ARM: Use loop to initialize locals when optimizing for size.

src/arm/full-codegen-arm.cc

Index: src/arm/full-codegen-arm.cc
diff --git a/src/arm/full-codegen-arm.cc b/src/arm/full-codegen-arm.cc
index c57c78559803c81e51137a32876688dff0ceebc0..fbea793efd6fbd962027b418074c40f236fb2b68 100644
--- a/src/arm/full-codegen-arm.cc
+++ b/src/arm/full-codegen-arm.cc
@@ -168,9 +168,21 @@ void FullCodeGenerator::Generate() {
     // Generators allocate locals, if any, in context slots.
     ASSERT(!info->function()->is_generator() || locals_count == 0);
     if (locals_count > 0) {
-      __ LoadRoot(ip, Heap::kUndefinedValueRootIndex);
-      for (int i = 0; i < locals_count; i++) {
+      // Emit a loop to initialize stack cells for locals when optimizing for
+      // size. Otherwise, unroll the loop for maximum performance.
+      if (FLAG_optimize_for_size && locals_count > 4) {
+        Label loop;
+        __ mov(r2, Operand(locals_count));
+        __ LoadRoot(ip, Heap::kUndefinedValueRootIndex);

[vincent.belliard.fr, 2013/10/30 08:45:02]

The two LoadRoot can be factorized before the if.

[Benedikt Meurer, 2013/10/30 08:46:53]

Hm, no, I think the mov(r2, Operand(locals_count)) could clobber ip.

[vincent.belliard.fr, 2013/10/30 08:53:56]

We can use an other register to load undefined (for example r9).

+        __ bind(&loop);
+        __ sub(r2, r2, Operand(1), SetCC);
         __ push(ip);
+        __ b(&loop, ne);
+      } else {
+        __ LoadRoot(ip, Heap::kUndefinedValueRootIndex);
+        for (int i = 0; i < locals_count; i++) {
+          __ push(ip);
+        }
       }
     }
   }