Implement -s SINGLE_FILE_BINARY_ENCODE=1 option to embed Wasm binary as binary-encoded form instead of base64 form in SINGLE_FILE mode. Continuation of emscripten-core#21478.

Author: juj

src/preamble.js

@@ -410,6 +410,10 @@ function instrumentWasmTableWithAbort() {
 }
 #endif
 
+#if SINGLE_FILE && SINGLE_FILE_BINARY_ENCODE && !WASM2JS
+#include "binaryDecode.js"
+#endif
+
 #if !SOURCE_PHASE_IMPORTS && !WASM_ESM_INTEGRATION
 var wasmBinaryFile;
 
@@ -423,27 +427,35 @@ function getWasmBinary(file) {}
 #else
 
 function findWasmBinary() {
-#if SINGLE_FILE
+#if SINGLE_FILE && SINGLE_FILE_BINARY_ENCODE && !WASM2JS
+  return binaryDecode('<<< WASM_BINARY_DATA >>>');
+#elif SINGLE_FILE
   return base64Decode('<<< WASM_BINARY_DATA >>>');
+#elif AUDIO_WORKLET || !EXPORT_ES6
+  // For an Audio Worklet, we cannot use `new URL()`.
+  return locateFile('{{{ WASM_BINARY_FILE }}}');
 #else
-#if EXPORT_ES6 && !AUDIO_WORKLET
-  if (Module['locateFile']) {
-#endif
-    return locateFile('{{{ WASM_BINARY_FILE }}}');
-#if EXPORT_ES6 && !AUDIO_WORKLET // For an Audio Worklet, we cannot use `new URL()`.
-  }
+
 #if ENVIRONMENT_MAY_BE_SHELL
   if (ENVIRONMENT_IS_SHELL) {
     return '{{{ WASM_BINARY_FILE }}}';
   }
 #endif
+
+  if (Module['locateFile']) {
+    return locateFile('{{{ WASM_BINARY_FILE }}}');
+  }
+
   // Use bundler-friendly `new URL(..., import.meta.url)` pattern; works in browsers too.
   return new URL('{{{ WASM_BINARY_FILE }}}', import.meta.url).href;
-#endif
+
 #endif
 }
 
 function getBinarySync(file) {
+#if SINGLE_FILE && SINGLE_FILE_BINARY_ENCODE
+  return file;
+#else
 #if SINGLE_FILE
   if (ArrayBuffer.isView(file)) {
     return file;
@@ -464,6 +476,7 @@ function getBinarySync(file) {
 #else
   throw 'sync fetching of the wasm failed: you can preload it to Module["wasmBinary"] manually, or emcc.py will do that for you when generating HTML (but not JS)';
 #endif
+#endif
 }
 
 async function getWasmBinary(binaryFile) {

src/settings.js

@@ -1853,6 +1853,13 @@ var WASMFS = false;
 // [link]
 var SINGLE_FILE = false;
 
+// If true, binary Wasm content is encoded using a custom UTF-8 embedding
+// instead of base64. This generates a smaller binary that compresses well.
+// Set this to false to revert back to earlier base64 encoding if you run into
+// issues with the binary encoding. (and please let us know of any such issues)
+// [link]
+var SINGLE_FILE_BINARY_ENCODE = true;
+
 // If set to 1, all JS libraries will be automatically available at link time.
 // This gets set to 0 in STRICT mode (or with MINIMAL_RUNTIME) which mean you
 // need to explicitly specify -lfoo.js in at link time in order to access

test/test_codesize.py

@@ -43,6 +43,7 @@ class codesize(RunnerCore):
     'random_printf_wasm2js': ('random_printf', True),
     'hello_webgl_wasm': ('hello_webgl', False),
     'hello_webgl_wasm2js': ('hello_webgl', True),
+    'hello_webgl2_wasm_singlefile': ('hello_webgl2_wasm_singlefile', False),
     'hello_webgl2_wasm': ('hello_webgl2', False),
     'hello_webgl2_wasm2js': ('hello_webgl2', True),
     'math': ('math', False),
@@ -87,6 +88,7 @@ def test_minimal_runtime_code_size(self, test_name, wasm2js, compare_js_output=F
                            '-lGL',
                            '-sMODULARIZE']
     hello_webgl2_sources = hello_webgl_sources + ['-sMAX_WEBGL_VERSION=2']
+    hello_webgl2_wasm_singlefile_sources = hello_webgl2_sources + ['-sSINGLE_FILE']
     hello_wasm_worker_sources = [test_file('wasm_worker/wasm_worker_code_size.c'), '-sWASM_WORKERS', '-sENVIRONMENT=web']
     audio_worklet_sources = [test_file('webaudio/audioworklet.c'), '-sWASM_WORKERS', '-sAUDIO_WORKLET', '-sENVIRONMENT=web', '-sTEXTDECODER=1']
     embind_hello_sources = [test_file('codesize/embind_hello_world.cpp'), '-lembind']
@@ -98,6 +100,7 @@ def test_minimal_runtime_code_size(self, test_name, wasm2js, compare_js_output=F
       'hello_webgl': hello_webgl_sources,
       'math': math_sources,
       'hello_webgl2': hello_webgl2_sources,
+      'hello_webgl2_wasm_singlefile': hello_webgl2_wasm_singlefile_sources,
       'hello_wasm_worker': hello_wasm_worker_sources,
       'audio_worklet': audio_worklet_sources,
       'embind_val': embind_val_sources,

test/test_other.py

@@ -9475,8 +9475,9 @@ def test_standalone_system_headers(self, prefix):
     'closure': (False, True),
   })
   @parameterized({
-    '': (True,),
-    'disabled': (False,),
+    '': (1,),
+    'disabled': (0,),
+    'binary_encode': (2,),
   })
   def test_single_file(self, debug_enabled, closure_enabled, single_file_enabled):
     cmd = [EMCC, test_file('hello_world.c')] + self.get_cflags()
@@ -9487,6 +9488,8 @@ def test_single_file(self, debug_enabled, closure_enabled, single_file_enabled):
     else:
       expect_wasm = self.is_wasm()
 
+    cmd += [f'-sSINGLE_FILE_BINARY_ENCODE={int(single_file_enabled == 2)}']
+
     if debug_enabled:
       cmd += ['-g']
     if closure_enabled:

tools/link.py

@@ -163,6 +163,10 @@ def base64_encode(filename):
   return b64.decode('ascii')
 
 
+def base64_or_binary_encode(b):
+  return binary_encode(b) if settings.SINGLE_FILE and settings.SINGLE_FILE_BINARY_ENCODE else base64_encode(b)
+
+
 def align_to_wasm_page_boundary(address):
   page_size = webassembly.WASM_PAGE_SIZE
   return ((address + (page_size - 1)) // page_size) * page_size
@@ -2435,7 +2439,7 @@ def phase_binaryen(target, options, wasm_target):
   if final_js and settings.SINGLE_FILE and not settings.WASM2JS:
     js = read_file(final_js)
 
-    js = do_replace(js, '<<< WASM_BINARY_DATA >>>', base64_encode(wasm_target))
+    js = do_replace(js, '<<< WASM_BINARY_DATA >>>', base64_or_binary_encode(wasm_target))
     delete_file(wasm_target)
     write_file(final_js, js)
 
@@ -2945,9 +2949,44 @@ def move_file(src, dst):
   shutil.move(src, dst)
 
 
+def binary_encode(filename):
+  """This function encodes the given binary byte array to a UTF-8 string, by
+  first adding +1 to all the bytes [0, 255] to form values [1, 256], and then
+  encoding each of those values as UTF-8, except for specific byte values that
+  are escaped as two bytes. This kind of encoding results in a string that will
+  compress well by both gzip and brotli, unlike base64 encoding binary data
+  would do, and avoids emitting the null byte inside a string.
+  """
+
+  data = utils.read_binary(filename)
+
+  out = bytearray(len(data) * 2) # Size output buffer conservatively
+  i = 0
+  for d in data:
+    d += 1 # Offset all bytes up by +1 to make zero (a very common value) be encoded with only one byte as 0x01. This is possible since we can encode 255 as 0x100 in UTF-8.
+    if d == ord("'"):
+      buf = [ord('\\'), d] # Escape single quote ' character with a backspace since we are writing a string inside single quotes. (' -> 2 bytes)
+    elif d == ord('"'):
+      buf = [ord('\\'), d] # Escape double quote " character with a backspace since optimizer may turn the string into being delimited with double quotes. (" -> 2 bytes)
+    elif d == ord('\r'):
+      buf = [ord('\\'), ord('r')] # Escape carriage return 0x0D as \r -> 2 bytes
+    elif d == ord('\n'):
+      buf = [ord('\\'), ord('n')] # Escape newline 0x0A as \n -> 2 bytes
+    elif d == ord('\\'):
+      buf = [ord('\\'), ord('\\')] # Escape backslash \ as \\ -> 2 bytes
+    else:
+      buf = chr(d).encode('utf-8') # Otherwise write the original value encoded in UTF-8 (1 or 2 bytes).
+    for b in buf: # Write the bytes to output buffer
+      out[i] = b
+      i += 1
+  return out[0:i].decode('utf-8') # Crop output buffer to the actual used size
+
+
 # Returns the subresource location for run-time access
 def get_subresource_location(path, mimetype='application/octet-stream'):
   if settings.SINGLE_FILE:
+    if settings.SINGLE_FILE_BINARY_ENCODE:
+      return binary_encode(utils.read_binary(path))
     return f'data:{mimetype};base64,{base64_encode(path)}'
   else:
     return os.path.basename(path)