[LLVMdev] Bug in InstCombiner::FoldAndOfFCmps

[Cameron McInally]

Hey guys,

I've come across a bug when combining an AND of FCMPs. This bug occurs in
my compiler as well as Clang built from trunk.

A reduced test case is:

int foo( float a, double b ) {
  return (a == a) & (b == b);
}

and the error is:

Assertion failed: (getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"), function
FCmpInst, file [snip]/clang/llvm/include/llvm/IR/Instructions.h, line 1117.


We begin with a float "a" and double "b":

> (fcmp oeq a, a) & (fcmp oeq b, b)

After canonicalization in InstCombine, we end up with the expression:

> (fcmp ord a, 0.0) & (fcmp ord b, 0.0)

The bug manifest is in InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst
*RHS), when trying to perform this transformation:

> (fcmp ord x, c) & (fcmp ord y, c)  -> (fcmp ord x, y)

This is all fine, except that the FCmpInst constructor in the IRBuilder
requires that both operands to the new fcmp are the same type, which they
are not.


Any suggestions on a proper fix for this issue? Some initial thoughts...

1) The transformation could be suppressed if the operand types do not match.

2) Cast the smaller operand to the type of the larger operand. Are there
any unforeseen implications here?

In any case, I suspect that there may be other transformations that need
similar checks also. If there is a clear solution, I will be happy to
submit a patch. Otherwise, I will file a Bug.

Thanks in advance,
Cameron
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130411/d5f2025b/attachment.html>