[ConstantFold] Fix incorrect index rewrites for GEPs

[anemet]

Summary

If an index for a vector or array type is out-of-range GEP constant
folding tries to factor it into preceding dimensions. The code however
does not consider addressing of structure field padding which should not
qualify as out-of-range index.

As demonstrated by the testcase, this can occur if the indexing
performed on a vector type and the preceding index is an array type.

SROA generates GEPs for example involving padding bytes as it slices an
alloca.

My fix disables this folding if the element type is a vector type. I
believe that this is the only way we can end up with padding. (We have
no access to DataLayout so I am not sure if there is actual robust way
of actually checking the presence of padding.)

[majnemer]

LGTM

lib/IR/ConstantFold.cpp

2196–2197 ↗

(On Diff #58568)

These conditions can now be merged.

2198 ↗

(On Diff #58568)

cast<SequentialType> can now be cast<ArrayType>

[anemet]

lib/IR/ConstantFold.cpp

2196–2197 ↗

(On Diff #58568)

Thanks! I'll do this in a separate commit because clang-format creates a big churn.

llvm/trunk/lib/IR/ConstantFold.cpp

static Constant *ConstantFoldGetElementPtrImpl(Type *PointeeTy, Constant *C,
   // factored out into preceding dimensions.
   SmallVector<Constant *, 8> NewIdxs;
   Type *Ty = PointeeTy;
   Type *Prev = C->getType();
   bool Unknown = !isa<ConstantInt>(Idxs[0]);
   for (unsigned i = 1, e = Idxs.size(); i != e;
        Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
     if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
-      if (isa<ArrayType>(Ty) || isa<VectorType>(Ty))
+      if (isa<ArrayType>(Ty))
         if (CI->getSExtValue() > 0 &&
             !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) {
           if (isa<SequentialType>(Prev)) {
             // It's out of range, but we can factor it into the prior
             // dimension.
             NewIdxs.resize(Idxs.size());
             uint64_t NumElements = 0;
             if (auto *ATy = dyn_cast<ArrayType>(Ty))

llvm/trunk/test/Transforms/InstCombine/getelementptr-folding.ll

+; RUN: opt -instcombine -S < %s | FileCheck %s
+
+%struct.matrix_float3x3 = type { [3 x <3 x float>] }
+
+; We used to fold this by rewriting the indices to 0, 0, 2, 0.  This is
+; invalid because there is a 4-byte padding after each <3 x float> field.
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.11.0"
+
+@matrix_identity_float3x3 = external global %struct.matrix_float3x3, align 16
+@bbb = global float* getelementptr inbounds (%struct.matrix_float3x3, %struct.matrix_float3x3* @matrix_identity_float3x3, i64 0, i32 0, i64 1, i64 3)
+; CHECK: @bbb = global float* getelementptr inbounds (%struct.matrix_float3x3, %struct.matrix_float3x3* @matrix_identity_float3x3, i64 0, i32 0, i64 1, i64 3)