create fast path and move old logic to standard path

Author: SakiTakamachi

ext/bcmath/libbcmath/src/sqrt.c

@@ -32,37 +32,89 @@
 #include "bcmath.h"
 #include <stdbool.h>
 #include <stddef.h>
+#include "private.h"
 
 /* Take the square root NUM and return it in NUM with SCALE digits
    after the decimal place. */
 
-bool bc_sqrt(bc_num *num, size_t scale)
+static inline BC_VECTOR bc_sqrt_get_pow_10(size_t exponent)
 {
-	/* Initial checks. */
-	if (bc_is_neg(*num)) {
-		/* Cannot take the square root of a negative number */
-		return false;
-	}
-	/* Square root of 0 is 0 */
-	if (bc_is_zero(*num)) {
-		bc_free_num (num);
-		*num = bc_copy_num(BCG(_zero_));
-		return true;
+	BC_VECTOR value = 1;
+	while (exponent >= 8) {
+		value *= BC_POW_10_LUT[8];
+		exponent -= 8;
 	}
+	value *= BC_POW_10_LUT[exponent];
+	return value;
+}
 
-	bcmath_compare_result num_cmp_one = bc_compare(*num, BCG(_one_), (*num)->n_scale);
-	/* Square root of 1 is 1 */
-	if (num_cmp_one == BCMATH_EQUAL) {
-		bc_free_num (num);
-		*num = bc_copy_num(BCG(_one_));
-		return true;
+static BC_VECTOR bc_fast_sqrt_vector(BC_VECTOR n_vector)
+{
+	/* Use a bitwise method for approximating the square root
+	 * as the initial guess for Newton's method. */
+	union {
+		uint64_t i;
+		double d;
+	} u;
+	u.d = (double) n_vector;
+	u.i = (1ULL << 61) + (u.i >> 1) - (1ULL << 50);
+	BC_VECTOR guess_vector = (BC_VECTOR) u.d;
+
+	/* Newton's algorithm. */
+	BC_VECTOR guess1_vector;
+	size_t diff;
+	do {
+		guess1_vector = guess_vector;
+		guess_vector = n_vector / guess_vector;
+		guess_vector += guess1_vector;
+		guess_vector /= 2;
+		diff = guess1_vector > guess_vector ? guess1_vector - guess_vector : guess_vector - guess1_vector;
+	} while (diff > 1);
+	return guess_vector;
+}
+
+static inline void *bc_fast_sqrt(bc_num *num, size_t rscale)
+{
+	BC_VECTOR n_vector = 0;
+	size_t i = 0;
+	for (; i < (*num)->n_len + (*num)->n_scale; i++) {
+		n_vector = n_vector * BASE + (*num)->n_value[i];
 	}
+	/* When calculating the square root of a number using only integer operations,
+		* need to adjust the digit scale accordingly.
+		* Considering that the original number is the square of the result,
+		* if the desired scale of the result is 5, the input number should be scaled
+		* by twice that, i.e., scale 10. */
+	n_vector *= bc_sqrt_get_pow_10((rscale + 1) * 2 - (*num)->n_scale);
+
+	/* Get sqrt */
+	BC_VECTOR guess_vector = bc_fast_sqrt_vector(n_vector);
+
+	size_t full_len = 0;
+	BC_VECTOR tmp_guess_vector = guess_vector;
+	do {
+		tmp_guess_vector /= BASE;
+		full_len++;
+	} while (tmp_guess_vector > 0);
+
+	size_t ret_ren = full_len > rscale + 1 ? full_len - (rscale + 1) : 1; /* for int zero */
+	bc_num ret = bc_new_num_nonzeroed(ret_ren, rscale);
+	char *rptr = ret->n_value;
+	char *rend = rptr + ret_ren + rscale - 1;
+
+	guess_vector /= BASE; /* Since the scale of guess_vector is rscale + 1, reduce the scale by 1. */
+	while (rend >= rptr) {
+		*rend-- = guess_vector % BASE;
+		guess_vector /= BASE;
+	}
+	bc_free_num(num);
+	*num = ret;
+}
 
-	/* Initialize the variables. */
-	size_t cscale;
+static inline void bc_standard_sqrt(bc_num *num, size_t rscale, bcmath_compare_result num_cmp_one)
+{
 	bc_num guess;
-	size_t rscale = MAX(scale, (*num)->n_scale);
-
+	size_t cscale;
 	/* Calculate the initial guess. */
 	if (num_cmp_one == BCMATH_RIGHT_GREATER) {
 		/* The number is between 0 and 1.  Guess should start at 1. */
@@ -84,7 +136,6 @@ bool bc_sqrt(bc_num *num, size_t scale)
 	point5->n_value[1] = 5;
 	bc_num diff = NULL;
 
-	/* Find the square root using Newton's algorithm. */
 	bool done = false;
 	while (!done) {
 		bc_free_num (&guess1);
@@ -109,5 +160,40 @@ bool bc_sqrt(bc_num *num, size_t scale)
 	bc_free_num (&guess1);
 	bc_free_num (&point5);
 	bc_free_num (&diff);
+}
+
+bool bc_sqrt(bc_num *num, size_t scale)
+{
+	/* Initial checks. */
+	if (bc_is_neg(*num)) {
+		/* Cannot take the square root of a negative number */
+		return false;
+	}
+	/* Square root of 0 is 0 */
+	if (bc_is_zero(*num)) {
+		bc_free_num (num);
+		*num = bc_copy_num(BCG(_zero_));
+		return true;
+	}
+
+	bcmath_compare_result num_cmp_one = bc_compare(*num, BCG(_one_), (*num)->n_scale);
+	/* Square root of 1 is 1 */
+	if (num_cmp_one == BCMATH_EQUAL) {
+		bc_free_num (num);
+		*num = bc_copy_num(BCG(_one_));
+		return true;
+	}
+
+	/* Initialize the variables. */
+	size_t rscale = MAX(scale, (*num)->n_scale);
+	size_t num_calc_full_len = (*num)->n_len + (rscale + 1) * 2;
+
+	/* Find the square root using Newton's algorithm. */
+	if (num_calc_full_len < MAX_LENGTH_OF_LONG) {
+		bc_fast_sqrt(num, rscale);
+	} else {
+		bc_standard_sqrt(num, rscale, num_cmp_one);
+	}
+
 	return true;
 }