Skip to content

R
rapidjson
Commit ca5000ba authored by Milo Yip's avatar Milo Yip
Browse files
Options

Merge pull request #108 from miloyip/ParsingOptimization 

Parsing optimization
parents df70ee82 4f81c873
Hide whitespace changes
Inline Side-by-side
Showing with 101 additions and 108 deletions
include/rapidjson/rapidjson.h
View file @ ca5000ba
...@@ -147,6 +147,18 @@ ...@@ -147,6 +147,18 @@
# endif # endif
#endif // RAPIDJSON_ENDIAN #endif // RAPIDJSON_ENDIAN
///////////////////////////////////////////////////////////////////////////////
// RAPIDJSON_64BIT
//! Whether using 64-bit architecture
#ifndef RAPIDJSON_64BIT
#if defined(__LP64__) || defined(_WIN64)
#define RAPIDJSON_64BIT 1
#else
#define RAPIDJSON_64BIT 0
#endif
#endif // RAPIDJSON_64BIT
/////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////
// RAPIDJSON_ALIGN // RAPIDJSON_ALIGN
......
include/rapidjson/reader.h
View file @ ca5000ba
...@@ -242,57 +242,34 @@ void SkipWhitespace(InputStream& is) { ...@@ -242,57 +242,34 @@ void SkipWhitespace(InputStream& is) {
#ifdef RAPIDJSON_SSE42 #ifdef RAPIDJSON_SSE42
//! Skip whitespace with SSE 4.2 pcmpistrm instruction, testing 16 8-byte characters at once. //! Skip whitespace with SSE 4.2 pcmpistrm instruction, testing 16 8-byte characters at once.
inline const char *SkipWhitespace_SIMD(const char* p) { inline const char *SkipWhitespace_SIMD(const char* p) {
static const char whitespace[16] = " \n\r\t"; // Fast return for single non-whitespace
static const char whitespaces[4][17] = { if (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
" ", ++p;
"\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n", else
"\r\r\r\r\r\r\r\r\r\r\r\r\r\r\r\r", return p;
"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t"};
// 16-byte align to the next boundary
// 16-byte align to the lower boundary const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & ~15);
const char* ap = reinterpret_cast<const char*>(reinterpret_cast<size_t>(p) & ~15); while (p != nextAligned)
if (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
// Test first unaligned characters ++p;
// Cannot make use of _mm_cmpistrm() because it stops when encounters '\0' before p else
if (ap != p) { return p;
const __m128i w0 = _mm_loadu_si128((const __m128i *)&whitespaces[0][0]);
const __m128i w1 = _mm_loadu_si128((const __m128i *)&whitespaces[1][0]); // The rest of string using SIMD
const __m128i w2 = _mm_loadu_si128((const __m128i *)&whitespaces[2][0]); static const char whitespace[16] = " \n\r\t";
const __m128i w3 = _mm_loadu_si128((const __m128i *)&whitespaces[3][0]); const __m128i w = _mm_loadu_si128((const __m128i *)&whitespace[0]);
unsigned char shift = reinterpret_cast<size_t>(p) & 15; for (;; p += 16) {
const __m128i s = _mm_load_si128(reinterpret_cast<const __m128i*>(ap)); const __m128i s = _mm_load_si128((const __m128i *)p);
__m128i x = _mm_cmpeq_epi8(s, w0);
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w1));
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w2));
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w3));
unsigned short r = (unsigned short)~_mm_movemask_epi8(x);
r = r >> shift << shift; // Clear results before p
if (r != 0) {
#ifdef _MSC_VER // Find the index of first non-whitespace
unsigned long offset;
_BitScanForward(&offset, r);
return ap + offset;
#else
return ap + __builtin_ffs(r) - 1;
#endif
}
ap += 16;
}
const __m128i w = _mm_loadu_si128((const __m128i *)&whitespace[0]);
// The rest of string
for (;; ap += 16) {
const __m128i s = _mm_load_si128((const __m128i *)ap);
const unsigned r = _mm_cvtsi128_si32(_mm_cmpistrm(w, s, _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_BIT_MASK | _SIDD_NEGATIVE_POLARITY)); const unsigned r = _mm_cvtsi128_si32(_mm_cmpistrm(w, s, _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_BIT_MASK | _SIDD_NEGATIVE_POLARITY));
if (r != 0) { // some of characters is non-whitespace if (r != 0) { // some of characters is non-whitespace
#ifdef _MSC_VER // Find the index of first non-whitespace #ifdef _MSC_VER // Find the index of first non-whitespace
unsigned long offset; unsigned long offset;
_BitScanForward(&offset, r); _BitScanForward(&offset, r);
return ap + offset; return p + offset;
#else #else
return ap + __builtin_ffs(r) - 1; return p + __builtin_ffs(r) - 1;
#endif #endif
} }
} }
...@@ -302,45 +279,34 @@ inline const char *SkipWhitespace_SIMD(const char* p) { ...@@ -302,45 +279,34 @@ inline const char *SkipWhitespace_SIMD(const char* p) {
//! Skip whitespace with SSE2 instructions, testing 16 8-byte characters at once. //! Skip whitespace with SSE2 instructions, testing 16 8-byte characters at once.
inline const char *SkipWhitespace_SIMD(const char* p) { inline const char *SkipWhitespace_SIMD(const char* p) {
static const char whitespaces[4][17] = { // Fast return for single non-whitespace
" ", if (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
"\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n", ++p;
"\r\r\r\r\r\r\r\r\r\r\r\r\r\r\r\r", else
"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t"}; return p;
const __m128i w0 = _mm_loadu_si128((const __m128i *)&whitespaces[0][0]); // 16-byte align to the next boundary
const __m128i w1 = _mm_loadu_si128((const __m128i *)&whitespaces[1][0]); const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & ~15);
const __m128i w2 = _mm_loadu_si128((const __m128i *)&whitespaces[2][0]); while (p != nextAligned)
const __m128i w3 = _mm_loadu_si128((const __m128i *)&whitespaces[3][0]); if (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
++p;
// 16-byte align to the lower boundary else
const char* ap = reinterpret_cast<const char*>(reinterpret_cast<size_t>(p) & ~15); return p;
// Test first unaligned characters
if (ap != p) {
unsigned char shift = reinterpret_cast<size_t>(p) & 15;
const __m128i s = _mm_load_si128(reinterpret_cast<const __m128i*>(ap));
__m128i x = _mm_cmpeq_epi8(s, w0);
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w1));
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w2));
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w3));
unsigned short r = (unsigned short)~_mm_movemask_epi8(x);
r = r >> shift << shift; // Clear results before p
if (r != 0) {
#ifdef _MSC_VER // Find the index of first non-whitespace
unsigned long offset;
_BitScanForward(&offset, r);
return ap + offset;
#else
return ap + __builtin_ffs(r) - 1;
#endif
}
ap += 16;
}
// The rest of string // The rest of string
for (;; ap += 16) { static const char whitespaces[4][17] = {
const __m128i s = _mm_load_si128((const __m128i *)ap); " ",
"\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
"\r\r\r\r\r\r\r\r\r\r\r\r\r\r\r\r",
"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t"};
const __m128i w0 = _mm_loadu_si128((const __m128i *)&whitespaces[0][0]);
const __m128i w1 = _mm_loadu_si128((const __m128i *)&whitespaces[1][0]);
const __m128i w2 = _mm_loadu_si128((const __m128i *)&whitespaces[2][0]);
const __m128i w3 = _mm_loadu_si128((const __m128i *)&whitespaces[3][0]);
for (;; p += 16) {
const __m128i s = _mm_load_si128((const __m128i *)p);
__m128i x = _mm_cmpeq_epi8(s, w0); __m128i x = _mm_cmpeq_epi8(s, w0);
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w1)); x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w1));
x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w2)); x = _mm_or_si128(x, _mm_cmpeq_epi8(s, w2));
...@@ -350,9 +316,9 @@ inline const char *SkipWhitespace_SIMD(const char* p) { ...@@ -350,9 +316,9 @@ inline const char *SkipWhitespace_SIMD(const char* p) {
#ifdef _MSC_VER // Find the index of first non-whitespace #ifdef _MSC_VER // Find the index of first non-whitespace
unsigned long offset; unsigned long offset;
_BitScanForward(&offset, r); _BitScanForward(&offset, r);
return ap + offset; return p + offset;
#else #else
return ap + __builtin_ffs(r) - 1; return p + __builtin_ffs(r) - 1;
#endif #endif
} }
} }
...@@ -760,7 +726,8 @@ private: ...@@ -760,7 +726,8 @@ private:
// Parse int: zero / ( digit1-9 *DIGIT ) // Parse int: zero / ( digit1-9 *DIGIT )
unsigned i = 0; unsigned i = 0;
bool try64bit = false; uint64_t i64 = 0;
bool use64bit = false;
if (s.Peek() == '0') { if (s.Peek() == '0') {
i = 0; i = 0;
s.Take(); s.Take();
...@@ -772,7 +739,8 @@ private: ...@@ -772,7 +739,8 @@ private:
while (s.Peek() >= '0' && s.Peek() <= '9') { while (s.Peek() >= '0' && s.Peek() <= '9') {
if (i >= 214748364) { // 2^31 = 2147483648 if (i >= 214748364) { // 2^31 = 2147483648
if (i != 214748364 || s.Peek() > '8') { if (i != 214748364 || s.Peek() > '8') {
try64bit = true; i64 = i;
use64bit = true;
break; break;
} }
} }
...@@ -782,7 +750,8 @@ private: ...@@ -782,7 +750,8 @@ private:
while (s.Peek() >= '0' && s.Peek() <= '9') { while (s.Peek() >= '0' && s.Peek() <= '9') {
if (i >= 429496729) { // 2^32 - 1 = 4294967295 if (i >= 429496729) { // 2^32 - 1 = 4294967295
if (i != 429496729 || s.Peek() > '5') { if (i != 429496729 || s.Peek() > '5') {
try64bit = true; i64 = i;
use64bit = true;
break; break;
} }
} }
...@@ -793,14 +762,14 @@ private: ...@@ -793,14 +762,14 @@ private:
RAPIDJSON_PARSE_ERROR(kParseErrorValueInvalid, s.Tell()); RAPIDJSON_PARSE_ERROR(kParseErrorValueInvalid, s.Tell());
// Parse 64bit int // Parse 64bit int
uint64_t i64 = 0; double d = 0.0;
bool useDouble = false; bool useDouble = false;
if (try64bit) { if (use64bit) {
i64 = i;
if (minus) if (minus)
while (s.Peek() >= '0' && s.Peek() <= '9') { while (s.Peek() >= '0' && s.Peek() <= '9') {
if (i64 >= RAPIDJSON_UINT64_C2(0x0CCCCCCC, 0xCCCCCCCC)) // 2^63 = 9223372036854775808 if (i64 >= RAPIDJSON_UINT64_C2(0x0CCCCCCC, 0xCCCCCCCC)) // 2^63 = 9223372036854775808
if (i64 != RAPIDJSON_UINT64_C2(0x0CCCCCCC, 0xCCCCCCCC) || s.Peek() > '8') { if (i64 != RAPIDJSON_UINT64_C2(0x0CCCCCCC, 0xCCCCCCCC) || s.Peek() > '8') {
d = (double)i64;
useDouble = true; useDouble = true;
break; break;
} }
...@@ -810,6 +779,7 @@ private: ...@@ -810,6 +779,7 @@ private:
while (s.Peek() >= '0' && s.Peek() <= '9') { while (s.Peek() >= '0' && s.Peek() <= '9') {
if (i64 >= RAPIDJSON_UINT64_C2(0x19999999, 0x99999999)) // 2^64 - 1 = 18446744073709551615 if (i64 >= RAPIDJSON_UINT64_C2(0x19999999, 0x99999999)) // 2^64 - 1 = 18446744073709551615
if (i64 != RAPIDJSON_UINT64_C2(0x19999999, 0x99999999) || s.Peek() > '5') { if (i64 != RAPIDJSON_UINT64_C2(0x19999999, 0x99999999) || s.Peek() > '5') {
d = (double)i64;
useDouble = true; useDouble = true;
break; break;
} }
...@@ -818,9 +788,7 @@ private: ...@@ -818,9 +788,7 @@ private:
} }
// Force double for big integer // Force double for big integer
double d = 0.0;
if (useDouble) { if (useDouble) {
d = (double)i64;
while (s.Peek() >= '0' && s.Peek() <= '9') { while (s.Peek() >= '0' && s.Peek() <= '9') {
if (d >= 1.7976931348623157e307) // DBL_MAX / 10.0 if (d >= 1.7976931348623157e307) // DBL_MAX / 10.0
RAPIDJSON_PARSE_ERROR(kParseErrorNumberTooBig, s.Tell()); RAPIDJSON_PARSE_ERROR(kParseErrorNumberTooBig, s.Tell());
...@@ -831,33 +799,46 @@ private: ...@@ -831,33 +799,46 @@ private:
// Parse frac = decimal-point 1*DIGIT // Parse frac = decimal-point 1*DIGIT
int expFrac = 0; int expFrac = 0;
if (s.Peek() == '.') { if (s.Peek() == '.') {
s.Take();
#if RAPIDJSON_64BIT
// Use i64 to store significand in 64-bit architecture
if (!useDouble) { if (!useDouble) {
d = try64bit ? (double)i64 : (double)i; if (!use64bit)
useDouble = true; i64 = i;
while (s.Peek() >= '0' && s.Peek() <= '9') {
if (i64 >= RAPIDJSON_UINT64_C2(0x19999999, 0x99999999))
break;
else {
i64 = i64 * 10 + static_cast<unsigned>(s.Take() - '0');
--expFrac;
}
}
d = (double)i64;
} }
s.Take(); #else
// Use double to store significand in 32-bit architecture
if (!useDouble)
d = use64bit ? (double)i64 : (double)i;
#endif
useDouble = true;
if (s.Peek() >= '0' && s.Peek() <= '9') { while (s.Peek() >= '0' && s.Peek() <= '9') {
d = d * 10 + (s.Take() - '0'); d = d * 10 + (s.Take() - '0');
--expFrac; --expFrac;
} }
else
RAPIDJSON_PARSE_ERROR(kParseErrorNumberMissFraction, s.Tell());
while (s.Peek() >= '0' && s.Peek() <= '9') { if (expFrac == 0)
if (expFrac > -16) { RAPIDJSON_PARSE_ERROR(kParseErrorNumberMissFraction, s.Tell());
d = d * 10 + (s.Peek() - '0');
--expFrac;
}
s.Take();
}
} }
// Parse exp = e [ minus / plus ] 1*DIGIT // Parse exp = e [ minus / plus ] 1*DIGIT
int exp = 0; int exp = 0;
if (s.Peek() == 'e' || s.Peek() == 'E') { if (s.Peek() == 'e' || s.Peek() == 'E') {
if (!useDouble) { if (!useDouble) {
d = try64bit ? (double)i64 : (double)i; d = use64bit ? (double)i64 : (double)i;
useDouble = true; useDouble = true;
} }
s.Take(); s.Take();
...@@ -900,7 +881,7 @@ private: ...@@ -900,7 +881,7 @@ private:
cont = handler.Double(minus ? -d : d); cont = handler.Double(minus ? -d : d);
} }
else { else {
if (try64bit) { if (use64bit) {
if (minus) if (minus)
cont = handler.Int64(-(int64_t)i64); cont = handler.Int64(-(int64_t)i64);
else else
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment