The current hot-reload implementation [0] will perform a reload by
first sending OOB messages to all isolates and waiting until those OOB
messages are being handled. The handler of the OOB message will block
the thread (and unschedule isolate) and notify the thread performing
reload it's ready.
This requires that all isolates within a group can actually run & block.
This is the case for the VM implementation of isolates (as they are
run an unlimited size thread pool).
Though flutter seems to multiplex several engine isolates on the same OS
thread. Reloading can then result in one engine isolate performing
reload waiting for another to act on the OOB message (which it will not
do as it's multiplexed on the same thread as the former).
Now that we have a more flexible safepointing mechanism (introduced in
[1]) we can utilize for hot reloading by introducing a new "reloading"
safepoint level.
Reload safepoints
-----------------------
We introduce a new safepoint level (SafepointLevel::kGCAndDeoptAndReload).
Being at a "reload safepoint" implies being at a "deopt safepoint"
which implies being at a "gc safepoint".
Code has to explicitly opt-into making safepoint checks participate /
check into "reload safepoints" using [ReloadParticipationScope]. We do
that at certain well-defined places where reload is possible (e.g. event
loop boundaries, descheduling of isolates, OOM message processing, ...).
While running under [NoReloadScope] we disable checking into "reload
safepoints".
Initiator of hot-reload
-----------------------
When a mutator initiates a reload operation (e.g. as part of a
`ReloadSources` `vm-service` API call) it will use a
[ReloadSafepointOperationScope] to get all other mutators to a
safepoint.
For mutators that aren't already at a "reload safepoint", we'll
notify them via an OOB message (instead of scheduling kVMInterrupt).
While waiting for all mutators to check into a "reload safepoint", the
thread is itself at a safepoint (as other mutators may perform lower
level safepoint operations - e.g. GC, Deopt, ...)
Once all mutators are at a "reload safepoint" the thread will take
ownership of all safepoint levels.
Other mutators
-----------------------
Mutators can be at a "reload safepoint" already (e.g. isolate is not
scheduled). If they try to exit safepoint they will block until the
reload operation is finished.
Mutators that are not at a "reload safepoint" (e.g. executing Dart or VM
code) will be sent an OOB message indicating it should check into a
"reload safepoint". We assume mutators make progress until they can
process OOB message.
Mutators may run under a [NoReloadScope] when handling the OOM message.
In that case they will not check into the "reload safepoint" and simply
ignore the message. To ensure the thread will eventually check-in,
we'll make the destructor of [~NoReloadScope] check & send itself a new OOB
message indicating reload should happen. Eventually getting the mutator
to process the OOM message (which is a well-defined place where we can
check into the reload safepoint).
Non-isolate mutators such as the background compiler do not react to OOB
messages. This means that either those mutators have to be stopped (e.g.
bg compiler) before initiating a reload safepoint operation, the
threads have to explicitly opt-into participating in reload safepoints
or the threads have to deschedule themselves eventually.
Misc
----
Owning a reload safepoint operation implies also owning the deopt &
gc safepoint operation. Yet some code would like to ensure it actually
runs under a [DeoptSafepointOperatoinScope]/[GCSafepointOperationScope].
=> The `Thread::OwnsGCSafepoint()` handles that.
While performing hot-reload we may exercise common code (e.g. kernel
loader, ...) that acquires safepoint locks. Normally it's disallows to
acquire safepoint locks while holding a safepoint operation (since
mutators may be stopped at places where they hold locks, creating
deadlock scenarios).
=> We explicitly opt code into participating in reload safepointing
requests. Those well-defined places aren't holding safepoint locks.
=> The `Thread::CanAcquireSafepointLocks()` will return `true` despite
owning a reload operation. (But if one also holds deopt/gc safepoint
operation it will return false)
Example where this matters: As part of hot-reload, we load kernel which
may create new symbols. The symbol creation code may acquire the symbol
lock and `InsertNewOrGet()` a symbol. This is safe as other mutators
don't hold the symbol lock at reload safepoints. The same cannot be said
for Deopt/GC safepoint operations - as they can interrupt code at many
more places where there's no guarantee that no locks are held.
[0] https://dart-review.googlesource.com/c/sdk/+/187461
[1] https://dart-review.googlesource.com/c/sdk/+/196927
Issue https://github.com/flutter/flutter/issues/124546
TEST=Newly added Reload_* tests.
Change-Id: I6842d7d2b284d043cc047fd702b7c5c7dd1fa3c5
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/296183
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
The sweeper threads are bypassing safepoints and have no need for normal
mutator state (e.g. storebuffer, marking stacks, .,..) on the [Thread]
object.
This fixes a TSAN report where sweeper would clear reusable handles
(which it didn't actually modify) and scavenger is reading those
handles (which arguably it doesn't have to either).
Issue https://github.com/dart-lang/sdk/issues/52125
TEST=ci
Change-Id: I03a36e8518b6c00eb7f3b57f65fd469dddba23e0
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/296860
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
An embedder (or the VM) can exit an isolate via `Thread::ExitIsolate()`
at a point where there's still active state (e.g. dart frames).
Because of this the VM has so far conservatively retained the [Thread]
object of dart mutators throughout the isolate's lifetime. After which
is was manually `delete`ed. We'd never re-use those [Thread] objects (we
do re-use [Thread] objects of non-dart-mutator threads).
When exiting via `Thread::ExitIsolate()` with active state, the mutator
was assumed to be at-safepoint at all levels. It was removed from the
thread registry's active threads. This also means that when e.g. GC runs
it can't use the thread registry to find all active threads it may
need to scan, instead it uses [Isolate::mutator_thread_] of all isolates.
This causes a variety of subtle issues, but the main one that motivated
this change is the following:
If a thread obtains a safepoint operation it means all other mutators
are parked. The thread owning the safepoint can do whatever it likes.
When introducing reload operation safepoints, a thread may want to
ReloadSafepointOperation reload(thread);
...
// Compile sources.
{
TransitionVMToNative transition(thread);
// Will temporarily exit & re-enter current isolate.
response_port = Dart_NewNativePort();
Dart_PostCObject(kernel_isolate_port, ...);
// Wait on [response_port] for response.
}
This will cause the reloading thread to own the reload safepoint
operation but still transition states and even exit/re-enter the
isolate. Though this is currently not possible in the way enter/exit is
implemented.
So we'll refactor this fragile code in the following way:
* Move thread enter/exit logic entirely to the [Thread] object.
* Keep used threads in the thread registry's active list.
=> This allows us to keep various state on the [Thread] and thereby
avoids clearing it when suspending & re-initialing it when resuming
=> It makes nested `Thread::ExitIsolate()` faster as we mainly have
to enter safepoint (avoid acquiring threads lock, avoid releasing
storebuffers, ...)
=> It makes nested `Thread::EnterIsolate()` faster as we mainly have
to leave the safepoint (avoid acquiring threads lock, avoid acquiring
storebuffers, ...).
=> A mutator can now own a safepoint operation (e.g. reload safepoint
operation) and still `ExitSafepoint()` / `EnterSafepoint()` safely -
as those are based on the normal `EnterSafepoint()` and
`LeaveSafepoint()` APIs.
* We separate
- Suspend & Resume of a dart mutator (possibly with active stack)
- Setup & Reset of state only relevant for dart mutators
- Setup & Reset of state relevant for any mutator
* We unify how the [Thread] objects are freed between dart mutator and
non-dart mutators: [Thread] objects without state can be given back to
the [ThreadRegistry] and re-used (instead of being deleted in
`Isolate::~Isolate`)
* We have capability to free [Thread] objects if a dart mutator has an
empty stack & re-use for another isolate of the same group.
(In future we may have N Thread objects for N cores and the threads
would even maintain their TLABs when switching between isolates)
* Since we allow reusing of [Thread] objects also for dart mutators now,
we have extensive asserts to ensure they are "clean" when they get
into the free list and come out "clean" again.
TEST=ci
Change-Id: Id85e8e484efd98d28e323b33795716420e619986
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/296585
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Commit-Queue: Martin Kustermann <kustermann@google.com>
Some cleanups factored out of a larger CL (which refactors enter/exit of threads):
* remove unused `#include "vm/thread_registry.h"`
* remove unused/unnecessary fields from [Thread] object
* rename IsMutator() -> IsDartMutator()
* make tests using setjmp() drain the sticky error
=> to ensure there's no sticky error on isolate shutdown
TEST=ci
Change-Id: I53935e8bd0628ab3768627d6d5e01c3f0e3a57ad
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/296582
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Alexander Aprelev <aam@google.com>
For every isolate there should be only one mutator with
a unique [Thread] object.
We change existing tests that use this functionality to instead use
`Thread::{Enter,Exit}IsolateGroupAsHelper`. It also results in a net
removal of code.
TEST=ci
Change-Id: Ic326e868a98ddedbab5b8c429252d38ea71bbf04
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/295940
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
During safepoint we can distinguish between
* owner of the safepoint operation (which is running code)
* everyone else (which are all blocked
Currently `Thread::IsAtSafepoint()` will return true for both. Since the
thread owning the safepoint operation is running, it's not actually
guaranteed that it's at "safe" point (e.g. to GC or to deopt) - it
really depends on what it's doing.
=> This CL will change it so that only actually parked threads will
have `Thread::IsAtSafepoint()`.
In order to do that we change varrious usages of `IsAtSafepoint()` to be
more precise:
* `Thread::OwnsSafepoint()`: True if this thread owns the
active safepoint. The thread is running.
* `Thread::OwnsGCSafepoint()`: True if the active safepoint is a GC
(or Deopt) safepoint and this thread owns it. The thread is running.
* `Thread::OwnsDeoptSafepoint()`: True if the active safepoint is a
Deopt safepoint and this thread owns it. The thread is running.
* `Thread::CanAcquireSafepointLocks()`: True if the thread is allowed
to acquire safepoint locks.
* `Thread::IsAtSafepoint()`: true if this thread is parked at a
safepoint
TEST=ci
Change-Id: I1a5a6727e84843ae79e0a344c438da19b7d6d916
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/295781
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
This CL makes some refactorings to the current safepoint mechanism:
* When owning safepoint level L we set current thread to be only at
safepoint level L (not Thread::Current()->current_safepoint_level()).
This ensures that after [WaitUntilThreadsReachedSafepointLevel] all
other threads are actually parked.
* When having nested safepoint scopes we increase operation count on
the level we own and all nested levels. This will (in later CL) allow
detection which closest scope we're in.
TEST=ci
Change-Id: Iffb2e9f4eea817a381acbd7a771bc75f5a89877b
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/295541
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Commit-Queue: Martin Kustermann <kustermann@google.com>
Our compiler shouldn't depend on current isolate, since it can
run on any isolate within an IG.
Doing this change, reveals two existing dependencies on current
isolate from compiler
- resolving native symbols in unoptimized compilations
- issuing of debug events for breakpoints
For the former we'll re-enter the currently active isolate that
triggered unoptimized compilation.
=> We may want to change that embedder API to not be based on
handles and instead give embedder a simple `const char*`.
For the ladder we'll enter the isolate corresponding to the
breakpoint debug event to be issued. We are at place where
all mutators are stopped, so that does seem okish.
=> Future could remove this by making Object Id Ring per-IG
Issue https://github.com/dart-lang/sdk/issues/48523
TEST=service_2/break_on_function_many_child_isolates_test/dds
Change-Id: Id246db5972ae505e82f637ce04bb2302bed76257
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/278901
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Right each `Pointer.fromFunction()` invocation will lead to creation of
a new ffi trampoline function & it's following JITed code. In AOT we
have exactly one ffi trampoline per target/native-signature/exceptional-return
combination.
=> This CL ensures we have only one such function.
Furthermore each `Pointer.fromFunction()` will currently perform 2
runtime calls in JIT: One to create a `Function` object, the other to
JIT that function & register callback metadata.
=> This CL ensures we won't do a runtime call to get a function, instead
do it at compile-time (as in AOT)
Furthermore we eagerly assign a callback-id to the unique/deduped ffi
trampoline callbacks. Only when the application requests a pointer, do
we populate metadata on the `Thread` object.
This CL doesn't (yet) change the fact that in JIT mode we have
isolate-specific jit trampolines (that will call now shared ffi trampoline
functions).
We also avoid baking in C++ runtime function pointers in generated
code. As a result we can now preserve ffi trampolines across AppJIT
serialization.
As a nice side-effect, we remove 100 lines of code.
TEST=ffi{,_2}/ffi_callback_unique_test
Issue https://github.com/dart-lang/sdk/issues/50611
Change-Id: I458831a47b041a088086f28f825de2a3849f6adc
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/273420
Reviewed-by: Daco Harkes <dacoharkes@google.com>
Commit-Queue: Martin Kustermann <kustermann@google.com>
This CL introduces new embedding APIs for supporting heap sample
profiling. A registered sampling callback is invoked approximately every
N bytes based on an exponential distribution, providing information
about the isolate group the allocation occurred in, the user visible
name of the allocated object type, a weak persistent handle to the
allocated object, and the size of the allocation.
Sampling is triggered using artificial TLAB boundaries to cause
allocations to be sampled to take the allocation slow path where the
registered callback can be invoked with the allocation information.
Only new space allocations are currently traced, with old space
allocation support to be added in a future CL.
TEST=Dart_HeapSampling
Change-Id: I22bcdeec6e823bc1ab44898d4c596fbed7169fa1
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/264520
Commit-Queue: Ben Konyi <bkonyi@google.com>
Reviewed-by: Siva Annamalai <asiva@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
This change introduces specialized stubs and IL instruction for
allocating records with 2 or 3 fields. This makes allocation of
small records slightly faster compared to a construction of similar
class instances and makes code size of record allocation smaller.
Benchmark:
MultipleReturns.NotInlined.Record(RunTime) 77150 -> 66222
MultipleReturns.NotInlined.RecordNamed(RunTime) 78073 -> 67044
MultipleReturns.Forwarded.Record(RunTime) 97130 -> 77635
MultipleReturns.Forwarded.RecordNamed(RunTime) 96495 -> 77904
TEST=ci
Issue: https://github.com/dart-lang/sdk/issues/49719
Change-Id: I8ed7add06b39ba79dfd78bbe2afaefe606cc505b
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/266420
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
Enable concurrent marking for IA32.
This removes the last write barrier relying on different alignment offsets for old and new space.
TEST=ci
Change-Id: Ib1c13124002392cf1c3ec264643325ec471a6918
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/262280
Commit-Queue: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Fix: Check handle contents for Smi.
Closes: https://github.com/flutter/flutter/issues/112726
Orignal CL description:
Makes `Dart_Handle` FFI returns behave as the following snippet:
```
Dart_Handle ExampleSnippet() {
Dart_Handle result = ...;
if (Dart_IsError(result)) {
Dart_PropagateError(result);
}
return result;
}
```
Also makes FFI consistent with Dart_NativeFunctions, which will
automatically throw upon return if Dart_SetReturnValue set the result
to an error.
`UnhandledExceptions` cannot flow out into Dart generated code. So,
the implementation needs to be in `FfiCallInstr::EmitNativeCode`.
Using `Dart_IsError` is slow compared to a machine code class id
check. So, we should do the handle unwrapping and class id check in
machine code.
Unwrapping Handles in machine code is only safe when the GC is
guaranteed to not run: Either (1) in `kThreadInGenerated`, or (2) in
`kThreadInNative`, but only when transitioned into safepoint. So, the
handle cannot be unwrapped immediately after the FFI call in machine code. We first need to transition back to generated.
This means we need to transition again to native to do the actual
`Dart_PropagateError` call. We can do so without the stub in JIT
because we never return with normal control flow.
Performance impact of this change is within benchmark noise in both
JIT and AOT.
Size impact is 42 bytes on x64, which is 10% in AOT and 12% in JIT.
For more numbers see: go/dart-ffi-handle-error
TEST=runtime/bin/ffi_test/ffi_test_functions_vmspecific.cc
TEST=tests/ffi/vmspecific_handle_test.dart
Closes: https://github.com/dart-lang/sdk/issues/49936
Change-Id: Id8edfd841a7d6246438386007d83747868a0a151
Cq-Include-Trybots: luci.dart.try:vm-canary-linux-debug-try,vm-ffi-android-debug-arm64c-try,vm-ffi-android-debug-arm-try,vm-kernel-gcc-linux-try,vm-kernel-linux-debug-x64-try,vm-kernel-linux-debug-x64c-try,vm-kernel-msvc-windows-try,vm-kernel-optcounter-threshold-linux-release-x64-try,vm-kernel-precomp-asan-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-reload-linux-debug-x64-try,vm-kernel-reload-rollback-linux-debug-x64-try,vm-kernel-tsan-linux-release-x64-try,vm-kernel-win-debug-x64-try,vm-kernel-win-debug-ia32-try,vm-precomp-ffi-qemu-linux-release-arm-try,vm-precomp-ffi-qemu-linux-release-riscv64-try,vm-kernel-linux-debug-ia32-try,vm-kernel-mac-release-arm64-try,vm-kernel-precomp-win-debug-x64c-try
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/262342
Reviewed-by: Martin Kustermann <kustermann@google.com>
Auto-Submit: Daco Harkes <dacoharkes@google.com>
Commit-Queue: Daco Harkes <dacoharkes@google.com>
This reverts commit d9c442bce8.
Reason for revert: https://github.com/flutter/flutter/issues/112726
Original change's description:
> [vm/ffi] Throw on returning `Error` in `Handle`
>
> Makes `Dart_Handle` FFI returns behave as the following snippet:
>
> ```
> Dart_Handle ExampleSnippet() {
> Dart_Handle result;
> if (Dart_IsError(result)) {
> Dart_PropagateError(result);
> }
> return result;
> }
> ```
>
> Also makes FFI consistent with Dart_NativeFunctions, which will
> automatically throw upon return if Dart_SetReturnValue set the result
> to an error.
>
> `UnhandledExceptions` cannot flow out into Dart generated code. So,
> the implementation needs to be in `FfiCallInstr::EmitNativeCode`.
>
> Using `Dart_IsError` is slow compared to a machine code class id
> check. So, we should do the handle unwrapping and class id check in
> machine code.
>
> Unwrapping Handles in machine code is only safe when the GC is
> guaranteed to not run: Either (1) in `kThreadInGenerated`, or (2) in
> `kThreadInNative`, but only when transitioned into safepoint. So, the
> handle cannot be unwrapped immediately after the FFI call in machine code. We first need to transition back to generated.
>
> This means we need to transition again to native to do the actual
> `Dart_PropagateError` call. We can do so without the stub in JIT
> because we never return with normal control flow.
>
> Performance impact of this change is within benchmark noise in both
> JIT and AOT.
> Size impact is 42 bytes on x64, which is 10% in AOT and 12% in JIT.
>
> For more numbers see: go/dart-ffi-handle-error
>
> TEST=runtime/bin/ffi_test/ffi_test_functions_vmspecific.cc
> TEST=tests/ffi/vmspecific_handle_test.dart
>
> Closes: https://github.com/dart-lang/sdk/issues/49936
> Change-Id: Ie8fabeb6d53bc80689541bc4470cb37ee2200581
> Cq-Include-Trybots: luci.dart.try:vm-canary-linux-debug-try,vm-ffi-android-debug-arm64c-try,vm-ffi-android-debug-arm-try,vm-kernel-gcc-linux-try,vm-kernel-linux-debug-x64-try,vm-kernel-linux-debug-x64c-try,vm-kernel-msvc-windows-try,vm-kernel-optcounter-threshold-linux-release-x64-try,vm-kernel-precomp-asan-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-reload-linux-debug-x64-try,vm-kernel-reload-rollback-linux-debug-x64-try,vm-kernel-tsan-linux-release-x64-try,vm-kernel-win-debug-x64-try,vm-kernel-win-debug-ia32-try,vm-precomp-ffi-qemu-linux-release-arm-try,vm-precomp-ffi-qemu-linux-release-riscv64-try,vm-kernel-linux-debug-ia32-try,vm-kernel-mac-release-arm64-try,vm-kernel-precomp-win-debug-x64c-try
> Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/261603
> Reviewed-by: Ryan Macnak <rmacnak@google.com>
> Reviewed-by: Martin Kustermann <kustermann@google.com>
> Commit-Queue: Daco Harkes <dacoharkes@google.com>
TBR=kustermann@google.com,rmacnak@google.com,dacoharkes@google.com,dart-scoped@luci-project-accounts.iam.gserviceaccount.com
Change-Id: I94cc63de16b54db2b0a4f92759c39a1e569b8e63
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Cq-Include-Trybots: luci.dart.try:vm-canary-linux-debug-try,vm-ffi-android-debug-arm64c-try,vm-ffi-android-debug-arm-try,vm-kernel-gcc-linux-try,vm-kernel-linux-debug-x64-try,vm-kernel-linux-debug-x64c-try,vm-kernel-msvc-windows-try,vm-kernel-optcounter-threshold-linux-release-x64-try,vm-kernel-precomp-asan-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-reload-linux-debug-x64-try,vm-kernel-reload-rollback-linux-debug-x64-try,vm-kernel-tsan-linux-release-x64-try,vm-kernel-win-debug-x64-try,vm-kernel-win-debug-ia32-try,vm-precomp-ffi-qemu-linux-release-arm-try,vm-precomp-ffi-qemu-linux-release-riscv64-try,vm-kernel-linux-debug-ia32-try,vm-kernel-mac-release-arm64-try,vm-kernel-precomp-win-debug-x64c-try
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/262270
Reviewed-by: Zach Anderson <zra@google.com>
Reviewed-by: Liam Appelbe <liama@google.com>
Commit-Queue: Zach Anderson <zra@google.com>
Makes `Dart_Handle` FFI returns behave as the following snippet:
```
Dart_Handle ExampleSnippet() {
Dart_Handle result;
if (Dart_IsError(result)) {
Dart_PropagateError(result);
}
return result;
}
```
Also makes FFI consistent with Dart_NativeFunctions, which will
automatically throw upon return if Dart_SetReturnValue set the result
to an error.
`UnhandledExceptions` cannot flow out into Dart generated code. So,
the implementation needs to be in `FfiCallInstr::EmitNativeCode`.
Using `Dart_IsError` is slow compared to a machine code class id
check. So, we should do the handle unwrapping and class id check in
machine code.
Unwrapping Handles in machine code is only safe when the GC is
guaranteed to not run: Either (1) in `kThreadInGenerated`, or (2) in
`kThreadInNative`, but only when transitioned into safepoint. So, the
handle cannot be unwrapped immediately after the FFI call in machine code. We first need to transition back to generated.
This means we need to transition again to native to do the actual
`Dart_PropagateError` call. We can do so without the stub in JIT
because we never return with normal control flow.
Performance impact of this change is within benchmark noise in both
JIT and AOT.
Size impact is 42 bytes on x64, which is 10% in AOT and 12% in JIT.
For more numbers see: go/dart-ffi-handle-error
TEST=runtime/bin/ffi_test/ffi_test_functions_vmspecific.cc
TEST=tests/ffi/vmspecific_handle_test.dart
Closes: https://github.com/dart-lang/sdk/issues/49936
Change-Id: Ie8fabeb6d53bc80689541bc4470cb37ee2200581
Cq-Include-Trybots: luci.dart.try:vm-canary-linux-debug-try,vm-ffi-android-debug-arm64c-try,vm-ffi-android-debug-arm-try,vm-kernel-gcc-linux-try,vm-kernel-linux-debug-x64-try,vm-kernel-linux-debug-x64c-try,vm-kernel-msvc-windows-try,vm-kernel-optcounter-threshold-linux-release-x64-try,vm-kernel-precomp-asan-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-reload-linux-debug-x64-try,vm-kernel-reload-rollback-linux-debug-x64-try,vm-kernel-tsan-linux-release-x64-try,vm-kernel-win-debug-x64-try,vm-kernel-win-debug-ia32-try,vm-precomp-ffi-qemu-linux-release-arm-try,vm-precomp-ffi-qemu-linux-release-riscv64-try,vm-kernel-linux-debug-ia32-try,vm-kernel-mac-release-arm64-try,vm-kernel-precomp-win-debug-x64c-try
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/261603
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Commit-Queue: Daco Harkes <dacoharkes@google.com>
Instead apply the same approach as we do in AOT: unbox based on the
static type information. There are no TFA results available in JIT,
but we could still unbox fields when running in sound null-safety.
TEST=ci
Cq-Include-Trybots: luci.dart.try:vm-kernel-reload-linux-release-x64-try,vm-kernel-reload-linux-debug-x64-try,vm-kernel-reload-rollback-linux-debug-x64-try,vm-kernel-reload-rollback-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-precomp-linux-product-x64-try,vm-kernel-precomp-linux-release-x64-try,vm-kernel-precomp-nnbd-linux-release-simarm64-try,vm-kernel-linux-debug-simriscv64-try,vm-kernel-precomp-linux-debug-simriscv64-try,vm-kernel-nnbd-linux-release-ia32-try,vm-kernel-nnbd-linux-debug-x64-try,vm-kernel-nnbd-linux-debug-ia32-try,vm-kernel-nnbd-linux-release-simarm-try,vm-kernel-nnbd-linux-release-simarm64-try
Change-Id: Ide2e78c6659261ef8d245a4586cf699ea0fbb459
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/256211
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Commit-Queue: Slava Egorov <vegorov@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
This change introduces separate stubs for suspending sync* functions
at start and at yield/yield*. Suspend stub for yield/yield*
no longer calls Dart callback (in order to make it faster).
Also, ReturnSyncStar stub is removed - sync* functions now directly
return false instead of going through the stub.
TEST=ci
Issue: https://github.com/dart-lang/sdk/issues/48378
Change-Id: Iee9a1f48cab2812cf0f9f0e4e6d8e847547e49f7
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/250420
Reviewed-by: Slava Egorov <vegorov@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Commit-Queue: Alexander Markov <alexmarkov@google.com>
The new implementation is based on suspend/resume stubs and doesn't
use desugaring of async functions on kernel AST.
Previously, new implementation of async/async* was only supported in
AOT mode. This change adds all necessary bits for the JIT mode:
* Suspending variable-length frames (for unoptimized code).
* Handling of Code and pool pointers in Dart stack frames.
* OSR.
* Deoptimization.
* Hot reload.
* Debugger.
The new implementation is not enabled in JIT mode yet.
Design doc: go/compact-async-await.
TEST=ci
Issue: https://github.com/dart-lang/sdk/issues/48378
Change-Id: I477d6684bdce7cbc1edb179ae2271ff598b7dcc5
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/246081
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Johnni Winther <johnniwinther@google.com>
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
The new implementation moves away from desugaring of async
functions on kernel AST, state machine generated in the flow graph and
capturing all local variables in the context.
Instead, async/await is implemented using a few stubs
(InitSuspendableFunction, Suspend, Resume, Return and
AsyncExceptionHandler). The stubs are implemented in a
platform-independent way using (macro-)assembler helpers.
When suspending a function, its frame is copied into a SuspendState
object, and when resuming a function it is copied back onto the stack.
No extra code is generated for accessing local variables.
Callback closures are created lazily on the first await.
Design doc: go/compact-async-await.
Part 1 (kernel): https://dart-review.googlesource.com/c/sdk/+/241842
TEST=ci
Issue: https://github.com/dart-lang/sdk/issues/48378
Change-Id: Ibad757035b7cc438ebdff80b460728b1d3eff1f5
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/242000
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
Right now all type literal usages will perform a runtime call which is
rather slow.
Flutter happens to use type literals such as `return T;` in hot code
which causes this to show up in the profile.
This CL adds fast paths for type parameter type literals if
the type parameter value (i.e. entry of TAV corresponding to T):
* is `null`: return `dynamic`
* is a non-FutureOr [Type] with compatible nullability: return value
* is [FunctionType] with compatible nullability: return value
otherwise fall back to runtime call.
It makes simple type literal uses 10x+ faster - the kinds that Flutter
is using.
Issue https://github.com/dart-lang/sdk/issues/48757
TEST=vm/dart{,_2}/instantiate_type_literal_test
Change-Id: I1139d6689aedbc68321f47ee6c9946a3323fbf6e
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/241968
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Commit-Queue: Martin Kustermann <kustermann@google.com>
- Fixes SIMD registers clobbered by write barrier on ARM64.
- Moves code generation out of non-compiler directory.
- Removes unnecessary building of Dart frames on leaf runtime calls.
- Removes unnecessary Threads slots for write barrier Code objects.
- Removes duplicate saves of SP in leaf runtime calls on ARM64 and RISC-V.
- Avoids some redundant SP updates on RISC-V.
TEST=ci
Change-Id: Idb92127658edc90b320923ef3d882a7219a450ae
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/236842
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
Commit-Queue: Ryan Macnak <rmacnak@google.com>
- refactor designated initializers as they are c++20
- unwrap #if/#endif in macro arguments
- remove taking pointer on purely intrinsic function (setjmp_)
- use DART_WARN_UNUSED_RESULT which is msvc/clang friendly
- provide StringRAII copying constructor
Bug: https://github.com/dart-lang/sdk/issues/48544
TEST=ci,msvc build
Change-Id: I1049589c96dc0cdf49e1d31d320b6804bf6e4558
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/237929
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Commit-Queue: Alexander Aprelev <aam@google.com>
Implements a backend targeting RV32GC and RV64GC, based on Linux standardizing around GC. The assembler is written to make it easy to disable usage of C, but because the sizes of some instruction sequences are compile-time constants, an additional build configuration would need to be defined to make use of it.
The assembler and disassembler cover every RV32/64GC instruction. The simulator covers all instructions except accessing CSRs and the floating point state accessible through such, include accrued exceptions and dynamic rounding mode.
Quirks:
- RISC-V is a compare-and-branch architecture, but some existing "architecture-independent" parts of the Dart compiler assume a condition code architecture. To avoid rewriting these parts, we use a peephole in the assembler to map to compare-and-branch. See Assembler::BranchIf. Luckily nothing depended on taking multiple branches on the same condition code set.
- There are no hardware overflow checks, so we must use Hacker's Delight style software checks. Often these are very cheap: if the sign of one operand is known, a single branch is needed.
- The ranges of RISC-V branches and jumps are such that we use 3 levels of generation for forward branches, instead of the 2 levels of near and far branches used on ARM[64]. Nearly all code is handled by the first two levels with 20-bits of range, with enormous regex matchers triggering the third level that uses aupic+jalr to get 32-bits of range.
- For PC-relative calls in AOT, we always generate auipc+jalr pairs with 32-bits of range, so we never generate trampolines.
- Only a subset of registers are available in some compressed instructions, so we assign the most popular uses to these registers. In particular, THR, TMP[2], CODE and PP. This has the effect of assigning CODE and PP to volatile registers in the C calling convention, whereas they are assigned preserved registers on the other architectures. As on ARM64, PP is untagged; this is so short indices can be accessed with a compressed instruction.
- There are no push or pop instructions, so combining pushes and pops is preferred so we can update SP once.
- The C calling convention has a strongly aligned stack, but unlike on ARM64 we don't need to use an alternate stack pointer. The author ensured language was added to the RISC-V psABI making the OS responsible for realigning the stack pointer for signal handlers, allowing Dart to leave the stack pointer misaligned from the C calling convention's point of view until a foreign call.
- We don't bother with the link register tracking done on ARM[64]. Instead we make use of an alternate link register to avoid inline spilling in the write barrier.
Unimplemented:
- non-trivial FFI cases
- Compressed pointers - No intention to implement.
- Unboxed SIMD - We might make use of the V extension registers when the V extension is ratified.
- BigInt intrinsics
TEST=existing tests for IL level, new tests for assembler/disassembler/simulator
Bug: https://github.com/dart-lang/sdk/issues/38587
Bug: https://github.com/dart-lang/sdk/issues/48164
Change-Id: I991d1df4be5bf55efec5371b767b332d37dfa3e0
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/217289
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Daco Harkes <dacoharkes@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
Commit-Queue: Ryan Macnak <rmacnak@google.com>
When ffi code calls Dart code that requests the isolate to exit(via Isolate.exit, for example), we ensure that isolate indeed exits on return from ffi call.
This is implemented by introduction of new safepoint bit which, once set, forces ExitSafepoint to propagate unwind error.
TEST=isolate_exit_sandwich_test
Change-Id: I2e8f5ecec7f4e59ae5f99b9525cc566f20d4b6a8
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/219846
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>
Commit-Queue: Alexander Aprelev <aam@google.com>
TSAN instruments C++ code by adding prologue/epilogue code which
maintains a shadow stack. Using setjmp()/longjmp() is intercepted by
TSAN and correspondingly unwinds the shadow stack.
When Dart VM throws exceptions we call the JumpToFrame stub
from C++ which will directly reset the stack to the exception handler
catch entry. This leaves the TSAN shadow stack unchanged.
This means whenever an exception is thrown we leak frames in TSAN's
shadow stack. Due to using a fixed-size shadow stack, it will cause a
buffer-overflow in TSAN when too many such frame leaks happen. This can
cause arbitrary memory to be overriden, leading to awkward crashes.
This is especially an issue on the "iso-stres" builder because it
launches - in the same process - *many* small tests, more easily hitting
that limit.
This CL will workaround the TSAN issue by making runtime call save it's
state via setjmp() and make exception throughing process go via
longjmp() (which TSAN will intercept) before actually calling the
JumpToFrame stub.
=> This will ensure the TSAN shadow stack is correctly maintained.
The [jmp_buf]'s encoding of register state is non-trivial (e.g. it uses
XOR'ing of the actual saved state under certain glibc versions). So we
store any state we need to pass to the target of the `longjmp()` on the
[Thread] instead of overriding the [jmp_buf]s register state with the
arguments.
Issue https://github.com/dart-lang/sdk/issues/47472#issuecomment-948235479
TEST=vm/dart{,_2}/regress47472_test.dart
Change-Id: Ifbf6580aa15bcce54d0584cdc3cd18cc19be0a9c
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/222300
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Commit-Queue: Martin Kustermann <kustermann@google.com>
Martin Kustermann