There are several performance optimizations:
* ByteBuffer/CharBuffer/StringBuilder objects pre-allocated and are
reused on each call to readLine.
* The state for readLine is lazily allocated via JVM classloading
(using a singleton object).
* There is an auto-detection heuristic for "directEOL" encodings which
represent LF ('\n') directly as the corresponding byte
(UTF-8 and many single-byte encodings are like that).
When "directEOL" encoding is used, then bytes are batched into
ByteBuffer for a single call to CharsetDecoder.decode which
results in higher throughput. Otherwise (UTF-16, etc), slower
byte-by-byte approach is used.
* Bytes and chars are directly moved in/out of byte/char arrays and
ByteBuffer/CharBuffer wrappers are used only to interface with
JVM CharsetDecoder class (which is the slowest piece).
* StringBuilder is not used at all for short lines (<=32 chars).
There are also some function improvements to readLine functionality:
* Restriction on "max chars per byte" is lifted, so readLine works with
all encodings that JVM supports.
* It support on-the-fly changes to system default charset, because
it rechecks current charset on each call and updates it decoder
when needed.
All the other features of readLine function are retained:
* It does not read more bytes from System.in than needed, so it
is compatible with other ways to read System.in. On-the-fly
changes to System.in are supported.
* It is thread-safe. Its internal mutable state is protected by
synchronization.
* There is an internal method for tests that supports explicit
charset specification, but the name of this method has changed.
There are additional tests:
* Check all supported encodings on JVM to make sure that readLine
works correctly with them all.
* Check unicode code points of different bits length with all standard
unicode encodings (UTF-8, UTF-16, and UTF-32 in LE/HE byte orders).
Benchmarks that compare different implementations of readLine,
including this one (readLine6NoLV in the set) can be found here:
https://github.com/elizarov/ReadLineBenchmark
Taking BufferedReader as 100% baseline we see that:
* Current readLine is 7.5 times slower than BufferedReader baseline.
* New implementation in this commit is 2.5 timer slower than baseline.
It is ~3 times faster than existing implementation of readLine.
Altogether these optimizations are enough to enable reading of
~500K lines in sports programming setting under 2s time-limit with
plenty of headroom in time. Example that is using this version of
readLine can be found here:
https://codeforces.com/contest/1322/submission/73005366
#KT-37416 Fixed
Instead of generating overrides for getOwner/getName/getSignature in
each anonymous class representing a callable reference, pass them to the
superclass' constructor and store as fields. This occupies some small
memory but helps to reduce the size of the generated class files, and
will be helpful for adding further runtime information to callable
references, such as information about implicit conversions this
reference has been subject to.
Represent owner as java.lang.Class + boolean instead of
KDeclarationContainer, so that the unnecessary wrapping Class->KClass
wouldn't happen before it's needed, and also to make sure all callable
references remain serializable.
Note that the argument type where the "is declaration container a class"
is passed is int instead of boolean. The plan is to pass the
aforementioned implicit conversion information as bits of this same
integer value.
#KT-27362 Fixed
Instead of non-existing `kotlin.KotlinPackage`, which led to
NoClassDefFoundError as soon as reflection tried to call getOwner on a
builtin callable reference, use a reference to a new physical class
`kotlin.jvm.internal.Intrinsics$Kotlin`. This will allow to support
KT-17151.
Note that for API version less than 1.4, this will still lead to
NoClassDefFoundError, but this is not worse than the current situation
where it happens anyway.
There is one failing test namely `ValByMapExtensionsTest.doTest`, which
is quite questionable because its checks the use of out projection and
Exact annotation (see KT-18789)
From now on, the old JVM backend will report an error by default when
compiling against class files produced by the JVM IR backend. This is
needed because we're not yet sure that the ABI generated by JVM IR is
fully correct and do not want to land in a 2-dimensional compatibility
situation where we'll need to consider twice more scenarios when
introducing any breaking change in the language. This is generally OK
since the JVM IR backend is still going to be experimental in 1.4.
However, for purposes of users which _do_ need to compile something with
the old backend against JVM IR, we provide two new compiler flags:
* -Xallow-jvm-ir-dependencies -- allows to suppress the error when
compiling with the old backend against JVM IR.
* -Xir-binary-with-stable-api -- allows to mark the generated binaries
as stable, when compiling anything with JVM IR, so that dependent
modules will compile even with the old backend automatically. In this
case, the author usually does not care for the generated ABI, or s/he
ensures that it's consistent with the one expected by the old compiler
with some external tools.
Internally, this is implemented by storing two new flags in
kotlin.Metadata: one tells if the class file was compiled with the JVM
IR, and another tells if the class file is stable (in case it's compiled
with JVM IR). Implementation is similar to the diagnostic reported by
the pre-release dependency checker.
Ilya Gorbunov