Am 14.07.2013 14:11, schrieb bearophile:
Benjamin Thaut:
http://d.puremagic.com/issues/show_bug.cgi?id=10226
I see there are codegen inefficiencies.
Until dmd uses the XMM registers correctly it doesn't make much sense
to add a special calling convetion for this purpose.
I don't agree, because:
- Even if DMD codegen is far from not perfect, it's a good idea to
improve all things in parallel. Generally improving things (or fixing
bugs) gives a better result if you adopt a pipelined development approach.
- A vector calling convention is meant to be usable on other compilers
too, like LDC2, that have better codegen. (The asm I have shown in this
thread comes from LDC2 because dmd doesn't even use SIMD registers on
Windows 32 bit).
Bye,
bearophile
I just wanted to say that there are currently bigger fish to fry then
micro optimization through calling convetions. (GC, allocators, all the
bugs...)
Did you compile the shown code with optimization enabled or is that a
debug build? If it is optimized I'M going to be disappointed by LDCs
codegen.
Kind Regards
Benjamin Thaut
