Currently, floating point to integral conversion requires special handling to support NaN and +/- Infinity.

 * If src is NaN, the result is 0.
 * If the src is negative infinity or any value less than or equal to the value of Integer.MIN_VALUE,
 * The result is equal to the value of Integer.MIN_VALUE.
 * If the src is positive infinity or any value greater than or equal to the value of Integer.MAX_VALUE,
 * The result is equal to the value of Integer.MAX_VALUE.

AVX10.2 adds new staturing conversion instructions "vcvttss2sis", "vcvttsd2sis", "vcvttps2dqs", "vcvttpd2qqs" which directly comply with Java and IEEE 754 semantics.


#include <stdio.h>
#include <math.h>

// compile command : clang -mavx10.2-256 -DAVX10 saturating-f2i.
// emulation command line : sde -dmr -itrace-execute-emulate -icount -ptr_raise -- ./a.out
// tags : arm parity, optimization over legacy CVTTSS2SI + SPECIAL-HANDLING.

void micro_f2i_sat_new(float src) {
  int dst = 0;
  asm volatile (
      "vcvttss2sis %1, %0 \n\t"
    : "=r"(dst)
    : "x"(src)
    :
  );
  printf("src = %f , dst = %d\n", src, dst);
}
int main() {
  printf("F2I SATURATING NEW\n");
  micro_f2i_sat_new(10.2f);
  micro_f2i_sat_new(INFINITY);
  micro_f2i_sat_new(-INFINITY);
  micro_f2i_sat_new(0x1.fffp+35f);
  micro_f2i_sat_new(0.0f);
  micro_f2i_sat_new(-0.0f);
  micro_f2i_sat_new(NAN);
  return 0;
}