llvm-test-suite SingleSource/UnitTests/Vectorizer/fmax-reduction miscompiles

[Meinersbur]

https://github.com/llvm/llvm-test-suite/blob/main/SingleSource/UnitTests/Vectorizer/fmax-reduction.cpp on x86-64

Expected output:

Checking fmaxnum_start_neg_2
Checking fmaxnum_start_min
Checking fmaxnum_start_denorm_min
Checking fmaxnum_start_is_nan
Checking fmax_non_strict_start_neg_2
Checking fmax_cmp_max_gt_start_neg_2
Checking fmax_cmp_max_lt_start_neg_2
Checking fmax_cmp_max_lt_start_denorm_min
Checking fmax_cmp_max_lt_start_neg_nan
exit 0

Actual output:

Checking fmaxnum_start_neg_2
Miscompare signed-zeros: -0 != 0
exit 1

The regression occured somewhere between clang 20 and clang 21:
https://godbolt.org/z/rnfqzc1Yj

@fhahn

[llvmbot]

@llvm/issue-subscribers-bug

Author: Michael Kruse (Meinersbur)

https://github.com/llvm/llvm-test-suite/blob/main/SingleSource/UnitTests/Vectorizer/fmax-reduction.cpp on x86-64

Expected output:

Checking fmaxnum_start_neg_2
Checking fmaxnum_start_min
Checking fmaxnum_start_denorm_min
Checking fmaxnum_start_is_nan
Checking fmax_non_strict_start_neg_2
Checking fmax_cmp_max_gt_start_neg_2
Checking fmax_cmp_max_lt_start_neg_2
Checking fmax_cmp_max_lt_start_denorm_min
Checking fmax_cmp_max_lt_start_neg_nan
exit 0

Actual output:

Checking fmaxnum_start_neg_2
Miscompare signed-zeros: -0 != 0
exit 1

The regression occured somehwere between clang 20 and clang 21:
https://godbolt.org/z/rnfqzc1Yj

@fhahn

[fhahn]

We now vectorize maxnum (std::fmax). I am not super familiar with X86 assembly, but it looks like the tail call <4 x float> @llvm.maxnum.v4f32(<4 x float> %a, <4 x float> %b) lowering is not handling signed zero correctly. According to LangRef, -0.0 < +0.0 for the intrinsic (https://llvm.org/docs/LangRef.html#llvm-maxnum-intrinsic)

But we generate MAXPS, which doesn't seem to match that behavior. From https://www.felixcloutier.com/x86/maxps

If the values being compared are both 0.0s (of either sign), the value in the second operand (source operand) is returned. If a value in the second operand is an SNaN, then SNaN is forwarded unchanged to the destination (that is, a QNaN version of the SNaN is not returned).

cc @RKSimon @arsenm

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Michael Kruse (Meinersbur)

https://github.com/llvm/llvm-test-suite/blob/main/SingleSource/UnitTests/Vectorizer/fmax-reduction.cpp on x86-64

Expected output:

Checking fmaxnum_start_neg_2
Checking fmaxnum_start_min
Checking fmaxnum_start_denorm_min
Checking fmaxnum_start_is_nan
Checking fmax_non_strict_start_neg_2
Checking fmax_cmp_max_gt_start_neg_2
Checking fmax_cmp_max_lt_start_neg_2
Checking fmax_cmp_max_lt_start_denorm_min
Checking fmax_cmp_max_lt_start_neg_nan
exit 0

Actual output:

Checking fmaxnum_start_neg_2
Miscompare signed-zeros: -0 != 0
exit 1

The regression occured somehwere between clang 20 and clang 21:
https://godbolt.org/z/rnfqzc1Yj

@fhahn

[fhahn]

Example for the lowering: https://clang.godbolt.org/z/esx98v1dh

[RKSimon]

Initial selection DAG: %bb.0 'foo:'
SelectionDAG has 10 nodes:
  t0: ch,glue = EntryToken
      t2: v4f32,ch = CopyFromReg t0, Register:v4f32 %0
      t4: v4f32,ch = CopyFromReg t0, Register:v4f32 %1
    t5: v4f32 = fmaxnum t2, t4
  t8: ch,glue = CopyToReg t0, Register:v4f32 $xmm0, t5
  t9: ch = X86ISD::RET_GLUE t8, TargetConstant:i32<0>, Register:v4f32 $xmm0, t8:1

I'm sorry but I haven't been keeping up with a lot of these changes - but if we need to handle signed zeros - shouldn't this be fmaxnum_ieee ?

[phoebewang]

I'm sorry but I haven't been keeping up with a lot of these changes - but if we need to handle signed zeros - shouldn't this be fmaxnum_ieee ?

No, llvm.maxnum.* distinguish +0 from -0, https://llvm.org/docs/LangRef.html#llvm-implementation. I'll take a look at this.

[fhahn]

great, thanks @RKSimon & @phoebewang