From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received-SPF: Pass (sender SPF authorized) identity=mailfrom; client-ip=17.151.62.66; helo=nwk-aaemail-lapp01.apple.com; envelope-from=afish@apple.com; receiver=edk2-devel@lists.01.org Received: from nwk-aaemail-lapp01.apple.com (nwk-aaemail-lapp01.apple.com [17.151.62.66]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ml01.01.org (Postfix) with ESMTPS id 2CCC62194D3B8 for ; Wed, 6 Mar 2019 21:56:30 -0800 (PST) Received: from pps.filterd (nwk-aaemail-lapp01.apple.com [127.0.0.1]) by nwk-aaemail-lapp01.apple.com (8.16.0.27/8.16.0.27) with SMTP id x275pv5b048871; Wed, 6 Mar 2019 21:56:29 -0800 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=apple.com; h=mime-version : content-type : sender : from : message-id : subject : date : in-reply-to : cc : to : references; s=20180706; bh=pYngo+kKs6aQUxg0Xt0elpWFHSdY4bPRlq4O4qFym2E=; b=KgxW0CHtdRZ1raSnmdp5ExdBU2kHcLXCm+uMscZiLBXjFSjYFg1TPxMABplW4V3QnmMT KkvKWV8QFoGtph91RMPzWUSNSW5VGn8hckxA3q2q57qDHxsUpS0SG6Gi/RMtq9TGewN8 rD1fvLwRYptbhslTXGNYU7GJV4xo2j+bVcERTLGSsvycWUBpB0nS6KzWiztjxoksZsYS TJRO3RmUQGMa+Pql+mut2ZaqXLY344exurPful3pPjoQfmSxwozTxmKDZVKRcVjIJeoS lxMV6YIRvgjk8hK+9l2kZPs3iZnlqPuz1GU98/ztH9gG7xexkMM2nnRVtvr8vddTV762 JQ== Received: from mr2-mtap-s01.rno.apple.com (mr2-mtap-s01.rno.apple.com [17.179.226.133]) by nwk-aaemail-lapp01.apple.com with ESMTP id 2qysn91dvg-9 (version=TLSv1.2 cipher=ECDHE-RSA-AES128-GCM-SHA256 bits=128 verify=NO); Wed, 06 Mar 2019 21:56:29 -0800 MIME-version: 1.0 Received: from ma1-mmpp-sz08.apple.com (ma1-mmpp-sz08.apple.com [17.171.128.176]) by mr2-mtap-s01.rno.apple.com (Oracle Communications Messaging Server 8.0.2.3.20181024 64bit (built Oct 24 2018)) with ESMTPS id <0PNZ00GIAF62ZK70@mr2-mtap-s01.rno.apple.com>; Wed, 06 Mar 2019 21:56:26 -0800 (PST) Received: from process_milters-daemon.ma1-mmpp-sz08.apple.com by ma1-mmpp-sz08.apple.com (Oracle Communications Messaging Server 8.0.2.3.20181024 64bit (built Oct 24 2018)) id <0PNZ00L00D5SU500@ma1-mmpp-sz08.apple.com>; Wed, 06 Mar 2019 21:56:26 -0800 (PST) X-Va-A: X-Va-T-CD: 81ca60fce39c2560b6c4a7e5841f9b8f X-Va-E-CD: 7d4c618f85f76966dfbaabb394b9e4f0 X-Va-R-CD: a66fd94a21bd5e42e01791934a036d45 X-Va-CD: 0 X-Va-ID: 43783121-ff2b-4149-892f-ac5bfc47879b X-V-A: X-V-T-CD: 81ca60fce39c2560b6c4a7e5841f9b8f X-V-E-CD: 7d4c618f85f76966dfbaabb394b9e4f0 X-V-R-CD: a66fd94a21bd5e42e01791934a036d45 X-V-CD: 0 X-V-ID: 37eb2cea-9e7d-4120-baa1-213b8bf27b45 X-Proofpoint-Virus-Version: vendor=fsecure engine=2.50.10434:,, definitions=2019-03-07_03:,, signatures=0 Received: from [17.234.172.123] (unknown [17.234.172.123]) by ma1-mmpp-sz08.apple.com (Oracle Communications Messaging Server 8.0.2.3.20181024 64bit (built Oct 24 2018)) with ESMTPSA id <0PNZ00IDJF5YGL10@ma1-mmpp-sz08.apple.com>; Wed, 06 Mar 2019 21:56:25 -0800 (PST) Sender: afish@apple.com From: Andrew Fish Message-id: Date: Wed, 06 Mar 2019 21:56:20 -0800 In-reply-to: <74D8A39837DF1E4DA445A8C0B3885C503F556F4E@shsmsx102.ccr.corp.intel.com> Cc: Mike Kinney , "Gao, Liming" , edk2-devel-01 To: "Yao, Jiewen" References: <20190305014059.17988-1-shenglei.zhang@intel.com> <20190305014059.17988-4-shenglei.zhang@intel.com> <4A89E2EF3DFEDB4C8BFDE51014F606A14E3FBBC2@SHSMSX104.ccr.corp.intel.com> <4A89E2EF3DFEDB4C8BFDE51014F606A14E3FCC47@SHSMSX104.ccr.corp.intel.com> <47E8FE4A-4BDC-48A3-B7E6-02F1682C0990@apple.com> <74D8A39837DF1E4DA445A8C0B3885C503F556F4E@shsmsx102.ccr.corp.intel.com> X-Mailer: Apple Mail (2.3445.6.18) X-Proofpoint-Virus-Version: vendor=fsecure engine=2.50.10434:, , definitions=2019-03-07_03:, , signatures=0 X-Content-Filtered-By: Mailman/MimeDel 2.1.29 Subject: Re: [PATCH 3/3] MdePkg/BaseSynchronizationLib: Remove inline X86 assembly code X-BeenThere: edk2-devel@lists.01.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: EDK II Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 07 Mar 2019 05:56:30 -0000 Content-Type: text/plain; CHARSET=US-ASCII Content-Transfer-Encoding: 7BIT > On Mar 6, 2019, at 9:06 PM, Yao, Jiewen wrote: > > HI Mike and Andrew > The problem is maintainability. > > If we have multiple copy of implementation, a developer need verify multiple copy of implementation, if we make update. Before that, a developer has to be aware that there is multiple copy of implementation. - That increases the complexity. > > If we have everything, there MAY be 5 copy - ASM, NASM, S, GCC inline, MS inline, theoretically. > Now, we remove ASM. It is good first step. > But we may still have 4 copies. I suggest we consider do more. > Jiewen, I think you are trying do the right thing, but are optimize the wrong thing. Most of the GCC/Clang inline assembler code is in Gccinline.c and since that code is mostly just abstracting an x86 instruction and the functions are very simple and thus it is NOT code that needs ongoing maintenance. Lets look at the history: https://github.com/tianocore/edk2/commits/master/MdePkg/Library/BaseLib/X64/GccInline.c The last logic fix was Jun 1, 2010 https://github.com/tianocore/edk2/commits/master/MdePkg/Library/BaseLib/Ia32/GccInline.c Ok so Mike had to make a fix in this file in 2015 to make something optional, due to an embedded CPU defeating an instruction. But prior to that it was 2010. The set of things that are C inline assembler we have should be static and not require much maintenance. More complex code should be written in C and use the C inline assembler functions we already have. If any more complex assembly code is required we should just write it in NASM. So I think it is fine to restrict new C inline assembler, or at least have a very high bar to add anything new. Any new inline assembler should also be simple and just be a function abstracting a CPU op-code that is not available to C. This is how we prevent the maintenance issues you are worrying about. I gave an example in this mail thread on how a Breakpoint goes from being 1 byte to 11 bytes if you remove the C inline assembler. For clang with LTO enabled a CpuBreakpoint will always get inlined into the callers code and it will only take the one byte for int 3 instruction. If that code moves to NASM then it get replaces with a 5 byte call instruction and an actual C ABI function for the breakpoint. VOID EFIAPI CpuBreakpoint ( VOID ) { __asm__ __volatile__ ("int $3"); } Today with clang LTO turned on calling CpuBreakpoint() looks like this from the callers point of view. a.out[0x1fa5] <+6>: cc int3 But if we move that to NASM: a.out[0x1fa6] <+7>: e8 07 00 00 00 calll 0x1fb2 ; CpuBreakpoint plus: a.out`CpuBreakpoint: a.out[0x1f99] <+0>: 55 pushl %ebp a.out[0x1f9a] <+1>: 89 e5 movl %esp, %ebp a.out[0x1f9c] <+3>: cc int3 a.out[0x1f9d] <+4>: 5d popl %ebp a.out[0x1f9e] <+5>: c3 retl For any compiler that emits the frame pointer if you move the INT 3 to assembly you need the frame pointer or the Exception Lib is not going to be able to print out the stack backtrace of the code when you hit a breakpoint. So this is how I get to 11 vs. 1 bytes. Thanks, Andrew Fish PS for clang LTO the compiler compiles to bitcode and that is an abstracted assembly language. At link time all that code gets optimized to together and then passed through the CPU specific code gen. For C inline assembler the assembly instructions end up in the bitcode with lots of information about the constraints. That is why these GccInline.c functions almost always get inlined with clang and LTO. The CpuBreakpoint() function looks like this in bitcode, but the function call gets optimized away by LTO in the code gen. ; Function Attrs: noinline nounwind optnone ssp uwtable define void @CpuBreakpoint() #0 { call void asm sideeffect "int $$3", "~{dirflag},~{fpsr},~{flags}"() #2, !srcloc !3 ret void } Even something more complex like AsmReadMsr64() can get inlined, but it contains a lot more info about the constraints: ; Function Attrs: noinline nounwind optnone ssp uwtable define i64 @AsmReadMsr64(i32) #0 { %2 = alloca i32, align 4 %3 = alloca i64, align 8 store i32 %0, i32* %2, align 4 %4 = load i32, i32* %2, align 4 %5 = call i64 asm sideeffect "rdmsr", "=A,{cx},~{dirflag},~{fpsr},~{flags}"(i32 %4) #2, !srcloc !4 store i64 %5, i64* %3, align 8 %6 = load i64, i64* %3, align 8 ret i64 %6 } The alloca, store, load, etc are the same bitcode instructions you would see with C code. > A recently case that SetJump/LongJump, I updated the NASM file for both IA32 and X64. > > Later, to my surprise, only X64 version NASM works, and IA32 version NASM does not work. > Then I notice that there is a different copy if IA32 Jump.c MS inline - I also need update. That is frustrated. > > I think there should be a balance between optimization and code readability/maintainability. > > Do we have data on what size benefit we can get with these inline function, from whole image level? > We can do evaluation at function level, case by case. > If we see a huge size benefit, I agree to keep this function. > If we see no size benefit, I suggest we do the cleanup for this function. > > > Thank you > Yao Jiewen > > > >> -----Original Message----- >> From: edk2-devel [mailto:edk2-devel-bounces@lists.01.org] On Behalf Of >> Andrew Fish via edk2-devel >> Sent: Wednesday, March 6, 2019 5:31 PM >> To: Kinney, Michael D >> Cc: edk2-devel-01 ; Gao, Liming >> >> Subject: Re: [edk2] [PATCH 3/3] MdePkg/BaseSynchronizationLib: Remove >> inline X86 assembly code >> >> >> >>> On Mar 6, 2019, at 4:41 PM, Kinney, Michael D >> wrote: >>> >>> Liming, >>> >>> That does not work either because inline assembly uses compiler specific >> syntax. >>> >>> Please update with the specific list of functions that you think the inline >> should be removed to improve maintenance with no impacts in size/perf. >>> >> >> Mike, >> >> It is easy to find the gcc/clang inline assembler, just `git grep "__asm__ >> __volatile__"` >> >> The main files are: >> https://github.com/tianocore/edk2/blob/master/MdePkg/Library/BaseIoLib >> Intrinsic/IoLibGcc.c >> https://github.com/tianocore/edk2/blob/master/MdePkg/Library/BaseLib/X >> 64/GccInline.c >> https://github.com/tianocore/edk2/blob/master/MdePkg/Library/BaseLib/I >> a32/GccInline.c >> https://github.com/tianocore/edk2/blob/master/MdePkg/Library/BaseSync >> hronizationLib/Ia32/GccInline.c >> https://github.com/tianocore/edk2/blob/master/MdePkg/Library/BaseSync >> hronizationLib/X64/GccInline.c >> >> This is really just compiler optimizer control. >> Library/BaseSynchronizationLib/SynchronizationGcc.c:21:#define >> _ReadWriteBarrier() do { __asm__ __volatile__ ("": : : "memory"); } while(0) >> >> Looks like this is working around the alignment ASSERT in BaseIoLibIntrinsic. >> OvmfPkg/QemuVideoDxe/UnalignedIoGcc.c:43: __asm__ __volatile__ >> ( "outl %0, %1" : : "a" (Value), "d" ((UINT16)Port) ); >> OvmfPkg/QemuVideoDxe/UnalignedIoGcc.c:67: __asm__ __volatile__ >> ( "inl %1, %0" : "=a" (Data) : "d" ((UINT16)Port) ); >> >> It seems like we have a reasonable set and should not use the inline >> assembly for any more complex code. >> >> Thanks, >> >> Andrew Fish >> >>> The issue you pointed to was around SetJump()/LongJump(). Can we limit >> this BZ to only those 2 functions? >>> >>> Mike >>> <> >>> From: Gao, Liming >>> Sent: Wednesday, March 6, 2019 4:28 PM >>> To: afish@apple.com >>> Cc: Zhang, Shenglei ; edk2-devel-01 >> ; Kinney, Michael D >> >>> Subject: RE: [edk2] [PATCH 3/3] MdePkg/BaseSynchronizationLib: Remove >> inline X86 assembly code >>> >>> Andrew: >>> I want to keep only one implementation. If inline assembly c source is >> preferred, I suggest to remove its nasm implementation. >>> >>> Thanks >>> Liming >>> <>From: afish@apple.com >> [mailto:afish@apple.com ] >>> Sent: Tuesday, March 5, 2019 2:44 PM >>> To: Gao, Liming > >>> Cc: Zhang, Shenglei > >; edk2-devel-01 >> >; Kinney, >> Michael D > > >>> Subject: Re: [edk2] [PATCH 3/3] MdePkg/BaseSynchronizationLib: Remove >> inline X86 assembly code >>> >>> >>> >>> >>> On Mar 5, 2019, at 1:32 PM, Gao, Liming > > wrote: >>> >>> Andrew: >>> BZ 1163 is to remove inline X86 assembly code in C source file. But, this >> patch is wrong. I have gave my comments to update this patch. >>> >>> >>> Why do we want to remove inline X86 assembly. As I mentioned it will lead >> to larger binaries, slower binaries, and less optimized code. >>> >>> For example take this simple inline assembly function: >>> >>> VOID >>> EFIAPI >>> CpuBreakpoint ( >>> VOID >>> ) >>> { >>> __asm__ __volatile__ ("int $3"); >>> } >>> >>> >>> Today with clang LTO turned on calling CpuBreakpoint() looks like this from >> the callers point of view. >>> >>> a.out[0x1fa5] <+6>: cc int3 >>> >>> >>> But if we move that to NASM: >>> >>> >>> a.out[0x1fa6] <+7>: e8 07 00 00 00 calll >> 0x1fb2 ; CpuBreakpoint >>> >>> >>> plus: >>> a.out`CpuBreakpoint: >>> a.out[0x1fb2] <+0>: cc int3 >>> a.out[0x1fb3] <+1>: c3 retl >>> >>> >>> And there is also an extra push and pop on the stack. The other issue is the >> call to the function that is unknown to the compiler will act like a >> _ReadWriteBarrier (Yikes I see _ReadWriteBarrier is deprecated in VC++ >> 2017). Is the depreciation of some of these intrinsics in VC++ driving the >> removal of inline assembly? For GCC inline assembly works great for local >> compile, and for clang LTO it works in entire program optimization. >>> >>> The LTO bitcode includes inline assembly and constraints so that the >> optimizer knows what to do so it can get optimized just like the abstract >> bitcode during the Link Time Optimization. >>> >>> This is CpuBreakpoint(): >>> ; Function Attrs: noinline nounwind optnone ssp uwtable >>> define void @CpuBreakpoint() #0 { >>> call void asm sideeffect "int $$3", "~{dirflag},~{fpsr},~{flags}"() >> #2, !srcloc !3 >>> ret void >>> } >>> >>> >>> This is AsmReadMsr64(): >>> ; Function Attrs: noinline nounwind optnone ssp uwtable >>> define i64 @AsmReadMsr64(i32) #0 { >>> %2 = alloca i32, align 4 >>> %3 = alloca i64, align 8 >>> store i32 %0, i32* %2, align 4 >>> %4 = load i32, i32* %2, align 4 >>> %5 = call i64 asm sideeffect "rdmsr", >> "=A,{cx},~{dirflag},~{fpsr},~{flags}"(i32 %4) #2, !srcloc !4 >>> store i64 %5, i64* %3, align 8 >>> %6 = load i64, i64* %3, align 8 >>> ret i64 %6 >>> } >>> >>> >>> >>> I agree it make sense to remove .S and .asm files and only have .nasm files. >>> >>> Thanks, >>> >>> Andrew Fish >>> >>> PS For the Xcode clang since it emits frame pointers by default we need to >> add the extra 4 bytes to save the assembly functions so the stack can get >> unwound. >>> >>> a.out`CpuBreakpoint: >>> a.out[0x1f99] <+0>: 55 pushl %ebp >>> a.out[0x1f9a] <+1>: 89 e5 movl %esp, %ebp >>> a.out[0x1f9c] <+3>: cc int3 >>> a.out[0x1f9d] <+4>: 5d popl %ebp >>> a.out[0x1f9e] <+5>: c3 retl >>> >>> >>> So breakpoint goes from 1 byte to 11 bytes if we get rid of the intrinsics. >>> >>> >>> The change is to reduce the duplicated implementation. It will be good >> on the code maintain. Recently, one patch has to update .c and .nasm both. >> Please see >> https://lists.01.org/pipermail/edk2-devel/2019-February/037165.html >> . >> Beside this change, I propose to remove all GAS assembly code for IA32 and >> X64 arch. After those change, the patch owner will collect the impact of the >> image size. >>> >>> Thanks >>> Liming >> >> _______________________________________________ >> edk2-devel mailing list >> edk2-devel@lists.01.org >> https://lists.01.org/mailman/listinfo/edk2-devel