LocalCSE: Consider pass options, both size and cost (#1840)

With this we can optimize redundant global accesses fairly well (at least locally; licm also works), see #1831

Author: kripken

src/ir/cost.h

@@ -17,6 +17,9 @@
 #ifndef wasm_ir_cost_h
 #define wasm_ir_cost_h
 
+#include <wasm.h>
+#include <wasm-traversal.h>
+
 namespace wasm {
 
 // Measure the execution cost of an AST. Very handwave-ey

src/passes/LocalCSE.cpp

@@ -42,6 +42,7 @@
 #include <wasm-traversal.h>
 #include <pass.h>
 #include <ir/effects.h>
+#include <ir/cost.h>
 #include <ir/equivalent_sets.h>
 #include <ir/hashed.h>
 
@@ -208,8 +209,19 @@ struct LocalCSE : public WalkerPass<LinearExecutionWalker<LocalCSE>> {
     if (EffectAnalyzer(getPassOptions(), value).hasSideEffects()) {
       return false; // we can't combine things with side effects
     }
-    // check what we care about TODO: use optimize/shrink levels?
-    return Measurer::measure(value) > 1;
+    auto& options = getPassRunner()->options;
+    // If the size is at least 3, then if we have two of them we have 6,
+    // and so adding one set+two gets and removing one of the items itself
+    // is not detrimental, and may be beneficial.
+    if (options.shrinkLevel > 0 && Measurer::measure(value) >= 3) {
+      return true;
+    }
+    // If we focus on speed, any reduction in cost is beneficial, as the
+    // cost of a get is essentially free.
+    if (options.shrinkLevel == 0 && CostAnalyzer(value).cost > 0) {
+      return true;
+    }
+    return false;
   }
 };
 

test/example/cpp-unit.cpp

@@ -0,0 +1,18 @@
+// test multiple uses of the threadPool
+
+#include <assert.h>
+
+#include <wasm.h>
+#include <ir/cost.h>
+
+using namespace wasm;
+
+int main() 
+{
+  // Some optimizations assume that the cost of a get is zero, e.g. local-cse.
+  GetLocal get;
+  assert(CostAnalyzer(&get).cost == 0);
+
+  std::cout << "Success.\n";
+  return 0;
+}

test/example/cpp-unit.txt

@@ -0,0 +1 @@
+Success.

test/passes/licm.txt

@@ -4,6 +4,7 @@
  (type $2 (func (result i64)))
  (type $3 (func (param i32)))
  (type $4 (func (param i32) (result i32)))
+ (global $glob (mut i32) (i32.const 1))
  (func $loop1 (; 0 ;) (type $0)
   (drop
    (i32.const 10)
@@ -674,4 +675,20 @@
    )
   )
  )
+ (func $global (; 40 ;) (type $0)
+  (local $x i32)
+  (set_local $x
+   (get_global $glob)
+  )
+  (drop
+   (get_local $x)
+  )
+  (loop $loop
+   (nop)
+   (nop)
+   (br_if $loop
+    (get_local $x)
+   )
+  )
+ )
 )

test/passes/licm.wast

@@ -1,4 +1,5 @@
 (module
+  (global $glob (mut i32) (i32.const 1))
   (func $loop1
     (loop $loop
       (drop (i32.const 10))
@@ -388,5 +389,13 @@
       (br_if $loop (i32.const 1))
     )
   )
+  (func $global
+    (local $x i32)
+    (loop $loop
+      (set_local $x (get_global $glob))
+      (drop (get_local $x))
+      (br_if $loop (get_local $x))
+    )
+  )
 )
 

test/passes/local-cse.txt

@@ -1,5 +1,7 @@
 (module
  (type $0 (func (result i64)))
+ (type $1 (func))
+ (global $glob (mut i32) (i32.const 1))
  (func $i64-shifts (; 0 ;) (type $0) (result i64)
   (local $temp i64)
   (set_local $temp
@@ -19,4 +21,17 @@
   )
   (get_local $temp)
  )
+ (func $global (; 1 ;) (type $1)
+  (local $x i32)
+  (local $y i32)
+  (set_local $x
+   (get_global $glob)
+  )
+  (set_local $y
+   (get_local $x)
+  )
+  (set_local $y
+   (get_local $x)
+  )
+ )
 )

test/passes/local-cse.wast

@@ -1,4 +1,5 @@
 (module
+ (global $glob (mut i32) (i32.const 1))
  (func $i64-shifts (result i64)
   (local $temp i64)
   (set_local $temp
@@ -18,4 +19,11 @@
   )
   (get_local $temp)
  )
+ (func $global
+  (local $x i32)
+  (local $y i32)
+  (set_local $x (get_global $glob))
+  (set_local $y (get_global $glob))
+  (set_local $y (get_global $glob))
+ )
 )