[PATCH v4 3/4] MdeModulePkg/EbxDxe AARCH64: use tail call for EBC to native thunk

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

Instead of pessimistically copying at least 64 bytes from the VM stack
to the native stack, and popping off the register arguments again
before doing the native call, try to avoid touching the stack completely
if the VM stack frame is <= 64 bytes. Also, if the stack frame does exceed
64 bytes, there is no need to copy the first 64 bytes, since we are passing
those in registers anyway.

Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
---
 MdeModulePkg/Universal/EbcDxe/AArch64/EbcLowLevel.S | 85 +++++++++++++++-----
 1 file changed, 65 insertions(+), 20 deletions(-)

diff --git a/MdeModulePkg/Universal/EbcDxe/AArch64/EbcLowLevel.S b/MdeModulePkg/Universal/EbcDxe/AArch64/EbcLowLevel.S
index b4b8531f1a01..34794c06a644 100644
--- a/MdeModulePkg/Universal/EbcDxe/AArch64/EbcLowLevel.S
+++ b/MdeModulePkg/Universal/EbcDxe/AArch64/EbcLowLevel.S
@@ -35,30 +35,75 @@ ASM_GLOBAL ASM_PFX(mEbcInstructionBufferTemplate)
 //****************************************************************************
 // UINTN EbcLLCALLEXNative(UINTN FuncAddr, UINTN NewStackPointer, VOID *FramePtr)
 ASM_PFX(EbcLLCALLEXNative):
-      stp  x19, x20, [sp, #-16]!
-      stp  x29, x30, [sp, #-16]!
+    mov     x8, x0                 // Preserve x0
+    mov     x9, x1                 // Preserve x1
 
-      mov  x19, x0
-      mov  x20, sp
-      sub  x2, x2, x1   // Length = NewStackPointer-FramePtr
-      sub  sp, sp, x2
-      sub  sp, sp, #64  // Make sure there is room for at least 8 args in the new stack
-      mov  x0, sp
-
-      bl   CopyMem      // Sp, NewStackPointer, Length
-
-      ldp  x0, x1, [sp], #16
-      ldp  x2, x3, [sp], #16
-      ldp  x4, x5, [sp], #16
-      ldp  x6, x7, [sp], #16
+    //
+    // If the EBC stack frame is smaller than or equal to 64 bytes, we know there
+    // are no stacked arguments #9 and beyond that we need to copy to the native
+    // stack. In this case, we can perform a tail call which is much more
+    // efficient, since there is no need to touch the native stack at all.
+    //
+    sub     x3, x2, x1              // Length = NewStackPointer - FramePtr
+    cmp     x3, #64
+    b.gt    1f
 
-      blr  x19
+    //
+    // While probably harmless in practice, we should not access the VM stack
+    // outside of the interval [NewStackPointer, FramePtr), which means we
+    // should not blindly fill all 8 argument registers with VM stack data.
+    // So instead, calculate how many argument registers we can fill based on
+    // the size of the VM stack frame, and skip the remaining ones.
+    //
+    adr     x0, 0f                  // Take address of 'br' instruction below
+    bic     x3, x3, #7              // Ensure correct alignment
+    sub     x0, x0, x3, lsr #1      // Subtract 4 bytes for each arg to unstack
+    br      x0                      // Skip remaining argument registers
+
+    ldr     x7, [x9, #56]           // Call with 8 arguments
+    ldr     x6, [x9, #48]           //  |
+    ldr     x5, [x9, #40]           //  |
+    ldr     x4, [x9, #32]           //  |
+    ldr     x3, [x9, #24]           //  |
+    ldr     x2, [x9, #16]           //  |
+    ldr     x1, [x9, #8]            //  V
+    ldr     x0, [x9]                // Call with 1 argument
+
+0:  br      x8                      // Call with no arguments
 
-      mov  sp,  x20
-      ldp  x29, x30, [sp], #16
-      ldp  x19, x20, [sp], #16
+    //
+    // More than 64 bytes: we need to build the full native stack frame and copy
+    // the part of the VM stack exceeding 64 bytes (which may contain stacked
+    // arguments) to the native stack
+    //
+1:  stp     x29, x30, [sp, #-16]!
+    mov     x29, sp
 
-      ret
+    //
+    // Ensure that the stack pointer remains 16 byte aligned,
+    // even if the size of the VM stack frame is not a multiple of 16
+    //
+    add     x1, x1, #64             // Skip over [potential] reg params
+    tbz     x3, #3, 2f              // Multiple of 16?
+    ldr     x4, [x2, #-8]!          // No? Then push one word
+    str     x4, [sp, #-16]!         // ... but use two slots
+    b       3f
+
+2:  ldp     x4, x5, [x2, #-16]!
+    stp     x4, x5, [sp, #-16]!
+3:  cmp     x2, x1
+    b.gt    2b
+
+    ldp     x0, x1, [x9]
+    ldp     x2, x3, [x9, #16]
+    ldp     x4, x5, [x9, #32]
+    ldp     x6, x7, [x9, #48]
+
+    blr     x8
+
+    mov     sp, x29
+    ldp     x29, x30, [sp], #16
+    ret
 
 //****************************************************************************
 // EbcLLEbcInterpret
-- 
2.7.4