wasm-emscripten-finalize: Do not skip the start function body (#3714)

When we can skip function bodies, we still need to parse the start function
for the pthreads case, see details in the comments. This still gives us 99%
of the speedup as the start function is just 1 function and it's not that big,
so with this we return to full speed after the reversion in #3705

2 files changed, 14 insertions, 4 deletions

diff --git a/src/tools/wasm-emscripten-finalize.cpp b/src/tools/wasm-emscripten-finalize.cpp
index b2f28dca0..38a55e5cb 100644
--- a/src/tools/wasm-emscripten-finalize.cpp
+++ b/src/tools/wasm-emscripten-finalize.cpp
@@ -220,9 +220,12 @@ int main(int argc, const char* argv[]) {
     // If we are not writing the output then all we are doing is simple parsing
     // of metadata from global parts of the wasm such as imports and exports. In
     // that case, it is unnecessary to parse function contents which are the
-    // great bulk of the work, and we can skip all that. However, the one
-    // exception is pthreads, which does require scanning the code, and so for
-    // now we cannot do reader.setSkipFunctionBodies(true); here yet.
+    // great bulk of the work, and we can skip all that.
+    // Note that the one case we do need function bodies for, pthreads + EM_ASM
+    // parsing, requires special handling. The start function has code that we
+    // parse in order to find the EM_ASMs, and for that reason the binary reader
+    // will still parse the start function even in this mode.
+    reader.setSkipFunctionBodies(true);
   }
   try {
     reader.read(infile, wasm, inputSourceMapFilename);
diff --git a/src/wasm/wasm-binary.cpp b/src/wasm/wasm-binary.cpp
index 37fa1ab78..6cb837296 100644
--- a/src/wasm/wasm-binary.cpp
+++ b/src/wasm/wasm-binary.cpp
@@ -2141,7 +2141,14 @@ void WasmBinaryBuilder::readFunctions() {
       assert(controlFlowStack.empty());
       assert(letStack.empty());
       assert(depth == 0);
-      if (!skipFunctionBodies) {
+      // Even if we are skipping function bodies we need to not skip the start
+      // function. That contains important code for wasm-emscripten-finalize in
+      // the form of pthread-related segment initializations. As this is just
+      // one function, it doesn't add significant time, so the optimization of
+      // skipping bodies is still very useful.
+      auto currFunctionIndex = functionImports.size() + functions.size();
+      bool isStart = startIndex == currFunctionIndex;
+      if (!skipFunctionBodies || isStart) {
         func->body = getBlockOrSingleton(func->sig.results);
       } else {
         // When skipping the function body we need to put something valid in