Mail Archives: djgpp-workers/2000/03/02/12:22:46
On Thu, 2 Mar 2000, Eli Zaretskii wrote:
> Funny, a similar problem seems to be bugging the GDB test suite. You
> might find the thread whose subject is "Regressions problem (200
> failures)" interesting. If you don't read gdb AT sourceware DOT cygnus DOT com, you
> can browse it via the Web at http://sourceware.cygnus.com/ml/gdb/2000-q1.
I'll see, if I find the time. Spare online time is harder to get at than
spare time at home, where I can brood over the GCC source tree as long as
I don't fall off my chair sleeping :-)
> > Compiling with '-g -O1 -mno-schedule-prologue' seemed to fix the problem.
>
> What happens if you use -O2 -mno-schedule-prologue?
No difference at all, in the test case. My test function is so simple
that -O2 makes no difference whatsoever, compared to -O1 (I removed the
sqrt() call and #include <math.h>, to get more easily diff-able output
from all stages, including the '-d<letter>' RTL debugging dumps.
> If that still
> produces correct debug info, perhaps we want to introduce
> -mno-schedule-prologue into specs?
I think the more important question would be: does the bug really only
happen in ultra-short (<= 2 active code lines) functions, where this
workaround would probably fix it? And, of course, _how_ does
'-mno-schedule-prologue' manage to work around the problem? Is it
an optimization <-> debug info problem, or is there a bug in the
RTL introduced to implement the pro/epilogues?
I've had a glance at this week's snapshot of gcc (egcs-20000228), and
there's been quite some activity in the i386 stack frame handling stuff,
recently, including the prologue/epilogue generator.
Hans-Bernhard Broeker (broeker AT physik DOT rwth-aachen DOT de)
Even if all the snow were burnt, ashes would remain.
- Raw text -