Only convert allocas to scalars if it is profitable. The profitability metric I

chose is having a non-memcpy/memset use and being larger than any native integer
type. Originally I chose having an access of a size smaller than the total size
of the alloca, but this caused some minor issues on the spirit benchmark where
SRoA runs again after some inlining.

This fixes <rdar://problem/8613163>.

llvm-svn: 127718

2 files changed, 43 insertions, 2 deletions

diff --git a/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index f70da11e9a5..3e7c640b9a8 100644
--- a/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -238,10 +238,15 @@ class ConvertToScalarInfo {
   /// also declared as a vector, we do want to promote to a vector.
   bool HadAVector;
 
+  /// HadAVector - True if there is at least one access to the alloca that is
+  /// not a MemTransferInst.  We don't want to turn structs into large integers
+  /// unless there is some potential for optimization.
+  bool HadNonMemTransferAccess;
+
 public:
   explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
     : AllocaSize(Size), TD(td), IsNotTrivial(false), VectorTy(0),
-      HadAVector(false) { }
+      HadAVector(false), HadNonMemTransferAccess(false) { }
 
   AllocaInst *TryConvert(AllocaInst *AI);
 
@@ -280,9 +285,14 @@ AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
           << *VectorTy << '\n');
     NewTy = VectorTy;  // Use the vector type.
   } else {
+    unsigned BitWidth = AllocaSize * 8;
+    if (!HadAVector && !HadNonMemTransferAccess &&
+        !TD.fitsInLegalInteger(BitWidth))
+      return 0;
+
     DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
     // Create and insert the integer alloca.
-    NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
+    NewTy = IntegerType::get(AI->getContext(), BitWidth);
   }
   AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
   ConvertUsesToScalar(AI, NewAI, 0);
@@ -431,6 +441,7 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
       // Don't touch MMX operations.
       if (LI->getType()->isX86_MMXTy())
         return false;
+      HadNonMemTransferAccess = true;
       MergeInType(LI->getType(), Offset);
       continue;
     }
@@ -441,6 +452,7 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
       // Don't touch MMX operations.
       if (SI->getOperand(0)->getType()->isX86_MMXTy())
         return false;
+      HadNonMemTransferAccess = true;
       MergeInType(SI->getOperand(0)->getType(), Offset);
       continue;
     }
@@ -465,6 +477,7 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
       if (!CanConvertToScalar(GEP, Offset+GEPOffset))
         return false;
       IsNotTrivial = true;  // Can't be mem2reg'd.
+      HadNonMemTransferAccess = true;
       continue;
     }
 
@@ -476,6 +489,7 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
           !isa<ConstantInt>(MSI->getLength()))
         return false;
       IsNotTrivial = true;  // Can't be mem2reg'd.
+      HadNonMemTransferAccess = true;
       continue;
     }
 
diff --git a/llvm/test/Transforms/ScalarRepl/only-memcpy-uses.ll b/llvm/test/Transforms/ScalarRepl/only-memcpy-uses.ll
new file mode 100644
index 00000000000..cfb88bd80d6
--- /dev/null
+++ b/llvm/test/Transforms/ScalarRepl/only-memcpy-uses.ll
@@ -0,0 +1,27 @@
+; RUN: opt < %s -scalarrepl -S | FileCheck %s
+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-darwin10.0.0"
+
+%struct.S = type { [12 x i32] }
+
+; CHECK: @bar4
+define void @bar4(%struct.S* byval %s) nounwind ssp {
+entry:
+; CHECK: alloca
+; CHECK-NOT: load
+; CHECK: memcpy
+  %t = alloca %struct.S, align 4
+  %agg.tmp = alloca %struct.S, align 4
+  %tmp = bitcast %struct.S* %t to i8*
+  %tmp1 = bitcast %struct.S* %s to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp, i8* %tmp1, i64 48, i32 4, i1 false)
+  %tmp2 = bitcast %struct.S* %agg.tmp to i8*
+  %tmp3 = bitcast %struct.S* %t to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp2, i8* %tmp3, i64 48, i32 4, i1 false)
+  %call = call i32 (...)* @bazz(%struct.S* byval %agg.tmp)
+  ret void
+}
+
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind
+
+declare i32 @bazz(...)