Huw Davies : gdi32: Eliminate a masking operation by generating a larger colour table.

[Alexandre Julliard]

Module: wine
Branch: master
Commit: 6c7ed1214c6d10b4cb598f976175431e2031bc91
URL:    http://source.winehq.org/git/wine.git/?a=commit;h=6c7ed1214c6d10b4cb598f976175431e2031bc91

Author: Huw Davies <huw at codeweavers.com>
Date:   Wed Aug 24 08:55:32 2016 +0100

gdi32: Eliminate a masking operation by generating a larger colour table.

This helps significantly for large rectangles.

This is only done for the 32-bpp case as typically the lower bit
depths would more likely be used to draw small icons and the like,
where the extra fixed cost of this optimization will potentially
outweigh the savings.

Signed-off-by: Huw Davies <huw at codeweavers.com>
Signed-off-by: Alexandre Julliard <julliard at winehq.org>

---

 dlls/gdi32/dibdrv/primitives.c | 26 +++++++++++++++-----------
 1 file changed, 15 insertions(+), 11 deletions(-)

diff --git a/dlls/gdi32/dibdrv/primitives.c b/dlls/gdi32/dibdrv/primitives.c
index e75792c..d512af2 100644
--- a/dlls/gdi32/dibdrv/primitives.c
+++ b/dlls/gdi32/dibdrv/primitives.c
@@ -5264,7 +5264,7 @@ static BOOL gradient_rect_null( const dib_info *dib, const RECT *rc, const TRIVE
 static void mask_rect_32( const dib_info *dst, const RECT *rc,
                           const dib_info *src, const POINT *origin, int rop2 )
 {
-    DWORD *dst_start = get_pixel_ptr_32(dst, rc->left, rc->top), dst_colors[2];
+    DWORD *dst_start = get_pixel_ptr_32(dst, rc->left, rc->top), dst_colors[256];
     DWORD src_val, bit_val, i, full, pos;
     struct rop_codes codes;
     int x, y;
@@ -5274,15 +5274,19 @@ static void mask_rect_32( const dib_info *dst, const RECT *rc,
     get_rop_codes( rop2, &codes );
 
     if (dst->funcs == &funcs_8888)
-        for (i = 0; i < sizeof(dst_colors) / sizeof(dst_colors[0]); i++)
+        for (i = 0; i < 2; i++)
             dst_colors[i] = color_table[i].rgbRed << 16 | color_table[i].rgbGreen << 8 |
                 color_table[i].rgbBlue;
     else
-        for (i = 0; i < sizeof(dst_colors) / sizeof(dst_colors[0]); i++)
+        for (i = 0; i < 2; i++)
             dst_colors[i] = put_field(color_table[i].rgbRed,   dst->red_shift,   dst->red_len) |
                             put_field(color_table[i].rgbGreen, dst->green_shift, dst->green_len) |
                             put_field(color_table[i].rgbBlue,  dst->blue_shift,  dst->blue_len);
 
+    /* Creating a BYTE-sized table so we don't need to mask the lsb of bit_val */
+    for (i = 2; i < sizeof(dst_colors) / sizeof(dst_colors[0]); i++)
+        dst_colors[i] = dst_colors[i & 1];
+
     for (y = rc->top; y < rc->bottom; y++)
     {
         pos = origin->x & 7;
@@ -5298,21 +5302,21 @@ static void mask_rect_32( const dib_info *dst, const RECT *rc,
         {
             src_val = src_start[pos / 8];
 
-            bit_val = (src_val >> 7) & 1;
+            bit_val = src_val >> 7;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = (src_val >> 6) & 1;
+            bit_val = src_val >> 6;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = (src_val >> 5) & 1;
+            bit_val = src_val >> 5;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = (src_val >> 4) & 1;
+            bit_val = src_val >> 4;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = (src_val >> 3) & 1;
+            bit_val = src_val >> 3;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = (src_val >> 2) & 1;
+            bit_val = src_val >> 2;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = (src_val >> 1) & 1;
+            bit_val = src_val >> 1;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
-            bit_val = src_val & 1;
+            bit_val = src_val;
             do_rop_codes_32( dst_start + x++, dst_colors[bit_val], &codes );
         }