[PATCH] D137862: InstCombine: Fold some identities for canonicalize

[Sanjay Patel via Phabricator via llvm-commits]

spatel added inline comments.


================
Comment at: llvm/test/Transforms/InstCombine/canonicalize.ll:284
-; CHECK-NEXT:    [[CANON_X:%.*]] = call float @llvm.canonicalize.f32(float [[X:%.*]])
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq float [[CANON_X]], [[X]]
 ; CHECK-NEXT:    ret i1 [[CMP]]
----------------
This isn't testing what was intended. The operands are canonicalized by complexity before we reach the transform in this patch.

It needs an extra instruction to create the expected pattern:

```
declare float @gen_f32()

define i1 @canonicalize_oeq_arg_f32_commute() {
  %x = call float @gen_f32() ; thwart complexity-based canonicalization
  %canon.x = call float @llvm.canonicalize.f32(float %x)
  %cmp = fcmp oeq float %x, %canon.x
  ret i1 %cmp
}

```


================
Comment at: llvm/test/Transforms/InstCombine/canonicalize.ll:306
-; CHECK-NEXT:    [[CANON_X:%.*]] = call <2 x float> @llvm.canonicalize.v2f32(<2 x float> [[X:%.*]])
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq <2 x float> [[CANON_X]], [[X]]
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
----------------
Similar to above - the operands got commuted.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D137862/new/

https://reviews.llvm.org/D137862