questions about COND_EXEC and SMS

[Steven Bosscher]

On Thu, Dec 5, 2013 at 12:11 PM, dxq wrote:
> hi all,
>
> *We found that COND_EXEC is better than IF_THEN_ELSE when used as expressing
> condition move insns, because in sched, IF_THEN_ELSE insn has a dependence
> on itself, and COND_EXEC has not.
> * Besides,  IF_THEN_ELSE is not good for SMS. some backend (frv) expands
> condition move as IF_THEN_ELSE pattern before reload and splits IF_THEN_ELSE
> into COND_EXEC at split2 pass, which is a post reolad pass. However, in SMS
> pass(pre-reload), we can't get the accurate information of IF_THEN_ELSE
> insns.
> * However, as far as i know, COND_EXEC is not supporting in pre-reload
> passes.
>
> So, I'm asking for some suggestions for supporiting COND_EXEC in pre-reload
> passes, for me, maybe from split1 to reload pass is good enough. what work
> need to do for supporting that, and is there any one who has done any work
> on that?

Supporting COND_EXEC before register allocation (RA) is in itself not
a big challenge. The problem is handling them *during* register
allocation: How to do spilling/reloading registers used in a
predicated instruction, and how to spill/reload the predicate
condition (condition code or BImode register).

If I would have to add support for this myself, I'd do it as follows:

1. Make your port work with LRA instead of reload.

2. Make pre-RA passes handle COND_EXEC with the predicate
(COND_EXEC_COND) explicitly assigned to a hard register and only
predicated instructions (COND_EXEC_EXPR) that don't need reloads (e.g.
only reg->reg and imm->reg moves).

3. Try to get the simple test cases from (1) to get through IRA/LRA. I
wouldn't try doing it with reload. With LRA it's likely simpler to add
support for this because you can spill to pseudo-registers.

4. Gradually relax the constraints on the kind of COND_EXEC
instructions you accept before reload, and play whack-a-mole on all
bugs that ubdoubtedly will pop up.

Supporting COND_EXEC before reload has been discussed before on this
mailing list, but I don't think anyone has every started a serious
effort to implement it. There are so few targets that use COND_EXPR
and the ones that do (ia64) are falling out of favor. I expect (or at
least: hope) that LRA makes it a lot easier to implement it.

Ciao!
Steven