Commit 3fa46106 authored by Ken Thompson's avatar Ken Thompson

multi precision floating point

R=r
OCL=20185
CL=20185
parent 37bdd3c3
......@@ -9,6 +9,7 @@ LIB=\
HFILES=\
go.h\
mparith.h\
y.tab.h\
YFILES=\
......
......@@ -126,10 +126,28 @@ convlit(Node *n, Type *t)
}
if(isfloat[et]) {
// float to float
if(mpcmpfltflt(n->val.u.fval, minfltval[et]) < 0)
double d;
float f;
Mpflt *fv;
fv = n->val.u.fval;
if(mpcmpfltflt(fv, minfltval[et]) < 0)
goto bad2;
if(mpcmpfltflt(n->val.u.fval, maxfltval[et]) > 0)
if(mpcmpfltflt(fv, maxfltval[et]) > 0)
goto bad2;
// switch(et) {
// case TFLOAT64:
// d = mpgetflt(fv);
// mpmovecflt(fv, d);
// break;
//
// case TFLOAT32:
// d = mpgetflt(fv);
// f = d;
// d = f;
// mpmovecflt(fv, d);
// break;
// }
break;
}
goto bad1;
......
......@@ -76,18 +76,18 @@ struct Array
enum
{
Mpscale = 29, /* safely smaller than bits in a long */
Mpprec = 10, /* Mpscale*Mpprec is max number of bits */
Mpbase = 1L<<Mpscale,
Mpscale = 29, // safely smaller than bits in a long
Mpprec = 16, // Mpscale*Mpprec is max number of bits
Mpnorm = Mpprec - 1, // significant words in a normalized float
Mpbase = 1L << Mpscale,
Mpsign = Mpbase >> 1,
Mpmask = Mpbase -1,
Debug = 1,
Mpmask = Mpbase - 1,
Mpdebug = 0,
};
typedef struct Mpint Mpint;
struct Mpint
{
vlong val;
long a[Mpprec];
uchar neg;
uchar ovf;
......@@ -96,8 +96,8 @@ struct Mpint
typedef struct Mpflt Mpflt;
struct Mpflt
{
double val;
uchar ovf;
Mpint val;
short exp;
};
typedef struct Val Val;
......@@ -551,7 +551,9 @@ void mpmovecfix(Mpint *a, vlong v);
int mptestfix(Mpint *a);
void mpaddfixfix(Mpint *a, Mpint *b);
void mpmulfixfix(Mpint *a, Mpint *b);
void mpmulfract(Mpint *a, Mpint *b);
void mpdivmodfixfix(Mpint *q, Mpint *r, Mpint *n, Mpint *d);
void mpdivfract(Mpint *a, Mpint *b);
void mpnegfix(Mpint *a);
void mpandfixfix(Mpint *a, Mpint *b);
void mplshfixfix(Mpint *a, Mpint *b);
......@@ -560,7 +562,7 @@ void mprshfixfix(Mpint *a, Mpint *b);
void mpxorfixfix(Mpint *a, Mpint *b);
void mpcomfix(Mpint *a);
vlong mpgetfix(Mpint *a);
double mpgetfixflt(Mpint *a);
void mpshiftfix(Mpint *a, int s);
/*
* mparith3.c
......@@ -573,6 +575,8 @@ void mpmulfltflt(Mpflt *a, Mpflt *b);
void mpdivfltflt(Mpflt *a, Mpflt *b);
void mpnegflt(Mpflt *a);
double mpgetflt(Mpflt *a);
int Fconv(Fmt*);
void mpnorm(Mpflt *a);
/*
* subr.c
......
......@@ -53,6 +53,7 @@ mainlex(int argc, char *argv[])
fmtinstall('Z', Zconv); // escaped string
fmtinstall('L', Lconv); // line number
fmtinstall('B', Bconv); // big numbers
fmtinstall('F', Fconv); // big float numbers
fmtinstall('W', Wconv); // whatis numbers (Wlitint)
lexinit();
......@@ -786,7 +787,7 @@ caseout:
yylval.val.u.fval = mal(sizeof(*yylval.val.u.fval));
mpatoflt(yylval.val.u.fval, namebuf);
if(yylval.val.u.fval->ovf) {
if(yylval.val.u.fval->val.ovf) {
yyerror("overflow in float constant");
mpmovecflt(yylval.val.u.fval, 0.0);
}
......
......@@ -2,9 +2,7 @@
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
#include <u.h>
#include <errno.h>
#include "go.h"
#include "go.h"
/// uses arithmetic
......@@ -14,7 +12,7 @@ mpcmpfixflt(Mpint *a, Mpflt *b)
char buf[500];
Mpflt c;
sprint(buf, "%B", a);
snprint(buf, sizeof(buf), "%B", a);
mpatoflt(&c, buf);
return mpcmpfltflt(&c, b);
}
......@@ -25,7 +23,7 @@ mpcmpfltfix(Mpflt *a, Mpint *b)
char buf[500];
Mpflt c;
sprint(buf, "%B", b);
snprint(buf, sizeof(buf), "%B", b);
mpatoflt(&c, buf);
return mpcmpfltflt(a, &c);
}
......@@ -71,17 +69,17 @@ mpcmpfltc(Mpflt *b, double c)
void
mpsubfixfix(Mpint *a, Mpint *b)
{
mpnegfix(b);
mpnegfix(a);
mpaddfixfix(a, b);
mpnegfix(b);
mpnegfix(a);
}
void
mpsubfltflt(Mpflt *a, Mpflt *b)
{
mpnegflt(b);
mpnegflt(a);
mpaddfltflt(a, b);
mpnegflt(b);
mpnegflt(a);
}
void
......@@ -151,7 +149,9 @@ mpcomfix(Mpint *a)
void
mpmovefixflt(Mpflt *a, Mpint *b)
{
mpmovecflt(a, mpgetfixflt(b));
a->val = *b;
a->exp = 0;
mpnorm(a);
}
void
......@@ -172,25 +172,18 @@ mpmovefltflt(Mpflt *a, Mpflt *b)
*a = *b;
}
//
// power of ten
//
static double
tentab[] = { 1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9 };
static double
dppow10(int n)
static double tab[] = { 1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7 };
static void
mppow10flt(Mpflt *a, int p)
{
int i;
if(n < 0)
return 1.0/dppow10(-n);
if(n < nelem(tentab))
return tentab[n];
i = n/2;
return dppow10(i) * dppow10(n-i);
if(p < nelem(tab)) {
mpmovecflt(a, tab[p]);
return;
}
mppow10flt(a, p>>1);
mpmulfltflt(a, a);
if(p & 1)
mpmulcflt(a, 10);
}
//
......@@ -200,17 +193,9 @@ dppow10(int n)
void
mpatoflt(Mpflt *a, char *as)
{
Mpflt b;
int dp, c, f, ef, ex, zer;
char *s;
double f64;
/* until Mpflt is really mp, use strtod to get rounding right */
errno = 0;
f64 = strtod(as, &s);
mpmovecflt(a, f64);
if(errno != 0)
a->ovf = 1;
return;
s = as;
dp = 0; /* digits after decimal point */
......@@ -283,21 +268,28 @@ mpatoflt(Mpflt *a, char *as)
if(dp)
dp--;
if(mpcmpfltc(a, 0.0) != 0)
mpmulcflt(a, dppow10(ex-dp));
if(mpcmpfltc(a, 0.0) != 0) {
if(ex >= dp) {
mppow10flt(&b, ex-dp);
mpmulfltflt(a, &b);
} else {
mppow10flt(&b, dp-ex);
mpdivfltflt(a, &b);
}
}
if(f)
mpnegflt(a);
return;
bad:
warn("set ovf in mpatof: %s", as);
warn("set ovf in mpatof");
mpmovecflt(a, 0.0);
}
//
// fixed point input
// required syntax is [+-][0[x]]d*
//
//
void
mpatofix(Mpint *a, char *as)
{
......@@ -410,3 +402,17 @@ Bconv(Fmt *fp)
*--p = '-';
return fmtstrcpy(fp, p);
}
int
Fconv(Fmt *fp)
{
char buf[500];
Mpflt *fval;
fval = va_arg(fp->args, Mpflt*);
if(fval->exp >= 0)
snprint(buf, sizeof(buf), "(%B*2^%d)", &fval->val, fval->exp);
else
snprint(buf, sizeof(buf), "(%B/2^%d)", &fval->val, -fval->exp);
return fmtstrcpy(fp, buf);
}
......@@ -2,7 +2,7 @@
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
#include "go.h"
#include "go.h"
//
// return the significant
......@@ -159,6 +159,31 @@ mpneg(Mpint *a)
}
}
void
mpshiftfix(Mpint *a, int s)
{
if(s >= 0) {
while(s >= Mpscale) {
mplshw(a);
s -= Mpscale;
}
while(s > 0) {
mplsh(a);
s--;
}
} else {
s = -s;
while(s >= Mpscale) {
mprshw(a);
s -= Mpscale;
}
while(s > 0) {
mprsh(a);
s--;
}
}
}
/// implements fix arihmetic
void
......@@ -274,6 +299,45 @@ mpmulfixfix(Mpint *a, Mpint *b)
warn("set ovf in mpmulfixfix");
}
void
mpmulfract(Mpint *a, Mpint *b)
{
int i, j;
long *a1, x;
Mpint s, q;
if(a->ovf || b->ovf) {
warn("ovf in mpmulflt");
a->ovf = 1;
return;
}
mpmovefixfix(&s, b);
a1 = &a->a[Mpprec];
s.neg = 0;
mpmovecfix(&q, 0);
for(i=0; i<Mpprec; i++) {
x = *--a1;
if(x == 0) {
mprshw(&s);
continue;
}
for(j=0; j<Mpscale; j++) {
x <<= 1;
if(x & Mpbase)
mpaddfixfix(&q, &s);
mprsh(&s);
}
}
q.neg = a->neg ^ b->neg;
mpmovefixfix(a, &q);
if(a->ovf)
warn("set ovf in mpmulflt");
}
void
mporfixfix(Mpint *a, Mpint *b)
{
......@@ -394,14 +458,7 @@ mplshfixfix(Mpint *a, Mpint *b)
return;
}
while(s >= Mpscale) {
mplshw(a);
s -= Mpscale;
}
while(s > 0) {
mplsh(a);
s--;
}
mpshiftfix(a, s);
}
void
......@@ -425,14 +482,7 @@ mprshfixfix(Mpint *a, Mpint *b)
return;
}
while(s >= Mpscale) {
mprshw(a);
s -= Mpscale;
}
while(s > 0) {
mprsh(a);
s--;
}
mpshiftfix(a, -s);
}
void
......@@ -459,17 +509,6 @@ mpgetfix(Mpint *a)
return v;
}
double
mpgetfixflt(Mpint *a)
{
// answer might not fit in intermediate vlong, so format
// to string and then let the string routine convert.
char buf[1000];
snprint(buf, sizeof buf, "%B", a);
return strtod(buf, nil);
}
void
mpmovecfix(Mpint *a, vlong c)
{
......@@ -532,6 +571,36 @@ mpdivmodfixfix(Mpint *q, Mpint *r, Mpint *n, Mpint *d)
}
}
void
mpdivfract(Mpint *a, Mpint *b)
{
Mpint n, d;
int i, j, neg;
long *a1, x;
mpmovefixfix(&n, a); // numerator
mpmovefixfix(&d, b); // denominator
a1 = &a->a[Mpprec]; // quotient
neg = n.neg ^ d.neg;
n.neg = 0;
d.neg = 0;
for(i=0; i<Mpprec; i++) {
x = 0;
for(j=0; j<Mpscale; j++) {
x <<= 1;
if(mpcmp(&d, &n) <= 0) {
x |= 1;
mpsubfixfix(&n, &d);
}
mprsh(&d);
}
*--a1 = x;
}
a->neg = neg;
}
int
mptestfix(Mpint *a)
{
......
......@@ -2,52 +2,266 @@
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
#include "go.h"
#include "go.h"
/*
* returns the leading non-zero
* word of the number
*/
int
sigfig(Mpflt *a)
{
int i;
for(i=Mpprec-1; i>=0; i--)
if(a->val.a[i] != 0)
break;
//print("sigfig %d %d\n", i-z+1, z);
return i+1;
}
/*
* shifts the leading non-zero
* word of the number to Mpnorm
*/
void
mpnorm(Mpflt *a)
{
int s;
s = sigfig(a);
if(s == 0) {
// zero
a->exp = 0;
a->val.neg = 0;
return;
}
s = (Mpnorm-s) * Mpscale;
mpshiftfix(&a->val, s);
a->exp -= s;
}
/// implements float arihmetic
void
mpaddfltflt(Mpflt *a, Mpflt *b)
{
a->val += b->val;
int sa, sb, s;
Mpflt c;
if(Mpdebug)
print("\n%F + %F", a, b);
sa = sigfig(a);
sb = sigfig(b);
if(sa == 0) {
if(sb == 0) {
// zero
a->exp = 0;
a->val.neg = 0;
return;
}
mpmovefltflt(a, b);
goto out;
}
if(sb == 0)
goto out;
s = a->exp - b->exp;
if(s > 0) {
// a is larger, shift b right
mpmovefltflt(&c, b);
mpshiftfix(&c.val, -s);
mpaddfixfix(&a->val, &c.val);
goto out;
}
if(s < 0) {
// b is larger, shift a right
mpshiftfix(&a->val, s);
a->exp -= s;
mpaddfixfix(&a->val, &b->val);
goto out;
}
mpaddfixfix(&a->val, &b->val);
out:
mpnorm(a);
if(Mpdebug)
print(" = %F\n\n", a);
}
void
mpmulfltflt(Mpflt *a, Mpflt *b)
{
a->val *= b->val;
int sa, sb;
if(Mpdebug)
print("%F\n * %F\n", a, b);
sa = sigfig(a);
sb = sigfig(b);
if(sa == 0 || sb == 0) {
// zero
a->exp = 0;
a->val.neg = 0;
return;
}
mpmulfract(&a->val, &b->val);
a->exp = (a->exp + b->exp) + Mpscale*Mpprec - 1;
mpnorm(a);
if(Mpdebug)
print(" = %F\n\n", a);
}
void
mpdivfltflt(Mpflt *a, Mpflt *b)
{
a->val /= b->val;
int sa, sb;
Mpflt c;
if(Mpdebug)
print("%F\n / %F\n", a, b);
sa = sigfig(a);
sb = sigfig(b);
if(sb == 0) {
// zero and ovfl
a->exp = 0;
a->val.neg = 0;
a->val.ovf = 1;
warn("mpdivfltflt divide by zero");
return;
}
if(sa == 0) {
// zero
a->exp = 0;
a->val.neg = 0;
return;
}
// adjust b to top
mpmovefltflt(&c, b);
mpshiftfix(&c.val, Mpscale);
// divide
mpdivfract(&a->val, &c.val);
a->exp = (a->exp-c.exp) - Mpscale*(Mpprec-1) + 1;
mpnorm(a);
if(Mpdebug)
print(" = %F\n\n", a);
}
double
mpgetflt(Mpflt *a)
{
return a->val;
int s, i;
uvlong v;
double f;
if(a->val.ovf)
warn("mpgetflt ovf");
s = sigfig(a);
if(s == 0)
return 0;
if(s != Mpnorm) {
warn("mpgetflt norm");
mpnorm(a);
}
while((a->val.a[Mpnorm-1] & (1L<<(Mpscale-1))) == 0) {
mpshiftfix(&a->val, 1);
a->exp -= 1;
}
// pick up the mantissa in a uvlong
s = 63;
v = 0;
for(i=Mpnorm-1; s>=Mpscale; i--) {
v = (v<<Mpscale) | a->val.a[i];
s -= Mpscale;
}
if(s > 0)
v = (v<<s) | (a->val.a[i]>>(Mpscale-s));
// should do this in multi precision
// 63 bits of mantissa being rounded to 53
if((v&0x3ffULL) != 0x200ULL || (v&0x400) != 0)
v += 0x200ULL; // round
v &= ~0x3ffULL;
f = (double)(v);
f = ldexp(f, Mpnorm*Mpscale + a->exp - 63);
if(a->val.neg)
f = -f;
return f;
}
void
mpmovecflt(Mpflt *a, double c)
{
a->val = c;
int i;
double f;
long l;
if(Mpdebug)
print("\nconst %g", c);
mpmovecfix(&a->val, 0);
a->exp = 0;
if(c == 0)
goto out;
if(c < 0) {
a->val.neg = 1;
c = -c;
}
f = frexp(c, &i);
a->exp = i;
for(i=0; i<10; i++) {
f = f*Mpbase;
l = floor(f);
f = f - l;
a->exp -= Mpscale;
a->val.a[0] = l;
if(f == 0)
break;
mpshiftfix(&a->val, Mpscale);
}
out:
mpnorm(a);
if(Mpdebug)
print(" = %F\n", a);
}
void
mpnegflt(Mpflt *a)
{
a->val = -a->val;
a->val.neg ^= 1;
}
int
mptestflt(Mpflt *a)
{
if(a->val < 0)
return -1;
if(a->val > 0)
return +1;
return 0;
int s;
if(Mpdebug)
print("\n%F?", a);
s = sigfig(a);
if(s != 0) {
s = +1;
if(a->val.neg)
s = -1;
}
if(Mpdebug)
print(" = %d\n", s);
return s;
}
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment