[OpenMP] [test] Skip the -mlong-double-80 test on MSVC ABI

[mstorsjo]

Within the MSVC ABI, long doubles are the same as regular 64 bit doubles. This test case, which is compiled with -mlong-double-80, cannot work when libomp has been compiled without that flag, as -mlong-double-80 changes the calling convention for the tested functions.

[llvmbot]

@llvm/pr-subscribers-openmp

Author: Martin Storsjö (mstorsjo)

Changes

Within the MSVC ABI, long doubles are the same as regular 64 bit doubles. This test case, which is compiled with -mlong-double-80, cannot work when libomp has been compiled without that flag, as -mlong-double-80 changes the calling convention for the tested functions.

---

Full diff: https://github.com/llvm/llvm-project/pull/81115.diff

1 Files Affected:

• (modified) openmp/runtime/test/atomic/kmp_atomic_float10_max_min.c (+2-2)

diff --git a/openmp/runtime/test/atomic/kmp_atomic_float10_max_min.c b/openmp/runtime/test/atomic/kmp_atomic_float10_max_min.c
index 82d8cbc19d7a16..72ba1eeb643ef1 100644
--- a/openmp/runtime/test/atomic/kmp_atomic_float10_max_min.c
+++ b/openmp/runtime/test/atomic/kmp_atomic_float10_max_min.c
@@ -28,7 +28,7 @@ extern long double __kmpc_atomic_float10_min_cpt(ident_t *id_ref, int gtid,
 
 int main() {
   int ret = 0;
-#if KMP_ARCH_X86 || KMP_ARCH_X86_64
+#if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !defined(_MSC_VER)
   long double s = 012.3456; // small
   long double e = 123.4567; // middle
   long double d = 234.5678; // big
@@ -158,6 +158,6 @@ int main() {
     printf("passed\n");
 #else
   printf("Unsupported architecture, skipping test...\n");
-#endif // KMP_ARCH_X86 || KMP_ARCH_X86_64
+#endif // (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !defined(_MSC_VER)
   return ret;
 }

[mstorsjo]

Ping @shiltian

[shiltian]

LG

[jhuber6]

Might be better to check for 64-bit doubles in general since that's present on various targets. But probably not a huge deal overall.

This should do it

#include <float.h>
#if (LDBL_MANT_DIG == 53)
// 64-bit double
#elif (LDBL_MANT_DIG == 64)
// 80-bit double
#else
// 128-bit double
#endif

[mstorsjo]

Might be better to check for 64-bit doubles in general since that's present on various targets. But probably not a huge deal overall.

This should do it

#include <float.h>
#if (LDBL_MANT_DIG == 53)
// 64-bit double
#elif (LDBL_MANT_DIG == 64)
// 80-bit double
#else
// 128-bit double
#endif

Hmm, so I kinda wonder how this test is supposed to work in the first place.

On the platforms where this test actually does work, the long doubles already are 80 bit from the start, so adding the -mlong-double-80 option is essentially a no-op. And for other platforms, where long doubles aren't 80 bit, passing -mlong-double-80 breaks the ABI between the calling code and libomp (libomp compiled as usual with whatever is the compiler default for long doubles), so I'm not really sure how that's meant to work at all.

Anyway, without taking on digging up what the point of this test was to begin with, are you ok with me merging this patch to hide the failure in MSVC build modes at least?