Re: conditionals

Re: conditionals

[Eljay Love-Jensen]

Hi Helge,

I don't think there's an "invariant test motion" optimization, which would consolidate the conditional code block bodies for you.

That could be useful in some cases.

Especially if your test for swap/don't swap was in a macro, and you really didn't have an easy means at the code level to pull the invariant test condition out of the macro -- so consolidating the conditional bodies by hand would be even more onerous.

--Eljay

Re: conditionals

[Helge Bahmann]

Hello Eljay,

> I don't think there's an "invariant test motion" optimization, which
> would consolidate the conditional code block bodies for you.
>
> That could be useful in some cases.
>
> Especially if your test for swap/don't swap was in a macro, and you
> really didn't have an easy means at the code level to pull the invariant
> test condition out of the macro -- so consolidating the conditional
> bodies by hand would be even more onerous.

Yes, this is the case in my application (inline functions instead of
macros, though).

How hard would it be for the compiler to do that optimization?
Intuitively, it looks obvious, at least not any more complicated than cse,
but then I really don't know anything.

Thanks
-- 
Helge Bahmann <hcb@chaoticmind.net>                     /| \__
The past: Smart users in front of dumb terminals       /_|____\
                                                     _/\ |   __)
$ ./configure                                        \\ \|__/__|
checking whether build environment is sane... yes     \\/___/ |
checking for AIX... no (we already did this)            |

Re: conditionals

[Eljay Love-Jensen]

Hi Helge,

>How hard would it be for the compiler to do that optimization?

I'm not sure.  I have to defer to the GNUrus who wrote GCC.

My experience in compiler writing goes only so far as Aho's dragon book and a computer science class.

:-(
--Eljay

conditionals

[Helge Bahmann]

Hello list,

Question about an optimization feature: I have a program that is full of
conditional variable assignments like

  if (swapped) x->a=bswap_32(a); else x->a=a;
  if (swapped) x->b=bswap_32(b); else x->b=b;
  if (swapped) x->c=bswap_32(c); else x->c=c;
  ...

(after inlining functions); I wonder if there is a way for the compiler to
make the obvious transformation into

  if (swapped) {
    x->a=bswap_32(a);
    x->b=bswap_32(b);
    ...
  } else {
    x->a=a;
    x->b=b;
    ...
  }

which is way more efficient; I tried various gcc versions (2.95, 3.2 and
3.3), even Intel icc, and neither compiler appears to be capable of doing
that (btw icc is _really_ stupid when it comes to byte-swapping)

Am I stuck with rewriting the code or is there hope that this can be
optimized by the compiler?

Best regards
-- 
Helge Bahmann <hcb@chaoticmind.net>                     /| \__
The past: Smart users in front of dumb terminals       /_|____\
                                                     _/\ |   __)
$ ./configure                                        \\ \|__/__|
checking whether build environment is sane... yes     \\/___/ |
checking for AIX... no (we already did this)            |