Add count-trailing-zero libcalls #597
Conversation
calc84maniac
force-pushed
the
cttz-libcalls
branch
from
2eae7db to
a61a932
Compare
calc84maniac
force-pushed
the
cttz-libcalls
branch
from
a61a932 to
98bec4b
Compare
| tst a, 030h | ||
| jr z, .high2 | ||
| dec a | ||
| and a, 014h | ||
| ret po | ||
| ld a, 5 | ||
| ret | ||
| .high2: | ||
| add a, a | ||
| sbc a, -8 | ||
| ret p | ||
| ld a, 6 | ||
| ret |
There was a problem hiding this comment.
This unfortunately breaks the optimization of the first trailing one functions relying on this to produce an output of 8 for an input of 0, but it might be worth it?
| tst a, 030h | |
| jr z, .high2 | |
| dec a | |
| and a, 014h | |
| ret po | |
| ld a, 5 | |
| ret | |
| .high2: | |
| add a, a | |
| sbc a, -8 | |
| ret p | |
| ld a, 6 | |
| ret | |
| add a, a | |
| add a, a | |
| jr z, .high2 | |
| add a, a | |
| add a, a | |
| sbc a, -5 | |
| ret | |
| .high2: | |
| sbc a, -7 | |
| ret |
There was a problem hiding this comment.
I guess this change would also hurt for maybe eventually implementing std::countr_zero, wouldn't it...
There was a problem hiding this comment.
Unfortunately, that assumption is also made by the compiler (this libcall implements CTTZ which is output by __builtin_ctz intrinsics on the Z80 target and not CTTZ_ZERO_UNDEF). Good optimization idea though, I'll have to think whether I can do something similar.
There was a problem hiding this comment.
I guess this change would also hurt for maybe eventually implementing
std::countr_zero, wouldn't it...
It's actually already been implemented in the toolchain recently, and makes use of the Z80 intrinsic behavior I mentioned.
There was a problem hiding this comment.
It occurs to me that, except for that pesky 0 case, you could potentially rework this whole routine to be based on branches after pairs of left shifts. I think whether that would be faster or not depends on the input distribution, but it would certainly be smaller.
| ld a, l | ||
| or a, a |
There was a problem hiding this comment.
Matching __icttz's first 3 bytes might compress better.
| ld a, l | |
| or a, a | |
| xor a, a | |
| or a, l |
There was a problem hiding this comment.
Good point, I always forget to think about compression alongside other size optimizations.
I've implemented optimized count-trailing-zero libcalls, and I have an ez80-clang commit ready to push to make it use them. They've also been added to ez80_builtin implementations where applicable, and tested via ez80_builtin/stdbit tests.