Merge branch 'lang/jolt-heap' into lang/foundations

packages/ludic.physics/native.ludic

@@ -45,3 +45,10 @@ extern function jph_body_angular_impulse(w: pointer, id: int, x: float, y: float
 extern function jph_body_spin(w: pointer, id: int, out: pointer) -> void = "jph_body_spin"
 extern function jph_body_set_pose(w: pointer, id: int, x: float, y: float, z: float, qx: float, qy: float, qz: float, qw: float, vx: float, vy: float, vz: float) -> void = "jph_body_set_pose"
 extern function jph_shape_heightfield_bspline(tex: pointer, n: int, ox: float, oz: float, cell: float) -> pointer = "jph_shape_heightfield_bspline"
+extern function jph_heap_bytes() -> long = "jph_heap_bytes"
+extern function jph_heap_peak() -> long = "jph_heap_peak"
+extern function jph_heap_allocs() -> long = "jph_heap_allocs"
+# what Jolt holds now, the most it has held, and the blocks it has asked for (counted in the shim)
+export function phys_heap_bytes() -> long { return jph_heap_bytes() }
+export function phys_heap_peak() -> long { return jph_heap_peak() }
+export function phys_heap_allocs() -> long { return jph_heap_allocs() }

packages/ludic.physics/native/shim/jph_alloc.inl

@@ -0,0 +1,68 @@
+// jph_alloc.inl - Jolt's allocations counted apart from the game's (plan 25): libc's malloc underneath,
+// each block carrying its size in a 16-byte header, so live bytes, the peak and the count are exact.
+// The fence reads them (extern_weak) to judge Jolt by its plateau, not per frame.
+#include <atomic>
+#include <cstdlib>
+#include <cstring>
+
+namespace {
+constexpr size_t kHead = 16;
+std::atomic<int64_t> g_live{0}, g_peak{0}, g_allocs{0};
+
+void heap_add(int64_t n) {
+  int64_t now = g_live.fetch_add(n, std::memory_order_relaxed) + n;
+  int64_t p = g_peak.load(std::memory_order_relaxed);
+  while (now > p && !g_peak.compare_exchange_weak(p, now, std::memory_order_relaxed)) {}
+}
+void *heap_alloc(size_t n) {
+  char *b = static_cast<char *>(std::malloc(n + kHead));
+  if (!b) return nullptr;
+  std::memcpy(b, &n, sizeof n);
+  g_allocs.fetch_add(1, std::memory_order_relaxed);
+  heap_add(int64_t(n));
+  return b + kHead;
+}
+void heap_free(void *p) {
+  if (!p) return;
+  char *b = static_cast<char *>(p) - kHead;
+  size_t n;
+  std::memcpy(&n, b, sizeof n);
+  heap_add(-int64_t(n));
+  std::free(b);
+}
+void *heap_realloc(void *p, size_t, size_t n) {
+  if (!p) return heap_alloc(n);
+  char *b = static_cast<char *>(p) - kHead;
+  size_t old;
+  std::memcpy(&old, b, sizeof old);
+  char *q = static_cast<char *>(std::realloc(b, n + kHead));
+  if (!q) return nullptr;
+  std::memcpy(q, &n, sizeof n);
+  heap_add(int64_t(n) - int64_t(old));
+  return q + kHead;
+}
+// aligned: over-asked, the counted block's pointer kept in the word before the aligned one
+void *heap_aligned(size_t n, size_t align) {
+  if (align < sizeof(void *)) align = sizeof(void *);
+  char *raw = static_cast<char *>(heap_alloc(n + align + sizeof(void *)));
+  if (!raw) return nullptr;
+  uintptr_t at = (reinterpret_cast<uintptr_t>(raw) + sizeof(void *) + align - 1) & ~(uintptr_t(align) - 1);
+  reinterpret_cast<void **>(at)[-1] = raw;
+  return reinterpret_cast<void *>(at);
+}
+void heap_aligned_free(void *p) {
+  if (p) heap_free(static_cast<void **>(p)[-1]);
+}
+void heap_register() {
+  JPH::Allocate = heap_alloc;
+  JPH::Reallocate = heap_realloc;
+  JPH::Free = heap_free;
+  JPH::AlignedAllocate = heap_aligned;
+  JPH::AlignedFree = heap_aligned_free;
+}
+} // namespace
+
+// bytes Jolt holds now, the most it ever held, and how many blocks it has asked for
+JPH_SHIM int64_t jph_heap_bytes(void) { return g_live.load(std::memory_order_relaxed); }
+JPH_SHIM int64_t jph_heap_peak(void) { return g_peak.load(std::memory_order_relaxed); }
+JPH_SHIM int64_t jph_heap_allocs(void) { return g_allocs.load(std::memory_order_relaxed); }

packages/ludic.physics/native/shim/jph_shim.cpp

@@ -120,10 +120,12 @@ void *keep(const ShapeSettings::ShapeResult &r) {
 }
 } // namespace
 
-// Jolt's allocator, factory and types, once per process; 0 when ready
+#include "jph_alloc.inl"
+
+// Jolt's allocator (counted, jph_alloc.inl), factory and types, once per process; 0 when ready
 JPH_SHIM int jph_init(void) {
   if (g_ready) return 0;
-  RegisterDefaultAllocator();
+  heap_register();
   Factory::sInstance = new Factory();
   RegisterTypes();
   g_ready = true;

packages/ludic.physics/tests/jolt_heap_test.ludic

@@ -0,0 +1,34 @@
+# jolt_heap_test.ludic - Jolt's own heap, counted in the shim (plan 25): a ground coming and going a
+# thousand times holds what it held after the first hundred, and a world closed gives back what it took
+import "ludic.physics"
+import "ludic.base"
+program JoltHeapTest {
+  numbers float
+  function cycle(physics_st: mut PhysicsState, h: []float, k: int) -> void {
+    let s = phys_heightfield(physics_st, h, 33, float(k % 7) * 64.0, 0.0, 2.0)
+    let b = phys_ground_add(physics_st, s)
+    phys_step(physics_st)
+    phys_remove(physics_st, b)
+    phys_shape_free(physics_st, s)
+    phys_settle(physics_st)
+  }
+
+  test "a ground in and out a thousand times holds Jolt's heap where it settled, and closing gives it back" (physics_st: mut PhysicsState) {
+    expect(phys_open(physics_st, 64, 1))
+    phys_close(physics_st)
+    let before = phys_heap_bytes()         # the factory and the types, made once per process
+    expect(phys_open(physics_st, 64, 1))
+    expect(phys_heap_bytes() > before)
+    let h = floats(33 * 33)
+    for k in 0 .. 100 { cycle(physics_st, h, k) }
+    let held = phys_heap_bytes()
+    let peak = phys_heap_peak()
+    for k in 0 .. 1000 { cycle(physics_st, h, k) }
+    print(`jolt heap: {held} bytes held after 100 grounds, {phys_heap_bytes()} after 1100, peak {phys_heap_peak()}, {phys_heap_allocs()} blocks asked for`)
+    expect(phys_heap_bytes() == held)
+    expect(phys_heap_peak() == peak)
+    phys_close(physics_st)
+    print(`jolt heap: {before} bytes before the world, {phys_heap_bytes()} after it closed`)
+    expect(phys_heap_bytes() == before)
+  }
+}