diff --git a/src/settings_internal.js b/src/settings_internal.js
index ae89b4b194730..d475180b40954 100644
--- a/src/settings_internal.js
+++ b/src/settings_internal.js
@@ -260,6 +260,11 @@ var ASYNCIFY_IMPORTS_EXCEPT_JS_LIBS = [];
 
 var WARN_DEPRECATED = true;
 
+// Enable fast math optimizations in wasm-opt when -ffast-math is passed.
+// This enables aggressive floating-point optimizations that may violate
+// IEEE 754 semantics but can improve performance.
+var FAST_MATH = 0;
+
 // WebGL 2 provides new garbage-free entry points to call to WebGL. Use
 // those always when possible.
 // We currently set this to false for certain browser when large memory sizes
diff --git a/test/test_other.py b/test/test_other.py
index ddcd572e56a23..1d287fadad283 100644
--- a/test/test_other.py
+++ b/test/test_other.py
@@ -15202,6 +15202,14 @@ def has_defined_function(file, func):
     self.assertIn('foo.cpp', out)
     self.assertIn('/emsdk/emscripten/system/lib/libc/musl/src/string/strcmp.c', out)
 
+  def test_binaryen_fast_math(self):
+    # Use a simple input; contents don't matter for -v flag inspection
+    err = self.run_process([EMCC, test_file('hello_world.c'), '-v', '-O2', '-ffast-math'], stderr=PIPE).stderr
+    self.assertContained('--fast-math', err)
+
+    err_no_fast = self.run_process([EMCC, test_file('hello_world.c'), '-v', '-O2'], stderr=PIPE).stderr
+    self.assertNotContained('--fast-math', err_no_fast)
+
   def test_relocatable(self):
     # This setting is due for removal:
     # https://github.com/emscripten-core/emscripten/issues/25262
diff --git a/tools/cmdline.py b/tools/cmdline.py
index 0afc105aa24e7..6fa339962d7ab 100644
--- a/tools/cmdline.py
+++ b/tools/cmdline.py
@@ -301,9 +301,8 @@ def consume_arg_file():
         settings.SHRINK_LEVEL = 0
         settings.DEBUG_LEVEL = max(settings.DEBUG_LEVEL, 1)
       elif requested_level == 'fast':
-        # TODO(https://github.com/emscripten-core/emscripten/issues/21497):
-        # If we ever map `-ffast-math` to `wasm-opt --fast-math` then
-        # then we should enable that too here.
+        # -Ofast typically includes -ffast-math semantics
+        settings.FAST_MATH = 1
         requested_level = 3
         settings.SHRINK_LEVEL = 0
       else:
@@ -552,6 +551,8 @@ def consume_arg_file():
       settings.WASM_EXCEPTIONS = 1
     elif arg == '-fignore-exceptions':
       settings.DISABLE_EXCEPTION_CATCHING = 1
+    elif arg == '-ffast-math':
+      settings.FAST_MATH = 1
     elif check_arg('--default-obj-ext'):
       exit_with_error('--default-obj-ext is no longer supported by emcc')
     elif arg.startswith('-fsanitize=cfi'):
diff --git a/tools/link.py b/tools/link.py
index 040764ccc5e72..0dc696e988ec2 100644
--- a/tools/link.py
+++ b/tools/link.py
@@ -392,6 +392,8 @@ def get_binaryen_passes():
       passes += ['--pass-arg=post-emscripten-side-module']
   if optimizing:
     passes += [building.opt_level_to_str(settings.OPT_LEVEL, settings.SHRINK_LEVEL)]
+  if settings.FAST_MATH:
+    passes += ['--fast-math']
   # when optimizing, use the fact that low memory is never used (1024 is a
   # hardcoded value in the binaryen pass). we also cannot do it when the stack
   # is first, as then the stack is in the low memory that should be unused.