Avoid StringUTF16View dispatch overhead for some bridged String methods #83529
Conversation
|
@swift-ci please test |
|
@swift-ci please Apple Silicon benchmark |
1 similar comment
|
@swift-ci please Apple Silicon benchmark |
|
That's very weird. The benchmark run consistently dies at -Onone only, the optimized ones pass. |
|
|
||
| @inline(__always) | ||
| final internal var utf16: String.UTF16View { | ||
| String.UTF16View(_StringGuts(self)) |
There was a problem hiding this comment.
Theoretically this should all inline out, but this skips going through the String initializer
| _uncheckedBounds: (aRange.location, aRange.location+aRange.length)) | ||
| let str = asString | ||
| unsafe str._copyUTF16CodeUnits( | ||
| unsafe utf16._nativeCopy( |
There was a problem hiding this comment.
Part of the theme of "move all the smarts into StringUTF16View.swift instead of having them scattered in 3-4 different files"
| "String index is out of bounds") | ||
| return unsafe UInt16((start + offset).pointee) | ||
| } else { | ||
| return utf16[nativeNonASCIIOffset: offset] |
There was a problem hiding this comment.
Part of the theme of "move all the smarts into StringUTF16View.swift instead of having them scattered in 3-4 different files"
| return _foreignSubscript(position: idx) | ||
| } | ||
|
|
||
| internal subscript(nativeNonASCIIOffset offset: Int) -> UTF16.CodeUnit { |
There was a problem hiding this comment.
This is pretty much just pulling the relevant bits out of the code it used to go through
| _precondition(alignedRange.lowerBound._encodedOffset < _guts.count && | ||
| alignedRange.upperBound._encodedOffset < _guts.count, | ||
| "String index is out of bounds") | ||
| unsafe _nativeCopy( |
There was a problem hiding this comment.
We should do a fast path in this for blocks of ascii (already have one for the entire thing being known-ascii), but I was having trouble getting it to generate reasonable looking code
| } | ||
|
|
||
| @inlinable | ||
| @inlinable @inline(__always) |
There was a problem hiding this comment.
It was actually generating a retain-release around this before?? ~18% overhead in my simple test
| unsafe d.initialize(from: s, count: n) | ||
| return (unsafe Iterator(_position: s + n, _end: s + count), n) | ||
| } | ||
|
|
There was a problem hiding this comment.
All the remaining changes are kinda unrelated, but I was getting frustrated with trivial pointer operations generating calls
|
Local -characterAtIndex: benchmark gives ASCII: ~2s -> ~0.63s |
|
@swift-ci please test |
|
@swift-ci please Apple Silicon benchmark |
|
For a shorter (3 character) non-ASCII string we go from ~2s to ~1.25s, which indicates that most of the win is removing constant overhead, as expected. |
|
lol for ASCII we're now very slightly faster than NSCFString |
|
@swift-ci please test |
|
@swift-ci please Apple Silicon benchmark |
|
@swift-ci please benchmark |
|
I should look into what's up with StringWalk, but overall this is looking pretty decent |
|
The Linux failure is funny. The expected error is moved to compile time so the test fails to build |
|
@swift-ci please test |
|
@swift-ci please test |
|
@swift-ci please Apple Silicon benchmark |
| let emptyAllocated = UnsafeMutablePointer<Float>.allocate(capacity: 0) | ||
| defer { emptyAllocated.deallocate() } | ||
|
|
||
| // Make sure we can't emit the error at compile time |
There was a problem hiding this comment.
I wonder… does converting a runtime trap into an error like this count as a source break?
There was a problem hiding this comment.
We have converted runtime traps to compilation errors before, without considering them source breaks. What it amounts to is speeding up the discovery of an incorrect program to compilation time.
|
StringWalk appears to be extremely sensitive to minor inlining differences. I've filed a followup bug for the optimizer folks about a missed optimization that occurs both before and after, and I'm hoping will more than recover the difference. |
2 participants
This removes a bunch of overhead on the UTF16 paths in String, as well as consolidating the complicated bits of the logic in one file.
Fixes rdar://157500258