This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: Local optimization on i386 ?


You do have to be careful with some of these simple peephole optimizations.
CodeWarrior has a peephole phase right before the internal assembler
produces object code, and we've had several released code generation bugs
because of what seemed harmless... consider the sequence mentioned already:

    subl %40, eax
    subl %56, eax

It may seem obvious at first to combine these into a single "subl %96, eax",
but it isn't harmless because the carry and overflow flags would now be set
differently than in the original set of instructions, for example if you
started with eax == 32.  When flags are involved, you need to follow the
flow of control after the instructions and ensure that the flags are killed
by a succeeding instruction before they are consumed.

(Hmmm... this kind of coding could be useful in a switch block to determine
a range of values -- have the two subs, then you know its in the range if
NF=1 and CF=1 after the two subs)

Most mov peepholes are OK, however, since the mov instructions tend to leave
the processor flags alone.

--
Ben Combee, x86/Win32/Novell/Linux CompilerWarrior
http://www.metrowerks.com/
----- Original Message -----
From: "Jamie Lokier" <egcs@tantalophile.demon.co.uk>
To: "Richard Henderson" <rth@cygnus.com>
Cc: "Marc Espie" <espie@quatramaran.ens.fr>; <law@cygnus.com>;
<egcs@egcs.cygnus.com>
Sent: Tuesday, November 30, 1999 19:29
Subject: Re: Local optimization on i386 ?


> I've examined GCC output for many years.  I've even done a little
> proprietary work on the x86 machine description.  GCC does many very
> clever optimisations.  Yet I have always been amazed that it still emits
> silly sequences, like "movl %eax,%edx; movl %edx,%eax", and stores
> followed by no loads to stack slots at the end of a function, for
> example.
>
> Such simple things that GAS, or even sed could peephole away.
>
> But I also know that after everything has been done, there are /obvious/
> redundancies that could be cleaned up by a peephole and aren't.  Really
> obvious things like redundant stores followed by no uses.



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]