Bug 43861

Summary: -Os creates larger binaries than before in some cases (-falign-... options enabled)
Product: gcc Reporter: rwahl
Component: middle-endAssignee: Not yet assigned to anyone <unassigned>
Status: UNCONFIRMED ---    
Severity: normal CC: gcc-bugs, hp
Priority: P3    
Version: 4.4.3   
Target Milestone: ---   
Host: Target:
Build: Known to work:
Known to fail: Last reconfirmed:
Attachments: Disable alignment when using -Os

Description rwahl 2010-04-22 23:48:23 UTC
Hi,

I just noticed that some of my libraries are greater when compiled with gcc 4.4.3 instead of gcc 4.3.2. Diffing the output of "gcc --help=optimizers -Q -Os" shows this in 4.3.2:

  -falign-jumps                        [disabled]
  -falign-labels                       [disabled]
  -falign-loops                        [enabled]

and this for 4.4.3:

  -falign-functions                     [enabled]
  -falign-jumps                         [enabled]
  -falign-labels                        [enabled]
  -falign-loops                         [disabled]

when explicitely disabling the alignment options then the code size shrinks. For libxml2 in my case from 704256 to 675584 bytes.

The manual says:

  -Os disables the following optimization flags: -falign-functions
           -falign-jumps  -falign-loops -falign-labels  -freorder-blocks
           -freorder-blocks-and-partition -fprefetch-loop-arrays
           -ftree-vect-loop-version

which is not true (even not for 4.3.2 since loop alignment was enabled). Some other opts like -freorder-blocks is enabled as well.

So is this actually a bug or is the documentation outdated or even both?
Comment 1 pinskia@gmail.com 2010-04-22 23:51:39 UTC
Subject: Re:   New: -Os creates larger binaries than before in some cases (-falign-... options enabled)

Which target is this for?

Sent from my iPhone

On Apr 22, 2010, at 4:48 PM, "rwahl at gmx dot de" <gcc-bugzilla@gcc.gnu.org 
 > wrote:

> Hi,
>
> I just noticed that some of my libraries are greater when compiled  
> with gcc
> 4.4.3 instead of gcc 4.3.2. Diffing the output of "gcc -- 
> help=optimizers -Q
> -Os" shows this in 4.3.2:
>
>  -falign-jumps                        [disabled]
>  -falign-labels                       [disabled]
>  -falign-loops                        [enabled]
>
> and this for 4.4.3:
>
>  -falign-functions                     [enabled]
>  -falign-jumps                         [enabled]
>  -falign-labels                        [enabled]
>  -falign-loops                         [disabled]
>
> when explicitely disabling the alignment options then the code size  
> shrinks.
> For libxml2 in my case from 704256 to 675584 bytes.
>
> The manual says:
>
>  -Os disables the following optimization flags: -falign-functions
>           -falign-jumps  -falign-loops -falign-labels  -freorder- 
> blocks
>           -freorder-blocks-and-partition -fprefetch-loop-arrays
>           -ftree-vect-loop-version
>
> which is not true (even not for 4.3.2 since loop alignment was  
> enabled). Some
> other opts like -freorder-blocks is enabled as well.
>
> So is this actually a bug or is the documentation outdated or even  
> both?
>
>
> -- 
>           Summary: -Os creates larger binaries than before in some  
> cases (-
>                    falign-... options enabled)
>           Product: gcc
>           Version: 4.4.3
>            Status: UNCONFIRMED
>          Severity: normal
>          Priority: P3
>         Component: middle-end
>        AssignedTo: unassigned at gcc dot gnu dot org
>        ReportedBy: rwahl at gmx dot de
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43861
>
Comment 2 rwahl 2010-04-23 00:10:32 UTC
All my own cross compilers for ARM and x86 hosted on x86 and the native gcc in Fedora 12 (x86) show this behavior.
Comment 3 rwahl 2010-05-02 16:55:19 UTC
Created attachment 20535 [details]
Disable alignment when using -Os

The attached patch fixes the problem. It could be written in a more elegant way by changing just opt1_max but this would modify the semantic of this variable and it should be renamed accordingly.

On ARM the fix is probably a NOP because the default alignment seems to be 1 anyway but I haven't checked this.

The output of "gcc -Q --help=optimizers" seems to be broken for the alignment options because on one hand these options are not boolean and on the other hand falign-loops is reported as enabled when it has been disabled and vice versa.