0000583: crash in LLDS .par grades

View Issue Details [ Jump to Notes ]

[ Issue History ] [ Print ]

Project

Category

View Status

Date Submitted

Last Update

0000583

mercury

Bug

public

2026-03-31 14:16

2026-04-02 17:48

Reporter

wangp

Assigned To

Priority

normal

Severity

crash

Reproducibility

always

Status

new

Resolution

open

Product Version

Target Version

Fixed in Version

Summary

0000583: crash in LLDS .par grades

Description

Here are some results of trying to bootcheck the asm_fast.par.gc grade on various Linux distributions. All tests were performed in a container on the same Linux host kernel, on a x86-64 machine.

--------

Debian 12
  GNU ld (GNU Binutils for Debian) 2.40
  gcc version 12.2.0 (Debian 12.2.0-14+deb12u1)
    - mercury-srcdist-rotd-2026-03-19 - ok

Debian 13
  GNU ld (GNU Binutils for Debian) 2.44
  gcc version 14.2.0 (Debian 14.2.0-19)
    - mercury-srcdist-rotd-2026-03-19 - ok

--------

Ubuntu 20.04
  GNU ld (GNU Binutils for Ubuntu) 2.34
  gcc version 9.4.0 (Ubuntu 9.4.0-1ubuntu1~20.04.2)
    - mercury-srcdist-rotd-2026-03-19 - ok

Ubuntu 22.04
  GNU ld (GNU Binutils for Ubuntu) 2.38
  gcc version 11.4.0 (Ubuntu 11.4.0-1ubuntu1~22.04.3)
    - mercury-srcdist-rotd-2026-03-19 - ok

Ubuntu 24.04
  GNU ld (GNU Binutils for Ubuntu) 2.42
  gcc version 11.5.0 (Ubuntu 11.5.0-1ubuntu1~24.04.1)
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful
  gcc version 12.4.0 (Ubuntu 12.4.0-2ubuntu1~24.04.1)
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful
  gcc version 13.3.0 (Ubuntu 13.3.0-6ubuntu2~24.04.1) - default
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful
  gcc version 14.2.0 (Ubuntu 14.2.0-4ubuntu2~24.04.1)
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful

--------

Alpine Linux 3.20
  GNU ld (GNU Binutils) 2.42
  gcc version 13.2.1 20240309 (Alpine 13.2.1_git20240309)
    - mercury-srcdist-rotd-2026-03-19 - ok

Alpine Linux 3.22
  GNU ld (GNU Binutils) 2.44
  gcc version 14.2.0 (Alpine 14.2.0)
    - mercury-srcdist-rotd-2026-03-19 - ok

Alpine Linux 3.23
  GNU ld (GNU Binutils) 2.45.1
  gcc version 15.2.0 (Alpine 15.2.0)
    - mercury-srcdist-rotd-2026-03-19 - ok

--------

AlmaLinux 9
  GNU ld version 2.35.2-67.el9_7.1
  gcc version 11.5.0 20240719 (Red Hat 11.5.0-11) (GCC)
    - mercury-srcdist-rotd-2026-03-19 - ok

AlmaLinux 10
  GNU ld version 2.41-58.el10_1.2.alma.1
  gcc version 14.3.1 20250617 (Red Hat 14.3.1-2) (GCC) - default
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful
  gcc version 15.1.1 20250521 (Red Hat 15.1.1-2) (GCC)
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful

--------

Fedora 43
  GNU ld version 2.45.1-4.fc43
  gcc version 14.3.1 20250808 (Red Hat 14.3.1-1) (GCC)
    - mercury-srcdist-rotd-2026-03-19 - ok
  gcc version 15.2.1 20260123 (Red Hat 15.2.1-7) (GCC) - default
    - mercury-srcdist-rotd-2026-03-19 - ok

--------

OpenSuSE Leap 15.6
  GNU ld (GNU Binutils; SUSE Linux Enterprise 15) 2.45.0.20251103-150100757
  gcc version 7.5.0 (SUSE Linux)
    - mercury-srcdist-rotd-2026-03-19 - ok

OpenSuSE Tumbleweed
  GNU ld (GNU Binutils; openSUSE Tumbleweed) 2.45.0.20251103-2
  gcc version 15.2.1 20260202 (SUSE Linux)
    - mercury-srcdist-rotd-2026-03-19 - building of stage 3 dependencies not successful

--------

Void Linux
  GNU ld (GNU Binutils) 2.44
  gcc version 14.2.1 20250405 (GCC)
    - mercury-srcdist-rotd-2026-03-19 - ok

--------

I'm not sure what conclusions we can draw. The problem appears to be something particular to a distribution, rather than an issue with certain gcc or binutils versions. (But keeping in mind that distributions will have applied their own patches to those software packages.)

Tags

No tags attached.

Attached Files

Relationships

Relationships

Notes
~0001244 wangp (developer) 2026-04-02 16:49	GC_DONT_GC=1 allows mmake depend to pass. The problem is likely premature collection of some object. GC_NPROCS=1 GC_MARKERS=1 doesn't help. Taking the working stage2 mercury_compile binary from Ubuntu 22.04, and running it on Ubuntu 24.04 does NOT work. Taking the working stage2 mercury_compile binary from AlmaLinux 9, and running it on AlmaLinux 10 does NOT work. Therefore, I suspect an issue in the interaction of Boehm GC + asm_fast.par.gc + glibc version (glibc is dynamically linked). Table of libc versions: Alpine Linux ok musl Ubuntu 20.04 ok glibc 2.31 AlmaLinux 9 ok glibc 2.34 Ubuntu 22.04 ok glibc 2.35 Debian 12 ok glibc 2.36 OpenSuSE Leap 15.6 ok glibc 2.38 AlmaLinux 10 crash glibc 2.39 Ubuntu 24.04 crash glibc 2.39 Debian 13 ok glibc 2.41 Fedora 43 ok glibc 2.42 OpenSuSE Tumbleweed crash glibc 2.42

~0001245 zs (developer) 2026-04-02 17:03	I remember that I diagnosed the issue years ago. It was that the various disjuncts did not flush their output variables to their own stack slots, because those variables did not have their "own" stack slots. This was because the stack slot allocation pass was not updated to tell the graph colouring algorithm "allocate the variables that are all live at the same time at the end of a parallel conjunction to different stack slots". The parallel conjunction will work - either if the output vars are allocated to distinct stack slots anyway by chance, - or if, of the variables incorrectly sharing a slot, only one is needed after the parallel concjunction, and it happens to be stored to the slot last. The first condition does happen very often in small benchmark programs, and the second also happens often (due to left-to-right flow of data), which is why this was not detected during initial development. There is no point in looking for any correlations between gc and crashes until this issue is fixed.

~0001246 wangp (developer) 2026-04-02 17:15	There is no parallel conjunction involved.

~0001247 juliensf (administrator) 2026-04-02 17:19	Is the problem restricted to glibc-based systems?

~0001249 zs (developer) 2026-04-02 17:43	A bootcheck in a parallel grade does involve parallel conjunctions. You just don't ordinarily don't notice them, because in non-parallel (or non-LLDS) grades they get converted to plain conjunctions. See for example library/integer.m, and specifically pos_mul_karatsuba. I think there are a couple in the compiler as well, though my bootcheck of a compiler that I modified to abort on parallel_conj does not get past the library due to integer.m.

~0001250 wangp (developer) 2026-04-02 17:48	I haven't seen the problem on Alpine Linux (which uses musl), and Void Linux (which has both glibc and musl variants, both test ok).

Notes

Date Modified	Username	Field	Change
Issue History
2026-03-31 14:16	wangp	New Issue
2026-04-01 16:52	wangp	Severity	minor => crash
2026-04-01 16:52	wangp	Reproducibility	sometimes => always
2026-04-02 16:49	wangp	Note Added: 0001244
2026-04-02 17:03	zs	Note Added: 0001245
2026-04-02 17:15	wangp	Note Added: 0001246
2026-04-02 17:19	juliensf	Note Added: 0001247
2026-04-02 17:43	zs	Note Added: 0001249
2026-04-02 17:48	wangp	Note Added: 0001250

Issue History