Re: [PATCH] Don't combine param and return value copies

[Segher Boessenkool <segher at kernel dot crashing dot org>]

On Mon, May 25, 2015 at 10:26:35AM +0930, Alan Modra wrote:
> On Sat, May 23, 2015 at 08:45:23AM -0500, Segher Boessenkool wrote:
> > Thanks.  Did you see improvements around return as well, or mostly /
> > only related to the function start?
> 
> The rlwinm vs. rldicl change was in dwarf2out.c add_ranges_num for r3
> one insn before a blr.  I'm sure that was due to not combining return
> value copies.  In fact, I'm sure all the improvements I saw were due
> to changing the exit..  See below.
> 
> > > +/* Twiddle BLOCK_FOR_INSN to TO for instructions in the first block BB
> > > +   that we don't want to combine with other instructions.  */
> > > +
> > > +static void
> > > +twiddle_first_block (basic_block bb, basic_block to)
> > 
> > I don't like this much.  Messing with global state makes things harder
> > to change later.  If it is hard to come up with a good name for a
> > factor, it usually means it is not such a good factor.
> > 
> > You can do these inside can_combine_{def,use}_p as far as I can see?
> > Probably need to give those an extra parameter then: for _def, a bool
> > that says "don't allow moves from hard regs", set when the scan has
> > encountered a NOTE_INSN_FUNCTION_BEG in this bb; and for _use, a regset
> > of those hard regs we don't want to allow moves to (those seen in USEs
> > later in that same block).  Or do it in the main loop itself, _{def,use}
> > are each called in one place only; whatever works best.
> 
> Huh, that's the way I started writing this patch..  The reason I went
> with modifying BLOCK_FOR_INSN is that the loop setting up LOG_LINKS
> needs to test that anyway.  Changing code in the main loop means
> every insn in a function will need to be tested for additional
> conditions.  I thought that might be slower.  I can see your point
> though, especially if someone wanted to wean combine off LOG_LINKS.

The setup of the LOG_LINKS is a simple fast linear loop, much less
work than the rest of combine.  So don't worry about performance too
much :-)

> > What makes return values different from CALL args here, btw?  Why do
> > you not want to combine return values, but you leave alone call args?
> 
> I don't think there is much difference between SETs for return values
> and SETs for call args.  The reason I deliberately left them out is
> that in the original discussion we were talking about parameters and
> return values.  So I thought I'd better restrict the change to just
> those SETs.
> 
> It would be a much simpler patch to make combine ignore all SETs
> that are a reg,reg copy with one of them a hard reg.  I was a little
> worried I'd regress some target if I tried that.

_All_ moves to/from hard regs includes much more (register asm, fixed
registers, maybe some targets do weird things as well?).  So yes I share
those worries.

Since we do not want any other pass (before RA) to foul up these
register moves either, maybe a better solution would be to mark them
some way at expand time?

> (Results on
> powerpc64le-linux for such a change look good.  A lot more cases where
> code is better, due to catching the parameter reg,reg copies.  In
> looking at gcc/*.o I haven't yet seen any regressions in code quality.)

That is good news :-)

> > > +/* Fill in log links field for all insns that we wish to combine.  */
> > 
> > Please don't change this comment; it was more correct before.
> 
> But it wasn't correct!  LOG_LINKS are not set up for *all* insns.

It "sets" all LOG_LINKS to some value ("sets", it doesn't actually
zero them at the start, it has an assert for that though).

> > > @@ -1103,7 +1192,7 @@ create_log_links (void)
> > >  		 we might wind up changing the semantics of the insn,
> > >  		 even if reload can make what appear to be valid
> > >  		 assignments later.  */
> > > -	      if (regno < FIRST_PSEUDO_REGISTER
> > > +	      if (HARD_REGISTER_NUM_P (regno)
> > >  		  && asm_noperands (PATTERN (use_insn)) >= 0)
> > >  		continue;
> > 
> > An independent change.
> 
> Yeah, I guess.  I tidy these if I'm working on a piece of code.

Pretty far away in this case ;-)

> Here's the whole file done.  Boostrapped and regression tested
> powerpc64le-linux and x86_64-linux.
> 
> 	* combine.c: Use HARD_REGISTER_P and HARD_REGISTER_NUM_P as
> 	appropriate throughout file in place of comparing regno
> 	against FIRST_PSEUDO_REGISTER.

Have we decided we want to convert the whole compiler to this?  It
is a pretty lousy interface IMO: HARD_REGISTER_P does not check if
its arg is a register; it does not check if its arg is a hard register
either (it checks if it is not a pseudo); it cannot be used in all
places in all passes because of that (which means, not in all macros
and hooks either).


Segher