/* vax_fpa.c - VAX f_, d_, g_floating instructions | |
Copyright (c) 1998-2011, Robert M Supnik | |
Permission is hereby granted, free of charge, to any person obtaining a | |
copy of this software and associated documentation files (the "Software"), | |
to deal in the Software without restriction, including without limitation | |
the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
and/or sell copies of the Software, and to permit persons to whom the | |
Software is furnished to do so, subject to the following conditions: | |
The above copyright notice and this permission notice shall be included in | |
all copies or substantial portions of the Software. | |
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
ROBERT M SUPNIK BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER | |
IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN | |
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. | |
Except as contained in this notice, the name of Robert M Supnik shall not be | |
used in advertising or otherwise to promote the sale, use or other dealings | |
in this Software without prior written authorization from Robert M Supnik. | |
23-Mar-12 RMS Fixed missing arguments in 32b floating add (Mark Pizzolato) | |
15-Sep-11 RMS Fixed integer overflow bug in EMODx | |
Fixed POLYx normalizing before add mask bug | |
(Camiel Vanderhoeven) | |
28-May-08 RMS Inlined physical memory routines | |
16-May-06 RMS Fixed bug in 32b floating multiply routine | |
Fixed bug in 64b extended modulus routine | |
03-May-06 RMS Fixed POLYD, POLYG to clear R4, R5 | |
Fixed POLYD, POLYG to set R3 correctly | |
Fixed POLYD, POLYG to not exit prematurely if arg = 0 | |
Fixed POLYD, POLYG to do full 64b multiply | |
Fixed POLYF, POLYD, POLYG to remove truncation on add | |
Fixed POLYF, POLYD, POLYG to mask mul reslt to 31b/63b/63b | |
Fixed fp add routine to test for zero via fraction | |
to support "denormal" argument from POLYF, POLYD, POLYG | |
(Tim Stark) | |
27-Sep-05 RMS Fixed bug in 32b structure definitions (Jason Stevens) | |
30-Sep-04 RMS Comment and formating changes based on vax_octa.c | |
18-Apr-04 RMS Moved format definitions to vax_defs.h | |
19-Jun-03 RMS Simplified add algorithm | |
16-May-03 RMS Fixed bug in floating to integer convert overflow | |
Fixed multiple bugs in EMODx | |
Integrated 32b only code | |
05-Jul-02 RMS Changed internal routine names for C library conflict | |
17-Apr-02 RMS Fixed bug in EDIV zero quotient | |
This module contains the instruction simulators for | |
- 64 bit arithmetic (ASHQ, EMUL, EDIV) | |
- single precision floating point | |
- double precision floating point, D and G format | |
*/ | |
#include "vax_defs.h" | |
#include <setjmp.h> | |
extern int32 R[16]; | |
extern int32 PSL; | |
extern int32 p1; | |
extern jmp_buf save_env; | |
#if defined (USE_INT64) | |
#define M64 0xFFFFFFFFFFFFFFFF /* 64b */ | |
#define FD_FRACW (0xFFFF & ~(FD_EXP | FPSIGN)) | |
#define FD_FRACL (FD_FRACW | 0xFFFF0000) /* f/d fraction */ | |
#define G_FRACW (0xFFFF & ~(G_EXP | FPSIGN)) | |
#define G_FRACL (G_FRACW | 0xFFFF0000) /* g fraction */ | |
#define UNSCRAM(h,l) (((((t_uint64) (h)) << 48) & 0xFFFF000000000000) | \ | |
((((t_uint64) (h)) << 16) & 0x0000FFFF00000000) | \ | |
((((t_uint64) (l)) << 16) & 0x00000000FFFF0000) | \ | |
((((t_uint64) (l)) >> 16) & 0x000000000000FFFF)) | |
#define CONCAT(h,l) ((((t_uint64) (h)) << 32) | ((uint32) (l))) | |
typedef struct { | |
int32 sign; | |
int32 exp; | |
t_uint64 frac; | |
} UFP; | |
#define UF_NM 0x8000000000000000 /* normalized */ | |
#define UF_FRND 0x0000008000000000 /* F round */ | |
#define UF_DRND 0x0000000000000080 /* D round */ | |
#define UF_GRND 0x0000000000000400 /* G round */ | |
#define UF_V_NM 63 | |
#define UF_V_FDHI 40 | |
#define UF_V_FDLO (UF_V_FDHI - 32) | |
#define UF_V_GHI 43 | |
#define UF_V_GLO (UF_V_GHI - 32) | |
#define UF_GETFDHI(x) (int32) ((((x) >> (16 + UF_V_FDHI)) & FD_FRACW) | \ | |
(((x) >> (UF_V_FDHI - 16)) & ~0xFFFF)) | |
#define UF_GETFDLO(x) (int32) ((((x) >> (16 + UF_V_FDLO)) & 0xFFFF) | \ | |
(((x) << (16 - UF_V_FDLO)) & ~0xFFFF)) | |
#define UF_GETGHI(x) (int32) ((((x) >> (16 + UF_V_GHI)) & G_FRACW) | \ | |
(((x) >> (UF_V_GHI - 16)) & ~0xFFFF)) | |
#define UF_GETGLO(x) (int32) ((((x) >> (16 + UF_V_GLO)) & 0xFFFF) | \ | |
(((x) << (16 - UF_V_GLO)) & ~0xFFFF)) | |
void unpackf (int32 hi, UFP *a); | |
void unpackd (int32 hi, int32 lo, UFP *a); | |
void unpackg (int32 hi, int32 lo, UFP *a); | |
void norm (UFP *a); | |
int32 rpackfd (UFP *a, int32 *rh); | |
int32 rpackg (UFP *a, int32 *rh); | |
void vax_fadd (UFP *a, UFP *b, uint32 mhi, uint32 mlo); | |
void vax_fmul (UFP *a, UFP *b, t_bool qd, int32 bias, uint32 mhi, uint32 mlo); | |
void vax_fdiv (UFP *b, UFP *a, int32 prec, int32 bias); | |
void vax_fmod (UFP *a, int32 bias, int32 *intgr, int32 *flg); | |
/* Quadword arithmetic shift | |
opnd[0] = shift count (cnt.rb) | |
opnd[1:2] = source (src.rq) | |
opnd[3:4] = destination (dst.wq) | |
*/ | |
int32 op_ashq (int32 *opnd, int32 *rh, int32 *flg) | |
{ | |
t_int64 src, r; | |
int32 sc = opnd[0]; | |
src = CONCAT (opnd[2], opnd[1]); /* build src */ | |
if (sc & BSIGN) { /* right shift? */ | |
*flg = 0; /* no ovflo */ | |
sc = 0x100 - sc; /* |shift| */ | |
if (sc > 63) /* sc > 63? */ | |
r = (opnd[2] & LSIGN)? -1: 0; | |
else r = src >> sc; | |
} | |
else { | |
if (sc > 63) { /* left shift */ | |
r = 0; /* sc > 63? */ | |
*flg = (src != 0); /* ovflo test */ | |
} | |
else { | |
r = src << sc; /* do shift */ | |
*flg = (src != (r >> sc)); /* ovflo test */ | |
} | |
} | |
*rh = (int32) ((r >> 32) & LMASK); /* hi result */ | |
return ((int32) (r & LMASK)); /* lo result */ | |
} | |
/* Extended multiply subroutine */ | |
int32 op_emul (int32 mpy, int32 mpc, int32 *rh) | |
{ | |
t_int64 lmpy = mpy; | |
t_int64 lmpc = mpc; | |
lmpy = lmpy * lmpc; | |
*rh = (int32) ((lmpy >> 32) & LMASK); | |
return ((int32) (lmpy & LMASK)); | |
} | |
/* Extended divide | |
opnd[0] = divisor (non-zero) | |
opnd[1:2] = dividend | |
*/ | |
int32 op_ediv (int32 *opnd, int32 *rh, int32 *flg) | |
{ | |
t_int64 ldvd, ldvr; | |
int32 quo, rem; | |
*flg = CC_V; /* assume error */ | |
*rh = 0; | |
ldvr = ((opnd[0] & LSIGN)? -opnd[0]: opnd[0]) & LMASK; /* |divisor| */ | |
ldvd = CONCAT (opnd[2], opnd[1]); /* 64b dividend */ | |
if (opnd[2] & LSIGN) /* |dividend| */ | |
ldvd = -ldvd; | |
if (((ldvd >> 32) & LMASK) >= ldvr) /* divide work? */ | |
return opnd[1]; | |
quo = (int32) (ldvd / ldvr); /* do divide */ | |
rem = (int32) (ldvd % ldvr); | |
if ((opnd[0] ^ opnd[2]) & LSIGN) { /* result -? */ | |
quo = -quo; /* negate */ | |
if (quo && ((quo & LSIGN) == 0)) /* right sign? */ | |
return opnd[1]; | |
} | |
else if (quo & LSIGN) | |
return opnd[1]; | |
if (opnd[2] & LSIGN) /* sign of rem */ | |
rem = -rem; | |
*flg = 0; /* no overflow */ | |
*rh = rem & LMASK; /* set rem */ | |
return (quo & LMASK); /* return quo */ | |
} | |
/* Compare floating */ | |
int32 op_cmpfd (int32 h1, int32 l1, int32 h2, int32 l2) | |
{ | |
t_uint64 n1, n2; | |
if ((h1 & FD_EXP) == 0) { | |
if (h1 & FPSIGN) | |
RSVD_OPND_FAULT; | |
h1 = l1 = 0; | |
} | |
if ((h2 & FD_EXP) == 0) { | |
if (h2 & FPSIGN) | |
RSVD_OPND_FAULT; | |
h2 = l2 = 0; | |
} | |
if ((h1 ^ h2) & FPSIGN) | |
return ((h1 & FPSIGN)? CC_N: 0); | |
n1 = UNSCRAM (h1, l1); | |
n2 = UNSCRAM (h2, l2); | |
if (n1 == n2) | |
return CC_Z; | |
return (((n1 < n2) ^ ((h1 & FPSIGN) != 0))? CC_N: 0); | |
} | |
int32 op_cmpg (int32 h1, int32 l1, int32 h2, int32 l2) | |
{ | |
t_uint64 n1, n2; | |
if ((h1 & G_EXP) == 0) { | |
if (h1 & FPSIGN) | |
RSVD_OPND_FAULT; | |
h1 = l1 = 0; | |
} | |
if ((h2 & G_EXP) == 0) { | |
if (h2 & FPSIGN) | |
RSVD_OPND_FAULT; | |
h2 = l2 = 0; | |
} | |
if ((h1 ^ h2) & FPSIGN) | |
return ((h1 & FPSIGN)? CC_N: 0); | |
n1 = UNSCRAM (h1, l1); | |
n2 = UNSCRAM (h2, l2); | |
if (n1 == n2) | |
return CC_Z; | |
return (((n1 < n2) ^ ((h1 & FPSIGN) != 0))? CC_N: 0); | |
} | |
/* Integer to floating convert */ | |
int32 op_cvtifdg (int32 val, int32 *rh, int32 opc) | |
{ | |
UFP a; | |
if (val == 0) { | |
if (rh) | |
*rh = 0; | |
return 0; | |
} | |
if (val < 0) { | |
a.sign = FPSIGN; | |
val = - val; | |
} | |
else a.sign = 0; | |
a.exp = 32 + ((opc & 0x100)? G_BIAS: FD_BIAS); | |
a.frac = ((t_uint64) val) << (UF_V_NM - 31); | |
norm (&a); | |
if (opc & 0x100) | |
return rpackg (&a, rh); | |
return rpackfd (&a, rh); | |
} | |
/* Floating to integer convert */ | |
int32 op_cvtfdgi (int32 *opnd, int32 *flg, int32 opc) | |
{ | |
UFP a; | |
int32 lnt = opc & 03; | |
int32 ubexp; | |
static t_uint64 maxv[4] = { 0x7F, 0x7FFF, 0x7FFFFFFF, 0x7FFFFFFF }; | |
*flg = 0; | |
if (opc & 0x100) { | |
unpackg (opnd[0], opnd[1], &a); | |
ubexp = a.exp - G_BIAS; | |
} | |
else { | |
if (opc & 0x20) | |
unpackd (opnd[0], opnd[1], &a); | |
else unpackf (opnd[0], &a); | |
ubexp = a.exp - FD_BIAS; | |
} | |
if ((a.exp == 0) || (ubexp < 0)) | |
return 0; | |
if (ubexp <= UF_V_NM) { | |
a.frac = a.frac >> (UF_V_NM - ubexp); /* leave rnd bit */ | |
if ((opc & 03) == 03) /* if CVTR, round */ | |
a.frac = a.frac + 1; | |
a.frac = a.frac >> 1; /* now justified */ | |
if (a.frac > (maxv[lnt] + (a.sign? 1: 0))) | |
*flg = CC_V; | |
} | |
else { | |
*flg = CC_V; /* set overflow */ | |
if (ubexp > (UF_V_NM + 32)) | |
return 0; | |
a.frac = a.frac << (ubexp - UF_V_NM - 1); /* no rnd bit */ | |
} | |
return ((int32) ((a.sign? (a.frac ^ LMASK) + 1: a.frac) & LMASK)); | |
} | |
/* Extended modularize | |
One of three floating point instructions dropped from the architecture, | |
EMOD presents two sets of complications. First, it requires an extended | |
fraction multiply, with precise (and unusual) truncation conditions. | |
Second, it has two write operands, a dubious distinction it shares | |
with EDIV. | |
*/ | |
int32 op_emodf (int32 *opnd, int32 *intgr, int32 *flg) | |
{ | |
UFP a, b; | |
unpackf (opnd[0], &a); /* unpack operands */ | |
unpackf (opnd[2], &b); | |
a.frac = a.frac | (((t_uint64) opnd[1]) << 32); /* extend src1 */ | |
vax_fmul (&a, &b, 0, FD_BIAS, 0, LMASK); /* multiply */ | |
vax_fmod (&a, FD_BIAS, intgr, flg); /* sep int & frac */ | |
return rpackfd (&a, NULL); /* return frac */ | |
} | |
int32 op_emodd (int32 *opnd, int32 *flo, int32 *intgr, int32 *flg) | |
{ | |
UFP a, b; | |
unpackd (opnd[0], opnd[1], &a); /* unpack operands */ | |
unpackd (opnd[3], opnd[4], &b); | |
a.frac = a.frac | opnd[2]; /* extend src1 */ | |
vax_fmul (&a, &b, 1, FD_BIAS, 0, 0); /* multiply */ | |
vax_fmod (&a, FD_BIAS, intgr, flg); /* sep int & frac */ | |
return rpackfd (&a, flo); /* return frac */ | |
} | |
int32 op_emodg (int32 *opnd, int32 *flo, int32 *intgr, int32 *flg) | |
{ | |
UFP a, b; | |
unpackg (opnd[0], opnd[1], &a); /* unpack operands */ | |
unpackg (opnd[3], opnd[4], &b); | |
a.frac = a.frac | (opnd[2] >> 5); /* extend src1 */ | |
vax_fmul (&a, &b, 1, G_BIAS, 0, 0); /* multiply */ | |
vax_fmod (&a, G_BIAS, intgr, flg); /* sep int & frac */ | |
return rpackg (&a, flo); /* return frac */ | |
} | |
/* Unpacked floating point routines */ | |
void vax_fadd (UFP *a, UFP *b, uint32 mhi, uint32 mlo) | |
{ | |
int32 ediff; | |
UFP t; | |
t_uint64 mask = (((t_uint64) mhi) << 32) | ((t_uint64) mlo); | |
if (a->frac == 0) { /* s1 = 0? */ | |
*a = *b; | |
return; | |
} | |
if (b->frac == 0) /* s2 = 0? */ | |
return; | |
if ((a->exp < b->exp) || /* |s1| < |s2|? swap */ | |
((a->exp == b->exp) && (a->frac < b->frac))) { | |
t = *a; | |
*a = *b; | |
*b = t; | |
} | |
ediff = a->exp - b->exp; /* exp diff */ | |
if (a->sign ^ b->sign) { /* eff sub? */ | |
if (ediff) { /* exp diff? */ | |
if (ediff > 63) /* retain sticky */ | |
b->frac = M64; | |
else b->frac = ((-((t_int64) b->frac) >> ediff) | /* denormalize */ | |
(M64 << (64 - ediff))); /* preserve sign */ | |
a->frac = a->frac + b->frac; /* add frac */ | |
} | |
else a->frac = a->frac - b->frac; /* sub frac */ | |
a->frac = a->frac & ~mask; /* mask before norm */ | |
norm (a); /* normalize */ | |
} | |
else { | |
if (ediff > 63) /* add */ | |
b->frac = 0; | |
else if (ediff) /* denormalize */ | |
b->frac = b->frac >> ediff; | |
a->frac = a->frac + b->frac; /* add frac */ | |
if (a->frac < b->frac) { /* chk for carry */ | |
a->frac = UF_NM | (a->frac >> 1); /* shift in carry */ | |
a->exp = a->exp + 1; /* skip norm */ | |
} | |
a->frac = a->frac & ~mask; /* mask */ | |
} | |
return; | |
} | |
/* Floating multiply - 64b * 64b with cross products */ | |
void vax_fmul (UFP *a, UFP *b, t_bool qd, int32 bias, uint32 mhi, uint32 mlo) | |
{ | |
t_uint64 ah, bh, al, bl, rhi, rlo, rmid1, rmid2; | |
t_uint64 mask = (((t_uint64) mhi) << 32) | ((t_uint64) mlo); | |
if ((a->exp == 0) || (b->exp == 0)) { /* zero argument? */ | |
a->frac = a->sign = a->exp = 0; /* result is zero */ | |
return; | |
} | |
a->sign = a->sign ^ b->sign; /* sign of result */ | |
a->exp = a->exp + b->exp - bias; /* add exponents */ | |
ah = (a->frac >> 32) & LMASK; /* split operands */ | |
bh = (b->frac >> 32) & LMASK; /* into 32b chunks */ | |
rhi = ah * bh; /* high result */ | |
if (qd) { /* 64b needed? */ | |
al = a->frac & LMASK; | |
bl = b->frac & LMASK; | |
rmid1 = ah * bl; | |
rmid2 = al * bh; | |
rlo = al * bl; | |
rhi = rhi + ((rmid1 >> 32) & LMASK) + ((rmid2 >> 32) & LMASK); | |
rmid1 = rlo + (rmid1 << 32); /* add mid1 to lo */ | |
if (rmid1 < rlo) /* carry? incr hi */ | |
rhi = rhi + 1; | |
rmid2 = rmid1 + (rmid2 << 32); /* add mid2 to lo */ | |
if (rmid2 < rmid1) /* carry? incr hi */ | |
rhi = rhi + 1; | |
} | |
a->frac = rhi & ~mask; | |
norm (a); /* normalize */ | |
return; | |
} | |
/* Floating modulus - there are three cases | |
exp <= bias - integer is 0, fraction is input, | |
no overflow | |
bias < exp <= bias+64 - separate integer and fraction, | |
integer overflow may occur | |
bias+64 < exp - result is integer, fraction is 0 | |
integer overflow | |
*/ | |
void vax_fmod (UFP *a, int32 bias, int32 *intgr, int32 *flg) | |
{ | |
if (a->exp <= bias) /* 0 or <1? int = 0 */ | |
*intgr = *flg = 0; | |
else if (a->exp <= (bias + 64)) { /* in range [1,64]? */ | |
*intgr = (int32) (a->frac >> (64 - (a->exp - bias))); | |
if ((a->exp > (bias + 32)) || /* test ovflo */ | |
((a->exp == (bias + 32)) && | |
(((uint32) *intgr) > (a->sign? 0x80000000: 0x7FFFFFFF)))) | |
*flg = CC_V; | |
else *flg = 0; | |
if (a->sign) /* -? comp int */ | |
*intgr = -*intgr; | |
if (a->exp == (bias + 64)) /* special case 64 */ | |
a->frac = 0; | |
else a->frac = a->frac << (a->exp - bias); | |
a->exp = bias; | |
} | |
else { | |
if (a->exp < (bias + 96)) /* need left shift? */ | |
*intgr = (int32) (a->frac << (a->exp - bias - 64)); | |
else *intgr = 0; /* out of range */ | |
if (a->sign) | |
*intgr = -*intgr; | |
a->frac = a->sign = a->exp = 0; /* result 0 */ | |
*flg = CC_V; /* overflow */ | |
} | |
norm (a); /* normalize */ | |
return; | |
} | |
/* Floating divide | |
Needs to develop at least one rounding bit. Since the first | |
divide step can fail, caller should specify 2 more bits than | |
the precision of the fraction. | |
*/ | |
void vax_fdiv (UFP *a, UFP *b, int32 prec, int32 bias) | |
{ | |
int32 i; | |
t_uint64 quo = 0; | |
if (a->exp == 0) /* divr = 0? */ | |
FLT_DZRO_FAULT; | |
if (b->exp == 0) /* divd = 0? */ | |
return; | |
b->sign = b->sign ^ a->sign; /* result sign */ | |
b->exp = b->exp - a->exp + bias + 1; /* unbiased exp */ | |
a->frac = a->frac >> 1; /* allow 1 bit left */ | |
b->frac = b->frac >> 1; | |
for (i = 0; (i < prec) && b->frac; i++) { /* divide loop */ | |
quo = quo << 1; /* shift quo */ | |
if (b->frac >= a->frac) { /* div step ok? */ | |
b->frac = b->frac - a->frac; /* subtract */ | |
quo = quo + 1; /* quo bit = 1 */ | |
} | |
b->frac = b->frac << 1; /* shift divd */ | |
} | |
b->frac = quo << (UF_V_NM - i + 1); /* shift quo */ | |
norm (b); /* normalize */ | |
return; | |
} | |
/* Support routines */ | |
void unpackf (int32 hi, UFP *r) | |
{ | |
r->sign = hi & FPSIGN; /* get sign */ | |
r->exp = FD_GETEXP (hi); /* get exponent */ | |
if (r->exp == 0) { /* exp = 0? */ | |
if (r->sign) /* if -, rsvd op */ | |
RSVD_OPND_FAULT; | |
r->frac = 0; /* else 0 */ | |
return; | |
} | |
hi = (((hi & FD_FRACW) | FD_HB) << 16) | ((hi >> 16) & 0xFFFF); | |
r->frac = ((t_uint64) hi) << (32 + UF_V_FDLO); | |
return; | |
} | |
void unpackd (int32 hi, int32 lo, UFP *r) | |
{ | |
r->sign = hi & FPSIGN; /* get sign */ | |
r->exp = FD_GETEXP (hi); /* get exponent */ | |
if (r->exp == 0) { /* exp = 0? */ | |
if (r->sign) /* if -, rsvd op */ | |
RSVD_OPND_FAULT; | |
r->frac = 0; /* else 0 */ | |
return; | |
} | |
hi = (hi & FD_FRACL) | FD_HB; /* canonical form */ | |
r->frac = UNSCRAM (hi, lo) << UF_V_FDLO; /* guard bits */ | |
return; | |
} | |
void unpackg (int32 hi, int32 lo, UFP *r) | |
{ | |
r->sign = hi & FPSIGN; /* get sign */ | |
r->exp = G_GETEXP (hi); /* get exponent */ | |
if (r->exp == 0) { /* exp = 0? */ | |
if (r->sign) /* if -, rsvd op */ | |
RSVD_OPND_FAULT; | |
r->frac = 0; /* else 0 */ | |
return; | |
} | |
hi = (hi & G_FRACL) | G_HB; /* canonical form */ | |
r->frac = UNSCRAM (hi, lo) << UF_V_GLO; /* guard bits */ | |
return; | |
} | |
void norm (UFP *r) | |
{ | |
int32 i; | |
static t_uint64 normmask[5] = { | |
0xc000000000000000, 0xf000000000000000, 0xff00000000000000, | |
0xffff000000000000, 0xffffffff00000000 | |
}; | |
static int32 normtab[6] = { 1, 2, 4, 8, 16, 32}; | |
if (r->frac == 0) { /* if fraction = 0 */ | |
r->sign = r->exp = 0; /* result is 0 */ | |
return; | |
} | |
while ((r->frac & UF_NM) == 0) { /* normalized? */ | |
for (i = 0; i < 5; i++) { /* find first 1 */ | |
if (r->frac & normmask[i]) | |
break; | |
} | |
r->frac = r->frac << normtab[i]; /* shift frac */ | |
r->exp = r->exp - normtab[i]; /* decr exp */ | |
} | |
return; | |
} | |
int32 rpackfd (UFP *r, int32 *rh) | |
{ | |
if (rh) /* assume 0 */ | |
*rh = 0; | |
if (r->frac == 0) /* result 0? */ | |
return 0; | |
r->frac = r->frac + (rh? UF_DRND: UF_FRND); /* round */ | |
if ((r->frac & UF_NM) == 0) { /* carry out? */ | |
r->frac = r->frac >> 1; /* renormalize */ | |
r->exp = r->exp + 1; | |
} | |
if (r->exp > (int32) FD_M_EXP) /* ovflo? fault */ | |
FLT_OVFL_FAULT; | |
if (r->exp <= 0) { /* underflow? */ | |
if (PSL & PSW_FU) /* fault if fu */ | |
FLT_UNFL_FAULT; | |
return 0; /* else 0 */ | |
} | |
if (rh) /* get low */ | |
*rh = UF_GETFDLO (r->frac); | |
return r->sign | (r->exp << FD_V_EXP) | UF_GETFDHI (r->frac); | |
} | |
int32 rpackg (UFP *r, int32 *rh) | |
{ | |
*rh = 0; /* assume 0 */ | |
if (r->frac == 0) /* result 0? */ | |
return 0; | |
r->frac = r->frac + UF_GRND; /* round */ | |
if ((r->frac & UF_NM) == 0) { /* carry out? */ | |
r->frac = r->frac >> 1; /* renormalize */ | |
r->exp = r->exp + 1; | |
} | |
if (r->exp > (int32) G_M_EXP) /* ovflo? fault */ | |
FLT_OVFL_FAULT; | |
if (r->exp <= 0) { /* underflow? */ | |
if (PSL & PSW_FU) /* fault if fu */ | |
FLT_UNFL_FAULT; | |
return 0; /* else 0 */ | |
} | |
*rh = UF_GETGLO (r->frac); /* get low */ | |
return r->sign | (r->exp << G_V_EXP) | UF_GETGHI (r->frac); | |
} | |
#else /* 32b code */ | |
#define WORDSWAP(x) ((((x) & WMASK) << 16) | (((x) >> 16) & WMASK)) | |
typedef struct { | |
uint32 lo; | |
uint32 hi; | |
} UDP; | |
typedef struct { | |
int32 sign; | |
int32 exp; | |
UDP frac; | |
} UFP; | |
#define UF_NM_H 0x80000000 /* normalized */ | |
#define UF_FRND_H 0x00000080 /* F round */ | |
#define UF_FRND_L 0x00000000 | |
#define UF_DRND_H 0x00000000 /* D round */ | |
#define UF_DRND_L 0x00000080 | |
#define UF_GRND_H 0x00000000 /* G round */ | |
#define UF_GRND_L 0x00000400 | |
#define UF_V_NM 63 | |
void unpackf (uint32 hi, UFP *a); | |
void unpackd (uint32 hi, uint32 lo, UFP *a); | |
void unpackg (uint32 hi, uint32 lo, UFP *a); | |
void norm (UFP *a); | |
int32 rpackfd (UFP *a, int32 *rh); | |
int32 rpackg (UFP *a, int32 *rh); | |
void vax_fadd (UFP *a, UFP *b, uint32 mhi, uint32 mlo); | |
void vax_fmul (UFP *a, UFP *b, t_bool qd, int32 bias, uint32 mhi, uint32 mlo); | |
void vax_fmod (UFP *a, int32 bias, int32 *intgr, int32 *flg); | |
void vax_fdiv (UFP *b, UFP *a, int32 prec, int32 bias); | |
void dp_add (UDP *a, UDP *b); | |
void dp_inc (UDP *a); | |
void dp_sub (UDP *a, UDP *b); | |
void dp_imul (uint32 a, uint32 b, UDP *r); | |
void dp_lsh (UDP *a, uint32 sc); | |
void dp_rsh (UDP *a, uint32 sc); | |
void dp_rsh_s (UDP *a, uint32 sc, uint32 neg); | |
void dp_neg (UDP *a); | |
int32 dp_cmp (UDP *a, UDP *b); | |
/* Quadword arithmetic shift | |
opnd[0] = shift count (cnt.rb) | |
opnd[1:2] = source (src.rq) | |
opnd[3:4] = destination (dst.wq) | |
*/ | |
int32 op_ashq (int32 *opnd, int32 *rh, int32 *flg) | |
{ | |
UDP r, sr; | |
uint32 sc = opnd[0]; | |
r.lo = opnd[1]; /* get source */ | |
r.hi = opnd[2]; | |
*flg = 0; /* assume no ovflo */ | |
if (sc & BSIGN) /* right shift? */ | |
dp_rsh_s (&r, 0x100 - sc, r.hi & LSIGN); /* signed right */ | |
else { | |
dp_lsh (&r, sc); /* left shift */ | |
sr = r; /* copy result */ | |
dp_rsh_s (&sr, sc, sr.hi & LSIGN); /* signed right */ | |
if ((sr.hi != ((uint32) opnd[2])) || /* reshift != orig? */ | |
(sr.lo != ((uint32) opnd[1]))) *flg = 1; /* overflow */ | |
} | |
*rh = r.hi; /* hi result */ | |
return r.lo; /* lo result */ | |
} | |
/* Extended multiply subroutine */ | |
int32 op_emul (int32 mpy, int32 mpc, int32 *rh) | |
{ | |
UDP r; | |
int32 sign = mpy ^ mpc; /* sign of result */ | |
if (mpy & LSIGN) /* abs value */ | |
mpy = -mpy; | |
if (mpc & LSIGN) | |
mpc = -mpc; | |
dp_imul (mpy & LMASK, mpc & LMASK, &r); /* 32b * 32b -> 64b */ | |
if (sign & LSIGN) /* negative result? */ | |
dp_neg (&r); | |
*rh = r.hi; | |
return r.lo; | |
} | |
/* Extended divide | |
opnd[0] = divisor (non-zero) | |
opnd[1:2] = dividend | |
*/ | |
int32 op_ediv (int32 *opnd, int32 *rh, int32 *flg) | |
{ | |
UDP dvd; | |
uint32 i, dvr, quo; | |
dvr = opnd[0]; /* get divisor */ | |
dvd.lo = opnd[1]; /* get dividend */ | |
dvd.hi = opnd[2]; | |
*flg = CC_V; /* assume error */ | |
*rh = 0; | |
if (dvd.hi & LSIGN) /* |dividend| */ | |
dp_neg (&dvd); | |
if (dvr & LSIGN) /* |divisor| */ | |
dvr = NEG (dvr); | |
if (dvd.hi >= dvr) /* divide work? */ | |
return opnd[1]; | |
for (i = quo = 0; i < 32; i++) { /* 32 iterations */ | |
quo = quo << 1; /* shift quotient */ | |
dp_lsh (&dvd, 1); /* shift dividend */ | |
if (dvd.hi >= dvr) { /* step work? */ | |
dvd.hi = (dvd.hi - dvr) & LMASK; /* subtract dvr */ | |
quo = quo + 1; | |
} | |
} | |
if ((opnd[0] ^ opnd[2]) & LSIGN) { /* result -? */ | |
quo = NEG (quo); /* negate */ | |
if (quo && ((quo & LSIGN) == 0)) /* right sign? */ | |
return opnd[1]; | |
} | |
else if (quo & LSIGN) | |
return opnd[1]; | |
if (opnd[2] & LSIGN) /* sign of rem */ | |
*rh = NEG (dvd.hi); | |
else *rh = dvd.hi; | |
*flg = 0; /* no overflow */ | |
return quo; /* return quo */ | |
} | |
/* Compare floating */ | |
int32 op_cmpfd (int32 h1, int32 l1, int32 h2, int32 l2) | |
{ | |
UFP a, b; | |
int32 r; | |
unpackd (h1, l1, &a); | |
unpackd (h2, l2, &b); | |
if (a.sign != b.sign) | |
return (a.sign? CC_N: 0); | |
r = a.exp - b.exp; | |
if (r == 0) | |
r = dp_cmp (&a.frac, &b.frac); | |
if (r < 0) | |
return (a.sign? 0: CC_N); | |
if (r > 0) | |
return (a.sign? CC_N: 0); | |
return CC_Z; | |
} | |
int32 op_cmpg (int32 h1, int32 l1, int32 h2, int32 l2) | |
{ | |
UFP a, b; | |
int32 r; | |
unpackg (h1, l1, &a); | |
unpackg (h2, l2, &b); | |
if (a.sign != b.sign) | |
return (a.sign? CC_N: 0); | |
r = a.exp - b.exp; | |
if (r == 0) | |
r = dp_cmp (&a.frac, &b.frac); | |
if (r < 0) | |
return (a.sign? 0: CC_N); | |
if (r > 0) | |
return (a.sign? CC_N: 0); | |
return CC_Z; | |
} | |
/* Integer to floating convert */ | |
int32 op_cvtifdg (int32 val, int32 *rh, int32 opc) | |
{ | |
UFP a; | |
if (val == 0) { /* zero? */ | |
if (rh) *rh = 0; /* return true 0 */ | |
return 0; | |
} | |
if (val < 0) { /* negative? */ | |
a.sign = FPSIGN; /* sign = - */ | |
val = -val; | |
} | |
else a.sign = 0; /* else sign = + */ | |
a.exp = 32 + ((opc & 0x100)? G_BIAS: FD_BIAS); /* initial exp */ | |
a.frac.hi = val & LMASK; /* fraction */ | |
a.frac.lo = 0; | |
norm (&a); /* normalize */ | |
if (opc & 0x100) /* pack and return */ | |
return rpackg (&a, rh); | |
return rpackfd (&a, rh); | |
} | |
/* Floating to integer convert */ | |
int32 op_cvtfdgi (int32 *opnd, int32 *flg, int32 opc) | |
{ | |
UFP a; | |
int32 lnt = opc & 03; | |
int32 ubexp; | |
static uint32 maxv[4] = { 0x7F, 0x7FFF, 0x7FFFFFFF, 0x7FFFFFFF }; | |
*flg = 0; | |
if (opc & 0x100) { /* G? */ | |
unpackg (opnd[0], opnd[1], &a); /* unpack */ | |
ubexp = a.exp - G_BIAS; /* unbiased exp */ | |
} | |
else { | |
if (opc & 0x20) /* F or D */ | |
unpackd (opnd[0], opnd[1], &a); | |
else unpackf (opnd[0], &a); /* unpack */ | |
ubexp = a.exp - FD_BIAS; /* unbiased exp */ | |
} | |
if ((a.exp == 0) || (ubexp < 0)) /* true zero or frac? */ | |
return 0; | |
if (ubexp <= UF_V_NM) { /* exp in range? */ | |
dp_rsh (&a.frac, UF_V_NM - ubexp); /* leave rnd bit */ | |
if (lnt == 03) /* if CVTR, round */ | |
dp_inc (&a.frac); | |
dp_rsh (&a.frac, 1); /* now justified */ | |
if ((a.frac.hi != 0) || | |
(a.frac.lo > (maxv[lnt] + (a.sign? 1: 0)))) | |
*flg = CC_V; | |
} | |
else { | |
*flg = CC_V; /* always ovflo */ | |
if (ubexp > (UF_V_NM + 32)) /* in ext range? */ | |
return 0; | |
dp_lsh (&a.frac, ubexp - UF_V_NM - 1); /* no rnd bit */ | |
} | |
return (a.sign? NEG (a.frac.lo): a.frac.lo); /* return lo frac */ | |
} | |
/* Extended modularize | |
One of three floating point instructions dropped from the architecture, | |
EMOD presents two sets of complications. First, it requires an extended | |
fraction multiply, with precise (and unusual) truncation conditions. | |
Second, it has two write operands, a dubious distinction it shares | |
with EDIV. | |
*/ | |
int32 op_emodf (int32 *opnd, int32 *intgr, int32 *flg) | |
{ | |
UFP a, b; | |
unpackf (opnd[0], &a); /* unpack operands */ | |
unpackf (opnd[2], &b); | |
a.frac.hi = a.frac.hi | opnd[1]; /* extend src1 */ | |
vax_fmul (&a, &b, 0, FD_BIAS, 0, LMASK); /* multiply */ | |
vax_fmod (&a, FD_BIAS, intgr, flg); /* sep int & frac */ | |
return rpackfd (&a, NULL); /* return frac */ | |
} | |
int32 op_emodd (int32 *opnd, int32 *flo, int32 *intgr, int32 *flg) | |
{ | |
UFP a, b; | |
unpackd (opnd[0], opnd[1], &a); /* unpack operands */ | |
unpackd (opnd[3], opnd[4], &b); | |
a.frac.lo = a.frac.lo | opnd[2]; /* extend src1 */ | |
vax_fmul (&a, &b, 1, FD_BIAS, 0, 0); /* multiply */ | |
vax_fmod (&a, FD_BIAS, intgr, flg); /* sep int & frac */ | |
return rpackfd (&a, flo); /* return frac */ | |
} | |
int32 op_emodg (int32 *opnd, int32 *flo, int32 *intgr, int32 *flg) | |
{ | |
UFP a, b; | |
unpackg (opnd[0], opnd[1], &a); /* unpack operands */ | |
unpackg (opnd[3], opnd[4], &b); | |
a.frac.lo = a.frac.lo | (opnd[2] >> 5); /* extend src1 */ | |
vax_fmul (&a, &b, 1, G_BIAS, 0, 0); /* multiply */ | |
vax_fmod (&a, G_BIAS, intgr, flg); /* sep int & frac */ | |
return rpackg (&a, flo); /* return frac */ | |
} | |
/* Unpacked floating point routines */ | |
/* Floating add */ | |
void vax_fadd (UFP *a, UFP *b, uint32 mhi, uint32 mlo) | |
{ | |
int32 ediff; | |
UFP t; | |
if ((a->frac.hi == 0) && (a->frac.lo == 0)) { /* s1 = 0? */ | |
*a = *b; | |
return; | |
} | |
if ((b->frac.hi == 0) && (b->frac.lo == 0)) /* s2 = 0? */ | |
return; | |
if ((a->exp < b->exp) || /* |s1| < |s2|? swap */ | |
((a->exp == b->exp) && (dp_cmp (&a->frac, &b->frac) < 0))) { | |
t = *a; | |
*a = *b; | |
*b = t; | |
} | |
ediff = a->exp - b->exp; /* exp diff */ | |
if (a->sign ^ b->sign) { /* eff sub? */ | |
if (ediff) { /* exp diff? */ | |
dp_neg (&b->frac); /* negate fraction */ | |
dp_rsh_s (&b->frac, ediff, 1); /* signed right */ | |
dp_add (&a->frac, &b->frac); /* "add" frac */ | |
} | |
else dp_sub (&a->frac, &b->frac); /* a >= b */ | |
a->frac.hi = a->frac.hi & ~mhi; /* mask before norm */ | |
a->frac.lo = a->frac.lo & ~mlo; | |
norm (a); /* normalize */ | |
} | |
else { | |
if (ediff) /* add, denormalize */ | |
dp_rsh (&b->frac, ediff); | |
dp_add (&a->frac, &b->frac); /* add frac */ | |
if (dp_cmp (&a->frac, &b->frac) < 0) { /* chk for carry */ | |
dp_rsh (&a->frac, 1); /* renormalize */ | |
a->frac.hi = a->frac.hi | UF_NM_H; /* add norm bit */ | |
a->exp = a->exp + 1; /* skip norm */ | |
} | |
a->frac.hi = a->frac.hi & ~mhi; /* mask */ | |
a->frac.lo = a->frac.lo & ~mlo; | |
} | |
return; | |
} | |
/* Floating multiply - 64b * 64b with cross products */ | |
void vax_fmul (UFP *a, UFP *b, t_bool qd, int32 bias, uint32 mhi, uint32 mlo) | |
{ | |
UDP rhi, rlo, rmid1, rmid2; | |
if ((a->exp == 0) || (b->exp == 0)) { /* zero argument? */ | |
a->frac.hi = a->frac.lo = 0; /* result is zero */ | |
a->sign = a->exp = 0; | |
return; | |
} | |
a->sign = a->sign ^ b->sign; /* sign of result */ | |
a->exp = a->exp + b->exp - bias; /* add exponents */ | |
dp_imul (a->frac.hi, b->frac.hi, &rhi); /* high result */ | |
if (qd) { /* 64b needed? */ | |
dp_imul (a->frac.hi, b->frac.lo, &rmid1); /* cross products */ | |
dp_imul (a->frac.lo, b->frac.hi, &rmid2); | |
dp_imul (a->frac.lo, b->frac.lo, &rlo); /* low result */ | |
rhi.lo = (rhi.lo + rmid1.hi) & LMASK; /* add hi cross */ | |
if (rhi.lo < rmid1.hi) /* to low high res */ | |
rhi.hi = (rhi.hi + 1) & LMASK; | |
rhi.lo = (rhi.lo + rmid2.hi) & LMASK; | |
if (rhi.lo < rmid2.hi) | |
rhi.hi = (rhi.hi + 1) & LMASK; | |
rlo.hi = (rlo.hi + rmid1.lo) & LMASK; /* add mid1 to low res */ | |
if (rlo.hi < rmid1.lo) /* carry? incr high res */ | |
dp_inc (&rhi); | |
rlo.hi = (rlo.hi + rmid2.lo) & LMASK; /* add mid2 to low res */ | |
if (rlo.hi < rmid2.lo) /* carry? incr high res */ | |
dp_inc (&rhi); | |
} | |
a->frac.hi = rhi.hi & ~mhi; /* mask fraction */ | |
a->frac.lo = rhi.lo & ~mlo; | |
norm (a); /* normalize */ | |
return; | |
} | |
/* Floating modulus - there are three cases | |
exp <= bias - integer is 0, fraction is input, | |
no overflow | |
bias < exp <= bias+64 - separate integer and fraction, | |
integer overflow may occur | |
bias+64 < exp - result is integer, fraction is 0 | |
integer overflow | |
*/ | |
void vax_fmod (UFP *a, int32 bias, int32 *intgr, int32 *flg) | |
{ | |
UDP ifr; | |
if (a->exp <= bias) /* 0 or <1? int = 0 */ | |
*intgr = *flg = 0; | |
else if (a->exp <= (bias + 64)) { /* in range [1,64]? */ | |
ifr = a->frac; | |
dp_rsh (&ifr, 64 - (a->exp - bias)); /* separate integer */ | |
if ((a->exp > (bias + 32)) || /* test ovflo */ | |
((a->exp == (bias + 32)) && | |
(ifr.lo > (a->sign? 0x80000000: 0x7FFFFFFF)))) | |
*flg = CC_V; | |
else *flg = 0; | |
*intgr = ifr.lo; | |
if (a->sign) /* -? comp int */ | |
*intgr = -*intgr; | |
dp_lsh (&a->frac, a->exp - bias); /* excise integer */ | |
a->exp = bias; | |
} | |
else { | |
if (a->exp < (bias + 96)) { /* need left shift? */ | |
ifr = a->frac; | |
dp_lsh (&ifr, a->exp - bias - 64); | |
*intgr = ifr.lo; | |
} | |
else *intgr = 0; /* out of range */ | |
if (a->sign) | |
*intgr = -*intgr; | |
a->frac.hi = a->frac.lo = a->sign = a->exp = 0; /* result 0 */ | |
*flg = CC_V; /* overflow */ | |
} | |
norm (a); /* normalize */ | |
return; | |
} | |
/* Floating divide | |
Needs to develop at least one rounding bit. Since the first | |
divide step can fail, caller should specify 2 more bits than | |
the precision of the fraction. | |
*/ | |
void vax_fdiv (UFP *a, UFP *b, int32 prec, int32 bias) | |
{ | |
int32 i; | |
UDP quo = { 0, 0 }; | |
if (a->exp == 0) /* divr = 0? */ | |
FLT_DZRO_FAULT; | |
if (b->exp == 0) /* divd = 0? */ | |
return; | |
b->sign = b->sign ^ a->sign; /* result sign */ | |
b->exp = b->exp - a->exp + bias + 1; /* unbiased exp */ | |
dp_rsh (&a->frac, 1); /* allow 1 bit left */ | |
dp_rsh (&b->frac, 1); | |
for (i = 0; i < prec; i++) { /* divide loop */ | |
dp_lsh (&quo, 1); /* shift quo */ | |
if (dp_cmp (&b->frac, &a->frac) >= 0) { /* div step ok? */ | |
dp_sub (&b->frac, &a->frac); /* subtract */ | |
quo.lo = quo.lo + 1; /* quo bit = 1 */ | |
} | |
dp_lsh (&b->frac, 1); /* shift divd */ | |
} | |
dp_lsh (&quo, UF_V_NM - prec + 1); /* put in position */ | |
b->frac = quo; | |
norm (b); /* normalize */ | |
return; | |
} | |
/* Double precision integer routines */ | |
int32 dp_cmp (UDP *a, UDP *b) | |
{ | |
if (a->hi < b->hi) /* compare hi */ | |
return -1; | |
if (a->hi > b->hi) | |
return +1; | |
if (a->lo < b->lo) /* hi =, compare lo */ | |
return -1; | |
if (a->lo > b->lo) | |
return +1; | |
return 0; /* hi, lo equal */ | |
} | |
void dp_add (UDP *a, UDP *b) | |
{ | |
a->lo = (a->lo + b->lo) & LMASK; /* add lo */ | |
if (a->lo < b->lo) /* carry? */ | |
a->hi = a->hi + 1; | |
a->hi = (a->hi + b->hi) & LMASK; /* add hi */ | |
return; | |
} | |
void dp_inc (UDP *a) | |
{ | |
a->lo = (a->lo + 1) & LMASK; /* inc lo */ | |
if (a->lo == 0) /* carry? inc hi */ | |
a->hi = (a->hi + 1) & LMASK; | |
return; | |
} | |
void dp_sub (UDP *a, UDP *b) | |
{ | |
if (a->lo < b->lo) /* borrow? decr hi */ | |
a->hi = a->hi - 1; | |
a->lo = (a->lo - b->lo) & LMASK; /* sub lo */ | |
a->hi = (a->hi - b->hi) & LMASK; /* sub hi */ | |
return; | |
} | |
void dp_lsh (UDP *r, uint32 sc) | |
{ | |
if (sc > 63) /* > 63? result 0 */ | |
r->hi = r->lo = 0; | |
else if (sc > 31) { /* [32,63]? */ | |
r->hi = (r->lo << (sc - 32)) & LMASK; | |
r->lo = 0; | |
} | |
else if (sc != 0) { | |
r->hi = ((r->hi << sc) | (r->lo >> (32 - sc))) & LMASK; | |
r->lo = (r->lo << sc) & LMASK; | |
} | |
return; | |
} | |
void dp_rsh (UDP *r, uint32 sc) | |
{ | |
if (sc > 63) /* > 63? result 0 */ | |
r->hi = r->lo = 0; | |
else if (sc > 31) { /* [32,63]? */ | |
r->lo = (r->hi >> (sc - 32)) & LMASK; | |
r->hi = 0; | |
} | |
else if (sc != 0) { | |
r->lo = ((r->lo >> sc) | (r->hi << (32 - sc))) & LMASK; | |
r->hi = (r->hi >> sc) & LMASK; | |
} | |
return; | |
} | |
void dp_rsh_s (UDP *r, uint32 sc, uint32 neg) | |
{ | |
dp_rsh (r, sc); /* do unsigned right */ | |
if (neg && sc) { /* negative? */ | |
if (sc > 63) /* > 63? result -1 */ | |
r->hi = r->lo = LMASK; | |
else { | |
UDP ones = { LMASK, LMASK }; | |
dp_lsh (&ones, 64 - sc); /* shift ones */ | |
r->hi = r->hi | ones.hi; /* or into result */ | |
r->lo = r->lo | ones.lo; | |
} | |
} | |
return; | |
} | |
void dp_imul (uint32 a, uint32 b, UDP *r) | |
{ | |
uint32 ah, bh, al, bl, rhi, rlo, rmid1, rmid2; | |
if ((a == 0) || (b == 0)) { /* zero argument? */ | |
r->hi = r->lo = 0; /* result is zero */ | |
return; | |
} | |
ah = (a >> 16) & WMASK; /* split operands */ | |
bh = (b >> 16) & WMASK; /* into 16b chunks */ | |
al = a & WMASK; | |
bl = b & WMASK; | |
rhi = ah * bh; /* high result */ | |
rmid1 = ah * bl; | |
rmid2 = al * bh; | |
rlo = al * bl; | |
rhi = rhi + ((rmid1 >> 16) & WMASK) + ((rmid2 >> 16) & WMASK); | |
rmid1 = (rlo + (rmid1 << 16)) & LMASK; /* add mid1 to lo */ | |
if (rmid1 < rlo) /* carry? incr hi */ | |
rhi = rhi + 1; | |
rmid2 = (rmid1 + (rmid2 << 16)) & LMASK; /* add mid2 to to */ | |
if (rmid2 < rmid1) /* carry? incr hi */ | |
rhi = rhi + 1; | |
r->hi = rhi & LMASK; /* mask result */ | |
r->lo = rmid2; | |
return; | |
} | |
void dp_neg (UDP *r) | |
{ | |
r->lo = NEG (r->lo); | |
r->hi = (~r->hi + (r->lo == 0)) & LMASK; | |
return; | |
} | |
/* Support routines */ | |
void unpackf (uint32 hi, UFP *r) | |
{ | |
r->sign = hi & FPSIGN; /* get sign */ | |
r->exp = FD_GETEXP (hi); /* get exponent */ | |
if (r->exp == 0) { /* exp = 0? */ | |
if (r->sign) /* if -, rsvd op */ | |
RSVD_OPND_FAULT; | |
r->frac.hi = r->frac.lo = 0; /* else 0 */ | |
return; | |
} | |
r->frac.hi = WORDSWAP ((hi & ~(FPSIGN | FD_EXP)) | FD_HB); | |
r->frac.lo = 0; | |
dp_lsh (&r->frac, FD_GUARD); | |
return; | |
} | |
void unpackd (uint32 hi, uint32 lo, UFP *r) | |
{ | |
r->sign = hi & FPSIGN; /* get sign */ | |
r->exp = FD_GETEXP (hi); /* get exponent */ | |
if (r->exp == 0) { /* exp = 0? */ | |
if (r->sign) /* if -, rsvd op */ | |
RSVD_OPND_FAULT; | |
r->frac.hi = r->frac.lo = 0; /* else 0 */ | |
return; | |
} | |
r->frac.hi = WORDSWAP ((hi & ~(FPSIGN | FD_EXP)) | FD_HB); | |
r->frac.lo = WORDSWAP (lo); | |
dp_lsh (&r->frac, FD_GUARD); | |
return; | |
} | |
void unpackg (uint32 hi, uint32 lo, UFP *r) | |
{ | |
r->sign = hi & FPSIGN; /* get sign */ | |
r->exp = G_GETEXP (hi); /* get exponent */ | |
if (r->exp == 0) { /* exp = 0? */ | |
if (r->sign) /* if -, rsvd op */ | |
RSVD_OPND_FAULT; | |
r->frac.hi = r->frac.lo = 0; /* else 0 */ | |
return; | |
} | |
r->frac.hi = WORDSWAP ((hi & ~(FPSIGN | G_EXP)) | G_HB); | |
r->frac.lo = WORDSWAP (lo); | |
dp_lsh (&r->frac, G_GUARD); | |
return; | |
} | |
void norm (UFP *r) | |
{ | |
int32 i; | |
static uint32 normmask[5] = { | |
0xc0000000, 0xf0000000, 0xff000000, 0xffff0000, 0xffffffff | |
}; | |
static int32 normtab[6] = { 1, 2, 4, 8, 16, 32}; | |
if ((r->frac.hi == 0) && (r->frac.lo == 0)) { /* if fraction = 0 */ | |
r->sign = r->exp = 0; /* result is 0 */ | |
return; | |
} | |
while ((r->frac.hi & UF_NM_H) == 0) { /* normalized? */ | |
for (i = 0; i < 5; i++) { /* find first 1 */ | |
if (r->frac.hi & normmask[i]) | |
break; | |
} | |
dp_lsh (&r->frac, normtab[i]); /* shift frac */ | |
r->exp = r->exp - normtab[i]; /* decr exp */ | |
} | |
return; | |
} | |
int32 rpackfd (UFP *r, int32 *rh) | |
{ | |
static UDP f_round = { UF_FRND_L, UF_FRND_H }; | |
static UDP d_round = { UF_DRND_L, UF_DRND_H }; | |
if (rh) /* assume 0 */ | |
*rh = 0; | |
if ((r->frac.hi == 0) && (r->frac.lo == 0)) /* result 0? */ | |
return 0; | |
if (rh) /* round */ | |
dp_add (&r->frac, &d_round); | |
else dp_add (&r->frac, &f_round); | |
if ((r->frac.hi & UF_NM_H) == 0) { /* carry out? */ | |
dp_rsh (&r->frac, 1); /* renormalize */ | |
r->exp = r->exp + 1; | |
} | |
if (r->exp > (int32) FD_M_EXP) /* ovflo? fault */ | |
FLT_OVFL_FAULT; | |
if (r->exp <= 0) { /* underflow? */ | |
if (PSL & PSW_FU) /* fault if fu */ | |
FLT_UNFL_FAULT; | |
return 0; /* else 0 */ | |
} | |
dp_rsh (&r->frac, FD_GUARD); /* remove guard */ | |
if (rh) /* get low */ | |
*rh = WORDSWAP (r->frac.lo); | |
return r->sign | (r->exp << FD_V_EXP) | | |
(WORDSWAP (r->frac.hi) & ~(FD_HB | FPSIGN | FD_EXP)); | |
} | |
int32 rpackg (UFP *r, int32 *rh) | |
{ | |
static UDP g_round = { UF_GRND_L, UF_GRND_H }; | |
*rh = 0; /* assume 0 */ | |
if ((r->frac.hi == 0) && (r->frac.lo == 0)) /* result 0? */ | |
return 0; | |
dp_add (&r->frac, &g_round); /* round */ | |
if ((r->frac.hi & UF_NM_H) == 0) { /* carry out? */ | |
dp_rsh (&r->frac, 1); /* renormalize */ | |
r->exp = r->exp + 1; | |
} | |
if (r->exp > (int32) G_M_EXP) /* ovflo? fault */ | |
FLT_OVFL_FAULT; | |
if (r->exp <= 0) { /* underflow? */ | |
if (PSL & PSW_FU) /* fault if fu */ | |
FLT_UNFL_FAULT; | |
return 0; /* else 0 */ | |
} | |
dp_rsh (&r->frac, G_GUARD); /* remove guard */ | |
*rh = WORDSWAP (r->frac.lo); /* get low */ | |
return r->sign | (r->exp << G_V_EXP) | | |
(WORDSWAP (r->frac.hi) & ~(G_HB | FPSIGN | G_EXP)); | |
} | |
#endif | |
/* Floating point instructions */ | |
/* Move/test/move negated floating | |
Note that only the high 32b is processed. | |
If the high 32b is not zero, it is unchanged. | |
*/ | |
int32 op_movfd (int32 val) | |
{ | |
if (val & FD_EXP) | |
return val; | |
if (val & FPSIGN) | |
RSVD_OPND_FAULT; | |
return 0; | |
} | |
int32 op_mnegfd (int32 val) | |
{ | |
if (val & FD_EXP) | |
return (val ^ FPSIGN); | |
if (val & FPSIGN) | |
RSVD_OPND_FAULT; | |
return 0; | |
} | |
int32 op_movg (int32 val) | |
{ | |
if (val & G_EXP) | |
return val; | |
if (val & FPSIGN) | |
RSVD_OPND_FAULT; | |
return 0; | |
} | |
int32 op_mnegg (int32 val) | |
{ | |
if (val & G_EXP) | |
return (val ^ FPSIGN); | |
if (val & FPSIGN) | |
RSVD_OPND_FAULT; | |
return 0; | |
} | |
/* Floating to floating convert - F to D is essentially done with MOVFD */ | |
int32 op_cvtdf (int32 *opnd) | |
{ | |
UFP a; | |
unpackd (opnd[0], opnd[1], &a); | |
return rpackfd (&a, NULL); | |
} | |
int32 op_cvtfg (int32 *opnd, int32 *rh) | |
{ | |
UFP a; | |
unpackf (opnd[0], &a); | |
a.exp = a.exp - FD_BIAS + G_BIAS; | |
return rpackg (&a, rh); | |
} | |
int32 op_cvtgf (int32 *opnd) | |
{ | |
UFP a; | |
unpackg (opnd[0], opnd[1], &a); | |
a.exp = a.exp - G_BIAS + FD_BIAS; | |
return rpackfd (&a, NULL); | |
} | |
/* Floating add and subtract */ | |
int32 op_addf (int32 *opnd, t_bool sub) | |
{ | |
UFP a, b; | |
unpackf (opnd[0], &a); /* F format */ | |
unpackf (opnd[1], &b); | |
if (sub) /* sub? -s1 */ | |
a.sign = a.sign ^ FPSIGN; | |
vax_fadd (&a, &b, 0, 0); /* add fractions */ | |
return rpackfd (&a, NULL); | |
} | |
int32 op_addd (int32 *opnd, int32 *rh, t_bool sub) | |
{ | |
UFP a, b; | |
unpackd (opnd[0], opnd[1], &a); | |
unpackd (opnd[2], opnd[3], &b); | |
if (sub) /* sub? -s1 */ | |
a.sign = a.sign ^ FPSIGN; | |
vax_fadd (&a, &b, 0, 0); /* add fractions */ | |
return rpackfd (&a, rh); | |
} | |
int32 op_addg (int32 *opnd, int32 *rh, t_bool sub) | |
{ | |
UFP a, b; | |
unpackg (opnd[0], opnd[1], &a); | |
unpackg (opnd[2], opnd[3], &b); | |
if (sub) /* sub? -s1 */ | |
a.sign = a.sign ^ FPSIGN; | |
vax_fadd (&a, &b, 0, 0); /* add fractions */ | |
return rpackg (&a, rh); /* round and pack */ | |
} | |
/* Floating multiply */ | |
int32 op_mulf (int32 *opnd) | |
{ | |
UFP a, b; | |
unpackf (opnd[0], &a); /* F format */ | |
unpackf (opnd[1], &b); | |
vax_fmul (&a, &b, 0, FD_BIAS, 0, 0); /* do multiply */ | |
return rpackfd (&a, NULL); /* round and pack */ | |
} | |
int32 op_muld (int32 *opnd, int32 *rh) | |
{ | |
UFP a, b; | |
unpackd (opnd[0], opnd[1], &a); /* D format */ | |
unpackd (opnd[2], opnd[3], &b); | |
vax_fmul (&a, &b, 1, FD_BIAS, 0, 0); /* do multiply */ | |
return rpackfd (&a, rh); /* round and pack */ | |
} | |
int32 op_mulg (int32 *opnd, int32 *rh) | |
{ | |
UFP a, b; | |
unpackg (opnd[0], opnd[1], &a); /* G format */ | |
unpackg (opnd[2], opnd[3], &b); | |
vax_fmul (&a, &b, 1, G_BIAS, 0, 0); /* do multiply */ | |
return rpackg (&a, rh); /* round and pack */ | |
} | |
/* Floating divide */ | |
int32 op_divf (int32 *opnd) | |
{ | |
UFP a, b; | |
unpackf (opnd[0], &a); /* F format */ | |
unpackf (opnd[1], &b); | |
vax_fdiv (&a, &b, 26, FD_BIAS); /* do divide */ | |
return rpackfd (&b, NULL); /* round and pack */ | |
} | |
int32 op_divd (int32 *opnd, int32 *rh) | |
{ | |
UFP a, b; | |
unpackd (opnd[0], opnd[1], &a); /* D format */ | |
unpackd (opnd[2], opnd[3], &b); | |
vax_fdiv (&a, &b, 58, FD_BIAS); /* do divide */ | |
return rpackfd (&b, rh); /* round and pack */ | |
} | |
int32 op_divg (int32 *opnd, int32 *rh) | |
{ | |
UFP a, b; | |
unpackg (opnd[0], opnd[1], &a); /* G format */ | |
unpackg (opnd[2], opnd[3], &b); | |
vax_fdiv (&a, &b, 55, G_BIAS); /* do divide */ | |
return rpackg (&b, rh); /* round and pack */ | |
} | |
/* Polynomial evaluation | |
The most mis-implemented instruction in the VAX (probably here too). | |
POLY requires a precise combination of masking versus normalizing | |
to achieve the desired answer. In particular, the multiply step | |
is masked prior to normalization. In addition, negative small | |
fractions must not be treated as 0 during denorm. | |
*/ | |
void op_polyf (int32 *opnd, int32 acc) | |
{ | |
UFP r, a, c; | |
int32 deg = opnd[1]; | |
int32 ptr = opnd[2]; | |
int32 i, wd, res; | |
if (deg > 31) /* degree > 31? fault */ | |
RSVD_OPND_FAULT; | |
unpackf (opnd[0], &a); /* unpack arg */ | |
wd = Read (ptr, L_LONG, RD); /* get C0 */ | |
ptr = ptr + 4; | |
unpackf (wd, &r); /* unpack C0 */ | |
res = rpackfd (&r, NULL); /* first result */ | |
for (i = 0; i < deg; i++) { /* loop */ | |
unpackf (res, &r); /* unpack result */ | |
vax_fmul (&r, &a, 0, FD_BIAS, 1, LMASK); /* r = r * arg, mask */ | |
wd = Read (ptr, L_LONG, RD); /* get Cnext */ | |
ptr = ptr + 4; | |
unpackf (wd, &c); /* unpack Cnext */ | |
vax_fadd (&r, &c, 1, LMASK); /* r = r + Cnext */ | |
res = rpackfd (&r, NULL); /* round and pack */ | |
} | |
R[0] = res; | |
R[1] = R[2] = 0; | |
R[3] = ptr; | |
return; | |
} | |
void op_polyd (int32 *opnd, int32 acc) | |
{ | |
UFP r, a, c; | |
int32 deg = opnd[2]; | |
int32 ptr = opnd[3]; | |
int32 i, wd, wd1, res, resh; | |
if (deg > 31) /* degree > 31? fault */ | |
RSVD_OPND_FAULT; | |
unpackd (opnd[0], opnd[1], &a); /* unpack arg */ | |
wd = Read (ptr, L_LONG, RD); /* get C0 */ | |
wd1 = Read (ptr + 4, L_LONG, RD); | |
ptr = ptr + 8; | |
unpackd (wd, wd1, &r); /* unpack C0 */ | |
res = rpackfd (&r, &resh); /* first result */ | |
for (i = 0; i < deg; i++) { /* loop */ | |
unpackd (res, resh, &r); /* unpack result */ | |
vax_fmul (&r, &a, 1, FD_BIAS, 0, 1); /* r = r * arg, mask */ | |
wd = Read (ptr, L_LONG, RD); /* get Cnext */ | |
wd1 = Read (ptr + 4, L_LONG, RD); | |
ptr = ptr + 8; | |
unpackd (wd, wd1, &c); /* unpack Cnext */ | |
vax_fadd (&r, &c, 0, 1); /* r = r + Cnext */ | |
res = rpackfd (&r, &resh); /* round and pack */ | |
} | |
R[0] = res; | |
R[1] = resh; | |
R[2] = 0; | |
R[3] = ptr; | |
R[4] = 0; | |
R[5] = 0; | |
return; | |
} | |
void op_polyg (int32 *opnd, int32 acc) | |
{ | |
UFP r, a, c; | |
int32 deg = opnd[2]; | |
int32 ptr = opnd[3]; | |
int32 i, wd, wd1, res, resh; | |
if (deg > 31) /* degree > 31? fault */ | |
RSVD_OPND_FAULT; | |
unpackg (opnd[0], opnd[1], &a); /* unpack arg */ | |
wd = Read (ptr, L_LONG, RD); /* get C0 */ | |
wd1 = Read (ptr + 4, L_LONG, RD); | |
ptr = ptr + 8; | |
unpackg (wd, wd1, &r); /* unpack C0 */ | |
res = rpackg (&r, &resh); /* first result */ | |
for (i = 0; i < deg; i++) { /* loop */ | |
unpackg (res, resh, &r); /* unpack result */ | |
vax_fmul (&r, &a, 1, G_BIAS, 0, 1); /* r = r * arg */ | |
wd = Read (ptr, L_LONG, RD); /* get Cnext */ | |
wd1 = Read (ptr + 4, L_LONG, RD); | |
ptr = ptr + 8; | |
unpackg (wd, wd1, &c); /* unpack Cnext */ | |
vax_fadd (&r, &c, 0, 1); /* r = r + Cnext */ | |
res = rpackg (&r, &resh); /* round and pack */ | |
} | |
R[0] = res; | |
R[1] = resh; | |
R[2] = 0; | |
R[3] = ptr; | |
R[4] = 0; | |
R[5] = 0; | |
return; | |
} |