[llvm] r282341 - [X86] Teach combineShuffle to avoid creating floating point operations with integer types and integer operations with floating point types. Seems isOperationLegal lies for mismatched types and operations.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Sat Sep 24 16:42:49 2016
New Revision: 282341

URL: http://llvm.org/viewvc/llvm-project?rev=282341&view=rev
Log:
[X86] Teach combineShuffle to avoid creating floating point operations with integer types and integer operations with floating point types. Seems isOperationLegal lies for mismatched types and operations.

Fixes PR30511.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=282341&r1=282340&r2=282341&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sat Sep 24 16:42:49 2016
@@ -26424,13 +26424,18 @@ static SDValue combineShuffle(SDNode *N,
       bool CanFold = false;
       switch (Opcode) {
       default : break;
-      case ISD::ADD :
-      case ISD::FADD :
-      case ISD::SUB :
-      case ISD::FSUB :
-      case ISD::MUL :
-      case ISD::FMUL :
-        CanFold = true;
+      case ISD::ADD:
+      case ISD::SUB:
+      case ISD::MUL:
+        // isOperationLegal lies for integer ops on floating point types.
+        CanFold = VT.isInteger();
+        break;
+      case ISD::FADD:
+      case ISD::FSUB:
+      case ISD::FMUL:
+        // isOperationLegal lies for floating point ops on integer types.
+        CanFold = VT.isFloatingPoint();
+        break;
       }
 
       unsigned SVTNumElts = SVT.getVectorNumElements();