Make CanXFormVExtractWithShuffleIntoLoad reject loads with multiple uses. Also make...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147428 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index baf39c270d0d90700c1825dcf6832e91b7b43953..2d0cd960394a4980eb7e0df7abe1fcbaaa2a647f 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -6268,31 +6268,27 @@ bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
   if (V.getOpcode() == ISD::BITCAST)
     V = V.getOperand(0);
 
-  if (ISD::isNormalLoad(V.getNode())) {
-    // Is the original load suitable?
-    LoadSDNode *LN0 = cast<LoadSDNode>(V);
-
-    // FIXME: avoid the multi-use bug that is preventing lots of
-    // of foldings to be detected, this is still wrong of course, but
-    // give the temporary desired behavior, and if it happens that
-    // the load has real more uses, during isel it will not fold, and
-    // will generate poor code.
-    if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
-      return false;
+  if (!ISD::isNormalLoad(V.getNode()))
+    return false;
 
-    if (!HasShuffleIntoBitcast)
-      return true;
+  // Is the original load suitable?
+  LoadSDNode *LN0 = cast<LoadSDNode>(V);
 
-    // If there's a bitcast before the shuffle, check if the load type and
-    // alignment is valid.
-    unsigned Align = LN0->getAlignment();
-    unsigned NewAlign =
-      TLI.getTargetData()->getABITypeAlignment(
-                                    VT.getTypeForEVT(*DAG.getContext()));
+  if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
+    return false;
 
-    if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
-      return false;
-  }
+  if (!HasShuffleIntoBitcast)
+    return true;
+
+  // If there's a bitcast before the shuffle, check if the load type and
+  // alignment is valid.
+  unsigned Align = LN0->getAlignment();
+  unsigned NewAlign =
+    TLI.getTargetData()->getABITypeAlignment(
+                                  VT.getTypeForEVT(*DAG.getContext()));
+
+  if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
+    return false;
 
   return true;
 }

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index d6ae3af0d4e94e62a676f23c9fe3f0a41a0ec045..4becf99bfa2b3151b609f3e288e9f30eef085c0e 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -1236,10 +1236,10 @@ let Predicates = [HasAVX] in {
 
   // Store patterns
   def : Pat<(store (f64 (vector_extract
-            (v2f64 (X86Unpckh VR128:$src, (undef))), (iPTR 0))), addr:$dst),
+            (v2f64 (X86Unpckh VR128:$src, VR128:$src)), (iPTR 0))), addr:$dst),
             (VMOVHPSmr addr:$dst, VR128:$src)>;
   def : Pat<(store (f64 (vector_extract
-            (v2f64 (X86Unpckh VR128:$src, (undef))), (iPTR 0))), addr:$dst),
+            (v2f64 (X86Unpckh VR128:$src, VR128:$src)), (iPTR 0))), addr:$dst),
             (VMOVHPDmr addr:$dst, VR128:$src)>;
 }
 
@@ -1259,7 +1259,7 @@ let Predicates = [HasSSE1] in {
 
   // Store patterns
   def : Pat<(store (f64 (vector_extract
-            (v2f64 (X86Unpckh VR128:$src, (undef))), (iPTR 0))), addr:$dst),
+            (v2f64 (X86Unpckh VR128:$src, VR128:$src)), (iPTR 0))), addr:$dst),
             (MOVHPSmr addr:$dst, VR128:$src)>;
 }
 
@@ -1279,7 +1279,7 @@ let Predicates = [HasSSE2] in {
 
   // Store patterns
   def : Pat<(store (f64 (vector_extract
-            (v2f64 (X86Unpckh VR128:$src, (undef))), (iPTR 0))),addr:$dst),
+            (v2f64 (X86Unpckh VR128:$src, VR128:$src)), (iPTR 0))),addr:$dst),
             (MOVHPDmr addr:$dst, VR128:$src)>;
 }