Run RemoveUnneededModuleElements early (#6620)

Doing it before anything else can help a lot if there is a significant amount of
dead code that can be removed, as it saves work for all the later passes. We
did run this pass if GC was enabled just a few passes later down, but even so
it is worthwhile to run it an additional time, and it makes sense to do even
without GC (though in typical optimized LLVM outputs there will be little
dead code).

If there is no dead code then this is wasted work, but this is a fairly fast pass,
and I measure no significant slowdown due to this. E.g. on the 35 MB clang.wasm
(which is already optimized, so little dead code) it takes around a second, while
all of -O2 takes almost two minutes, so the difference is just 1%.

On J2CL I measure a 15% speedup in -O3 --closed-world -tnh, and also the
binary is 2.5% smaller, which means there is less work for later cycles of -O3.

1 files changed, 6 insertions, 0 deletions

diff --git a/src/passes/pass.cpp b/src/passes/pass.cpp
index 8f954e5e3..ad0cc448c 100644
--- a/src/passes/pass.cpp
+++ b/src/passes/pass.cpp
@@ -669,7 +669,13 @@ void PassRunner::addDefaultFunctionOptimizationPasses() {
 }
 
 void PassRunner::addDefaultGlobalOptimizationPrePasses() {
+  // Removing duplicate functions is fast and saves work later.
   addIfNoDWARFIssues("duplicate-function-elimination");
+  // Do a global cleanup before anything heavy, as it is fairly fast and can
+  // save a lot of work if there is a significant amount of dead code.
+  if (options.optimizeLevel >= 2) {
+    addIfNoDWARFIssues("remove-unused-module-elements");
+  }
   addIfNoDWARFIssues("memory-packing");
   if (options.optimizeLevel >= 2) {
     addIfNoDWARFIssues("once-reduction");