]> git.ipfire.org Git - thirdparty/Python/cpython.git/commitdiff
mods by Andrew Kuchling to implement
authorGuido van Rossum <guido@python.org>
Mon, 29 Aug 1994 12:48:32 +0000 (12:48 +0000)
committerGuido van Rossum <guido@python.org>
Mon, 29 Aug 1994 12:48:32 +0000 (12:48 +0000)
        pow(x,y,z) == pow(x,y)%z, but without incurring overflow
Correct problems found by THINK C 6.0

Objects/intobject.c

index c1d750cf0b5beb7f1de5d8cbca61a91b72609607..7a4708e5f805d757eb63c7d83f6776f023b7d2d6 100644 (file)
@@ -1,5 +1,5 @@
 /***********************************************************
-Copyright 1991, 1992, 1993 by Stichting Mathematisch Centrum,
+Copyright 1991, 1992, 1993, 1994 by Stichting Mathematisch Centrum,
 Amsterdam, The Netherlands.
 
                         All Rights Reserved
@@ -27,7 +27,7 @@ OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
 #include "allobjects.h"
 #include "modsupport.h"
 
-#ifdef __STDC__
+#ifdef HAVE_LIMITS_H
 #include <limits.h>
 #endif
 
@@ -168,12 +168,31 @@ long
 getintvalue(op)
        register object *op;
 {
-       if (!is_intobject(op)) {
-               err_badcall();
+       number_methods *nb;
+       intobject *io;
+       long val;
+       
+       if (op && is_intobject(op))
+               return GETINTVALUE((intobject*) op);
+       
+       if (op == NULL || (nb = op->ob_type->tp_as_number) == NULL ||
+           nb->nb_int == NULL) {
+               err_badarg();
                return -1;
        }
-       else
-               return ((intobject *)op) -> ob_ival;
+       
+       io = (intobject*) (*nb->nb_int) (op);
+       if (io == NULL)
+               return -1;
+       if (!is_intobject(io)) {
+               err_setstr(TypeError, "nb_int should return int object");
+               return -1;
+       }
+       
+       val = GETINTVALUE(io);
+       DECREF(io);
+       
+       return val;
 }
 
 /* Methods */
@@ -245,19 +264,134 @@ int_sub(v, w)
        return newintobject(x);
 }
 
+/*
+Integer overflow checking used to be done using a double, but on 64
+bit machines (where both long and double are 64 bit) this fails
+because the double doesn't have enouvg precision.  John Tromp suggests
+the following algorithm:
+
+Suppose again we normalize a and b to be nonnegative.
+Let ah and al (bh and bl) be the high and low 32 bits of a (b, resp.).
+Now we test ah and bh against zero and get essentially 3 possible outcomes.
+
+1) both ah and bh > 0 : then report overflow
+
+2) both ah and bh = 0 : then compute a*b and report overflow if it comes out
+                        negative
+
+3) ah > 0 and bh = 0  : compute ah*bl and report overflow if it's >= 2^31
+                        compute al*bl and report overflow if it's negative
+                        add (ah*bl)<<32 to al*bl and report overflow if
+                        it's negative
+
+In case of no overflow the result is then negated if necessary.
+
+The majority of cases will be 2), in which case this method is the same as
+what I suggested before. If multiplication is expensive enough, then the
+other method is faster on case 3), but also more work to program, so I
+guess the above is the preferred solution.
+
+*/
+
 static object *
 int_mul(v, w)
        intobject *v;
        intobject *w;
 {
-       register long a, b;
-       double x;
+       long a, b, ah, bh, x, y;
+       int s = 1;
+
        a = v->ob_ival;
        b = w->ob_ival;
-       x = (double)a * (double)b;
-       if (x > LONG_MAX || x < (double) (long) (LONG_MIN))
-               return err_ovf("integer multiplication");
-       return newintobject(a * b);
+       ah = a >> (LONG_BIT/2);
+       bh = b >> (LONG_BIT/2);
+
+       /* Quick test for common case: two small positive ints */
+
+       if (ah == 0 && bh == 0) {
+               x = a*b;
+               if (x < 0)
+                       goto bad;
+               return newintobject(x);
+       }
+
+       /* Arrange that a >= b >= 0 */
+
+       if (a < 0) {
+               a = -a;
+               if (a < 0) {
+                       /* Largest negative */
+                       if (b == 0 || b == 1) {
+                               x = a*b;
+                               goto ok;
+                       }
+                       else
+                               goto bad;
+               }
+               s = -s;
+               ah = a >> (LONG_BIT/2);
+       }
+       if (b < 0) {
+               b = -b;
+               if (b < 0) {
+                       /* Largest negative */
+                       if (a == 0 || a == 1 && s == 1) {
+                               x = a*b;
+                               goto ok;
+                       }
+                       else
+                               goto bad;
+               }
+               s = -s;
+               bh = b >> (LONG_BIT/2);
+       }
+
+       /* 1) both ah and bh > 0 : then report overflow */
+
+       if (ah != 0 && bh != 0)
+               goto bad;
+
+       /* 2) both ah and bh = 0 : then compute a*b and report
+                                  overflow if it comes out negative */
+
+       if (ah == 0 && bh == 0) {
+               x = a*b;
+               if (x < 0)
+                       goto bad;
+               return newintobject(x*s);
+       }
+
+       if (a < b) {
+               /* Swap */
+               x = a;
+               a = b;
+               b = x;
+               ah = bh;
+               /* bh not used beyond this point */
+       }
+
+       /* 3) ah > 0 and bh = 0  : compute ah*bl and report overflow if
+                                  it's >= 2^31
+                        compute al*bl and report overflow if it's negative
+                        add (ah*bl)<<32 to al*bl and report overflow if
+                        it's negative
+                       (NB b == bl in this case, and we make a = al) */
+
+       y = ah*b;
+       if (y >= (1L << (LONG_BIT/2)))
+               goto bad;
+       a &= (1L << (LONG_BIT/2)) - 1;
+       x = a*b;
+       if (x < 0)
+               goto bad;
+       x += y << LONG_BIT/2;
+       if (x < 0)
+               goto bad;
+ ok:
+       return newintobject(x * s);
+
+ bad:
+       return err_ovf("integer multiplication");
 }
 
 static int
@@ -330,10 +464,70 @@ int_divmod(x, y)
 }
 
 static object *
-int_pow(v, w)
+int_pow(v, w, z)
        intobject *v;
        intobject *w;
+       intobject *z;
 {
+#if 1
+       register long iv, iw, iz, ix, temp, prev;
+       int zset = 0;
+       iv = v->ob_ival;
+       iw = w->ob_ival;
+       if (iw < 0) {
+               err_setstr(ValueError, "integer to the negative power");
+               return NULL;
+       }
+       if ((object *)z != None) {
+               iz = z->ob_ival;
+               zset = 1;
+       }
+       /*
+        * XXX: The original exponentiation code stopped looping
+        * when temp hit zero; this code will continue onwards
+        * unnecessarily, but at least it won't cause any errors.
+        * Hopefully the speed improvement from the fast exponentiation
+        * will compensate for the slight inefficiency.
+        * XXX: Better handling of overflows is desperately needed.
+        */
+       temp = iv;
+       ix = 1;
+       while (iw > 0) {
+               prev = ix;      /* Save value for overflow check */
+               if (iw & 1) {   
+                       ix = ix*temp;
+                       if (temp == 0)
+                               break; /* Avoid ix / 0 */
+                       if (ix / temp != prev)
+                               return err_ovf("integer pow()");
+               }
+               iw >>= 1;       /* Shift exponent down by 1 bit */
+               if (iw==0) break;
+               prev = temp;
+               temp *= temp;   /* Square the value of temp */
+               if (prev!=0 && temp/prev!=prev)
+                       return err_ovf("integer pow()");
+               if (zset) {
+                       /* If we did a multiplication, perform a modulo */
+                       ix = ix % iz;
+                       temp = temp % iz;
+               }
+       }
+       if (zset) {
+               object *t1, *t2;
+               long int div, mod;
+               t1=newintobject(ix); 
+               t2=newintobject(iz);
+               if (t1==NULL || t2==NULL ||
+                       i_divmod((intobject *)t1, (intobject *)t2, &div, &mod)<0) {
+                       XDECREF(t1);
+                       XDECREF(t2);
+                       return(NULL);
+               }
+               ix=mod;
+       }
+       return newintobject(ix);
+#else
        register long iv, iw, ix;
        iv = v->ob_ival;
        iw = w->ob_ival;
@@ -341,6 +535,10 @@ int_pow(v, w)
                err_setstr(ValueError, "integer to the negative power");
                return NULL;
        }
+       if ((object *)z != None) {
+               err_setstr(TypeError, "pow(int, int, int) not yet supported");
+               return NULL;
+       }
        ix = 1;
        while (--iw >= 0) {
                long prev = ix;
@@ -351,7 +549,8 @@ int_pow(v, w)
                        return err_ovf("integer pow()");
        }
        return newintobject(ix);
-}
+#endif
+}                              
 
 static object *
 int_neg(v)
@@ -535,29 +734,29 @@ int_hex(v)
 }
 
 static number_methods int_as_number = {
-       int_add,        /*nb_add*/
-       int_sub,        /*nb_subtract*/
-       int_mul,        /*nb_multiply*/
-       int_div,        /*nb_divide*/
-       int_mod,        /*nb_remainder*/
-       int_divmod,     /*nb_divmod*/
-       int_pow,        /*nb_power*/
-       int_neg,        /*nb_negative*/
-       int_pos,        /*nb_positive*/
-       int_abs,        /*nb_absolute*/
-       int_nonzero,    /*nb_nonzero*/
-       int_invert,     /*nb_invert*/
-       int_lshift,     /*nb_lshift*/
-       int_rshift,     /*nb_rshift*/
-       int_and,        /*nb_and*/
-       int_xor,        /*nb_xor*/
-       int_or,         /*nb_or*/
+       (binaryfunc)int_add, /*nb_add*/
+       (binaryfunc)int_sub, /*nb_subtract*/
+       (binaryfunc)int_mul, /*nb_multiply*/
+       (binaryfunc)int_div, /*nb_divide*/
+       (binaryfunc)int_mod, /*nb_remainder*/
+       (binaryfunc)int_divmod, /*nb_divmod*/
+       (ternaryfunc)int_pow, /*nb_power*/
+       (unaryfunc)int_neg, /*nb_negative*/
+       (unaryfunc)int_pos, /*nb_positive*/
+       (unaryfunc)int_abs, /*nb_absolute*/
+       (inquiry)int_nonzero, /*nb_nonzero*/
+       (unaryfunc)int_invert, /*nb_invert*/
+       (binaryfunc)int_lshift, /*nb_lshift*/
+       (binaryfunc)int_rshift, /*nb_rshift*/
+       (binaryfunc)int_and, /*nb_and*/
+       (binaryfunc)int_xor, /*nb_xor*/
+       (binaryfunc)int_or, /*nb_or*/
        0,              /*nb_coerce*/
-       int_int,        /*nb_int*/
-       int_long,       /*nb_long*/
-       int_float,      /*nb_float*/
-       int_oct,        /*nb_oct*/
-       int_hex,        /*nb_hex*/
+       (unaryfunc)int_int, /*nb_int*/
+       (unaryfunc)int_long, /*nb_long*/
+       (unaryfunc)int_float, /*nb_float*/
+       (unaryfunc)int_oct, /*nb_oct*/
+       (unaryfunc)int_hex, /*nb_hex*/
 };
 
 typeobject Inttype = {
@@ -566,14 +765,14 @@ typeobject Inttype = {
        "int",
        sizeof(intobject),
        0,
-       int_dealloc,    /*tp_dealloc*/
-       int_print,      /*tp_print*/
+       (destructor)int_dealloc, /*tp_dealloc*/
+       (printfunc)int_print, /*tp_print*/
        0,              /*tp_getattr*/
        0,              /*tp_setattr*/
-       int_compare,    /*tp_compare*/
-       int_repr,       /*tp_repr*/
+       (cmpfunc)int_compare, /*tp_compare*/
+       (reprfunc)int_repr, /*tp_repr*/
        &int_as_number, /*tp_as_number*/
        0,              /*tp_as_sequence*/
        0,              /*tp_as_mapping*/
-       int_hash,       /*tp_hash*/
+       (hashfunc)int_hash, /*tp_hash*/
 };