[Bug c/45804] New: ARM: unnecessary sign extension

[Bug c/45804] New: ARM: unnecessary sign extension

[felipe.contreras at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45804

           Summary: ARM: unnecessary sign extension
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: felipe.contreras@gmail.com


In the following code gcc add an unnecessary uxth operation.

static inline uint16_t read16_be(const uint8_t *p)
{
    uint16_t v;
    __asm__("ldrh %0, %1" : "=r"(v) : "m"(*(const uint16_t *)p));
    __asm__("rev16 %0, %0" : "+r"(v));
    return v;
}

Many other people have experienced this issue:
http://article.gmane.org/gmane.comp.video.ffmpeg.cvs/33470
http://article.gmane.org/gmane.comp.video.ffmpeg.devel/106781
https://bugs.launchpad.net/gcc-linaro/+bug/634682

[Bug target/45804] ARM: unnecessary sign extension

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45804

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Target|                            |arm-*-*
          Component|c                           |target
           Severity|normal                      |enhancement

[Bug target/45804] ARM: unnecessary sign extension

[rearnsha at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45804

Richard Earnshaw <rearnsha at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
                 CC|                            |rearnsha at gcc dot gnu.org
         Resolution|                            |INVALID

--- Comment #1 from Richard Earnshaw <rearnsha at gcc dot gnu.org> 2010-09-27 16:30:35 UTC ---
This particular case is invalid.  

The ARM ABI requires that short and char result types are extended to 32 bits
(using signed or unsigned extension as appropriate) before being returned. 
Since the compiler can't see inside the asm block to work out what is being
done it cannot remove the uxth instruction on the result.

This does not mean that some of your x-refed reports are invalid though.

[Bug target/45804] ARM: unnecessary sign extension

[felipe.contreras at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45804

--- Comment #2 from Felipe Contreras <felipe.contreras at gmail dot com> 2010-09-27 19:02:19 UTC ---
(In reply to comment #1)
> This particular case is invalid.  
> 
> The ARM ABI requires that short and char result types are extended to 32 bits
> (using signed or unsigned extension as appropriate) before being returned. 
> Since the compiler can't see inside the asm block to work out what is being
> done it cannot remove the uxth instruction on the result.

What do you mean by "returned"? This is an inline function, so it depends
entirely on what the referencing code is doing whether or not this value is
returned, as function return.

Say:

if ((read16_be(data) & 0xfffe) == 0xfff8)

> This does not mean that some of your x-refed reports are invalid though.

Come again?

[Bug target/45804] ARM: unnecessary sign extension

[rearnsha at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45804

--- Comment #3 from Richard Earnshaw <rearnsha at gcc dot gnu.org> 2010-09-27 21:52:38 UTC ---
(In reply to comment #2)
> 
> What do you mean by "returned"? This is an inline function, so it depends
> entirely on what the referencing code is doing whether or not this value is
> returned, as function return.

You didn't cite any context of the use.  Just a single function.