From 69ea9d230865425ed7422a5c5fe9dbb6ea612433 Mon Sep 17 00:00:00 2001
From: Chris Lattner <sabre@nondot.org>
Date: Wed, 30 Apr 2008 06:39:11 +0000
Subject: [PATCH] move lowering of llvm.memset -> store from simplify libcalls
 to instcombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50472 91177308-0d34-0410-b5e6-96231b3b80d8
---
 .../Scalar/InstructionCombining.cpp           | 56 ++++++++++++++++---
 test/Transforms/InstCombine/memset.ll         | 15 +++++
 2 files changed, 64 insertions(+), 7 deletions(-)
 create mode 100644 test/Transforms/InstCombine/memset.ll

diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp
index 82c67d03d93..507e6f2a854 100644
--- a/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -370,6 +370,7 @@ namespace {
     Instruction *MatchBSwap(BinaryOperator &I);
     bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
     Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
+    Instruction *SimplifyMemSet(MemSetInst *MI);
 
 
     Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
@@ -8491,7 +8492,9 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
   // A single load+store correctly handles overlapping memory in the memmove
   // case.
   unsigned Size = MemOpLength->getZExtValue();
-  if (Size == 0 || Size > 8 || (Size&(Size-1)))
+  if (Size == 0) return MI;  // Delete this mem transfer.
+  
+  if (Size > 8 || (Size&(Size-1)))
     return 0;  // If not 1/2/4/8 bytes, exit.
   
   // Use an integer load+store unless we can find something better.
@@ -8545,6 +8548,48 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
   return MI;
 }
 
+Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
+  unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
+  if (MI->getAlignment()->getZExtValue() < Alignment) {
+    MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
+    return MI;
+  }
+  
+  // Extract the length and alignment and fill if they are constant.
+  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
+  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
+  if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
+    return 0;
+  uint64_t Len = LenC->getZExtValue();
+  Alignment = MI->getAlignment()->getZExtValue();
+  
+  // If the length is zero, this is a no-op
+  if (Len == 0) return MI; // memset(d,c,0,a) -> noop
+  
+  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
+  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
+    const Type *ITy = IntegerType::get(Len*8);  // n=1 -> i8.
+    
+    Value *Dest = MI->getDest();
+    Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
+
+    // Alignment 0 is identity for alignment 1 for memset, but not store.
+    if (Alignment == 0) Alignment = 1;
+    
+    // Extract the fill value and store.
+    uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
+    InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
+                                      Alignment), *MI);
+    
+    // Set the size of the copy to 0, it will be deleted on the next iteration.
+    MI->setLength(Constant::getNullValue(LenC->getType()));
+    return MI;
+  }
+
+  return 0;
+}
+
+
 /// visitCallInst - CallInst simplification.  This mostly only handles folding 
 /// of intrinsic instructions.  For normal calls, it allows visitCallSite to do
 /// the heavy lifting.
@@ -8592,12 +8637,9 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
       if (Instruction *I = SimplifyMemTransfer(MI))
         return I;
-    } else if (isa<MemSetInst>(MI)) {
-      unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
-      if (MI->getAlignment()->getZExtValue() < Alignment) {
-        MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
-        Changed = true;
-      }
+    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
+      if (Instruction *I = SimplifyMemSet(MSI))
+        return I;
     }
           
     if (Changed) return II;
diff --git a/test/Transforms/InstCombine/memset.ll b/test/Transforms/InstCombine/memset.ll
new file mode 100644
index 00000000000..27a5b60cade
--- /dev/null
+++ b/test/Transforms/InstCombine/memset.ll
@@ -0,0 +1,15 @@
+; RUN: llvm-as < %s | opt -instcombine | llvm-dis | not grep {call.*llvm.memset}
+
+declare void @llvm.memset.i32(i8*, i8, i32, i32)
+
+define i32 @main() {
+	%target = alloca [1024 x i8]		; <[1024 x i8]*> [#uses=1]
+	%target_p = getelementptr [1024 x i8]* %target, i32 0, i32 0		; <i8*> [#uses=5]
+	call void @llvm.memset.i32( i8* %target_p, i8 1, i32 0, i32 1 )
+	call void @llvm.memset.i32( i8* %target_p, i8 1, i32 1, i32 1 )
+	call void @llvm.memset.i32( i8* %target_p, i8 1, i32 2, i32 2 )
+	call void @llvm.memset.i32( i8* %target_p, i8 1, i32 4, i32 4 )
+	call void @llvm.memset.i32( i8* %target_p, i8 1, i32 8, i32 8 )
+	ret i32 0
+}
+
-- 
2.34.1