CMake: Fix illegal instruction on MSVC

[khuiqel]

Problem

CMakeLists.txt checks SNAPPY_HAVE_BMI2 by compiling a function with _bzhi_u32. On Clang and GCC, SNAPPY_HAVE_BMI2 will be 0 if BMI2 is not enabled. On MSVC, the result will always be 1 on x86 regardless of the microarchitecture level. On CPUs without AVX2/BMI2, this will result in an illegal instruction (in snappy::ExtractLowBytes()). Presumably this also occurs with SNAPPY_HAVE_SSSE3 and SNAPPY_HAVE_X86_CRC32, as well as SNAPPY_HAVE_NEON_CRC32 and SNAPPY_HAVE_NEON on ARM, but I couldn't verify that. Can be reproduced with the BM_UFlat benchmarks.

The BMI2 check was added in 4f0adca, which noted the __BMI2__ preprocessor macro doesn't exist on MSVC. Did not test MSVC+Clang.

Solution

It's very much a band-aid, but wrap the instruction checks in if(NOT MSVC). This solves the illegal instruction issue, but SSSE3/CRC32/BMI2 intrinsics will no longer generate.

+if(NOT MSVC)
+
/* ARM and other x86 intrinsics checks */

check_cxx_source_compiles("
#include <immintrin.h>
int main() {
  return _bzhi_u32(0, 1);
}" SNAPPY_HAVE_BMI2)
 
+endif(NOT MSVC)
+

All 25 test pass with this patch. The following tests fail without this patch (on a CPU without BMI2):

[  FAILED  ] 14 tests, listed below:
[  FAILED  ] CorruptedTest.VerifyCorrupted
[  FAILED  ] Snappy.SimpleTests
[  FAILED  ] Snappy.AppendSelfPatternExtensionEdgeCases
[  FAILED  ] Snappy.AppendSelfPatternExtensionEdgeCasesExhaustive
[  FAILED  ] Snappy.MaxBlowup
[  FAILED  ] Snappy.RandomData
[  FAILED  ] Snappy.CompressionContext
[  FAILED  ] Snappy.CompressionContextStaticWorkspace
[  FAILED  ] Snappy.IOVecSinkEdgeCases
[  FAILED  ] Snappy.IOVecCopyOverflow
[  FAILED  ] Snappy.ZeroOffsetCopy
[  FAILED  ] Snappy.ZeroOffsetCopyValidation
[  FAILED  ] Snappy.LiteralLengthU32Overflow
[  FAILED  ] Snappy.TestBenchmarkFiles

Checking for SSSE3 and SSE4.2 (x86 CRC32) support on MSVC is impossible because MSVC only defines __AVX__, __AVX2__, and AVX-512 macros with no plan to change. So a proper solution for MSVC would involve adding SNAPPY_HAVE_XXXX options like AVX/AVX2 for SSSE3 and SSE4.2, like Highway does.

Somewhat related to #254