From 19a22232513de54674f03b0dd1d71da58e287a21 Mon Sep 17 00:00:00 2001 From: Eric Biggers Date: Tue, 12 Sep 2023 22:18:19 -0700 Subject: [PATCH] Don't use bzhi intrinsics on old MSVC versions A test failure, narrowed down to a problem in deflate_decompress_bmi2(), was reported when libdeflate is built in release mode with MSVC from VS2017. VS2022 works fine. Currently, in MSVC builds of libdeflate, the only difference between the bmi2 and default decompression functions is whether the bzhi instrinsics (_bzhi_u64() and _bzhi_u32()) are used. But as far as I can tell, libdeflate uses these intrinsics correctly. Avoid this issue by disabling the use of bzhi intrinsics if the compiler is MSVC before VS2022. MSVC is closed source, so root causing this issue is not possible. But one hypothesis is that MSVC may have overlooked that the bzhi intrinsics are supposed to always truncate the index argument to 8 bits, matching the corresponding CPU instructions and as specified by Intel at https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-8/bzhi-u32-64.html. Most software that uses these intrinsics probably doesn't rely on this behavior, but libdeflate does. So a compiler bug seems plausible here, especially considering the issue went away in newer compiler versions. Fixes https://github.com/ebiggers/libdeflate/issues/325 --- lib/x86/cpu_features.h | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/lib/x86/cpu_features.h b/lib/x86/cpu_features.h index 3cb3f465..ad14e435 100644 --- a/lib/x86/cpu_features.h +++ b/lib/x86/cpu_features.h @@ -145,6 +145,16 @@ static inline u32 get_x86_cpu_features(void) { return 0; } #else # define HAVE_BMI2_INTRIN 0 #endif +/* + * MSVC from VS2017 (toolset v141) apparently miscompiles the _bzhi_*() + * intrinsics. It seems to be fixed in VS2022. + */ +#if defined(_MSC_VER) && _MSC_VER < 1930 /* older than VS2022 (toolset v143) */ +# undef HAVE_BMI2_NATIVE +# undef HAVE_BMI2_INTRIN +# define HAVE_BMI2_NATIVE 0 +# define HAVE_BMI2_INTRIN 0 +#endif #endif /* ARCH_X86_32 || ARCH_X86_64 */