Also, quick note: I've seen linking blowing up because the ABIs being used in the bare metal code and in libgcc were different; this was in riscv64. I know x86 doesn't do such checks but a few others do, including RISCV (they tag object files with the ABI). This makes libgcc unusable in such architectures.

If we want to provide intrinsics, then possibly shipping our own compiler-rt would be the only solid option.

On Fri, 28 Jan 2022, 14:07 Gerd Hoffmann, <kraxel@redhat.com> wrote:
  Hi,

Oops, dropped the list by mistake, forwarding ...

----- Forwarded message from "kraxel@redhat.com" <kraxel@redhat.com> -----

Date: Fri, 28 Jan 2022 10:35:10 +0100
Subject: Re: [edk2-devel] [PATCH 00/24] CryptoPkg/openssl: update openssl
 submodule to v3.0
From: "kraxel@redhat.com" <kraxel@redhat.com>
To: Kilian Kegel <kilian_kegel@outlook.com>
Message-ID: <20220128093510.atupc4ly6bvwinlk@sirius.home.kraxel.org>
Content-Type: text/plain; charset=utf-8

  Hi,

> On my 32Bit Ubuntu standard installation I ran
>
>   1.  cc - Xlinker -Map=static.map hello.c -static
>   2.  cc  -Xlinker -Map=shared.map hello.c
>
> The first .OBJ file mentioned in the .MAP file is in both cases:
> /usr/lib/gcc/i686-linux-gnu/6/libgcc.a(_udivdi3.o)

Yes, you are correct.  gcc provides both shared and static intrinsics.
There is a command line switch to pick which one you want
(-static-libgcc, -shared-libgcc).

> It seems to me that GNU holds the intrinsic functions in a separate library
> that can be used without any change, and is always correct by definition.

>   1.  add libgcc.a as a search library, adjust the conf\tools_def.txt like:
>
> DEBUG_GCCxx_IA32_DLINK_FLAGS   = …predefined parameter … /usr/lib/gcc/i686-linux-gnu/6/libgcc.a

gcc documentation suggests to use just '-lgcc' (should pick the correct
library no matter what the compiler version and architecture is), so I
tried this:

-DEFINE GCC_DLINK2_FLAGS_COMMON     = -Wl,--script=$(EDK_TOOLS_PATH)/Scripts/GccBase.lds
+DEFINE GCC_DLINK2_FLAGS_COMMON     = -Wl,--script=$(EDK_TOOLS_PATH)/Scripts/GccBase.lds -lgcc

Build doesn't come very far.  Looks like the gcc intrinsics are not
free-standing but want call into libc:

/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_absvsi2.o): in function `__absvdi2':
(.text+0x18): undefined reference to `abort'
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_absvsi2.o): in function `__absvsi2':
(.text+0x32): undefined reference to `abort'
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_absvdi2.o): in function `__absvti2.cold':
(.text.unlikely+0x2): undefined reference to `abort'
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_addvsi3.o): in function `__addvdi3':
(.text+0xf): undefined reference to `abort'
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_addvsi3.o): in function `__addvsi3':
(.text+0x2d): undefined reference to `abort'
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_addvdi3.o):(.text.unlikely+0x2): more undefined references to `abort' follow
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(_eprintf.o): in function `__eprintf':
(.text+0x8): undefined reference to `stderr'
/usr/bin/ld: (.text+0x1d): undefined reference to `fprintf'
/usr/bin/ld: (.text+0x25): undefined reference to `fflush'
/usr/bin/ld: (.text+0x2a): undefined reference to `abort'
/usr/bin/ld: /usr/lib/gcc/x86_64-redhat-linux/11/libgcc.a(generic-morestack.o): in function `__morestack_fail':
(.text+0xbc): undefined reference to `writev'
[ ... more errors snipped ... ]

The generic-morestack.o issues should be solvable, that shouldn't be
something which tianocore actually needs.  Not sure why the linker tries
to resolve symbols for object files which should not be needed in the
first place.  Possibly something else is fishy here, any hints are
welcome.  Something in the linker script maybe?

But the math intrinsics apparently having error paths which print a
message and abort doesn't look very promising to me.

Also: When trying arm cross-builds I run into the ABI problem already
mentioned elsewhere in this thread:

/usr/bin/arm-linux-gnu-ld: error: /usr/lib/gcc/arm-linux-gnueabi/11/libgcc.a(_muldi3.o) uses VFP register arguments, /home/kraxel/projects/edk2/Build/ArmVirtQemu-ARM/DEBUG_GCC5/ARM/OvmfPkg/VirtioBlkDxe/VirtioBlk/DEBUG/VirtioBlkDxe.dll does not

Patches are here:
  https://github.com/kraxel/edk2/commits/intrinsics-playground

> >* I have my doubts that compiler's builtin libraries are optimized for
> >   size, so I'd suspect we would see a noticeable size grow from that.
> Please check the size of __udivdi3() and whether the tianocore reimplementation is smaller or not

I'll rather check the size of the final binaries, but I can only do that
once the build works ...

> The intrinsic library belongs to the compiler not to the build system.

I'm open to explore that path, but apparently we have a number of road
blocks along the way.  Seems neither gcc nor xcode (see other reply)
provide a usable free-standing intrinsic library ...

take care,
  Gerd