This corrects the type given to GEP2 when initializing a stacklocal
array. This issue only cause problems when bitcode gets written before
any of the optimization passes, in which case it will fail on LLVM
validation when reading the bitcode back in on a subsequent pass.
Kotlin code for Objective-C interop is always included to stdlib.
It depends on some C++ code. The latter is compiled conditionally -
only for targets supporting Objective-C interop. That might cause
undefined symbols problems when DCE is not enabled.
To deal with it, we usually have trivial stubs in C++ code, used
for targets not supporting Objective-C interop.
This commit adds a couple of such missing stubs.
The -Xcoverage feature has not worked and has been disabled for a while.
This fix removes it, and all of its uses.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-821
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
LLVMConstGEP2 and LLVMConstInBoundsGEP2 were forward declared in
LLVM-11, but not implemented until LLVM-14. This patch adds these along
with the missing LLVMAddAlias2. All three implementations are copied
from llvm branch release/14.x
This commit specializes the existing Kotlin/Native stdlib ArrayList into
two subclasses:
* ArrayList
* ArraySubList
This avoids repeatedly checking whether a basic ArrayList is created as a
sublist of another ArrayList.
In the iterators, checkForComodification is marked for inlining, since
this significantly improves iterations performance.
A number of benchmarks are added to the native ring benchmark suite, to
test whether the changed runtime type of ArrayList.subList(...) has an
impact.
This change reorders declared fields such that LLVM can pack each class
space optimally, minimizing the padding required. The guarantee is that
an objects body will take up space equal to the sum of the sizes of its
fields (inherited and declared) rounded up to a multiple of 8 bytes.
This also implies that no object will contain more than 7 bytes of
padding.
This increases the boundary for the next GC scheduling by how much is
estimated is currently being wasted by the allocator due to paging. This
is necessary, since the gc scheduling boundary is set based on the size
of the live set at the last GC, while the current memory use is only
tracked in whole allocator pages. This patch avoids some of the
situations where the GC would be triggered again immediately upon
completion, which happens with many sparsely filled pages.
After the change to track allocations in big chunks (KT-57773), the
scheduler no longer had information about how much of the allocated
memory was used on live objects, which lead to a lot of extra memory
usage (KT-61914). This patch adds the tracking of live objects size to
sweeping, which is enough to mitigate the problem.
co-authored by Alexander Shabalin
^KT-61914
Also enables the integration tests for custom allocator and the
currently available GCs.
Co-authored-by: Troels Bjerre Lund <troels@google.com>
Merge-request: KT-MR-11199
Merged-by: Alexander Shabalin <Alexander.Shabalin@jetbrains.com>
The command line argument parser is using between 0.25s and 0.5s
(depending on platform) on finding annotated properties. This fix
replaces the slow kotlin reflection with java reflection, which is an
order of magnitude faster.
#KT-58183 Fixed
This avoids that the thread sweeping the base object can get stale
locals, in the case that the finalizer thread unlinks the extraobject at
the same time. This is done by moving the responsibility of unlinking
the extraobject to the sweeping of the base object.
Co-authored-by: Troels Bjerre Lund <troels@google.com>
Merge-request: KT-MR-10518
Merged-by: Alexander Shabalin <Alexander.Shabalin@jetbrains.com>
On linux, some memory access patterns cause a cascade of page faults. By
asking mmap for MAP_POPULATE on linux, the page faults happen up front
in a much more efficient manner.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-697
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
Swept and kept objects in NextFitPages are double counted; first in
GCApi::SweepObject and then again inside NextFitPage::Sweep. This fix
removes the latter.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-691
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
The existing threshold for determining between medium and large page
does not take the dummy block into account, meaning that an allocation
on the threshold will lead to a loop that only terminates when the
process goes out of memory.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-618
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
This commit adds finalizers to the custom allocator. Compared to the
existing solution:
* The finalizer queue is replaced by an AtomicStack<ExtraObjectData>.
* All objects with finalizers get meta objects attached. This was not
previously the case for CleanerImpl, but is now needed to link objects
together in the finalization queue.
* The finalizer queue is built during SweepExtraObjects, instead of
during regular sweeping.
* Cleaners are executed by the finalizer thread, and no longer by a
separate worker thread.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-592
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
This change avoids looping through the source stack, looking for the
last element, in the case where the target stack is empty. This could
matter in two places:
* when merging ready_ and used_ into unswept_ in PrepareForGC. The first
of these will be faster with this change.
* when merging finalizer queues (in CR-592, not merged yet), where we
expect the target queue to be empty.
Since we can expect used_ to be larger than ready_, since we are about
to do a GC, the order of these two has also been reversed.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-599
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
The default behavior on a sweep is to push empty pages into a separate
stack, which will be freed if not used before the next GC starts. This
serves two purposes: it reduces the number of system allocations, and it
avoids a race condition inside AtomicStack::Pop.
Neither of these are relevant for LargePages, since LargePages are never
reused and it is only the GC thread that calls
AtomicStack<LargePage>::Pop. The change is to free LargePages
immediately instead of waiting for the next GC cycle.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-598
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
Enable custom allocator with -Xallocator=custom.
If the gc is cms (default):
* a patched version of the memory manager is used that does not build a
list of allocated objects.
* a patched version of the cms is used that defers to the allocator for
sweeping.
Otherwise, a warning is printed, and the allocator is used using the
standard api.
Design doc:
https://docs.google.com/document/d/15xMp-nE-DWL8OrtOc8DoXB80AHUphFICEGjj5K0aNFc
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-546
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
The bootstrap advance to 1.8.20-dev-2904 broke the K/N performance
benchmark build, when built in non-Java8 environments.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-576
Merged-by: Pavel Punegov <pavel.punegov@jetbrains.com>
This unlinks the benchmarksAnalyzer build from being configured automatically
whenever any benchmark build is.
Co-authored-by: Troels Lund <troels@google.com>
If the count argument is above 2**29, then memberSize does not fit in
uint32_t.
Co-authored-by: Troels Lund <troels@google.com>
Merge-request: KOTLIN-MR-533
Merged-by: Alexander Shabalin <alexander.shabalin@jetbrains.com>
When loading default libraries, the code assumed that all subdirectories
of <kotlin native home>/klib/*/ corresponded to valid libraries. The bug
report only considers the case of .DS_STORE, which is generated by
Finder on MacOS, but this case has already been accounted for in the
code.
The fix filters out all subdirectories that are not stdlib and don’t have the
“org.jetbrains.kotlin” prefix.
Co-authored-by: Troels Lund <troels@google.com>
Currently, :kotlin-native:endorsedLibraries does not have its build
directory deleted during a clean.
This task should probably be inherited from a plugin instead, but it is
not clear to me which one.
The lowering that inserts explicit returns into returnable blocks uses
the character offset of the returnable block for the newly created
return statement. This is wrong, but only noticeable when the callee is
an inlined function in another file, where the statements of the block
belong to callee rather than the caller.
The fix takes the offset of the last statement in the block. In the
case the block does not contain any statements, it falls back to using
the offset of the block itself.
The test does not impose any specific stepping order, since this would
be too fragile. It only tests that LLDB does not step to empty (or fully
//-commented) lines.
Co-authored-by: Troels Lund <troels@google.com>
Troels Bjerre Lund