Re: PATCH COMMITTED: Use conditional moves in groups of assignments

[Joern RENNECKE <joern dot rennecke at st dot com>]

Ian Lance Taylor wrote:

> My patch handles code which is different in the then/else branches, as
> long as everything is an assignment to a pseudo-register.
My code would handle the RHSs of these assignments as inputs.

>  In
> particular, the then/else branches can set different registers. 
That is something not covered by my patch.  It would require extra data
structures to describe where such sets are.

> I
> initially only handled the case of constant or register inputs, but
> it's relatively straightforward to extend to general operands, which
> need not be the same in the then/else branches, with more careful
> checking of which code affects which other code.
>  
My code could easily be adopted to allow a wider range of possible 
inputs for
if-conversion.  There would have to be a flag in the struct equiv_info 
to tell if
such more general inputs are allowed, since this cannot easily be 
extended to
cross-jumping - we'd first have to find registers where to put the 
alternate inputs.
(This would already be true for constant inputs.)

However, even for if-conversion, I wonder what kind of inputs you could 
safely
yet profitably use - in general, memory operands will be unsuitable 
because they
can trap.  Or would this optimization be restricted to targets that can 
use such
sources directly in a conditional move such that no trap occurs if the 
move is
not actually done?