Extend memcpy expansion in Transform/Utils to handle wider operand types.

Adds loop expansions for known-size and unknown-sized memcpy calls, allowing the
target to provide the operand types through TTI callbacks. The default values
for the TTI callbacks use int8 operand types and matches the existing behaviour
if they aren't overridden by the target.

Differential revision: https://reviews.llvm.org/D32536

llvm-svn: 307346

2 files changed, 36 insertions, 12 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULowerIntrinsics.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerIntrinsics.cpp
index 846e7dff5f8..7e0e9802c0e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerIntrinsics.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerIntrinsics.cpp
@@ -10,6 +10,7 @@
 #include "AMDGPU.h"
 #include "AMDGPUSubtarget.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
@@ -34,9 +35,14 @@ public:
   AMDGPULowerIntrinsics() : ModulePass(ID) {}
 
   bool runOnModule(Module &M) override;
+  bool expandMemIntrinsicUses(Function &F);
   StringRef getPassName() const override {
     return "AMDGPU Lower Intrinsics";
   }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<TargetTransformInfoWrapperPass>();
+  }
 };
 
 }
@@ -55,7 +61,7 @@ static bool shouldExpandOperationWithSize(Value *Size) {
   return !CI || (CI->getZExtValue() > MaxStaticSize);
 }
 
-static bool expandMemIntrinsicUses(Function &F) {
+bool AMDGPULowerIntrinsics::expandMemIntrinsicUses(Function &F) {
   Intrinsic::ID ID = F.getIntrinsicID();
   bool Changed = false;
 
@@ -67,7 +73,10 @@ static bool expandMemIntrinsicUses(Function &F) {
     case Intrinsic::memcpy: {
       auto *Memcpy = cast<MemCpyInst>(Inst);
       if (shouldExpandOperationWithSize(Memcpy->getLength())) {
-        expandMemCpyAsLoop(Memcpy);
+        Function *ParentFunc = Memcpy->getParent()->getParent();
+        const TargetTransformInfo &TTI =
+            getAnalysis<TargetTransformInfoWrapperPass>().getTTI(*ParentFunc);
+        expandMemCpyAsLoop(Memcpy, TTI);
         Changed = true;
         Memcpy->eraseFromParent();
       }
diff --git a/llvm/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp b/llvm/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp
index 3be291b48b8..989f0a3aba2 100644
--- a/llvm/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp
@@ -14,6 +14,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "NVPTXLowerAggrCopies.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/StackProtector.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DataLayout.h"
@@ -42,6 +43,7 @@ struct NVPTXLowerAggrCopies : public FunctionPass {
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addPreserved<StackProtector>();
+    AU.addRequired<TargetTransformInfoWrapperPass>();
   }
 
   bool runOnFunction(Function &F) override;
@@ -61,6 +63,8 @@ bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
 
   const DataLayout &DL = F.getParent()->getDataLayout();
   LLVMContext &Context = F.getParent()->getContext();
+  const TargetTransformInfo &TTI =
+      getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
 
   // Collect all aggregate loads and mem* calls.
   for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
@@ -104,15 +108,26 @@ bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
     Value *SrcAddr = LI->getOperand(0);
     Value *DstAddr = SI->getOperand(1);
     unsigned NumLoads = DL.getTypeStoreSize(LI->getType());
-    Value *CopyLen = ConstantInt::get(Type::getInt32Ty(Context), NumLoads);
-
-    createMemCpyLoop(/* ConvertedInst */ SI,
-                     /* SrcAddr */ SrcAddr, /* DstAddr */ DstAddr,
-                     /* CopyLen */ CopyLen,
-                     /* SrcAlign */ LI->getAlignment(),
-                     /* DestAlign */ SI->getAlignment(),
-                     /* SrcIsVolatile */ LI->isVolatile(),
-                     /* DstIsVolatile */ SI->isVolatile());
+    ConstantInt *CopyLen =
+        ConstantInt::get(Type::getInt32Ty(Context), NumLoads);
+
+    if (!TTI.useWideIRMemcpyLoopLowering()) {
+      createMemCpyLoop(/* ConvertedInst */ SI,
+                       /* SrcAddr */ SrcAddr, /* DstAddr */ DstAddr,
+                       /* CopyLen */ CopyLen,
+                       /* SrcAlign */ LI->getAlignment(),
+                       /* DestAlign */ SI->getAlignment(),
+                       /* SrcIsVolatile */ LI->isVolatile(),
+                       /* DstIsVolatile */ SI->isVolatile());
+    } else {
+      createMemCpyLoopKnownSize(/* ConvertedInst */ SI,
+                                /* SrcAddr */ SrcAddr, /* DstAddr */ DstAddr,
+                                /* CopyLen */ CopyLen,
+                                /* SrcAlign */ LI->getAlignment(),
+                                /* DestAlign */ SI->getAlignment(),
+                                /* SrcIsVolatile */ LI->isVolatile(),
+                                /* DstIsVolatile */ SI->isVolatile(), TTI);
+    }
 
     SI->eraseFromParent();
     LI->eraseFromParent();
@@ -121,7 +136,7 @@ bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
   // Transform mem* intrinsic calls.
   for (MemIntrinsic *MemCall : MemCalls) {
     if (MemCpyInst *Memcpy = dyn_cast<MemCpyInst>(MemCall)) {
-      expandMemCpyAsLoop(Memcpy);
+      expandMemCpyAsLoop(Memcpy, TTI);
     } else if (MemMoveInst *Memmove = dyn_cast<MemMoveInst>(MemCall)) {
       expandMemMoveAsLoop(Memmove);
     } else if (MemSetInst *Memset = dyn_cast<MemSetInst>(MemCall)) {