[Bug fortran/37536] New: a mfcr is produced instead of branches (when it is used in branches later)

[Bug fortran/37536] New: a mfcr is produced instead of branches (when it is used in branches later)

[pinskia at gcc dot gnu dot org]

When looking into PR 37534, I noticed that in some cases a compare and then a
mfcr and then again a compare happens.  This is the most reduced testcase I
could come up with:
Function f(MaxDev, a, b, Step)
        Integer :: Step, NewX, NewY
        Integer :: MaxDev, a, b
        Do NewX = a-MaxDev, b+MaxDev, Step
enddo
MaxDev = NewX
end
--- CUT --
Currently we produce:
.L2:
        cmpw 7,10,9
        neg 8,0
        subf 10,10,9
        mfcr 11
        rlwinm 11,11,30,1
        divwu 10,10,8
        b .L3
--- CUT ---
This is caused by the following tree level code:
    if (D.692 > 0 ? D.691 < D.690 : D.691 > D.690) goto L.2;

I think the front-end might be better at optimizing this.


-- 
           Summary: a mfcr is produced instead of branches (when it is used
                    in branches later)
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: pinskia at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] a mfcr is produced instead of branches (when it is used in branches later)

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2008-09-16 01:32 -------
>    if (D.692 > 0 ? D.691 < D.690 : D.691 > D.690) goto L.2;
> I think the front-end might be better at optimizing this.

That is produce:
if (D.692 > 0)
  {
    if (D.691 < D.690) goto L.2;
  }
  else if (D.691 > D.690) goto L.2;

Or maybe the gimplifier should do that.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] a mfcr is produced instead of branches (when it is used in branches later)

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2008-09-16 01:39 -------
This was caused by PR 31399.
(In reply to comment #1)
> Or maybe the gimplifier should do that.

As a stop gap, I think the front-end should do it for 4.4, I might want to look
into that soon.  I will file a C bug version.  The Fortran version is caused by
the front-end and the user did not do it.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
OtherBugsDependingO|                            |31399
              nThis|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|a mfcr is produced instead  |[4.3/4.4 Regression] a mfcr
                   |of branches (when it is used|is produced instead of
                   |in branches later)          |branches for DO loops
   Target Milestone|---                         |4.3.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[jakub at gcc dot gnu dot org]

------- Comment #3 from jakub at gcc dot gnu dot org  2008-09-16 13:02 -------
Created an attachment (id=16338)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16338&action=view)
gcc44-pr37536.patch

So far untested patch to optimize slightly integer type non-simple do loops.
At least on x86_64 it results in smaller and hopefully faster code.
Still testing it with make check-gfortran, but will probably need a SPEC
testing too.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[rguenth at gcc dot gnu dot org]

------- Comment #4 from rguenth at gcc dot gnu dot org  2008-09-16 13:10 -------
ifcombine would be a place to do this CFG re-writing.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu dot
                   |                            |org
           Severity|enhancement                 |normal
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
      Known to work|                            |4.2.4
   Last reconfirmed|0000-00-00 00:00:00         |2008-09-16 13:10:01
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[pault at gcc dot gnu dot org]

------- Comment #5 from pault at gcc dot gnu dot org  2008-09-16 15:29 -------
(In reply to comment #2)
Hi Andrew,

Does this do it for you or is this all to rarified for this
Bear-of-little-brain to follow?

    D.1359 = *start - *offset;
    D.1360 = *end + *offset;
    D.1361 = *step;
    D.1364 = D.1361 > 0 ? D.1360 < D.1359 : D.1360 > D.1359; /* Fix the
result*/
    if (D.1364) goto L.2;                                    /* Then branch  */
    countm1.0 = (D.1361 > 0 ? (character(kind=4)) (D.1360 - D.1359) :
(character(kind=4)) (D.1359 - D.1360)) / (character(kind=4)) ABS_EXPR <D.1361>;
    i = D.1359;
    while (1)

Cheers

Paul


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2008-09-16 15:55 -------
(In reply to comment #5)
> (In reply to comment #2)
> Hi Andrew,
> 
> Does this do it for you or is this all to rarified for this
> Bear-of-little-brain to follow?
> 
>     D.1359 = *start - *offset;
>     D.1360 = *end + *offset;
>     D.1361 = *step;
>     D.1364 = D.1361 > 0 ? D.1360 < D.1359 : D.1360 > D.1359; /* Fix the
> result*/
>     if (D.1364) goto L.2;                                    /* Then branch  */

The issue is that D.1364 is used to the store of the result of a comparison. 
Since GCC does not currently optimize that case, the front-end could rewrite
the if statement to begin with; this is what Jabuk's patch does (Thanks
Jabuk!).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[pault at gcc dot gnu dot org]

------- Comment #7 from pault at gcc dot gnu dot org  2008-09-16 17:51 -------
(In reply to comment #2)
Hi Andrew,

Does this do it for you or is this all to rarified for this
Bear-of-little-brain to follow?

    D.1359 = *start - *offset;
    D.1360 = *end + *offset;
    D.1361 = *step;
    D.1364 = D.1361 > 0 ? D.1360 < D.1359 : D.1360 > D.1359; /* Fix the
result*/
    if (D.1364) goto L.2;                                    /* Then branch  */
    countm1.0 = (D.1361 > 0 ? (character(kind=4)) (D.1360 - D.1359) :
(character(kind=4)) (D.1359 - D.1360)) / (character(kind=4)) ABS_EXPR <D.1361>;
    i = D.1359;
    while (1)

Cheers

Paul

(In reply to comment #6)
> (In reply to comment #5)
> > (In reply to comment #2)
> > Hi Andrew,
> > 
> > Does this do it for you or is this all to rarified for this
> > Bear-of-little-brain to follow?
> > 
> >     D.1359 = *start - *offset;
> >     D.1360 = *end + *offset;
> >     D.1361 = *step;
> >     D.1364 = D.1361 > 0 ? D.1360 < D.1359 : D.1360 > D.1359; /* Fix the
> > result*/
> >     if (D.1364) goto L.2;                                    /* Then branch  */
> 
> The issue is that D.1364 is used to the store of the result of a comparison. 
> Since GCC does not currently optimize that case, the front-end could rewrite
> the if statement to begin with; this is what Jabuk's patch does (Thanks
> Jabuk!).
> 

Ahhhhh....!

Thanks

P.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[jakub at gcc dot gnu dot org]

------- Comment #8 from jakub at gcc dot gnu dot org  2008-09-16 22:17 -------
An ifcombine optimization is certainly desirable, still IMHO it makes sense to
do something about it in the FE too.  The vanilla f951 generated code contains
essentially 3 different conditional expressions on step > 0, and one of those
is assigned to a var to be used by another COND_EXPR.  One is the empty
COND_EXPR,
another is step > 0 ? to - from : from - to and another is ABS_EXPR <step>.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[luisgpm at linux dot vnet dot ibm dot com]

------- Comment #9 from luisgpm at linux dot vnet dot ibm dot com  2008-09-17 15:22 -------
Gathered some PPC 32/64 performance numbers with the patch (based on 140409).
No noticeable performance regressions were found. 32-bit swin and 64-bit art
had a little boost on speed (7.8% and 3.4% respectivelly).


-- 

luisgpm at linux dot vnet dot ibm dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |luisgpm at linux dot vnet
                   |                            |dot ibm dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[jakub at gcc dot gnu dot org]

------- Comment #11 from jakub at gcc dot gnu dot org  2008-09-17 21:53 -------
Fixed on the trunk.


-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|4.3.3                       |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536

[Bug fortran/37536] [4.3/4.4 Regression] a mfcr is produced instead of branches for DO loops

[jakub at gcc dot gnu dot org]

------- Comment #10 from jakub at gcc dot gnu dot org  2008-09-17 21:53 -------
Subject: Bug 37536

Author: jakub
Date: Wed Sep 17 21:51:50 2008
New Revision: 140433

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=140433
Log:
        PR fortran/37536
        * trans-stmt.c (gfc_trans_do): Optimize integer type non-simple
        do loop initialization.

Modified:
    trunk/gcc/fortran/ChangeLog
    trunk/gcc/fortran/trans-stmt.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37536