Re: [Caml-list] Accessing record fields

[Goswin von Brederlow <goswin-v-b@web.de>]

On Sun, Sep 15, 2013 at 12:38:50PM +0200, Jacques-Henri Jourdan wrote:
> I think both are very close.
> 
> I would say that f is faster, because r.a is loaded only once. However,
> this can increase register pressure...
> 
> I would say to use whichever is more easy to use a a given context, and
> to tune specifically the code if this is the bottleneck: this is very
> unlikely to be the performance bottleneck, and there won't be a big
> performance improvement anyway...
> 
> 
> -- 
> JH
> 
> Le 15/09/2013 12:30, José Romildo Malaquias a écrit :
> > Hello.
> > 
> > OCaml offers at least two ways of accessing a record field: using the
> > dot notation, and doing pattern matching.
> > 
> > Does one of them deliver better performance than the other?
> > 
> > This may be relevant when a field is accessed multiple times.
> > 
> > For instance:
> > 
> >    type trec = { a : int; mutable b: int }
> > 
> >    let f {a;b} = a * a + b
> > 
> >    let g r = r.a * r.a + r.b
> > 
> > Which one would be preferred in this case: f or g?
> > 
> > Romildo

Why not look at the assembly code they generate?
(here OCaml version 4.00.1 on amd64)

0000000000404030 <camlBla__test_f_1011>:
  404030:       48 89 c3                mov    %rax,%rbx
  404033:       48 8b 03                mov    (%rbx),%rax
  404036:       48 8b 7b 08             mov    0x8(%rbx),%rdi
  40403a:       48 89 c3                mov    %rax,%rbx
  40403d:       48 d1 fb                sar    %rbx
  404040:       48 ff c8                dec    %rax
  404043:       48 0f af c3             imul   %rbx,%rax
  404047:       48 01 f8                add    %rdi,%rax
  40404a:       c3                      retq   
  40404b:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)

0000000000404050 <camlBla__test_g_1014>:
  404050:       48 8b 78 08             mov    0x8(%rax),%rdi
  404054:       48 8b 18                mov    (%rax),%rbx
  404057:       48 d1 fb                sar    %rbx
  40405a:       48 8b 00                mov    (%rax),%rax
  40405d:       48 ff c8                dec    %rax
  404060:       48 0f af c3             imul   %rbx,%rax
  404064:       48 01 f8                add    %rdi,%rax
  404067:       c3                      retq   
  404068:       0f 1f 84 00 00 00 00    nopl   0x0(%rax,%rax,1)
  40406f:       00 

Other than reordering there are 2 differences:

1) In "f" the record is first copied to %rbx, then deconstructed while
   in "g" it is deconstructed directly from %rax saving one instruction.

2) In "f" r.a is extracted once and then copied while in "g" it is
   extracted twice. Not sure how much faster a "mov    %rax,%rbx"
   is over a "mov    (%rax),%rax" if at all.

MfG
	Goswin