🚀 go-pugleaf

jbuck@galileo.berkeley.edu (Joe Buck) writes:

>Texas Instruments supplies a C compiler; I used it extensively.
>As of version 2.10, the code quality was very poor.  Because of
>the close resemblance of the register sets and addressing modes
>of the TMS320C30 family to the Motorola 68000 family, TI could
>have done far better by starting with a good 68000 compiler and
>porting it, or with any other C compiler, since whoever wrote
>it made no attempt at any kind of optimization at all.

>It is possible to use the compiler for some purposes by heavy
>use of register variable declarations and inline assembly code.
>But many of the excuses given for poor DSP C compilers simply
>don't apply in the case of the C30; it is close enough to a
>general purpose CPU that a good optimizing C compiler could be
>written for it.

>If TI has improved the compiler I'd like to hear about it;
>however, version 2.10 is nothing to be proud of.

Agreed.  I know next to nothing about compiler optimizations, but I do know
that the code the C compiler put out was lousy.  It tends to use much fewer
registers than it has available (used ~3-4 of the Auxiliary Regs when it had
8 available on the C30, extended floating-point register usage was similar). 
Instead, it kept putting results on a local stack it maintained on top of
the current stack at (subroutine) entry.

Also, memory usage on the simulator is limited to about 1 64K segment.  This
is really annoying if you want to try to crunch on a few decent size images
simultaneously.

There's much more, but my brain is in a fog today.


------------------------------------------------------------------------------
Rick Fanta 	PAR Technologies (richf@adiron.uucp)

"... She said that she was working for the ABC news,
     it was as much of the alphabet as she knew how to use ..."
	- Elvis Costello