diff options
| -rw-r--r-- | lib/Transforms/Scalar/ScalarReplAggregates.cpp | 66 | ||||
| -rw-r--r-- | test/Transforms/ScalarRepl/2008-08-22-out-of-range-array-promote.ll | 22 |
2 files changed, 53 insertions, 35 deletions
diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp index 4e289e3d69..c5ca22145e 100644 --- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp +++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp @@ -511,42 +511,12 @@ void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, bool IsAllZeroIndices = true; - // If this is a use of an array allocation, do a bit more checking for sanity. + // If the first index is a non-constant index into an array, see if we can + // handle it as a special case. if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { - uint64_t NumElements = AT->getNumElements(); - - if (ConstantInt *Idx = dyn_cast<ConstantInt>(I.getOperand())) { - IsAllZeroIndices &= Idx->isZero(); - - // Check to make sure that index falls within the array. If not, - // something funny is going on, so we won't do the optimization. - // - if (Idx->getZExtValue() >= NumElements) - return MarkUnsafe(Info); - - // We cannot scalar repl this level of the array unless any array - // sub-indices are in-range constants. In particular, consider: - // A[0][i]. We cannot know that the user isn't doing invalid things like - // allowing i to index an out-of-range subscript that accesses A[1]. - // - // Scalar replacing *just* the outer index of the array is probably not - // going to be a win anyway, so just give up. - for (++I; I != E && (isa<ArrayType>(*I) || isa<VectorType>(*I)); ++I) { - uint64_t NumElements; - if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*I)) - NumElements = SubArrayTy->getNumElements(); - else - NumElements = cast<VectorType>(*I)->getNumElements(); - - ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); - if (!IdxVal) return MarkUnsafe(Info); - if (IdxVal->getZExtValue() >= NumElements) - return MarkUnsafe(Info); - IsAllZeroIndices &= IdxVal->isZero(); - } - - } else { + if (!isa<ConstantInt>(I.getOperand())) { IsAllZeroIndices = 0; + uint64_t NumElements = AT->getNumElements(); // If this is an array index and the index is not constant, we cannot // promote... that is unless the array has exactly one or two elements in @@ -560,7 +530,33 @@ void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, return MarkUnsafe(Info); } } - + + + // Walk through the GEP type indices, checking the types that this indexes + // into. + for (; I != E; ++I) { + // Ignore struct elements, no extra checking needed for these. + if (isa<StructType>(*I)) + continue; + + // Don't SROA pointers into vectors. + if (isa<VectorType>(*I)) + return MarkUnsafe(Info); + + // Otherwise, we must have an index into an array type. Verify that this is + // an in-range constant integer. Specifically, consider A[0][i]. We + // cannot know that the user isn't doing invalid things like allowing i to + // index an out-of-range subscript that accesses A[1]. Because of this, we + // have to reject SROA of any accesses into structs where any of the + // components are variables. + ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); + if (!IdxVal) return MarkUnsafe(Info); + if (IdxVal->getZExtValue() >= cast<ArrayType>(*I)->getNumElements()) + return MarkUnsafe(Info); + + IsAllZeroIndices &= IdxVal->isZero(); + } + // If there are any non-simple uses of this getelementptr, make sure to reject // them. return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); diff --git a/test/Transforms/ScalarRepl/2008-08-22-out-of-range-array-promote.ll b/test/Transforms/ScalarRepl/2008-08-22-out-of-range-array-promote.ll new file mode 100644 index 0000000000..a2386fdedc --- /dev/null +++ b/test/Transforms/ScalarRepl/2008-08-22-out-of-range-array-promote.ll @@ -0,0 +1,22 @@ +; RUN: llvm-as < %s | opt -scalarrepl | llvm-dis | grep {s = alloca .struct.x} +; PR2423 +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" + %struct.x = type { [1 x i32], i32, i32 } + +define i32 @b() nounwind { +entry: + %s = alloca %struct.x ; <%struct.x*> [#uses=2] + %r = alloca %struct.x ; <%struct.x*> [#uses=2] + call i32 @a( %struct.x* %s ) nounwind ; <i32>:0 [#uses=0] + %r1 = bitcast %struct.x* %r to i8* ; <i8*> [#uses=1] + %s2 = bitcast %struct.x* %s to i8* ; <i8*> [#uses=1] + call void @llvm.memcpy.i32( i8* %r1, i8* %s2, i32 12, i32 8 ) + getelementptr %struct.x* %r, i32 0, i32 0, i32 1 ; <i32*>:1 [#uses=1] + load i32* %1, align 4 ; <i32>:2 [#uses=1] + ret i32 %2 +} + +declare i32 @a(%struct.x*) + +declare void @llvm.memcpy.i32(i8*, i8*, i32, i32) nounwind |