Broken instruction pre-fetch hueristics on the 7030?

Thread View: alt.folklore.computers

3 messages

[T] Back to Threads [V] Tree View [G] Group Articles

3 total messages Started by cos... Bernie Cosell Tue, 09 Oct 1990 14:08

#5286

Author: cos... Bernie Cosell
Date: Tue, 09 Oct 1990 14:08

20 lines
1488 bytes

Perhaps this is the place to get a bit of myth/rumor I picked up years and
years ago straightened out.  

What i heard was that in an attempt to get the 7030 to go as fast as
possible, IBM gave it the first (?) instruction lookahead/prefetch
logic.  Not only did they prefetch the next instruction on 'ordinary'
instructions, but they also prefetched across GOTOs.  The tricky part,
as I heard it, was that for _conditional_ transfers, they put in a fair
pile of heuristics to try to *guess* which branch a conditional would
take.

The rumor basically asserted that they got the heuristics almost always
wrong [like 70+% of the time], and so when it came time to make a
similar lookahead box for the 360, they just used the SAME heuristics,
but simply reversed the 'which branch' decision... and presto: went
from being WRONG 70% of the time to being RIGHT 70% of the time.

Is any of this grounded in reality?

  /Bernie\

Re: Broken instruction pre-fetch hueristics on the 7030?

#5287

Author: Boyd Roberts
Date: Thu, 11 Oct 1990 01:36

8 lines
911 bytes

Didn't the CRISP (develloped at Bell Labs) have a branch direction bit?
The upshot being that most of the time the branch direction was wrong.
So negating the logic proved a big win.


Boyd Roberts			bo...@necisa.ho.necisa.oz.au

``When the going gets wierd, the weird turn pro...''

Re: Broken instruction pre-fetch hueristics on the 7030?

#5288

Author: jcmo... Joe Morris
Date: Thu, 11 Oct 1990 14:09

22 lines
1717 bytes

In a recent article boyd@necisa.ho.necisa.oz (Boyd Roberts) writes:
>Didn't the CRISP (develloped at Bell Labs) have a branch direction bit?
>The upshot being that most of the time the branch direction was wrong.
>So negating the logic proved a big win.

It's not a hardware item, but the first FORTRAN I ever used (FORTRAN II
under FMS -- Fortran Monitor System -- on a 7090) included a statement
to tell the compiler the anticipated distribution of results in an
arithmetic IF.  My old manual is buried where I haven't seen it in years,
but the format was something like:

      FREQUENCY  mm,zz,pp

for the three possibilities.  (Don't ask me what happened if mm+zz+pp != 100.
Maybe there were just two parameters.)  The idea was that the compiler could
optimize the machine code and save a few cycles.

As I recall the FMS system documentation stated that the FREQUENCY statement
would be accepted and syntax-checked but otherwise ignored.  I never saw
any other implementation of the statement.

Joe Morris

Thread Navigation

This is a paginated view of messages in the thread with full content displayed inline.

Messages are displayed in chronological order, with the original post highlighted in green.

Use pagination controls to navigate through all messages in large threads.

Back to All Threads