Re: [rfc] Make regrename call recog

[Bernd Schmidt <bernds_cb1 at t-online dot de>]

> working on a S/390 patch optimizing load/store multiple I've
> noticed that regrename replaces registers without calling
> recog afterwards. So it happened that a single register out
> of a PARALLEL block of SETs gets renamed. Because the
> load_multiple patterns only accept SETs with consecutive
> registers this insn is invalid after regrename.
>
> The attached patch turns "do_replace" in "try_replace" using
> validate_change to replace the register along the def-use chain.

IIRC the ARM used to have a similar problem.

> Is this the way to go?

IMO, no.  The renamer is careful to modify insns only in such a way that 
the constraints are still met.  If that isn't the case on S/390, it 
means that the constraint strings are incorrect.  In the ARM case, I 
think it happened for an insn that only ever got generated with hard 
regs in place and thus never needed any reloading, which is why this 
wasn't detected before.  Is that true for your pattern as well?


Bernd