Issue 241317
Summary [PGO] clang asserts in ProfileSummaryInfo::computeThresholds when a corrupted summary inverts the hot/cold MinCounts
Labels clang
Assignees
Reporter cactusBalll
    **Generated by Fuzzer.** Compiling with `-fprofile-instr-use=<profdata>` aborts in
`llvm::ProfileSummaryInfo::computeThresholds()` when a corrupted profile summary makes the
hot-cutoff entry's `MinBlockCount` greater than or equal to the cold-cutoff entry's
`MinBlockCount`. The reader accepts the summary structurally, and the analysis layer then
asserts `ColdCountThreshold < HotCountThreshold` on untrusted data.

## Environment

- clang `24.0.0git`, commit `0db002877564df53575e997e985538a949a03327`
- build config `+assertions,+asan,+ubsan`, `x86_64-unknown-linux-gnu`

## Reproduction

The crash does not depend on the compiled source contents; a one-line translation unit is
enough. Save the following as `repro.c`:

```c
int main(void) { return 0; }
```

Save this corrupted indexed profile as `mut.profdata`:

```bash
base64 -d > mut.profdata <<'EOF'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EOF
```

Then run:

```bash
clang -O2 -fprofile-instr-use=mut.profdata -c repro.c
```

The corrupted summary contains `cutoff 990000 -> MinBlockCount 76` and
`cutoff 999999 -> MinBlockCount 172`, so `computeThresholds()` sees a cold threshold above the
hot threshold.

## Output

```text
clang-24: llvm/lib/Analysis/ProfileSummaryInfo.cpp:124:
void llvm::ProfileSummaryInfo::computeThresholds():
Assertion `ColdCountThreshold < HotCountThreshold &&
"Cold count threshold should be less than hot count threshold!"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/
Stack dump:
...
#17 llvm::ProfileSummaryInfo::computeThresholds()
 llvm/lib/Analysis/ProfileSummaryInfo.cpp:123
#18 annotateAllFunctions(...)
 llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp:2099
#19 llvm::PGOInstrumentationUse::run(...)
 llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp:2283
```

## Root cause

`llvm/lib/Analysis/ProfileSummaryInfo.cpp`:

```cpp
HotCountThreshold =
 ProfileSummaryBuilder::getHotCountThreshold(DetailedSummary);
ColdCountThreshold =
    ProfileSummaryBuilder::getColdCountThreshold(DetailedSummary);

if (HotCountThreshold == ColdCountThreshold) {
  if (ColdCountThreshold > 0)
 (*ColdCountThreshold)--;
  else
 (*HotCountThreshold)++;
}
assert(ColdCountThreshold < HotCountThreshold &&
       "Cold count threshold should be less than hot count threshold!");
```

`getHotCountThreshold()` / `getColdCountThreshold()` just copy `MinCount` from the entries at
the hot (990000) and cold (999999) cutoffs. The reader does not validate that `MinCount` is
monotonic with respect to the cutoff, so a corrupted profile can invert the two thresholds.

In release builds the assert is compiled out and the compiler silently uses inconsistent
hot/cold thresholds, which can affect inlining, ICP, loop optimizations, and other PGO-guided
decisions.

## Relation to existing work

- #230018 / PR #231727 validate the structural size/count fields of the summary, not the
  semantic relationship between cutoffs and `MinCount`.
- #225435 covers on-disk hash-table decoding.


_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to