Copy ASCII strings in the view's dump with ensure_ascii (#5879)

With ensure_ascii, the view's dump wrote every string through the escaping
path, which decodes UTF-8 byte by byte, even strings of the source that are
plain ASCII. A string of the source without escape sequences has no quote,
backslash, or control character, so ensure_ascii leaves it unchanged
exactly when it has no byte >= 0x7F. The dump now checks that eight bytes
at a time (one add and one or per word) and copies such strings like the
dump without ensure_ascii.

The compact walk (dump_compact) now also serves ensure_ascii, through a
separate out-of-line instantiation, so the walk without ensure_ascii keeps
its speed.

Parsing is unchanged. A flag set by the parser would let the dump skip the
check and was about twice as fast for compact output, but it costs every
parse of string-heavy input 1-2% with GCC.

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

include/nlohmann/detail/view/serializer.hpp

@@ -171,6 +171,34 @@ inline std::size_t plain_output_run(const unsigned char* s, std::size_t n) noexc
     return n;
 }
 
+/// Whether dump() with ensure_ascii writes the string at s unchanged, for a
+/// string of the source without escape sequences (which has no quote,
+/// backslash, or control character): no byte is DEL (0x7F) or above. Adding 1
+/// to each byte sets its high bit exactly for 0x7F, and a carry into the next
+/// byte only comes from 0xFF, whose own high bit is set anyway.
+inline bool is_ascii_without_delete(const char* s, std::size_t n) noexcept
+{
+    constexpr std::uint64_t ones = 0x0101010101010101ull;
+    constexpr std::uint64_t high = 0x8080808080808080ull;
+    std::size_t i = 0;
+    for (; i + 8 <= n; i += 8)
+    {
+        const std::uint64_t v = read_eight_bytes(s + i);
+        if (((v | (v + ones)) & high) != 0)
+        {
+            return false;
+        }
+    }
+    for (; i < n; ++i)
+    {
+        if (static_cast<unsigned char>(s[i]) >= 0x7F)
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
 /// A stack that starts in a buffer of the caller (a local array) and moves to
 /// the heap (a vector of the caller) only when that is full, so that dumps of
 /// shallow documents need no allocation. The top is a pointer, as in
@@ -286,15 +314,19 @@ class view_serializer
 
     void dump(const node* root)
     {
-        if (!m_style.pretty && !m_style.ensure_ascii)
+        if (!m_style.pretty)
         {
-            if (m_style.source_numbers)
+            if (m_style.ensure_ascii)
             {
-                dump_compact<true>(root);
+                dump_compact_ascii(root);
+            }
+            else if (m_style.source_numbers)
+            {
+                dump_compact<true, false>(root);
             }
             else
             {
-                dump_compact<false>(root);
+                dump_compact<false, false>(root);
             }
             return;
         }
@@ -400,6 +432,41 @@ class view_serializer
         m_out.put('"');
     }
 
+    /// dump_compact() with ensure_ascii, out of line: the walk without it is
+    /// faster when it is not merged with this one into dump()
+    NLOHMANN_VIEW_NOINLINE void dump_compact_ascii(const node* root)
+    {
+        if (m_style.source_numbers)
+        {
+            dump_compact<true, true>(root);
+        }
+        else
+        {
+            dump_compact<false, true>(root);
+        }
+    }
+
+    /// whether dump_compact() copies a string or key from the source: one of
+    /// the source, and with ensure_ascii one without bytes >= 0x7F (an overload
+    /// each, so that the walk without ensure_ascii is compiled exactly as
+    /// before: with GCC, a condition that only folds away later changed its
+    /// block layout and made it 4-6% slower)
+    static bool copied(const node& n, const char* /*src*/, std::false_type /*ensure_ascii*/) noexcept
+    {
+        return (n.flags & node_flags::storage) == 0;
+    }
+
+    static bool copied(const node& n, const char* src, std::true_type /*ensure_ascii*/) noexcept
+    {
+        return (n.flags & node_flags::storage) == 0 && is_ascii_without_delete(src + n.off, n.len);
+    }
+
+    /// a string with ensure_ascii (dump_compact()), out of line
+    NLOHMANN_VIEW_NOINLINE void write_ascii_string(const node& n)
+    {
+        write_string(n);
+    }
+
     /// the copies of dump_compact() that are not fixed-size moves (long
     /// strings, or near the end of the source); out of line, so that the
     /// compiler does not merge the fixed-size moves into this call
@@ -417,16 +484,18 @@ class view_serializer
     }
 
     /*!
-    @brief the compact output without ensure_ascii (the default dump())
+    @brief the compact output (the default dump(), or with ensure_ascii)
 
     The same walk as dump(), with the write position in a local variable
     (stores through char pointers would otherwise force a reload of the
     buffer's members after each one), and with strings and number tokens of
     the source copied by fixed-size moves of 32 bytes where the source has
     that many bytes left, instead of a library call per token. The buffer
-    keeps 64 bytes of slack for the overshoot.
+    keeps 64 bytes of slack for the overshoot. With ensure_ascii (Ascii), a
+    string of the source is copied only if it has no byte >= 0x7F (see
+    is_ascii_without_delete()); any other string is written by write_string().
     */
-    template<bool SourceNumbers>
+    template<bool SourceNumbers, bool Ascii>
     void dump_compact(const node* root)
     {
         struct frame
@@ -478,7 +547,14 @@ class view_serializer
         const auto escaped = [&](const node & n)
         {
             m_out.set_cursor(w);
-            write_decoded(n);
+            if (enabled(Ascii))
+            {
+                write_ascii_string(n);
+            }
+            else
+            {
+                write_decoded(n);
+            }
             w = m_out.cursor();
             lim = m_out.limit();
         };
@@ -501,7 +577,7 @@ class view_serializer
             switch (static_cast<value_t>(n->kind))
             {
                 case value_t::string:
-                    if ((n->flags & node_flags::storage) == 0)
+                    if (copied(*n, src, std::integral_constant<bool, Ascii>()))
                     {
                         room(n->len + 2);
                         *w++ = '"';
@@ -632,7 +708,7 @@ class view_serializer
             if (cur.object)
             {
                 const node& key = *at;
-                if ((key.flags & node_flags::storage) == 0)
+                if (copied(key, src, std::integral_constant<bool, Ascii>()))
                 {
                     room(key.len + 3);
                     *w++ = '"';
@@ -843,9 +919,10 @@ class view_serializer
     {
         const char* const s = m_doc.str(n);
         m_out.put('"');
-        if ((n.flags & node_flags::storage) == 0 && !m_style.ensure_ascii)
+        if ((n.flags & node_flags::storage) == 0 && (!m_style.ensure_ascii || is_ascii_without_delete(s, n.len)))
         {
-            // a string of the source without escape sequences has nothing to escape
+            // a string of the source without escape sequences has nothing to
+            // escape (with ensure_ascii, if it is printable ASCII)
             m_out.put(s, n.len);
         }
         else if (m_style.ensure_ascii)

single_include/nlohmann/json_view.hpp

@@ -6237,6 +6237,34 @@ inline std::size_t plain_output_run(const unsigned char* s, std::size_t n) noexc
     return n;
 }
 
+/// Whether dump() with ensure_ascii writes the string at s unchanged, for a
+/// string of the source without escape sequences (which has no quote,
+/// backslash, or control character): no byte is DEL (0x7F) or above. Adding 1
+/// to each byte sets its high bit exactly for 0x7F, and a carry into the next
+/// byte only comes from 0xFF, whose own high bit is set anyway.
+inline bool is_ascii_without_delete(const char* s, std::size_t n) noexcept
+{
+    constexpr std::uint64_t ones = 0x0101010101010101ull;
+    constexpr std::uint64_t high = 0x8080808080808080ull;
+    std::size_t i = 0;
+    for (; i + 8 <= n; i += 8)
+    {
+        const std::uint64_t v = read_eight_bytes(s + i);
+        if (((v | (v + ones)) & high) != 0)
+        {
+            return false;
+        }
+    }
+    for (; i < n; ++i)
+    {
+        if (static_cast<unsigned char>(s[i]) >= 0x7F)
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
 /// A stack that starts in a buffer of the caller (a local array) and moves to
 /// the heap (a vector of the caller) only when that is full, so that dumps of
 /// shallow documents need no allocation. The top is a pointer, as in
@@ -6352,15 +6380,19 @@ class view_serializer
 
     void dump(const node* root)
     {
-        if (!m_style.pretty && !m_style.ensure_ascii)
+        if (!m_style.pretty)
         {
-            if (m_style.source_numbers)
+            if (m_style.ensure_ascii)
             {
-                dump_compact<true>(root);
+                dump_compact_ascii(root);
+            }
+            else if (m_style.source_numbers)
+            {
+                dump_compact<true, false>(root);
             }
             else
             {
-                dump_compact<false>(root);
+                dump_compact<false, false>(root);
             }
             return;
         }
@@ -6466,6 +6498,41 @@ class view_serializer
         m_out.put('"');
     }
 
+    /// dump_compact() with ensure_ascii, out of line: the walk without it is
+    /// faster when it is not merged with this one into dump()
+    NLOHMANN_VIEW_NOINLINE void dump_compact_ascii(const node* root)
+    {
+        if (m_style.source_numbers)
+        {
+            dump_compact<true, true>(root);
+        }
+        else
+        {
+            dump_compact<false, true>(root);
+        }
+    }
+
+    /// whether dump_compact() copies a string or key from the source: one of
+    /// the source, and with ensure_ascii one without bytes >= 0x7F (an overload
+    /// each, so that the walk without ensure_ascii is compiled exactly as
+    /// before: with GCC, a condition that only folds away later changed its
+    /// block layout and made it 4-6% slower)
+    static bool copied(const node& n, const char* /*src*/, std::false_type /*ensure_ascii*/) noexcept
+    {
+        return (n.flags & node_flags::storage) == 0;
+    }
+
+    static bool copied(const node& n, const char* src, std::true_type /*ensure_ascii*/) noexcept
+    {
+        return (n.flags & node_flags::storage) == 0 && is_ascii_without_delete(src + n.off, n.len);
+    }
+
+    /// a string with ensure_ascii (dump_compact()), out of line
+    NLOHMANN_VIEW_NOINLINE void write_ascii_string(const node& n)
+    {
+        write_string(n);
+    }
+
     /// the copies of dump_compact() that are not fixed-size moves (long
     /// strings, or near the end of the source); out of line, so that the
     /// compiler does not merge the fixed-size moves into this call
@@ -6483,16 +6550,18 @@ class view_serializer
     }
 
     /*!
-    @brief the compact output without ensure_ascii (the default dump())
+    @brief the compact output (the default dump(), or with ensure_ascii)
 
     The same walk as dump(), with the write position in a local variable
     (stores through char pointers would otherwise force a reload of the
     buffer's members after each one), and with strings and number tokens of
     the source copied by fixed-size moves of 32 bytes where the source has
     that many bytes left, instead of a library call per token. The buffer
-    keeps 64 bytes of slack for the overshoot.
+    keeps 64 bytes of slack for the overshoot. With ensure_ascii (Ascii), a
+    string of the source is copied only if it has no byte >= 0x7F (see
+    is_ascii_without_delete()); any other string is written by write_string().
     */
-    template<bool SourceNumbers>
+    template<bool SourceNumbers, bool Ascii>
     void dump_compact(const node* root)
     {
         struct frame
@@ -6544,7 +6613,14 @@ class view_serializer
         const auto escaped = [&](const node & n)
         {
             m_out.set_cursor(w);
-            write_decoded(n);
+            if (enabled(Ascii))
+            {
+                write_ascii_string(n);
+            }
+            else
+            {
+                write_decoded(n);
+            }
             w = m_out.cursor();
             lim = m_out.limit();
         };
@@ -6567,7 +6643,7 @@ class view_serializer
             switch (static_cast<value_t>(n->kind))
             {
                 case value_t::string:
-                    if ((n->flags & node_flags::storage) == 0)
+                    if (copied(*n, src, std::integral_constant<bool, Ascii>()))
                     {
                         room(n->len + 2);
                         *w++ = '"';
@@ -6698,7 +6774,7 @@ class view_serializer
             if (cur.object)
             {
                 const node& key = *at;
-                if ((key.flags & node_flags::storage) == 0)
+                if (copied(key, src, std::integral_constant<bool, Ascii>()))
                 {
                     room(key.len + 3);
                     *w++ = '"';
@@ -6909,9 +6985,10 @@ class view_serializer
     {
         const char* const s = m_doc.str(n);
         m_out.put('"');
-        if ((n.flags & node_flags::storage) == 0 && !m_style.ensure_ascii)
+        if ((n.flags & node_flags::storage) == 0 && (!m_style.ensure_ascii || is_ascii_without_delete(s, n.len)))
         {
-            // a string of the source without escape sequences has nothing to escape
+            // a string of the source without escape sequences has nothing to
+            // escape (with ensure_ascii, if it is printable ASCII)
             m_out.put(s, n.len);
         }
         else if (m_style.ensure_ascii)

tests/src/unit-json_view_dump.cpp

@@ -12,6 +12,7 @@
 using nlohmann::json;
 using nlohmann::ordered_json;
 using nlohmann::json_document;
+using nlohmann::ordered_json_editable_document;
 using nlohmann::json_view;
 using nlohmann::ordered_json_document;
 using nlohmann::ordered_json_view;
@@ -219,6 +220,76 @@ TEST_CASE("json_view dump")
         CHECK(d.root()["a"].dump(-1, ' ', true).size() == 2 + (5000 * 6));
     }
 
+    SECTION("ensure_ascii with strings of the source, decoded strings, and edited strings")
+    {
+        // plain ASCII, non-ASCII, escaped (decoded), and mixed strings, as values and as keys;
+        // the second text has a DEL, which ensure_ascii writes as an escape
+        for (const char* const del :
+                {"", "\x7F"
+                })
+        {
+            const std::string text = std::string("{\"plain\":\"abc\",\"caf\xC3\xA9\":\"\xC3\xA9\xE2\x82\xAC\xF0\x9F\x98\x80\",")
+                                     + "\"esc\\n\\\"\":\"a\\tb\\u00e4\\ud83d\\ude00\",\"del\":\"a" + del + "b\",\"\":\"\"," // NOLINT(modernize-raw-string-literal)
+                                     + "\"list\":[\"x\",\"\xC3\xBC\",\"\\\\\",\"long ASCII string with more than thirty-two bytes in it\","
+                                     + "\"long string with more than thirty-two bytes and one non-ASCII byte \xC3\xB6\"],"
+                                     + "\"n\":[1,2.5,-0,true,null]}";
+            const ordered_json j = ordered_json::parse(text);
+            const ordered_json_document d = ordered_json_document::parse(text);
+            CHECK(d.root().dump(-1, ' ', true) == j.dump(-1, ' ', true));
+            CHECK(d.root().dump(4, ' ', true) == j.dump(4, ' ', true));
+            CHECK(d.root().dump() == j.dump());
+            CHECK(d.root()["list"].dump(-1, ' ', true) == j["list"].dump(-1, ' ', true));
+            CHECK(d.root()["del"].dump(-1, ' ', true) == j["del"].dump(-1, ' ', true));
+
+            // strings that an edit wrote (the source strings that they replace were plain)
+            ordered_json_editable_document e = ordered_json_editable_document::parse(text);
+            ordered_json k = j;
+            e.set(e.root()["plain"], std::string("\xC3\xA4gypten"));
+            k["plain"] = "\xC3\xA4gypten";
+            e.set(e.root()["list"][0], std::string("ascii after edit"));
+            k["list"][0] = "ascii after edit";
+            e.set(e.root()["list"][1], std::string("\xC3\xA4"));
+            k["list"][1] = "\xC3\xA4";
+            e.set(e.root()["del"], std::string("tab\there"));
+            k["del"] = "tab\there";
+            CHECK(e.root().dump(-1, ' ', true) == k.dump(-1, ' ', true));
+            CHECK(e.root().dump(2, ' ', true) == k.dump(2, ' ', true));
+            CHECK(e.root().dump() == k.dump());
+
+            // strings that were edited from or to ASCII text
+            e.set(e.root()["list"][2], std::string("\xC3\xA4"));
+            k["list"][2] = "\xC3\xA4";
+            e.set(e.root()["list"][4], std::string("plain again"));
+            k["list"][4] = "plain again";
+            CHECK(e.root().dump(-1, ' ', true) == k.dump(-1, ' ', true));
+        }
+    }
+
+    SECTION("ensure_ascii with DEL or a non-ASCII character at every position")
+    {
+        // the check whether a string of the source has no byte >= 0x7F reads
+        // eight bytes at a time and then the rest one by one
+        for (std::size_t length = 1; length <= 20; ++length)
+        {
+            for (std::size_t pos = 0; pos < length; ++pos)
+            {
+                for (const char* const c :
+                        {
+                            "\x7F", "\x7E", "\xC3\xA4", "\xC3\xBF" // NOLINT(modernize-raw-string-literal)
+                        })
+                {
+                    const std::string s = std::string(pos, 'a') + c + std::string(length - pos - 1, 'b');
+                    std::string text = "[\"";
+                    text.append(s).append("\",{\"").append(s).append("\":1}]");
+                    const json_document d = json_document::parse(text);
+                    const json j = json::parse(text);
+                    CHECK(d.root().dump(-1, ' ', true) == j.dump(-1, ' ', true));
+                    CHECK(d.root().dump(1, ' ', true) == j.dump(1, ' ', true));
+                }
+            }
+        }
+    }
+
     SECTION("streams and discarded views")
     {
         const json_document d = json_document::parse(R"({"a": [1, 2]})");