[SLP]Fix PR87358: broken module, Instruction does not dominate all uses.

If the first node is a gather node with extractelement instructions,
still need to put the vector value after all instructions, not after the
very first one.

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -10736,9 +10736,13 @@ Instruction &BoUpSLP::getLastInstructionInBundle(const TreeEntry *E) {
                 [](Value *V) {
                   return !isa<GetElementPtrInst>(V) && isa<Instruction>(V);
                 })) ||
-        all_of(E->Scalars, [](Value *V) {
-          return !isVectorLikeInstWithConstOps(V) && isUsedOutsideBlock(V);
-        }))
+        all_of(E->Scalars,
+               [](Value *V) {
+                 return !isVectorLikeInstWithConstOps(V) &&
+                        isUsedOutsideBlock(V);
+               }) ||
+        (E->State == TreeEntry::NeedToGather && E->Idx == 0 &&
+         all_of(E->Scalars, IsaPred<ExtractElementInst, UndefValue>)))
       Res.second = FindLastInst();
     else
       Res.second = FindFirstInst();

llvm/test/Transforms/SLPVectorizer/X86/extractlements-gathered-first-node.ll

@@ -0,0 +1,18 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S --passes=slp-vectorizer -slp-threshold=-99999 -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
+
+define void @test() {
+; CHECK-LABEL: define void @test() {
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    [[TMP0:%.*]] = extractelement <4 x i32> zeroinitializer, i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x i32> zeroinitializer, i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x i32> <i32 0, i32 undef>, i32 [[TMP1]], i32 1
+; CHECK-NEXT:    [[ICMP:%.*]] = icmp ult i32 [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    ret void
+;
+bb:
+  %0 = extractelement <4 x i32> zeroinitializer, i32 0
+  %1 = extractelement <2 x i32> zeroinitializer, i32 0
+  %icmp = icmp ult i32 %0, %1
+  ret void
+}