[Bug] --enable-insert-sync hangs on gemm_incore_0.pto in v0.22, works in v0.17

[zhangqi-chen]

Component

Passes / Transforms (sync insertion via --enable-insert-sync)

Description

There appears to be a regression in sync instruction insertion triggered by --enable-insert-sync.

For the same input file:

• PTOAS v0.17 works
• PTOAS v0.22 hangs indefinitely

The reporter can attach the exact .pto file after the issue is created.

Reproduction (minimal)

Run PTOAS with sync insertion enabled on the same input file:

ptoas build_output/GemmProgram_20260407_154637/ptoas/gemm_incore_0.pto \
  -o build_output/GemmProgram_20260407_154637/ptoas/gemm_incore_0.cpp \
  --enable-insert-sync \
  --pto-level=level3

Observed behavior:

1. PTOAS v0.17 completes successfully
2. PTOAS v0.22 hangs indefinitely on the same input

Current file path in the reporter workspace:

gemm_incore_0.pto.txt

The reporter can attach this file to the issue after creation.

Expected behavior

v0.22 should handle this input the same way as v0.17, or fail with a diagnostic. It should not hang.

Actual behavior / error logs

With --enable-insert-sync, v0.22 gets stuck on this input and does not finish. No useful diagnostic has been captured yet.

Environment

• Host platform: Linux (aarch64)
• Host kernel: Linux liteserver-hps-148e-00001 5.10.0-182.0.0.95.r1941_123.hce2.aarch64
• Affected versions: v0.17 (good), v0.22 (hang)
• Local PTOAS checkout HEAD: 602136d7cd271a787bbfb2cb150fcdb4d09ed10e

[TaoTao-real]

Repro update from our side (Apr 7, 2026): with the attached file, we cannot reproduce the hang yet.

Tested input

• gemm_incore_0.pto.txt from this issue attachment
• sha256: 4ffcf4e4b16ce424de5d6c1ef56a2114e16a939cb1812dfc90148dfb017da5fb

1. macOS arm64 (local build from v0.22 tag)

• Host: macOS 15.7.4, arm64
• v0.22 tag commit: 346d908
• Command:

ptoas gemm_incore_0.pto.txt -o out.cpp --enable-insert-sync --pto-level=level3

• Result: RC=0, finished in ~0.48s (no hang)

2. Linux x86_64 (official release binary)

• Host: Ubuntu 24.04 x86_64
• Binary: ptoas-bin-x86_64.tar.gz from release v0.22
• sha256: df30a90c7dbcd25499d2625b6486e06a049bd51eade35d303108c97ec8e00886
• Same command as above
• Result: RC=0, finished in ~0.27s (no hang)

3. Cross-check with v0.17 on same machines

• Also finishes normally on the same input.

Current conclusion

• With the attached input, we cannot reproduce a v0.22 insert-sync hang on either macOS arm64 (v0.22-tag build) or Linux x86_64 (official v0.22 release binary).

Could you share the exact runtime context where it hangs on Linux aarch64?

• ptoas --version output (or binary checksum)
• exact input file checksum used at runtime
• command line (full)
• last stdout/stderr lines before hang (or where it stalls)

Once we get those, we can continue narrowing this down quickly.