Update uxn core and fix some new ppu bugs

1 files changed, 134 insertions, 28 deletions

diff --git a/src/ppu.c b/src/ppu.c
index c258fe2..afde963 100644
--- a/src/ppu.c
+++ b/src/ppu.c
@@ -247,6 +247,118 @@ ppu_pixel(u32 *layer, u16 x, u16 y, u8 clr) {
     dirty_tiles[tile_y] |= 1 << tile_x;
 }
 
+IWRAM_CODE
+void
+ppu_rect(u32 *layer, size_t x0, size_t y0, size_t x1, size_t y1, u8 clr) {
+    BOUNDCHECK_SCREEN(x0, y0);
+    BOUNDCHECK_SCREEN(x1, y1);
+
+    // Find row positions for the given x/y coordinates.
+    size_t tile_x0 = x0 / 8;
+    size_t tile_y0 = y0 / 8;
+    size_t tile_x1 = x1 / 8;
+    size_t tile_y1 = y1 / 8;
+    size_t start_col0 = x0 % 8;
+    size_t start_col1 = x1 % 8;
+    size_t start_row0 = y0 % 8;
+    size_t start_row1 = y1 % 8;
+
+    // Get a pointer to the backbuffer and the tile row.
+    u32 *buf_top = &layer[start_row0 + (tile_x0 + tile_y0 * 32) * 8];
+    u32 *buf_bot = &layer[start_row1 + (tile_x0 + tile_y1 * 32) * 8];
+
+    size_t dx = tile_x1 - tile_x0;
+    size_t dy = tile_y1 - tile_y0;
+
+    // We can update two lines at a time, which is faster than calling draw_line
+    // four times.
+    if (dx < 1) {
+        u32 row_mask = 0xFFFFFFFF;
+        row_mask >>= (7 - start_col1 - dx) * 4;
+        row_mask &= 0xFFFFFFFF << start_col0 * 4;
+        u32 row = (0x11111111 * clr) & row_mask;
+        buf_top[0] = (buf_top[0] & ~row_mask) | row;
+        buf_bot[0] = (buf_bot[0] & ~row_mask) | row;
+        dirty_tiles[tile_y0] |= 1 << tile_x0;
+        dirty_tiles[tile_y1] |= 1 << tile_x0;
+    } else {
+        size_t shift_left = start_col0 * 4;
+        size_t shift_right = (7 - start_col1) * 4;
+        u32 row_mask = 0xFFFFFFFF;
+        u32 row = 0x11111111 * clr;
+        buf_top[0] = buf_top[0] & ~(row_mask << shift_left);
+        buf_top[0] |= row << shift_left;
+        buf_bot[0] = buf_bot[0] & ~(row_mask << shift_left);
+        buf_bot[0] |= row << shift_left;
+        dirty_tiles[tile_y0] |= 1 << tile_x0;
+        dirty_tiles[tile_y1] |= 1 << tile_x0;
+        for (size_t i = 1; i < dx; i++) {
+            buf_top[i * 8] = row;
+            buf_bot[i * 8] = row;
+            dirty_tiles[tile_y0] |= 1 << (tile_x0 + i);
+            dirty_tiles[tile_y1] |= 1 << (tile_x0 + i);
+        }
+        buf_top[dx * 8] = buf_top[dx * 8] & ~(row_mask >> shift_right);
+        buf_top[dx * 8] |= row >> shift_right;
+        buf_bot[dx * 8] = buf_bot[dx * 8] & ~(row_mask >> shift_right);
+        buf_bot[dx * 8] |= row >> shift_right;
+        dirty_tiles[tile_y0] |= 1 << (tile_x0 + dx);
+        dirty_tiles[tile_y1] |= 1 << (tile_x0 + dx);
+    }
+    u32 row_mask_left = 0xF << start_col0 * 4;
+    u32 row_mask_right = 0xF << start_col1 * 4;
+    u32 row_left = (0x11111111 * clr) & row_mask_left;
+    u32 row_right = (0x11111111 * clr) & row_mask_right;
+    if (dy < 1) {
+        for (size_t i = 1; i < y1 - y0; i++, buf_top++) {
+            buf_top[1] = buf_top[1] & ~row_mask_left;
+            buf_top[1] |= row_left;
+            buf_top[1 + 8 * dx] = buf_top[1 + 8 * dx] & ~row_mask_right;
+            buf_top[1 + 8 * dx] |= row_right;
+        }
+    } else {
+        for (size_t i = 1; i < (8 - start_row0); i++, buf_top++) {
+            buf_top[1] = buf_top[1] & ~row_mask_left;
+            buf_top[1] |= row_left;
+            buf_top[1 + 8 * dx] = buf_top[1 + 8 * dx] & ~row_mask_right;
+            buf_top[1 + 8 * dx] |= row_right;
+        }
+        buf_top += 8 * 31;
+        for (size_t j = 1; j < dy; j++) {
+            for (size_t i = 0; i < 8; i++, buf_top++) {
+                buf_top[1] = buf_top[1] & ~row_mask_left;
+                buf_top[1] |= row_left;
+                buf_top[1 + 8 * dx] = buf_top[1 + 8 * dx] & ~row_mask_right;
+                buf_top[1 + 8 * dx] |= row_right;
+            }
+            buf_top += 8 * 31;
+            dirty_tiles[tile_y0 + j] |= 1 << tile_x0;
+            dirty_tiles[tile_y0 + j] |= 1 << (tile_x0 + dx);
+        }
+        for (size_t i = 0; i < start_row1; i++, buf_top++) {
+            buf_top[1] = buf_top[1] & ~row_mask_left;
+            buf_top[1] |= row_left;
+            buf_top[1 + 8 * dx] = buf_top[1 + 8 * dx] & ~row_mask_right;
+            buf_top[1 + 8 * dx] |= row_right;
+        }
+    }
+}
+
+IWRAM_CODE
+void
+screen_fill(u32 *layer, u16 x0, u16 y0, u16 x1, u16 y1, u8 clr) {
+    BOUNDCHECK_SCREEN(x0, y0);
+    BOUNDCHECK_SCREEN(x1, y1);
+
+    size_t dx = x1 - x0;
+    size_t dy = y1 - y0;
+    size_t n_rect = MIN(dx, dy);
+    n_rect = n_rect / 2 + 1;
+    for (size_t i = 0; i < n_rect; i++) {
+        ppu_rect(layer, x0 + i, y0 + i, x1 - i, y1 - i, clr);
+    }
+}
+
 #if NEW_PPU == 0
 IWRAM_CODE
 void
@@ -315,34 +427,28 @@ ppu_1bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
     u32 *dst = &layer[start_row + (tile_x + tile_y * 32) * 8];
     u32 *lut = flip_x ? dec_byte_flip_x : dec_byte;
     if (blending[4][clr]) {
-        u64 mask = ~((u64)0xFFFFFFFF);
+        u32 mask = 0xFFFFFFFF;
         if (!flip_y) {
             for(size_t v = 0; v < 8; v++, dst++) {
                 if ((y + v) >= SCREEN_HEIGHT) break;
                 u8 ch1 = sprite[v];
                 u32 color_1 = lut[ch1];
-                u32 color_2 = (color_1 ^ 0xffffffff) & 0x11111111;
+                u32 color_2 = (color_1 ^ 0xFFFFFFFF) & 0x11111111;
                 u32 color = (color_1 * (clr & 3)) | (color_2 * (clr >> 2));
-                if (start_col == 0) {
-                    dst[0] = (dst[0] & mask) | color;
-                } else {
-                    dst[0] = (dst[0] & (mask << shift_left)) | color;
-                    dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
-                }
+                dst[0] = (dst[0] & ~(mask << shift_left)) | (color << shift_left);
+                dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         } else {
             for(size_t v = 0; v < 8; v++, dst++) {
                 if ((y + v) >= SCREEN_HEIGHT) break;
                 u8 ch1 = sprite[(7 - v)];
                 u32 color_1 = lut[ch1];
-                u32 color_2 = (color_1 ^ 0xffffffff) & 0x11111111;
+                u32 color_2 = (color_1 ^ 0xFFFFFFFF) & 0x11111111;
                 u32 color = (color_1 * (clr & 3)) | (color_2 * (clr >> 2));
-                if (start_col == 0) {
-                    dst[0] = (dst[0] & mask) | color;
-                } else {
-                    dst[0] = (dst[0] & (mask << shift_left)) | color;
-                    dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
-                }
+                dst[0] = (dst[0] & ~(mask << shift_left)) | (color << shift_left);
+                dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         }
     } else {
@@ -359,6 +465,7 @@ ppu_1bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                     dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
                     dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
                 }
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         } else {
             for(size_t v = 0; v < 8; v++, dst++) {
@@ -373,6 +480,7 @@ ppu_1bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                     dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
                     dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
                 }
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         }
     }
@@ -529,6 +637,7 @@ ppu_2bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                     dst[0] = (dst[0] & (mask << shift_left)) | color;
                     dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
                 }
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         } else {
             for(size_t v = 0; v < 8; v++, dst++) {
@@ -542,10 +651,11 @@ ppu_2bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                     dst[0] = (dst[0] & (mask << shift_left)) | color;
                     dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
                 }
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         }
     } else if (blending[4][clr]) {
-        u64 mask = ~((u64)0xFFFFFFFF << shift_left);
+        u32 mask = 0xFFFFFFFF;
         u8 clr0 = blending[0][clr];
         u8 clr1 = blending[1][clr];
         u8 clr2 = blending[2][clr];
@@ -567,12 +677,9 @@ ppu_2bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                         (clr1 * col1mask) |
                         (clr2 * col2mask) |
                         (clr3 * col3mask);
-                if (start_col == 0) {
-                    dst[0] = (dst[0] & mask) | color;
-                } else {
-                    dst[0] = (dst[0] & (mask << shift_left)) | color;
-                    dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
-                }
+                dst[0] = (dst[0] & ~(mask << shift_left)) | (color << shift_left);
+                dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         } else {
             for(size_t v = 0; v < 8; v++, dst++) {
@@ -591,12 +698,9 @@ ppu_2bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                         (clr1 * col1mask) |
                         (clr2 * col2mask) |
                         (clr3 * col3mask);
-                if (start_col == 0) {
-                    dst[0] = (dst[0] & mask) | color;
-                } else {
-                    dst[0] = (dst[0] & (mask << shift_left)) | color;
-                    dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
-                }
+                dst[0] = (dst[0] & ~(mask << shift_left)) | (color << shift_left);
+                dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         }
     } else {
@@ -625,6 +729,7 @@ ppu_2bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                     dst[0] = (dst[0] & (mask << shift_left)) | color;
                     dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
                 }
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         } else {
             for(size_t v = 0; v < 8; v++, dst++) {
@@ -648,6 +753,7 @@ ppu_2bpp(u32 *layer, u16 x, u16 y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
                     dst[0] = (dst[0] & (mask << shift_left)) | color;
                     dst[8] = (dst[8] & (mask >> shift_right)) | (color >> shift_right);
                 }
+                if ((start_row + v)  == 7) dst += (32 - 1) * 8;
             }
         }
     }