verafx: fix copy() and clear() unrolling

compiler/res/prog8lib/cx16/verafx.p8

@@ -42,25 +42,19 @@ verafx {
         cx16.VERA_FX_MULT = 0
         cx16.VERA_FX_CTRL = %01000000    ; cache write enable
 
-        if (num_longwords & %1111110000000011) == 0 {
-            repeat lsb(num_longwords >> 2)
-                unroll 4 cx16.VERA_DATA0=0       ; write 4*4 bytes at a time, unrolled
+        cx16.r0 = num_longwords>>3
+        if cx16.r0H==0 {
+            repeat cx16.r0L {
+                unroll 8 cx16.VERA_DATA0=0       ; write 8*4 bytes at a time, unrolled
+            }
+        } else {
+            repeat cx16.r0 {
+                unroll 8 cx16.VERA_DATA0=0       ; write 8*4 bytes at a time, unrolled
+            }
         }
-        else if (num_longwords & %1111111000000001) == 0 {
-            repeat lsb(num_longwords >> 1)
-                unroll 2 cx16.VERA_DATA0=0       ; write 2*4 bytes at a time, unrolled
-        }
-        else if (lsb(num_longwords) & 3) == 0 {
-            repeat num_longwords >> 2
-                unroll 4 cx16.VERA_DATA0=0       ; write 4*4 bytes at a time, unrolled
-        }
-        else if (lsb(num_longwords) & 1) == 0 {
-            repeat num_longwords >> 1
-                unroll 2 cx16.VERA_DATA0=0       ; write 2*4 bytes at a time, unrolled
-        }
-        else {
-            repeat num_longwords
-                cx16.VERA_DATA0=0       ; write 4 bytes at a time
+
+        repeat lsb(num_longwords) & 7 {
+            cx16.VERA_DATA0=0       ; write 4 bytes at a time (remaining longs)
         }
 
         cx16.VERA_FX_CTRL = 0       ; cache write disable
@@ -81,35 +75,41 @@ verafx {
         cx16.VERA_CTRL = 2<<1       ; dcsel = 2
         cx16.VERA_FX_MULT = 0
         cx16.VERA_FX_CTRL = %01100000    ; cache write enable + cache fill enable
-        cx16.r0 = num_longwords
 
-        if (cx16.r0L & 1) == 0 {
-            repeat cx16.r0>>1 {
-                %asm {{
-                    lda  cx16.VERA_DATA1    ; fill cache with 4 source bytes...
+        cx16.r0 = num_longwords>>1
+
+        if cx16.r0H==0 {
+            repeat cx16.r0L {
+                unroll 2 %asm {{
                     lda  cx16.VERA_DATA1
                     lda  cx16.VERA_DATA1
                     lda  cx16.VERA_DATA1
-                    stz  cx16.VERA_DATA0    ; write 4 bytes at once.
-                    lda  cx16.VERA_DATA1    ; fill cache with 4 source bytes...
                     lda  cx16.VERA_DATA1
-                    lda  cx16.VERA_DATA1
-                    lda  cx16.VERA_DATA1
-                    stz  cx16.VERA_DATA0    ; write 4 bytes at once.
+                    stz  cx16.VERA_DATA0
                 }}
             }
         } else {
             repeat cx16.r0 {
-                %asm {{
-                    lda  cx16.VERA_DATA1    ; fill cache with 4 source bytes...
+                unroll 2 %asm {{
                     lda  cx16.VERA_DATA1
                     lda  cx16.VERA_DATA1
                     lda  cx16.VERA_DATA1
-                    stz  cx16.VERA_DATA0    ; write 4 bytes at once.
+                    lda  cx16.VERA_DATA1
+                    stz  cx16.VERA_DATA0
                 }}
             }
         }
 
+        if lsb(num_longwords) & 1 == 1 {
+            %asm {{
+                lda  cx16.VERA_DATA1
+                lda  cx16.VERA_DATA1
+                lda  cx16.VERA_DATA1
+                lda  cx16.VERA_DATA1
+                stz  cx16.VERA_DATA0
+            }}
+        }
+
         cx16.VERA_FX_CTRL = 0    ; cache write disable
         cx16.VERA_CTRL = 0
     }

compiler/test/TestCompilerOnExamples.kt

@@ -221,7 +221,7 @@ class TestCompilerOnExamplesCx16: FunSpec({
             ExampleSizes("charfade", cx16SizeOptimized=1054, cx16SizeUnoptimized=1371),
             ExampleSizes("charsets", cx16SizeOptimized=1125, cx16SizeUnoptimized=1417),
             ExampleSizes("circles", cx16SizeOptimized=2409, cx16SizeUnoptimized=5209),
-            ExampleSizes("cobramk3-gfx", cx16SizeOptimized=7842, cx16SizeUnoptimized=11048),
+            ExampleSizes("cobramk3-gfx", cx16SizeOptimized=7722, cx16SizeUnoptimized=10930),
             ExampleSizes("colorbars", cx16SizeOptimized=1606, cx16SizeUnoptimized=1904),
             ExampleSizes("cxlogo", cx16SizeOptimized=599, cx16SizeUnoptimized=895),
             ExampleSizes("diskspeed", cx16SizeOptimized=6298, cx16SizeUnoptimized=8376),

docs/source/todo.rst

@@ -81,7 +81,6 @@ Libraries
 Optimizations
 -------------
 
-- repeat 256 {...} doesn't use STZ on the 65C02 (maybe other repeat values too?)
 - inliner: extend multi-value return inlining to support parameterized subroutines. Currently only works for parameterless subroutines. (Void calls with parameters already work if the parameters are unused in the body.)
 - bind types in the Ast much sooner than the simplifiedAst creation, so that we maybe could get rid of InferredType ?
 - Port more benchmarks from https://thred.github.io/c-bench-64/  to prog8 and see how it stacks up. (see benchmark-c/ directory)

examples/test.p8

@@ -1,26 +1,14 @@
-%import textio
-
-plane {
-   struct Point {
-       ubyte x
-       ubyte y
-   }
-}
-
-txt {
-    %option merge
-    sub print_pt(^^plane.Point p) {
-        txt.chrout('(')
-        txt.print_ub(p.x)
-        txt.chrout(',')
-        txt.print_ub(p.y)
-        txt.chrout(')')
-    }
-}
+%import monogfx
+%import verafx
 
 main {
     sub start() {
-        ^^plane.Point origin = ^^plane.Point:[0,0]
-        txt.print_pt(origin)
+        monogfx.lores()
+        monogfx.circle(160, 60, 59, true)
+
+        verafx.clear(0, 0, 255, 320*10/8/4)
+        verafx.copy(0,0, 0, 320*120/8, 320*120/8/4)
+
+        repeat {}
     }
 }