Fix the second half of PR7437: scalarrepl wasn't preserving

address spaces when SRoA'ing memcpy's.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107846 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 44 insertions, 21 deletions

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index 506e9823c0..dd445f6332 100644
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -1272,6 +1272,8 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
   // If there is an other pointer, we want to convert it to the same pointer
   // type as AI has, so we can GEP through it safely.
   if (OtherPtr) {
+    unsigned AddrSpace =
+      cast<PointerType>(OtherPtr->getType())->getAddressSpace();
 
     // Remove bitcasts and all-zero GEPs from OtherPtr.  This is an
     // optimization, but it's also required to detect the corner case where
@@ -1279,20 +1281,8 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
     // OtherPtr may be a bitcast or GEP that currently being rewritten.  (This
     // function is only called for mem intrinsics that access the whole
     // aggregate, so non-zero GEPs are not an issue here.)
-    while (1) {
-      if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) {
-        OtherPtr = BC->getOperand(0);
-        continue;
-      }
-      if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) {
-        // All zero GEPs are effectively bitcasts.
-        if (GEP->hasAllZeroIndices()) {
-          OtherPtr = GEP->getOperand(0);
-          continue;
-        }
-      }
-      break;
-    }
+    OtherPtr = OtherPtr->stripPointerCasts();
+    
     // Copying the alloca to itself is a no-op: just delete it.
     if (OtherPtr == AI || OtherPtr == NewElts[0]) {
       // This code will run twice for a no-op memcpy -- once for each operand.
@@ -1304,15 +1294,13 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
       return;
     }
     
-    if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
-      if (BCE->getOpcode() == Instruction::BitCast)
-        OtherPtr = BCE->getOperand(0);
-    
     // If the pointer is not the right type, insert a bitcast to the right
     // type.
-    if (OtherPtr->getType() != AI->getType())
-      OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
-                                 MI);
+    const Type *NewTy =
+      PointerType::get(AI->getType()->getElementType(), AddrSpace);
+    
+    if (OtherPtr->getType() != NewTy)
+      OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
   }
   
   // Process each element of the aggregate.
diff --git a/test/Transforms/ScalarRepl/address-space.ll b/test/Transforms/ScalarRepl/address-space.ll
new file mode 100644
index 0000000000..318d4e7590
--- /dev/null
+++ b/test/Transforms/ScalarRepl/address-space.ll
@@ -0,0 +1,35 @@
+; RUN: opt -S -scalarrepl < %s | FileCheck %s
+; PR7437 - Make sure SROA preserves address space of memcpy when
+; hacking on it.
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-darwin10"
+
+%struct.anon = type { [1 x float] }
+
+; CHECK: define void @Test(
+; CHECK: load float addrspace(2)*
+; CHECK-NEXT: fsub float
+; CHECK: store float {{.*}}, float addrspace(2)* 
+define void @Test(%struct.anon addrspace(2)* %pPtr) nounwind {
+entry:
+  %s = alloca %struct.anon, align 4               ; <%struct.anon*> [#uses=3]
+  %arrayidx = getelementptr inbounds %struct.anon addrspace(2)* %pPtr, i64 0 ; <%struct.anon addrspace(2)*> [#uses=1]
+  %tmp1 = bitcast %struct.anon* %s to i8*         ; <i8*> [#uses=1]
+  %tmp2 = bitcast %struct.anon addrspace(2)* %arrayidx to i8 addrspace(2)* ; <i8 addrspace(2)*> [#uses=1]
+  call void @llvm.memcpy.p0i8.p2i8.i64(i8* %tmp1, i8 addrspace(2)* %tmp2, i64 4, i32 4, i1 false)
+  %tmp3 = getelementptr inbounds %struct.anon* %s, i32 0, i32 0 ; <[1 x float]*> [#uses=1]
+  %arrayidx4 = getelementptr inbounds [1 x float]* %tmp3, i32 0, i64 0 ; <float*> [#uses=2]
+  %tmp5 = load float* %arrayidx4                  ; <float> [#uses=1]
+  %sub = fsub float %tmp5, 5.000000e+00           ; <float> [#uses=1]
+  store float %sub, float* %arrayidx4
+  %arrayidx7 = getelementptr inbounds %struct.anon addrspace(2)* %pPtr, i64 0 ; <%struct.anon addrspace(2)*> [#uses=1]
+  %tmp8 = bitcast %struct.anon addrspace(2)* %arrayidx7 to i8 addrspace(2)* ; <i8 addrspace(2)*> [#uses=1]
+  %tmp9 = bitcast %struct.anon* %s to i8*         ; <i8*> [#uses=1]
+  call void @llvm.memcpy.p2i8.p0i8.i64(i8 addrspace(2)* %tmp8, i8* %tmp9, i64 4, i32 4, i1 false)
+  ret void
+}
+
+declare void @llvm.memcpy.p0i8.p2i8.i64(i8* nocapture, i8 addrspace(2)* nocapture, i64, i32, i1) nounwind
+
+declare void @llvm.memcpy.p2i8.p0i8.i64(i8 addrspace(2)* nocapture, i8* nocapture, i64, i32, i1) nounwind
+