ByteBuffer: small & medium sized string copies over 10% faster #640
Conversation
|
spurious test error |
|
@swift-nio-bot test this please |
Sources/NIO/ByteBuffer-core.swift
Outdated
| let newBytesPtr = UnsafeMutableRawBufferPointer(rebasing: self._slicedStorageBuffer | ||
| .dropFirst(Int(index)) | ||
| .prefix(capacity)) | ||
| let newBytesPtr = UnsafeMutableRawBufferPointer(rebasing: self._slicedStorageBuffer[Int(index) ..< Int(index) &+ Int(capacity)]) |
There was a problem hiding this comment.
Why the unchecked addition? None of the others perform it.
There was a problem hiding this comment.
@Lukasa just to double check I understand you are talking about &+ (no overflow check) ?
There was a problem hiding this comment.
I think I am generally fine with doing it here but if we do so we should also do it in other places where we are sure overflow can never happen to be consist. So maybe we should just not do it in this PR and then open another PR with only the "unchecked" stuff.
There was a problem hiding this comment.
I think this is mostly my point: if we believe the unchecked math here is safe we should a) do it more consistently, and b) put comments explaining why we think that, as we've done for force-unwrapping.
There was a problem hiding this comment.
benchmarks don't show that the &+ make a diff here, removing
Motivation: urbp[start ..< end] is quite a bit faster than when done in two operations urbp.dropFirst(start).prefix(length). This improves small (4 bytes) & medium sized (24 bytes) String copies by over 10%. Good for HTTP headers etc. Modifications: replace urbp.dropFirst(start).prefix(length) by urbp[start ..< start + length] Result: small & medium sized String copies faster
weissi
force-pushed
the
jw-faster-string-blit
branch
from
608dcc1 to
8c8a65a
Compare
Motivation:
urbp[start ..< end] is quite a bit faster than when done in two
operations urbp.dropFirst(start).prefix(length). This improves small (4
bytes) & medium sized (24 bytes) String copies by over 10%. Good for
HTTP headers etc.
Modifications:
replace urbp.dropFirst(start).prefix(length) by urbp[start ..< start +
length]
Result:
small & medium sized String copies faster