Merge remote-tracking branch 'origin/develop' into claude/parser-performance-research-4hkdf8

Resolves the number-parsing conflict between the discard_number_values
fast path (an early return for accept()-only calls that skips value
conversion for short digit runs) and the contiguous number scanner's
convert_number()/convert_integer() split: the discard check now runs
as the first step of convert_number(), so both the byte-at-a-time
scanner and the bulk contiguous path benefit from it.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

include/nlohmann/detail/input/lexer.hpp

@@ -178,10 +178,11 @@ class lexer : public lexer_base<BasicJsonType>
   public:
     using token_type = typename lexer_base<BasicJsonType>::token_type;
 
-    explicit lexer(InputAdapterType&& adapter, bool ignore_comments_ = false) noexcept
+    explicit lexer(InputAdapterType&& adapter, bool ignore_comments_ = false, bool discard_number_values_ = false) noexcept
         : ia(std::move(adapter))
         , ignore_comments(ignore_comments_)
         , decimal_point_char(static_cast<char_int_type>(get_decimal_point()))
+        , discard_number_values(discard_number_values_)
     {}
 
     // deleted because of pointer members
@@ -1465,6 +1466,58 @@ scan_number_done:
     */
     token_type convert_number(token_type number_type, std::size_t mantissa_end)
     {
+        // If the caller does not need the converted value (only whether the
+        // input is syntactically valid; see json_sax_acceptor/accept()), an
+        // unsigned/integer token can be reported without calling
+        // strtoull()/strtoll() at all, *provided* we can already tell from
+        // the digit count alone that the conversion cannot overflow 64 bits.
+        // Such tokens are always finite and are accepted unconditionally by
+        // the parser regardless of their actual value (parser::sax_parse_internal()
+        // never checks finiteness for value_unsigned/value_integer), so the
+        // classification below is all that is needed.
+        //
+        // A decimal number with up to 18 digits is always representable in
+        // both std::uint64_t and std::int64_t (18 nines is ~1e18, well below
+        // both UINT64_MAX ~1.8e19 and INT64_MAX ~9.2e18), so strtoull()/strtoll()
+        // could not have set errno to ERANGE for it. Numbers with more digits
+        // (rare in practice) fall through to the exact code below, unchanged,
+        // so their handling -- including reclassification to value_float when
+        // the value overflows 64 bits, and rejection when it is not even
+        // finite as a double -- is bit-for-bit identical to before this
+        // optimization.
+        //
+        // Note this reasons about std::uint64_t/std::int64_t, not about
+        // number_unsigned_t/number_integer_t (BasicJsonType's own, possibly
+        // narrower, template parameters -- e.g. std::uint32_t). That is fine
+        // *only* because discard_number_values is exclusively set by
+        // accept() (see json.hpp), and accept() always parses through the
+        // library's own json_sax_acceptor -- never a user-supplied SAX
+        // consumer -- whose number_unsigned()/number_integer()/number_float()
+        // callbacks unconditionally discard their argument and return true.
+        // So for every caller that can reach this branch, neither the token
+        // classification below nor the eventual (possibly narrowed, and on
+        // this fast path left stale/unset) value_unsigned/value_integer is
+        // ever consulted -- an unsigned/integer token is accepted outright,
+        // and even a >18-digit token that this fast path deliberately falls
+        // through for is, once reclassified to value_float, still finite
+        // (and thus accepted) for any digit count that fits in number_unsigned_t
+        // or number_integer_t regardless of that type's width. If this
+        // function is ever taught to run with discard_number_values true for
+        // a caller that *does* read the converted value, this reasoning (and
+        // the fast path below) would need to be revisited.
+        if (discard_number_values)
+        {
+            constexpr std::size_t safe_digit_count = 18;
+            if (number_type == token_type::value_unsigned && token_buffer.size() <= safe_digit_count)
+            {
+                return token_type::value_unsigned;
+            }
+            if (number_type == token_type::value_integer && token_buffer.size() - 1 <= safe_digit_count)
+            {
+                return token_type::value_integer;
+            }
+        }
+
         const char* const num_begin = token_buffer.data();
         const char* const num_end = num_begin + token_buffer.size();
 
@@ -1723,8 +1776,7 @@ scan_number_done:
     */
     char_int_type get()
     {
-        ++position.chars_read_total;
-        ++position.chars_read_current_line;
+        advance_position();
 
         if (next_unget)
         {
@@ -1736,6 +1788,23 @@ scan_number_done:
             current = ia.get_character();
         }
 
+        return track_after_read();
+    }
+
+    /// shared head of get() / get_ignoring_pending_unget(): bump the
+    /// per-character position counters (line-count-on-'\n' bookkeeping is
+    /// handled afterwards, in track_after_read(), once `current` is known)
+    void advance_position() noexcept
+    {
+        ++position.chars_read_total;
+        ++position.chars_read_current_line;
+    }
+
+    /// shared tail of get() / get_ignoring_pending_unget(): capture the
+    /// character for error messages (if needed) and update line/column
+    /// bookkeeping for the character now in `current`
+    char_int_type track_after_read()
+    {
         // seekable adapters reconstruct the token lazily on error (see
         // get_token_string), so the eager per-character copy is skipped
         capture_char(std::integral_constant<bool, lazy_token_string> {});
@@ -1752,6 +1821,29 @@ scan_number_done:
         return current;
     }
 
+    /*!
+    @brief like get(), but for call sites that can prove no unget() is pending
+
+    get() has to check the `next_unget` flag on every call, because a
+    previous token may have ended with unget() (e.g. scan_number() always
+    ungets the character that terminated the number, so the next call to
+    scan() can see it again). skip_whitespace() reads that first,
+    possibly-ungotten character via a plain get(), but every further
+    character it reads is guaranteed to be a fresh read: nothing between
+    those calls invokes unget(). This variant skips the (otherwise always
+    false) next_unget branch for those calls; it is not a general
+    replacement for get().
+    */
+    char_int_type get_ignoring_pending_unget()
+    {
+        JSON_ASSERT(!next_unget);
+
+        advance_position();
+        current = ia.get_character();
+
+        return track_after_read();
+    }
+
     /// seekable adapter: nothing to capture, the token is rebuilt on error
     void capture_char(std::true_type /*lazy*/) const noexcept {}
 
@@ -1953,13 +2045,37 @@ scan_number_done:
         return true;
     }
 
+    /// whether `current` is one of the four JSON whitespace characters
+    bool current_is_whitespace() const noexcept
+    {
+        return current == ' ' || current == '\t' || current == '\n' || current == '\r';
+    }
+
     void skip_whitespace()
     {
+        // the first character may be a pending unget() left over from the
+        // previous token (see get_ignoring_pending_unget()); every
+        // subsequent character read by this loop is guaranteed fresh, since
+        // nothing below calls unget()
+        get();
+
+        if (!current_is_whitespace())
+        {
+            return;
+        }
+
+        // this is written as an if-guarded do-while (rather than a plain
+        // while loop) because that shape is what lets both GCC and Clang
+        // keep the input adapter's read pointer in a register across
+        // iterations; the equivalent while-loop measurably defeated that
+        // optimization in testing, turning long whitespace runs (e.g. the
+        // indentation of pretty-printed JSON) from a register-only loop
+        // into one that reloads the pointer from memory every character
         do
         {
-            get();
+            get_ignoring_pending_unget();
         }
-        while (current == ' ' || current == '\t' || current == '\n' || current == '\r');
+        while (current_is_whitespace());
     }
 
     token_type scan()
@@ -2099,6 +2215,13 @@ scan_number_done:
     const char_int_type decimal_point_char = '.';
     /// the position of the decimal point in the input
     std::size_t decimal_point_position = std::string::npos;
+
+    /// whether the caller (e.g. accept()/json_sax_acceptor) only needs the
+    /// token classification and never looks at the converted numeric value;
+    /// when set, scan_number() may skip strtoull()/strtoll() for
+    /// value_unsigned/value_integer tokens whose digit count guarantees they
+    /// fit into 64 bits (see scan_number())
+    const bool discard_number_values = false;
 };
 
 }  // namespace detail

include/nlohmann/detail/input/parser.hpp

@@ -72,9 +72,10 @@ class parser
                     parser_callback_t<BasicJsonType> cb = nullptr,
                     const bool allow_exceptions_ = true,
                     const bool ignore_comments = false,
-                    const bool ignore_trailing_commas_ = false)
+                    const bool ignore_trailing_commas_ = false,
+                    const bool discard_number_values_ = false)
         : callback(std::move(cb))
-        , m_lexer(std::move(adapter), ignore_comments)
+        , m_lexer(std::move(adapter), ignore_comments, discard_number_values_)
         , allow_exceptions(allow_exceptions_)
         , ignore_trailing_commas(ignore_trailing_commas_)
     {