[llvm] r222265 - InstCombine: Fix another infinite loop caused by visitFPTrunc
Steven Wu
stevenwu at apple.com
Thu Dec 11 14:13:29 PST 2014
Hi David
Here is another infinite loop fall out from this patch. Here is a test case:
define float @test(float %arg) {
entry:
%0 = fpext float %arg to double
%1 = frem double %0, 0xBAD
%2 = fptrunc double %1 to float
ret float %2
}
The test case has exactly the same pattern as the instruction generated from instcombine which suggests probably it shouldn’t generate a truncation of same type? Maybe even a smaller type?
Here is a patch. Let me know what you think.
diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp
index aba77bb..357bf24 100644
--- a/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1269,16 +1269,19 @@ Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
// type of OpI doesn't enter into things at all. We simply evaluate
// in whichever source type is larger, then convert to the
// destination type.
- if (LHSWidth < SrcWidth)
- LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
- else if (RHSWidth <= SrcWidth)
- RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
- if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
- Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
- if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
- RI->copyFastMathFlags(OpI);
- return CastInst::CreateFPCast(ExactResult, CI.getType());
+ if (SrcWidth != OpWidth) {
+ if (LHSWidth < SrcWidth)
+ LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
+ else if (RHSWidth <= SrcWidth)
+ RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
+ if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
+ Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
+ if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
+ RI->copyFastMathFlags(OpI);
+ return CastInst::CreateFPCast(ExactResult, CI.getType());
+ }
}
+ break;
}
Thanks
Steven
> On Nov 18, 2014, at 2:06 PM, David Majnemer <david.majnemer at gmail.com> wrote:
>
> majnemer
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141211/075c718a/attachment.html>
More information about the llvm-commits
mailing list