Use Binaryen StackIR at the end of the pipeline

test/code_size/embind_val_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 380,
   "a.js": 7072,
   "a.js.gz": 2997,
-  "a.wasm": 11625,
-  "a.wasm.gz": 5814,
-  "total": 19249,
-  "total_gz": 9191
+  "a.wasm": 11588,
+  "a.wasm.gz": 5778,
+  "total": 19212,
+  "total_gz": 9155
 }

test/code_size/hello_webgl2_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 328,
   "a.js": 4521,
   "a.js.gz": 2312,
-  "a.wasm": 10427,
-  "a.wasm.gz": 6719,
-  "total": 15402,
-  "total_gz": 9359
+  "a.wasm": 10399,
+  "a.wasm.gz": 6695,
+  "total": 15374,
+  "total_gz": 9335
 }

test/code_size/hello_webgl2_wasm2js.json

@@ -1,8 +1,8 @@
 {
   "a.html": 346,
   "a.html.gz": 262,
-  "a.js": 22191,
-  "a.js.gz": 11586,
-  "total": 22537,
-  "total_gz": 11848
+  "a.js": 22193,
+  "a.js.gz": 11584,
+  "total": 22539,
+  "total_gz": 11846
 }

test/code_size/hello_webgl_wasm.json

@@ -3,8 +3,8 @@
   "a.html.gz": 328,
   "a.js": 4059,
   "a.js.gz": 2153,
-  "a.wasm": 10427,
-  "a.wasm.gz": 6719,
-  "total": 14940,
-  "total_gz": 9200
+  "a.wasm": 10399,
+  "a.wasm.gz": 6695,
+  "total": 14912,
+  "total_gz": 9176
 }

test/code_size/hello_webgl_wasm2js.json

@@ -1,8 +1,8 @@
 {
   "a.html": 346,
   "a.html.gz": 262,
-  "a.js": 21717,
-  "a.js.gz": 11425,
-  "total": 22063,
-  "total_gz": 11687
+  "a.js": 21719,
+  "a.js.gz": 11420,
+  "total": 22065,
+  "total_gz": 11682
 }

test/code_size/random_printf_wasm.json

@@ -1,6 +1,6 @@
 {
-  "a.html": 12746,
-  "a.html.gz": 6918,
-  "total": 12746,
-  "total_gz": 6918
+  "a.html": 12694,
+  "a.html.gz": 6891,
+  "total": 12694,
+  "total_gz": 6891
 }

test/other/metadce/test_metadce_files_wasmfs.size

@@ -1 +1 @@
-51018
+50957

test/other/metadce/test_metadce_hello_O3.size

@@ -1 +1 @@
-1741
+1737

test/other/metadce/test_metadce_hello_Os.size

@@ -1 +1 @@
-1732
+1728

test/other/metadce/test_metadce_hello_Oz.size

@@ -1 +1 @@
-1290
+1286

test/other/metadce/test_metadce_hello_dylink.size

@@ -1 +1 @@
-9355
+9347

test/other/metadce/test_metadce_hello_wasmfs.size

@@ -1 +1 @@
-1741
+1737

test/other/metadce/test_metadce_mem_O3.size

@@ -1 +1 @@
-5265
+5261

test/other/metadce/test_metadce_mem_O3_grow.size

@@ -1 +1 @@
-5266
+5262

test/other/metadce/test_metadce_mem_O3_grow_standalone.size

@@ -1 +1 @@
-5589
+5585

test/other/metadce/test_metadce_mem_O3_standalone.size

@@ -1 +1 @@
-5514
+5510

test/other/metadce/test_metadce_mem_O3_standalone_lib.size

@@ -1 +1 @@
-5277
+5273

test/other/metadce/test_metadce_mem_O3_standalone_narg.size

@@ -1 +1 @@
-5307
+5303

test/other/metadce/test_metadce_mem_O3_standalone_narg_flto.size

@@ -1 +1 @@
-4124
+4120

test/other/metadce/test_metadce_minimal_pthreads.size

@@ -1 +1 @@
-19481
+19421

test/test_other.py

@@ -8038,6 +8038,26 @@ def test_eval_ctors_debug_output(self):
     self.assertNotContained('ctor_evaller: not successful', err)
     self.assertContained('could not eval: call import: env.external_thing', err)
 
+  @uses_canonical_tmp
+  @with_env_modify({'EMCC_DEBUG': '1'})
+  @parameterized({
+    # StackIR optimizations should happen once at the end, and not multiple
+    # times in the middle. In -O2 we simply do them in the main wasm-opt
+    # invocation, while in -O3 we run wasm-opt later, and so we should disable
+    # StackIR first and enable it at the end.
+    'O2': (['-O2'], False),
+    'O3': (['-O3'], True),
+  })
+  def test_binaryen_stack_ir(self, opts, disable_and_enable):
+    err = self.run_process([EMXX, test_file('hello_world.c')] + opts, stderr=PIPE).stderr
+    DISABLE = '--no-stack-ir'
+    ENABLE = '--optimize-stack-ir'
+    self.assertContainedIf(DISABLE, err, disable_and_enable)
+    self.assertContainedIf(ENABLE, err, disable_and_enable)
+    # they should also appear at most once each
+    self.assertLess(err.count(DISABLE), 2)
+    self.assertLess(err.count(ENABLE), 2)
+
   def test_override_js_execution_environment(self):
     create_file('main.c', r'''
       #include <emscripten.h>

tools/building.py

@@ -719,7 +719,10 @@ def minify_wasm_js(js_file, wasm_file, expensive_optimizations, debug_info):
     # if we are optimizing for size, shrink the combined wasm+JS
     # TODO: support this when a symbol map is used
     if expensive_optimizations:
-      js_file = metadce(js_file, wasm_file, debug_info=debug_info)
+      js_file = metadce(js_file,
+                        wasm_file,
+                        debug_info=debug_info,
+                        last=not settings.MINIFY_WASM_IMPORTS_AND_EXPORTS)
       # now that we removed unneeded communication between js and wasm, we can clean up
       # the js some more.
       passes = ['AJSDCE']
@@ -743,8 +746,16 @@ def is_internal_global(name):
   return name in internal_start_stop_symbols or any(name.startswith(p) for p in internal_prefixes)
 
 
+# get the flags to pass into the very last binaryen tool invocation, that runs
+# the final set of optimizations
+def get_last_binaryen_opts():
+  return [f'--optimize-level={settings.OPT_LEVEL}',
+          f'--shrink-level={settings.SHRINK_LEVEL}',
+          '--optimize-stack-ir']
+
+
 # run binaryen's wasm-metadce to dce both js and wasm
-def metadce(js_file, wasm_file, debug_info):
+def metadce(js_file, wasm_file, debug_info, last):
   logger.debug('running meta-DCE')
   temp_files = shared.get_temp_files()
   # first, get the JS part of the graph
@@ -810,10 +821,13 @@ def metadce(js_file, wasm_file, debug_info):
   temp = temp_files.get('.json', prefix='emcc_dce_graph_').name
   utils.write_file(temp, json.dumps(graph, indent=2))
   # run wasm-metadce
+  args = ['--graph-file=' + temp]
+  if last:
+    args += get_last_binaryen_opts()
   out = run_binaryen_command('wasm-metadce',
                              wasm_file,
                              wasm_file,
-                             ['--graph-file=' + temp],
+                             args,
                              debug=debug_info,
                              stdout=PIPE)
   # find the unused things in js
@@ -878,23 +892,19 @@ def asyncify_lazy_load_code(wasm_target, debug):
 def minify_wasm_imports_and_exports(js_file, wasm_file, minify_exports, debug_info):
   logger.debug('minifying wasm imports and exports')
   # run the pass
+  args = []
   if minify_exports:
     # standalone wasm mode means we need to emit a wasi import module.
     # otherwise, minify even the imported module names.
     if settings.MINIFY_WASM_IMPORTED_MODULES:
-      pass_name = '--minify-imports-and-exports-and-modules'
+      args.append('--minify-imports-and-exports-and-modules')
     else:
-      pass_name = '--minify-imports-and-exports'
+      args.append('--minify-imports-and-exports')
   else:
-    pass_name = '--minify-imports'
-  out = run_wasm_opt(wasm_file, wasm_file,
-                     [pass_name],
-                     debug=debug_info,
-                     stdout=PIPE)
-  # TODO this is the last tool we run, after normal opts and metadce. it
-  # might make sense to run Stack IR optimizations here or even -O (as
-  # metadce which runs before us might open up new general optimization
-  # opportunities). however, the benefit is less than 0.5%.
+    args.append('--minify-imports')
+  # this is always the last tool we run (if we run it)
+  args += get_last_binaryen_opts()
+  out = run_wasm_opt(wasm_file, wasm_file, args, debug=debug_info, stdout=PIPE)
 
   # get the mapping
   SEP = ' => '
@@ -1057,15 +1067,18 @@ def handle_final_wasm_symbols(wasm_file, symbols_file, debug_info):
   args = []
   if symbols_file:
     args += ['--print-function-map']
-  if not debug_info:
-    # to remove debug info, we just write to that same file, and without -g
-    args += ['-o', wasm_file]
   else:
     # suppress the wasm-opt warning regarding "no output file specified"
     args += ['--quiet']
   output = run_wasm_opt(wasm_file, args=args, stdout=PIPE)
   if symbols_file:
     utils.write_file(symbols_file, output)
+  if not debug_info:
+    # strip the names section using llvm-objcopy. this is slightly slower than
+    # using wasm-opt (we could run wasm-opt without -g here and just tell it to
+    # write the file back out), but running wasm-opt would undo StackIR
+    # optimizations, if we did those.
+    strip(wasm_file, wasm_file, sections=['name'])
 
 
 def is_ar(filename):

tools/link.py

@@ -424,6 +424,17 @@ def check_human_readable_list(items):
   if settings.WASM_EXNREF:
     passes += ['--emit-exnref']
 
+  # If we are going to run metadce then that means we will be running binaryen
+  # tools after the main invocation, whose flags are determined here
+  # (specifically we will run metadce and possibly also wasm-opt for import/
+  # export minification). And when we run such a tool it will "undo" any
+  # StackIR optimizations (since the conversion to BinaryenIR undoes them as it
+  # restructures the code). We could re-run those opts, but it is most efficient
+  # to just not do them now if we'll invoke other tools later, and we'll do them
+  # only in the very last invocation.
+  if will_metadce():
+    passes += ['--no-stack-ir']
+
   return passes