1  /* { dg-do compile } */
       2  /* { dg-options "-O2 -mcpu=ev67" } */
       3  
       4  typedef short INT16;
       5  typedef unsigned int CARD32;
       6  typedef unsigned short CARD16;
       7  typedef unsigned char CARD8;
       8  typedef struct _Picture *PicturePtr;
       9  typedef int FbStride;
      10  typedef unsigned long __m64;
      11  extern __m64 load8888 (__m64);
      12  static __inline __m64 _mm_setzero_si64(void)
      13  {
      14    return (__m64)0L;
      15  }
      16  static __inline __m64 _mm_adds_pu8(__m64 __m1, __m64 __m2)
      17  {
      18      return __m1 + __builtin_alpha_minsb8(__m2, ~__m1);
      19  }
      20  static __inline __m64 _mm_packs_pu16(__m64 __m1, __m64 __m2)
      21  {
      22      __m1 = __builtin_alpha_minuw4(__m1, 0x00ff00ff00ff00ff);
      23      __m2 = __builtin_alpha_minuw4(__m2, 0x00ff00ff00ff00ff);
      24      return __m1 | (__m2 << 32);
      25  }
      26  typedef unsigned long long ullong;
      27  static __inline__ __m64 pix_multiply(__m64 a)
      28  {
      29      if (a)
      30  	return a;
      31  }
      32  static __inline__ __m64 over(__m64 src, __m64 srca, __m64 dest)
      33  {
      34      return _mm_adds_pu8(src, pix_multiply(dest));
      35  }
      36  
      37  void fbCompositeSolid_nx8888mmx(CARD8 op, PicturePtr pSrc, PicturePtr pMask,
      38  				INT16 yDst, CARD16 width, CARD16 height)
      39  {
      40      CARD32 src;
      41      CARD32 *dstLine, *dst;
      42      CARD16 w;
      43      FbStride dstStride;
      44      __m64 vsrc, vsrca;
      45      vsrc = load8888(src);
      46      while (height--) {
      47  	dst = dstLine;
      48  	dstLine += dstStride;
      49  	while (w && (unsigned long) dst & 7) {
      50  	    *dst = _mm_packs_pu16(_mm_adds_pu8(vsrc, load8888(*dst)),
      51  				  _mm_setzero_si64());
      52  	    dst++;
      53  	}
      54  	while (w >= 2) {
      55  	    __m64 dest0, dest1;
      56  	    *(__m64 *) dst = _mm_packs_pu16(dest0, dest1);
      57  	    w -= 2;
      58  	}
      59  	while (w) {
      60  	    *dst = _mm_packs_pu16(_mm_adds_pu8(vsrc, pix_multiply(0)), 0);
      61  	    w--;
      62  	}
      63      }
      64  }