The backend never generates iinc

[scabug]

In the slightly shocking fact of the day dept. It is a decadent waste of bytecode to be generating four instructions including a load and a store when a single instruction can increment in-place.

A typical example - this is a single class:

git diff -w head^ head -- scala/collection/immutable/Range.class
 diff --git a/scala/collection/immutable/Range.class b/scala/collection/immutable/Range.class
 index 6fe5a67ab1..7bf47a582e 100644
 --- a/scala/collection/immutable/Range.class
 +++ b/scala/collection/immutable/Range.class
 @@ -371,10 +371,7 @@ public class scala.collection.immutable.Range extends scala.collection.AbstractS
         : invokestatic  #                 // Method scala/runtime/BoxesRunTime.boxToInteger:(I)Ljava/lang/Integer;
         : invokeinterface # ,             // InterfaceMethod scala/Function .apply:(Ljava/lang/Object;)Ljava/lang/Object;
         : pop
 -       : iload_2
 -       : iconst_1
 -       : iadd
 -       : istore_2
 +       : iinc           ,
         : iload_3
         : invokevirtual #                 // Method step:()I
         : iadd
 @@ -431,10 +428,7 @@ public class scala.collection.immutable.Range extends scala.collection.AbstractS
         : invokestatic  #                 // Method scala/runtime/BoxesRunTime.boxToInteger:(I)Ljava/lang/Integer;
         : invokeinterface # ,             // InterfaceMethod scala/Function .apply:(Ljava/lang/Object;)Ljava/lang/Object;
         : pop
 -       : iload
 -       : iconst_1
 -       : iadd
 -       : istore
 +       : iinc           ,
         : iload_3
         : iload
        : iadd
 @@ -529,10 +523,7 @@ public class scala.collection.immutable.Range extends scala.collection.AbstractS
         : iload_2
         : invokeinterface # ,             // InterfaceMethod scala/Function .apply$mcZI$sp:(I)Z
         : ifeq
 -       : iload_3
 -       : iconst_1
 -       : iadd
 -       : istore_3
 +       : iinc           ,
         : iload_2
         : invokevirtual #                 // Method step:()I
         : iadd
 @@ -814,10 +805,7 @@ public class scala.collection.immutable.Range extends scala.collection.AbstractS
         : aload_1
         : iload_3
         : invokeinterface # ,             // InterfaceMethod scala/Function .apply$mcVI$sp:(I)V
 -       : iload
 -       : iconst_1
 -       : iadd
 -       : istore
 +       : iinc           ,
         : iload_3
         : iload
         : iadd

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-7452?orig=1
Reporter: @paulp

[scabug]

@paulp said (edited on May 3, 2013 8:26:57 AM UTC):
Unfortunately this is one of those things which ought to be trivial but is made somewhat difficult by the implementation. I guess the easiest thing to do would be a peephole optimization pass right before genasm gets down to emitting bytecode. It will have to have (at least) a four instruction window, looking for sequences like this, where the locals have the same symbol and the second argument to add is constant and fits in a short.

LOAD_LOCAL(variable idx)
CONSTANT(1)
CALL_PRIMITIVE(Arithmetic(ADD,INT))
STORE_LOCAL(variable idx)

Given the local's index and the constant increment, asm has a method to do the rest:

jmethod.visitIincInsn(idx, inc)

[scabug]

@magarciaEPFL said:
Fair game for the new backend.

[scabug]

@magarciaEPFL said:
Additionally, with GenASM alone it's easier than pre-tweaking ICode: peephole right after genDefDef() has built a MethodNode.