From: HAO CHEN GUI <guihaoc@linux.ibm.com>
To: Segher Boessenkool <segher@kernel.crashing.org>
Cc: gcc-patches <gcc-patches@gcc.gnu.org>,
Bill Schmidt <wschmidt@linux.ibm.com>, David <dje.gcc@gmail.com>
Subject: Re: [PATCH, rs6000] punish reload of lfiwzx when loading an int variable [PR102169, PR102146]
Date: Thu, 14 Oct 2021 16:37:11 +0800 [thread overview]
Message-ID: <250fc105-306f-5b87-a0d1-d35c7554c993@linux.ibm.com> (raw)
In-Reply-To: <20211014001222.GE614@gate.crashing.org>
On 14/10/2021 上午 8:12, Segher Boessenkool wrote:
> On Wed, Sep 29, 2021 at 04:32:19PM +0800, HAO CHEN GUI wrote:
>> The patch punishes reload of alternative pair of "d, Z" for
>> movsi_internal1. The reload occurs if 'Z' doesn't match and generates an
>> additional insn. So the memory reload should be punished.
> As David says, why only for loads? But also, why not for lxsiwzx (and
> stxsiwx) as well?
>
> But, what for all other uses of lfiwzx? And lfiwax?
>
> We need to find out why the register allocator considers it a good idea
> to use FP regs here, and fix *that*?
Let me explain why it uses FP regs.
In ira pass, the cost of general, float and altivec registers are all zero. So it prefers 'GEN_OR_VSX_REGS' class and is finally assigned register vs32. Not sure if the altivec registers are preferable for 'GEN_OR_VSX_REGS'.
r120: preferred GEN_OR_VSX_REGS, alternative NO_REGS, allocno GEN_OR_VSX_REGS
a0(r120,l0) costs: BASE_REGS:0,0 GENERAL_REGS:0,0 FLOAT_REGS:0,0 ALTIVEC_REGS:0,0 VSX_REGS:4000,4000 GEN_OR_FLOAT_REGS:6000,6000 GEN_OR_VSX_REGS:6000,6000 LINK_REGS:12000,12000 CTR_REGS:12000,12000 LINK_OR_CTR_REGS:12000,12000 SPEC_OR_GEN_REGS:12000,12000 ALL_REGS:36000,36000 MEM:8000,8000
Allocno a0r120 of GEN_OR_VSX_REGS(93) has 93 avail. regs 0 3-12 14-95, node: 0 3-12 14-95 (confl regs = 1-2 13 96-110)
Forming thread from colorable bucket:
Pushing a0(r120,l0)(cost 0)
Popping a0(r120,l0) -- assign reg 32
Disposition:
0:r120 l0 32
In reload pass, the third alternative pair is "r,m" and the forth pair is "d,Z". As the 'r' doesn't match the class of reg vs32, it needs a output reload and got a "reject++" as well as a "losers". For pair "d,Z", the second operands 'Z' doesn't match. It needs a input reload and got a "losers". But the memory reload is not punished if there is no addition modifies with the alternative. So it only got a "addr_losers". Overall cost of "r,m" is great than "d,Z". So it picked up FP register and 'lfiwzx' instruction.
0 Non input pseudo reload: reject++
alt=2,overall=7,losers=1,rld_nregs=1
alt=3,overall=6,losers=1,rld_nregs=0
Gui Haochen
>
> The extra insn you talk about is because this insn only allows indexed
> addressing ([reg+reg] or [reg] addressing). That is true for very many
> insns. Reload (well, LRA in the modern world) should know about such
> extra costs. Does it not?
>
>> gcc/
>> * gcc/config/rs6000/rs6000.md (movsi_internal1): disparages
>> slightly the alternative 'Z' of "lfiwzx" when reload is
>> needed.
> "Disparage", no "s". Changelog entries are written in the imperative.
>
>
> Segher
next prev parent reply other threads:[~2021-10-14 8:37 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-09-29 8:32 HAO CHEN GUI
2021-10-11 5:46 ` Ping^1 " HAO CHEN GUI
2021-10-14 0:12 ` Segher Boessenkool
2021-10-14 8:37 ` HAO CHEN GUI [this message]
2021-10-13 15:21 David Edelsohn
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=250fc105-306f-5b87-a0d1-d35c7554c993@linux.ibm.com \
--to=guihaoc@linux.ibm.com \
--cc=dje.gcc@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=segher@kernel.crashing.org \
--cc=wschmidt@linux.ibm.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).