/*
- * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
#include "xmalloc.h"
+/** The number of extra precesion rounding bits */
+#define ROUNDING_BITS 2
+
typedef uint32_t UINT32;
#ifdef HAVE_LONG_DOUBLE
#define _shift_left(x, y, b) sc_shl((x), (y), value_size*4, 0, (b))
-#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
+/** 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 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) {
+static void fail_char(const char *str, unsigned int len, int pos) {
if (*(str+pos))
printf("ERROR: Unexpected character '%c'\n", *(str + pos));
else
#endif
/** pack machine-like */
-static void *_pack(const fp_value *int_float, void *packed) {
+static void *pack(const fp_value *int_float, void *packed) {
char *shift_val;
char *temp;
fp_value *val_buffer;
sc_or(temp, packed, packed);
/* extract mantissa */
- /* remove 2 rounding bits */
- sc_val_from_ulong(2, shift_val);
+ /* 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) */
return packed;
}
-static void _normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
+/**
+ * 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);
- /* +2: save two rounding bits at the end */
- hsb = 2 + in_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(in_val)) - 1;
+ /* 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;
out_val->desc.clss = NORMAL;
- /* mantissa all zeros, so zero exponent (because of explicit one)*/
- if (hsb == 2 + in_val->desc.mantissa_size) {
+ /* 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_right(_mant(in_val), temp, _mant(out_val));
/* remember if some bits were shifted away */
- if (!sticky) sticky = sc_had_carry();
-
+ if (sc_had_carry()) {
+ exact = 0;
+ sticky = 1;
+ }
sc_add(_exp(in_val), temp, _exp(out_val));
} else if (hsb > -1) {
/* shift left */
sc_sub(temp, _exp(out_val), NULL);
_shift_right(_mant(out_val), sc_get_buffer(), _mant(out_val));
- if (!sticky) sticky = sc_had_carry();
+ if (sc_had_carry()) {
+ exact = 0;
+ sticky = 1;
+ }
/* denormalized means exponent of zero */
sc_val_from_ulong(0, _exp(out_val));
/* 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 + 2, 0) & 0x7;
+ lsb = sc_sub_bits(_mant(out_val), out_val->desc.mantissa_size + ROUNDING_BITS, 0) & 0x7;
guard = (lsb&0x2)>>1;
round = lsb&0x1;
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 */
out_val->desc.clss = ZERO;
/* check for rounding overflow */
- hsb = 2 + out_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1;
+ 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,
}
}
}
+ return exact;
}
/**
if (a != result) memcpy(result, a, calc_buffer_size); \
return; \
} \
- if (b->desc.clss == NAN) { \
+ if (b->desc.clss == NAN) { \
if (b != result) memcpy(result, b, calc_buffer_size); \
return; \
} \
char sign, res_sign;
char sticky;
+ fc_exact = 1;
+
handle_NAN(a, b, result);
/* make sure result has a descriptor */
_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
/* resulting exponent is the bigger one */
memmove(_exp(result), _exp(a), value_size);
- _normalize(result, result, sticky);
+ fc_exact &= normalize(result, result, sticky);
}
/**
* calculate a * b
*/
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);
* 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 + result->desc.mantissa_size, temp);
+ * 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;
- _normalize(result, result, sc_had_carry());
+ fc_exact &= normalize(result, result, sticky);
}
/**
* calculate a / b
*/
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);
result->sign = res_sign = a->sign ^ b->sign;
- /* produce nan on 0/0 and inf/inf */
+ /* produce NAN on 0/0 and inf/inf */
if (a->desc.clss == ZERO) {
if (b->desc.clss == ZERO)
/* 0/0 -> nan */
* 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 + result->desc.mantissa_size, temp);
+ sc_val_from_ulong(ROUNDING_BITS + result->desc.mantissa_size, temp);
_shift_left(_mant(a), temp, dividend);
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;
}
- _normalize(result, result, sc_had_carry());
+ fc_exact &= normalize(result, result, sticky);
}
#if 0
build = alloca(value_size);
temp = alloca(value_size);
- sc_val_from_ulong((1 << result->desc.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 */
_save_result(build);
/* temp is amount of left shift needed to put the value left of the radix point */
- sc_val_from_ulong(result->desc.mantissa_size + 2, temp);
+ sc_val_from_ulong(result->desc.mantissa_size + ROUNDING_BITS, temp);
_shift_left(build, temp, _mant(result));
/**
* Truncate the fractional part away.
*
- * This does not clip to any integer rang.
+ * This does not clip to any integer range.
*/
static void _trunc(const fp_value *a, fp_value *result) {
/*
int exp_bias, exp_val;
char *temp;
+ /* fixme: can be exact */
+ fc_exact = 0;
+
temp = alloca(value_size);
if (a != result)
break;
default:
- _fail_char(old_str, len, str - old_str);
+ fail_char(old_str, len, str - old_str);
}
break;
goto done;
default:
- _fail_char(old_str, len, str - old_str);
+ fail_char(old_str, len, str - old_str);
}
break;
goto done;
default:
- _fail_char(old_str, len, str - old_str);
+ fail_char(old_str, len, str - old_str);
}
break;
exp_sign = 1;
/* fall through */
case '+':
- if (*(str-1) != 'e' && *(str-1) != 'E') _fail_char(old_str, len, str - old_str);
+ if (*(str-1) != 'e' && *(str-1) != 'E') fail_char(old_str, len, str - old_str);
str++;
break;
break;
default:
- _fail_char(old_str, len, str - old_str);
+ fail_char(old_str, len, str - old_str);
}
break;
case '\0': goto done;
default:
- _fail_char(old_str, len, str - old_str);
+ fail_char(old_str, len, str - old_str);
}
}
} /* switch(state) */
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);
+ sc_val_from_ulong(mant_size + ROUNDING_BITS, exp_val);
_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);
UINT32 sign, exponent, mantissa0, mantissa1;
srcval.d = l;
- bias_res = ((1<<exp_size)/2-1);
+ bias_res = ((1 << (exp_size - 1)) - 1);
#ifdef HAVE_LONG_DOUBLE
mant_val = 64;
if (result == NULL) result = calc_buffer;
temp = alloca(value_size);
+ /* CLEAR the buffer, else some bits might be uninitialised */
+ memset(result, 0, fc_get_buffer_length());
+
result->desc.exponent_size = exp_size;
result->desc.mantissa_size = mant_size;
if (exponent != 0) {
/* insert the hidden bit */
sc_val_from_ulong(1, temp);
- sc_val_from_ulong(mant_val + 2, NULL);
+ sc_val_from_ulong(mant_val + ROUNDING_BITS, NULL);
_shift_left(temp, sc_get_buffer(), NULL);
}
else
/* bits from the lower word */
sc_val_from_ulong(mantissa1, temp);
- sc_val_from_ulong(2, NULL);
+ sc_val_from_ulong(ROUNDING_BITS, NULL);
_shift_left(temp, sc_get_buffer(), temp);
sc_or(_mant(result), temp, _mant(result));
_shift_left(_mant(result), sc_get_buffer(), _mant(result));
}
- _normalize(result, result, 0);
+ normalize(result, result, 0);
TRACEPRINTF(("val_from_float results in %s\n", fc_print(result, temp, calc_buffer_size, FC_PACKED)));
* lead to wrong results */
exponent = sc_val_to_long(_exp(value)) ;
- sc_val_from_ulong(2, NULL);
+ sc_val_from_ulong(ROUNDING_BITS, NULL);
_shift_right(_mant(value), sc_get_buffer(), _mant(value));
mantissa0 = 0;
return result;
}
+ if (value->desc.clss == NAN) {
+ if (sc_get_highest_set_bit(_mant(value)) == value->desc.mantissa_size + 1)
+ return fc_get_qnan(exp_size, mant_size, result);
+ else
+ return fc_get_snan(exp_size, mant_size, result);
+ }
+
/* set the descriptor of the new value */
result->desc.exponent_size = exp_size;
result->desc.mantissa_size = mant_size;
memmove(_mant(result), _mant(value), value_size);
}
- _normalize(result, result, 0);
+ normalize(result, result, 0);
TRACEPRINTF(("Cast results in %s\n", fc_print(result, temp, value_size, FC_PACKED)));
return result;
}
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);
+ sc_val_from_ulong(ROUNDING_BITS, NULL);
_shift_left(_mant(result), sc_get_buffer(), _mant(result));
return result;
case FC_PACKED:
default:
- snprintf(buf, buflen, "%s", sc_print(_pack(val, mul_1), value_size*4, SC_HEX, 0));
+ snprintf(buf, buflen, "%s", sc_print(pack(val, mul_1), value_size*4, SC_HEX, 0));
buf[buflen - 1] = '\0';
break;
}
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 *pack = NULL;
+ static char *packed_value = NULL;
- if (pack == NULL) pack = xmalloc(value_size);
+ if (packed_value == NULL) packed_value = xmalloc(value_size);
if (value != NULL)
- _pack(value, pack);
+ pack(value, packed_value);
- return sc_sub_bits(pack, num_bits, byte_ofs);
+ 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_is_zero(_mant(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 non-zero if a given value can be converted lossless into another precision */
+int fc_can_lossless_conv_to(const fp_value *value, char exp_size, char mant_size) {
+ 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 << (exp_size - 1)) - 1;
+ v = fc_get_exponent(value) + exp_bias;
+ if (0 < v && v < (1 << exp_size) - 1) {
+ /* check the mantissa */
+ v = value->desc.mantissa_size + ROUNDING_BITS - sc_get_lowest_set_bit(_mant(value));
+ return v < mant_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)
* addition overflow */
max_precision = sc_get_precision() - 4;
if (max_precision < precision)
- printf("WARING: not enough precision available, using %d\n", max_precision);
+ printf("WARNING: not enough precision available, using %d\n", max_precision);
rounding_mode = FC_TONEAREST;
value_size = sc_get_buffer_length();
free(calc_buffer); calc_buffer = NULL;
}
+#ifdef FLTCALC_TRACE_CALC
+static char buffer[100];
+#endif
+
/* definition of interface functions */
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, 100, FC_PACKED)));
- TRACEPRINTF(("+ %s ", fc_print(b, buffer, 100, FC_PACKED)));
+ 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)
else
_fadd(a, b, result);
- TRACEPRINTF(("= %s\n", fc_print(result, buffer, 100, FC_PACKED)));
+ TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED)));
return result;
}
if (result == NULL) result = calc_buffer;
- TRACEPRINTF(("%s ", fc_print(a, buffer, 100, FC_PACKED)));
- TRACEPRINTF(("- %s ", fc_print(b, buffer, 100, FC_PACKED)));
+ 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);
else
_fadd(a, temp, result);
- TRACEPRINTF(("= %s\n", fc_print(result, buffer, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
- TRACEPRINTF(("* %s ", fc_print(b, buffer, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
- TRACEPRINTF(("/ %s ", fc_print(b, buffer, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
+ 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, 100, FC_PACKED)));
+ TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
TRACEPRINTF(("truncated to integer "));
_trunc(a, result);
- TRACEPRINTF(("= %s\n", fc_print(result, buffer, 100, FC_PACKED)));
+ 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;
- TRACEPRINTF(("%s ", fc_print(a, buffer, 100, FC_PACKED)));
+ (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, 100, FC_PACKED)));
+ 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;
+}