Make SROA more aggressive with allocas containing padding.

SROA only split up structs and arrays one level at a time, so padding can
only cause trouble if it is located in between the struct or array elements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@123380 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index e0674492f892e678a69abe693297546b0fa03cbe..ec298e7316d80aabd3796a6022354a7e874447a7 100644 (file)
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -1676,46 +1676,39 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
 }
 
 /// HasPadding - Return true if the specified type has any structure or
-/// alignment padding, false otherwise.
+/// alignment padding in between the elements that would be split apart
+/// by SROA; return false otherwise.
 static bool HasPadding(const Type *Ty, const TargetData &TD) {
-  if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
-    return HasPadding(ATy->getElementType(), TD);
-  
-  if (const VectorType *VTy = dyn_cast<VectorType>(Ty))
-    return HasPadding(VTy->getElementType(), TD);
-  
-  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
-    const StructLayout *SL = TD.getStructLayout(STy);
-    unsigned PrevFieldBitOffset = 0;
-    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
-      unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
-
-      // Padding in sub-elements?
-      if (HasPadding(STy->getElementType(i), TD))
-        return true;
+  if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
+    Ty = ATy->getElementType();
+    return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
+  }
 
-      // Check to see if there is any padding between this element and the
-      // previous one.
-      if (i) {
-        unsigned PrevFieldEnd =
+  // SROA currently handles only Arrays and Structs.
+  const StructType *STy = cast<StructType>(Ty);
+  const StructLayout *SL = TD.getStructLayout(STy);
+  unsigned PrevFieldBitOffset = 0;
+  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
+    unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
+
+    // Check to see if there is any padding between this element and the
+    // previous one.
+    if (i) {
+      unsigned PrevFieldEnd =
         PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
-        if (PrevFieldEnd < FieldBitOffset)
-          return true;
-      }
-
-      PrevFieldBitOffset = FieldBitOffset;
-    }
-
-    //  Check for tail padding.
-    if (unsigned EltCount = STy->getNumElements()) {
-      unsigned PrevFieldEnd = PrevFieldBitOffset +
-                   TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
-      if (PrevFieldEnd < SL->getSizeInBits())
+      if (PrevFieldEnd < FieldBitOffset)
         return true;
     }
+    PrevFieldBitOffset = FieldBitOffset;
   }
-  
-  return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
+  // Check for tail padding.
+  if (unsigned EltCount = STy->getNumElements()) {
+    unsigned PrevFieldEnd = PrevFieldBitOffset +
+      TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
+    if (PrevFieldEnd < SL->getSizeInBits())
+      return true;
+  }
+  return false;
 }
 
 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of

diff --git a/test/Transforms/ScalarRepl/copy-aggregate.ll b/test/Transforms/ScalarRepl/copy-aggregate.ll
index 29924138762fc4e142b298f98e587d65b3d0fc46..6e01aa3c51840662a5beba95f2c343bbf2daa55d 100644 (file)
--- a/test/Transforms/ScalarRepl/copy-aggregate.ll
+++ b/test/Transforms/ScalarRepl/copy-aggregate.ll
@@ -1,9 +1,11 @@
-; RUN: opt < %s -scalarrepl -S | not grep alloca
+; RUN: opt < %s -scalarrepl -S | FileCheck %s
 ; PR3290
 target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64"
 
 ;; Store of integer to whole alloca struct.
 define i32 @test1(i64 %V) nounwind {
+; CHECK: test1
+; CHECK-NOT: alloca
        %X = alloca {{i32, i32}}
        %Y = bitcast {{i32,i32}}* %X to i64*
        store i64 %V, i64* %Y
@@ -18,6 +20,8 @@ define i32 @test1(i64 %V) nounwind {
 
 ;; Store of integer to whole struct/array alloca.
 define float @test2(i128 %V) nounwind {
+; CHECK: test2
+; CHECK-NOT: alloca
        %X = alloca {[4 x float]}
        %Y = bitcast {[4 x float]}* %X to i128*
        store i128 %V, i128* %Y
@@ -32,6 +36,8 @@ define float @test2(i128 %V) nounwind {
 
 ;; Load of whole alloca struct as integer
 define i64 @test3(i32 %a, i32 %b) nounwind {
+; CHECK: test3
+; CHECK-NOT: alloca
        %X = alloca {{i32, i32}}
 
        %A = getelementptr {{i32,i32}}* %X, i32 0, i32 0, i32 0
@@ -46,6 +52,8 @@ define i64 @test3(i32 %a, i32 %b) nounwind {
 
 ;; load of integer from whole struct/array alloca.
 define i128 @test4(float %a, float %b) nounwind {
+; CHECK: test4
+; CHECK-NOT: alloca
        %X = alloca {[4 x float]}
        %A = getelementptr {[4 x float]}* %X, i32 0, i32 0, i32 0
        %B = getelementptr {[4 x float]}* %X, i32 0, i32 0, i32 3
@@ -56,3 +64,22 @@ define i128 @test4(float %a, float %b) nounwind {
        %V = load i128* %Y
        ret i128 %V
 }
+
+;; If the elements of a struct or array alloca contain padding, SROA can still
+;; split up the alloca as long as there is no padding between the elements.
+%padded = type { i16, i8 }
+%arr = type [4 x %padded]
+define void @test5(%arr* %p, %arr* %q) {
+entry:
+; CHECK: test5
+; CHECK-NOT: i128
+  %var = alloca %arr, align 4
+  %vari8 = bitcast %arr* %var to i8*
+  %pi8 = bitcast %arr* %p to i8*
+  call void @llvm.memcpy.i32(i8* %vari8, i8* %pi8, i32 16, i32 4)
+  %qi8 = bitcast %arr* %q to i8*
+  call void @llvm.memcpy.i32(i8* %qi8, i8* %vari8, i32 16, i32 4)
+  ret void
+}
+
+declare void @llvm.memcpy.i32(i8* nocapture, i8* nocapture, i32, i32) nounwind