gcc/testsuite/gcc.target/i386/xop-vpermil2ps-256-1.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target xop } */
   3 /* { dg-options "-O2 -mxop" } */
   4
   5 #include "xop-check.h"
   6
   7 #include <x86intrin.h>
   8
   9 #ifndef ZERO_MATCH
  10 #define ZERO_MATCH  3
  11 #endif
  12
  13 static float
  14 select2sp(float *src1, float *src2, int sel)
  15 {
  16     float tmp;
  17
  18     if ((sel & 0x7) == 0) tmp = src1[0];
  19     if ((sel & 0x7) == 1) tmp = src1[1];
  20     if ((sel & 0x7) == 2) tmp = src1[2];
  21     if ((sel & 0x7) == 3) tmp = src1[3];
  22     if ((sel & 0x7) == 4) tmp = src2[0];
  23     if ((sel & 0x7) == 5) tmp = src2[1];
  24     if ((sel & 0x7) == 6) tmp = src2[2];
  25     if ((sel & 0x7) == 7) tmp = src2[3];
  26
  27     return tmp;
  28 }
  29 static float
  30 sel_and_condzerosp(float *src1, float *src2, int sel, int imm8)
  31 {
  32     float tmp;
  33
  34     tmp = select2sp(src1, src2, sel & 0x7);
  35
  36     if (((imm8 & 0x3) == 2) && ((sel & 0x8) == 0x8)) tmp = 0;
  37     if (((imm8 & 0x3) == 3) && ((sel & 0x8) == 0x0)) tmp = 0;
  38
  39     return tmp;
  40 }
  41
  42 void static
  43 xop_test ()
  44 {
  45     int i;
  46     union256  source1, source2, u;
  47     union256i_d source3;
  48     float s1[8]={1, 2, 3, 4, 5, 6, 7, 8};
  49     float s2[8]={9, 10, 11, 12, 13, 14, 15, 16};
  50     int   s3[8]={11, 2, 3, 15, 5, 12, 7, 8};
  51     float  e[8];
  52
  53     source1.x = _mm256_loadu_ps(s1);
  54     source2.x = _mm256_loadu_ps(s2);
  55     source3.x = _mm256_loadu_si256((__m256i*) s3);
  56     u.x = _mm256_permute2_ps(source1.x, source2.x, source3.x, ZERO_MATCH);
  57
  58     for (i = 0; i < 8; ++i) {
  59         e[i] = sel_and_condzerosp(&s1[i & 0x4], &s2[i & 0x4], s3[i] & 0xf, ZERO_MATCH & 0x3);
  60     }
  61
  62    if (check_union256(u, e))
  63      abort ();
  64 }