Walter Bright wrote:
Don wrote:
Having looked at the DMD optimiser, I'm a bit surprised that it's
competitive at all (especially in floating point). There is so much
low-hanging fruit, it's practically an orchard <g>.
I believe that's because it has reached the point of diminishing returns.
I presume you're talking about integer optimisation, because that's
definitely not the case for DMD's floating point. Eg, there's a comment
in the code somewhere saying that CSE (complex sub expression)
optimisations are not done for floating point because the code generator
can't cope with it. It does loops particularly poorly, it never seems to
keep a variable on the FP stack, so it keeps loading an saving them. The
potential speedups are enormous.
