1/* { dg-do compile } */
2/* { dg-options "-O2 -msse2" } */
3/* { dg-require-effective-target sse2 } */
4
5typedef float __v4sf __attribute__ ((__vector_size__ (16)));
6typedef float __m128 __attribute__ ((__vector_size__ (16)));
7typedef long long __v2di __attribute__ ((__vector_size__ (16)));
8
9static __inline __m128
10_mm_cmpeq_ps (__m128 __A, __m128 __B)
11{
12  return (__m128) __builtin_ia32_cmpeqps ((__v4sf)__A, (__v4sf)__B);
13}
14
15static __inline __m128
16_mm_setr_ps (float __Z, float __Y, float __X, float __W)
17{
18  return __extension__ (__m128)(__v4sf){__Z, __Y, __X, __W };
19}
20
21static __inline __m128
22_mm_and_si128 (__m128 __A, __m128 __B)
23{
24  return (__m128)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);
25}
26
27static __inline __m128
28_mm_or_si128 (__m128 __A, __m128 __B)
29{
30  return (__m128)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);
31}
32
33typedef union
34{
35  __m128 xmmi;
36  int si[4];
37}
38__attribute__ ((aligned (16))) um128;
39
40um128 u;
41
42static inline int
43sse_max_abs_indexf (float *v, int step, int n)
44{
45  __m128 m1, mm;
46  __m128 mim, mi, msk;
47  um128 u, ui;
48  int n4, step2, step3;
49  mm = __builtin_ia32_andps ((__m128) (__v4sf)
50			     { 0.0, v[step], v[step2], v[step3] }
51			     , u.xmmi);
52  if (n4)
53    {
54      int i;
55      for (i = 0; i < n4; ++i);
56      msk = (__m128) _mm_cmpeq_ps (m1, mm);
57      mim = _mm_or_si128 (_mm_and_si128 (msk, mi), mim);
58    }
59  ui.xmmi = (__m128) mim;
60  return ui.si[n];
61}
62
63static void
64sse_swap_rowf (float *r1, float *r2, int n)
65{
66  int n4 = (n / 4) * 4;
67  float *r14end = r1 + n4;
68  while (r1 < r14end)
69    {
70      *r1 = *r2;
71      r1++;
72    }
73}
74
75void
76ludcompf (float *m, int nw, int *prow, int n)
77{
78  int i, s = 0;
79  float *pm;
80  for (i = 0, pm = m; i < n - 1; ++i, pm += nw)
81    {
82      int vi = sse_max_abs_indexf (pm + i, nw, n - i);
83      float *pt;
84      int j;
85      if (vi != 0)
86	{
87	  sse_swap_rowf (pm, pm + vi * nw, nw);
88	  swap_index (prow, i, i + vi);
89	}
90      for (j = i + 1, pt = pm + nw; j < n; ++j, pt += nw)
91	sse_add_rowf (pt + i + 1, pm + i + 1, -1.0, n - i - 1);
92    }
93}
94