Re: ifcvt/crossjump patch: Fix PR 42496, 21803

[Jeff Law <law at redhat dot com>]

 On 07/27/10 16:38, Eric Botcazou wrote:
> > Well, I don't know what if anything Eric has in mind, but assuming we have
> >
> > BB1
> > lots of stuff
> > if (x) goto A;
> > BB2
> > y = 1;
> > goto C;
> > BB3
> > A: y = 1;
> > goto D;
> >
> > how can we possibly avoid code movement?
> Split BB2 and BB3 after "y = 1;" and redirect the edges from BB1.  Then split
> BB1 before the test and insert one instance of the common heads.
Which to me seems more convoluted than just moving insns implementing 
the common code.  And I think that's the whole point behind the 
disagreement.  While we *can* formulate this as a series of CFG 
manipulations I think it actually makes the resulting transformation 
more difficult to understand.

Also note that trying to pluck common insns that aren't at the head of 
the blocks is more difficult to do as a pure CFG transformation (it's 
doable, just ugly) while it ought to be trivial to just move them around.


Jeff