[CodeGenPrepare] Move extractelement close to store if they can be combined.

[oota-llvm.git] / lib / Target / ARM / ARMISelLowering.cpp

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index df3b016b3ac7d26a3e8dffafe79115480d5aa469..fc7a02f5e6bd645d7df61c45f0a187b480f05404 100644 (file)
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -11123,6 +11123,35 @@ bool ARMTargetLowering::useLoadStackGuardNode() const {
   return Subtarget->getTargetTriple().getObjectFormat() == Triple::MachO;
 }
 
+bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
+                                                  unsigned &Cost) const {
+  // If we do not have NEON, vector types are not natively supported.
+  if (!Subtarget->hasNEON())
+    return false;
+
+  // Floating point values and vector values map to the same register file.
+  // Therefore, althought we could do a store extract of a vector type, this is
+  // better to leave at float as we have more freedom in the addressing mode for
+  // those.
+  if (VectorTy->isFPOrFPVectorTy())
+    return false;
+
+  // If the index is unknown at compile time, this is very expensive to lower
+  // and it is not possible to combine the store with the extract.
+  if (!isa<ConstantInt>(Idx))
+    return false;
+
+  assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
+  unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
+  // We can do a store + vector extract on any vector that fits perfectly in a D
+  // or Q register.
+  if (BitWidth == 64 || BitWidth == 128) {
+    Cost = 0;
+    return true;
+  }
+  return false;
+}
+
 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
                                          AtomicOrdering Ord) const {
   Module *M = Builder.GetInsertBlock()->getParent()->getParent();