[Bug c++/33604] New: significantly slower results with 4.3 compared to 4.2

[Bug c++/33604] New: significantly slower results with 4.3 compared to 4.2

[michael dot olbrich at gmx dot net]

In a C++ template library (a matrix library with expression templates)
upgrading from g++-4.2 to g++-4.3 results in 3x slower programs.

Compiler versions:
g++-4.2 (GCC) 4.2.1 (Debian 4.2.1-5)
g++-4.3 (Debian 4.3-20070902-1) 4.3.0 20070902 (experimental) [trunk revision
128028]

$ g++-4.2 -O3 -DNDEBUG benchmark.cpp -o benchmark && time ./benchmark
1.0001 0 0
0 1.0001 0
0 0 1.0001

real    0m4.495s
user    0m4.416s
sys     0m0.003s

$ g++-4.3 -O3 -DNDEBUG benchmark.cpp -o benchmark && time ./benchmark
1.0001 0 0
0 1.0001 0
0 0 1.0001

real    0m15.891s
user    0m15.595s
sys     0m0.018s

I looked at the assembler code but I didn't see anything obvious (I don't
know much about assembler so I may have missed something.
I did notice that adding -fforce-addr changes the result for 4.3 but not
for 4.2.

$ g++-4.3 -O3 -DNDEBUG benchmark.cpp -fforce-addr -o benchmark && time
./benchmark
1.0001 0 0
0 1.0001 0
0 0 1.0001

real    0m8.779s
user    0m8.662s
sys     0m0.007s


-- 
           Summary: significantly slower results with 4.3 compared to 4.2
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: michael dot olbrich at gmx dot net
 GCC build triplet: i486-linux-gnu
  GCC host triplet: i486-linux-gnu
GCC target triplet: i486-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] significantly slower results with 4.3 compared to 4.2

[michael dot olbrich at gmx dot net]

------- Comment #1 from michael dot olbrich at gmx dot net  2007-09-30 21:15 -------
Created an attachment (id=14273)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14273&action=view)
The archive allowing to reproduce the bug.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2007-10-01 09:46 -------
With a more recent 4.3 version I get

     4.2   4.3
-O2  6.1s  5.1s
-O3  4.1s  5.1s

both -O3 variants have one function not inlined in the main loop.  For 4.2
it is

EiEval<EiMatrixProduct<EiMatrix<double, 3, 3>, EiMatrix<double, 3, 3> > >
operator*<double, EiMatrix<double, 3, 3>, EiMatrix<double, 3, 3>
>(EiObject<double, EiMatrix<double, 3, 3> > const&, EiObject<double,
EiMatrix<double, 3, 3> > const&)

for 4.3 it is

EiMatrix<double, 3, 3>& EiObject<double, EiMatrix<double, 3, 3>
>::operator=<EiSum<EiMatrix<double, 3, 3>,
EiScalarProduct<EiSum<EiMatrix<double, 3, 3>, EiMatrix<double, 3, 3> > > >
>(EiObject<double, EiSum<EiMatrix<double, 3, 3>,
EiScalarProduct<EiSum<EiMatrix<double, 3, 3>, EiMatrix<double, 3, 3> > > > >
const&)

but even complete inlining does not improve numbers much.  There's also
nothing obvious in the asm - we simply have too little registers and load/store
from/to memory very often.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[hjl at lucon dot org]

------- Comment #3 from hjl at lucon dot org  2007-10-01 21:03 -------
I have verified that revision 119502:

http://gcc.gnu.org/ml/gcc-cvs/2006-12/msg00119.html

is the cause.


-- 

hjl at lucon dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dberlin at gcc dot gnu dot
                   |                            |org
  GCC build triplet|i486-linux-gnu              |
   GCC host triplet|i486-linux-gnu              |
 GCC target triplet|i486-linux-gnu              |
            Summary|significantly slower results|[4.3 Regression] Revision
                   |with 4.3 compared to 4.2    |119502 causes significantly
                   |                            |slower results with 4.3
                   |                            |compared to 4.2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[hjl at lucon dot org]

------- Comment #4 from hjl at lucon dot org  2007-10-01 21:04 -------
I saw 40% performance regression at -O3 with testcase in comment #1 on
Linux/x86-64.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[dberlin at dberlin dot org]

------- Comment #5 from dberlin at gcc dot gnu dot org  2007-10-01 21:07 -------
Subject: Re:  [4.3 Regression] Revision 119502 causes significantly slower
results with 4.3 compared to 4.2

I'm not fixing this until someone can tell me what exactly is going
wrong.  There have been *so* many changes to PTA since that revision
that the majority of the code it touched doesn't even do the same
thing anymore.
My guess is that this is a case where adding extra vdefs/vuses made
some dumb optimizer able to see something it can't when the chains are
separate like they should be.


On 1 Oct 2007 21:04:40 -0000, hjl at lucon dot org
<gcc-bugzilla@gcc.gnu.org> wrote:
>
>
> ------- Comment #4 from hjl at lucon dot org  2007-10-01 21:04 -------
> I saw 40% performance regression at -O3 with testcase in comment #1 on
> Linux/x86-64.
>
>
> --
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604
>
>


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

Re: [Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[Daniel Berlin]

I'm not fixing this until someone can tell me what exactly is going
wrong.  There have been *so* many changes to PTA since that revision
that the majority of the code it touched doesn't even do the same
thing anymore.
My guess is that this is a case where adding extra vdefs/vuses made
some dumb optimizer able to see something it can't when the chains are
separate like they should be.


On 1 Oct 2007 21:04:40 -0000, hjl at lucon dot org
<gcc-bugzilla@gcc.gnu.org> wrote:
>
>
> ------- Comment #4 from hjl at lucon dot org  2007-10-01 21:04 -------
> I saw 40% performance regression at -O3 with testcase in comment #1 on
> Linux/x86-64.
>
>
> --
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604
>
>

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2007-10-22 08:32 -------
Anyone wants to rerun the results?


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu dot
                   |                            |org
   Target Milestone|---                         |4.3.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[mmitchel at gcc dot gnu dot org]

-- 

mmitchel at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #7 from bonzini at gnu dot org  2007-11-06 17:10 -------
I would also like to know the status, because I would like to get rid of
-fforce-addr. :-)


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bonzini at gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[michael dot olbrich at gmx dot net]

------- Comment #8 from michael dot olbrich at gmx dot net  2007-11-06 18:55 -------
A rerun with Debian's latest versions:
g++-4.2 (GCC) 4.2.3 20071014 (prerelease) (Debian 4.2.2-3)
g++-4.3 (Debian 4.3-20071020-1) 4.3.0 20071020 (experimental) [trunk revision
129510]

  4.2   4.3 |
---------------
 21.2s  19.4s |-O2
  4.4s  15.9s |-O3
  4.4s   8.5s |-O3 -fforce-addr

This is more or less the same as my original results, but very different from
Richard Guenther's results. Very strange.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #9 from bonzini at gnu dot org  2007-11-06 20:08 -------
Can you try -O3 -fno-predictive-commoning and -O3 -fno-tree-vectorize?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[michael dot olbrich at gmx dot net]

------- Comment #10 from michael dot olbrich at gmx dot net  2007-11-06 21:01 -------
  4.2   4.3   |
---------------
  4.4s  15.9s |-O3
  4.6s  16.3s |-O3 -fno-tree-vectorize
   X    18.9s |-O3 -fno-predictive-commoning


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug c++/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #11 from bonzini at gnu dot org  2007-11-06 21:23 -------
michael dot olbrich at gmx dot net wrote:
> ------- Comment #10 from michael dot olbrich at gmx dot net  2007-11-06 21:01 -------
>   4.2   4.3   |
> ---------------
>   4.4s  15.9s |-O3
>   4.6s  16.3s |-O3 -fno-tree-vectorize
>    X    18.9s |-O3 -fno-predictive-commoning

This is *good* news. :-)  New optimizations in GCC 4.3 save 15% of the runtime.
 It is then only an aliasing problem, maybe due to extra vdefs/vuses.  I
haven't look at the code but, anyway, Danny/Diego---would a patch like this
make sense?

   /* Since we clearly cannot lower the number of virtual operators
      below the total number of memory statements in the function, we
      may need to adjust MAX_ALIASED_VOPS beforehand.  */
-  if (MAX_ALIASED_VOPS < mem_ref_stats->num_mem_stmts)
-    MAX_ALIASED_VOPS = mem_ref_stats->num_mem_stmts;
+  if (MAX_ALIASED_VOPS < mem_ref_stats->num_mem_stmts * AVG_ALIASED_VOPS)
+    MAX_ALIASED_VOPS = mem_ref_stats->num_mem_stmts * AVG_ALIASED_VOPS;


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dnovillo at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #12 from rguenth at gcc dot gnu dot org  2007-11-06 21:51 -------
I don't think we want to start playing with the heuristics ;)  That patch
certainly will cause compile-time and memory usage regressions.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[paolo dot bonzini at lu dot unisi dot ch]

------- Comment #13 from paolo dot bonzini at lu dot unisi dot ch  2007-11-07 06:03 -------
Subject: Re:  [4.3 Regression] Revision 119502
 causes significantly slower results with 4.3 compared to 4.2


> I don't think we want to start playing with the heuristics ;)  That patch
> certainly will cause compile-time and memory usage regressions.

Not for -O2, since the default AVG_ALIASED_VOPS is 1.  For -O3 it is 3, 
not huge either.

Paolo


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[dnovillo at google dot com]

------- Comment #14 from dnovillo at google dot com  2007-11-07 12:14 -------
Subject: Re:  [4.3 Regression] Revision 119502 causes significantly slower
results with 4.3 compared to 4.2

On 7 Nov 2007 06:03:09 -0000, paolo dot bonzini at lu dot unisi dot ch
<gcc-bugzilla@gcc.gnu.org> wrote:

> > I don't think we want to start playing with the heuristics ;)  That patch
> > certainly will cause compile-time and memory usage regressions.
>
> Not for -O2, since the default AVG_ALIASED_VOPS is 1.  For -O3 it is 3,
> not huge either.

You'd be surprised.  AVG_ALIASED_VOPS is fairly sensitive wrt
compile-time effects.  You need to test the effects of this change
carefully.


Diego.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #15 from bonzini at gnu dot org  2007-11-07 13:09 -------
Created an attachment (id=14496)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14496&action=view)
self contained testcase

                       4.2          4.3
-O2                    14.5s        11.4s
-O3                     4.0s         8.7s
-O2 -finline-functions  4.0s         9.6s


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #14273|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-11-07 13:09:53
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #16 from bonzini at gnu dot org  2007-11-07 14:22 -------
4.3 does much less SRA.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #17 from bonzini at gnu dot org  2007-11-07 16:03 -------
Created an attachment (id=14497)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14497&action=view)
"minimal" testcase

This is as small as I could get it (130 lines).  Can be made smaller, but the
performance difference is much less sensitive.  I get 0.8s with 4.2, 1.4s with
4.3.

The problem is that the MatRef classes (called EiMatrixConstRef in the
original, I renamed stuff and detemplatized it to simplify dumps) are not
scalarized.


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #14496|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #18 from bonzini at gnu dot org  2007-11-07 16:26 -------
Created an attachment (id=14498)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14498&action=view)
never say minimal!

So this is a testcase constructed from scratch.  4.3 is 4.5x slower (0.8s vs.
0.18s).  Furthermore, if you "#define const", the performance pitfall goes
away.

For whoever wants to work on this, I would double check with the previous
minimal testcase, but this has the same symptoms (no SRA of the Ref instances)
and is much much smaller, so it is more suited to go in GCC's testsuite.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #19 from bonzini at gnu dot org  2007-11-07 16:35 -------
Might be tree-level forwprop, CCing richi...


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #20 from rguenth at gcc dot gnu dot org  2007-11-07 17:28 -------
Basically things like

  D.2196_13 = (struct Ref *) &D.2150.lhs;
  D.2196_13->m ={v} &I;

where D.2150.lhs is const and thus there is a (not useless) cast from
(const struct Ref *) to (struct Ref *).  So what the code does is to
store to "const" qualified memory:

  D.2150.lhs.m = &I;

of course only as part of constructing it, so it is legal, but the middle-end
cannot tell apart legal from illegal cases.

So we can just declare all of them legal based on the fact that otherwise
this would certainly invoke undefined behavior.  In which case tree-level
forwprop can be teached to disregard conversions that change constness
only on address-forwarding.

Something like a more careful

Index: tree-ssa-forwprop.c
===================================================================
--- tree-ssa-forwprop.c (revision 129897)
+++ tree-ssa-forwprop.c (working copy)
@@ -952,8 +952,14 @@ tree_ssa_forward_propagate_single_use_va
                  continue;
                }

-             if (TREE_CODE (rhs) == ADDR_EXPR)
+             if (TREE_CODE (rhs) == ADDR_EXPR
+                 || ((TREE_CODE (rhs) == NOP_EXPR
+                      || TREE_CODE (rhs) == CONVERT_EXPR)
+                     && TREE_CODE (TREE_OPERAND (rhs, 0)) == ADDR_EXPR
+                     && (TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (rhs)))
+                         == TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE
(TREE_OPERAND (rhs, 0)))))))
                {
+                 STRIP_NOPS (rhs);
                  if (forward_propagate_addr_expr (lhs, rhs))
                    {
                      release_defs (stmt);


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |rguenth at gcc dot gnu dot
                   |dot org                     |org
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2007-11-07 13:09:53         |2007-11-07 17:28:11
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[jacob at math dot jussieu dot fr]

------- Comment #21 from jacob at math dot jussieu dot fr  2007-11-07 20:58 -------
Hi, I'm the guy behind Eigen, from which the initial testcase is taken.

Would it help the compiler if I removed most of the "const" keywords in my
code?

I am asking because of the following. As Eigen has Lvalue expression
templates(*), it is very cumbersome to enforce constness as many methods have
to be written twice, once const-qualified and once not. Therefore I have given
up enforcing constness, and am doing const_cast's at several places. Therefore
many of the "const" keywords in Eigen are rather meaningless and could be
removed.

(*) for instance the following is allowed:
matrix.row(i) += a * matrix.row(j).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[paolo dot bonzini at lu dot unisi dot ch]

------- Comment #22 from paolo dot bonzini at lu dot unisi dot ch  2007-11-08 06:12 -------
Subject: Re:  [4.3 Regression] Revision 119502
 causes significantly slower results with 4.3 compared to 4.2

jacob at math dot jussieu dot fr wrote:
> ------- Comment #21 from jacob at math dot jussieu dot fr  2007-11-07 20:58 -------
> Hi, I'm the guy behind Eigen, from which the initial testcase is taken.
> 
> Would it help the compiler if I removed most of the "const" keywords in my
> code?

I'm not the expert on expression templates, but I don't think so. 
"const"'s are mostly ignored by the compiler for optimization purposes, 
but removing those would be bad for the users, I think.

Paolo


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #23 from rguenth at gcc dot gnu dot org  2007-11-09 12:32 -------
Subject: Bug 33604

Author: rguenth
Date: Fri Nov  9 12:32:20 2007
New Revision: 130040

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130040
Log:
2007-11-09  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/33604
        * tree-ssa-forwprop.c (tree_ssa_forward_propagate_single_use_vars):
        Disregard changes in CV qualifiers of pointed to types for
        forward propagating ADDR_EXPRs.
        * tree-ssa-ccp.c (fold_stmt_r): Preserve volatileness of the original
        expression.

        * g++.dg/tree-ssa/pr33604.C: New testcase.
        * gcc.dg/pr32721.c: Adjust pattern.

Added:
    trunk/gcc/testsuite/g++.dg/tree-ssa/pr33604.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/gcc.dg/pr32721.c
    trunk/gcc/tree-ssa-ccp.c
    trunk/gcc/tree-ssa-forwprop.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #24 from rguenth at gcc dot gnu dot org  2007-11-09 12:33 -------
Fixed.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[jacob at math dot jussieu dot fr]

------- Comment #25 from jacob at math dot jussieu dot fr  2007-11-09 12:49 -------
Huge thanks to Richard Guenther for the fix!

I don't have a 4.3 installed here but I trust you if you say it's fixed :)

Also thanks to everybody who helped, especially to Paolo Bonzini for the
minimal testcase!


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenther at suse dot de]

------- Comment #26 from rguenther at suse dot de  2007-11-09 12:51 -------
Subject: Re:  [4.3 Regression] Revision 119502
 causes significantly slower results with 4.3 compared to 4.2

On Fri, 9 Nov 2007, jacob at math dot jussieu dot fr wrote:

> Huge thanks to Richard Guenther for the fix!
> 
> I don't have a 4.3 installed here but I trust you if you say it's fixed :)
> 
> Also thanks to everybody who helped, especially to Paolo Bonzini for the
> minimal testcase!

Note that I only looked at the generated assembler ;)  You still might
want to check performance at some point.

Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[jacob at math dot jussieu dot fr]

------- Comment #27 from jacob at math dot jussieu dot fr  2007-11-09 12:55 -------
OK, I'm using ubuntu and there's no 4.3 package, and last time I tried
compiling a snapshot it failed. I hope someone else will be able to run the
benchmark and post the result.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[michael dot olbrich at gmx dot net]

------- Comment #28 from michael dot olbrich at gmx dot net  2007-12-04 10:56 -------
With the latest debian packages:
g++-4.2 (GCC) 4.2.3 20071123 (prerelease) (Debian 4.2.2-4)
g++-4.3 (Debian 4.3-20071130-1) 4.3.0 20071130 (experimental) [trunk revision
130545]

Same speed for benchmark.cpp but _not_ for the original gcc-bug-report.tar.gz:
g++-4.3 is a bit faster than when I first tested (0m14.248s vs. 0m15.891s) but
nowhere near g++-4.2 (0m4.495s).


-- 

michael dot olbrich at gmx dot net changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
         Resolution|FIXED                       |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[bonzini at gnu dot org]

------- Comment #29 from bonzini at gnu dot org  2007-12-04 11:17 -------
Could you try 4.2 vs. 4.3 on:

- http://gcc.gnu.org/bugzilla/attachment.cgi?id=14496
- http://gcc.gnu.org/bugzilla/attachment.cgi?id=14497
- http://gcc.gnu.org/bugzilla/attachment.cgi?id=14497 with the first #if 0
enabled

Thanks!


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #30 from rguenth at gcc dot gnu dot org  2007-12-04 11:21 -------
It is likely that for the bigger testcases we run into our usual
alias-partitioning problems, so also try with --param max-aliased-vops=100000.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[michael dot olbrich at gmx dot net]

------- Comment #31 from michael dot olbrich at gmx dot net  2007-12-04 12:17 -------
Subject: Re:  [4.3 Regression] Revision 119502
        causes significantly slower results with 4.3 compared to 4.2

On Tue, Dec 04, 2007 at 11:17:46AM -0000, bonzini at gnu dot org wrote:
> - http://gcc.gnu.org/bugzilla/attachment.cgi?id=14496

same as gcc-bug-report.tar.gz

> - http://gcc.gnu.org/bugzilla/attachment.cgi?id=14497
> - http://gcc.gnu.org/bugzilla/attachment.cgi?id=14497 with the first #if 0 enabled

The difference between 4.2 and 4.3 is not as big but is still there:
0.7s vs. 1.6s

No significant change with "--param max-aliased-vops=100000".


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[paolo dot bonzini at lu dot unisi dot ch]

------- Comment #32 from paolo dot bonzini at lu dot unisi dot ch  2007-12-04 12:38 -------
Subject: Re:  [4.3 Regression] Revision 119502
 causes significantly slower results with 4.3 compared to 4.2


> The difference between 4.2 and 4.3 is not as big but is still there:
> 0.7s vs. 1.6s

Well, that's more than 2x.  So, we could say that the penultimate 
testcase is good enough.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3 Regression] Revision 119502 causes significantly slower results with 4.3 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #33 from rguenth at gcc dot gnu dot org  2008-01-02 11:00 -------
Unassigning.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|rguenth at gcc dot gnu dot  |unassigned at gcc dot gnu
                   |org                         |dot org
             Status|REOPENED                    |NEW


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug tree-optimization/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[jacob at math dot jussieu dot fr]

------- Comment #34 from jacob at math dot jussieu dot fr  2008-02-27 16:36 -------
I retried today with

g++-4.3 (Ubuntu 4.3-20080126-1ubuntu1) 4.3.0 20080126 (experimental) [trunk
revision 131874]

and my benchmark ran in 25s compared to 20s with g++-4.2.3.

So there is certainly a lot of improvement though it's still 20% slower than
4.2.3. (This might have further improved since my 4.3 snapshot is 1 month old).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[pinskia at gcc dot gnu dot org]

------- Comment #35 from pinskia at gcc dot gnu dot org  2008-02-27 16:43 -------
We get:
<bb 3>:
  m__valuem_I_lsm.28 = 1.0e+0 - m__valuem_I_lsm.28;
  ivtmp.30 = ivtmp.30 + 1;
  if (ivtmp.30 != 100000000)
    goto <bb 3>;
  else
    goto <bb 4>;

or:
L2:
        addl    $1, %eax
        movapd  %xmm1, %xmm2
        subsd   %xmm0, %xmm2
        cmpl    $100000000, %eax
        movapd  %xmm2, %xmm0
        jne     L2

or :
L2:
        addl    $1, %eax
        cmpl    $100000000, %eax
        fsub    %st, %st(1)
        jne     L2

All are fast.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |target


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[jacob at math dot jussieu dot fr]

------- Comment #36 from jacob at math dot jussieu dot fr  2008-02-27 16:58 -------
That's great; from the assembly code I take it that you are referring tothe
last benchmark.cpp; I was referring to the first one. Again, my 4.3 is one
month old so maybe things have further improved since.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[bonzini at gnu dot org]

------- Comment #37 from bonzini at gnu dot org  2008-02-27 17:05 -------
Subject: Re:  [4.3/4.4 Regression] Revision 119502 causes
 significantly slower results with 4.3/4.4 compared to 4.2

jacob at math dot jussieu dot fr wrote:
> ------- Comment #36 from jacob at math dot jussieu dot fr  2008-02-27 16:58 -------
> That's great; from the assembly code I take it that you are referring tothe
> last benchmark.cpp; I was referring to the first one. Again, my 4.3 is one
> month old so maybe things have further improved since.

No, I doubt.  The last benchmark.cpp is now fully optimized, but we 
might be missing something.

Paolo


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[michael dot olbrich at gmx dot net]

------- Comment #38 from michael dot olbrich at gmx dot net  2008-03-02 12:14 -------
I tried again with 
g++-4.2 (GCC) 4.2.3 (Debian 4.2.3-2)
g++-4.3 (Debian 4.3-20080227-1) 4.3.0 20080227 (prerelease) [gcc-4_3-branch
revision 132730]

The problem still exists for the first two test cases.
As I noted in comment #8 there is a significant speedup from -O2 to -O3 for
g++-4.2 (18s -> 5s)
With the current g++-4.3 there is no difference between -O2 and -O3 (both 14s)
"-fforce-addr" which produced significant speedup does not exist anymore.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[bonzini at gnu dot org]

------- Comment #39 from bonzini at gnu dot org  2008-03-02 12:26 -------
Subject: Re:  [4.3/4.4 Regression] Revision 119502 causes
 significantly slower results with 4.3/4.4 compared to 4.2


> The problem still exists for the first two test cases.
> As I noted in comment #8 there is a significant speedup from -O2 to -O3 for
> g++-4.2 (18s -> 5s)
> With the current g++-4.3 there is no difference between -O2 and -O3 (both 14s)
> "-fforce-addr" which produced significant speedup does not exist anymore.

So maybe we need to restore part of -fforce-addr's behavior, but not the 
one that caused regressions.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #40 from rguenth at gcc dot gnu dot org  2008-03-02 14:00 -------
I think new analysis is necessary first -- what is exactly causing the speed
difference?


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
 GCC target triplet|                            |i?86-*-* x86_64-*-*
           Keywords|                            |missed-optimization


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[rguenth at gcc dot gnu dot org]

-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|                            |4.3.0
   Target Milestone|4.3.0                       |4.3.1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #41 from rguenth at gcc dot gnu dot org  2008-06-06 14:58 -------
4.3.1 is being released, adjusting target milestone.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.3.1                       |4.3.2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[jsm28 at gcc dot gnu dot org]

------- Comment #42 from jsm28 at gcc dot gnu dot org  2008-08-27 22:02 -------
4.3.2 is released, changing milestones to 4.3.3.


-- 

jsm28 at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.3.2                       |4.3.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[amonakov at gcc dot gnu dot org]

------- Comment #43 from amonakov at gcc dot gnu dot org  2008-08-29 13:12 -------
Checking original testcase times on x86_64 prescott with gentoo 4.2, 4.3 and
today's trunk:
2.960s    g++-4.2.4 (GCC) 4.2.4 (Gentoo 4.2.4 p1.0)
2.916s    g++-4.3.1 (Gentoo 4.3.1-r1 p1.1) 4.3.1
3.993s    g++ (GCC) 4.4.0 20080829 (experimental)
2.796s    g++ (GCC) 4.4.0 20080829 (experimental) with --param
max-inline-insns-auto=126

So I believe lack of inlining is the biggest 4.4's problem. We do not inline
3x3 matrix multiplication in benchmark loop.

While looking at it I found that einline2 dump does not always show the reason
for not inlining. I would like to propose the following patch:

--- a/gcc/ipa-inline.c
+++ b/gcc/ipa-inline.c
@@ -1494,6 +1494,8 @@ cgraph_decide_inlining_incrementally (struct cgraph_node
*node,
          }
        if (cgraph_default_inline_p (e->callee, &failed_reason))
          inlined |= try_inline (e, mode, depth);
+       else if (dump_file)
+         fprintf (dump_file, "Not inlining: %s.\n", failed_reason);
       }
   node->aux = (void *)(size_t) old_mode;
   return inlined;


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[pinskia at gcc dot gnu dot org]

------- Comment #44 from pinskia at gcc dot gnu dot org  2008-12-25 18:13 -------
Does anyone have new numbers?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[Joey dot ye at intel dot com]

------- Comment #45 from Joey dot ye at intel dot com  2008-12-30 01:49 -------
(In reply to comment #44)
> Does anyone have new numbers?
Fixed on both i386/x86_64:
x86_64:
4.4 (trunk 142847): 5.4s
4.3.2 release:      5.4s
4.2.4 release:      5.4s

i386:
4.4 (trunk 142847): 2.7s
4.3.2 release:      2.8s
4.2.4 release:      2.7s


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[bonzini at gnu dot org]

------- Comment #46 from bonzini at gnu dot org  2008-12-30 08:02 -------
What benchmark.cpp was that? And did you test -O2 or -O3?

Thanks!


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[rguenth at gcc dot gnu dot org]

------- Comment #47 from rguenth at gcc dot gnu dot org  2009-01-24 10:19 -------
GCC 4.3.3 is being released, adjusting target milestone.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.3.3                       |4.3.4


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[bonzini at gnu dot org]

------- Comment #48 from bonzini at gnu dot org  2009-02-01 08:14 -------
Fixed on the trunk with the original testcase:

4.2 -O2       0m13.897s
4.2 -O3       miscompiled
4.4 -O2/-O3   0m8.714s


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604

[Bug target/33604] [4.3/4.4 Regression] Revision 119502 causes significantly slower results with 4.3/4.4 compared to 4.2

[rguenth at gcc dot gnu dot org]

-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.3.4                       |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33604