https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53513
Oleg Endo <olegendo at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #33691|0 |1
is obsolete| |
--- Comment #23 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Created attachment 33716
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=33716&action=edit
Using virtual FPSCR registers to model insn dependencies
This is a somewhat larger patch which does a couple of things to address the
deficits of the previous approach.
1) The hard reg set is extended by two more registers:
- FPSCR_MODES_REG represents any mode bits in FPSCR, which are used by fp
insns.
- FPSCR_STAT_REG represents the status bits in FPSCR, which are written by fp
insns.
All fp insns that previously had a
(use (match_operand:PSI 2 "fpscr_operand" "")
now get a
(use (reg:SI FPSCR_MODES_REG))
2) The 'fpu_switch' insn, which is just a FPSCR load/store, now uses and sets
the virtual regs FPSCR_MODES_REG and FPSCR_STAT_REG. This creates a sort of
reordering barrier due to the reg uses/sets.
3) Two new dummy insns extend_psi_si and truncate_si_psi are added to convert
PSIMode <-> SImode, because we can do only logic on SImode, but FPSCR is
described as PSImode. This hack could be removed later maybe.
4) The insns toggle_sz and toggle_pr are now also setting the
virtual reg FPSCR_MODES_REG.
5) The fsca insn needed some adjustments for the operand matching, as combine
started going different paths. Some of the fsca tests in gcc.target/sh were
failing and are fixed by this.
6) sh_emit_set_t_insn is adjusted to emit the correct comparison insns with
use/set of virtual FPSCR regs.
7) calc_live_regs is adjusted to ignore virtual regs FPSCR_MODES_REG and
FPSCR_STAT_REG, or else it will try to generate push/pop insns for those regs.
8) sh_emit_mode_set now emits the FPSCR store-modify-load insn sequence instead
of loading FPSCR from __fpscr_values.
9) Some redundant fp insns and related functions are deleted.
I've tried to keep the patch as short as possible and not do too many unrelated
changes. I haven't fully tested the patch, so there are probably a couple of
fallouts.
There are two regressions with this patch. The FPSCR loads are not put into
delay slots anymore. Probably because the 'fpu_switch' pattern now has too
many sets and the DBR rejects that as a slot candidate. The other thing are
the failures in gcc.target/sh w.r.t. interrupt functions, which can be
addressed later.
Kaz, could you please have an early look at it?
