wasm2js: more coercion optimization (#2109)

5 files changed, 209 insertions, 19 deletions

diff --git a/src/tools/wasm2js.cpp b/src/tools/wasm2js.cpp
index b3f8068e2..731efbee0 100644
--- a/src/tools/wasm2js.cpp
+++ b/src/tools/wasm2js.cpp
@@ -145,9 +145,20 @@ static void replaceInPlaceIfPossible(Ref target, Ref value) {
 static void optimizeJS(Ref ast) {
   // Helpers
 
-  auto isOrZero = [](Ref node) {
+  auto isBinary = [](Ref node, IString op) {
     return node->isArray() && !node->empty() && node[0] == BINARY &&
-           node[1] == OR && node[3]->isNumber() && node[3]->getNumber() == 0;
+           node[1] == op;
+  };
+
+  auto isConstantBinary = [&](Ref node, IString op, int num) {
+    return isBinary(node, op) && node[3]->isNumber() &&
+           node[3]->getNumber() == num;
+  };
+
+  auto isOrZero = [&](Ref node) { return isConstantBinary(node, OR, 0); };
+
+  auto isTrshiftZero = [&](Ref node) {
+    return isConstantBinary(node, TRSHIFT, 0);
   };
 
   auto isPlus = [](Ref node) {
@@ -174,11 +185,6 @@ static void optimizeJS(Ref ast) {
            node[1] == op;
   };
 
-  auto isConstantBitwise = [](Ref node, IString op, int num) {
-    return node->isArray() && !node->empty() && node[0] == BINARY &&
-           node[1] == op && node[3]->isNumber() && node[3]->getNumber() == num;
-  };
-
   auto isWhile = [](Ref node) {
     return node->isArray() && !node->empty() && node[0] == WHILE;
   };
@@ -243,11 +249,14 @@ static void optimizeJS(Ref ast) {
   };
 
   auto optimizeBoolean = [&](Ref node) {
-    // x ^ 1  =>  !x
-    if (isConstantBitwise(node, XOR, 1)) {
+    if (isConstantBinary(node, XOR, 1)) {
+      // x ^ 1  =>  !x
       node[0]->setString(UNARY_PREFIX);
       node[1]->setString(L_NOT);
       node[3]->setNull();
+    } else if (isOrZero(node) || isTrshiftZero(node)) {
+      // Just being different from 0 is enough, casts don't matter.
+      return node[2];
     }
     return node;
   };
@@ -257,7 +266,7 @@ static void optimizeJS(Ref ast) {
   // Pre-simplification
   traversePost(ast, [&](Ref node) {
     // x >> 0  =>  x | 0
-    if (isConstantBitwise(node, RSHIFT, 0)) {
+    if (isConstantBinary(node, RSHIFT, 0)) {
       node[1]->setString(OR);
     }
   });
@@ -270,10 +279,7 @@ static void optimizeJS(Ref ast) {
       // x | 0 | 0  =>  x | 0
       if (isOrZero(node)) {
         if (isBitwise(node[2])) {
-          auto child = node[2];
-          node[1] = child[1];
-          node[2] = child[2];
-          node[3] = child[3];
+          replaceInPlace(node, node[2]);
         }
       }
       if (isHeapAccess(node[2])) {
@@ -284,7 +290,7 @@ static void optimizeJS(Ref ast) {
           if (isIntegerHeap(heap)) {
             replacementHeap = heap;
           }
-        } else if (isConstantBitwise(node, TRSHIFT, 0)) {
+        } else if (isTrshiftZero(node)) {
           // For signed or unsigned loads smaller than 32 bits, doing an | 0
           // was safe either way - they aren't in the range an | 0 can affect.
           // For >>> 0 however, a negative value would change, so we still
@@ -308,7 +314,7 @@ static void optimizeJS(Ref ast) {
         // better performance (perhaps since HEAPU32 is at risk of not being a
         // smallint).
         if (node[1] == AND) {
-          if (isConstantBitwise(node, AND, 1)) {
+          if (isConstantBinary(node, AND, 1)) {
             if (heap == HEAPU8) {
               setHeapOnAccess(node[2], HEAP8);
             } else if (heap == HEAPU16) {
@@ -348,6 +354,21 @@ static void optimizeJS(Ref ast) {
     } else if (isUnary(node, L_NOT)) {
       node[2] = optimizeBoolean(node[2]);
     }
+    // Add/subtract can merge coercions up.
+    else if (isBinary(node, PLUS) || isBinary(node, MINUS)) {
+      auto left = node[2];
+      auto right = node[3];
+      if (isOrZero(left) && isOrZero(right)) {
+        auto op = node[1]->getIString();
+        // Add a coercion on top.
+        node[1]->setString(OR);
+        node[2] = left;
+        node[3] = ValueBuilder::makeNum(0);
+        // Add/subtract the inner uncoerced values.
+        left[1]->setString(op);
+        left[3] = right[2];
+      }
+    }
     // Assignment into a heap coerces.
     else if (node->isAssign()) {
       auto assign = node->asAssign();
@@ -357,12 +378,12 @@ static void optimizeJS(Ref ast) {
         if (isIntegerHeap(heap)) {
           if (heap == HEAP8 || heap == HEAPU8) {
             while (isOrZero(assign->value()) ||
-                   isConstantBitwise(assign->value(), AND, 255)) {
+                   isConstantBinary(assign->value(), AND, 255)) {
               assign->value() = assign->value()[2];
             }
           } else if (heap == HEAP16 || heap == HEAPU16) {
             while (isOrZero(assign->value()) ||
-                   isConstantBitwise(assign->value(), AND, 65535)) {
+                   isConstantBinary(assign->value(), AND, 65535)) {
               assign->value() = assign->value()[2];
             }
           } else {
diff --git a/test/wasm2js/br_table_temp.2asm.js.opt b/test/wasm2js/br_table_temp.2asm.js.opt
index 5620f829f..4633a99e3 100644
--- a/test/wasm2js/br_table_temp.2asm.js.opt
+++ b/test/wasm2js/br_table_temp.2asm.js.opt
@@ -12662,7 +12662,7 @@ function asmFunc(global, env, buffer) {
  
  function $63($0) {
   $0 = $0 | 0;
-  if ($0 - 1 | 0) {
+  if ($0 - 1) {
    $0 = 9
   } else {
    $0 = 8
diff --git a/test/wasm2js/emscripten.2asm.js b/test/wasm2js/emscripten.2asm.js
index 3cbb33a35..6304e9b9d 100644
--- a/test/wasm2js/emscripten.2asm.js
+++ b/test/wasm2js/emscripten.2asm.js
@@ -90,6 +90,48 @@ function asmFunc(global, env, buffer) {
   if ((HEAPU8[144 >> 0] | 0) >>> 0 < 3 >>> 0) {
    bar()
   }
+  if ((bools(314159 | 0) | 0) >>> 7 | 0) {
+   bar()
+  }
+  if ((bools(314159 | 0) | 0) >> 8 | 0) {
+   bar()
+  }
+  if (~~Math_fround(getf32()) >>> 0) {
+   bar()
+  }
+  if (~~Math_fround(getf32())) {
+   bar()
+  }
+  if (~~+getf64() >>> 0) {
+   bar()
+  }
+  if (~~+getf64()) {
+   bar()
+  }
+  if (((geti32() | 0) + (geti32() | 0) | 0) + (geti32() | 0) | 0) {
+   bar()
+  }
+  if ((geti32() | 0) + ((geti32() | 0) + (geti32() | 0) | 0) | 0) {
+   bar()
+  }
+  if (((geti32() | 0) + (geti32() | 0) | 0) + ((geti32() | 0) + (geti32() | 0) | 0) | 0) {
+   bar()
+  }
+  if ((((geti32() | 0) + (geti32() | 0) | 0) + ((geti32() | 0) + (geti32() | 0) | 0) | 0) + (((geti32() | 0) + (geti32() | 0) | 0) + ((geti32() | 0) + (geti32() | 0) | 0) | 0) | 0) {
+   bar()
+  }
+ }
+ 
+ function geti32() {
+  return geti32() | 0 | 0;
+ }
+ 
+ function getf32() {
+  return Math_fround(Math_fround(getf32()));
+ }
+ 
+ function getf64() {
+  return +(+getf64());
  }
  
  function __growWasmMemory($0) {
diff --git a/test/wasm2js/emscripten.2asm.js.opt b/test/wasm2js/emscripten.2asm.js.opt
index 682daaf85..69251d9bb 100644
--- a/test/wasm2js/emscripten.2asm.js.opt
+++ b/test/wasm2js/emscripten.2asm.js.opt
@@ -83,6 +83,48 @@ function asmFunc(global, env, buffer) {
   if (HEAPU8[144] < 3) {
    bar()
   }
+  if (bools(314159) >>> 7) {
+   bar()
+  }
+  if (bools(314159) >> 8) {
+   bar()
+  }
+  if (~~getf32()) {
+   bar()
+  }
+  if (~~getf32()) {
+   bar()
+  }
+  if (~~getf64()) {
+   bar()
+  }
+  if (~~getf64()) {
+   bar()
+  }
+  if ((geti32() + geti32() | 0) + geti32()) {
+   bar()
+  }
+  if (geti32() + (geti32() + geti32() | 0)) {
+   bar()
+  }
+  if (geti32() + geti32() + (geti32() + geti32())) {
+   bar()
+  }
+  if (geti32() + geti32() + (geti32() + geti32()) + (geti32() + geti32() + (geti32() + geti32()))) {
+   bar()
+  }
+ }
+ 
+ function geti32() {
+  return geti32();
+ }
+ 
+ function getf32() {
+  return getf32();
+ }
+ 
+ function getf64() {
+  return getf64();
  }
  
  function __growWasmMemory($0) {
diff --git a/test/wasm2js/emscripten.wast b/test/wasm2js/emscripten.wast
index f6ad7aae7..5b1b81507 100644
--- a/test/wasm2js/emscripten.wast
+++ b/test/wasm2js/emscripten.wast
@@ -110,6 +110,91 @@
   (if (i32.lt_u (i32.load8_u (i32.const 144)) (i32.const 3))
    (call  $bar)
   )
+  (if (i32.shr_u (call $bools (i32.const 314159)) (i32.const 7))
+   (call  $bar)
+  )
+  (if (i32.shr_s (call $bools (i32.const 314159)) (i32.const 8))
+   (call  $bar)
+  )
+  (if (i32.trunc_f32_u (call $getf32))
+   (call  $bar)
+  )
+  (if (i32.trunc_f32_s (call $getf32))
+   (call  $bar)
+  )
+  (if (i32.trunc_f64_u (call $getf64))
+   (call  $bar)
+  )
+  (if (i32.trunc_f64_s (call $getf64))
+   (call  $bar)
+  )
+  (if
+    (i32.add
+      (i32.add
+        (call $geti32)
+        (call $geti32)
+      )
+      (call $geti32)
+    )
+    (call $bar)
+  )
+  (if
+    (i32.add
+      (call $geti32)
+      (i32.add
+        (call $geti32)
+        (call $geti32)
+      )
+    )
+    (call $bar)
+  )
+  (if
+    (i32.add
+      (i32.add
+        (call $geti32)
+        (call $geti32)
+      )
+      (i32.add
+        (call $geti32)
+        (call $geti32)
+      )
+    )
+    (call $bar)
+  )
+  (if
+    (i32.add
+      (i32.add
+        (i32.add
+          (call $geti32)
+          (call $geti32)
+        )
+        (i32.add
+          (call $geti32)
+          (call $geti32)
+        )
+      )
+      (i32.add
+        (i32.add
+          (call $geti32)
+          (call $geti32)
+        )
+        (i32.add
+          (call $geti32)
+          (call $geti32)
+        )
+      )
+    )
+    (call $bar)
+  )
+ )
+ (func $geti32 (result i32)
+  (call $geti32)
+ )
+ (func $getf32 (result f32)
+  (call $getf32)
+ )
+ (func $getf64 (result f64)
+  (call $getf64)
  )
  (func $__growWasmMemory (param $0 i32) (result i32)
   (grow_memory