From:	luke AT jet DOT msk DOT su (Oleg Yu. Polyanski)
Newsgroups:	comp.os.msdos.djgpp
Subject:	Re: Compiling with optimizations...
Date:	27 May 1998 16:35:46 +0400
Organization:	Jet Infosystems
Lines:	37
Message-ID:	<huu7m37c2n1.fsf@goliath.jet.msk.su>
References:	<356c0953 DOT 171134608 AT 1 DOT 0 DOT 0 DOT 2>
NNTP-Posting-Host:	goliath.service.jet.msk.su
Mime-Version:	1.0 (generated by tm-edit 7.108)
To:	djgpp AT delorie DOT com
DJ-Gateway:	from newsgroup comp.os.msdos.djgpp

>>>>> "J" == J Schaerer <jscha AT swbell DOT net> writes:

J> I would like to know (specifically) how to optimize a compile for
J> the following criteria: 1) Speed 2) Size

	first of all, gcc has no special options for speed or for size
optimization. in common cases it is enough to add 

`-O2 -fomit-frame-pointer -m486' (*)

and in computation extensive code also add `-fstrength-reduce'
to your command line. you may say -O3 and some of your simple functions
will be "inlined" (even in C). btw, `-O3' option has nothing with
pentium optimizations). if you want to get the binaries especially 
optimized for p5 or p6 add `-mpentium'/`-m586' or `-mpentiumpro'/`-m686'
respectively. also, gcc 2.8.x optimizes for p6 in much higher degree than
for p5 so you may want to try pgcc (pentium gcc), it has special pentium
oriented optimization facilities. but in general, the options marked
with (*) would be enough for all your code. i would like to recommend some
general rules: select the optimization degree (-O2 or -O3); add cpu type
qualifier (-m386/-m486/-m586/-m686) - it will tune optimizer in an 
architecture specific way; it is useful to say `-fomit-frame-pointer' due
to on intel cpus we have very small number of general use registers and 
extra one will help to us great; trust in optimizer and do not add extra
options (data aligning options may be one good exception in my opinion).

p.s. there are, of course, other extra options to tune the optimizer more
precisely: aligning data in memory (significantly speeds up access to data
stored in memory), you may want to pass parameters to functions in the
registers instead of stack (it is not easy though). it depends on how
good quality optimization you *really* need and how good you understand
code generation technique.
p.p.s. also remember: bigger code does not mean slow execution speed.
p.p.p.s. type `info gcc invo opti' and also `info gcc invo submodel i386'
for other details.

/Oleg

- Raw text -