[PATCH v5 6/7] math: Remove the SVID error handling from tgammaf

[Wilco Dijkstra]

Hi Adhemerval,

> It improves latency for about 1.5% and throughput for about 2-4%.

@@ -89,30 +82,32 @@ __ieee754_gammaf_r (float x, int *signgamp)
         * z - 0x1.2788cfc6fb619p-1;
       double f = 1.0 / z + d;
       float r = f;
+
       uint64_t rt = asuint64 (f);
       if (((rt + 2) & 0xfffffff) < 4)
         {
-         for (unsigned i = 0; i < sizeof (tb) / sizeof (tb[0]); i++)
+         for (unsigned i = 0; i < array_length (tb); i++)
             if (t == tb[i].u)
-             return tb[i].f + tb[i].df;
+             {
+               r = tb[i].f + tb[i].df;
+               break;
+             }

It's not clear it is useful to assume the table has entries that can
overflow since it is fixed (and we know it doesn't).

         }
+      if (isinf (r))
+       errno = ERANGE;
+
       return r;

That's just:

  if (isinf (r)) return __math_oflowf (t >> 31);

Cheers,
Wilco