gcc/testsuite/gcc.dg/vect/pr115841.c

   1 /* { dg-do compile } */
   2 /* { dg-additional-options "-Ofast -fcommon -fvect-cost-model=dynamic --param vect-partial-vector-usage=1" } */
   3 /* { dg-additional-options "-mavx512vl" { target avx512vl } } */
   4
   5 /* To trigger the bug costing needs to determine that aligning the A170
   6    accesses with a prologue is good and there should be a vectorized
   7    epilogue with a smaller vector size, re-using the vector accumulator
   8    from the vectorized main loop that's statically known to execute
   9    but the epilogue loop is not.  */
  10
  11 static unsigned char xl[192];
  12 unsigned char A170[192*3];
  13
  14 void jerate (unsigned char *, unsigned char *);
  15 float foo (unsigned n)
  16 {
  17   jerate (xl, A170);
  18
  19   unsigned i = 32;
  20   int kr = 1;
  21   float sfn11s = 0.f;
  22   float sfn12s = 0.f;
  23   do
  24     {
  25       int krm1 = kr - 1;
  26       long j = krm1;
  27       float a = (*(float(*)[n])A170)[j];
  28       float b = (*(float(*)[n])xl)[j];
  29       float c = a * b;
  30       float d = c * 6.93149983882904052734375e-1f;
  31       float e = (*(float(*)[n])A170)[j+48];
  32       float f = (*(float(*)[n])A170)[j+96];
  33       float g = d * e;
  34       sfn11s = sfn11s + g;
  35       float h = f * d;
  36       sfn12s = sfn12s + h;
  37       kr++;
  38     }
  39   while (--i != 0);
  40   float tem = sfn11s + sfn12s;
  41   return tem;
  42 }