From: Hans-Peter Nilsson <hans-peter.nilsson@axis.com>
To: rdsandiford@googlemail.com
Cc: iant@google.com, zadeck@naturalbridge.com,
gcc-patches@gcc.gnu.org, mikestump@comcast.net,
Kenneth.Zadeck@naturalbridge.com, avr@gjlay.de,
ramana.radhakrishnan@linaro.org
Subject: Re: Fix gcc.dg/lower-subreg-1.c failure
Date: Wed, 09 May 2012 16:37:00 -0000 [thread overview]
Message-ID: <201205091636.q49Gar0j017309@ignucius.se.axis.com> (raw)
In-Reply-To: <g4zk9hd8c6.fsf@richards-thinkpad.stglab.manchester.uk.ibm.com> (message from Richard Sandiford on Wed, 9 May 2012 11:14:49 +0200)
> From: Richard Sandiford <rdsandiford@googlemail.com>
> Date: Wed, 9 May 2012 11:14:49 +0200
> Hans-Peter Nilsson <hans-peter.nilsson@axis.com> writes:
> >> From: Richard Sandiford <rdsandiford@googlemail.com>
> >> Date: Tue, 1 May 2012 16:46:38 +0200
> >
> >> To repeat: as things stand, very few targets define proper rtx costs
> >> for SET.
> >
> > IMHO it's wrong to start blaming targets when rtx_cost doesn't
> > take the mode in account in the first place, for the default
> > cost. (Well, except for the modes-tieable subreg special-case.)
> > The targets where an operation in N * word_mode costs no more
> > than one in word_mode, if there even is one, is a minority,
> > let's adjust the defaults to that.
>
> I'll pass on approving or disapproving this patch, but for the record:
> a factor of word_mode seems a bit too simplistic.
I'd say it's the right level: simplistic enough for the default,
not to mention now linear, without being plainly ignorant as now.
> It's OK for moves
> and logic ops, but addition of multiword modes is a bit more complicated.
How about (same factor) factor*COSTS_N_INSNS (1)*3/2 to account
for carry? Or is 2*factor a better default? Further
improvements are welcome, but I see the patch as a strict
improvement and I hope it will not be shot down by requests for
perfection - at least not without detailing said perfection.
> Multiplication and division by multiword modes is even more
> so, of course.
Suggestions are welcome, but in the absence of that, I'd say any
factor larger than one is is a good start, like in the patch.
> > I think there should be a gating check whether the target
> > implements that kind of shift in that mode at all, before
> > checking the cost. Not sure whether it's generally best to put
> > that test here, or to make the rtx_cost function return the cost
> > of a libcall for that mode when that happens. Similar for the
> > other insns.
>
> This has come up a few times in past discussions about rtx_cost
> (as I'm sure you remember :-)). On the one hand, I agree it might
> be nice to shield the backend from invalid insns. That would
> effectively mean calling expand on each insn though, which would be
> rather expensive.
No, nothing that complicated. I'm thinking of just basically
checking that there's an operation in that mode, like:
if (direct_optab_handler (code_to_optab [GET_CODE (x)], mode)
== CODE_FOR_nothing)
{
... return tabled default cost; for libcall or open-code ...
}
Restricting the validity-gating to checking that the mode is
valid for the operation wouldn't interfere with fancy pipeline
speculative use.
> So I think this patch is using rtx_cost according to its current
> interface.
The "interface" use previously ignored the mode for most uses
(QED), so that's not completely correct. ;)
> If someone wants to change or restrict that interface,
> than that's a separate change IMO. And it should be done consistently
> rather than in this one place.
>
> In this case it doesn't matter anyway. If we never see a shift
> in mode M by amount X, we'll never need to make a decision about
> whether to split it.
If it's never used, then I don't mind it being wrong if that
simplifies the computation. :)
> > Isn't the below better than doing virtually the same in each
> > target's rtx_costs?
>
> FWIW, MIPS, SH and x86 all used something slightly different (and more
> complicated). I imagine PowerPC and SPARC will too. So "each" seems
> a bit strong.
>
> That's not an objection to the patch though. I realise some ports do
> have very regular architectures where every register is the same width
> and has the same move cost.
Hence the default should follow a very regular model...
brgds, H-P
next prev parent reply other threads:[~2012-05-09 16:37 UTC|newest]
Thread overview: 35+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <4F67D595.9030700@naturalbridge.com>
[not found] ` <mcrd387n9bf.fsf@dhcp-172-18-216-180.mtv.corp.google.com>
[not found] ` <4F6881EA.9080306@naturalbridge.com>
[not found] ` <mcrvclzl7b7.fsf@dhcp-172-18-216-180.mtv.corp.google.com>
[not found] ` <4F6889CC.8080302@naturalbridge.com>
[not found] ` <mcrr4wnl6lb.fsf@dhcp-172-18-216-180.mtv.corp.google.com>
2012-03-29 21:10 ` [C Patch]: pr52543 Kenneth Zadeck
2012-03-30 8:34 ` Ramana Radhakrishnan
2012-03-30 10:09 ` Richard Sandiford
2012-03-30 15:25 ` Kenneth Zadeck
2012-03-30 15:41 ` Richard Sandiford
2012-03-31 16:21 ` Kenneth Zadeck
2012-04-03 13:53 ` Richard Sandiford
2012-04-03 15:33 ` Kenneth Zadeck
2012-04-03 19:20 ` Richard Sandiford
2012-04-03 20:36 ` Ian Lance Taylor
2012-05-01 14:47 ` Richard Sandiford
2012-05-01 17:51 ` H.J. Lu
2012-05-03 19:52 ` Georg-Johann Lay
2012-05-03 22:14 ` Mike Stump
2012-05-04 23:02 ` Georg-Johann Lay
2012-05-06 18:55 ` [committed] Fix lower-subreg cost calculation Richard Sandiford
2012-05-08 15:26 ` Richard Earnshaw
2012-05-08 21:42 ` Richard Sandiford
2012-05-09 9:48 ` Richard Earnshaw
2012-05-07 19:01 ` [C Patch]: pr52543 Mike Stump
2012-05-09 6:02 ` Fix gcc.dg/lower-subreg-1.c failure (was: [C Patch]: pr52543) Hans-Peter Nilsson
2012-05-09 9:15 ` Fix gcc.dg/lower-subreg-1.c failure Richard Sandiford
2012-05-09 16:37 ` Hans-Peter Nilsson [this message]
2012-07-07 23:03 ` Fix gcc.dg/lower-subreg-1.c failure, revisited Hans-Peter Nilsson
2012-07-08 12:14 ` Richard Sandiford
2012-07-12 21:14 ` Hans-Peter Nilsson
2012-05-16 6:25 ` ping: Fix gcc.dg/lower-subreg-1.c failure (was: [C Patch]: pr52543) Hans-Peter Nilsson
2012-05-23 4:42 ` ping*2: " Hans-Peter Nilsson
2012-05-30 2:49 ` ping*3: " Hans-Peter Nilsson
2012-06-07 5:39 ` ping*4: " Hans-Peter Nilsson
2012-04-03 16:22 ` [C Patch]: pr52543 Ian Lance Taylor
2012-05-10 6:45 ` Paolo Bonzini
2012-05-10 6:54 ` Paolo Bonzini
[not found] ` <CACUk7=XVO=yHrPBFtAVfPxMtViYthtGGfuQSVGHNqHE7ibER0g@mail.gmail.com>
2012-03-30 19:29 ` Kenneth Zadeck
2012-03-30 22:38 ` Ramana Radhakrishnan
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=201205091636.q49Gar0j017309@ignucius.se.axis.com \
--to=hans-peter.nilsson@axis.com \
--cc=Kenneth.Zadeck@naturalbridge.com \
--cc=avr@gjlay.de \
--cc=gcc-patches@gcc.gnu.org \
--cc=iant@google.com \
--cc=mikestump@comcast.net \
--cc=ramana.radhakrishnan@linaro.org \
--cc=rdsandiford@googlemail.com \
--cc=zadeck@naturalbridge.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).