Tools for Analyzing Chrome's Binary Size

These tools currently focus on Android compiled with GCC. They somewhat work for Android + Clang, and Linux builds, but not as well. As for Windows, some great tools already exist and are documented here:

There is also a dedicated mailing-list for binary size discussions:

Bugs are tracked here:

diagnose_bloat.py

Determine the cause of binary size bloat between two commits. Works for Android and Linux (although Linux symbol diffs have issues, as noted below).

How it Works

  1. Builds multiple revisions using release GN args.
    • Default is to build just two revisions (before & after commit)
    • Rather than building, can fetch build artifacts and .size files from perf bots (--cloud)
  2. Measures all outputs using resource_size.py and supersize.
  3. Saves & displays a breakdown of the difference in binary sizes.

Example Usage

# Build and diff HEAD^ and HEAD.
tools/binary_size/diagnose_bloat.py HEAD -v

# Diff BEFORE_REV and AFTER_REV using build artifacts downloaded from perf bots.
tools/binary_size/diagnose_bloat.py AFTER_REV --reference-rev BEFORE_REV --cloud -v

# Fetch a single .size and .apk from the bots:
tools/binary_size/diagnose_bloat.py AFTER_REV --cloud --single

# Build and diff all contiguous revs in range BEFORE_REV..AFTER_REV for src/v8.
tools/binary_size/diagnose_bloat.py AFTER_REV --reference-rev BEFORE_REV --subrepo v8 --all -v

# Display detailed usage info (there are many options).
tools/binary_size/diagnose_bloat.py -h

Super Size

Collect, archive, and analyze Chrome's binary size. Supports Android and Linux (although Linux has issues).

.size files are archived on perf builders so that regressions can be quickly analyzed (via diagnose_bloat.py --cloud).

.size files are archived on official builders so that symbols can be diff'ed between milestones.

Technical Details

What's in a .size File?

.size files are gzipped plain text files that contain:

  1. A list of .so section sizes, as reported by readelf -S,
  2. Metadata (GN args, filenames, timestamps, git revision, build id),
  3. A list of symbols, including name, address, size, padding (caused by alignment), and associated .o / .cc files.

How are Symbols Collected?

  1. Symbol list is Extracted from linker .map file.
    • Map files contain some unique pieces of information, such as ** merge strings entries, and the odd unnamed symbol (which map at least lists a .o path).
  2. .o files are mapped to .cc files by parsing .ninja files.
    • This means that .h files are never listed as sources. No information about inlined symbols is gathered.
  3. Symbol aliases (when multiple symbols share an address) are collected from debug information via nm elf-file.
    • Aliases are created by identical code folding (linker optimization).
    • Aliases have the same address and size, but report their .pss as .size / .num_aliases.
  4. Paths for shared symbols (those found in multiple .o files) are collected by running nm on every .o file.

What Other Processing Happens?

  1. Path normalization:

    • Prefixes are removed: out/Release/, gen/, obj/
    • Archive names made more pathy: foo/bar.a(baz.o) -> foo/bar.a/baz.o
    • Shared symbols do not store the complete source paths. Instead, the common ancestor is computed and stored as the path.
      • Example: base/{shared}/3 (the “3” means three different files contain the symbol)

  2. Name normalization:

    • (anonymous::) is removed from names (and stored as a symbol flag).
    • [clone] suffix removed (and stored as a symbol flag).
    • vtable for FOO -> Foo [vtable]
    • Mangling done by linkers is undone (e.g. prefixing with “unlikely.”)
    • Names are processed into:
      • name: Name without template and argument parameters
      • template_name: Name without argument parameters.
      • full_name: Name with all parameters.

  3. Clustering

    • Compiler & linker optimizations can cause symbols to be broken into multiple parts to become candidates for inlining (“partial inlining”).
    • These symbols are sometimes suffixed with “[clone]” (removed by normalization).
    • Clustering creates groups containing all pieces of a symbol (in the case where multiple pieces remain after inlining).
    • Clustering is done by default on SizeInfo.symbols. To view unclustered symbols, use SizeInfo.raw_symbols.

  4. Diffing

    • Some heuristics for matching up before/after symbols.

Is Super Size a Generic Tool?

No. Most of the logic is would could work for any ELF executable. However, being a generic tool is not a goal. Some examples of existing Chrome-specific logic:

  • Assumes .ninja build rules are available.
  • Heuristic for locating .so given .apk.
  • Roadmap includes .pak file analysis.

Usage: archive

Collect size information and dump it into a .size file.

Note: Refer to diagnose_bloat.py for list of GN args to build a Release binary (or just use the tool with --single).

Example Usage:

# Android:
ninja -C out/Release -j 1000 apks/ChromePublic.apk
tools/binary_size/supersize archive chrome.size --apk-file out/Release/apks/ChromePublic.apk -v

# Linux:
LLVM_DOWNLOAD_GOLD_PLUGIN=1 gclient runhooks  # One-time download.
ninja -C out/Release -j 1000 chrome
tools/binary_size/supersize archive chrome.size --elf-file out/Release/chrome -v

Usage: html_report

Creates an interactive size breakdown (by source path) as a stand-alone html report.

Example output: https://agrieve.github.io/chrome/

Example Usage:

tools/binary_size/supersize html_report chrome.size --report-dir size-report -v
xdg-open size-report/index.html

Usage: diff

A convenience command equivalent to: console before.size after.size --query='Print(Diff(size_info1, size_info2))'

Example Usage:

tools/binary_size/supersize diff before.size after.size --all

Usage: console

Starts a Python interpreter where you can run custom queries, or run pre-made queries from canned_queries.py.

Example Usage:

# Prints size infomation and exits (does not enter interactive mode).
tools/binary_size/supersize console chrome.size --query='Print(size_info)'

# Enters a Python REPL (it will print more guidance).
tools/binary_size/supersize console chrome.size

Example session:

>>> ShowExamples()  # Get some inspiration.
...
>>> sorted = size_info.symbols.WhereInSection('t').Sorted()
>>> Print(sorted)  # Have a look at the largest symbols.
...
>>> sym = sorted.WhereNameMatches('TrellisQuantizeBlock')[0]
>>> Disassemble(sym)  # Time to learn assembly.
...
>>> help(canned_queries)
...
>>> Print(canned_queries.TemplatesByName(depth=-1))

Roadmap

  1. Better Linux support (clang+lld+lto vs gcc+gold).
  2. More archive features:
    • Find out more about 0xffffffffffffffff addresses, and why such large gaps exist after them. (crbug/709050)
    • Collect .pak file information (using .o.whitelist files)
    • Collect java symbol information
    • Collect .apk entry information
  3. More console features:
    • CSV output (for pasting into a spreadsheet).
    • Add SplitByName() - Like GroupByName(), but recursive.
    • A canned query, that does what ShowGlobals does (as described in Windows Binary Sizes).
    • Show symbol counts by bucket size.
      • 3 symbols < 64 bytes. 10 symbols < 128, 3 < 256, 5 < 512, 0 < 1024, 3 < 2048
  4. More html_report features:
    • Able to render size diffs (tint negative size red).
    • Break down by other groupings (Create from result of SplitByName())
    • Render as simple tree view rather than 2d boxes
  5. Integrate with resource_sizes.py so that it tracks size of major components separately: chrome vs blink vs skia vs v8.
  6. Add dependency graph info, perhaps just on a per-file basis.
    • No idea how to do this, but Windows can do it via tools\win\linker_verbose_tracking.py