remove a pointless restriction from memcpyopt. It was

refusing to optimize two memcpy's like this:

copy A <- B
copy C <- A

if it couldn't prove that noalias(B,C).  We can eliminate
the copy by producing a memmove instead of memcpy.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@119694 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Transforms/Scalar/MemCpyOptimizer.cpp

@@ -688,11 +688,14 @@ bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
   if (DepSize < MSize)
     return false;
   
-  // Finally, we have to make sure that the dest of the second does not
+  Intrinsic::ID ResultFn = Intrinsic::memcpy;
-  // alias the source of the first.
+  
+  // If the dest of the second might alias the source of the first, then the
+  // source and dest might overlap.  We still want to eliminate the intermediate
+  // value, but we have to generate a memmove instead of memcpy.
   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
   if (!AA.isNoAlias(M->getRawDest(), MSize, MDep->getRawSource(), DepSize))
-    return false;
+    ResultFn = Intrinsic::memmove;
   
   // If all checks passed, then we can transform these memcpy's
   const Type *ArgTys[3] = {
@@ -702,7 +705,7 @@ bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
   };
   Function *MemCpyFun =
     Intrinsic::getDeclaration(M->getParent()->getParent()->getParent(),
-                              M->getIntrinsicID(), ArgTys, 3);
+                              ResultFn, ArgTys, 3);
   
   // Make sure to use the lesser of the alignment of the source and the dest
   // since we're changing where we're reading from, but don't want to increase

test/Transforms/MemCpyOpt/2008-04-29-SRetRemoval.ll

@@ -1,17 +0,0 @@
-; RUN: opt < %s -memcpyopt -S | grep {call.*memcpy.*agg.result}
-
-target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
-target triple = "i386-apple-darwin8"
-@x = external global { x86_fp80, x86_fp80 }		; <{ x86_fp80, x86_fp80 }*> [#uses=1]
-
-define void @foo({ x86_fp80, x86_fp80 }* noalias sret %agg.result) nounwind  {
-entry:
-	%x.0 = alloca { x86_fp80, x86_fp80 }		; <{ x86_fp80, x86_fp80 }*> [#uses=1]
-	%x.01 = bitcast { x86_fp80, x86_fp80 }* %x.0 to i8*		; <i8*> [#uses=2]
-	call void @llvm.memcpy.i32( i8* %x.01, i8* bitcast ({ x86_fp80, x86_fp80 }* @x to i8*), i32 32, i32 16 )
-	%agg.result2 = bitcast { x86_fp80, x86_fp80 }* %agg.result to i8*		; <i8*> [#uses=1]
-	call void @llvm.memcpy.i32( i8* %agg.result2, i8* %x.01, i32 32, i32 16 )
-	ret void
-}
-
-declare void @llvm.memcpy.i32(i8*, i8*, i32, i32) nounwind 

test/Transforms/MemCpyOpt/memcpy.ll

@@ -1,4 +1,4 @@
-; RUN: opt < %s -basicaa -memcpyopt -dse -S | grep {call.*memcpy} | count 1
+; RUN: opt < %s -basicaa -memcpyopt -dse -S | FileCheck %s
 
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
 target triple = "i686-apple-darwin9"
@@ -20,7 +20,7 @@ entry:
 
 ; CHECK: @test1
 ; CHECK: call void @ccoshl
-; CHECK: call @llvm.memcpy
+; CHECK: call void @llvm.memcpy
 ; CHECK-NOT: llvm.memcpy
 ; CHECK: ret void
 	ret void
@@ -29,3 +29,36 @@ entry:
 declare void @ccoshl({ x86_fp80, x86_fp80 }* sret , x86_fp80, x86_fp80) nounwind 
 
 declare void @llvm.memcpy.i32(i8*, i8*, i32, i32) nounwind 
+
+
+; The intermediate alloca and one of the memcpy's should be eliminated, the
+; other should be related with a memmove.
+define void @test2(i8* %P, i8* %Q) nounwind  {
+	%memtmp = alloca { x86_fp80, x86_fp80 }, align 16
+	%R = bitcast { x86_fp80, x86_fp80 }* %memtmp to i8*
+	call void @llvm.memcpy.i32( i8* %R, i8* %P, i32 32, i32 16 )
+	call void @llvm.memcpy.i32( i8* %Q, i8* %R, i32 32, i32 16 )
+        ret void
+        
+; CHECK: @test2
+; CHECK-NEXT: call void @llvm.memmove{{.*}}(i8* %Q, i8* %P
+; CHECK-NEXT: ret void
+}
+
+
+
+
+@x = external global { x86_fp80, x86_fp80 }
+
+define void @test3({ x86_fp80, x86_fp80 }* noalias sret %agg.result) nounwind  {
+	%x.0 = alloca { x86_fp80, x86_fp80 }
+	%x.01 = bitcast { x86_fp80, x86_fp80 }* %x.0 to i8*
+	call void @llvm.memcpy.i32( i8* %x.01, i8* bitcast ({ x86_fp80, x86_fp80 }* @x to i8*), i32 32, i32 16 )
+	%agg.result2 = bitcast { x86_fp80, x86_fp80 }* %agg.result to i8*
+	call void @llvm.memcpy.i32( i8* %agg.result2, i8* %x.01, i32 32, i32 16 )
+	ret void
+; CHECK: @test3
+; CHECK-NEXT: %agg.result2 = bitcast 
+; CHECK-NEXT: call void @llvm.memcpy
+; CHECK-NEXT: ret void
+}