Compile glimmerhmm from source to dodge bioconda's AVX2 baseline (Rosetta 2 SIGILL)

[nextgenusfs]

Summary

Same root cause we already addressed for augustus (PR #70) and pytantan: bioconda's glimmerhmm 3.0.4 binary is compiled with -march=x86-64-v3 (AVX2/BMI2 baseline), so it SIGILLs under Rosetta 2 on Apple Silicon and on pre-Haswell x86_64 hardware. In the linux/amd64 Docker image this surfaces as:

[May 27 03:17 PM] ERROR 43: /app/.pixi/envs/default/bin/../share/glimmerhmm/train/build1 exited funny: 4 at /app/.pixi/envs/default/bin/trainGlimmerHMM line 338.

(perl $? = 4 == SIGILL → illegal instruction → build1 was compiled with instructions Rosetta 2 can't emulate.)

Unlike augustus, glimmerhmm is not packaged in Ubuntu apt, so the apt-substitution trick used in PR #70 isn't available. Instead, this PR adds a new Docker build stage that compiles glimmerhmm 3.0.4 from the upstream tarball with -march=x86-64 -mtune=generic (the x86-64 v1 baseline) and layers it onto PATH ahead of the bioconda binary in the pixi env.

Changes

• Dockerfile

  • New glimmerhmm-build stage (Ubuntu noble + build-essential) that:
    • Downloads GlimmerHMM-3.0.4.tar.gz from ccb.jhu.edu (sha256-pinned).
    • Applies the same upstream patches the bioconda recipe carries — makefile typo fixes (escoreSTOP2 → scoreSTOP2, rfapp → erfapp, drop the nonexistent trainGlimmerHMM make-clean target) and trainGlimmerHMM / glimmhmm.pl patched to #!/usr/bin/env perl + FindBin qw($RealBin) so they self-locate support binaries via $RealBin/../share/glimmerhmm/train.
    • Compiles sources/ and train/ with CFLAGS/CXXFLAGS="-O3 -march=x86-64 -mtune=generic ...".
    • Installs into /opt/glimmerhmm using bioconda's bin/ + share/glimmerhmm/{train,trained_dir} layout.
    • Defense-in-depth check: objdump scans the compiled binaries for AVX/AVX2/AVX512 mnemonics (ymm/zmm refs, vpbroadcast, vextracti128, vfmadd, …) and fails the build if any appear. Same approach pytantan uses.
  • Final stage:
    • COPY --from=glimmerhmm-build /opt/glimmerhmm /opt/glimmerhmm.
    • ENV PATH=/opt/glimmerhmm/bin:/app/.pixi/envs/default/bin:... so the self-compiled binary wins over the pixi-env bioconda binary.
    • entrypoint.sh re-prepends /opt/glimmerhmm/bin to PATH after the pixi shell-hook activation — necessary because the shell-hook re-prepends /app/.pixi/envs/default/bin and would otherwise put the bioconda binary back in front.

• pixi.toml

  • Comment-only change explaining why glimmerhmm = "*" is intentionally kept in [dependencies] despite being shadowed in the docker image: on osx-arm64 it provides the arm64-native binary for outside-docker dev work, and inside the image it sits behind /opt/glimmerhmm on PATH as a shadowed fallback.

Why a fallback?

funannotate2 invokes glimmerhmm via PATH-resolved binary names (trainGlimmerHMM, glimmerhmm, glimmhmm.pl) — confirmed by grepping funannotate2/abinitio.py, train.py, predict.py. With /opt/glimmerhmm/bin at the front of PATH, the self-compiled v1-baseline binaries win. If /opt/glimmerhmm ever goes missing (corrupted COPY, manual edit), the bioconda binary in the pixi env will be picked up automatically — at which point Rosetta 2 will SIGILL again, but on a native x86_64 host it'll just keep working.

Verification

• git diff confirms only Dockerfile and pixi.toml change; pixi.lock is untouched.
• Smoke test (Docker build + trainGlimmerHMM run under Rosetta 2) still needs to happen before merging.

Related

• PR Fix CPU-baseline SIGILLs under Rosetta 2 (apt augustus + pytantan no-SIMD) #70 — augustus / pytantan CPU-baseline SIGILL fix (merged).

---

Pull Request opened by Augment Code with guidance from the PR author