Read-only data handling: a proposal

[Martin Mares]

Hello,

> Hmm, wait.  I thought we already had code to keep/discard static constants
> (-fkeep-static-constants or some such)  Or are we discussing global constants?
> If so it seems to me -fkeep-unused-constants might be the right name for
> such an option.

   OK.

>   >    I was suggesting deferring it until the end of a function -- it's doesn't
>   > have the problems you've mentioned and it can produce better code on some
>   > architectures if you have a smart assembler which can convert absolute refs
>   > to PC-relative refs whenever possible.
> It certainly avoids those problems.  Of course you lose the opportunity to
> merge more constants.  But I guess something is better than nothing :-)

   Well, I suggested deferring the output, not freeing of the memory. The
values we need to defer are already copied to permanently allocated memory
and since the tree nodes I propose to add are not allocated automatically
by the parser, but in the varasm code instead, there seems to be no problem
in allocating them permanently, too.

> The PC-rel stuff is useful, but the number of architectures where such tricks
> help are diminishing. 

  E.g., MC680x0, but AFAIK gas is not capable of such optimization yet.

				Have a nice fortnight
-- 
Martin `MJ' Mares   <mj@gts.cz>   http://atrey.karlin.mff.cuni.cz/~mj/
Faculty of Math and Physics, Charles University, Prague, Czech Rep., Earth
"Foolproof operation:  All parameters are hard coded."