Minor pass pipeline improvements (#1424)

Author: MaxGraey

src/module.ts

@@ -1506,7 +1506,14 @@ export class Module {
       // --- PassRunner::addDefaultFunctionOptimizationPasses ---
 
       if (optimizeLevel >= 3 || shrinkLevel >= 1) {
+        passes.push("simplify-locals-notee-nostructure");
+        passes.push("rse");
+        passes.push("vacuum");
         passes.push("ssa-nomerge");
+        passes.push("simplify-globals-optimizing");
+        passes.push("remove-unused-brs");
+        passes.push("remove-unused-names");
+        passes.push("merge-blocks");
       }
       if (optimizeLevel >= 3) {
         passes.push("flatten");
@@ -1518,10 +1525,6 @@ export class Module {
         passes.push("local-cse");
         passes.push("reorder-locals");
       }
-      if (optimizeLevel >= 2 || shrinkLevel >= 1) {
-        passes.push("rse");
-        passes.push("vacuum");
-      }
       // FIXME: see issue #1288
       // if (usesARC) {
       //   if (optimizeLevel < 3) {
@@ -1537,16 +1540,16 @@ export class Module {
       passes.push("remove-unused-brs");
       passes.push("remove-unused-names");
       passes.push("inlining-optimizing");
-      if (optimizeLevel >= 2 || shrinkLevel >= 1) {
-        passes.push("pick-load-signs");
-        passes.push("simplify-globals-optimizing");
-      }
       if (optimizeLevel >= 3 || shrinkLevel >= 2) {
         passes.push("precompute-propagate");
       } else {
         passes.push("precompute");
       }
       passes.push("vacuum");
+      if (optimizeLevel >= 2 || shrinkLevel >= 1) {
+        passes.push("pick-load-signs");
+        passes.push("simplify-globals-optimizing");
+      }
       if (optimizeLevel >= 3 && shrinkLevel <= 1) {
         passes.push("licm");
       }

tests/compiler/binary.optimized.wat

@@ -50,6 +50,10 @@
   local.get $0
   i32.reinterpret_f32
   local.tee $1
+  i32.const -2147483648
+  i32.and
+  local.set $4
+  local.get $1
   i32.const 23
   i32.shr_u
   i32.const 255
@@ -63,10 +67,6 @@
    f32.div
    return
   end
-  local.get $1
-  i32.const -2147483648
-  i32.and
-  local.set $4
   block $folding-inner0
    local.get $1
    i32.const 1
@@ -191,6 +191,10 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $1
+  i64.const 63
+  i64.shr_u
+  local.set $4
+  local.get $1
   i64.const 52
   i64.shr_u
   i64.const 2047
@@ -204,10 +208,6 @@
    f64.div
    return
   end
-  local.get $1
-  i64.const 63
-  i64.shr_u
-  local.set $4
   block $folding-inner0
    local.get $1
    i64.const 1

tests/compiler/comma.optimized.wat

@@ -80,17 +80,6 @@
   global.set $comma/b
   i32.const 2
   global.set $comma/a
-  global.get $comma/b
-  i32.const 2
-  i32.ne
-  if
-   i32.const 0
-   i32.const 1040
-   i32.const 19
-   i32.const 1
-   call $~lib/builtins/abort
-   unreachable
-  end
   loop $for-loop|0
    local.get $0
    global.get $comma/a