binary-size-analysis

name

binary-size-analysis

description

Analyze agent-data-plane binary size to identify top crates and modules contributing to binary size, and evaluate optimization opportunities. TRIGGER when: user asks about binary size, code size, shrinking the binary, binary optimization, reducing binary footprint, evaluating what contributes to binary size, or asks to find ways to make the binary smaller. DO NOT TRIGGER when: user asks about runtime memory usage, memory leaks, allocation profiling, or performance tuning.

Binary Size Analysis for agent-data-plane

You are performing a standalone binary size analysis of the agent-data-plane binary. Your goal is to identify the top 3-5 contributors to binary size and assess whether each can feasibly be reduced.

Step 1: Build with optimized-release profile

Build the binary using the production release profile. Warn the user this may take several minutes due to fat LTO and single codegen unit.

cargo build --profile optimized-release -p agent-data-plane

The optimized-release profile is defined in the workspace root Cargo.toml and inherits from release with lto = "fat" and codegen-units = 1. This matches what ships to users.

Step 2: Separate debug info and strip the binary

Create a separate debug file and strip the binary, so we analyze the actual shipped artifact size while retaining symbols for attribution. This follows the established CI pattern from .gitlab/benchmark.yml.

cp target/optimized-release/agent-data-plane target/optimized-release/agent-data-plane.analyzed
objcopy --only-keep-debug target/optimized-release/agent-data-plane.analyzed target/optimized-release/agent-data-plane.debug
strip --strip-debug target/optimized-release/agent-data-plane.analyzed

We work on a copy (.analyzed) so we don't destroy the original build artifact. Report the stripped binary size to the user as the headline number.

Step 3: Multi-tool analysis

Run these analyses to build a comprehensive picture.

IMPORTANT — Bloaty DWARF attribution is unreliable with fat LTO + single CGU. Bloaty uses DWARF debug info ranges to attribute sizes to symbols. With fat LTO and codegen-units = 1, these DWARF ranges become wildly inaccurate — a small function's debug range can span a huge address range that actually belongs to other functions. In testing, bloaty reported a 21 KiB function as 1.79 MiB (88x inflation), and multiple symbols showed 5-176x inflation. Large async state machine functions (>25 KiB) tended to be accurate within ~20%, but smaller functions were often grossly inflated.

Use ELF symbol sizes as ground truth for per-symbol sizes. The ELF symbol table (readelf -Ws or nm --size-sort) records the actual function sizes set by the compiler/linker and is not affected by DWARF range issues. Always cross-reference bloaty's per-symbol claims against nm.

cargo bloat --crates is the most trustworthy crate-level view. It uses ELF symbol sizes (not DWARF), so it is not affected by the LTO attribution problem.

Bloaty's section breakdown (-d sections) remains accurate since it reads ELF section headers directly, not DWARF. Bloaty's compile unit and symbol breakdowns should be treated as directional hints only — always validate specific claims against nm/readelf.

3a. cargo bloat — crate-level attribution (primary source of truth)

cargo bloat --profile optimized-release -p agent-data-plane --crates -n 20

This gives the most trustworthy crate-level view of .text section contributions, using ELF symbol sizes. For the top crates, drill in:

cargo bloat --profile optimized-release -p agent-data-plane --filter <crate_name> -n 20

3b. Largest symbols via nm (ground truth for per-symbol sizes)

Use nm on the unstripped binary (not the .analyzed copy) for accurate ELF symbol sizes:

nm --size-sort --reverse-sort target/optimized-release/agent-data-plane | rustfilt | head -60

The first column is the symbol size in hex. This is the authoritative size for each symbol.

3c. Bloaty — section breakdown (accurate)

bloaty -d sections \
  --debug-file=target/optimized-release/agent-data-plane.debug \
  target/optimized-release/agent-data-plane.analyzed

This reveals the proportion of .text (code), .rodata (read-only data), .data, .bss, etc. Section-level data comes from ELF headers and is reliable.

3d. Bloaty — compile unit breakdown (directional only)

bloaty -d compileunits -n 30 \
  --debug-file=target/optimized-release/agent-data-plane.debug \
  target/optimized-release/agent-data-plane.analyzed

This shows which source files / compile units contribute the most code. Treat as directional — with fat LTO, compile unit attribution can be imprecise.

3e. Bloaty — symbol breakdown (cross-reference with nm)

bloaty -d symbols -n 50 \
  --debug-file=target/optimized-release/agent-data-plane.debug \
  target/optimized-release/agent-data-plane.analyzed | rustfilt

Always cross-reference the top symbols against nm --size-sort to validate. If bloaty reports a symbol as significantly larger than nm does (>1.5x), trust nm. You can batch-validate like this:

# Get bloaty's CSV output
bloaty -d symbols -n 50 --csv \
  --debug-file=target/optimized-release/agent-data-plane.debug \
  target/optimized-release/agent-data-plane.analyzed > /tmp/bloaty_symbols.csv

# Compare against nm for the same symbols (match by hash suffix)
nm --size-sort --reverse-sort target/optimized-release/agent-data-plane | rustfilt > /tmp/nm_symbols.txt

Then compare the sizes for each symbol. Flag any where bloaty exceeds nm by >1.5x as a DWARF attribution artifact — do not report those inflated sizes to the user.

3f. Dependency tree

cargo tree -p agent-data-plane -e normal --depth 1

For heavy crates identified in the analysis, investigate their transitive dependencies:

cargo tree -p agent-data-plane -i <crate_name>

Step 4: Investigate top contributors

For each of the top 3-5 contributors identified above, perform targeted investigation.

Distinguishing workspace vs third-party crates

Use cargo metadata --no-deps --format-version 1 to get the workspace crate list (packages with source: null). The existing script at ci/tooling/analyze-binary-size.py has this logic as a reference. Workspace crates use underscores in symbols but hyphens in Cargo.toml.

For third-party dependencies

• Check if heavy transitive deps are being pulled in: cargo tree -i <crate>
• Check if unnecessary Cargo features are enabled: cargo tree -p agent-data-plane -f '{p} {f}' | grep <crate>
• Research whether a lighter alternative crate exists
• Check if feature flags could reduce scope (for example, disabling default features)

For workspace crates

• Examine whether size comes from monomorphized generics: look for many instantiations of the same function with different type parameters in the symbol list. A tell-tale sign is seeing the same function name repeated with different type suffixes.
• Check for large inline code or embedded data (for example, HTML templates, lookup tables).
• For monomorphization issues: evaluate whether type-erasure (dyn Trait, Box<dyn ...>) or consolidating generic instantiations could reduce code size. Reference the recent work on saluki-io where HttpClient was refactored to use type-erased ClientBody to reduce monomorphization of HTTP client/server types.

Monomorphization deep-dive

When you suspect monomorphization is a significant contributor for a crate, use nm (not bloaty) to get accurate symbol sizes and grep for that crate's symbols:

nm --size-sort --reverse-sort target/optimized-release/agent-data-plane | rustfilt | grep '<crate_name>' | head -30

The first column is the hex size. Look for the same function appearing with different generic instantiations (different hash suffixes but same base name). Count unique instantiations vs unique base function names to quantify the monomorphization overhead.

Step 5: Present findings

Produce a summary table like:

Rank	Crate/Module	Size	Category	Feasibility	Notes
1	example_crate	~1.5 MiB	Third-party dep	Medium	Could disable unused features X, Y
2	saluki_io::net	~800 KiB	Monomorphization	Easy	12 instantiations of send_request<T>, type-erasure viable
...

Categories: third-party dep, workspace code, monomorphization, embedded data, protocol buffers, serialization/deserialization

Feasibility: easy / medium / hard with a one-line rationale.

After presenting the table, ask the user:

"Would you like me to create a plan to tackle any of these specific areas?"

Do NOT proceed to implementation without user direction. The goal of this skill is diagnosis and feasibility assessment, not automatic refactoring.