Allow SROA to split up an array of vectors into multiple vectors, even when the vectors are dynamically indexed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@158529 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 112 insertions, 3 deletions

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index a6c3a5e3e9..9fc4164cc0 100644
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -1642,6 +1642,8 @@ void SROA::isSafeGEP(GetElementPtrInst *GEPI,
   gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
   if (GEPIt == E)
     return;
+  bool NonConstant = false;
+  unsigned NonConstantIdxSize = 0;
 
   // Walk through the GEP type indices, checking the types that this indexes
   // into.
@@ -1651,15 +1653,30 @@ void SROA::isSafeGEP(GetElementPtrInst *GEPI,
       continue;
 
     ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
-    if (!IdxVal)
-      return MarkUnsafe(Info, GEPI);
+    if (!IdxVal) {
+      // Non constant GEPs are only a problem on arrays, structs, and pointers
+      // Vectors can be dynamically indexed.
+      // FIXME: Add support for dynamic indexing on arrays.  This should be
+      // ok on any subarrays of the alloca array, eg, a[0][i] is ok, but a[i][0]
+      // isn't.
+      if (!(*GEPIt)->isVectorTy())
+        return MarkUnsafe(Info, GEPI);
+      NonConstant = true;
+      NonConstantIdxSize = TD->getTypeAllocSize(*GEPIt);
+    }
   }
 
   // Compute the offset due to this GEP and check if the alloca has a
   // component element at that offset.
   SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
+  // If this GEP is non constant then the last operand must have been a
+  // dynamic index into a vector.  Pop this now as it has no impact on the
+  // constant part of the offset.
+  if (NonConstant)
+    Indices.pop_back();
   Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
-  if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
+  if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset,
+                        NonConstantIdxSize))
     MarkUnsafe(Info, GEPI);
 }
 
@@ -1961,6 +1978,13 @@ void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
                       SmallVector<AllocaInst*, 32> &NewElts) {
   uint64_t OldOffset = Offset;
   SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
+  // If the GEP was dynamic then it must have been a dynamic vector lookup.
+  // In this case, it must be the last GEP operand which is dynamic so keep that
+  // aside until we've found the constant GEP offset then add it back in at the
+  // end.
+  Value* NonConstantIdx = 0;
+  if (!GEPI->hasAllConstantIndices())
+    NonConstantIdx = Indices.pop_back_val();
   Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
 
   RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
@@ -1987,6 +2011,8 @@ void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
     uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
     NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
   }
+  if (NonConstantIdx)
+    NewArgs.push_back(NonConstantIdx);
   Instruction *Val = NewElts[Idx];
   if (NewArgs.size() > 1) {
     Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
diff --git a/test/Transforms/ScalarRepl/dynamic-vector-gep.ll b/test/Transforms/ScalarRepl/dynamic-vector-gep.ll
new file mode 100644
index 0000000000..48a0da9dab
--- /dev/null
+++ b/test/Transforms/ScalarRepl/dynamic-vector-gep.ll
@@ -0,0 +1,83 @@
+; 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-i64:32:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64"
+target triple = "x86_64-apple-darwin10.0.0"
+
+; CHECK: @test1
+; CHECK: %[[alloc0:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: %[[alloc1:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: store <4 x float> zeroinitializer, <4 x float>* %[[alloc0]]
+
+; Split the array but don't replace the memset with an insert
+; element as its not a constant offset.
+define float @test1(i32 %idx1, i32 %idx2) {
+entry:
+  %0 = alloca [4 x <4 x float>]
+  store [4 x <4 x float>] zeroinitializer, [4 x <4 x float>]* %0
+  %ptr1 = getelementptr [4 x <4 x float>]* %0, i32 0, i32 0, i32 %idx1
+  %cast = bitcast float* %ptr1 to i8*
+  call void @llvm.memset.p0i8.i32(i8* %cast, i8 0, i32 4, i32 4, i1 false)
+  %ptr2 = getelementptr [4 x <4 x float>]* %0, i32 0, i32 1, i32 %idx2
+  %ret = load float* %ptr2
+  ret float %ret
+}
+
+; CHECK: @test2
+; CHECK: %[[alloc:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: store <4 x float> zeroinitializer, <4 x float>* %[[alloc]]
+; CHECK: %ptr1 = getelementptr inbounds <4 x float>* %[[alloc]], i32 0, i32 %idx1
+; CHECK: store float 1.000000e+00, float* %ptr1
+; CHECK: %ptr2 = getelementptr inbounds <4 x float>* %[[alloc]], i32 0, i32 %idx2
+; CHECK: %ret = load float* %ptr2
+; CHECK: ret float %ret
+
+; Do SROA on the array when it has dynamic vector reads and writes.
+define float @test2(i32 %idx1, i32 %idx2) {
+entry:
+  %0 = alloca [4 x <4 x float>]
+  store [4 x <4 x float>] zeroinitializer, [4 x <4 x float>]* %0
+  %ptr1 = getelementptr [4 x <4 x float>]* %0, i32 0, i32 0, i32 %idx1
+  store float 1.0, float* %ptr1
+  %ptr2 = getelementptr [4 x <4 x float>]* %0, i32 0, i32 0, i32 %idx2
+  %ret = load float* %ptr2
+  ret float %ret
+}
+
+; CHECK: test3
+; CHECK: %0 = alloca [4 x <4 x float>]
+; CHECK-NOT: alloca
+
+; Don't do SROA on a dynamically indexed vector when it spans
+; more than one array element of the alloca array it is within.
+define float @test3(i32 %idx1, i32 %idx2) {
+entry:
+  %0 = alloca [4 x <4 x float>]
+  store [4 x <4 x float>] zeroinitializer, [4 x <4 x float>]* %0
+  %bigvec = bitcast [4 x <4 x float>]* %0 to <16 x float>*
+  %ptr1 = getelementptr <16 x float>* %bigvec, i32 0, i32 %idx1
+  store float 1.0, float* %ptr1
+  %ptr2 = getelementptr <16 x float>* %bigvec, i32 0, i32 %idx2
+  %ret = load float* %ptr2
+  ret float %ret
+}
+
+; CHECK: @test4
+; CHECK: %0 = alloca [4 x <4 x float>]
+; CHECK-NOT: alloca
+
+; Don't do SROA as the is a second dynamically indexed array
+; which may span multiple elements of the alloca.
+define float @test4(i32 %idx1, i32 %idx2) {
+entry:
+  %0 = alloca [4 x <4 x float>]
+  store [4 x <4 x float>] zeroinitializer, [4 x <4 x float>]* %0
+  %ptr1 = getelementptr [4 x <4 x float>]* %0, i32 0, i32 0, i32 %idx1
+  %ptr2 = bitcast float* %ptr1 to [1 x <2 x float>]*
+  %ptr3 = getelementptr [1 x <2 x float>]* %ptr2, i32 0, i32 0, i32 %idx1
+  store float 1.0, float* %ptr1
+  %ptr4 = getelementptr [4 x <4 x float>]* %0, i32 0, i32 0, i32 %idx2
+  %ret = load float* %ptr4
+  ret float %ret
+}
+
+declare void @llvm.memset.p0i8.i32(i8*, i8, i32, i32, i1)