| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| | |
|
| | |
|
| |
|
|
| |
That way it is possible to check if a ListMap has one element by checking x.tail.isEmpty. Size is O(1), so size==1 won't do!
|
| | |
|
| |
|
|
| |
elements
|
| | |
|
| |
|
|
|
|
| |
The old approach first created a broken HashTrieSet (with elems(0) == null) and then fixed it afterwards by assigning elems(0) from the outside. But that won't work with the assertion.
The new method is recursive, but the maximum recursion depth is 7, and on average the method won't recurse at all.
|
| |
|
|
| |
If this is the case we must return subNew because a HashTrieMap with one leaf element is not allowed by the assertions
|
| |
|
|
| |
of it from HashMapCollision1.updated0 as well
|
| |
|
|
| |
While we still have to check for merger eq null in updated0 because updated0 gets called with a nu
|
| |
|
|
| |
HashTrieMap itself
|
| | |
|
| |\
| |
| | |
Compiler optimizations (take 3, hopefully final)
|
| | |
| |
| |
| |
| |
| |
| | |
Reduced method size of map/flatMap to under 35 bytes to make them
inlinable. Also specialized filterNot code so that it does not create
a second closure. Finally, avoided default argument for sizeHint to
reduce size of methods calling it.
|
| | |
| |
| |
| | |
All statistics code is now subject to a switch at each use site, as opposed to a compile-time constant before. We should veryfy that this does not affect performance. Statistics is a useful ability to have for trouble-shooting compile time speeds on specific code bases. But we should not pay a lot for having it on normal runs.
|
| | |
| |
| |
| |
| |
| | |
If our theory wrt map is correct (i.e. that it gets inlined by JVM), then closure hoisting is counter-productive because it migh well prevent the closure from being inlined. This commit removes all hoisted closures that are passed to map and leaves only those boolean closures that are passed to exists and forall.
I should have split the early optimizations including closures into separate commits. As that train has left, I am now reverting some bits to see whether the reverts affect performance at all, and in what direction.
|
| | | |
|
| | | |
|
| | |
| |
| |
| | |
Profile data actually showed a small slowdown in steady state. We keep the reordering but merge all submethods into one.
|
| | |
| |
| |
| | |
It seems any gains were more then compensated by the loss of making map polymorphic, probably preventing inlining decisions in Hotspot.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Inline `UndoLog.undoUnless` into `isSubType`
and `isSameType` methods. They are responsible
for 20% of all Function0 allocations.
I had to do manual inlining because inlier
doesn't cope with exception handlers well.
Let it be noted that I almost cried while doing
this ugly change. I'll want to see inliner
being fixed so we can revert this change ASAP.
|
| | |
| |
| |
| |
| |
| |
| | |
Two optimizations:
(1) Make SymTree a class rather than a trait to allow faster access to symbol
(2) Split itransform into several smaller methods and order cases according to expected frequency.
|
| | | |
|
| | |
| |
| |
| |
| |
| |
| |
| | |
Curiously, profiling showed that these methods (in particular map) got slower when specialized in LinearSeqOptimized from their generic impl in TraversableLike.
I speculate that the virtual calls to head/tail/nonEmpty are more expensive than the closure creation which we save (Note that foreach IS specialized in List,
so head/tail are monomorphic, even though isEmpty still has two implementations.
If we confirm a speedup in this commit it would make sense to specialize more methods in LinearSeqOptimized again in List.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Provided specialized version of hashing algorithm
for List that does not allocate the closure.
This saves us 124k closures allocated in a compiler
when compiling `Vector.scala` from std library.
That's aroudn 5% of all closures allocated for that
compiler run.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| | |
Introduced explicit locking mechanism that allows us to
make public methods of `UndoLog` that take thunks as
arguments final and thus `@inline`. Once inliner is able
to cope with inlining methods with exception handlers
we'll be able to get rid of significant (around 5%)
number of thunk allocation.
|
| | |
| |
| |
| |
| | |
Those two methods contribute 2% of Function0 allocation
when compiling `Vector.scala`.
|
| | |
| |
| |
| |
| | |
Overriden ClassTags in places where we allocate
a lot of Arrays.
|
| | |
| |
| |
| | |
Driven by profile data.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| | |
profiling data shows that accessing the hashCode field has a cost of about 10% of fully hot running times. We speculate that it's megamorphic dispatch which is costing us this much.
Rewrote code so that UniqueType is not an abstract base class with the hashCode field.
Also removed fingerPrinting bloom filter on findMember because it caused complexity and did not gain us anything accdoring to the tests.
Made TypeVar an abstract case class to avoid extractor calls in TypeMap.
|
| | |
| |
| |
| | |
now the same and faster.
|
| | |
| |
| |
| | |
Moved costly hashMap updates into SuperAccessors from general tree transformers and replaced by immutable map.
|
| | | |
|
| | | |
|
| | | |
|
| | |
| |
| |
| | |
Also, added inlined List.foreach
|
| | |
| |
| |
| | |
All cases factored out into separate methods. Cases ordered according to node frequency.
|
| | | |
|
| | | |
|
| | |
| |
| |
| |
| | |
Also avoided systematically to map (_.tpe) on parameters in favor of
lazy val paramTypes.
|
| |/
|
|
|
|
|
|
|
|
|
|
| |
It turns out that exists is not inlinable, even if put into List.
We try to eliminate or hoist most closures passed to exists in Types.
There are some other small improvements as well.
--
(@gkossakowski): This commit contains also a fix to crasher
prepared by @paulp. I squashed that commit and kept the
test-case that came with it.
|
| |\
| |
| | |
even more cleanup in Macros.scala
|
| | | |
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Error reporting is moved to ContextErrors to disentangle stuff in Macros.scala.
With logics and error reporting intertwined it was an awful mess.
Exceptions are used for the same reason. Instead of threading failures through
the code polluting it with options/ifs, I outline the success path.
It worked much better for typedMacroBody, but I'm also happy with the resulting
code of macroExpand. To me a major factor towards applicability of exceptions
was that they are short-lived and that there might be max one error per domain,
after which we unconditionally bail.
|
| |\ \
| | |
| | | |
Fix for community build blocker.
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
As two character bugfixes go, this was a doozy.
I will forego elaborating at length and offer generic
instructions for elucidation:
% git log --grep=tpeHK
|
| |\ \ \
| |/ /
|/| | |
SI-5788 correct test and symbol update
|
| |/ /
| |
| |
| |
| |
| | |
Previously I thought it's fixed because I didn't include the correct
testcase. Now it's the correct testcase and this prompted me to
change the code a bit to make it work properly.
|
| |\ \
| | |
| | | |
Use rawflags for checking FINAL flag.
|
| | | |
| | |
| | |
| | | |
Partially reverts 18efdedfb97de7ca9f6
|
