[CodeGen] Fix crash in CodeGenPrepare::optimizeGatherScatterInst.

[craig.topper]

llvm/test/CodeGen/Generic/masked-gather-struct-gep.ll

3 ↗

(On Diff #524435)

Can we just put the test in the X86 directory?

[jcranmer-intel]

llvm/test/CodeGen/Generic/masked-gather-struct-gep.ll

3 ↗

(On Diff #524435)

Sure, I just put it there because CodeGenPrepare isn't really target-specific, and I was cribbing the requires line from another test in this directory.

[jcranmer-intel]

Move test to CodeGen/X86

[craig.topper]

llvm/test/CodeGen/X86/masked-gather-struct-gep.ll

2 ↗

(On Diff #524447)

Thinking about it again, don't we have test/Transforms/CodeGenPrepare/X86 ?

[craig.topper]

llvm/test/CodeGen/X86/masked-gather-struct-gep.ll

22 ↗

(On Diff #524447)

Should we return %gather so its not trivially dead?

[jcranmer-intel]

Move test again, add return.

[craig.topper]

LGTM

llvm/lib/CodeGen/CodeGenPrepare.cpp

if (!Ops[FinalIndex]->getType()->isVectorTy()) {
           Builder.CreateGEP(SecondTy, NewAddr, Constant::getNullValue(IndexTy));
     } else {
       Value *Base = Ops[0];
       Value *Index = Ops[FinalIndex];
 
       // Create a scalar GEP if there are more than 2 operands.
       if (Ops.size() != 2) {
         // Replace the last index with 0.
-        Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy);
+        Ops[FinalIndex] =
+            Constant::getNullValue(Ops[FinalIndex]->getType()->getScalarType());
         Base = Builder.CreateGEP(SourceTy, Base, ArrayRef(Ops).drop_front());
         SourceTy = GetElementPtrInst::getIndexedType(
             SourceTy, ArrayRef(Ops).drop_front());
       }
 
       // Now create the GEP with scalar pointer and vector index.
       NewAddr = Builder.CreateGEP(SourceTy, Base, Index);
     }

llvm/test/Transforms/CodeGenPrepare/X86/masked-gather-struct-gep.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt -S -codegenprepare < %s | FileCheck %s
+; REQUIRES: x86-registered-target
+target triple = "x86_64-pc-linux"
+
+%s = type <{ float, i32, i8, [3 x i8] }>
+
+declare <4 x float> @llvm.masked.gather.v4f32.v4p0(<4 x ptr>, i32 immarg, <4 x i1>, <4 x float>)
+
+define <4 x float> @foo(ptr %p) {
+; CHECK-LABEL: define <4 x float> @foo
+; CHECK-SAME: (ptr [[P:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast ptr [[P]] to ptr
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, ptr [[TMP1]], <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[GATHER:%.*]] = call <4 x float> @llvm.masked.gather.v4f32.v4p0(<4 x ptr> [[TMP2]], i32 0, <4 x i1> zeroinitializer, <4 x float> zeroinitializer)
+; CHECK-NEXT:    ret <4 x float> [[GATHER]]
+;
+  %base.splatinsert = insertelement <4 x ptr> poison, ptr %p, i32 0
+  %base = shufflevector <4 x ptr> %base.splatinsert, <4 x ptr> poison, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
+  %gep = getelementptr %s, <4 x ptr> %base, <4 x i64> zeroinitializer, <4 x i32> zeroinitializer
+  %gather = call <4 x float> @llvm.masked.gather.v4f32.v4p0(<4 x ptr> %gep, i32 0, <4 x i1> zeroinitializer, <4 x float> zeroinitializer)
+  ret <4 x float> %gather
+}