feat(bb): faster create_circuit — arithmetic, tiled gate storage, containers, cold assert paths (#537)

* feat(bb): faster create_circuit — wide arithmetic, tiled (AoSoA) gate storage, size hints

Barrett fast paths for scalar-field divmod, native Montgomery + safegcd
bigfield division, constexpr folding for small field literals, gate storage
as 8-gate tiles with contiguous per-selector arrays (single append_gate
surface, bulk copy_into extraction), and builder size hints from ACIR.

* feat(bb): container and allocation optimizations (polynomial backing, copy cycles, builder reserves)

Lazy zero-initialized (anonymous mmap) polynomial backing memory, CSR copy
cycles in trace population, and reserved per-variable bookkeeping vectors in
the ACIR builder path.

* feat(bb): compile BB_BENCH instrumentation out of shipped wasm

Sets ENABLE_WASM_BENCH=OFF in the wasm-threads preset so the released
wasm no longer carries op-counting overhead. Wasm bench runs keep
wall-clock and peak-memory measurements; hierarchical breakdowns and
memory profiles are native-only, and the CI breakdown upload is skipped
for wasm. Configure with -DENABLE_WASM_BENCH=ON to get wasm op counts
for dev runs.

* refactor(bb): review feedback — name magic numbers, drop stale comment, GateRowT alias

- databus selector array sized by NUM_BUS_COLUMNS instead of a literal 7
- named constants for the ACIR block size hints and variable reserve
- remove orphaned doc block for the deleted check_selector_length_consistency
- replace 19 'typename std::decay_t<decltype(...)>::Row' sites with the existing GateRowT alias

* refactor(bb): single NON_GATE_SELECTORS table drives gate-storage selector wiring

The non-gate selector list (q_m, q_c, q_1..q_5) was duplicated across GateRow,
GateTile, append_gate's copies, the columns_ wiring, and a literal 7. A single
constexpr member-pointer table now defines the list and its order; the count,
append_gate, and column construction derive from it. Also: GATE_TILE_SHIFT
derived from GATE_TILE_SIZE (power-of-two asserted), tile-size comment expressed
in terms of sizeof(FF), and shared gate_selector_or_zero read helper.

Author: ludamad

barretenberg/.claude/skills/benchmark-chonk/SKILL.md

@@ -133,7 +133,7 @@ HARDWARE_CONCURRENCY=8 BB_BENCH=1 \
 The wasmtime wrapper sets:
 - `-Wthreads=y -Sthreads=y` — enable WASM threads and shared memory
 - `--env HARDWARE_CONCURRENCY` — thread count
-- `--env BB_BENCH` — enable operation counting (`ENABLE_WASM_BENCH=ON` is set by the `wasm-threads` preset)
+- `--env BB_BENCH` — enable operation counting (requires configuring with `-DENABLE_WASM_BENCH=ON`; it is OFF by default in the `wasm-threads` preset to keep the shipped wasm fast)
 - `--dir=$HOME/.bb-crs --dir=.` — filesystem access for CRS and working directory
 
 ## Local runs are noisy — average 3 runs
@@ -322,7 +322,7 @@ The generic Google-Benchmark A/B scripts still exist for non-Chonk targets:
 - **Use `./bootstrap.sh` for initial builds** — it downloads cached artifacts and avoids build issues. Use `cmake --preset clang20 && cd build && ninja bb` for incremental rebuilds after code changes.
 - **Build dir is `build/`** — the `clang20` preset outputs to `build/`, not `build-no-avm`. The `clang20-no-avm` preset also uses `build/` (it disables AVM at cmake level, not via directory name).
 - **If the zig cache breaks** (missing `libubsan_rt.a` errors), delete `build/` and reconfigure: `rm -rf build && cmake --preset clang20`.
-- **WASM preset:** `wasm-threads`. Build dir is `build-wasm-threads/`. The preset enables `ENABLE_WASM_BENCH=ON` automatically.
+- **WASM preset:** `wasm-threads`. Build dir is `build-wasm-threads/`. `ENABLE_WASM_BENCH` is OFF by default (the shipped wasm carries no BB_BENCH instrumentation); configure with `cmake --preset wasm-threads -DENABLE_WASM_BENCH=ON` when you need wasm op counts, and reconfigure without it afterwards.
 - **WASM is ~2.8x slower than native** — this ratio is consistent across all circuit types.
 - **CRS:** Ensure `~/.bb-crs` exists. For WASM, wasmtime needs `--dir=$HOME/.bb-crs`.
 - **`BB_BENCH=1` vs `--print_bench`:** Either activates profiling. `--print_bench` also triggers the hierarchical tree output to stderr. In Google-Benchmark targets that wrap their loops with `GOOGLE_BB_BENCH_REPORTER`, the same activation happens automatically when `BB_BENCH=1` is set.

barretenberg/cpp/CLAUDE.md

@@ -97,6 +97,8 @@ These are load-bearing: violating them will compile on native but break WASM, or
 
 **IMPORTANT**: In the barretenberg context, "bench" or "benchmark" almost always means running `benchmark_remote.sh` for the given target on a remote benchmarking machine.
 
+**Never benchmark against test binaries (`*_tests`) — the results will always be wrong.** Test circuits are small mocks whose cost profile does not resemble real proving workloads. Benchmark against real inputs: the pinned Chonk flows (`scripts/chonk_inputs.sh download`, then `bb prove --scheme chonk --ivc_inputs_path chonk-pinned-flows/<flow>/ivc-inputs.msgpack`) or the dedicated `*_bench` targets.
+
 To run benchmarks for a specific target:
 ```bash
 cd barretenberg/cpp
@@ -196,7 +198,7 @@ For bb.js, run:
 
 ```bash
 barretenberg/cpp/scripts/chonk_inputs.sh download
-barretenberg/ts/bb.js/scripts/run_test.sh bbapi/chonk_pinned_inputs.test.js
+barretenberg/ts/scripts/run_test.sh bbapi/chonk_pinned_inputs.test.js
 ```
 
 Typical workflow

barretenberg/cpp/CMakePresets.json

@@ -434,7 +434,7 @@
       "cacheVariables": {
         "CMAKE_BUILD_TYPE": "Release",
         "MULTITHREADING": "ON",
-        "ENABLE_WASM_BENCH": "ON"
+        "ENABLE_WASM_BENCH": "OFF"
       }
     },
     {

barretenberg/cpp/scripts/benchmark_wasm.sh

@@ -9,6 +9,8 @@ HARDWARE_CONCURRENCY=${HARDWARE_CONCURRENCY:-16}
 cd $(dirname $0)/..
 
 # Configure and build.
+# BB_BENCH op counts are compiled out of wasm by default; add -DENABLE_WASM_BENCH=ON here if you
+# need them (timings are unaffected either way).
 cmake --preset wasm-threads
 cmake --build --preset wasm-threads --target $BENCHMARK
 

barretenberg/cpp/scripts/ci_benchmark_ivc_flows.sh

@@ -167,10 +167,12 @@ function run_bb_cli_bench {
     }
   else # wasm
     export WASMTIME_ALLOWED_DIRS="--dir=$flow_folder --dir=$output"
-    # Add --bench_out_hierarchical flag for wasm builds to capture hierarchical op counts and timings
-    # Note: --memory_profile_out is native-only (getrusage not available in wasm)
-    memusage scripts/wasmtime.sh $WASMTIME_ALLOWED_DIRS ./build-wasm-threads/bin/bb "$@" "--bench_out_hierarchical" "$output/benchmark_breakdown.json" || {
-      echo "bb wasm failed with args: $@ --bench_out_hierarchical $output/benchmark_breakdown.json"
+    # No --bench_out_hierarchical or --memory_profile_out here: BB_BENCH instrumentation is compiled out
+    # of wasm builds (ENABLE_WASM_BENCH is OFF in the wasm-threads preset to keep the shipped wasm fast),
+    # and --memory_profile_out is native-only (getrusage not available in wasm).
+    # Only wall-clock time and peak memory are measured for wasm.
+    memusage scripts/wasmtime.sh $WASMTIME_ALLOWED_DIRS ./build-wasm-threads/bin/bb "$@" || {
+      echo "bb wasm failed with args: $@"
       print_chonk_flow_reuse_hint "$runtime" "$flow_folder"
       exit 1
     }
@@ -262,7 +264,8 @@ chonk_flow "$runtime_arg" "$resolved_chonk_flow_folder"
 runtime="$runtime_arg"
 flow_name="$(basename "$resolved_chonk_flow_folder")"
 
-if [[ "${CI:-}" == "1" ]] && [[ "${CI_USE_BUILD_INSTANCE_KEY:-0}" == "1" ]]; then
+# Breakdown and memory profile only exist for native runs (BB_BENCH is compiled out of wasm builds).
+if [[ "$runtime" == "native" ]] && [[ "${CI:-}" == "1" ]] && [[ "${CI_USE_BUILD_INSTANCE_KEY:-0}" == "1" ]]; then
   echo_header "Uploading Barretenberg benchmark breakdowns for $flow_name"
 
   current_sha=$(git rev-parse HEAD)

barretenberg/cpp/src/barretenberg/benchmark/ultra_bench/mega_honk.bench.cpp

@@ -94,16 +94,20 @@ BENCHMARK(construct_proof_megahonk_power_of_2)
 
 int main(int argc, char** argv)
 {
+#if !defined(__wasm__) || defined(ENABLE_WASM_BENCH)
     bb::detail::use_bb_bench = true;
+#endif
 
     ::benchmark::Initialize(&argc, argv);
     if (::benchmark::ReportUnrecognizedArguments(argc, argv))
         return 1;
     ::benchmark::RunSpecifiedBenchmarks();
     ::benchmark::Shutdown();
 
+#if !defined(__wasm__) || defined(ENABLE_WASM_BENCH)
     std::cout << "\n=== Detailed BB_BENCH Profiling Stats ===\n";
     bb::detail::GLOBAL_BENCH_STATS.print_aggregate_counts_hierarchical(std::cout);
+#endif
 
     return 0;
 }

barretenberg/cpp/src/barretenberg/benchmark/ultra_bench/ultra_honk.bench.cpp

@@ -173,8 +173,10 @@ BENCHMARK(construct_proof_ultrahonk_1M_gates_dyadic_2_21)->Unit(kMillisecond);
 
 int main(int argc, char** argv)
 {
+#if !defined(__wasm__) || defined(ENABLE_WASM_BENCH)
     // Enable BB_BENCH profiling
     bb::detail::use_bb_bench = true;
+#endif
 
     // Run benchmarks
     ::benchmark::Initialize(&argc, argv);
@@ -183,9 +185,11 @@ int main(int argc, char** argv)
     ::benchmark::RunSpecifiedBenchmarks();
     ::benchmark::Shutdown();
 
+#if !defined(__wasm__) || defined(ENABLE_WASM_BENCH)
     // Print detailed profiling stats
     std::cout << "\n=== Detailed BB_BENCH Profiling Stats ===\n";
     bb::detail::GLOBAL_BENCH_STATS.print_aggregate_counts_hierarchical(std::cout);
+#endif
 
     return 0;
 }

barretenberg/cpp/src/barretenberg/circuit_checker/ultra_circuit_builder_arithmetic.test.cpp

@@ -543,29 +543,22 @@ TEST_F(UltraCircuitBuilderArithmetic, QArith3Gate)
         uint32_t w4_next_idx = builder.add_variable(w_4_next);
 
         // Gate 1: q_arith = 3
-        builder.blocks.arithmetic.populate_wires(w1_idx, w2_idx, w3_idx, w4_idx);
-        builder.blocks.arithmetic.q_m().emplace_back(q_m);
-        builder.blocks.arithmetic.q_1().emplace_back(q_1);
-        builder.blocks.arithmetic.q_2().emplace_back(q_2);
-        builder.blocks.arithmetic.q_3().emplace_back(q_3);
-        builder.blocks.arithmetic.q_4().emplace_back(q_4);
-        builder.blocks.arithmetic.q_5().emplace_back(0);
-        builder.blocks.arithmetic.q_c().emplace_back(q_c);
-        builder.blocks.arithmetic.set_gate_selector(GateKind::Arith, 3);
-        builder.check_selector_length_consistency();
+        builder.blocks.arithmetic.append_gate({ .wires = { w1_idx, w2_idx, w3_idx, w4_idx },
+                                                .q_m = q_m,
+                                                .q_c = q_c,
+                                                .q_1 = q_1,
+                                                .q_2 = q_2,
+                                                .q_3 = q_3,
+                                                .q_4 = q_4,
+                                                .gate_kind = GateKind::Arith,
+                                                .gate_value = 3 });
         builder.increment_num_gates();
 
         // Gate 2: provides w_1_shift and w_4_shift
-        builder.blocks.arithmetic.populate_wires(w1_next_idx, builder.zero_idx(), builder.zero_idx(), w4_next_idx);
-        builder.blocks.arithmetic.q_m().emplace_back(0);
-        builder.blocks.arithmetic.q_1().emplace_back(0);
-        builder.blocks.arithmetic.q_2().emplace_back(0);
-        builder.blocks.arithmetic.q_3().emplace_back(0);
-        builder.blocks.arithmetic.q_4().emplace_back(0);
-        builder.blocks.arithmetic.q_5().emplace_back(0);
-        builder.blocks.arithmetic.q_c().emplace_back(0);
-        builder.blocks.arithmetic.set_gate_selector(GateKind::Arith, 1);
-        builder.check_selector_length_consistency();
+        builder.blocks.arithmetic.append_gate(
+            { .wires = { w1_next_idx, builder.zero_idx(), builder.zero_idx(), w4_next_idx },
+              .gate_kind = GateKind::Arith,
+              .gate_value = 1 });
         builder.increment_num_gates();
 
         if (expect_valid) {