Re: [Mesa3d-dev] codegen again

[Josh V. <ho...@na...>]

Allen Akin <ak...@po...> writes:

> Also, my personal bias would be to use an input representation that's
> more abstract than MIPS-like virtual-machine instructions.  That makes
> it easier to optimize the output code if you like, and needn't have a
> huge impact on compilation time if optimization is turned off.  I
> acknowledge this can become a tarpit, though.

I'm surprised they went for only 6 registers.  Most virtual machines
go the other way and give you a huge number of registers.  It seems
like it would be hard to get good performance with such a small number
of registers.  (Although it would be possible to create a pentium pro
style register renamer to at least get some use of the extra registers
available on non x86 processors.)

Even with more registers, a generic assembly language is still at a
large disadvantage for optimization just because of the amount of
information missing vs. a higher level language.  I've read (in the
GCC IA-64 summit minutes *) that that even gcc's RTL representation
throws away too much information for some optimizations, so they are
going to make some of their latest optimizations work on trees instead
of RTL.

Josh

* http://linuxia64.org/gcc_summit.2001.06.06.html