Re: No NRA Bootstrap on Alpha, was: Re: A few SPEC95 results forthe new register allocator

[Michael Matz <matzmich at cs dot tu-berlin dot de>]

Hi,

On Sat, 24 Feb 2001, Toon Moene wrote:
> > Hmm, here the stage1 compiler crashes, which is compiled with the system
> > compiler, so at least this can't be a miscompilation resulting from the
> > new reg-allocator.  It still can be caused by it of course.  To test this,
> > you can compile the above file (in the gcc dir, with the exact same
> > command as above), with the additional flag -fno-new-ra.  If it still
> > crashes, it's not caused by new-RA in any way, but has other reasons (if
> > that's the case we might consider to merge HEAD again into new-RA, if HEAD
> > bootstraps on Alpha).
>
> Alas - adding -fno-new-ra to the compiler flag in the above makes it
> compile; so the ball is in your court :-)

OK. Hmm, actually not OK ;)

> When I try to continue with the stage2 compiler built that way, it
> hangs while compiling libgcc/./_divdi3.

If you let it run really long, does it somewhen breaks with a "Couldn't
find a coloring"?  I guess so, it probably is spilling a Bad Web over and
over.  I should lowering the cutoff, where it's aborting (in reg_alloc()).
40 passes is too much, probably 10 or so is enough.  If it takes more,
then it anyway is somehow broken.

> Anything that I can do to help it track down ?

Hmm. Yes.  Please lower the number 40 in ra.c:reg_alloc() to 5 (you will
see where I mean), so it somewhen stops.  Then produce a divdi3.i (from
the command compiling divdi3, with -E replacing any -c).  Then compile the
above thing (divdi3.i) with -dl (with the broken compiler of course).
Send me the resulting divdi3.i.18.lreg, and the preprocessed divdi3.i.
(All compressed possibly)  This hopefully allows me to see the
circumstances and which webs are the bad ones.  Many thanks.


Ciao,
Michael.