vacuum drops better

7 files changed, 138 insertions, 122 deletions

diff --git a/src/passes/Vacuum.cpp b/src/passes/Vacuum.cpp
index db42a994a..53b9eb397 100644
--- a/src/passes/Vacuum.cpp
+++ b/src/passes/Vacuum.cpp
@@ -47,26 +47,31 @@ struct Vacuum : public WalkerPass<ExpressionStackWalker<Vacuum, Visitor<Vacuum>>
         case Expression::Id::CallImportId:
         case Expression::Id::CallIndirectId:
         case Expression::Id::SetLocalId:
-        case Expression::Id::LoadId:
         case Expression::Id::StoreId:
         case Expression::Id::ReturnId:
-        case Expression::Id::GetGlobalId:
         case Expression::Id::SetGlobalId:
         case Expression::Id::HostId:
         case Expression::Id::UnreachableId: return curr; // always needed
 
+        case Expression::Id::LoadId: {
+          if (!resultUsed) {
+            return curr->cast<Load>()->ptr;
+          }
+          return curr;
+        }
         case Expression::Id::ConstId:
         case Expression::Id::GetLocalId:
+        case Expression::Id::GetGlobalId: {
+          if (!resultUsed) return nullptr;
+          return curr;
+        }
+
         case Expression::Id::UnaryId:
         case Expression::Id::BinaryId:
         case Expression::Id::SelectId: {
           if (resultUsed) {
             return curr; // used, keep it
           }
-          // result is not used, perhaps it is dead
-          if (curr->is<Const>() || curr->is<GetLocal>()) {
-            return nullptr;
-          }
           // for unary, binary, and select, we need to check their arguments for side effects
           if (auto* unary = curr->dynCast<Unary>()) {
             if (EffectAnalyzer(unary->value).hasSideEffects()) {
@@ -200,11 +205,19 @@ struct Vacuum : public WalkerPass<ExpressionStackWalker<Vacuum, Visitor<Vacuum>>
   }
 
   void visitDrop(Drop* curr) {
-    // if the drop input has no side effects, it can be wiped out
-    if (!EffectAnalyzer(curr->value).hasSideEffects()) {
+    // optimize the dropped value, maybe leaving nothing
+    curr->value = optimize(curr->value, false);
+    if (curr->value == nullptr) {
       ExpressionManipulator::nop(curr);
       return;
     }
+    // a drop of a tee is a set
+    if (auto* set = curr->value->dynCast<SetLocal>()) {
+      assert(set->isTee());
+      set->setTee(false);
+      replaceCurrent(set);
+      return;
+    }
     // sink a drop into an arm of an if-else if the other arm ends in an unreachable, as it if is a branch, this can make that branch optimizable and more vaccuming possible
     auto* iff = curr->value->dynCast<If>();
     if (iff && iff->ifFalse && isConcreteWasmType(iff->type)) {
diff --git a/test/emcc_hello_world.fromasm b/test/emcc_hello_world.fromasm
index 4c032627d..3b93c26fd 100644
--- a/test/emcc_hello_world.fromasm
+++ b/test/emcc_hello_world.fromasm
@@ -3325,21 +3325,17 @@
                       )
                       (i32.const 10)
                     )
-                    (drop
-                      (i32.load offset=4
-                        (tee_local $1
+                    (set_local $1
+                      (i32.add
+                        (get_local $3)
+                        (i32.shl
                           (i32.add
-                            (get_local $3)
-                            (i32.shl
-                              (i32.add
-                                (i32.load8_s
-                                  (get_local $6)
-                                )
-                                (i32.const -48)
-                              )
-                              (i32.const 3)
+                            (i32.load8_s
+                              (get_local $6)
                             )
+                            (i32.const -48)
                           )
+                          (i32.const 3)
                         )
                       )
                     )
@@ -3702,21 +3698,17 @@
                         )
                         (i32.const 10)
                       )
-                      (drop
-                        (i32.load offset=4
-                          (tee_local $1
+                      (set_local $1
+                        (i32.add
+                          (get_local $3)
+                          (i32.shl
                             (i32.add
-                              (get_local $3)
-                              (i32.shl
-                                (i32.add
-                                  (i32.load8_s
-                                    (get_local $6)
-                                  )
-                                  (i32.const -48)
-                                )
-                                (i32.const 3)
+                              (i32.load8_s
+                                (get_local $6)
                               )
+                              (i32.const -48)
                             )
+                            (i32.const 3)
                           )
                         )
                       )
@@ -4539,12 +4531,8 @@
                         )
                         (br $switch$24)
                       )
-                      (drop
-                        (i32.load offset=4
-                          (tee_local $1
-                            (get_local $18)
-                          )
-                        )
+                      (set_local $1
+                        (get_local $18)
                       )
                       (i32.store8
                         (get_local $71)
@@ -4606,12 +4594,8 @@
                   )
                   (br $switch$24)
                 )
-                (drop
-                  (i32.load offset=4
-                    (tee_local $1
-                      (get_local $18)
-                    )
-                  )
+                (set_local $1
+                  (get_local $18)
                 )
                 (i32.store
                   (get_local $72)
diff --git a/test/emcc_hello_world.fromasm.imprecise b/test/emcc_hello_world.fromasm.imprecise
index d5dd39c95..0d1653a62 100644
--- a/test/emcc_hello_world.fromasm.imprecise
+++ b/test/emcc_hello_world.fromasm.imprecise
@@ -3318,21 +3318,17 @@
                       )
                       (i32.const 10)
                     )
-                    (drop
-                      (i32.load offset=4
-                        (tee_local $1
+                    (set_local $1
+                      (i32.add
+                        (get_local $3)
+                        (i32.shl
                           (i32.add
-                            (get_local $3)
-                            (i32.shl
-                              (i32.add
-                                (i32.load8_s
-                                  (get_local $6)
-                                )
-                                (i32.const -48)
-                              )
-                              (i32.const 3)
+                            (i32.load8_s
+                              (get_local $6)
                             )
+                            (i32.const -48)
                           )
+                          (i32.const 3)
                         )
                       )
                     )
@@ -3695,21 +3691,17 @@
                         )
                         (i32.const 10)
                       )
-                      (drop
-                        (i32.load offset=4
-                          (tee_local $1
+                      (set_local $1
+                        (i32.add
+                          (get_local $3)
+                          (i32.shl
                             (i32.add
-                              (get_local $3)
-                              (i32.shl
-                                (i32.add
-                                  (i32.load8_s
-                                    (get_local $6)
-                                  )
-                                  (i32.const -48)
-                                )
-                                (i32.const 3)
+                              (i32.load8_s
+                                (get_local $6)
                               )
+                              (i32.const -48)
                             )
+                            (i32.const 3)
                           )
                         )
                       )
@@ -4532,12 +4524,8 @@
                         )
                         (br $switch$24)
                       )
-                      (drop
-                        (i32.load offset=4
-                          (tee_local $1
-                            (get_local $18)
-                          )
-                        )
+                      (set_local $1
+                        (get_local $18)
                       )
                       (i32.store8
                         (get_local $71)
@@ -4599,12 +4587,8 @@
                   )
                   (br $switch$24)
                 )
-                (drop
-                  (i32.load offset=4
-                    (tee_local $1
-                      (get_local $18)
-                    )
-                  )
+                (set_local $1
+                  (get_local $18)
                 )
                 (i32.store
                   (get_local $72)
diff --git a/test/passes/vacuum.txt b/test/passes/vacuum.txt
index 1adcc3e6f..f64db152f 100644
--- a/test/passes/vacuum.txt
+++ b/test/passes/vacuum.txt
@@ -42,16 +42,10 @@
   )
   (func $binary (type $2) (result f32)
     (drop
-      (f32.add
-        (unreachable)
-        (f32.const 3)
-      )
+      (unreachable)
     )
     (drop
-      (f32.add
-        (f32.const 4)
-        (unreachable)
-      )
+      (unreachable)
     )
     (f32.add
       (unreachable)
@@ -64,25 +58,13 @@
   )
   (func $select (type $3) (result i32)
     (drop
-      (select
-        (unreachable)
-        (i32.const 4)
-        (i32.const 5)
-      )
+      (unreachable)
     )
     (drop
-      (select
-        (i32.const 6)
-        (unreachable)
-        (i32.const 7)
-      )
+      (unreachable)
     )
     (drop
-      (select
-        (i32.const 8)
-        (i32.const 9)
-        (unreachable)
-      )
+      (unreachable)
     )
     (select
       (unreachable)
@@ -178,4 +160,21 @@
       )
     )
   )
+  (func $drop-silly (type $0)
+    (drop
+      (call_import $int)
+    )
+    (drop
+      (call_import $int)
+    )
+    (drop
+      (call_import $int)
+    )
+    (drop
+      (i32.add
+        (call_import $int)
+        (call_import $int)
+      )
+    )
+  )
 )
diff --git a/test/passes/vacuum.wast b/test/passes/vacuum.wast
index 3630512b4..9da5fe6bc 100644
--- a/test/passes/vacuum.wast
+++ b/test/passes/vacuum.wast
@@ -314,4 +314,44 @@
       )
     )
   )
+  (func $drop-silly
+    (drop
+      (i32.eqz
+        (i32.eqz
+          (i32.const 1)
+        )
+      )
+    )
+    (drop
+      (i32.eqz
+        (i32.eqz
+          (call_import $int)
+        )
+      )
+    )
+    (drop
+      (i32.add
+        (i32.const 2)
+        (i32.const 3)
+      )
+    )
+    (drop
+      (i32.add
+        (i32.const 4)
+        (call_import $int)
+      )
+    )
+    (drop
+      (i32.add
+        (call_import $int)
+        (i32.const 5)
+      )
+    )
+    (drop
+      (i32.add
+        (call_import $int)
+        (call_import $int)
+      )
+    )
+  )
 )
diff --git a/test/unit.fromasm b/test/unit.fromasm
index c6e11e0ef..42b67a72c 100644
--- a/test/unit.fromasm
+++ b/test/unit.fromasm
@@ -109,14 +109,10 @@
         (get_local $0)
       )
     )
-    (drop
-      (f64.convert_s/i32
-        (tee_local $2
-          (call_import $f64-to-int
-            (f64.promote/f32
-              (get_local $1)
-            )
-          )
+    (set_local $2
+      (call_import $f64-to-int
+        (f64.promote/f32
+          (get_local $1)
         )
       )
     )
@@ -538,8 +534,10 @@
     (call $phi)
   )
   (func $useSetGlobal (result i32)
-    (set_global $Int
-      (i32.const 10)
+    (drop
+      (set_global $Int
+        (i32.const 10)
+      )
     )
     (set_global $Int
       (i32.const 20)
diff --git a/test/unit.fromasm.imprecise b/test/unit.fromasm.imprecise
index 2715b551b..634356f9f 100644
--- a/test/unit.fromasm.imprecise
+++ b/test/unit.fromasm.imprecise
@@ -98,13 +98,9 @@
   (func $conversions
     (local $0 f32)
     (local $1 i32)
-    (drop
-      (f64.convert_s/i32
-        (tee_local $1
-          (i32.trunc_s/f32
-            (get_local $0)
-          )
-        )
+    (set_local $1
+      (i32.trunc_s/f32
+        (get_local $0)
       )
     )
   )
@@ -519,8 +515,10 @@
     (call $phi)
   )
   (func $useSetGlobal (result i32)
-    (set_global $Int
-      (i32.const 10)
+    (drop
+      (set_global $Int
+        (i32.const 10)
+      )
     )
     (set_global $Int
       (i32.const 20)