[Bug rtl-optimization/19001] New: Loops with power of two step and variable bounds not unrolled

[Bug rtl-optimization/19001] New: Loops with power of two step and variable bounds not unrolled

[rakdver at gcc dot gnu dot org]

Loop of the following form

int check(int a,int b, char *c)
{
for(;a<b;a+=4)
  if(c[a]==1) return a;
return a;
}

(with step being power of two and > 1, and bounds variable) does not get unrolled.
It may happen that this loop is infinite, but this should not prevent unrolling.

-- 
           Summary: Loops with power of two step and variable bounds not
                    unrolled
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P2
         Component: rtl-optimization
        AssignedTo: rakdver at gcc dot gnu dot org
        ReportedBy: rakdver at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-12-15 04:55 -------
Confirmed as a regression from 3.3.2, this happens in libstdc++ as mentioned other places.
And I also check that 3.3.2 did not unroll a loop with non power of two step aka:
int check(int a,int b, char *c)
{
for(;a<b;a+=5)
  if(c[a]==1) return a;
return a;
}
is not unrolled in 3.3.2.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
      Known to work|                            |3.3.2
   Last reconfirmed|0000-00-00 00:00:00         |2004-12-15 04:55:28
               date|                            |
            Summary|Loops with power of two step|[3.4/4.0 Regression] Loops
                   |and variable bounds not     |with power of two step and
                   |unrolled                    |variable bounds not unrolled
   Target Milestone|---                         |3.4.4


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[pcarlini at suse dot de]

------- Additional Comments From pcarlini at suse dot de  2004-12-15 10:20 -------
To Zdenek and Andrew it's obvious, but maybe it's worth mentioning that the
following kind of loop, not involving explicitly power of two steps, pointers
instead, and *very* simple, is also not unrolled anymore:

int*
check(int* a, int* b)
{
 for (; a < b; ++a)
   if (*a == 1)
     return a;
 return a;
}

Maybe the Summary could be made a tad clearer

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pcarlini at suse dot de


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[chris at bubblescope dot net]

------- Additional Comments From chris at bubblescope dot net  2004-12-15 12:28 -------
Also (stating the perhaps blindingly obvious), this bug is not being caused by
the fact that the loop contains a return statement.

int check(int* a,int *b)
{
int* returnval=b;
for(;a<b;++a)
  if(*a) returnval=a;
return returnval;
}

Also is not optimised.

Out of interest, surely when the loop actually involves pointers, then "looping
a pointer around the end of memory" is undefined behaviour, and therefore this
doesn't have to be considered in terms of optimising (although this won't help
the case in the original bug report).

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[rakdver at atrey dot karlin dot mff dot cuni dot cz]

------- Additional Comments From rakdver at atrey dot karlin dot mff dot cuni dot cz  2004-12-15 12:52 -------
Subject: Re:  [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

> ------- Additional Comments From chris at bubblescope dot net  2004-12-15 12:28 -------
> Also (stating the perhaps blindingly obvious), this bug is not being caused by
> the fact that the loop contains a return statement.
> 
> int check(int* a,int *b)
> {
> int* returnval=b;
> for(;a<b;++a)
>   if(*a) returnval=a;
> return returnval;
> }
> 
> Also is not optimised.
> 
> Out of interest, surely when the loop actually involves pointers, then "looping
> a pointer around the end of memory" is undefined behaviour, and therefore this
> doesn't have to be considered in terms of optimising (although this won't help
> the case in the original bug report).

And it won't help at all, since on rtl we no longer have a reliable
information about pointers.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[rakdver at gcc dot gnu dot org]

------- Additional Comments From rakdver at gcc dot gnu dot org  2004-12-16 09:18 -------
Patch:

http://gcc.gnu.org/ml/gcc-patches/2004-12/msg01203.html

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |patch


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[chris at bubblescope dot net]

------- Additional Comments From chris at bubblescope dot net  2004-12-18 19:20 -------
I thought I'd post this here rather than as a new bug.. I will do if there isn't
a reply within a week or so.

A very similar regression is:

int check(int a,int b, char* c)
{
for(;a!=b;a+=4)
  if(c[a]==1) return a;
return a;
}

ie where "a<b" is replaced with "a!=b". Is there a simple / similar fix for this?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-12-18 20:14 -------
Subject: Bug 19001

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	rakdver@gcc.gnu.org	2004-12-18 20:14:25

Modified files:
	gcc            : ChangeLog loop-iv.c 

Log message:
	PR rtl-optimization/19001
	* loop-iv.c (iv_number_of_iterations): Record assumptions for loops
	with power of two step to 'infinite' field.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.6880&r2=2.6881
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/loop-iv.c.diff?cvsroot=gcc&r1=2.24&r2=2.25



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4/4.0 Regression] Loops with power of two step and variable bounds not unrolled

[rakdver at gcc dot gnu dot org]

------- Additional Comments From rakdver at gcc dot gnu dot org  2004-12-18 20:28 -------
What regression do you observe in that testcase? It gets unrolled just fine for 
me.

If there is some other problem, please be more specific (and probably also 
create a new PR for it).

Closing this one as fixed.

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4 Regression] Loops with power of two step and variable bounds not unrolled

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-12-18 20:42 -------
Reopening as it is a still a 3.4 regression.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
      Known to work|3.3.2                       |3.3.2 4.0.0
         Resolution|FIXED                       |
            Summary|[3.4/4.0 Regression] Loops  |[3.4 Regression] Loops with
                   |with power of two step and  |power of two step and
                   |variable bounds not unrolled|variable bounds not unrolled


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4 Regression] Loops with power of two step and variable bounds not unrolled

[chris at bubblescope dot net]

------- Additional Comments From chris at bubblescope dot net  2004-12-19 10:59 -------
Sorry, I was testing for the existance of loop-unrolling by timing, rather than
looking at the code.

The compiler is unrolling the code I provided, but then optimised quite badly.
As this might not be the fault of the loop optimiser, but some other piece of
code, I have opened a new bug report for it: 19078

Sorry for the inconvience.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4 Regression] Loops with power of two step and variable bounds not unrolled

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-12-20 11:46 -------
Notice that no testcase was added to 4.0 either. 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4 Regression] Loops with power of two step and variable bounds not unrolled

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|REOPENED                    |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001

[Bug rtl-optimization/19001] [3.4 Regression] Loops with power of two step and variable bounds not unrolled

[mmitchel at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|3.4.4                       |3.4.5


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19001