Improved planned sequence for A->B->C sequences where B is conditional successor of A.

src/main/java/dk/camelot64/kickc/Compiler.java

@@ -377,7 +377,7 @@ public class Compiler {
       new Pass3AssertConstants(program).check();
       new Pass3AssertArrayLengths(program).check();
       new Pass3AssertNoMulDivMod(program).check();
-      new PassNBlockSequencePlanner(program).step();
+      //new PassNBlockSequencePlanner(program).step();
       // Phi lifting ensures that all variables in phi-blocks are in different live range equivalence classes
       new Pass3PhiLifting(program).perform();
       new PassNBlockSequencePlanner(program).step();

src/main/java/dk/camelot64/kickc/passes/PassNBlockSequencePlanner.java

@@ -39,11 +39,25 @@ public class PassNBlockSequencePlanner extends Pass2SsaOptimization {
          if(block.getCallSuccessor() != null) {
             pushTodo(getGraph().getCallSuccessor(block));
          }
-         if(block.getConditionalSuccessor() != null) {
+         ControlFlowBlock conditionalSuccessor = getGraph().getConditionalSuccessor(block);
+         ControlFlowBlock defaultSuccessor = getGraph().getDefaultSuccessor(block);
+         if(conditionalSuccessor != null && defaultSuccessor != null) {
+            // Both conditional and default successor
+            if(conditionalSuccessor.getDefaultSuccessor().equals(defaultSuccessor.getLabel())) {
+               // Best sequence is cond->default (resulting in better code locality)
+               pushTodo(defaultSuccessor);
+               pushTodo(getGraph().getConditionalSuccessor(block));
+            } else {
+               // Best sequence is default->cond
+               pushTodo(getGraph().getConditionalSuccessor(block));
+               pushTodo(defaultSuccessor);
+            }
+         } else if(conditionalSuccessor != null) {
+            // Only a conditional successor
             pushTodo(getGraph().getConditionalSuccessor(block));
-         }
-         if(getGraph().getDefaultSuccessor(block) != null) {
-            pushTodo(getGraph().getDefaultSuccessor(block));
+         } else if(defaultSuccessor != null) {
+            // Only a default successor
+            pushTodo(defaultSuccessor);
          }
 
       }

src/test/ref/bool-ifs.asm

@@ -17,6 +17,10 @@ bool_complex: {
     and #1
     cpx #$a
     bcc b6
+    jmp b5
+  b6:
+    cmp #0
+    beq b2
   b5:
     cpx #$a
     bcc b4
@@ -34,10 +38,6 @@ bool_complex: {
     lda #' '
     sta screen,x
     jmp b3
-  b6:
-    cmp #0
-    beq b2
-    jmp b5
 }
 bool_not: {
     .label screen = $450
@@ -91,6 +91,10 @@ bool_and: {
     and #1
     cpx #$a
     bcc b5
+    jmp b4
+  b5:
+    cmp #0
+    beq b2
   b4:
     lda #' '
     sta screen,x
@@ -99,10 +103,6 @@ bool_and: {
     cpx #$15
     bne b1
     rts
-  b5:
-    cmp #0
-    beq b2
-    jmp b4
   b2:
     lda #'*'
     sta screen,x

src/test/ref/bool-vars.asm

@@ -31,6 +31,11 @@ bool_complex: {
     lda o1
     cmp #0
     bne b6
+    jmp b5
+  b6:
+    lda o2
+    cmp #0
+    bne b2
   b5:
     lda o1
     cmp #0
@@ -50,11 +55,6 @@ bool_complex: {
     lda #' '
     sta screen,x
     jmp b3
-  b6:
-    lda o2
-    cmp #0
-    bne b2
-    jmp b5
 }
 bool_not: {
     .label screen = $450
@@ -108,6 +108,10 @@ bool_and: {
     and #1
     cpx #$a
     bcc b5
+    jmp b4
+  b5:
+    cmp #0
+    beq b2
   b4:
     lda #' '
     sta screen,x
@@ -116,10 +120,6 @@ bool_and: {
     cpx #$15
     bne b1
     rts
-  b5:
-    cmp #0
-    beq b2
-    jmp b4
   b2:
     lda #'*'
     sta screen,x

src/test/ref/complex-conditional-problem.asm

@@ -11,10 +11,10 @@ main: {
     bcs b3
     cmp #$40
     bcc b3
+    jmp b2
+  b3:
+    lda #0
   b2:
     sta SCREEN
     jmp b1
-  b3:
-    lda #0
-    jmp b2
 }

src/test/ref/complex/tetris/tetris.asm

@@ -145,18 +145,18 @@
   .label current_xpos_59 = $a
   .label current_piece_gfx_64 = 5
   .label current_piece_char_68 = $b
-  .label render_screen_render_69 = 9
-  .label current_xpos_130 = $a
-  .label current_xpos_131 = $a
-  .label current_piece_gfx_120 = 5
-  .label current_piece_gfx_121 = 5
-  .label current_piece_char_108 = $b
-  .label current_piece_char_109 = $b
-  .label current_piece_100 = 5
+  .label render_screen_render_68 = 9
+  .label current_xpos_127 = $a
+  .label current_xpos_128 = $a
+  .label current_piece_gfx_117 = 5
+  .label current_piece_gfx_118 = 5
+  .label current_piece_char_105 = $b
+  .label current_piece_char_106 = $b
   .label current_piece_101 = 5
   .label current_piece_102 = 5
   .label current_piece_103 = 5
   .label current_piece_104 = 5
+  .label current_piece_105 = 5
 bbegin:
   // The screen currently being showed to the user. $00 for screen 1 / $20 for screen 2.
   lda #0
@@ -192,13 +192,13 @@ main: {
     jsr render_playfield
     ldx current_ypos
     lda current_xpos
-    sta current_xpos_130
+    sta current_xpos_127
     lda current_piece_gfx
-    sta current_piece_gfx_120
+    sta current_piece_gfx_117
     lda current_piece_gfx+1
-    sta current_piece_gfx_120+1
+    sta current_piece_gfx_117+1
     lda current_piece_char
-    sta current_piece_char_108
+    sta current_piece_char_105
     lda #$20
     sta render_screen_render_33
     jsr render_moving
@@ -250,15 +250,15 @@ main: {
     jsr render_playfield
     ldx current_ypos
     lda render_screen_render
-    sta render_screen_render_69
+    sta render_screen_render_68
     lda current_xpos
-    sta current_xpos_131
+    sta current_xpos_128
     lda current_piece_gfx
-    sta current_piece_gfx_121
+    sta current_piece_gfx_118
     lda current_piece_gfx+1
-    sta current_piece_gfx_121+1
+    sta current_piece_gfx_118+1
     lda current_piece_char
-    sta current_piece_char_109
+    sta current_piece_char_106
     jsr render_moving
     lda render_screen_render
     ldx next_piece_idx
@@ -608,9 +608,9 @@ play_move_rotate: {
     sta play_collision.ypos
     ldx orientation
     lda current_piece
-    sta current_piece_103
+    sta current_piece_104
     lda current_piece+1
-    sta current_piece_103+1
+    sta current_piece_104+1
     jsr play_collision
     cmp #COLLISION_NONE
     bne b4
@@ -738,9 +738,9 @@ play_move_leftright: {
     sta play_collision.ypos
     ldx current_orientation
     lda current_piece
-    sta current_piece_102
+    sta current_piece_103
     lda current_piece+1
-    sta current_piece_102+1
+    sta current_piece_103+1
     jsr play_collision
     cmp #COLLISION_NONE
     bne b3
@@ -759,9 +759,9 @@ play_move_leftright: {
     sta play_collision.ypos
     ldx current_orientation
     lda current_piece
-    sta current_piece_101
+    sta current_piece_102
     lda current_piece+1
-    sta current_piece_101+1
+    sta current_piece_102+1
     jsr play_collision
     cmp #COLLISION_NONE
     bne b3
@@ -804,9 +804,9 @@ play_move_down: {
     sta play_collision.xpos
     ldx current_orientation
     lda current_piece
-    sta current_piece_100
+    sta current_piece_101
     lda current_piece+1
-    sta current_piece_100+1
+    sta current_piece_101+1
     jsr play_collision
     cmp #COLLISION_NONE
     beq b10
@@ -858,9 +858,9 @@ play_spawn_current: {
     lda current_ypos
     sta play_collision.ypos
     lda current_piece_gfx
-    sta current_piece_104
+    sta current_piece_105
     lda current_piece_gfx+1
-    sta current_piece_104+1
+    sta current_piece_105+1
     ldx #0
     jsr play_collision
     cmp #COLLISION_PLAYFIELD

src/test/ref/examples/multiplexer/simple-multiplexer.asm

@@ -142,10 +142,11 @@ plexShowSprite: {
     asl plex_sprite_msb
     lda plex_sprite_msb
     cmp #0
-    bne breturn
+    bne b5
     lda #1
     sta plex_sprite_msb
-  breturn:
+    rts
+  b5:
     rts
   b1:
     lda SPRITES_XMSB
@@ -185,6 +186,12 @@ plexSort: {
     dex
     cpx #$ff
     bne b5
+    jmp b4
+  b5:
+    lda nxt_y
+    ldy PLEX_SORTED_IDX,x
+    cmp PLEX_YPOS,y
+    bcc b3
   b4:
     inx
     lda nxt_idx
@@ -202,12 +209,6 @@ plexSort: {
     cpx #8
     bne plexFreePrepare1_b1
     rts
-  b5:
-    lda nxt_y
-    ldy PLEX_SORTED_IDX,x
-    cmp PLEX_YPOS,y
-    bcc b3
-    jmp b4
 }
 // Initialize the program
 init: {

src/test/ref/loop-break-continue.asm

@@ -19,11 +19,7 @@ main: {
     lda str,x
     cmp #' '
     bne b3
-  b4:
-    inx
-    cpx #0
-    bne b1
-    rts
+    jmp b4
   b3:
     lda str,x
     ldy #0
@@ -32,6 +28,10 @@ main: {
     bne !+
     inc screen+1
   !:
-    jmp b4
+  b4:
+    inx
+    cpx #0
+    bne b1
+    rts
     str: .text "hello brave new world@"
 }

src/test/ref/loop-break-nested.asm

@@ -20,6 +20,13 @@ main: {
     lda (line),y
     cmp #'a'
     bne b3
+    jmp b4
+  b3:
+    lda #'a'
+    sta (line),y
+    iny
+    cpy #$28
+    bne b2
   b4:
     lda #$28
     clc
@@ -37,11 +44,4 @@ main: {
     bcc b1
   !:
     rts
-  b3:
-    lda #'a'
-    sta (line),y
-    iny
-    cpy #$28
-    bne b2
-    jmp b4
 }

src/test/ref/loop-continue.asm

@@ -9,12 +9,12 @@ main: {
     lda SCREEN,x
     cmp #' '
     bne b2
+    jmp b3
+  b2:
+    inc SCREEN,x
   b3:
     inx
     cpx #$28*6+1
     bne b1
     rts
-  b2:
-    inc SCREEN,x
-    jmp b3
 }

src/test/ref/multiplexer-irq/simple-multiplexer-irq.asm

@@ -130,6 +130,12 @@ plexSort: {
     dex
     cpx #$ff
     bne b6
+    jmp b4
+  b6:
+    lda nxt_y
+    ldy PLEX_SORTED_IDX,x
+    cmp PLEX_YPOS,y
+    bcc b3
   b4:
     inx
     lda nxt_idx
@@ -154,12 +160,6 @@ plexSort: {
     bne plexFreePrepare1_b1
     sta plex_free_next
     rts
-  b6:
-    lda nxt_y
-    ldy PLEX_SORTED_IDX,x
-    cmp PLEX_YPOS,y
-    bcc b3
-    jmp b4
 }
 // Initialize the program
 init: {
@@ -242,6 +242,10 @@ plex_irq: {
     lda plex_show_idx
     cmp #PLEX_COUNT
     bcc b7
+    jmp b4
+  b7:
+    cpx _4
+    bcc b3
   b4:
     lda #IRQ_RASTER
     sta IRQ_STATUS
@@ -257,10 +261,6 @@ plex_irq: {
   b1:
     stx RASTER
     jmp b2
-  b7:
-    cpx _4
-    bcc b3
-    jmp b4
 }
 // Show the next sprite.
 // plexSort() prepares showing the sprites

src/test/ref/no-recursion-heavy.asm

@@ -270,5 +270,8 @@ fc: {
     cmp #7
     cmp #8
     cmp #9
+    bne b10
+    rts
+  b10:
     rts
 }

src/test/ref/sandbox.asm

@@ -202,9 +202,16 @@ myprintf: {
     rts
   b3:
     cpx #'1'
-    bcc !b37+
-    jmp b37
-  !b37:
+    bcs b37
+    jmp b4
+  b37:
+    cpx #'9'
+    bcs !b23+
+    jmp b23
+  !b23:
+    bne !b23+
+    jmp b23
+  !b23:
   b4:
     cpx #'-'
     bne b5
@@ -280,6 +287,11 @@ myprintf: {
     lda bTrailing
     cmp #0
     beq b39
+    jmp b15
+  b39:
+    lda b
+    cmp bDigits
+    bcc b16
   b15:
     ldx #0
   b19:
@@ -309,18 +321,13 @@ myprintf: {
     cmp bDigits
     bcc b21
     jmp b22
-  b39:
-    lda b
-    cmp bDigits
-    bcc b16
-    jmp b15
   b16:
     lda bLeadZero
     cmp #0
-    beq b14
+    beq b17
     lda #'0'
     jmp b18
-  b14:
+  b17:
     lda #' '
   b18:
     ldy bLen
@@ -341,11 +348,6 @@ myprintf: {
     sta strTemp,y
     inc bLen
     jmp b22
-  b37:
-    cpx #'9'
-    bcc b23
-    beq b23
-    jmp b4
   b23:
     txa
     axs #'0'
@@ -392,13 +394,7 @@ myprintf: {
   b28:
     cpx #$41
     bcs b41
-  b30:
-    // swap 0x41 / 0x61 when in lower case mode
-    ldy bLen
-    txa
-    sta strTemp,y
-    inc bLen
-    jmp b27
+    jmp b30
   b41:
     cpx #$5a+1
     bcc b35
@@ -406,7 +402,13 @@ myprintf: {
   b35:
     txa
     axs #-[$20]
-    jmp b30
+  b30:
+    // swap 0x41 / 0x61 when in lower case mode
+    ldy bLen
+    txa
+    sta strTemp,y
+    inc bLen
+    jmp b27
     buf6: .fill 6, 0
 }
 // utoa(word zeropage($12) value, byte* zeropage($14) dst)

src/test/ref/signed-words.asm

@@ -22,7 +22,7 @@
   .label yvel_10 = 6
   .label xvel = 2
   .label yvel_12 = 6
-  .label yvel_22 = 6
+  .label yvel_21 = 6
 main: {
     jsr init
     lda #$64
@@ -93,9 +93,9 @@ anim: {
     sta yvel+1
   b3:
     lda yvel
-    sta yvel_22
+    sta yvel_21
     lda yvel+1
-    sta yvel_22+1
+    sta yvel_21+1
     lda #0
     sta ypos
     sta ypos+1

src/test/ref/travis1.asm

@@ -22,6 +22,17 @@ main: {
     jsr game_ready
     cmp #0
     bne b3
+    jmp b2
+  b3:
+    lda print_line_cursor
+    sta print_char_cursor
+    lda print_line_cursor+1
+    sta print_char_cursor+1
+    lda #<str
+    sta print_str_ln.str
+    lda #>str
+    sta print_str_ln.str+1
+    jsr print_str_ln
   b2:
     inc i
     lda #6
@@ -34,17 +45,6 @@ main: {
     lda print_line_cursor+1
     sta print_char_cursor+1
     jmp b1
-  b3:
-    lda print_line_cursor
-    sta print_char_cursor
-    lda print_line_cursor+1
-    sta print_char_cursor+1
-    lda #<str
-    sta print_str_ln.str
-    lda #>str
-    sta print_str_ln.str+1
-    jsr print_str_ln
-    jmp b2
     str: .text "ready!@"
 }
 // Print a zero-terminated string followed by a newline