> On Fri, May 27, 2011 at 10:54 AM, Jan Hubicka <hubi...@ucw.cz> wrote: > > Hi, > > here is updated patch incorporating the feedback. I now use variant of > > uleb format, > > just do 4 bit chunks instead of 8bit. It works better in my tests and is > > also used > > by LLVM so it must be cool ;) > > > > I also fixed off by one error in filename streaming. While debugging this I > > noticed that bitpack will have random effect when the value passed does not > > fit > > in range (i.e. the bits will end up in next packed values). Adding assert > > on this > > I found one place where we try to pack negative value in it. Fixed by > > using my > > new functions, but wonder what the logic of code is, given that we always > > pack 0 or -1 as alias set. > > Yeah, I think diego noticed this as well. > > Micha has a point in that everything should be a single bitpack which > suitable points of re-alignment (to also make optimizing the packing > possible). But that's more work and can be done on top of this.
Yep, I tought of this a bit. It will be fun to implement it in a way that resulting code will be good (i.e. optimize well across the byte alignment points with word sized bitpack output buffer) Honza