"H.J. Lu" <hjl.to...@gmail.com> writes: > On Fri, Oct 4, 2019 at 11:03 AM H.J. Lu <hjl.to...@gmail.com> wrote: >> >> On Wed, Sep 11, 2019 at 12:14 PM Richard Sandiford >> <richard.sandif...@arm.com> wrote: >> > >> > lra_reg has an actual_call_used_reg_set field that is only used during >> > inheritance. This in turn required a special lra_create_live_ranges >> > pass for flag_ipa_ra to set up this field. This patch instead makes >> > the inheritance code do its own live register tracking, using the >> > same ABI-mask-and-clobber-set pair as for IRA. >> > >> > Tracking ABIs simplifies (and cheapens) the logic in lra-lives.c and >> > means we no longer need a separate path for -fipa-ra. It also means >> > we can remove TARGET_RETURN_CALL_WITH_MAX_CLOBBERS. >> > >> > The patch also strengthens the sanity check in lra_assigns so that >> > we check that reg_renumber is consistent with the whole conflict set, >> > not just the call-clobbered registers. >> > >> > >> > 2019-09-11 Richard Sandiford <richard.sandif...@arm.com> >> > >> > gcc/ >> > * target.def (return_call_with_max_clobbers): Delete. >> > * doc/tm.texi.in (TARGET_RETURN_CALL_WITH_MAX_CLOBBERS): Delete. >> > * doc/tm.texi: Regenerate. >> > * config/aarch64/aarch64.c (aarch64_return_call_with_max_clobbers) >> > (TARGET_RETURN_CALL_WITH_MAX_CLOBBERS): Delete. >> > * lra-int.h (lra_reg::actual_call_used_reg_set): Delete. >> > (lra_reg::call_insn): Delete. >> > * lra.c: Include function-abi.h. >> > (initialize_lra_reg_info_element): Don't initialize the fields >> > above. >> > (lra): Use crtl->abi to test whether the current function needs to >> > save a register in the prologue. Remove special pre-inheritance >> > lra_create_live_ranges pass for flag_ipa_ra. >> > * lra-assigns.c: Include function-abi.h >> > (find_hard_regno_for_1): Use crtl->abi to test whether the current >> > function needs to save a register in the prologue. >> > (lra_assign): Assert that registers aren't allocated to a >> > conflicting register, rather than checking only for overlaps >> > with call_used_or_fixed_regs. Do this even for flag_ipa_ra, >> > and for registers that are not live across a call. >> > * lra-constraints.c (last_call_for_abi): New variable. >> > (full_and_partial_call_clobbers): Likewise. >> > (setup_next_usage_insn): Remove the register from >> > full_and_partial_call_clobbers. >> > (need_for_call_save_p): Use call_clobbered_in_region_p to test >> > whether the register needs a caller save. >> > (need_for_split_p): Use full_and_partial_reg_clobbers instead >> > of call_used_or_fixed_regs. >> > (inherit_in_ebb): Initialize and maintain last_call_for_abi and >> > full_and_partial_call_clobbers. >> > * lra-lives.c (check_pseudos_live_through_calls): Replace >> > last_call_used_reg_set and call_insn arguments with an abi >> > argument. >> > Remove handling of lra_reg::call_insn. Use >> > function_abi::mode_clobbers >> > as the set of conflicting registers. >> > (calls_have_same_clobbers_p): Delete. >> > (process_bb_lives): Track the ABI of the last call instead of an >> > insn/HARD_REG_SET pair. Update calls to >> > check_pseudos_live_through_calls. Use eh_edge_abi to calculate >> > the set of registers that could be clobbered by an EH edge. >> > Include partially-clobbered as well as fully-clobbered registers. >> > (lra_create_live_ranges_1): Don't initialize lra_reg::call_insn. >> > * lra-remat.c: Include function-abi.h. >> > (call_used_regs_arr_len, call_used_regs_arr): Delete. >> > (set_bb_regs): Use call_insn_abi to get the set of call-clobbered >> > registers and bitmap_view to combine them into dead_regs. >> > (call_used_input_regno_present_p): Take a function_abi argument >> > and use it to test whether a register is call-clobbered. >> > (calculate_gen_cands): Use call_insn_abi to get the ABI of the >> > call insn target. Update tje call to >> > call_used_input_regno_present_p. >> > (do_remat): Likewise. >> > (lra_remat): Remove the initialization of call_used_regs_arr_len >> > and call_used_regs_arr. >> >> This caused: >> >> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91994
Thanks for reducing & tracking down the underlying cause. > This change doesn't work with -mzeroupper. When -mzeroupper is used, > upper bits of vector registers are clobbered upon callee return if any > MM/ZMM registers are used in callee. Even if YMM7 isn't used, upper > bits of YMM7 can still be clobbered by vzeroupper when YMM1 is used. The problem here really is that the pattern is just: (define_insn "avx_vzeroupper" [(unspec_volatile [(const_int 0)] UNSPECV_VZEROUPPER)] "TARGET_AVX" "vzeroupper" ...) and so its effect on the registers isn't modelled at all in rtl. Maybe one option would be to add a parallel: (set (reg:V2DI N) (reg:V2DI N)) for each register. Or we could do something like I did for the SVE tlsdesc calls, although here that would mean using a call pattern for something that isn't really a call. Or we could reinstate clobber_high and use that, but that's very much third out of three. I don't think we should add target hooks to get around this, since that's IMO papering over the issue. I'll try the parallel set thing first. Richard
