clang-cl can not support function targets

[carewolf]

The immintrin.h headers has a bug where it does not include sub-arch intrinsics headers if _MSC_VER is defined unless the subarch target is current.

This is inconsistent with MSVC which always defines intrinsics regardless of active arch, and is also inconsistent with normal-clang (gcc-style) which defines them for use with subarch function targets. The result is there is no way of using intrinsics for subarch targets with clang-cl, since neither gcc nor msvc style works.

This has forced us to disable many optimizations in Qt with clang-cl, see https://bugreports.qt.io/browse/QTBUG-88081, https://bugreports.qt.io/browse/QTBUG-88434 and https://bugreports.qt.io/browse/QTBUG-98253

I suggest at least allowing gcc style, and the intrinsics working in target attributed functions, if you can't support the MSVC style.

[zero9178]

This is indeed a nasty divergence from MSVC and one that IIRC is actually quite hard to fix as its also an issue in LLVM, not just clang. There is a workaround however as you can use the /clang: flag to pass GNU style command line options. That way you can eg. pass /clang:-mrdseed to enable that sub-arch feature.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

[carewolf]

That won't help. The problem as I see it is that immintrin.h does not include non-current subarch targets when _MSC_VER is defined. I think if that was removed it still wouldn't work as MSVC does, but it would atleast work in functions with an appropiate target function attribute (clang/gcc style).

[zero9178]

I am not quite sure as to how my above suggestion differs from clang/gcc style. Using the GCC style -m options will lead to the definition of the various feature test macros and therefore the inclusion of the sub arch headers despite _MSC_VER being defined. As an example from Qt source I wrote:

#include <immintrin.h>

int main()
{
        unsigned int value;
        _rdrand32_step(&value);
}

and then compiled it using clang-cl test.cpp /clang:-mrdrnd. As far as I could tell this should work with all the intrinsics.

But yes this won't match MSVC behaviour, which is clearly a bug. The above would simply serve as a workaround

[carewolf]

I think we are talking past each other. I am talking about using intrinsics without having similar commandline flags for runtime target detection

MSVC style:

#include <immintrin.h>
void foo_avx2(args) {
 _mm256_avx_command();
}

clang/gcc style:

#include <immintrin.h>
__attribute__(__target__("arch=haswell")) // or __attribute__(__target__("haswell")), can't remember which is gcc and which is clang
void foo_avx2(args) {
 _mm256_avx_command();
}

This is for runtime detection of CPU feature, not anything enabled at compile time.

In MSVC non-target intrinsics always works (somehow), in clang/gcc they work if appropriate function target has been declared where used. In clang-cl neither works because the compile-time flags are checked before defining the functions in the immintrin.h header.

[wangwenx190]

Please fix this bug in clang-cl, this issue is really annoying.

[ThiagoIze]

I think this has already been explained by Allan, but just to make sure it's clear. This bug is preventing us from writing a single Windows binary that can be optimized for various architectures. For instance, we'd like the binary to work on older CPUs that don't have AVX and on newer machines with AVX we can run AVX optimized functions instead of the generic SSE2 functions. We might even have optimizations for AVX-512.

Fixing this can give applications a 4x speedup on Windows while still allowing the application to run on older CPUs. It's not everyday a compiler can do something that gives a 4x speedup, so I think this should be a high priority to fix.

[wangwenx190]

@ThiagoIze is right, this is performance critical for many applications, please prioritize this issue.

[wangwenx190]

Any progress on this?

[wangwenx190]

🤔