Trying to fix issue 465 (#466)

* Trying to fix issue 465

* Actually testing

* Refreshing amal.

* Removing spurious ;

.circleci/config.yml

@@ -29,6 +29,9 @@ commands:
       - run: make checkperf
       - run: make clean
       - run: ARCHFLAGS=-march=haswell make test # this breaks runtime dispatch, but see https://github.com/lemire/simdjson/issues/444... this is a code robustness test
+      - run: make clean
+      - run: EXTRAFLAGS=-DSIMDJSON_NO_COMPUTED_GOTO=true make test # this should run tests with computed gotos disabled
+  
   cmake_test:
     steps:
       - run: apt-get update -qq

README.md

@@ -90,6 +90,15 @@ On x64 hardware, you should typically build your code by specifying the oldest/l
 
 We also support 64-bit ARM. We assume NEON support. There is no runtime dispatch on ARM.
 
+## Computed GOTOs
+
+For best performance, we use a technique called "computed goto", it is also sometimes described as "Labels as Values".
+Though it is not part of the C++ standard, it is supported by many major compilers and it brings measurable performance benefits that
+are difficult to achieve otherwise.
+The computed gotos are  automatically disabled under Visual Studio.
+If you wish to forcefully disable computed gotos, you can do so by compiling the code with the macro `SIMDJSON_NO_COMPUTED_GOTO`
+defined. It is not recommended to disable computed gotos if your compiler supports it. In fact, you should almost never need to
+be concerned with computed gotos.
 
 
 ## Thread safety

include/simdjson/common_defs.h

@@ -25,7 +25,7 @@
 #define DEBUG_BLOCK(name, block)
 #endif
 
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) && !defined(SIMDJSON_NO_COMPUTED_GOTO)
 // Implemented using Labels as Values which works in GCC and CLANG (and maybe
 // also in Intel's compiler), but won't work in MSVC.
 #define SIMDJSON_USE_COMPUTED_GOTO

include/simdjson/stage1_find_marks.h

@@ -34,6 +34,6 @@ int find_structural_bits(const char *buf, size_t len, simdjson::ParsedJson &pj)
   return find_structural_bits<T>((const uint8_t *)buf, len, pj);
 }
 
-}; // namespace simdjson
+} // namespace simdjson
 
 #endif

singleheader/amalgamation_demo.cpp

@@ -1,4 +1,4 @@
-/* auto-generated on Wed Jan 15 15:50:50 EST 2020. Do not edit! */
+/* auto-generated on Mon Jan 27 10:35:34 EST 2020. Do not edit! */
 
 #include <iostream>
 #include "simdjson.h"

singleheader/simdjson.cpp

@@ -1,4 +1,4 @@
-/* auto-generated on Wed Jan 15 15:50:50 EST 2020. Do not edit! */
+/* auto-generated on Mon Jan 27 10:35:34 EST 2020. Do not edit! */
 #include "simdjson.h"
 
 /* used for http://dmalloc.com/ Dmalloc - Debug Malloc Library */
@@ -1129,7 +1129,7 @@ int json_parse_dispatch(const uint8_t *buf, size_t len, ParsedJson &pj,
     break;
 #endif
   default:
-    std::cerr << "The processor is not supported by simdjson." << std::endl;
+    // The processor is not supported by simdjson.
     return simdjson::UNEXPECTED_ERROR;
   }
 
@@ -1145,9 +1145,7 @@ ParsedJson build_parsed_json(const uint8_t *buf, size_t len,
   bool ok = pj.allocate_capacity(len);
   if (ok) {
     json_parse(buf, len, pj, realloc);
-  } else {
-    std::cerr << "failure during memory allocation " << std::endl;
-  }
+  } 
   return pj;
 }
 } // namespace simdjson
@@ -1384,7 +1382,6 @@ void find_the_best_supported_implementation() {
         return;
     }
 #endif
-    std::cerr << "The processor is not supported by simdjson." << std::endl;
     // we throw an exception since this should not be recoverable
     throw new std::runtime_error("unsupported architecture");
 }
@@ -3661,7 +3658,7 @@ namespace simdjson::arm64::simd {
     }
 
     // Store to array
-    really_inline void store(uint8_t dst[16]) { return vst1q_u8(dst, *this); }
+    really_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }
 
     // Saturated math
     really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return vqaddq_u8(*this, other); }
@@ -3761,7 +3758,7 @@ namespace simdjson::arm64::simd {
     }
 
     // Store to array
-    really_inline void store(int8_t dst[16]) { return vst1q_s8(dst, *this); }
+    really_inline void store(int8_t dst[16]) const { return vst1q_s8(dst, *this); }
 
     // Explicit conversion to/from unsigned
     really_inline explicit simd8(const uint8x16_t other): simd8(vreinterpretq_s8_u8(other)) {}
@@ -3819,7 +3816,7 @@ namespace simdjson::arm64::simd {
     really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
     really_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
 
-    really_inline void store(T ptr[64]) {
+    really_inline void store(T ptr[64]) const {
       this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
       this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
       this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
@@ -5125,44 +5122,68 @@ public:
   really_inline void scan(const uint8_t *buf, const size_t len, utf8_checker &utf8_checker);
 };
 
-// return a bitvector indicating where we have characters that end an odd-length
-// sequence of backslashes (and thus change the behavior of the next character
-// to follow). A even-length sequence of backslashes, and, for that matter, the
-// largest even-length prefix of our odd-length sequence of backslashes, simply
-// modify the behavior of the backslashes themselves.
-// We also update the prev_iter_ends_odd_backslash reference parameter to
-// indicate whether we end an iteration on an odd-length sequence of
-// backslashes, which modifies our subsequent search for odd-length
-// sequences of backslashes in an obvious way.
-really_inline uint64_t follows_odd_sequence_of(const uint64_t match, uint64_t &overflow) {
+// Routines to print masks and text for debugging bitmask operations
+UNUSED static char * format_input_text(const simd8x64<uint8_t> in) {
+  static char *buf = (char*)malloc(sizeof(simd8x64<uint8_t>) + 1);
+  in.store((uint8_t*)buf);
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+UNUSED static char * format_mask(uint64_t mask) {
+  static char *buf = (char*)malloc(64 + 1);
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+//
+// Finds escaped characters (characters following \).
+//
+// Handles runs of backslashes like \\\" and \\\\" correctly (yielding 0101 and 01010, respectively).
+//
+// Does this by:
+// - Shift the escape mask to get potentially escaped characters (characters after backslashes).
+// - Mask escaped sequences that start on *even* bits with 1010101010 (odd bits are escaped, even bits are not)
+// - Mask escaped sequences that start on *odd* bits with 0101010101 (even bits are escaped, odd bits are not)
+//
+// To distinguish between escaped sequences starting on even/odd bits, it finds the start of all
+// escape sequences, filters out the ones that start on even bits, and adds that to the mask of
+// escape sequences. This causes the addition to clear out the sequences starting on odd bits (since
+// the start bit causes a carry), and leaves even-bit sequences alone.
+//
+// Example:
+//
+// text           |  \\\ | \\\"\\\" \\\" \\"\\" |
+// escape         |  xxx |  xx xxx  xxx  xx xx  | Removed overflow backslash; will | it into follows_escape
+// odd_starts     |  x   |  x       x       x   | escape & ~even_bits & ~follows_escape
+// even_seq       |     c|    cxxx     c xx   c | c = carry bit -- will be masked out later
+// invert_mask    |      |     cxxx     c xx   c| even_seq << 1
+// follows_escape |   xx | x xx xxx  xxx  xx xx | Includes overflow bit
+// escaped        |   x  | x x  x x  x x  x  x  |
+// desired        |   x  | x x  x x  x x  x  x  |
+// text           |  \\\ | \\\"\\\" \\\" \\"\\" |
+//
+really_inline uint64_t find_escaped(uint64_t escape, uint64_t &escaped_overflow) {
+  // If there was overflow, pretend the first character isn't a backslash
+  escape &= ~escaped_overflow;
+  uint64_t follows_escape = escape << 1 | escaped_overflow;
+
+  // Get sequences starting on even bits by clearing out the odd series using +
   const uint64_t even_bits = 0x5555555555555555ULL;
-  const uint64_t odd_bits = ~even_bits;
-  uint64_t start_edges = match & ~(match << 1);
-  /* flip lowest if we have an odd-length run at the end of the prior
-  * iteration */
-  uint64_t even_start_mask = even_bits ^ overflow;
-  uint64_t even_starts = start_edges & even_start_mask;
-  uint64_t odd_starts = start_edges & ~even_start_mask;
-  uint64_t even_carries = match + even_starts;
+  uint64_t odd_sequence_starts = escape & ~even_bits & ~follows_escape;
+  uint64_t sequences_starting_on_even_bits;
+  escaped_overflow = add_overflow(odd_sequence_starts, escape, &sequences_starting_on_even_bits);
+  uint64_t invert_mask = sequences_starting_on_even_bits << 1; // The mask we want to return is the *escaped* bits, not escapes.
 
-  uint64_t odd_carries;
-  /* must record the carry-out of our odd-carries out of bit 63; this
-  * indicates whether the sense of any edge going to the next iteration
-  * should be flipped */
-  bool new_overflow = add_overflow(match, odd_starts, &odd_carries);
-
-  odd_carries |= overflow; /* push in bit zero as a
-                              * potential end if we had an
-                              * odd-numbered run at the
-                              * end of the previous
-                              * iteration */
-  overflow = new_overflow ? 0x1ULL : 0x0ULL;
-  uint64_t even_carry_ends = even_carries & ~match;
-  uint64_t odd_carry_ends = odd_carries & ~match;
-  uint64_t even_start_odd_end = even_carry_ends & odd_bits;
-  uint64_t odd_start_even_end = odd_carry_ends & even_bits;
-  uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
-  return odd_ends;
+  // Mask every other backslashed character as an escaped character
+  // Flip the mask for sequences that start on even bits, to correct them
+  return (even_bits ^ invert_mask) & follows_escape;
 }
 
 //
@@ -5211,7 +5232,7 @@ really_inline ErrorValues json_structural_scanner::detect_errors_on_eof(bool str
 //
 really_inline uint64_t json_structural_scanner::find_strings(const simd::simd8x64<uint8_t> in) {
   const uint64_t backslash = in.eq('\\');
-  const uint64_t escaped = follows_odd_sequence_of(backslash, prev_escaped);
+  const uint64_t escaped = find_escaped(backslash, prev_escaped);
   const uint64_t quote = in.eq('"') & ~escaped;
   // prefix_xor flips on bits inside the string (and flips off the end quote).
   const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
@@ -5371,9 +5392,6 @@ really_inline void json_structural_scanner::scan(const uint8_t *buf, const size_
 template<size_t STEP_SIZE>
 int find_structural_bits(const uint8_t *buf, size_t len, simdjson::ParsedJson &pj, bool streaming) {
   if (unlikely(len > pj.byte_capacity)) {
-    std::cerr << "Your ParsedJson object only supports documents up to "
-              << pj.byte_capacity << " bytes but you are trying to process "
-              << len << " bytes" << std::endl;
     return simdjson::CAPACITY;
   }
   utf8_checker utf8_checker{};
@@ -6012,44 +6030,68 @@ public:
   really_inline void scan(const uint8_t *buf, const size_t len, utf8_checker &utf8_checker);
 };
 
-// return a bitvector indicating where we have characters that end an odd-length
-// sequence of backslashes (and thus change the behavior of the next character
-// to follow). A even-length sequence of backslashes, and, for that matter, the
-// largest even-length prefix of our odd-length sequence of backslashes, simply
-// modify the behavior of the backslashes themselves.
-// We also update the prev_iter_ends_odd_backslash reference parameter to
-// indicate whether we end an iteration on an odd-length sequence of
-// backslashes, which modifies our subsequent search for odd-length
-// sequences of backslashes in an obvious way.
-really_inline uint64_t follows_odd_sequence_of(const uint64_t match, uint64_t &overflow) {
+// Routines to print masks and text for debugging bitmask operations
+UNUSED static char * format_input_text(const simd8x64<uint8_t> in) {
+  static char *buf = (char*)malloc(sizeof(simd8x64<uint8_t>) + 1);
+  in.store((uint8_t*)buf);
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+UNUSED static char * format_mask(uint64_t mask) {
+  static char *buf = (char*)malloc(64 + 1);
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+//
+// Finds escaped characters (characters following \).
+//
+// Handles runs of backslashes like \\\" and \\\\" correctly (yielding 0101 and 01010, respectively).
+//
+// Does this by:
+// - Shift the escape mask to get potentially escaped characters (characters after backslashes).
+// - Mask escaped sequences that start on *even* bits with 1010101010 (odd bits are escaped, even bits are not)
+// - Mask escaped sequences that start on *odd* bits with 0101010101 (even bits are escaped, odd bits are not)
+//
+// To distinguish between escaped sequences starting on even/odd bits, it finds the start of all
+// escape sequences, filters out the ones that start on even bits, and adds that to the mask of
+// escape sequences. This causes the addition to clear out the sequences starting on odd bits (since
+// the start bit causes a carry), and leaves even-bit sequences alone.
+//
+// Example:
+//
+// text           |  \\\ | \\\"\\\" \\\" \\"\\" |
+// escape         |  xxx |  xx xxx  xxx  xx xx  | Removed overflow backslash; will | it into follows_escape
+// odd_starts     |  x   |  x       x       x   | escape & ~even_bits & ~follows_escape
+// even_seq       |     c|    cxxx     c xx   c | c = carry bit -- will be masked out later
+// invert_mask    |      |     cxxx     c xx   c| even_seq << 1
+// follows_escape |   xx | x xx xxx  xxx  xx xx | Includes overflow bit
+// escaped        |   x  | x x  x x  x x  x  x  |
+// desired        |   x  | x x  x x  x x  x  x  |
+// text           |  \\\ | \\\"\\\" \\\" \\"\\" |
+//
+really_inline uint64_t find_escaped(uint64_t escape, uint64_t &escaped_overflow) {
+  // If there was overflow, pretend the first character isn't a backslash
+  escape &= ~escaped_overflow;
+  uint64_t follows_escape = escape << 1 | escaped_overflow;
+
+  // Get sequences starting on even bits by clearing out the odd series using +
   const uint64_t even_bits = 0x5555555555555555ULL;
-  const uint64_t odd_bits = ~even_bits;
-  uint64_t start_edges = match & ~(match << 1);
-  /* flip lowest if we have an odd-length run at the end of the prior
-  * iteration */
-  uint64_t even_start_mask = even_bits ^ overflow;
-  uint64_t even_starts = start_edges & even_start_mask;
-  uint64_t odd_starts = start_edges & ~even_start_mask;
-  uint64_t even_carries = match + even_starts;
+  uint64_t odd_sequence_starts = escape & ~even_bits & ~follows_escape;
+  uint64_t sequences_starting_on_even_bits;
+  escaped_overflow = add_overflow(odd_sequence_starts, escape, &sequences_starting_on_even_bits);
+  uint64_t invert_mask = sequences_starting_on_even_bits << 1; // The mask we want to return is the *escaped* bits, not escapes.
 
-  uint64_t odd_carries;
-  /* must record the carry-out of our odd-carries out of bit 63; this
-  * indicates whether the sense of any edge going to the next iteration
-  * should be flipped */
-  bool new_overflow = add_overflow(match, odd_starts, &odd_carries);
-
-  odd_carries |= overflow; /* push in bit zero as a
-                              * potential end if we had an
-                              * odd-numbered run at the
-                              * end of the previous
-                              * iteration */
-  overflow = new_overflow ? 0x1ULL : 0x0ULL;
-  uint64_t even_carry_ends = even_carries & ~match;
-  uint64_t odd_carry_ends = odd_carries & ~match;
-  uint64_t even_start_odd_end = even_carry_ends & odd_bits;
-  uint64_t odd_start_even_end = odd_carry_ends & even_bits;
-  uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
-  return odd_ends;
+  // Mask every other backslashed character as an escaped character
+  // Flip the mask for sequences that start on even bits, to correct them
+  return (even_bits ^ invert_mask) & follows_escape;
 }
 
 //
@@ -6098,7 +6140,7 @@ really_inline ErrorValues json_structural_scanner::detect_errors_on_eof(bool str
 //
 really_inline uint64_t json_structural_scanner::find_strings(const simd::simd8x64<uint8_t> in) {
   const uint64_t backslash = in.eq('\\');
-  const uint64_t escaped = follows_odd_sequence_of(backslash, prev_escaped);
+  const uint64_t escaped = find_escaped(backslash, prev_escaped);
   const uint64_t quote = in.eq('"') & ~escaped;
   // prefix_xor flips on bits inside the string (and flips off the end quote).
   const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
@@ -6258,9 +6300,6 @@ really_inline void json_structural_scanner::scan(const uint8_t *buf, const size_
 template<size_t STEP_SIZE>
 int find_structural_bits(const uint8_t *buf, size_t len, simdjson::ParsedJson &pj, bool streaming) {
   if (unlikely(len > pj.byte_capacity)) {
-    std::cerr << "Your ParsedJson object only supports documents up to "
-              << pj.byte_capacity << " bytes but you are trying to process "
-              << len << " bytes" << std::endl;
     return simdjson::CAPACITY;
   }
   utf8_checker utf8_checker{};
@@ -6904,44 +6943,68 @@ public:
   really_inline void scan(const uint8_t *buf, const size_t len, utf8_checker &utf8_checker);
 };
 
-// return a bitvector indicating where we have characters that end an odd-length
-// sequence of backslashes (and thus change the behavior of the next character
-// to follow). A even-length sequence of backslashes, and, for that matter, the
-// largest even-length prefix of our odd-length sequence of backslashes, simply
-// modify the behavior of the backslashes themselves.
-// We also update the prev_iter_ends_odd_backslash reference parameter to
-// indicate whether we end an iteration on an odd-length sequence of
-// backslashes, which modifies our subsequent search for odd-length
-// sequences of backslashes in an obvious way.
-really_inline uint64_t follows_odd_sequence_of(const uint64_t match, uint64_t &overflow) {
+// Routines to print masks and text for debugging bitmask operations
+UNUSED static char * format_input_text(const simd8x64<uint8_t> in) {
+  static char *buf = (char*)malloc(sizeof(simd8x64<uint8_t>) + 1);
+  in.store((uint8_t*)buf);
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+UNUSED static char * format_mask(uint64_t mask) {
+  static char *buf = (char*)malloc(64 + 1);
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+//
+// Finds escaped characters (characters following \).
+//
+// Handles runs of backslashes like \\\" and \\\\" correctly (yielding 0101 and 01010, respectively).
+//
+// Does this by:
+// - Shift the escape mask to get potentially escaped characters (characters after backslashes).
+// - Mask escaped sequences that start on *even* bits with 1010101010 (odd bits are escaped, even bits are not)
+// - Mask escaped sequences that start on *odd* bits with 0101010101 (even bits are escaped, odd bits are not)
+//
+// To distinguish between escaped sequences starting on even/odd bits, it finds the start of all
+// escape sequences, filters out the ones that start on even bits, and adds that to the mask of
+// escape sequences. This causes the addition to clear out the sequences starting on odd bits (since
+// the start bit causes a carry), and leaves even-bit sequences alone.
+//
+// Example:
+//
+// text           |  \\\ | \\\"\\\" \\\" \\"\\" |
+// escape         |  xxx |  xx xxx  xxx  xx xx  | Removed overflow backslash; will | it into follows_escape
+// odd_starts     |  x   |  x       x       x   | escape & ~even_bits & ~follows_escape
+// even_seq       |     c|    cxxx     c xx   c | c = carry bit -- will be masked out later
+// invert_mask    |      |     cxxx     c xx   c| even_seq << 1
+// follows_escape |   xx | x xx xxx  xxx  xx xx | Includes overflow bit
+// escaped        |   x  | x x  x x  x x  x  x  |
+// desired        |   x  | x x  x x  x x  x  x  |
+// text           |  \\\ | \\\"\\\" \\\" \\"\\" |
+//
+really_inline uint64_t find_escaped(uint64_t escape, uint64_t &escaped_overflow) {
+  // If there was overflow, pretend the first character isn't a backslash
+  escape &= ~escaped_overflow;
+  uint64_t follows_escape = escape << 1 | escaped_overflow;
+
+  // Get sequences starting on even bits by clearing out the odd series using +
   const uint64_t even_bits = 0x5555555555555555ULL;
-  const uint64_t odd_bits = ~even_bits;
-  uint64_t start_edges = match & ~(match << 1);
-  /* flip lowest if we have an odd-length run at the end of the prior
-  * iteration */
-  uint64_t even_start_mask = even_bits ^ overflow;
-  uint64_t even_starts = start_edges & even_start_mask;
-  uint64_t odd_starts = start_edges & ~even_start_mask;
-  uint64_t even_carries = match + even_starts;
+  uint64_t odd_sequence_starts = escape & ~even_bits & ~follows_escape;
+  uint64_t sequences_starting_on_even_bits;
+  escaped_overflow = add_overflow(odd_sequence_starts, escape, &sequences_starting_on_even_bits);
+  uint64_t invert_mask = sequences_starting_on_even_bits << 1; // The mask we want to return is the *escaped* bits, not escapes.
 
-  uint64_t odd_carries;
-  /* must record the carry-out of our odd-carries out of bit 63; this
-  * indicates whether the sense of any edge going to the next iteration
-  * should be flipped */
-  bool new_overflow = add_overflow(match, odd_starts, &odd_carries);
-
-  odd_carries |= overflow; /* push in bit zero as a
-                              * potential end if we had an
-                              * odd-numbered run at the
-                              * end of the previous
-                              * iteration */
-  overflow = new_overflow ? 0x1ULL : 0x0ULL;
-  uint64_t even_carry_ends = even_carries & ~match;
-  uint64_t odd_carry_ends = odd_carries & ~match;
-  uint64_t even_start_odd_end = even_carry_ends & odd_bits;
-  uint64_t odd_start_even_end = odd_carry_ends & even_bits;
-  uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
-  return odd_ends;
+  // Mask every other backslashed character as an escaped character
+  // Flip the mask for sequences that start on even bits, to correct them
+  return (even_bits ^ invert_mask) & follows_escape;
 }
 
 //
@@ -6990,7 +7053,7 @@ really_inline ErrorValues json_structural_scanner::detect_errors_on_eof(bool str
 //
 really_inline uint64_t json_structural_scanner::find_strings(const simd::simd8x64<uint8_t> in) {
   const uint64_t backslash = in.eq('\\');
-  const uint64_t escaped = follows_odd_sequence_of(backslash, prev_escaped);
+  const uint64_t escaped = find_escaped(backslash, prev_escaped);
   const uint64_t quote = in.eq('"') & ~escaped;
   // prefix_xor flips on bits inside the string (and flips off the end quote).
   const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
@@ -7150,9 +7213,6 @@ really_inline void json_structural_scanner::scan(const uint8_t *buf, const size_
 template<size_t STEP_SIZE>
 int find_structural_bits(const uint8_t *buf, size_t len, simdjson::ParsedJson &pj, bool streaming) {
   if (unlikely(len > pj.byte_capacity)) {
-    std::cerr << "Your ParsedJson object only supports documents up to "
-              << pj.byte_capacity << " bytes but you are trying to process "
-              << len << " bytes" << std::endl;
     return simdjson::CAPACITY;
   }
   utf8_checker utf8_checker{};
@@ -7769,7 +7829,6 @@ UNTARGET_REGION
 /* begin file src/stage2_build_tape.cpp */
 #include <cassert>
 #include <cstring>
-#include <iostream>
 
 
 using namespace simdjson;
@@ -9628,7 +9687,7 @@ bool ParsedJson::allocate_capacity(size_t len, size_t max_depth) {
   if (!string_buf || !tape ||
       !containing_scope_offset || !ret_address ||
       !structural_indexes) {
-    std::cerr << "Could not allocate memory" << std::endl;
+    // Could not allocate memory
     return false;
   }
   /*
@@ -9685,13 +9744,11 @@ bool ParsedJson::print_json(std::ostream &os) const {
   if (type == 'r') {
     how_many = tape_val & JSON_VALUE_MASK;
   } else {
-    fprintf(stderr, "Error: no starting root node?");
+    // Error: no starting root node?
     return false;
   }
   if (how_many > tape_capacity) {
-    fprintf(
-        stderr,
-        "We may be exceeding the tape capacity. Is this a valid document?\n");
+    // We may be exceeding the tape capacity. Is this a valid document?
     return false;
   }
   tape_idx++;
@@ -9778,10 +9835,10 @@ bool ParsedJson::print_json(std::ostream &os) const {
       os << ']';
       break;
     case 'r': // we start and end with the root node
-      fprintf(stderr, "should we be hitting the root node?\n");
+      // should we be hitting the root node?
       return false;
     default:
-      fprintf(stderr, "bug %c\n", type);
+      // bug?
       return false;
     }
   }
@@ -9803,7 +9860,7 @@ bool ParsedJson::dump_raw_tape(std::ostream &os) const {
   if (type == 'r') {
     how_many = tape_val & JSON_VALUE_MASK;
   } else {
-    fprintf(stderr, "Error: no starting root node?");
+    // Error: no starting root node?
     return false;
   }
   os << "\t// pointing to " << how_many << " (right after last node)\n";
@@ -9871,7 +9928,7 @@ bool ParsedJson::dump_raw_tape(std::ostream &os) const {
          << " (start of the scope) \n";
       break;
     case 'r': // we start and end with the root node
-      fprintf(stderr, "should we be hitting the root node?\n");
+      // should we be hitting the root node?
       return false;
     default:
       return false;

singleheader/simdjson.h

@@ -1,4 +1,4 @@
-/* auto-generated on Wed Jan 15 15:50:50 EST 2020. Do not edit! */
+/* auto-generated on Mon Jan 27 10:35:34 EST 2020. Do not edit! */
 /* begin file include/simdjson/simdjson_version.h */
 // /include/simdjson/simdjson_version.h automatically generated by release.py,
 // do not change by hand
@@ -20,7 +20,6 @@ enum {
 
 #include <cstddef>
 #include <cstdint>
-#include <cstdio>
 #include <cstdlib>
 #ifdef _MSC_VER
 #include <iso646.h>
@@ -66,7 +65,7 @@ enum {
 #endif
 
 // under GCC and CLANG, we use these two macros
-#define TARGET_HASWELL TARGET_REGION("avx2,bmi,pclmul")
+#define TARGET_HASWELL TARGET_REGION("avx2,bmi,pclmul,lzcnt")
 #define TARGET_WESTMERE TARGET_REGION("sse4.2,pclmul")
 #define TARGET_ARM64
 
@@ -294,53 +293,11 @@ static inline uint32_t detect_supported_architectures() {
 #ifndef SIMDJSON_JSONFORMATUTILS_H
 #define SIMDJSON_JSONFORMATUTILS_H
 
-#include <cstdio>
 #include <iomanip>
 #include <iostream>
 
 namespace simdjson {
-// ends with zero char
-static inline void print_with_escapes(const unsigned char *src) {
-  while (*src) {
-    switch (*src) {
-    case '\b':
-      putchar('\\');
-      putchar('b');
-      break;
-    case '\f':
-      putchar('\\');
-      putchar('f');
-      break;
-    case '\n':
-      putchar('\\');
-      putchar('n');
-      break;
-    case '\r':
-      putchar('\\');
-      putchar('r');
-      break;
-    case '\"':
-      putchar('\\');
-      putchar('"');
-      break;
-    case '\t':
-      putchar('\\');
-      putchar('t');
-      break;
-    case '\\':
-      putchar('\\');
-      putchar('\\');
-      break;
-    default:
-      if (*src <= 0x1F) {
-        printf("\\u%04x", *src);
-      } else {
-        putchar(*src);
-      }
-    }
-    src++;
-  }
-}
+
 
 // ends with zero char
 static inline void print_with_escapes(const unsigned char *src,
@@ -389,49 +346,6 @@ static inline void print_with_escapes(const unsigned char *src,
   }
 }
 
-// print len chars
-static inline void print_with_escapes(const unsigned char *src, size_t len) {
-  const unsigned char *finalsrc = src + len;
-  while (src < finalsrc) {
-    switch (*src) {
-    case '\b':
-      putchar('\\');
-      putchar('b');
-      break;
-    case '\f':
-      putchar('\\');
-      putchar('f');
-      break;
-    case '\n':
-      putchar('\\');
-      putchar('n');
-      break;
-    case '\r':
-      putchar('\\');
-      putchar('r');
-      break;
-    case '\"':
-      putchar('\\');
-      putchar('"');
-      break;
-    case '\t':
-      putchar('\\');
-      putchar('t');
-      break;
-    case '\\':
-      putchar('\\');
-      putchar('\\');
-      break;
-    default:
-      if (*src <= 0x1F) {
-        printf("\\u%04x", *src);
-      } else {
-        putchar(*src);
-      }
-    }
-    src++;
-  }
-}
 
 // print len chars
 static inline void print_with_escapes(const unsigned char *src,
@@ -586,7 +500,7 @@ const std::string &error_message(const int);
 #define DEBUG_BLOCK(name, block)
 #endif
 
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) && !defined(SIMDJSON_NO_COMPUTED_GOTO)
 // Implemented using Labels as Values which works in GCC and CLANG (and maybe
 // also in Intel's compiler), but won't work in MSVC.
 #define SIMDJSON_USE_COMPUTED_GOTO
@@ -642,7 +556,7 @@ const std::string &error_message(const int);
 
 
 namespace simdjson {
-// low-level function to allocate memory with padding so we can read passed the
+// low-level function to allocate memory with padding so we can read past the
 // "length" bytes safely. if you must provide a pointer to some data, create it
 // with this function: length is the max. size in bytes of the string caller is
 // responsible to free the memory (free(...))
@@ -815,7 +729,6 @@ static inline size_t json_minify(const padded_string &p, char *out) {
 #define SIMDJSON_PARSEDJSON_H
 
 #include <cstring>
-#include <iostream>
 #include <memory>
 
 #define JSON_VALUE_MASK 0xFFFFFFFFFFFFFF
@@ -864,7 +777,7 @@ public:
   // this should be called when parsing (right before writing the tapes)
   void init();
 
-  // print the json to stdout (should be valid)
+  // print the json to std::ostream (should be valid)
   // return false if the tape is likely wrong (e.g., you did not parse a valid
   // JSON).
   WARN_UNUSED
@@ -915,8 +828,8 @@ public:
     tape[saved_loc] |= val;
   }
 
-  class InvalidJSON : public std::exception {
-    const char *what() const throw() { return "JSON document is invalid"; }
+  struct InvalidJSON : public std::exception {
+    const char *what() const noexcept { return "JSON document is invalid"; }
   };
 
   template <size_t max_depth> class BasicIterator;
@@ -948,20 +861,7 @@ public:
 
 };
 
-// dump bits low to high
-inline void dumpbits_always(uint64_t v, const std::string &msg) {
-  for (uint32_t i = 0; i < 64; i++) {
-    std::cout << (((v >> static_cast<uint64_t>(i)) & 0x1ULL) ? "1" : "_");
-  }
-  std::cout << " " << msg.c_str() << "\n";
-}
 
-inline void dumpbits32_always(uint32_t v, const std::string &msg) {
-  for (uint32_t i = 0; i < 32; i++) {
-    std::cout << (((v >> i) & 0x1ULL) ? "1" : "_");
-  }
-  std::cout << " " << msg.c_str() << "\n";
-}
 } // namespace simdjson
 #endif
 /* end file include/simdjson/parsedjson.h */
@@ -1744,7 +1644,7 @@ int find_structural_bits(const char *buf, size_t len, simdjson::ParsedJson &pj)
   return find_structural_bits<T>((const uint8_t *)buf, len, pj);
 }
 
-}; // namespace simdjson
+} // namespace simdjson
 
 #endif
 /* end file include/simdjson/stage1_find_marks.h */
@@ -2127,7 +2027,9 @@ namespace simdjson {
         size_t next_json{0};
         bool load_next_batch{true};
         size_t current_buffer_loc{0};
+#ifdef SIMDJSON_THREADS_ENABLED
         size_t last_json_buffer_loc{0};
+#endif
         size_t n_parsed_docs{0};
         size_t n_bytes_parsed{0};
 #ifdef SIMDJSON_THREADS_ENABLED