We no longer create intermediate NSString copies to compare and hash swift
Strings. Instead we call directly into the ICU library.
I measured a 1.2 to 2x improvement on dictionary benchmarks as a result of this.
The SuperChars benchmark is also about 1.2x faster because of this.
Pure ASCII comparison has gotten a little bit slower (20% on a pure comparison
micro-benchmark) because we no longer do a memcmp. Doing a memcmp on ASCII is
not the same as the default unicode collation. Instead we have to a string scan.
The default unicode collation does not order like ASCII does and ignores
characters (for example the \0 character).
rdar://18992510
Swift SVN r31474
cache instead.
Create two tables. One that caches the ASCII ordering as returned by the root
collator. This table is static and we can use this for comparison and ordering.
The second table is cached the first time it is queried. This table caches the
actual collation values. We use this table for hashing. Collation values might
change between versions of the ICU dylib and so we can't store this in a static
table.
Use the second table to create a unicode hash for ascii strings.
The first table will be used in a follow-up commit.
Swift SVN r31472
Some characters are not considered when comparing strings. Hashing in the length
would cause a difference of hash values in equal strings.
Swift SVN r31468
Thanks to this, we can get rid of these two dirty symbols:
19f0f24e0 __ZZN17MetadataAllocator5allocEmE12pagesizeMask 8
19f0f24e8 __ZGVZN17MetadataAllocator5allocEmE12pagesizeMask 8
This is part of the effort to put the stdlib into a shared cache.
rdar://22375554
Swift SVN r31388
The reason for this patch:
- Currently, when we invoke swift_conformsToProtocol for the first time, we load ALL protocol conformances into conformances cache (to avoid rescanning the conformances section next time), which consumes about 50-70KB of memory even per minimal Swift application (e.g. hello-world app). While it may seem not so much, if you think about apps running on watchOS or writing daemons in Swift (and we have about 120 of those running on iOS), it is a lot of memory.
- In reality, only a couple of those conformances are used by applications, which means that most of the loaded conformances are useless and just consume the memory.
The implemented solution:
- Load only the conformances which were queried by calling swift_conformsToProtocol. Don’t try to load any other conformances, when not asked to do so.
- Use std::vector instead of std::deqeue for SectionsToScan. This shaves off another 4KB of memory, because std::deque reserves at least 4KB by default.
This patch does not seem to produce any detectable performance hit on our benchmark suite.
With these changes, the minimal "hello world" application consumes only 9.8KB, whereas before it used to consume 60KB.
rdar://22331482
Swift SVN r31310
Un-revert the below commits with the following addition:
add declarations for posix_spawn related APIs to SwiftPrivateDarwinExtras.
posix_spawn-related APIs aren't available in the public SDKs, so force past
the availability by creating our own stubs in the internal DarwinExtras
library.
r31244, r31245
CMake: build all platforms except watchOS using the public SDK
Covers rdar://problem/21145996.
A step towards rdar://problem/21099318.
Switch SDK overlays to use the public SDK
I had to cut the dependency on CrashReporterClient.h and reimplement
some of that code inline in the Swift runtime. This shoud be OK (even
though not very clean), since the layout of CrashReporter sections is
ABI.
rdar://21099318
Swift SVN r31252
I had to cut the dependency on CrashReporterClient.h and reimplement
some of that code inline in the Swift runtime. This shoud be OK (even
though not very clean), since the layout of CrashReporter sections is
ABI.
rdar://21099318
Swift SVN r31245
dealloc_ref [destructor] is the existing behavior. It expects the
reference count to have reached zero and the isDeallocating bit to
be set.
The new [constructor] variant first drops the initial strong
reference.
This allows DI to properly free uninitialized instances in
constructors. Previously this would fail with an assertion if the
runtime was built with debugging enabled.
Progress on <rdar://problem/21991742>.
Swift SVN r31142
This reverts commit cd3f1ba7d1ee2397817e1a165209fdeab8a1c004.
Reverting this b/c it is breaking buildbots with the following:
CMake Error at CMakeLists.txt:522 (message):
CrashReporterClient library is required, but it was not found
Swift SVN r31047
This got flagged by the ASan bot once "Enable reflection for multi-payload
enums with non-trivial layout" went in, but the problem existed all along.
The field types array is only as large as the number of payload cases, but
we were loading from it unconditionally. We would set payloadType to nullptr
afterwards anyway in this case, but indirect was potentially wrong.
Swift SVN r30533
This reverts commit r30215.
Fixes a bunch of problems on the ASAN bot.
Before:
Swift :: 1_stdlib/ErrorType.swift
Swift :: 1_stdlib/Runtime.swift
Swift :: Constraints/bridging.swift
Swift :: Constraints/diagnostics.swift
Swift :: Constraints/lvalues.swift
Swift :: DebugInfo/variables-repl.swift
Swift :: Interpreter/enum_runtime_alignment.swift
Swift :: Interpreter/nil_error_value.swift
Swift :: Interpreter/return_from_main.swift
Swift :: Misc/misc_diagnostics.swift
Swift :: Prototypes/Result.swift
Swift :: expr/expressions.swift
Swift-Unit :: runtime/SwiftRuntimeTests/MetadataTest.installCommonValueWitnesses_pod_indirect
After:
Swift :: Constraints/bridging.swift
Swift :: Constraints/diagnostics.swift
Swift :: Constraints/lvalues.swift
Swift :: Misc/misc_diagnostics.swift
Swift :: expr/expressions.swift
Swift-Unit :: runtime/SwiftRuntimeTests/MetadataTest.installCommonValueWitnesses_pod_indirect
Swift SVN r30396
Full type metadata isn't necessary to calculate the runtime layout of a dependent struct or enum; we only need the non-function data from the value witness table (size, alignment, extra inhabitant count, and POD/BT/etc. flags). This can be generated more efficiently than the type metadata for many types--if we know a specific instantiation is fixed-layout, we can regenerate the layout information, or if we know the type has the same layout as another well-known type, we can get the layout from a common value witness table. This breaks a deadlock in most (but not all) cases where a value type is recursive using classes or fixed-layout indirected structs like UnsafePointer. rdar://problem/19898165
This time, factor out the ObjC-dependent parts of the tests so they only run with ObjC interop.
Swift SVN r30266
Full type metadata isn't necessary to calculate the runtime layout of a dependent struct or enum; we only need the non-function data from the value witness table (size, alignment, extra inhabitant count, and POD/BT/etc. flags). This can be generated more efficiently than the type metadata for many types--if we know a specific instantiation is fixed-layout, we can regenerate the layout information, or if we know the type has the same layout as another well-known type, we can get the layout from a common value witness table. This breaks a deadlock in most (but not all) cases where a value type is recursive using classes or fixed-layout indirected structs like UnsafePointer. rdar://problem/19898165
Swift SVN r30243
This came up for multi-payload enums without generic parameters, eg
enum MyError {
case BusError
case TrainError(Int)
case DataLoss(String)
}
Fixes <rdar://problem/21739870>.
Swift SVN r30215
These will be used for reflection, and eventually to speed up generic
operations on single payload enums as well.
Progress on <rdar://problem/21739870>.
Swift SVN r30214
Leave the qualification off of enum cases and type names when 'print'-ing them, but keep them on 'debugPrint'. (At least, at the outermost level; since ad-hoc printing of structs and tuples uses debugPrint, we'll still get qualification at depth, which kind of sucks but needs more invasive state management in print to make possible.) Implements rdar://problem/21788604.
Swift SVN r30166
We incorrectly tested the uninitialized "next" pointer against MAP_FAILED, instead of the real result of mmap. Fixes rdar://problem/21659505.
Swift SVN r30030
Change ProtocolConformanceRecord::getCanonicalTypeMetadata() to return null if
the class metadata in the protocol conformance record is null. This fixes a
runtime crash in swift_conformsToProtocol when there is a protocol extension of
a missing weakly-linked class.
rdar://problem/21541766
Swift SVN r30012
This brings the David Owens benchmark from http://owensd.io/2015/06/27/performance-xcode7-beta-2.html from parity with simd.h-based C to 3x faster.
Before:
RenderGradient ([UInt32].withUnsafeMutablePointer (SIMD)) │ 7.035851 │ 6.304739 │ 9.815832 │ 1.212 │
After:
RenderGradient ([UInt32].withUnsafeMutablePointer (SIMD)) │ 2.318357 │ 2.223325 │ 2.697981 │ 0.1490 │
This also addresses rdar://problem/21574425, since Builtin.add_VecNxIntM isn't overflow-checked, and overflow checks really aren't wanted when working with vector types directly.
Reapplying now that Nadav's fixed the ARM64 SelectionDAG issue this exposed before, and Arnold's fixed
yet another SelectionDAG issue exposed after that.
Swift SVN r30006
Metatypes can't directly conform to _ObjectiveCBridgeable, but we can pretend they do by making a struct with the same ABI conform and returning that conformance when we call findBridgeWitness on a metatype. Fixes rdar://problem/16238475.
Swift SVN r29999
The stdlib uses this condition to recognize types that represent classes without representation changing, which isn't true for metatypes. They will natively be pointers to the Swift type metadata instead of the ObjC class object, so a conversion step is necessary. This doesn't directly fix container bridging, but it prevents the runtime from trying to bridge verbatim metatypes without first changing them to ObjC representation.
Swift SVN r29998
Implement casting support for taking an AnyObject and conditionally converting it to a T.Type for some class type. Fix some memory management bugs too, where we used swift_release to release an object not known to have Swift refcounting. This mostly fixes rdar://problem/16238475, though the SIL optimizer still improperly folds away attempted casts from NSObject to a class metatype, and I haven't yet validated bridging support for NSArray<Class>*.
Swift SVN r29956
Revert "simd overlay: Use LLVM vector types."
This reverts commit r29922 and r29924.
More arm64 instruction selection errors.
rdar://21703486
Swift SVN r29941
This brings the David Owens benchmark from http://owensd.io/2015/06/27/performance-xcode7-beta-2.html from parity with simd.h-based C to 3x faster.
Before:
RenderGradient ([UInt32].withUnsafeMutablePointer (SIMD)) │ 7.035851 │ 6.304739 │ 9.815832 │ 1.212 │
After:
RenderGradient ([UInt32].withUnsafeMutablePointer (SIMD)) │ 2.318357 │ 2.223325 │ 2.697981 │ 0.1490 │
This also addresses rdar://problem/21574425, since Builtin.add_VecNxIntM isn't overflow-checked, and overflow checks really aren't wanted when working with vector types directly.
Reapplying now that Nadav's fixed the ARM64 SelectionDAG issue this exposed before.
Swift SVN r29922
