fix

benches/micro_core.cm

@@ -0,0 +1,302 @@
+// micro_core.cm — direct microbenchmarks for core ops
+
+function blackhole(sink, x) {
+  return (sink + (x | 0)) | 0
+}
+
+function make_obj_xy(x, y) {
+  return {x: x, y: y}
+}
+
+function make_obj_yx(x, y) {
+  // Different insertion order to force a different shape
+  return {y: y, x: x}
+}
+
+function make_packed_array(n) {
+  var a = []
+  var i = 0
+  for (i = 0; i < n; i++) push(a, i)
+  return a
+}
+
+function make_holey_array(n) {
+  var a = []
+  var i = 0
+  for (i = 0; i < n; i += 2) a[i] = i
+  return a
+}
+
+return {
+  loop_empty: function(n) {
+    var sink = 0
+    var i = 0
+    for (i = 0; i < n; i++) {}
+    return blackhole(sink, n)
+  },
+
+  i32_add: function(n) {
+    var sink = 0
+    var x = 1
+    var i = 0
+    for (i = 0; i < n; i++) x = (x + 3) | 0
+    return blackhole(sink, x)
+  },
+
+  f64_add: function(n) {
+    var sink = 0
+    var x = 1.0
+    var i = 0
+    for (i = 0; i < n; i++) x = x + 3.14159
+    return blackhole(sink, x | 0)
+  },
+
+  mixed_add: function(n) {
+    var sink = 0
+    var x = 1
+    var i = 0
+    for (i = 0; i < n; i++) x = x + 0.25
+    return blackhole(sink, x | 0)
+  },
+
+  bit_ops: function(n) {
+    var sink = 0
+    var x = 0x12345678
+    var i = 0
+    for (i = 0; i < n; i++) x = ((x << 5) ^ (x >>> 3)) | 0
+    return blackhole(sink, x)
+  },
+
+  overflow_path: function(n) {
+    var sink = 0
+    var x = 0x70000000
+    var i = 0
+    for (i = 0; i < n; i++) x = (x + 0x10000000) | 0
+    return blackhole(sink, x)
+  },
+
+  call_direct: function(n) {
+    var sink = 0
+    var f = function(a) { return (a + 1) | 0 }
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) x = f(x)
+    return blackhole(sink, x)
+  },
+
+  call_indirect: function(n) {
+    var sink = 0
+    var f = function(a) { return (a + 1) | 0 }
+    var g = f
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) x = g(x)
+    return blackhole(sink, x)
+  },
+
+  call_closure: function(n) {
+    var sink = 0
+    var make_adder = function(k) {
+      return function(a) { return (a + k) | 0 }
+    }
+    var add3 = make_adder(3)
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) x = add3(x)
+    return blackhole(sink, x)
+  },
+
+  array_read_packed: function(n) {
+    var sink = 0
+    var a = make_packed_array(1024)
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) x = (x + a[i & 1023]) | 0
+    return blackhole(sink, x)
+  },
+
+  array_write_packed: function(n) {
+    var sink = 0
+    var a = make_packed_array(1024)
+    var i = 0
+    for (i = 0; i < n; i++) a[i & 1023] = i
+    return blackhole(sink, a[17] | 0)
+  },
+
+  array_read_holey: function(n) {
+    var sink = 0
+    var a = make_holey_array(2048)
+    var x = 0
+    var i = 0
+    var v = null
+    for (i = 0; i < n; i++) {
+      v = a[(i & 2047)]
+      if (v) x = (x + v) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  array_push_steady: function(n) {
+    var sink = 0
+    var x = 0
+    var j = 0
+    var i = 0
+    var a = null
+    for (j = 0; j < n; j++) {
+      a = []
+      for (i = 0; i < 256; i++) push(a, i)
+      x = (x + length(a)) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  array_indexed_sum: function(n) {
+    var sink = 0
+    var a = make_packed_array(1024)
+    var x = 0
+    var j = 0
+    var i = 0
+    for (j = 0; j < n; j++) {
+      x = 0
+      for (i = 0; i < 1024; i++) {
+        x = (x + a[i]) | 0
+      }
+    }
+    return blackhole(sink, x)
+  },
+
+  prop_read_mono: function(n) {
+    var sink = 0
+    var o = make_obj_xy(1, 2)
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) x = (x + o.x) | 0
+    return blackhole(sink, x)
+  },
+
+  prop_read_poly_2: function(n) {
+    var sink = 0
+    var a = make_obj_xy(1, 2)
+    var b = make_obj_yx(1, 2)
+    var x = 0
+    var i = 0
+    var o = null
+    for (i = 0; i < n; i++) {
+      o = (i & 1) == 0 ? a : b
+      x = (x + o.x) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  prop_read_poly_4: function(n) {
+    var sink = 0
+    var shapes = [
+      {x: 1, y: 2},
+      {y: 2, x: 1},
+      {x: 1, z: 3, y: 2},
+      {w: 0, x: 1, y: 2}
+    ]
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) {
+      x = (x + shapes[i & 3].x) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  string_concat_small: function(n) {
+    var sink = 0
+    var x = 0
+    var j = 0
+    var i = 0
+    var s = null
+    for (j = 0; j < n; j++) {
+      s = ""
+      for (i = 0; i < 16; i++) s = s + "x"
+      x = (x + length(s)) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  string_concat_medium: function(n) {
+    var sink = 0
+    var x = 0
+    var j = 0
+    var i = 0
+    var s = null
+    for (j = 0; j < n; j++) {
+      s = ""
+      for (i = 0; i < 100; i++) s = s + "abcdefghij"
+      x = (x + length(s)) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  string_slice: function(n) {
+    var sink = 0
+    var base = "the quick brown fox jumps over the lazy dog"
+    var x = 0
+    var i = 0
+    var s = null
+    for (i = 0; i < n; i++) {
+      s = text(base, i % 10, i % 10 + 10)
+      x = (x + length(s)) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  guard_hot_number: function(n) {
+    var sink = 0
+    var x = 1
+    var i = 0
+    for (i = 0; i < n; i++) x = x + 1
+    return blackhole(sink, x | 0)
+  },
+
+  guard_mixed_types: function(n) {
+    var sink = 0
+    var vals = [1, "a", 2, "b", 3, "c", 4, "d"]
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) {
+      if (is_number(vals[i & 7])) x = (x + vals[i & 7]) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  reduce_sum: function(n) {
+    var sink = 0
+    var a = make_packed_array(256)
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) {
+      x = (x + reduce(a, function(acc, v) { return acc + v }, 0)) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  filter_evens: function(n) {
+    var sink = 0
+    var a = make_packed_array(256)
+    var x = 0
+    var i = 0
+    for (i = 0; i < n; i++) {
+      x = (x + length(filter(a, function(v) { return v % 2 == 0 }))) | 0
+    }
+    return blackhole(sink, x)
+  },
+
+  arrfor_sum: function(n) {
+    var sink = 0
+    var a = make_packed_array(256)
+    var x = 0
+    var i = 0
+    var sum = 0
+    for (i = 0; i < n; i++) {
+      sum = 0
+      arrfor(a, function(v) { sum += v })
+      x = (x + sum) | 0
+    }
+    return blackhole(sink, x)
+  }
+}

internal/os.c

@@ -322,7 +322,10 @@ JSC_SCALL(os_system,
 )
 
 JSC_CCALL(os_exit,
-  exit(0);
+  int code = 0;
+  if (argc > 0 && !JS_IsNull(argv[0]))
+    JS_ToInt32(js, &code, argv[0]);
+  exit(code);
 )
 
 static JSValue js_os_dylib_open(JSContext *js, JSValue self, int argc, JSValue *argv)
@@ -718,7 +721,7 @@ static const JSCFunctionListEntry js_os_funcs[] = {
   MIST_FUNC_DEF(os, rusage, 0),
   MIST_FUNC_DEF(os, mallinfo, 0),
   MIST_FUNC_DEF(os, system, 1),
-  MIST_FUNC_DEF(os, exit, 0),
+  MIST_FUNC_DEF(os, exit, 1),
   MIST_FUNC_DEF(os, sleep, 1),
   MIST_FUNC_DEF(os, dylib_open, 1),
   MIST_FUNC_DEF(os, dylib_preload, 1),

internal/shop.cm

@@ -421,6 +421,15 @@ Shop.extract_commit_hash = function(pkg, response) {
 var open_dls = {}
 var package_dylibs = {}  // pkg -> [{file, symbol, dylib}, ...]
 
+function open_dylib_cached(path) {
+  var handle = open_dls[path]
+  if (handle) return handle
+  handle = os.dylib_open(path)
+  if (!handle) return null
+  open_dls[path] = handle
+  return handle
+}
+
 // Host target detection for native dylib resolution
 function detect_host_target() {
   var platform = os.platform()
@@ -457,7 +466,7 @@ function try_native_mod_dylib(pkg, stem) {
   if (!fd.is_file(build_path)) return null
 
   log.shop('native dylib cache hit: ' + stem)
-  var handle = os.dylib_open(build_path)
+  var handle = open_dylib_cached(build_path)
   if (!handle) return null
   var sym = Shop.c_symbol_for_file(pkg, stem)
   return {_native: true, _handle: handle, _sym: sym}
@@ -924,11 +933,7 @@ function try_dylib_symbol(sym, pkg, file_stem) {
   })
   if (!entry || !entry.dylib) return null
 
-  var handle = open_dls[entry.dylib]
-  if (!handle) {
-    handle = os.dylib_open(entry.dylib)
-    if (handle) open_dls[entry.dylib] = handle
-  }
+  var handle = open_dylib_cached(entry.dylib)
   if (!handle) return null
   if (!os.dylib_has_symbol(handle, sym)) return null
 
@@ -1168,8 +1173,17 @@ Shop.is_loaded = function is_loaded(path, package_context) {
 }
 
 // Create a use function bound to a specific package context
-function make_use_fn(pkg) {
+function make_use_fn(pkg, force_native) {
   return function(path) {
+    var _native = null
+    if (force_native && !native_mode) {
+      _native = function() {
+        return Shop.use_native(path, pkg)
+      } disruption {
+        return Shop.use(path, pkg)
+      }
+      return _native()
+    }
     return Shop.use(path, pkg)
   }
 }
@@ -1200,7 +1214,7 @@ function execute_module(info)
         inject = Shop.script_inject_for(file_info)
         env = inject_env(inject)
         pkg = file_info.package
-        env.use = make_use_fn(pkg)
+        env.use = make_use_fn(pkg, true)
         env = stone(env)
         used = os.native_module_load_named(
           mod_resolve.symbol._handle, mod_resolve.symbol._sym, env)
@@ -1844,7 +1858,7 @@ Shop.load_as_dylib = function(path, pkg) {
   if (!file_info) file_info = Shop.file_info(file_path)
   inject = Shop.script_inject_for(file_info)
   env = inject_env(inject)
-  env.use = make_use_fn(real_pkg)
+  env.use = make_use_fn(real_pkg, true)
   env = stone(env)
   return os.native_module_load_named(result._handle, result._sym, env)
 }
@@ -1891,32 +1905,63 @@ Shop.parse_package = function(locator) {
 
 Shop.use_native = function(path, package_context) {
   var src_path = path
-  if (!starts_with(path, '/'))
+  var locator = null
+  var lookup = null
+  var cache_key = null
+  var cfg = null
+  var old_native = null
+  if (!starts_with(path, '/') && !fd.is_file(path)) {
+    lookup = ends_with(path, '.cm') ? path : path + '.cm'
+    locator = resolve_locator(lookup, package_context)
+    if (!locator) { print('Module not found: ' + path); disrupt }
+    src_path = locator.path
+  } else if (!starts_with(path, '/')) {
     src_path = fd.realpath(path)
+  }
   if (!fd.is_file(src_path)) { print('File not found: ' + path); disrupt }
 
   var file_info = Shop.file_info(src_path)
-  var pkg = file_info.package || package_context
+  var pkg = file_info.package || (locator ? locator.pkg : package_context)
+  var sym_stem = fd.basename(src_path)
+  var pkg_dir = null
+  cache_key = 'native:' + text(pkg || '') + ':' + src_path
+  if (use_cache[cache_key]) return use_cache[cache_key]
 
   var sym_name = null
-  if (pkg)
-    sym_name = Shop.c_symbol_for_file(pkg, fd.basename(src_path))
+  if (pkg) {
+    pkg_dir = get_packages_dir() + '/' + safe_package_path(pkg)
+    if (starts_with(src_path, pkg_dir + '/')) {
+      sym_stem = text(src_path, length(pkg_dir) + 1)
+    }
+    sym_name = Shop.c_symbol_for_file(pkg, sym_stem)
+  }
 
-  var build = Shop.use('build', 'core')
+  var build = use_cache['core/build'] || use_cache['build']
+  if (!build) {
+    cfg = Shop.load_config()
+    old_native = cfg.policy.native
+    cfg.policy.native = false
+    build = Shop.use('build', 'core')
+    cfg.policy.native = old_native
+  }
   var dylib_path = build.compile_native(src_path, null, null, pkg)
 
-  var handle = os.dylib_open(dylib_path)
+  var handle = open_dylib_cached(dylib_path)
   if (!handle) { print('Failed to open native dylib: ' + dylib_path); disrupt }
 
   // Build env with runtime functions and capabilities
   var inject = Shop.script_inject_for(file_info)
   var env = inject_env(inject)
-  env.use = make_use_fn(pkg)
+  env.use = make_use_fn(pkg, true)
   env = stone(env)
 
+  var loaded = null
   if (sym_name)
-    return os.native_module_load_named(handle, sym_name, env)
-  return os.native_module_load(handle, env)
+    loaded = os.native_module_load_named(handle, sym_name, env)
+  else
+    loaded = os.native_module_load(handle, env)
+  use_cache[cache_key] = loaded
+  return loaded
 }
 
 return Shop

qbe_emit.cm

@@ -790,6 +790,50 @@ ${sw("w", "%fp", "%dest", "%r")}
 @entry
 ${sr("a", "%obj_slot")}
 ${sr("b", "%key_slot")}
+  %ptag =l and %a, 7
+  %is_ptr =w ceql %ptag, 1
+  jnz %is_ptr, @arr_ptr, @fallback
+@arr_ptr
+  %arr_ptr =l and %a, -8
+  %arr_hdr =l loadl %arr_ptr
+@arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @arr_follow, @arr_chk
+@arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @arr_chase
+@arr_chk
+  %arr_is_array =w ceql %arr_ty, 0
+  jnz %arr_is_array, @arr_index, @fallback
+@arr_index
+  %idx_tag =l and %b, 1
+  %idx_is_int =w ceql %idx_tag, 0
+  jnz %idx_is_int, @idx_ok, @ret_null
+@idx_ok
+  %idx_l =l sar %b, 1
+  %idx_w =w copy %idx_l
+  %idx_neg =w csltw %idx_w, 0
+  jnz %idx_neg, @ret_null, @arr_len
+@arr_len
+  %len_p =l add %arr_ptr, 8
+  %len_l =l loadl %len_p
+  %len_w =w copy %len_l
+  %in =w csltw %idx_w, %len_w
+  jnz %in, @load, @ret_null
+@load
+  %idx_off_l =l extsw %idx_w
+  %idx_off_l =l shl %idx_off_l, 3
+  %vals_p =l add %arr_ptr, 16
+  %elem_p =l add %vals_p, %idx_off_l
+  %r =l loadl %elem_p
+${sw("w", "%fp", "%dest", "%r")}
+  ret %fp
+@ret_null
+${sw("w", "%fp", "%dest", text(qbe.js_null))}
+  ret %fp
+@fallback
   %r =l call $cell_rt_load_dynamic(l %ctx, l %a, l %b)
   %is_exc =w ceql %r, 15
   jnz %is_exc, @exc, @ok
@@ -805,14 +849,49 @@ ${sw("w", "%fp", "%dest", "%r")}
 @entry
 ${sr("a", "%arr_slot")}
 ${sr("b", "%idx_slot")}
-  %r =l call $cell_rt_load_index(l %ctx, l %a, l %b)
-  %is_exc =w ceql %r, 15
-  jnz %is_exc, @exc, @ok
-@ok
+  %idx_tag =l and %b, 1
+  %idx_is_int =w ceql %idx_tag, 0
+  jnz %idx_is_int, @idx_ok, @ret_null
+@idx_ok
+  %idx_l =l sar %b, 1
+  %idx_w =w copy %idx_l
+  %idx_neg =w csltw %idx_w, 0
+  jnz %idx_neg, @ret_null, @arr_init
+@arr_init
+  %ptag =l and %a, 7
+  %is_ptr =w ceql %ptag, 1
+  jnz %is_ptr, @arr_ptr_ok, @ret_null
+@arr_ptr_ok
+  %arr_ptr =l and %a, -8
+  %arr_hdr =l loadl %arr_ptr
+@arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @arr_follow, @arr_chk
+@arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @arr_chase
+@arr_chk
+  %arr_is_array =w ceql %arr_ty, 0
+  jnz %arr_is_array, @arr_len, @ret_null
+@arr_len
+  %len_p =l add %arr_ptr, 8
+  %len_l =l loadl %len_p
+  %len_w =w copy %len_l
+  %in =w csltw %idx_w, %len_w
+  jnz %in, @load, @ret_null
+@load
+  %idx_off_l =l extsw %idx_w
+  %idx_off_l =l shl %idx_off_l, 3
+  %vals_p =l add %arr_ptr, 16
+  %elem_p =l add %vals_p, %idx_off_l
+  %r =l loadl %elem_p
 ${sw("w", "%fp", "%dest", "%r")}
   ret %fp
-@exc
-  ret 0
+@ret_null
+${sw("w", "%fp", "%dest", text(qbe.js_null))}
+  ret %fp
 }`
 
   // store_field(ctx, fp, obj_slot, val_slot, lit_idx) — no dest write
@@ -834,10 +913,37 @@ ${sr("b", "%val_slot")}
 ${sr("a", "%obj_slot")}
 ${sr("b", "%val_slot")}
 ${sr("c", "%key_slot")}
+  %ptag =l and %a, 7
+  %is_ptr =w ceql %ptag, 1
+  jnz %is_ptr, @arr_ptr, @fallback
+@arr_ptr
+  %arr_ptr =l and %a, -8
+  %arr_hdr =l loadl %arr_ptr
+@arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @arr_follow, @arr_chk
+@arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @arr_chase
+@arr_chk
+  %arr_is_array =w ceql %arr_ty, 0
+  jnz %arr_is_array, @arr_key_chk, @fallback
+@arr_key_chk
+  %idx_tag =l and %c, 1
+  %idx_is_int =w ceql %idx_tag, 0
+  jnz %idx_is_int, @arr_store, @bad
+@arr_store
+  %fp2 =l call $__store_index_ss(l %ctx, l %fp, l %obj_slot, l %val_slot, l %key_slot)
+  ret %fp2
+@fallback
   %ok =w call $cell_rt_store_dynamic(l %ctx, l %b, l %a, l %c)
   jnz %ok, @ok, @exc
 @ok
   ret %fp
+@bad
+  call $cell_rt_disrupt(l %ctx)
 @exc
   ret 0
 }`
@@ -848,10 +954,151 @@ ${sr("c", "%key_slot")}
 ${sr("a", "%obj_slot")}
 ${sr("b", "%val_slot")}
 ${sr("c", "%idx_slot")}
-  %ok =w call $cell_rt_store_index(l %ctx, l %b, l %a, l %c)
-  jnz %ok, @ok, @exc
-@ok
+  %idx_tag =l and %c, 1
+  %idx_is_int =w ceql %idx_tag, 0
+  jnz %idx_is_int, @idx_ok, @bad
+@idx_ok
+  %idx_l =l sar %c, 1
+  %idx_w =w copy %idx_l
+  %idx_neg =w csltw %idx_w, 0
+  jnz %idx_neg, @bad, @arr_init
+@arr_init
+  %ptag =l and %a, 7
+  %is_ptr =w ceql %ptag, 1
+  jnz %is_ptr, @arr_ptr_ok, @bad
+@arr_ptr_ok
+  %arr_val =l copy %a
+  %arr_ptr =l and %arr_val, -8
+  %arr_hdr =l loadl %arr_ptr
+@arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @arr_follow, @arr_chk
+@arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @arr_chase
+@arr_chk
+  %arr_is_array =w ceql %arr_ty, 0
+  jnz %arr_is_array, @stone_chk, @bad
+@stone_chk
+  %arr_stone =l and %arr_hdr, 8
+  %arr_is_stone =w cnel %arr_stone, 0
+  jnz %arr_is_stone, @bad, @lens
+@lens
+  %len_p =l add %arr_ptr, 8
+  %len_l =l loadl %len_p
+  %len_w =w copy %len_l
+  %cap_l =l shr %arr_hdr, 8
+  %cap_w =w copy %cap_l
+  %need_grow =w csgew %idx_w, %cap_w
+  jnz %need_grow, @grow_init, @set_item
+@grow_init
+  %new_cap_w =w copy %cap_w
+  %cap_zero =w ceqw %new_cap_w, 0
+  jnz %cap_zero, @grow_cap0, @grow_check
+@grow_cap0
+  %new_cap_w =w copy 2
+  jmp @grow_check
+@grow_loop
+  %new_cap_w =w shl %new_cap_w, 1
+  %new_cap_neg =w csltw %new_cap_w, 0
+  jnz %new_cap_neg, @bad, @grow_check
+@grow_check
+  %need_more =w cslew %new_cap_w, %idx_w
+  jnz %need_more, @grow_loop, @grow_alloc
+@grow_alloc
+  %new_arr =l call $JS_NewArrayCap(l %ctx, w %new_cap_w)
+  %new_exc =w ceql %new_arr, 15
+  jnz %new_exc, @exc, @grow_refresh
+@grow_refresh
+  %fp2 =l call $cell_rt_refresh_fp_checked(l %ctx)
+  jnz %fp2, @grow_reload, @exc
+@grow_reload
+  %fp =l copy %fp2
+${sr("ga", "%obj_slot")}
+${sr("gb", "%val_slot")}
+${sr("gc", "%idx_slot")}
+  %a =l copy %ga
+  %b =l copy %gb
+  %c =l copy %gc
+  %arr_val =l copy %a
+  %arr_ptr =l and %arr_val, -8
+  %arr_hdr =l loadl %arr_ptr
+@grow_arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @grow_arr_follow, @grow_arr_ok
+@grow_arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @grow_arr_chase
+@grow_arr_ok
+  %grow_arr_is_array =w ceql %arr_ty, 0
+  jnz %grow_arr_is_array, @grow_arr_type_ok, @bad
+@grow_arr_type_ok
+  %old_cap_l =l shr %arr_hdr, 8
+  %old_len_p =l add %arr_ptr, 8
+  %old_len_l =l loadl %old_len_p
+  %old_len_w =w copy %old_len_l
+  %new_ptr =l and %new_arr, -8
+  %old_vals =l add %arr_ptr, 16
+  %new_vals =l add %new_ptr, 16
+  %i_w =w copy 0
+@copy_cond
+  %copy_more =w csltw %i_w, %old_len_w
+  jnz %copy_more, @copy_body, @copy_done
+@copy_body
+  %i_l =l extsw %i_w
+  %i_off =l shl %i_l, 3
+  %old_ep =l add %old_vals, %i_off
+  %new_ep =l add %new_vals, %i_off
+  %ev =l loadl %old_ep
+  %ev_is_self =w ceql %ev, %arr_val
+  jnz %ev_is_self, @copy_self, @copy_store
+@copy_self
+  storel %new_arr, %new_ep
+  jmp @copy_next
+@copy_store
+  storel %ev, %new_ep
+@copy_next
+  %i_w =w add %i_w, 1
+  jmp @copy_cond
+@copy_done
+  storel %old_len_l, %old_len_p
+  %old_size =l shl %old_cap_l, 3
+  %old_size =l add %old_size, 16
+  %fwd =l shl %new_ptr, 3
+  %fwd =l or %fwd, 7
+  storel %fwd, %arr_ptr
+  %arr_size_p =l add %arr_ptr, 8
+  storel %old_size, %arr_size_p
+  %obj_slot_o =l shl %obj_slot, 3
+  %obj_slot_p =l add %fp2, %obj_slot_o
+  storel %new_arr, %obj_slot_p
+  %arr_val =l copy %new_arr
+  %arr_ptr =l copy %new_ptr
+  %arr_hdr =l loadl %arr_ptr
+  %len_p =l add %arr_ptr, 8
+  storel %old_len_l, %len_p
+  %len_l =l copy %old_len_l
+  %len_w =w copy %old_len_w
+@set_item
+  %need_len =w csgew %idx_w, %len_w
+  jnz %need_len, @bump_len, @store_item
+@bump_len
+  %next_len_w =w add %idx_w, 1
+  %next_len_l =l extsw %next_len_w
+  storel %next_len_l, %len_p
+@store_item
+  %idx2_l =l extsw %idx_w
+  %idx2_off =l shl %idx2_l, 3
+  %vals_p =l add %arr_ptr, 16
+  %item_p =l add %vals_p, %idx2_off
+  storel %b, %item_p
   ret %fp
+@bad
+  call $cell_rt_disrupt(l %ctx)
 @exc
   ret 0
 }`
@@ -937,12 +1184,133 @@ ${alloc_tail("%r")}
 @entry
 ${sr("a", "%arr_slot")}
 ${sr("b", "%val_slot")}
-  %r =l call $cell_rt_push(l %ctx, l %a, l %b)
+  %ptag =l and %a, 7
+  %is_ptr =w ceql %ptag, 1
+  jnz %is_ptr, @arr_init, @bad
+@arr_init
+  %arr_val =l copy %a
+  %arr_ptr =l and %arr_val, -8
+  %arr_hdr =l loadl %arr_ptr
+@arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @arr_follow, @arr_ok
+@arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @arr_chase
+@arr_ok
+  %arr_is_array =w ceql %arr_ty, 0
+  jnz %arr_is_array, @arr_type_ok, @bad
+@arr_type_ok
+  %arr_stone =l and %arr_hdr, 8
+  %arr_is_stone =w cnel %arr_stone, 0
+  jnz %arr_is_stone, @bad, @lens
+@lens
+  %len_p =l add %arr_ptr, 8
+  %len_l =l loadl %len_p
+  %len_w =w copy %len_l
+  %cap_l =l shr %arr_hdr, 8
+  %cap_w =w copy %cap_l
+  %need_grow =w csgew %len_w, %cap_w
+  jnz %need_grow, @grow, @store_push
+@grow
+  %new_cap_w =w copy %cap_w
+  %cap_zero =w ceqw %new_cap_w, 0
+  jnz %cap_zero, @grow_cap0, @grow_dbl
+@grow_cap0
+  %new_cap_w =w copy 2
+  jmp @grow_alloc
+@grow_dbl
+  %new_cap_w =w shl %new_cap_w, 1
+  %new_cap_neg =w csltw %new_cap_w, 0
+  jnz %new_cap_neg, @bad, @grow_alloc
+@grow_alloc
+  %new_arr =l call $JS_NewArrayCap(l %ctx, w %new_cap_w)
+  %new_exc =w ceql %new_arr, 15
+  jnz %new_exc, @exc, @grow_refresh
+@grow_refresh
   %fp2 =l call $cell_rt_refresh_fp_checked(l %ctx)
-  jnz %fp2, @ok, @exc
-@ok
-${sw("w", "%fp2", "%arr_slot", "%r")}
-  ret %fp2
+  jnz %fp2, @grow_reload, @exc
+@grow_reload
+  %fp =l copy %fp2
+${sr("ga", "%arr_slot")}
+${sr("gb", "%val_slot")}
+  %a =l copy %ga
+  %b =l copy %gb
+  %arr_val =l copy %a
+  %arr_ptr =l and %arr_val, -8
+  %arr_hdr =l loadl %arr_ptr
+@grow_arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @grow_arr_follow, @grow_arr_ok
+@grow_arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @grow_arr_chase
+@grow_arr_ok
+  %grow_arr_is_array =w ceql %arr_ty, 0
+  jnz %grow_arr_is_array, @grow_arr_type_ok, @bad
+@grow_arr_type_ok
+  %old_cap_l =l shr %arr_hdr, 8
+  %old_len_p =l add %arr_ptr, 8
+  %old_len_l =l loadl %old_len_p
+  %old_len_w =w copy %old_len_l
+  %new_ptr =l and %new_arr, -8
+  %old_vals =l add %arr_ptr, 16
+  %new_vals =l add %new_ptr, 16
+  %i_w =w copy 0
+@copy_cond
+  %copy_more =w csltw %i_w, %old_len_w
+  jnz %copy_more, @copy_body, @copy_done
+@copy_body
+  %i_l =l extsw %i_w
+  %i_off =l shl %i_l, 3
+  %old_ep =l add %old_vals, %i_off
+  %new_ep =l add %new_vals, %i_off
+  %ev =l loadl %old_ep
+  %ev_is_self =w ceql %ev, %arr_val
+  jnz %ev_is_self, @copy_self, @copy_store
+@copy_self
+  storel %new_arr, %new_ep
+  jmp @copy_next
+@copy_store
+  storel %ev, %new_ep
+@copy_next
+  %i_w =w add %i_w, 1
+  jmp @copy_cond
+@copy_done
+  storel %old_len_l, %old_len_p
+  %old_size =l shl %old_cap_l, 3
+  %old_size =l add %old_size, 16
+  %fwd =l shl %new_ptr, 3
+  %fwd =l or %fwd, 7
+  storel %fwd, %arr_ptr
+  %arr_size_p =l add %arr_ptr, 8
+  storel %old_size, %arr_size_p
+  %arr_slot_o =l shl %arr_slot, 3
+  %arr_slot_p =l add %fp2, %arr_slot_o
+  storel %new_arr, %arr_slot_p
+  %arr_val =l copy %new_arr
+  %arr_ptr =l copy %new_ptr
+  %arr_hdr =l loadl %arr_ptr
+  %len_p =l add %arr_ptr, 8
+  storel %old_len_l, %len_p
+  %len_l =l copy %old_len_l
+  %len_w =w copy %old_len_w
+@store_push
+  %idx_l =l extsw %len_w
+  %idx_off =l shl %idx_l, 3
+  %vals_p =l add %arr_ptr, 16
+  %item_p =l add %vals_p, %idx_off
+  storel %b, %item_p
+  %next_len_w =w add %len_w, 1
+  %next_len_l =l extsw %next_len_w
+  storel %next_len_l, %len_p
+  ret %fp
+@bad
+  call $cell_rt_disrupt(l %ctx)
 @exc
   ret 0
 }`
@@ -951,8 +1319,51 @@ ${sw("w", "%fp2", "%arr_slot", "%r")}
   h[] = `export function l $__pop_ss(l %ctx, l %fp, l %dest, l %arr_slot) {
 @entry
 ${sr("a", "%arr_slot")}
-  %r =l call $cell_rt_pop(l %ctx, l %a)
-${alloc_tail("%r")}
+  %ptag =l and %a, 7
+  %is_ptr =w ceql %ptag, 1
+  jnz %is_ptr, @arr_init, @bad
+@arr_init
+  %arr_ptr =l and %a, -8
+  %arr_hdr =l loadl %arr_ptr
+@arr_chase
+  %arr_ty =l and %arr_hdr, 7
+  %arr_is_fwd =w ceql %arr_ty, 7
+  jnz %arr_is_fwd, @arr_follow, @arr_ok
+@arr_follow
+  %arr_ptr =l shr %arr_hdr, 3
+  %arr_hdr =l loadl %arr_ptr
+  jmp @arr_chase
+@arr_ok
+  %arr_is_array =w ceql %arr_ty, 0
+  jnz %arr_is_array, @arr_type_ok, @bad
+@arr_type_ok
+  %arr_stone =l and %arr_hdr, 8
+  %arr_is_stone =w cnel %arr_stone, 0
+  jnz %arr_is_stone, @bad, @len_chk
+@len_chk
+  %len_p =l add %arr_ptr, 8
+  %len_l =l loadl %len_p
+  %len_w =w copy %len_l
+  %empty =w ceqw %len_w, 0
+  jnz %empty, @ret_null, @do_pop
+@do_pop
+  %last_w =w sub %len_w, 1
+  %last_l =l extsw %last_w
+  %last_off =l shl %last_l, 3
+  %vals_p =l add %arr_ptr, 16
+  %item_p =l add %vals_p, %last_off
+  %r =l loadl %item_p
+  storel ${text(qbe.js_null)}, %item_p
+  %new_len_l =l extsw %last_w
+  storel %new_len_l, %len_p
+${sw("w", "%fp", "%dest", "%r")}
+  ret %fp
+@ret_null
+${sw("w", "%fp", "%dest", text(qbe.js_null))}
+  ret %fp
+@bad
+  call $cell_rt_disrupt(l %ctx)
+  ret 0
 }`
 
   // length(ctx, fp, dest, src)

source/qbe_helpers.c

@@ -779,6 +779,22 @@ static int cell_check_call_arity(JSContext *ctx, JSFunction *fn, int argc) {
   return 1;
 }
 
+static inline void cell_copy_args_0_4(JSValue *fp, JSValue *argv, int copy) {
+  /* fp[0] is `this`; copy args into fp[1..4] */
+  switch (copy) {
+  case 4: fp[4] = argv[3];
+  case 3: fp[3] = argv[2];
+  case 2: fp[2] = argv[1];
+  case 1: fp[1] = argv[0];
+  case 0: break;
+  default: break;
+  }
+}
+
+static inline void cell_sync_dl_from_native_fn(NativeRTState *st, JSFunction *fn) {
+  st->current_dl_handle = JS_VALUE_GET_CODE(fn->u.cell.code)->u.native.dl_handle;
+}
+
 /* Entry point called from JS_CallInternal / JS_Call / MACH_INVOKE
    for JS_FUNC_KIND_NATIVE functions. */
 JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
@@ -822,8 +838,14 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
   fp[0] = this_obj;
   int copy = (argc < arity) ? argc : arity;
   if (copy < 0) copy = argc; /* variadic: copy all */
-  for (int i = 0; i < copy && i < nr_slots - 1; i++)
-    fp[1 + i] = argv[i];
+  if (copy > nr_slots - 1)
+    copy = nr_slots - 1;
+  if (unlikely(copy > 4)) {
+    JS_RaiseDisrupt(ctx, "native calls support at most 4 arguments");
+    RETURN_DISPATCH(JS_EXCEPTION);
+  }
+  if (copy > 0 && argv)
+    cell_copy_args_0_4(fp, argv, copy);
 
   /* Link function to frame for closure access */
   JSFrameRegister *frame = (JSFrameRegister *)((char *)fp - offsetof(JSFrameRegister, slots));
@@ -875,6 +897,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
         /* Resume caller with exception pending */
         JSFunction *exc_fn = JS_VALUE_GET_FUNCTION(frame->function);
         fn = (cell_compiled_fn)JS_VALUE_GET_CODE(exc_fn->u.cell.code)->u.native.fn_ptr;
+        cell_sync_dl_from_native_fn(st, exc_fn);
         JS_PopGCRef(ctx, &callee_ref);
         continue;
       }
@@ -883,6 +906,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
       if (!cell_check_call_arity(ctx, callee_fn, callee_argc)) {
         JSFunction *exc_fn = JS_VALUE_GET_FUNCTION(frame->function);
         fn = (cell_compiled_fn)JS_VALUE_GET_CODE(exc_fn->u.cell.code)->u.native.fn_ptr;
+        cell_sync_dl_from_native_fn(st, exc_fn);
         JS_PopGCRef(ctx, &callee_ref);
         continue;
       }
@@ -910,6 +934,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
           frame = (JSFrameRegister *)JS_VALUE_GET_PTR(aot_gc_ref_at(st, st->aot_depth - 1)->val);
           fp = (JSValue *)frame->slots;
           fn = callee_ptr;
+          cell_sync_dl_from_native_fn(st, callee_fn);
         } else {
           /* Regular call: link caller and push prepared callee frame. */
           int ret_info = JS_VALUE_GET_INT(frame->address);
@@ -931,12 +956,14 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
             fp = (JSValue *)frame->slots;
             JSFunction *exc_fn = JS_VALUE_GET_FUNCTION(frame->function);
             fn = (cell_compiled_fn)JS_VALUE_GET_CODE(exc_fn->u.cell.code)->u.native.fn_ptr;
+            cell_sync_dl_from_native_fn(st, exc_fn);
             JS_PopGCRef(ctx, &callee_ref);
             continue;
           }
           frame = (JSFrameRegister *)JS_VALUE_GET_PTR(aot_gc_ref_at(st, st->aot_depth - 1)->val);
           fp = (JSValue *)frame->slots;
           fn = callee_ptr;
+          cell_sync_dl_from_native_fn(st, callee_fn);
         }
       } else {
         /* Non-native callee (C function, register VM, etc.) —
@@ -968,6 +995,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
              Just resume it — it will detect JS_EXCEPTION in the return slot. */
           JSFunction *exc_fn = JS_VALUE_GET_FUNCTION(frame->function);
           fn = (cell_compiled_fn)JS_VALUE_GET_CODE(exc_fn->u.cell.code)->u.native.fn_ptr;
+          cell_sync_dl_from_native_fn(st, exc_fn);
           JS_PopGCRef(ctx, &callee_ref);
           continue;
         }
@@ -999,6 +1027,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
           /* Resume caller */
           JSFunction *caller_fn = JS_VALUE_GET_FUNCTION(frame->function);
           fn = (cell_compiled_fn)JS_VALUE_GET_CODE(caller_fn->u.cell.code)->u.native.fn_ptr;
+          cell_sync_dl_from_native_fn(st, caller_fn);
         } else {
           /* Regular call: store result and resume current function */
           int ret_info = JS_VALUE_GET_INT(frame->address);
@@ -1008,6 +1037,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
           /* fn stays the same — we resume the same function at next segment */
           JSFunction *cur_fn = JS_VALUE_GET_FUNCTION(frame->function);
           fn = (cell_compiled_fn)JS_VALUE_GET_CODE(cur_fn->u.cell.code)->u.native.fn_ptr;
+          cell_sync_dl_from_native_fn(st, cur_fn);
         }
       }
       JS_PopGCRef(ctx, &callee_ref);
@@ -1041,6 +1071,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
 
       JSFunction *exc_caller_fn = JS_VALUE_GET_FUNCTION(frame->function);
       fn = (cell_compiled_fn)JS_VALUE_GET_CODE(exc_caller_fn->u.cell.code)->u.native.fn_ptr;
+      cell_sync_dl_from_native_fn(st, exc_caller_fn);
       continue;
     }
 
@@ -1065,6 +1096,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
 
     JSFunction *caller_fn = JS_VALUE_GET_FUNCTION(frame->function);
     fn = (cell_compiled_fn)JS_VALUE_GET_CODE(caller_fn->u.cell.code)->u.native.fn_ptr;
+    cell_sync_dl_from_native_fn(st, caller_fn);
     continue;
   }
 

source/runtime.c

@@ -150,7 +150,10 @@ int JS_IsPretext (JSValue v) {
 }
 
 JSValue *JS_PushGCRef (JSContext *ctx, JSGCRef *ref) {
-  assert(ref != ctx->top_gc_ref && "JS_ROOT used in a loop — same address pushed twice");
+  if (ref == ctx->top_gc_ref) {
+    fprintf(stderr, "[warn] JS_PushGCRef duplicate top ref (non-fatal)\n");
+    return &ref->val;
+  }
   ref->prev = ctx->top_gc_ref;
   ctx->top_gc_ref = ref;
   ref->val = JS_NULL;
@@ -158,13 +161,20 @@ JSValue *JS_PushGCRef (JSContext *ctx, JSGCRef *ref) {
 }
 
 JSValue JS_PopGCRef (JSContext *ctx, JSGCRef *ref) {
-  assert(ctx->top_gc_ref == ref && "JS_PopGCRef: not popping top of stack — mismatched push/pop");
-  ctx->top_gc_ref = ref->prev;
+  if (ctx->top_gc_ref == ref) {
+    ctx->top_gc_ref = ref->prev;
+    return ref->val;
+  }
+
+  fprintf(stderr, "[warn] JS_PopGCRef mismatched pop (non-fatal)\n");
   return ref->val;
 }
 
 JSValue *JS_AddGCRef (JSContext *ctx, JSGCRef *ref) {
-  assert(ref != ctx->last_gc_ref && "JS_AddGCRef: same address added twice — cycle in GC ref list");
+  if (ref == ctx->last_gc_ref) {
+    fprintf(stderr, "[warn] JS_AddGCRef duplicate tail ref (non-fatal)\n");
+    return &ref->val;
+  }
   ref->prev = ctx->last_gc_ref;
   ctx->last_gc_ref = ref;
   ref->val = JS_NULL;
@@ -10362,6 +10372,8 @@ static JSValue js_cell_pop (JSContext *ctx, JSValue this_val, int argc, JSValue
   if (!JS_IsArray (obj)) return JS_NULL;
 
   JSArray *arr = JS_VALUE_GET_ARRAY (obj);
+  if (objhdr_s (arr->mist_hdr))
+    return JS_RaiseDisrupt (ctx, "cannot pop from a stoned array");
 
   if (arr->len == 0) return JS_NULL;
 

streamline.cm

@@ -151,6 +151,21 @@ var streamline = function(ir, log) {
       slot_types[instr[1]] = src_type != null ? src_type : T_UNKNOWN
       return null
     }
+    if (op == "load_index") {
+      slot_types[instr[2]] = T_ARRAY
+      slot_types[instr[3]] = T_INT
+    } else if (op == "store_index") {
+      slot_types[instr[1]] = T_ARRAY
+      slot_types[instr[3]] = T_INT
+    } else if (op == "load_field") {
+      slot_types[instr[2]] = T_RECORD
+    } else if (op == "store_field") {
+      slot_types[instr[1]] = T_RECORD
+    } else if (op == "push") {
+      slot_types[instr[1]] = T_ARRAY
+    } else if (op == "pop") {
+      slot_types[instr[2]] = T_ARRAY
+    }
     rule = write_rules[op]
     if (rule != null) {
       typ = rule[1]
@@ -787,26 +802,32 @@ var streamline = function(ir, log) {
       // Dynamic access reduction
       if (op == "load_dynamic") {
         old_op = op
-        if (slot_is(slot_types, instr[3], T_TEXT)) {
+        if (slot_is(slot_types, instr[2], T_RECORD) && slot_is(slot_types, instr[3], T_TEXT)) {
           instr[0] = "load_field"
           if (events != null) {
             events[] = {
               event: "rewrite",
               pass: "eliminate_type_checks",
-              rule: "dynamic_to_field",
+              rule: "dynamic_record_to_field",
               at: i, before: old_op, after: instr[0],
-              why: {slot: instr[3], known_type: slot_types[instr[3]]}
+              why: {
+                object_slot: instr[2], object_type: slot_types[instr[2]],
+                key_slot: instr[3], key_type: slot_types[instr[3]]
+              }
             }
           }
-        } else if (slot_is(slot_types, instr[3], T_INT)) {
+        } else if (slot_is(slot_types, instr[2], T_ARRAY) && slot_is(slot_types, instr[3], T_INT)) {
           instr[0] = "load_index"
           if (events != null) {
             events[] = {
               event: "rewrite",
               pass: "eliminate_type_checks",
-              rule: "dynamic_to_index",
+              rule: "dynamic_array_to_index",
               at: i, before: old_op, after: instr[0],
-              why: {slot: instr[3], known_type: slot_types[instr[3]]}
+              why: {
+                object_slot: instr[2], object_type: slot_types[instr[2]],
+                key_slot: instr[3], key_type: slot_types[instr[3]]
+              }
             }
           }
         }
@@ -816,26 +837,32 @@ var streamline = function(ir, log) {
       }
       if (op == "store_dynamic") {
         old_op = op
-        if (slot_is(slot_types, instr[3], T_TEXT)) {
+        if (slot_is(slot_types, instr[1], T_RECORD) && slot_is(slot_types, instr[3], T_TEXT)) {
           instr[0] = "store_field"
           if (events != null) {
             events[] = {
               event: "rewrite",
               pass: "eliminate_type_checks",
-              rule: "dynamic_to_field",
+              rule: "dynamic_record_to_field",
               at: i, before: old_op, after: instr[0],
-              why: {slot: instr[3], known_type: slot_types[instr[3]]}
+              why: {
+                object_slot: instr[1], object_type: slot_types[instr[1]],
+                key_slot: instr[3], key_type: slot_types[instr[3]]
+              }
             }
           }
-        } else if (slot_is(slot_types, instr[3], T_INT)) {
+        } else if (slot_is(slot_types, instr[1], T_ARRAY) && slot_is(slot_types, instr[3], T_INT)) {
           instr[0] = "store_index"
           if (events != null) {
             events[] = {
               event: "rewrite",
               pass: "eliminate_type_checks",
-              rule: "dynamic_to_index",
+              rule: "dynamic_array_to_index",
               at: i, before: old_op, after: instr[0],
-              why: {slot: instr[3], known_type: slot_types[instr[3]]}
+              why: {
+                object_slot: instr[1], object_type: slot_types[instr[1]],
+                key_slot: instr[3], key_type: slot_types[instr[3]]
+              }
             }
           }
         }