[openmp] __kmp_x86_cpuid fix for i386/PIC builds.

[devnexen]

No description provided.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[brad0]

cc @MaskRay

[mgorny]

I'm sorry to say but this change actually broke 32-bit builds on Gentoo, when using GCC. Most of the tests segfault now, e.g.:

0/1] cd /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test && /tmp/portage/sys-libs/libomp-19.0.0.9999/temp/python3.11/bin/python3 /usr/bin/lit -vv -j 12 /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test
-- Testing: 325 tests, 12 workers --
FAIL: libomp :: affinity/format/api.c (1 of 325)
******************** TEST 'libomp :: affinity/format/api.c' FAILED ********************
Exit Code: -11

Command Output (stdout):
--
# RUN: at line 1
/usr/lib/ccache/bin/i686-pc-linux-gnu-clang -fopenmp   -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test -L /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src  -fno-omit-frame-pointer -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/ompt /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/affinity/format/api.c -o /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp -lm -latomic && /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp
# executed command: /usr/lib/ccache/bin/i686-pc-linux-gnu-clang -fopenmp -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test -L /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src -fno-omit-frame-pointer -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/ompt /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/affinity/format/api.c -o /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp -lm -latomic
# executed command: /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp
# note: command had no output on stdout or stderr
# error: command failed with exit status: -11

--

********************

My educated guess is that you aren't listing edi in clobbers. If I do that, most of the tests start passing again but I still get a handful of segfaults.

[devnexen]

I'm sorry to say but this change actually broke 32-bit builds on Gentoo, when using GCC. Most of the tests segfault now, e.g.:

0/1] cd /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test && /tmp/portage/sys-libs/libomp-19.0.0.9999/temp/python3.11/bin/python3 /usr/bin/lit -vv -j 12 /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test
-- Testing: 325 tests, 12 workers --
FAIL: libomp :: affinity/format/api.c (1 of 325)
******************** TEST 'libomp :: affinity/format/api.c' FAILED ********************
Exit Code: -11

Command Output (stdout):
--
# RUN: at line 1
/usr/lib/ccache/bin/i686-pc-linux-gnu-clang -fopenmp   -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test -L /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src  -fno-omit-frame-pointer -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/ompt /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/affinity/format/api.c -o /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp -lm -latomic && /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp
# executed command: /usr/lib/ccache/bin/i686-pc-linux-gnu-clang -fopenmp -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test -L /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/src -fno-omit-frame-pointer -I /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/ompt /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp/runtime/test/affinity/format/api.c -o /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp -lm -latomic
# executed command: /tmp/portage/sys-libs/libomp-19.0.0.9999/work/openmp_build-abi_x86_32.x86/runtime/test/affinity/format/Output/api.c.tmp
# note: command had no output on stdout or stderr
# error: command failed with exit status: -11

--

********************

My educated guess is that you aren't listing edi in clobbers. If I do that, most of the tests start passing again but I still get a handful of segfaults.

What happens if you do the following instead ?

__asm__ __volatile__("pushl  %%ebx\n"
		"cpuid\n"
		"mov    %%ebx,%1\n"
		"popl   %%ebx"
                 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
                 : "a"(leaf), "c"(subleaf));

[mgorny]

Actually, this change makes no sense to me (either the committed or the proposed variant). You're explicitly specifying ebx as the output operand, so the compiler should assume it will be overwritten. I've tested it both with clang and gcc, and indeed it does pushl %ebx before the assembly: https://godbolt.org/z/Tchze37ja

On top of that, you're effectively clobbering %ebx set by cpuid, so p->ebx ends up containing "junk" prior to cpuid call rather than the value returned by cpuid.

[mgorny]

Adding a revert on top of dbbdee2 (this is the commit I've originally tested) makes all tests pass again.

[devnexen]

ok let s revert it then.