Always use bulkmemory memcpy and memset when available (#20144)

sbc100 · web-flow · commit 60b3d4edc1b7 · 2023-08-28T14:46:50.000-07:00
Unlike the JS versions of these function there is no need to only use these for small inputs. Results of running the test_memcpy_128b test before and after this change: ``` v8-bulk: mean: 1.536 (+-0.071) secs median: 1.495 range: 1.471-1.650 (noise: 4.630%) (5 runs) size: 149291, compressed: 54249 ``` -> ``` v8-bulk: mean: 1.489 (+-0.117) secs median: 1.535 range: 1.268-1.606 (noise: 7.871%) (5 runs)- size: 148387, compressed: 53813- ``` See comments in #19128
diff --git a/src/library.js b/src/library.js
@@ -385,8 +385,8 @@ addToLibrary({
   $ENV: {},
 
   // In -Oz builds, we replace memcpy() altogether with a non-unrolled wasm
-  // variant, so we should never emit emscripten_memcpy_big() in the build.
-  // In STANDALONE_WASM we avoid the emscripten_memcpy_big dependency so keep
+  // 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.
@@ -407,11 +407,11 @@ addToLibrary({
   //   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_big: `= Uint8Array.prototype.copyWithin
+  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_big: (dest, src, num) => HEAPU8.copyWithin(dest, src, src + num),
+  emscripten_memcpy_js: (dest, src, num) => HEAPU8.copyWithin(dest, src, src + num),
 #endif
 
 #endif
diff --git a/src/library_sigs.js b/src/library_sigs.js
@@ -690,7 +690,7 @@ sigs = {
   emscripten_math_sqrt__sig: 'dd',
   emscripten_math_tan__sig: 'dd',
   emscripten_math_tanh__sig: 'dd',
-  emscripten_memcpy_big__sig: 'vppp',
+  emscripten_memcpy_js__sig: 'vppp',
   emscripten_navigator_hardware_concurrency__sig: 'i',
   emscripten_notify_memory_growth__sig: 'vp',
   emscripten_num_logical_cores__sig: 'i',
diff --git a/system/lib/libc/emscripten_internal.h b/system/lib/libc/emscripten_internal.h
@@ -28,10 +28,14 @@ extern "C" {
 
 // An external JS implementation that is efficient for very large copies, using
 // HEAPU8.set()
-void emscripten_memcpy_big(void* __restrict__ dest,
+void emscripten_memcpy_js(void* __restrict__ dest,
                            const void* __restrict__ src,
-                           size_t n) EM_IMPORT(emscripten_memcpy_big);
-void emscripten_memset_big(void* ptr, char value, size_t n);
+                           size_t n) EM_IMPORT(emscripten_memcpy_js);
+
+void* emscripten_memcpy_bulkmem(void* __restrict__ dest,
+                                const void* __restrict__ src,
+                                size_t n);
+void* emscripten_memset_bulkmem(void* ptr, char value, size_t n);
 
 void emscripten_notify_memory_growth(size_t memory_index);
 
diff --git a/system/lib/libc/emscripten_memcpy.c b/system/lib/libc/emscripten_memcpy.c
@@ -19,6 +19,12 @@ static void *__memcpy(void *dest, const void *src, size_t n) {
   return dest;
 }
 
+#elif defined(__wasm_bulk_memory__)
+
+static void *__memcpy(void *restrict dest, const void *restrict src, size_t n) {
+  return emscripten_memcpy_bulkmem(dest, src, n);
+}
+
 #else
 
 static void *__memcpy(void *restrict dest, const void *restrict src, size_t n) {
@@ -29,9 +35,9 @@ 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) || defined(__wasm_bulk_memory__)
+#if !defined(EMSCRIPTEN_STANDALONE_WASM)
   if (n >= 512) {
-    emscripten_memcpy_big(dest, src, n);
+    emscripten_memcpy_js(dest, src, n);
     return dest;
   }
 #endif
diff --git a/system/lib/libc/emscripten_memcpy_bulkmem.S b/system/lib/libc/emscripten_memcpy_bulkmem.S
@@ -4,11 +4,12 @@
 #define PTR i32
 #endif
 
-.globl emscripten_memcpy_big
-emscripten_memcpy_big:
-  .functype emscripten_memcpy_big (PTR, PTR, PTR) -> ()
+.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
+  local.get 0
   end_function
diff --git a/system/lib/libc/emscripten_memset.c b/system/lib/libc/emscripten_memset.c
@@ -20,16 +20,8 @@ void *__memset(void *str, int c, size_t n) {
 
 #elif defined(__wasm_bulk_memory__)
 
-#define memset __musl_memset
-#include "musl/src/string/memset.c"
-#undef memset
-
 void *__memset(void *str, int c, size_t n) {
-  if (n >= 512) {
-    emscripten_memset_big(str, c, n);
-    return str;
-  }
-  return __musl_memset(str, c, n);
+  return emscripten_memset_bulkmem(str, c, n);
 }
 
 #else
diff --git a/system/lib/libc/emscripten_memset_bulkmem.S b/system/lib/libc/emscripten_memset_bulkmem.S
@@ -4,11 +4,12 @@
 #define PTR i32
 #endif
 
-.globl emscripten_memset_big
-emscripten_memset_big:
-  .functype emscripten_memset_big (PTR, i32, PTR) -> ()
+.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
+  local.get 0
   end_function
diff --git a/system/lib/standalone/standalone.c b/system/lib/standalone/standalone.c
@@ -142,11 +142,6 @@ weak int getentropy(void* buffer, size_t length) {
 
 // Emscripten additions
 
-// Should never be called in standalone mode
-weak void emscripten_memcpy_big(void *restrict dest, const void *restrict src, size_t n) {
-  __builtin_unreachable();
-}
-
 size_t emscripten_get_heap_max() {
   // In standalone mode we don't have any wasm instructions to access the max
   // memory size so the best we can do (without calling an import) is return
diff --git a/test/other/metadce/test_metadce_cxx_ctors1.imports b/test/other/metadce/test_metadce_cxx_ctors1.imports
@@ -1,6 +1,6 @@
 env.__cxa_throw
 env.abort
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.strftime_l
 wasi_snapshot_preview1.environ_get
diff --git a/test/other/metadce/test_metadce_cxx_ctors1.jssize b/test/other/metadce/test_metadce_cxx_ctors1.jssize
@@ -1 +1 @@
-25060
+25059
diff --git a/test/other/metadce/test_metadce_cxx_ctors1.sent b/test/other/metadce/test_metadce_cxx_ctors1.sent
@@ -1,6 +1,6 @@
 __cxa_throw
 abort
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 environ_get
 environ_sizes_get
diff --git a/test/other/metadce/test_metadce_cxx_ctors1.size b/test/other/metadce/test_metadce_cxx_ctors1.size
@@ -1 +1 @@
-123498
+123497
diff --git a/test/other/metadce/test_metadce_cxx_ctors2.imports b/test/other/metadce/test_metadce_cxx_ctors2.imports
@@ -1,6 +1,6 @@
 env.__cxa_throw
 env.abort
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.strftime_l
 wasi_snapshot_preview1.fd_close
diff --git a/test/other/metadce/test_metadce_cxx_ctors2.jssize b/test/other/metadce/test_metadce_cxx_ctors2.jssize
@@ -1 +1 @@
-25028
+25027
diff --git a/test/other/metadce/test_metadce_cxx_ctors2.sent b/test/other/metadce/test_metadce_cxx_ctors2.sent
@@ -1,6 +1,6 @@
 __cxa_throw
 abort
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 environ_get
 environ_sizes_get
diff --git a/test/other/metadce/test_metadce_cxx_ctors2.size b/test/other/metadce/test_metadce_cxx_ctors2.size
@@ -1 +1 @@
-123403
+123402
diff --git a/test/other/metadce/test_metadce_cxx_except.imports b/test/other/metadce/test_metadce_cxx_except.imports
@@ -7,7 +7,7 @@ env.__cxa_throw
 env.__cxa_uncaught_exceptions
 env.__resumeException
 env.abort
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.invoke_diii
 env.invoke_fiii
diff --git a/test/other/metadce/test_metadce_cxx_except.jssize b/test/other/metadce/test_metadce_cxx_except.jssize
@@ -1 +1 @@
-29193
+29192
diff --git a/test/other/metadce/test_metadce_cxx_except.sent b/test/other/metadce/test_metadce_cxx_except.sent
@@ -7,7 +7,7 @@ __cxa_throw
 __cxa_uncaught_exceptions
 __resumeException
 abort
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 environ_get
 environ_sizes_get
diff --git a/test/other/metadce/test_metadce_cxx_except.size b/test/other/metadce/test_metadce_cxx_except.size
@@ -1 +1 @@
-166328
+166327
diff --git a/test/other/metadce/test_metadce_cxx_except_wasm.imports b/test/other/metadce/test_metadce_cxx_except_wasm.imports
@@ -1,5 +1,5 @@
 env.abort
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.strftime_l
 wasi_snapshot_preview1.environ_get
diff --git a/test/other/metadce/test_metadce_cxx_except_wasm.jssize b/test/other/metadce/test_metadce_cxx_except_wasm.jssize
@@ -1 +1 @@
-24834
+24833
diff --git a/test/other/metadce/test_metadce_cxx_except_wasm.sent b/test/other/metadce/test_metadce_cxx_except_wasm.sent
@@ -1,5 +1,5 @@
 abort
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 environ_get
 environ_sizes_get
diff --git a/test/other/metadce/test_metadce_cxx_except_wasm.size b/test/other/metadce/test_metadce_cxx_except_wasm.size
@@ -1 +1 @@
-137117
+137116
diff --git a/test/other/metadce/test_metadce_cxx_mangle.imports b/test/other/metadce/test_metadce_cxx_mangle.imports
@@ -7,7 +7,7 @@ env.__cxa_throw
 env.__cxa_uncaught_exceptions
 env.__resumeException
 env.abort
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.invoke_diii
 env.invoke_fiii
diff --git a/test/other/metadce/test_metadce_cxx_mangle.jssize b/test/other/metadce/test_metadce_cxx_mangle.jssize
@@ -1 +1 @@
-29192
+29191
diff --git a/test/other/metadce/test_metadce_cxx_mangle.sent b/test/other/metadce/test_metadce_cxx_mangle.sent
@@ -7,7 +7,7 @@ __cxa_throw
 __cxa_uncaught_exceptions
 __resumeException
 abort
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 environ_get
 environ_sizes_get
diff --git a/test/other/metadce/test_metadce_cxx_mangle.size b/test/other/metadce/test_metadce_cxx_mangle.size
@@ -1 +1 @@
-221419
+221418
diff --git a/test/other/metadce/test_metadce_cxx_noexcept.imports b/test/other/metadce/test_metadce_cxx_noexcept.imports
@@ -1,6 +1,6 @@
 env.__cxa_throw
 env.abort
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.strftime_l
 wasi_snapshot_preview1.environ_get
diff --git a/test/other/metadce/test_metadce_cxx_noexcept.jssize b/test/other/metadce/test_metadce_cxx_noexcept.jssize
@@ -1 +1 @@
-25060
+25059
diff --git a/test/other/metadce/test_metadce_cxx_noexcept.sent b/test/other/metadce/test_metadce_cxx_noexcept.sent
@@ -1,6 +1,6 @@
 __cxa_throw
 abort
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 environ_get
 environ_sizes_get
diff --git a/test/other/metadce/test_metadce_cxx_noexcept.size b/test/other/metadce/test_metadce_cxx_noexcept.size
@@ -1 +1 @@
-126300
+126299
diff --git a/test/other/metadce/test_metadce_cxx_wasmfs.imports b/test/other/metadce/test_metadce_cxx_wasmfs.imports
@@ -11,7 +11,7 @@ env._wasmfs_stdin_get_char
 env.abort
 env.emscripten_date_now
 env.emscripten_err
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_out
 env.emscripten_resize_heap
 env.getentropy
diff --git a/test/other/metadce/test_metadce_cxx_wasmfs.jssize b/test/other/metadce/test_metadce_cxx_wasmfs.jssize
@@ -1 +1 @@
-12469
+12468
diff --git a/test/other/metadce/test_metadce_cxx_wasmfs.sent b/test/other/metadce/test_metadce_cxx_wasmfs.sent
@@ -11,7 +11,7 @@ _wasmfs_stdin_get_char
 abort
 emscripten_date_now
 emscripten_err
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_out
 emscripten_resize_heap
 environ_get
diff --git a/test/other/metadce/test_metadce_cxx_wasmfs.size b/test/other/metadce/test_metadce_cxx_wasmfs.size
@@ -1 +1 @@
-163625
+163624
diff --git a/test/other/metadce/test_metadce_hello_O0.imports b/test/other/metadce/test_metadce_hello_O0.imports
@@ -1,2 +1,2 @@
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 wasi_snapshot_preview1.fd_write
diff --git a/test/other/metadce/test_metadce_hello_O0.jssize b/test/other/metadce/test_metadce_hello_O0.jssize
@@ -1 +1 @@
-23682
+23681
diff --git a/test/other/metadce/test_metadce_hello_O0.sent b/test/other/metadce/test_metadce_hello_O0.sent
@@ -1,2 +1,2 @@
-emscripten_memcpy_big
+emscripten_memcpy_js
 fd_write
diff --git a/test/other/metadce/test_metadce_hello_O0.size b/test/other/metadce/test_metadce_hello_O0.size
@@ -1 +1 @@
-12156
+12155
diff --git a/test/other/metadce/test_metadce_hello_O1.imports b/test/other/metadce/test_metadce_hello_O1.imports
@@ -1,2 +1,2 @@
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 wasi_snapshot_preview1.fd_write
diff --git a/test/other/metadce/test_metadce_hello_O1.jssize b/test/other/metadce/test_metadce_hello_O1.jssize
@@ -1 +1 @@
-8094
+8093
diff --git a/test/other/metadce/test_metadce_hello_O1.sent b/test/other/metadce/test_metadce_hello_O1.sent
@@ -1,2 +1,2 @@
-emscripten_memcpy_big
+emscripten_memcpy_js
 fd_write
diff --git a/test/other/metadce/test_metadce_hello_O1.size b/test/other/metadce/test_metadce_hello_O1.size
@@ -1 +1 @@
-2413
+2412
diff --git a/test/other/metadce/test_metadce_hello_O2.imports b/test/other/metadce/test_metadce_hello_O2.imports
@@ -1,2 +1,2 @@
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 wasi_snapshot_preview1.fd_write
diff --git a/test/other/metadce/test_metadce_hello_O2.jssize b/test/other/metadce/test_metadce_hello_O2.jssize
@@ -1 +1 @@
-5753
+5752
diff --git a/test/other/metadce/test_metadce_hello_O2.sent b/test/other/metadce/test_metadce_hello_O2.sent
@@ -1,2 +1,2 @@
-emscripten_memcpy_big
+emscripten_memcpy_js
 fd_write
diff --git a/test/other/metadce/test_metadce_hello_O2.size b/test/other/metadce/test_metadce_hello_O2.size
@@ -1 +1 @@
-2054
+2053
diff --git a/test/other/metadce/test_metadce_hello_dylink.imports b/test/other/metadce/test_metadce_hello_dylink.imports
@@ -3,7 +3,7 @@ env.__indirect_function_table
 env.__memory_base
 env.__stack_pointer
 env.__table_base
-env.emscripten_memcpy_big
+env.emscripten_memcpy_js
 env.emscripten_resize_heap
 env.memory
 wasi_snapshot_preview1.fd_write
diff --git a/test/other/metadce/test_metadce_hello_dylink.jssize b/test/other/metadce/test_metadce_hello_dylink.jssize
@@ -1 +1 @@
-14820
+14819
diff --git a/test/other/metadce/test_metadce_hello_dylink.sent b/test/other/metadce/test_metadce_hello_dylink.sent
@@ -3,7 +3,7 @@ __indirect_function_table
 __memory_base
 __stack_pointer
 __table_base
-emscripten_memcpy_big
+emscripten_memcpy_js
 emscripten_resize_heap
 fd_write
 memory
diff --git a/test/other/metadce/test_metadce_hello_dylink.size b/test/other/metadce/test_metadce_hello_dylink.size
@@ -1 +1 @@
-9377
+9376
diff --git a/test/other/test_unoptimized_code_size.js.size b/test/other/test_unoptimized_code_size.js.size
@@ -1 +1 @@
-58712
+58709
diff --git a/test/other/test_unoptimized_code_size.wasm.size b/test/other/test_unoptimized_code_size.wasm.size
@@ -1 +1 @@
-12153
+12152
diff --git a/test/other/test_unoptimized_code_size_no_asserts.js.size b/test/other/test_unoptimized_code_size_no_asserts.js.size
@@ -1 +1 @@
-32008
+32005
diff --git a/test/other/test_unoptimized_code_size_no_asserts.wasm.size b/test/other/test_unoptimized_code_size_no_asserts.wasm.size
@@ -1 +1 @@
-11600
+11599
diff --git a/test/other/test_unoptimized_code_size_strict.js.size b/test/other/test_unoptimized_code_size_strict.js.size
@@ -1 +1 @@
-57399
+57396
diff --git a/test/other/test_unoptimized_code_size_strict.wasm.size b/test/other/test_unoptimized_code_size_strict.wasm.size
@@ -1 +1 @@
-12153
+12152
diff --git a/test/test_other.py b/test/test_other.py
@@ -13486,11 +13486,11 @@ def run(args, expect_bulk_mem):
       funcs = self.parse_wasm('test_memops_bulk_memory.wasm')[2]
       js = read_file('test_memops_bulk_memory.js')
       if expect_bulk_mem:
-        self.assertNotContained('_emscripten_memcpy_big', js)
-        self.assertIn('$emscripten_memcpy_big', funcs)
+        self.assertNotContained('_emscripten_memcpy_js', js)
+        self.assertIn('$emscripten_memcpy_bulkmem', funcs)
       else:
-        self.assertContained('_emscripten_memcpy_big', js)
-        self.assertNotIn('$emscripten_memcpy_big', funcs)
+        self.assertContained('_emscripten_memcpy_js', js)
+        self.assertNotIn('$emscripten_memcpy_bulkmem', funcs)
 
     # By default we expect to find _emscripten_memcpy_big in the generaed JS and not in the
     # native code.
diff --git a/tools/system_libs.py b/tools/system_libs.py

Original file line number	Diff line number	Diff line change
`@@ -1,2 +1,2 @@`
`1`		`-env.emscripten_memcpy_big`
	`1`	`+env.emscripten_memcpy_js`
`2`	`2`	`wasi_snapshot_preview1.fd_write`
Original file line number	Diff line number	Diff line change
`@@ -1,2 +1,2 @@`
`1`		`-emscripten_memcpy_big`
	`1`	`+emscripten_memcpy_js`
`2`	`2`	`fd_write`