[SLP]Fix a crash for the replaced vectorized value.

If two nodes share the same value, which is replaced in one of the
nodes, need to automatically replace same value in all nodes. Btter to
use WeakTrackingVH for this to fix compiler crash.

Change-Id: I6724034c0540e0b933807a20126d60b7d9d2019c

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -2611,7 +2611,7 @@ class BoUpSLP {
     ValueList Scalars;
 
     /// The Scalars are vectorized into this value. It is initialized to Null.
-    Value *VectorizedValue = nullptr;
+    WeakTrackingVH VectorizedValue = nullptr;
 
     /// Do we need to gather this sequence or vectorize it
     /// (either with vector instruction or with scatter/gather

llvm/test/Transforms/SLPVectorizer/X86/postponed_gathers.ll

@@ -0,0 +1,75 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt < %s -passes=slp-vectorizer -slp-threshold=-10 -mtriple=x86_64-unknown-linux-gnu -S | FileCheck %s
+
+define void @foo() {
+; CHECK-LABEL: define void @foo() {
+; CHECK-NEXT:  bci_0:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr addrspace(1) null, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x i32> <i32 poison, i32 0>, i32 [[TMP0]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> poison, <2 x i32> zeroinitializer
+; CHECK-NEXT:    br label [[BCI_252:%.*]]
+; CHECK:       bci_252:
+; CHECK-NEXT:    [[TMP3:%.*]] = phi <2 x i32> [ zeroinitializer, [[BCI_0:%.*]] ], [ [[TMP16:%.*]], [[BCI_252_1:%.*]] ]
+; CHECK-NEXT:    [[TMP4:%.*]] = mul <2 x i32> zeroinitializer, [[TMP1]]
+; CHECK-NEXT:    [[TMP5:%.*]] = or <2 x i32> [[TMP3]], zeroinitializer
+; CHECK-NEXT:    [[TMP6:%.*]] = or <2 x i32> [[TMP2]], [[TMP4]]
+; CHECK-NEXT:    [[TMP7:%.*]] = or <2 x i32> [[TMP6]], zeroinitializer
+; CHECK-NEXT:    [[TMP8:%.*]] = or <2 x i32> [[TMP5]], [[TMP7]]
+; CHECK-NEXT:    [[TMP9:%.*]] = or <2 x i32> [[TMP8]], zeroinitializer
+; CHECK-NEXT:    br i1 false, label [[NOT_ZERO70:%.*]], label [[BCI_252_1]]
+; CHECK:       bci_252.1:
+; CHECK-NEXT:    [[TMP10:%.*]] = or <2 x i32> [[TMP2]], zeroinitializer
+; CHECK-NEXT:    [[TMP11:%.*]] = or <2 x i32> [[TMP9]], zeroinitializer
+; CHECK-NEXT:    [[TMP12:%.*]] = mul <2 x i32> [[TMP10]], zeroinitializer
+; CHECK-NEXT:    [[TMP13:%.*]] = or <2 x i32> [[TMP2]], [[TMP12]]
+; CHECK-NEXT:    [[TMP14:%.*]] = or <2 x i32> [[TMP13]], zeroinitializer
+; CHECK-NEXT:    [[TMP15:%.*]] = or <2 x i32> [[TMP11]], [[TMP14]]
+; CHECK-NEXT:    [[TMP16]] = or <2 x i32> [[TMP15]], zeroinitializer
+; CHECK-NEXT:    br label [[BCI_252]]
+; CHECK:       not_zero70:
+; CHECK-NEXT:    [[TMP17:%.*]] = phi <2 x i32> [ [[TMP9]], [[BCI_252]] ]
+; CHECK-NEXT:    ret void
+;
+bci_0:
+  %0 = load i32, ptr addrspace(1) null, align 8
+  br label %bci_252
+
+bci_252:
+  %1 = phi i32 [ 0, %bci_0 ], [ %20, %bci_252.1 ]
+  %2 = phi i32 [ 0, %bci_0 ], [ %15, %bci_252.1 ]
+  %3 = mul i32 %0, 0
+  %4 = or i32 %0, %3
+  %5 = or i32 %4, 0
+  %.neg91.neg = or i32 %2, 0
+  %.neg446 = or i32 %.neg91.neg, %5
+  %6 = or i32 %.neg446, 0
+  %7 = mul i32 0, 0
+  %8 = or i32 %0, %7
+  %9 = or i32 %8, 0
+  %.neg91.1.neg = or i32 %1, 0
+  %.neg448 = or i32 %.neg91.1.neg, %9
+  %10 = or i32 %.neg448, 0
+  br i1 false, label %not_zero70, label %bci_252.1
+
+bci_252.1:
+  %11 = or i32 %0, 0
+  %12 = mul i32 %11, 0
+  %13 = or i32 %0, %12
+  %14 = or i32 %13, 0
+  %.neg91.neg.1 = or i32 %6, 0
+  %.neg446.1 = or i32 %.neg91.neg.1, %14
+  %15 = or i32 %.neg446.1, 0
+  %16 = or i32 %0, 0
+  %17 = mul i32 %16, 0
+  %18 = or i32 %0, %17
+  %19 = or i32 %18, 0
+  %.neg91.1.neg.1 = or i32 %10, 0
+  %.neg448.1 = or i32 %.neg91.1.neg.1, %19
+  %20 = or i32 %.neg448.1, 0
+  br label %bci_252
+
+not_zero70:
+  %.lcssa546 = phi i32 [ %6, %bci_252 ]
+  %.lcssa545 = phi i32 [ %10, %bci_252 ]
+  ret void
+}