[SLP] Fix infinite loop in isUndefVector.

[fhahn]

Summary

This fixes an infinite loop if isa<T>(II->getOperand(1)) is true.
Return an all-0 bitvector, which is done along other error paths instead
of the continue, which leads to an infinite loop because Base
is not updated.

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

386–387

Move Base = II->getOperand(0); before the if statement instead, it is not an error path

[fhahn]

Keep continue, move updating Base to top of loop.

[fhahn]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

386–387

Updated, it looks like we are still missing test coverage for cases where it actually makes a difference though.

[ABataev]

LG

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

static SmallBitVector isUndefVector(const Value *V,
   auto *VecTy = dyn_cast<FixedVectorType>(V->getType());
   if (!VecTy)
     return Res.reset();
   auto *C = dyn_cast<Constant>(V);
   if (!C) {
     if (!UseMask.empty()) {
       const Value *Base = V;
       while (auto *II = dyn_cast<InsertElementInst>(Base)) {
+        Base = II->getOperand(0);
         if (isa<T>(II->getOperand(1)))
           continue;
-        Base = II->getOperand(0);
         std::optional<unsigned> Idx = getInsertIndex(II);
         if (!Idx)
           continue;
         if (*Idx < UseMask.size() && !UseMask.test(*Idx))
           Res.reset(*Idx);
       }
       // TODO: Add analysis for shuffles here too.
       if (V == Base) {
         Res.reset();

llvm/test/Transforms/SLPVectorizer/AArch64/insertelement.ll

;
   %f1 = tail call fast float @llvm.fabs.f32(float undef)
   %i0 = insertelement <vscale x 2 x float> undef, float %f0, i32 0
   %i1 = insertelement <vscale x 2 x float> %i0, float %f1, i32 1
   ret <vscale x 2 x float> %i1
 }
 
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare float @llvm.fabs.f32(float)
+
+
+define <4 x float> @insertelement_poison_lanes(ptr %0) {
+; CHECK-LABEL: @insertelement_poison_lanes(
+; CHECK-NEXT:    [[INS_1:%.*]] = insertelement <4 x float> zeroinitializer, float poison, i64 0
+; CHECK-NEXT:    [[INS_2:%.*]] = insertelement <4 x float> [[INS_1]], float 0.000000e+00, i64 0
+; CHECK-NEXT:    [[GEP_1:%.*]] = getelementptr double, ptr [[TMP0:%.*]], i64 1
+; CHECK-NEXT:    store <2 x double> <double 0.000000e+00, double 1.000000e+00>, ptr [[GEP_1]], align 8
+; CHECK-NEXT:    ret <4 x float> [[INS_2]]
+;
+  %trunc.1 = fptrunc double 0.000000e+00 to float
+  %trunc.2 = fptrunc double 1.000000e+00 to float
+  %ins.1 = insertelement <4 x float> zeroinitializer, float poison, i64 0
+  %ins.2 = insertelement <4 x float> %ins.1, float %trunc.1, i64 0
+  %ext.1 = fpext float %trunc.1 to double
+  %gep.1  = getelementptr double, ptr %0, i64 1
+  store double %ext.1, ptr %gep.1, align 8
+  %ext.2 = fpext float %trunc.2 to double
+  %gep.2 = getelementptr double, ptr %0, i64 2
+  store double %ext.2, ptr %gep.2, align 8
+  ret <4 x float> %ins.2
+}