[clang] __MMX__ defined on non-SSE2 x86 targets

[thrimbor]

Since #96540 MMX intrinsics require SSE2, but __MMX__ is still defined on targets that technically support MMX but not SSE2, like pentium3, causing compilation issues with e.g. SDL2 for those targets.

[applesauceman]

SDL opted to keep MMX support by building the three affected functions
with __attribute__((target("mmx"))), see libsdl-org/SDL@69de696

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Stefan Schmidt (thrimbor)

Since https://github.com//pull/96540 MMX intrinsics require SSE2, but `__MMX__` is still defined on targets that technically support MMX but not SSE2, like `pentium3`, causing compilation issues with e.g. SDL2 for those targets.

[llvmbot]

@llvm/issue-subscribers-clang-driver

Author: Stefan Schmidt (thrimbor)

Since https://github.com//pull/96540 MMX intrinsics require SSE2, but `__MMX__` is still defined on targets that technically support MMX but not SSE2, like `pentium3`, causing compilation issues with e.g. SDL2 for those targets.

[RKSimon]

This could be a really dumb idea, but we could just drop the MMX/SSE2 requirement entirely for MMX intrinsics that are just generic implementations (the simple arithmetic, logic, shuffles, compares) - on non-SSE2 targets they'd just scalarize. I'm not sure if it'd avoid all the issues that SDL et al. have encountered though.

@phoebewang @jyknight WDYT?

[jyknight]

Adjusting the attributes only on the generic functions won't really help. We'd also need to introduce generic implementations for e.g. _mm_slli_si64 too, which doesn't really seem worthwhile, IMO.

While we could consider disabling the __MMX__ define for non-sse2 targets, my previous exploration found that most pure-MMX code in the wild is using assembly (either inline or out of line), rather than the intrinsics. And that usage does continue to be supported on pentium3. Disabling __MMX__ would (in many instances) break that usage. So I don't think we should.

@thrimbor It looks like SDL2 already tests that MMX intrinsics are supported, via its CMakeLists.txt "COMPILER_SUPPORTS_MMX" detection since 905b892314d1e3950912ac24d417a729ae57f010 (Mar 21 2023, released in 3.2.0). Are you using a version of SDL2 from before that? If so, I'd suggest to update your SDL2.

[thrimbor]

While we could consider disabling the __MMX__ define for non-sse2 targets, my previous exploration found that most pure-MMX code in the wild is using assembly (either inline or out of line), rather than the intrinsics. And that usage does continue to be supported on pentium3. Disabling __MMX__ would (in many instances) break that usage. So I don't think we should.

@thrimbor It looks like SDL2 already tests that MMX intrinsics are supported, via its CMakeLists.txt "COMPILER_SUPPORTS_MMX" detection since 905b892314d1e3950912ac24d417a729ae57f010 (Mar 21 2023, released in 3.2.0). Are you using a version of SDL2 from before that? If so, I'd suggest to update your SDL2.

That is SDL3, for reasons unrelated to LLVM or SDL we are currently using an older version of SDL2.

My point however wasn't that I can't compile SDL - it's trivial to patch - it's that I believed that __MMX__'s intent is to show availability of the MMX intrinsics. If the intent instead is to just show availability of the instruction set (or there's no good way to show that MMX is supported but the intrinsics are not) then I'm fine with that and we can close this issue.