X-Git-Url: http://plrg.eecs.uci.edu/git/?p=folly.git;a=blobdiff_plain;f=folly%2FString.cpp;h=097e5bbfd4c7b2cd86ed2b388ee557bea5124952;hp=cfcb3582104b67a210cfa847b3830d0b1197e090;hb=18de341f84035f76395347f77a8cc71d0461ab37;hpb=5c77fedbef46995a71ffa268c9fcaf49efddd01b diff --git a/folly/String.cpp b/folly/String.cpp index cfcb3582..097e5bbf 100644 --- a/folly/String.cpp +++ b/folly/String.cpp @@ -1,5 +1,5 @@ /* - * Copyright 2013 Facebook, Inc. + * Copyright 2017 Facebook, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -14,86 +14,124 @@ * limitations under the License. */ -#include "folly/String.h" -#include "folly/Format.h" +#include +#include #include #include #include -#include #include +#include + #include -#undef FOLLY_DEMANGLE -#if defined(__GNUG__) && __GNUG__ >= 4 -# include -# define FOLLY_DEMANGLE 1 -#endif +#include namespace folly { -namespace { +static inline bool is_oddspace(char c) { + return c == '\n' || c == '\t' || c == '\r'; +} + +StringPiece ltrimWhitespace(StringPiece sp) { + // Spaces other than ' ' characters are less common but should be + // checked. This configuration where we loop on the ' ' + // separately from oddspaces was empirically fastest. + +loop: + for (; !sp.empty() && sp.front() == ' '; sp.pop_front()) { + } + if (!sp.empty() && is_oddspace(sp.front())) { + sp.pop_front(); + goto loop; + } -inline void stringPrintfImpl(std::string& output, const char* format, - va_list args) { - // Tru to the space at the end of output for our output buffer. - // Find out write point then inflate its size temporarily to its - // capacity; we will later shrink it to the size needed to represent - // the formatted string. If this buffer isn't large enough, we do a - // resize and try again. + return sp; +} + +StringPiece rtrimWhitespace(StringPiece sp) { + // Spaces other than ' ' characters are less common but should be + // checked. This configuration where we loop on the ' ' + // separately from oddspaces was empirically fastest. + +loop: + for (; !sp.empty() && sp.back() == ' '; sp.pop_back()) { + } + if (!sp.empty() && is_oddspace(sp.back())) { + sp.pop_back(); + goto loop; + } - const auto write_point = output.size(); - auto remaining = output.capacity() - write_point; - output.resize(output.capacity()); + return sp; +} + +namespace { +int stringAppendfImplHelper(char* buf, + size_t bufsize, + const char* format, + va_list args) { va_list args_copy; va_copy(args_copy, args); - int bytes_used = vsnprintf(&output[write_point], remaining, format, - args_copy); + int bytes_used = vsnprintf(buf, bufsize, format, args_copy); va_end(args_copy); + return bytes_used; +} + +void stringAppendfImpl(std::string& output, const char* format, va_list args) { + // Very simple; first, try to avoid an allocation by using an inline + // buffer. If that fails to hold the output string, allocate one on + // the heap, use it instead. + // + // It is hard to guess the proper size of this buffer; some + // heuristics could be based on the number of format characters, or + // static analysis of a codebase. Or, we can just pick a number + // that seems big enough for simple cases (say, one line of text on + // a terminal) without being large enough to be concerning as a + // stack variable. + std::array inline_buffer; + + int bytes_used = stringAppendfImplHelper( + inline_buffer.data(), inline_buffer.size(), format, args); if (bytes_used < 0) { - throw std::runtime_error( - to("Invalid format string; snprintf returned negative " - "with format string: ", format)); - } else if (bytes_used < remaining) { - // There was enough room, just shrink and return. - output.resize(write_point + bytes_used); - } else { - output.resize(write_point + bytes_used + 1); - remaining = bytes_used + 1; - va_list args_copy; - va_copy(args_copy, args); - bytes_used = vsnprintf(&output[write_point], remaining, format, - args_copy); - va_end(args_copy); - if (bytes_used + 1 != remaining) { - throw std::runtime_error( - to("vsnprint retry did not manage to work " - "with format string: ", format)); - } - output.resize(write_point + bytes_used); + throw std::runtime_error(to( + "Invalid format string; snprintf returned negative " + "with format string: ", + format)); } + + if (static_cast(bytes_used) < inline_buffer.size()) { + output.append(inline_buffer.data(), size_t(bytes_used)); + return; + } + + // Couldn't fit. Heap allocate a buffer, oh well. + std::unique_ptr heap_buffer(new char[size_t(bytes_used + 1)]); + int final_bytes_used = stringAppendfImplHelper( + heap_buffer.get(), size_t(bytes_used + 1), format, args); + // The second call can take fewer bytes if, for example, we were printing a + // string buffer with null-terminating char using a width specifier - + // vsnprintf("%.*s", buf.size(), buf) + CHECK(bytes_used >= final_bytes_used); + + // We don't keep the trailing '\0' in our output string + output.append(heap_buffer.get(), size_t(final_bytes_used)); } -} // anon namespace +} // namespace std::string stringPrintf(const char* format, ...) { - // snprintf will tell us how large the output buffer should be, but - // we then have to call it a second time, which is costly. By - // guestimating the final size, we avoid the double snprintf in many - // cases, resulting in a performance win. We use this constructor - // of std::string to avoid a double allocation, though it does pad - // the resulting string with nul bytes. Our guestimation is twice - // the format string size, or 32 bytes, whichever is larger. This - // is a hueristic that doesn't affect correctness but attempts to be - // reasonably fast for the most common cases. - std::string ret(std::max(32UL, strlen(format) * 2), '\0'); - ret.resize(0); - va_list ap; va_start(ap, format); - stringPrintfImpl(ret, format, ap); - va_end(ap); + SCOPE_EXIT { + va_end(ap); + }; + return stringVPrintf(format, ap); +} + +std::string stringVPrintf(const char* format, va_list ap) { + std::string ret; + stringAppendfImpl(ret, format, ap); return ret; } @@ -102,17 +140,31 @@ std::string stringPrintf(const char* format, ...) { std::string& stringAppendf(std::string* output, const char* format, ...) { va_list ap; va_start(ap, format); - stringPrintfImpl(*output, format, ap); - va_end(ap); + SCOPE_EXIT { + va_end(ap); + }; + return stringVAppendf(output, format, ap); +} + +std::string& stringVAppendf(std::string* output, + const char* format, + va_list ap) { + stringAppendfImpl(*output, format, ap); return *output; } void stringPrintf(std::string* output, const char* format, ...) { - output->clear(); va_list ap; va_start(ap, format); - stringPrintfImpl(*output, format, ap); - va_end(ap); + SCOPE_EXIT { + va_end(ap); + }; + return stringVPrintf(output, format, ap); +} + +void stringVPrintf(std::string* output, const char* format, va_list ap) { + output->clear(); + stringAppendfImpl(*output, format, ap); }; namespace { @@ -129,7 +181,7 @@ const PrettySuffix kPrettyTimeSuffixes[] = { { "ns", 1e-9L }, { "ps", 1e-12L }, { "s ", 0 }, - { 0, 0 }, + { nullptr, 0 }, }; const PrettySuffix kPrettyBytesMetricSuffixes[] = { @@ -138,7 +190,7 @@ const PrettySuffix kPrettyBytesMetricSuffixes[] = { { "MB", 1e6L }, { "kB", 1e3L }, { "B ", 0L }, - { 0, 0 }, + { nullptr, 0 }, }; const PrettySuffix kPrettyBytesBinarySuffixes[] = { @@ -147,7 +199,7 @@ const PrettySuffix kPrettyBytesBinarySuffixes[] = { { "MB", int64_t(1) << 20 }, { "kB", int64_t(1) << 10 }, { "B ", 0L }, - { 0, 0 }, + { nullptr, 0 }, }; const PrettySuffix kPrettyBytesBinaryIECSuffixes[] = { @@ -156,7 +208,7 @@ const PrettySuffix kPrettyBytesBinaryIECSuffixes[] = { { "MiB", int64_t(1) << 20 }, { "KiB", int64_t(1) << 10 }, { "B ", 0L }, - { 0, 0 }, + { nullptr, 0 }, }; const PrettySuffix kPrettyUnitsMetricSuffixes[] = { @@ -165,7 +217,7 @@ const PrettySuffix kPrettyUnitsMetricSuffixes[] = { { "M", 1e6L }, { "k", 1e3L }, { " ", 0 }, - { 0, 0 }, + { nullptr, 0 }, }; const PrettySuffix kPrettyUnitsBinarySuffixes[] = { @@ -174,7 +226,7 @@ const PrettySuffix kPrettyUnitsBinarySuffixes[] = { { "M", int64_t(1) << 20 }, { "k", int64_t(1) << 10 }, { " ", 0 }, - { 0, 0 }, + { nullptr, 0 }, }; const PrettySuffix kPrettyUnitsBinaryIECSuffixes[] = { @@ -183,7 +235,32 @@ const PrettySuffix kPrettyUnitsBinaryIECSuffixes[] = { { "Mi", int64_t(1) << 20 }, { "Ki", int64_t(1) << 10 }, { " ", 0 }, - { 0, 0 }, + { nullptr, 0 }, +}; + +const PrettySuffix kPrettySISuffixes[] = { + { "Y", 1e24L }, + { "Z", 1e21L }, + { "E", 1e18L }, + { "P", 1e15L }, + { "T", 1e12L }, + { "G", 1e9L }, + { "M", 1e6L }, + { "k", 1e3L }, + { "h", 1e2L }, + { "da", 1e1L }, + { "d", 1e-1L }, + { "c", 1e-2L }, + { "m", 1e-3L }, + { "u", 1e-6L }, + { "n", 1e-9L }, + { "p", 1e-12L }, + { "f", 1e-15L }, + { "a", 1e-18L }, + { "z", 1e-21L }, + { "y", 1e-24L }, + { " ", 0 }, + { nullptr, 0} }; const PrettySuffix* const kPrettySuffixes[PRETTY_NUM_TYPES] = { @@ -194,9 +271,10 @@ const PrettySuffix* const kPrettySuffixes[PRETTY_NUM_TYPES] = { kPrettyUnitsMetricSuffixes, kPrettyUnitsBinarySuffixes, kPrettyUnitsBinaryIECSuffixes, + kPrettySISuffixes, }; -} // namespace +} // namespace std::string prettyPrint(double val, PrettyType type, bool addSpace) { char buf[100]; @@ -224,6 +302,49 @@ std::string prettyPrint(double val, PrettyType type, bool addSpace) { return std::string(buf); } +//TODO: +//1) Benchmark & optimize +double prettyToDouble(folly::StringPiece *const prettyString, + const PrettyType type) { + double value = folly::to(prettyString); + while (prettyString->size() > 0 && std::isspace(prettyString->front())) { + prettyString->advance(1); //Skipping spaces between number and suffix + } + const PrettySuffix* suffixes = kPrettySuffixes[type]; + int longestPrefixLen = -1; + int bestPrefixId = -1; + for (int j = 0 ; suffixes[j].suffix; ++j) { + if (suffixes[j].suffix[0] == ' '){//Checking for " " -> number rule. + if (longestPrefixLen == -1) { + longestPrefixLen = 0; //No characters to skip + bestPrefixId = j; + } + } else if (prettyString->startsWith(suffixes[j].suffix)) { + int suffixLen = int(strlen(suffixes[j].suffix)); + //We are looking for a longest suffix matching prefix of the string + //after numeric value. We need this in case suffixes have common prefix. + if (suffixLen > longestPrefixLen) { + longestPrefixLen = suffixLen; + bestPrefixId = j; + } + } + } + if (bestPrefixId == -1) { //No valid suffix rule found + throw std::invalid_argument(folly::to( + "Unable to parse suffix \"", + prettyString->toString(), "\"")); + } + prettyString->advance(size_t(longestPrefixLen)); + return suffixes[bestPrefixId].val ? value * suffixes[bestPrefixId].val : + value; +} + +double prettyToDouble(folly::StringPiece prettyString, const PrettyType type){ + double result = prettyToDouble(&prettyString, type); + detail::enforceWhitespace(prettyString); + return result; +} + std::string hexDump(const void* ptr, size_t size) { std::ostringstream os; hexDump(ptr, size, std::ostream_iterator(os, "\n")); @@ -241,13 +362,27 @@ fbstring errnoStr(int err) { fbstring result; + // https://developer.apple.com/library/mac/documentation/Darwin/Reference/ManPages/man3/strerror_r.3.html // http://www.kernel.org/doc/man-pages/online/pages/man3/strerror.3.html -#if (_POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600 || \ - !FOLLY_HAVE_FEATURES_H) && !_GNU_SOURCE +#if defined(_WIN32) && (defined(__MINGW32__) || defined(_MSC_VER)) + // mingw64 has no strerror_r, but Windows has strerror_s, which C11 added + // as well. So maybe we should use this across all platforms (together + // with strerrorlen_s). Note strerror_r and _s have swapped args. + int r = strerror_s(buf, sizeof(buf), err); + if (r != 0) { + result = to( + "Unknown error ", err, + " (strerror_r failed with error ", errno, ")"); + } else { + result.assign(buf); + } +#elif defined(FOLLY_HAVE_XSI_STRERROR_R) || \ + defined(__APPLE__) || defined(__ANDROID__) // Using XSI-compatible strerror_r int r = strerror_r(err, buf, sizeof(buf)); - if (r == -1) { + // OSX/FreeBSD use EINVAL and Linux uses -1 so just check for non-zero + if (r != 0) { result = to( "Unknown error ", err, " (strerror_r failed with error ", errno, ")"); @@ -262,34 +397,152 @@ fbstring errnoStr(int err) { return result; } -#ifdef FOLLY_DEMANGLE +namespace { -fbstring demangle(const char* name) { - int status; - size_t len = 0; - // malloc() memory for the demangled type name - char* demangled = abi::__cxa_demangle(name, nullptr, &len, &status); - if (status != 0) { - return name; - } - // len is the length of the buffer (including NUL terminator and maybe - // other junk) - return fbstring(demangled, strlen(demangled), len, AcquireMallocatedString()); +void toLowerAscii8(char& c) { + // Branchless tolower, based on the input-rotating trick described + // at http://www.azillionmonkeys.com/qed/asmexample.html + // + // This algorithm depends on an observation: each uppercase + // ASCII character can be converted to its lowercase equivalent + // by adding 0x20. + + // Step 1: Clear the high order bit. We'll deal with it in Step 5. + uint8_t rotated = uint8_t(c & 0x7f); + // Currently, the value of rotated, as a function of the original c is: + // below 'A': 0- 64 + // 'A'-'Z': 65- 90 + // above 'Z': 91-127 + + // Step 2: Add 0x25 (37) + rotated += 0x25; + // Now the value of rotated, as a function of the original c is: + // below 'A': 37-101 + // 'A'-'Z': 102-127 + // above 'Z': 128-164 + + // Step 3: clear the high order bit + rotated &= 0x7f; + // below 'A': 37-101 + // 'A'-'Z': 102-127 + // above 'Z': 0- 36 + + // Step 4: Add 0x1a (26) + rotated += 0x1a; + // below 'A': 63-127 + // 'A'-'Z': 128-153 + // above 'Z': 25- 62 + + // At this point, note that only the uppercase letters have been + // transformed into values with the high order bit set (128 and above). + + // Step 5: Shift the high order bit 2 spaces to the right: the spot + // where the only 1 bit in 0x20 is. But first, how we ignored the + // high order bit of the original c in step 1? If that bit was set, + // we may have just gotten a false match on a value in the range + // 128+'A' to 128+'Z'. To correct this, need to clear the high order + // bit of rotated if the high order bit of c is set. Since we don't + // care about the other bits in rotated, the easiest thing to do + // is invert all the bits in c and bitwise-and them with rotated. + rotated &= ~c; + rotated >>= 2; + + // Step 6: Apply a mask to clear everything except the 0x20 bit + // in rotated. + rotated &= 0x20; + + // At this point, rotated is 0x20 if c is 'A'-'Z' and 0x00 otherwise + + // Step 7: Add rotated to c + c += char(rotated); } -#else +void toLowerAscii32(uint32_t& c) { + // Besides being branchless, the algorithm in toLowerAscii8() has another + // interesting property: None of the addition operations will cause + // an overflow in the 8-bit value. So we can pack four 8-bit values + // into a uint32_t and run each operation on all four values in parallel + // without having to use any CPU-specific SIMD instructions. + uint32_t rotated = c & uint32_t(0x7f7f7f7fL); + rotated += uint32_t(0x25252525L); + rotated &= uint32_t(0x7f7f7f7fL); + rotated += uint32_t(0x1a1a1a1aL); + + // Step 5 involves a shift, so some bits will spill over from each + // 8-bit value into the next. But that's okay, because they're bits + // that will be cleared by the mask in step 6 anyway. + rotated &= ~c; + rotated >>= 2; + rotated &= uint32_t(0x20202020L); + c += rotated; +} -fbstring demangle(const char* name) { - return name; +void toLowerAscii64(uint64_t& c) { + // 64-bit version of toLower32 + uint64_t rotated = c & uint64_t(0x7f7f7f7f7f7f7f7fL); + rotated += uint64_t(0x2525252525252525L); + rotated &= uint64_t(0x7f7f7f7f7f7f7f7fL); + rotated += uint64_t(0x1a1a1a1a1a1a1a1aL); + rotated &= ~c; + rotated >>= 2; + rotated &= uint64_t(0x2020202020202020L); + c += rotated; } -#endif -#undef FOLLY_DEMANGLE +} // namespace + +void toLowerAscii(char* str, size_t length) { + static const size_t kAlignMask64 = 7; + static const size_t kAlignMask32 = 3; + + // Convert a character at a time until we reach an address that + // is at least 32-bit aligned + size_t n = (size_t)str; + n &= kAlignMask32; + n = std::min(n, length); + size_t offset = 0; + if (n != 0) { + n = std::min(4 - n, length); + do { + toLowerAscii8(str[offset]); + offset++; + } while (offset < n); + } + + n = (size_t)(str + offset); + n &= kAlignMask64; + if ((n != 0) && (offset + 4 <= length)) { + // The next address is 32-bit aligned but not 64-bit aligned. + // Convert the next 4 bytes in order to get to the 64-bit aligned + // part of the input. + toLowerAscii32(*(uint32_t*)(str + offset)); + offset += 4; + } + + // Convert 8 characters at a time + while (offset + 8 <= length) { + toLowerAscii64(*(uint64_t*)(str + offset)); + offset += 8; + } + + // Convert 4 characters at a time + while (offset + 4 <= length) { + toLowerAscii32(*(uint32_t*)(str + offset)); + offset += 4; + } + + // Convert any characters remaining after the last 4-byte aligned group + while (offset < length) { + toLowerAscii8(str[offset]); + offset++; + } +} namespace detail { size_t hexDumpLine(const void* ptr, size_t offset, size_t size, std::string& line) { + static char hexValues[] = "0123456789abcdef"; // Line layout: // 8: address // 1: space @@ -303,13 +556,24 @@ size_t hexDumpLine(const void* ptr, size_t offset, size_t size, line.reserve(78); const uint8_t* p = reinterpret_cast(ptr) + offset; size_t n = std::min(size - offset, size_t(16)); - format("{:08x} ", offset).appendTo(line); + line.push_back(hexValues[(offset >> 28) & 0xf]); + line.push_back(hexValues[(offset >> 24) & 0xf]); + line.push_back(hexValues[(offset >> 20) & 0xf]); + line.push_back(hexValues[(offset >> 16) & 0xf]); + line.push_back(hexValues[(offset >> 12) & 0xf]); + line.push_back(hexValues[(offset >> 8) & 0xf]); + line.push_back(hexValues[(offset >> 4) & 0xf]); + line.push_back(hexValues[offset & 0xf]); + line.push_back(' '); for (size_t i = 0; i < n; i++) { if (i == 8) { line.push_back(' '); } - format(" {:02x}", p[i]).appendTo(line); + + line.push_back(' '); + line.push_back(hexValues[(p[i] >> 4) & 0xf]); + line.push_back(hexValues[p[i] & 0xf]); } // 3 spaces for each byte we're not printing, one separating the halves @@ -323,11 +587,66 @@ size_t hexDumpLine(const void* ptr, size_t offset, size_t size, } line.append(16 - n, ' '); line.push_back('|'); - DCHECK_EQ(line.size(), 78); + DCHECK_EQ(line.size(), 78u); return n; } } // namespace detail -} // namespace folly +std::string stripLeftMargin(std::string s) { + std::vector pieces; + split("\n", s, pieces); + auto piecer = range(pieces); + + auto piece = (piecer.end() - 1); + auto needle = std::find_if(piece->begin(), + piece->end(), + [](char c) { return c != ' ' && c != '\t'; }); + if (needle == piece->end()) { + (piecer.end() - 1)->clear(); + } + piece = piecer.begin(); + needle = std::find_if(piece->begin(), + piece->end(), + [](char c) { return c != ' ' && c != '\t'; }); + if (needle == piece->end()) { + piecer.erase(piecer.begin(), piecer.begin() + 1); + } + + const auto sentinel = std::numeric_limits::max(); + auto indent = sentinel; + size_t max_length = 0; + for (piece = piecer.begin(); piece != piecer.end(); piece++) { + needle = std::find_if(piece->begin(), + piece->end(), + [](char c) { return c != ' ' && c != '\t'; }); + if (needle != piece->end()) { + indent = std::min(indent, size_t(needle - piece->begin())); + } else { + max_length = std::max(piece->size(), max_length); + } + } + indent = indent == sentinel ? max_length : indent; + for (piece = piecer.begin(); piece != piecer.end(); piece++) { + if (piece->size() < indent) { + piece->clear(); + } else { + piece->erase(piece->begin(), piece->begin() + indent); + } + } + return join("\n", piecer); +} + +} // namespace folly + +#ifdef FOLLY_DEFINED_DMGL +# undef FOLLY_DEFINED_DMGL +# undef DMGL_NO_OPTS +# undef DMGL_PARAMS +# undef DMGL_ANSI +# undef DMGL_JAVA +# undef DMGL_VERBOSE +# undef DMGL_TYPES +# undef DMGL_RET_POSTFIX +#endif