time profiling and nested function inlining

time profiling and nested function inlining

[Quôc Peyrot]

Hello, I have two questions, sorry if they have already been asked,  
but I searched
through the archives and couldn't find the answers:

- I tried to do some time profiling (Mac OS X, ppc G4) but for some  
reasons it doesn't
seem to work. I compiled with OCamlMakefile using the command line  
"make profiling-native-code".
When I execute the program, it does generate gmon.out, but when I run  
gprof the only thing I get is:

                                   called/total       parents
index  %time    self descendents  called+self    name           index
                                   called/total       children

                 0.00        0.00       1/1           ___fixunsdfdi  
[476]
[21]     0.0    0.00        0.00       1         ___stub_getrealaddr  
[21]

-----------------------------------------------

and

%   cumulative   self              self     total
time   seconds   seconds    calls  ms/call  ms/call  name
0.0       0.00     0.00        1     0.00     0.00   
___stub_getrealaddr [21]

Thus, I'm wondering whether or not time profiling is supported on PPC  
G4.
And if it is,  can someone give me some clue to debug this issue?
If it isn't, I would appreciate if someone could give me alternative  
solutions (apart from using an intel computer ;) )

- I was looking at the asm output to get familiar with efficient  
coding style, and I tried the following example:

let rec log2_acc value acc =
   if value = 0
   then acc
   else log2_acc (value lsr 1) (acc + 1)

let log2 value =
   log2_acc value 0

which compiles to (using "ocamlopt -inline 100 -unsafe -S)

_camlTest_regular__log2_acc_57:
L101:
         cmpwi   r3, 1
         bne     L100
         mr      r3, r4
         blr
L100:
         srwi    r5, r3, 1
         ori     r3, r5, 1
         addi    r4, r4, 2
         b       L101
         .globl  _camlTest_regular__log2_60
         .text
         .align  2
_camlTest_regular__log2_60:
L102:
         li      r4, 1
         b       _camlTest_regular__log2_acc_57

Although log2_acc could have been inlined (which might not be  
beneficial in this case anyway), it looks quite ok.

But when I tried with a nested function:

let log2 value =
   let rec log2_acc value acc =
     if value = 0
     then acc
     else log2_acc (value lsr 1) (acc + 1)
   in
   log2_acc value 0

I got the following output:
_camlTest_nested__2:
         .long   _caml_curry2
         .long   5
         .long   _camlTest_nested__log2_acc_59
         .globl  _camlTest_nested__log2_acc_59
         .text
         .align  2
_camlTest_nested__log2_acc_59:
L101:
         cmpwi   r3, 1
         bne     L100
         mr      r3, r4
         blr
L100:
         srwi    r5, r3, 1
         ori     r3, r5, 1
         addi    r4, r4, 2
         b       L101
         .globl  _camlTest_nested__log2_57
         .text
         .align  2
_camlTest_nested__log2_57:
L102:
         addis   r4, 0, ha16(_camlTest_nested__2)
         addi    r4, r4, lo16(_camlTest_nested__2)
         li      r4, 1
         b       _camlTest_nested__log2_acc_59

I'm wondering what these computations before the call are (frame?)  
and why the compiler couldn't get rid of them.
Not that I am utterly concerned by these small extra computations...  
I'm just curious.

Thanks for the help/explanations,

-- 
Best Regards,
Quôc

Re: [Caml-list] time profiling and nested function inlining

[Daniel Bünzli]

If you have the development tools installed, you can profile on  
macosx using shark (you don't even need to compile with -p). Invoke  
it from the command line on your executable as follows:

 > shark -i -1 -q ./yourexec.opt args

This will write a .mshark file in the directory that you can open  
with Shark.app.

 > open *.mshark

Best,

Daniel

P.S. I remember having problems if the executable did not run for  
long enough, shark couldn't take any samples.

Re: [Caml-list] time profiling and nested function inlining

[Daniel Bünzli]

Le 6 déc. 06 à 09:55, Daniel Bünzli a écrit :

> If you have the development tools installed, you can profile on  
> macosx using shark (you don't even need to compile with -p).

However if you want to see the name of system function calls in the  
profile you need to compile with -g.

Daniel

Re: [Caml-list] time profiling and nested function inlining

[Quôc Peyrot]

On Dec 6, 2006, at 12:55 AM, Daniel Bünzli wrote:

> If you have the development tools installed, you can profile on  
> macosx using shark (you don't even need to compile with -p). Invoke  
> it from the command line on your executable as follows:
>
> > shark -i -1 -q ./yourexec.opt args
>
> This will write a .mshark file in the directory that you can open  
> with Shark.app.
>
> > open *.mshark

Thanks, it worked like a charm.

Anyone for my second question from my original email (about the  
nested function)?

The more I look at the assembly output, the more I am puzzled.  
Another simple example:

for i = 0 to len - 1 do
   for j = 0 to len - 1 do
     array.(i).(j) <- 0;
   done;
done;

doesn't seem to be equivalent to

for i = 0 to len - 1 do
   let array_i = array.(i) in
   for j = 0 to len - 1 do
     array_i.(j) <- 0;
   done;
done;

in the former,  the compiler doesn't detect the invariant "array.(i)"  
and keeps it inside the inner loop.
I tried to pass -ccopt -O3 to ocamlopt but it didn't seem to change  
anything.

-- 
Best Regards,
Quôc

Re: [Caml-list] time profiling and nested function inlining

[Jon Harrop]

On Thursday 07 December 2006 10:43, Quôc Peyrot wrote:
> in the former,  the compiler doesn't detect the invariant "array.(i)"
> and keeps it inside the inner loop.
> I tried to pass -ccopt -O3 to ocamlopt but it didn't seem to change
> anything.

Ocamlopt doesn't do that optimisation and -ccopt won't have any effect unless 
you're compiling C code.

-- 
Dr Jon D Harrop, Flying Frog Consultancy Ltd.
Objective CAML for Scientists
http://www.ffconsultancy.com/products/ocaml_for_scientists