Re: PowerPC build breakage caused by 8/25/98 loop.c change

Re: PowerPC build breakage caused by 8/25/98 loop.c change

[Jeffrey A Law]

  In message < 9809182157.AA33180@marc.watson.ibm.com >you write:

  > 	I don't have a lot of time to analyze this.  Performing various
  > checkouts and attempting to bootstrap was a very simple, repetitive
  > operation to narrow down the source of the failure.  Maybe the
  > check_dbra_loop changes that were interleaved are causing the current
  > problems. 
Quite possible.  The check_dbra_loop changes caused a number of problems
that took about a week or so to hash out.

jeff

Re: PowerPC build breakage caused by 8/25/98 loop.c change

[Jeffrey A Law]

  In message < 9809182120.AA35006@marc.watson.ibm.com >you write:
  > 	I tried building a recent CVS checkout with loop.c reverted to
  > version 1.65 on 8/25/98 prior to Mark's loop.c changes.  Bootstrapping is
  > successful.  Something about those changes (possibly the
  > AVOID_CC_MODE_COPIES tweak?) causes PowerPC failures.  Using loop.c prior
  > to that date which includes Jeff's and David's loop.c changes still is
  > okay.
Note that AVOID_CCMODE_COPIES is only defined by one port I believe
(MIPS).  So I'd be a little suprised if it was causing problems on
PPC ports.

jeff

Re: PowerPC build breakage caused by 8/25/98 loop.c change

[Mark Mitchell]

>>>>> "David" == David Edelsohn <dje@watson.ibm.com> writes:


    David> 	I don't know enough about these changes to determine
    David> what is wrong.  Mark, I would appreciate if you could find
    David> time to look into this.  I can try to provide some help and
    David> dump traces, if you want to see intermediate output.

I'd like to help.  Unfortunately, I've not got access to a PowerPC
platform.  Can you isolate which function is being miscompiled? 

I doubt the AVOID_CCMODE_COPIES bit is to blame; if anything went
wrong with that code it would probably a trigger an immediate crash as
load_mems_and_recount_loop_regs_set runs.  It sounds like you're
seeing a failure further down the line, right?  But, you can easily
test that hypothesis; take your working version, and substitute just
the AVOID_CCMODE_COPIES in load_mems_and_recount_loop_regs_set.
You'll have to change `VARRAY_CHAR (may_not_optimize, i)' to
`may_not_optimize[i] = 1' since you'll not have the VARRAY stuff.  If
the resulting compiler doesn't bootstrap, then we'll know that's the
problem; otherwise, it's elsewhere.

I'll take a look at the rest, but you can too: it was a
straightforward replacement of `x[i] = j' with `VARRAY_TYPE (x, i) =
j'.  One possibility is that TYPE (which should be CHAR if the thing
was an array of chars, INT if ints, etc.) is wrong somewhere.

-- 
Mark Mitchell 			mark@markmitchell.com
Mark Mitchell Consulting	http://www.markmitchell.com

Re: PowerPC build breakage caused by 8/25/98 loop.c change

[David Edelsohn]

>>>>> Mark Mitchell writes:

Mark> I doubt the AVOID_CCMODE_COPIES bit is to blame; if anything went
Mark> wrong with that code it would probably a trigger an immediate crash as
Mark> load_mems_and_recount_loop_regs_set runs.  It sounds like you're
Mark> seeing a failure further down the line, right?  

	There are a rapid succession of loop.c changes betwen 8/25 and
8/28.  With all of the compiler changes around that time, the compiler
generates an internal error.  Within a day, the compiler produces
executables which crash.  Then finally by 8/28, the compiler simply
produces code which behaves incorrectly.  It is hard for me to disentangle
which loop.c changes are new which could be causing the current failure
from those which are incremental fixes to the previous patches.

	I don't have a lot of time to analyze this.  Performing various
checkouts and attempting to bootstrap was a very simple, repetitive
operation to narrow down the source of the failure.  Maybe the
check_dbra_loop changes that were interleaved are causing the current
problems. 

	I'm trying to bootstrap again with loop.c r1.66 prior to the
check_dbra_loop changes, prior to Jeff's scan_loop tweak.

David

PowerPC build breakage caused by 8/25/98 loop.c change

[David Edelsohn]

	I tried building a recent CVS checkout with loop.c reverted to
version 1.65 on 8/25/98 prior to Mark's loop.c changes.  Bootstrapping is
successful.  Something about those changes (possibly the
AVOID_CC_MODE_COPIES tweak?) causes PowerPC failures.  Using loop.c prior
to that date which includes Jeff's and David's loop.c changes still is
okay.

	I don't know enough about these changes to determine what is
wrong.  Mark, I would appreciate if you could find time to look into
this.  I can try to provide some help and dump traces, if you want to see
intermediate output.

Thanks, David
===============================================================================
David Edelsohn                                      T.J. Watson Research Center
dje@watson.ibm.com                                  P.O. Box 218
+1 914 945 4364 (TL 862)                            Yorktown Heights, NY 10598