx
+ xorl %edi,%esi
+ movl 96(%esp),%ebx
+ shrdl $5,%ecx,%ecx
+ andl %edx,%esi
+ movl %edx,20(%esp)
+ xorl %ecx,%edx
+ addl 32(%esp),%ebx
+ xorl %edi,%esi
+ shrdl $6,%edx,%edx
+ movl %eax,%ecx
+ addl %esi,%ebx
+ shrdl $9,%ecx,%ecx
+ addl %edx,%ebx
+ movl 8(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,4(%esp)
+ leal -4(%esp),%esp
+ shrdl $11,%ecx,%ecx
+ movl (%ebp),%esi
+ xorl %eax,%ecx
+ movl 20(%esp),%edx
+ xorl %edi,%eax
+ shrdl $2,%ecx,%ecx
+ addl %esi,%ebx
+ movl %eax,(%esp)
+ addl %ebx,%edx
+ andl 4(%esp),%eax
+ addl %ecx,%ebx
+ xorl %edi,%eax
+ addl $4,%ebp
+ addl %ebx,%eax
+ cmpl $3248222580,%esi
+ jne .L00600_15_shrd
+ movl 156(%esp),%ecx
+ jmp .L00716_63_shrd
+.align 16
+.L00716_63_shrd:
+ movl %ecx,%ebx
+ movl 104(%esp),%esi
+ shrdl $11,%ecx,%ecx
+ movl %esi,%edi
+ shrdl $2,%esi,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ shrdl $7,%ecx,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ shrdl $17,%esi,%esi
+ addl 160(%esp),%ebx
+ shrl $10,%edi
+ addl 124(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 24(%esp),%esi
+ shrdl $14,%ecx,%ecx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %edx,%ecx
+ xorl %edi,%esi
+ movl %ebx,96(%esp)
+ shrdl $5,%ecx,%ecx
+ andl %edx,%esi
+ movl %edx,20(%esp)
+ xorl %ecx,%edx
+ addl 32(%esp),%ebx
+ xorl %edi,%esi
+ shrdl $6,%edx,%edx
+ movl %eax,%ecx
+ addl %esi,%ebx
+ shrdl $9,%ecx,%ecx
+ addl %edx,%ebx
+ movl 8(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,4(%esp)
+ leal -4(%esp),%esp
+ shrdl $11,%ecx,%ecx
+ movl (%ebp),%esi
+ xorl %eax,%ecx
+ movl 20(%esp),%edx
+ xorl %edi,%eax
+ shrdl $2,%ecx,%ecx
+ addl %esi,%ebx
+ movl %eax,(%esp)
+ addl %ebx,%edx
+ andl 4(%esp),%eax
+ addl %ecx,%ebx
+ xorl %edi,%eax
+ movl 156(%esp),%ecx
+ addl $4,%ebp
+ addl %ebx,%eax
+ cmpl $3329325298,%esi
+ jne .L00716_63_shrd
+ movl 356(%esp),%esi
+ movl 8(%esp),%ebx
+ movl 16(%esp),%ecx
+ addl (%esi),%eax
+ addl 4(%esi),%ebx
+ addl 8(%esi),%edi
+ addl 12(%esi),%ecx
+ movl %eax,(%esi)
+ movl %ebx,4(%esi)
+ movl %edi,8(%esi)
+ movl %ecx,12(%esi)
+ movl 24(%esp),%eax
+ movl 28(%esp),%ebx
+ movl 32(%esp),%ecx
+ movl 360(%esp),%edi
+ addl 16(%esi),%edx
+ addl 20(%esi),%eax
+ addl 24(%esi),%ebx
+ addl 28(%esi),%ecx
+ movl %edx,16(%esi)
+ movl %eax,20(%esi)
+ movl %ebx,24(%esi)
+ movl %ecx,28(%esi)
+ leal 356(%esp),%esp
+ subl $256,%ebp
+ cmpl 8(%esp),%edi
+ jb .L005loop_shrd
+ movl 12(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.align 64
+.L001K256:
+.long 1116352408,1899447441,3049323471,3921009573,961987163,1508970993,2453635748,2870763221,3624381080,310598401,607225278,1426881987,1925078388,2162078206,2614888103,3248222580,3835390401,4022224774,264347078,604807628,770255983,1249150122,1555081692,1996064986,2554220882,2821834349,2952996808,3210313671,3336571891,3584528711,113926993,338241895,666307205,773529912,1294757372,1396182291,1695183700,1986661051,2177026350,2456956037,2730485921,2820302411,3259730800,3345764771,3516065817,3600352804,4094571909,275423344,430227734,506948616,659060556,883997877,958139571,1322822218,1537002063,1747873779,1955562222,2024104815,2227730452,2361852424,2428436474,2756734187,3204031479,3329325298
+.long 66051,67438087,134810123,202182159
+.byte 83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97
+.byte 110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+.byte 67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte 112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte 62,0
+.align 16
+.L008unrolled:
+ leal -96(%esp),%esp
+ movl (%esi),%eax
+ movl 4(%esi),%ebp
+ movl 8(%esi),%ecx
+ movl 12(%esi),%ebx
+ movl %ebp,4(%esp)
+ xorl %ecx,%ebp
+ movl %ecx,8(%esp)
+ movl %ebx,12(%esp)
+ movl 16(%esi),%edx
+ movl 20(%esi),%ebx
+ movl 24(%esi),%ecx
+ movl 28(%esi),%esi
+ movl %ebx,20(%esp)
+ movl %ecx,24(%esp)
+ movl %esi,28(%esp)
+ jmp .L009grand_loop
+.align 16
+.L009grand_loop:
+ movl (%edi),%ebx
+ movl 4(%edi),%ecx
+ bswap %ebx
+ movl 8(%edi),%esi
+ bswap %ecx
+ movl %ebx,32(%esp)
+ bswap %esi
+ movl %ecx,36(%esp)
+ movl %esi,40(%esp)
+ movl 12(%edi),%ebx
+ movl 16(%edi),%ecx
+ bswap %ebx
+ movl 20(%edi),%esi
+ bswap %ecx
+ movl %ebx,44(%esp)
+ bswap %esi
+ movl %ecx,48(%esp)
+ movl %esi,52(%esp)
+ movl 24(%edi),%ebx
+ movl 28(%edi),%ecx
+ bswap %ebx
+ movl 32(%edi),%esi
+ bswap %ecx
+ movl %ebx,56(%esp)
+ bswap %esi
+ movl %ecx,60(%esp)
+ movl %esi,64(%esp)
+ movl 36(%edi),%ebx
+ movl 40(%edi),%ecx
+ bswap %ebx
+ movl 44(%edi),%esi
+ bswap %ecx
+ movl %ebx,68(%esp)
+ bswap %esi
+ movl %ecx,72(%esp)
+ movl %esi,76(%esp)
+ movl 48(%edi),%ebx
+ movl 52(%edi),%ecx
+ bswap %ebx
+ movl 56(%edi),%esi
+ bswap %ecx
+ movl %ebx,80(%esp)
+ bswap %esi
+ movl %ecx,84(%esp)
+ movl %esi,88(%esp)
+ movl 60(%edi),%ebx
+ addl $64,%edi
+ bswap %ebx
+ movl %edi,100(%esp)
+ movl %ebx,92(%esp)
+ movl %edx,%ecx
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl 32(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1116352408(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl 36(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1899447441(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl 40(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3049323471(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl 44(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3921009573(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl 48(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 961987163(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl 52(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1508970993(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl 56(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2453635748(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl 60(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2870763221(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl 64(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3624381080(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl 68(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 310598401(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl 72(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 607225278(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl 76(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1426881987(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl 80(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1925078388(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl 84(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2162078206(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl %edx,%ecx
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl 88(%esp),%ebx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2614888103(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl %edx,%esi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl 92(%esp),%ebx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3248222580(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 36(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 88(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 32(%esp),%ebx
+ shrl $10,%edi
+ addl 68(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,32(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3835390401(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 40(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl 92(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 36(%esp),%ebx
+ shrl $10,%edi
+ addl 72(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,36(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 4022224774(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 44(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl 32(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 40(%esp),%ebx
+ shrl $10,%edi
+ addl 76(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,40(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 264347078(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 48(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl 36(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 44(%esp),%ebx
+ shrl $10,%edi
+ addl 80(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,44(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 604807628(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 52(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl 40(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 48(%esp),%ebx
+ shrl $10,%edi
+ addl 84(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,48(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 770255983(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 56(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl 44(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 52(%esp),%ebx
+ shrl $10,%edi
+ addl 88(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,52(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1249150122(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 60(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl 48(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 56(%esp),%ebx
+ shrl $10,%edi
+ addl 92(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,56(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1555081692(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 64(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl 52(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 60(%esp),%ebx
+ shrl $10,%edi
+ addl 32(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,60(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1996064986(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 68(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 56(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 64(%esp),%ebx
+ shrl $10,%edi
+ addl 36(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,64(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2554220882(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 72(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl 60(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 68(%esp),%ebx
+ shrl $10,%edi
+ addl 40(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,68(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2821834349(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 76(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl 64(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 72(%esp),%ebx
+ shrl $10,%edi
+ addl 44(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,72(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2952996808(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 80(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl 68(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 76(%esp),%ebx
+ shrl $10,%edi
+ addl 48(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,76(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3210313671(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 84(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl 72(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 80(%esp),%ebx
+ shrl $10,%edi
+ addl 52(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,80(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3336571891(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 88(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl 76(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 84(%esp),%ebx
+ shrl $10,%edi
+ addl 56(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,84(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3584528711(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 92(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl 80(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 88(%esp),%ebx
+ shrl $10,%edi
+ addl 60(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,88(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 113926993(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 32(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl 84(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 92(%esp),%ebx
+ shrl $10,%edi
+ addl 64(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,92(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 338241895(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 36(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 88(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 32(%esp),%ebx
+ shrl $10,%edi
+ addl 68(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,32(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 666307205(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 40(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl 92(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 36(%esp),%ebx
+ shrl $10,%edi
+ addl 72(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,36(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 773529912(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 44(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl 32(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 40(%esp),%ebx
+ shrl $10,%edi
+ addl 76(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,40(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1294757372(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 48(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl 36(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 44(%esp),%ebx
+ shrl $10,%edi
+ addl 80(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,44(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1396182291(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 52(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl 40(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 48(%esp),%ebx
+ shrl $10,%edi
+ addl 84(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,48(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1695183700(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 56(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl 44(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 52(%esp),%ebx
+ shrl $10,%edi
+ addl 88(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,52(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1986661051(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 60(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl 48(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 56(%esp),%ebx
+ shrl $10,%edi
+ addl 92(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,56(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2177026350(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 64(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl 52(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 60(%esp),%ebx
+ shrl $10,%edi
+ addl 32(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,60(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2456956037(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 68(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 56(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 64(%esp),%ebx
+ shrl $10,%edi
+ addl 36(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,64(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2730485921(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 72(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl 60(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 68(%esp),%ebx
+ shrl $10,%edi
+ addl 40(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,68(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2820302411(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 76(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl 64(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 72(%esp),%ebx
+ shrl $10,%edi
+ addl 44(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,72(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3259730800(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 80(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl 68(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 76(%esp),%ebx
+ shrl $10,%edi
+ addl 48(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,76(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3345764771(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 84(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl 72(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 80(%esp),%ebx
+ shrl $10,%edi
+ addl 52(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,80(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3516065817(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 88(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl 76(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 84(%esp),%ebx
+ shrl $10,%edi
+ addl 56(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,84(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3600352804(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 92(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl 80(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 88(%esp),%ebx
+ shrl $10,%edi
+ addl 60(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,88(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 4094571909(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 32(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl 84(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 92(%esp),%ebx
+ shrl $10,%edi
+ addl 64(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,92(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 275423344(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 36(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 88(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 32(%esp),%ebx
+ shrl $10,%edi
+ addl 68(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,32(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 430227734(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 40(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl 92(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 36(%esp),%ebx
+ shrl $10,%edi
+ addl 72(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,36(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 506948616(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 44(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl 32(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 40(%esp),%ebx
+ shrl $10,%edi
+ addl 76(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,40(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 659060556(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 48(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl 36(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 44(%esp),%ebx
+ shrl $10,%edi
+ addl 80(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,44(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 883997877(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 52(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl 40(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 48(%esp),%ebx
+ shrl $10,%edi
+ addl 84(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,48(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 958139571(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 56(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl 44(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 52(%esp),%ebx
+ shrl $10,%edi
+ addl 88(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,52(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1322822218(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 60(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl 48(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 56(%esp),%ebx
+ shrl $10,%edi
+ addl 92(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,56(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1537002063(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 64(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl 52(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 60(%esp),%ebx
+ shrl $10,%edi
+ addl 32(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,60(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 1747873779(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 68(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 56(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 64(%esp),%ebx
+ shrl $10,%edi
+ addl 36(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 20(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 24(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,64(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,16(%esp)
+ xorl %ecx,%edx
+ addl 28(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 4(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 1955562222(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 72(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 12(%esp),%edx
+ addl %ecx,%ebp
+ movl 60(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 68(%esp),%ebx
+ shrl $10,%edi
+ addl 40(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 16(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 20(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,68(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,12(%esp)
+ xorl %esi,%edx
+ addl 24(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl (%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,28(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2024104815(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 76(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 8(%esp),%edx
+ addl %esi,%eax
+ movl 64(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 72(%esp),%ebx
+ shrl $10,%edi
+ addl 44(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 12(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 16(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,72(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,8(%esp)
+ xorl %ecx,%edx
+ addl 20(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 28(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,24(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2227730452(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 80(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 4(%esp),%edx
+ addl %ecx,%ebp
+ movl 68(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 76(%esp),%ebx
+ shrl $10,%edi
+ addl 48(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 8(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 12(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,76(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,4(%esp)
+ xorl %esi,%edx
+ addl 16(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 24(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,20(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2361852424(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 84(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl (%esp),%edx
+ addl %esi,%eax
+ movl 72(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 80(%esp),%ebx
+ shrl $10,%edi
+ addl 52(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 4(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 8(%esp),%edi
+ xorl %ecx,%edx
+ movl %ebx,80(%esp)
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,(%esp)
+ xorl %ecx,%edx
+ addl 12(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 20(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,16(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 2428436474(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 88(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 28(%esp),%edx
+ addl %ecx,%ebp
+ movl 76(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 84(%esp),%ebx
+ shrl $10,%edi
+ addl 56(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl (%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 4(%esp),%edi
+ xorl %esi,%edx
+ movl %ebx,84(%esp)
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,28(%esp)
+ xorl %esi,%edx
+ addl 8(%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 16(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,12(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 2756734187(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ movl 92(%esp),%ecx
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 24(%esp),%edx
+ addl %esi,%eax
+ movl 80(%esp),%esi
+ movl %ecx,%ebx
+ rorl $11,%ecx
+ movl %esi,%edi
+ rorl $2,%esi
+ xorl %ebx,%ecx
+ shrl $3,%ebx
+ rorl $7,%ecx
+ xorl %edi,%esi
+ xorl %ecx,%ebx
+ rorl $17,%esi
+ addl 88(%esp),%ebx
+ shrl $10,%edi
+ addl 60(%esp),%ebx
+ movl %edx,%ecx
+ xorl %esi,%edi
+ movl 28(%esp),%esi
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl (%esp),%edi
+ xorl %ecx,%edx
+ xorl %edi,%esi
+ rorl $5,%edx
+ andl %ecx,%esi
+ movl %ecx,24(%esp)
+ xorl %ecx,%edx
+ addl 4(%esp),%ebx
+ xorl %esi,%edi
+ rorl $6,%edx
+ movl %eax,%ecx
+ addl %edi,%ebx
+ rorl $9,%ecx
+ movl %eax,%esi
+ movl 12(%esp),%edi
+ xorl %eax,%ecx
+ movl %eax,8(%esp)
+ xorl %edi,%eax
+ rorl $11,%ecx
+ andl %eax,%ebp
+ leal 3204031479(%ebx,%edx,1),%edx
+ xorl %esi,%ecx
+ xorl %edi,%ebp
+ movl 32(%esp),%esi
+ rorl $2,%ecx
+ addl %edx,%ebp
+ addl 20(%esp),%edx
+ addl %ecx,%ebp
+ movl 84(%esp),%ecx
+ movl %esi,%ebx
+ rorl $11,%esi
+ movl %ecx,%edi
+ rorl $2,%ecx
+ xorl %ebx,%esi
+ shrl $3,%ebx
+ rorl $7,%esi
+ xorl %edi,%ecx
+ xorl %esi,%ebx
+ rorl $17,%ecx
+ addl 92(%esp),%ebx
+ shrl $10,%edi
+ addl 64(%esp),%ebx
+ movl %edx,%esi
+ xorl %ecx,%edi
+ movl 24(%esp),%ecx
+ rorl $14,%edx
+ addl %edi,%ebx
+ movl 28(%esp),%edi
+ xorl %esi,%edx
+ xorl %edi,%ecx
+ rorl $5,%edx
+ andl %esi,%ecx
+ movl %esi,20(%esp)
+ xorl %esi,%edx
+ addl (%esp),%ebx
+ xorl %ecx,%edi
+ rorl $6,%edx
+ movl %ebp,%esi
+ addl %edi,%ebx
+ rorl $9,%esi
+ movl %ebp,%ecx
+ movl 8(%esp),%edi
+ xorl %ebp,%esi
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ rorl $11,%esi
+ andl %ebp,%eax
+ leal 3329325298(%ebx,%edx,1),%edx
+ xorl %ecx,%esi
+ xorl %edi,%eax
+ rorl $2,%esi
+ addl %edx,%eax
+ addl 16(%esp),%edx
+ addl %esi,%eax
+ movl 96(%esp),%esi
+ xorl %edi,%ebp
+ movl 12(%esp),%ecx
+ addl (%esi),%eax
+ addl 4(%esi),%ebp
+ addl 8(%esi),%edi
+ addl 12(%esi),%ecx
+ movl %eax,(%esi)
+ movl %ebp,4(%esi)
+ movl %edi,8(%esi)
+ movl %ecx,12(%esi)
+ movl %ebp,4(%esp)
+ xorl %edi,%ebp
+ movl %edi,8(%esp)
+ movl %ecx,12(%esp)
+ movl 20(%esp),%edi
+ movl 24(%esp),%ebx
+ movl 28(%esp),%ecx
+ addl 16(%esi),%edx
+ addl 20(%esi),%edi
+ addl 24(%esi),%ebx
+ addl 28(%esi),%ecx
+ movl %edx,16(%esi)
+ movl %edi,20(%esi)
+ movl %ebx,24(%esi)
+ movl %ecx,28(%esi)
+ movl %edi,20(%esp)
+ movl 100(%esp),%edi
+ movl %ebx,24(%esp)
+ movl %ecx,28(%esp)
+ cmpl 104(%esp),%edi
+ jb .L009grand_loop
+ movl 108(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size sha256_block_data_order,.-.L_sha256_block_data_order_begin
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha512-586.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha512-586.S
new file mode 100644
index 0000000000..c01cc2bed8
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/sha/sha512-586.S
@@ -0,0 +1,574 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/sha/asm/sha512-586.pl
+#
+# Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.globl sha512_block_data_order
+.type sha512_block_data_order,@function
+.align 16
+sha512_block_data_order:
+.L_sha512_block_data_order_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl 20(%esp),%esi
+ movl 24(%esp),%edi
+ movl 28(%esp),%eax
+ movl %esp,%ebx
+ call .L000pic_point
+.L000pic_point:
+ popl %ebp
+ leal .L001K512-.L000pic_point(%ebp),%ebp
+ subl $16,%esp
+ andl $-64,%esp
+ shll $7,%eax
+ addl %edi,%eax
+ movl %esi,(%esp)
+ movl %edi,4(%esp)
+ movl %eax,8(%esp)
+ movl %ebx,12(%esp)
+.align 16
+.L002loop_x86:
+ movl (%edi),%eax
+ movl 4(%edi),%ebx
+ movl 8(%edi),%ecx
+ movl 12(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 16(%edi),%eax
+ movl 20(%edi),%ebx
+ movl 24(%edi),%ecx
+ movl 28(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 32(%edi),%eax
+ movl 36(%edi),%ebx
+ movl 40(%edi),%ecx
+ movl 44(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 48(%edi),%eax
+ movl 52(%edi),%ebx
+ movl 56(%edi),%ecx
+ movl 60(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 64(%edi),%eax
+ movl 68(%edi),%ebx
+ movl 72(%edi),%ecx
+ movl 76(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 80(%edi),%eax
+ movl 84(%edi),%ebx
+ movl 88(%edi),%ecx
+ movl 92(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 96(%edi),%eax
+ movl 100(%edi),%ebx
+ movl 104(%edi),%ecx
+ movl 108(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ movl 112(%edi),%eax
+ movl 116(%edi),%ebx
+ movl 120(%edi),%ecx
+ movl 124(%edi),%edx
+ bswap %eax
+ bswap %ebx
+ bswap %ecx
+ bswap %edx
+ pushl %eax
+ pushl %ebx
+ pushl %ecx
+ pushl %edx
+ addl $128,%edi
+ subl $72,%esp
+ movl %edi,204(%esp)
+ leal 8(%esp),%edi
+ movl $16,%ecx
+.long 2784229001
+.align 16
+.L00300_15_x86:
+ movl 40(%esp),%ecx
+ movl 44(%esp),%edx
+ movl %ecx,%esi
+ shrl $9,%ecx
+ movl %edx,%edi
+ shrl $9,%edx
+ movl %ecx,%ebx
+ shll $14,%esi
+ movl %edx,%eax
+ shll $14,%edi
+ xorl %esi,%ebx
+ shrl $5,%ecx
+ xorl %edi,%eax
+ shrl $5,%edx
+ xorl %ecx,%eax
+ shll $4,%esi
+ xorl %edx,%ebx
+ shll $4,%edi
+ xorl %esi,%ebx
+ shrl $4,%ecx
+ xorl %edi,%eax
+ shrl $4,%edx
+ xorl %ecx,%eax
+ shll $5,%esi
+ xorl %edx,%ebx
+ shll $5,%edi
+ xorl %esi,%eax
+ xorl %edi,%ebx
+ movl 48(%esp),%ecx
+ movl 52(%esp),%edx
+ movl 56(%esp),%esi
+ movl 60(%esp),%edi
+ addl 64(%esp),%eax
+ adcl 68(%esp),%ebx
+ xorl %esi,%ecx
+ xorl %edi,%edx
+ andl 40(%esp),%ecx
+ andl 44(%esp),%edx
+ addl 192(%esp),%eax
+ adcl 196(%esp),%ebx
+ xorl %esi,%ecx
+ xorl %edi,%edx
+ movl (%ebp),%esi
+ movl 4(%ebp),%edi
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ movl 32(%esp),%ecx
+ movl 36(%esp),%edx
+ addl %esi,%eax
+ adcl %edi,%ebx
+ movl %eax,(%esp)
+ movl %ebx,4(%esp)
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ movl 8(%esp),%ecx
+ movl 12(%esp),%edx
+ movl %eax,32(%esp)
+ movl %ebx,36(%esp)
+ movl %ecx,%esi
+ shrl $2,%ecx
+ movl %edx,%edi
+ shrl $2,%edx
+ movl %ecx,%ebx
+ shll $4,%esi
+ movl %edx,%eax
+ shll $4,%edi
+ xorl %esi,%ebx
+ shrl $5,%ecx
+ xorl %edi,%eax
+ shrl $5,%edx
+ xorl %ecx,%ebx
+ shll $21,%esi
+ xorl %edx,%eax
+ shll $21,%edi
+ xorl %esi,%eax
+ shrl $21,%ecx
+ xorl %edi,%ebx
+ shrl $21,%edx
+ xorl %ecx,%eax
+ shll $5,%esi
+ xorl %edx,%ebx
+ shll $5,%edi
+ xorl %esi,%eax
+ xorl %edi,%ebx
+ movl 8(%esp),%ecx
+ movl 12(%esp),%edx
+ movl 16(%esp),%esi
+ movl 20(%esp),%edi
+ addl (%esp),%eax
+ adcl 4(%esp),%ebx
+ orl %esi,%ecx
+ orl %edi,%edx
+ andl 24(%esp),%ecx
+ andl 28(%esp),%edx
+ andl 8(%esp),%esi
+ andl 12(%esp),%edi
+ orl %esi,%ecx
+ orl %edi,%edx
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ movl %eax,(%esp)
+ movl %ebx,4(%esp)
+ movb (%ebp),%dl
+ subl $8,%esp
+ leal 8(%ebp),%ebp
+ cmpb $148,%dl
+ jne .L00300_15_x86
+.align 16
+.L00416_79_x86:
+ movl 312(%esp),%ecx
+ movl 316(%esp),%edx
+ movl %ecx,%esi
+ shrl $1,%ecx
+ movl %edx,%edi
+ shrl $1,%edx
+ movl %ecx,%eax
+ shll $24,%esi
+ movl %edx,%ebx
+ shll $24,%edi
+ xorl %esi,%ebx
+ shrl $6,%ecx
+ xorl %edi,%eax
+ shrl $6,%edx
+ xorl %ecx,%eax
+ shll $7,%esi
+ xorl %edx,%ebx
+ shll $1,%edi
+ xorl %esi,%ebx
+ shrl $1,%ecx
+ xorl %edi,%eax
+ shrl $1,%edx
+ xorl %ecx,%eax
+ shll $6,%edi
+ xorl %edx,%ebx
+ xorl %edi,%eax
+ movl %eax,(%esp)
+ movl %ebx,4(%esp)
+ movl 208(%esp),%ecx
+ movl 212(%esp),%edx
+ movl %ecx,%esi
+ shrl $6,%ecx
+ movl %edx,%edi
+ shrl $6,%edx
+ movl %ecx,%eax
+ shll $3,%esi
+ movl %edx,%ebx
+ shll $3,%edi
+ xorl %esi,%eax
+ shrl $13,%ecx
+ xorl %edi,%ebx
+ shrl $13,%edx
+ xorl %ecx,%eax
+ shll $10,%esi
+ xorl %edx,%ebx
+ shll $10,%edi
+ xorl %esi,%ebx
+ shrl $10,%ecx
+ xorl %edi,%eax
+ shrl $10,%edx
+ xorl %ecx,%ebx
+ shll $13,%edi
+ xorl %edx,%eax
+ xorl %edi,%eax
+ movl 320(%esp),%ecx
+ movl 324(%esp),%edx
+ addl (%esp),%eax
+ adcl 4(%esp),%ebx
+ movl 248(%esp),%esi
+ movl 252(%esp),%edi
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ addl %esi,%eax
+ adcl %edi,%ebx
+ movl %eax,192(%esp)
+ movl %ebx,196(%esp)
+ movl 40(%esp),%ecx
+ movl 44(%esp),%edx
+ movl %ecx,%esi
+ shrl $9,%ecx
+ movl %edx,%edi
+ shrl $9,%edx
+ movl %ecx,%ebx
+ shll $14,%esi
+ movl %edx,%eax
+ shll $14,%edi
+ xorl %esi,%ebx
+ shrl $5,%ecx
+ xorl %edi,%eax
+ shrl $5,%edx
+ xorl %ecx,%eax
+ shll $4,%esi
+ xorl %edx,%ebx
+ shll $4,%edi
+ xorl %esi,%ebx
+ shrl $4,%ecx
+ xorl %edi,%eax
+ shrl $4,%edx
+ xorl %ecx,%eax
+ shll $5,%esi
+ xorl %edx,%ebx
+ shll $5,%edi
+ xorl %esi,%eax
+ xorl %edi,%ebx
+ movl 48(%esp),%ecx
+ movl 52(%esp),%edx
+ movl 56(%esp),%esi
+ movl 60(%esp),%edi
+ addl 64(%esp),%eax
+ adcl 68(%esp),%ebx
+ xorl %esi,%ecx
+ xorl %edi,%edx
+ andl 40(%esp),%ecx
+ andl 44(%esp),%edx
+ addl 192(%esp),%eax
+ adcl 196(%esp),%ebx
+ xorl %esi,%ecx
+ xorl %edi,%edx
+ movl (%ebp),%esi
+ movl 4(%ebp),%edi
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ movl 32(%esp),%ecx
+ movl 36(%esp),%edx
+ addl %esi,%eax
+ adcl %edi,%ebx
+ movl %eax,(%esp)
+ movl %ebx,4(%esp)
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ movl 8(%esp),%ecx
+ movl 12(%esp),%edx
+ movl %eax,32(%esp)
+ movl %ebx,36(%esp)
+ movl %ecx,%esi
+ shrl $2,%ecx
+ movl %edx,%edi
+ shrl $2,%edx
+ movl %ecx,%ebx
+ shll $4,%esi
+ movl %edx,%eax
+ shll $4,%edi
+ xorl %esi,%ebx
+ shrl $5,%ecx
+ xorl %edi,%eax
+ shrl $5,%edx
+ xorl %ecx,%ebx
+ shll $21,%esi
+ xorl %edx,%eax
+ shll $21,%edi
+ xorl %esi,%eax
+ shrl $21,%ecx
+ xorl %edi,%ebx
+ shrl $21,%edx
+ xorl %ecx,%eax
+ shll $5,%esi
+ xorl %edx,%ebx
+ shll $5,%edi
+ xorl %esi,%eax
+ xorl %edi,%ebx
+ movl 8(%esp),%ecx
+ movl 12(%esp),%edx
+ movl 16(%esp),%esi
+ movl 20(%esp),%edi
+ addl (%esp),%eax
+ adcl 4(%esp),%ebx
+ orl %esi,%ecx
+ orl %edi,%edx
+ andl 24(%esp),%ecx
+ andl 28(%esp),%edx
+ andl 8(%esp),%esi
+ andl 12(%esp),%edi
+ orl %esi,%ecx
+ orl %edi,%edx
+ addl %ecx,%eax
+ adcl %edx,%ebx
+ movl %eax,(%esp)
+ movl %ebx,4(%esp)
+ movb (%ebp),%dl
+ subl $8,%esp
+ leal 8(%ebp),%ebp
+ cmpb $23,%dl
+ jne .L00416_79_x86
+ movl 840(%esp),%esi
+ movl 844(%esp),%edi
+ movl (%esi),%eax
+ movl 4(%esi),%ebx
+ movl 8(%esi),%ecx
+ movl 12(%esi),%edx
+ addl 8(%esp),%eax
+ adcl 12(%esp),%ebx
+ movl %eax,(%esi)
+ movl %ebx,4(%esi)
+ addl 16(%esp),%ecx
+ adcl 20(%esp),%edx
+ movl %ecx,8(%esi)
+ movl %edx,12(%esi)
+ movl 16(%esi),%eax
+ movl 20(%esi),%ebx
+ movl 24(%esi),%ecx
+ movl 28(%esi),%edx
+ addl 24(%esp),%eax
+ adcl 28(%esp),%ebx
+ movl %eax,16(%esi)
+ movl %ebx,20(%esi)
+ addl 32(%esp),%ecx
+ adcl 36(%esp),%edx
+ movl %ecx,24(%esi)
+ movl %edx,28(%esi)
+ movl 32(%esi),%eax
+ movl 36(%esi),%ebx
+ movl 40(%esi),%ecx
+ movl 44(%esi),%edx
+ addl 40(%esp),%eax
+ adcl 44(%esp),%ebx
+ movl %eax,32(%esi)
+ movl %ebx,36(%esi)
+ addl 48(%esp),%ecx
+ adcl 52(%esp),%edx
+ movl %ecx,40(%esi)
+ movl %edx,44(%esi)
+ movl 48(%esi),%eax
+ movl 52(%esi),%ebx
+ movl 56(%esi),%ecx
+ movl 60(%esi),%edx
+ addl 56(%esp),%eax
+ adcl 60(%esp),%ebx
+ movl %eax,48(%esi)
+ movl %ebx,52(%esi)
+ addl 64(%esp),%ecx
+ adcl 68(%esp),%edx
+ movl %ecx,56(%esi)
+ movl %edx,60(%esi)
+ addl $840,%esp
+ subl $640,%ebp
+ cmpl 8(%esp),%edi
+ jb .L002loop_x86
+ movl 12(%esp),%esp
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.align 64
+.L001K512:
+.long 3609767458,1116352408
+.long 602891725,1899447441
+.long 3964484399,3049323471
+.long 2173295548,3921009573
+.long 4081628472,961987163
+.long 3053834265,1508970993
+.long 2937671579,2453635748
+.long 3664609560,2870763221
+.long 2734883394,3624381080
+.long 1164996542,310598401
+.long 1323610764,607225278
+.long 3590304994,1426881987
+.long 4068182383,1925078388
+.long 991336113,2162078206
+.long 633803317,2614888103
+.long 3479774868,3248222580
+.long 2666613458,3835390401
+.long 944711139,4022224774
+.long 2341262773,264347078
+.long 2007800933,604807628
+.long 1495990901,770255983
+.long 1856431235,1249150122
+.long 3175218132,1555081692
+.long 2198950837,1996064986
+.long 3999719339,2554220882
+.long 766784016,2821834349
+.long 2566594879,2952996808
+.long 3203337956,3210313671
+.long 1034457026,3336571891
+.long 2466948901,3584528711
+.long 3758326383,113926993
+.long 168717936,338241895
+.long 1188179964,666307205
+.long 1546045734,773529912
+.long 1522805485,1294757372
+.long 2643833823,1396182291
+.long 2343527390,1695183700
+.long 1014477480,1986661051
+.long 1206759142,2177026350
+.long 344077627,2456956037
+.long 1290863460,2730485921
+.long 3158454273,2820302411
+.long 3505952657,3259730800
+.long 106217008,3345764771
+.long 3606008344,3516065817
+.long 1432725776,3600352804
+.long 1467031594,4094571909
+.long 851169720,275423344
+.long 3100823752,430227734
+.long 1363258195,506948616
+.long 3750685593,659060556
+.long 3785050280,883997877
+.long 3318307427,958139571
+.long 3812723403,1322822218
+.long 2003034995,1537002063
+.long 3602036899,1747873779
+.long 1575990012,1955562222
+.long 1125592928,2024104815
+.long 2716904306,2227730452
+.long 442776044,2361852424
+.long 593698344,2428436474
+.long 3733110249,2756734187
+.long 2999351573,3204031479
+.long 3815920427,3329325298
+.long 3928383900,3391569614
+.long 566280711,3515267271
+.long 3454069534,3940187606
+.long 4000239992,4118630271
+.long 1914138554,116418474
+.long 2731055270,174292421
+.long 3203993006,289380356
+.long 320620315,460393269
+.long 587496836,685471733
+.long 1086792851,852142971
+.long 365543100,1017036298
+.long 2618297676,1126000580
+.long 3409855158,1288033470
+.long 4234509866,1501505948
+.long 987167468,1607167915
+.long 1246189591,1816402316
+.long 67438087,66051
+.long 202182159,134810123
+.size sha512_block_data_order,.-.L_sha512_block_data_order_begin
+.byte 83,72,65,53,49,50,32,98,108,111,99,107,32,116,114,97
+.byte 110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+.byte 67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte 112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte 62,0
diff --git a/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/x86cpuid.S b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/x86cpuid.S
new file mode 100644
index 0000000000..a1934cbc58
--- /dev/null
+++ b/CryptoPkg/Library/OpensslLib/IA32Gcc/crypto/x86cpuid.S
@@ -0,0 +1,449 @@
+# WARNING: do not edit!
+# Generated from openssl/crypto/x86cpuid.pl
+#
+# Copyright 2004-2020 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License"). You may not use
+# this file except in compliance with the License. You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+.text
+.globl OPENSSL_ia32_cpuid
+.type OPENSSL_ia32_cpuid,@function
+.align 16
+OPENSSL_ia32_cpuid:
+.L_OPENSSL_ia32_cpuid_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ xorl %edx,%edx
+ pushfl
+ popl %eax
+ movl %eax,%ecx
+ xorl $2097152,%eax
+ pushl %eax
+ popfl
+ pushfl
+ popl %eax
+ xorl %eax,%ecx
+ xorl %eax,%eax
+ movl 20(%esp),%esi
+ movl %eax,8(%esi)
+ btl $21,%ecx
+ jnc .L000nocpuid
+ .byte 0x0f,0xa2
+ movl %eax,%edi
+ xorl %eax,%eax
+ cmpl $1970169159,%ebx
+ setne %al
+ movl %eax,%ebp
+ cmpl $1231384169,%edx
+ setne %al
+ orl %eax,%ebp
+ cmpl $1818588270,%ecx
+ setne %al
+ orl %eax,%ebp
+ jz .L001intel
+ cmpl $1752462657,%ebx
+ setne %al
+ movl %eax,%esi
+ cmpl $1769238117,%edx
+ setne %al
+ orl %eax,%esi
+ cmpl $1145913699,%ecx
+ setne %al
+ orl %eax,%esi
+ jnz .L001intel
+ movl $2147483648,%eax
+ .byte 0x0f,0xa2
+ cmpl $2147483649,%eax
+ jb .L001intel
+ movl %eax,%esi
+ movl $2147483649,%eax
+ .byte 0x0f,0xa2
+ orl %ecx,%ebp
+ andl $2049,%ebp
+ cmpl $2147483656,%esi
+ jb .L001intel
+ movl $2147483656,%eax
+ .byte 0x0f,0xa2
+ movzbl %cl,%esi
+ incl %esi
+ movl $1,%eax
+ xorl %ecx,%ecx
+ .byte 0x0f,0xa2
+ btl $28,%edx
+ jnc .L002generic
+ shrl $16,%ebx
+ andl $255,%ebx
+ cmpl %esi,%ebx
+ ja .L002generic
+ andl $4026531839,%edx
+ jmp .L002generic
+.L001intel:
+ cmpl $4,%edi
+ movl $-1,%esi
+ jb .L003nocacheinfo
+ movl $4,%eax
+ movl $0,%ecx
+ .byte 0x0f,0xa2
+ movl %eax,%esi
+ shrl $14,%esi
+ andl $4095,%esi
+.L003nocacheinfo:
+ movl $1,%eax
+ xorl %ecx,%ecx
+ .byte 0x0f,0xa2
+ andl $3220176895,%edx
+ cmpl $0,%ebp
+ jne .L004notintel
+ orl $1073741824,%edx
+ andb $15,%ah
+ cmpb $15,%ah
+ jne .L004notintel
+ orl $1048576,%edx
+.L004notintel:
+ btl $28,%edx
+ jnc .L002generic
+ andl $4026531839,%edx
+ cmpl $0,%esi
+ je .L002generic
+ orl $268435456,%edx
+ shrl $16,%ebx
+ cmpb $1,%bl
+ ja .L002generic
+ andl $4026531839,%edx
+.L002generic:
+ andl $2048,%ebp
+ andl $4294965247,%ecx
+ movl %edx,%esi
+ orl %ecx,%ebp
+ cmpl $7,%edi
+ movl 20(%esp),%edi
+ jb .L005no_extended_info
+ movl $7,%eax
+ xorl %ecx,%ecx
+ .byte 0x0f,0xa2
+ movl %ebx,8(%edi)
+.L005no_extended_info:
+ btl $27,%ebp
+ jnc .L006clear_avx
+ xorl %ecx,%ecx
+.byte 15,1,208
+ andl $6,%eax
+ cmpl $6,%eax
+ je .L007done
+ cmpl $2,%eax
+ je .L006clear_avx
+.L008clear_xmm:
+ andl $4261412861,%ebp
+ andl $4278190079,%esi
+.L006clear_avx:
+ andl $4026525695,%ebp
+ andl $4294967263,8(%edi)
+.L007done:
+ movl %esi,%eax
+ movl %ebp,%edx
+.L000nocpuid:
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size OPENSSL_ia32_cpuid,.-.L_OPENSSL_ia32_cpuid_begin
+.globl OPENSSL_rdtsc
+.type OPENSSL_rdtsc,@function
+.align 16
+OPENSSL_rdtsc:
+.L_OPENSSL_rdtsc_begin:
+ xorl %eax,%eax
+ xorl %edx,%edx
+ leal OPENSSL_ia32cap_P,%ecx
+ btl $4,(%ecx)
+ jnc .L009notsc
+ .byte 0x0f,0x31
+.L009notsc:
+ ret
+.size OPENSSL_rdtsc,.-.L_OPENSSL_rdtsc_begin
+.globl OPENSSL_instrument_halt
+.type OPENSSL_instrument_halt,@function
+.align 16
+OPENSSL_instrument_halt:
+.L_OPENSSL_instrument_halt_begin:
+ leal OPENSSL_ia32cap_P,%ecx
+ btl $4,(%ecx)
+ jnc .L010nohalt
+.long 2421723150
+ andl $3,%eax
+ jnz .L010nohalt
+ pushfl
+ popl %eax
+ btl $9,%eax
+ jnc .L010nohalt
+ .byte 0x0f,0x31
+ pushl %edx
+ pushl %eax
+ hlt
+ .byte 0x0f,0x31
+ subl (%esp),%eax
+ sbbl 4(%esp),%edx
+ addl $8,%esp
+ ret
+.L010nohalt:
+ xorl %eax,%eax
+ xorl %edx,%edx
+ ret
+.size OPENSSL_instrument_halt,.-.L_OPENSSL_instrument_halt_begin
+.globl OPENSSL_far_spin
+.type OPENSSL_far_spin,@function
+.align 16
+OPENSSL_far_spin:
+.L_OPENSSL_far_spin_begin:
+ pushfl
+ popl %eax
+ btl $9,%eax
+ jnc .L011nospin
+ movl 4(%esp),%eax
+ movl 8(%esp),%ecx
+.long 2430111262
+ xorl %eax,%eax
+ movl (%ecx),%edx
+ jmp .L012spin
+.align 16
+.L012spin:
+ incl %eax
+ cmpl (%ecx),%edx
+ je .L012spin
+.long 529567888
+ ret
+.L011nospin:
+ xorl %eax,%eax
+ xorl %edx,%edx
+ ret
+.size OPENSSL_far_spin,.-.L_OPENSSL_far_spin_begin
+.globl OPENSSL_wipe_cpu
+.type OPENSSL_wipe_cpu,@function
+.align 16
+OPENSSL_wipe_cpu:
+.L_OPENSSL_wipe_cpu_begin:
+ xorl %eax,%eax
+ xorl %edx,%edx
+ leal OPENSSL_ia32cap_P,%ecx
+ movl (%ecx),%ecx
+ btl $1,(%ecx)
+ jnc .L013no_x87
+.long 4007259865,4007259865,4007259865,4007259865,2430851995
+.L013no_x87:
+ leal 4(%esp),%eax
+ ret
+.size OPENSSL_wipe_cpu,.-.L_OPENSSL_wipe_cpu_begin
+.globl OPENSSL_atomic_add
+.type OPENSSL_atomic_add,@function
+.align 16
+OPENSSL_atomic_add:
+.L_OPENSSL_atomic_add_begin:
+ movl 4(%esp),%edx
+ movl 8(%esp),%ecx
+ pushl %ebx
+ nop
+ movl (%edx),%eax
+.L014spin:
+ leal (%eax,%ecx,1),%ebx
+ nop
+.long 447811568
+ jne .L014spin
+ movl %ebx,%eax
+ popl %ebx
+ ret
+.size OPENSSL_atomic_add,.-.L_OPENSSL_atomic_add_begin
+.globl OPENSSL_cleanse
+.type OPENSSL_cleanse,@function
+.align 16
+OPENSSL_cleanse:
+.L_OPENSSL_cleanse_begin:
+ movl 4(%esp),%edx
+ movl 8(%esp),%ecx
+ xorl %eax,%eax
+ cmpl $7,%ecx
+ jae .L015lot
+ cmpl $0,%ecx
+ je .L016ret
+.L017little:
+ movb %al,(%edx)
+ subl $1,%ecx
+ leal 1(%edx),%edx
+ jnz .L017little
+.L016ret:
+ ret
+.align 16
+.L015lot:
+ testl $3,%edx
+ jz .L018aligned
+ movb %al,(%edx)
+ leal -1(%ecx),%ecx
+ leal 1(%edx),%edx
+ jmp .L015lot
+.L018aligned:
+ movl %eax,(%edx)
+ leal -4(%ecx),%ecx
+ testl $-4,%ecx
+ leal 4(%edx),%edx
+ jnz .L018aligned
+ cmpl $0,%ecx
+ jne .L017little
+ ret
+.size OPENSSL_cleanse,.-.L_OPENSSL_cleanse_begin
+.globl CRYPTO_memcmp
+.type CRYPTO_memcmp,@function
+.align 16
+CRYPTO_memcmp:
+.L_CRYPTO_memcmp_begin:
+ pushl %esi
+ pushl %edi
+ movl 12(%esp),%esi
+ movl 16(%esp),%edi
+ movl 20(%esp),%ecx
+ xorl %eax,%eax
+ xorl %edx,%edx
+ cmpl $0,%ecx
+ je .L019no_data
+.L020loop:
+ movb (%esi),%dl
+ leal 1(%esi),%esi
+ xorb (%edi),%dl
+ leal 1(%edi),%edi
+ orb %dl,%al
+ decl %ecx
+ jnz .L020loop
+ negl %eax
+ shrl $31,%eax
+.L019no_data:
+ popl %edi
+ popl %esi
+ ret
+.size CRYPTO_memcmp,.-.L_CRYPTO_memcmp_begin
+.globl OPENSSL_instrument_bus
+.type OPENSSL_instrument_bus,@function
+.align 16
+OPENSSL_instrument_bus:
+.L_OPENSSL_instrument_bus_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl $0,%eax
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size OPENSSL_instrument_bus,.-.L_OPENSSL_instrument_bus_begin
+.globl OPENSSL_instrument_bus2
+.type OPENSSL_instrument_bus2,@function
+.align 16
+OPENSSL_instrument_bus2:
+.L_OPENSSL_instrument_bus2_begin:
+ pushl %ebp
+ pushl %ebx
+ pushl %esi
+ pushl %edi
+ movl $0,%eax
+ popl %edi
+ popl %esi
+ popl %ebx
+ popl %ebp
+ ret
+.size OPENSSL_instrument_bus2,.-.L_OPENSSL_instrument_bus2_begin
+.globl OPENSSL_ia32_rdrand_bytes
+.type OPENSSL_ia32_rdrand_bytes,@function
+.align 16
+OPENSSL_ia32_rdrand_bytes:
+.L_OPENSSL_ia32_rdrand_bytes_begin:
+ pushl %edi
+ pushl %ebx
+ xorl %eax,%eax
+ movl 12(%esp),%edi
+ movl 16(%esp),%ebx
+ cmpl $0,%ebx
+ je .L021done
+ movl $8,%ecx
+.L022loop:
+.byte 15,199,242
+ jc .L023break
+ loop .L022loop
+ jmp .L021done
+.align 16
+.L023break:
+ cmpl $4,%ebx
+ jb .L024tail
+ movl %edx,(%edi)
+ leal 4(%edi),%edi
+ addl $4,%eax
+ subl $4,%ebx
+ jz .L021done
+ movl $8,%ecx
+ jmp .L022loop
+.align 16
+.L024tail:
+ movb %dl,(%edi)
+ leal 1(%edi),%edi
+ incl %eax
+ shrl $8,%edx
+ decl %ebx
+ jnz .L024tail
+.L021done:
+ xorl %edx,%edx
+ popl %ebx
+ popl %edi
+ ret
+.size OPENSSL_ia32_rdrand_bytes,.-.L_OPENSSL_ia32_rdrand_bytes_begin
+.globl OPENSSL_ia32_rdseed_bytes
+.type OPENSSL_ia32_rdseed_bytes,@function
+.align 16
+OPENSSL_ia32_rdseed_bytes:
+.L_OPENSSL_ia32_rdseed_bytes_begin:
+ pushl %edi
+ pushl %ebx
+ xorl %eax,%eax
+ movl 12(%esp),%edi
+ movl 16(%esp),%ebx
+ cmpl $0,%ebx
+ je .L025done
+ movl $8,%ecx
+.L026loop:
+.byte 15,199,250
+ jc .L027break
+ loop .L026loop
+ jmp .L025done
+.align 16
+.L027break:
+ cmpl $4,%ebx
+ jb .L028tail
+ movl %edx,(%edi)
+ leal 4(%edi),%edi
+ addl $4,%eax
+ subl $4,%ebx
+ jz .L025done
+ movl $8,%ecx
+ jmp .L026loop
+.align 16
+.L028tail:
+ movb %dl,(%edi)
+ leal 1(%edi),%edi
+ incl %eax
+ shrl $8,%edx
+ decl %ebx
+ jnz .L028tail
+.L025done:
+ xorl %edx,%edx
+ popl %ebx
+ popl %edi
+ ret
+.size OPENSSL_ia32_rdseed_bytes,.-.L_OPENSSL_ia32_rdseed_bytes_begin
+.hidden OPENSSL_cpuid_setup
+.hidden OPENSSL_ia32cap_P
+.comm OPENSSL_ia32cap_P,16,4
+.section .init
+ call OPENSSL_cpuid_setup
--
2.29.2.windows.2
^ permalink raw reply related [flat|nested] 8+ messages in thread