https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120347
--- Comment #9 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The trunk branch has been updated by Richard Sandiford <rsand...@gcc.gnu.org>: https://gcc.gnu.org/g:e322dff09d011f65f5cae4e95c3a24ccfae7b1e1 commit r16-984-ge322dff09d011f65f5cae4e95c3a24ccfae7b1e1 Author: Richard Sandiford <richard.sandif...@arm.com> Date: Fri May 30 09:36:35 2025 +0100 rtl-ssa: Reject non-address uses of autoinc regs [PR120347] As the rtl.texi documentation of RTX_AUTOINC expressions says: If a register used as the operand of these expressions is used in another address in an insn, the original value of the register is used. Uses of the register outside of an address are not permitted within the same insn as a use in an embedded side effect expression because such insns behave differently on different machines and hence must be treated as ambiguous and disallowed. late-combine was failing to follow this rule. One option would have been to enforce it during the substitution phase, like combine does. This could either be a dedicated condition in the substitution code or, more generally, an extra condition in can_merge_accesses. (The latter would include extending is_pre_post_modify to uses.) However, since the restriction applies to patterns rather than to actions on patterns, the more robust fix seemed to be test and reject this case in (a subroutine of) rtl_ssa::recog. We already do something similar for hard-coded register clobbers. Using vec_rtx_properties isn't the lightest-weight operation out there. I did wonder about relying on the is_pre_post_modify flag of the definitions in the new_defs array, but that would require callers that create new autoincs to set the flag before calling recog. Normally these flags are instead updated automatically based on the final pattern. Besides, recog itself has had to traverse the whole pattern, and it is even less light-weight than vec_rtx_properties. At least the pattern should be in cache. The rtl-ssa fix showed up a mistake (of mine) in the rtl_properties walker: try_to_add_src would drop all flags except IN_NOTE before recursing into RTX_AUTOINC addresses. RTX_AUTOINCs only occur in addresses, and so for them, the flags coming into try_to_add_src are set by: unsigned int base_flags = flags & rtx_obj_flags::STICKY_FLAGS; ... if (MEM_P (x)) { ... unsigned int addr_flags = base_flags | rtx_obj_flags::IN_MEM_STORE; if (flags & rtx_obj_flags::IS_READ) addr_flags |= rtx_obj_flags::IN_MEM_LOAD; try_to_add_src (XEXP (x, 0), addr_flags); return; } This means that the only flags that can be set are: - IN_NOTE (the sole member of STICKY_FLAGS) - IN_MEM_STORE - IN_MEM_LOAD Thus dropping all flags except IN_NOTE had the effect of dropping IN_MEM_STORE and IN_MEM_LOAD, and nothing else. But those flags are the ones that mark something as being part of a mem address. The exclusion was therefore exactly wrong. gcc/ PR rtl-optimization/120347 * rtlanal.cc (rtx_properties::try_to_add_src): Don't drop the IN_MEM_LOAD and IN_MEM_STORE flags for autoinc registers. * rtl-ssa/changes.cc (recog_level2): Check whether an RTX_AUTOINCed register also appears outside of an address. gcc/testsuite/ PR rtl-optimization/120347 * gcc.dg/torture/pr120347.c: New test.
