The "poor" performance of the Caltech C compiler.

[Eugene D. Brooks III]

With regards to the performance of the Caltech C compiler with single
precision floating point operations.

This compiler implemented both single and double precision register
variables and delivered quite a substantial speed improvement over
the standard Unix compiler.  I recorded speed increases of up to a
factor of 2.5 for some vector operations implemented as unrolled loops
in C.

	register float *a, *b, *c;

	int dim;

	dim /= 8;

	do {
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
		*a++ = *b++ + *c++;
	while(dim-- > 0);

With the apropo handling of the dim%8 part, went like the devil.  Other
operations, dot product, etc obtained similar speed improvements.  Inspection
of the resulting assembler code showed that one could not do better by
writing in assembler.