This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: hard register reload patch
- To: law at cygnus dot com
- Subject: Re: hard register reload patch
- From: Joern Rennecke <amylaar at cygnus dot co dot uk>
- Date: Mon, 13 Dec 1999 20:15:20 +0000 (GMT)
- Cc: amylaar at cygnus dot co dot uk, gcc-patches at gcc dot gnu dot org
> Don't we usually write this loop using a "while" so that we strip off any
> nested SUBREGs?
We use a "while" when we also strip STRICT_LOW_PARTs / ZERO_EXTRACTs /
SIGN_EXTRACTs . We don't expect these in a call insn, so we don't handle
them here. SUBREGs are not supposed to be nested.
> > + regno = REGNO (reg) + word;
> > + endregno = regno + HARD_REGNO_NREGS (regno, mode);
> Is "mode" really correct here? You're going to get the mode of the outer
> SUBREG. I have no idea if that's what you wanted or not.
It is what I wanted.
> I didn't see what ChangeLog entry corresponded to pulling that code out
> of the loop itself [ I'm glad you pulled that code out of the loop; I have
> no idea why I wrote the code in that manner in the first place. ]
It is this one:
(insert_restore, insert_save): Add restored / saved registers to
dead_or_set.
I had to insert the insert_one_insn call before the loop over the hard
registers, so I had the choice between putting more non-looping code
into the loop or pulling some out.
> In insert_one_insn, do you need to update the REG_DEAD notes, or does reload
> not care about them?
>
> It also seems to me that you got the live_throughout wrong. The register
> dies at the save insn, right? Then why set new->live_throughout. I think
> we've got the mirror case to consider at the restore. [ I'm talking about
> the code in insert_one_insn still. ]
The register might continue to be live till the call_insn, if it is used
as a function argument. We could test for that, but that would be another
optimization that we don't currently do. Likewise for adding REG_DEAD notes.
> In reg_becomes_live, when setting regs_set you special cased fixed registers
> when the original register was a hard register, but not when the original was
> a pseudo, why? Similarly in reg_dies.
We don't allocate pseudos to fixed registers.
> Have you bootstrapped with this change? Have you done any timing tests to
> see if reload/global allocation is faster?
I have bootstrapped with this change, but I haven't run any timing tests.