Re: [PATCH v3 1/4] ifcvt: handle sequences that clobber flags in noce_convert_multiple_sets

[Jeff Law <jeffreyalaw@gmail.com>]

On 10/20/23 01:04, Robin Dapp wrote:
> 
>> But I'm not sure which cases this code is trying to catch.  Is it trying
>> to catch cases where seq2 "spontaneously" uses registers that happen to
>> overlap with cond?  If so, then when does that happen?  And if it does
>> happen, wouldn't the sequence also have to set the registers first?
> 
> In order for sequence costing to be superior to just counting "conditional"
> instructions we need to make sure that as few redundant instructions as
> possible are present in the costed sequences. (redundant as in "will be
> removed in a subsequent pass").
[ ... ]
Sounds a lot like a scenario we had with threading.  Threading will 
often generate code in duplicated blocks that will trivially be eliminated.

IIRC I had Alex O. tackle that in threading several years back with good 
success.  I don't think he tried to be exhaustive, just sensible in what 
was likely to be dead after threading.  It helped numerous cases where 
we clearly should have threaded, but weren't because of artificially 
high costing.

It's not a trivial balancing act.

Jeff