dart-bytecode, arm64: +4.742% geomean
dart-bytecode-jit-unopt, arm64: +12.73% geomean
dart2js-compile, x64: +3.635% geomean
In the polymorphic and unlinked cases, call to a stub the does a linear scan against an ICData.
In the monomorphic case, call to a prologue of the expected target function that checks the expected receiver class. There is additional indirection in the JIT version compared to the AOT version to also tick a usage counter so the inliner can make good decisions.
In the megamorphic case, call to a stub that does a hash table lookup against a MegamorphicCache.
Megamorphic call sites face a loss of precision in usage counts. The call site count is not recorded and the usage counter of the target function is used as an approximation.
Monomorphic and megamorphic calls sites are reset to the polymorphic/unlinked state on hot reload.
Monomorphic and megamorphic calls sites do not check the stepping state, so they are reset to the polymorphic/unlinked state when stepping begins and disabled.
Back-edges now increment the usage counter in addition to checking it. This ensures function with loops containing monomorphic calls will eventually cross the optimization threshold.
Fixed backwards use of kMonomorphicEntryOffset and kPolymorphicEntryOffset.
Fixed C stack overflow when bouncing between the KBC interpreter and a simulator.
Bug: https://github.com/dart-lang/sdk/issues/26780
Bug: https://github.com/dart-lang/sdk/issues/36409
Bug: https://github.com/dart-lang/sdk/issues/36731
Change-Id: I78a49cccd962703a459288e71ce246ed845df474
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/102820
Commit-Queue: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Rationale:
A basic peephole optimizer with window size one that avoids
redundant push-pop sequences already results is substantial
savings in code size and runtime, without any noticeable impact
on compile time (the peephole is very, very fast).
Performance:
Golem unoptimized code (which typically runs -90% compared to
optimized code, sees 5-60% improvements
https://github.com/dart-lang/sdk/issues/36409
Change-Id: I08db4b3dbc92377d89340a4969db6e664e54bceb
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/102980
Commit-Queue: Aart Bik <ajcbik@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Previously we would generate the prologue in FunctionEntryInstr or OsrEntryInstr
on X64/ARM, and in CompileGraph on ARM64/IA32/DBC. This caused a confusing asymmetry
in the code.
Now we generate it in FunctionEntryInstr/OsrEntryInstr on all backends.
Partially addresses dartbug.com/34162.
This also fixes a regression on a protobuf benchmark due to ICDatas not being saved.
Change-Id: Ie270d759234b197141025dde07a27546d221f6c8
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/101297
Commit-Queue: Samir Jindel <sjindel@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Previously these instructions were not safe to use on JIT, because they
crated catch-entry moves which are not used in JIT.
Now we assign deopt-infos to these slow paths so they can handle exceptions
correctly in JIT.
I considered removing uses of CheckNull in JIT mode, but I did not find a simple
replacement for all the existing uses.
Fixes#35887.
Change-Id: I05dfb46200679934d0ff6114cae8723f4908af8e
Reviewed-on: https://dart-review.googlesource.com/c/92783
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Commit-Queue: Samir Jindel <sjindel@google.com>
Auto-Submit: Samir Jindel <sjindel@google.com>
ARM write barrier goes from
ldrne lr, [thr, #+508]
blxne lr
to
blne <offset>
ARM64 write barrier goes from (similarly X64)
beq +8
ldrx lr, [thr, #1056]
blr lr
to
beq +4
bl <offset>
It reduces RX on arm/arm64 by around 0.9%
Though the write barrier wrappers stub has multiple entrypoints (one for
each available register). Because of this, we modify the relocation
logic to support per-call offsets into the target.
To avoid making the assembler code depend on StubCode/FlowGraphCompiler,
we set a closure, which the assembler can call.
Issue https://github.com/dart-lang/sdk/issues/33274
Change-Id: I9e3d68260cab7ef19ea88f1235c78d6031819d6d
Reviewed-on: https://dart-review.googlesource.com/c/90063
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
This CL improves AOT code for StackOverflowInstr/CheckNullInstr:
* On ARM we can do a conditional pc-relative calls for the stack overflow
checks, getting rid of the slow-paths entirely.
* On ARM64 we can do pc-relative calls on the slow path, avoiding an
extra load.
Flutter gallery size impact (in bare instructions mode):
* ARM: -3.7% RX
* ARM64: -1.4% RX
Issue https://github.com/dart-lang/sdk/issues/33274
Change-Id: Ia1acd76ac6efa26642f99e1ce3e417100aa357f3
Reviewed-on: https://dart-review.googlesource.com/c/89620
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
This is the final CL which adds a new --use-bare-instructions flag to
the VM.
If this flag is set during AOT compilation, we will:
* Build one global object pool (abbr: GOP) which all code objects
share. This gop will be stored in the object store. The PP register
is populated in the enter dart stub and it is restored when
returning from native calls.
* Gets rid of the CODE_REG/PP slots from the dart frames. Instead the
compiled code uses the global object pool, which is always in PP.
* Starts emitting pc-relative calls for calls between two dart
functions or when invoking a stub.
Limitation: We only emit pc-relative calls between two code objects
in the same isolate (this is because the image writer is writing
instruction objects for vm-isolate/main-isolate seperately)
* We do compile-time relocation of those static calls after the
precompiler has finished its work, but before writing the snapshot.
This patches all the instruction objects with pc-relative calls to
have the right .text distance.
* We emit a sorted list of code objects in ObjectStore::reverse_code_table,
which will be used by the AOT runtime to go back from PC to Code
objects (where all metadata, e.g. stack maps, catch entry moves, pc
descriptors are available).
Issue https://github.com/dart-lang/sdk/issues/33274
Change-Id: I6c5dd2b1571e3a889b27e804a24c2986c71e03b6
Reviewed-on: https://dart-review.googlesource.com/c/85769
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Similar to how we treat catch entry instructions, this cl adds new
function and osr entry instructions. The [FunctionEntry] and
[OsrEntry] - just like [CatchBlockEntry] - have now their own initial
definitions. The [GraphEntry] has only initial definitions for
constants.
Explicit phis are inserted for all parameter / special parameter
instructions if necessary.
Future work is:
a) Minimize parallel moves due to the phis on parameters
b) Cleanup frame setup: Move it entirely into FunctionEntry/CatchEntry
(instead of the split version we have now)
Fixes https://github.com/dart-lang/sdk/issues/34435
Fixes https://github.com/dart-lang/sdk/issues/34287
Change-Id: Iefa0280a709716f748d6fb0523b8d0f4d8de1fec
Reviewed-on: https://dart-review.googlesource.com/c/74782
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Previously we tried to rely on the assumption that all variables would be
boxed - so the machinery for setting correct catch-entry state only
supported tagged values and constants. However this both leads to worse code
and is not entirely correct assumption.
This also:
- renames various confusingly named classes: we move away from talking
about "catch entry state" to "catch entry moves" - because we only
record a subset of moves that needs to be performed and that does
not describe the whole state;
- refactors a bunch of associated code to be more readable and maintainable;
- adds documentation about catch implementation in optimized code
to runtime/docs/compiler;
Fixes https://github.com/flutter/flutter/issues/21685.
Change-Id: I03ae361a1bb7710acbd9f661ae014e663a163c59
Reviewed-on: https://dart-review.googlesource.com/74860
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Test Plan:
Behavioral correctness should be ensured by existing tests. Tests in vm/dart/entrypoints
ensure that the unchecked entrypoint is used in cases where the optimization should trigger.
Bug: https://github.com/dart-lang/sdk/issues/31798
Change-Id: I5b880b2dfa6343b4bb0a96ad23562facff73e41f
Cq-Include-Trybots: luci.dart.try:vm-kernel-win-release-x64-try,vm-kernel-optcounter-threshold-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-precomp-linux-release-simarm-try,vm-kernel-precomp-linux-release-simarm64-try,vm-kernel-precomp-linux-release-x64-try,vm-kernel-precomp-win-release-x64-try
Reviewed-on: https://dart-review.googlesource.com/69741
Commit-Queue: Samir Jindel <sjindel@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Test Plan:
Will be tested by following revisions which introduce optimizations that invoke the
new entrypoint.
Change-Id: I56d714bceb92c44a13c46d52457154b82577aa8c
Cq-Include-Trybots: luci.dart.try:vm-kernel-win-release-x64-try,vm-kernel-optcounter-threshold-linux-release-x64-try,vm-kernel-precomp-linux-debug-x64-try,vm-kernel-precomp-linux-release-simarm-try,vm-kernel-precomp-linux-release-simarm64-try,vm-kernel-precomp-linux-release-x64-try,vm-kernel-precomp-win-release-x64-try
Reviewed-on: https://dart-review.googlesource.com/67345
Commit-Queue: Samir Jindel <sjindel@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
After this change, ICData map is always unpacked in CompileParsedFunctionHelper::Compile
regardless of the optimizing mode. The unpacked ICData array is passed
both to FlowGraphBuilder and to FlowGraphCompiler.
FlowGraphCompiler::FlowGraphCompiler no longer unpacks ICData array, and
the duplicated code for unpacking of ICData map is replaced with debug asserts.
Change-Id: Ib6864acdb6bcd02fe7acd1cb5bcf70d6c69808af
Reviewed-on: https://dart-review.googlesource.com/67505
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
The stub code for calling into the runtime is now responsible for saving any
live registers. Since it doesn't know which registers are live (it's shared
across many safepoints), it saves all the registers. The stackmaps and
environments for safepoints which call this stub are extended to cover the stack
slots in the stub's frame which correspond to registers which are actually live
at the safepoint.
Currently this is only supported in precompiled X64 for the `CheckNull`
slow-path. By creating a separate stub for the CheckNull slow path we are able
to eliminate the size overhead of preparing arguments for the runtime call stub,
and reduce the slow-path size on X64 to 1 instruction.
# Test Plan
The stackmap for a safepoint in the `CheckNull` slowpath is only used when
handling the exception causes a GC, which is extremely unlikely. To induce this,
we add a flag to force this behavior in debug-mode and a test which exercises
this. The rest of the changes inherit existing coverage.
Change-Id: I3d8f2a7c3560c432599344d7420d42ddbee7a89d
Reviewed-on: https://dart-review.googlesource.com/59660
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Commit-Queue: Samir Jindel <sjindel@google.com>
So far the frontend (parser, flow graph builder, ssa construction) were
aware of the actual frame layout.
This CL makes the indices we assign to [LocalVariable]s logical
indices, assigning:
* M parameters the indices 1 ... M
* N local variables the indices 0 -1 ... -(N-1)
The scope building, flow graph builder and ssa construction operate on
those logical indices.
When emitting actual code, the backend will translate those indices into
actual FP relative indices. This allows us to be more flexible in the
backend which frame layout we choose.
Issue https://github.com/dart-lang/sdk/issues/33274
Change-Id: I9a504bf97821c257aafd2b3430df9f4c9da4b442
Reviewed-on: https://dart-review.googlesource.com/57321
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Relanding 4be50d6fa1 with fixes to DBC
and location summaries: AssertAssignable must save FPU registers.
For now we are limiting this to type checks against type parameter types.
In Dart 1 mode Dart2JS compiles itself in 28s when running from source
and in 23s when running from ideal app-jit snapshot (trained on the
same workload).
Before this change in Dart 2 mode numbers were 51s and 57s respectively.
After this change in Dart 2 mode numbers are 38s and 32s. Meaning
that regression is reduced by 50%.
Issue https://github.com/dart-lang/sdk/issues/31798
Issue https://github.com/dart-lang/sdk/issues/33257
Change-Id: Ifb55f86453bfdf36a2e03bcd7f3197cfde257103
Reviewed-on: https://dart-review.googlesource.com/57980
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
This reverts commit 4be50d6fa1.
Reason for revert: Failures on SIMDBC64 and Analyzer bots.
Original change's description:
> [vm] Enable type stubs based type checks in JIT mode for some types.
>
> For now we are limiting this to type checks against type parameter types.
>
> # Performance improvements
>
> In Dart 1 mode Dart2JS compiles itself in 28s when running from source
> and in 23s when running from ideal app-jit snapshot (trained on the
> same workload).
>
> Before this change in Dart 2 mode numbers were 51s and 57s respectively.
>
> After this change in Dart 2 mode numbers are 38s and 32s. Meaning
> that regression is reduced by 50%.
>
> Issue https://github.com/dart-lang/sdk/issues/31798
> Issue https://github.com/dart-lang/sdk/issues/33257
>
> Change-Id: I34bf5385a5cc3c7702dc281c6dfa89da85d3dde1
> Reviewed-on: https://dart-review.googlesource.com/57601
> Reviewed-by: Régis Crelier <regis@google.com>
> Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
TBR=vegorov@google.com,kustermann@google.com,regis@google.com
Change-Id: I85a30c962b0cd556310e19193f5993ab76ecf2e7
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://dart-review.googlesource.com/57840
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
For now we are limiting this to type checks against type parameter types.
# Performance improvements
In Dart 1 mode Dart2JS compiles itself in 28s when running from source
and in 23s when running from ideal app-jit snapshot (trained on the
same workload).
Before this change in Dart 2 mode numbers were 51s and 57s respectively.
After this change in Dart 2 mode numbers are 38s and 32s. Meaning
that regression is reduced by 50%.
Issue https://github.com/dart-lang/sdk/issues/31798
Issue https://github.com/dart-lang/sdk/issues/33257
Change-Id: I34bf5385a5cc3c7702dc281c6dfa89da85d3dde1
Reviewed-on: https://dart-review.googlesource.com/57601
Reviewed-by: Régis Crelier <regis@google.com>
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
This CL fixes error message in NoSuchMethodError thrown by null checks
and line numbers in corresponding stack traces (in AOT).
Name of the called function is placed into object pool, and metadata
for null check site is generated in CodeSourceMap.
Size of flutter gallery in release mode:
Before After
RW 2165286 2201642 (+1.68%)
RO 2122192 2168224 (+2.17%)
RX 6871072 6871072 (+0.00%)
Total 11163079 11245467 (+0.74%)
Closes https://github.com/dart-lang/sdk/issues/32863
Change-Id: I5ad1190f2ec9452a669863f7dd114ea5f9092d52
Reviewed-on: https://dart-review.googlesource.com/52703
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Relands 165c583d57
[VM] Introduction of type testing stubs - Part 1
This CL:
* Adds a field to [RawAbstractType] which will always hold a pointer
to the entrypoint of a type testing stub
* Makes this new field be initialized to a default stub whenever a
instances are created (e.g. via Type::New(), snapshot reader, ...)
* Makes the clustered snapshotter write a reference to the
corresponding [RawInstructions] object when writing the field and do
the reverse when reading it.
* Makes us call the type testing stub for performing assert-assignable
checks.
To reduce unnecessary loads on callsites, we store the entrypoint of the
type testing stubs directly in the type objects. This means that the
caller of type testing stubs can simply branch there without populating
a code object first. This also means that the type testing stubs
themselves have no access to a pool and we therefore also don't hold on
to the [Code] object, only the [Instruction] object is necessary.
The type testing stubs do not setup a frame themselves and also have no
safepoint. In the case when the type testing stubs could not determine
a positive answer they will tail-call a general-purpose stub.
The general-purpose stub sets up a stub frame, tries to consult a
[SubtypeTestCache] and bails out to runtime if this was unsuccessful.
This CL is just the the first, for ease of reviewing. The actual
type-specialized type testing stubs will be generated in later CLs.
Reviewed-on: https://dart-review.googlesource.com/44787
Relands f226c22424
[VM] Introduction of type testing stubs - Part 2
This CL starts building type testing stubs specialzed for [Type] objects
we test against.
More specifically, it adds support for:
* Handling obvious fast cases on the call sites (while still having a
call to stub for negative case)
* Handling type tests against type parameters, by loading the value
of the type parameter on the call sites and invoking it's type testing stub.
* Specialzed type testing stubs for instantiated types where we can
do [CidRange]-based subtype-checks.
==> e.g. String/List<dynamic>
* Specialzed type testing stubs for instantiated types where we can
do [CidRange]-based subclass-checks for the class and
[CidRange]-based subtype-checks for the type arguments.
==> e.g. Widget<State>, where we know [Widget] is only extended and not
implemented.
* Specialzed type testing stubs for certain non-instantiated types where we
can do [CidRange]-based subclass-checks for the class and
[CidRange]-based subtype-checks for the instantiated type arguments and
cid based comparisons for type parameters. (Note that this fast-case migth
result in some false-negatives!)
==> e.g. _HashMapEntry<K, V>, where we know [_HashMapEntry] is only
extended and not implemented.
This optimizes cases where the caller uses `new HashMap<A, B>()` and only
uses `A` and `B` as key/values (and not subclasses of it). The false-negative
can occur when subtypes of A or B are used. In such cases we fall back to the
[SubtypeTestCache]-based imlementation.
Reviewed-on: https://dart-review.googlesource.com/44788
Relands 25f98bcc75
[VM] Introduction of type testing stubs - Part 3
The changes include:
* Make AssertAssignableInstr no longer have a call-summary, which
helps methods with several parameter checks by not having to
re-load/re-initialize type arguments registers
* Lazily create SubtypeTestCaches: We already go to runtime to warm up
the caches, so we now also create the caches on the first runtime
call and patch the pool entries.
* No longer load the destination name into a register: We only need
the name when we throw an exception, so it is not on the hot path.
Instead we let the runtime look at the call site, decoding a pool
index from the instructions stream. The destination name will be
available in the pool, at a consecutive index to the subtype cache.
* Remove the fall-through to N=1 case for probing subtypeing tests,
since those will always be handled by the optimized stubs.
* Do not generate optimized stubs for FutureOr<T> (so far it just
falled-through to TTS). We can make optimzed version of that later,
but it requires special subtyping rules.
* Local code quality improvement in the type-testing-stubs: Avoid
extra jump at last case of cid-class-range checks.
There are still a number of optimization opportunities we can do in
future changes.
Reviewed-on: https://dart-review.googlesource.com/46984
Relands 2c52480ec8
[VM] Introduction of type testing stubs - Part 4
In order to avoid generating type testing stubs for too many types in
the system - and thereby potentially cause an increase in code size -
this change introduces a smarter way to decide for which types we should
generate optimized type testing stubs.
The precompiler creates a [TypeUsageInfo] which we use to collect
information. More specifically:
a) We collect the destination types for all type checks we emit
(we do this inside AssertAssignableInstr::EmitNativeCode).
-> These are types we might want to generate optimized type testing
stubs for.
b) We collect type argument vectors used in instance creations (we do
this inside AllocateObjectInstr::EmitNativeCode) and keep a set of
of used type argument vectors for each class.
After the precompiler has finished compiling normal code we scan the set
of destination types collected in a) for uninstantiated types (or more
specifically, type parameter types).
We then propagate the type argument vectors used on object allocation sites,
which were collected in b), in order to find out what kind of types are flowing
into those type parameters.
This allows us to extend the set of types which we test against, by
adding the types that flow into type parameters.
We use this final augmented set of destination types as a "filter" when
making the decision whether to generate an optimized type testing stub
for a given type.
Reviewed-on: https://dart-review.googlesource.com/48640
Issue https://github.com/dart-lang/sdk/issues/32603
Closes https://github.com/dart-lang/sdk/issues/32852
Change-Id: Ib79fbe7f043aa88f32bddad62d7656c638914b44
Reviewed-on: https://dart-review.googlesource.com/50944
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
We add two things:
* --print_instruction_stats makes compiler dump per IL instruction size
breakdown (how many bytes of code were produced from specific instruction
kinds). This was largely implemented by kustermann@ in
https://codereview.chromium.org/2584613002/ and this CL does only few changes
to the original implementation, namely more uniform handling of slow-path code
and puts statistics object into RawInstructions (which has free space due to
alignment) instead of RawCode.
* --print_instructions_sizes_to=symbols.json makes compiler dump per Instruction
object size breakdown into a JSON file. This JSON file can later be processed
with pkg/vm/tool/run_binary_size_analysis.dart script to produce interactive
binary size diagram similar to runtime/third_party/binary_size tool.
Change-Id: Ied4965b9a0a91b3025eefbe981ecd47cdcf782d6
Reviewed-on: https://dart-review.googlesource.com/50501
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Relands 165c583d57
[VM] Introduction of type testing stubs - Part 1
This CL:
* Adds a field to [RawAbstractType] which will always hold a pointer
to the entrypoint of a type testing stub
* Makes this new field be initialized to a default stub whenever a
instances are created (e.g. via Type::New(), snapshot reader, ...)
* Makes the clustered snapshotter write a reference to the
corresponding [RawInstructions] object when writing the field and do
the reverse when reading it.
* Makes us call the type testing stub for performing assert-assignable
checks.
To reduce unnecessary loads on callsites, we store the entrypoint of the
type testing stubs directly in the type objects. This means that the
caller of type testing stubs can simply branch there without populating
a code object first. This also means that the type testing stubs
themselves have no access to a pool and we therefore also don't hold on
to the [Code] object, only the [Instruction] object is necessary.
The type testing stubs do not setup a frame themselves and also have no
safepoint. In the case when the type testing stubs could not determine
a positive answer they will tail-call a general-purpose stub.
The general-purpose stub sets up a stub frame, tries to consult a
[SubtypeTestCache] and bails out to runtime if this was unsuccessful.
This CL is just the the first, for ease of reviewing. The actual
type-specialized type testing stubs will be generated in later CLs.
Reviewed-on: https://dart-review.googlesource.com/44787
Relands f226c22424
[VM] Introduction of type testing stubs - Part 2
This CL starts building type testing stubs specialzed for [Type] objects
we test against.
More specifically, it adds support for:
* Handling obvious fast cases on the call sites (while still having a
call to stub for negative case)
* Handling type tests against type parameters, by loading the value
of the type parameter on the call sites and invoking it's type testing stub.
* Specialzed type testing stubs for instantiated types where we can
do [CidRange]-based subtype-checks.
==> e.g. String/List<dynamic>
* Specialzed type testing stubs for instantiated types where we can
do [CidRange]-based subclass-checks for the class and
[CidRange]-based subtype-checks for the type arguments.
==> e.g. Widget<State>, where we know [Widget] is only extended and not
implemented.
* Specialzed type testing stubs for certain non-instantiated types where we
can do [CidRange]-based subclass-checks for the class and
[CidRange]-based subtype-checks for the instantiated type arguments and
cid based comparisons for type parameters. (Note that this fast-case migth
result in some false-negatives!)
==> e.g. _HashMapEntry<K, V>, where we know [_HashMapEntry] is only
extended and not implemented.
This optimizes cases where the caller uses `new HashMap<A, B>()` and only
uses `A` and `B` as key/values (and not subclasses of it). The false-negative
can occur when subtypes of A or B are used. In such cases we fall back to the
[SubtypeTestCache]-based imlementation.
Reviewed-on: https://dart-review.googlesource.com/44788
Relands 25f98bcc75
[VM] Introduction of type testing stubs - Part 3
The changes include:
* Make AssertAssignableInstr no longer have a call-summary, which
helps methods with several parameter checks by not having to
re-load/re-initialize type arguments registers
* Lazily create SubtypeTestCaches: We already go to runtime to warm up
the caches, so we now also create the caches on the first runtime
call and patch the pool entries.
* No longer load the destination name into a register: We only need
the name when we throw an exception, so it is not on the hot path.
Instead we let the runtime look at the call site, decoding a pool
index from the instructions stream. The destination name will be
available in the pool, at a consecutive index to the subtype cache.
* Remove the fall-through to N=1 case for probing subtypeing tests,
since those will always be handled by the optimized stubs.
* Do not generate optimized stubs for FutureOr<T> (so far it just
falled-through to TTS). We can make optimzed version of that later,
but it requires special subtyping rules.
* Local code quality improvement in the type-testing-stubs: Avoid
extra jump at last case of cid-class-range checks.
There are still a number of optimization opportunities we can do in
future changes.
Reviewed-on: https://dart-review.googlesource.com/46984
Relands 2c52480ec8
[VM] Introduction of type testing stubs - Part 4
In order to avoid generating type testing stubs for too many types in
the system - and thereby potentially cause an increase in code size -
this change introduces a smarter way to decide for which types we should
generate optimized type testing stubs.
The precompiler creates a [TypeUsageInfo] which we use to collect
information. More specifically:
a) We collect the destination types for all type checks we emit
(we do this inside AssertAssignableInstr::EmitNativeCode).
-> These are types we might want to generate optimized type testing
stubs for.
b) We collect type argument vectors used in instance creations (we do
this inside AllocateObjectInstr::EmitNativeCode) and keep a set of
of used type argument vectors for each class.
After the precompiler has finished compiling normal code we scan the set
of destination types collected in a) for uninstantiated types (or more
specifically, type parameter types).
We then propagate the type argument vectors used on object allocation sites,
which were collected in b), in order to find out what kind of types are flowing
into those type parameters.
This allows us to extend the set of types which we test against, by
adding the types that flow into type parameters.
We use this final augmented set of destination types as a "filter" when
making the decision whether to generate an optimized type testing stub
for a given type.
Reviewed-on: https://dart-review.googlesource.com/48640
Issue https://github.com/dart-lang/sdk/issues/32603
Change-Id: I6d33d4ca3d5187a1eb1664078c003061855f0160
Reviewed-on: https://dart-review.googlesource.com/50482
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Commit-Queue: Martin Kustermann <kustermann@google.com>
Relands 165c583d57
[VM] Introduction of type testing stubs - Part 1
This CL:
* Adds a field to [RawAbstractType] which will always hold a pointer
to the entrypoint of a type testing stub
* Makes this new field be initialized to a default stub whenever a
instances are created (e.g. via Type::New(), snapshot reader, ...)
* Makes the clustered snapshotter write a reference to the
corresponding [RawInstructions] object when writing the field and do
the reverse when reading it.
* Makes us call the type testing stub for performing assert-assignable
checks.
To reduce unnecessary loads on callsites, we store the entrypoint of the
type testing stubs directly in the type objects. This means that the
caller of type testing stubs can simply branch there without populating
a code object first. This also means that the type testing stubs
themselves have no access to a pool and we therefore also don't hold on
to the [Code] object, only the [Instruction] object is necessary.
The type testing stubs do not setup a frame themselves and also have no
safepoint. In the case when the type testing stubs could not determine
a positive answer they will tail-call a general-purpose stub.
The general-purpose stub sets up a stub frame, tries to consult a
[SubtypeTestCache] and bails out to runtime if this was unsuccessful.
This CL is just the the first, for ease of reviewing. The actual
type-specialized type testing stubs will be generated in later CLs.
Reviewed-on: https://dart-review.googlesource.com/44787
Relands f226c22424
[VM] Introduction of type testing stubs - Part 2
This CL starts building type testing stubs specialzed for [Type] objects
we test against.
More specifically, it adds support for:
* Handling obvious fast cases on the call sites (while still having a
call to stub for negative case)
* Handling type tests against type parameters, by loading the value
of the type parameter on the call sites and invoking it's type testing stub.
* Specialzed type testing stubs for instantiated types where we can
do [CidRange]-based subtype-checks.
==> e.g. String/List<dynamic>
* Specialzed type testing stubs for instantiated types where we can
do [CidRange]-based subclass-checks for the class and
[CidRange]-based subtype-checks for the type arguments.
==> e.g. Widget<State>, where we know [Widget] is only extended and not
implemented.
* Specialzed type testing stubs for certain non-instantiated types where we
can do [CidRange]-based subclass-checks for the class and
[CidRange]-based subtype-checks for the instantiated type arguments and
cid based comparisons for type parameters. (Note that this fast-case migth
result in some false-negatives!)
==> e.g. _HashMapEntry<K, V>, where we know [_HashMapEntry] is only
extended and not implemented.
This optimizes cases where the caller uses `new HashMap<A, B>()` and only
uses `A` and `B` as key/values (and not subclasses of it). The false-negative
can occur when subtypes of A or B are used. In such cases we fall back to the
[SubtypeTestCache]-based imlementation.
Reviewed-on: https://dart-review.googlesource.com/44788
Relands 25f98bcc75
[VM] Introduction of type testing stubs - Part 3
The changes include:
* Make AssertAssignableInstr no longer have a call-summary, which
helps methods with several parameter checks by not having to
re-load/re-initialize type arguments registers
* Lazily create SubtypeTestCaches: We already go to runtime to warm up
the caches, so we now also create the caches on the first runtime
call and patch the pool entries.
* No longer load the destination name into a register: We only need
the name when we throw an exception, so it is not on the hot path.
Instead we let the runtime look at the call site, decoding a pool
index from the instructions stream. The destination name will be
available in the pool, at a consecutive index to the subtype cache.
* Remove the fall-through to N=1 case for probing subtypeing tests,
since those will always be handled by the optimized stubs.
* Do not generate optimized stubs for FutureOr<T> (so far it just
falled-through to TTS). We can make optimzed version of that later,
but it requires special subtyping rules.
* Local code quality improvement in the type-testing-stubs: Avoid
extra jump at last case of cid-class-range checks.
There are still a number of optimization opportunities we can do in
future changes.
Reviewed-on: https://dart-review.googlesource.com/46984
Relands 2c52480ec8
[VM] Introduction of type testing stubs - Part 4
In order to avoid generating type testing stubs for too many types in
the system - and thereby potentially cause an increase in code size -
this change introduces a smarter way to decide for which types we should
generate optimized type testing stubs.
The precompiler creates a [TypeUsageInfo] which we use to collect
information. More specifically:
a) We collect the destination types for all type checks we emit
(we do this inside AssertAssignableInstr::EmitNativeCode).
-> These are types we might want to generate optimized type testing
stubs for.
b) We collect type argument vectors used in instance creations (we do
this inside AllocateObjectInstr::EmitNativeCode) and keep a set of
of used type argument vectors for each class.
After the precompiler has finished compiling normal code we scan the set
of destination types collected in a) for uninstantiated types (or more
specifically, type parameter types).
We then propagate the type argument vectors used on object allocation sites,
which were collected in b), in order to find out what kind of types are flowing
into those type parameters.
This allows us to extend the set of types which we test against, by
adding the types that flow into type parameters.
We use this final augmented set of destination types as a "filter" when
making the decision whether to generate an optimized type testing stub
for a given type.
Reviewed-on: https://dart-review.googlesource.com/48640
Issue https://github.com/dart-lang/sdk/issues/32603
Change-Id: I44a1d5d4b27454ae026aef2a301aada3dd399ea0
Reviewed-on: https://dart-review.googlesource.com/49861
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
In order to avoid generating type testing stubs for too many types in
the system - and thereby potentially cause an increase in code size -
this change introduces a smarter way to decide for which types we should
generate optimized type testing stubs.
The precompiler creates a [TypeUsageInfo] which we use to collect
information. More specifically:
a) We collect the destination types for all type checks we emit
(we do this inside AssertAssignableInstr::EmitNativeCode).
-> These are types we might want to generate optimized type testing
stubs for.
b) We collect type argument vectors used in instance creations (we do
this inside AllocateObjectInstr::EmitNativeCode) and keep a set of
of used type argument vectors for each class.
After the precompiler has finished compiling normal code we scan the set
of destination types collected in a) for uninstantiated types (or more
specifically, type parameter types).
We then propagate the type argument vectors used on object allocation sites,
which were collected in b), in order to find out what kind of types are flowing
into those type parameters.
This allows us to extend the set of types which we test against, by
adding the types that flow into type parameters.
We use this final augmented set of destination types as a "filter" when
making the decision whether to generate an optimized type testing stub
for a given type.
Issue https://github.com/dart-lang/sdk/issues/32603
Measured impact on flutter HEAD-HEAD-HEAD with TTS Part 1 - 4 applied (2018-04-03):
* stock build benchmark: around 4% improvement
* gallery app.so size: -2.68% (13987348 -> 13612928)
* gallery memory: no sigificant changes:
- SubtypeTestCache: - 10kb
- ObjectPool: + 6 kb
- Type: no change (probably due to wasted alignment slot before)
- TypeParameter: + 4 kb (can get rid of the field here later)
* gallery AOT compile-time: measured +1.3%, inside flakiness range
Change-Id: I12a398d18f970ba2db741913bb47b0f36ae38d58
Reviewed-on: https://dart-review.googlesource.com/48640
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
The changes include:
* Make AssertAssignableInstr no longer have a call-summary, which
helps methods with several parameter checks by not having to
re-load/re-initialize type arguments registers
* Lazily create SubtypeTestCaches: We already go to runtime to warm up
the caches, so we now also create the caches on the first runtime
call and patch the pool entries.
* No longer load the destination name into a register: We only need
the name when we throw an exception, so it is not on the hot path.
Instead we let the runtime look at the call site, decoding a pool
index from the instructions stream. The destination name will be
available in the pool, at a consecutive index to the subtype cache.
* Remove the fall-through to N=1 case for probing subtypeing tests,
since those will always be handled by the optimized stubs.
* Do not generate optimized stubs for FutureOr<T> (so far it just
falled-through to TTS). We can make optimzed version of that later,
but it requires special subtyping rules.
* Local code quality improvement in the type-testing-stubs: Avoid
extra jump at last case of cid-class-range checks.
There are still a number of optimization opportunities we can do in
future changes.
Issue https://github.com/dart-lang/sdk/issues/31798
Change-Id: I4dc5a8a49f939178fe74d44736ef69e4b9088e46
Reviewed-on: https://dart-review.googlesource.com/46984
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
This CL:
* Adds a field to [RawAbstractType] which will always hold a pointer
to the entrypoint of a type testing stub
* Makes this new field be initialized to a default stub whenever a
instances are created (e.g. via Type::New(), snapshot reader, ...)
* Makes the clustered snapshotter write a reference to the
corresponding [RawInstructions] object when writing the field and do
the reverse when reading it.
* Makes us call the type testing stub for performing assert-assignable
checks.
To reduce unnecessary loads on callsites, we store the entrypoint of the
type testing stubs directly in the type objects. This means that the
caller of type testing stubs can simply branch there without populating
a code object first. This also means that the type testing stubs
themselves have no access to a pool and we therefore also don't hold on
to the [Code] object, only the [Instruction] object is necessary.
The type testing stubs do not setup a frame themselves and also have no
safepoint. In the case when the type testing stubs could not determine
a positive answer they will tail-call a general-purpose stub.
The general-purpose stub sets up a stub frame, tries to consult a
[SubtypeTestCache] and bails out to runtime if this was unsuccessful.
This CL is just the the first, for ease of reviewing. The actual
type-specialized type testing stubs will be generated in later CLs.
Issue https://github.com/dart-lang/sdk/issues/31798
Change-Id: I174a11b3b812799f399a60af799144c2ba3c26ec
Reviewed-on: https://dart-review.googlesource.com/44787
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
When the type to test against is instantiated and has no type arguments
there is a high probability that we receive instances of that class or
subclasses at runtime.
This CL therefore extends the fast-path of AssertAssignable/InstanceOf
by checking whether the instance class id is within the cid ranges that
directly/indirectly implement/extend the type to test against.
Currently we have an almost depth-first preorder numbering of class ids
in AOT, but there are exceptions. So each class can have a number of
cid-ranges as subclasses / classes which implement it's interface.
This seems to improve performance of dart-aot-v2
* flutter stock build by 15+%
* DeltaBlueClosures by 10+%
and reduces code size on
* flutter gallery by -3%
Issue https://github.com/dart-lang/sdk/issues/31798
Change-Id: I07dd91589cc3fcd8c5952bdba339e2e2a459e08e
Reviewed-on: https://dart-review.googlesource.com/35620
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
This CL extends support of mixed operations in strong-mode AOT
from double/smi to double/int (if underlying code generator supports
conversions from int64 to double).
* MintToDouble instruction is renamed to Int64ToDouble.
* Code generation of Int64ToDouble is implemented on x64.
Issue: https://github.com/dart-lang/sdk/issues/30480
Change-Id: I3142302e80c2785be71d7a8baff42d40f0b5b1b8
Reviewed-on: https://dart-review.googlesource.com/35382
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Ryan Macnak