Size of a large application:
Before: 26628600
After: 21480120 (-19.3%)
Size of snapshots:
isolate_snapshot_framework.bin
Before: 9322496
After: 6782976 (-27.2%)
isolate_snapshot_product_framework.bin
Before: 9166848
After: 6602752 (-27.9%)
Regressions in tests:
1) Test language_2/type_alias_equality_test/04 fails similarly to default mode, as VM
does not implement comparison of function types according to the specification.
Previously this test was passing as function types were canonicalized in bytecode,
which was not always correct. This CL fixes the problem with canonicalization of
function types in bytecode and the test starts failing again.
2) Tests standalone_2/entrypoints_verification_test, standalone_2/io/test_extension_test,
standalone_2/io/test_extension_fail_test fail as native extensions are not supported
in bytecode yet. These tests start passing after df5e7aac17,
which switched bytecode tests to kernel service (on x64), because kernel service doesn't
drop ASTs. This CL switches from reading AST library declarations to bytecode even if
AST is not removed, so tests fail again.
Change-Id: I8b7ba44bfa49d0b1599b2509553ff7c831a4e244
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/104700
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
Commit-Queue: Alexander Markov <alexmarkov@google.com>
This change replaces kernel AST declarations of fields and functions
with bytecode declarations.
Size of dilp files is reduced by 11-12%.
Startup latency:
Time to the first full frame: 1.945s -> 1.687s
FinalizeClass: 554ms -> 277ms
FinishClassLoading: 296ms -> 156ms
There are following regressions in bytecode mode, which will be fixed
in future:
* dart:mirrors are not supported yet (implementation of mirrors relies
on reading kernel AST in certain cases).
As the result, lib_2/mirrors/* tests fail.
* native extensions are not supported yet (annotations on libraries
and classes in AST are cleaned up as they could reference members
which are now removed from AST).
As the result, standalone_2/entrypoints_verification_test test fails.
* language_2/spread_collections/const_error_test/* tests fail
due to https://github.com/dart-lang/sdk/issues/36286.
Change-Id: I5130f401fd7b84038b136136e7ccc1a6e51b6cea
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/97561
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Before this change we allowed to serialize some references as null,
and would only complain when trying to read the dill back in.
This CL changes that so it complains up front.
It is unknown if any of this would ever occur in practise.
Change-Id: Id5e1af89abfb88a2d4249bd439b53b062d7aeffa
Reviewed-on: https://dart-review.googlesource.com/c/85392
Reviewed-by: Kevin Millikin <kmillikin@google.com>
Commit-Queue: Jens Johansen <jensj@google.com>
This is a major revamp of bytecode metadata format. Now bytecode has
its own serialization mechanisms.
This CL adds 'bytecode component' metadata, which contains bytecode
object table and string table. All references from bytecode (constant
pools) to libraries, classes, members, types and strings now have a new
format. References to frequently used objects are represented as indices
in object table, while rarely used objects are written inline.
This allows VM to cache frequently used objects while reading bytecode.
Representation of strings is aligned with VM - string characters are
stored in separate pools of one-byte and two-byte strings. This allows
VM to avoid UTF-8 decoding and extra copying.
Closure declarations are now explicit. Type parameters no longer require
enslosing scopes when reading/writing them.
Benchmarks:
GenKernelKernelReadAllBytecode (Intel Core i5) +29.84%
GenKernelKernelReadAllBytecode (Intel Xeon) +28.74%
Change-Id: I4b80009733a8f8c038264af74f97c4e094b9e311
Reviewed-on: https://dart-review.googlesource.com/c/85469
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
Reviewed-by: Zach Anderson <zra@google.com>
If a field is initialized with null (either explicitly or implicitly),
field store can be omitted in bytecode. In such case, bytecode should
still convey the information about this initialization to VM for
field guards to work correctly.
Change-Id: I1fd45b858c3c521b97fa5dbffe0e15b1ea75d92f
Reviewed-on: https://dart-review.googlesource.com/76060
Reviewed-by: Zach Anderson <zra@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
This change improves type flow analysis to infer int types, in addition
to concrete classes (int type is abstract and has 2 concrete subclasses,
_Smi and _Mint).
Also, this CL enables devirtualization of dynamic calls (previously, only
calls with known interface target were devirtualized).
DartMicroBench.PrimeNumber:
x64: +221%
armv8: +461%
armv7h: +47.98%
No measurable impact on Flutter gallery build time in release mode.
Issue: https://github.com/flutter/flutter/issues/19677
Change-Id: I34db195f52f2a434218ee11eee7f308d4661738b
Reviewed-on: https://dart-review.googlesource.com/69520
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Metadata is no longer written ahead of all nodes. Instead, metadata for
each node is written in the same context as the node itself (into a separate
buffer). This allows metadata to contain (serialize) arbitrary nodes
(for example, arbitrary DartTypes) and use serialization context of parent
nodes (such as declared type parameters).
However, with this change metadata looses the ability to reference
arbitrary AST nodes. This ability was overly restricted and had no
practical uses. (It was not possible to reference nodes which are not
reachable from root Component. As a consequence, it was not possible to
write references to arbitrary DartTypes.)
This change aligns the serialization capabilities of metadata with
how kernel AST nodes are serialized.
Change-Id: I027299a33b599b62572eccd4aa7083ad1dd2b3b3
Reviewed-on: https://dart-review.googlesource.com/54481
Commit-Queue: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Jens Johansen <jensj@google.com>
To enable bytecode generation, toggle kEnableKernelBytecode flag at
the beginning of pkg/vm/lib/bytecode/gen_bytecode.dart.
This will also enable generation of bytecode in platform dill files.
Also, bytecode generation can be enabled using --gen-bytecode option of
pkg/vm/tool/gen_kernel.
In kernel binaries, the generated bytecode and constant pool are
placed into 'vm.bytecode' metadata attached to members.
pkg/vm/tool/dump_kernel tool can be used to disassemble
bytecode and print constant pool.
Differences between generated bytecode and original DBC are
described in pkg/vm/lib/bytecode/dbc.dart.
Format of constant pool is described in
pkg/vm/lib/bytecode/constant_pool.dart.
Currently, only a small subset of Dart language is supported.
Bytecode generator will not emit vm.bytecode metadata for a
member if its body has an unsupported operation.
Change-Id: I27d9a361dc779ea115e5676508ee757c1754e05d
Reviewed-on: https://dart-review.googlesource.com/49600
Reviewed-by: Samir Jindel <sjindel@google.com>
Reviewed-by: Martin Kustermann <kustermann@google.com>
Reviewed-by: Zach Anderson <zra@google.com>
Reviewed-by: Régis Crelier <regis@google.com>
Commit-Queue: Alexander Markov <alexmarkov@google.com>
This relands 75a9579ea0.
The approach works as follows:
Step 1: Kernel transform. Under the closed world assumption compute the
set of selectors dispatched dynamically, then mark all procedures that don't
match any of those selectors as 'not-dispatched-dynamically'.
Step 2: VM backend. When building IR for a function if this function was
marked as not-dispatched-dynamically then omit type checks for any parameter
that is not marked as generic-covariant-impl, as such arguments are guaranteed
to be checked on the caller side (by front-end).
+------------------------+------------+----------+-----------+--------------+
| benchmark | baseline | current | with opt | improved by |
+------------------------+------------+----------+-----------+--------------+
| stock_layout_iteration | 2366.3786 | 2724.3 | 2562.75 | -5.93% |
| stock_build_iteration | 3824.3 | 4914.8 | 4681 | -4.76% |
+------------------------+------------+----------+-----------+--------------+
* Flutter gallery Instructions size is reduced by 11% (8748720 bytes to 7846368 bytes).
Baseline is at 6196496 bytes.
Alternatively to annotating individual procedures, I considered annotating Program node
with a set of dynamically dispatched selectors. Decoding and passing this information
around proved to be quite cumbersome in the "streaming" world, so I opted for a simpler
approach where all individual procedures are annotated.
Bug: https://github.com/dart-lang/sdk/issues/3179
Change-Id: I2f32a609e3872c74d5ae7bbd97555453aaedf15f
Reviewed-on: https://dart-review.googlesource.com/38125
Commit-Queue: Samir Jindel <sjindel@google.com>
Reviewed-by: Samir Jindel <sjindel@google.com>
This reverts commit 75a9579ea0.
Reason for revert: function_subtype_bound_closure7_test fails in darkp
Original change's description:
> [vm/kernel/aot] Skip unnecessary type checks on parameters of instance methods.
>
> The approach works as follows:
>
> Step 1: Kernel transform. Under the closed world assumption compute the
> set of selectors dispatched dynamically, then mark all procedures that don't
> match any of those selectors as 'not-dispatched-dynamically'.
>
> Step 2: VM backend. When building IR for a function if this function was
> marked as not-dispatched-dynamically then omit type checks for any parameter
> that is not marked as generic-covariant-impl, as such arguments are guaranteed
> to be checked on the caller side (by front-end).
>
> # Performance Impact
>
> +------------------------+------------+----------+-----------+--------------+
> | benchmark | baseline | current | with opt | improved by |
> +------------------------+------------+----------+-----------+--------------+
> | stock_layout_iteration | 2366.3786 | 2724.3 | 2562.75 | -5.93% |
> | stock_build_iteration | 3824.3 | 4914.8 | 4681 | -4.76% |
> +------------------------+------------+----------+-----------+--------------+
>
> * Flutter gallery Instructions size is reduced by 11% (8748720 bytes to 7846368 bytes).
> Baseline is at 6196496 bytes.
>
> # Alternative Implementations
>
> Alternatively to annotating individual procedures, I considered annotating Program node
> with a set of dynamically dispatched selectors. Decoding and passing this information
> around proved to be quite cumbersome in the "streaming" world, so I opted for a simpler
> approach where all individual procedures are annotated.
>
> Change-Id: I363db6d5dd1138fe25917646313c16d0b213c3b4
> Reviewed-on: https://dart-review.googlesource.com/37822
> Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
> Reviewed-by: Alexander Markov <alexmarkov@google.com>
> Reviewed-by: Samir Jindel <sjindel@google.com>
TBR=vegorov@google.com,alexmarkov@google.com,sjindel@google.com
Change-Id: I0fd6bf3e6edfc3e605da2b996f9a0da6c409d01c
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://dart-review.googlesource.com/37802
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
The approach works as follows:
Step 1: Kernel transform. Under the closed world assumption compute the
set of selectors dispatched dynamically, then mark all procedures that don't
match any of those selectors as 'not-dispatched-dynamically'.
Step 2: VM backend. When building IR for a function if this function was
marked as not-dispatched-dynamically then omit type checks for any parameter
that is not marked as generic-covariant-impl, as such arguments are guaranteed
to be checked on the caller side (by front-end).
# Performance Impact
+------------------------+------------+----------+-----------+--------------+
| benchmark | baseline | current | with opt | improved by |
+------------------------+------------+----------+-----------+--------------+
| stock_layout_iteration | 2366.3786 | 2724.3 | 2562.75 | -5.93% |
| stock_build_iteration | 3824.3 | 4914.8 | 4681 | -4.76% |
+------------------------+------------+----------+-----------+--------------+
* Flutter gallery Instructions size is reduced by 11% (8748720 bytes to 7846368 bytes).
Baseline is at 6196496 bytes.
# Alternative Implementations
Alternatively to annotating individual procedures, I considered annotating Program node
with a set of dynamically dispatched selectors. Decoding and passing this information
around proved to be quite cumbersome in the "streaming" world, so I opted for a simpler
approach where all individual procedures are annotated.
Change-Id: I363db6d5dd1138fe25917646313c16d0b213c3b4
Reviewed-on: https://dart-review.googlesource.com/37822
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
Reviewed-by: Samir Jindel <sjindel@google.com>
* The package:kernel/transformations/precompiler.dart is moved to
package:vm and split into 2 libraries: metadata/direct_call and
transformations/cha_devrtualization.
* Fasta 'target-options' command line argument and
package:kernel/target/implementation_option.dart are cleaned up
as they are no longer used.
Issue: https://github.com/dart-lang/sdk/issues/30480
Change-Id: I21a7b6bc62a036a9694e62a5dae890f6120d79ab
Reviewed-on: https://dart-review.googlesource.com/26360
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Alexander Markov