From: Jeff Law <jlaw@ventanamicro.com>
To: Manolis Tsamis <manolis.tsamis@vrull.eu>,
Jeff Law <jeffreyalaw@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>,
gcc-patches@gcc.gnu.org, Vineet Gupta <vineetg@rivosinc.com>,
Kito Cheng <kito.cheng@gmail.com>,
philipp.tomsich@vrull.eu
Subject: Re: [PATCH] Enable shrink wrapping for the RISC-V target.
Date: Wed, 2 Nov 2022 09:02:51 -0600 [thread overview]
Message-ID: <2cebf19f-d465-73f3-0334-523a79d5b306@ventanamicro.com> (raw)
In-Reply-To: <CAM3yNXrGybbD_o06dR+sG-2Ffo2PEEbQR9zERU3K7GmOEB2Puw@mail.gmail.com>
On 11/2/22 08:12, Manolis Tsamis wrote:
> On Wed, Oct 19, 2022 at 8:16 PM Jeff Law via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>>
>> On 10/18/22 11:35, Palmer Dabbelt wrote:
>>>> I would have expected things to work fine with libcalls, perhaps with
>>>> the exception of the save/restore libcalls. So that needs deeper
>>>> investigation.
>>> The save/restore libcalls only support saving/restoring a handful of
>>> register configurations (just the saved X registers in the order
>>> they're usually saved in by GCC). It should be OK for correctness to
>>> over-save registers, but it kind of just un-does the shrink wrapping
>>> so not sure it's worth worrying about at that point.
>>>
>>> There's also some oddness around the save/restore libcall ABI, it's
>>> not the standard function ABI but instead a GCC-internal one. IIRC it
>>> just uses the alternate link register (ie, t0 instead of ra) but I may
>>> have forgotten something else.
>> I hadn't really dug into it -- I was pretty sure they weren't following
>> the standard ABI based on its name and how I've used similar routines to
>> save space on some targets in the past. So if we're having problems
>> with shrink-wrapping and libcalls, those two might be worth investigating.
>>
>>
>> But I think the most important takeaway is that shrink wrapping should
>> work with libcalls, there's nothing radically different about libcalls
>> that would make them inherently interact poorly with shrink-wrapping.
>> So that aspect of the shrink-wrapping patch needs deeper investigation.
>>
>> Jeff
> I think I miscommunicated the issue previously because my understanding
> of libcalls wasn't very solid. The guard is against the save/restore libcalls
> specifically; other than that shrink wrapping and libcalls are fine.I think it
> makes sense to leave this check because the prologue/epilogue does
> something similar when using libcall save/restore:
> frame->mask = 0; /* Temporarily fib that we need not save GPRs. */
Looking more closely, yea, it might have been a miscommunication between
us WRT libcalls.
You're testing riscv_use_save_libcall, which is only going to kick in
when we're using that special function to do register saves/restores.
While we could, in theory, shrink-wrap that as well, I don't think it's
worth the additional headache.
>
> Since shrink wrap components are marked by testing frame->mask then
> no registers should be wrapped with the libcall save/restore if I understand
> correctly.
That's my understanding as well after looking at the code more closely.
Essentially the mask is set to zero for the duration of the call to
riscv_for_each_saved_regs which will effectively avoid shrink wrapping
in that case.
>
> Nonetheless, I tested what happens if this guard condition is removed
> and the result is that a RISCV test fails (riscv/pr95252.c). In that case
> a unnecessary save/restore of a register is emitted together with
> inconsistent cfi notes that make dwarf2cfi abort.
ACK. This is the kind of thing I was referring to above when I said we
probably could shrink wrap the call to save/restore registers, but it's
probably not worth the headache right now. I could see someone in the
embedded space one day trying to tackle this problem, but I don't think
we need to for this patch to go forward.
>
> To conclude, I believe that this makes the code in the commit fine since
> it only guards against the libcall save/restore case. But I may be still
> missing something about this.
I think you've got it figured out reasonably well.
So the final conclusion is libcalls are resolved to my satisfaction.
Jeff
next prev parent reply other threads:[~2022-11-02 15:02 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-09-06 10:39 mtsamis
2022-10-02 20:32 ` Palmer Dabbelt
2022-10-18 15:18 ` Manolis Tsamis
2022-10-18 15:57 ` Jeff Law
2022-10-18 17:35 ` Palmer Dabbelt
2022-10-19 17:15 ` Jeff Law
2022-10-20 7:42 ` Manolis Tsamis
2022-11-02 14:12 ` Manolis Tsamis
2022-11-02 15:02 ` Jeff Law [this message]
2022-10-20 7:35 ` Manolis Tsamis
2022-11-02 13:54 ` Manolis Tsamis
2022-11-02 15:06 ` Jeff Law
2022-11-03 0:26 ` Palmer Dabbelt
2022-11-03 22:23 ` Jeff Law
2022-11-07 22:07 ` Palmer Dabbelt
2022-11-13 1:32 ` Jeff Law
2022-11-16 10:26 ` Manolis Tsamis
2022-11-17 2:09 ` Jeff Law
2022-11-17 10:54 ` Manolis Tsamis
2022-11-17 11:59 ` Philipp Tomsich
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=2cebf19f-d465-73f3-0334-523a79d5b306@ventanamicro.com \
--to=jlaw@ventanamicro.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=jeffreyalaw@gmail.com \
--cc=kito.cheng@gmail.com \
--cc=manolis.tsamis@vrull.eu \
--cc=palmer@dabbelt.com \
--cc=philipp.tomsich@vrull.eu \
--cc=vineetg@rivosinc.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).