Intel ITANIUM ARCHITECTURE - SOFTWARE DEVELOPERS VOLUME 3 REV 2.3, Manual

Page: 1027 / 1898

3:128 Volume 3: Instruction Reference
frcpa
frcpa — Floating-point Reciprocal Approximation
Format: (
qp
) frcpa.
sf f
1
,
p
2
=
f
2
,
f
3
F6
Description:
If PR
qp
is 0, PR
p
2
is cleared and FR
f
1
remains unchanged.
If PR
qp
is 1, the following will occur:
• FR
f
1
is either set to an approximation (with a relative error < 2
-8.886
) of the
reciprocal of FR
f
3
, or to the IEEE-754 mandated quotient of FR
f
2
/FR
f
3
— if either
FR
f
2
or FR
f
3
is in the set {-Infinity, -0, Pseudo-zero, +0, +Infinity, NaN,
Unsupported}.
• If FR
f
1
is set to the approximation of the reciprocal of FR
f
3
, then PR
p
2
is set to 1;
otherwise, it is set to 0.
• If FR
f
2
and FR
f
3
are such that the approximation of FR
f
3
’s reciprocal may cause the
Newton-Raphson iterations to fail to produce the correct IEEE-754 result of FR
f
2
/FR
f
3
, then a Floating-point Exception fault for Software Assist occurs.
System software is expected to compute the IEEE-754 quotient (FR
f
2
/FR
f
3
), return
the result in FR
f
1
, and set PR
p
2
to 0.
• If either FR
f
2
or FR
f
3
is a NaTVal, FR
f
1
is set to NaTVal instead of the computed
result, and PR
p
2
is cleared.
The mnemonic values for
sf
are given in Table 2-23 on page 3:56 .
Operation: if (PR[ qp ]) {
fp_check_target_register( f
1
);
if (tmp_isrcode = fp_reg_disabled( f
1
, f
2
, f
3
, 0))
disabled_fp_register_fault(tmp_isrcode, 0);
if (fp_is_natval(FR[ f
2
]) || fp_is_natval(FR[ f
3
])) {
FR[ f
1
] = NATVAL;
PR[ p
2
] = 0;
} else {
tmp_default_result = frcpa_exception_fault_check( f
2
, f
3
, sf ,
&tmp_fp_env);
if (fp_raise_fault(tmp_fp_env))
fp_exception_fault(fp_decode_fault(tmp_fp_env));
if (fp_is_nan_or_inf(tmp_default_result)) {
FR[ f
1
] = tmp_default_result;
PR[ p
2
] = 0;
} else {
num = fp_normalize(fp_reg_read(FR[ f
2
]));
den = fp_normalize(fp_reg_read(FR[ f
3
]));
if (fp_is_inf(num) && fp_is_finite(den)) {
FR[ f
1
] = FP_INFINITY;
FR[ f
1
].sign = num.sign ^ den.sign;
PR[ p
2
] = 0;
} else if (fp_is_finite(num) && fp_is_inf(den)) {
FR[ f
1
] = FP_ZERO;
FR[ f
1
].sign = num.sign ^ den.sign;
PR[ p
2
] = 0;
} else if (fp_is_zero(num) && fp_is_finite(den)) {
FR[ f
1
] = FP_ZERO;
FR[ f
1
].sign = num.sign ^ den.sign;
PR[ p
2
] = 0;