Enable bulk-memory by default

embuilder.py

@@ -30,7 +30,6 @@
 
 # Minimal subset of targets used by CI systems to build enough to be useful
 MINIMAL_TASKS = [
-    'libbulkmemory',
     'libcompiler_rt',
     'libcompiler_rt-wasm-sjlj',
     'libcompiler_rt-ww',

emcc.py

@@ -385,14 +385,6 @@ def get_clang_flags(user_args):
     if '-mbulk-memory' not in user_args:
       flags.append('-mbulk-memory')
 
-  # In emscripten we currently disable bulk memory by default.
-  # This should be removed/updated when we als update the default browser targets.
-  if '-mbulk-memory' not in user_args and '-mno-bulk-memory' not in user_args:
-    # Bulk memory may be enabled via threads or directly via -s.
-    if not settings.BULK_MEMORY:
-      flags.append('-mno-bulk-memory')
-      flags.append('-mno-bulk-memory-opt')
-
   if settings.RELOCATABLE and '-fPIC' not in user_args:
     flags.append('-fPIC')
 

src/library.js

@@ -405,38 +405,6 @@ addToLibrary({
   // the initial values of the environment accessible by getenv.
   $ENV: {},
 
-  // In -Oz builds, we replace memcpy() altogether with a non-unrolled wasm
-  // variant, so we should never emit _emscripten_memcpy_js() in the build.
-  // In STANDALONE_WASM we avoid the _emscripten_memcpy_js dependency so keep
-  // the wasm file standalone.
-  // In BULK_MEMORY mode we include native versions of these functions based
-  // on memory.fill and memory.copy.
-  // In MAIN_MODULE=1 or EMCC_FORCE_STDLIBS mode all of libc is force included
-  // so we cannot override parts of it, and therefore cannot use libc_optz.
-#if (SHRINK_LEVEL < 2 || LINKABLE || process.env.EMCC_FORCE_STDLIBS) && !STANDALONE_WASM && !BULK_MEMORY
-
-#if MIN_CHROME_VERSION < 45 || MIN_FIREFOX_VERSION < 34 || MIN_SAFARI_VERSION < 100101
-  // https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/TypedArray/copyWithin lists browsers that support TypedArray.prototype.copyWithin, but it
-  // has outdated information for Safari, saying it would not support it.
-  // https://github.com/WebKit/webkit/commit/24a800eea4d82d6d595cdfec69d0f68e733b5c52#diff-c484911d8df319ba75fce0d8e7296333R1 suggests support was added on Aug 28, 2015.
-  // Manual testing suggests:
-  //   Safari/601.1 Version/9.0 on iPhone 4s with iOS 9.3.6 (released September 30, 2015) does not support copyWithin.
-  // but the following systems do:
-  //   AppleWebKit/602.2.14 Safari/602.1 Version/10.0 Mobile/14B100 iPhone OS 10_1_1 on iPhone 5s with iOS 10.1.1 (released October 31, 2016)
-  //   AppleWebKit/603.3.8 Safari/602.1 Version/10.0 on iPhone 5 with iOS 10.3.4 (released July 22, 2019)
-  //   AppleWebKit/605.1.15 iPhone OS 12_3_1 Version/12.1.1 Safari/604.1 on iPhone SE with iOS 12.3.1
-  //   AppleWebKit/605.1.15 Safari/604.1 Version/13.0.4 iPhone OS 13_3 on iPhone 6s with iOS 13.3
-  //   AppleWebKit/605.1.15 Version/13.0.3 Intel Mac OS X 10_15_1 on Safari 13.0.3 (15608.3.10.1.4) on macOS Catalina 10.15.1
-  // Hence the support status of .copyWithin() for Safari version range [10.0.0, 10.1.0] is unknown.
-  _emscripten_memcpy_js: `= Uint8Array.prototype.copyWithin
-    ? (dest, src, num) => HEAPU8.copyWithin(dest, src, src + num)
-    : (dest, src, num) => HEAPU8.set(HEAPU8.subarray(src, src+num), dest)`,
-#else
-  _emscripten_memcpy_js: (dest, src, num) => HEAPU8.copyWithin(dest, src, src + num),
-#endif
-
-#endif
-
 #if !STANDALONE_WASM
   // ==========================================================================
   // assert.h

src/library_sigs.js

@@ -322,7 +322,6 @@ sigs = {
   _emscripten_get_progname__sig: 'vpi',
   _emscripten_init_main_thread_js__sig: 'vp',
   _emscripten_lookup_name__sig: 'ip',
-  _emscripten_memcpy_js__sig: 'vppp',
   _emscripten_notify_mailbox_postmessage__sig: 'vpp',
   _emscripten_push_main_loop_blocker__sig: 'vppp',
   _emscripten_push_uncounted_main_loop_blocker__sig: 'vppp',

system/lib/libc/emscripten_internal.h

@@ -30,12 +30,6 @@ _Noreturn void _abort_js(void);
 
 void setThrew(uintptr_t threw, int value);
 
-// An external JS implementation that is efficient for very large copies, using
-// HEAPU8.set()
-void _emscripten_memcpy_js(void* __restrict__ dest,
-                           const void* __restrict__ src,
-                           size_t n) EM_IMPORT(_emscripten_memcpy_js);
-
 void* _emscripten_memcpy_bulkmem(void* __restrict__ dest,
                                  const void* __restrict__ src,
                                  size_t n);

system/lib/libc/emscripten_memcpy.c

@@ -8,21 +8,25 @@
 #include "libc.h"
 #include "emscripten_internal.h"
 
+#if !defined(__wasm_bulk_memory__)
+#error "This file must be compile with bulk memory enabled"
+#endif
+
 // Use the simple/naive version of memcpy when building with asan
-#if defined(EMSCRIPTEN_OPTIMIZE_FOR_OZ) || __has_feature(address_sanitizer)
+#if __has_feature(address_sanitizer)
 
 static void *__memcpy(void *dest, const void *src, size_t n) {
   unsigned char *d = (unsigned char *)dest;
   const unsigned char *s = (const unsigned char *)src;
-#pragma clang loop unroll(disable)
   while(n--) *d++ = *s++;
   return dest;
 }
 
-#elif defined(__wasm_bulk_memory__)
+#elif defined(EMSCRIPTEN_OPTIMIZE_FOR_OZ)
 
 static void *__memcpy(void *restrict dest, const void *restrict src, size_t n) {
-  return n ? _emscripten_memcpy_bulkmem(dest, src, n) : dest;
+  // TODO: Ensure this is inlined with Binaryen or inline asm
+  return _emscripten_memcpy_bulkmem(dest, src, n);
 }
 
 #else
@@ -35,12 +39,10 @@ static void *__memcpy(void *restrict dest, const void *restrict src, size_t n) {
   unsigned char *block_aligned_d_end;
   unsigned char *d_end;
 
-#if !defined(EMSCRIPTEN_STANDALONE_WASM)
   if (n >= 512) {
-    _emscripten_memcpy_js(dest, src, n);
-    return dest;
+    // TODO: Re-investigate the size threshold to enable this
+    return _emscripten_memcpy_bulkmem(dest, src, n);
   }
-#endif
 
   d_end = d + n;
   if ((((uintptr_t)d) & 3) == (((uintptr_t)s) & 3)) {

system/lib/libc/emscripten_memcpy_bulkmem.S

@@ -7,10 +7,17 @@
 .globl _emscripten_memcpy_bulkmem
 _emscripten_memcpy_bulkmem:
   .functype _emscripten_memcpy_bulkmem (PTR, PTR, PTR) -> (PTR)
-  local.get 0
-  local.get 1
   local.get 2
-  memory.copy 0, 0
+#ifdef __wasm64__
+  i32.wrap_i64
+#endif
+  // memory.copy traps on OOB zero-length copies, but memcpy must not.
+  if
+    local.get 0
+    local.get 1
+    local.get 2
+    memory.copy 0, 0
+  end_if
   local.get 0
   end_function
 

system/lib/libc/emscripten_memset.c

@@ -9,16 +9,11 @@ __attribute__((no_sanitize("address"))) void *__memset(void *str, int c, size_t
 __attribute__((__weak__)) void *__musl_memset(void *str, int c, size_t n);
 __attribute__((__weak__)) void *__memset(void *str, int c, size_t n);
 
-#ifdef EMSCRIPTEN_OPTIMIZE_FOR_OZ
-
-void *__memset(void *str, int c, size_t n) {
-  unsigned char *s = (unsigned char *)str;
-#pragma clang loop unroll(disable)
-  while(n--) *s++ = c;
-  return str;
-}
+#if !defined(__wasm_bulk_memory__)
+#error "This file must be compiled with bulk memory enabled"
+#endif
 
-#elif defined(__wasm_bulk_memory__)
+#if defined(EMSCRIPTEN_OPTIMIZE_FOR_OZ)
 
 void *__memset(void *str, int c, size_t n) {
   return _emscripten_memset_bulkmem(str, c, n);

system/lib/libc/emscripten_memset_bulkmem.S

@@ -7,10 +7,17 @@
 .globl _emscripten_memset_bulkmem
 _emscripten_memset_bulkmem:
   .functype _emscripten_memset_bulkmem (PTR, i32, PTR) -> (PTR)
-  local.get 0
-  local.get 1
   local.get 2
-  memory.fill 0
+#ifdef __wasm64__
+  i32.wrap_i64
+#endif
+  // memory.fill traps on OOB zero-length sets, but memset must not.
+  if
+   local.get 0
+   local.get 1
+   local.get 2
+   memory.fill 0
+  end_if
   local.get 0
   end_function
 

test/browser/test_small_js_flags.js.size

@@ -1 +1 @@
-4374
+4341

test/code_size/embind_hello_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 380,
   "a.js": 9879,
   "a.js.gz": 4288,
-  "a.wasm": 7580,
-  "a.wasm.gz": 3449,
-  "total": 18011,
-  "total_gz": 8117
+  "a.wasm": 7451,
+  "a.wasm.gz": 3375,
+  "total": 17882,
+  "total_gz": 8043
 }

test/code_size/embind_val_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 380,
   "a.js": 7153,
   "a.js.gz": 3042,
-  "a.wasm": 9493,
-  "a.wasm.gz": 4872,
-  "total": 17198,
-  "total_gz": 8294
+  "a.wasm": 9381,
+  "a.wasm.gz": 4818,
+  "total": 17086,
+  "total_gz": 8240
 }

test/code_size/hello_webgl2_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 328,
   "a.js": 4538,
   "a.js.gz": 2320,
-  "a.wasm": 10206,
-  "a.wasm.gz": 6663,
-  "total": 15198,
-  "total_gz": 9311
+  "a.wasm": 10148,
+  "a.wasm.gz": 6641,
+  "total": 15140,
+  "total_gz": 9289
 }

test/code_size/hello_webgl2_wasm2js.json

@@ -2,7 +2,7 @@
   "a.html": 346,
   "a.html.gz": 262,
   "a.js": 22202,
-  "a.js.gz": 11604,
+  "a.js.gz": 11636,
   "total": 22548,
-  "total_gz": 11866
+  "total_gz": 11898
 }

test/code_size/hello_webgl_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 328,
   "a.js": 4076,
   "a.js.gz": 2163,
-  "a.wasm": 10206,
-  "a.wasm.gz": 6663,
-  "total": 14736,
-  "total_gz": 9154
+  "a.wasm": 10148,
+  "a.wasm.gz": 6641,
+  "total": 14678,
+  "total_gz": 9132
 }

test/code_size/hello_webgl_wasm2js.json

@@ -2,7 +2,7 @@
   "a.html": 346,
   "a.html.gz": 262,
   "a.js": 21728,
-  "a.js.gz": 11435,
+  "a.js.gz": 11470,
   "total": 22074,
-  "total_gz": 11697
+  "total_gz": 11732
 }

test/code_size/hello_world_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 380,
   "a.js": 283,
   "a.js.gz": 244,
-  "a.wasm": 103,
+  "a.wasm": 106,
   "a.wasm.gz": 113,
-  "total": 938,
+  "total": 941,
   "total_gz": 737
 }

test/code_size/hello_world_wasm2js.js

@@ -15,6 +15,7 @@ function e(b) {
             for (var v, p = 0, t = a, w = f.length, y = a + (3 * w >> 2) - ("=" == f[w - 2]) - ("=" == f[w - 1]); p < w; p += 4) a = m[f.charCodeAt(p + 1)], 
             v = m[f.charCodeAt(p + 2)], c[t++] = m[f.charCodeAt(p)] << 2 | a >> 4, t < y && (c[t++] = a << 4 | v >> 2), 
             t < y && (c[t++] = v << 6 | m[f.charCodeAt(p + 3)]);
+            return c;
         }
         for (var q, m = new Uint8Array(123), n = 25; 0 <= n; --n) m[48 + n] = 52 + n, m[65 + n] = n, 
         m[97 + n] = 26 + n;

test/code_size/hello_world_wasm2js.json

@@ -1,8 +1,8 @@
 {
   "a.html": 323,
   "a.html.gz": 253,
-  "a.js": 1052,
-  "a.js.gz": 630,
-  "total": 1375,
-  "total_gz": 883
+  "a.js": 1061,
+  "a.js.gz": 629,
+  "total": 1384,
+  "total_gz": 882
 }

test/code_size/math_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 380,
   "a.js": 110,
   "a.js.gz": 125,
-  "a.wasm": 2695,
-  "a.wasm.gz": 1664,
-  "total": 3357,
-  "total_gz": 2169
+  "a.wasm": 2698,
+  "a.wasm.gz": 1666,
+  "total": 3360,
+  "total_gz": 2171
 }

test/code_size/random_printf_wasm.json

@@ -1,6 +1,6 @@
 {
-  "a.html": 12517,
-  "a.html.gz": 6871,
-  "total": 12517,
-  "total_gz": 6871
+  "a.html": 12449,
+  "a.html.gz": 6811,
+  "total": 12449,
+  "total_gz": 6811
 }

test/code_size/random_printf_wasm2js.json

@@ -1,6 +1,6 @@
 {
-  "a.html": 17195,
-  "a.html.gz": 7480,
-  "total": 17195,
-  "total_gz": 7480
+  "a.html": 17233,
+  "a.html.gz": 7535,
+  "total": 17233,
+  "total_gz": 7535
 }

test/other/codesize/test_codesize_cxx_ctors1.gzsize

@@ -1 +1 @@
-8408
+8390

test/other/codesize/test_codesize_cxx_ctors1.imports

@@ -1,6 +1,5 @@
 env.__cxa_throw
 env._abort_js
-env._emscripten_memcpy_js
 env._tzset_js
 env.emscripten_resize_heap
 wasi_snapshot_preview1.environ_get

test/other/codesize/test_codesize_cxx_ctors1.jssize

@@ -1 +1 @@
-20486
+20433

test/other/codesize/test_codesize_cxx_ctors1.sent

@@ -1,6 +1,5 @@
 __cxa_throw
 _abort_js
-_emscripten_memcpy_js
 _tzset_js
 emscripten_resize_heap
 environ_get

test/other/codesize/test_codesize_cxx_ctors1.size

@@ -1 +1 @@
-128914
+128556

test/other/codesize/test_codesize_cxx_ctors2.gzsize

@@ -1 +1 @@
-8392
+8374

test/other/codesize/test_codesize_cxx_ctors2.imports

@@ -1,5 +1,4 @@
 env._abort_js
-env._emscripten_memcpy_js
 env._tzset_js
 env.emscripten_resize_heap
 wasi_snapshot_preview1.fd_close

test/other/codesize/test_codesize_cxx_ctors2.jssize

@@ -1 +1 @@
-20454
+20401

test/other/codesize/test_codesize_cxx_ctors2.sent

@@ -1,6 +1,5 @@
 __cxa_throw
 _abort_js
-_emscripten_memcpy_js
 _tzset_js
 emscripten_resize_heap
 environ_get

test/other/codesize/test_codesize_cxx_ctors2.size

@@ -1 +1 @@
-128363
+127960

test/other/codesize/test_codesize_cxx_except.gzsize

@@ -1 +1 @@
-9413
+9393

test/other/codesize/test_codesize_cxx_except.imports

@@ -7,7 +7,6 @@ env.__cxa_throw
 env.__cxa_uncaught_exceptions
 env.__resumeException
 env._abort_js
-env._emscripten_memcpy_js
 env._tzset_js
 env.emscripten_resize_heap
 env.invoke_diii
@@ -17,7 +16,6 @@ env.invoke_ii
 env.invoke_iii
 env.invoke_iiii
 env.invoke_iiiii
-env.invoke_iiiiii
 env.invoke_iiiiiii
 env.invoke_iiiiiiii
 env.invoke_iiiiiiiiiiii

test/other/codesize/test_codesize_cxx_except.jssize

@@ -1 +1 @@
-24255
+24202