Efficiently obtain the initial value for the standard path.

Author: SakiTakamachi

ext/bcmath/libbcmath/src/sqrt.c

@@ -140,20 +140,34 @@ static inline void bc_standard_sqrt(bc_num *num, size_t rscale, size_t num_calc_
 
 	bc_convert_to_vector_with_zero_pad(n_vector, nend, n_full_len, n_extend_zeros);
 
-	/* Prepare guess_vector (Temporary implementation) */
-	for (size_t i = 0; i < guess_vector_size - 2; i++) {
-		guess_vector[i] = BC_VECTOR_BOUNDARY_NUM - 1;
+	/* Prepare guess_vector. Use bc_fast_sqrt_vector to quickly obtain a highly accurate initial value. */
+	size_t n_top_len_for_initial_guess = SIZEOF_SIZE_T == 8 ? 18 : 10;
+
+	/* Set the number of digits of num to be used as the initial value for Newton's method.
+	 * Just as the square roots of 1000 and 100 differ significantly, the number of digits
+	 * to "ignore" here must be even. */
+	if (num_calc_full_len & 1) {
+		n_top_len_for_initial_guess--;
 	}
-	if (guess_full_len % BC_VECTOR_SIZE == 0) {
-		guess_vector[guess_vector_size - 2] = BC_VECTOR_BOUNDARY_NUM - 1;
-	} else {
-		guess_vector[guess_vector_size - 2] = 0;
-		for (size_t i = 0; i < guess_full_len % BC_VECTOR_SIZE; i++) {
-			guess_vector[guess_vector_size - 2] *= BASE;
-			guess_vector[guess_vector_size - 2] += 9;
-		}
+	const char *nptr = (*num)->n_value;
+	BC_VECTOR n_top = 0;
+	for (size_t i = 0; i < n_top_len_for_initial_guess; i++) {
+		n_top *= BASE;
+		n_top += *nptr++;
+	}
+
+	for (size_t i = 0; i < guess_vector_size; i++) {
+		guess_vector[i] = 0;
 	}
-	guess_vector[guess_vector_size - 1] = 0;
+	BC_VECTOR initial_guess = bc_fast_sqrt_vector(n_top);
+
+	size_t initial_guess_len = SIZEOF_SIZE_T == 8 ? 9 : 5;
+	size_t guess_top_vector_len = guess_full_len % BC_VECTOR_SIZE == 0 ? BC_VECTOR_SIZE : guess_full_len % BC_VECTOR_SIZE;
+	size_t guess_len_diff = initial_guess_len - guess_top_vector_len;
+	guess_vector[guess_vector_size - 2] = initial_guess / BC_POW_10_LUT[guess_len_diff];
+	initial_guess %= BC_POW_10_LUT[guess_len_diff];
+	guess_vector[guess_vector_size - 3] = initial_guess * BC_POW_10_LUT[BC_VECTOR_SIZE - guess_len_diff];
+
 	guess1_vector[guess_vector_size - 1] = 0;
 
 	size_t quot_size = n_vector_size - (guess_vector_size - 1) + 1;