I think I'll wait for Aztec C, hopefully it will be a bit better at
getting things right.
It is supposed to generate better code than Lattice too. I know
from a piece of one of my routines, the code was supposed to assign
one structure to another, it so happens that the structure is 8 bytes
long. On my SUN at work, its C compiler generates two move.l
instructions: move.l -8(a6),-16(a6)
move.l -12(a6),-20(a6) That take 56 clock cycles.
Lattice generates
lea -8(a6),a0
lea -16(a6),a1
moveq.l #7,d0
loop: move.b (a0)+,(a1)+
dbf d0,loop
which means 18 instructions executed at 186 clock cycles. Not very
good performance.