summaryrefslogtreecommitdiffstats
path: root/src/common/d2s_intrinsics.h
blob: ae0f28dbb2f75c5f0b25f71fe46b4999958bd604 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
/*---------------------------------------------------------------------------
 *
 * Ryu floating-point output for double precision.
 *
 * Portions Copyright (c) 2018-2023, PostgreSQL Global Development Group
 *
 * IDENTIFICATION
 *	  src/common/d2s_intrinsics.h
 *
 * This is a modification of code taken from github.com/ulfjack/ryu under the
 * terms of the Boost license (not the Apache license). The original copyright
 * notice follows:
 *
 * Copyright 2018 Ulf Adams
 *
 * The contents of this file may be used under the terms of the Apache
 * License, Version 2.0.
 *
 *     (See accompanying file LICENSE-Apache or copy at
 *      http://www.apache.org/licenses/LICENSE-2.0)
 *
 * Alternatively, the contents of this file may be used under the terms of the
 * Boost Software License, Version 1.0.
 *
 *     (See accompanying file LICENSE-Boost or copy at
 *      https://www.boost.org/LICENSE_1_0.txt)
 *
 * Unless required by applicable law or agreed to in writing, this software is
 * distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
 * KIND, either express or implied.
 *
 *---------------------------------------------------------------------------
 */
#ifndef RYU_D2S_INTRINSICS_H
#define RYU_D2S_INTRINSICS_H

#if defined(HAS_64_BIT_INTRINSICS)

#include <intrin.h>

static inline uint64
umul128(const uint64 a, const uint64 b, uint64 *const productHi)
{
	return _umul128(a, b, productHi);
}

static inline uint64
shiftright128(const uint64 lo, const uint64 hi, const uint32 dist)
{
	/*
	 * For the __shiftright128 intrinsic, the shift value is always modulo 64.
	 * In the current implementation of the double-precision version of Ryu,
	 * the shift value is always < 64. (In the case RYU_OPTIMIZE_SIZE == 0,
	 * the shift value is in the range [49, 58]. Otherwise in the range [2,
	 * 59].) Check this here in case a future change requires larger shift
	 * values. In this case this function needs to be adjusted.
	 */
	Assert(dist < 64);
	return __shiftright128(lo, hi, (unsigned char) dist);
}

#else							/* defined(HAS_64_BIT_INTRINSICS) */

static inline uint64
umul128(const uint64 a, const uint64 b, uint64 *const productHi)
{
	/*
	 * The casts here help MSVC to avoid calls to the __allmul library
	 * function.
	 */
	const uint32 aLo = (uint32) a;
	const uint32 aHi = (uint32) (a >> 32);
	const uint32 bLo = (uint32) b;
	const uint32 bHi = (uint32) (b >> 32);

	const uint64 b00 = (uint64) aLo * bLo;
	const uint64 b01 = (uint64) aLo * bHi;
	const uint64 b10 = (uint64) aHi * bLo;
	const uint64 b11 = (uint64) aHi * bHi;

	const uint32 b00Lo = (uint32) b00;
	const uint32 b00Hi = (uint32) (b00 >> 32);

	const uint64 mid1 = b10 + b00Hi;
	const uint32 mid1Lo = (uint32) (mid1);
	const uint32 mid1Hi = (uint32) (mid1 >> 32);

	const uint64 mid2 = b01 + mid1Lo;
	const uint32 mid2Lo = (uint32) (mid2);
	const uint32 mid2Hi = (uint32) (mid2 >> 32);

	const uint64 pHi = b11 + mid1Hi + mid2Hi;
	const uint64 pLo = ((uint64) mid2Lo << 32) + b00Lo;

	*productHi = pHi;
	return pLo;
}

static inline uint64
shiftright128(const uint64 lo, const uint64 hi, const uint32 dist)
{
	/* We don't need to handle the case dist >= 64 here (see above). */
	Assert(dist < 64);
#if !defined(RYU_32_BIT_PLATFORM)
	Assert(dist > 0);
	return (hi << (64 - dist)) | (lo >> dist);
#else
	/* Avoid a 64-bit shift by taking advantage of the range of shift values. */
	Assert(dist >= 32);
	return (hi << (64 - dist)) | ((uint32) (lo >> 32) >> (dist - 32));
#endif
}

#endif							/* // defined(HAS_64_BIT_INTRINSICS) */

#ifdef RYU_32_BIT_PLATFORM

/*  Returns the high 64 bits of the 128-bit product of a and b. */
static inline uint64
umulh(const uint64 a, const uint64 b)
{
	/*
	 * Reuse the umul128 implementation. Optimizers will likely eliminate the
	 * instructions used to compute the low part of the product.
	 */
	uint64		hi;

	umul128(a, b, &hi);
	return hi;
}

/*----
 *  On 32-bit platforms, compilers typically generate calls to library
 *  functions for 64-bit divisions, even if the divisor is a constant.
 *
 *  E.g.:
 *  https://bugs.llvm.org/show_bug.cgi?id=37932
 *  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=17958
 *  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=37443
 *
 *  The functions here perform division-by-constant using multiplications
 *  in the same way as 64-bit compilers would do.
 *
 *  NB:
 *  The multipliers and shift values are the ones generated by clang x64
 *  for expressions like x/5, x/10, etc.
 *----
 */

static inline uint64
div5(const uint64 x)
{
	return umulh(x, UINT64CONST(0xCCCCCCCCCCCCCCCD)) >> 2;
}

static inline uint64
div10(const uint64 x)
{
	return umulh(x, UINT64CONST(0xCCCCCCCCCCCCCCCD)) >> 3;
}

static inline uint64
div100(const uint64 x)
{
	return umulh(x >> 2, UINT64CONST(0x28F5C28F5C28F5C3)) >> 2;
}

static inline uint64
div1e8(const uint64 x)
{
	return umulh(x, UINT64CONST(0xABCC77118461CEFD)) >> 26;
}

#else							/* RYU_32_BIT_PLATFORM */

static inline uint64
div5(const uint64 x)
{
	return x / 5;
}

static inline uint64
div10(const uint64 x)
{
	return x / 10;
}

static inline uint64
div100(const uint64 x)
{
	return x / 100;
}

static inline uint64
div1e8(const uint64 x)
{
	return x / 100000000;
}

#endif							/* RYU_32_BIT_PLATFORM */

#endif							/* RYU_D2S_INTRINSICS_H */