Found the following inefficiencies in fjbg:

  ldc being used instead of bipush or sipush. The cheaper instructions
  were only being used when the argument was typed as byte or short, but
  an Int in the byte or short range can as easily use them. This also
  saves an entry in the constant pool.

  iconst_n not being used except on Ints. Same issue in reverse: pushing
  short 3 on the stack is cheaper with iconst_3 than with sipush.

Example:

  class A {
    def f1: Int = 100   // was ldc #X, now bipush 100
    def f2: Int = 5000  // was ldc #X, now sipush 5000
    def f3: Byte = 2    // was bipush 2, now iconst_2
    def f4: Short = 2   // was sipush 2, now iconst_2
  }

Review by dragos.

1 files changed, 1 insertions, 1 deletions

diff --git a/lib/fjbg.jar.desired.sha1 b/lib/fjbg.jar.desired.sha1
index 7e1cb1c13e..862868d030 100644
--- a/lib/fjbg.jar.desired.sha1
+++ b/lib/fjbg.jar.desired.sha1
@@ -1 +1 @@
-b73c0c9115d83849bada9a5c361f77745da22e0d ?fjbg.jar
+3997a32211461f0f7de1e4eb37e964cd134ea003 ?fjbg.jar