From f6136891cc7b775cb2bbfe27acdd8108c9e076cc Mon Sep 17 00:00:00 2001
From: Irmen de Jong <irmen@razorvine.net>
Date: Sat, 13 Feb 2021 13:27:06 +0100
Subject: [PATCH] optimized for loop over const bytes, fixed downto 1

---
 .../target/c64/codegen/ForLoopsAsmGen.kt      | 38 ++++++++-----------
 docs/source/todo.rst                          |  1 -
 examples/test.p8                              | 35 +++++++++++------
 3 files changed, 40 insertions(+), 34 deletions(-)

diff --git a/compiler/src/prog8/compiler/target/c64/codegen/ForLoopsAsmGen.kt b/compiler/src/prog8/compiler/target/c64/codegen/ForLoopsAsmGen.kt
index f32429a04..7cfc8f506 100644
--- a/compiler/src/prog8/compiler/target/c64/codegen/ForLoopsAsmGen.kt
+++ b/compiler/src/prog8/compiler/target/c64/codegen/ForLoopsAsmGen.kt
@@ -369,7 +369,7 @@ $loopLabel""")
                         throw AssemblyError("step 0, 1 and -1 should have been handled specifically  $stmt")
                     }
                     2 -> {
-                        if(range.last==255) {
+                        if(range.last==255 || range.last==254) {
                             asmgen.out("""
                                 inc  $varname
                                 beq  $endLabel
@@ -377,12 +377,11 @@ $loopLabel""")
                                 bne  $loopLabel""")
                         } else {
                             asmgen.out("""
+                                inc  $varname
+                                inc  $varname
                                 lda  $varname
-                                cmp  #${range.last}
-                                beq  $endLabel
-                                inc  $varname
-                                inc  $varname
-                                jmp  $loopLabel""")
+                                cmp  #${range.last+2}
+                                bne  $loopLabel""")
                         }
                     }
                     -2 -> {
@@ -399,12 +398,11 @@ $loopLabel""")
                                 dec  $varname
                                 bne  $loopLabel""")
                             else -> asmgen.out("""
+                                dec  $varname
+                                dec  $varname
                                 lda  $varname
-                                cmp  #${range.last}
-                                beq  $endLabel
-                                dec  $varname
-                                dec  $varname
-                                jmp  $loopLabel""")
+                                cmp  #${range.last-2}
+                                bne  $loopLabel""")
                         }
                     }
                     else -> {
@@ -480,11 +478,10 @@ $loopLabel""")
 $endLabel""")
         } else {
             asmgen.out("""
-                lda  $varname
-                cmp  #${range.last}
-                beq  $endLabel
                 inc  $varname
-                jmp  $loopLabel
+                lda  $varname
+                cmp  #${range.last+1}
+                bne  $loopLabel
 $endLabel""")
         }
         asmgen.loopEndLabels.pop()
@@ -512,16 +509,15 @@ $endLabel""")
             1 -> {
                 asmgen.out("""
                     dec  $varname
-                    jmp  $loopLabel
+                    bne  $loopLabel
 $endLabel""")
             }
             else -> {
                 asmgen.out("""
-                    lda  $varname
-                    cmp  #${range.last}
-                    beq  $endLabel
                     dec  $varname
-                    jmp  $loopLabel
+                    lda  $varname
+                    cmp  #${range.last-1}
+                    bne  $loopLabel
 $endLabel""")
             }
         }
@@ -546,7 +542,6 @@ $loopLabel""")
             bne  +
             lda  $varname+1
             cmp  #>${range.last}
-            bne  +
             beq  $endLabel
 +           inc  $varname
             bne  $loopLabel
@@ -574,7 +569,6 @@ $loopLabel""")
             bne  +
             lda  $varname+1
             cmp  #>${range.last}
-            bne  +
             beq  $endLabel
 +           lda  $varname
             bne  +
diff --git a/docs/source/todo.rst b/docs/source/todo.rst
index b7a03c2de..9929b0c63 100644
--- a/docs/source/todo.rst
+++ b/docs/source/todo.rst
@@ -2,7 +2,6 @@
 TODO
 ====
 
-- optimize for loop iterations better to allow proper inx, cpx #value, bne loop  instructions  (like repeat loop)
 - implement the linked_list millfork benchmark
 
 - refactor the asmgen into their own submodule?
diff --git a/examples/test.p8 b/examples/test.p8
index d07ae9c86..39c0ba4aa 100644
--- a/examples/test.p8
+++ b/examples/test.p8
@@ -4,16 +4,29 @@
 main {
 
     sub start() {
-        txt.print("hello\n")
-        txt.column(3)
-        txt.print("hello2\n")
-        txt.column(8)
-        txt.print("hello3\n")
-        txt.column(34)
-        txt.print("hello4\n")
-        txt.column(1)
-        txt.print("hello5\n")
-        txt.column(0)
-        txt.print("hello6\n")
+        ubyte ib
+        uword iw
+        ubyte xx
+
+        xx=0
+
+        for ib in 241 to 253 step 2 {
+            txt.print_ub(ib)
+            txt.nl()
+            xx++
+        }
+
+        for ib in 10 downto 2 step -2 {
+            txt.print_ub(ib)
+            txt.nl()
+            xx--
+        }
+        for ib in 6 downto 0 step -2 {
+            txt.print_ub(ib)
+            txt.nl()
+            xx--
+        }
+
+        txt.print_ub(xx)
     }
 }