[Reassociation][PR21205] Don't canonicalize X - undef to X + (-undef).

[mcrosier]

Summary

This isn't going to expose additional opportunities for reassociation, AFAICT, and it's currently causing an assertion failure.
For further details, see: http://llvm.org/bugs/show_bug.cgi?id=21205
Please have a look.
Regards,

Chad

[mcrosier]

test/Transforms/Reassociate/pr21205.ll

1

I missed the FileCheck.. one moment.

[mcrosier]

With correct FileCheck.

[hfinkel]

LGTM.

[mcrosier]

Committed r219434.

lib/Transforms/Scalar/Reassociate.cpp

   if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
     return false;
 
+  // Don't breakup X - undef.
+  if (isa<UndefValue>(Sub->getOperand(1)))
+    return false;
+
   // Don't bother to break this up unless either the LHS is an associable add or
   // subtract or if this is only used by one.
   Value *V0 = Sub->getOperand(0);

test/Transforms/Reassociate/pr21205.ll

+; RUN: opt < %s -reassociate -disable-output
+; PR21205
+
+@a = common global i32 0, align 4
+@b = common global i32 0, align 4
+
+; Don't canonicalize %conv - undef into %conv + (-undef).
+; CHECK-LABEL: @test1
+; CHECK: %sub = fsub fast float %conv, undef
+; CHECK:  %sub1 = fadd fast float %sub, -1.000000e+00
+
+define i32 @test1() {
+entry:
+  %0 = load i32* @a, align 4
+  %conv = sitofp i32 %0 to float
+  %sub = fsub fast float %conv, undef
+  %sub1 = fadd fast float %sub, -1.000000e+00
+  %conv2 = fptosi float %sub1 to i32
+  store i32 %conv2, i32* @b, align 4
+  ret i32 undef
+}