[SLP]Fix a crash on attempt to revectorize vectorized phi.

If the PHI node is vectorized during vectorization of its operands, no
need to try to vectorize its operands once again.

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -18732,6 +18732,11 @@ bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
         },
         /*MaxVFOnly=*/true, R);
     Changed |= HaveVectorizedPhiNodes;
+    if (HaveVectorizedPhiNodes && any_of(PHIToOpcodes, [&](const auto &P) {
+          auto *PHI = dyn_cast<PHINode>(P.first);
+          return !PHI || R.isDeleted(PHI);
+        }))
+      PHIToOpcodes.clear();
     VisitedInstrs.insert(Incoming.begin(), Incoming.end());
   } while (HaveVectorizedPhiNodes);
 
@@ -18804,7 +18809,7 @@ bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
       }
       // Try to vectorize the incoming values of the PHI, to catch reductions
       // that feed into PHIs.
-      for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; I++) {
+      for (unsigned I : seq<unsigned>(P->getNumIncomingValues())) {
         // Skip if the incoming block is the current BB for now. Also, bypass
         // unreachable IR for efficiency and to avoid crashing.
         // TODO: Collect the skipped incoming values and try to vectorize them
@@ -18816,9 +18821,16 @@ bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
         // Postponed instructions should not be vectorized here, delay their
         // vectorization.
         if (auto *PI = dyn_cast<Instruction>(P->getIncomingValue(I));
-            PI && !IsInPostProcessInstrs(PI))
-          Changed |= vectorizeRootInstruction(nullptr, PI,
+            PI && !IsInPostProcessInstrs(PI)) {
+          bool Res = vectorizeRootInstruction(nullptr, PI,
                                               P->getIncomingBlock(I), R, TTI);
+          Changed |= Res;
+          if (Res && R.isDeleted(P)) {
+            It = BB->begin();
+            E = BB->end();
+            break;
+          }
+        }
       }
       continue;
     }

llvm/test/Transforms/SLPVectorizer/X86/phi-removed-on-operand-vectorization.ll

@@ -0,0 +1,37 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S --passes=slp-vectorizer -mtriple=i386-pc-windows-msvc19.34.0 -mcpu=pentium4 < %s | FileCheck %s
+
+define i32 @test(double %mul321.i) {
+; CHECK-LABEL: define i32 @test(
+; CHECK-SAME: double [[MUL321_I:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <2 x double> poison, double [[MUL321_I]], i32 0
+; CHECK-NEXT:    br label %[[DO_BODY220_I:.*]]
+; CHECK:       [[DO_BODY220_I]]:
+; CHECK-NEXT:    [[TMP1:%.*]] = phi <2 x double> [ [[TMP6:%.*]], %[[DO_BODY221_I:.*]] ], [ zeroinitializer, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[DO_BODY221_I]]
+; CHECK:       [[DO_BODY221_I]]:
+; CHECK-NEXT:    [[TMP2:%.*]] = fadd <2 x double> [[TMP1]], zeroinitializer
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul <2 x double> [[TMP2]], zeroinitializer
+; CHECK-NEXT:    [[TMP4:%.*]] = extractelement <2 x double> [[TMP3]], i32 0
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <2 x double> [[TMP3]], i32 1
+; CHECK-NEXT:    [[ADD318_I:%.*]] = fadd double [[TMP4]], [[TMP5]]
+; CHECK-NEXT:    [[TMP6]] = insertelement <2 x double> [[TMP0]], double [[ADD318_I]], i32 1
+; CHECK-NEXT:    br label %[[DO_BODY220_I]]
+;
+entry:
+  br label %do.body220.i
+
+do.body220.i:
+  %c1.2.i = phi double [ %mul321.i, %do.body221.i ], [ 0.000000e+00, %entry ]
+  %s1.1.i = phi double [ %add318.i, %do.body221.i ], [ 0.000000e+00, %entry ]
+  br label %do.body221.i
+
+do.body221.i:                                     ; preds = %do.body220.i
+  %sub311.i1 = fadd double %c1.2.i, 0.000000e+00
+  %add315.i = fadd double %s1.1.i, 0.000000e+00
+  %mul316.i = fmul double %sub311.i1, 0.000000e+00
+  %mul317.i = fmul double %add315.i, 0.000000e+00
+  %add318.i = fadd double %mul316.i, %mul317.i
+  br label %do.body220.i
+}