DAG combine incorrectly optimize (i32 vextract (v4i16 load $addr), c) to

(i16 load $addr+c*sizeof(i16)) and replace uses of (i32 vextract) with the
i16 load. It should issue an extload instead: (i32 extload $addr+c*sizeof(i16)).

rdar://11035895

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@152675 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index d93af729322e9baa0e97180e85dc9c89edd52e86..66c04b42a92aec2cd4f4e260df556167d54ea2e2 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -7202,6 +7202,11 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
     EVT ExtVT = VT.getVectorElementType();
     EVT LVT = ExtVT;
 
+    // If the result of load has to be truncated, then it's not necessarily
+    // profitable.
+    if (NVT.bitsLT(LVT))
+      return SDValue();
+
     if (InVec.getOpcode() == ISD::BITCAST) {
       // Don't duplicate a load with other uses.
       if (!InVec.hasOneUse())
@@ -7302,10 +7307,20 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
     // Note that this replacement assumes that the extractvalue is the only
     // use of the load; that's okay because we don't want to perform this
     // transformation in other cases anyway.
-    SDValue Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
-                               LN0->getPointerInfo().getWithOffset(PtrOff),
-                               LN0->isVolatile(), LN0->isNonTemporal(), 
-                               LN0->isInvariant(), Align);
+    SDValue Load;
+    if (NVT.bitsGT(LVT)) {
+      // If the result type of vextract is wider than the load, then issue an
+      // extending load instead.
+      ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
+        ? ISD::ZEXTLOAD : ISD::EXTLOAD;
+      Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
+                            NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
+                            LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
+    } else
+      Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
+                         LN0->getPointerInfo().getWithOffset(PtrOff),
+                         LN0->isVolatile(), LN0->isNonTemporal(), 
+                         LN0->isInvariant(), Align);
     WorkListRemover DeadNodes(*this);
     SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
     SDValue To[] = { Load.getValue(0), Load.getValue(1) };

diff --git a/test/CodeGen/ARM/2012-03-13-DAGCombineBug.ll b/test/CodeGen/ARM/2012-03-13-DAGCombineBug.ll
new file mode 100644 (file)
index 0000000..6d596df
--- /dev/null
+++ b/test/CodeGen/ARM/2012-03-13-DAGCombineBug.ll
@@ -0,0 +1,16 @@
+; RUN: llc < %s -mtriple=thumbv7-apple-ios | FileCheck %s
+; rdar://11035895
+
+; DAG combine incorrectly optimize (i32 vextract (v4i16 load $addr), c) to
+; (i16 load $addr+c*sizeof(i16)). It should have issued an extload instead. i.e.
+; (i32 extload $addr+c*sizeof(i16)
+define void @test_hi_short3(<3 x i16> * nocapture %srcA, <2 x i16> * nocapture %dst) nounwind {
+entry:
+; CHECK: ldrh [[REG:r[0-9]+]]
+; CHECK: strh [[REG]]
+  %0 = load <3 x i16> * %srcA, align 8
+  %1 = shufflevector <3 x i16> %0, <3 x i16> undef, <2 x i32> <i32 2, i32 undef>
+  store <2 x i16> %1, <2 x i16> * %dst, align 4
+  ret void
+}
+