dart-sdk/pkg/vm_snapshot_analysis

vm_snapshot_analysis

This package provides libraries and a utility for analysing the size and contents of Dart VM AOT snapshots based on the output of --print-instructions-sizes-to, --write-v8-snapshot-profile-to and --trace-precompiler-to VM flags.

AOT Snapshot Basics

Dart VM AOT snapshot is simply a serialized representation of the Dart VM heap graph. It consists of two parts: data (e.g. strings, const instances, objects representing classes, libraries, functions and runtime metadata) and executable code (machine code generated from Dart sources). Some nodes in this graph have clean and direct relationship to the original program (e.g. objects representing libraries, classes, functions), while other nodes don't. Bitwise equivalent objects can be deduplicated and shared (e.g. two functions with the same body will end up using the same machine code object). This makes impossible to attribute of every single byte from the snapshot to a particular place in the program with a 100% accuracy.

• --print-instructions-sizes-to attributes executable code from the snapshot to a particular Dart function (or internal stub) from which this code originated (ignoring deduplication). Executable code usually constitutes around half of the snapshot, those this varies depending on the application.
• --write-v8-snapshot-profile-to is a graph representation of the snapshot, it attributes bytes written into a snapshot to a node in the heap graph. This format covers both data and code sections of the snapshot.
• --trace-precompiler-to gives information about dependencies between compiled functions, allowing to determine why certain function was pulled into the snapshot.

Passing flags to the AOT compiler

Both in dart2native and Flutter you can use --extra-gen-snapshot-options to pass flags to the AOT compiler:

$ flutter build aot --release --extra-gen-snapshot-options=--write-v8-snapshot-profile-to=profile.json

$ dart2native --extra-gen-snapshot-options=--write-v8-snapshot-profile-to=profile.json -o binary input.dart

Similarly with --print-instructions-sizes-to.

If you are working on the Dart SDK you can use the pkg/vm/tool/precompiler2 script, in which case you can just pass these flags directly:

$ pkg/vm/tool/precompiler2 --write-v8-snapshot-profile-to=profile.json input.dart binary

CLI

The command line interface to the tools in this package is provided by a single entry point bin/analyse.dart. It consumes output of --print-instructions-sizes-to and --write-v8-snapshot-profile-to flags and presents it in different human readable ways.

This script can be installed globally as snapshot_analysis using

$ pub global activate vm_snapshot_analysis

snapshot_analysis supports the following subcommands:

summary

$ snapshot_analysis summary [-b granularity] [-w filter] <input-profile.json>

This command shows breakdown of snapshot bytes at the given granularity (e.g. method, class, library or package), filtered by the given substring filter.

For example, here is a output showing how many bytes from a snapshot can be attributed to classes in the dart:core library:

$ pkg/vm/bin/snapshot_analysis.dart summary -b class -w dart:core profile.json
+-----------+------------------------+--------------+---------+----------+
| Library   | Class                  | Size (Bytes) | Percent | Of total |
+-----------+------------------------+--------------+---------+----------+
| dart:core | _Uri                   |        43563 |  15.53% |    5.70% |
| dart:core | _StringBase            |        28831 |  10.28% |    3.77% |
| dart:core | ::                     |        27559 |   9.83% |    3.60% |
| @other    |                        |        25467 |   9.08% |    3.33% |
| dart:core | Uri                    |        14936 |   5.33% |    1.95% |
| dart:core | int                    |        12276 |   4.38% |    1.61% |
| dart:core | NoSuchMethodError      |        12222 |   4.36% |    1.60% |
...

Here objects which can be attributed to _Uri take 5.7% of the snapshot, at the same time objects which can be attributed to dart:core library but not to any specific class within this library take 3.33% of the snapshot.

This command also supports estimating cumulative impact of a library or a package together with its dependencies - which can be computed from a precompiler trace (generated by adding --trace-precompiler-to=path/to/trace.json to --extra-gen-snapshot-options=).

$ snapshot_analysis summary [-b granularity] [-w filter] [-s dep-tree-start-depth] [-d dep-tree-display-depth] --precompiler-trace=<input-trace.json> <input-profile.json>

For example:

$ snapshot_analysis summary -b package /tmp/profile.json
+-----------------------------+--------------+---------+
| Package                     | Size (Bytes) | Percent |
+-----------------------------+--------------+---------+
| package:compiler            |      5369933 |  38.93% |
| package:front_end           |      2644942 |  19.18% |
| package:kernel              |      1443568 |  10.47% |
| package:_fe_analyzer_shared |       944555 |   6.85% |
...
$ snapshot_analysis summary -b package -s 1 -d 3 --precompiler-trace=/tmp/trace.json /tmp/profile.json
+------------------------------+--------------+---------+
| Package                      | Size (Bytes) | Percent |
+------------------------------+--------------+---------+
| package:compiler (+ 8 deps)  |      5762761 |  41.78% |
| package:front_end (+ 1 deps) |      2708981 |  19.64% |
| package:kernel               |      1443568 |  10.47% |
| package:_fe_analyzer_shared  |       944555 |   6.85% |
...
Dependency trees:

package:compiler (total 5762761 bytes)
├── package:js_ast (total 242490 bytes)
├── package:dart2js_info (total 101280 bytes)
├── package:crypto (total 27434 bytes)
│   ├── package:typed_data (total 11850 bytes)
│   └── package:convert (total 5185 bytes)
├── package:collection (total 15182 bytes)
├── package:_js_interop_checks (total 4627 bytes)
└── package:js_runtime (total 1815 bytes)

package:front_end (total 2708981 bytes)
└── package:package_config (total 64039 bytes)

compare

$ snapshot_analysis compare [-b granularity] <old.json> <new.json>

This command shows comparison between two size profiles, allowing to understand changes to which part of the program contributed most to the change in the overall snapshot size.

$ pkg/vm/bin/snapshot_analysis.dart compare -b class old.json new.json
+------------------------+--------------------------+--------------+---------+
| Library                | Class                    | Diff (Bytes) | Percent |
+------------------------+--------------------------+--------------+---------+
| dart:core              | _SimpleUri               |        11519 |  22.34% |
| dart:core              | _Uri                     |         6563 |  12.73% |
| dart:io                | _RandomAccessFile        |         5337 |  10.35% |
| @other                 |                          |         4009 |   7.78% |
...

In this example 11519 more bytes can be attributed to _SimpleUri class in new.json compared to old.json.

treemap

$ snapshot_analysis treemap [--format <format>] <input.json> <output-dir>
$ google-chrome <output-dir>/index.html

This command generates treemap representation of the information from the profile input.json and stores it in output-dir directory. Treemap can later be viewed by opening <output-dir>/index.html in the browser of your choice.

--format flag allows to control granularity of the output when input.json is a V8 snapshot profile, available options are:

• collapsed essentially renders ProgramInfo as a treemap, individual snapshot nodes are ignored.
• simplified same as collapsed, but also folds size information from nested functions into outermost function (e.g. top level function or a method) producing easy to consume output.
• data-and-code collapses snapshot nodes based on whether they represent data or executable code.
• object-type (default) collapses snapshot nodes based on their type only.

explain

explain dynamic-calls

$ snapshot_analysis explain dynamic-calls <profile.json> <trace.json>

This command generates a report listing dynamically dispatched selectors and their approximate impact on the code size.

snapshot_analysis explain dynamic-calls /tmp/profile.json /tmp/trace.json
+------------------------------+--------------+---------+----------+
| Selector                     | Size (Bytes) | Percent | Of total |
+------------------------------+--------------+---------+----------+
| set:requestHeader            |        10054 |  28.00% |    0.03% |
| get:scale                    |         3630 |  10.11% |    0.01% |
...
Dynamic call to set:requestHeader (retaining ~10054 bytes) occurs in:
    package:my-super-app/src/injector.dart::Injector.handle{body}

Dynamic call to get:scale (retaining ~3630 bytes) occurs in:
    package:some-dependency/src/image.dart::Image.==

API

This package can also be used as a building block for other packages which want to analyse VM AOT snapshots.

• package:vm_snapshot_analysis/instruction_sizes.dart provides helpers to read output of --print-instructions-sizes-to=...
• package:vm_snapshot_analysis/v8_profile.dart provides helpers to read output of --write-v8-snapshot-profile-to=...

Both formats can be converted into a ProgramInfo structure which attempts to breakdown snapshot size into hierarchical representation of the program structure which can be understood by a Dart developer, attributing bytes to packages, libraries, classes and functions.

• package:vm_snapshot_analysis/utils.dart contains helper method loadProgramInfo which automatically detects format of the input JSON file and creates ProgramInfo in an appropriate way, allowing to write code which works in the same way with both formats.

Precompiler Trace Format (--write-precompiler-trace-to=...)

AOT compiler can produce a JSON file containing information about compiled functions and dependencies between them. This file has the following structure:

{
    "trace": traceArray,
    "entities": entitiesArray,
    "strings": stringsArray,
}

• stringsArray is an array of strings referenced by other parts of the trace by their index in this array.

• entitiesArray is an flattened array of entities:

  • "C", <library-uri-idx>, <class-name-idx>, 0 - class record;
  • "V", <class-idx>, <name-idx>, 0 - static field record;
  • "F"|"S", <class-idx>, <name-idx>, <selector-id> - function record (F for dynamic functions and S for static functions);

  Note that all records in this array occupy the same amount of elements (4) to make random access by index possible.

• traceArray is an flattened array of precompilation events:

  • "R" - root event (always the first element)
  • "E" - end event (always the last element)
  • "C", <function-idx> - function compilation event

  Root and function compilation events can additionally be followed by a sequence of references which enumerate outgoing dependencies discovered by the AOT compiler:

  • <entity-idx> - a reference to a function or a static field;
  • "S", <selector-idx> - a dynamic call with the given selector;
  • "T", <selector-id> - dispatch table call with the given selector id;

Flattened array is an array of records formed by consecutive elements: [R0_0, R0_1, R0_2, R1_0, R1_1, R1_2, ...] here R0_* is the first record and R1_* is the second record and so on.

Features and bugs

Please file feature requests and bugs at the issue tracker.