X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Ftv%2Ffltcalc.c;h=0f162da3f2dafbba4776be2c704cadb23c4baabf;hb=ec3df9a8f58fda0fdfcbd2930f28fa9b96bfc8e1;hp=105d9623f74d343020ec5f33882d9693703ac42b;hpb=8fb3ce50cde68cff60b14725b205a7c80702a115;p=libfirm diff --git a/ir/tv/fltcalc.c b/ir/tv/fltcalc.c index 105d9623f..0f162da3f 100644 --- a/ir/tv/fltcalc.c +++ b/ir/tv/fltcalc.c @@ -1,20 +1,31 @@ /* - * Project: libFIRM - * File name: ir/tv/fltcalc.c - * Purpose: - * Author: - * Modified by: - * Created: 2003 - * CVS-ID: $Id$ - * Copyright: (c) 2003 Universität Karlsruhe - * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * + * This file is part of libFirm. + * + * This file may be distributed and/or modified under the terms of the + * GNU General Public License version 2 as published by the Free Software + * Foundation and appearing in the file LICENSE.GPL included in the + * packaging of this file. + * + * Licensees holding valid libFirm Professional Edition licenses may use + * this file in accordance with the libFirm Commercial License. + * Agreement provided with the Software. + * + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. */ +/** + * @file + * @brief tarval floating point calculations + * @date 2003 + * @author Mathias Heil + * @version $Id$ + */ -#ifdef HAVE_CONFIG_H -# include "config.h" -#endif - +#include "config.h" #include "fltcalc.h" #include "strcalc.h" @@ -28,412 +39,374 @@ #ifdef HAVE_INTTYPES_H # include #endif -#ifdef HAVE_STRING_H -# include -#endif -#ifdef HAVE_STDLIB_H -# include -#endif -#ifdef HAVE_ALLOCA_H -# include -#endif -#ifdef HAVE_MALLOC_H -# include -#endif +#include +#include #include #include #include "xmalloc.h" +/** The number of extra precision rounding bits */ +#define ROUNDING_BITS 2 + typedef uint32_t UINT32; #ifdef HAVE_LONG_DOUBLE #ifdef WORDS_BIGENDIAN typedef union { - struct { - UINT32 high; - UINT32 mid; - UINT32 low; - } val; - volatile long double d; + struct { + UINT32 high; + UINT32 mid; + UINT32 low; + } val; + volatile long double d; } value_t; #else typedef union { - struct { - UINT32 low; - UINT32 mid; - UINT32 high; - } val; - volatile long double d; + struct { + UINT32 low; + UINT32 mid; + UINT32 high; + } val; + volatile long double d; } value_t; #endif #else #ifdef WORDS_BIGENDIAN typedef union { - struct { - UINT32 high; - UINT32 low; - } val; - volatile double d; + struct { + UINT32 high; + UINT32 low; + } val; + volatile double d; } value_t; #else typedef union { - struct { - UINT32 low; - UINT32 high; - } val; - volatile double d; + struct { + UINT32 low; + UINT32 high; + } val; + volatile double d; } value_t; #endif #endif -/** - * possible float states - */ -typedef enum { - NORMAL, /**< normal representation, implicit 1 */ - ZERO, /**< +/-0 */ - SUBNORMAL, /**< denormals, implicit 0 */ - INF, /**< +/-oo */ - NAN, /**< Not A Number */ -} value_class_t; - -/** A descriptor for an IEEE float value. */ -typedef struct { - unsigned char exponent_size; /**< size of exponent in bits */ - unsigned char mantissa_size; /**< size of mantissa in bits */ - value_class_t clss; /**< state of this float */ -} descriptor_t; - #define CLEAR_BUFFER(buffer) memset(buffer, 0, calc_buffer_size) -/* because variable sized structs are impossible, the internal - * value is represented as a pseudo-struct char array, addressed - * by macros - * struct { - * char sign; // 0 for positive, 1 for negative - * char exp[value_size]; - * char mant[value_size]; - * descriptor_t desc; - * }; - */ -#define _sign(a) (((char*)a)[SIGN_POS]) -#define _exp(a) (&((char*)a)[EXPONENT_POS]) -#define _mant(a) (&((char*)a)[MANTISSA_POS]) -#define _desc(a) (*(descriptor_t *)&((char *)a)[DESCRIPTOR_POS]) +/* our floating point value */ +struct _fp_value { + ieee_descriptor_t desc; + char sign; + char value[1]; /* exp[value_size] + mant[value_size] */ +}; -#define _save_result(x) memcpy((x), sc_get_buffer(), value_size) -#define _shift_right(x, y, b) sc_shr((x), (y), value_size*4, 0, (b)) -#define _shift_left(x, y, b) sc_shl((x), (y), value_size*4, 0, (b)) +#define _exp(a) &((a)->value[0]) +#define _mant(a) &((a)->value[value_size]) -#define FC_DEFINE1(code) \ -char *fc_##code(const void *a, void *result) { \ - return _calc((const char*)a, NULL, FC_##code, (char*)result); \ -} - -#define FC_DEFINE2(code) \ -char *fc_##code(const void *a, const void *b, void *result) { \ - return _calc((const char*)a, (const char*)b, FC_##code, (char*)result); \ -} +#define _save_result(x) memcpy((x), sc_get_buffer(), value_size) +#define _shift_right(x, y, res) sc_shr((x), (y), value_size*4, 0, (res)) +#define _shift_left(x, y, res) sc_shl((x), (y), value_size*4, 0, (res)) -#define FUNC_PTR(code) fc_##code -#if FLTCALC_DEBUG +#ifdef FLTCALC_DEBUG # define DEBUGPRINTF(x) printf x #else # define DEBUGPRINTF(x) ((void)0) #endif -#if FLTCALC_TRACE_CALC +#ifdef FLTCALC_TRACE_CALC # define TRACEPRINTF(x) printf x #else # define TRACEPRINTF(x) ((void)0) #endif -static char *calc_buffer = NULL; +/** The immediate precision. */ +static unsigned immediate_prec = 0; + +/** A temporal buffer. */ +static fp_value *calc_buffer = NULL; +/** Current rounding mode.*/ static fc_rounding_mode_t rounding_mode; static int calc_buffer_size; static int value_size; -static int SIGN_POS; -static int EXPONENT_POS; -static int MANTISSA_POS; -static int DESCRIPTOR_POS; - static int max_precision; -/******** - * private functions - ********/ + +/** Exact flag. */ +static int fc_exact = 1; + #if 0 -static void _fail_char(const char *str, unsigned int len, int pos) -{ - if (*(str+pos)) - printf("ERROR: Unexpected character '%c'\n", *(str + pos)); - else - printf("ERROR: Unexpected end of string\n"); - while (len-- && *str) printf("%c", *str++); printf("\n"); - while (pos--) printf(" "); printf("^\n"); - /* the front end has to to check constant strings */ - exit(-1); +static void fail_char(const char *str, unsigned int len, int pos) { + if (*(str+pos)) + printf("ERROR: Unexpected character '%c'\n", *(str + pos)); + else + printf("ERROR: Unexpected end of string\n"); + while (len-- && *str) printf("%c", *str++); printf("\n"); + while (pos--) printf(" "); printf("^\n"); + /* the front end has to to check constant strings */ + exit(-1); } #endif /** pack machine-like */ -static char* _pack(const char *int_float, char *packed) -{ - char *shift_val; - char *temp; - char *val_buffer; - - temp = alloca(value_size); - shift_val = alloca(value_size); - - switch (_desc(int_float).clss) { - case NAN: - val_buffer = alloca(calc_buffer_size); - fc_get_qnan(_desc(int_float).exponent_size, _desc(int_float).mantissa_size, val_buffer); - int_float = val_buffer; - break; - - case INF: - val_buffer = alloca(calc_buffer_size); - fc_get_plusinf(_desc(int_float).exponent_size, _desc(int_float).mantissa_size, val_buffer); - _sign(val_buffer) = _sign(int_float); - int_float = val_buffer; - break; - - default: - break; - } - /* pack sign */ - sc_val_from_ulong(_sign(int_float), temp); - - sc_val_from_ulong(_desc(int_float).exponent_size + _desc(int_float).mantissa_size, NULL); - _shift_left(temp, sc_get_buffer(), packed); - - /* extract exponent */ - sc_val_from_ulong(_desc(int_float).mantissa_size, shift_val); - - _shift_left(_exp(int_float), shift_val, temp); - - sc_or(temp, packed, packed); - - /* extract mantissa */ - /* remove 2 rounding bits */ - sc_val_from_ulong(2, shift_val); - _shift_right(_mant(int_float), shift_val, temp); - - /* remove leading 1 (or 0 if denormalized) */ - sc_max_from_bits(_desc(int_float).mantissa_size, 0, shift_val); /* all mantissa bits are 1's */ - sc_and(temp, shift_val, temp); - - /* save result */ - sc_or(temp, packed, packed); - - return packed; +static void *pack(const fp_value *int_float, void *packed) { + char *shift_val; + char *temp; + fp_value *val_buffer; + int pos; + + temp = alloca(value_size); + shift_val = alloca(value_size); + + switch ((value_class_t)int_float->desc.clss) { + case NAN: + val_buffer = alloca(calc_buffer_size); + fc_get_qnan(&int_float->desc, val_buffer); + int_float = val_buffer; + break; + + case INF: + val_buffer = alloca(calc_buffer_size); + fc_get_plusinf(&int_float->desc, val_buffer); + val_buffer->sign = int_float->sign; + int_float = val_buffer; + break; + + default: + break; + } + assert(int_float->desc.explicit_one <= 1); + + /* pack sign: move it to the left after exponent AND mantissa */ + sc_val_from_ulong(int_float->sign, temp); + + pos = int_float->desc.exponent_size + int_float->desc.mantissa_size + int_float->desc.explicit_one; + sc_val_from_ulong(pos, NULL); + _shift_left(temp, sc_get_buffer(), packed); + + /* pack exponent: move it to the left after mantissa */ + pos = int_float->desc.mantissa_size + int_float->desc.explicit_one; + sc_val_from_ulong(pos, shift_val); + _shift_left(_exp(int_float), shift_val, temp); + + /* combine sign|exponent */ + sc_or(temp, packed, packed); + + /* extract mantissa */ + /* remove rounding bits */ + sc_val_from_ulong(ROUNDING_BITS, shift_val); + _shift_right(_mant(int_float), shift_val, temp); + + /* remove leading 1 (or 0 if denormalized) */ + sc_max_from_bits(pos, 0, shift_val); /* all mantissa bits are 1's */ + sc_and(temp, shift_val, temp); + + /* combine sign|exponent|mantissa */ + sc_or(temp, packed, packed); + + return packed; } -char* _normalize(const char *in_val, char *out_val, int sticky) -{ - int hsb; - char lsb, guard, round, round_dir = 0; - char *temp; - - temp = alloca(value_size); - - /* +2: save two rounding bits at the end */ - hsb = 2 + _desc(in_val).mantissa_size - sc_get_highest_set_bit(_mant(in_val)) - 1; - - if (in_val != out_val) - { - _sign(out_val) = _sign(in_val); - memcpy(&_desc(out_val), &_desc(in_val), sizeof(descriptor_t)); - } - - _desc(out_val).clss = NORMAL; - - /* mantissa all zeros, so zero exponent (because of explicit one)*/ - if (hsb == 2 + _desc(in_val).mantissa_size) - { - sc_val_from_ulong(0, _exp(out_val)); - hsb = -1; - } - - /* shift the first 1 into the left of the radix point (i.e. hsb == -1) */ - if (hsb < -1) - { - /* shift right */ - sc_val_from_ulong(-hsb-1, temp); - - _shift_right(_mant(in_val), temp, _mant(out_val)); - - /* remember if some bits were shifted away */ - if (!sticky) sticky = sc_had_carry(); - - sc_add(_exp(in_val), temp, _exp(out_val)); - } - else if (hsb > -1) - { - /* shift left */ - sc_val_from_ulong(hsb+1, temp); - - _shift_left(_mant(in_val), temp, _mant(out_val)); - - sc_sub(_exp(in_val), temp, _exp(out_val)); - } - - /* check for exponent underflow */ - if (sc_is_negative(_exp(out_val)) || sc_is_zero(_exp(out_val))) { - DEBUGPRINTF(("Exponent underflow!\n")); - /* exponent underflow */ - /* shift the mantissa right to have a zero exponent */ - sc_val_from_ulong(1, temp); - sc_sub(temp, _exp(out_val), NULL); - - _shift_right(_mant(out_val), sc_get_buffer(), _mant(out_val)); - if (!sticky) sticky = sc_had_carry(); - /* denormalized means exponent of zero */ - sc_val_from_ulong(0, _exp(out_val)); - - _desc(out_val).clss = SUBNORMAL; - } - - /* perform rounding by adding a value that clears the guard bit and the round bit - * and either causes a carry to round up or not */ - /* get the last 3 bits of the value */ - lsb = sc_sub_bits(_mant(out_val), _desc(out_val).mantissa_size + 2, 0) & 0x7; - guard = (lsb&0x2)>>1; - round = lsb&0x1; - - switch (rounding_mode) - { - case FC_TONEAREST: - /* round to nearest representable value, if in doubt choose the version - * with lsb == 0 */ - round_dir = guard && (sticky || round || lsb>>2); - break; - case FC_TOPOSITIVE: - /* if positive: round to one if the exact value is bigger, else to zero */ - round_dir = (!_sign(out_val) && (guard || round || sticky)); - break; - case FC_TONEGATIVE: - /* if negative: round to one if the exact value is bigger, else to zero */ - round_dir = (_sign(out_val) && (guard || round || sticky)); - break; - case FC_TOZERO: - /* always round to 0 (chopping mode) */ - round_dir = 0; - break; - } - DEBUGPRINTF(("Rounding (s%d, l%d, g%d, r%d, s%d) %s\n", _sign(out_val), lsb>>2, guard, round, sticky, (round_dir)?"up":"down")); - - if (round_dir == 1) - { - guard = (round^guard)<<1; - lsb = !(round || guard)<<2 | guard | round; - } - else - { - lsb = -((guard<<1) | round); - } - - /* add the rounded value */ - if (lsb != 0) { - sc_val_from_long(lsb, temp); - sc_add(_mant(out_val), temp, _mant(out_val)); - } - - /* could have rounded down to zero */ - if (sc_is_zero(_mant(out_val)) && (_desc(out_val).clss == SUBNORMAL)) - _desc(out_val).clss = ZERO; - - /* check for rounding overflow */ - hsb = 2 + _desc(out_val).mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1; - if ((_desc(out_val).clss != SUBNORMAL) && (hsb < -1)) - { - sc_val_from_ulong(1, temp); - _shift_right(_mant(out_val), temp, _mant(out_val)); - - sc_add(_exp(out_val), temp, _exp(out_val)); - } - else if ((_desc(out_val).clss == SUBNORMAL) && (hsb == -1)) - { - /* overflow caused the mantissa to be normal again, - * so adapt the exponent accordingly */ - sc_val_from_ulong(1, temp); - sc_add(_exp(out_val), temp, _exp(out_val)); - - _desc(out_val).clss = NORMAL; - } - /* no further rounding is needed, because rounding overflow means - * the carry of the original rounding was propagated all the way - * up to the bit left of the radix point. This implies the bits - * to the right are all zeros (rounding is +1) */ - - /* check for exponent overflow */ - sc_val_from_ulong((1 << _desc(out_val).exponent_size) - 1, temp); - if (sc_comp(_exp(out_val), temp) != -1) { - DEBUGPRINTF(("Exponent overflow!\n")); - /* exponent overflow, reaction depends on rounding method: - * - * mode | sign of value | result - *-------------------------------------------------------------- - * TO_NEAREST | + | +inf - * | - | -inf - *-------------------------------------------------------------- - * TO_POSITIVE | + | +inf - * | - | smallest representable value - *-------------------------------------------------------------- - * TO_NEAGTIVE | + | largest representable value - * | - | -inf - *-------------------------------------------------------------- - * TO_ZERO | + | largest representable value - * | - | smallest representable value - *--------------------------------------------------------------*/ - if (_sign(out_val) == 0) - { - /* value is positive */ - switch (rounding_mode) { - case FC_TONEAREST: - case FC_TOPOSITIVE: - _desc(out_val).clss = INF; - break; - - case FC_TONEGATIVE: - case FC_TOZERO: - fc_get_max(_desc(out_val).exponent_size, _desc(out_val).mantissa_size, out_val); - } - } else { - /* value is negative */ - switch (rounding_mode) { - case FC_TONEAREST: - case FC_TONEGATIVE: - _desc(out_val).clss = INF; - break; - - case FC_TOPOSITIVE: - case FC_TOZERO: - fc_get_min(_desc(out_val).exponent_size, _desc(out_val).mantissa_size, out_val); - } - } - } - - return out_val; +/** + * Normalize a fp_value. + * + * @return non-zero if result is exact + */ +static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) { + int exact = 1; + int hsb; + char lsb, guard, round, round_dir = 0; + char *temp = alloca(value_size); + + /* save rounding bits at the end */ + hsb = ROUNDING_BITS + in_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(in_val)) - 1; + + if (in_val != out_val) { + out_val->sign = in_val->sign; + memcpy(&out_val->desc, &in_val->desc, sizeof(out_val->desc)); + } + + out_val->desc.clss = NORMAL; + + /* mantissa all zeros, so zero exponent (because of explicit one) */ + if (hsb == ROUNDING_BITS + in_val->desc.mantissa_size) { + sc_val_from_ulong(0, _exp(out_val)); + hsb = -1; + } + + /* shift the first 1 into the left of the radix point (i.e. hsb == -1) */ + if (hsb < -1) { + /* shift right */ + sc_val_from_ulong(-hsb-1, temp); + + _shift_right(_mant(in_val), temp, _mant(out_val)); + + /* remember if some bits were shifted away */ + if (sc_had_carry()) { + exact = 0; + sticky = 1; + } + sc_add(_exp(in_val), temp, _exp(out_val)); + } else if (hsb > -1) { + /* shift left */ + sc_val_from_ulong(hsb+1, temp); + + _shift_left(_mant(in_val), temp, _mant(out_val)); + + sc_sub(_exp(in_val), temp, _exp(out_val)); + } + + /* check for exponent underflow */ + if (sc_is_negative(_exp(out_val)) || sc_is_zero(_exp(out_val))) { + DEBUGPRINTF(("Exponent underflow!\n")); + /* exponent underflow */ + /* shift the mantissa right to have a zero exponent */ + sc_val_from_ulong(1, temp); + sc_sub(temp, _exp(out_val), NULL); + + _shift_right(_mant(out_val), sc_get_buffer(), _mant(out_val)); + if (sc_had_carry()) { + exact = 0; + sticky = 1; + } + /* denormalized means exponent of zero */ + sc_val_from_ulong(0, _exp(out_val)); + + out_val->desc.clss = SUBNORMAL; + } + + /* perform rounding by adding a value that clears the guard bit and the round bit + * and either causes a carry to round up or not */ + /* get the last 3 bits of the value */ + lsb = sc_sub_bits(_mant(out_val), out_val->desc.mantissa_size + ROUNDING_BITS, 0) & 0x7; + guard = (lsb&0x2)>>1; + round = lsb&0x1; + + switch (rounding_mode) { + case FC_TONEAREST: + /* round to nearest representable value, if in doubt choose the version + * with lsb == 0 */ + round_dir = guard && (sticky || round || lsb>>2); + break; + case FC_TOPOSITIVE: + /* if positive: round to one if the exact value is bigger, else to zero */ + round_dir = (!out_val->sign && (guard || round || sticky)); + break; + case FC_TONEGATIVE: + /* if negative: round to one if the exact value is bigger, else to zero */ + round_dir = (out_val->sign && (guard || round || sticky)); + break; + case FC_TOZERO: + /* always round to 0 (chopping mode) */ + round_dir = 0; + break; + } + DEBUGPRINTF(("Rounding (s%d, l%d, g%d, r%d, s%d) %s\n", out_val->sign, lsb>>2, guard, round, sticky, (round_dir)?"up":"down")); + + if (round_dir == 1) { + guard = (round^guard)<<1; + lsb = !(round || guard)<<2 | guard | round; + } else { + lsb = -((guard<<1) | round); + } + + /* add the rounded value */ + if (lsb != 0) { + sc_val_from_long(lsb, temp); + sc_add(_mant(out_val), temp, _mant(out_val)); + exact = 0; + } + + /* could have rounded down to zero */ + if (sc_is_zero(_mant(out_val)) && (out_val->desc.clss == SUBNORMAL)) + out_val->desc.clss = ZERO; + + /* check for rounding overflow */ + hsb = ROUNDING_BITS + out_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1; + if ((out_val->desc.clss != SUBNORMAL) && (hsb < -1)) { + sc_val_from_ulong(1, temp); + _shift_right(_mant(out_val), temp, _mant(out_val)); + if (exact && sc_had_carry()) + exact = 0; + sc_add(_exp(out_val), temp, _exp(out_val)); + } else if ((out_val->desc.clss == SUBNORMAL) && (hsb == -1)) { + /* overflow caused the mantissa to be normal again, + * so adapt the exponent accordingly */ + sc_val_from_ulong(1, temp); + sc_add(_exp(out_val), temp, _exp(out_val)); + + out_val->desc.clss = NORMAL; + } + /* no further rounding is needed, because rounding overflow means + * the carry of the original rounding was propagated all the way + * up to the bit left of the radix point. This implies the bits + * to the right are all zeros (rounding is +1) */ + + /* check for exponent overflow */ + sc_val_from_ulong((1 << out_val->desc.exponent_size) - 1, temp); + if (sc_comp(_exp(out_val), temp) != -1) { + DEBUGPRINTF(("Exponent overflow!\n")); + /* exponent overflow, reaction depends on rounding method: + * + * mode | sign of value | result + *-------------------------------------------------------------- + * TO_NEAREST | + | +inf + * | - | -inf + *-------------------------------------------------------------- + * TO_POSITIVE | + | +inf + * | - | smallest representable value + *-------------------------------------------------------------- + * TO_NEAGTIVE | + | largest representable value + * | - | -inf + *-------------------------------------------------------------- + * TO_ZERO | + | largest representable value + * | - | smallest representable value + *--------------------------------------------------------------*/ + if (out_val->sign == 0) { + /* value is positive */ + switch (rounding_mode) { + case FC_TONEAREST: + case FC_TOPOSITIVE: + out_val->desc.clss = INF; + break; + + case FC_TONEGATIVE: + case FC_TOZERO: + fc_get_max(&out_val->desc, out_val); + } + } else { + /* value is negative */ + switch (rounding_mode) { + case FC_TONEAREST: + case FC_TONEGATIVE: + out_val->desc.clss = INF; + break; + + case FC_TOPOSITIVE: + case FC_TOZERO: + fc_get_min(&out_val->desc, out_val); + } + } + } + return exact; } /** - * Operations involving NaN's must return NaN + * Operations involving NaN's must return NaN. + * They are NOT exact. */ #define handle_NAN(a, b, result) \ do { \ - if (_desc(a).clss == NAN) { \ + if (a->desc.clss == NAN) { \ if (a != result) memcpy(result, a, calc_buffer_size); \ - return result; \ + fc_exact = 0; \ + return; \ } \ - if (_desc(b).clss == NAN) { \ + if (b->desc.clss == NAN) { \ if (b != result) memcpy(result, b, calc_buffer_size); \ - return result; \ + fc_exact = 0; \ + return; \ } \ }while (0) @@ -441,1173 +414,1337 @@ do { \ /** * calculate a + b, where a is the value with the bigger exponent */ -static char* _fadd(const char* a, const char* b, char* result) -{ - char *temp; - char *exp_diff; - - char sign; - char sticky; - - handle_NAN(a, b, result); - - /* make sure result has a descriptor */ - if (result != a && result != b) - memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t)); - - /* determine if this is an addition or subtraction */ - sign = _sign(a) ^ _sign(b); - - /* produce NaN on inf - inf */ - if (sign && (_desc(a).clss == INF) && (_desc(b).clss == INF)) - return fc_get_qnan(_desc(a).exponent_size, _desc(b).mantissa_size, result); - - temp = alloca(value_size); - exp_diff = alloca(value_size); - - /* get exponent difference */ - sc_sub(_exp(a), _exp(b), exp_diff); - - /* initially set sign to be the sign of a, special treatment of subtraction - * when exponents are equal is required though. - * Also special care about the sign is needed when the mantissas are equal - * (+/- 0 ?) */ - if (sign && sc_val_to_long(exp_diff) == 0) { - switch (sc_comp(_mant(a), _mant(b))) { - case 1: /* a > b */ - _sign(result) = _sign(a); /* abs(a) is bigger and a is negative */ - break; - case 0: /* a == b */ - _sign(result) = (rounding_mode == FC_TONEGATIVE); - break; - case -1: /* a < b */ - _sign(result) = _sign(b); /* abs(b) is bigger and b is negative */ - break; - default: - /* can't be reached */ - break; - } - } - else - _sign(result) = _sign(a); - - /* sign has been taken care of, check for special cases */ - if (_desc(a).clss == ZERO) { - if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, calc_buffer_size-SIGN_POS-1); - return result; - } - if (_desc(b).clss == ZERO) { - if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, calc_buffer_size-SIGN_POS-1); - return result; - } - - if (_desc(a).clss == INF) { - if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, calc_buffer_size-SIGN_POS-1); - return result; - } - if (_desc(b).clss == INF) { - if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, calc_buffer_size-SIGN_POS-1); - return result; - } - - /* shift the smaller value to the right to align the radix point */ - /* subnormals have their radix point shifted to the right, - * take care of this first */ - if ((_desc(b).clss == SUBNORMAL) && (_desc(a).clss != SUBNORMAL)) - { - sc_val_from_ulong(1, temp); - sc_sub(exp_diff, temp, exp_diff); - } - - _shift_right(_mant(b), exp_diff, temp); - sticky = sc_had_carry(); - - if (sticky && sign) - { - /* if subtracting a little more than the represented value or adding a little - * more than the represented value to a negative value this, in addition to the - * still set sticky bit, takes account of the 'little more' */ - char *temp1 = alloca(calc_buffer_size); - sc_val_from_ulong(1, temp1); - sc_add(temp, temp1, temp); - } - - if (sign) { - if (sc_comp(_mant(a), temp) == -1) - sc_sub(temp, _mant(a), _mant(result)); - else - sc_sub(_mant(a), temp, _mant(result)); - } else { - sc_add(_mant(a), temp, _mant(result)); - } - - /* _normalize expects a 'normal' radix point, adding two subnormals - * results in a subnormal radix point -> shifting before normalizing */ - if ((_desc(a).clss == SUBNORMAL) && (_desc(b).clss == SUBNORMAL)) - { - sc_val_from_ulong(1, NULL); - _shift_left(_mant(result), sc_get_buffer(), _mant(result)); - } - - /* resulting exponent is the bigger one */ - memmove(_exp(result), _exp(a), value_size); - - return _normalize(result, result, sticky); +static void _fadd(const fp_value *a, const fp_value *b, fp_value *result) { + char *temp; + char *exp_diff; + + char sign, res_sign; + char sticky; + + fc_exact = 1; + + handle_NAN(a, b, result); + + /* make sure result has a descriptor */ + if (result != a && result != b) + result->desc = a->desc; + + /* determine if this is an addition or subtraction */ + sign = a->sign ^ b->sign; + + /* produce NaN on inf - inf */ + if (sign && (a->desc.clss == INF) && (b->desc.clss == INF)) { + fc_exact = 0; + fc_get_qnan(&a->desc, result); + return; + } + + temp = alloca(value_size); + exp_diff = alloca(value_size); + + /* get exponent difference */ + sc_sub(_exp(a), _exp(b), exp_diff); + + /* initially set sign to be the sign of a, special treatment of subtraction + * when exponents are equal is required though. + * Also special care about the sign is needed when the mantissas are equal + * (+/- 0 ?) */ + if (sign && sc_val_to_long(exp_diff) == 0) { + switch (sc_comp(_mant(a), _mant(b))) { + case 1: /* a > b */ + res_sign = a->sign; /* abs(a) is bigger and a is negative */ + break; + case 0: /* a == b */ + res_sign = (rounding_mode == FC_TONEGATIVE); + break; + case -1: /* a < b */ + res_sign = b->sign; /* abs(b) is bigger and b is negative */ + break; + default: + /* can't be reached */ + res_sign = 0; + break; + } + } + else + res_sign = a->sign; + result->sign = res_sign; + + /* sign has been taken care of, check for special cases */ + if (a->desc.clss == ZERO || b->desc.clss == INF) { + if (b != result) + memcpy(result, b, calc_buffer_size); + fc_exact = b->desc.clss == NORMAL; + result->sign = res_sign; + return; + } + if (b->desc.clss == ZERO || a->desc.clss == INF) { + if (a != result) + memcpy(result, a, calc_buffer_size); + fc_exact = a->desc.clss == NORMAL; + result->sign = res_sign; + return; + } + + /* shift the smaller value to the right to align the radix point */ + /* subnormals have their radix point shifted to the right, + * take care of this first */ + if ((b->desc.clss == SUBNORMAL) && (a->desc.clss != SUBNORMAL)) { + sc_val_from_ulong(1, temp); + sc_sub(exp_diff, temp, exp_diff); + } + + _shift_right(_mant(b), exp_diff, temp); + sticky = sc_had_carry(); + fc_exact &= !sticky; + + if (sticky && sign) { + /* if subtracting a little more than the represented value or adding a little + * more than the represented value to a negative value this, in addition to the + * still set sticky bit, takes account of the 'little more' */ + char *temp1 = alloca(calc_buffer_size); + sc_val_from_ulong(1, temp1); + sc_add(temp, temp1, temp); + } + + if (sign) { + if (sc_comp(_mant(a), temp) == -1) + sc_sub(temp, _mant(a), _mant(result)); + else + sc_sub(_mant(a), temp, _mant(result)); + } else { + sc_add(_mant(a), temp, _mant(result)); + } + + /* _normalize expects a 'normal' radix point, adding two subnormals + * results in a subnormal radix point -> shifting before normalizing */ + if ((a->desc.clss == SUBNORMAL) && (b->desc.clss == SUBNORMAL)) { + sc_val_from_ulong(1, NULL); + _shift_left(_mant(result), sc_get_buffer(), _mant(result)); + } + + /* resulting exponent is the bigger one */ + memmove(_exp(result), _exp(a), value_size); + + fc_exact &= normalize(result, result, sticky); } /** * calculate a * b */ -static char* _fmul(const char* a, const char* b, char* result) -{ - char *temp; - - handle_NAN(a, b, result); - - temp = alloca(value_size); - - if (result != a && result != b) - memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t)); - - _sign(result) = _sign(a) ^ _sign(b); - - /* produce NaN on 0 * inf */ - if (_desc(a).clss == ZERO) { - if (_desc(b).clss == INF) - fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result); - else - if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, calc_buffer_size-1); - return result; - } - if (_desc(b).clss == ZERO) { - if (_desc(a).clss == INF) - fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result); - else - if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, calc_buffer_size-1); - return result; - } - - if (_desc(a).clss == INF) { - if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, calc_buffer_size-1); - return result; - } - if (_desc(b).clss == INF) { - if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, calc_buffer_size-1); - return result; - } - - /* exp = exp(a) + exp(b) - excess */ - sc_add(_exp(a), _exp(b), _exp(result)); - - sc_val_from_ulong((1<<_desc(a).exponent_size)/2-1, temp); - sc_sub(_exp(result), temp, _exp(result)); - - /* mixed normal, subnormal values introduce an error of 1, correct it */ - if ((_desc(a).clss == SUBNORMAL) ^ (_desc(b).clss == SUBNORMAL)) - { - sc_val_from_ulong(1, temp); - sc_add(_exp(result), temp, _exp(result)); - } - - sc_mul(_mant(a), _mant(b), _mant(result)); - - /* realign result: after a multiplication the digits right of the radix - * point are the sum of the factors' digits after the radix point. As all - * values are normalized they both have the same amount of these digits, - * which has to be restored by proper shifting - * +2 because of the two rounding bits */ - sc_val_from_ulong(2 + _desc(result).mantissa_size, temp); - - _shift_right(_mant(result), temp, _mant(result)); - - return _normalize(result, result, sc_had_carry()); +static void _fmul(const fp_value *a, const fp_value *b, fp_value *result) { + int sticky; + char *temp; + char res_sign; + + fc_exact = 1; + + handle_NAN(a, b, result); + + temp = alloca(value_size); + + if (result != a && result != b) + result->desc = a->desc; + + result->sign = res_sign = a->sign ^ b->sign; + + /* produce NaN on 0 * inf */ + if (a->desc.clss == ZERO) { + if (b->desc.clss == INF) { + fc_get_qnan(&a->desc, result); + fc_exact = 0; + } else { + if (a != result) + memcpy(result, a, calc_buffer_size); + result->sign = res_sign; + } + return; + } + if (b->desc.clss == ZERO) { + if (a->desc.clss == INF) { + fc_get_qnan(&a->desc, result); + fc_exact = 0; + } else { + if (b != result) + memcpy(result, b, calc_buffer_size); + result->sign = res_sign; + } + return; + } + + if (a->desc.clss == INF) { + fc_exact = 0; + if (a != result) + memcpy(result, a, calc_buffer_size); + result->sign = res_sign; + return; + } + if (b->desc.clss == INF) { + fc_exact = 0; + if (b != result) + memcpy(result, b, calc_buffer_size); + result->sign = res_sign; + return; + } + + /* exp = exp(a) + exp(b) - excess */ + sc_add(_exp(a), _exp(b), _exp(result)); + + sc_val_from_ulong((1 << (a->desc.exponent_size - 1)) - 1, temp); + sc_sub(_exp(result), temp, _exp(result)); + + /* mixed normal, subnormal values introduce an error of 1, correct it */ + if ((a->desc.clss == SUBNORMAL) ^ (b->desc.clss == SUBNORMAL)) { + sc_val_from_ulong(1, temp); + sc_add(_exp(result), temp, _exp(result)); + } + + sc_mul(_mant(a), _mant(b), _mant(result)); + + /* realign result: after a multiplication the digits right of the radix + * point are the sum of the factors' digits after the radix point. As all + * values are normalized they both have the same amount of these digits, + * which has to be restored by proper shifting + * because of the rounding bits */ + sc_val_from_ulong(ROUNDING_BITS + result->desc.mantissa_size, temp); + + _shift_right(_mant(result), temp, _mant(result)); + sticky = sc_had_carry(); + fc_exact &= !sticky; + + fc_exact &= normalize(result, result, sticky); } /** * calculate a / b */ -static char* _fdiv(const char* a, const char* b, char* result) -{ - char *temp, *dividend; - - handle_NAN(a, b, result); - - temp = alloca(value_size); - dividend = alloca(value_size); - - if (result != a && result != b) - memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t)); - - _sign(result) = _sign(a) ^ _sign(b); - - /* produce nan on 0/0 and inf/inf */ - if (_desc(a).clss == ZERO) { - if (_desc(b).clss == ZERO) - /* 0/0 -> nan */ - fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result); - else - /* 0/x -> a */ - if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, calc_buffer_size-1); - return result; - } - - if (_desc(b).clss == INF) { - if (_desc(a).clss == INF) - /* inf/inf -> nan */ - fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result); - else { - /* x/inf -> 0 */ - sc_val_from_ulong(0, NULL); - _save_result(_exp(result)); - _save_result(_mant(result)); - _desc(result).clss = ZERO; - } - return result; - } - - if (_desc(a).clss == INF) { - /* inf/x -> inf */ - if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, calc_buffer_size-1); - return result; - } - if (_desc(b).clss == ZERO) { - /* division by zero */ - if (_sign(result)) - fc_get_minusinf(_desc(a).exponent_size, _desc(a).mantissa_size, result); - else - fc_get_plusinf(_desc(a).exponent_size, _desc(a).mantissa_size, result); - return result; - } - - /* exp = exp(a) - exp(b) + excess - 1*/ - sc_sub(_exp(a), _exp(b), _exp(result)); - sc_val_from_ulong((1 << _desc(a).exponent_size)/2-2, temp); - sc_add(_exp(result), temp, _exp(result)); - - /* mixed normal, subnormal values introduce an error of 1, correct it */ - if ((_desc(a).clss == SUBNORMAL) ^ (_desc(b).clss == SUBNORMAL)) - { - sc_val_from_ulong(1, temp); - sc_add(_exp(result), temp, _exp(result)); - } - - /* mant(res) = mant(a) / 1/2mant(b) */ - /* to gain more bits of precision in the result the dividend could be - * shifted left, as this operation does not loose bits. This would not - * fit into the integer precision, but due to the rounding bits (which - * are always zero because the values are all normalized) the divisor - * can be shifted right instead to achieve the same result */ - sc_val_from_ulong(2 + _desc(result).mantissa_size, temp); - - _shift_left(_mant(a), temp, dividend); - - { - char *divisor = alloca(calc_buffer_size); - sc_val_from_ulong(1, divisor); - _shift_right(_mant(b), divisor, divisor); - sc_div(dividend, divisor, _mant(result)); - } - - return _normalize(result, result, sc_had_carry()); +static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) { + int sticky; + char *temp, *dividend; + char res_sign; + + fc_exact = 1; + + handle_NAN(a, b, result); + + temp = alloca(value_size); + dividend = alloca(value_size); + + if (result != a && result != b) + result->desc = a->desc; + + result->sign = res_sign = a->sign ^ b->sign; + + /* produce NAN on 0/0 and inf/inf */ + if (a->desc.clss == ZERO) { + if (b->desc.clss == ZERO) { + /* 0/0 -> NaN */ + fc_get_qnan(&a->desc, result); + fc_exact = 0; + } else { + /* 0/x -> a */ + if (a != result) + memcpy(result, a, calc_buffer_size); + result->sign = res_sign; + } + return; + } + + if (b->desc.clss == INF) { + fc_exact = 0; + if (a->desc.clss == INF) { + /* inf/inf -> NaN */ + fc_get_qnan(&a->desc, result); + } else { + /* x/inf -> 0 */ + sc_val_from_ulong(0, NULL); + _save_result(_exp(result)); + _save_result(_mant(result)); + result->desc.clss = ZERO; + } + return; + } + + if (a->desc.clss == INF) { + fc_exact = 0; + /* inf/x -> inf */ + if (a != result) + memcpy(result, a, calc_buffer_size); + result->sign = res_sign; + return; + } + if (b->desc.clss == ZERO) { + fc_exact = 0; + /* division by zero */ + if (result->sign) + fc_get_minusinf(&a->desc, result); + else + fc_get_plusinf(&a->desc, result); + return; + } + + /* exp = exp(a) - exp(b) + excess - 1*/ + sc_sub(_exp(a), _exp(b), _exp(result)); + sc_val_from_ulong((1 << (a->desc.exponent_size - 1)) - 2, temp); + sc_add(_exp(result), temp, _exp(result)); + + /* mixed normal, subnormal values introduce an error of 1, correct it */ + if ((a->desc.clss == SUBNORMAL) ^ (b->desc.clss == SUBNORMAL)) { + sc_val_from_ulong(1, temp); + sc_add(_exp(result), temp, _exp(result)); + } + + /* mant(res) = mant(a) / 1/2mant(b) */ + /* to gain more bits of precision in the result the dividend could be + * shifted left, as this operation does not loose bits. This would not + * fit into the integer precision, but due to the rounding bits (which + * are always zero because the values are all normalized) the divisor + * can be shifted right instead to achieve the same result */ + sc_val_from_ulong(ROUNDING_BITS + result->desc.mantissa_size, temp); + + _shift_left(_mant(a), temp, dividend); + + { + char *divisor = alloca(calc_buffer_size); + sc_val_from_ulong(1, divisor); + _shift_right(_mant(b), divisor, divisor); + sc_div(dividend, divisor, _mant(result)); + sticky = sc_had_carry(); + fc_exact &= !sticky; + } + + fc_exact &= normalize(result, result, sticky); } -static void _power_of_ten(int exp, descriptor_t *desc, char *result) -{ - char *build; - char *temp; +#if 0 +static void _power_of_ten(int exp, ieee_descriptor_t *desc, char *result) { + char *build; + char *temp; - /* positive sign */ - _sign(result) = 0; + /* positive sign */ + result->sign = 0; - /* set new descriptor (else result is supposed to already have one) */ - if (desc != NULL) - memcpy(&_desc(result), desc, sizeof(descriptor_t)); + /* set new descriptor (else result is supposed to already have one) */ + if (desc != NULL) + result->desc = *desc; - build = alloca(value_size); - temp = alloca(value_size); + build = alloca(value_size); + temp = alloca(value_size); - sc_val_from_ulong((1 << _desc(result).exponent_size)/2-1, _exp(result)); + sc_val_from_ulong((1 << (result->desc.exponent_size - 1)) - 1, _exp(result)); - if (exp > 0) - { - /* temp is value of ten now */ - sc_val_from_ulong(10, NULL); - _save_result(temp); + if (exp > 0) { + /* temp is value of ten now */ + sc_val_from_ulong(10, NULL); + _save_result(temp); - for (exp--; exp > 0; exp--) { - _save_result(build); - sc_mul(build, temp, NULL); - } - _save_result(build); + for (exp--; exp > 0; exp--) { + _save_result(build); + sc_mul(build, temp, NULL); + } + _save_result(build); - /* temp is amount of left shift needed to put the value left of the radix point */ - sc_val_from_ulong(_desc(result).mantissa_size + 2, temp); + /* temp is amount of left shift needed to put the value left of the radix point */ + sc_val_from_ulong(result->desc.mantissa_size + ROUNDING_BITS, temp); - _shift_left(build, temp, _mant(result)); + _shift_left(build, temp, _mant(result)); - _normalize(result, result, 0); - } + _normalize(result, result, 0); + } } +#endif /** * Truncate the fractional part away. * - * This does not clip to any integer rang. + * This does not clip to any integer range. */ -static char* _trunc(const char *a, char *result) -{ - /* - * When exponent == 0 all bits left of the radix point - * are the integral part of the value. For 15bit exp_size - * this would require a left shift of max. 16383 bits which - * is too much. - * But it is enough to ensure that no bit right of the radix - * point remains set. This restricts the interesting - * exponents to the interval [0, mant_size-1]. - * Outside this interval the truncated value is either 0 or - * it does not have fractional parts. - */ - - int exp_bias, exp_val; - char *temp; - - temp = alloca(value_size); - - if (a != result) - memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t)); - - exp_bias = (1<<_desc(a).exponent_size)/2-1; - exp_val = sc_val_to_long(_exp(a)) - exp_bias; - - if (exp_val < 0) { - sc_val_from_ulong(0, NULL); - _save_result(_exp(result)); - _save_result(_mant(result)); - _desc(result).clss = ZERO; - - return result; - } - - if (exp_val > _desc(a).mantissa_size) { - if (a != result) - memcpy(result, a, calc_buffer_size); - - return result; - } - - /* set up a proper mask to delete all bits right of the - * radix point if the mantissa had been shifted until exp == 0 */ - sc_max_from_bits(1 + exp_val, 0, temp); - sc_val_from_long(_desc(a).mantissa_size - exp_val + 2, NULL); - _shift_left(temp, sc_get_buffer(), temp); - - /* and the mask and return the result */ - sc_and(_mant(a), temp, _mant(result)); - - if (a != result) memcpy(_exp(result), _exp(a), value_size); - - return result; -} - -/* - * This does value sanity checking(or should do it), sets up any prerequisites, - * calls the proper internal functions, clears up and returns - * the result - */ -char* _calc(const char *a, const char *b, int opcode, char *result) -{ - char *temp; -#ifdef FLTCALC_TRACE_CALC - char *buffer; - - buffer = alloca(100); -#endif - - if (result == NULL) result = calc_buffer; - - TRACEPRINTF(("%s ", fc_print(a, buffer, 100, FC_PACKED))); - switch (opcode) - { - case FC_add: - /* make the value with the bigger exponent the first one */ - TRACEPRINTF(("+ %s ", fc_print(b, buffer, 100, FC_PACKED))); - if (sc_comp(_exp(a), _exp(b)) == -1) - _fadd(b, a, result); - else - _fadd(a, b, result); - break; - case FC_sub: - TRACEPRINTF(("- %s ", fc_print(b, buffer, 100, FC_PACKED))); - temp = alloca(calc_buffer_size); - memcpy(temp, b, calc_buffer_size); - _sign(temp) = !_sign(b); - if (sc_comp(_exp(a), _exp(temp)) == -1) - _fadd(temp, a, result); - else - _fadd(a, temp, result); - break; - case FC_mul: - TRACEPRINTF(("* %s ", fc_print(b, buffer, 100, FC_PACKED))); - _fmul(a, b, result); - break; - case FC_div: - TRACEPRINTF(("/ %s ", fc_print(b, buffer, 100, FC_PACKED))); - _fdiv(a, b, result); - break; - case FC_neg: - TRACEPRINTF(("negated ")); - if (a != result) memcpy(result, a, calc_buffer_size); - _sign(result) = !_sign(a); - break; - case FC_int: - TRACEPRINTF(("truncated to integer ")); - _trunc(a, result); - break; - case FC_rnd: - TRACEPRINTF(("rounded to integer ")); - assert(0); - break; - } - - TRACEPRINTF(("= %s\n", fc_print(result, buffer, 100, FC_PACKED))); - return result; +static void _trunc(const fp_value *a, fp_value *result) { + /* + * When exponent == 0 all bits left of the radix point + * are the integral part of the value. For 15bit exp_size + * this would require a left shift of max. 16383 bits which + * is too much. + * But it is enough to ensure that no bit right of the radix + * point remains set. This restricts the interesting + * exponents to the interval [0, mant_size-1]. + * Outside this interval the truncated value is either 0 or + * it does not have fractional parts. + */ + + int exp_bias, exp_val; + char *temp; + + /* fixme: can be exact */ + fc_exact = 0; + + temp = alloca(value_size); + + if (a != result) + result->desc = a->desc; + + exp_bias = (1 << (a->desc.exponent_size - 1)) - 1; + exp_val = sc_val_to_long(_exp(a)) - exp_bias; + + if (exp_val < 0) { + sc_val_from_ulong(0, NULL); + _save_result(_exp(result)); + _save_result(_mant(result)); + result->desc.clss = ZERO; + + return; + } + + if (exp_val > a->desc.mantissa_size) { + if (a != result) + memcpy(result, a, calc_buffer_size); + + return; + } + + /* set up a proper mask to delete all bits right of the + * radix point if the mantissa had been shifted until exp == 0 */ + sc_max_from_bits(1 + exp_val, 0, temp); + sc_val_from_long(a->desc.mantissa_size - exp_val + 2, NULL); + _shift_left(temp, sc_get_buffer(), temp); + + /* and the mask and return the result */ + sc_and(_mant(a), temp, _mant(result)); + + if (a != result) memcpy(_exp(result), _exp(a), value_size); } /******** * functions defined in fltcalc.h ********/ -const void *fc_get_buffer(void) -{ - return calc_buffer; +const void *fc_get_buffer(void) { + return calc_buffer; } -int fc_get_buffer_length(void) -{ - return calc_buffer_size; +int fc_get_buffer_length(void) { + return calc_buffer_size; } -char* fc_val_from_str(const char *str, unsigned int len, char exp_size, char mant_size, char *result) -{ +void *fc_val_from_str(const char *str, unsigned int len, const ieee_descriptor_t *desc, void *result) { #if 0 - enum { - START, - LEFT_OF_DOT, - RIGHT_OF_DOT, - EXP_START, - EXPONENT, - END - }; - - char exp_sign; - int exp_int, hsb, state; - - const char *old_str; - - int pos; - char *mant_str, *exp_val, *power_val; - - if (result == NULL) result = calc_buffer; - - exp_val = alloca(value_size); - power_val = alloca(calc_buffer_size); - mant_str = alloca((len)?(len):(strlen(str))); - - _desc(result).exponent_size = exp_size; - _desc(result).mantissa_size = mant_size; - _desc(result).clss = NORMAL; - - old_str = str; - pos = 0; - exp_int = 0; - state = START; - - while (len == 0 || str-old_str < len) - { - switch (state) { - case START: - switch (*str) { - case '+': - _sign(result) = 0; - state = LEFT_OF_DOT; - str++; - break; - - case '-': - _sign(result) = 1; - state = LEFT_OF_DOT; - str++; - break; - - case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': - _sign(result) = 0; - state = LEFT_OF_DOT; - break; - - case '.': - _sign(result) = 0; - state = RIGHT_OF_DOT; - str++; - break; - - case 'n': - case 'N': - case 'i': - case 'I': - break; - - default: - _fail_char(old_str, len, str - old_str); - } - break; - - case LEFT_OF_DOT: - switch (*str) { - case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': - mant_str[pos++] = *(str++); - break; - - case '.': - state = RIGHT_OF_DOT; - str++; - break; - - case 'e': - case 'E': - state = EXP_START; - str++; - break; - - case '\0': - mant_str[pos] = '\0'; - goto done; - - default: - _fail_char(old_str, len, str - old_str); - } - break; - - case RIGHT_OF_DOT: - switch (*str) { - case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': - mant_str[pos++] = *(str++); - exp_int++; - break; - - case 'e': - case 'E': - state = EXP_START; - str++; - break; - - case '\0': - mant_str[pos] = '\0'; - goto done; - - default: - _fail_char(old_str, len, str - old_str); - } - break; - - case EXP_START: - switch (*str) { - case '-': - exp_sign = 1; - /* fall through */ - case '+': - if (*(str-1) != 'e' && *(str-1) != 'E') _fail_char(old_str, len, str - old_str); - str++; - break; - - case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': - mant_str[pos] = '\0'; - pos = 1; - str++; - state = EXPONENT; - break; - - default: - _fail_char(old_str, len, str - old_str); - } - break; - - case EXPONENT: - switch (*str) { - case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': - pos++; - str++; - break; - - case '\0': goto done; - - default: - _fail_char(old_str, len, str - old_str); - } - } - } /* switch(state) */ + enum { + START, + LEFT_OF_DOT, + RIGHT_OF_DOT, + EXP_START, + EXPONENT, + END + }; + + char exp_sign; + int exp_int, hsb, state; + + const char *old_str; + + int pos; + char *mant_str, *exp_val, *power_val; + + (void) len; + if (result == NULL) result = calc_buffer; + + exp_val = alloca(value_size); + power_val = alloca(calc_buffer_size); + mant_str = alloca((len)?(len):(strlen(str))); + + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + result->desc.clss = NORMAL; + + old_str = str; + pos = 0; + exp_int = 0; + state = START; + + while (len == 0 || str-old_str < len) { + switch (state) { + case START: + switch (*str) { + case '+': + result->sign = 0; + state = LEFT_OF_DOT; + str++; + break; + + case '-': + result->sign = 1; + state = LEFT_OF_DOT; + str++; + break; + + case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': + result->sign = 0; + state = LEFT_OF_DOT; + break; + + case '.': + result->sign = 0; + state = RIGHT_OF_DOT; + str++; + break; + + case 'n': + case 'N': + case 'i': + case 'I': + break; + + default: + fail_char(old_str, len, str - old_str); + } + break; + + case LEFT_OF_DOT: + switch (*str) { + case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': + mant_str[pos++] = *(str++); + break; + + case '.': + state = RIGHT_OF_DOT; + str++; + break; + + case 'e': + case 'E': + state = EXP_START; + str++; + break; + + case '\0': + mant_str[pos] = '\0'; + goto done; + + default: + fail_char(old_str, len, str - old_str); + } + break; + + case RIGHT_OF_DOT: + switch (*str) { + case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': + mant_str[pos++] = *(str++); + exp_int++; + break; + + case 'e': + case 'E': + state = EXP_START; + str++; + break; + + case '\0': + mant_str[pos] = '\0'; + goto done; + + default: + fail_char(old_str, len, str - old_str); + } + break; + + case EXP_START: + switch (*str) { + case '-': + exp_sign = 1; + /* fall through */ + case '+': + if (*(str-1) != 'e' && *(str-1) != 'E') fail_char(old_str, len, str - old_str); + str++; + break; + + case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': + mant_str[pos] = '\0'; + pos = 1; + str++; + state = EXPONENT; + break; + + default: + fail_char(old_str, len, str - old_str); + } + break; + + case EXPONENT: + switch (*str) { + case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': + pos++; + str++; + break; + + case '\0': goto done; + + default: + fail_char(old_str, len, str - old_str); + } + } + } /* switch(state) */ done: - sc_val_from_str(mant_str, strlen(mant_str), _mant(result)); + sc_val_from_str(mant_str, strlen(mant_str), _mant(result)); - /* shift to put value left of radix point */ - sc_val_from_ulong(mant_size + 2, exp_val); + /* shift to put value left of radix point */ + sc_val_from_ulong(mant_size + ROUNDING_BITS, exp_val); - _shift_left(_mant(result), exp_val, _mant(result)); + _shift_left(_mant(result), exp_val, _mant(result)); - sc_val_from_ulong((1 << exp_size)/2-1, _exp(result)); + sc_val_from_ulong((1 << (exp_size - 1)) - 1, _exp(result)); - _normalize(result, result, 0); + _normalize(result, result, 0); - if (state == EXPONENT) { - exp_int -= atoi(str-pos); - } + if (state == EXPONENT) { + exp_int -= atoi(str-pos); + } - _power_of_ten(exp_int, &_desc(result), power_val); + _power_of_ten(exp_int, &result->desc, power_val); - _fdiv(result, power_val, result); + _fdiv(result, power_val, result); - return result; + return result; #else + /* XXX excuse of an implementation to make things work */ + LLDBL val; + fp_value *tmp = alloca(calc_buffer_size); + ieee_descriptor_t tmp_desc; + (void) len; - /* XXX excuse of an implementation to make things work */ - LLDBL val; #ifdef HAVE_LONG_DOUBLE - val = strtold(str, NULL); + val = strtold(str, NULL); + DEBUGPRINTF(("val_from_str(%s)\n", str)); + tmp_desc.exponent_size = 15; + tmp_desc.mantissa_size = 63; + tmp_desc.explicit_one = 1; + tmp_desc.clss = NORMAL; + fc_val_from_ieee754(val, &tmp_desc, tmp); #else - val = strtod(str, NULL); -#endif - - DEBUGPRINTF(("val_from_str(%s)\n", str)); - return fc_val_from_float(val, exp_size, mant_size, result); + val = strtod(str, NULL); + DEBUGPRINTF(("val_from_str(%s)\n", str)); + tmp_desc.exponent_size = 11; + tmp_desc.mantissa_size = 52; + tmp_desc.explicit_one = 0; + tmp_desc.clss = NORMAL; + fc_val_from_ieee754(val, &tmp_desc, tmp); +#endif /* HAVE_LONG_DOUBLE */ + return fc_cast(tmp, desc, result); #endif } -char* fc_val_from_float(LLDBL l, char exp_size, char mant_size, char* result) -{ - char *temp; - int bias_res, bias_val, mant_val; - value_t srcval; - UINT32 sign, exponent, mantissa0, mantissa1; +fp_value *fc_val_from_ieee754(LLDBL l, const ieee_descriptor_t *desc, fp_value *result) { + char *temp; + int bias_res, bias_val, mant_val; + value_t srcval; + char sign; + UINT32 exponent, mantissa0, mantissa1; - srcval.d = l; - bias_res = ((1<exponent_size - 1)) - 1); #ifdef HAVE_LONG_DOUBLE - mant_val = 64; - bias_val = 0x3fff; - sign = (srcval.val.high & 0x00008000) != 0; - exponent = (srcval.val.high & 0x00007FFF) ; - mantissa0 = srcval.val.mid; - mantissa1 = srcval.val.low; + mant_val = 63; + bias_val = 0x3fff; + sign = (srcval.val.high & 0x00008000) != 0; + exponent = (srcval.val.high & 0x00007FFF) ; + mantissa0 = srcval.val.mid; + mantissa1 = srcval.val.low; #else /* no long double */ - mant_val = 52; - bias_val = 0x3ff; - sign = (srcval.val.high & 0x80000000) != 0; - exponent = (srcval.val.high & 0x7FF00000) >> 20; - mantissa0 = srcval.val.high & 0x000FFFFF; - mantissa1 = srcval.val.low; + mant_val = 52; + bias_val = 0x3ff; + sign = (srcval.val.high & 0x80000000) != 0; + exponent = (srcval.val.high & 0x7FF00000) >> 20; + mantissa0 = srcval.val.high & 0x000FFFFF; + mantissa1 = srcval.val.low; #endif #ifdef HAVE_LONG_DOUBLE - TRACEPRINTF(("val_from_float(%.8X%.8X%.8X)\n", ((int*)&l)[2], ((int*)&l)[1], ((int*)&l)[0]));/* srcval.val.high, srcval.val.mid, srcval.val.low)); */ - DEBUGPRINTF(("(%d-%.4X-%.8X%.8X)\n", sign, exponent, mantissa0, mantissa1)); + TRACEPRINTF(("val_from_float(%.8X%.8X%.8X)\n", ((int*)&l)[2], ((int*)&l)[1], ((int*)&l)[0]));/* srcval.val.high, srcval.val.mid, srcval.val.low)); */ + DEBUGPRINTF(("(%d-%.4X-%.8X%.8X)\n", sign, exponent, mantissa0, mantissa1)); #else - TRACEPRINTF(("val_from_float(%.8X%.8X)\n", srcval.val.high, srcval.val.low)); - DEBUGPRINTF(("(%d-%.3X-%.5X%.8X)\n", sign, exponent, mantissa0, mantissa1)); + TRACEPRINTF(("val_from_float(%.8X%.8X)\n", srcval.val.high, srcval.val.low)); + DEBUGPRINTF(("(%d-%.3X-%.5X%.8X)\n", sign, exponent, mantissa0, mantissa1)); #endif - if (result == NULL) result = calc_buffer; - temp = alloca(value_size); - - _desc(result).exponent_size = exp_size; - _desc(result).mantissa_size = mant_size; - - /* extract sign */ - _sign(result) = sign; - - /* sign and flag suffice to identify nan or inf, no exponent/mantissa - * encoding is needed. the function can return immediately in these cases */ - if (isnan(l)) { - _desc(result).clss = NAN; - TRACEPRINTF(("val_from_float resulted in NAN\n")); - return result; - } - else if (isinf(l)) { - _desc(result).clss = INF; - TRACEPRINTF(("val_from_float resulted in %sINF\n", (_sign(result)==1)?"-":"")); - return result; - } - - /* build exponent, because input and output exponent and mantissa sizes may differ - * this looks more complicated than it is: unbiased input exponent + output bias, - * minus the mantissa difference which is added again later when the output float - * becomes normalized */ -#ifdef HAVE_EXPLICIT_ONE - sc_val_from_long((exponent-bias_val+bias_res)-(mant_val-mant_size-1), _exp(result)); -#else - sc_val_from_long((exponent-bias_val+bias_res)-(mant_val-mant_size), _exp(result)); -#endif - - /* build mantissa representation */ -#ifndef HAVE_EXPLICIT_ONE - if (exponent != 0) - { - /* insert the hidden bit */ - sc_val_from_ulong(1, temp); - sc_val_from_ulong(mant_val + 2, NULL); - _shift_left(temp, sc_get_buffer(), NULL); - } - else -#endif - { - sc_val_from_ulong(0, NULL); - } - - _save_result(_mant(result)); - - /* bits from the upper word */ - sc_val_from_ulong(mantissa0, temp); - sc_val_from_ulong(34, NULL); - _shift_left(temp, sc_get_buffer(), temp); - sc_or(_mant(result), temp, _mant(result)); - - /* bits from the lower word */ - sc_val_from_ulong(mantissa1, temp); - sc_val_from_ulong(2, NULL); - _shift_left(temp, sc_get_buffer(), temp); - sc_or(_mant(result), temp, _mant(result)); - - /* _normalize expects the radix point to be normal, so shift mantissa of subnormal - * origin one to the left */ - if (exponent == 0) - { - sc_val_from_ulong(1, NULL); - _shift_left(_mant(result), sc_get_buffer(), _mant(result)); - } - - _normalize(result, result, 0); - - TRACEPRINTF(("val_from_float results in %s\n", fc_print(result, temp, calc_buffer_size, FC_PACKED))); - - return result; + if (result == NULL) result = calc_buffer; + temp = alloca(value_size); + + /* CLEAR the buffer, else some bits might be uninitialized */ + memset(result, 0, fc_get_buffer_length()); + + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + + /* extract sign */ + result->sign = sign; + + /* sign and flag suffice to identify NaN or inf, no exponent/mantissa + * encoding is needed. the function can return immediately in these cases */ + if (isnan(l)) { + result->desc.clss = NAN; + TRACEPRINTF(("val_from_float resulted in NAN\n")); + return result; + } + else if (isinf(l)) { + result->desc.clss = INF; + TRACEPRINTF(("val_from_float resulted in %sINF\n", (result->sign == 1) ? "-" : "")); + return result; + } + + /* build exponent, because input and output exponent and mantissa sizes may differ + * this looks more complicated than it is: unbiased input exponent + output bias, + * minus the mantissa difference which is added again later when the output float + * becomes normalized */ + sc_val_from_long((exponent - bias_val + bias_res) - (mant_val - desc->mantissa_size), _exp(result)); + + /* build mantissa representation */ + if (exponent != 0) { + /* insert the hidden bit */ + sc_val_from_ulong(1, temp); + sc_val_from_ulong(mant_val + ROUNDING_BITS, NULL); + _shift_left(temp, sc_get_buffer(), NULL); + } + else { + sc_val_from_ulong(0, NULL); + } + + _save_result(_mant(result)); + + /* bits from the upper word */ + sc_val_from_ulong(mantissa0, temp); + sc_val_from_ulong(34, NULL); + _shift_left(temp, sc_get_buffer(), temp); + sc_or(_mant(result), temp, _mant(result)); + + /* bits from the lower word */ + sc_val_from_ulong(mantissa1, temp); + sc_val_from_ulong(ROUNDING_BITS, NULL); + _shift_left(temp, sc_get_buffer(), temp); + sc_or(_mant(result), temp, _mant(result)); + + /* _normalize expects the radix point to be normal, so shift mantissa of subnormal + * origin one to the left */ + if (exponent == 0) { + sc_val_from_ulong(1, NULL); + _shift_left(_mant(result), sc_get_buffer(), _mant(result)); + } + + normalize(result, result, 0); + + TRACEPRINTF(("val_from_float results in %s\n", fc_print(result, temp, calc_buffer_size, FC_PACKED))); + + return result; } -LLDBL fc_val_to_float(const void *val) -{ - const char *value; - char *temp = NULL; +LLDBL fc_val_to_ieee754(const fp_value *val) { + fp_value *value; + fp_value *temp = NULL; - int byte_offset; + int byte_offset; - UINT32 sign; - UINT32 exponent; - UINT32 mantissa0; - UINT32 mantissa1; + UINT32 sign; + UINT32 exponent; + UINT32 mantissa0; + UINT32 mantissa1; - value_t buildval; + value_t buildval; + ieee_descriptor_t desc; + unsigned mantissa_size; #ifdef HAVE_LONG_DOUBLE - char result_exponent = 15; - char result_mantissa = 64; + desc.exponent_size = 15; + desc.mantissa_size = 63; + desc.explicit_one = 1; + desc.clss = NORMAL; #else - char result_exponent = 11; - char result_mantissa = 52; + desc.exponent_size = 11; + desc.mantissa_size = 52; + desc.explicit_one = 0; + desc.clss = NORMAL; #endif + mantissa_size = desc.mantissa_size + desc.explicit_one; - temp = alloca(calc_buffer_size); -#ifdef HAVE_EXPLICIT_ONE - value = fc_cast(val, result_exponent, result_mantissa-1, temp); -#else - value = fc_cast(val, result_exponent, result_mantissa, temp); -#endif + temp = alloca(calc_buffer_size); + value = fc_cast(val, &desc, temp); - sign = _sign(value); + sign = value->sign; - /* @@@ long double exponent is 15bit, so the use of sc_val_to_long should not - * lead to wrong results */ - exponent = sc_val_to_long(_exp(value)) ; + /* @@@ long double exponent is 15bit, so the use of sc_val_to_long should not + * lead to wrong results */ + exponent = sc_val_to_long(_exp(value)) ; - sc_val_from_ulong(2, NULL); - _shift_right(_mant(value), sc_get_buffer(), _mant(value)); + sc_val_from_ulong(ROUNDING_BITS, NULL); + _shift_right(_mant(value), sc_get_buffer(), _mant(value)); - mantissa0 = 0; - mantissa1 = 0; + mantissa0 = 0; + mantissa1 = 0; - for (byte_offset = 0; byte_offset < 4; byte_offset++) - mantissa1 |= sc_sub_bits(_mant(value), result_mantissa, byte_offset) << (byte_offset<<3); + for (byte_offset = 0; byte_offset < 4; byte_offset++) + mantissa1 |= sc_sub_bits(_mant(value), mantissa_size, byte_offset) << (byte_offset << 3); - for (; (byte_offset<<3) < result_mantissa; byte_offset++) - mantissa0 |= sc_sub_bits(_mant(value), result_mantissa, byte_offset) << ((byte_offset-4)<<3); + for (; (byte_offset<<3) < desc.mantissa_size; byte_offset++) + mantissa0 |= sc_sub_bits(_mant(value), mantissa_size, byte_offset) << ((byte_offset - 4) << 3); #ifdef HAVE_LONG_DOUBLE - buildval.val.high = sign << 15; - buildval.val.high |= exponent; - buildval.val.mid = mantissa0; - buildval.val.low = mantissa1; + buildval.val.high = sign << 15; + buildval.val.high |= exponent; + buildval.val.mid = mantissa0; + buildval.val.low = mantissa1; #else /* no long double */ - mantissa0 &= 0x000FFFFF; /* get rid of garbage */ - buildval.val.high = sign << 31; - buildval.val.high |= exponent << 20; - buildval.val.high |= mantissa0; - buildval.val.low = mantissa1; + mantissa0 &= 0x000FFFFF; /* get rid of garbage */ + buildval.val.high = sign << 31; + buildval.val.high |= exponent << 20; + buildval.val.high |= mantissa0; + buildval.val.low = mantissa1; #endif - TRACEPRINTF(("val_to_float: %d-%x-%x%x\n", sign, exponent, mantissa0, mantissa1)); - return buildval.d; + TRACEPRINTF(("val_to_float: %d-%x-%x%x\n", sign, exponent, mantissa0, mantissa1)); + return buildval.d; } -char* fc_cast(const void *val, char exp_size, char mant_size, char *result) -{ - const char *value = (const char*) val; - char *temp; - int exp_offset, val_bias, res_bias; - - if (result == NULL) result = calc_buffer; - temp = alloca(value_size); - - if (_desc(value).exponent_size == exp_size && _desc(value).mantissa_size == mant_size) - { - if (value != result) memcpy(result, value, calc_buffer_size); - return result; - } - - /* set the descriptor of the new value */ - _desc(result).exponent_size = exp_size; - _desc(result).mantissa_size = mant_size; - _desc(result).clss = _desc(value).clss; - - _sign(result) = _sign(value); - - /* when the mantissa sizes differ normalizing has to shift to align it. - * this would change the exponent, which is unwanted. So calculate this - * offset and add it */ - val_bias = (1<<_desc(value).exponent_size)/2-1; - res_bias = (1<desc.exponent_size == desc->exponent_size && + value->desc.mantissa_size == desc->mantissa_size && + value->desc.explicit_one == desc->explicit_one) { + if (value != result) + memcpy(result, value, calc_buffer_size); + return result; + } + + if (value->desc.clss == NAN) { + if (sc_get_highest_set_bit(_mant(value)) == value->desc.mantissa_size + 1) + return fc_get_qnan(desc, result); + else + return fc_get_snan(desc, result); + } + + /* set the descriptor of the new value */ + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + result->desc.clss = value->desc.clss; + + result->sign = value->sign; + + /* when the mantissa sizes differ normalizing has to shift to align it. + * this would change the exponent, which is unwanted. So calculate this + * offset and add it */ + val_bias = (1 << (value->desc.exponent_size - 1)) - 1; + res_bias = (1 << (desc->exponent_size - 1)) - 1; + + exp_offset = (res_bias - val_bias) - (value->desc.mantissa_size - desc->mantissa_size); + sc_val_from_long(exp_offset, temp); + sc_add(_exp(value), temp, _exp(result)); + + /* _normalize expects normalized radix point */ + if (value->desc.clss == SUBNORMAL) { + sc_val_from_ulong(1, NULL); + _shift_left(_mant(value), sc_get_buffer(), _mant(result)); + } else if (value != result) { + memcpy(_mant(result), _mant(value), value_size); + } else { + memmove(_mant(result), _mant(value), value_size); + } + + normalize(result, result, 0); + TRACEPRINTF(("Cast results in %s\n", fc_print(result, temp, value_size, FC_PACKED))); + return result; } -char* fc_get_max(unsigned int exponent_size, unsigned int mantissa_size, char* result) -{ - if (result == NULL) result = calc_buffer; +fp_value *fc_get_max(const ieee_descriptor_t *desc, fp_value *result) { + if (result == NULL) result = calc_buffer; - _desc(result).exponent_size = exponent_size; - _desc(result).mantissa_size = mantissa_size; - _desc(result).clss = NORMAL; + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + result->desc.clss = NORMAL; - _sign(result) = 0; + result->sign = 0; - sc_val_from_ulong((1<exponent_size) - 2, _exp(result)); - sc_max_from_bits(mantissa_size + 1, 0, _mant(result)); - sc_val_from_ulong(2, NULL); - _shift_left(_mant(result), sc_get_buffer(), _mant(result)); + sc_max_from_bits(desc->mantissa_size + 1, 0, _mant(result)); + sc_val_from_ulong(ROUNDING_BITS, NULL); + _shift_left(_mant(result), sc_get_buffer(), _mant(result)); - return result; + return result; } -char* fc_get_min(unsigned int exponent_size, unsigned int mantissa_size, char *result) -{ - if (result == NULL) result = calc_buffer; +fp_value *fc_get_min(const ieee_descriptor_t *desc, fp_value *result) { + if (result == NULL) result = calc_buffer; - fc_get_max(exponent_size, mantissa_size, result); - _sign(result) = 1; + fc_get_max(desc, result); + result->sign = 1; - return result; + return result; } -char* fc_get_snan(unsigned int exponent_size, unsigned int mantissa_size, char *result) -{ - if (result == NULL) result = calc_buffer; +fp_value *fc_get_snan(const ieee_descriptor_t *desc, fp_value *result) { + if (result == NULL) result = calc_buffer; - _desc(result).exponent_size = exponent_size; - _desc(result).mantissa_size = mantissa_size; - _desc(result).clss = NAN; + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + result->desc.clss = NAN; - _sign(result) = 0; + result->sign = 0; - sc_val_from_ulong((1<exponent_size) - 1, _exp(result)); - /* signaling NaN has non-zero mantissa with msb not set */ - sc_val_from_ulong(1, _mant(result)); + /* signaling NaN has non-zero mantissa with msb not set */ + sc_val_from_ulong(1, _mant(result)); - return result; + return result; } -char* fc_get_qnan(unsigned int exponent_size, unsigned int mantissa_size, char *result) -{ - if (result == NULL) result = calc_buffer; +fp_value *fc_get_qnan(const ieee_descriptor_t *desc, fp_value *result) { + if (result == NULL) result = calc_buffer; - _desc(result).exponent_size = exponent_size; - _desc(result).mantissa_size = mantissa_size; - _desc(result).clss = NAN; + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + result->desc.clss = NAN; - _sign(result) = 0; + result->sign = 0; - sc_val_from_ulong((1<exponent_size) - 1, _exp(result)); - /* quiet NaN has the msb of the mantissa set, so shift one there */ - sc_val_from_ulong(1, _mant(result)); - /* mantissa_size >+< 1 because of two extra rounding bits */ - sc_val_from_ulong(mantissa_size + 1, NULL); - _shift_left(_mant(result), sc_get_buffer(), _mant(result)); + /* quiet NaN has the msb of the mantissa set, so shift one there */ + sc_val_from_ulong(1, _mant(result)); + /* mantissa_size >+< 1 because of two extra rounding bits */ + sc_val_from_ulong(desc->mantissa_size + 1, NULL); + _shift_left(_mant(result), sc_get_buffer(), _mant(result)); - return result; + return result; } -char* fc_get_plusinf(unsigned int exponent_size, unsigned int mantissa_size, char *result) -{ - if (result == NULL) result = calc_buffer; +fp_value *fc_get_plusinf(const ieee_descriptor_t *desc, fp_value *result) { + if (result == NULL) result = calc_buffer; - _desc(result).exponent_size = exponent_size; - _desc(result).mantissa_size = mantissa_size; - _desc(result).clss = NORMAL; + result->desc.exponent_size = desc->exponent_size; + result->desc.mantissa_size = desc->mantissa_size; + result->desc.explicit_one = desc->explicit_one; + result->desc.clss = INF; - _sign(result) = 0; + result->sign = 0; - sc_val_from_ulong((1<exponent_size) - 1, _exp(result)); - sc_val_from_ulong(0, _mant(result)); + sc_val_from_ulong(0, _mant(result)); - return result; + return result; } -char* fc_get_minusinf(unsigned int exponent_size, unsigned int mantissa_size, char *result) -{ - if (result == NULL) result = calc_buffer; +fp_value *fc_get_minusinf(const ieee_descriptor_t *desc, fp_value *result) { + if (result == NULL) result = calc_buffer; - fc_get_plusinf(exponent_size, mantissa_size, result); - _sign(result) = 1; + fc_get_plusinf(desc, result); + result->sign = 1; - return result; + return result; } -int fc_comp(const void *a, const void *b) -{ - const char *val_a = (const char*)a; - const char *val_b = (const char*)b; - int mul = 1; - - /* - * shortcut: if both values are identical, they are either - * Unordered if NaN or equal - */ - if (a == b) - return _desc(val_a).clss == NAN ? 2 : 0; - - /* unordered if one is a NaN */ - if (_desc(val_a).clss == NAN || _desc(val_b).clss == NAN) - return 2; - - /* zero is equal independent of sign */ - if ((_desc(val_a).clss == ZERO) && (_desc(val_b).clss == ZERO)) - return 0; - - /* different signs make compare easy */ - if (_sign(val_a) != _sign(val_b)) - return (_sign(val_a)==0)?(1):(-1); - - mul = _sign(a) ? -1 : 1; - - /* both infinity means equality */ - if ((_desc(val_a).clss == INF) && (_desc(val_b).clss == INF)) - return 0; - - /* infinity is bigger than the rest */ - if (_desc(val_a).clss == INF) - return 1 * mul; - if (_desc(val_b).clss == INF) - return -1 * mul; - - /* check first exponent, that mantissa if equal */ - switch (sc_comp(_exp(val_a), _exp(val_b))) { - case -1: - return -1 * mul; - case 1: - return 1 * mul; - case 0: - return sc_comp(_mant(val_a), _mant(val_b)) * mul; - default: - return 2; - } +int fc_comp(const fp_value *val_a, const fp_value *val_b) { + int mul = 1; + + /* + * shortcut: if both values are identical, they are either + * Unordered if NaN or equal + */ + if (val_a == val_b) + return val_a->desc.clss == NAN ? 2 : 0; + + /* unordered if one is a NaN */ + if (val_a->desc.clss == NAN || val_b->desc.clss == NAN) + return 2; + + /* zero is equal independent of sign */ + if ((val_a->desc.clss == ZERO) && (val_b->desc.clss == ZERO)) + return 0; + + /* different signs make compare easy */ + if (val_a->sign != val_b->sign) + return (val_a->sign == 0) ? (1) : (-1); + + mul = val_a->sign ? -1 : 1; + + /* both infinity means equality */ + if ((val_a->desc.clss == INF) && (val_b->desc.clss == INF)) + return 0; + + /* infinity is bigger than the rest */ + if (val_a->desc.clss == INF) + return 1 * mul; + if (val_b->desc.clss == INF) + return -1 * mul; + + /* check first exponent, that mantissa if equal */ + switch (sc_comp(_exp(val_a), _exp(val_b))) { + case -1: + return -1 * mul; + case 1: + return 1 * mul; + case 0: + return sc_comp(_mant(val_a), _mant(val_b)) * mul; + default: + return 2; + } } -int fc_is_zero(const void *a) -{ - return _desc(a).clss == ZERO; +int fc_is_zero(const fp_value *a) { + return a->desc.clss == ZERO; } -int fc_is_negative(const void *a) -{ - return _sign(a); +int fc_is_negative(const fp_value *a) { + return a->sign; } -int fc_is_inf(const void *a) -{ - return _desc(a).clss == INF; +int fc_is_inf(const fp_value *a) { + return a->desc.clss == INF; } -int fc_is_nan(const void *a) -{ - return _desc(a).clss == NAN; +int fc_is_nan(const fp_value *a) { + return a->desc.clss == NAN; } -int fc_is_subnormal(const void *a) -{ - return _desc(a).clss == SUBNORMAL; +int fc_is_subnormal(const fp_value *a) { + return a->desc.clss == SUBNORMAL; } -char *fc_print(const void *a, char *buf, int buflen, unsigned base) -{ - const char *val; - char *mul_1; - - val = (const char*)a; - - mul_1 = alloca(calc_buffer_size); - - switch (base) { - case FC_DEC: - switch (_desc(val).clss) { - case INF: - if (buflen >= 8+_sign(val)) sprintf(buf, "%sINFINITY", _sign(val)?"-":""); - else snprintf(buf, buflen, "%sINF", _sign(val)?"-":NULL); - break; - case NAN: - snprintf(buf, buflen, "NAN"); - break; - case ZERO: - snprintf(buf, buflen, "0.0"); - break; - default: - /* XXX to be implemented */ +char *fc_print(const fp_value *val, char *buf, int buflen, unsigned base) { + char *mul_1; + LLDBL flt_val; + + mul_1 = alloca(calc_buffer_size); + + switch (base) { + case FC_DEC: + switch ((value_class_t)val->desc.clss) { + case INF: + snprintf(buf, buflen, "%cINF", val->sign ? '-' : '+'); + break; + case NAN: + snprintf(buf, buflen, "NaN"); + break; + case ZERO: + snprintf(buf, buflen, "0.0"); + break; + default: + flt_val = fc_val_to_ieee754(val); #ifdef HAVE_LONG_DOUBLE - /* XXX 30 is arbitrary */ - snprintf(buf, buflen, "%.30LE", fc_val_to_float(val)); + /* XXX 30 is arbitrary */ + snprintf(buf, buflen, "%.30LE", flt_val); #else - snprintf(buf, buflen, "%.18E", fc_val_to_float(val)); + snprintf(buf, buflen, "%.18E", flt_val); #endif - } - break; - - case FC_HEX: - switch (_desc(val).clss) { - case INF: - if (buflen >= 8+_sign(val)) sprintf(buf, "%sINFINITY", _sign(val)?"-":""); - else snprintf(buf, buflen, "%sINF", _sign(val)?"-":NULL); - break; - case NAN: - snprintf(buf, buflen, "NAN"); - break; - case ZERO: - snprintf(buf, buflen, "0.0"); - break; - default: + } + break; + + case FC_HEX: + switch ((value_class_t)val->desc.clss) { + case INF: + snprintf(buf, buflen, "%cINF", val->sign ? '-' : '+'); + break; + case NAN: + snprintf(buf, buflen, "NAN"); + break; + case ZERO: + snprintf(buf, buflen, "0.0"); + break; + default: + flt_val = fc_val_to_ieee754(val); #ifdef HAVE_LONG_DOUBLE - snprintf(buf, buflen, "%LA", fc_val_to_float(val)); + snprintf(buf, buflen, "%LA", flt_val); #else - snprintf(buf, buflen, "%A", fc_val_to_float(val)); + snprintf(buf, buflen, "%A", flt_val); #endif - } - break; - - case FC_PACKED: - default: - snprintf(buf, buflen, "%s", sc_print(_pack(val, mul_1), value_size*4, SC_HEX, 0)); - buf[buflen - 1] = '\0'; - break; - } - return buf; + } + break; + + case FC_PACKED: + default: + snprintf(buf, buflen, "%s", sc_print(pack(val, mul_1), value_size*4, SC_HEX, 0)); + buf[buflen - 1] = '\0'; + break; + } + return buf; } -unsigned char fc_sub_bits(const void *value, unsigned num_bits, unsigned byte_ofs) -{ - /* this is used to cache the packed version of the value */ - static char *pack = NULL; +unsigned char fc_sub_bits(const fp_value *value, unsigned num_bits, unsigned byte_ofs) { + /* this is used to cache the packed version of the value */ + static char *packed_value = NULL; + + if (packed_value == NULL) packed_value = XMALLOCN(char, value_size); - if (pack == NULL) pack = xmalloc(value_size); + if (value != NULL) + pack(value, packed_value); - if (value != NULL) - _pack((const char*)value, pack); + return sc_sub_bits(packed_value, num_bits, byte_ofs); +} + +/* Returns non-zero if the mantissa is zero, i.e. 1.0Exxx */ +int fc_zero_mantissa(const fp_value *value) { + return sc_get_lowest_set_bit(_mant(value)) == ROUNDING_BITS + value->desc.mantissa_size; +} + +/* Returns the exponent of a value. */ +int fc_get_exponent(const fp_value *value) { + int exp_bias = (1 << (value->desc.exponent_size - 1)) - 1; + return sc_val_to_long(_exp(value)) - exp_bias; +} - return sc_sub_bits(pack, num_bits, byte_ofs); +/* Return non-zero if a given value can be converted lossless into another precision */ +int fc_can_lossless_conv_to(const fp_value *value, const ieee_descriptor_t *desc) { + int v; + int exp_bias; + + /* handle some special cases first */ + switch (value->desc.clss) { + case ZERO: + case INF: + case NAN: + return 1; + default: + break; + } + + /* check if the exponent can be encoded: note, 0 and all ones are reserved for the exponent */ + exp_bias = (1 << (desc->exponent_size - 1)) - 1; + v = fc_get_exponent(value) + exp_bias; + if (0 < v && v < (1 << desc->exponent_size) - 1) { + /* exponent can be encoded, now check the mantissa */ + v = value->desc.mantissa_size + ROUNDING_BITS - sc_get_lowest_set_bit(_mant(value)); + return v < desc->mantissa_size; + } + return 0; } -fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode) -{ - if (mode == FC_TONEAREST || mode == FC_TOPOSITIVE || mode == FC_TONEGATIVE || mode == FC_TOZERO) - rounding_mode = mode; - return rounding_mode; +fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode) { + if (mode == FC_TONEAREST || mode == FC_TOPOSITIVE || mode == FC_TONEGATIVE || mode == FC_TOZERO) + rounding_mode = mode; + + return rounding_mode; } -fc_rounding_mode_t fc_get_rounding_mode(void) -{ - return rounding_mode; +fc_rounding_mode_t fc_get_rounding_mode(void) { + return rounding_mode; } -void init_fltcalc(int precision) -{ - if (calc_buffer == NULL) { - /* does nothing if already init */ - if (precision == 0) precision = FC_DEFAULT_PRECISION; - - init_strcalc(precision + 4); - - /* needs additionally two bits to round, a bit as explicit 1., and one for - * addition overflow */ - max_precision = sc_get_precision() - 4; - if (max_precision < precision) - printf("WARING: not enough precision available, using %d\n", max_precision); - - rounding_mode = FC_TONEAREST; - value_size = sc_get_buffer_length(); - SIGN_POS = 0; - EXPONENT_POS = SIGN_POS + sizeof(char); - MANTISSA_POS = EXPONENT_POS + value_size; - DESCRIPTOR_POS = MANTISSA_POS + value_size; - calc_buffer_size = DESCRIPTOR_POS + sizeof(descriptor_t); - - calc_buffer = xmalloc(calc_buffer_size); - memset(calc_buffer, 0, calc_buffer_size); - DEBUGPRINTF(("init fltcalc:\n\tVALUE_SIZE = %d\n\tSIGN_POS = %d\n\tEXPONENT_POS = %d\n\tMANTISSA_POS = %d\n\tDESCRIPTOR_POS = %d\n\tCALC_BUFFER_SIZE = %d\n\tcalc_buffer = %p\n\n", value_size, SIGN_POS, EXPONENT_POS, MANTISSA_POS, DESCRIPTOR_POS, calc_buffer_size, calc_buffer)); +void init_fltcalc(int precision) { + if (calc_buffer == NULL) { + /* does nothing if already init */ + if (precision == 0) precision = FC_DEFAULT_PRECISION; + + init_strcalc(precision + 2 + ROUNDING_BITS); + + /* needs additionally rounding bits, one bit as explicit 1., and one for + * addition overflow */ + max_precision = sc_get_precision() - (2 + ROUNDING_BITS); + if (max_precision < precision) + printf("WARNING: not enough precision available, using %d\n", max_precision); + + rounding_mode = FC_TONEAREST; + value_size = sc_get_buffer_length(); + calc_buffer_size = sizeof(fp_value) + 2*value_size - 1; + + calc_buffer = xmalloc(calc_buffer_size); + memset(calc_buffer, 0, calc_buffer_size); + DEBUGPRINTF(("init fltcalc:\n\tVALUE_SIZE = %d\ntCALC_BUFFER_SIZE = %d\n\tcalc_buffer = %p\n\n", value_size, calc_buffer_size, calc_buffer)); #ifdef HAVE_LONG_DOUBLE - DEBUGPRINTF(("\tUsing long double (1-15-64) interface\n")); + DEBUGPRINTF(("\tUsing long double (1-15-64) interface\n")); #else - DEBUGPRINTF(("\tUsing double (1-11-52) interface\n")); + DEBUGPRINTF(("\tUsing double (1-11-52) interface\n")); #endif #ifdef WORDS_BIGENDIAN - DEBUGPRINTF(("\tWord order is big endian\n\n")); + DEBUGPRINTF(("\tWord order is big endian\n\n")); #else - DEBUGPRINTF(("\tWord order is little endian\n\n")); + DEBUGPRINTF(("\tWord order is little endian\n\n")); #endif - } + } } void finish_fltcalc (void) { - free(calc_buffer); calc_buffer = NULL; + free(calc_buffer); calc_buffer = NULL; } +#ifdef FLTCALC_TRACE_CALC +static char buffer[100]; +#endif + /* definition of interface functions */ -FC_DEFINE2(add) -FC_DEFINE2(sub) -FC_DEFINE2(mul) -FC_DEFINE2(div) -FC_DEFINE1(neg) -FC_DEFINE1(int) -FC_DEFINE1(rnd) +fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result) { + if (result == NULL) result = calc_buffer; + + TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + TRACEPRINTF(("+ %s ", fc_print(b, buffer, sizeof(buffer), FC_PACKED))); + + /* make the value with the bigger exponent the first one */ + if (sc_comp(_exp(a), _exp(b)) == -1) + _fadd(b, a, result); + else + _fadd(a, b, result); + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result) { + fp_value *temp; + + if (result == NULL) result = calc_buffer; + + TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + TRACEPRINTF(("- %s ", fc_print(b, buffer, sizeof(buffer), FC_PACKED))); + + temp = alloca(calc_buffer_size); + memcpy(temp, b, calc_buffer_size); + temp->sign = !b->sign; + if (sc_comp(_exp(a), _exp(temp)) == -1) + _fadd(temp, a, result); + else + _fadd(a, temp, result); + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result) { + if (result == NULL) result = calc_buffer; + + TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + TRACEPRINTF(("* %s ", fc_print(b, buffer, sizeof(buffer), FC_PACKED))); + + _fmul(a, b, result); + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result) { + if (result == NULL) result = calc_buffer; + + TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + TRACEPRINTF(("/ %s ", fc_print(b, buffer, sizeof(buffer), FC_PACKED))); + + _fdiv(a, b, result); + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +fp_value *fc_neg(const fp_value *a, fp_value *result) { + if (result == NULL) result = calc_buffer; + + TRACEPRINTF(("- %s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + + if (a != result) + memcpy(result, a, calc_buffer_size); + result->sign = !a->sign; + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +fp_value *fc_int(const fp_value *a, fp_value *result) { + if (result == NULL) result = calc_buffer; + + TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + TRACEPRINTF(("truncated to integer ")); + + _trunc(a, result); + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +fp_value *fc_rnd(const fp_value *a, fp_value *result) { + if (result == NULL) result = calc_buffer; + + (void) a; + TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED))); + TRACEPRINTF(("rounded to integer ")); + + assert(!"fc_rnd() not yet implemented"); + + TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED))); + return result; +} + +/* + * convert a floating point value into an sc value ... + */ +int fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode) { + if (a->desc.clss == NORMAL) { + int exp_bias = (1 << (a->desc.exponent_size - 1)) - 1; + int exp_val = sc_val_to_long(_exp(a)) - exp_bias; + int shift, highest; + + if (a->sign && !mode_is_signed(dst_mode)) { + /* FIXME: for now we cannot convert this */ + return 0; + } + + assert(exp_val >= 0 && "floating point value not integral before fc_flt2int() call"); + shift = exp_val - (a->desc.mantissa_size + ROUNDING_BITS); + + if (shift > 0) { + sc_shlI(_mant(a), shift, 64, 0, result); + } else { + sc_shrI(_mant(a), -shift, 64, 0, result); + } + + /* check for overflow */ + highest = sc_get_highest_set_bit(result); + + if (mode_is_signed(dst_mode)) { + if (highest == sc_get_lowest_set_bit(result)) { + /* need extra test for MIN_INT */ + if (highest >= (int) get_mode_size_bits(dst_mode)) { + /* FIXME: handle overflow */ + return 0; + } + } else { + if (highest >= (int) get_mode_size_bits(dst_mode) - 1) { + /* FIXME: handle overflow */ + return 0; + } + } + } else { + if (highest >= (int) get_mode_size_bits(dst_mode)) { + /* FIXME: handle overflow */ + return 0; + } + } + + if (a->sign) + sc_neg(result, result); + + return 1; + } + else if (a->desc.clss == ZERO) { + sc_zero(result); + return 1; + } + return 0; +} + + +unsigned fc_set_immediate_precision(unsigned bits) { + unsigned old = immediate_prec; + + immediate_prec = bits; + return old; +} + +int fc_is_exact(void) { + return fc_exact; +}