recipes/xorg-lib/pixman-0.17.8/0005-Support-of-overlapping-src-dst-for-pixman_blt_sse2.patch

   1 From edc80b41c6480b7c80ec5f7c835c92b2debb3774 Mon Sep 17 00:00:00 2001
   2 From: Siarhei Siamashka <siarhei.siamashka@nokia.com>
   3 Date: Thu, 22 Oct 2009 05:45:54 +0300
   4 Subject: [PATCH 5/6] Support of overlapping src/dst for pixman_blt_sse2
   5
   6 ---
   7  pixman/pixman-sse2.c |   55 +++++++++++++++++++++++++++++--------------------
   8  1 files changed, 32 insertions(+), 23 deletions(-)
   9
  10 diff --git a/pixman/pixman-sse2.c b/pixman/pixman-sse2.c
  11 index 78b0ad1..b84636b 100644
  12 --- a/pixman/pixman-sse2.c
  13 +++ b/pixman/pixman-sse2.c
  14 @@ -5300,34 +5300,43 @@ pixman_blt_sse2 (uint32_t *src_bits,
  15  {
  16      uint8_t *   src_bytes;
  17      uint8_t *   dst_bytes;
  18 -    int byte_width;
  19 +    int         bpp;
  20
  21 -    if (src_bpp != dst_bpp)
  22 +    if (src_bpp != dst_bpp || src_bpp & 7)
  23         return FALSE;
  24
  25 -    if (src_bpp == 16)
  26 -    {
  27 -       src_stride = src_stride * (int) sizeof (uint32_t) / 2;
  28 -       dst_stride = dst_stride * (int) sizeof (uint32_t) / 2;
  29 -       src_bytes =(uint8_t *)(((uint16_t *)src_bits) + src_stride * (src_y) + (src_x));
  30 -       dst_bytes = (uint8_t *)(((uint16_t *)dst_bits) + dst_stride * (dst_y) + (dst_x));
  31 -       byte_width = 2 * width;
  32 -       src_stride *= 2;
  33 -       dst_stride *= 2;
  34 -    }
  35 -    else if (src_bpp == 32)
  36 +    bpp = src_bpp >> 3;
  37 +    width *= bpp;
  38 +    src_stride *= 4;
  39 +    dst_stride *= 4;
  40 +    src_bytes = (uint8_t *)src_bits + src_y * src_stride + src_x * bpp;
  41 +    dst_bytes = (uint8_t *)dst_bits + dst_y * dst_stride + dst_x * bpp;
  42 +
  43 +    if (src_bpp != 16 && src_bpp != 32)
  44      {
  45 -       src_stride = src_stride * (int) sizeof (uint32_t) / 4;
  46 -       dst_stride = dst_stride * (int) sizeof (uint32_t) / 4;
  47 -       src_bytes = (uint8_t *)(((uint32_t *)src_bits) + src_stride * (src_y) + (src_x));
  48 -       dst_bytes = (uint8_t *)(((uint32_t *)dst_bits) + dst_stride * (dst_y) + (dst_x));
  49 -       byte_width = 4 * width;
  50 -       src_stride *= 4;
  51 -       dst_stride *= 4;
  52 +       pixman_blt_helper (src_bytes, dst_bytes, src_stride, dst_stride,
  53 +                          width, height);
  54 +       return TRUE;
  55      }
  56 -    else
  57 +
  58 +    if (src_bytes < dst_bytes && src_bytes + src_stride * height > dst_bytes)
  59      {
  60 -       return FALSE;
  61 +       src_bytes += src_stride * height - src_stride;
  62 +       dst_bytes += dst_stride * height - dst_stride;
  63 +       dst_stride = -dst_stride;
  64 +       src_stride = -src_stride;
  65 +
  66 +       if (src_bytes + width > dst_bytes)
  67 +       {
  68 +           /* TODO: reverse scanline copy using SSE2 */
  69 +           while (--height >= 0)
  70 +           {
  71 +               memmove (dst_bytes, src_bytes, width);
  72 +               dst_bytes += dst_stride;
  73 +               src_bytes += src_stride;
  74 +           }
  75 +           return TRUE;
  76 +       }
  77      }
  78
  79      cache_prefetch ((__m128i*)src_bytes);
  80 @@ -5340,7 +5349,7 @@ pixman_blt_sse2 (uint32_t *src_bits,
  81         uint8_t *d = dst_bytes;
  82         src_bytes += src_stride;
  83         dst_bytes += dst_stride;
  84 -       w = byte_width;
  85 +       w = width;
  86
  87         cache_prefetch_next ((__m128i*)s);
  88         cache_prefetch_next ((__m128i*)d);
  89 --
  90 1.6.2.4
  91