From db9e697725e81edb4c5cb80f8dc7b412431be0d0 Mon Sep 17 00:00:00 2001
From: Richard Osborne <richard@xmos.com>
Date: Thu, 16 Jul 2009 12:50:48 +0000
Subject: [PATCH] Combine an unaligned store of unaligned load into a memmove.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75908 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/XCore/XCoreISelLowering.cpp        | 52 +++++++++++++++++++
 lib/Target/XCore/XCoreISelLowering.h          |  2 +
 test/CodeGen/XCore/unaligned_store_combine.ll | 12 +++++
 3 files changed, 66 insertions(+)
 create mode 100644 test/CodeGen/XCore/unaligned_store_combine.ll

diff --git a/lib/Target/XCore/XCoreISelLowering.cpp b/lib/Target/XCore/XCoreISelLowering.cpp
index 0f503bc7a64..113bfbf8dbe 100644
--- a/lib/Target/XCore/XCoreISelLowering.cpp
+++ b/lib/Target/XCore/XCoreISelLowering.cpp
@@ -153,6 +153,9 @@ XCoreTargetLowering::XCoreTargetLowering(XCoreTargetMachine &XTM)
 
   maxStoresPerMemset = 4;
   maxStoresPerMemmove = maxStoresPerMemcpy = 2;
+
+  // We have target-specific dag combine patterns for the following nodes:
+  setTargetDAGCombine(ISD::STORE);
 }
 
 SDValue XCoreTargetLowering::
@@ -1049,6 +1052,55 @@ XCoreTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   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
 //===----------------------------------------------------------------------===//
diff --git a/lib/Target/XCore/XCoreISelLowering.h b/lib/Target/XCore/XCoreISelLowering.h
index e7a48a97265..4b244660ccf 100644
--- a/lib/Target/XCore/XCoreISelLowering.h
+++ b/lib/Target/XCore/XCoreISelLowering.h
@@ -122,6 +122,8 @@ namespace llvm {
   
     // Expand specifics
     SDValue ExpandADDSUB(SDNode *Op, SelectionDAG &DAG);
+
+    virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   };
 }
 
diff --git a/test/CodeGen/XCore/unaligned_store_combine.ll b/test/CodeGen/XCore/unaligned_store_combine.ll
new file mode 100644
index 00000000000..b61cf86988a
--- /dev/null
+++ b/test/CodeGen/XCore/unaligned_store_combine.ll
@@ -0,0 +1,12 @@
+; 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
+}