optimize out 0-x, a zero only used to negate an int, when possible (#1365)

kripken · web-flow · commit 14cb0c01ee22 · 2018-01-17T21:20:53.000-08:00
diff --git a/src/passes/OptimizeInstructions.cpp b/src/passes/OptimizeInstructions.cpp
@@ -506,7 +506,43 @@ struct OptimizeInstructions : public WalkerPass<PostWalker<OptimizeInstructions,
           }
         }
         // note that both left and right may be consts, but then we let precompute compute the constant result
-      } else if (binary->op == AddInt32 || binary->op == SubInt32) {
+      } else if (binary->op == AddInt32) {
+        // try to get rid of (0 - ..), that is, a zero only used to negate an
+        // int. an add of a subtract can be flipped in order to remove it:
+        //   (i32.add
+        //     (i32.sub
+        //       (i32.const 0)
+        //       X
+        //     )
+        //     Y
+        //   )
+        // =>
+        //   (i32.sub
+        //     Y
+        //     X
+        //   )
+        if (auto* sub = binary->left->dynCast<Binary>()) {
+          if (sub->op == SubInt32) {
+            if (auto* subZero = sub->left->dynCast<Const>()) {
+              if (subZero->value.geti32() == 0) {
+                sub->left = binary->right;
+                return sub;
+              }
+            }
+          }
+        }
+        if (auto* sub = binary->right->dynCast<Binary>()) {
+          if (sub->op == SubInt32) {
+            if (auto* subZero = sub->left->dynCast<Const>()) {
+              if (subZero->value.geti32() == 0) {
+                sub->left = binary->left;
+                return sub;
+              }
+            }
+          }
+        }
+        return optimizeAddedConstants(binary);
+      } else if (binary->op == SubInt32) {
         return optimizeAddedConstants(binary);
       }
       // a bunch of operations on a constant right side can be simplified
diff --git a/test/emcc_O2_hello_world.fromasm b/test/emcc_O2_hello_world.fromasm
@@ -5966,12 +5966,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $11)
-         )
+         (get_local $11)
         )
        )
        (get_local $14)
diff --git a/test/emcc_O2_hello_world.fromasm.clamp b/test/emcc_O2_hello_world.fromasm.clamp
@@ -5966,12 +5966,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $11)
-         )
+         (get_local $11)
         )
        )
        (get_local $14)
diff --git a/test/emcc_O2_hello_world.fromasm.imprecise b/test/emcc_O2_hello_world.fromasm.imprecise
@@ -5965,12 +5965,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $11)
-         )
+         (get_local $11)
         )
        )
        (get_local $14)
diff --git a/test/emcc_hello_world.fromasm b/test/emcc_hello_world.fromasm
@@ -13329,12 +13329,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $7)
-         )
+         (get_local $7)
         )
        )
        (get_local $11)
diff --git a/test/emcc_hello_world.fromasm.clamp b/test/emcc_hello_world.fromasm.clamp
@@ -13379,12 +13379,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $7)
-         )
+         (get_local $7)
         )
        )
        (get_local $11)
diff --git a/test/emcc_hello_world.fromasm.imprecise b/test/emcc_hello_world.fromasm.imprecise
@@ -13263,14 +13263,11 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (tee_local $8
-           (i32.load
-            (get_local $1)
-           )
+         (tee_local $8
+          (i32.load
+           (get_local $1)
           )
          )
         )
diff --git a/test/memorygrowth.fromasm b/test/memorygrowth.fromasm
@@ -6017,12 +6017,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $10)
-         )
+         (get_local $10)
         )
        )
        (get_local $14)
diff --git a/test/memorygrowth.fromasm.clamp b/test/memorygrowth.fromasm.clamp
@@ -6017,12 +6017,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $10)
-         )
+         (get_local $10)
         )
        )
        (get_local $14)
diff --git a/test/memorygrowth.fromasm.imprecise b/test/memorygrowth.fromasm.imprecise
@@ -6015,12 +6015,9 @@
      (if
       (i32.lt_u
        (tee_local $1
-        (i32.add
+        (i32.sub
          (get_local $1)
-         (i32.sub
-          (i32.const 0)
-          (get_local $10)
-         )
+         (get_local $10)
         )
        )
        (get_local $14)
diff --git a/test/passes/optimize-instructions.txt b/test/passes/optimize-instructions.txt
@@ -2264,6 +2264,38 @@
    )
   )
  )
+ (func $subzero1 (; 56 ;) (type $3) (param $0 i32) (result i32)
+  (i32.sub
+   (i32.const 32)
+   (i32.clz
+    (get_local $0)
+   )
+  )
+ )
+ (func $subzero2 (; 57 ;) (type $3) (param $0 i32) (result i32)
+  (i32.sub
+   (i32.const 32)
+   (i32.clz
+    (get_local $0)
+   )
+  )
+ )
+ (func $subzero3 (; 58 ;) (type $6) (param $0 i32) (param $1 i32) (result i32)
+  (i32.sub
+   (get_local $1)
+   (i32.clz
+    (get_local $0)
+   )
+  )
+ )
+ (func $subzero4 (; 59 ;) (type $6) (param $0 i32) (param $1 i32) (result i32)
+  (i32.sub
+   (get_local $0)
+   (i32.clz
+    (get_local $1)
+   )
+  )
+ )
 )
 (module
  (type $0 (func))
diff --git a/test/passes/optimize-instructions.wast b/test/passes/optimize-instructions.wast
@@ -2663,6 +2663,50 @@
       (i32.and (i32.wrap/i64 (i64.const 1)) (i32.eqz (get_local $y)))
     )
   )
+  (func $subzero1 (param $0 i32) (result i32)
+    (i32.add
+     (i32.sub
+      (i32.const 1)
+      (i32.clz
+       (get_local $0)
+      )
+     )
+     (i32.const 31)
+    )
+  )
+  (func $subzero2 (param $0 i32) (result i32)
+    (i32.add
+     (i32.const 31)
+     (i32.sub
+      (i32.const 1)
+      (i32.clz
+       (get_local $0)
+      )
+     )
+    )
+  )
+  (func $subzero3 (param $0 i32) (param $1 i32) (result i32)
+    (i32.add
+     (i32.sub
+      (i32.const 0)
+      (i32.clz
+       (get_local $0)
+      )
+     )
+     (get_local $1)
+    )
+  )
+  (func $subzero4 (param $0 i32) (param $1 i32) (result i32)
+    (i32.add
+     (get_local $0)
+     (i32.sub
+      (i32.const 0)
+      (i32.clz
+       (get_local $1)
+      )
+     )
+    )
+  )
 )
 (module
   (import "env" "memory" (memory $0 (shared 256 256)))
@@ -2680,3 +2724,4 @@
    )
   )
 )
+

Original file line number	Diff line number	Diff line change
`@@ -13263,14 +13263,11 @@`
`13263`	`13263`	`(if`
`13264`	`13264`	`(i32.lt_u`
`13265`	`13265`	`(tee_local $1`
`13266`		`- (i32.add`
	`13266`	`+ (i32.sub`
`13267`	`13267`	`(get_local $1)`
`13268`		`- (i32.sub`
`13269`		`- (i32.const 0)`
`13270`		`- (tee_local $8`
`13271`		`- (i32.load`
`13272`		`- (get_local $1)`
`13273`		`- )`
	`13268`	`+ (tee_local $8`
	`13269`	`+ (i32.load`
	`13270`	`+ (get_local $1)`
`13274`	`13271`	`)`
`13275`	`13272`	`)`
`13276`	`13273`	`)`