Path: utzoo!attcan!uunet!mcsun!hp4nl!charon!dik
From: dik@cwi.nl (Dik T. Winter)
Newsgroups: comp.arch
Subject: Re: speculative execution
Message-ID: <2354@charon.cwi.nl>
Date: 13 Oct 90 03:23:59 GMT
References: <1990Oct10.164353.21070@rice.edu> <PHILIP.90Oct11102731@beeblebrox.dle.dg.com> <22587@grebyn.com>
Sender: news@cwi.nl
Organization: CWI, Amsterdam
Lines: 32

In article <22587@grebyn.com> ckp@grebyn.UUCP (Checkpoint Technologies) writes:
 > Why not include condition code bits with each register? Store these when
 > the register value is stored, when it pops out the end of a pipeline.
 > Include a flag for NAN (not a new idea, right?), and store this for
 > things like divide-by-0.  Then you can trap when the register is used as
 > a source operand, and you can code explicit tests (and maybe traps) when
 > you want to, on operations that could very well be taking place
 > simultaneously.
Something similar is done in all machines designed by Seymour Cray.  There
is no instruction that generates a result trap, only source traps are
generated.  There are advantages and disadvantages.
Advantages:
1.  It is possible to initialize all variables at a trapping value.  This
    will very fast eleminate the use of uninitialized values.  (CDC users
    will know this:  LDSET(PRESET=UNDEF), although this would not work
    with integer variables.)
2.  It makes it possible for the compiler to generate spurious results.
    I.e. if the array B has never been initialized the following loop:
	for I from 1 to N do if A[I] != 0.0 then B[I] = 1.0/A[I]
    can be modified by omitting the test without adverse effects.
    This can be beneficial on vector processors.  In fact the Fortran
    compiler on the NEC SX/2 generates spurious results in some cases;
    alas, they trigger a result trap.
Disadvantage:
1.  Debugging is very problematical.  I know, I have debugged a number of
    programs on the CDC Cyber.  It is very difficult to see where a bad
    result originates.  This is even worse if (as on the Cyber) there are
    multiple functional units so that it is even problematical which unit
    did generate the trap.
--
dik t. winter, cwi, amsterdam, nederland
dik@cwi.nl