Pervasives.{min,max} not specialized to floats by ocamlopt

Pervasives.{min,max} not specialized to floats by ocamlopt

[Will M Farr]

Hello all,

I was profiling some numerical code compiled with ocamlopt today (for  
the record: PPC G4, Mac OS X 10.4.7, ocamlopt 3.09.2), and I noticed  
that a lot of time was taken up in the compare_val function, even  
though type inference asserted that all values I was comparing were  
floats.  Turns out the culprits were Pervasives.min and  
Pervasives.max---these functions do not specialize to the appropriate  
assembly instructions if their arguments are known to be floats.   
Including my own min and max functions in the module to shadow these  
removed the calls to compare_val.  I'm curious if this is well-known  
behavior (a google search didn't turn up anything).  It's, of course,  
not a bug, but it's definitely surprising when all the other  
polymorphic comparisons are special-cased for floats!

I don't know how hard it would be to change this in ocamlopt, but it  
seems like it would be pretty simple.  If it is simple, it would be a  
good candidate to change in the next release for those of us who do a  
lot of float-comparing.

Thanks!
Will

Re: [Caml-list] Pervasives.{min, max} not specialized to floats by ocamlopt

[Olivier Andrieu]

 Will M Farr [Thursday 17 August 2006] :
 >
 > Hello all,
 > 
 > I was profiling some numerical code compiled with ocamlopt today (for  
 > the record: PPC G4, Mac OS X 10.4.7, ocamlopt 3.09.2), and I noticed  
 > that a lot of time was taken up in the compare_val function, even  
 > though type inference asserted that all values I was comparing were  
 > floats.  Turns out the culprits were Pervasives.min and  
 > Pervasives.max---these functions do not specialize to the appropriate  
 > assembly instructions if their arguments are known to be floats.   

Indeed. That's because min and max are not primitives provided by the
runtime library but regular caml functions from pervasives.ml :
,----
| let min x y = if x <= y then x else y
| let max x y = if x >= y then x else y
`----

 > I don't know how hard it would be to change this in ocamlopt, but
 > it seems like it would be pretty simple.

I don't know, float comparisons tend not to be 'pretty simple' usually
(because of NaNs).
-- 
   Olivier

Re: [Caml-list] Pervasives.{min, max} not specialized to floats by ocamlopt

[Will M Farr]

On Aug 18, 2006, at 11:34 AM, Olivier Andrieu wrote:

> Indeed. That's because min and max are not primitives provided by the
> runtime library but regular caml functions from pervasives.ml :
> ,----
> | let min x y = if x <= y then x else y
> | let max x y = if x >= y then x else y
> `----

In this case, it seems strange that they weren't inlined---when I  
defined my own min and max functions, *they* were inlined.  Could it  
be because I was working inside a functor, or am I mistaken in my  
belief that ocaml does cross-module inlining?

Will

Re: [Caml-list] Pervasives.{min, max} not specialized to floats by ocamlopt

[Shawn]

Olivier Andrieu wrote:
>  Will M Farr [Thursday 17 August 2006] :
>
>   
>  > I don't know how hard it would be to change this in ocamlopt, but
>  > it seems like it would be pretty simple.
>
> I don't know, float comparisons tend not to be 'pretty simple' usually
> (because of NaNs).
>   

So only do it with -ffast-math, which already special-cases Pervasives 
functions like sin and cos...

It looks like only the x86 port supports that option right now, though. 
Hmm. It really should be present on all architectures, even if it 
doesn't actually do anything.