Fix bug in DAGBuilder for getelementptr with expanded vector.

[aymanmus]

Summary

Replacing the usage of MVT with EVT in case the vector type is expanded.

[mkuper]

test/CodeGen/X86/gep-expanded-vector.ll

1 ↗

(On Diff #67314)

Why does this test require SKX? If we're crashing in the builder - because v64i64 isn't simple - this should crash regardless of feature-set. If you do need AVX512 of some sort, it would be better to specify it explicitly with -mattr.
On the other hand, you probably want to specify a triple - the issue is x86_64-only, right?

7 ↗

(On Diff #67314)

Can you modify the test so that it (a) produces some kind of result, and (b) we check that we get the right result (as opposed to just "not crashing")?

[aymanmus]

Updating the test to actually check that code is emitted with right address calculation.

[mkuper]

The test looks better now - but, as I wrote on the last iteration, please specify a triple (either as part of the IR, or, better yet, pass -mtriple to llc).
Otherwise, the test is dependent on the host triple - so you'll break every build that's not x86_64.

[aymanmus]

Specifying triple for test llc run.

[mkuper]

LGTM

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

if (StructType *StTy = dyn_cast<StructType>(*GTI)) {
       if (!CI && isa<ConstantDataVector>(Idx) &&
           cast<ConstantDataVector>(Idx)->getSplatValue())
         CI = cast<ConstantInt>(cast<ConstantDataVector>(Idx)->getSplatValue());
 
       if (CI) {
         if (CI->isZero())
           continue;
         APInt Offs = ElementSize * CI->getValue().sextOrTrunc(PtrSize);
+        LLVMContext &Context = *DAG.getContext();
         SDValue OffsVal = VectorWidth ?
-          DAG.getConstant(Offs, dl, MVT::getVectorVT(PtrTy, VectorWidth)) :
+          DAG.getConstant(Offs, dl, EVT::getVectorVT(Context, PtrTy, VectorWidth)) :
           DAG.getConstant(Offs, dl, PtrTy);
 
         // In an inbouds GEP with an offset that is nonnegative even when
         // interpreted as signed, assume there is no unsigned overflow.
         SDNodeFlags Flags;
         if (Offs.isNonNegative() && cast<GEPOperator>(I).isInBounds())
           Flags.setNoUnsignedWrap(true);
 

llvm/trunk/test/CodeGen/X86/gep-expanded-vector.ll

+; RUN: llc < %s -O2 -mattr=avx512f -mtriple=x86_64-unknown | FileCheck %s
+
+%struct.S1 = type { %struct.S1*, %struct.S1* }
+
+define %struct.S1** @malloc_init_state(<64 x %struct.S1**> %tmp, i32 %ind) {
+entry:
+  %Vec = getelementptr inbounds %struct.S1*, <64 x %struct.S1**> %tmp , i64 2
+  %ptr = extractelement <64 x %struct.S1**> %Vec, i32 %ind
+  ret %struct.S1** %ptr
+}
+
+; CHECK: .LCPI0_0:
+; CHECK: .quad 16
+; CHECK: vpbroadcastq    .LCPI0_0(%rip), [[Z1:%zmm[0-9]]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z2:%zmm[0-9]]], [[Z2]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z3:%zmm[0-9]]], [[Z3]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z4:%zmm[0-9]]], [[Z4]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z5:%zmm[0-9]]], [[Z5]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z6:%zmm[0-9]]], [[Z6]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z7:%zmm[0-9]]], [[Z7]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z8:%zmm[0-9]]], [[Z8]]
+; CHECK-NEXT: vpaddq  [[Z1]], [[Z9:%zmm[0-9]]], [[Z9]]
+
+