From: Andrew Crabtree <andrewc AT typhoon DOT rose DOT hp DOT com>
Message-Id: <199802101830.AA273515458@typhoon.rose.hp.com>
Subject: Re: src/gcc.opt in alpha 980101
To: eldredge AT ap DOT net (Nate Eldredge)
Date: Tue, 10 Feb 1998 10:30:58 PST
Cc: djgpp-workers AT delorie DOT com
In-Reply-To: <199802100441.UAA02942@adit.ap.net>; from "Nate Eldredge" at Feb 9, 98 8:41 pm
Reply-To: andrewc AT rosemail DOT rose DOT hp DOT com
Precedence: bulk


> >> Why do the switches used to compile the library still include 
> >> -fno-strength-reduce?  Doesn't this belong to pre-2.7.2.1 gcc and is 
> >> solved in later versions?
Hmm - I thought that 2.7.2.1 jsut disabled it by default but that the 
bug still existed.  It was supposed to be rather rare to come across 
it though.
 
> and see section 4.22. Basically, the theory is that strength reduction tends
> to replace multiplication with shifts and adds, costing more registers which
> the x86 can ill afford.
This depends on the cost table used.  If you use gcc 2.7.2.x which 
only knows about 386/486 then it will probably not make the best 
decisions here.  Gcc 2.8, and of course pgcc are much better at strengh
reduction determination.
 
> They do give the caveat, however, that one should try different flags and
> see. Has anyone done this with the libc, or do they plan to?
The only place it makes a huge difference is floating point code.  It
would be worthwhile to try unrolling loops w/ strength reduction there.