allocator stats inline + mremap XL fast path + bench harness honesty

Stats accounting moved out of the Allocator.c _dyn dispatch wrappers
and into each typed backend body (Heap, Page, Arena, Slab, Budget,
Debug). The dyn wrappers no longer carry parallel updates -- duplicate
writes would double-count. Reads use new direct accessor macros in
<Misra/Std/Allocator.h>:

    AllocatorBytesInUse(a), AllocatorBytesRequested(a),
    AllocatorPeakBytesInUse(a), AllocatorAllocations(a),
    AllocatorReallocations(a), AllocatorDeallocations(a),
    AllocatorFailedAllocations(a)

Same `((void)0, ptr->field)` shape as BufLength / VecCapacity. No
AllocatorGetStats(...) function call, no _Generic-dispatched reader.
ALLOCATOR_OF(a) does the typed -> base cast at the macro level so each
read compiles down to a single field load. AllocatorResetStats stays
a function (it mutates).

Heap remap gains an mremap fast path on the XL (mmap-per-alloc) class.
On Linux direct-syscall builds, os_page_remap calls mremap with
MREMAP_MAYMOVE and the xl[] descriptor gets re-inserted at the new
sorted position. Darwin / Windows have no equivalent; those fall back
to alloc-new + memcpy. Drops BM_ReallocGrow on heap-based misra
columns from 247 us to 16 us (15x speedup), matching mimalloc and
tcmalloc.

Bench harness corrections:
- Allocator_misra_correct.c picks PageAllocator above 4 KiB instead
  of falling back to HeapAllocator -- the latter was the "wrong tool"
  path that misra-correct is supposed to avoid. AllocFreePair at
  16/64 KiB drops 2400 ns -> 17 ns (matches misra-page).
- BM_ArenaBumpReset now SkipWithMessage("backend has no bulk reset")
  for non-arena backends instead of duplicating BM_BatchAllocFree
  numbers in the row. Honest n/a.
- Benchmark/meson.build documents the LTO requirement: cross-TU
  inlining of the trivial bench_alloc / bench_free / bench_realloc
  wrappers into Bench.cpp's hot loop relies on -Db_lto=true. README
  reproduce block already shows both setup invocations.
- Benchmark/README.md regenerated with stats inline + LTO; misra
  columns now report bytes_in_use in the fragmentation table for the
  first time (lowest values in the table, exact live MB).

CODING-CONVENTIONS.md: replaced the "typed-direct skips stats" rule
(now wrong) with three general patterns -- shared per-instance state
on the base, one writer layer only, read-only fields exposed through
the ((void)0, ptr->field) accessor macro shape.

Windows test runner: SetJmp / LongJmp asm pair gated on __clang__ so
plain MSVC cl.exe falls back to libc setjmp/longjmp -- cl doesn't
understand __attribute__((naked)) or GNU-style __asm__. clang-cl path
unchanged.

Author: brightprogrammer

Benchmark/README.md

@@ -10,14 +10,14 @@ Single alloc/free pair, 16 B:
 
 | backend | time |
 |---|---:|
-| tcmalloc | 3.5 ns |
-| glibc    | 7.4 ns |
-| jemalloc | 3.9 ns |
-| mimalloc | 3.2 ns |
+| tcmalloc | 3.4 ns |
+| glibc    | 7.2 ns |
+| jemalloc | 3.4 ns |
+| mimalloc | 2.9 ns |
 | misra (Heap, validate-full) | 13.4 ns |
-| misra (Heap, validate-fast) | 13.6 ns |
-| misra-correct (Slab, validate-full) | 12.4 ns |
-| misra-correct (Slab, validate-fast) | 7.0 ns |
+| misra (Heap, validate-fast) | 13.0 ns |
+| misra-correct (Slab, validate-full) | 12.8 ns |
+| misra-correct (Slab, validate-fast) | 8.8 ns |
 
 ## Timing
 
@@ -27,13 +27,13 @@ One `alloc(size)` immediately followed by `free(ptr)`, repeated. Hot reuse — t
 
 | benchmark | glibc | jemalloc | mimalloc | tcmalloc | misra-full | misra-fast | misra-correct-full | misra-correct-fast | misra-arena | misra-page |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| 16 B | 7.4 | 3.9 | 3.2 | 3.5 | 13.4 | 13.6 | 12.4 | 7.0 | 9.5 | n/a |
-| 64 B | 7.2 | 3.9 | 3.3 | 3.5 | 13.2 | 13.3 | 12.4 | 6.9 | 9.3 | n/a |
-| 256 B | 7.2 | 4.3 | 14.0 | 3.5 | 13.2 | 13.6 | 12.4 | 7.0 | 9.3 | n/a |
-| 1 KiB | 22.1 | 6.0 | 15.5 | 3.5 | 13.4 | 13.4 | 12.5 | 7.0 | 9.3 | n/a |
-| 4 KiB | 15.2 | 12.4 | 11.1 | 3.7 | 2481.2 | 2445.8 | 12.2 | 6.8 | 9.3 | 17.0 |
-| 16 KiB | 16.2 | 20.8 | 20.2 | 3.7 | 1214.3 | 1005.3 | 2463.5 | 2389.7 | 9.2 | 16.9 |
-| 64 KiB | 17.4 | 288.3 | 20.3 | 3.8 | 1201.8 | 1017.4 | 2496.4 | 2380.8 | 9.2 | 16.9 |
+| 16 B | 7.2 | 3.4 | 2.9 | 3.4 | 13.4 | 13.0 | 12.8 | 8.8 | 9.9 | n/a |
+| 64 B | 7.0 | 3.5 | 3.0 | 3.3 | 13.5 | 13.3 | 12.7 | 8.7 | 9.9 | n/a |
+| 256 B | 7.2 | 3.9 | 13.5 | 3.3 | 13.7 | 13.3 | 12.7 | 8.7 | 9.9 | n/a |
+| 1 KiB | 7.2 | 5.2 | 14.8 | 3.3 | 13.2 | 13.2 | 12.7 | 8.7 | 9.9 | n/a |
+| 4 KiB | 22.9 | 9.7 | 10.6 | 3.3 | 1514.5 | 1517.9 | 12.7 | 8.7 | 9.9 | 16.6 |
+| 16 KiB | 23.4 | 18.7 | 19.9 | 3.4 | 1159.8 | 1152.9 | 16.9 | 16.9 | 9.9 | 16.7 |
+| 64 KiB | 25.6 | 270.3 | 20.0 | 3.4 | 1529.7 | 1531.6 | 16.9 | 16.9 | 9.9 | 16.8 |
 
 _Values in ns._
 
@@ -43,9 +43,9 @@ _Values in ns._
 
 | benchmark | glibc | jemalloc | mimalloc | tcmalloc | misra-full | misra-fast | misra-correct-full | misra-correct-fast | misra-arena | misra-page |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| 128 × 64 B | 1.4 | 1.1 | 0.4 | 0.5 | 1.4 | 1.4 | 1.7 | 1.0 | n/a | n/a |
-| 1024 × 64 B | 12.3 | 12.0 | 3.6 | 4.0 | 11.2 | 10.9 | 18.7 | 12.9 | n/a | n/a |
-| 8192 × 64 B | 101.2 | 95.5 | 34.4 | 39.8 | 87.6 | 86.7 | 419.4 | 361.1 | n/a | n/a |
+| 128 × 64 B | 1.2 | 1.0 | 0.5 | 0.5 | 1.5 | 1.4 | 1.7 | 1.0 | n/a | n/a |
+| 1024 × 64 B | 9.6 | 10.9 | 4.0 | 4.3 | 11.8 | 11.7 | 17.5 | 12.2 | n/a | n/a |
+| 8192 × 64 B | 77.9 | 88.9 | 34.3 | 39.1 | 94.7 | 94.6 | 390.2 | 358.8 | n/a | n/a |
 
 _Values in us._
 
@@ -55,9 +55,9 @@ Same shape as the pair test but writes every byte of the allocation before freei
 
 | benchmark | glibc | jemalloc | mimalloc | tcmalloc | misra-full | misra-fast | misra-correct-full | misra-correct-fast | misra-arena | misra-page |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| 64 B | 7.9 | 5.2 | 3.3 | 4.6 | 15.1 | 14.3 | 14.8 | 9.3 | 11.6 | n/a |
-| 4 KiB | 36.7 | 29.4 | 25.5 | 19.5 | 3377.2 | 2435.7 | 31.9 | 26.2 | 52.2 | 35.3 |
-| 64 KiB | 973.6 | 1418.4 | 866.3 | 841.0 | 13475.6 | 13614.5 | 14605.9 | 14411.7 | 914.2 | 847.3 |
+| 64 B | 7.0 | 4.3 | 3.3 | 4.1 | 14.4 | 14.3 | 14.4 | 9.9 | 10.7 | n/a |
+| 4 KiB | 33.0 | 23.1 | 24.1 | 18.8 | 2348.6 | 2364.9 | 30.9 | 25.0 | 56.3 | 33.7 |
+| 64 KiB | 864.7 | 1063.5 | 865.2 | 839.8 | 13202.5 | 13183.3 | 847.7 | 847.2 | 949.9 | 847.0 |
 
 _Values in ns._
 
@@ -67,7 +67,7 @@ _Values in ns._
 
 | benchmark | glibc | jemalloc | mimalloc | tcmalloc | misra-full | misra-fast | misra-correct-full | misra-correct-fast | misra-arena | misra-page |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| Pareto(1.16, 24) | 31.1 | 12.0 | 11.4 | 11.5 | 32.5 | 31.3 | 33.3 | 32.5 | n/a | n/a |
+| Pareto(1.16, 24) | 28.6 | 10.9 | 11.3 | 11.1 | 32.5 | 32.6 | 32.2 | 31.8 | n/a | n/a |
 
 _Values in us._
 
@@ -77,7 +77,7 @@ Geometric realloc ladder from 8 B up to 1 MiB. Time per full ladder, lower is be
 
 | benchmark | glibc | jemalloc | mimalloc | tcmalloc | misra-full | misra-fast | misra-correct-full | misra-correct-fast | misra-arena | misra-page |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| 8 B → 1 MiB | 4614.3 | 2825.7 | 18236.5 | 15248.1 | 247616.4 | 249123.4 | 247637.3 | 245527.6 | 114.4 | n/a |
+| 8 B → 1 MiB | 8265.9 | 4054.7 | 18185.7 | 15349.6 | 15787.3 | 16312.1 | 16030.5 | 16180.3 | 88.3 | n/a |
 
 _Values in ns._
 
@@ -87,9 +87,9 @@ Allocate N small (32 B) objects, then release them all. Arena does this as one O
 
 | benchmark | glibc | jemalloc | mimalloc | tcmalloc | misra-full | misra-fast | misra-correct-full | misra-correct-fast | misra-arena | misra-page |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| 128 × 32 B | 1.5 | 0.6 | 0.4 | 0.5 | 1.4 | 1.3 | 1.4 | 1.4 | 0.7 | n/a |
-| 1024 × 32 B | 12.8 | 10.2 | 3.3 | 4.1 | 11.3 | 10.6 | 11.3 | 11.3 | 5.3 | n/a |
-| 8192 × 32 B | 100.2 | 101.7 | 31.1 | 32.8 | 91.1 | 84.9 | 89.8 | 88.3 | 42.5 | n/a |
+| 128 × 32 B | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 0.6 | n/a |
+| 1024 × 32 B | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 4.9 | n/a |
+| 8192 × 32 B | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 39.5 | n/a |
 
 _Values in us._
 
@@ -99,11 +99,11 @@ Each allocator's own introspection API reports committed bytes after the workloa
 
 | benchmark | live MB | glibc MB | jemalloc MB | mimalloc MB | tcmalloc MB | misra-full MB | misra-fast MB | misra-correct-full MB | misra-correct-fast MB | misra-arena MB | misra-page MB |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
-| Checkerboard (4 K small) | 0.6 | 2.4 | 9.2 | n/a | 7.0 | n/a | n/a | n/a | n/a | n/a | n/a |
-| Checkerboard (16 K small) | 2.5 | 4.1 | 11.9 | n/a | 7.0 | n/a | n/a | n/a | n/a | n/a | n/a |
-| Checkerboard (64 K small) | 10.0 | 15.3 | 21.8 | n/a | 16.0 | n/a | n/a | n/a | n/a | n/a | n/a |
-| Lifetime mix | 4.0 | 5.1 | 19.0 | n/a | 16.0 | n/a | n/a | n/a | n/a | n/a | n/a |
-| Page overhang | 18.2 | 20.5 | 32.3 | n/a | 29.0 | n/a | n/a | n/a | n/a | n/a | n/a |
+| Checkerboard (4 K small) | 0.6 | 1.1 | 9.2 | n/a | 2.0 | 0.6 | 0.6 | n/a | n/a | n/a | n/a |
+| Checkerboard (16 K small) | 2.5 | 3.9 | 11.8 | n/a | 4.0 | 2.5 | 2.5 | n/a | n/a | n/a | n/a |
+| Checkerboard (64 K small) | 10.0 | 15.1 | 21.8 | n/a | 14.0 | 10.0 | 10.0 | n/a | n/a | n/a | n/a |
+| Lifetime mix | 4.0 | 4.9 | 20.2 | n/a | 14.0 | 4.0 | 4.0 | n/a | n/a | n/a | n/a |
+| Page overhang | 18.2 | 20.3 | 32.4 | n/a | 27.0 | 18.2 | 18.2 | n/a | n/a | n/a | n/a |
 
 ## How to read
 
@@ -143,10 +143,10 @@ Regenerates this file with measurements from the host.
 
 | | |
 |---|---|
-| timestamp | 2026-05-29 00:47:52 UTC |
-| git rev   | db0e77fcc603 (master) |
+| timestamp | 2026-05-29 02:14:47 UTC |
+| git rev   | af143d4679e6 (master) |
 | host CPU  | Intel(R) Core(TM) Ultra 7 165U |
 | kernel    | Linux 6.18.28 |
 | compiler  | gcc 14.3.0 |
-| build     | validate-full: buildtype=release optimization=3 b_sanitize=[] b_lto=False alloc_debug=False heap_validate_full=True  |  validate-fast: buildtype=release optimization=3 b_lto=False b_sanitize=[] alloc_debug=False heap_validate_full=False |
+| build     | validate-full: buildtype=release optimization=3 b_lto=True b_sanitize=[] alloc_debug=False heap_validate_full=True  |  validate-fast: buildtype=release optimization=3 b_lto=True b_sanitize=[] alloc_debug=False heap_validate_full=False |
 | reps      | 10 per row (median reported) |

Benchmark/Source/Allocator_misra.c

@@ -92,8 +92,7 @@ void bench_free(void *p) {
 
 uint64_t bench_live_bytes(void) {
 #if FEATURE_ALLOC_STATS
-    AllocatorStats s = AllocatorGetStats(g_alloc);
-    return (uint64_t)s.bytes_in_use;
+    return (uint64_t)AllocatorBytesInUse(g_alloc);
 #else
     return 0;
 #endif

Benchmark/Source/Allocator_misra_arena.c

@@ -89,8 +89,7 @@ void bench_free(void *p) {
 
 uint64_t bench_live_bytes(void) {
 #if FEATURE_ALLOC_STATS
-    AllocatorStats s = AllocatorGetStats(ALLOCATOR_OF(&g_arena));
-    return (uint64_t)s.bytes_in_use;
+    return (uint64_t)AllocatorBytesInUse(&g_arena);
 #else
     return 0;
 #endif

Benchmark/Source/Allocator_misra_correct.c

@@ -24,21 +24,23 @@
 
 #include <Misra/Std/Allocator.h>
 #include <Misra/Std/Allocator/Heap.h>
+#include <Misra/Std/Allocator/Page.h>
 #include <Misra/Std/Allocator/Slab.h>
 
-// MODE_NONE = pre-init / post-teardown. MODE_HEAP / MODE_SLAB choose
-// which of the two static allocator instances below is currently
-// live. The two are unioned by usage, not by struct -- only one is
-// initialised at any moment. We pay the (~150 + ~80 B) BSS cost to
-// avoid a heap alloc on bench_init.
+// MODE_NONE = pre-init / post-teardown. MODE_HEAP / MODE_SLAB / MODE_PAGE
+// choose which of the static allocator instances below is currently
+// live. They're unioned by usage, not by struct -- only one is
+// initialised at any moment.
 typedef enum {
     MODE_NONE = 0,
     MODE_HEAP = 1,
     MODE_SLAB = 2,
+    MODE_PAGE = 3,
 } bench_mode;
 
 static HeapAllocator g_heap;
 static SlabAllocator g_slab;
+static PageAllocator g_page;
 static bench_mode     g_mode = MODE_NONE;
 
 Zstr bench_backend_name(void) {
@@ -56,6 +58,8 @@ static void tear_current(void) {
         HeapAllocatorDeinit(&g_heap);
     } else if (g_mode == MODE_SLAB) {
         SlabAllocatorDeinit(&g_slab);
+    } else if (g_mode == MODE_PAGE) {
+        PageAllocatorDeinit(&g_page);
     }
     g_mode = MODE_NONE;
 }
@@ -69,27 +73,20 @@ void bench_teardown(void) {
     tear_current();
 }
 
-// Threshold above which the slab's slot-size design (one slab = one OS
-// page, slot_size capped at the page) can't fit a single slot. Above
-// this, the "correct" backend falls back to HeapAllocator so the bench
-// still measures a real allocation rather than NULL returns.
-//
-// Matches MisraStdC's slab max (currently 4096 B, the smallest OS page
-// size we target). If the slab grows multi-page support later, raise
-// this threshold to match.
+// SlabAllocator caps at MisraStdC's slab max (currently 4096 B, the
+// smallest OS page size we target). Above this, the "right tool" is
+// PageAllocator -- the request is already page-aligned in size, so
+// one mmap per alloc is the honest cost, no rounding waste, no
+// general-heap dispatch noise. The previous fallback to Heap measured
+// the wrong-tool case, which is exactly what the `misra` (Heap-only)
+// column is for.
 #define BENCH_SLAB_MAX_SLOT 4096u
 
 void bench_use_fixed_size(size_t slot) {
     tear_current();
     if (slot > BENCH_SLAB_MAX_SLOT) {
-        // Slab can't hold a slot this big; fall back to Heap and let
-        // it route through XL (mmap-per-alloc) the way the upstream
-        // "wrong tool" backend does. The bench's
-        // wrong-vs-right-allocator comparison still works for sizes
-        // within the slab's range; for larger sizes both backends
-        // converge on the same Heap path.
-        g_heap = HeapAllocatorInit();
-        g_mode = MODE_HEAP;
+        g_page = PageAllocatorInit();
+        g_mode = MODE_PAGE;
         return;
     }
     g_slab = SlabAllocatorInit((size)slot);
@@ -108,27 +105,26 @@ int  bench_can_reset(void) { return 0; }
 void bench_reset(void)     {}
 
 void *bench_alloc(size_t n) {
-    // _Generic in AllocatorAlloc routes to slab_allocator_allocate or
-    // heap_allocator_allocate based on the static type at each call
-    // site. The g_mode branch is perfectly predicted within a
-    // benchmark because the mode is set once before the iteration
-    // loop and never changes inside it.
+    // _Generic in AllocatorAlloc routes to the typed *_allocator_allocate
+    // statically. The g_mode branch is perfectly predicted within a
+    // benchmark -- mode is set once before the iteration loop and
+    // never changes inside it.
     if (g_mode == MODE_SLAB) {
         return AllocatorAlloc(&g_slab, (size)n, 0);
     }
+    if (g_mode == MODE_PAGE) {
+        return AllocatorAlloc(&g_page, (size)n, 0);
+    }
     return AllocatorAlloc(&g_heap, (size)n, 0);
 }
 
 void *bench_realloc(void *p, size_t n) {
-    // SlabAllocator slot size is fixed at init; realloc is not a
-    // sensible operation in slab mode. The bench's only realloc
-    // workload (BM_ReallocGrow) runs in MODE_HEAP, so this branch
-    // should never fire in slab mode. Return NULL (caller treats as
-    // alloc failure -- the bench then reads the size-class check
-    // and moves on); we avoid LOG_FATAL here because aborting the
-    // whole bench process on a routing mismatch would make the
-    // failure mode hostile to interactive debugging.
-    if (g_mode == MODE_SLAB) {
+    // Realloc is sensible only on the general-purpose Heap path.
+    // BM_ReallocGrow is the sole realloc workload and runs in
+    // MODE_HEAP, so the slab / page branches should never fire here.
+    // Return NULL (caller treats as alloc failure) instead of
+    // LOG_FATAL so a routing mismatch doesn't kill interactive runs.
+    if (g_mode == MODE_SLAB || g_mode == MODE_PAGE) {
         (void)p;
         (void)n;
         return NULL;
@@ -139,34 +135,30 @@ void *bench_realloc(void *p, size_t n) {
 void bench_free(void *p) {
     if (g_mode == MODE_SLAB) {
         AllocatorFree(&g_slab, p);
+    } else if (g_mode == MODE_PAGE) {
+        AllocatorFree(&g_page, p);
     } else {
         AllocatorFree(&g_heap, p);
     }
 }
 
 uint64_t bench_live_bytes(void) {
 #if FEATURE_ALLOC_STATS
-    // AllocatorGetStats takes a base `Allocator *`, not a typed
-    // pointer -- both HeapAllocator and SlabAllocator embed
-    // `Allocator base` as their first field, so the cast is layout-safe.
-    AllocatorStats s;
-    if (g_mode == MODE_SLAB) {
-        s = AllocatorGetStats(ALLOCATOR_OF(&g_slab));
-    } else if (g_mode == MODE_HEAP) {
-        s = AllocatorGetStats(ALLOCATOR_OF(&g_heap));
-    } else {
-        return 0;
-    }
-    return (uint64_t)s.bytes_in_use;
+    if (g_mode == MODE_SLAB) return (uint64_t)AllocatorBytesInUse(&g_slab);
+    if (g_mode == MODE_HEAP) return (uint64_t)AllocatorBytesInUse(&g_heap);
+    if (g_mode == MODE_PAGE) return (uint64_t)AllocatorBytesInUse(&g_page);
+    return 0;
 #else
     return 0;
 #endif
 }
 
 uint64_t bench_footprint_bytes(void) {
-    // HeapAllocator and SlabAllocator no longer embed a PageAllocator;
-    // each talks to the kernel directly. Footprint measurement via the
-    // PageAllocator entries[] vector is no longer available here.
-    (void)g_mode;
+    // Only the page-backed mode can report a meaningful footprint
+    // today; Heap and Slab manage their own pages with no public
+    // committed-bytes accessor yet.
+    if (g_mode == MODE_PAGE) {
+        return (uint64_t)PageAllocatorFootprintBytes(&g_page);
+    }
     return 0;
 }

Benchmark/Source/Allocator_misra_page.c

@@ -72,8 +72,7 @@ void bench_free(void *p) {
 
 uint64_t bench_live_bytes(void) {
 #if FEATURE_ALLOC_STATS
-    AllocatorStats s = AllocatorGetStats(ALLOCATOR_OF(&g_page));
-    return (uint64_t)s.bytes_in_use;
+    return (uint64_t)AllocatorBytesInUse(&g_page);
 #else
     return 0;
 #endif

Benchmark/Source/Bench.cpp

@@ -120,22 +120,27 @@ BENCHMARK(BM_MixedPareto);
 // ---------------------------------------------------------------------------
 // 5. arena bump + bulk reset
 // ---------------------------------------------------------------------------
-// Allocate N small objects, then either reset the backing arena (bulk
-// O(1) free) or fall back to freeing each pointer individually. Same
-// workload on every backend; arena-shaped backends short-circuit to
-// reset. Shows the cost differential of "1 reset" vs "N individual
-// frees" for the same bump-N workload.
+// Allocate N small objects, then reset the backing arena (bulk O(1)
+// free). Reports the cost of "1 reset" for a bump-N workload. Only
+// runs on backends that actually expose a reset; everyone else gets
+// n/a so the column isn't accidentally compared against per-pointer
+// free (that comparison lives in BM_BatchAllocFree).
 static void BM_ArenaBumpReset(benchmark::State &state) {
+    if (!bench_can_reset()) {
+        // This bench is specifically the bulk-reset workload. On
+        // backends that don't support reset, running per-pointer free
+        // instead would silently duplicate BM_BatchAllocFree numbers
+        // and pretend they're an arena-vs-heap comparison. Skip so the
+        // row shows n/a on those columns honestly.
+        state.SkipWithMessage("backend has no bulk reset");
+        return;
+    }
     const size_t n  = static_cast<size_t>(state.range(0));
     const size_t sz = 32;
     std::vector<void *> ptrs(n);
     for (auto _ : state) {
         for (size_t i = 0; i < n; i++) ptrs[i] = bench_alloc(sz);
-        if (bench_can_reset()) {
-            bench_reset();
-        } else {
-            for (size_t i = 0; i < n; i++) bench_free(ptrs[i]);
-        }
+        bench_reset();
     }
     state.SetItemsProcessed(int64_t(state.iterations()) * int64_t(n));
 }

Benchmark/meson.build

@@ -76,6 +76,10 @@ foreach b : bench_libc_backends
 endforeach
 
 # Misra backends. Each picks up the in-tree misra_std built above.
+# Cross-TU inlining of the trivial `bench_alloc` / `bench_free` /
+# `bench_realloc` wrappers into `Bench.cpp`'s hot loop relies on the
+# project being configured with `-Db_lto=true`; without that the
+# wrapper costs a real function call per iteration.
 executable(
     'bench-misra',
     files('Source/Allocator_misra.c') + bench_cpp_sources,