[Bug tree-optimization/37312] New: -Os significantly faster than -O2 on test case

[Bug tree-optimization/37312] New: -Os significantly faster than -O2 on test case

[andi-gcc at firstfloor dot org]

[component might be wrong]

The appended test case is significantly faster with -Os -funroll-all-loops
(~5%) versus -O2 -funroll-all-loops in gcc 4.4 ( gcc version 4.4.0 20080829;
that
is shortly after the IRA merge) on a Core2 (Merom) 

In earlier gcc versions they are about the same performance. The -Os
improvement
is against all earlier versions (good!) but it should be in -O2 too.

I tried -fno-tree-pre as it was suggested and it didn't make a difference.


-- 
           Summary: -Os significantly faster than -O2 on test case
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: andi-gcc at firstfloor dot org
  GCC host triplet: x86_64-linux
GCC target triplet: x86-64-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312

[Bug tree-optimization/37312] -Os significantly faster than -O2 on test case

[andi-gcc at firstfloor dot org]

------- Comment #1 from andi-gcc at firstfloor dot org  2008-09-01 11:22 -------
Created an attachment (id=16178)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16178&action=view)
test case

checksum functions extracted from the Linux kernel.

Not preprocessed, but should compile on any x86 ISO-C system


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312

[Bug tree-optimization/37312] -Os significantly faster than -O2 on test case

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2008-09-01 13:42 -------
Uh, well.  The code ist mostly inline assembly which doesn't give GCC much
freedom to do something.  I guess -O2 simply optimizes "too much" around the
asm.  Try not using inline assembly instead.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312

[Bug tree-optimization/37312] -Os significantly faster than -O2 on test case

[andi-gcc at firstfloor dot org]

------- Comment #3 from andi-gcc at firstfloor dot org  2008-09-01 14:20 -------
Thanks for the us^whelpful comment. If you can suggest a way to do carry
preserving addition without inline assembler that would be fine, otherwise not.

-Os seems to do something that improves it at least (and that is new in 4.4,
4.3 didn't do that)

I suppose -O2 does something more that makes it then worse again. 

I merely filled it because I thought it would be interesting to fix that
something to not pessimize code.




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312

[Bug tree-optimization/37312] -Os significantly faster than -O2 on test case

[rguenth at gcc dot gnu dot org]

------- Comment #4 from rguenth at gcc dot gnu dot org  2008-09-01 14:36 -------
Well, now -Os -funroll-all-loops doesn't do any unrolling anymore while it did
before.  With -O2 you get what you ask for - unrolled loops.

-funroll-all-loops isn't really a flag to be used in general.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312

Re: [Bug tree-optimization/37312] -Os significantly faster than -O2 on test case

[Andrew Thomas Pinski]

This is mostly because of extra register moves that IRA some times  
introduces. There is another bug about Inline-asm and the return  
register.

Sent from my iPhone

On Sep 1, 2008, at 7:36, "rguenth at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org 
 > wrote:

>
>
> ------- Comment #4 from rguenth at gcc dot gnu dot org  2008-09-01  
> 14:36 -------
> Well, now -Os -funroll-all-loops doesn't do any unrolling anymore  
> while it did
> before.  With -O2 you get what you ask for - unrolled loops.
>
> -funroll-all-loops isn't really a flag to be used in general.
>
>
> -- 
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312
>

[Bug tree-optimization/37312] -Os significantly faster than -O2 on test case wiht -funroll-all-loops

[pinskia at gmail dot com]

------- Comment #5 from pinskia at gmail dot com  2008-09-01 20:41 -------
Subject: Re:  -Os significantly faster than -O2 on test case

This is mostly because of extra register moves that IRA some times  
introduces. There is another bug about Inline-asm and the return  
register.

Sent from my iPhone

On Sep 1, 2008, at 7:36, "rguenth at gcc dot gnu dot org"
<gcc-bugzilla@gcc.gnu.org 
 > wrote:

>
>
> ------- Comment #4 from rguenth at gcc dot gnu dot org  2008-09-01  
> 14:36 -------
> Well, now -Os -funroll-all-loops doesn't do any unrolling anymore  
> while it did
> before.  With -O2 you get what you ask for - unrolled loops.
>
> -funroll-all-loops isn't really a flag to be used in general.
>
>
> -- 
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312
>


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312

[Bug tree-optimization/37312] -Os significantly faster than -O2 on test case wiht -funroll-all-loops

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2008-09-02 20:45 -------
The main difference between -O2 and -Os is that csum_partial is inlined for -Os
and unrolling is disabled for -Os.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37312