bearophile wrote:
Andrei:
Cristi's GSoC project may target linear algebra operations.
That benchmark code is not a realistic example because don't write matrix
multiplications manually like that. But the benchmark has found a weakness in
array management that's important anyway.
Bye,
bearophile
It's very well known. DMD's floating point code is poor because it NEVER
stores floating-point loop variables in registers. Ever.
Disassembling almost any piece of DMD floating point code will show
that. The floating-point backend is about the simplest you could
imagine. (By contrast, the integer backend is extremely sophisticated).