* limitations under the License.
*/
-#include "folly/String.h"
-#include "folly/Format.h"
+#include <folly/String.h>
+#include <folly/Format.h>
#include <cerrno>
#include <cstdarg>
{ "z", 1e-21L },
{ "y", 1e-24L },
{ " ", 0 },
- { 0, 0}
+ { 0, 0}
};
const PrettySuffix* const kPrettySuffixes[PRETTY_NUM_TYPES] = {
//TODO:
//1) Benchmark & optimize
-double prettyToDouble(folly::StringPiece *const prettyString,
+double prettyToDouble(folly::StringPiece *const prettyString,
const PrettyType type) {
double value = folly::to<double>(prettyString);
while (prettyString->size() > 0 && std::isspace(prettyString->front())) {
prettyString->toString(), "\""));
}
prettyString->advance(longestPrefixLen);
- return suffixes[bestPrefixId].val ? value * suffixes[bestPrefixId].val :
+ 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.data(),
+ detail::enforceWhitespace(prettyString.data(),
prettyString.data() + prettyString.size());
return 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 defined(__APPLE__) || defined(__FreeBSD__) || \
+#if defined(__APPLE__) || defined(__FreeBSD__) || defined(__CYGWIN__) ||\
((_POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600) && !_GNU_SOURCE)
// Using XSI-compatible strerror_r
int r = strerror_r(err, buf, sizeof(buf));
return result;
}
+StringPiece skipWhitespace(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.
+ auto oddspace = [] (char c) {
+ return c == '\n' || c == '\t' || c == '\r';
+ };
+
+loop:
+ for (; !sp.empty() && sp.front() == ' '; sp.pop_front()) {
+ }
+ if (!sp.empty() && oddspace(sp.front())) {
+ sp.pop_front();
+ goto loop;
+ }
+
+ return sp;
+}
+
+namespace {
+
+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.
+ unsigned char rotated = 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 += rotated;
+}
+
+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;
+}
+
+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;
+}
+
+} // anon 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,
# undef DMGL_TYPES
# undef DMGL_RET_POSTFIX
#endif
-