[Bug bootstrap/49815] [4.7 regression] ICE in cselib_record_set, at cselib.c:2241 compiling 64-bit libjava on SPARC

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49815

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-07-22 14:46:49 UTC ---
(In reply to comment #9)
> > BTW, wouldn't it be better to model the insn completely rather than partially?
> > In particular, in addition to the arguments being copied have also the
> > copy/clobber for all the other arguments, except for the sp = sp - something
> > which would be another set.
> 
> What other arguments?  Does vt_add_function_parameters not see them all?  I'm

I meant registers.  Especially with the planned shrink wrapping, say if code
before the prologue stores some argument into %o5 register that doesn't hold
any parameter and then prologue does perform a save, that value is now live in
%i5 rather than %o5 and var-tracking should handle that transparently.

> not sure I understand the need to model the stack decrement either given that
> the CFA register is the frame pointer.

Var-tracking should be told that this window save instruction initializes the
hard frame pointer (%i6 <= %o6) and that %o6 has been decremented, without that
it is possible it will give wrong answers where values live in.