1 | // Copyright 2016 the V8 project authors. All rights reserved. |
2 | // Use of this source code is governed by a BSD-style license that can be |
3 | // found in the LICENSE file. |
4 | |
5 | #include "src/string-case.h" |
6 | |
7 | #include "src/assert-scope.h" |
8 | #include "src/base/logging.h" |
9 | #include "src/globals.h" |
10 | #include "src/utils.h" |
11 | |
12 | namespace v8 { |
13 | namespace internal { |
14 | |
15 | // FastAsciiConvert tries to do character processing on a word_t basis if |
16 | // source and destination strings are properly aligned. Natural alignment of |
17 | // string data depends on kTaggedSize so we define word_t via Tagged_t. |
18 | using word_t = std::make_unsigned<Tagged_t>::type; |
19 | |
20 | const word_t kWordTAllBitsSet = std::numeric_limits<word_t>::max(); |
21 | const word_t kOneInEveryByte = kWordTAllBitsSet / 0xFF; |
22 | const word_t kAsciiMask = kOneInEveryByte << 7; |
23 | |
24 | #ifdef DEBUG |
25 | bool CheckFastAsciiConvert(char* dst, const char* src, int length, bool changed, |
26 | bool is_to_lower) { |
27 | bool expected_changed = false; |
28 | for (int i = 0; i < length; i++) { |
29 | if (dst[i] == src[i]) continue; |
30 | expected_changed = true; |
31 | if (is_to_lower) { |
32 | DCHECK('A' <= src[i] && src[i] <= 'Z'); |
33 | DCHECK(dst[i] == src[i] + ('a' - 'A')); |
34 | } else { |
35 | DCHECK('a' <= src[i] && src[i] <= 'z'); |
36 | DCHECK(dst[i] == src[i] - ('a' - 'A')); |
37 | } |
38 | } |
39 | return (expected_changed == changed); |
40 | } |
41 | #endif |
42 | |
43 | // Given a word and two range boundaries returns a word with high bit |
44 | // set in every byte iff the corresponding input byte was strictly in |
45 | // the range (m, n). All the other bits in the result are cleared. |
46 | // This function is only useful when it can be inlined and the |
47 | // boundaries are statically known. |
48 | // Requires: all bytes in the input word and the boundaries must be |
49 | // ASCII (less than 0x7F). |
50 | static inline word_t AsciiRangeMask(word_t w, char m, char n) { |
51 | // Use strict inequalities since in edge cases the function could be |
52 | // further simplified. |
53 | DCHECK(0 < m && m < n); |
54 | // Has high bit set in every w byte less than n. |
55 | word_t tmp1 = kOneInEveryByte * (0x7F + n) - w; |
56 | // Has high bit set in every w byte greater than m. |
57 | word_t tmp2 = w + kOneInEveryByte * (0x7F - m); |
58 | return (tmp1 & tmp2 & (kOneInEveryByte * 0x80)); |
59 | } |
60 | |
61 | template <bool is_lower> |
62 | int FastAsciiConvert(char* dst, const char* src, int length, |
63 | bool* changed_out) { |
64 | #ifdef DEBUG |
65 | char* saved_dst = dst; |
66 | #endif |
67 | const char* saved_src = src; |
68 | DisallowHeapAllocation no_gc; |
69 | // We rely on the distance between upper and lower case letters |
70 | // being a known power of 2. |
71 | DCHECK_EQ('a' - 'A', 1 << 5); |
72 | // Boundaries for the range of input characters than require conversion. |
73 | static const char lo = is_lower ? 'A' - 1 : 'a' - 1; |
74 | static const char hi = is_lower ? 'Z' + 1 : 'z' + 1; |
75 | bool changed = false; |
76 | const char* const limit = src + length; |
77 | |
78 | // dst is newly allocated and always aligned. |
79 | DCHECK(IsAligned(reinterpret_cast<Address>(dst), sizeof(word_t))); |
80 | // Only attempt processing one word at a time if src is also aligned. |
81 | if (IsAligned(reinterpret_cast<Address>(src), sizeof(word_t))) { |
82 | // Process the prefix of the input that requires no conversion one aligned |
83 | // (machine) word at a time. |
84 | while (src <= limit - sizeof(word_t)) { |
85 | const word_t w = *reinterpret_cast<const word_t*>(src); |
86 | if ((w & kAsciiMask) != 0) return static_cast<int>(src - saved_src); |
87 | if (AsciiRangeMask(w, lo, hi) != 0) { |
88 | changed = true; |
89 | break; |
90 | } |
91 | *reinterpret_cast<word_t*>(dst) = w; |
92 | src += sizeof(word_t); |
93 | dst += sizeof(word_t); |
94 | } |
95 | // Process the remainder of the input performing conversion when |
96 | // required one word at a time. |
97 | while (src <= limit - sizeof(word_t)) { |
98 | const word_t w = *reinterpret_cast<const word_t*>(src); |
99 | if ((w & kAsciiMask) != 0) return static_cast<int>(src - saved_src); |
100 | word_t m = AsciiRangeMask(w, lo, hi); |
101 | // The mask has high (7th) bit set in every byte that needs |
102 | // conversion and we know that the distance between cases is |
103 | // 1 << 5. |
104 | *reinterpret_cast<word_t*>(dst) = w ^ (m >> 2); |
105 | src += sizeof(word_t); |
106 | dst += sizeof(word_t); |
107 | } |
108 | } |
109 | // Process the last few bytes of the input (or the whole input if |
110 | // unaligned access is not supported). |
111 | while (src < limit) { |
112 | char c = *src; |
113 | if ((c & kAsciiMask) != 0) return static_cast<int>(src - saved_src); |
114 | if (lo < c && c < hi) { |
115 | c ^= (1 << 5); |
116 | changed = true; |
117 | } |
118 | *dst = c; |
119 | ++src; |
120 | ++dst; |
121 | } |
122 | |
123 | DCHECK( |
124 | CheckFastAsciiConvert(saved_dst, saved_src, length, changed, is_lower)); |
125 | |
126 | *changed_out = changed; |
127 | return length; |
128 | } |
129 | |
130 | template int FastAsciiConvert<false>(char* dst, const char* src, int length, |
131 | bool* changed_out); |
132 | template int FastAsciiConvert<true>(char* dst, const char* src, int length, |
133 | bool* changed_out); |
134 | |
135 | } // namespace internal |
136 | } // namespace v8 |
137 | |