[PATCH] PR fortran/31593: Speed up loops with DO variables as procedure arguments

[Tobias Schlüter]

Steven Bosscher wrote:
> 2009/8/24 Tobias Schlüter <tobias.schlueter@physik.uni-muenchen.de>:
>> diff -r 395e36e913de gcc/fortran/trans-stmt.c
> 
> Patches are often a bit more readable if you use diff -p (shows the
> function you're patching)...

I've not yet managed to teach mercurial that bit, but I'll look again. 
It makes ChangeLog writing harder as well :)

>> --- a/gcc/fortran/trans-stmt.c  Sun Aug 23 18:13:56 2009 +0000
>> +++ b/gcc/fortran/trans-stmt.c  Mon Aug 24 21:43:32 2009 +0200
>> @@ -756,6 +756,49 @@
>>  }
>>
>>
> 
> Comment here to explain what you are doing (stack of DO loop vars) and
> why (show the transformations you intend to do)?
> 
>> +typedef struct do_var_stack {
>> +  struct do_var_stack *prev;
>> +  gfc_symbol *sym;
>> +} do_var_stack;
>> +
>> +static do_var_stack *dvs_top = NULL;
>> +
>> +
>> +static void
>> +push_do_var (gfc_symbol *s)
> 
> Comment before the function please.  Same for pop_do_var() and gfc_is_do_var()

Sorry, yes, will do.  I didn't write them, when I wrote this originally, 
and forgot when I warmed the patch up.

>> +{
>> +  do_var_stack *p
>> +    = (do_var_stack *)gfc_getmem(sizeof(do_var_stack));
> 
> Since you're putting them on a stack -- why not use an obstack of
> objects instead? See obstack examples throughout the middle-end.

I'll have a look, I take it that I could just search linearly instead of 
traversing a linked list?

> Ah, you explained it here.  Maybe just point to here from the other site.

I thought about that, but I thought that other optimizations may come to 
our collective minds.

>> +      if (gfc_is_do_var (expr->symtree->n.sym)
>> +         && !expr->symtree->n.sym->attr.target)
> 
> I would put the TARGET check before the gfc_is_do_var() check.

Makes sense.

> This patch speeds up one of my toy codes by ~12%. Nice :-)

Heh, I'm afraid that it only speeds up toy code, Thomas's contrived 
example gains some 40%.

Cheers,
- Tobi