toggle quoted messageShow quoted text
This issue is introduced by the commit 5f6ecaa398ba902f724be504663a0156c9e24570 OvmfPkg/VirtioGpuDxe: use GopQueryMode in GopSetMode from you. Can you provide the hot fix for this stable tag 202205? This stable tag will be released on tomorrow.
发件人: email@example.com <firstname.lastname@example.org> 代表 Yao, Jiewen
发送时间: 2022年5月26日 8:06
收件人: Pedro Falcato <pedro.falcato@...>; edk2-devel-groups-io <email@example.com>
抄送: Ard Biesheuvel <ardb@...>; Rebecca Cran <quic_rcran@...>; Ard Biesheuvel <ardb+tianocore@...>; Justen, Jordan L <jordan.l.justen@...>; Gerd Hoffmann <kraxel@...>
主题: Re: [edk2-devel] OvmfPkgX64 doesn't build with CLANG38 (clang 14.0.3) NOOPT - undefined reference to `memcpy'
Sounds good. :-)
From: Pedro Falcato <pedro.falcato@...>
Sent: Thursday, May 26, 2022 7:54 AM
To: edk2-devel-groups-io <firstname.lastname@example.org>; Yao, Jiewen <jiewen.yao@...>
Cc: Ard Biesheuvel <ardb@...>; Rebecca Cran <quic_rcran@...>; Ard Biesheuvel <ardb+tianocore@...>; Justen, Jordan L <jordan.l.justen@...>; Gerd Hoffmann <kraxel@...>
Subject: Re: [edk2-devel] OvmfPkgX64 doesn't build with CLANG38 (clang 14.0.3) NOOPT - undefined reference to `memcpy'
I'm definitely talking about the long term. Fixing this by just replacing it with CopyMem is definitely the right call for the stable tag.
OK. Just want to clarify:
- Are you discussing the solution now to resolve this issue, maybe in this stable tag?
- Or are you discussing the long term solution?
For me, I am focusing on #1 how to resolve this issue in this stable tag.
I am not talking long term solution #2 here. I think that need wider discussion in EDKII.
Right. But as I said, GCC and Clang depend on 4 standard C library functions. This behavior is documented, and has been stable for probably the past decade or more. These are not intrinsics, but bog standard functions you can totally redirect to the existing UEFI ones. I talked with the LLVM people on IRC and they confirmed that that was essentially the reason why it used a memcpy call.
It's, in my opinion, OK to fix this particular issue if for example Visual Studio does something fancier or requires actual intrinsics, but having support for this would probably better solidify future compilation using GNU-like toolchains.
Slightly offtopic, but I would dare to say that moving towards memcpy() and others(versus CopyMem(), etc) would benefit EDK2 generated ASM quality as the compiler has more of an idea of what you want to accomplish, if you pass -fbuiltin and let it assume your standard C library has the correct behavior; and indeed, GCC and Clang can, for instance, transform memcpy() calls into rep movsb's or whatnot based on the CPU microarch it's compiling for, or completely alias them if it sees you're doing it because of pointer aliasing.
“Don’t use structure assignment” is our best known method to avoid compiler intrinsic in EDKII coding style since the project is created.
A third part library may include structure assignment, then we have to link intrinsic.
If you only talk about memcpy, that is easy. But there might be others. The problem is: We don’t know how many intrinsic should be there. It is fully based upon compiler.
Gerd proposed to move to intrinsic from cryptopkg to mdepkg. But it is not done yet.
To me, using MemCopy is much easier to resolve this problem.
Is there a legitimate reason not to define memcpy? It'd be easier to do so and comply to the compiler's requirements.
Would you please use CopyMem() for the structure assignment?
VgpuGop->GopModeInfo = *GopModeInfo;
That is best known method to resolve memcpy symbol issue.
On Wed, May 25, 2022 at 5:45 PM Ard Biesheuvel <ardb@...> wrote:
On Wed, 25 May 2022 at 18:44, Pedro Falcato <pedro.falcato@...> wrote:
> On Wed, May 25, 2022 at 4:50 PM Ard Biesheuvel <ardb@...> wrote:
>> On Wed, 25 May 2022 at 17:08, Rebecca Cran <quic_rcran@...> wrote:
>> > I noticed OvmfPkg/OvmfPkgX64.dsc doesn't build with `-t CLANG38 -b
>> > NOOPT` (with clang version 14.0.2) with the latest edk2 master
>> > (07c0c2eb0a5970db614ebce1060fc79d6904bdfd):
>> > make: Nothing to be done for 'tbuild'.
>> > /usr/bin/ld: /usr/bin/ld: DWARF error: invalid or unhandled FORM value:
>> > 0x23
>> > /home/bcran/src/upstream/uefi/edk2/Build/OvmfX64/NOOPT_CLANG38/X64/OvmfPkg/VirtioGpuDxe/VirtioGpu/OUTPUT/VirtioGpuDxe.lib(Gop.obj):
>> > in function `GopSetMode':
>> > Gop.c:(.text.GopSetMode+0x418): undefined reference to `memcpy'
>> Can you dump the object file to see where the memcpy() call is emitted?
> I think I found the smoking gun: https://github.com/tianocore/edk2/blob/916f90baa547b3ebef8fa87c530e2f0c8e35e1e3/OvmfPkg/VirtioGpuDxe/Gop.c#L512
Indeed. We don't support struct assignment in Tianocore code, exactly
for this reason.
Passing -ffreestanding would also be a good idea as to let the compiler know it's dealing with a freestanding environment vs a hosted user-space one.