linux-arm-kernel.lists.infradead.org archive mirror
 help / color / mirror / Atom feed
From: w@1wt.eu (Willy Tarreau)
To: linux-arm-kernel@lists.infradead.org
Subject: gcc miscompiles csum_tcpudp_magic() on ARMv5
Date: Thu, 12 Dec 2013 16:28:49 +0100	[thread overview]
Message-ID: <20131212152849.GC31816@1wt.eu> (raw)
In-Reply-To: <yw1xsityhzd1.fsf@unicorn.mansr.com>

On Thu, Dec 12, 2013 at 03:18:18PM +0000, M?ns Rullg?rd wrote:
> Willy Tarreau <w@1wt.eu> writes:
> 
> >> >> Hmmm aren't you passing a 16-bit register directly to the ASM for
> >> >> being used as a 32-bit one ? This seems hasardous to me since
> >> >> nowhere you tell gcc how you're going to use the register.
> >> >
> >> > this is exactly what I'm complaining about, the arm code for
> >> > csum_tcpudp_nofold() in the kernel does exactly that.
> >> >
> >> >> Could you check if that fixes it :
> >> >> 
> >> >>  static inline uint32_t asm_add(uint16_t len, uint32_t sum)
> >> >>  {
> >> >>          uint32_t len32 = len;
> >> >
> >> > or change the asm_add() proto to take an "uint32_t len" instead, and yes
> >> > of course that fixes it.
> >> 
> >> It's a bug.  Please report it to the gcc developers.
> >
> > Here I don't agree with the generalization (and believe me I swear all the
> > day about gcc's bugs). It's a matter of ABI and availability or not of 16
> > bit registers or not. If the ASM supports 16-bit regs and the compiler is
> > allowed to emit 16-bit regs, then gcc will have no way to know whether it
> > must zero-extend the value first. If it's specified that "r" is necessarily
> > a 32-bit register then it should.
> 
> ARM has *only* 32-bit registers.
> 
> > Maybe the issue is in the ABI itself, I don't know if 16-bit values
> > are supposed to be zero-extended only when they're converted to 32-bit
> > or also when they're passed as arguments. The fact that it works
> > without inline may simply be a side effect of the different code (eg:
> > 16 lower bits of the register copied into another one).
> >
> > So one needs to look at the specs of the ABI to know where the 16-bit value
> > is supposed to be converted to 32-bit, then the faulty component must be
> > fixed (gcc or kernel code).
> 
> The ABI for function calls sign/zero-extends all arguments prior to the
> call.

OK then that's pretty clear, there's no ambiguity, thanks for the precision.

Willy

  reply	other threads:[~2013-12-12 15:28 UTC|newest]

Thread overview: 37+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2013-12-12 12:14 gcc miscompiles csum_tcpudp_magic() on ARMv5 Maxime Bizon
2013-12-12 12:40 ` Russell King - ARM Linux
2013-12-12 13:36   ` Maxime Bizon
2013-12-12 13:48     ` Måns Rullgård
2013-12-12 14:10       ` Maxime Bizon
2013-12-12 14:19         ` Willy Tarreau
2013-12-12 14:28           ` Maxime Bizon
2013-12-12 14:42             ` Måns Rullgård
2013-12-12 14:52               ` Maxime Bizon
2013-12-12 14:58                 ` Måns Rullgård
2013-12-12 15:00                 ` Russell King - ARM Linux
2013-12-12 15:26                   ` Maxime Bizon
2013-12-12 15:07               ` Willy Tarreau
2013-12-12 15:18                 ` Måns Rullgård
2013-12-12 15:28                   ` Willy Tarreau [this message]
2013-12-12 15:43                     ` Russell King - ARM Linux
2013-12-12 15:50                       ` Måns Rullgård
2013-12-12 14:37           ` Måns Rullgård
2013-12-12 14:40             ` Maxime Bizon
2013-12-12 14:47               ` Måns Rullgård
2013-12-12 14:26         ` Måns Rullgård
2013-12-12 14:48     ` Russell King - ARM Linux
2013-12-12 15:00       ` Måns Rullgård
2013-12-12 15:04       ` Maxime Bizon
2013-12-12 15:41         ` Russell King - ARM Linux
2013-12-12 16:04           ` Måns Rullgård
2013-12-12 16:04           ` Willy Tarreau
2013-12-12 16:47             ` Russell King - ARM Linux
2013-12-12 17:11               ` Willy Tarreau
2013-12-12 17:20                 ` Russell King - ARM Linux
2013-12-12 17:35                   ` Willy Tarreau
2013-12-12 18:07                   ` Nicolas Pitre
2013-12-12 22:30               ` Maxime Bizon
2013-12-12 22:36                 ` Russell King - ARM Linux
2013-12-12 22:44                   ` Maxime Bizon
2013-12-12 22:48                     ` Russell King - ARM Linux
2013-12-12 17:34           ` Maxime Bizon

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=20131212152849.GC31816@1wt.eu \
    --to=w@1wt.eu \
    --cc=linux-arm-kernel@lists.infradead.org \
    /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).