bigdecimal 2.0.0 → 4.1.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,219 @@
1
+ struct vp_settings {
2
+ size_t prec_limit;
3
+ unsigned short rounding_mode;
4
+ };
5
+
6
+ static VALUE
7
+ restore_vp_settings(VALUE saved_ptr)
8
+ {
9
+ struct vp_settings *saved = (struct vp_settings *)saved_ptr;
10
+ VpSetPrecLimit(saved->prec_limit);
11
+ VpSetRoundMode(saved->rounding_mode);
12
+ return Qnil;
13
+ }
14
+
15
+ // Calculate the inverse of x using the Newton-Raphson method.
16
+ // Assumes no precision limit and ROUND_HALF_UP. See VpDivdNewton.
17
+ static VALUE
18
+ newton_raphson_inverse(VALUE x, size_t prec) {
19
+ BDVALUE bdone = NewZeroWrap(1, 1);
20
+ VpSetOne(bdone.real);
21
+ VALUE one = bdone.bigdecimal;
22
+
23
+ // Initial approximation: 10^18 / (leading 9 digits of x), calculated in 64-bit integer.
24
+ // Truncating x to 9 digits and the quotient to an integer keep the relative error below 1.2e-8,
25
+ // so it has at least 7 correct digits.
26
+ const size_t initial_digits = 7;
27
+ const DECDIG_DBL base_sq = (DECDIG_DBL)BIGDECIMAL_BASE * BIGDECIMAL_BASE;
28
+ BDVALUE bdx = GetBDValueMust(x);
29
+ DECDIG_DBL x_lead = (DECDIG_DBL)bdx.real->frac[0] * BIGDECIMAL_BASE + (bdx.real->Prec >= 2 ? bdx.real->frac[1] : 0);
30
+ int shift = 0;
31
+ while (x_lead < base_sq / 10) {
32
+ x_lead *= 10;
33
+ shift++;
34
+ }
35
+ // Dividing base_sq - 1 instead of base_sq keeps the quotient below 10 * BASE,
36
+ // so that frac[0] of inv0 never reaches BASE.
37
+ DECDIG_DBL inv_lead = (base_sq - 1) / (x_lead / BIGDECIMAL_BASE);
38
+ for (int i = 0; i < shift; i++) inv_lead *= 10;
39
+ BDVALUE inv0 = NewZeroWrap(1, 2 * BIGDECIMAL_COMPONENT_FIGURES);
40
+ VpSetOne(inv0.real);
41
+ inv0.real->frac[0] = (DECDIG)(inv_lead / BIGDECIMAL_BASE);
42
+ inv0.real->frac[1] = (DECDIG)(inv_lead % BIGDECIMAL_BASE);
43
+ inv0.real->Prec = 2;
44
+ inv0.real->exponent = 1 - bdx.real->exponent;
45
+ VpNmlz(inv0.real);
46
+ RB_GC_GUARD(bdx.bigdecimal);
47
+ VALUE inv = inv0.bigdecimal;
48
+
49
+ int bl = 1;
50
+ while (((size_t)1 << bl) < prec) bl++;
51
+
52
+ // Each iteration doubles the number of correct digits, and the rounding error of
53
+ // the previous iteration is squared into the new digits. Margin digits keep the
54
+ // squared error below the last digit, otherwise it can grow over iterations.
55
+ const size_t margin = 4;
56
+ for (int i = bl; i >= 0; i--) {
57
+ size_t n = (prec >> i) + margin;
58
+ if (n > prec) n = prec;
59
+ // inv0 already has this precision. The last iteration is kept to round inv to prec digits.
60
+ if (n <= initial_digits && i > 0) continue;
61
+ // Newton-Raphson iteration: inv_next = inv + inv * (1 - x * inv)
62
+ // (1 - x * inv) is about 10^(-n/2), so calculating it in n/2 digits is enough for inv_next in n digits.
63
+ VALUE one_minus_x_inv = BigDecimal_sub2(
64
+ one,
65
+ BigDecimal_mult(BigDecimal_mult2(x, one, SIZET2NUM(n + 1)), inv),
66
+ SIZET2NUM(n / 2 + margin)
67
+ );
68
+ inv = BigDecimal_add2(
69
+ inv,
70
+ BigDecimal_mult(inv, one_minus_x_inv),
71
+ SIZET2NUM(n)
72
+ );
73
+ }
74
+ return inv;
75
+ }
76
+
77
+ // Calculates divmod by multiplying approximate reciprocal of y
78
+ static void
79
+ divmod_by_inv_mul(VALUE x, VALUE y, VALUE inv, VALUE *res_div, VALUE *res_mod) {
80
+ VALUE div = BigDecimal_fix(BigDecimal_mult(x, inv));
81
+ VALUE mod = BigDecimal_sub(x, BigDecimal_mult(div, y));
82
+ while (RTEST(BigDecimal_lt(mod, INT2FIX(0)))) {
83
+ mod = BigDecimal_add(mod, y);
84
+ div = BigDecimal_sub(div, INT2FIX(1));
85
+ }
86
+ while (RTEST(BigDecimal_ge(mod, y))) {
87
+ mod = BigDecimal_sub(mod, y);
88
+ div = BigDecimal_add(div, INT2FIX(1));
89
+ }
90
+ *res_div = div;
91
+ *res_mod = mod;
92
+ }
93
+
94
+ static void
95
+ slice_copy(DECDIG *dest, Real *src, size_t rshift, size_t length) {
96
+ ssize_t start = src->exponent - (ssize_t)rshift - (ssize_t)length;
97
+ ssize_t from = Max(start, 0);
98
+ ssize_t to = Min(start + (ssize_t)length, (ssize_t)src->Prec);
99
+ if (from >= to) return;
100
+ memcpy(dest + (from - start), src->frac + from, (size_t)(to - from) * sizeof(DECDIG));
101
+ }
102
+
103
+ /* Calculates divmod using Newton-Raphson method.
104
+ * x and y must be a BigDecimal representing an integer value.
105
+ *
106
+ * To calculate with low cost, we need to split x into blocks and perform divmod for each block.
107
+ * x_digits = remaining_digits(<= y_digits) + block_digits * num_blocks
108
+ *
109
+ * Example:
110
+ * xxx_xxxxx_xxxxx_xxxxx(18 digits) / yyyyy(5 digits)
111
+ * remaining_digits = 3, block_digits = 5, num_blocks = 3
112
+ * repeating xxxxx_xxxxxx.divmod(yyyyy) calculation 3 times.
113
+ *
114
+ * In each divmod step, dividend is at most (y_digits + block_digits) digits and divisor is y_digits digits.
115
+ * Reciprocal of y needs block_digits + 1 precision.
116
+ */
117
+ static void
118
+ divmod_newton(VALUE x, VALUE y, VALUE *div_out, VALUE *mod_out) {
119
+ size_t x_digits = NUM2SIZET(BigDecimal_exponent(x));
120
+ size_t y_digits = NUM2SIZET(BigDecimal_exponent(y));
121
+ if (x_digits <= y_digits) x_digits = y_digits + 1;
122
+
123
+ size_t n = x_digits / y_digits;
124
+ size_t block_figs = (x_digits - y_digits) / n / BIGDECIMAL_COMPONENT_FIGURES + 1;
125
+ size_t block_digits = block_figs * BIGDECIMAL_COMPONENT_FIGURES;
126
+ size_t num_blocks = (x_digits - y_digits + block_digits - 1) / block_digits;
127
+ size_t y_figs = (y_digits - 1) / BIGDECIMAL_COMPONENT_FIGURES + 1;
128
+ VALUE yinv = newton_raphson_inverse(y, block_digits + 1);
129
+
130
+ BDVALUE divident = NewZeroWrap(1, BIGDECIMAL_COMPONENT_FIGURES * (y_figs + block_figs));
131
+ BDVALUE div_result = NewZeroWrap(1, BIGDECIMAL_COMPONENT_FIGURES * (num_blocks * block_figs + 1));
132
+ BDVALUE bdx = GetBDValueMust(x);
133
+
134
+ VALUE mod = BigDecimal_fix(BigDecimal_decimal_shift(x, SSIZET2NUM(-(ssize_t)(num_blocks * block_digits))));
135
+ for (ssize_t i = (ssize_t)(num_blocks - 1); i >= 0; i--) {
136
+ memset(divident.real->frac, 0, (y_figs + block_figs) * sizeof(DECDIG));
137
+
138
+ BDVALUE bdmod = GetBDValueMust(mod);
139
+ slice_copy(divident.real->frac, bdmod.real, 0, y_figs);
140
+ slice_copy(divident.real->frac + y_figs, bdx.real, (size_t)i * block_figs, block_figs);
141
+ RB_GC_GUARD(bdmod.bigdecimal);
142
+
143
+ VpSetSign(divident.real, 1);
144
+ divident.real->exponent = (ssize_t)(y_figs + block_figs);
145
+ divident.real->Prec = y_figs + block_figs;
146
+ VpNmlz(divident.real);
147
+
148
+ VALUE div;
149
+ divmod_by_inv_mul(divident.bigdecimal, y, yinv, &div, &mod);
150
+ BDVALUE bddiv = GetBDValueMust(div);
151
+ slice_copy(div_result.real->frac + (num_blocks - (size_t)i - 1) * block_figs, bddiv.real, 0, block_figs + 1);
152
+ RB_GC_GUARD(bddiv.bigdecimal);
153
+ }
154
+ VpSetSign(div_result.real, 1);
155
+ div_result.real->exponent = (ssize_t)(num_blocks * block_figs + 1);
156
+ div_result.real->Prec = num_blocks * block_figs + 1;
157
+ VpNmlz(div_result.real);
158
+ RB_GC_GUARD(bdx.bigdecimal);
159
+ RB_GC_GUARD(divident.bigdecimal);
160
+ RB_GC_GUARD(div_result.bigdecimal);
161
+ *div_out = div_result.bigdecimal;
162
+ *mod_out = mod;
163
+ }
164
+
165
+ struct vp_divd_newton_args {
166
+ Real *c, *r, *a, *b;
167
+ };
168
+
169
+ static VALUE
170
+ VpDivdNewtonInner(VALUE args_ptr)
171
+ {
172
+ struct vp_divd_newton_args *args = (struct vp_divd_newton_args *)args_ptr;
173
+ Real *c = args->c, *r = args->r, *a = args->a, *b = args->b;
174
+ BDVALUE a2, b2, c2, r2;
175
+ VALUE div, mod, a2_frac = Qnil;
176
+ size_t div_prec = c->MaxPrec - 1;
177
+ size_t base_prec = b->Prec;
178
+
179
+ a2 = NewZeroWrap(1, a->Prec * BIGDECIMAL_COMPONENT_FIGURES);
180
+ b2 = NewZeroWrap(1, b->Prec * BIGDECIMAL_COMPONENT_FIGURES);
181
+ VpAsgn(a2.real, a, 1);
182
+ VpAsgn(b2.real, b, 1);
183
+ VpSetSign(a2.real, 1);
184
+ VpSetSign(b2.real, 1);
185
+ a2.real->exponent = (ssize_t)(base_prec + div_prec);
186
+ b2.real->exponent = (ssize_t)base_prec;
187
+
188
+ if ((ssize_t)a2.real->Prec > a2.real->exponent) {
189
+ a2_frac = BigDecimal_frac(a2.bigdecimal);
190
+ VpMidRound(a2.real, VP_ROUND_DOWN, 0);
191
+ }
192
+ divmod_newton(a2.bigdecimal, b2.bigdecimal, &div, &mod);
193
+ if (a2_frac != Qnil) mod = BigDecimal_add(mod, a2_frac);
194
+
195
+ c2 = GetBDValueMust(div);
196
+ r2 = GetBDValueMust(mod);
197
+ VpAsgn(c, c2.real, VpGetSign(a) * VpGetSign(b));
198
+ VpAsgn(r, r2.real, VpGetSign(a));
199
+ if (!VpIsZero(c)) AddExponent(c, a->exponent - b->exponent - (ssize_t)div_prec);
200
+ if (!VpIsZero(r)) AddExponent(r, a->exponent - (ssize_t)(base_prec + div_prec));
201
+ RB_GC_GUARD(a2.bigdecimal);
202
+ RB_GC_GUARD(b2.bigdecimal);
203
+ RB_GC_GUARD(c2.bigdecimal);
204
+ RB_GC_GUARD(r2.bigdecimal);
205
+ return Qnil;
206
+ }
207
+
208
+ // Newton-Raphson iteration converges from below, so a rounding mode that rounds toward zero
209
+ // (ROUND_DOWN, ROUND_FLOOR) adds error in the same direction at every iteration.
210
+ // Calculate with ROUND_HALF_UP and without precision limit, and restore them even if an exception is raised.
211
+ static void
212
+ VpDivdNewton(Real *c, Real *r, Real *a, Real *b)
213
+ {
214
+ struct vp_divd_newton_args args = {c, r, a, b};
215
+ struct vp_settings saved = {VpGetPrecLimit(), VpGetRoundMode()};
216
+ VpSetPrecLimit(0);
217
+ VpSetRoundMode(VP_ROUND_HALF_UP);
218
+ rb_ensure(VpDivdNewtonInner, (VALUE)&args, restore_vp_settings, (VALUE)&saved);
219
+ }
@@ -1,47 +1,54 @@
1
- # frozen_string_literal: false
1
+ # frozen_string_literal: true
2
2
  require 'mkmf'
3
3
 
4
- def check_bigdecimal_version(gemspec_path)
5
- message "checking RUBY_BIGDECIMAL_VERSION... "
4
+ def have_builtin_func(name, check_expr, opt = "", &b)
5
+ checking_for checking_message(name.funcall_style, nil, opt) do
6
+ if try_compile(<<SRC, opt, &b)
7
+ int foo;
8
+ int main() { #{check_expr}; return 0; }
9
+ SRC
10
+ $defs.push(format("-DHAVE_BUILTIN_%s", name.tr_cpp))
11
+ true
12
+ else
13
+ false
14
+ end
15
+ end
16
+ end
6
17
 
7
- bigdecimal_version =
8
- IO.readlines(gemspec_path)
9
- .grep(/\Abigdecimal_version\s+=\s+/)[0][/\'([^\']+)\'/, 1]
18
+ have_builtin_func("__builtin_clz", "__builtin_clz(0)")
19
+ have_builtin_func("__builtin_clzl", "__builtin_clzl(0)")
20
+ have_builtin_func("__builtin_clzll", "__builtin_clzll(0)")
10
21
 
11
- version_components = bigdecimal_version.split('.')
12
- bigdecimal_version = version_components[0, 3].join('.')
13
- bigdecimal_version << "-#{version_components[3]}" if version_components[3]
14
- $defs << %Q[-DRUBY_BIGDECIMAL_VERSION=\\"#{bigdecimal_version}\\"]
22
+ have_header("float.h")
23
+ have_header("math.h")
24
+ have_header("stdbool.h")
25
+ have_header("stdlib.h")
15
26
 
16
- message "#{bigdecimal_version}\n"
27
+ if have_header("x86intrin.h")
28
+ have_func("_lzcnt_u32", "x86intrin.h")
29
+ have_func("_lzcnt_u64", "x86intrin.h")
17
30
  end
18
31
 
19
- gemspec_name = gemspec_path = nil
20
- unless ['', '../../'].any? {|dir|
21
- gemspec_name = "#{dir}bigdecimal.gemspec"
22
- gemspec_path = File.expand_path("../#{gemspec_name}", __FILE__)
23
- File.file?(gemspec_path)
24
- }
25
- $stderr.puts "Unable to find bigdecimal.gemspec"
26
- abort
32
+ if have_header("intrin.h")
33
+ have_func("__lzcnt", "intrin.h")
34
+ have_func("__lzcnt64", "intrin.h")
35
+ have_func("_BitScanReverse", "intrin.h")
36
+ have_func("_BitScanReverse64", "intrin.h")
27
37
  end
28
38
 
29
- check_bigdecimal_version(gemspec_path)
30
-
31
- have_func("labs", "stdlib.h")
32
- have_func("llabs", "stdlib.h")
33
- have_func("finite", "math.h")
34
- have_func("isfinite", "math.h")
39
+ have_header("ruby/atomic.h")
40
+ have_header("ruby/internal/has/builtin.h")
41
+ have_header("ruby/internal/static_assert.h")
35
42
 
36
- have_type("struct RRational", "ruby.h")
37
- have_func("rb_rational_num", "ruby.h")
38
- have_func("rb_rational_den", "ruby.h")
39
- have_type("struct RComplex", "ruby.h")
40
43
  have_func("rb_complex_real", "ruby.h")
41
44
  have_func("rb_complex_imag", "ruby.h")
42
- have_func("rb_array_const_ptr", "ruby.h")
43
- have_func("rb_sym2str", "ruby.h")
44
45
  have_func("rb_opts_exception_p", "ruby.h")
46
+ have_func("rb_category_warn", "ruby.h")
47
+ have_const("RB_WARN_CATEGORY_DEPRECATED", "ruby.h")
48
+
49
+ if RUBY_ENGINE == "ruby"
50
+ have_const("RUBY_TYPED_EMBEDDABLE", "ruby.h") # RUBY_VERSION >= 3.3
51
+ end
45
52
 
46
53
  if File.file?(File.expand_path('../lib/bigdecimal.rb', __FILE__))
47
54
  bigdecimal_rb = "$(srcdir)/lib/bigdecimal.rb"
@@ -49,7 +56,8 @@ else
49
56
  bigdecimal_rb = "$(srcdir)/../../lib/bigdecimal.rb"
50
57
  end
51
58
 
59
+ $defs.push '-DBIGDECIMAL_USE_VP_TEST_METHODS' if ENV['BIGDECIMAL_USE_VP_TEST_METHODS'] == 'true'
60
+
52
61
  create_makefile('bigdecimal') {|mf|
53
- mf << "GEMSPEC = #{gemspec_name}\n"
54
62
  mf << "BIGDECIMAL_RB = #{bigdecimal_rb}\n"
55
63
  }
@@ -0,0 +1,68 @@
1
+ #ifndef BIGDECIMAL_HAS_FEATURE_H
2
+ #define BIGDECIMAL_HAS_FEATURE_H
3
+
4
+ /* ======== __has_feature ======== */
5
+
6
+ #ifndef __has_feature
7
+ # define __has_feature(_) 0
8
+ #endif
9
+
10
+ /* ======== __has_extension ======== */
11
+
12
+ #ifndef __has_extension
13
+ # define __has_extension __has_feature
14
+ #endif
15
+
16
+ /* ======== __has_builtin ======== */
17
+
18
+ #ifdef HAVE_RUBY_INTERNAL_HAS_BUILTIN_H
19
+ # include <ruby/internal/has/builtin.h>
20
+ #endif
21
+
22
+ #ifdef RBIMPL_HAS_BUILTIN
23
+ # define BIGDECIMAL_HAS_BUILTIN(...) RBIMPL_HAS_BUILTIN(__VA_ARGS__)
24
+
25
+ #else
26
+ # /* The following section is copied from CRuby's builtin.h */
27
+ #
28
+ # ifdef __has_builtin
29
+ # if defined(__INTEL_COMPILER)
30
+ # /* :TODO: Intel C Compiler has __has_builtin (since 19.1 maybe?), and is
31
+ # * reportedly broken. We have to skip them. However the situation can
32
+ # * change. They might improve someday. We need to revisit here later. */
33
+ # elif defined(__GNUC__) && ! __has_builtin(__builtin_alloca)
34
+ # /* FreeBSD's <sys/cdefs.h> defines its own *broken* version of
35
+ # * __has_builtin. Cygwin copied that content to be a victim of the
36
+ # * broken-ness. We don't take them into account. */
37
+ # else
38
+ # define HAVE___HAS_BUILTIN 1
39
+ # endif
40
+ # endif
41
+ #
42
+ # if defined(HAVE___HAS_BUILTIN)
43
+ # define BIGDECIMAL_HAS_BUILTIN(_) __has_builtin(_)
44
+ #
45
+ # elif defined(__GNUC__)
46
+ # define BIGDECIMAL_HAS_BUILTIN(_) BIGDECIMAL_HAS_BUILTIN_ ## _
47
+ # if defined(__GNUC__) && (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 6))
48
+ # define BIGDECIMAL_HAS_BUILTIN___builtin_clz 1
49
+ # define BIGDECIMAL_HAS_BUILTIN___builtin_clzl 1
50
+ # else
51
+ # define BIGDECIMAL_HAS_BUILTIN___builtin_clz 0
52
+ # define BIGDECIMAL_HAS_BUILTIN___builtin_clzl 0
53
+ # endif
54
+ # elif defined(_MSC_VER)
55
+ # define BIGDECIMAL_HAS_BUILTIN(_) 0
56
+ #
57
+ # else
58
+ # define BIGDECIMAL_HAS_BUILTIN(_) BIGDECIMAL_HAS_BUILTIN_ ## _
59
+ # define BIGDECIMAL_HAS_BUILTIN___builtin_clz HAVE_BUILTIN___BUILTIN_CLZ
60
+ # define BIGDECIMAL_HAS_BUILTIN___builtin_clzl HAVE_BUILTIN___BUILTIN_CLZL
61
+ # endif
62
+ #endif /* RBIMPL_HAS_BUILTIN */
63
+
64
+ #ifndef __has_builtin
65
+ # define __has_builtin(...) BIGDECIMAL_HAS_BUILTIN(__VA_ARGS__)
66
+ #endif
67
+
68
+ #endif /* BIGDECIMAL_HAS_FEATURE_H */