🚀 go-pugleaf

    I did an application that decompressed G4 images and then scaled
    the images for a for a 300 page per minute printer . The C
    compiler was GNU and the risc was the Intel 80960CA. The code had
    beed optimized a lot in C and was still not fast enough.

300 pages per minute?  5 pages per second???  Yow!

    My usual experience with C compilers on DSP's that a 8 to 10 x
    increase in speed is possible in the c to asm port. After looking
    at the output of the GNU 80960 C compiler, I was impressed with
    the efficiency of the assembler output. I estimated that the
    increase in speed in the c to asm port would be a factor of 2 at
    best. I was able to attain a 3x speed increase on the G4
    decompression code and A 8X IMPROVEMENT on the scaling code.

    This speed increase was due to 1) using a machine instruction the
    complier was clueless about, and 2) an optimization of the scaling
    algorithmn that would have been not obvious at the compiler level.

You should report the missing machine instruction to the maintainers
of your compiler (I'm not sure which of the several available 960
GCC ports you were using), or simply extend the compiler yourself
to know about the extra instruction.  Having the compiler sources
is great.