For Arm BFDOT and BFMMLA, we need a version of round-to-odd that overflows to infinity, instead of the max normal number. Cc: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20210525225817.400336-6-richard.henderson@linaro.org Reviewed-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
820 lines
22 KiB
C++
820 lines
22 KiB
C++
/*
|
|
* QEMU float support
|
|
*
|
|
* The code in this source file is derived from release 2a of the SoftFloat
|
|
* IEC/IEEE Floating-point Arithmetic Package. Those parts of the code (and
|
|
* some later contributions) are provided under that license, as detailed below.
|
|
* It has subsequently been modified by contributors to the QEMU Project,
|
|
* so some portions are provided under:
|
|
* the SoftFloat-2a license
|
|
* the BSD license
|
|
* GPL-v2-or-later
|
|
*
|
|
* Any future contributions to this file after December 1st 2014 will be
|
|
* taken to be licensed under the Softfloat-2a license unless specifically
|
|
* indicated otherwise.
|
|
*/
|
|
|
|
static void partsN(return_nan)(FloatPartsN *a, float_status *s)
|
|
{
|
|
switch (a->cls) {
|
|
case float_class_snan:
|
|
float_raise(float_flag_invalid, s);
|
|
if (s->default_nan_mode) {
|
|
parts_default_nan(a, s);
|
|
} else {
|
|
parts_silence_nan(a, s);
|
|
}
|
|
break;
|
|
case float_class_qnan:
|
|
if (s->default_nan_mode) {
|
|
parts_default_nan(a, s);
|
|
}
|
|
break;
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
}
|
|
|
|
static FloatPartsN *partsN(pick_nan)(FloatPartsN *a, FloatPartsN *b,
|
|
float_status *s)
|
|
{
|
|
if (is_snan(a->cls) || is_snan(b->cls)) {
|
|
float_raise(float_flag_invalid, s);
|
|
}
|
|
|
|
if (s->default_nan_mode) {
|
|
parts_default_nan(a, s);
|
|
} else {
|
|
int cmp = frac_cmp(a, b);
|
|
if (cmp == 0) {
|
|
cmp = a->sign < b->sign;
|
|
}
|
|
|
|
if (pickNaN(a->cls, b->cls, cmp > 0, s)) {
|
|
a = b;
|
|
}
|
|
if (is_snan(a->cls)) {
|
|
parts_silence_nan(a, s);
|
|
}
|
|
}
|
|
return a;
|
|
}
|
|
|
|
static FloatPartsN *partsN(pick_nan_muladd)(FloatPartsN *a, FloatPartsN *b,
|
|
FloatPartsN *c, float_status *s,
|
|
int ab_mask, int abc_mask)
|
|
{
|
|
int which;
|
|
|
|
if (unlikely(abc_mask & float_cmask_snan)) {
|
|
float_raise(float_flag_invalid, s);
|
|
}
|
|
|
|
which = pickNaNMulAdd(a->cls, b->cls, c->cls,
|
|
ab_mask == float_cmask_infzero, s);
|
|
|
|
if (s->default_nan_mode || which == 3) {
|
|
/*
|
|
* Note that this check is after pickNaNMulAdd so that function
|
|
* has an opportunity to set the Invalid flag for infzero.
|
|
*/
|
|
parts_default_nan(a, s);
|
|
return a;
|
|
}
|
|
|
|
switch (which) {
|
|
case 0:
|
|
break;
|
|
case 1:
|
|
a = b;
|
|
break;
|
|
case 2:
|
|
a = c;
|
|
break;
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
if (is_snan(a->cls)) {
|
|
parts_silence_nan(a, s);
|
|
}
|
|
return a;
|
|
}
|
|
|
|
/*
|
|
* Canonicalize the FloatParts structure. Determine the class,
|
|
* unbias the exponent, and normalize the fraction.
|
|
*/
|
|
static void partsN(canonicalize)(FloatPartsN *p, float_status *status,
|
|
const FloatFmt *fmt)
|
|
{
|
|
if (unlikely(p->exp == 0)) {
|
|
if (likely(frac_eqz(p))) {
|
|
p->cls = float_class_zero;
|
|
} else if (status->flush_inputs_to_zero) {
|
|
float_raise(float_flag_input_denormal, status);
|
|
p->cls = float_class_zero;
|
|
frac_clear(p);
|
|
} else {
|
|
int shift = frac_normalize(p);
|
|
p->cls = float_class_normal;
|
|
p->exp = fmt->frac_shift - fmt->exp_bias - shift + 1;
|
|
}
|
|
} else if (likely(p->exp < fmt->exp_max) || fmt->arm_althp) {
|
|
p->cls = float_class_normal;
|
|
p->exp -= fmt->exp_bias;
|
|
frac_shl(p, fmt->frac_shift);
|
|
p->frac_hi |= DECOMPOSED_IMPLICIT_BIT;
|
|
} else if (likely(frac_eqz(p))) {
|
|
p->cls = float_class_inf;
|
|
} else {
|
|
frac_shl(p, fmt->frac_shift);
|
|
p->cls = (parts_is_snan_frac(p->frac_hi, status)
|
|
? float_class_snan : float_class_qnan);
|
|
}
|
|
}
|
|
|
|
/*
|
|
* Round and uncanonicalize a floating-point number by parts. There
|
|
* are FRAC_SHIFT bits that may require rounding at the bottom of the
|
|
* fraction; these bits will be removed. The exponent will be biased
|
|
* by EXP_BIAS and must be bounded by [EXP_MAX-1, 0].
|
|
*/
|
|
static void partsN(uncanon)(FloatPartsN *p, float_status *s,
|
|
const FloatFmt *fmt)
|
|
{
|
|
const int exp_max = fmt->exp_max;
|
|
const int frac_shift = fmt->frac_shift;
|
|
const uint64_t frac_lsb = fmt->frac_lsb;
|
|
const uint64_t frac_lsbm1 = fmt->frac_lsbm1;
|
|
const uint64_t round_mask = fmt->round_mask;
|
|
const uint64_t roundeven_mask = fmt->roundeven_mask;
|
|
uint64_t inc;
|
|
bool overflow_norm;
|
|
int exp, flags = 0;
|
|
|
|
if (unlikely(p->cls != float_class_normal)) {
|
|
switch (p->cls) {
|
|
case float_class_zero:
|
|
p->exp = 0;
|
|
frac_clear(p);
|
|
return;
|
|
case float_class_inf:
|
|
g_assert(!fmt->arm_althp);
|
|
p->exp = fmt->exp_max;
|
|
frac_clear(p);
|
|
return;
|
|
case float_class_qnan:
|
|
case float_class_snan:
|
|
g_assert(!fmt->arm_althp);
|
|
p->exp = fmt->exp_max;
|
|
frac_shr(p, fmt->frac_shift);
|
|
return;
|
|
default:
|
|
break;
|
|
}
|
|
g_assert_not_reached();
|
|
}
|
|
|
|
overflow_norm = false;
|
|
switch (s->float_rounding_mode) {
|
|
case float_round_nearest_even:
|
|
inc = ((p->frac_lo & roundeven_mask) != frac_lsbm1 ? frac_lsbm1 : 0);
|
|
break;
|
|
case float_round_ties_away:
|
|
inc = frac_lsbm1;
|
|
break;
|
|
case float_round_to_zero:
|
|
overflow_norm = true;
|
|
inc = 0;
|
|
break;
|
|
case float_round_up:
|
|
inc = p->sign ? 0 : round_mask;
|
|
overflow_norm = p->sign;
|
|
break;
|
|
case float_round_down:
|
|
inc = p->sign ? round_mask : 0;
|
|
overflow_norm = !p->sign;
|
|
break;
|
|
case float_round_to_odd:
|
|
overflow_norm = true;
|
|
/* fall through */
|
|
case float_round_to_odd_inf:
|
|
inc = p->frac_lo & frac_lsb ? 0 : round_mask;
|
|
break;
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
|
|
exp = p->exp + fmt->exp_bias;
|
|
if (likely(exp > 0)) {
|
|
if (p->frac_lo & round_mask) {
|
|
flags |= float_flag_inexact;
|
|
if (frac_addi(p, p, inc)) {
|
|
frac_shr(p, 1);
|
|
p->frac_hi |= DECOMPOSED_IMPLICIT_BIT;
|
|
exp++;
|
|
}
|
|
}
|
|
frac_shr(p, frac_shift);
|
|
|
|
if (fmt->arm_althp) {
|
|
/* ARM Alt HP eschews Inf and NaN for a wider exponent. */
|
|
if (unlikely(exp > exp_max)) {
|
|
/* Overflow. Return the maximum normal. */
|
|
flags = float_flag_invalid;
|
|
exp = exp_max;
|
|
frac_allones(p);
|
|
}
|
|
} else if (unlikely(exp >= exp_max)) {
|
|
flags |= float_flag_overflow | float_flag_inexact;
|
|
if (overflow_norm) {
|
|
exp = exp_max - 1;
|
|
frac_allones(p);
|
|
} else {
|
|
p->cls = float_class_inf;
|
|
exp = exp_max;
|
|
frac_clear(p);
|
|
}
|
|
}
|
|
} else if (s->flush_to_zero) {
|
|
flags |= float_flag_output_denormal;
|
|
p->cls = float_class_zero;
|
|
exp = 0;
|
|
frac_clear(p);
|
|
} else {
|
|
bool is_tiny = s->tininess_before_rounding || exp < 0;
|
|
|
|
if (!is_tiny) {
|
|
FloatPartsN discard;
|
|
is_tiny = !frac_addi(&discard, p, inc);
|
|
}
|
|
|
|
frac_shrjam(p, 1 - exp);
|
|
|
|
if (p->frac_lo & round_mask) {
|
|
/* Need to recompute round-to-even/round-to-odd. */
|
|
switch (s->float_rounding_mode) {
|
|
case float_round_nearest_even:
|
|
inc = ((p->frac_lo & roundeven_mask) != frac_lsbm1
|
|
? frac_lsbm1 : 0);
|
|
break;
|
|
case float_round_to_odd:
|
|
case float_round_to_odd_inf:
|
|
inc = p->frac_lo & frac_lsb ? 0 : round_mask;
|
|
break;
|
|
default:
|
|
break;
|
|
}
|
|
flags |= float_flag_inexact;
|
|
frac_addi(p, p, inc);
|
|
}
|
|
|
|
exp = (p->frac_hi & DECOMPOSED_IMPLICIT_BIT) != 0;
|
|
frac_shr(p, frac_shift);
|
|
|
|
if (is_tiny && (flags & float_flag_inexact)) {
|
|
flags |= float_flag_underflow;
|
|
}
|
|
if (exp == 0 && frac_eqz(p)) {
|
|
p->cls = float_class_zero;
|
|
}
|
|
}
|
|
p->exp = exp;
|
|
float_raise(flags, s);
|
|
}
|
|
|
|
/*
|
|
* Returns the result of adding or subtracting the values of the
|
|
* floating-point values `a' and `b'. The operation is performed
|
|
* according to the IEC/IEEE Standard for Binary Floating-Point
|
|
* Arithmetic.
|
|
*/
|
|
static FloatPartsN *partsN(addsub)(FloatPartsN *a, FloatPartsN *b,
|
|
float_status *s, bool subtract)
|
|
{
|
|
bool b_sign = b->sign ^ subtract;
|
|
int ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
|
|
|
|
if (a->sign != b_sign) {
|
|
/* Subtraction */
|
|
if (likely(ab_mask == float_cmask_normal)) {
|
|
if (parts_sub_normal(a, b)) {
|
|
return a;
|
|
}
|
|
/* Subtract was exact, fall through to set sign. */
|
|
ab_mask = float_cmask_zero;
|
|
}
|
|
|
|
if (ab_mask == float_cmask_zero) {
|
|
a->sign = s->float_rounding_mode == float_round_down;
|
|
return a;
|
|
}
|
|
|
|
if (unlikely(ab_mask & float_cmask_anynan)) {
|
|
goto p_nan;
|
|
}
|
|
|
|
if (ab_mask & float_cmask_inf) {
|
|
if (a->cls != float_class_inf) {
|
|
/* N - Inf */
|
|
goto return_b;
|
|
}
|
|
if (b->cls != float_class_inf) {
|
|
/* Inf - N */
|
|
return a;
|
|
}
|
|
/* Inf - Inf */
|
|
float_raise(float_flag_invalid, s);
|
|
parts_default_nan(a, s);
|
|
return a;
|
|
}
|
|
} else {
|
|
/* Addition */
|
|
if (likely(ab_mask == float_cmask_normal)) {
|
|
parts_add_normal(a, b);
|
|
return a;
|
|
}
|
|
|
|
if (ab_mask == float_cmask_zero) {
|
|
return a;
|
|
}
|
|
|
|
if (unlikely(ab_mask & float_cmask_anynan)) {
|
|
goto p_nan;
|
|
}
|
|
|
|
if (ab_mask & float_cmask_inf) {
|
|
a->cls = float_class_inf;
|
|
return a;
|
|
}
|
|
}
|
|
|
|
if (b->cls == float_class_zero) {
|
|
g_assert(a->cls == float_class_normal);
|
|
return a;
|
|
}
|
|
|
|
g_assert(a->cls == float_class_zero);
|
|
g_assert(b->cls == float_class_normal);
|
|
return_b:
|
|
b->sign = b_sign;
|
|
return b;
|
|
|
|
p_nan:
|
|
return parts_pick_nan(a, b, s);
|
|
}
|
|
|
|
/*
|
|
* Returns the result of multiplying the floating-point values `a' and
|
|
* `b'. The operation is performed according to the IEC/IEEE Standard
|
|
* for Binary Floating-Point Arithmetic.
|
|
*/
|
|
static FloatPartsN *partsN(mul)(FloatPartsN *a, FloatPartsN *b,
|
|
float_status *s)
|
|
{
|
|
int ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
|
|
bool sign = a->sign ^ b->sign;
|
|
|
|
if (likely(ab_mask == float_cmask_normal)) {
|
|
FloatPartsW tmp;
|
|
|
|
frac_mulw(&tmp, a, b);
|
|
frac_truncjam(a, &tmp);
|
|
|
|
a->exp += b->exp + 1;
|
|
if (!(a->frac_hi & DECOMPOSED_IMPLICIT_BIT)) {
|
|
frac_add(a, a, a);
|
|
a->exp -= 1;
|
|
}
|
|
|
|
a->sign = sign;
|
|
return a;
|
|
}
|
|
|
|
/* Inf * Zero == NaN */
|
|
if (unlikely(ab_mask == float_cmask_infzero)) {
|
|
float_raise(float_flag_invalid, s);
|
|
parts_default_nan(a, s);
|
|
return a;
|
|
}
|
|
|
|
if (unlikely(ab_mask & float_cmask_anynan)) {
|
|
return parts_pick_nan(a, b, s);
|
|
}
|
|
|
|
/* Multiply by 0 or Inf */
|
|
if (ab_mask & float_cmask_inf) {
|
|
a->cls = float_class_inf;
|
|
a->sign = sign;
|
|
return a;
|
|
}
|
|
|
|
g_assert(ab_mask & float_cmask_zero);
|
|
a->cls = float_class_zero;
|
|
a->sign = sign;
|
|
return a;
|
|
}
|
|
|
|
/*
|
|
* Returns the result of multiplying the floating-point values `a' and
|
|
* `b' then adding 'c', with no intermediate rounding step after the
|
|
* multiplication. The operation is performed according to the
|
|
* IEC/IEEE Standard for Binary Floating-Point Arithmetic 754-2008.
|
|
* The flags argument allows the caller to select negation of the
|
|
* addend, the intermediate product, or the final result. (The
|
|
* difference between this and having the caller do a separate
|
|
* negation is that negating externally will flip the sign bit on NaNs.)
|
|
*
|
|
* Requires A and C extracted into a double-sized structure to provide the
|
|
* extra space for the widening multiply.
|
|
*/
|
|
static FloatPartsN *partsN(muladd)(FloatPartsN *a, FloatPartsN *b,
|
|
FloatPartsN *c, int flags, float_status *s)
|
|
{
|
|
int ab_mask, abc_mask;
|
|
FloatPartsW p_widen, c_widen;
|
|
|
|
ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
|
|
abc_mask = float_cmask(c->cls) | ab_mask;
|
|
|
|
/*
|
|
* It is implementation-defined whether the cases of (0,inf,qnan)
|
|
* and (inf,0,qnan) raise InvalidOperation or not (and what QNaN
|
|
* they return if they do), so we have to hand this information
|
|
* off to the target-specific pick-a-NaN routine.
|
|
*/
|
|
if (unlikely(abc_mask & float_cmask_anynan)) {
|
|
return parts_pick_nan_muladd(a, b, c, s, ab_mask, abc_mask);
|
|
}
|
|
|
|
if (flags & float_muladd_negate_c) {
|
|
c->sign ^= 1;
|
|
}
|
|
|
|
/* Compute the sign of the product into A. */
|
|
a->sign ^= b->sign;
|
|
if (flags & float_muladd_negate_product) {
|
|
a->sign ^= 1;
|
|
}
|
|
|
|
if (unlikely(ab_mask != float_cmask_normal)) {
|
|
if (unlikely(ab_mask == float_cmask_infzero)) {
|
|
goto d_nan;
|
|
}
|
|
|
|
if (ab_mask & float_cmask_inf) {
|
|
if (c->cls == float_class_inf && a->sign != c->sign) {
|
|
goto d_nan;
|
|
}
|
|
goto return_inf;
|
|
}
|
|
|
|
g_assert(ab_mask & float_cmask_zero);
|
|
if (c->cls == float_class_normal) {
|
|
*a = *c;
|
|
goto return_normal;
|
|
}
|
|
if (c->cls == float_class_zero) {
|
|
if (a->sign != c->sign) {
|
|
goto return_sub_zero;
|
|
}
|
|
goto return_zero;
|
|
}
|
|
g_assert(c->cls == float_class_inf);
|
|
}
|
|
|
|
if (unlikely(c->cls == float_class_inf)) {
|
|
a->sign = c->sign;
|
|
goto return_inf;
|
|
}
|
|
|
|
/* Perform the multiplication step. */
|
|
p_widen.sign = a->sign;
|
|
p_widen.exp = a->exp + b->exp + 1;
|
|
frac_mulw(&p_widen, a, b);
|
|
if (!(p_widen.frac_hi & DECOMPOSED_IMPLICIT_BIT)) {
|
|
frac_add(&p_widen, &p_widen, &p_widen);
|
|
p_widen.exp -= 1;
|
|
}
|
|
|
|
/* Perform the addition step. */
|
|
if (c->cls != float_class_zero) {
|
|
/* Zero-extend C to less significant bits. */
|
|
frac_widen(&c_widen, c);
|
|
c_widen.exp = c->exp;
|
|
|
|
if (a->sign == c->sign) {
|
|
parts_add_normal(&p_widen, &c_widen);
|
|
} else if (!parts_sub_normal(&p_widen, &c_widen)) {
|
|
goto return_sub_zero;
|
|
}
|
|
}
|
|
|
|
/* Narrow with sticky bit, for proper rounding later. */
|
|
frac_truncjam(a, &p_widen);
|
|
a->sign = p_widen.sign;
|
|
a->exp = p_widen.exp;
|
|
|
|
return_normal:
|
|
if (flags & float_muladd_halve_result) {
|
|
a->exp -= 1;
|
|
}
|
|
finish_sign:
|
|
if (flags & float_muladd_negate_result) {
|
|
a->sign ^= 1;
|
|
}
|
|
return a;
|
|
|
|
return_sub_zero:
|
|
a->sign = s->float_rounding_mode == float_round_down;
|
|
return_zero:
|
|
a->cls = float_class_zero;
|
|
goto finish_sign;
|
|
|
|
return_inf:
|
|
a->cls = float_class_inf;
|
|
goto finish_sign;
|
|
|
|
d_nan:
|
|
float_raise(float_flag_invalid, s);
|
|
parts_default_nan(a, s);
|
|
return a;
|
|
}
|
|
|
|
/*
|
|
* Returns the result of dividing the floating-point value `a' by the
|
|
* corresponding value `b'. The operation is performed according to
|
|
* the IEC/IEEE Standard for Binary Floating-Point Arithmetic.
|
|
*/
|
|
static FloatPartsN *partsN(div)(FloatPartsN *a, FloatPartsN *b,
|
|
float_status *s)
|
|
{
|
|
int ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
|
|
bool sign = a->sign ^ b->sign;
|
|
|
|
if (likely(ab_mask == float_cmask_normal)) {
|
|
a->sign = sign;
|
|
a->exp -= b->exp + frac_div(a, b);
|
|
return a;
|
|
}
|
|
|
|
/* 0/0 or Inf/Inf => NaN */
|
|
if (unlikely(ab_mask == float_cmask_zero) ||
|
|
unlikely(ab_mask == float_cmask_inf)) {
|
|
float_raise(float_flag_invalid, s);
|
|
parts_default_nan(a, s);
|
|
return a;
|
|
}
|
|
|
|
/* All the NaN cases */
|
|
if (unlikely(ab_mask & float_cmask_anynan)) {
|
|
return parts_pick_nan(a, b, s);
|
|
}
|
|
|
|
a->sign = sign;
|
|
|
|
/* Inf / X */
|
|
if (a->cls == float_class_inf) {
|
|
return a;
|
|
}
|
|
|
|
/* 0 / X */
|
|
if (a->cls == float_class_zero) {
|
|
return a;
|
|
}
|
|
|
|
/* X / Inf */
|
|
if (b->cls == float_class_inf) {
|
|
a->cls = float_class_zero;
|
|
return a;
|
|
}
|
|
|
|
/* X / 0 => Inf */
|
|
g_assert(b->cls == float_class_zero);
|
|
float_raise(float_flag_divbyzero, s);
|
|
a->cls = float_class_inf;
|
|
return a;
|
|
}
|
|
|
|
/*
|
|
* Rounds the floating-point value `a' to an integer, and returns the
|
|
* result as a floating-point value. The operation is performed
|
|
* according to the IEC/IEEE Standard for Binary Floating-Point
|
|
* Arithmetic.
|
|
*
|
|
* parts_round_to_int_normal is an internal helper function for
|
|
* normal numbers only, returning true for inexact but not directly
|
|
* raising float_flag_inexact.
|
|
*/
|
|
static bool partsN(round_to_int_normal)(FloatPartsN *a, FloatRoundMode rmode,
|
|
int scale, int frac_size)
|
|
{
|
|
uint64_t frac_lsb, frac_lsbm1, rnd_even_mask, rnd_mask, inc;
|
|
int shift_adj;
|
|
|
|
scale = MIN(MAX(scale, -0x10000), 0x10000);
|
|
a->exp += scale;
|
|
|
|
if (a->exp < 0) {
|
|
bool one;
|
|
|
|
/* All fractional */
|
|
switch (rmode) {
|
|
case float_round_nearest_even:
|
|
one = false;
|
|
if (a->exp == -1) {
|
|
FloatPartsN tmp;
|
|
/* Shift left one, discarding DECOMPOSED_IMPLICIT_BIT */
|
|
frac_add(&tmp, a, a);
|
|
/* Anything remaining means frac > 0.5. */
|
|
one = !frac_eqz(&tmp);
|
|
}
|
|
break;
|
|
case float_round_ties_away:
|
|
one = a->exp == -1;
|
|
break;
|
|
case float_round_to_zero:
|
|
one = false;
|
|
break;
|
|
case float_round_up:
|
|
one = !a->sign;
|
|
break;
|
|
case float_round_down:
|
|
one = a->sign;
|
|
break;
|
|
case float_round_to_odd:
|
|
one = true;
|
|
break;
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
|
|
frac_clear(a);
|
|
a->exp = 0;
|
|
if (one) {
|
|
a->frac_hi = DECOMPOSED_IMPLICIT_BIT;
|
|
} else {
|
|
a->cls = float_class_zero;
|
|
}
|
|
return true;
|
|
}
|
|
|
|
if (a->exp >= frac_size) {
|
|
/* All integral */
|
|
return false;
|
|
}
|
|
|
|
if (N > 64 && a->exp < N - 64) {
|
|
/*
|
|
* Rounding is not in the low word -- shift lsb to bit 2,
|
|
* which leaves room for sticky and rounding bit.
|
|
*/
|
|
shift_adj = (N - 1) - (a->exp + 2);
|
|
frac_shrjam(a, shift_adj);
|
|
frac_lsb = 1 << 2;
|
|
} else {
|
|
shift_adj = 0;
|
|
frac_lsb = DECOMPOSED_IMPLICIT_BIT >> (a->exp & 63);
|
|
}
|
|
|
|
frac_lsbm1 = frac_lsb >> 1;
|
|
rnd_mask = frac_lsb - 1;
|
|
rnd_even_mask = rnd_mask | frac_lsb;
|
|
|
|
if (!(a->frac_lo & rnd_mask)) {
|
|
/* Fractional bits already clear, undo the shift above. */
|
|
frac_shl(a, shift_adj);
|
|
return false;
|
|
}
|
|
|
|
switch (rmode) {
|
|
case float_round_nearest_even:
|
|
inc = ((a->frac_lo & rnd_even_mask) != frac_lsbm1 ? frac_lsbm1 : 0);
|
|
break;
|
|
case float_round_ties_away:
|
|
inc = frac_lsbm1;
|
|
break;
|
|
case float_round_to_zero:
|
|
inc = 0;
|
|
break;
|
|
case float_round_up:
|
|
inc = a->sign ? 0 : rnd_mask;
|
|
break;
|
|
case float_round_down:
|
|
inc = a->sign ? rnd_mask : 0;
|
|
break;
|
|
case float_round_to_odd:
|
|
inc = a->frac_lo & frac_lsb ? 0 : rnd_mask;
|
|
break;
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
|
|
if (shift_adj == 0) {
|
|
if (frac_addi(a, a, inc)) {
|
|
frac_shr(a, 1);
|
|
a->frac_hi |= DECOMPOSED_IMPLICIT_BIT;
|
|
a->exp++;
|
|
}
|
|
a->frac_lo &= ~rnd_mask;
|
|
} else {
|
|
frac_addi(a, a, inc);
|
|
a->frac_lo &= ~rnd_mask;
|
|
/* Be careful shifting back, not to overflow */
|
|
frac_shl(a, shift_adj - 1);
|
|
if (a->frac_hi & DECOMPOSED_IMPLICIT_BIT) {
|
|
a->exp++;
|
|
} else {
|
|
frac_add(a, a, a);
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
static void partsN(round_to_int)(FloatPartsN *a, FloatRoundMode rmode,
|
|
int scale, float_status *s,
|
|
const FloatFmt *fmt)
|
|
{
|
|
switch (a->cls) {
|
|
case float_class_qnan:
|
|
case float_class_snan:
|
|
parts_return_nan(a, s);
|
|
break;
|
|
case float_class_zero:
|
|
case float_class_inf:
|
|
break;
|
|
case float_class_normal:
|
|
if (parts_round_to_int_normal(a, rmode, scale, fmt->frac_size)) {
|
|
float_raise(float_flag_inexact, s);
|
|
}
|
|
break;
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
}
|
|
|
|
/*
|
|
* Returns the result of converting the floating-point value `a' to
|
|
* the two's complement integer format. The conversion is performed
|
|
* according to the IEC/IEEE Standard for Binary Floating-Point
|
|
* Arithmetic---which means in particular that the conversion is
|
|
* rounded according to the current rounding mode. If `a' is a NaN,
|
|
* the largest positive integer is returned. Otherwise, if the
|
|
* conversion overflows, the largest integer with the same sign as `a'
|
|
* is returned.
|
|
*/
|
|
static int64_t partsN(float_to_sint)(FloatPartsN *p, FloatRoundMode rmode,
|
|
int scale, int64_t min, int64_t max,
|
|
float_status *s)
|
|
{
|
|
int flags = 0;
|
|
uint64_t r;
|
|
|
|
switch (p->cls) {
|
|
case float_class_snan:
|
|
case float_class_qnan:
|
|
flags = float_flag_invalid;
|
|
r = max;
|
|
break;
|
|
|
|
case float_class_inf:
|
|
flags = float_flag_invalid;
|
|
r = p->sign ? min : max;
|
|
break;
|
|
|
|
case float_class_zero:
|
|
return 0;
|
|
|
|
case float_class_normal:
|
|
/* TODO: N - 2 is frac_size for rounding; could use input fmt. */
|
|
if (parts_round_to_int_normal(p, rmode, scale, N - 2)) {
|
|
flags = float_flag_inexact;
|
|
}
|
|
|
|
if (p->exp <= DECOMPOSED_BINARY_POINT) {
|
|
r = p->frac_hi >> (DECOMPOSED_BINARY_POINT - p->exp);
|
|
} else {
|
|
r = UINT64_MAX;
|
|
}
|
|
if (p->sign) {
|
|
if (r <= -(uint64_t)min) {
|
|
r = -r;
|
|
} else {
|
|
flags = float_flag_invalid;
|
|
r = min;
|
|
}
|
|
} else if (r > max) {
|
|
flags = float_flag_invalid;
|
|
r = max;
|
|
}
|
|
break;
|
|
|
|
default:
|
|
g_assert_not_reached();
|
|
}
|
|
|
|
float_raise(flags, s);
|
|
return r;
|
|
}
|