Simplify the implementation of HARD_REG_SET
authorRichard Sandiford <richard.sandiford@arm.com>
Mon, 9 Sep 2019 17:58:36 +0000 (17:58 +0000)
committerRichard Sandiford <rsandifo@gcc.gnu.org>
Mon, 9 Sep 2019 17:58:36 +0000 (17:58 +0000)
commit504279ae0a0ce28ad37f820dcdb7f6557aabef7c
treec15d0cea7850e25709d4fa0dd813b4a270bbb203
parent812b3c62a7c8b3c006196635d0da08acc5674d98
Simplify the implementation of HARD_REG_SET

We have two styles of HARD_REG_SET: a single integer based on
HOST_WIDEST_FAST_INT (used when FIRST_PSEUDO_REGISTER is small enough)
or an array of integers.  One of the nice properties of this arrangement
is that:

  void foo (const HARD_REG_SET);

is passed by value as an integer when the set is small enough and
by reference otherwise.

(This is in constrast to "const HARD_REG_SET &", which would always
be passed by reference, and in contrast to passing a structure wrapper
like "struct s { T elts[1]; }" by value, where the structure might be
passed like a T or by reference, depending on the ABI.)

However, one of the disadvantages of using an array is that simple
assignment doesn't work.  We need to use COPY_HARD_REG_SET instead.

This patch uses a structure wrapper around the array, and preserves
the above "nice property" using a new const_hard_reg_set typedef.
The patch also removes the manual unrolling for small array sizes;
I think these days we can rely on the compiler to do that for us.

This meant fixing two port-specific quirks:

- epiphany passed NULL as a HARD_REG_SET whose value doesn't matter.
  The patch passes the NO_REGS set instead.

- ia64 reused TEST_HARD_REG_BIT and SET_HARD_REG_BIT for arrays that
  are bigger than HARD_REG_SET.  The patch just open-codes them.

The patch is probably being too conservative.  Very few places actually
take advantage of the "nice property" above, and we could have a
cleaner interface if we used a structure wrapper for all cases.

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

gcc/
* hard-reg-set.h (HARD_REG_SET): Define using a typedef rather
than a #define.  Use a structure rather than an array as the
fallback definition.  Remove special cases for low array sizes.
(const_hard_reg_set): New typedef.
(hard_reg_set_subset_p): Use it instead of "const HARD_REG_SET".
(hard_reg_set_equal_p, hard_reg_set_intersect_p): Likewise.
(hard_reg_set_empty_p): Likewise.
(SET_HARD_REG_BIT): Use a function rather than a macro to
handle the case in which HARD_REG_SET is a structure.
(CLEAR_HARD_REG_BIT, TEST_HARD_REG_BIT, CLEAR_HARD_REG_SET)
(SET_HARD_REG_SET, COPY_HARD_REG_SET, COMPL_HARD_REG_SET)
(AND_HARD_REG_SET, AND_COMPL_HARD_REG_SET, IOR_HARD_REG_SET)
(IOR_COMPL_HARD_REG_SET): Likewise.
(hard_reg_set_iterator::pset): Constify the pointer target.
(hard_reg_set_iter_init): Take a const_hard_reg_set instead
of a "const HARD_REG_SET".  Update the handling of non-integer
HARD_REG_SETs.
* recog.h: Test HARD_CONST instead of CLEAR_HARD_REG_SET.
* reload.h: Likewise.
* rtl.h (choose_hard_reg_mode): Remove unnecessary line break.
* regs.h (in_hard_reg_set_p): Take a const_hard_reg_set instead
of a "const HARD_REG_SET".
(overlaps_hard_reg_set_p, range_overlaps_hard_reg_set_p): Likewise.
(range_in_hard_reg_set_p): Likewise.
* ira-costs.c (restrict_cost_classes): Likewise.
* shrink-wrap.c (move_insn_for_shrink_wrap): Likewise.
* config/epiphany/resolve-sw-modes.c (pass_resolve_sw_modes::execute):
Pass a NO_REGS HARD_REG_SET rather than NULL to emit_set_fp_mode.
* config/ia64/ia64.c (rws_insn): In the CHECKING_P version,
use unsigned HOST_WIDEST_FAST_INT rather than HARD_REG_ELT_TYPE.
(rws_insn_set, rws_insn_test): In the CHECKING_P version,
take an unsigned int and open-code the HARD_REG_SET operations.

From-SVN: r275526
gcc/ChangeLog
gcc/config/epiphany/resolve-sw-modes.c
gcc/config/ia64/ia64.c
gcc/hard-reg-set.h
gcc/ira-costs.c
gcc/recog.h
gcc/regs.h
gcc/reload.h
gcc/rtl.h
gcc/shrink-wrap.c