From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
To: Aleksandar Markovic <aleksandar.m.mail@gmail.com>
Cc: stefan.brankovic@rt-rk.com,
"qemu-ppc@nongnu.org" <qemu-ppc@nongnu.org>,
Paul Clarke <pc@us.ibm.com>, qemu-devel <qemu-devel@nongnu.org>
Subject: Re: target/ppc: bug in optimised vsl/vsr implementation?
Date: Tue, 1 Oct 2019 19:24:42 +0100 [thread overview]
Message-ID: <c9679b01-91c3-3d69-fb38-dfef1602dcf4@ilande.co.uk> (raw)
In-Reply-To: <CAL1e-=gcK2mdtrt9vibHGpbm4_FZgQWTA91+p=9ouuMYmZwPqQ@mail.gmail.com>
On 28/09/2019 18:45, Aleksandar Markovic wrote:
Hi Aleksandar,
Thanks for taking a look at this!
> Mark and Paul (and Stefan),
>
> Thanks for spotting this and pinpointing the culprit commit. I guess Stefan is going
> to respond soon, but, in the meantime, I took a look at the commit in question:
>
> https://github.com/qemu/qemu/commit/4e6d0920e7547e6af4bbac5ffe9adfe6ea621822
>
> I don't have at the moment any dev/test environment handy, but I did manual
> inspection of the code, and here is what I found (in order of importance, perceived
> by me):
>
> 1. The code will not work correctly if the shift ammount (variable 'sh') is 0. This
> is because, in that case, one of succeeding invocations of TCG shift functions will
> be required to shift a 64-bit TCG variable by 64 bits, and the result of such TCG
> operation is undefined (shift amount must be 63 or less) - see
> https://github.com/qemu/qemu/blob/master/tcg/README.
Yes I think you're right here - the old helper got around this by doing an explicit
copy from a to r if the shift value is zero. In fact the case that Paul reported is
exactly this:
vsl VRT, VRA, VRB
=> 0x100006e0 <vec_slq+132>: vsl v0,v0,v1
(gdb) p $vr0.uint128
$21 = 0x10111213141516172021222324252650
(gdb) p $vr1.uint128
$22 = 0x0
(gdb) stepi
0x00000000100006e4 in vec_slq ()
1: x/i $pc
=> 0x100006e4 <vec_slq+136>: xxlor vs0,vs32,vs32
(gdb) p $vr0.uint128
$23 = 0x10111213141516172021222324252650
I guess the solution is check for sh == 0 and if this is the case, execute a copy
instead.
> 2. Variable naming is better in the old helper than in the new translator. In that
> light, I would advise Stefan to change 'sh' to 'shift', and 'shifted' to 'carry'.
It looks like the name "sh" comes from the ISA documentation, so whilst it's a little
tricky to compare with the previous implementation it does make sense when comparing
with the algorithm shown there. Note: this implementation also drops the check for
each byte of VRB having the same shift value - should we care about this?
> 3. Lines
>
> tcg_gen_andi_i64(shifted, shifted, 0x7fULL);
>
> and
>
> tcg_gen_andi_i64(shifted, shifted, 0xfe00000000000000ULL);
>
> appear to be spurious (albait in a harmless way). Therefore, they should be deleted,
> or, alternatevely, a justification for them should be provided.
I'm not sure why they are needed either - there's certainly no mention of it in the
ISA documentation. Stefan?
> 4. In the commit message, variable names were used without quotation mark, resulting
> in puzzling and unclear wording.
>
> 5. (a question for Mark) After all recent changes, does get_avr64(..., ..., true)
> always (for any endian configuration) return the "high" half of an Altivec register,
> and get_avr64(..., ..., false) the "low" one?
Yes - the new functions always return the MSB (high) and LSB (low) correctly
regardless of host endian.
> Given all these circumstances, perhaps the most reasonable solution would be to
> revert the commit in question, and allow Stefan enough dev and test time to hopefully
> submit a new, better, version later on.
Given that it has been broken for 3 months now, I don't think we're in any major rush
to revert ASAP. I'd prefer to give Stefan a bit more time first since he does report
some substantial speed improvements from these new implementations.
ATB,
Mark.
next prev parent reply other threads:[~2019-10-01 18:26 UTC|newest]
Thread overview: 13+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-09-26 18:04 target/ppc: bug in optimised vsl/vsr implementation? Mark Cave-Ayland
2019-09-28 17:45 ` Aleksandar Markovic
2019-09-28 22:17 ` Aleksandar Markovic
2019-09-30 14:34 ` Paul Clarke
2019-09-30 14:53 ` Aleksandar Markovic
2019-09-30 14:37 ` Aleksandar Markovic
2019-10-01 18:24 ` Mark Cave-Ayland [this message]
2019-10-02 14:08 ` Stefan Brankovic
2019-10-03 11:11 ` Stefan Brankovic
2019-10-02 17:38 ` Alex Bennée
2019-10-02 19:40 ` Richard Henderson
2019-10-02 19:55 ` Paul Clarke
2019-10-04 19:32 ` Aleksandar Markovic
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=c9679b01-91c3-3d69-fb38-dfef1602dcf4@ilande.co.uk \
--to=mark.cave-ayland@ilande.co.uk \
--cc=aleksandar.m.mail@gmail.com \
--cc=pc@us.ibm.com \
--cc=qemu-devel@nongnu.org \
--cc=qemu-ppc@nongnu.org \
--cc=stefan.brankovic@rt-rk.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 NNTP newsgroup(s).