https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111231
--- Comment #32 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Richard Earnshaw from comment #31) > While that does seem to fix the bug, it's at the cost of 6 additional stores > in the problematic test that are redundant other than changing the alias set > view. The alternative is to alter the earlier store MEM_ATTRs to use an alias-set covering both which usually means using alias-set zero. This will pessimize followup optimizations around the store though but it might be a good trade-off if done only late - I'd say after sched2 but it doesn't look like theres CSE/DSE after it. So maybe after sched1 which effectively means after reload, but there's no regular CSE after reload either. The latest CSE is pass_cse2. IIRC a minor complication is that the earlier insn isn't readily available - IIRC 'dest' is copied/mangled and not necessarily the single origial RTX of the earlier SET_DEST (IIRC - it's been some time). OTOH I think that correctness trumps optimization and if this is the problematical transform then I don't see much options here. In the place CSE applies the transform we'd have to set MEM_ALIAS_SET to zero if the alias set condition doesn't hold and clear MEM_EXPR if the MEM_EXPR condition doesn't hold. Note I can't get the cse.cc code to trigger with the full preprocessed source and a cross to arm and using -O2 -fno-exceptions -march=armv7-a -mfpu=neon-vfpv4 -mfloat-abi=hard -mfp16-format=ieee -fmath-errno You mention at one point an insn removed by postreload, but that doesn't use alias_set_subset_of. I also don't remember postreload doing redundant store removal.
