Public bug reported: Ubuntu 26.04 LTS currently ships GCC 15.2.0 (`15.2.0-16ubuntu1`), which predates the upstream fix for GCC PR tree-optimization/124802.
This regression causes incorrect code generation for AVX2 double-complex operations at `-O2`/`-O3`. The upstream fix was backported to the GCC 15 release branch on April 14, 2026: - Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124802 - Fix: https://gcc.gnu.org/g:8d8e8cd6fbed9029e996a8d7298aa82fd453345e We have also encountered a related, reliably reproducible issue when building ELPA with GCC 15.2: its double-complex AVX/AVX2 BLOCK2 kernels can return severely incorrect and nonorthogonal eigenvectors without reporting an error. Adding `-fno-tree-slp-vectorize` resolves the issue, while GCC 15.3 builds work correctly. See https://github.com/marekandreas/elpa/issues/77 for details. Since this class of regression can silently corrupt scientific computing results (which is terrible), could the upstream fix for bug 124802 be backported to Ubuntu 26.04 through an SRU? The compiler change itself is minimal (one source-code line, plus a regression test). ** Affects: gcc-15 (Ubuntu) Importance: Undecided Status: New ** Description changed: Ubuntu 26.04 LTS currently ships GCC 15.2.0 (`15.2.0-16ubuntu1`), which predates the upstream fix for GCC PR tree-optimization/124802. This regression causes incorrect code generation for AVX2 double-complex operations at `-O2`/`-O3`. The upstream fix was backported to the GCC 15 release branch on April 14, 2026: - - Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124802 - - Fix: https://gcc.gnu.org/g:8d8e8cd6fbed9029e996a8d7298aa82fd453345e + - Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122408 + - Fix: https://gcc.gnu.org/cgit/gcc/commit/?id=5b7639b0de2f1d610e16eec628033fe3650d4763 We have also encountered a related, reliably reproducible issue when building ELPA with GCC 15.2: its double-complex AVX/AVX2 BLOCK2 kernels can return severely incorrect and nonorthogonal eigenvectors without reporting an error. Adding `-fno-tree-slp-vectorize` resolves the issue, while GCC 15.3 builds work correctly. See https://github.com/marekandreas/elpa/issues/77 for details. Since this class of regression can silently corrupt scientific computing results (which is terrible), could the upstream fix for PR 124802 be backported to Ubuntu 26.04 through an SRU? The compiler change itself is minimal (one source-code line, plus a regression test). ** Description changed: Ubuntu 26.04 LTS currently ships GCC 15.2.0 (`15.2.0-16ubuntu1`), which predates the upstream fix for GCC PR tree-optimization/124802. This regression causes incorrect code generation for AVX2 double-complex operations at `-O2`/`-O3`. The upstream fix was backported to the GCC 15 release branch on April 14, 2026: - Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122408 - Fix: https://gcc.gnu.org/cgit/gcc/commit/?id=5b7639b0de2f1d610e16eec628033fe3650d4763 We have also encountered a related, reliably reproducible issue when building ELPA with GCC 15.2: its double-complex AVX/AVX2 BLOCK2 kernels can return severely incorrect and nonorthogonal eigenvectors without reporting an error. Adding `-fno-tree-slp-vectorize` resolves the issue, while GCC 15.3 builds work correctly. See https://github.com/marekandreas/elpa/issues/77 for details. Since this class of regression can silently corrupt scientific computing - results (which is terrible), could the upstream fix for PR 124802 be + results (which is terrible), could the upstream fix for bug 122408 be backported to Ubuntu 26.04 through an SRU? The compiler change itself is minimal (one source-code line, plus a regression test). ** Description changed: Ubuntu 26.04 LTS currently ships GCC 15.2.0 (`15.2.0-16ubuntu1`), which predates the upstream fix for GCC PR tree-optimization/124802. This regression causes incorrect code generation for AVX2 double-complex operations at `-O2`/`-O3`. The upstream fix was backported to the GCC 15 release branch on April 14, 2026: - - Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122408 - - Fix: https://gcc.gnu.org/cgit/gcc/commit/?id=5b7639b0de2f1d610e16eec628033fe3650d4763 + - Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124802 + - Fix: https://gcc.gnu.org/g:8d8e8cd6fbed9029e996a8d7298aa82fd453345e We have also encountered a related, reliably reproducible issue when building ELPA with GCC 15.2: its double-complex AVX/AVX2 BLOCK2 kernels can return severely incorrect and nonorthogonal eigenvectors without reporting an error. Adding `-fno-tree-slp-vectorize` resolves the issue, while GCC 15.3 builds work correctly. See https://github.com/marekandreas/elpa/issues/77 for details. Since this class of regression can silently corrupt scientific computing - results (which is terrible), could the upstream fix for bug 122408 be + results (which is terrible), could the upstream fix for bug 124802 be backported to Ubuntu 26.04 through an SRU? The compiler change itself is minimal (one source-code line, plus a regression test). -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2175491 Title: [SRU] Backport GCC PR 124802 SLP wrong-code fix to Ubuntu 26.04 To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/gcc-15/+bug/2175491/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
