AMDGPU: Make better use of op_sel with high components
Handle more general swizzles.
llvm-svn: 303296
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 286be35..4b27bba 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -2660,6 +2660,15 @@
SDValue Vec = Op.getOperand(0);
SDValue Idx = Op.getOperand(1);
+ DAGCombinerInfo DCI(DAG, AfterLegalizeVectorOps, true, nullptr);
+
+ // Make sure we we do any optimizations that will make it easier to fold
+ // source modifiers before obscuring it with bit operations.
+
+ // XXX - Why doesn't this get called when vector_shuffle is expanded?
+ if (SDValue Combined = performExtractVectorEltCombine(Op.getNode(), DCI))
+ return Combined;
+
if (const ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx)) {
SDValue Result = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Vec);