Path: utzoo!mnetor!uunet!husc6!think!ames!lll-tis!lll-lcc!pyramid!prls!mips!earl
From: earl@mips.UUCP (Earl Killian)
Newsgroups: comp.arch
Subject: Re: taken -vs- untaken branches, Fortran FREQUENCY declaration
Message-ID: <1365@gumby.UUCP>
Date: 20 Jan 88 06:40:24 GMT
References: <496@cresswell.quintus.UUCP> <638@l.cc.purdue.edu> <836@ima.ISC.COM> <2652@mmintl.UUCP>
Lines: 36

In article <2652@mmintl.UUCP>, franka@mmintl.UUCP (Frank Adams) writes:
> I will note in passing that the assumption that continuing a loop is more
> likely than terminating it is often wrong.  Loops of the form:
> 
> while (error is present) {
> 	try to fix it;
> }
> 
> are not all that uncommon.

True, they are common, but precisely because such things execute 0 or
1 times means they don't weight branch statistics as heavily as the
"real" loop that executes 10 times.

Also,
	while (t) {
	  body
	}
is usually compiled as
	if (t) {
	  do {
	    body
	  } while (t);
	}
by a good compiler.  Thus the backward branch in this loop is never
executed in your case, and so doesn't affect the statistics of
backward branches.

When I last measured backward conditional branches on 13 programs, I
found an average of 85% are taken.  The data on forward branches was
52% (i.e. 50-50).  That's the average behavior; individual programs
tended to vary a whole lot on the frequency of forward branches.

One final statistic: backward branches were only 28% of the total on
average.  Another way of thinking about this is that the average loop
contains 2.6 forward branches (e.g. executed if statements).