[LLVM] llvm.fptosi.sat.* and llvm.fptoui.sat.* generate suboptimal code in some cases on x86

[johnplatts]

Here is a snippet that demonstrates suboptimal code generation for llvm.fptosi.sat.* and llvm.fptoui.sat.* on x86 with SSE2 or AVX512F enabled but not with AVX10.2 enabled in the case where floating point exceptions are ignored and masked:
https://godbolt.org/z/bPY4orsK5

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: John Platts (johnplatts)

Here is a snippet that demonstrates suboptimal code generation for llvm.fptosi.sat.* and llvm.fptoui.sat.* on x86 with SSE2 or AVX512F enabled but not with AVX10.2 enabled in the case where floating point exceptions are ignored and masked: https://godbolt.org/z/bPY4orsK5

[phoebewang]

Can you please give an optimal assemble for reference?

[MaxGraey]

Btw I think MSVC's intrinsic _cvt_dtoull_sat produce little bit more optimal code for f64 -> u64 saturate conversion:
https://godbolt.org/z/vzzTqKe1G

The key feature is that they check the specific case of x >= 2^63 where they work directly with the mantissa:

... check / adjust bounds ...

if (x >= 2^63) { // unlikely
   // Handle values between 2^63 and 2^64
   uint64_t bits = *(uint64_t*)&x; // reinterpret double as uint64_t
   bits |= (1ULL << 52); // Set implicit bit
   bits <<= 11; // Shift to align mantissa
   return bits;
} else {
   return (uint64_t)(int64_t)x;
}

[johnplatts]

@phoebewang Here is a comparison between two different saturated scalar F64 to I32/I64 implementations, along with a llvm-mc analysis of the generated code:

• https://godbolt.org/z/vczP8YEos (llvm.fptosi.sat.*/llvm.fptoui.sat.*)
• https://godbolt.org/z/xTxEefzsr (alternative implementation for SSE2/AVX512F)

[johnplatts]

@phoebewang - I reported a similar issue with suboptimal llvm.fptosi.sat.*/llvm.fptoui.sat.* codegen on PowerPC in issue #136315.

[johnplatts]

@phoebewang Suboptimal code is being generated for @llvm.fptosi.sat.v2i32.v2f64, @llvm.fptosi.sat.v4i32.v4f32, and @llvm.fptoui.sat.v4i32.v4f32 on x86 targets that support SSE2 according to a snippet at https://godbolt.org/z/TYzrW6e3f.

A more optimal alternative for these ops on x86 with SSE2 can be found at https://godbolt.org/z/s6nexGx64.