[AArch64][GlobalISel] llvm.fptoui.sat.i32.f128 miscompile

[nikic]

Per rust-lang/compiler-builtins#760, LLVM 20 appears to be miscompiling fptoui.sat.i32.f128 on AArch64 when GlobalISel is used.

Assembly for LLVM 19 and LLVM 20: https://llvm.godbolt.org/z/n3K4MEG7r

Likely related to #96297, cc @davemgreen

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Nikita Popov (nikic)

Per https://github.com/rust-lang/compiler-builtins/pull/760, LLVM 20 appears to be miscompiling fptoui.sat.i32.f128 on AArch64 when GlobalISel is used.

Assembly for LLVM 19 and LLVM 20: https://llvm.godbolt.org/z/n3K4MEG7r

Likely related to #96297, cc @davemgreen

[nikic]

Possibly I'm misreading the code, but I suspect that the select operands are inverted?

llvm-project/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

Lines 7818 to 7821 in 760ec2c

	auto MaxC = MIRBuilder.buildFConstant(SrcTy, MinFloat);
	auto MaxP = MIRBuilder.buildFCmp(CmpInst::FCMP_ULT,
	SrcTy.changeElementSize(1), Src, MaxC);
	auto Max = MIRBuilder.buildSelect(SrcTy, MaxP, Src, MaxC);

This is doing Src < MinFloat ? Src : MinFloat, but surely it should be Src < MinFloat ? MinFloat : Src?

It doesn't really help that MaxC is referring to the minimum float...

[davemgreen]

Hi - Yeah it sounds like I translated the code from using fminnum/fmaxnum badly..

Do you want me to put together a patch?

[nikic]

Hi - Yeah it sounds like I translated the code from using fminnum/fmaxnum badly..

Do you want me to put together a patch?

Yes, please :)

[davemgreen]

/cherry-pick 70ed381