Bench: GC overhead via bump-allocator comparison

Adds --alloc=<gc|bump> to ail build/run. Bump path links a 256MB
no-free arena C stub instead of libgc; IR is byte-identical except
for the @GC_malloc → @bump_malloc symbol swap. Bench harness times
two allocation-heavy workloads (list cons/sum and balanced tree
build/walk) under both modes.

Numbers (RUNS=5, median of 4):
  bench_list_sum   gc 0.141s  bump 0.048s  +194%
  bench_tree_walk  gc 0.103s  bump 0.041s  +151%

Bucket: large. ~60% of runtime is Boehm on these workloads —
upper bound for any realistic program. Both fixtures hold the
heap fully live, so the cost we're seeing is Boehm's allocate
path itself, not collection work; that fact narrows the design
space for the GC discussion.

- crates/ailang-codegen: AllocStrategy enum, three callsites and
  the IR header parameterised.
- crates/ail/src/main.rs: --alloc flag plumbed; bump runtime
  located + compiled on demand.
- runtime/bump.c: 256MB static arena, abort-on-overflow.
- examples/bench_list_sum, bench_tree_walk: accumulator-form
  fixtures (textbook recursive sum was constructor-blocked).
- bench/run.sh: harness with Python timing helper (Arch's
  /usr/bin/time isn't part of the base install).

No language-level changes; default --alloc=gc, all 141 workspace
tests green, all 5 IR snapshots unchanged, 11 prior fixtures
produce identical stdout.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

bench/run.sh

@@ -0,0 +1,169 @@
+#!/usr/bin/env bash
+#
+# GC-overhead bench harness (Bench iter).
+#
+# Builds each fixture twice — `--alloc=gc` (Boehm conservative GC) and
+# `--alloc=bump` (no-free 256 MB arena from `runtime/bump.c`). Runs each
+# binary N times, drops the slowest run, takes the median wall time.
+# The bump number minus the gc number is the upper-bound cost of GC.
+#
+# Output: a table with gc-median, bump-median, overhead %, and max RSS
+# for both modes. Designed to be captured verbatim into a JOURNAL entry.
+#
+# Requirements: bash, /usr/bin/time -v (GNU coreutils), bc, sort, awk,
+# a release-mode `ail` binary.
+#
+# Usage: bench/run.sh [-n RUNS]
+#   -n RUNS   number of timed runs per binary (default 5; min 3 so we
+#             can drop the slowest and still take a median over 4).
+
+set -euo pipefail
+
+RUNS=5
+while getopts "n:" opt; do
+    case $opt in
+        n) RUNS="$OPTARG" ;;
+        *) echo "usage: $0 [-n RUNS]" >&2; exit 2 ;;
+    esac
+done
+
+if (( RUNS < 3 )); then
+    echo "RUNS must be >= 3" >&2
+    exit 2
+fi
+
+# Anchor at the workspace root regardless of CWD.
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+cd "$ROOT"
+
+# We measure wall-clock + max RSS via a small Python helper that wraps
+# the binary, calls `time.monotonic()` around `os.waitpid`, and reads
+# `getrusage(RUSAGE_CHILDREN).ru_maxrss` (KB on Linux). This avoids a
+# /usr/bin/time dependency (Arch / minimal containers often don't have
+# the GNU coreutils `time` binary installed) without sacrificing
+# either signal: monotonic clocks for wall time, kernel-reported peak
+# resident set for RSS.
+PY="$(command -v python3 || true)"
+if [[ -z "$PY" ]]; then
+    echo "error: python3 is required for the timing helper" >&2
+    exit 2
+fi
+
+# Build the release `ail` binary if needed.
+echo ">>> ensuring release ail binary"
+cargo build --release -p ail >/dev/null
+AIL="$ROOT/target/release/ail"
+[[ -x "$AIL" ]] || { echo "ail binary missing: $AIL" >&2; exit 1; }
+
+OUTDIR="$ROOT/target/bench"
+mkdir -p "$OUTDIR"
+
+# Compile both modes for both fixtures up front so the bench loop only
+# measures runtime, not build time.
+fixtures=(bench_list_sum bench_tree_walk)
+modes=(gc bump)
+echo ">>> compiling fixtures (-O2)"
+for f in "${fixtures[@]}"; do
+    src="$ROOT/examples/$f.ail.json"
+    [[ -f "$src" ]] || { echo "missing fixture: $src" >&2; exit 1; }
+    for m in "${modes[@]}"; do
+        bin="$OUTDIR/${f}_${m}"
+        echo "    $f --alloc=$m -> $bin"
+        "$AIL" build --opt=-O2 --alloc="$m" "$src" -o "$bin" >/dev/null
+    done
+done
+
+# Time one binary one time. Wraps the binary in a Python helper that
+# measures wall-clock via time.monotonic() and max RSS (KB) via
+# getrusage(RUSAGE_CHILDREN).ru_maxrss after the child exits. Stdout
+# of the binary is discarded; we already verified correctness via a
+# smoke run earlier. Output: "wall_seconds rss_kb" on a single line.
+time_one() {
+    local bin="$1"
+    "$PY" -c '
+import os, resource, subprocess, sys, time
+bin_path = sys.argv[1]
+t0 = time.monotonic()
+p = subprocess.Popen([bin_path], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+p.wait()
+t1 = time.monotonic()
+ru = resource.getrusage(resource.RUSAGE_CHILDREN)
+# ru_maxrss is in KB on Linux. We want the peak across only this child;
+# RUSAGE_CHILDREN is cumulative across all children of the helper, but
+# the helper only spawns this one child per invocation, so the value is
+# this run.
+print(f"{t1 - t0:.3f} {ru.ru_maxrss}")
+sys.exit(0 if p.returncode == 0 else 1)
+' "$bin"
+}
+
+# Run a binary RUNS times, drop the slowest run by wall time, return
+# median of the rest as `wall rss` (rss = max across the kept runs).
+median_run() {
+    local bin="$1"
+    local times=()
+    local rsses=()
+    for ((i = 0; i < RUNS; i++)); do
+        read -r w r < <(time_one "$bin")
+        times+=("$w")
+        rsses+=("$r")
+    done
+    # Compute index of slowest (largest wall) and drop it.
+    local slowest_idx=0
+    for ((i = 1; i < ${#times[@]}; i++)); do
+        if [[ $(awk -v a="${times[$i]}" -v b="${times[$slowest_idx]}" 'BEGIN { print (a > b) ? 1 : 0 }') == 1 ]]; then
+            slowest_idx=$i
+        fi
+    done
+    local kept_t=()
+    local kept_r=()
+    for ((i = 0; i < ${#times[@]}; i++)); do
+        if [[ $i -ne $slowest_idx ]]; then
+            kept_t+=("${times[$i]}")
+            kept_r+=("${rsses[$i]}")
+        fi
+    done
+    # Median wall over kept runs.
+    local sorted_t
+    sorted_t=$(printf "%s\n" "${kept_t[@]}" | sort -g)
+    local n=${#kept_t[@]}
+    local mid=$((n / 2))
+    local median_t
+    if (( n % 2 == 1 )); then
+        median_t=$(echo "$sorted_t" | sed -n "$((mid + 1))p")
+    else
+        local a b
+        a=$(echo "$sorted_t" | sed -n "${mid}p")
+        b=$(echo "$sorted_t" | sed -n "$((mid + 1))p")
+        median_t=$(awk -v a="$a" -v b="$b" 'BEGIN { printf "%.3f", (a + b) / 2 }')
+    fi
+    # Max RSS across kept runs (peak memory is the natural per-run agg).
+    local max_r=0
+    for r in "${kept_r[@]}"; do
+        if (( r > max_r )); then max_r=$r; fi
+    done
+    printf "%s %s\n" "$median_t" "$max_r"
+}
+
+echo
+echo ">>> timing (RUNS=$RUNS, drop slowest, median of $((RUNS - 1)))"
+echo
+
+# Header.
+printf "%-22s | %12s | %12s | %12s | %14s | %14s\n" \
+    "workload" "gc median(s)" "bump median(s)" "overhead %" "gc max RSS(KB)" "bump max RSS(KB)"
+printf -- "-----------------------+--------------+--------------+--------------+----------------+----------------\n"
+
+for f in "${fixtures[@]}"; do
+    read -r gc_t gc_r < <(median_run "$OUTDIR/${f}_gc")
+    read -r bp_t bp_r < <(median_run "$OUTDIR/${f}_bump")
+    # overhead = (gc - bump) / bump * 100. Negative would mean GC
+    # faster, which would itself be a finding worth reporting.
+    overhead=$(awk -v g="$gc_t" -v b="$bp_t" 'BEGIN { printf "%.1f", (g - b) / b * 100 }')
+    printf "%-22s | %12s | %12s | %12s | %14s | %14s\n" \
+        "$f" "$gc_t" "$bp_t" "$overhead" "$gc_r" "$bp_r"
+done
+
+echo
+echo ">>> done"

crates/ail/src/main.rs

@@ -125,6 +125,12 @@ enum Cmd {
         /// Optimization (e.g. `-O2`); default `-O0` for debuggability.
         #[arg(long, default_value = "-O0")]
         opt: String,
+        /// Bench iter: heap allocator. `gc` (default) is Boehm
+        /// conservative GC; `bump` swaps every `@GC_malloc` for a
+        /// no-free 256 MB bump-allocator stub from `runtime/bump.c`.
+        /// The bump path is bench-only — it leaks every allocation.
+        #[arg(long, default_value = "gc", value_parser = ["gc", "bump"])]
+        alloc: String,
     },
     /// Build into a tempdir and execute. Exits with the binary's exit code.
     /// Convenience wrapper around `build` + invocation of the resulting
@@ -135,6 +141,9 @@ enum Cmd {
         /// Optimization (e.g. `-O2`); default `-O0` for debuggability.
         #[arg(long, default_value = "-O0")]
         opt: String,
+        /// Bench iter: heap allocator. See `build --alloc` for details.
+        #[arg(long, default_value = "gc", value_parser = ["gc", "bump"])]
+        alloc: String,
         /// Args passed through to the compiled program.
         #[arg(last = true)]
         args: Vec<String>,
@@ -427,21 +436,23 @@ fn main() -> Result<()> {
                 None => print!("{ir}"),
             }
         }
-        Cmd::Build { path, out, opt } => {
-            let bin = build_to(&path, out, &opt)?;
+        Cmd::Build { path, out, opt, alloc } => {
+            let strategy = parse_alloc_strategy(&alloc)?;
+            let bin = build_to(&path, out, &opt, strategy)?;
             eprintln!("built {}", bin.display());
         }
-        Cmd::Run { path, opt, args } => {
+        Cmd::Run { path, opt, alloc, args } => {
             // Iter 9b: build into a fresh tempdir per run, exec, propagate
             // exit code. The artefact dir is left around (no cleanup) so
             // it can be inspected in case of a crash; OS temp policy
             // collects them.
+            let strategy = parse_alloc_strategy(&alloc)?;
             let tmpdir = std::env::temp_dir().join(format!(
                 "ailang-run-{}",
                 std::process::id()
             ));
             std::fs::create_dir_all(&tmpdir)?;
-            let bin = build_to(&path, Some(tmpdir.join("bin")), &opt)?;
+            let bin = build_to(&path, Some(tmpdir.join("bin")), &opt, strategy)?;
             let status = std::process::Command::new(&bin)
                 .args(&args)
                 .status()
@@ -1507,6 +1518,49 @@ fn render_workspace_diff_text(r: &WorkspaceDiffReport) -> String {
     out
 }
 
+fn parse_alloc_strategy(s: &str) -> Result<ailang_codegen::AllocStrategy> {
+    match s {
+        "gc" => Ok(ailang_codegen::AllocStrategy::Gc),
+        "bump" => Ok(ailang_codegen::AllocStrategy::Bump),
+        other => anyhow::bail!("unknown --alloc value `{other}` (expected `gc` or `bump`)"),
+    }
+}
+
+/// Locate the workspace-root `runtime/bump.c` file relative to the
+/// `ail` binary. We search upwards for a directory that contains
+/// `runtime/bump.c`; that's the bench-only allocator stub. If we
+/// can't find it, the build aborts with a clear message — the
+/// `--alloc=bump` path is opt-in, so this only fires when the user
+/// asked for it.
+fn locate_bump_runtime() -> Result<PathBuf> {
+    // Two anchors we try in order:
+    //   1. the directory containing the running `ail` binary, walked
+    //      up to find `runtime/bump.c` (handles `target/release/ail`
+    //      and `target/debug/ail` cleanly).
+    //   2. the current working directory, walked up.
+    let candidates = [
+        std::env::current_exe().ok(),
+        std::env::current_dir().ok(),
+    ];
+    for start in candidates.iter().flatten() {
+        let mut cur: &Path = start.as_path();
+        loop {
+            let candidate = cur.join("runtime").join("bump.c");
+            if candidate.exists() {
+                return Ok(candidate);
+            }
+            match cur.parent() {
+                Some(p) => cur = p,
+                None => break,
+            }
+        }
+    }
+    anyhow::bail!(
+        "could not locate `runtime/bump.c` (required for --alloc=bump). \
+         Run `ail` from inside the AILang workspace."
+    )
+}
+
 /// Iter 9b: shared build helper for `Cmd::Build` and `Cmd::Run`.
 /// Loads the workspace, runs the typechecker, emits IR, and links via
 /// clang. On typecheck failure, prints diagnostics to stderr and exits
@@ -1519,7 +1573,18 @@ fn render_workspace_diff_text(r: &WorkspaceDiffReport) -> String {
 /// LetRecs that capture `Term::Let`-bound names, whose types are
 /// only known after typecheck). The lifted workspace then goes to
 /// codegen unchanged.
-fn build_to(path: &Path, out: Option<PathBuf>, opt: &str) -> Result<PathBuf> {
+///
+/// Bench iter: `alloc` selects the heap allocator the emitted IR
+/// targets. Default `Gc` keeps the entire pipeline (IR text, link
+/// command) byte-identical to pre-bench. `Bump` declares
+/// `@bump_malloc` instead of `@GC_malloc` and links `runtime/bump.c`
+/// in lieu of `-lgc`.
+fn build_to(
+    path: &Path,
+    out: Option<PathBuf>,
+    opt: &str,
+    alloc: ailang_codegen::AllocStrategy,
+) -> Result<PathBuf> {
     let ws = ailang_core::load_workspace(path)?;
     let diags = ailang_check::check_workspace(&ws);
     if !diags.is_empty() {
@@ -1556,23 +1621,49 @@ fn build_to(path: &Path, out: Option<PathBuf>, opt: &str) -> Result<PathBuf> {
         modules: lifted_modules,
         root_dir: ws.root_dir.clone(),
     };
-    let ir = ailang_codegen::lower_workspace(&ws)?;
+    let ir = ailang_codegen::lower_workspace_with_alloc(&ws, alloc)?;
     let tmpdir = std::env::temp_dir().join(format!("ailang-{}", std::process::id()));
     std::fs::create_dir_all(&tmpdir)?;
     let ll_path = tmpdir.join(format!("{}.ll", ws.entry));
     std::fs::write(&ll_path, &ir)?;
     let out_bin = out.unwrap_or_else(|| Path::new(".").join(&ws.entry).with_extension(""));
-    let status = std::process::Command::new("clang")
-        .arg(opt)
-        .arg("-o")
-        .arg(&out_bin)
-        .arg(&ll_path)
-        // Boehm conservative GC (Decision 9 / Iter 14f). The lowered IR
-        // calls @GC_malloc; libgc supplies it. Pthread/dl are pulled in
-        // transitively via libgc.so on Linux, so a single -lgc suffices.
-        .arg("-lgc")
-        .status()
-        .context("running clang")?;
+    let mut clang = std::process::Command::new("clang");
+    clang.arg(opt).arg("-o").arg(&out_bin).arg(&ll_path);
+    match alloc {
+        ailang_codegen::AllocStrategy::Gc => {
+            // Boehm conservative GC (Decision 9 / Iter 14f). The lowered
+            // IR calls @GC_malloc; libgc supplies it. Pthread/dl are
+            // pulled in transitively via libgc.so on Linux, so a single
+            // -lgc suffices.
+            clang.arg("-lgc");
+        }
+        ailang_codegen::AllocStrategy::Bump => {
+            // Bench iter: link the no-free arena stub from
+            // `runtime/bump.c` instead of libgc. We compile the stub
+            // inline at -O2 (its body is small, the .o is cached at
+            // <tmpdir>/bump.o per build invocation; no global cache
+            // because the bench harness rebuilds binaries top-to-bottom
+            // anyway).
+            let bump_src = locate_bump_runtime()?;
+            let bump_obj = tmpdir.join("bump.o");
+            let cstatus = std::process::Command::new("clang")
+                .arg("-O2")
+                .arg("-c")
+                .arg(&bump_src)
+                .arg("-o")
+                .arg(&bump_obj)
+                .status()
+                .context("compiling runtime/bump.c")?;
+            if !cstatus.success() {
+                anyhow::bail!(
+                    "clang failed compiling bump.c (status {})",
+                    cstatus
+                );
+            }
+            clang.arg(&bump_obj);
+        }
+    }
+    let status = clang.status().context("running clang")?;
     if !status.success() {
         anyhow::bail!(
             "clang failed (status {}); ll at {}",

crates/ailang-codegen/src/lib.rs

@@ -110,6 +110,35 @@ pub enum CodegenError {
 
 type Result<T> = std::result::Result<T, CodegenError>;
 
+/// Bench iter: which heap-allocation runtime the emitted IR targets.
+///
+/// `Gc` is the default (Boehm conservative GC, Decision 9 / Iter 14f).
+/// `Bump` swaps every `@GC_malloc` for `@bump_malloc`, which is supplied
+/// by `runtime/bump.c` — a no-free, statically-sized arena allocator
+/// used purely to quantify the GC's overhead via an A/B comparison.
+/// The IR is otherwise byte-identical between the two strategies.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum AllocStrategy {
+    Gc,
+    Bump,
+}
+
+impl Default for AllocStrategy {
+    fn default() -> Self {
+        AllocStrategy::Gc
+    }
+}
+
+impl AllocStrategy {
+    /// LLVM IR-level name of the allocator fn (without leading `@`).
+    fn fn_name(self) -> &'static str {
+        match self {
+            AllocStrategy::Gc => "GC_malloc",
+            AllocStrategy::Bump => "bump_malloc",
+        }
+    }
+}
+
 /// Single-module entry point. Lowers `m` to a `.ll` string with `m`
 /// itself as the entry module. Returns the full LLVM IR text, ready to
 /// be written to disk and handed to `clang`.
@@ -145,6 +174,16 @@ pub fn emit_ir(m: &Module) -> Result<String> {
     lower_workspace(&ws)
 }
 
+/// Bench iter: variant of [`lower_workspace`] that selects the heap
+/// allocator at codegen time. `AllocStrategy::Gc` produces IR
+/// byte-identical to [`lower_workspace`]; `AllocStrategy::Bump` swaps
+/// every `@GC_malloc` site for `@bump_malloc` (supplied by
+/// `runtime/bump.c`). Used by `ail build --alloc=bump` to quantify the
+/// GC's runtime overhead via an A/B comparison.
+pub fn lower_workspace_with_alloc(ws: &Workspace, alloc: AllocStrategy) -> Result<String> {
+    lower_workspace_inner(ws, alloc)
+}
+
 /// Multi-module entry point. Lowers an entire [`Workspace`] (entry
 /// module plus its transitive imports, as produced by
 /// `ailang_core::load_workspace`) to a single `.ll` string and emits
@@ -171,6 +210,10 @@ pub fn emit_ir(m: &Module) -> Result<String> {
 /// Use [`emit_ir`] for the single-file shortcut when there are no
 /// imports.
 pub fn lower_workspace(ws: &Workspace) -> Result<String> {
+    lower_workspace_inner(ws, AllocStrategy::Gc)
+}
+
+fn lower_workspace_inner(ws: &Workspace, alloc: AllocStrategy) -> Result<String> {
     // Iter 16a: desugar every module before any lowering work runs.
     // The pass is idempotent and structurally identical to what
     // `ailang-check` runs at its public entries, so the codegen
@@ -293,6 +336,7 @@ pub fn lower_workspace(ws: &Workspace) -> Result<String> {
             &module_ctor_index,
             &module_consts,
             import_map,
+            alloc,
         );
         emitter
             .emit_module()
@@ -354,7 +398,11 @@ pub fn lower_workspace(ws: &Workspace) -> Result<String> {
 
     out.push_str("declare i32 @printf(ptr, ...)\n");
     out.push_str("declare i32 @puts(ptr)\n");
-    out.push_str("declare ptr @GC_malloc(i64)\n");
+    // Bench iter: the allocator declaration name follows `alloc`.
+    // Default `Gc` keeps the emitted IR byte-identical to the pre-bench
+    // pipeline; `Bump` declares `@bump_malloc` instead, supplied by
+    // `runtime/bump.c` and linked in lieu of `-lgc`.
+    out.push_str(&format!("declare ptr @{}(i64)\n", alloc.fn_name()));
     // Iter 16e: `==` on `Str` lowers to `@strcmp` followed by
     // `icmp eq i32 0`. NUL-terminated strings make this a one-liner;
     // libc supplies `strcmp` so no extra link flag is needed.
@@ -465,6 +513,10 @@ struct Emitter<'a> {
     /// Populated by `analyze_fn_body` at the start of `emit_fn` and at
     /// the start of every lambda thunk emission inside `lower_lambda`.
     non_escape: NonEscapeSet,
+    /// Bench iter: which allocator the heap-allocation paths target.
+    /// Decided at the top-level entry point (`lower_workspace_inner`)
+    /// and propagated to every site that emits a `call ptr @<alloc>(...)`.
+    alloc: AllocStrategy,
 }
 
 #[derive(Debug, Clone)]
@@ -514,6 +566,7 @@ impl<'a> Emitter<'a> {
         module_ctor_index: &'a BTreeMap<String, BTreeMap<String, CtorRef>>,
         module_consts: &'a BTreeMap<String, BTreeMap<String, ConstDef>>,
         import_map: BTreeMap<String, String>,
+        alloc: AllocStrategy,
     ) -> Self {
         let mut types: BTreeMap<String, Vec<CtorInfo>> = BTreeMap::new();
         for def in &module.defs {
@@ -568,6 +621,7 @@ impl<'a> Emitter<'a> {
             lam_counter: 0,
             deferred_thunks: Vec::new(),
             non_escape: NonEscapeSet::new(),
+            alloc,
         }
     }
 
@@ -1296,7 +1350,8 @@ impl<'a> Emitter<'a> {
             ));
         } else {
             self.body.push_str(&format!(
-                "  {p} = call ptr @GC_malloc(i64 {size_bytes})\n"
+                "  {p} = call ptr @{}(i64 {size_bytes})\n",
+                self.alloc.fn_name()
             ));
         }
         // Write tag.
@@ -2093,7 +2148,8 @@ impl<'a> Emitter<'a> {
                 ));
             } else {
                 self.body.push_str(&format!(
-                    "  {env} = call ptr @GC_malloc(i64 {env_size})\n"
+                    "  {env} = call ptr @{}(i64 {env_size})\n",
+                    self.alloc.fn_name()
                 ));
             }
             for (i, (_cname, outer_ssa, cty, _c_ail, _sig)) in cap_meta.iter().enumerate() {
@@ -2116,8 +2172,10 @@ impl<'a> Emitter<'a> {
             self.body
                 .push_str(&format!("  {clos} = alloca i8, i64 16, align 8\n"));
         } else {
-            self.body
-                .push_str(&format!("  {clos} = call ptr @GC_malloc(i64 16)\n"));
+            self.body.push_str(&format!(
+                "  {clos} = call ptr @{}(i64 16)\n",
+                self.alloc.fn_name()
+            ));
         }
         let cs_t = self.fresh_ssa();
         self.body.push_str(&format!(

docs/JOURNAL.md

@@ -5129,3 +5129,196 @@ the Observations section above are explicitly NOT queued —
 they require user sign-off on whether the project's GC
 direction is "improve precision of stack-alloca", "replace
 Boehm with a precise collector", or something else entirely.
+
+## Bench — GC overhead via bump-allocator comparison
+
+Single-purpose data-gathering iter, not a feature. Goal: quantify
+how much of the runtime spent by AILang programs is paid to the
+Boehm conservative collector by comparing the same program built
+two ways — `--alloc=gc` (default, current behavior, links `-lgc`)
+and `--alloc=bump` (a no-free 256 MB statically-allocated bump
+arena from `runtime/bump.c`). The IR text for the two builds is
+byte-identical except that every `@GC_malloc` callsite and the
+`declare ptr @GC_malloc(i64)` declaration become `@bump_malloc`.
+The link command swaps `-lgc` for `runtime/bump.o`. Nothing else
+changes.
+
+### Methodology
+
+Two fixtures, both designed to drive heap allocation hard enough
+that the collector / arena is on the hot path:
+
+- **`examples/bench_list_sum.ail.json`**. Local `IntList` ADT.
+  Builds three lists (lengths 100k / 1M / 3M) by tail-recursive
+  `cons_n_acc`, sums each via tail-recursive `sum_acc`, prints
+  the three sums. Both build and sum are written in accumulator
+  form with `tail-app` because at 3M elements a non-tail
+  recursion overflows the 8 MB system stack. Each `ICons` cell is
+  24 B; total heap traffic ≈ 99 MB across the run (the bump
+  arena's 256 MB ceiling was the constraint that capped the
+  largest size at 3M, not 10M).
+- **`examples/bench_tree_walk.ail.json`**. Local `Tree` ADT
+  (`Leaf | Node Int Tree Tree`). Builds and sums balanced trees
+  of depth 16 / 18 / 20. At depth 20 the tree has 2^20 − 1 nodes,
+  ~32 B per `Node`, ~64 MB heap traffic for the depth-20 phase
+  alone. Recursion in `build_tree` / `sum_tree` is constructor-
+  blocked so it cannot be `tail-app`'d, but the recursion depth
+  equals the tree depth (≤ 20), so it fits trivially.
+
+Build configuration: `clang -O2`, both modes. The harness
+(`bench/run.sh`) runs each binary 5 times under a Python wrapper
+that reads `getrusage(RUSAGE_CHILDREN).ru_maxrss` for peak RSS
+and `time.monotonic()` deltas around `subprocess.Popen.wait` for
+wall time. Slowest run is dropped, median wall over the kept 4 is
+reported. The harness runs `cargo build --release -p ail` first,
+then compiles each `(fixture, mode)` pair once before the timing
+loop, so build time is excluded from measurements.
+
+### Numbers (Linux 7.0.3-1-cachyos, single machine, RUNS=5)
+
+```
+workload               | gc median(s) | bump median(s) |   overhead % | gc max RSS(KB) | bump max RSS(KB)
+-----------------------+--------------+--------------+--------------+----------------+----------------
+bench_list_sum         |        0.145 |        0.050 |        190.0 |         103788 |          97640
+bench_tree_walk        |        0.105 |        0.038 |        176.3 |          73452 |          55452
+```
+
+A second run with RUNS=9 (median of 8) corroborates within noise:
+
+```
+bench_list_sum         |        0.141 |        0.048 |        193.7 |         103784 |          97884
+bench_tree_walk        |        0.103 |        0.039 |        164.1 |          73448 |          55396
+```
+
+Overhead = `(gc - bump) / bump * 100` — i.e. the GC-mode runtime is
+~2.7–2.9× the bump-mode runtime. Equivalently, ~63–65 % of the
+GC-mode wall time is GC overhead (collector pauses + write
+barriers + allocation-path complexity vs. a single bump pointer).
+
+### Bucket
+
+**Large.** GC takes roughly two-thirds of total runtime on these
+allocation-heavy workloads. For comparison, the typical Boehm
+conservative-GC overhead reported in the literature on
+allocation-heavy workloads sits in the 20–60 % range; ~190 % puts
+this firmly past that envelope. Caveat below.
+
+### Caveats
+
+- **Single-machine measurement.** No cross-machine confirmation,
+  no isolation from background load. Variance across the kept-4
+  runs was ≤ 5 ms in absolute terms, but a bigger machine /
+  smaller machine / different libgc version could shift these
+  numbers materially.
+- **Allocation-heavy workloads.** Both fixtures spend almost their
+  entire runtime in the allocator (Cons cell construction, Node
+  cell construction). Real programs that compute as well as
+  allocate would have a smaller GC-overhead share. The numbers
+  here are therefore an *upper bound* on the GC's share of any
+  realistic workload.
+- **Bump leaks everything.** The bump-mode binary never frees a
+  byte; max RSS reflects the working-set after every allocation
+  the program ever made, plus committed pages from the 256 MB
+  arena. For `bench_list_sum`'s 99 MB heap traffic, GC's heap
+  (~100 MB RSS) is essentially identical to bump's (~97 MB).
+  Where the workload actually leaks past bump's arena (~256 MB
+  cells × any factor), GC would win on RSS by reusing freed
+  memory; this bench does not exhibit that regime.
+- **No warmup theatrics.** Each timed run is a cold process start.
+  AILang has no JIT and no per-process allocation-path tuning,
+  so first-run / steady-state distinction does not apply here.
+  Variance was within noise even on the first kept run.
+- **Hardcoded N.** No env-var / argv plumbing in AILang yet, so
+  the workload sizes are baked into the source. The three sizes
+  per fixture provide enough variety to detect a wildly
+  size-dependent overhead (none observed — both fixtures show a
+  flat ~2.8x ratio across all three calls).
+- **The bench measures `GC_malloc` overhead, not full GC.** Boehm's
+  collector runs inline on allocation when the heap grows past a
+  threshold; we never observe it as a separate cost. A program
+  with a long-lived heap that causes repeated full marks would
+  see a different (likely larger) overhead share. Neither fixture
+  here triggers that.
+
+### Implementation summary
+
+- `crates/ailang-codegen/src/lib.rs`: new public `AllocStrategy`
+  enum (`Gc` / `Bump`); new public entry `lower_workspace_with_alloc`;
+  `lower_workspace` delegates to it with `Gc`. The single
+  declaration line and the three `@GC_malloc` callsites
+  (`lower_ctor`, lambda env, closure pair) all read the
+  Emitter's `alloc` field.
+- `crates/ail/src/main.rs`: `--alloc=<gc|bump>` flag added to
+  both `build` and `run`, default `gc`. Threaded into a now-four-
+  arg `build_to`; on `Bump`, the helper `locate_bump_runtime()`
+  walks up from the binary path / cwd to find `runtime/bump.c`,
+  compiles it inline (`clang -O2 -c`) into a tempdir-scoped
+  `bump.o`, and links that instead of `-lgc`.
+- `runtime/bump.c`: 256 MB static arena, single bump pointer,
+  8-byte alignment, `abort()` on overflow. Single function
+  `void *bump_malloc(size_t)`.
+- `examples/bench_list_sum.{ailx,ail.json}` and
+  `examples/bench_tree_walk.{ailx,ail.json}`: the two fixtures
+  described above. List builder rewritten to accumulator form
+  to fit in 8 MB stack at 3M elements.
+- `bench/run.sh`: harness as specified. Python helper for
+  monotonic clock + RUSAGE_CHILDREN max RSS (avoids the
+  `/usr/bin/time` dependency, which is not on Arch by default).
+  `awk` replaces `bc` for the same reason.
+- `docs/DESIGN.md` not touched. The CLI flag is opt-in, the
+  default behavior is identical to pre-bench, and the bump path
+  is bench-only — it does not deserve language-spec status.
+
+### Cross-iter regression verified
+
+- Default `--alloc=gc` is byte-identical to pre-bench. The five
+  IR snapshots (`hello`, `sum`, `list`, `max3`, `ws_main`) pass
+  unchanged. All workspace tests pass: 141 → 141 (no test
+  count delta from this iter; no e2e additions).
+- Manual smoke run of representative existing fixtures
+  (`sum`, `list`, `list_map`, `gc_stress`, `std_list_demo`,
+  `escape_local_demo`) under `--alloc=gc` produces identical
+  stdout to the documented expected outputs.
+- The bump-mode IR, after a textual `s/GC_malloc/bump_malloc/g`
+  on the gc-mode IR, is `diff`-clean against a real `--alloc=bump`
+  build. The IR is byte-identical except for the allocator
+  symbol name.
+
+### Did anything surprise
+
+- **The overhead is large.** ~2.8x slowdown is at the high end of
+  what one expects for a modern conservative collector on
+  allocation-heavy code. Two factors likely contributing: (a)
+  Boehm's `GC_malloc` does conservative root scanning of the
+  C stack on every collection — for workloads that allocate
+  heavily, the collector triggers often; (b) AILang's escape
+  analysis (Iter 17a) flags 0 of 270 ctor sites in shipped code
+  as non-escaping, and 0 of the allocations in either bench
+  fixture, so the entire allocation traffic goes through the
+  collector. Workloads that converted more allocations to
+  `alloca` would see a smaller GC share.
+- **No segfault from the bump leak.** The bump arena is
+  256 MB; the heaviest workload (3M-element list) consumes
+  ~99 MB. We have headroom even at the largest configured size.
+  10M elements (the original spec value) would have been
+  240 MB — uncomfortably close to the ceiling, justified the
+  reduction to 3M.
+- **GC's max RSS is barely larger than bump's.** I expected GC's
+  max RSS to be substantially smaller than bump's (because GC
+  reclaims dead memory). It isn't — the bump fixtures' working
+  sets are simply not large enough to pressure the collector
+  into reclaiming much. The list fixture builds the entire
+  3M-element list before summing, so all allocations are live
+  at once anyway. Different workloads (e.g. a fold that builds
+  intermediate lists discarded between iterations) would surface
+  the RSS gap.
+- **Tail-call discipline matters.** The original spec's "tail-
+  recursive sum" is a misnomer for `sum_list (Cons h t) = h +
+  sum_list t` — that's constructor-blocked, not tail-recursive.
+  Naïvely transcribing the spec produced a binary that
+  segfaulted at 3M elements. Both fixtures' linear-recursion fns
+  had to be rewritten in accumulator form with explicit
+  `tail-app` markers. Captured here because it is a real
+  consequence of how AILang is structured: an LLM author who
+  ports a textbook recursive sum into AILang at scale will
+  hit the stack ceiling unless they know about Decision 8.

examples/bench_list_sum.ail.json

@@ -0,0 +1 @@
+{"defs":[{"ctors":[{"fields":[],"name":"INil"},{"fields":[{"k":"con","name":"Int"},{"k":"con","name":"IntList"}],"name":"ICons"}],"kind":"type","name":"IntList"},{"body":{"cond":{"args":[{"name":"n","t":"var"},{"lit":{"kind":"int","value":0},"t":"lit"}],"fn":{"name":"==","t":"var"},"t":"app"},"else":{"args":[{"args":[{"name":"n","t":"var"},{"lit":{"kind":"int","value":1},"t":"lit"}],"fn":{"name":"-","t":"var"},"t":"app"},{"args":[{"args":[{"name":"n","t":"var"},{"lit":{"kind":"int","value":1},"t":"lit"}],"fn":{"name":"-","t":"var"},"t":"app"},{"name":"acc","t":"var"}],"ctor":"ICons","t":"ctor","type":"IntList"}],"fn":{"name":"cons_n_acc","t":"var"},"t":"app","tail":true},"t":"if","then":{"name":"acc","t":"var"}},"doc":"Tail-recursive list builder. Result = accumulator-prepended list.","kind":"fn","name":"cons_n_acc","params":["n","acc"],"type":{"effects":[],"k":"fn","params":[{"k":"con","name":"Int"},{"k":"con","name":"IntList"}],"ret":{"k":"con","name":"IntList"}}},{"body":{"args":[{"name":"n","t":"var"},{"args":[],"ctor":"INil","t":"ctor","type":"IntList"}],"fn":{"name":"cons_n_acc","t":"var"},"t":"app"},"doc":"Build [0, 1, ..., n-1] :: IntList. Order doesn't matter for sum.","kind":"fn","name":"cons_n","params":["n"],"type":{"effects":[],"k":"fn","params":[{"k":"con","name":"Int"}],"ret":{"k":"con","name":"IntList"}}},{"body":{"arms":[{"body":{"name":"acc","t":"var"},"pat":{"ctor":"INil","fields":[],"p":"ctor"}},{"body":{"args":[{"name":"t","t":"var"},{"args":[{"name":"acc","t":"var"},{"name":"h","t":"var"}],"fn":{"name":"+","t":"var"},"t":"app"}],"fn":{"name":"sum_acc","t":"var"},"t":"app","tail":true},"pat":{"ctor":"ICons","fields":[{"name":"h","p":"var"},{"name":"t","p":"var"}],"p":"ctor"}}],"scrutinee":{"name":"xs","t":"var"},"t":"match"},"doc":"Tail-recursive sum.","kind":"fn","name":"sum_acc","params":["xs","acc"],"type":{"effects":[],"k":"fn","params":[{"k":"con","name":"IntList"},{"k":"con","name":"Int"}],"ret":{"k":"con","name":"Int"}}},{"body":{"args":[{"name":"xs","t":"var"},{"lit":{"kind":"int","value":0},"t":"lit"}],"fn":{"name":"sum_acc","t":"var"},"t":"app"},"doc":"Sum every element. Calls sum_acc with seed 0.","kind":"fn","name":"sum_list","params":["xs"],"type":{"effects":[],"k":"fn","params":[{"k":"con","name":"IntList"}],"ret":{"k":"con","name":"Int"}}},{"body":{"args":[{"args":[{"args":[{"name":"n","t":"var"}],"fn":{"name":"cons_n","t":"var"},"t":"app"}],"fn":{"name":"sum_list","t":"var"},"t":"app"}],"op":"io/print_int","t":"do"},"doc":"Build a list of length n, sum it, print the sum.","kind":"fn","name":"run_one","params":["n"],"type":{"effects":["IO"],"k":"fn","params":[{"k":"con","name":"Int"}],"ret":{"k":"con","name":"Unit"}}},{"body":{"lhs":{"args":[{"lit":{"kind":"int","value":100000},"t":"lit"}],"fn":{"name":"run_one","t":"var"},"t":"app"},"rhs":{"lhs":{"args":[{"lit":{"kind":"int","value":1000000},"t":"lit"}],"fn":{"name":"run_one","t":"var"},"t":"app"},"rhs":{"args":[{"lit":{"kind":"int","value":3000000},"t":"lit"}],"fn":{"name":"run_one","t":"var"},"t":"app"},"t":"seq"},"t":"seq"},"kind":"fn","name":"main","params":[],"type":{"effects":["IO"],"k":"fn","params":[],"ret":{"k":"con","name":"Unit"}}}],"imports":[],"name":"bench_list_sum","schema":"ailang/v0"}

examples/bench_list_sum.ailx

@@ -0,0 +1,99 @@
+; Bench fixture (Bench iter): linked-list build + sum.
+;
+; Drives the heap allocator hard via a recursive Cons spine. Every
+; ICons cell is one allocation (24 bytes: tag + Int payload + tail).
+; A list of length N therefore costs N allocations. We invoke the
+; workload at three different sizes within a single program run to
+; cover small / medium / large territory.
+;
+; Both `cons_n` (build) and `sum_list` (traverse) are written in
+; accumulator form so the recursive call sits in tail position and
+; can be marked `tail-app`. This is essential at the sizes used here:
+; without `musttail`, three million stack frames overflow the default
+; thread stack and segfault.
+;
+; Workload sizes (hardcoded — AILang has no env-var/argv pipeline):
+;
+;   100_000   * 24 B =   2.4 MB
+;   1_000_000 * 24 B =  24   MB
+;   3_000_000 * 24 B =  72   MB
+;
+; (3M was chosen as the largest size that comfortably fits inside the
+; bump allocator's 256 MB arena with headroom for closure pairs and
+; misc allocations.)
+;
+; Build is allocation-heavy; sum is pure traversal of already-allocated
+; heap (the interesting one for GC pressure / barrier overhead).
+;
+; Expected stdout (one int per line, the sum 0+1+...+(N-1) = N*(N-1)/2):
+;   100_000   -> 4999950000
+;   1_000_000 -> 499999500000
+;   3_000_000 -> 4499998500000
+
+(module bench_list_sum
+
+  (data IntList
+    (ctor INil)
+    (ctor ICons (con Int) (con IntList)))
+
+  (fn cons_n_acc
+    (doc "Tail-recursive list builder. Result = accumulator-prepended list.")
+    (type
+      (fn-type
+        (params (con Int) (con IntList))
+        (ret (con IntList))))
+    (params n acc)
+    (body
+      (if (app == n 0)
+        acc
+        (tail-app cons_n_acc
+          (app - n 1)
+          (term-ctor IntList ICons (app - n 1) acc)))))
+
+  (fn cons_n
+    (doc "Build [0, 1, ..., n-1] :: IntList. Order doesn't matter for sum.")
+    (type
+      (fn-type
+        (params (con Int))
+        (ret (con IntList))))
+    (params n)
+    (body
+      (app cons_n_acc n (term-ctor IntList INil))))
+
+  (fn sum_acc
+    (doc "Tail-recursive sum.")
+    (type
+      (fn-type
+        (params (con IntList) (con Int))
+        (ret (con Int))))
+    (params xs acc)
+    (body
+      (match xs
+        (case (pat-ctor INil) acc)
+        (case (pat-ctor ICons h t)
+          (tail-app sum_acc t (app + acc h))))))
+
+  (fn sum_list
+    (doc "Sum every element. Calls sum_acc with seed 0.")
+    (type
+      (fn-type
+        (params (con IntList))
+        (ret (con Int))))
+    (params xs)
+    (body
+      (app sum_acc xs 0)))
+
+  (fn run_one
+    (doc "Build a list of length n, sum it, print the sum.")
+    (type (fn-type (params (con Int)) (ret (con Unit)) (effects IO)))
+    (params n)
+    (body
+      (do io/print_int (app sum_list (app cons_n n)))))
+
+  (fn main
+    (type (fn-type (params) (ret (con Unit)) (effects IO)))
+    (params)
+    (body
+      (seq (app run_one 100000)
+      (seq (app run_one 1000000)
+           (app run_one 3000000))))))

examples/bench_tree_walk.ail.json

@@ -0,0 +1 @@
+{"defs":[{"ctors":[{"fields":[],"name":"Leaf"},{"fields":[{"k":"con","name":"Int"},{"k":"con","name":"Tree"},{"k":"con","name":"Tree"}],"name":"Node"}],"kind":"type","name":"Tree"},{"body":{"cond":{"args":[{"name":"depth","t":"var"},{"lit":{"kind":"int","value":0},"t":"lit"}],"fn":{"name":"==","t":"var"},"t":"app"},"else":{"args":[{"lit":{"kind":"int","value":1},"t":"lit"},{"args":[{"args":[{"name":"depth","t":"var"},{"lit":{"kind":"int","value":1},"t":"lit"}],"fn":{"name":"-","t":"var"},"t":"app"}],"fn":{"name":"build_tree","t":"var"},"t":"app"},{"args":[{"args":[{"name":"depth","t":"var"},{"lit":{"kind":"int","value":1},"t":"lit"}],"fn":{"name":"-","t":"var"},"t":"app"}],"fn":{"name":"build_tree","t":"var"},"t":"app"}],"ctor":"Node","t":"ctor","type":"Tree"},"t":"if","then":{"args":[],"ctor":"Leaf","t":"ctor","type":"Tree"}},"doc":"Balanced binary tree of given depth, every value = 1.","kind":"fn","name":"build_tree","params":["depth"],"type":{"effects":[],"k":"fn","params":[{"k":"con","name":"Int"}],"ret":{"k":"con","name":"Tree"}}},{"body":{"arms":[{"body":{"lit":{"kind":"int","value":0},"t":"lit"},"pat":{"ctor":"Leaf","fields":[],"p":"ctor"}},{"body":{"args":[{"name":"v","t":"var"},{"args":[{"args":[{"name":"l","t":"var"}],"fn":{"name":"sum_tree","t":"var"},"t":"app"},{"args":[{"name":"r","t":"var"}],"fn":{"name":"sum_tree","t":"var"},"t":"app"}],"fn":{"name":"+","t":"var"},"t":"app"}],"fn":{"name":"+","t":"var"},"t":"app"},"pat":{"ctor":"Node","fields":[{"name":"v","p":"var"},{"name":"l","p":"var"},{"name":"r","p":"var"}],"p":"ctor"}}],"scrutinee":{"name":"t","t":"var"},"t":"match"},"doc":"Sum every Node value via match recursion. Constructor-blocked: not tail-recursive, but recursion depth = tree depth so fits.","kind":"fn","name":"sum_tree","params":["t"],"type":{"effects":[],"k":"fn","params":[{"k":"con","name":"Tree"}],"ret":{"k":"con","name":"Int"}}},{"body":{"args":[{"args":[{"args":[{"name":"depth","t":"var"}],"fn":{"name":"build_tree","t":"var"},"t":"app"}],"fn":{"name":"sum_tree","t":"var"},"t":"app"}],"op":"io/print_int","t":"do"},"doc":"Build a tree of given depth, sum it, print the sum.","kind":"fn","name":"run_one","params":["depth"],"type":{"effects":["IO"],"k":"fn","params":[{"k":"con","name":"Int"}],"ret":{"k":"con","name":"Unit"}}},{"body":{"lhs":{"args":[{"lit":{"kind":"int","value":16},"t":"lit"}],"fn":{"name":"run_one","t":"var"},"t":"app"},"rhs":{"lhs":{"args":[{"lit":{"kind":"int","value":18},"t":"lit"}],"fn":{"name":"run_one","t":"var"},"t":"app"},"rhs":{"args":[{"lit":{"kind":"int","value":20},"t":"lit"}],"fn":{"name":"run_one","t":"var"},"t":"app"},"t":"seq"},"t":"seq"},"kind":"fn","name":"main","params":[],"type":{"effects":["IO"],"k":"fn","params":[],"ret":{"k":"con","name":"Unit"}}}],"imports":[],"name":"bench_tree_walk","schema":"ailang/v0"}

examples/bench_tree_walk.ailx

@@ -0,0 +1,78 @@
+; Bench fixture (Bench iter): balanced-tree build + sum.
+;
+; Allocates a balanced binary tree of `Node value left right` cells,
+; then walks it summing the value field. Distinct from
+; `bench_list_sum` in two ways:
+;   1. Each tree node has an additional pointer field versus a list
+;      cell — 32-byte alloc instead of 24-byte. The branching shape
+;      means the recursion structure is genuinely tree-shaped: the
+;      build cannot be made tail-recursive without explicit
+;      continuation passing, so this fixture is constrained to depths
+;      where the recursion stack fits.
+;   2. The traversal pattern hits two children per node, exercising
+;      the GC's mark-phase pointer-chasing heuristics differently from
+;      a plain linked-list walk.
+;
+; Depth picked: 20  -> 2^20 - 1 = 1_048_575 nodes -> 32 MB heap usage.
+; Recursion depth in `build_tree` and `sum_tree` matches `depth`,
+; which fits comfortably in the default 8 MB system stack.
+;
+; Hardcoded multi-call form: build/sum the same tree thrice for
+; signal averaging. The depths are different per call so the GC has
+; to deal with three independent live-set sizes.
+;
+; Expected stdout (one int per line):
+;   depth 16: 2^16 - 1 = 65535 nodes, sum = 65535
+;   depth 18: 2^18 - 1 = 262143 nodes, sum = 262143
+;   depth 20: 2^20 - 1 = 1048575 nodes, sum = 1048575
+;
+; (Each node stores literal `1`; sum is therefore node count.)
+
+(module bench_tree_walk
+
+  (data Tree
+    (ctor Leaf)
+    (ctor Node (con Int) (con Tree) (con Tree)))
+
+  (fn build_tree
+    (doc "Balanced binary tree of given depth, every value = 1.")
+    (type
+      (fn-type
+        (params (con Int))
+        (ret (con Tree))))
+    (params depth)
+    (body
+      (if (app == depth 0)
+        (term-ctor Tree Leaf)
+        (term-ctor Tree Node
+          1
+          (app build_tree (app - depth 1))
+          (app build_tree (app - depth 1))))))
+
+  (fn sum_tree
+    (doc "Sum every Node value via match recursion. Constructor-blocked: not tail-recursive, but recursion depth = tree depth so fits.")
+    (type
+      (fn-type
+        (params (con Tree))
+        (ret (con Int))))
+    (params t)
+    (body
+      (match t
+        (case (pat-ctor Leaf) 0)
+        (case (pat-ctor Node v l r)
+          (app + v (app + (app sum_tree l) (app sum_tree r)))))))
+
+  (fn run_one
+    (doc "Build a tree of given depth, sum it, print the sum.")
+    (type (fn-type (params (con Int)) (ret (con Unit)) (effects IO)))
+    (params depth)
+    (body
+      (do io/print_int (app sum_tree (app build_tree depth)))))
+
+  (fn main
+    (type (fn-type (params) (ret (con Unit)) (effects IO)))
+    (params)
+    (body
+      (seq (app run_one 16)
+      (seq (app run_one 18)
+           (app run_one 20))))))

runtime/bump.c

@@ -0,0 +1,40 @@
+/* Bench-only bump allocator stub.
+ *
+ * Used by `ail build --alloc=bump` (Bench iter) to A/B compare AILang
+ * binaries against the default Boehm-GC build. The whole runtime is a
+ * 256 MB statically-allocated arena and a single bump pointer; there
+ * is no `free`, no scan, no anything. If the workload exceeds 256 MB
+ * we abort — this is bench code, the right response to overflow is to
+ * notice and pick a smaller workload.
+ *
+ * The signature mirrors `GC_malloc` from libgc: `void *bump_malloc(size_t)`.
+ * The codegen replaces every `call ptr @GC_malloc` with
+ * `call ptr @bump_malloc` when `--alloc=bump` is set, so the AILang IR
+ * is otherwise byte-identical between the two strategies.
+ */
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#define ARENA_BYTES (256ul * 1024ul * 1024ul)
+
+static uint8_t arena[ARENA_BYTES];
+static size_t cursor = 0;
+
+void *bump_malloc(size_t n) {
+    /* Align bump pointer up to 8 bytes — AILang's allocations are all
+     * 8-byte aligned (tag + 8-byte fields, env pointers, closure pairs
+     * of two `ptr`s). Matches the alignment Boehm gives us. */
+    size_t aligned = (cursor + 7ul) & ~((size_t)7ul);
+    if (aligned + n > ARENA_BYTES) {
+        fprintf(stderr,
+                "bump_malloc: arena exhausted (cursor=%zu, requested=%zu, arena=%zu)\n",
+                aligned, n, (size_t)ARENA_BYTES);
+        abort();
+    }
+    void *p = (void *)(arena + aligned);
+    cursor = aligned + n;
+    return p;
+}