withUTF8 currently vends a typed UInt8 pointer to the underlying
SmallString. That pointer type differs from SmallString's
representation. It should simply vend a raw pointer, which would be
both type safe and convenient for UTF8 data. However, since this
method is already @inlinable, I added calls to bindMemory to prevent
the optimizer from reasoning about access to the typed pointer that we
vend.
rdar://67983613 (Undefinied behavior in SmallString.withUTF8 is miscompiled)
Additional commentary:
SmallString creates a situation where there are two types, the
in-memory type, (UInt64, UInt64), vs. the element type,
UInt8. `UnsafePointer<T>` specifies the in-memory type of the pointee,
because that's how C works. If you want to specify an element type,
not the in-memory type, then you need to use something other than
UnsafePointer to view the memory. A trivial `BufferView<UInt8>` would
be fine, although, frankly, I think UnsafeRawPointer is a perfectly
good type on its own for UTF8 bytes.
Unfortunately, a lot of the UTF8 helper code is ABI-exposed, so to
work around this, we need to insert calls to bindMemory at strategic
points to avoid undefined behavior. This is high-risk and can
negatively affect performance. So far, I was able to resolve the
regressions in our microbenchmarks just by tweaking the inliner.
* Don't allocate breadrumbs pointer if under threshold
* Increase breadrumbs threshold
* Linear 16-byte bucketing until 128 bytes, malloc_size after
* Allow cap less than _SmallString.capacity (bridging non-ASCII)
This change decreases the amount of heap usage for moderate-length
strings (< 64 UTF-8 code units in length) and increases the amount of
spare code unit capacity available (less growth needed).
Average improvements for moderate-length strings:
* 64-bit: on average, 8 bytes saved and 4 bytes of extra capacity
* 32-bit: on average, 4 bytes saved and 6 bytes of extra capacity
Additionally, on 32-bit, large-length strings also gain an average of
6 bytes of extra spare capacity.
Details:
On 64-bit, half of moderate-length allocations will save 16 bytes
while the other half get an extra 8 bytes of spare capacity.
On 32-bit, a quarter of moderate-length allocations will save 16
bytes, and the rest get an extra 4 bytes of spare
capacity. Additionally, 32-bit string's storage class now claims its
full allocation, which is its birthright. Prior to this change, we'd
have on average 1.5 bytes of spare capacity, and now we have 7.5 bytes
of spare capacity.
Breadcrumbs threshold is increased from the super-conservative 32 to
the pretty-conservative 64. Some speed improvements are incorporated
in this change, but more are in flight. Even without those eventual
improvements, this is a worthwhile change (ASCII is still fast-pathed
and irrelevant to breadcrumbing).
For a complex real-world workload, this amounts to around a 5%
improvement to transient heap usage due to all strings and a 4%
improvement to peak heap usage due to all strings. For moderate-length
strings specifically, this gives around 11% improvement to both.
Since scalar-alignment is set in inlinable code, switch the alignment
bit to one of the previously-reserved bits rather than a grapheme
cache bit. Setting a grapheme cache bit in inlinable would break
backward deployment, as older versions would interpret it as a cached
value.
Also adjust the name to "scalar-aligned", which is clearer, and
removed assertion (which should be a real precondition).
Fixes a general category (pun intended) of scalar-alignment bugs
surrounding exchanging non-scalar-aligned indices between views and
for slicing.
SE-0180 unifies the Index type of String and all its views and allows
non-scalar-aligned indices to be used across views. In order to
guarantee behavior, we often have to check and perform scalar
alignment. To speed up these checks, we allocate a bit denoting
known-to-be-aligned, so that the alignment check can skip the
load. The below shows what views need to check for alignment before
they can operate, and whether the indices they produce are aligned.
┌───────────────╥────────────────────┬──────────────────────────┐
│ View ║ Requires Alignment │ Produces Aligned Indices │
╞═══════════════╬════════════════════╪══════════════════════════╡
│ Native UTF8 ║ no │ no │
├───────────────╫────────────────────┼──────────────────────────┤
│ Native UTF16 ║ yes │ no │
╞═══════════════╬════════════════════╪══════════════════════════╡
│ Foreign UTF8 ║ yes │ no │
├───────────────╫────────────────────┼──────────────────────────┤
│ Foreign UTF16 ║ no │ no │
╞═══════════════╬════════════════════╪══════════════════════════╡
│ UnicodeScalar ║ yes │ yes │
├───────────────╫────────────────────┼──────────────────────────┤
│ Character ║ yes │ yes │
└───────────────╨────────────────────┴──────────────────────────┘
The "requires alignment" applies to any operation taking a
String.Index that's not defined entirely in terms of other operations
taking a String.Index. These include:
* index(after:)
* index(before:)
* subscript
* distance(from:to:) (since `to` is compared against directly)
* UTF16View._nativeGetOffset(for:)
String.Index has an encodedOffset-based initializer and computed
property that exists for serialization purposes. It was documented as
UTF-16 in the SE proposal introducing it, which was String's
underlying encoding at the time, but the dream of String even then was
to abstract away whatever encoding happend to be used.
Serialization needs an explicit encoding for serialized indices to
make sense: the offsets need to align with the view. With String
utilizing UTF-8 encoding for native contents in Swift 5, serialization
isn't necessarily the most efficient in UTF-16.
Furthermore, the majority of usage of encodedOffset in the wild is
buggy and operates under the assumption that a UTF-16 code unit was a
Swift Character, which isn't even valid if the String is known to be
all-ASCII (because CR-LF).
This change introduces a pair of semantics-preserving alternatives to
encodedOffset that explicitly call out the UTF-16 assumption. These
serve as a gentle off-ramp for current mis-uses of encodedOffset.
Old Swift and new Swift runtimes and overlays need to coexist in the same process. This means there must not be any classes which have the same ObjC runtime name in old and new, because the ObjC runtime doesn't like name collisions.
When possible without breaking source compatibility, classes were renamed in Swift, which results in a different ObjC name.
Public classes were renamed only on the ObjC side using the @_objcRuntimeName attribute.
This is similar to the work done in pull request #19295. That only renamed @objc classes. This renames all of the others, since even pure Swift classes still get an ObjC name.
rdar://problem/46646438
In anticipation of potential future HW features, e.g. armv8.5 memory
tagging, only use the high 4 bytes as discriminator bits in
_BridgeObject rather than the top 8 bits. Utilize two perf flags to
cover this instead. This requires shifting around a fair amount of
internal complexity.
Include some tuning and tweaking to reduce the constant factors
involved in string comparison. This yields considerable improvement on
our micro-benchmarks, and allows us to make less inlinable code and
have a smaller ABI surface area.
Adds more extensive testing of corner cases in our existing
fast-paths.
After rebasing on master and incorporating more 32-bit support,
perform a bunch of cleanup, documentation updates, comments, move code
back to String declaration, etc.
Refactor and rename _StringGutsSlice, apply NFC-aware fast paths to a
new buffered iterator.
Also, fix bug in _typeName which used to assume ASCIIness and better
SIL optimizations on StringObject.
Add inlinability annotations to restore performance parity with 4.2 String.
Take advantage of known NFC as a fast-path for comparison, and
overhaul comparison dispatch.
RRC improvements and optmizations.
* Refactor out RRC implementation into dedicated file.
* Change our `_invariantCheck` pattern to generate efficient code in
asserts builds and make the optimizer job's easier.
* Drop a few Bidi shims we no longer need.
* Restore View decls to String, workaround no longer needed
* Cleaner unicode helper facilities
This is a giant squashing of a lot of individual changes prototyping a
switch of String in Swift 5 to be natively encoded as UTF-8. It
includes what's necessary for a functional prototype, dropping some
history, but still leaves plenty of history available for future
commits.
My apologies to anyone trying to do code archeology between this
commit and the one prior. This was the lesser of evils.
Exclusively store small strings in little-endian byte order. This
will insert byte swaps when accessing small strings on big endian
platforms, however these are usually extremely cheap.
This approach means that the layout of the code points and count
in memory will be the same on both big and little endian machines
simplifying future development. Prior to this change this code
was broken on big endian machines because the memory layout was
different (the count ending up in the middle of the string).
_StringGuts shouldn't expose a subscript, implying efficient
access. Switch to the explicit code unit fetch method. Update tests
accordingly, and switch off of deprecated typealiases.
Add an isSmall query to Character so testing doesn't have to bake in
internal format. Clarify the purpose of the invalid UTF-16 backdoor
creation method.
Create a _StringRepresentation struct to standardize internal testing
on. Internalize much of _StringGuts, except for some SPI hacks, and
update tests to use _StringRepresentation.
@effects is too low a level, and not meant for general usage outside
the standard library. Therefore it deserves to be underscored like
other such attributes.
_StringGuts is not meant to be an abstraction across all the forms a
String may take. It's meant to abstract the book-keeping and the
visitor is a parameterization over operations.
Extract slow paths into non-inlinable functions so that fast-paths can
be faster and we don't pay the large code bloat for the Unicode
parsers.
Some tests proactively extended to highlight UTF8View of multiple
kinds of Strings.
Drop append-related @inlinable annotations for String, StringGuts,
StringStorage, and the Views. Drop several for larger operations, such
as case conversion. Drop as many as we can from StringGuts for now.
Beside the general goal to remove inlinable functions, this reduces code size and also improves performance for several benchmarks.
The performance problem was that by inlining top-level String API functions into client code (like String.count) it ended up calling non-inlinable internal String functions eventually.
This is much slower than to make a single call at the top-level API boundary into the library. Inside the library all the internal String functions can be specialized and inlined.
rdar://problem/39921548
StringStorage.create is the primary means of allocating storage for a
string, so drop inlinability to allow for future evolution.
StringStorage also exposes some .appendInPlace methods, which we
currently need to keep inlinable for benchmark performance. We'd
really like to drop inlinability for these for evolution purposes
(e.g. imagine a future version that adjusts nul-termination or changes
in coordination with create). These are flagged with:
` // TODO(inlinability): @usableFromInline - P3`
Where "P3" reflects urgency on a scale from 1 (stop the presses) to 5
(whatevs).
