Remove call_fixed_reg_set
authorRichard Sandiford <richard.sandiford@arm.com>
Tue, 10 Sep 2019 18:56:26 +0000 (18:56 +0000)
committerRichard Sandiford <rsandifo@gcc.gnu.org>
Tue, 10 Sep 2019 18:56:26 +0000 (18:56 +0000)
commitdf1f0eef67939274e9ddd3df426e8dfc5184086b
tree5c82b08c38893696032489e881bf32e860f507a2
parent031e8857886b91da290ffb03d9027424566e5da3
Remove call_fixed_reg_set

On targets that use reload, call_fixed_reg_set is structurally:

  fixed_reg_set                           -- reginfo.c
  | (call_used_reg_set & ~have_save_mode) -- first loop in init_caller_save
  | ~have_save_insn                       -- final loop in init_caller_save

(where "have_save_mode" and "have_save_insn" are just my names).
But the final loop in init_caller_save does:

  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
    for (j = 1; j <= MOVE_MAX_WORDS; j++)
      if (reg_save_code (i,regno_save_mode[i][j]) == -1)

This last condition ought to be true whenever:

  regno_save_mode[i][j] == VOIDmode

since either targetm.hard_regno_mode_ok (i, VOIDmode) should be
false or the VOIDmode save & restore shouldn't match any move insn.
And after the first loop, regno_save_mode[i][j] == VOIDmode whenever
!call_used_regs[i].  So the above is actually:

  fixed_reg_set
  | (call_used_reg_set & ~have_save_mode)
  | (~call_used_reg_set | ~have_save_insn)

which simplifies to:

  fixed_reg_set                        -- reginfo.c
  | ~have_save_mode                    -- first loop in init_caller_save
  | ~have_save_insn                    -- final loop in init_caller_save
  | ~call_used_reg_set                 -- final loop in init_caller_save

So:

  ~call_fixed_reg_set == (~fixed_reg_set
          & have_save_mode
  & have_save_insn
  & call_used_reg_set)  [A]

All users have the form:

  (call_used_reg_set or some subset) & ~(call_fixed_reg_set | ...)

i.e.:

  (call_used_reg_set or some subset) & ~call_fixed_reg_set & ~(...)

We can therefore drop the "& call_used_reg_set" from [A], leaving:

  ~fixed_reg_set & have_save_mode & have_save_insn

This patch combines have_save_mode & have_save_insn into a single
condition "a save is possible", represented as savable_regs.
We can then substitute:

  ~call_fixed_reg_set --> ~fixed_reg_set & savable_regs
                          (registers we can actually save around calls)

The patch also sets regno_save_mode[i][j] for all registers,
in case non-default ABIs require a save when the default ABI
doesn't.  This ensures that savable_regs (like fixed_reg_set but
unlike call_fixed_reg_set) isn't affected by the ABI.  This only
becomes significant with later patches and at this point is just
a simplification.

Since init_caller_save is only called for reload targets,
the default assumption for LRA is that all registers are savable,
just like the default assumption before the patch was that
(~)call_fixed_reg_set == (~)fixed_reg_set.

2019-09-10  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* hard-reg-set.h (target_hard_regs::x_call_fixed_reg_set): Delete.
(target_hard_regs::x_savable_regs): New field.
(call_fixed_reg_set): Delete.
(savable_regs): New macro,
* reginfo.c (globalize_reg): Don't set call_fixed_reg_set.
(init_reg_sets_1): Likewise.  Initialize savable_regs.
* caller-save.c (init_caller_save): Invoke HARD_REGNO_CALLER_SAVE_MODE
for all registers.  Set savable_regs instead of call_fixed_reg_set.
(setup_save_areas, save_call_clobbered_regs): Replace uses of
~call_fixed_reg_set with ~fixed_reg_set & savable_regs.
* config/sh/sh.c (output_stack_adjust): Likewise.

From-SVN: r275598
gcc/ChangeLog
gcc/caller-save.c
gcc/config/sh/sh.c
gcc/hard-reg-set.h
gcc/reginfo.c