optimized bc_raise

Author: SakiTakamachi

ext/bcmath/libbcmath/src/raise.c

@@ -30,24 +30,163 @@
 *************************************************************************/
 
 #include "bcmath.h"
+#include "convert.h"
+#include "private.h"
 #include <assert.h>
 #include <stdbool.h>
 #include <stddef.h>
 
-void bc_square_ex(bc_num n1, bc_num *result, size_t scale_min) {
-	bc_num square_ex = bc_square(n1, scale_min);
-	bc_free_num(result);
-	*(result) = square_ex;
+static inline size_t bc_multiply_vector_ex(
+	BC_VECTOR **n1_vector, size_t n1_arr_size, BC_VECTOR *n2_vector, size_t n2_arr_size, BC_VECTOR **result_vector)
+{
+	size_t result_arr_size = n1_arr_size + n2_arr_size;
+	bc_multiply_vector(*n1_vector, n1_arr_size, n2_vector, n2_arr_size, *result_vector, result_arr_size);
+
+	/* Eliminate extra zeros because they increase the number of calculations. */
+	while ((*result_vector)[result_arr_size - 1] == 0) {
+		result_arr_size--;
+	}
+
+	/* Swap n1_vector and result_vector. */
+	BC_VECTOR *tmp = *n1_vector;
+	*n1_vector = *result_vector;
+	*result_vector = tmp;
+
+	return result_arr_size;
+}
+
+static inline size_t bc_square_vector_ex(BC_VECTOR **base_vector, size_t base_arr_size, BC_VECTOR **result_vector)
+{
+	return bc_multiply_vector_ex(base_vector, base_arr_size, *base_vector, base_arr_size, result_vector);
+}
+
+/* Use "exponentiation by squaring". This is the fast path when the results are small. */
+static inline bc_num bc_fast_raise(
+	const char *base_end, long exponent, size_t base_len, size_t power_len, size_t power_scale, size_t power_full_len)
+{
+	BC_VECTOR base_vector = 0;
+
+	/* Convert to BC_VECTOR[] */
+	bc_convert_to_vector(&base_vector, base_end, base_len);
+
+	while ((exponent & 1) == 0) {
+		base_vector *= base_vector;
+		exponent >>= 1;
+	}
+
+	/* copy base to power */
+	BC_VECTOR power_vector = base_vector;
+	exponent >>= 1;
+
+	while (exponent > 0) {
+		base_vector *= base_vector;
+		if ((exponent & 1) == 1) {
+			power_vector *= base_vector;
+		}
+		exponent >>= 1;
+	}
+
+	bc_num power = bc_new_num_nonzeroed(power_len, power_scale);
+	char *pptr = power->n_value;
+	char *pend = pptr + power_full_len - 1;
+
+	while (pend >= pptr) {
+		*pend-- = power_vector % BASE;
+		power_vector /= BASE;
+	}
+	return power;
+}
+
+/* Use "exponentiation by squaring". This is the standard path. */
+static bc_num bc_standard_raise(
+	const char *base_end, long exponent, size_t base_len, size_t power_scale)
+{
+	size_t base_arr_size = (base_len + BC_VECTOR_SIZE - 1) / BC_VECTOR_SIZE;
+	size_t max_power_arr_size =	base_arr_size * exponent;
+
+	/* The allocated memory area is reused on a rotational basis, so the same size is required. */
+	BC_VECTOR *buf = safe_emalloc(max_power_arr_size * 3, sizeof(BC_VECTOR), 0);
+	BC_VECTOR *base_vector = buf;
+	BC_VECTOR *power_vector = base_vector + max_power_arr_size;
+	BC_VECTOR *tmp_result_vector = power_vector + max_power_arr_size;
+
+	/* Convert to BC_VECTOR[] */
+	bc_convert_to_vector(base_vector, base_end, base_len);
+
+	while ((exponent & 1) == 0) {
+		base_arr_size = bc_square_vector_ex(&base_vector, base_arr_size, &tmp_result_vector);
+		exponent >>= 1;
+	}
+
+	/* copy base to power */
+	size_t power_arr_size = base_arr_size;
+	for (size_t i = 0; i < base_arr_size; i++) {
+		power_vector[i] = base_vector[i];
+	}
+	exponent >>= 1;
+
+	while (exponent > 0) {
+		base_arr_size = bc_square_vector_ex(&base_vector, base_arr_size, &tmp_result_vector);
+		if ((exponent & 1) == 1) {
+			power_arr_size = bc_multiply_vector_ex(&power_vector, power_arr_size, base_vector, base_arr_size, &tmp_result_vector);
+		}
+		exponent >>= 1;
+	}
+
+	/* Convert to bc_num */
+	size_t power_leading_zeros = 0;
+	size_t power_len;
+	size_t power_full_len = power_arr_size * BC_VECTOR_SIZE;
+	if (power_full_len > power_scale) {
+		power_len = power_full_len - power_scale;
+	} else {
+		power_len = 1;
+		power_leading_zeros = power_scale - power_full_len + 1;
+		power_full_len = power_scale + 1;
+	}
+	bc_num power = bc_new_num_nonzeroed(power_len, power_scale);
+
+	char *pptr = power->n_value;
+	char *pend = pptr + power_full_len - 1;
+
+	/* Pad with leading zeros if necessary. */
+	while (power_leading_zeros > sizeof(uint32_t)) {
+		bc_write_bcd_representation(0, pptr);
+		pptr += sizeof(uint32_t);
+		power_leading_zeros -= sizeof(uint32_t);
+	}
+	for (size_t i = 0; i < power_leading_zeros; i++) {
+		*pptr++ = 0;
+	}
+
+	size_t i = 0;
+	while (i < power_arr_size - 1) {
+#if BC_VECTOR_SIZE == 4
+		bc_write_bcd_representation(power_vector[i], pend - 3);
+		pend -= 4;
+#else
+		bc_write_bcd_representation(power_vector[i] / 10000, pend - 7);
+		bc_write_bcd_representation(power_vector[i] % 10000, pend - 3);
+		pend -= 8;
+#endif
+		i++;
+	}
+
+	while (pend >= pptr) {
+		*pend-- = power_vector[i] % BASE;
+		power_vector[i] /= BASE;
+	}
+
+	efree(buf);
+
+	return power;
 }
 
 /* Raise "base" to the "exponent" power.  The result is placed in RESULT.
    Maximum exponent is LONG_MAX.  If a "exponent" is not an integer,
    only the integer part is used.  */
 bool bc_raise(bc_num base, long exponent, bc_num *result, size_t scale) {
-	bc_num temp, power;
 	size_t rscale;
-	size_t pwrscale;
-	size_t calcscale;
 	bool is_neg;
 
 	/* Special case if exponent is a zero. */
@@ -74,43 +213,47 @@ bool bc_raise(bc_num base, long exponent, bc_num *result, size_t scale) {
 		return !is_neg;
 	}
 
-	/* Set initial value of temp. */
-	power = bc_copy_num(base);
-	pwrscale = base->n_scale;
-	while ((exponent & 1) == 0) {
-		pwrscale = 2 * pwrscale;
-		bc_square_ex(power, &power, pwrscale);
-		exponent = exponent >> 1;
+	size_t base_len = base->n_len + base->n_scale;
+	size_t power_len = base->n_len * exponent;
+	size_t power_scale = base->n_scale * exponent;
+	size_t power_full_len = power_len + power_scale;
+
+	sign power_sign;
+	if (base->n_sign == MINUS && (exponent & 1) == 1) {
+		power_sign = MINUS;
+	} else {
+		power_sign = PLUS;
 	}
-	temp = bc_copy_num(power);
-	calcscale = pwrscale;
-	exponent = exponent >> 1;
 
-	/* Do the calculation. */
-	while (exponent > 0) {
-		pwrscale = 2 * pwrscale;
-		bc_square_ex(power, &power, pwrscale);
-		if ((exponent & 1) == 1) {
-			calcscale = pwrscale + calcscale;
-			bc_multiply_ex(temp, power, &temp, calcscale);
-		}
-		exponent = exponent >> 1;
+	const char *base_end = base->n_value + base_len - 1;
+
+	bc_num power;
+	if (base_len <= BC_VECTOR_SIZE && power_full_len <= BC_VECTOR_SIZE * 2) {
+		power = bc_fast_raise(base_end, exponent, base_len, power_len, power_scale, power_full_len);
+	} else {
+		power = bc_standard_raise(base_end, exponent, base_len, power_scale);
+	}
+
+	_bc_rm_leading_zeros(power);
+	if (bc_is_zero(power)) {
+		power->n_sign = PLUS;
+		power->n_scale = 0;
+	} else {
+		power->n_sign = power_sign;
 	}
 
 	/* Assign the value. */
 	if (is_neg) {
-		if (bc_divide(BCG(_one_), temp, result, rscale) == false) {
-			bc_free_num (&temp);
+		if (bc_divide(BCG(_one_), power, result, rscale) == false) {
 			bc_free_num (&power);
 			return false;
 		}
-		bc_free_num (&temp);
+		bc_free_num (&power);
 	} else {
 		bc_free_num (result);
-		*result = temp;
+		*result = power;
 		(*result)->n_scale = MIN(scale, (*result)->n_scale);
 	}
-	bc_free_num (&power);
 	return true;
 }