Re: static estimation of conditional branches?

markh@csd4.csd.uwm.edu (Mark)
Wed, 9 Dec 1992 04:43:41 GMT

          From comp.compilers

Related articles
static estimation of conditional branches? mahlke@crhc.uiuc.edu (Scott Mahlke) (1992-12-08)
Re: static estimation of conditional branches? markh@csd4.csd.uwm.edu (1992-12-09)
Re: static estimation of conditional branches? bill@amber.csd.harris.com (1992-12-09)
Re: static estimation of conditional branches? tom@derby.cs.wisc.edu (1992-12-09)
Re: static estimation of conditional branches? tsych@sedona.intel.com (1992-12-09)
Re: static estimation of conditional branches? hagerman@ece.cmu.edu (1992-12-09)
Re: static estimation of conditional branches? bill@amber.csd.harris.com (1992-12-10)
Re: static estimation of conditional branches? hrubin@pop.stat.purdue.edu (1992-12-11)
[11 later articles]
| List of all articles for this month |

Newsgroups: comp.compilers
From: markh@csd4.csd.uwm.edu (Mark)
Organization: Computing Services Division, University of Wisconsin - Milwaukee
Date: Wed, 9 Dec 1992 04:43:41 GMT
Keywords: optimize, comment
References: 92-12-029

Scott Mahlke <mahlke@crhc.uiuc.edu> writes:
>I was wondering if anyone is aware of any studies performed to staticly
>estimate the direction of non-loop back conditional branches in C
>programs. ...


How will the proposed method handle this case?


I = 0;
if (I < 0) {
      Branch 1
} else {
      Branch 2
}


The probability of branching in each direction is clear.


How will it handle this case?


I = Function();
if (I < 0) {
      Branch 1
} else {
      Branch 2
}


No method can give you sensible results on both cases without also
deriving probability distributions for all the variables in the program.


The method has to know what the variables are doing because these are what
determine the probabilities of the branches.
[Perhaps one could collect statistics of real programs and try to make
estimates based on measurements of similar code in the measured programs.
Of course, ``similar'' is the hard part. -John]
--


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.