0

Arm Architecture Reference Manual for A-profile architecture (issue I.a) (emphasis added):

FIZ, bit [0]

When FEAT_AFP is implemented:

0b1 Denormalized single-precision and double-precision inputs to most floating-point instructions flushed to zero.

A simple question: for which floating-point instructions inputs are not flushed to zero under FIZ=1?

pmor
  • 5,392
  • 4
  • 17
  • 36
  • 2
    It says right below: "For more information, see Flushing denormalized numbers to zero on page A1-58 and the pseudocode of the floating-point instructions." That section gives a complete explanation. As in your other question, and for the same reasons, the principal exceptions are FNEG, FABS, FMIN, FMAX. – Nate Eldredge May 29 '23 at 16:57
  • Thanks! I've overlooked the "For more information, see ...". – pmor May 30 '23 at 12:42
  • As for the _inputs_ I see that the exceptions are FABS and FNEG, but not FMIN / FMAX. – pmor May 30 '23 at 12:46
  • Yes, you're right. FMAX/FMIN have the behavior that *outputs* are not flushed. – Nate Eldredge May 30 '23 at 17:10

1 Answers1

0

A simple question: for which floating-point instructions inputs are not flushed to zero under FIZ=1?

FNEG, FABS.

pmor
  • 5,392
  • 4
  • 17
  • 36