zig c++ overrides explicitly set cpu features

[happyalu]

I was trying to build tensorflow with zig c++ on an intel kabylake processor and building a dependency (xnnpack) failed when compiling a source that was using XOP intrinsics. Their cmake build was adding -mxop when compiling the appropriate files. The build worked with clang++. I was able to narrow this down to a small test case:

cat > test.cc <<EOF
#include <x86intrin.h>

int main(){
        __m128i x;
        _mm_macc_epi32(x, x, x);
        return 0;
}
EOF

$ clang++ -mxop test.cc; echo $?
0

$ zig c++ -mxop test.cc
test.cc:5:9: error: always_inline function '_mm_macc_epi32' requires target feature 'xop', but would be inlined into function 'main' that is compiled without support for 'xop'
        _mm_macc_epi32(x, x, x);
        ^
1 error generated.

Running with -v shows that zig sets the xop features twice (once +xop and later -xop) : ... -target-cpu x86-64 -target-feature +xop -tune-cpu generic ... -target-cpu skylake ... -target-feature -xop ...

zig -cc1 -triple x86_64-unknown-linux-gnu -emit-obj --mrelax-relocations -disable-free -disable-llvm-verifier -discard-value-names -main-file-name test.cc -mrelocation-model pic -pic-level 2 -fhalf-no-semantic-interposition -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -target-feature +xop -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -v -nostdsysteminc -nobuiltininc -resource-dir /path/lib/clang/12.0.1 -dependency-file /path/.cache/zig/tmp/52ed2834085a3e15-test.o.d -MT /path/.cache/zig/tmp/52ed2834085a3e15-test.o -sys-header-deps -MV -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libcxx/include -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libcxxabi/include -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libunwind/include -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/include -isystem /path/code/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libc/include/x86_64-linux-gnu -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libc/include/generic-glibc -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libc/include/x86_64-linux-any -isystem /path/zig-linux-x86_64-0.9.0-dev.4+ddf9c40bc/lib/libc/include/any-linux-any -isystem /usr/local/include -isystem /usr/include/x86_64-linux-gnu -isystem /usr/include -D _LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS -D _LIBCXXABI_DISABLE_VISIBILITY_ANNOTATIONS -D _LIBCPP_HAS_NO_VENDOR_AVAILABILITY_ANNOTATIONS -D _DEBUG -Og -fdeprecated-macro -fdebug-compilation-dir /path/tflite_build/xnnpack -ferror-limit 19 -fsanitize=alignment,array-bounds,bool,builtin,enum,float-cast-overflow,function,integer-divide-by-zero,nonnull-attribute,null,object-size,pointer-overflow,return,returns-nonnull-attribute,shift-base,shift-exponent,signed-integer-overflow,unreachable,vla-bound -fsanitize-trap=alignment,array-bounds,bool,builtin,enum,float-cast-overflow,function,integer-divide-by-zero,nonnull-attribute,null,object-size,pointer-overflow,return,returns-nonnull-attribute,shift-base,shift-exponent,signed-integer-overflow,unreachable,vla-bound -stack-protector 2 -stack-protector-buffer-size 4 -fgnuc-version=4.2.1 -fcxx-exceptions -fexceptions -fcolor-diagnostics -fno-spell-checking -target-cpu skylake -target-feature -16bit-mode -target-feature -32bit-mode -target-feature -3dnow -target-feature -3dnowa -target-feature +64bit -target-feature +adx -target-feature +aes -target-feature -amx-bf16 -target-feature -amx-int8 -target-feature -amx-tile -target-feature +avx -target-feature +avx2 -target-feature -avx512bf16 -target-feature -avx512bitalg -target-feature -avx512bw -target-feature -avx512cd -target-feature -avx512dq -target-feature -avx512er -target-feature -avx512f -target-feature -avx512ifma -target-feature -avx512pf -target-feature -avx512vbmi -target-feature -avx512vbmi2 -target-feature -avx512vl -target-feature -avx512vnni -target-feature -avx512vp2intersect -target-feature -avx512vpopcntdq -target-feature -avxvnni -target-feature +bmi -target-feature +bmi2 -target-feature -branchfusion -target-feature -cldemote -target-feature +clflushopt -target-feature -clwb -target-feature -clzero -target-feature +cmov -target-feature +cx16 -target-feature +cx8 -target-feature -enqcmd -target-feature +ermsb -target-feature +f16c -target-feature -false-deps-lzcnt-tzcnt -target-feature +false-deps-popcnt -target-feature -fast-11bytenop -target-feature +fast-15bytenop -target-feature -fast-7bytenop -target-feature -fast-bextr -target-feature +fast-gather -target-feature -fast-hops -target-feature -fast-lzcnt -target-feature +fast-scalar-fsqrt -target-feature -fast-scalar-shift-masks -target-feature +fast-shld-rotate -target-feature +fast-variable-shuffle -target-feature +fast-vector-fsqrt -target-feature -fast-vector-shift-masks -target-feature +fma -target-feature -fma4 -target-feature +fsgsbase -target-feature -fsrm -target-feature +fxsr -target-feature -gfni -target-feature -hreset -target-feature -idivl-to-divb -target-feature +idivq-to-divl -target-feature +invpcid -target-feature -kl -target-feature -lea-sp -target-feature -lea-uses-ag -target-feature -lvi-cfi -target-feature -lvi-load-hardening -target-feature -lwp -target-feature +lzcnt -target-feature +macrofusion -target-feature +mmx -target-feature +movbe -target-feature -movdir64b -target-feature -movdiri -target-feature -mwaitx -target-feature +nopl -target-feature -pad-short-functions -target-feature +pclmul -target-feature -pconfig -target-feature -pku -target-feature +popcnt -target-feature -prefer-128-bit -target-feature -prefer-256-bit -target-feature -prefer-mask-registers -target-feature -prefetchwt1 -target-feature +prfchw -target-feature -ptwrite -target-feature -rdpid -target-feature +rdrnd -target-feature +rdseed -target-feature -retpoline -target-feature -retpoline-external-thunk -target-feature -retpoline-indirect-branches -target-feature -retpoline-indirect-calls -target-feature -rtm -target-feature +sahf -target-feature -serialize -target-feature -seses -target-feature +sgx -target-feature -sha -target-feature -shstk -target-feature +slow-3ops-lea -target-feature -slow-incdec -target-feature -slow-lea -target-feature -slow-pmaddwd -target-feature -slow-pmulld -target-feature -slow-shld -target-feature -slow-two-mem-ops -target-feature -slow-unaligned-mem-16 -target-feature -slow-unaligned-mem-32 -target-feature -soft-float -target-feature +sse -target-feature +sse2 -target-feature +sse3 -target-feature +sse4.1 -target-feature +sse4.2 -target-feature -sse4a -target-feature -sse-unaligned-mem -target-feature +ssse3 -target-feature -tbm -target-feature -tsxldtrk -target-feature -uintr -target-feature -use-aa -target-feature -use-glm-div-sqrt-costs -target-feature -vaes -target-feature -vpclmulqdq -target-feature +vzeroupper -target-feature -waitpkg -target-feature -wbnoinvd -target-feature -widekl -target-feature +x87 -target-feature -xop -target-feature +xsave -target-feature +xsavec -target-feature +xsaveopt -target-feature +xsaves -faddrsig -o /path/.cache/zig/tmp/52ed2834085a3e15-test.o -x c++ test.cc

Not sure if this was happening due to native cpu detection, but I tried adding -target x86_64-linux-gnu to pretend to cross-compile, but it didn't help. Am I using this wrong? Is there a workaround or fix I could try?

Thanks!

[andrewrk]

I think this is a matter of missing support for -m flag detection in general. It would be near trivial to solve it for -mxop, but the real question is how do we systematically solve it for all such parameters?

[happyalu]

There's something that is correctly adding +xop and something else later adding rest of the flags including -xop. Perhaps I'll try to figure out where these are happening. Not sure if we can detect that +xop was set earlier, and then not reset it when emitting rest of the flags.

Tried the same thing with -mavx512f that my CPU doesn't support, and it generated -target-cpu x86-64 -target-feature +avx512f -tune-cpu generic ... -target-cpu skylake ... -target-feature -avx512f.

Is there a way to disable emitting flags for the detected cpu?

(Update): It seems the -mxop gets passed over to clang, and then we also add the target cpu features here

[matpow2]

We ran into this issue while trying to build libvpx with zig cc.
Essentially we have something like this in our CMake script:

        set_source_files_properties(
            ${VPX_GLOB_CPP_AVX}
            PROPERTIES COMPILE_FLAGS
            "-mavx2")
        set_source_files_properties(
            ${VPX_GLOB_CPP_SSSE3}
            PROPERTIES COMPILE_FLAGS
            "-mssse3")
        set_source_files_properties(
            ${VPX_GLOB_CPP_SSE2}
            PROPERTIES COMPILE_FLAGS
            "-msse2")

but zig cc only supports -mcpu and not -mxxx.

We worked around this by creating a wrapper for zig.exe and setting -mcpu=baseline+x+y+z if one of the CPU feature flags were set, but would be great if this was supported without the workaround.

[andrewrk]

Fixed now:

$ cat test.c
#include <x86intrin.h>

int main(){
        __m128i x;
        _mm_macc_epi32(x, x, x);
        return 0;
}

$ ./zig cc test.c
test.c:5:9: error: always_inline function '_mm_macc_epi32' requires target feature 'xop', but would be inlined into function 'main' that is compiled without support for 'xop'
        _mm_macc_epi32(x, x, x);
        ^
1 error generated.

$ ./zig cc -mxop test.c

$ 

[andrewrk]

Ah this was a duplicate of #4912 btw.