[llvm] [InstCombine] Fold fcmp ogt (x - y), 0 into fcmp ogt x, y #85245 (PR #85506)

[via llvm-commits]

SahilPatidar wrote:

After updating the test, it is breaking `tail call fast <4 x float> @llvm.fabs.v4f32(<4 x float> [[TMP2]])`.
```llvm
define i32 @TestVectorsEqualFP(ptr noalias %Vec0, ptr noalias %Vec1, float %Tole
 ; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x float>, ptr [[VEC0:%.*]], align 4
 ; CHECK-NEXT:    [[TMP1:%.*]] = load <4 x float>, ptr [[VEC1:%.*]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = fsub fast <4 x float> [[TMP0]], [[TMP1]]
-; CHECK-NEXT:    [[TMP3:%.*]] = tail call fast <4 x float> @llvm.fabs.v4f32(<4 x float> [[TMP2]])
-; CHECK-NEXT:    [[TMP4:%.*]] = tail call fast float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[TMP3]])
-; CHECK-NEXT:    [[CMP4:%.*]] = fcmp fast ole float [[TMP4]], [[TOLERANCE:%.*]]
+; CHECK-NEXT:    [[TMP3:%.*]] = fcmp fast oge <4 x float> [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = fneg fast <4 x float> [[TMP2]]
+; CHECK-NEXT:    [[TMP5:%.*]] = select <4 x i1> [[TMP3]], <4 x float> [[TMP2]], <4 x float> [[TMP4]]
+; CHECK-NEXT:    [[TMP6:%.*]] = tail call fast float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[TMP5]])
+; CHECK-NEXT:    [[CMP4:%.*]] = fcmp fast ole float [[TMP6]], [[TOLERANCE:%.*]]
 ; CHECK-NEXT:    [[COND5:%.*]] = zext i1 [[CMP4]] to i32
 ; CHECK-NEXT:    ret i32 [[COND5]]
 ;

```

https://github.com/llvm/llvm-project/pull/85506