Re: Reload patch to improve 386 code

[meissner@cygnus.com <meissner@cygnus.com>]

| >  I have come up with a patch for the reload pass that can
| >  improve the generated code on the i386 quite dramatically.
| >  The problem with reload is that when it decides to spill
| >  a hard register, it spills _all_ pseudos allocated to
| >  this hard register.  This is often not necessary, it is
| >  possible to use register life information to determine
| >  that certain pseudos are not live across any instructions
| >  that need reloads, and thus need not be spilled.
| 
| Isn't this something that would have been solved [ coniunctivus  
| irrealis ! ] by live range splitting ?
| 
| "pseudo's that are not live across instructions" sounds to me as  
| the definition of the place to split the liveness range.
| 
| Or am I missing something ?

My opinion is reload needs to be rewritten completely.  I'm in the middle of
working on a beginning of live range splitting right now, but it probably will
be some time before it moves to egcs (ie, its not ready yet).

| OTOH, "beter een vogel in de hand dan tien in de lucht" (sorry,  
| couldn't resist).
| 
| Cheers,
| Toon.
|