Re: forall and do loop
- From: highegg <highegg@xxxxxxxxx>
- Date: Fri, 27 Jul 2007 12:53:37 -0000
On Jul 27, 2:33 pm, Tobias Burnus <bur...@xxxxxxxx> wrote:
On Jul 27, 11:55 am, highegg <high...@xxxxxxxxx> wrote:That's cool - I didn't know.
Commercial compilers (Intel, EkoPath) even offer the possibility to
use the profiling information for subsequent optimizations, which is
potentially very powerful approach.
GCC has the same - does this make GCC a commercial compiler? ;-)
GCC has the options: -fprofile-generate and -fprofile-use.
Depending on the program, performance might be gained or lost. For the
current gfortran 4.3 and the Polyhedron test suite, seehttp://physik.fu-berlin.de/~tburnus/gcc-trunk/benchmark/#rt
As one test gets 24% slower, the few per cent gains for the other
tests average out (in the geometric mean).
(I filled a bugreport about it, #32913.)
Tobias
That's why I called it a "potentially powerful" technique, I know that
it does not always work well. Nevertheless, it seems to me that
the knowledge of where the hot spots lie may enable the compiler to do
some aggresive optimizations that would otherwise, if used everywhere
throughout the code, lead to unacceptable code bloat.
For instance, generating a separate unit-stride and non-unit-stride
code for array expressions (I'm not sure if gfortran does this).
.
- Follow-Ups:
- Re: forall and do loop
- From: Janne Blomqvist
- Re: forall and do loop
- From: Tim Prince
- Re: forall and do loop
- References:
- forall and do loop
- From: aeroguy
- Re: forall and do loop
- From: glen herrmannsfeldt
- Re: forall and do loop
- From: aeroguy
- Re: forall and do loop
- From: highegg
- Re: forall and do loop
- From: aeroguy
- Re: forall and do loop
- From: highegg
- Re: forall and do loop
- From: Tobias Burnus
- forall and do loop
- Prev by Date: Re: forall and do loop
- Next by Date: Re: forall and do loop
- Previous by thread: Re: forall and do loop
- Next by thread: Re: forall and do loop
- Index(es):
Relevant Pages
|
|