aha/arch/m68k/math-emu/fp_arith.c
Linus Torvalds 1da177e4c3 Linux-2.6.12-rc2
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!
2005-04-16 15:20:36 -07:00

701 lines
14 KiB
C

/*
fp_arith.c: floating-point math routines for the Linux-m68k
floating point emulator.
Copyright (c) 1998-1999 David Huggins-Daines.
Somewhat based on the AlphaLinux floating point emulator, by David
Mosberger-Tang.
You may copy, modify, and redistribute this file under the terms of
the GNU General Public License, version 2, or any later version, at
your convenience.
*/
#include "fp_emu.h"
#include "multi_arith.h"
#include "fp_arith.h"
const struct fp_ext fp_QNaN =
{
.exp = 0x7fff,
.mant = { .m64 = ~0 }
};
const struct fp_ext fp_Inf =
{
.exp = 0x7fff,
};
/* let's start with the easy ones */
struct fp_ext *
fp_fabs(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fabs\n");
fp_monadic_check(dest, src);
dest->sign = 0;
return dest;
}
struct fp_ext *
fp_fneg(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fneg\n");
fp_monadic_check(dest, src);
dest->sign = !dest->sign;
return dest;
}
/* Now, the slightly harder ones */
/* fp_fadd: Implements the kernel of the FADD, FSADD, FDADD, FSUB,
FDSUB, and FCMP instructions. */
struct fp_ext *
fp_fadd(struct fp_ext *dest, struct fp_ext *src)
{
int diff;
dprint(PINSTR, "fadd\n");
fp_dyadic_check(dest, src);
if (IS_INF(dest)) {
/* infinity - infinity == NaN */
if (IS_INF(src) && (src->sign != dest->sign))
fp_set_nan(dest);
return dest;
}
if (IS_INF(src)) {
fp_copy_ext(dest, src);
return dest;
}
if (IS_ZERO(dest)) {
if (IS_ZERO(src)) {
if (src->sign != dest->sign) {
if (FPDATA->rnd == FPCR_ROUND_RM)
dest->sign = 1;
else
dest->sign = 0;
}
} else
fp_copy_ext(dest, src);
return dest;
}
dest->lowmant = src->lowmant = 0;
if ((diff = dest->exp - src->exp) > 0)
fp_denormalize(src, diff);
else if ((diff = -diff) > 0)
fp_denormalize(dest, diff);
if (dest->sign == src->sign) {
if (fp_addmant(dest, src))
if (!fp_addcarry(dest))
return dest;
} else {
if (dest->mant.m64 < src->mant.m64) {
fp_submant(dest, src, dest);
dest->sign = !dest->sign;
} else
fp_submant(dest, dest, src);
}
return dest;
}
/* fp_fsub: Implements the kernel of the FSUB, FSSUB, and FDSUB
instructions.
Remember that the arguments are in assembler-syntax order! */
struct fp_ext *
fp_fsub(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fsub ");
src->sign = !src->sign;
return fp_fadd(dest, src);
}
struct fp_ext *
fp_fcmp(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fcmp ");
FPDATA->temp[1] = *dest;
src->sign = !src->sign;
return fp_fadd(&FPDATA->temp[1], src);
}
struct fp_ext *
fp_ftst(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "ftst\n");
(void)dest;
return src;
}
struct fp_ext *
fp_fmul(struct fp_ext *dest, struct fp_ext *src)
{
union fp_mant128 temp;
int exp;
dprint(PINSTR, "fmul\n");
fp_dyadic_check(dest, src);
/* calculate the correct sign now, as it's necessary for infinities */
dest->sign = src->sign ^ dest->sign;
/* Handle infinities */
if (IS_INF(dest)) {
if (IS_ZERO(src))
fp_set_nan(dest);
return dest;
}
if (IS_INF(src)) {
if (IS_ZERO(dest))
fp_set_nan(dest);
else
fp_copy_ext(dest, src);
return dest;
}
/* Of course, as we all know, zero * anything = zero. You may
not have known that it might be a positive or negative
zero... */
if (IS_ZERO(dest) || IS_ZERO(src)) {
dest->exp = 0;
dest->mant.m64 = 0;
dest->lowmant = 0;
return dest;
}
exp = dest->exp + src->exp - 0x3ffe;
/* shift up the mantissa for denormalized numbers,
so that the highest bit is set, this makes the
shift of the result below easier */
if ((long)dest->mant.m32[0] >= 0)
exp -= fp_overnormalize(dest);
if ((long)src->mant.m32[0] >= 0)
exp -= fp_overnormalize(src);
/* now, do a 64-bit multiply with expansion */
fp_multiplymant(&temp, dest, src);
/* normalize it back to 64 bits and stuff it back into the
destination struct */
if ((long)temp.m32[0] > 0) {
exp--;
fp_putmant128(dest, &temp, 1);
} else
fp_putmant128(dest, &temp, 0);
if (exp >= 0x7fff) {
fp_set_ovrflw(dest);
return dest;
}
dest->exp = exp;
if (exp < 0) {
fp_set_sr(FPSR_EXC_UNFL);
fp_denormalize(dest, -exp);
}
return dest;
}
/* fp_fdiv: Implements the "kernel" of the FDIV, FSDIV, FDDIV and
FSGLDIV instructions.
Note that the order of the operands is counter-intuitive: instead
of src / dest, the result is actually dest / src. */
struct fp_ext *
fp_fdiv(struct fp_ext *dest, struct fp_ext *src)
{
union fp_mant128 temp;
int exp;
dprint(PINSTR, "fdiv\n");
fp_dyadic_check(dest, src);
/* calculate the correct sign now, as it's necessary for infinities */
dest->sign = src->sign ^ dest->sign;
/* Handle infinities */
if (IS_INF(dest)) {
/* infinity / infinity = NaN (quiet, as always) */
if (IS_INF(src))
fp_set_nan(dest);
/* infinity / anything else = infinity (with approprate sign) */
return dest;
}
if (IS_INF(src)) {
/* anything / infinity = zero (with appropriate sign) */
dest->exp = 0;
dest->mant.m64 = 0;
dest->lowmant = 0;
return dest;
}
/* zeroes */
if (IS_ZERO(dest)) {
/* zero / zero = NaN */
if (IS_ZERO(src))
fp_set_nan(dest);
/* zero / anything else = zero */
return dest;
}
if (IS_ZERO(src)) {
/* anything / zero = infinity (with appropriate sign) */
fp_set_sr(FPSR_EXC_DZ);
dest->exp = 0x7fff;
dest->mant.m64 = 0;
return dest;
}
exp = dest->exp - src->exp + 0x3fff;
/* shift up the mantissa for denormalized numbers,
so that the highest bit is set, this makes lots
of things below easier */
if ((long)dest->mant.m32[0] >= 0)
exp -= fp_overnormalize(dest);
if ((long)src->mant.m32[0] >= 0)
exp -= fp_overnormalize(src);
/* now, do the 64-bit divide */
fp_dividemant(&temp, dest, src);
/* normalize it back to 64 bits and stuff it back into the
destination struct */
if (!temp.m32[0]) {
exp--;
fp_putmant128(dest, &temp, 32);
} else
fp_putmant128(dest, &temp, 31);
if (exp >= 0x7fff) {
fp_set_ovrflw(dest);
return dest;
}
dest->exp = exp;
if (exp < 0) {
fp_set_sr(FPSR_EXC_UNFL);
fp_denormalize(dest, -exp);
}
return dest;
}
struct fp_ext *
fp_fsglmul(struct fp_ext *dest, struct fp_ext *src)
{
int exp;
dprint(PINSTR, "fsglmul\n");
fp_dyadic_check(dest, src);
/* calculate the correct sign now, as it's necessary for infinities */
dest->sign = src->sign ^ dest->sign;
/* Handle infinities */
if (IS_INF(dest)) {
if (IS_ZERO(src))
fp_set_nan(dest);
return dest;
}
if (IS_INF(src)) {
if (IS_ZERO(dest))
fp_set_nan(dest);
else
fp_copy_ext(dest, src);
return dest;
}
/* Of course, as we all know, zero * anything = zero. You may
not have known that it might be a positive or negative
zero... */
if (IS_ZERO(dest) || IS_ZERO(src)) {
dest->exp = 0;
dest->mant.m64 = 0;
dest->lowmant = 0;
return dest;
}
exp = dest->exp + src->exp - 0x3ffe;
/* do a 32-bit multiply */
fp_mul64(dest->mant.m32[0], dest->mant.m32[1],
dest->mant.m32[0] & 0xffffff00,
src->mant.m32[0] & 0xffffff00);
if (exp >= 0x7fff) {
fp_set_ovrflw(dest);
return dest;
}
dest->exp = exp;
if (exp < 0) {
fp_set_sr(FPSR_EXC_UNFL);
fp_denormalize(dest, -exp);
}
return dest;
}
struct fp_ext *
fp_fsgldiv(struct fp_ext *dest, struct fp_ext *src)
{
int exp;
unsigned long quot, rem;
dprint(PINSTR, "fsgldiv\n");
fp_dyadic_check(dest, src);
/* calculate the correct sign now, as it's necessary for infinities */
dest->sign = src->sign ^ dest->sign;
/* Handle infinities */
if (IS_INF(dest)) {
/* infinity / infinity = NaN (quiet, as always) */
if (IS_INF(src))
fp_set_nan(dest);
/* infinity / anything else = infinity (with approprate sign) */
return dest;
}
if (IS_INF(src)) {
/* anything / infinity = zero (with appropriate sign) */
dest->exp = 0;
dest->mant.m64 = 0;
dest->lowmant = 0;
return dest;
}
/* zeroes */
if (IS_ZERO(dest)) {
/* zero / zero = NaN */
if (IS_ZERO(src))
fp_set_nan(dest);
/* zero / anything else = zero */
return dest;
}
if (IS_ZERO(src)) {
/* anything / zero = infinity (with appropriate sign) */
fp_set_sr(FPSR_EXC_DZ);
dest->exp = 0x7fff;
dest->mant.m64 = 0;
return dest;
}
exp = dest->exp - src->exp + 0x3fff;
dest->mant.m32[0] &= 0xffffff00;
src->mant.m32[0] &= 0xffffff00;
/* do the 32-bit divide */
if (dest->mant.m32[0] >= src->mant.m32[0]) {
fp_sub64(dest->mant, src->mant);
fp_div64(quot, rem, dest->mant.m32[0], 0, src->mant.m32[0]);
dest->mant.m32[0] = 0x80000000 | (quot >> 1);
dest->mant.m32[1] = (quot & 1) | rem; /* only for rounding */
} else {
fp_div64(quot, rem, dest->mant.m32[0], 0, src->mant.m32[0]);
dest->mant.m32[0] = quot;
dest->mant.m32[1] = rem; /* only for rounding */
exp--;
}
if (exp >= 0x7fff) {
fp_set_ovrflw(dest);
return dest;
}
dest->exp = exp;
if (exp < 0) {
fp_set_sr(FPSR_EXC_UNFL);
fp_denormalize(dest, -exp);
}
return dest;
}
/* fp_roundint: Internal rounding function for use by several of these
emulated instructions.
This one rounds off the fractional part using the rounding mode
specified. */
static void fp_roundint(struct fp_ext *dest, int mode)
{
union fp_mant64 oldmant;
unsigned long mask;
if (!fp_normalize_ext(dest))
return;
/* infinities and zeroes */
if (IS_INF(dest) || IS_ZERO(dest))
return;
/* first truncate the lower bits */
oldmant = dest->mant;
switch (dest->exp) {
case 0 ... 0x3ffe:
dest->mant.m64 = 0;
break;
case 0x3fff ... 0x401e:
dest->mant.m32[0] &= 0xffffffffU << (0x401e - dest->exp);
dest->mant.m32[1] = 0;
if (oldmant.m64 == dest->mant.m64)
return;
break;
case 0x401f ... 0x403e:
dest->mant.m32[1] &= 0xffffffffU << (0x403e - dest->exp);
if (oldmant.m32[1] == dest->mant.m32[1])
return;
break;
default:
return;
}
fp_set_sr(FPSR_EXC_INEX2);
/* We might want to normalize upwards here... however, since
we know that this is only called on the output of fp_fdiv,
or with the input to fp_fint or fp_fintrz, and the inputs
to all these functions are either normal or denormalized
(no subnormals allowed!), there's really no need.
In the case of fp_fdiv, observe that 0x80000000 / 0xffff =
0xffff8000, and the same holds for 128-bit / 64-bit. (i.e. the
smallest possible normal dividend and the largest possible normal
divisor will still produce a normal quotient, therefore, (normal
<< 64) / normal is normal in all cases) */
switch (mode) {
case FPCR_ROUND_RN:
switch (dest->exp) {
case 0 ... 0x3ffd:
return;
case 0x3ffe:
/* As noted above, the input is always normal, so the
guard bit (bit 63) is always set. therefore, the
only case in which we will NOT round to 1.0 is when
the input is exactly 0.5. */
if (oldmant.m64 == (1ULL << 63))
return;
break;
case 0x3fff ... 0x401d:
mask = 1 << (0x401d - dest->exp);
if (!(oldmant.m32[0] & mask))
return;
if (oldmant.m32[0] & (mask << 1))
break;
if (!(oldmant.m32[0] << (dest->exp - 0x3ffd)) &&
!oldmant.m32[1])
return;
break;
case 0x401e:
if (!(oldmant.m32[1] >= 0))
return;
if (oldmant.m32[0] & 1)
break;
if (!(oldmant.m32[1] << 1))
return;
break;
case 0x401f ... 0x403d:
mask = 1 << (0x403d - dest->exp);
if (!(oldmant.m32[1] & mask))
return;
if (oldmant.m32[1] & (mask << 1))
break;
if (!(oldmant.m32[1] << (dest->exp - 0x401d)))
return;
break;
default:
return;
}
break;
case FPCR_ROUND_RZ:
return;
default:
if (dest->sign ^ (mode - FPCR_ROUND_RM))
break;
return;
}
switch (dest->exp) {
case 0 ... 0x3ffe:
dest->exp = 0x3fff;
dest->mant.m64 = 1ULL << 63;
break;
case 0x3fff ... 0x401e:
mask = 1 << (0x401e - dest->exp);
if (dest->mant.m32[0] += mask)
break;
dest->mant.m32[0] = 0x80000000;
dest->exp++;
break;
case 0x401f ... 0x403e:
mask = 1 << (0x403e - dest->exp);
if (dest->mant.m32[1] += mask)
break;
if (dest->mant.m32[0] += 1)
break;
dest->mant.m32[0] = 0x80000000;
dest->exp++;
break;
}
}
/* modrem_kernel: Implementation of the FREM and FMOD instructions
(which are exactly the same, except for the rounding used on the
intermediate value) */
static struct fp_ext *
modrem_kernel(struct fp_ext *dest, struct fp_ext *src, int mode)
{
struct fp_ext tmp;
fp_dyadic_check(dest, src);
/* Infinities and zeros */
if (IS_INF(dest) || IS_ZERO(src)) {
fp_set_nan(dest);
return dest;
}
if (IS_ZERO(dest) || IS_INF(src))
return dest;
/* FIXME: there is almost certainly a smarter way to do this */
fp_copy_ext(&tmp, dest);
fp_fdiv(&tmp, src); /* NOTE: src might be modified */
fp_roundint(&tmp, mode);
fp_fmul(&tmp, src);
fp_fsub(dest, &tmp);
/* set the quotient byte */
fp_set_quotient((dest->mant.m64 & 0x7f) | (dest->sign << 7));
return dest;
}
/* fp_fmod: Implements the kernel of the FMOD instruction.
Again, the argument order is backwards. The result, as defined in
the Motorola manuals, is:
fmod(src,dest) = (dest - (src * floor(dest / src))) */
struct fp_ext *
fp_fmod(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fmod\n");
return modrem_kernel(dest, src, FPCR_ROUND_RZ);
}
/* fp_frem: Implements the kernel of the FREM instruction.
frem(src,dest) = (dest - (src * round(dest / src)))
*/
struct fp_ext *
fp_frem(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "frem\n");
return modrem_kernel(dest, src, FPCR_ROUND_RN);
}
struct fp_ext *
fp_fint(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fint\n");
fp_copy_ext(dest, src);
fp_roundint(dest, FPDATA->rnd);
return dest;
}
struct fp_ext *
fp_fintrz(struct fp_ext *dest, struct fp_ext *src)
{
dprint(PINSTR, "fintrz\n");
fp_copy_ext(dest, src);
fp_roundint(dest, FPCR_ROUND_RZ);
return dest;
}
struct fp_ext *
fp_fscale(struct fp_ext *dest, struct fp_ext *src)
{
int scale, oldround;
dprint(PINSTR, "fscale\n");
fp_dyadic_check(dest, src);
/* Infinities */
if (IS_INF(src)) {
fp_set_nan(dest);
return dest;
}
if (IS_INF(dest))
return dest;
/* zeroes */
if (IS_ZERO(src) || IS_ZERO(dest))
return dest;
/* Source exponent out of range */
if (src->exp >= 0x400c) {
fp_set_ovrflw(dest);
return dest;
}
/* src must be rounded with round to zero. */
oldround = FPDATA->rnd;
FPDATA->rnd = FPCR_ROUND_RZ;
scale = fp_conv_ext2long(src);
FPDATA->rnd = oldround;
/* new exponent */
scale += dest->exp;
if (scale >= 0x7fff) {
fp_set_ovrflw(dest);
} else if (scale <= 0) {
fp_set_sr(FPSR_EXC_UNFL);
fp_denormalize(dest, -scale);
} else
dest->exp = scale;
return dest;
}