maxStoresPerMemset = 4;
maxStoresPerMemmove = maxStoresPerMemcpy = 2;
+
+ // We have target-specific dag combine patterns for the following nodes:
+ setTargetDAGCombine(ISD::STORE);
}
SDValue XCoreTargetLowering::
return BB;
}
+//===----------------------------------------------------------------------===//
+// Target Optimization Hooks
+//===----------------------------------------------------------------------===//
+
+SDValue XCoreTargetLowering::PerformDAGCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+ SelectionDAG &DAG = DCI.DAG;
+ DebugLoc dl = N->getDebugLoc();
+ switch (N->getOpcode()) {
+ default: break;
+ case ISD::STORE: {
+ // Replace unaligned store of unaligned load with memmove.
+ StoreSDNode *ST = cast<StoreSDNode>(N);
+ if (!DCI.isBeforeLegalize() || allowsUnalignedMemoryAccesses() ||
+ ST->isVolatile() || ST->isIndexed()) {
+ break;
+ }
+ SDValue Chain = ST->getChain();
+
+ unsigned StoreBits = ST->getMemoryVT().getStoreSizeInBits();
+ if (StoreBits % 8) {
+ break;
+ }
+ unsigned ABIAlignment = getTargetData()->
+ getABITypeAlignment(ST->getMemoryVT().getTypeForMVT(*DAG.getContext()));
+ unsigned Alignment = ST->getAlignment();
+ if (Alignment >= ABIAlignment) {
+ break;
+ }
+
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(ST->getValue())) {
+ if (LD->hasNUsesOfValue(1, 0) && ST->getMemoryVT() == LD->getMemoryVT() &&
+ LD->getAlignment() == Alignment &&
+ !LD->isVolatile() && !LD->isIndexed() &&
+ Chain.reachesChainWithoutSideEffects(SDValue(LD, 1))) {
+ return DAG.getMemmove(Chain, dl, ST->getBasePtr(),
+ LD->getBasePtr(),
+ DAG.getConstant(StoreBits/8, MVT::i32),
+ Alignment, ST->getSrcValue(),
+ ST->getSrcValueOffset(), LD->getSrcValue(),
+ LD->getSrcValueOffset());
+ }
+ }
+ break;
+ }
+ }
+ return SDValue();
+}
+
//===----------------------------------------------------------------------===//
// Addressing mode description hooks
//===----------------------------------------------------------------------===//
--- /dev/null
+; RUN: llvm-as < %s | llc -march=xcore > %t1.s
+; RUN: grep "bl memmove" %t1.s | count 1
+; RUN: grep "ldc r., 8" %t1.s | count 1
+
+; Unaligned load / store pair. Should be combined into a memmove
+; of size 8
+define void @f(i64* %dst, i64* %src) nounwind {
+entry:
+ %0 = load i64* %src, align 1
+ store i64 %0, i64* %dst, align 1
+ ret void
+}