Enable bulk-memory by default (emscripten-core#22873)

* Remove libbulkmemory `_emscripten_memcpy_js` and fold memcpy and
memset into libc
* Use bulk memcpy/memset for Oz builds, but keep ASan behavior the same.
* Move the zero-length check in memcpy from C into assembly, and add one
for memset
* Remove the use of `-mno-bulk-memory` at compile time (enabling it in
object files)
* Temporarily set the Safari version required to use bulk memory to a
14.1 (which has the effect of enabling it by default without enabling
the other 14.1 features by default). This will be reverted when
nontrapping-fptoint and bigint are also enabled by default.

See emscripten-core#23184

Author: dschuff

.circleci/config.yml

@@ -240,6 +240,10 @@ commands:
         description: "Name of given test suite"
         type: string
         default: ""
+      extra-cflags:
+        description: "Extra EMCC_CFLAGS"
+        type: string
+        default: ""
     steps:
       - when:
           # We only set EMTEST_RETRY_FLAKY on pull requests.  When we run
@@ -250,6 +254,8 @@ commands:
             - set-retry-flaky-tests
       - run:
           name: run tests (<< parameters.title >>)
+          environment:
+            EMCC_CFLAGS: << parameters.extra-cflags >>
           command: |
             env
             ./test/runner << parameters.test_targets >>
@@ -765,6 +771,7 @@ jobs:
           node_version: "10.19.0"
       - run-tests:
           title: "node (oldest / 10.19.0)"
+          extra-cflags: "-sMIN_NODE_VERSION=101900"
           test_targets: "
             other.test_gen_struct_info
             other.test_native_call_before_init

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

@@ -321,7 +321,6 @@ 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

@@ -9,20 +9,20 @@
 #include "emscripten_internal.h"
 
 // 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 +35,10 @@ 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 +9,7 @@ __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;
-}
-
-#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