[Bug middle-end/38518] Excessive compile time with -O3

[Bug middle-end/38518] Excessive compile time with -O3

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-02-02
     Ever Confirmed|0                           |1
      Known to fail|                            |4.7.0

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-02-02 19:06:47 UTC ---
-ftime-report will report where the extra time is happening.

Anyways confirmed.
 loop unswitching        : 308.19 (70%) usr   0.05 ( 4%) sys 308.17 (70%) wall 
   105 kB ( 0%) ggc

[Bug middle-end/38518] Excessive compile time with -O3

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
With GCC 4.8 we see

 loop invariant motion   :  16.79 (32%) usr   0.02 ( 3%) sys  16.65 (31%) wall 
   148 kB ( 0%) ggc
 loop unswitching        :  10.43 (20%) usr   0.02 ( 3%) sys  10.42 (20%) wall 
     0 kB ( 0%) ggc
 TOTAL                 :  52.07             0.67            53.00            
363360 kB

which is RTL loop optimizers.

[Bug rtl-optimization/38518] Excessive compile time with -O3

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixing this DF issue get's us down to

 loop invariant motion   :   0.14 ( 0%) usr   0.01 ( 1%) sys   0.14 ( 0%) wall 
   148 kB ( 0%) ggc
 loop unswitching        :  15.44 (30%) usr   0.02 ( 3%) sys  15.48 (29%) wall 
     0 kB ( 0%) ggc
 TOTAL                 :  52.26             0.73            52.99            
456534 kB

so finally a RTL unswitching slowness ;)

[Bug rtl-optimization/38518] Excessive compile time with -O3

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #7)
> Fixing this DF issue get's us down to
> 
>  loop invariant motion   :   0.14 ( 0%) usr   0.01 ( 1%) sys   0.14 ( 0%)
> wall     148 kB ( 0%) ggc
>  loop unswitching        :  15.44 (30%) usr   0.02 ( 3%) sys  15.48 (29%)
> wall       0 kB ( 0%) ggc
>  TOTAL                 :  52.26             0.73            52.99           
> 456534 kB
> 
> so finally a RTL unswitching slowness ;)

Fixing the same issue in loop-iv.c fixes that, too.

 loop invariant motion   :   0.18 ( 0%) usr   0.00 ( 0%) sys   0.16 ( 0%) wall 
   148 kB ( 0%) ggc
 loop unswitching        :   0.01 ( 0%) usr   0.00 ( 0%) sys   0.01 ( 0%) wall 
     0 kB ( 0%) ggc
 TOTAL                 :  37.14             0.67            37.80            
456534 kB

[Bug rtl-optimization/38518] Excessive compile time with -O3

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Mine.

[Bug rtl-optimization/38518] Excessive compile time with -O3

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
Author: rguenth
Date: Fri Jan 17 10:47:59 2014
New Revision: 206702

URL: http://gcc.gnu.org/viewcvs?rev=206702&root=gcc&view=rev
Log:
2014-01-17  Richard Biener  <rguenther@suse.de>

    PR rtl-optimization/38518
    * df.h (df_analyze_loop): Declare.
    * df-core.c: Include cfgloop.h.
    (df_analyze_1): Split out main part of df_analyze.
    (df_analyze): Adjust.
    (loop_inverted_post_order_compute): New function.
    (loop_post_order_compute): Likewise.
    (df_analyze_loop): New function avoiding whole-function
    postorder computes.
    * loop-invariant.c (find_defs): Use df_analyze_loop.
    (find_invariants): Adjust.
    * loop-iv.c (iv_analysis_loop_init): Use df_analyze_loop.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/df-core.c
    trunk/gcc/df.h
    trunk/gcc/loop-invariant.c
    trunk/gcc/loop-iv.c

[Bug rtl-optimization/38518] Excessive compile time with -O3

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ra
             Status|ASSIGNED                    |NEW
                 CC|                            |vmakarov at gcc dot gnu.org
           Assignee|rguenth at gcc dot gnu.org         |unassigned at gcc dot gnu.org

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
With 4.9.0

-O1: 5s
-O2: 7s
-O3: 41s

the profile at -O3 is mostly flat, worst offender is LRA hard-reg assignment
which may also contribute to the high DF time.

Execution times (seconds)
 phase opt and generate  :  41.00 (100%) usr   0.68 (94%) sys  41.75 (100%)
wall  451991 kB (99%) ggc
 df reaching defs        :   5.87 (14%) usr   0.03 ( 4%) sys   5.82 (14%) wall 
     0 kB ( 0%) ggc
 dominance frontiers     :   2.52 ( 6%) usr   0.00 ( 0%) sys   2.51 ( 6%) wall 
     0 kB ( 0%) ggc
 LRA hard reg assignment :   4.94 (12%) usr   0.04 ( 6%) sys   4.96 (12%) wall 
     0 kB ( 0%) ggc
 TOTAL                 :  41.05             0.72            41.84            
456534 kB

It's now a RA issue, I'm out of here (though I'd even consider closing this
as fixed - I've added the testcase to daily monitoring).

[Bug middle-end/38518] Excessive compile time with -O3

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=38518

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
               Host|suse-linux-x86_64           |
          Component|rtl-optimization            |middle-end
             Target|                            |x86_64-*-*

--- Comment #16 from Andrew Pinski <pinskia at gcc dot gnu.org> ---

> callgraph functions expansion      :  38.69 ( 98%)   0.88 ( 83%)  39.60 ( 97%)   431M ( 93%)

I know the above is a catch all though.


At least one issue is this for the trunk:
 machine dep reorg                  :   7.23 ( 18%)   0.00 (  0%)   7.23 ( 18%)
    0  (  0%)

This is with checking enabled but with -fno-checking.

So the x86 maintainer should look into why for that.

Note the trunk with -O2 -fno-checking (compiled with --enable-checking=yes) is
faster than 7.3 compiled with --enable-checking=release :).


Nothing else stands out for the trunk even though -O3 is 39s while -O2 is 8s.