Inline jsoncharutils per-implementation

src/arm64/dom_parser_implementation.cpp

@@ -1,5 +1,6 @@
 #include "arm64/begin_implementation.h"
 #include "arm64/dom_parser_implementation.h"
+#include "generic/stage2/jsoncharutils.h"
 
 //
 // Stage 1

src/fallback/dom_parser_implementation.cpp

@@ -1,5 +1,6 @@
 #include "fallback/begin_implementation.h"
 #include "fallback/dom_parser_implementation.h"
+#include "generic/stage2/jsoncharutils.h"
 
 //
 // Stage 1

src/generic/stage2/jsoncharutils.h

@@ -0,0 +1,134 @@
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace stage2 {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+really_inline uint32_t is_not_structural_or_whitespace_or_null(uint8_t c) {
+  return structural_or_whitespace_or_null_negated[c];
+}
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return structural_or_whitespace_negated[c];
+}
+
+really_inline uint32_t is_structural_or_whitespace_or_null(uint8_t c) {
+  return structural_or_whitespace_or_null[c];
+}
+
+really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = digit_to_val32[630 + src[0]];
+  uint32_t v2 = digit_to_val32[420 + src[1]];
+  uint32_t v3 = digit_to_val32[210 + src[2]];
+  uint32_t v4 = digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+////
+// The following code is used in number parsing. It is not
+// properly "char utils" stuff, but we move it here so that
+// it does not get copied multiple times in the binaries (once
+// per instruction set).
+///
+
+constexpr int FASTFLOAT_SMALLEST_POWER = -325;
+constexpr int FASTFLOAT_LARGEST_POWER = 308;
+
+struct value128 {
+  uint64_t low;
+  uint64_t high;
+};
+
+#ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  value128 answer;
+#if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
+#ifdef _M_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // _M_ARM64
+#else // defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
+  __uint128_t r = ((__uint128_t)value1) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace stage2
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson

src/generic/stage2/numberparsing.h

@@ -1,4 +1,3 @@
-#include "jsoncharutils.h"
 #include <cmath>
 #include <limits>
 

src/generic/stage2/stringparsing.h

@@ -1,8 +1,6 @@
 // This file contains the common code every implementation uses
 // It is intended to be included multiple times and compiled multiple times
 
-#include "jsoncharutils.h"
-
 namespace simdjson {
 namespace SIMDJSON_IMPLEMENTATION {
 namespace stage2 {

src/haswell/dom_parser_implementation.cpp

@@ -1,5 +1,6 @@
 #include "haswell/begin_implementation.h"
 #include "haswell/dom_parser_implementation.h"
+#include "generic/stage2/jsoncharutils.h"
 
 //
 // Stage 1

src/jsoncharutils_tables.h

@@ -1,5 +1,5 @@
-#ifndef SIMDJSON_JSONCHARUTILS_H
+#ifndef SIMDJSON_JSONCHARUTILS_TABLES_H
-#define SIMDJSON_JSONCHARUTILS_H
+#define SIMDJSON_JSONCHARUTILS_TABLES_H
 
 #include "simdjson.h"
 
@@ -34,12 +34,6 @@ const uint32_t structural_or_whitespace_or_null_negated[256] = {
     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
 
-// return non-zero if not a structural or whitespace char
-// zero otherwise
-really_inline uint32_t is_not_structural_or_whitespace_or_null(uint8_t c) {
-  return structural_or_whitespace_or_null_negated[c];
-}
-
 const uint32_t structural_or_whitespace_negated[256] = {
     1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
@@ -57,12 +51,6 @@ const uint32_t structural_or_whitespace_negated[256] = {
     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
 
-// return non-zero if not a structural or whitespace char
-// zero otherwise
-really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
-  return structural_or_whitespace_negated[c];
-}
-
 const uint32_t structural_or_whitespace_or_null[256] = {
     1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
@@ -76,10 +64,6 @@ const uint32_t structural_or_whitespace_or_null[256] = {
     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
 
-really_inline uint32_t is_structural_or_whitespace_or_null(uint8_t c) {
-  return structural_or_whitespace_or_null[c];
-}
-
 const uint32_t structural_or_whitespace[256] = {
     0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
@@ -93,10 +77,6 @@ const uint32_t structural_or_whitespace[256] = {
     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
 
-really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
-  return structural_or_whitespace[c];
-}
-
 const uint32_t digit_to_val32[886] = {
     0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
     0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
@@ -246,62 +226,6 @@ const uint32_t digit_to_val32[886] = {
     0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
     0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
     0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF};
-// returns a value with the high 16 bits set if not valid
-// otherwise returns the conversion of the 4 hex digits at src into the bottom
-// 16 bits of the 32-bit return register
-//
-// see
-// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
-static inline uint32_t hex_to_u32_nocheck(
-    const uint8_t *src) { // strictly speaking, static inline is a C-ism
-  uint32_t v1 = digit_to_val32[630 + src[0]];
-  uint32_t v2 = digit_to_val32[420 + src[1]];
-  uint32_t v3 = digit_to_val32[210 + src[2]];
-  uint32_t v4 = digit_to_val32[0 + src[3]];
-  return v1 | v2 | v3 | v4;
-}
-
-// given a code point cp, writes to c
-// the utf-8 code, outputting the length in
-// bytes, if the length is zero, the code point
-// is invalid
-//
-// This can possibly be made faster using pdep
-// and clz and table lookups, but JSON documents
-// have few escaped code points, and the following
-// function looks cheap.
-//
-// Note: we assume that surrogates are treated separately
-//
-inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
-  if (cp <= 0x7F) {
-    c[0] = uint8_t(cp);
-    return 1; // ascii
-  }
-  if (cp <= 0x7FF) {
-    c[0] = uint8_t((cp >> 6) + 192);
-    c[1] = uint8_t((cp & 63) + 128);
-    return 2; // universal plane
-    //  Surrogates are treated elsewhere...
-    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
-    //  return 0; // surrogates // could put assert here
-  } else if (cp <= 0xFFFF) {
-    c[0] = uint8_t((cp >> 12) + 224);
-    c[1] = uint8_t(((cp >> 6) & 63) + 128);
-    c[2] = uint8_t((cp & 63) + 128);
-    return 3;
-  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
-                               // is not needed
-    c[0] = uint8_t((cp >> 18) + 240);
-    c[1] = uint8_t(((cp >> 12) & 63) + 128);
-    c[2] = uint8_t(((cp >> 6) & 63) + 128);
-    c[3] = uint8_t((cp & 63) + 128);
-    return 4;
-  }
-  // will return 0 when the code point was too large.
-  return 0; // bad r
-}
-
 ////
 // The following code is used in number parsing. It is not
 // properly "char utils" stuff, but we move it here so that
@@ -317,42 +241,6 @@ struct value128 {
   uint64_t high;
 };
 
-#ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
-// this is a slow emulation routine for 32-bit
-//
-static inline uint64_t __emulu(uint32_t x, uint32_t y) {
-  return x * (uint64_t)y;
-}
-static inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
-  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
-  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
-  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
-  uint64_t adbc_carry = !!(adbc < ad);
-  uint64_t lo = bd + (adbc << 32);
-  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
-        (adbc_carry << 32) + !!(lo < bd);
-  return lo;
-}
-#endif
-
-really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
-  value128 answer;
-#if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
-#ifdef _M_ARM64
-  // ARM64 has native support for 64-bit multiplications, no need to emultate
-  answer.high = __umulh(value1, value2);
-  answer.low = value1 * value2;
-#else
-  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
-#endif // _M_ARM64
-#else // defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
-  __uint128_t r = ((__uint128_t)value1) * value2;
-  answer.low = uint64_t(r);
-  answer.high = uint64_t(r >> 64);
-#endif
-  return answer;
-}
-
 // Precomputed powers of ten from 10^0 to 10^22. These
 // can be represented exactly using the double type.
 static const double power_of_ten[] = {
@@ -1333,4 +1221,4 @@ const uint64_t mantissa_128[] = {
 
 } // namespace simdjson
 
-#endif // SIMDJSON_JSONCHARUTILS_H
+#endif // SIMDJSON_JSONCHARUTILS_TABLES_H

src/simdjson.cpp

@@ -9,7 +9,7 @@ SIMDJSON_DISABLE_UNDESIRED_WARNINGS
 // Anything in the top level directory MUST be included outside of the #if statements
 // below, or amalgamation will screw them up!
 #include "isadetection.h"
-#include "jsoncharutils.h"
+#include "jsoncharutils_tables.h"
 #include "simdprune_tables.h"
 
 #if SIMDJSON_IMPLEMENTATION_ARM64

src/westmere/dom_parser_implementation.cpp

@@ -1,4 +1,6 @@
 #include "westmere/begin_implementation.h"
+#include "westmere/dom_parser_implementation.h"
+#include "generic/stage2/jsoncharutils.h"
 
 //
 // Stage 1