Re: EGCS-1.1.2 compile error on -O3

[Franz Sirl <Franz dot Sirl-kernel at lauterbach dot com>]

At 19:17 13.04.99 , David Edelsohn wrote:
>>>>>> Jens Tingleff writes:
>
>Jens> I have just survived compiling installing Berkeley Ptolemy 7.1p1.
>
>Jens> After upgrading to egcs-1.1.2 I got all the files to compile ;-) I 
>did have
>Jens> problems with the optimisation on three files (but, unlike egc-1.1, 
>I could
>Jens> compile with little/no optimisation - Thank You, everyone who works 
>on EGCS
>
>Jens> http://www.imaginet.fr/~jensting/EGCS-err/index.html
>Jens> complete with (GZIP'ed) preprocessed files.
>Jens> First few lines of said page:
>
>Jens> egcs-1.1.2 stops with internal error--unrecognizable insn when being 
>asked
>Jens> to optimise with -O3 but works fine with little or
>Jens> no optimisation.
>
>        I don't know for sure, but I think that part of the problem is
>that egcs-1.1 uses "register_operand" predicate for a number of GOT
>patterns in the machine description which should be "gpc_reg_operand".
>This change already had been applied to the EGCS development sources a
>while ago, but was not change in the egcs-1.1 updates.  It will be
>included in the forthcoming egcs-1.2 release.
>
>        Maybe someone who has built a recent version of the development
>sources for powerpc-linux can try to recompile your examples to see if
>they still fail.

I can compile all 3 examples fine with -O3 and -O3 -fpic (-fpic is probably 
the option which makes the difference here) and egcs-1.1.2 on a 
glibc-2.1.1pre system. The egcs-1.1.2 in use contains the usual reload hack 
from Geoff Keating and some minor patches to fix obscure loop optimization 
bugs. AFAIK the R4 RPM's from Gary Thomas contain the reload hack as well, 
which I would consider the only patch I know of that would fix some 
"unrecognizeable insn" problems with -fpic.

Franz.