[InstCombine] Only fold bitcast(fptrunc) if destination type matches fptrunc result type.

[vmustya]

It's not enough to just make sure destination type is floating point,
because the following chain may be incorrectly optimized:

  %trunc = fptrunc float %src to bfloat
  %cast = bitcast bfloat %trunc to half

Before the fix, the instruction sequence mentioned above used to be
translated into single fptrunc instruction as follows:

  %trunc = fptrunc float %src to half

Such transformation was semantically incorrect.

[github-actions]

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be
notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write
permissions for the repository. In which case you can instead tag reviewers by
name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review
by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate
is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

[llvmbot]

@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-llvm-transforms

Author: Victor Mustya (vmustya)

Changes

It's not enough to just make sure destination type is floating point,
because the following chain may be incorrectly optimized:

  %trunc = fptrunc float %src to bfloat
  %cast = bitcast bfloat %trunc to half

Before the fix, the instruction sequence mentioned above used to be
translated into single fptrunc instruction as follows:

  %trunc = fptrunc float %src to half

Such transformation was semantically incorrect.

---

Full diff: https://github.com/llvm/llvm-project/pull/77046.diff

2 Files Affected:

• (modified) llvm/lib/IR/Instructions.cpp (+2-2)
• (modified) llvm/test/Transforms/InstCombine/fptrunc.ll (+13)

diff --git a/llvm/lib/IR/Instructions.cpp b/llvm/lib/IR/Instructions.cpp
index 299b4e74677dcc..87874c3abc4680 100644
--- a/llvm/lib/IR/Instructions.cpp
+++ b/llvm/lib/IR/Instructions.cpp
@@ -3203,8 +3203,8 @@ unsigned CastInst::isEliminableCastPair(
       return 0;
     case 4:
       // No-op cast in second op implies firstOp as long as the DestTy
-      // is floating point.
-      if (DstTy->isFloatingPointTy())
+      // matches MidTy.
+      if (DstTy == MidTy)
         return firstOp;
       return 0;
     case 5:
diff --git a/llvm/test/Transforms/InstCombine/fptrunc.ll b/llvm/test/Transforms/InstCombine/fptrunc.ll
index d3e153f12106e0..c78df0b83d9cdf 100644
--- a/llvm/test/Transforms/InstCombine/fptrunc.ll
+++ b/llvm/test/Transforms/InstCombine/fptrunc.ll
@@ -190,3 +190,16 @@ define half @ItoFtoF_u25_f32_f16(i25 %i) {
   %r = fptrunc float %x to half
   ret half %r
 }
+
+; Negative test - bitcast bfloat to half is not optimized
+
+define half @fptrunc_to_bfloat_bitcast_to_half(float %src) {
+; CHECK-LABEL: @fptrunc_to_bfloat_bitcast_to_half(
+; CHECK-NEXT:    [[TRUNC:%.*]] = fptrunc float [[SRC:%.*]] to bfloat
+; CHECK-NEXT:    [[CAST:%.*]] = bitcast bfloat [[TRUNC]] to half
+; CHECK-NEXT:    ret half [[CAST]]
+;
+  %trunc = fptrunc float %src to bfloat
+  %cast = bitcast bfloat %trunc to half
+  ret half %cast
+}