From: Uros Bizjak <ubizjak@gmail.com>
To: "H.J. Lu" <hjl.tools@gmail.com>
Cc: gcc-patches@gcc.gnu.org, Jakub Jelinek <jakub@redhat.com>
Subject: Re: PATCH: PR target/40838: gcc shouldn't assume that the stack is aligned
Date: Thu, 15 Oct 2009 19:53:00 -0000 [thread overview]
Message-ID: <4AD77D3A.2040502@gmail.com> (raw)
In-Reply-To: <6dc9ffc80910151232t20f3fbbfw97b063a0ab33c5ce@mail.gmail.com>
On 10/15/2009 09:32 PM, H.J. Lu wrote:
>>>> We should use PREFERRED_STACK_BOUNDARY_DEFAULT instead of hardcoded 128
>>>> here. At least the comment for P_S_B_D says:
>>>>
>>>> /* It should be MIN_STACK_BOUNDARY. �But we set it to 128 bits for
>>>> � both 32bit and 64bit, to support codes that need 128 bit stack
>>>> � alignment for SSE instructions, but can't realign the stack. �*/
>>>> #define PREFERRED_STACK_BOUNDARY_DEFAULT 128
>>>>
>>>>
>>> In forced_stack_alignment, we collect the hard alignment requirement
>>> on stack. The decision if stack should be aligned depends on the
>>> incoming stack alignment and the hard stack alignment. It is separate
>>> from PREFERRED_STACK_BOUNDARY_DEFAULT.
>>>
>>>
>>>
>>>> OTOH, I think that this whole stuff should depend on -mstackrealing
>>>> somehow,
>>>> according to the comment:
>>>>
>>>> /* 1 if -mstackrealign should be turned on by default. �It will
>>>> � generate an alternate prologue and epilogue that realigns the
>>>> � runtime stack if nessary. �This supports mixing codes that keep a
>>>> � 4-byte aligned stack, as specified by i386 psABI, with codes that
>>>> � need a 16-byte aligned stack, as required by SSE instructions. �If
>>>> � STACK_REALIGN_DEFAULT is 1 and PREFERRED_STACK_BOUNDARY_DEFAULT is
>>>> � 128, stacks for all functions may be realigned. �*/
>>>> #define STACK_REALIGN_DEFAULT 0
>>>>
>>>> As far as I understand from many comments from the PR40838 trail
>>>> (especially
>>>> comment #51), -mstackrealing is not effective in some cases involving
>>>> automatic SSE variables on the stack. I think that -mstackrealign should
>>>> be
>>>> fixed in the way your patch outlines, so old/broken sources that assume
>>>> 4byte alignment can be compiled using this option without penalizing
>>>> new/fixed code that assumes 16byte alignment.
>>>>
>>>>
>>>>
>>> -mstackrealign is very effective. If we turn it on by default, it works
>>> fine:
>>>
>>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40838#c45
>>>
>>> It is just that we don't want to make it the default since we will realign
>>> the stack for almost all functions. In most cases, misaligned stack won't
>>> cause any problems.
>>>
>>> What my patch does is to assume 4byte incoming stack alignment
>>> for functions containing SSE instructions which require hard 16byte
>>> stack alignment.
>>>
>>> After my patch is applied, we can update -mstackrealign to make it an
>>> no-op since all the problems it tries to solve:
>>>
>>> http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00854.html
>>> http://gcc.gnu.org/ml/gcc-patches/2006-05/msg00526.html
>>>
>>> disappeared with my patch.
>>>
>>>
>> Then we should do realignment the other way around: instead of using
>> -mstackrealing for all the code (including where it has no effect), let's
>> use -mstackrealign to activate realignment functionality that is introduced
>> by your patch.
>>
> That defeats the whole purpose of my patch, which automatically
> realigns the stack when there is a hard alignment requirement. If it
> isn't turned on by default, it is not very useful.
>
>
>> IOW, lightweight -mstackrealign, firing up only when there is the
>> possibility of unaligned access in the code it precedes.
>>
>>
> That is what my patch does, but turned it on by default.
>
I think that we have the same situation here as was with the infamous
-mcld option. A specific functionality is needed for compatibility with
some [broken] code and this functionality has non-negligible impact on
the performance. AFAICS, there will always be opinions for this
functionality as well as opinions against it.
I propose that we implement new option -mhard-stackrealign [we can
bikeshed about this option name a bit ;) ] with corresponding
--enable-hard-stackrealign as a configure option. This way, both groups
can have whatever they prefer - compatibility vs. performance.
Looking at -mcld discussions, is this acceptable solution?
Uros.
next prev parent reply other threads:[~2009-10-15 19:51 UTC|newest]
Thread overview: 59+ messages / expand[flat|nested] mbox.gz Atom feed top
2009-08-06 21:42 H.J. Lu
2009-08-06 22:26 ` Jakub Jelinek
2009-08-06 22:52 ` H.J. Lu
2009-10-15 15:58 ` H.J. Lu
2009-10-15 18:45 ` Uros Bizjak
2009-10-15 19:22 ` H.J. Lu
2009-10-15 19:32 ` Uros Bizjak
2009-10-15 19:43 ` H.J. Lu
2009-10-15 19:48 ` Jakub Jelinek
2009-10-15 20:11 ` H.J. Lu
2009-10-15 19:53 ` Uros Bizjak [this message]
2009-10-15 21:01 ` H.J. Lu
2009-10-15 21:41 ` Uros Bizjak
2009-10-16 20:27 ` H.J. Lu
2009-10-17 1:03 ` Ian Lance Taylor
2009-10-17 18:22 ` H.J. Lu
2009-10-17 19:02 ` Richard Guenther
2009-10-17 19:21 ` H.J. Lu
2009-10-17 19:29 ` Richard Guenther
2009-10-17 19:35 ` H.J. Lu
2009-10-17 19:46 ` Richard Guenther
2009-10-17 20:01 ` H.J. Lu
2009-10-17 20:59 ` Richard Guenther
2009-10-18 19:21 ` Michael Matz
2009-10-18 19:45 ` Richard Guenther
2009-10-19 16:36 ` H.J. Lu
2009-10-20 1:12 ` Michael Matz
2009-10-20 19:10 ` H.J. Lu
2009-10-19 16:38 ` H.J. Lu
2009-10-19 17:08 ` Ian Lance Taylor
2009-10-19 17:26 ` H.J. Lu
2009-10-19 17:33 ` Ian Lance Taylor
2009-10-19 17:46 ` H.J. Lu
2009-10-19 17:55 ` Ian Lance Taylor
2009-10-19 19:16 ` H.J. Lu
2009-10-19 21:15 ` Ian Lance Taylor
2009-10-20 19:00 ` H.J. Lu
2009-10-20 1:23 ` Michael Matz
2009-10-20 19:12 ` H.J. Lu
2009-10-20 1:53 ` Michael Matz
2009-10-20 21:15 ` H.J. Lu
2009-10-21 1:10 ` H.J. Lu
2009-10-21 9:54 ` Michael Matz
2009-10-21 16:56 ` H.J. Lu
2009-10-30 10:08 ` Richard Guenther
2009-10-17 7:09 ` Uros Bizjak
2009-08-07 0:54 Mikulas Patocka
2009-08-07 7:13 ` Jakub Jelinek
2009-08-07 12:53 ` H.J. Lu
2009-08-07 22:30 ` H.J. Lu
2009-08-08 17:35 ` Mikulas Patocka
2009-08-16 21:25 ` H.J. Lu
2009-08-24 17:39 ` H.J. Lu
2009-09-12 23:32 ` Mikulas Patocka
2009-09-12 23:42 ` Mikulas Patocka
2009-09-13 1:55 ` H.J. Lu
2009-09-13 14:10 ` Mikulas Patocka
2009-08-07 21:08 ` Mikulas Patocka
2009-08-07 21:25 ` Richard Guenther
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=4AD77D3A.2040502@gmail.com \
--to=ubizjak@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hjl.tools@gmail.com \
--cc=jakub@redhat.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).