/* arith03.c Copyright (C) 1990-2002 Codemist Ltd */ /* * Arithmetic functions. * quotient. * */ /* * This code may be used and modified, and redistributed in binary * or source form, subject to the "CCL Public License", which should * accompany it. This license is a variant on the BSD license, and thus * permits use of code derived from this in either open and commercial * projects: but it does require that updates to this code be made * available back to the originators of the package. * Before merging other code in with this or linking this code * with other packages or libraries please check that the license terms * of the other material are compatible with those of this. */ /* Signature: 4fa193f7 08-Apr-2002 */ #include #include #include #include #include "machine.h" #include "tags.h" #include "cslerror.h" #include "externs.h" #include "arith.h" #ifdef TIMEOUT #include "timeout.h" #endif /* * Division */ #ifndef IDIVIDE #ifdef MULDIV64 unsigned32 Idivide(unsigned32 *qp, unsigned32 a, unsigned32 b, unsigned32 c) /* * *qp = (a,b) / c, return the remainder * * The double-length value (a,b) is given as a 62-bit positive number. Its * low order 31 bits are in b, and the 0x80000000 bit of b is zero. The * high order part (in a) is also represented with the 0x80000000 bit zero. * The divisor c is a positive value that is at most 0x7fffffff, and the * procedure will only be called when the correct quotient is in the * (inclusive) range 0 to 0x7fffffff. The above constraints can be thought * of in two ways: (a) Idivide used a 31-bit word and is working on * unsigned values. The 31-bit quantities happen to be being passed around * in 32 bit registers, but the top bit of those registers will never be used * and will contain zero, or (b) the range of values used is such that * a 64 by 32-bit division can be performed, and by constraining the range * of values this 64 by 32-bit division only has to handle positive numbers, * but depending on the hardware of your computer you are entitled to use * either signed or unsigned arithmetic. */ { unsigned64 p = ((unsigned64)a << 31) | (unsigned64)b; *qp = (unsigned32)(p / (unsigned64)c); return (unsigned32)(p % (unsigned64)c); } #else unsigned32 Idivide(unsigned32 *qp, unsigned32 a, unsigned32 b, unsigned32 c) /* * *qp = (a,b) / c, return the remainder * * The double-length value (a,b) is given as a 62-bit positive number. Its * low order 31 bits are in b, and the 0x80000000 bit of b is zero. The * high order part (in a) is also represented with the 0x80000000 bit zero. * The divisor c is a positive value that is at most 0x7fffffff, and the * procedure will only be called when the correct quotient is in the * (inclusive) range 0 to 0x7fffffff. The above constraints can be thought * of in two ways: (a) Idivide used a 31-bit word and is working on * unsigned values. The 31-bit quantities happen to be being passed around * in 32 bit registers, but the top bit of those registers will never be used * and will contain zero, or (b) the range of values used is such that * a 64 by 32-bit division can be performed, and by constraining the range * of values this 64 by 32-bit division only has to handle positive numbers, * but depending on the hardware of your computer you are entitled to use * either signed or unsigned arithmetic. */ { int i; unsigned32 q = 0; /* * I have a loop that needs to be executed 32 times, with just * slightly different behaviour the last time around. Since it is * fairly short, I unwind it three-fold into a loop executed 10 times * plus a postlude. I also do a test at the start that decides if the * quotient will be small (less than about 16 bits) an in that case * loop rather less often - my reasoning is that a plausible distribution * of quotients is exponential so the short route will be taken about * half of the time, and will save almost half of the work done here at the * cost of a not-too-expensive extra test to begin with. */ if (a < (c >> 15)) { a = (a << 15) | (b >> 16); b = (b << 15) & 0x7fffffff; i = 5; } else i = 0; do { q = q << 3; /* accumulate quotient 3 bits at a time */ if (a >= c) /* bit 1 of 3... */ { a = a - c; q = q + 4; } a = a << 1; /* shift (a,b) doubleword left 1 bit */ b = b << 1; if (((int32)b) < 0) a = a + 1; if (a >= c) /* bit 2 of 3... */ { a = a - c; q = q + 2; } a = a << 1; b = b << 1; if (((int32)b) < 0) a = a + 1; if (a >= c) /* bit 3 of 3... */ { a = a - c; q = q + 1; } a = a << 1; b = b << 1; if (((int32)b) < 0) a = a + 1; i++; } while (i < 10); q = q << 2; /* 2 more bits to be included */ if (a >= c) { a = a - c; q = q + 2; } a = a << 1; b = b << 1; if (((int32)b) < 0) a = a + 1; if (a >= c) /* the final bit of the quotient */ { a = a - c; /* leave remainder in a, b should be zero */ q = q + 1; } *qp = q; return a; } #endif #endif /* IDIVIDE */ #ifdef COMMON static Lisp_Object quotis(Lisp_Object a, Lisp_Object b) { Float_union bb; bb.i = b - TAG_SFLOAT; if (bb.f == 0.0) return aerror2("bad arg for quotient", a, b); bb.f = (float) ((float)int_of_fixnum(a) / bb.f); return (bb.i & ~(int32)0xf) + TAG_SFLOAT; } #endif static Lisp_Object quotib(Lisp_Object a, Lisp_Object b) { /* * a fixnum divided by a bignum always gives 0, regardless of signs etc., * save in the one case of (-#x8000000)/(#x8000000) which must yield -1 */ if (int_of_fixnum(a) == fix_mask && bignum_length(b) == 8 && bignum_digits(b)[0] == 0x08000000) return fixnum_of_int(-1); else return fixnum_of_int(0); } #ifdef COMMON static Lisp_Object CLquotib(Lisp_Object a, Lisp_Object b) { Lisp_Object g, nil = C_nil; CSLbool w; push2(a, b); w = minusp(b); errexitn(2); g = gcd(stack[0], stack[-1]); errexitn(2); if (w) { g = negate(g); errexitn(2); } a = stack[-1]; push(g); a = quot2(a, g); errexitn(3); pop2(g, b); stack[0] = a; b = quot2(b, g); pop(a); errexit(); return make_ratio(a, b); } static Lisp_Object CLquotbi(Lisp_Object a, Lisp_Object b) { return CLquotib(a, b); } static Lisp_Object CLquotbb(Lisp_Object a, Lisp_Object b) { return CLquotib(a, b); } static Lisp_Object quotir(Lisp_Object a, Lisp_Object b) { Lisp_Object w, nil; if (a == fixnum_of_int(0)) return a; push3(b, a, C_nil); #define g stack[0] #define a stack[-1] #define b stack[-2] g = gcd(a, numerator(b)); /* * the gcd() function guarantees to hand back a positive result. */ nil = C_nil; if (exception_pending()) goto fail; a = quot2(a, g); nil = C_nil; if (exception_pending()) goto fail; w = minusp(numerator(b)); nil = C_nil; if (exception_pending()) goto fail; if (w) { g = negate(g); nil = C_nil; if (exception_pending()) goto fail; } g = quot2(numerator(b), g); /* denominator of result will be +ve */ nil = C_nil; if (exception_pending()) goto fail; a = times2(a, denominator(b)); nil = C_nil; if (exception_pending()) goto fail; w = make_ratio(a, g); popv(3); return w; fail: popv(3); return nil; #undef a #undef b #undef g } static Lisp_Object quotic(Lisp_Object a, Lisp_Object b) /* * Used for all cases of xxx/. This is coded in a fairly naive * way, multiplying both numerator and denominator of what will end up * as the result by the complex conjugate of b. If floating point * arithmetic is used this can lead to grossly premature overflow. For * the moment I will ignore that miserable fact */ { Lisp_Object u, v, nil; push2(a, b); #define b stack[0] #define a stack[-1] /* * a / (p + iq) is computed as follows: * (a * (p - iq)) / (p^2 + q^2) */ u = negate(imag_part(b)); nil = C_nil; if (exception_pending()) goto fail; u = make_complex(real_part(b), u); nil = C_nil; if (exception_pending()) goto fail; a = times2(a, u); u = real_part(b); u = times2(u, u); nil = C_nil; if (exception_pending()) goto fail; v = imag_part(b); b = u; u = times2(v, v); nil = C_nil; if (exception_pending()) goto fail; u = plus2(u, b); nil = C_nil; if (exception_pending()) goto fail; v = a; popv(2); return quot2(v, u); #undef a #undef b fail: popv(2); return nil; } #endif static Lisp_Object quotif(Lisp_Object a, Lisp_Object b) { double d = float_of_number(b); if (d == 0.0) return aerror2("bad arg for quotient", a, b); d = (double)int_of_fixnum(a) / d; return make_boxfloat(d, type_of_header(flthdr(b))); } #ifdef COMMON static Lisp_Object quotsi(Lisp_Object a, Lisp_Object b) { Float_union aa; if (b == fixnum_of_int(0)) return aerror2("bad arg for quotient", a, b); aa.i = a - TAG_SFLOAT; aa.f = (float) (aa.f / (float)int_of_fixnum(b)); return (aa.i & ~(int32)0xf) + TAG_SFLOAT; } static Lisp_Object quotsb(Lisp_Object a, Lisp_Object b) { double d = float_of_number(b); if (d == 0.0) return aerror2("bad arg for quotient", a, b); d = float_of_number(a) / d; return make_sfloat(d); } #define quotsr(a, b) quotsb(a, b) #define quotsc(a, b) quotic(a, b) #endif static Lisp_Object quotsf(Lisp_Object a, Lisp_Object b) { double d = float_of_number(b); if (d == 0.0) return aerror2("bad arg for quotient", a, b); d = float_of_number(a) / d; return make_boxfloat(d, type_of_header(flthdr(b))); } Lisp_Object quotbn(Lisp_Object a, int32 n) /* * Divide a bignum by an integer, where the integer is (by now) * a natural C int32 but limited to 31 not 32 bits active. I.e. * this can be used when dividing by a fixnum or by dividing by * a one-word bignum. I will not call this with n=-1, which would * be the one case where it could cause the quotient to be bigger * than the dividend. Leaves nwork set to the remainder. */ { Lisp_Object nil; int sign; int32 lena = (bignum_length(a)-CELL)/4-1, i, lenc, lenx; unsigned32 carry; if (lena == 0) /* one-word bignum as numerator */ { int32 p = (int32)bignum_digits(a)[0], w; nil = C_nil; /* So I can access nwork */ nwork = p % n; /* * C does not define what happens on non-exact division involving * negative quantities, so I adjust things here so that the remainder * has the sign that I want and the division I do is exact. */ if (p < 0) { if (nwork > 0) nwork -= n; } else if (nwork < 0) nwork += n; /* Remainder should be correct (with correct sign) by now, regardless */ p = (p - nwork) / n; w = p & fix_mask; if (w == 0 || w == fix_mask) return fixnum_of_int(p); else return make_one_word_bignum(p); } else if (lena == 1) /* * I treat division of a 2-word bignum by a fixnum or 1-word bignum as * a special case since it can lead to a fixnum result - if the divisor is * just one word long and the dividend is 3 or more words I would * certainly have a bignum result. Thus by separating off the code here * I (a) remove the need for test relating to big- to fixnum conversion * later on and (b) avoid allocating heap in this tolerably common case * when sometimes I will not need to. */ { unsigned32 a0 = bignum_digits(a)[0]; int32 a1 = (int32)bignum_digits(a)[1]; if (a1 < 0) { sign = 3; if (a0 == 0) a1 = -a1; else { a0 = clear_top_bit(-(int32)a0); a1 = ~a1; } } else sign = 0; if (n < 0) sign ^= 1, n = -n; if (a1 < n) /* result can be calculated in one word - good! */ { unsigned32 q, r; Ddivide(r, q, (unsigned32)a1, a0, n); if ((sign & 2) != 0) r = -(int32)r; nil = C_nil; nwork = r; a0 = q; if (a0 == 0) return fixnum_of_int(0); if ((sign & 1) == 0) { if ((a0 & fix_mask) == 0) return fixnum_of_int(a0); else if ((a0 & 0x40000000) == 0) return make_one_word_bignum(a0); else return make_two_word_bignum(0, a0); } a0 = -(int32)a0; if ((a0 & fix_mask) == fix_mask) return fixnum_of_int(a0); else if ((a0 & 0xc0000000U) == 0xc0000000U) return make_one_word_bignum(a0); else return make_two_word_bignum(-1, clear_top_bit(a0)); } /* * here the quotient will involve exactly two digits. */ else { unsigned32 q1, q0, r; Ddivide(r, q1, 0, (unsigned32)a1, n); Ddivide(r, q0, r, a0, n); if ((sign & 2) != 0) r = -(int32)r; nil = C_nil; nwork = r; if ((sign & 1) != 0) { if (q0 == 0) q1 = -(int32)q1; else { q1 = ~q1; q0 = clear_top_bit(-(int32)q0); } } if (q1 == 0 && (q0 & 0x40000000) == 0) return make_one_word_bignum(q0); return make_two_word_bignum(q1, q0); } } /* * That has dealt with the special cases - now the input is a bignum * with at least 3 digits and the quotient will certainly be a bignum. * Start by allocating a workspace copy of the dividend. negateb will * leave a a bignum, although it may change its length. */ if ((int32)bignum_digits(a)[lena] < 0) a = negateb(a), sign = 3; else a = copyb(a), sign = 0; errexit(); if (n < 0) { sign ^= 1; n = -n; } /* * Now both a and n have been made positive, and their original signs * have been recorded so that I can tidy up at the end. The 1 bit in sign * tells me what sign the quotient will have, the 2 bit gives the sign * for the remainder. */ lena = (bignum_length(a)-CELL)/4; carry = 0; for (i=lena-1; i>=0; i--) Ddivide(carry, bignum_digits(a)[i], carry, bignum_digits(a)[i], n); if ((sign & 2) != 0) carry = -(int32)carry; nil = C_nil; nwork = carry; /* leave remainder available to caller */ lena--; while (bignum_digits(a)[lena] == 0) lena--; if ((bignum_digits(a)[lena] & 0x40000000) != 0) lena++; if ((sign & 1) != 0) /* quotient will be negative */ { carry = 0xffffffffU; for (i=0; i 0) nwork -= n; } else if (nwork < 0) nwork += n; /* Remainder should be correct (with correct sign) by now, regardless */ else return nil; } else if (lena == 1) /* * I treat division of a 2-word bignum by a fixnum or 1-word bignum as * a special case since it can lead to a fixnum result - if the divisor is * just one word long and the dividend is 3 or more words I would * certainly have a bignum result. Thus by separating off the code here * I (a) remove the need for test relating to big- to fixnum conversion * later on and (b) avoid allocating heap in this tolerably common case * when sometimes I will not need to. */ { unsigned32 a0 = bignum_digits(a)[0]; int32 a1 = (int32)bignum_digits(a)[1]; if (a1 < 0) { sign = 3; if (a0 == 0) a1 = -a1; else { a0 = clear_top_bit(-(int32)a0); a1 = ~a1; } } else sign = 0; if (n < 0) sign ^= 1, n = -n; if (a1 < n) /* result can be calculated in one word - good! */ { unsigned32 q, r; Ddivide(r, q, (unsigned32)a1, a0, n); if ((sign & 2) != 0) r = -(int32)r; nil = C_nil; nwork = r; return nil; } /* * here the quotient will involve two digits. */ else { unsigned32 q1, q0, r; Ddivide(r, q1, 0, (unsigned32)a1, n); Ddivide(r, q0, r, a0, n); if ((sign & 2) != 0) r = -(int32)r; nil = C_nil; nwork = r; return nil; } } /* * That has dealt with the special cases - now the input is a bignum * with at least 3 digits and the quotient will certainly be a bignum. * Start by allocating a workspace copy of the dividend. negateb will * leave a a bignum, although it may change its length. */ /* * Also note that I could fold the negateb() in with the division, and * thereby save allocating a big hunk of extra memory. */ if ((int32)bignum_digits(a)[lena] < 0) a = negateb(a), sign = 3; else a = copyb(a), sign = 0; errexit(); if (n < 0) { sign ^= 1; n = -n; } /* * Now both a and n have been made positive, and their original signs * have been recorded so that I can tidy up at the end. The 1 bit in sign * tells me what sign the quotient will have, the 2 bit gives the sign * for the remainder. */ lena = (bignum_length(a)-CELL)/4; carry = 0; for (i=lena-1; i>=0; i--) Ddivide(carry, bignum_digits(a)[i], carry, bignum_digits(a)[i], n); if ((sign & 2) != 0) carry = -(int32)carry; nil = C_nil; nwork = carry; /* leave remainder available to caller */ return nil; } static Lisp_Object quotbi(Lisp_Object a, Lisp_Object b) { /* * dividing by 1 or -1 seems worth optimising. */ if (b == fixnum_of_int(1)) return a; else if (b == fixnum_of_int(-1)) return negateb(a); else if (b == fixnum_of_int(0)) return aerror2("bad arg for quotient", a, b); else return quotbn(a, int_of_fixnum(b)); } #ifdef COMMON #define quotbs(a, b) quotsb(a, b) #endif #ifdef __powerc /* If you have trouble compiling this just comment it out, please */ #pragma options(!global_optimizer) #endif Lisp_Object quotbb(Lisp_Object a, Lisp_Object b) /* * Divide one bignum by another. This can compute the * remainder at the same time as the quotient, and leaves same around * in mv_2. If it is not needed then setting up the remainder is * a cost - but usually a modest one (in context), so I think that * the simplification is worth-while. */ { Lisp_Object nil, olda; int sign; int32 lena, lenb, lenc, lenx, i; unsigned32 carry, scale, v1; Lisp_Object q, r; /* * If I am dividing by a one-word bignum I can go a bit faster... */ lenb = (bignum_length(b)-CELL-4)/4; if (lenb == 0) { q = quotbn(a, bignum_digits(b)[0]); errexit(); /* * Now lots of frivolity packing up the remainder... */ r = (Lisp_Object)(nwork & fix_mask); if ((int32)r == 0 || (int32)r == fix_mask) mv_2 = fixnum_of_int(nwork); else { push(q); a = make_one_word_bignum(nwork); pop(q); errexit(); mv_2 = a; } return q; } /* * Convert to sign and magnitude representation */ push2(a, b); lena = (bignum_length(a)-CELL-4)/4; if ((int32)bignum_digits(a)[lena] < 0) { a = negateb(a); sign = 3; } else { a = copyb(a); sign = 0; } pop(b); errexit(); lena = (bignum_length(a)-CELL-4)/4; push(a); if ((int32)bignum_digits(b)[lenb] < 0) { b = negateb(b); /* * I MUST NOT do lenb = (bignum_length(b)-CELL-4)/4; here because the * negateb may have failed and therefore handed back junk rather than * a bignum. E.g. an interrupt provoked by the user or a store-jam might * lead to a failure report here. So I must defer re-finding the length until * I have checked for exceptions. */ sign ^= 1; } else b = copyb(b); pop2(a, olda); /* original a, with original sign */ errexit(); lenb = (bignum_length(b)-CELL-4)/4; /* * Now that the numbers are unsigned it could be that I can drop * a leading zero that had previously been necessary. That could reveal * that I have a one-word divisor after all... */ if (bignum_digits(b)[lenb] == 0) lenb--; if (lenb == 0) { a = quotbn(a, bignum_digits(b)[0]); errexit(); if ((sign & 2) != 0) nwork = -nwork; if ((sign & 1) != 0) a = negate(a); errexit(); r = (Lisp_Object)(nwork & fix_mask); if (r == 0 || (int32)r == fix_mask) mv_2 = fixnum_of_int(nwork); else { push(a); if (signed_overflow(nwork)) { if ((int32)nwork < 0) b = make_two_word_bignum(-1, clear_top_bit(nwork)); else b = make_two_word_bignum(0, nwork); } else b = make_one_word_bignum(nwork); pop(a); errexit(); mv_2 = b; } return a; } /* * Now the divisor is at least two digits long. */ if (bignum_digits(a)[lena] == 0) lena--; /* * I will take a special case when the lengths of the two numbers * match since in that case the quotient will be only one digit long. * Also after I have filtered the lena<=lenb cases out, and have treated * one-word cases of b specially I will know that the numerator a has * at least two digits. */ if (lena < lenb) { mv_2 = olda; return fixnum_of_int(0); } else if (lenb == lena) { unsigned32 p0 = bignum_digits(a)[lena-1], p1 = bignum_digits(a)[lena], q0 = bignum_digits(b)[lenb-1], q1 = bignum_digits(b)[lena], r0, r1, q, w, carry1; /* * If the dividend is smaller than the divisor I can return zero right now. */ if (p1 < q1 || (p1 == q1 && p0 < q0)) { mv_2 = olda; return fixnum_of_int(0); } /* * I scale the divisor so that it will have its top bit set (top wrt the * 31 bit field that is in use, that is), and scale the top two digits * of the dividend to match. The resulting values can then be divided to get * a pretty good estimate for the true quotient. Note that the division on * the next line is UNSIGNED arithmetic. */ scale = 0x80000000U / ((unsigned32)1 + q1); Dmultiply(carry, w, q0, scale, 0); Dmultiply(q, w, q1, scale, carry); q = w; Dmultiply(carry, w, p0, scale, 0); Dmultiply(carry, w, p1, scale, carry); if (carry == q) q = 0x7fffffff; else { Ddivide(q, w, carry, w, q); q = w; } /* * q is now my estimated quotient, based on a quick look at high digits. */ Dmultiply(carry, w, q0, q, 0); r0 = w; Dmultiply(carry, w, q1, q, carry); r1 = w; r0 = r0 - p0; if ((int32)r0 < 0) { r0 = clear_top_bit(r0); r1 = r1 - p1 - 1; } else r1 = r1 - p1; /* /* * the next line is a cop-out for now - if my estimated quotient * was close enough to the true value than the residual I get here * ought to be fairly small - if it is not I have bungled. Over several years * of testa and use I have not seen the disaster message triggered, but the * code should stay in until I can write the paragraph of comment that * should go here explaing why all is well... */ if (carry != 0 && (int32)r1 < 0 && carry != 1 && r1 != ~0x7fffffff) { err_printf("\n+++!!!+++ carry needed (line %d of \"%s\")\n", __LINE__, __FILE__); my_exit(EXIT_FAILURE); } /* * That obtained the remainder from (p1,p0)/(q1,q0) - now adjust q until * the remainder has the sign I want it to have. */ /* /* I do not look at carry here - I have a nasty suspicion that I should.. */ while ((int32)r1 > 0 || (r1 == 0 && r0 != 0)) { q--; r0 = r0 - q0; if ((int32)r0 < 0) { r0 = clear_top_bit(r0); r1 = r1 - q1 - 1; } else r1 = r1 - q1; } /* * Now q is (p1,p0)/(q1,q0), which is an over-estimate for the true * quotient, but by at worst 1. Compute a proper full-length remainder * now, using the original unscaled input numbers. */ carry = 0; carry1 = 0xffffffffU; for (i=0; i<=lena; i++) { Dmultiply(carry, w, bignum_digits(b)[i], q, carry); carry1 = bignum_digits(a)[i] + clear_top_bit(~w) + top_bit(carry1); bignum_digits(a)[i] = clear_top_bit(carry1); } if ((int32)carry1 >= 0) /* need to correct it! */ { q--; for (i=0; i<=lena; i++) { carry1 = bignum_digits(a)[i] + bignum_digits(b)[i] + top_bit(carry1); bignum_digits(a)[i] = clear_top_bit(carry1); } } /* * Now the quotient is correct - but since I want to hand back a neat * remainder I have more to do - I must convert the value stored in a * into a legitimate result, and store it in mv_2. */ while (lena > 0 && bignum_digits(a)[lena] == 0) lena--; if ((bignum_digits(a)[lena] & 0x40000000) != 0) lena++; if (lena == 0) /* Maybe fixnum remainder? */ { int32 rr = bignum_digits(a)[0]; if (rr != 0 && (sign & 2) != 0) { rr = -rr; if ((rr & fix_mask) == fix_mask) { mv_2 = fixnum_of_int(rr); goto return_q; } } else if ((rr & fix_mask) == 0) { mv_2 = fixnum_of_int(rr); goto return_q; } } if ((sign & 2) != 0) { carry = 0xffffffffU; for (i=0; i= lenb) { unsigned32 w, h0, h1, v2, u2, carry1, carry2, qq, rr; if (carry == v1) { qq = 0x7fffffff; rr = carry + bignum_digits(a)[lena]; } else { Ddivide(rr, w, carry, bignum_digits(a)[lena], v1); qq = w; } v2 = bignum_digits(b)[lenb-1]; Dmultiply(h1, h0, v2, qq, 0); u2 = bignum_digits(a)[lena-1]; if (h1 > rr || (h1 == rr && h0 > u2)) { qq--; h0 -= v2; if ((int32)h0 < 0) { h0 = clear_top_bit(h0); h1--; } rr += v1; if (h1 > rr || (h1 == rr && h0 > u2)) qq--; } /* * Now subtract off q*b from a, up as far as lena. */ carry1 = 0; carry2 = 0xffffffffU; for (i=0; i<=lenb; i++) { Dmultiply(carry1, w, bignum_digits(b)[i], qq, carry1); carry2 = bignum_digits(a)[lena-lenb+i] + clear_top_bit(~w) + top_bit(carry2); bignum_digits(a)[lena-lenb+i] = clear_top_bit(carry2); } carry2 = carry + clear_top_bit(~carry1) + top_bit(carry2); if ((int32)carry2 >= 0) /* overshot.. */ { qq--; carry2 = 0; for (i=0; i<=lenb; i++) { carry2 = bignum_digits(a)[lena-lenb+i] + bignum_digits(b)[i] + top_bit(carry2); bignum_digits(a)[lena-lenb+i] = clear_top_bit(carry2); } } carry = bignum_digits(a)[lena]; bignum_digits(a)[lena] = qq; /* good place to store the quotient */ lena--; } lena = (bignum_length(a)-CELL-4)/4; /* * Now the quotient is in the top digits of a, and the remainder is left * (scaled up) in the low bit (plus carry) */ bignum_digits(b)[lenb] = carry/scale; carry = carry%scale; for (i=lenb-1; i>=0; i--) Ddivide(carry, bignum_digits(b)[i], carry, bignum_digits(a)[i], scale); if (carry != 0) { err_printf("+++!!!+++ Incorrect unscaling line %d file \"%s\" (%ld)\n", __LINE__, __FILE__, (long)carry); my_exit(EXIT_FAILURE); } for (i=0; i<=lena-lenb; i++) bignum_digits(a)[i] = bignum_digits(a)[i+lenb]; for (;i < lena; i++) bignum_digits(a)[i] = 0; lena = lena - lenb; while (bignum_digits(a)[lena] == 0) lena--; /* * the above loop terminates properly since the quotient will be non-zero, * which in turn is because I was dividing p by q where p had strictly * more digits than q, and was hence strictly greater than q. */ if ((bignum_digits(a)[lena] & 0x40000000) != 0) lena++; if ((sign & 1) != 0) /* quotient will be negative */ { carry = 0xffffffffU; for (i=0; i= 0 && bignum_digits(b)[lenb] == 0) lenb--; if (lenb < 0) mv_2 = fixnum_of_int(0); /* hah - zero remainder! */ else { if ((bignum_digits(b)[lenb] & 0x40000000) != 0) lenb++; if ((sign & 2) != 0) /* need to negate the remainder */ { carry = 0xffffffffU; for (i=0; i 0) r -= bb; } else if (r < 0) r += bb; /* then the division can be an exact one, as here */ r = (aa - r)/bb; /* * the only case where dividing one fixnum by another can lead to * a bignum result is (-0x08000000/(-1)) which JUST overflows. */ if (r != 0x08000000) return fixnum_of_int(r); else return make_one_word_bignum(r); } #ifdef COMMON case TAG_SFLOAT: return quotis(a, b); #endif case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotib(a, b); #ifdef COMMON case TYPE_RATNUM: return quotir(a, b); case TYPE_COMPLEX_NUM: return quotic(a, b); #endif default: return aerror1("bad arg for quotient", b); } } case TAG_BOXFLOAT: return quotif(a, b); default: return aerror1("bad arg for quotient", b); } #ifdef COMMON case TAG_SFLOAT: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotsi(a, b); case TAG_SFLOAT: { Float_union aa, bb; aa.i = a - TAG_SFLOAT; bb.i = b - TAG_SFLOAT; aa.f = (float) (aa.f / bb.f); return (aa.i & ~(int32)0xf) + TAG_SFLOAT; } case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotsb(a, b); case TYPE_RATNUM: return quotsr(a, b); case TYPE_COMPLEX_NUM: return quotsc(a, b); default: return aerror1("bad arg for quotient", b); } } case TAG_BOXFLOAT: return quotsf(a, b); default: return aerror1("bad arg for quotient", b); } #endif case TAG_NUMBERS: { int32 ha = type_of_header(numhdr(a)); switch (ha) { case TYPE_BIGNUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotbi(a, b); #ifdef COMMON case TAG_SFLOAT: return quotbs(a, b); #endif case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotbb(a, b); #ifdef COMMON case TYPE_RATNUM: return quotbr(a, b); case TYPE_COMPLEX_NUM: return quotbc(a, b); #endif default: return aerror1("bad arg for quotient", b); } } case TAG_BOXFLOAT: return quotbf(a, b); default: return aerror1("bad arg for quotient", b); } #ifdef COMMON case TYPE_RATNUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotri(a, b); case TAG_SFLOAT: return quotrs(a, b); case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotrb(a, b); case TYPE_RATNUM: return quotrr(a, b); case TYPE_COMPLEX_NUM: return quotrc(a, b); default: return aerror1("bad arg for quotient", b); } } case TAG_BOXFLOAT: return quotrf(a, b); default: return aerror1("bad arg for quotient", b); } case TYPE_COMPLEX_NUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotci(a, b); case TAG_SFLOAT: return quotcs(a, b); case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotcb(a, b); case TYPE_RATNUM: return quotcr(a, b); case TYPE_COMPLEX_NUM: return quotcc(a, b); default: return aerror1("bad arg for quotient", b); } } case TAG_BOXFLOAT: return quotcf(a, b); default: return aerror1("bad arg for quotient", b); } #endif default: return aerror1("bad arg for quotient", a); } } case TAG_BOXFLOAT: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotfi(a, b); #ifdef COMMON case TAG_SFLOAT: return quotfs(a, b); #endif case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotfb(a, b); #ifdef COMMON case TYPE_RATNUM: return quotfr(a, b); case TYPE_COMPLEX_NUM: return quotfc(a, b); #endif default: return aerror1("bad arg for quotient", b); } } case TAG_BOXFLOAT: return quotff(a, b); default: return aerror1("bad arg for quotient", b); } default: return aerror1("bad arg for quotient", a); } } #ifdef COMMON /* * The above version of quot2 is as required for the Standard Lisp QUOTIENT * function, and it is called from various internal places in CSL/CCL, eg * from within the code for TRUNCATE. Next I have something that will be very * similar in behaviour, but which turns quotients of integers into * rational numbers when that is necessary. */ /***** Not reconstructed yet!! */ Lisp_Object CLquot2(Lisp_Object a, Lisp_Object b) { switch ((int)a & TAG_BITS) { case TAG_FIXNUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: /* * This is where fixnum / fixnum arithmetic happens - the case I most want to * make efficient. */ if (b == fixnum_of_int(0)) return aerror2("bad arg for /", a, b); else { int32 r, aa, bb, w; aa = int_of_fixnum(a); bb = int_of_fixnum(b); if (bb < 0) aa = -aa, bb = -bb; r = aa % bb; if (r == 0) /* Exact division case */ { r = aa/bb; /* * the only case where dividing one fixnum by another can lead to * a bignum result is (-0x08000000/(-1)) which JUST overflows. */ if (r != 0x08000000) return fixnum_of_int(r); else return make_one_word_bignum(r); } w = bb; if (r < 0) r = -r; while (r != 0) { int32 w1 = w % r; w = r; r = w1; } aa = aa / w; bb = bb / w; return make_ratio(fixnum_of_int(aa), fixnum_of_int(bb)); } #ifdef COMMON case TAG_SFLOAT: return quotis(a, b); #endif case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return CLquotib(a, b); #ifdef COMMON case TYPE_RATNUM: return quotir(a, b); case TYPE_COMPLEX_NUM: return quotic(a, b); #endif default: return aerror1("bad arg for /", b); } } case TAG_BOXFLOAT: return quotif(a, b); default: return aerror1("bad arg for /", b); } #ifdef COMMON case TAG_SFLOAT: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotsi(a, b); case TAG_SFLOAT: { Float_union aa, bb; aa.i = a - TAG_SFLOAT; bb.i = b - TAG_SFLOAT; aa.f = (float) (aa.f / bb.f); return (aa.i & ~(int32)0xf) + TAG_SFLOAT; } case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotsb(a, b); case TYPE_RATNUM: return quotsr(a, b); case TYPE_COMPLEX_NUM: return quotsc(a, b); default: return aerror1("bad arg for /", b); } } case TAG_BOXFLOAT: return quotsf(a, b); default: return aerror1("bad arg for /", b); } #endif case TAG_NUMBERS: { int32 ha = type_of_header(numhdr(a)); switch (ha) { case TYPE_BIGNUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return CLquotbi(a, b); #ifdef COMMON case TAG_SFLOAT: return quotbs(a, b); #endif case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return CLquotbb(a, b); #ifdef COMMON case TYPE_RATNUM: return quotbr(a, b); case TYPE_COMPLEX_NUM: return quotbc(a, b); #endif default: return aerror1("bad arg for /", b); } } case TAG_BOXFLOAT: return quotbf(a, b); default: return aerror1("bad arg for /", b); } #ifdef COMMON case TYPE_RATNUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotri(a, b); case TAG_SFLOAT: return quotrs(a, b); case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotrb(a, b); case TYPE_RATNUM: return quotrr(a, b); case TYPE_COMPLEX_NUM: return quotrc(a, b); default: return aerror1("bad arg for /", b); } } case TAG_BOXFLOAT: return quotrf(a, b); default: return aerror1("bad arg for /", b); } case TYPE_COMPLEX_NUM: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotci(a, b); case TAG_SFLOAT: return quotcs(a, b); case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotcb(a, b); case TYPE_RATNUM: return quotcr(a, b); case TYPE_COMPLEX_NUM: return quotcc(a, b); default: return aerror1("bad arg for /", b); } } case TAG_BOXFLOAT: return quotcf(a, b); default: return aerror1("bad arg for /", b); } #endif default: return aerror1("bad arg for /", a); } } case TAG_BOXFLOAT: switch ((int)b & TAG_BITS) { case TAG_FIXNUM: return quotfi(a, b); #ifdef COMMON case TAG_SFLOAT: return quotfs(a, b); #endif case TAG_NUMBERS: { int32 hb = type_of_header(numhdr(b)); switch (hb) { case TYPE_BIGNUM: return quotfb(a, b); #ifdef COMMON case TYPE_RATNUM: return quotfr(a, b); case TYPE_COMPLEX_NUM: return quotfc(a, b); #endif default: return aerror1("bad arg for /", b); } } case TAG_BOXFLOAT: return quotff(a, b); default: return aerror1("bad arg for /", b); } default: return aerror1("bad arg for /", a); } } #endif /* end of arith03.c */