Counting A WebAssembly Module

wasm-counts.wasm turns a WebAssembly module into deterministic long-form CSV. It reports factual static counts for comparing builds; it does not assign a score or decide whether a module is acceptable.

make -j qip components/application/wasm/wasm-counts.wasm

qip run \
  -i components/text/e164.wasm \
  -- components/application/wasm/wasm-counts.wasm

The output has one integer measurement per row:

metric,value
module_bytes,304
functions_defined,5
functions_imported,0
instructions,75
loops,1
branches,5
simd_instructions,0

Rows are always emitted in the same order, including zero values. This makes the output useful with diff, spreadsheets, and databases without treating a missing row as zero.

What It Counts #

The report covers module structure, functions, control flow, calls, SIMD, declared memory, data segments, and instructions that can trap directly based on their operands or bounds.

• loops counts structured loop instructions.
• function_instructions counts instructions in function bodies. instructions also includes instructions in constant expressions used by globals and active data or element segments.
• branches counts br, br_if, and br_table instructions.
• conditional_branches counts if, br_if, and br_table instructions.
• simd_instructions counts instructions in the SIMD opcode space.
• v128_types counts v128 occurrences in function and global types.
• table_initial_slots and table_maximum_slots sum declared table limits. element_initializers counts values supplied by element segments; it is not the table's capacity.
• tables_funcref, tables_externref, and tables_typed_reference count declared table reference types. tables_fixed_size counts tables whose declared minimum and maximum are equal.
• Element rows separate active, passive, and declarative segments. Active segments are also split between table zero and nonzero tables. Their offset expressions distinguish i32.const 0, i32.const 1, other i32.const values, global.get, and other expressions. Initializers are split between function-index and expression encodings.
• call_indirect, return_call_indirect, and call_ref count those exact instructions. calls_indirect remains their aggregate. Table-indexed calls are also split between table zero and nonzero tables.
• table_get, table_set, table_init, elem_drop, table_copy, table_grow, table_size, and table_fill count those exact instructions. The ref_null, ref_is_null, and ref_func rows do the same for reference instructions.
• memory_loads, memory_stores, memory_copies, and memory_fills count instruction sites, not runtime accesses.
• potentially_trapping_instructions combines explicit traps, integer division and remainder, trapping float-to-integer conversion, bounded memory and table operations, and indirect or reference calls.
• explicit_traps, integer_divisions, integer_remainders, trapping_float_to_int, potentially_trapping_memory, potentially_trapping_table, and call_ref expose that total's parts.

The trapping count describes instruction sites, not executions. A load that is provably in bounds still counts, while an ordinary direct call does not count merely because its callee might trap.

All rows report facts from the binary. The component does not decide whether a table shape, instruction, or count is acceptable. Apply those decisions in a separate checker or in the system that consumes the CSV.

Memory rows report declared capacity and initial data. They do not measure allocator use, working set, stack depth, or peak memory; those require running or instrumenting the module.

Load It Into SQLite #

SQLite's shell can import the output directly:

qip run -i a.wasm -- components/application/wasm/wasm-counts.wasm > a.csv

sqlite3 counts.db

Then use the shell's CSV mode:

.mode csv
.headers on
CREATE TABLE a_counts (metric TEXT PRIMARY KEY, value INTEGER NOT NULL);
.import --skip 1 a.csv a_counts
SELECT * FROM a_counts WHERE metric IN ('loops', 'branches', 'simd_instructions');

For comparisons across many modules, add the module name as a column while loading each file into a shared (module, metric, value) table.

When Not To Use It #

Use WABT's wasm-stats when you need a full frequency table for every opcode and immediate value. Use runtime profiling when you need to know which code is hot or how much memory an execution actually touches. Counts alone do not predict either result.