[IROutliner][IRSim] Add all outlined region basic blocks to canonical numbering to add basic blocks to generated PHINode numbering generation.

[AndrewLitteken]

Summary

Issue: https://github.com/llvm/llvm-project/issues/54431

PHINodes that need to be generated to accommodate a PHINode outside the region due to different output paths need to have their own numbering to determine the number of output schemes required to properly handle all the outlined regions. This numbering was previously only determined by the order and values of the incoming values, as well as the parent block of the PHINode. This adds the incoming blocks to the calculation of a hash value for these PHINodes as well, and the supporting infrastructure to give each block in a region a corresponding canonical numbering.

[paquette]

llvm/lib/Analysis/IRSimilarityIdentifier.cpp

1025

1032

I think we can have more descriptive names for the variables here?

1034

1039

Can we use a more descriptive name than V?

1046

Nit: We can use the ternary operator for the assignment here.

Value *V = BB == getStartBB() ? frontInstruction() : &BB->front();

1047

I think we can be more descriptive than This<Thing> for the names here. Is there a way to link the name back to what you described in the comment above?

llvm/lib/Transforms/IPO/IROutliner.cpp

1192

nit: Maybe in a follow-up, it'd be good to rename OGVN to something like CanonicalNum, just to be clearer as to what it really is.

[paquette]

I think this is fine.

[paquette]

I applied this patch locally and tried to run the test suite at -O2 with it with asserts enabled. It fell over building fpcmp.

Assertion failed: (V != nullptr && "Value is a nullptr?"), function getGVN, file IRSimilarityIdentifier.h, line 896.

Reduced reproducer here: https://godbolt.org/z/4jW3KsEns

[AndrewLitteken]

Reduced reproducer here: https://godbolt.org/z/4jW3KsEns

I tried to replicate this on my side, but I can't. I applied the patch posted here and applied it to the current main branch as well rather than using my branch and still couldn't reproduce it on a standard debug build of Clang. I also tried to build the test suite with -O2 and the outliner with and without the cost model on, but still couldn't reproduce the bug shown here :/. I'm not sure if I'm missing something, or what I am doing differently.

[paquette]

You know what, I think I wonked something up somehow on my end, because I rebuilt + reran the test suite overnight with -Oz + -O2 and it seems fine.

I think this is good to go!

[bjope]

Seen assertion failures downstream with this patch. Here is a reproducer:

; ModuleID = 'irsimilarity_crash.ll'
source_filename = "irsimilarity_crash.ll"

@v_13 = external dso_local global ptr, align 1

; Function Attrs: nocallback nofree nosync nounwind readnone speculatable willreturn
declare void @llvm.dbg.declare(metadata, metadata, metadata) #0

define dso_local i16 @main() {
entry:
  %indirect-arg-temp537 = alloca { float, float }, align 1
  %0 = load ptr, ptr @v_13, align 1
  br label %for.cond108

for.cond108:                                      ; preds = %for.body110, %entry
  br i1 undef, label %for.body110, label %for.end122

for.body110:                                      ; preds = %for.cond108
  br label %for.cond108

for.end122:                                       ; preds = %for.cond108
  call void @llvm.dbg.declare(metadata ptr undef, metadata !1, metadata !DIExpression()), !dbg !11
  store i32 30, ptr undef, align 1
  br label %for.cond123

for.cond123:                                      ; preds = %for.end122
  br i1 undef, label %for.body125, label %for.end146

for.body125:                                      ; preds = %for.cond123
  unreachable

for.end146:                                       ; preds = %for.cond123
  br label %for.cond148

for.cond148:                                      ; preds = %for.end146
  br label %for.cond167

for.cond167:                                      ; preds = %for.body169, %for.cond148
  br i1 undef, label %for.body169, label %for.end246

for.body169:                                      ; preds = %for.cond167
  br label %for.cond167

for.end246:                                       ; preds = %for.cond167
  store i32 0, ptr undef, align 1
  unreachable
}

attributes #0 = { nocallback nofree nosync nounwind readnone speculatable willreturn }

!llvm.dbg.cu = !{}
!llvm.module.flags = !{!0}

!0 = !{i32 2, !"Debug Info Version", i32 3}
!1 = !DILocalVariable(name: "v_68", scope: !2, file: !3, line: 522, type: !10)
!2 = distinct !DILexicalBlock(scope: !4, file: !3, line: 522, column: 9)
!3 = !DIFile(filename: "41097217.c", directory: "rt.outdir")
!4 = distinct !DISubprogram(name: "main", scope: !3, file: !3, line: 480, type: !5, scopeLine: 481, spFlags: DISPFlagDefinition, unit: !8, retainedNodes: !9)
!5 = !DISubroutineType(types: !6)
!6 = !{!7}
!7 = !DIBasicType(name: "int", size: 16, encoding: DW_ATE_signed)
!8 = distinct !DICompileUnit(language: DW_LANG_C99, file: !3, producer: "clang version 15.0.0.prerel", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, retainedTypes: !9, globals: !9, splitDebugInlining: false, nameTableKind: None)
!9 = !{}
!10 = !DIBasicType(name: "long", size: 32, encoding: DW_ATE_signed)
!11 = !DILocation(line: 522, column: 23, scope: !2)

Hits an assertion if using a build with asserts when running

opt -passes='require<ir-similarity>' -S -o - irsimilarity_crash.ll

Part of stack trace:

opt: ../include/llvm/ADT/Optional.h:195: T &llvm::optional_detail::OptionalStorage<unsigned int, true>::getValue() & [T = unsigned int]: Assertion `hasVal' failed.
Stack dump:
0.      Program arguments: build-all/bin/opt -passes=require<ir-similarity> -S -o - irsimilarity_crash.ll
 #0 0x0000000002f2be08 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (build-all/bin/opt+0x2f2be08)
 #1 0x0000000002f29a7e llvm::sys::RunSignalHandlers() (build-all/bin/opt+0x2f29a7e)
 #2 0x0000000002f2c4a6 SignalHandler(int) Signals.cpp:0:0
 #3 0x00007f76ba600630 __restore_rt sigaction.c:0:0
 #4 0x00007f76b7d47387 raise (/lib64/libc.so.6+0x36387)
 #5 0x00007f76b7d48a78 abort (/lib64/libc.so.6+0x37a78)
 #6 0x00007f76b7d401a6 __assert_fail_base (/lib64/libc.so.6+0x2f1a6)
 #7 0x00007f76b7d40252 (/lib64/libc.so.6+0x2f252)
 #8 0x0000000001d57f8c llvm::IRSimilarity::IRSimilarityCandidate::createCanonicalRelationFrom(llvm::IRSimilarity::IRSimilarityCandidate&, llvm::DenseMap<unsigned int, llvm::DenseSet<unsigned int, llvm::DenseMapInfo<unsigned int, void> >, llvm::DenseMapInfo<unsigned int, void>, llvm::detail::DenseMapPair<unsigned int, llvm::DenseSet<unsigned int, llvm::DenseMapInfo<unsigned int, void> > > >&, llvm::DenseMap<unsigned int, llvm::DenseSet<unsigned int, llvm::DenseMapInfo<unsigned int, void> >, llvm::DenseMapInfo<unsigned int, void>, llvm::detail::DenseMapPair<unsigned int, llvm::DenseSet<unsigned int, llvm::DenseMapInfo<unsigned int, void> > > >&) (build-all/bin/opt+0x1d57f8c)
 #9 0x0000000001d59141 llvm::IRSimilarity::IRSimilarityIdentifier::findCandidates(std::vector<llvm::IRSimilarity::IRInstructionData*, std::allocator<llvm::IRSimilarity::IRInstructionData*> >&, std::vector<unsigned int, std::allocator<unsigned int> >&) (build-all/bin/opt+0x1d59141)
#10 0x0000000001d59fbc llvm::IRSimilarity::IRSimilarityIdentifier::findSimilarity(llvm::Module&) (build-all/bin/opt+0x1d59fbc)

[AndrewLitteken]

In D122207#3455485, @bjope wrote:

Seen assertion failures downstream with this patch. Here is a reproducer:

Here is a patch that fixes this issue.

[uabelho]

Hi!

The following starts to crash with this patch, and it still crashes on current top of tree 036aeac36c:

opt -passes='require<ir-similarity>' -o /dev/null bbi-68950.ll

We get

opt: ../include/llvm/Support/Casting.h:269: typename cast_retty<X, Y *>::ret_type llvm::cast(Y *) [X = llvm::Instruction, Y = llvm::Value]: Assertion `isa<X>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: build-all/bin/opt -passes=require<ir-similarity> -o /dev/null /home/uabelho/bbi-68950.ll
Stack dump without symbol names (ensure you have llvm-symbolizer in your PATH or set the environment var `LLVM_SYMBOLIZER_PATH` to point to it):
build-all/bin/opt(_ZN4llvm3sys15PrintStackTraceERNS_11raw_ostreamEi+0x23)[0x2cbf503]
build-all/bin/opt(_ZN4llvm3sys17RunSignalHandlersEv+0xee)[0x2cbd17e]
build-all/bin/opt[0x2cbf886]
/lib64/libpthread.so.0(+0xf630)[0x7f7bebb82630]
/lib64/libc.so.6(gsignal+0x37)[0x7f7be92c9387]
/lib64/libc.so.6(abort+0x148)[0x7f7be92caa78]
/lib64/libc.so.6(+0x2f1a6)[0x7f7be92c21a6]
/lib64/libc.so.6(+0x2f252)[0x7f7be92c2252]
build-all/bin/opt(_ZN4llvm12IRSimilarity21IRSimilarityCandidate27createCanonicalRelationFromERS1_RNS_8DenseMapIjNS_8DenseSetIjNS_12DenseMapInfoIjvEEEES6_NS_6detail12DenseMapPairIjS7_EEEESC_+0x11e6)[0x1b835d6]
build-all/bin/opt(_ZN4llvm12IRSimilarity22IRSimilarityIdentifier14findCandidatesERSt6vectorIPNS0_17IRInstructionDataESaIS4_EERS2_IjSaIjEE+0xa51)[0x1b84631]
build-all/bin/opt(_ZN4llvm12IRSimilarity22IRSimilarityIdentifier14findSimilarityERNS_6ModuleE+0xec)[0x1b854ac]
build-all/bin/opt(_ZN4llvm20IRSimilarityAnalysis3runERNS_6ModuleERNS_15AnalysisManagerIS1_JEEE+0x181)[0x1b85a81]
build-all/bin/opt[0x2fe4320]
build-all/bin/opt(_ZN4llvm15AnalysisManagerINS_6ModuleEJEE13getResultImplEPNS_11AnalysisKeyERS1_+0x20e)[0x245fb1e]
build-all/bin/opt[0x2ff1258]
build-all/bin/opt[0x2ff109d]
build-all/bin/opt(_ZN4llvm11PassManagerINS_6ModuleENS_15AnalysisManagerIS1_JEEEJEE3runERS1_RS3_+0x208)[0x245d3d8]
build-all/bin/opt(_ZN4llvm15runPassPipelineENS_9StringRefERNS_6ModuleEPNS_13TargetMachineEPNS_21TargetLibraryInfoImplEPNS_14ToolOutputFileES8_S8_S0_NS_8ArrayRefIS0_EENS9_INS_10PassPluginEEENS_8opt_tool10OutputKindENSD_12VerifierKindEbbbbb+0x3f9f)[0x78657f]
build-all/bin/opt(main+0x3fad)[0x7972cd]
/lib64/libc.so.6(__libc_start_main+0xf5)[0x7f7be92b5555]
build-all/bin/opt[0x780ec0]

bbi-68950.ll544 BDownload

[uabelho]

In D122207#3466709, @uabelho wrote:

Hi!

The following starts to crash with this patch, and it still crashes on current top of tree 036aeac36c:

opt -passes='require<ir-similarity>' -o /dev/null bbi-68950.ll

We get

opt: ../include/llvm/Support/Casting.h:269: typename cast_retty<X, Y *>::ret_type llvm::cast(Y *) [X = llvm::Instruction, Y = llvm::Value]: Assertion `isa<X>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: build-all/bin/opt -passes=require<ir-similarity> -o /dev/null /home/uabelho/bbi-68950.ll
Stack dump without symbol names (ensure you have llvm-symbolizer in your PATH or set the environment var `LLVM_SYMBOLIZER_PATH` to point to it):
build-all/bin/opt(_ZN4llvm3sys15PrintStackTraceERNS_11raw_ostreamEi+0x23)[0x2cbf503]
build-all/bin/opt(_ZN4llvm3sys17RunSignalHandlersEv+0xee)[0x2cbd17e]
build-all/bin/opt[0x2cbf886]
/lib64/libpthread.so.0(+0xf630)[0x7f7bebb82630]
/lib64/libc.so.6(gsignal+0x37)[0x7f7be92c9387]
/lib64/libc.so.6(abort+0x148)[0x7f7be92caa78]
/lib64/libc.so.6(+0x2f1a6)[0x7f7be92c21a6]
/lib64/libc.so.6(+0x2f252)[0x7f7be92c2252]
build-all/bin/opt(_ZN4llvm12IRSimilarity21IRSimilarityCandidate27createCanonicalRelationFromERS1_RNS_8DenseMapIjNS_8DenseSetIjNS_12DenseMapInfoIjvEEEES6_NS_6detail12DenseMapPairIjS7_EEEESC_+0x11e6)[0x1b835d6]
build-all/bin/opt(_ZN4llvm12IRSimilarity22IRSimilarityIdentifier14findCandidatesERSt6vectorIPNS0_17IRInstructionDataESaIS4_EERS2_IjSaIjEE+0xa51)[0x1b84631]
build-all/bin/opt(_ZN4llvm12IRSimilarity22IRSimilarityIdentifier14findSimilarityERNS_6ModuleE+0xec)[0x1b854ac]
build-all/bin/opt(_ZN4llvm20IRSimilarityAnalysis3runERNS_6ModuleERNS_15AnalysisManagerIS1_JEEE+0x181)[0x1b85a81]
build-all/bin/opt[0x2fe4320]
build-all/bin/opt(_ZN4llvm15AnalysisManagerINS_6ModuleEJEE13getResultImplEPNS_11AnalysisKeyERS1_+0x20e)[0x245fb1e]
build-all/bin/opt[0x2ff1258]
build-all/bin/opt[0x2ff109d]
build-all/bin/opt(_ZN4llvm11PassManagerINS_6ModuleENS_15AnalysisManagerIS1_JEEEJEE3runERS1_RS3_+0x208)[0x245d3d8]
build-all/bin/opt(_ZN4llvm15runPassPipelineENS_9StringRefERNS_6ModuleEPNS_13TargetMachineEPNS_21TargetLibraryInfoImplEPNS_14ToolOutputFileES8_S8_S0_NS_8ArrayRefIS0_EENS9_INS_10PassPluginEEENS_8opt_tool10OutputKindENSD_12VerifierKindEbbbbb+0x3f9f)[0x78657f]
build-all/bin/opt(main+0x3fad)[0x7972cd]
/lib64/libc.so.6(__libc_start_main+0xf5)[0x7f7be92b5555]
build-all/bin/opt[0x780ec0]

bbi-68950.ll544 BDownload

@AndrewLitteken: I wrote an issue about this new crash here:
https://github.com/llvm/llvm-project/issues/55131

[uabelho]

llvm/lib/Analysis/IRSimilarityIdentifier.cpp

1049

Hi @AndrewLitteken ,

The following command crashes on this cast:

opt -passes="function(loop-mssa(loop-rotate,licm<allowspeculation>,simple-loop-unswitch<nontrivial;trivial>)),require<ir-similarity>" bbi-80271.ll -o /dev/null

When it crashes SourceV is this:

(gdb) p SourceV->dump()
i16 1

The input is reduced with llvm-reduce and I've tried to shorten the command with utils/reduce_pipeline.py but it's still rather large and ugly.

bbi-80271.ll4 KBDownload

[AndrewLitteken]

llvm/lib/Analysis/IRSimilarityIdentifier.cpp

1049

Can you check if this patch: https://reviews.llvm.org/D139337 fixes the issue? This was fixed in a future refactoring for compilation speed and seems to work on my build.

[AndrewLitteken]

llvm/lib/Analysis/IRSimilarityIdentifier.cpp

1049

Sorry, this is not correct, it should be this patch: https://reviews.llvm.org/D139336

[uabelho]

llvm/lib/Analysis/IRSimilarityIdentifier.cpp

1049

Yes, with https://reviews.llvm.org/D139336 the crash I see goes away.

llvm/lib/Analysis/IRSimilarityIdentifier.cpp

for (unsigned Loc = StartIdx; Loc < StartIdx + Len; Loc++, ID++) {
     }
   }
 
   // Setting the first and last instruction data pointers for the candidate.  If
   // we got through the entire for loop without hitting an assert, we know
   // that both of these instructions are not nullptrs.
   FirstInst = FirstInstIt;
   LastInst = LastInstIt;
+
+  // Add the basic blocks contained in the set into the global value numbering.
+  DenseSet<BasicBlock *> BBSet;
+  getBasicBlocks(BBSet);
+  for (BasicBlock *BB : BBSet) {
+    if (ValueToNumber.find(BB) != ValueToNumber.end())
+      continue;
+    
+    ValueToNumber.try_emplace(BB, LocalValNumber);
+    NumberToValue.try_emplace(LocalValNumber, BB);
+    LocalValNumber++;
+  }
 }
 
 bool IRSimilarityCandidate::isSimilar(const IRSimilarityCandidate &A,
                                       const IRSimilarityCandidate &B) {
   if (A.getLength() != B.getLength())
     return false;
 
   auto InstrDataForBoth =
for (std::pair<unsigned, DenseSet<unsigned>> &GVNMapping : ToSourceMapping) {
 
     // Whatever GVN is found, we mark it as used.
     UsedGVNs.insert(ResultGVN);
 
     unsigned CanonNum = *SourceCand.getCanonicalNum(ResultGVN);
     CanonNumToNumber.insert(std::make_pair(CanonNum, SourceGVN));
     NumberToCanonNum.insert(std::make_pair(SourceGVN, CanonNum));
   }
+
+  DenseSet<BasicBlock *> BBSet;
+  getBasicBlocks(BBSet);
+  // Find canonical numbers for the BasicBlocks in the current candidate.
+  // This is done by finding the corresponding value for the first instruction
+  // in the block in the current candidate, finding the matching value in the
+  // source candidate.  Then by finding the parent of this value, use the
+  // canonical number of the block in the source candidate for the canonical
+  // number in the current candidate.
+  for (BasicBlock *BB : BBSet) {
+    unsigned BBGVNForCurrCand = ValueToNumber.find(BB)->second;
+
+    // We can skip the BasicBlock if the canonical numbering has already been
+    // found in a separate instruction.
+    if (NumberToCanonNum.find(BBGVNForCurrCand) != NumberToCanonNum.end())
+      continue;
+
+    // If the basic block is the starting block, then the shared instruction may
+    // not be the first instruction in the block, it will be the first
+    // instruction in the similarity region.
+    Value *FirstOutlineInst =
+        BB == getStartBB() ? frontInstruction() : &BB->front();
+
+    unsigned FirstInstGVN = *getGVN(FirstOutlineInst);
+    unsigned FirstInstCanonNum = *getCanonicalNum(FirstInstGVN);
+    unsigned SourceGVN = *SourceCand.fromCanonicalNum(FirstInstCanonNum);
+    Value *SourceV = *SourceCand.fromGVN(SourceGVN);
+    BasicBlock *SourceBB = cast<Instruction>(SourceV)->getParent();
+    unsigned SourceBBGVN = *SourceCand.getGVN(SourceBB);
+    unsigned SourceCanonBBGVN = *SourceCand.getCanonicalNum(SourceBBGVN);
+    CanonNumToNumber.insert(std::make_pair(SourceCanonBBGVN, BBGVNForCurrCand));
+    NumberToCanonNum.insert(std::make_pair(BBGVNForCurrCand, SourceCanonBBGVN));
+  }
 }
 
 void IRSimilarityCandidate::createCanonicalMappingFor(
     IRSimilarityCandidate &CurrCand) {
   assert(CurrCand.CanonNumToNumber.size() == 0 &&
          "Canonical Relationship is non-empty");
   assert(CurrCand.NumberToCanonNum.size() == 0 &&
          "Canonical Relationship is non-empty");

llvm/lib/Transforms/IPO/IROutliner.cpp

for (unsigned Idx = 0, EIdx = PN->getNumIncomingValues(); Idx < EIdx; Idx++) {
     if (Blocks.find(IncomingBlock) == Blocks.end())
       continue;
 
     // Collect the canonical numbers of the values in the PHINode.
     unsigned GVN = OGVN.getValue();
     OGVN = Cand.getCanonicalNum(GVN);
     assert(OGVN.hasValue() && "No GVN found for incoming value?");
     PHIGVNs.push_back(*OGVN);
+
+    // Find the incoming block and use the canonical numbering as well to define
+    // the hash for the PHINode.
+    OGVN = Cand.getGVN(IncomingBlock);
+
+    // If there is no number for the incoming block, it is becaause we have
+    // split the candidate basic blocks.  So we use the previous block that it
+    // was split from to find the valid global value numbering for the PHINode.
+    if (!OGVN.hasValue()) {
+      assert(Cand.getStartBB() == IncomingBlock &&
+             "Unknown basic block used in exit path PHINode.");
+
+      BasicBlock *PrevBlock = IncomingBlock->getSinglePredecessor();
+      OGVN = Cand.getGVN(PrevBlock);
+    }
+    GVN = OGVN.getValue();
+    OGVN = Cand.getCanonicalNum(GVN);
+    assert(OGVN.hasValue() && "No GVN found for incoming block?");
+    PHIGVNs.push_back(*OGVN);
   }
 
   // Now that we have the GVNs for the incoming values, we are going to combine
   // them with the GVN of the incoming bock, and the output location of the
   // PHINode to generate a hash value representing this instance of the PHINode.
   DenseMap<hash_code, unsigned>::iterator GVNToPHIIt;
   DenseMap<unsigned, PHINodeData>::iterator PHIToGVNIt;
   Optional<unsigned> BBGVN = Cand.getGVN(PHIBB);

llvm/test/Transforms/IROutliner/phi-node-exit-path-order.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --include-generated-funcs
+; RUN: opt -S -verify -iroutliner -ir-outlining-no-cost < %s | FileCheck %s
+
+; A PHINode defines the global value number of a split phi node for
+; an exit paths based on the canonical number for the incoming values, and
+; the canonical number for the basic block.  This checks that we accurately
+; capture a different numbering for the same incoming value but with different
+; blocks.
+
+define void @func1(i32 %0, i32 %1) local_unnamed_addr #0 {
+bb1:
+  br label %bb5
+
+bb2:
+  %a = add i32 %0, %1
+  %b = add i32 %0, %1
+  %c = icmp eq i32 %b, 1
+  br i1 %c, label %bb5, label %bb3
+
+bb3:
+  %d = add i32 %0, %1
+  br label %bb5
+
+bb4:
+  %e = sub i32 %0, %1
+  br label %bb2
+
+bb5:
+  %f = phi i32 [ 0, %bb1 ], [ 1, %bb2 ], [ 1, %bb3 ]
+  ret void
+}
+
+define void @func2(i32 %0, i32 %1) local_unnamed_addr #0 {
+bb1:
+  br label %bb5
+
+bb2:
+  %a = sub i32 %0, %1
+  %b = add i32 %0, %1
+  %c = icmp eq i32 %b, 1
+  br i1 %c, label %bb5, label %bb3
+
+bb3:
+  %d = add i32 %0, %1
+  br label %bb5
+
+bb4:
+  %e = add i32 %0, %1
+  br label %bb2
+
+bb5:
+  %f = phi i32 [ 0, %bb1 ], [ 1, %bb3 ], [ 1, %bb2 ]
+  ret void
+}
+; CHECK-LABEL: @func1(
+; CHECK-NEXT:  bb1:
+; CHECK-NEXT:    [[F_CE_LOC:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    br label [[BB5:%.*]]
+; CHECK:       bb2:
+; CHECK-NEXT:    [[A:%.*]] = add i32 [[TMP0:%.*]], [[TMP1:%.*]]
+; CHECK-NEXT:    [[LT_CAST:%.*]] = bitcast i32* [[F_CE_LOC]] to i8*
+; CHECK-NEXT:    call void @llvm.lifetime.start.p0i8(i64 -1, i8* [[LT_CAST]])
+; CHECK-NEXT:    call void @outlined_ir_func_0(i32 [[TMP0]], i32 [[TMP1]], i32* [[F_CE_LOC]], i32 0)
+; CHECK-NEXT:    [[F_CE_RELOAD:%.*]] = load i32, i32* [[F_CE_LOC]], align 4
+; CHECK-NEXT:    call void @llvm.lifetime.end.p0i8(i64 -1, i8* [[LT_CAST]])
+; CHECK-NEXT:    br label [[BB5]]
+; CHECK:       bb4:
+; CHECK-NEXT:    [[E:%.*]] = sub i32 [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    br label [[BB2:%.*]]
+; CHECK:       bb5:
+; CHECK-NEXT:    [[F:%.*]] = phi i32 [ 0, [[BB1:%.*]] ], [ [[F_CE_RELOAD]], [[BB2]] ]
+; CHECK-NEXT:    ret void
+;
+;
+; CHECK-LABEL: @func2(
+; CHECK-NEXT:  bb1:
+; CHECK-NEXT:    [[F_CE_LOC:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    br label [[BB5:%.*]]
+; CHECK:       bb2:
+; CHECK-NEXT:    [[A:%.*]] = sub i32 [[TMP0:%.*]], [[TMP1:%.*]]
+; CHECK-NEXT:    [[LT_CAST:%.*]] = bitcast i32* [[F_CE_LOC]] to i8*
+; CHECK-NEXT:    call void @llvm.lifetime.start.p0i8(i64 -1, i8* [[LT_CAST]])
+; CHECK-NEXT:    call void @outlined_ir_func_0(i32 [[TMP0]], i32 [[TMP1]], i32* [[F_CE_LOC]], i32 1)
+; CHECK-NEXT:    [[F_CE_RELOAD:%.*]] = load i32, i32* [[F_CE_LOC]], align 4
+; CHECK-NEXT:    call void @llvm.lifetime.end.p0i8(i64 -1, i8* [[LT_CAST]])
+; CHECK-NEXT:    br label [[BB5]]
+; CHECK:       bb4:
+; CHECK-NEXT:    [[E:%.*]] = add i32 [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    br label [[BB2:%.*]]
+; CHECK:       bb5:
+; CHECK-NEXT:    [[F:%.*]] = phi i32 [ 0, [[BB1:%.*]] ], [ [[F_CE_RELOAD]], [[BB2]] ]
+; CHECK-NEXT:    ret void
+;
+;
+; CHECK-LABEL: define internal void @outlined_ir_func_0(
+; CHECK-NEXT:  newFuncRoot:
+; CHECK-NEXT:    br label [[BB2_TO_OUTLINE:%.*]]
+; CHECK:       bb2_to_outline:
+; CHECK-NEXT:    [[B:%.*]] = add i32 [[TMP0:%.*]], [[TMP1:%.*]]
+; CHECK-NEXT:    [[C:%.*]] = icmp eq i32 [[B]], 1
+; CHECK-NEXT:    br i1 [[C]], label [[BB5_SPLIT:%.*]], label [[BB3:%.*]]
+; CHECK:       bb3:
+; CHECK-NEXT:    [[D:%.*]] = add i32 [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    br label [[BB5_SPLIT]]
+; CHECK:       bb5.split:
+; CHECK-NEXT:    [[TMP4:%.*]] = phi i32 [ 1, [[BB3]] ], [ 1, [[BB2_TO_OUTLINE]] ]
+; CHECK-NEXT:    [[F_CE:%.*]] = phi i32 [ 1, [[BB2_TO_OUTLINE]] ], [ 1, [[BB3]] ]
+; CHECK-NEXT:    br label [[BB5_EXITSTUB:%.*]]
+; CHECK:       bb5.exitStub:
+; CHECK-NEXT:    switch i32 [[TMP3:%.*]], label [[FINAL_BLOCK_0:%.*]] [
+; CHECK-NEXT:    i32 0, label [[OUTPUT_BLOCK_0_0:%.*]]
+; CHECK-NEXT:    i32 1, label [[OUTPUT_BLOCK_1_0:%.*]]
+; CHECK-NEXT:    ]
+; CHECK:       output_block_0_0:
+; CHECK-NEXT:    store i32 [[F_CE]], i32* [[TMP2:%.*]], align 4
+; CHECK-NEXT:    br label [[FINAL_BLOCK_0]]
+; CHECK:       output_block_1_0:
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[TMP2]], align 4
+; CHECK-NEXT:    br label [[FINAL_BLOCK_0]]
+; CHECK:       final_block_0:
+; CHECK-NEXT:    ret void
+;