Vectorize UTF16 offset calculations #41866
Conversation
|
@swift-ci please smoke test |
|
@swift-ci Please test Windows platform |
|
@swift-ci please benchmark |
| let fourBytes = U.zero.replacing(with: U.one, where: uValue .>= 0b11110000) | ||
| let fourByteCount = Int(fourBytes.wrappedSum()) | ||
|
|
||
| utf16Count &+= (U.scalarCount - continuationCount) + fourByteCount |
There was a problem hiding this comment.
Horizontal operations (wrappedSum) are expensive; usually in SIMD code we keep the accumulation in vector until forced to map down to a scalar rather, than accumulating every loop iteration. It's a little bit of a pain with Int8 because you can only do 127 accumulations before you have to worry about overflow, but still doing one horizontal operation every N iterations is much better than 2/iteration.
The easy change here would be to do:
&+= U.scalarCount + (fourByteCount &- continuationCount).wrappedSum()
so that you only do half as many as currently.
There was a problem hiding this comment.
Sure, I'll make a followup patch
There was a problem hiding this comment.
(The optimizer might be managing to pull this apart for you already, but it would be nice to make explicit anyway.)
There was a problem hiding this comment.
(I'm hoping once wider vectors are feasible that the common case here, post-breadcrumb offsets, will only ever iterate once because we can just do the entire 64 bytes at once)
Cleaned up PR and commit history for #41684