https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119722
--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jakub Jelinek <ja...@gcc.gnu.org>: https://gcc.gnu.org/g:3f9dfb94eab1ab1bbf9a2b5e20d1f61e36516063 commit r15-9397-g3f9dfb94eab1ab1bbf9a2b5e20d1f61e36516063 Author: Jakub Jelinek <ja...@redhat.com> Date: Sat Apr 12 13:13:53 2025 +0200 bitintlower: Fix up handling of SSA_NAME copies in coalescing [PR119722] The following patch is miscompiled, because during the limited SSA name coalescing the bitintlower pass does we incorrectly don't register a conflict. This is on <bb 4> [local count: 1073741824]: # b_17 = PHI <b_19(3), 8(2)> g.4_13 = g; _14 = g.4_13 >> 50; _15 = (unsigned int) _14; _21 = b_17; _16 = (unsigned int) _21; s_22 = _15 + _16; return s_22; basic block where in the map->bitint bitmap we track 14, 17 and 19. The build_bitint_stmt_ssa_conflicts "hook" has special code where it tracks uses at the final statements of mergeable operations, so e.g. the _16 = (unsigned int) _21; statement is considered to be use of b_17 because _21 is not in map->bitmap (or large_huge.m_names), i.e. is mergeable. The problem is that build_ssa_conflict_graph has special code to handle SSA_NAME copies and _21 = b_17; is gimple_assign_copy_p. In such cases it calls live_track_clear_var on the rhs1. The problem is that on the above bb, after we note in the _16 = (unsigned int) _21; stmt we need b_17 the generic code makes us forget that because of the copy statement, and then build_bitint_stmt_ssa_conflicts ignores it completely (because _21 is large/huge bitint and is not in map->bitint, so assumed to be handled by a later stmt in the bb, for backwards walk like this before this one). As the b_17 use is ignored, the coalescing thinks it can put all of b_17, b_19 and _14 into the same partition, which is wrong, while we can and should coalesce b_17 and b_19, _14 needs to be a different temporary because b_17 is set before and used after _14 has been written. The following patch fixes it by handling gimple_assign_copy_p in two separate spots, move the generic coalesce handling of it after build_ssa_conflict_graph (where build_ssa_conflict_graph handling doesn't fall through to that, it does continue after the call) and inside of build_ssa_conflict_graph it performs it too, but only if the lhs is not mergeable large/huge bitint. 2025-04-12 Jakub Jelinek <ja...@redhat.com> PR tree-optimization/119722 * gimple-lower-bitint.h (build_bitint_stmt_ssa_conflicts): Add CLEAR argument. * gimple-lower-bitint.cc (build_bitint_stmt_ssa_conflicts): Add CLEAR argument. Call clear on gimple_assign_copy_p rhs1 if lhs is large/huge bitint unless lhs is not in names. * tree-ssa-coalesce.cc (build_ssa_conflict_graph): Adjust build_bitint_stmt_ssa_conflicts caller. Move gimple_assign_copy_p handling to after the build_bitint_stmt_ssa_conflicts call. * gcc.dg/torture/bitint-77.c: New test.
