Re: RFC -- update_equiv_regs and friends

["John David Anglin" <dave at hiauly1 dot hia dot nrc dot ca>]

> Not surprisingly I've been thinking more about this code in the last couple
> days, and I can't convince myself the code to deal with dying registers
> in notes for local-alloc is correct either.
> 
> While I believe it handles the case where the death note is on the equiv
> insn, I don't believe it handles the case where the death note is on a
> later insn.  ie consider something like this
> 
> insn A where we have an equivalence referencing register X
> ...
> insn B where register X dies
> ...
> insn C where we use the result of insn A.
> 
> If we substitute the equivalence from A into C, then we lose because we've
> changed the lifetime of register X.

In looking at the problems arising from the substitution og LO_SUM MEMs
generated by the PA port for pic references, I had come to the same
conclusion.

The function rtx_varies_p used in update_equiv_regs only determines
whether a rtx varies at the location where the rtx is used in the
code.  There is no guarantee that an equivalence determined at A
will still be valid at C.  I previously submitted a patch to change
at least some of the code in update_equiv_regs to use rtx_unstable_p
instead of rtx_varies_p.  This provides a proper test to determine
whether a equivalence valid at A is still valid at C.  See
<http://gcc.gnu.org/ml/gcc-patches/2001-02/msg01187.html>.

See also
<http://gcc.gnu.org/ml/gcc/2001-02/msg00769.html>.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)