Constant prop fails on vector code

[llvmbot]

Bugzilla Link	8994
Resolution	FIXED
Resolved on	Jan 20, 2011 20:32
Version	trunk
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

The instruction combine pass carshes when trying to fold a bitcast from a vector type to a scalar type. Reproduce with:

define i48 @​test(<3 x i1> %icmp1, <3 x i1> %icmp2, i16 addrspace(1)* %dst) nounwind {
entry:
%select1 = select <3 x i1> %icmp1, <3 x i16> zeroinitializer, <3 x i16> <i16 -1, i16 -1, i16 -1>
%select2 = select <3 x i1> %icmp2, <3 x i16> zeroinitializer, <3 x i16> %select1
%tmp2 = bitcast <3 x i16> %select2 to i48
ret i48 %tmp2
}

here's a fix:

Index: InstructionCombining.cpp

--- InstructionCombining.cpp (revision xxx)
+++ InstructionCombining.cpp (revision xxx)
@@ -221,6 +221,12 @@
Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
// Don't modify shared select instructions
if (!SI->hasOneUse()) return 0;
+

• // we cannot fold bitcasts which transform a vector to scalars
• if ( BitCastInst* BI = dyn_cast(&Op) ) {

• if ( BI->getSrcTy()->isVectorTy() ) return 0;
  

• }
• Value *TV = SI->getOperand(1);
  Value *FV = SI->getOperand(2);

[llvmbot]

Your patch is good but overly aggressive: it would forbid optimizing a select between two <4 x i32>s that are then casted to a <4 x float>.

Fixed in http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20110117/115543.html . Thanks for the bug report, especially with testcase!