Re: PATCH COMMITTED: Permit splitters to create new pseudo regs in combine

[Ian Lance Taylor <iant@google.com>]

Adam Nemet <anemet@caviumnetworks.com> writes:

> Ian Lance Taylor <iant@google.com> writes:
> > I have committed this patch to fix the problem.  This simply changes
> > reg_stat to be a VEC, and grows the VEC when a splitter creates a new
> > pseudo.
> 
> I must be missing something.  Who updates the old-style dataflow
> information that combine uses for the new pseudos created?
> 
> For example in try_combine, don't we now call record_value_for_reg on
> the pseudo that was originally offered by combine instead of the one
> the backend picked?

Sorry, I missed this note earlier.

When we call record_value_for_reg in try_combine, it is resetting the
value recorded for the register to a value set in an earlier insn,
which is correct.

Nothing records the values for the new registers created by the
splitters.  That is OK, as no other insns refer to them.  It is
theoretically possible that this could cause a missed optimization if
the insns generated by a splitter can themselves be combined.  But
that is a rather unlikely scenario, as nobody would write a splitter
that way.

Ian