[LLVM] llvm.fptosi.sat.* and llvm.fptoui.sat.* generate suboptimal code on PowerPC targets

[johnplatts]

The @llvm.fptosi.sat.* and @llvm.fptoui.sat.* intrinsics generate suboptimal code on PowerPC targets as demonstrated in a snippet over at https://godbolt.org/z/r6ajjTM3v.

Equivalent, more optimal alternatives on 64-bit PowerPC, POWER8, and POWER9 can be found at https://godbolt.org/z/5hKP9zhsa.

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: John Platts (johnplatts)

The @llvm.fptosi.sat.* and @llvm.fptoui.sat.* intrinsics generate suboptimal code on PowerPC targets as demonstrated in a snippet over at https://godbolt.org/z/r6ajjTM3v.

Equivalent, more optimal alternatives on 64-bit PowerPC, POWER8, and POWER9 can be found at https://godbolt.org/z/4rezzhGna.

[johnplatts]

Here are links to LLVM IR snippets for saturated scalar F64 to I32/I64/U32/U64 conversions on PowerPC with generated assembly and llvm-mca machine code analysis:

• https://godbolt.org/z/hPh4dGjr1 (llvm.fptosi.sat.*/llvm.fptoui.sat.*)
• https://godbolt.org/z/7PPffczaq (alternative implementation using PowerPC or VSX instructions)