void main()
{
a = _m_from_int( k );
printf( "int=%8.8lx m=%8.8lx%8.8lx\n",
k, a._32[1], a._32[0] );
}
__m64 test90(int a) {
// CHECK: movd
return _m_from_int(a);
}
__m64 test_m_from_int(int a) {
// CHECK-LABEL: test_m_from_int
// CHECK: insertelement <2 x i32>
return _m_from_int(a);
}
/* Fill a surface with a gradient which is generated by bilinearly
interpolating between four corner color values. Can take
a source surface and multiply it into the gradient, but if
'src' is NULL, it will generate the gradient without multiplying */
static void fillBlend(
SDL_Surface *dst, SDL_Surface *src, BROGUE_DRAW_COLOR *color)
{
int x, y;
int lr, lg, lb, rr, rg, rb;
int ldr, ldg, ldb, rdr, rdg, rdb;
int w, h;
BROGUE_DRAW_COLOR ul = color[0];
BROGUE_DRAW_COLOR ur = color[1];
BROGUE_DRAW_COLOR bl = color[2];
BROGUE_DRAW_COLOR br = color[3];
#if defined(__MMX__)
int mmx = SDL_HasMMX();
#endif
w = dst->w;
h = dst->h;
if (src != NULL)
{
assert(dst->w == src->w);
assert(dst->h == src->h);
}
lr = clamp(ul.red * 0xFFFF, 0, 0xFFFF);
lg = clamp(ul.green * 0xFFFF, 0, 0xFFFF);
lb = clamp(ul.blue * 0xFFFF, 0, 0xFFFF);
rr = clamp(ur.red * 0xFFFF, 0, 0xFFFF);
rg = clamp(ur.green * 0xFFFF, 0, 0xFFFF);
rb = clamp(ur.blue * 0xFFFF, 0, 0xFFFF);
ldr = (clamp(bl.red * 0xFFFF, 0, 0xFFFF) - lr) / h;
ldg = (clamp(bl.green * 0xFFFF, 0, 0xFFFF) - lg) / h;
ldb = (clamp(bl.blue * 0xFFFF, 0, 0xFFFF) - lb) / h;
rdr = (clamp(br.red * 0xFFFF, 0, 0xFFFF) - rr) / h;
rdg = (clamp(br.green * 0xFFFF, 0, 0xFFFF) - rg) / h;
rdb = (clamp(br.blue * 0xFFFF, 0, 0xFFFF) - rb) / h;
for (y = 0; y < h; y++)
{
unsigned char *pix;
int dr, dg, db;
int rpp, gpp, bpp, raccum, gaccum, baccum;
pix = (unsigned char *)dst->pixels + dst->pitch * y;
dr = rr - lr;
dg = rg - lg;
db = rb - lb;
rpp = dr / w;
gpp = dg / w;
bpp = db / w;
raccum = lr;
gaccum = lg;
baccum = lb;
lr += ldr;
lg += ldg;
lb += ldb;
rr += rdr;
rg += rdg;
rb += rdb;
if (src != NULL)
{
unsigned char *src_pix = (unsigned char *)src->pixels
+ src->pitch * y;
x = w;
#if defined(__MMX__)
/* MMX is significantly faster. Use it if the CPU supports it */
if (mmx)
{
__m64 mmx_zero = _m_from_int(0);
long long ll_color =
((long long)0xFFFF << 48)
| ((long long)raccum << 32)
| ((long long)gaccum << 16)
| ((long long)baccum);
__m64 mmx_color = *(__m64 *)&ll_color;
long long ll_pp =
((long long)(rpp & 0xFFFF) << 32)
| ((long long)(gpp & 0xFFFF) << 16)
| ((long long)(bpp & 0xFFFF));
__m64 mmx_pp = *(__m64 *)&ll_pp;
while (x >= 2)
{
__m64 src_pair = *(__m64 *)src_pix;
/* Separate the left pixel and right pixel */
__m64 left_pix = _mm_unpacklo_pi8(src_pair, mmx_zero);
__m64 right_pix = _mm_unpackhi_pi8(src_pair, mmx_zero);
/* Multiply the left source by the gradient color */
left_pix = _mm_mullo_pi16(left_pix,
_mm_srli_pi16(mmx_color, 8));
/* Advance the gradient color for the next pixel */
mmx_color = _mm_add_pi16(mmx_color, mmx_pp);
/* Multiply the right source by the gradient color */
right_pix = _mm_mullo_pi16(right_pix,
_mm_srli_pi16(mmx_color, 8));
/* Advance the gradient */
mmx_color = _mm_add_pi16(mmx_color, mmx_pp);
/* Recombine the pixels */
__m64 result_pix = _mm_packs_pu16(
_mm_srli_pi16(left_pix, 8),
_mm_srli_pi16(right_pix, 8));
*(__m64 *)pix = result_pix;
src_pix += 8;
pix += 8;
x -= 2;
}
/* Extract the accumulated gradient value for the potential
odd remaining pixel */
short *s_color = (short *)&mmx_color;
raccum = s_color[2];
gaccum = s_color[1];
baccum = s_color[0];
}
#endif
/* The equivalent slow loop for odd pixels or CPUs without MMX */
while (x > 0)
{
#if SDL_BYTEORDER == SDL_LIL_ENDIAN
pix[3] = src_pix[3];
pix[2] = (src_pix[2] * raccum) >> 16;
pix[1] = (src_pix[1] * gaccum) >> 16;
pix[0] = (src_pix[0] * baccum) >> 16;
#else
pix[0] = src_pix[0];
pix[1] = (src_pix[1] * raccum) >> 16;
pix[2] = (src_pix[2] * gaccum) >> 16;
pix[3] = (src_pix[3] * baccum) >> 16;
#endif
raccum += rpp;
gaccum += gpp;
baccum += bpp;
src_pix += 4;
pix += 4;
x--;
}
}
else
{
void lines_scale2(const unsigned char *src, unsigned y, unsigned char *dst1, unsigned char *dst2, unsigned nPix)
{
const unsigned char
*u = src + ((y-1) & 7)*sc2lines_width,
*m = src + ((y+0) & 7)*sc2lines_width,
*l = src + ((y+1) & 7)*sc2lines_width;
for (unsigned i = 0; i < nPix; i += 4) {
if (*(unsigned*)(u+i) ^ *(unsigned*)(l+i)) {
__m64 mm = *(__m64*)(m+i-2);
__m64 uu = *(__m64*)(u+i-2);
__m64 ll = *(__m64*)(l+i-2);
__m64 md = _mm_slli_si64(mm,8);
__m64 mf = _mm_srli_si64(mm,8);
__m64 maskall = _mm_or_si64(_mm_cmpeq_pi8(md,mf), _mm_cmpeq_pi8(uu,ll));
__m64 e0, e1, v1, v2;
e0 = _mm_cmpeq_pi8(md,uu);
e0 = _mm_andnot_si64(maskall, e0);
e0 = _mm_srli_si64(e0,16);
e0 = _mm_unpacklo_pi8(e0, _mm_setzero_si64());
e1 = _mm_cmpeq_pi8(mf,uu);
e1 = _mm_andnot_si64(maskall, e1);
e1 = _mm_srli_si64(e1,16);
e1 = _mm_unpacklo_pi8(_mm_setzero_si64(), e1);
e0 = _mm_or_si64(e0, e1);
v1 = _m_from_int(*(unsigned*)(m+i));
v2 = _m_from_int(*(unsigned*)(u+i));
v1 = _mm_unpacklo_pi8(v1,v1);
v2 = _mm_unpacklo_pi8(v2,v2);
*(__m64*)(dst1 + 2*i) = _mm_or_si64( _mm_and_si64(e0,v2), _mm_andnot_si64(e0,v1) );
e0 = _mm_cmpeq_pi8(md,ll);
e0 = _mm_andnot_si64(maskall, e0);
e0 = _mm_srli_si64(e0,16);
e0 = _mm_unpacklo_pi8(e0, _mm_setzero_si64());
e1 = _mm_cmpeq_pi8(mf,ll);
e1 = _mm_andnot_si64(maskall, e1);
e1 = _mm_srli_si64(e1,16);
e1 = _mm_unpacklo_pi8(_mm_setzero_si64(), e1);
e0 = _mm_or_si64(e0, e1);
v1 = _m_from_int(*(unsigned*)(m+i));
v2 = _m_from_int(*(unsigned*)(l+i));
v1 = _mm_unpacklo_pi8(v1,v1);
v2 = _mm_unpacklo_pi8(v2,v2);
*(__m64*)(dst2 + 2*i) = _mm_or_si64( _mm_and_si64(e0,v2), _mm_andnot_si64(e0,v1) );
} else {
__m64 v1 = _m_from_int(*(unsigned*)(m+i));
v1 = _mm_unpacklo_pi8(v1,v1);
*(__m64*)(dst1 + 2*i) = v1;
*(__m64*)(dst2 + 2*i) = v1;
}
}
}