x86: improve SIMD detection #257
Conversation
Since 56dbde0 (Add xgetbv feature detection support on x86, 2024-03-23), AVX support can be "rescinded" if it has been partially disabled, but the check added didn't include the possibility that OSXSAVE could be also disabled (ex: using the `noxsave` kernel parameter in Linux). Correct that and while at it, add a similar check for SSE2 (only relevant in 32bit) and for the possibility that if not enabled (ex: the OS never set CR4.OSFXSR), then SSE41 support should also be rescinded.
| if ((feature_list & CPU_FEATURE_OSXSAVE) && (execute_get_xcr0_low() & 0x4) == 0) | ||
| #if defined(SLJIT_CONFIG_X86_32) && SLJIT_CONFIG_X86_32 | ||
| if (!(feature_list & CPU_FEATURE_SSE2)) | ||
| feature_list &= ~(sljit_u32)CPU_FEATURE_SSE41; |
There was a problem hiding this comment.
Who sets this? Does the compiler runs on that old hardware?
There was a problem hiding this comment.
Just like AVX2; SSE2 also needs OS support, which could be missing (ex: Debian 3).
When gcc targets a CPU with SSE2 it defines __SSE2__ just like MSVC does with _M_IX86_FP
Note that in this case, if saving of the XMM registers is not supported, any operation that uses them will SIGILL, hence why this code disables SSE41 (hence SIMD) in that case.
There was a problem hiding this comment.
The question was how is it possible to set SSE4.1 without setting SSE2. I always worry that we accidentally disables features for newer systems.
| if (info[3] & 0x8000) | ||
| feature_list |= CPU_FEATURE_CMOV; | ||
| #if (defined(SLJIT_DETECT_SSE2) && SLJIT_DETECT_SSE2) \ | ||
| && (defined(__SSE2__) || (defined(_WIN64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 2))) |
There was a problem hiding this comment.
This will disable SSE2 with SUNPRO, TINYC or other compilers than won't provide the defines.
Avoid SIGILL when AVX has been disabled by removing OSXSAVE, as well as when SSE2 is not enabled (most likely by running a really old OS in modern hardware)