[PGOInstrUse] "dereferencing end() iterator" on unreachable block

[aeubanks]

diff --git a/llvm/test/Transforms/PGOProfile/Inputs/musttail_pgo-instr-use.proftext b/llvm/test/Transforms/PGOProfile/Inputs/musttail_pgo-instr-use.proftext
new file mode 100644
index 000000000000..9be174fb32bc
--- /dev/null
+++ b/llvm/test/Transforms/PGOProfile/Inputs/musttail_pgo-instr-use.proftext
@@ -0,0 +1,9 @@
+# IR level Instrumentation Flag
+:ir
+foo
+# Func Hash:
+742261418966908927
+# Num Counters:
+1
+# Counter Values:
+1
diff --git a/llvm/test/Transforms/PGOProfile/musttail_pgo-instr-use.ll b/llvm/test/Transforms/PGOProfile/musttail_pgo-instr-use.ll
new file mode 100644
index 000000000000..1d1aa8747971
--- /dev/null
+++ b/llvm/test/Transforms/PGOProfile/musttail_pgo-instr-use.ll
@@ -0,0 +1,12 @@
+; RUN: llvm-profdata merge %S/Inputs/musttail_pgo-instr-use.proftext -o %t.profdata
+; RUN: opt < %s -passes=pgo-instr-use -pgo-test-profile-file=%t.profdata -S
+
+declare ptr @bar()
+
+define ptr @foo() {
+entry:
+  ret ptr null
+
+2:
+  ret ptr null
+}

we get an error:

clang++: ../../llvm/include/llvm/ADT/DenseMap.h:1245: pointer llvm::DenseMapIterator<const llvm::BasicBlock *, std::unique_ptr<(anonymous namespace)::PGOUseBBInfo>, llvm::DenseMapInfo<const llvm::BasicBlock *>, llvm::detail::DenseMapPair<const llvm::BasicBlock *, std::unique_ptr<(anonymous namespace)::PGOUseBBInfo>>, true>::operator->() const [KeyT = const llvm::BasicBlock *, V
alueT = std::unique_ptr<(anonymous namespace)::PGOUseBBInfo>, KeyInfoT = llvm::DenseMapInfo<const llvm::BasicBlock *>, Bucket = llvm::detail::DenseMapPair<const llvm::BasicBlock *, std::uniq
ue_ptr<(anonymous namespace)::PGOUseBBInfo>>, IsConst = true]: Assertion `Ptr != End && "dereferencing end() iterator"' failed.

[aeubanks]

@xur-llvm @mtrofin @mingmingl-llvm

[aeubanks]

likely something just needs to handle/ignore unreachable blocks

[mingmingl-llvm]

I think this happens when the function has one BB (the entry block), and the rest of BBs are unreachable. In this case, the compiler runs

llvm-project/llvm/include/llvm/Transforms/Instrumentation/CFGMST.h

Lines 136 to 139 in f7bdf30

	// Special handling for single BB functions.
	if (succ_empty(Entry)) {
	addEdge(Entry, nullptr, EntryWeight);
	return;

, which skips creating BBInfo for the rest of unreachable BBs (the loop at L144)

Running opt -passes=pgo-instr-use -pgo-test-profile-file=<file>.profdata -S -debug <input.ll> shows the edge creation logs, and there are no edges to the unreachable block.

Build Edge on foo
  Edge: from fake node to entry w = 18014398509481984
Function Hash Computation for foo:
 CRC = 4294967295, Selects = 0, Edges = 2, ICSites = 0, Memops = 0, High32 CRC = 2765139247, Hash = 742261418966908927
Dump Function foo Hash: 742261418966908927	after CFGMST
  Number of Basic Blocks: 2
  BB: FakeNode  Index=0
  BB: entry  Index=1
  Number of Edges: 2 (*: Instrument, C: CriticalEdge, -: Removed)
  Edge 0: 0-->1 *   W=18014398509481984
  Edge 1: 1-->0     W=18014398509481984
1 counts
  0: 1
SUM =  1
Populate counts in 2 passes.
Dump Function foo Hash: 742261418966908927	after reading profile.
  Number of Basic Blocks: 2
  BB: FakeNode  Index=0
  BB: entry  Index=1  Count=1
  Number of Edges: 2 (*: Instrument, C: CriticalEdge, -: Removed)
  Edge 0: 0-->1 *   W=18014398509481984  Count=1
  Edge 1: 1-->0     W=18014398509481984  Count=1

[zmodem]

Is anyone looking into fixing this?