Add minor optimization removing branching

2 files changed, 13 insertions, 44 deletions

diff --git a/src/renderer_m0.c b/src/renderer_m0.c
index eaf4658..7e598c3 100644
--- a/src/renderer_m0.c
+++ b/src/renderer_m0.c
@@ -472,10 +472,7 @@ draw_2bpp_row(size_t x, size_t y, u8 a, u8 b, u8 flip_x) {
 
     u32 *dst = &backbuf[start_row + (tile_x + tile_y * 32) * 8];
 #if DEC_BIG_LUT
-    u32 *lut = dec_byte;
-    if (flip_x) {
-        lut = dec_byte_flip_x;
-    }
+    u32 *lut = flip_x ? dec_byte_flip_x : dec_byte;
     u32 clr_a = lut[a];
     u32 clr_b = lut[b];
 #else
@@ -486,12 +483,8 @@ draw_2bpp_row(size_t x, size_t y, u8 a, u8 b, u8 flip_x) {
     u32 mask_b = (clr_b * 0xF);
     u32 mask = (mask_a | mask_b);
     u32 color = clr_a + (clr_b << 1);
-    if (start_col == 0) {
-        dst[0] = (dst[0] & ~mask) | color;
-    } else {
-        dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
-        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);
 
     // TODO: different blend modes?
 }
@@ -513,12 +506,8 @@ draw_1bpp_row(size_t x, size_t y, u8 a, u8 clr, u8 flip_x) {
     u32 color = decode_1bpp(a, flip_x);
     u32 mask = (color * 0xF);
     color *= clr;
-    if (start_col == 0) {
-        dst[0] = (dst[0] & ~mask) | color;
-    } else {
-        dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
-        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);
 
     // TODO: different blend modes?
 }
diff --git a/src/renderer_m4.c b/src/renderer_m4.c
index 2a5e9fe..73b51d5 100644
--- a/src/renderer_m4.c
+++ b/src/renderer_m4.c
@@ -449,10 +449,7 @@ draw_2bpp_row(size_t x, size_t y, u8 a, u8 b, u8 flip_x) {
 
     u64 *dst = &backbuf[(y * 30 + tile_x) * 8 / 2];
 #if DEC_BIG_LUT
-    u64 *lut = dec_byte;
-    if (flip_x) {
-        lut = dec_byte_flip_x;
-    }
+    u64 *lut = flip_x ? dec_byte_flip_x : dec_byte;
     u64 clr_a = lut[a];
     u64 clr_b = lut[b];
 #else
@@ -463,15 +460,8 @@ draw_2bpp_row(size_t x, size_t y, u8 a, u8 b, u8 flip_x) {
     u64 mask_b = (clr_b * 0xF);
     u64 mask = (mask_a | mask_b);
     u64 color = clr_a + (clr_b << 1);
-    if (start_col == 0) {
-        dst[0] = (dst[0] & ~mask) | color;
-    } else {
-        dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
-        if ((x + 7) > (SCREEN_WIDTH)) {
-            return;
-        }
-        dst[1] = (dst[1] & ~(mask >> shift_right)) | (color >> shift_right);
-    }
+    dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
+    dst[1] = (dst[1] & ~(mask >> shift_right)) | (color >> shift_right);
 
     // TODO: different blend modes?
 }
@@ -488,21 +478,11 @@ draw_1bpp_row(size_t x, size_t y, u8 a, u8 clr, u8 flip_x) {
     size_t shift_right = (8 - start_col) * 8;
 
     u64 *dst = &backbuf[(y * 30 + tile_x) * 8 / 2];
-    if (start_col == 0) {
-        u64 color = decode_1bpp(a, flip_x);
-        u64 mask = (color * 0xF);
-        color *= clr;
-        dst[0] = (dst[0] & ~mask) | color;
-    } else {
-        u64 color = decode_1bpp(a, flip_x);
-        u64 mask = (color * 0xF);
-        color *= clr;
-        dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
-        if ((x + 7) > (SCREEN_WIDTH)) {
-            return;
-        }
-        dst[1] = (dst[1] & ~(mask >> shift_right)) | (color >> shift_right);
-    }
+    u64 color = decode_1bpp(a, flip_x);
+    u64 mask = (color * 0xF);
+    color *= clr;
+    dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
+    dst[1] = (dst[1] & ~(mask >> shift_right)) | (color >> shift_right);
 
     // TODO: different blend modes?
 }