version 0.34.0

Works by keeping the old prompt unreflowed rather than clearing it.
There may still be some flicker for people using long or right side
prompts, but that cant be avoided, since we cannot know how the shell
will redraw after the resize. But in the common case of a left side
smallish prompt that fits in the resized window, the flicker is
eliminated.

It means we have to do some more copying work on resize, but the nicer
visuals is worth it, IMO.

.gitattributes

@@ -16,11 +16,17 @@ kittens/diff/options/types.py linguist-generated=true
 kittens/diff/options/parse.py linguist-generated=true
 glfw/*.c linguist-vendored=true
 glfw/*.h linguist-vendored=true
+3rdparty/** linguist-vendored=true
 kittens/unicode_input/names.h linguist-generated=true
 tools/wcswidth/std.go linguist-generated=true
 tools/unicode_names/names.txt linguist-generated=true
 terminfo/kitty.term* linguist-generated=true
 terminfo/x/* linguist-generated=true
+*_generated.h linguist-generated=true
+*_generated.go linguist-generated=true
+*_generated_test.go linguist-generated=true
+*_generated_test.s linguist-generated=true
+*_generated.s linguist-generated=true
 
 *.py  text diff=python
 *.m  text diff=objc

.github/workflows/ci.py

@@ -2,6 +2,7 @@
 # vim:fileencoding=utf-8
 # License: GPLv3 Copyright: 2020, Kovid Goyal <kovid at kovidgoyal.net>
 
+import glob
 import io
 import os
 import shlex
@@ -9,6 +10,7 @@ import shutil
 import subprocess
 import sys
 import tarfile
+import time
 from urllib.request import urlopen
 
 BUNDLE_URL = 'https://download.calibre-ebook.com/ci/kitty/{}-64.tar.xz'
@@ -17,14 +19,44 @@ is_macos = 'darwin' in sys.platform.lower()
 SW = None
 
 
-def run(*a):
+def do_print_crash_reports():
+    print('Printing available crash reports...')
+    if is_macos:
+        end_time = time.monotonic() + 90
+        while time.monotonic() < end_time:
+            time.sleep(1)
+            items = glob.glob(os.path.join(os.path.expanduser('~/Library/Logs/DiagnosticReports'), 'kitty-*.ips'))
+            if items:
+                break
+        if items:
+            time.sleep(1)
+            print(os.path.basename(items[0]))
+            sdir = os.path.dirname(os.path.abspath(__file__))
+            subprocess.check_call([sys.executable, os.path.join(sdir, 'macos_crash_report.py'), items[0]])
+    else:
+        run('sh -c "echo bt | coredumpctl debug"')
+    print(flush=True)
+
+
+def run(*a, print_crash_reports=False):
     if len(a) == 1:
         a = shlex.split(a[0])
-    print(' '.join(map(shlex.quote, a)))
+    cmd = ' '.join(map(shlex.quote, a))
+    print(cmd)
     sys.stdout.flush()
     ret = subprocess.Popen(a).wait()
     if ret != 0:
-        raise SystemExit(ret)
+        if ret < 0:
+            import signal
+            try:
+                sig = signal.Signals(-ret)
+            except ValueError:
+                pass
+            else:
+                if print_crash_reports:
+                    do_print_crash_reports()
+                raise SystemExit(f'The following process was killed by signal: {sig.name}:\n{cmd}')
+        raise SystemExit(f'The following process failed with exit code: {ret}:\n{cmd}')
 
 
 def install_deps():
@@ -37,13 +69,13 @@ def install_deps():
         import ssl
         if ssl.OPENSSL_VERSION_INFO[0] == 1:
             openssl += '@1.1'
-        run('brew', 'install', 'fish', openssl, *items)
+        run('brew', 'install', 'fish', 'simde', openssl, *items)
     else:
         run('sudo apt-get update')
         run('sudo apt-get install -y libgl1-mesa-dev libxi-dev libxrandr-dev libxinerama-dev ca-certificates'
             ' libxcursor-dev libxcb-xkb-dev libdbus-1-dev libxkbcommon-dev libharfbuzz-dev libx11-xcb-dev zsh'
             ' libpng-dev liblcms2-dev libfontconfig-dev libxkbcommon-x11-dev libcanberra-dev libxxhash-dev uuid-dev'
-            ' zsh bash dash')
+            ' libsimde-dev zsh bash dash systemd-coredump gdb')
         # for some reason these directories are world writable which causes zsh
         # compinit to break
         run('sudo chmod -R og-w /usr/share/zsh')
@@ -59,13 +91,18 @@ def install_deps():
 def build_kitty():
     python = shutil.which('python3') if is_bundle else sys.executable
     cmd = f'{python} setup.py build --verbose'
+    if is_macos:
+        cmd += ' --debug'  # for better crash report to debug SIGILL issue
     if os.environ.get('KITTY_SANITIZE') == '1':
         cmd += ' --debug --sanitize'
     run(cmd)
 
 
 def test_kitty():
-    run('./test.py')
+    if is_macos:
+        run('ulimit -c unlimited')
+        run('sudo chmod -R 777 /cores')
+    run('./test.py', print_crash_reports=True)
 
 
 def package_kitty():

.github/workflows/ci.yml

@@ -2,7 +2,7 @@ name: CI
 on: [push, pull_request]
 env:
     CI: 'true'
-    ASAN_OPTIONS: leak_check_at_exit=0
+    ASAN_OPTIONS: detect_leaks=0
     LC_ALL: en_US.UTF-8
     LANG: en_US.UTF-8
 

.github/workflows/macos_crash_report.py

@@ -0,0 +1,450 @@
+#!/usr/bin/env python
+# License: GPLv3 Copyright: 2024, Kovid Goyal <kovid at kovidgoyal.net>
+
+import json
+import posixpath
+import sys
+from collections import namedtuple
+from datetime import datetime
+from enum import Enum
+from functools import cached_property
+from typing import IO, List, Mapping, Optional
+
+Frame = namedtuple('Frame', 'image_name image_base image_offset symbol symbol_offset')
+Register = namedtuple('Register', 'name value')
+
+
+def surround(x: str, start: int, end: int) -> str:
+    if sys.stdout.isatty():
+        x = f'\033[{start}m{x}\033[{end}m'
+    return x
+
+
+def cyan(x: str) -> str:
+    return surround(x, 96, 39)
+
+
+def bold(x: str) -> str:
+    return surround(x, 1, 22)
+
+
+class BugType(Enum):
+    WatchdogTimeout = '28'
+    BasebandStats = '195'
+    GPUEvent = '284'
+    Sandbox = '187'
+    TerminatingStackshot = '509'
+    ServiceWatchdogTimeout = '29'
+    Session = '179'
+    LegacyStackshot = '188'
+    MACorrelation = '197'
+    iMessages = '189'
+    log_power = '278'
+    PowerLog = 'powerlog'
+    DuetKnowledgeCollector2 = '58'
+    BridgeRestore = '83'
+    LegacyJetsam = '198'
+    ExcResource_385 = '385'
+    Modem = '199'
+    Stackshot = '288'
+    SystemInformation = 'system_profile'
+    Jetsam_298 = '298'
+    MemoryResource = '30'
+    Bridge = '31'
+    DifferentialPrivacy = 'diff_privacy'
+    FirmwareIntegrity = '32'
+    CoreAnalytics_33 = '33'
+    AutoBugCapture = '34'
+    EfiFirmwareIntegrity = '35'
+    SystemStats = '36'
+    AnonSystemStats = '37'
+    Crash_9 = '9'
+    Jetsam_98 = '98'
+    LDCM = '100'
+    Panic_10 = '10'
+    Spin = '11'
+    CLTM = '101'
+    Hang = '12'
+    Panic_110 = '110'
+    ConnectionFailure = '13'
+    MessageTracer = '14'
+    LowBattery = '120'
+    Siri = '201'
+    ShutdownStall = '17'
+    Panic_210 = '210'
+    SymptomsCPUUsage = '202'
+    AssumptionViolation = '18'
+    CoreHandwriting = 'chw'
+    IOMicroStackShot = '44'
+    CoreAnalytics_211 = '211'
+    SiriAppPrediction = '203'
+    spin_45 = '45'
+    PowerMicroStackshots = '220'
+    BTMetadata = '212'
+    SystemMemoryReset = '301'
+    ResetCount = '115'
+    AutoBugCapture_204 = '204'
+    WifiCrashBinary = '221'
+    MicroRunloopHang = '310'
+    Rosetta = '213'
+    glitchyspin = '302'
+    System = '116'
+    IOPowerSources = '141'
+    PanicStats = '205'
+    PowerLog_230 = '230'
+    LongRunloopHang = '222'
+    HomeProductsAnalytics = '311'
+    DifferentialPrivacy_150 = '150'
+    Rhodes = '214'
+    ProactiveEventTrackerTransparency = '303'
+    WiFi = '117'
+    SymptomsCPUWakes = '142'
+    SymptomsCPUUsageFatal = '206'
+    Crash_109 = '109'
+    ShortRunloopHang = '223'
+    CoreHandwriting_231 = '231'
+    ForceReset = '151'
+    SiriAppSelection = '215'
+    PrivateFederatedLearning = '304'
+    Bluetooth = '118'
+    SCPMotion = '143'
+    HangSpin = '207'
+    StepCount = '160'
+    RTCTransparency = '224'
+    DiagnosticRequest = '312'
+    MemorySnapshot = '152'
+    Rosetta_B = '216'
+    AudioAccessory = '305'
+    General = '119'
+    HotSpotIOMicroSS = '144'
+    GeoServicesTransparency = '233'
+    MotionState = '161'
+    AppStoreTransparency = '225'
+    SiriSearchFeedback = '313'
+    BearTrapReserved = '153'
+    Portrait = '217'
+    AWDMetricLog = 'metriclog'
+    SymptomsIO = '145'
+    SubmissionReserved = '170'
+    WifiCrash = '209'
+    Natalies = '162'
+    SecurityTransparency = '226'
+    BiomeMapReduce = '234'
+    MemoryGraph = '154'
+    MultichannelAudio = '218'
+    honeybee_payload = '146'
+    MesaReserved = '171'
+    WifiSensing = '235'
+    SiriMiss = '163'
+    ExcResourceThreads_227 = '227'
+    TestA = 'T01'
+    NetworkUsage = '155'
+    WifiReserved = '180'
+    SiriActionPrediction = '219'
+    honeybee_heartbeat = '147'
+    ECCEvent = '172'
+    KeyTransparency = '236'
+    SubDiagHeartBeat = '164'
+    ThirdPartyHang = '228'
+    OSFault = '308'
+    CoreTime = '156'
+    WifiDriverReserved = '181'
+    Crash_309 = '309'
+    honeybee_issue = '148'
+    CellularPerfReserved = '173'
+    TestB = 'T02'
+    StorageStatus = '165'
+    SiriNotificationTransparency = '229'
+    TestC = 'T03'
+    CPUMicroSS = '157'
+    AccessoryUpdate = '182'
+    xprotect = '20'
+    MultitouchFirmware = '149'
+    MicroStackshot = '174'
+    AppLaunchDiagnostics = '238'
+    KeyboardAccuracy = '166'
+    GPURestart = '21'
+    FaceTime = '191'
+    DuetKnowledgeCollector = '158'
+    OTASUpdate = '183'
+    ExcResourceThreads_327 = '327'
+    ExcResource_22 = '22'
+    DuetDB = '175'
+    ThirdPartyHangDeveloper = '328'
+    PrivacySettings = '167'
+    GasGauge = '192'
+    MicroStackShots = '23'
+    BasebandCrash = '159'
+    GPURestart_184 = '184'
+    SystemWatchdogCrash = '409'
+    FlashStatus = '176'
+    SleepWakeFailure = '24'
+    CarouselEvent = '168'
+    AggregateD = '193'
+    WakeupsMonitorViolation = '25'
+    DifferentialPrivacy_50 = '50'
+    ExcResource_185 = '185'
+    UIAutomation = '177'
+    ping = '26'
+    SiriTransaction = '169'
+    SURestore = '194'
+    KtraceStackshot = '186'
+    WirelessDiagnostics = '27'
+    PowerLogLite = '178'
+    SKAdNetworkAnalytics = '237'
+    HangWorkflowResponsiveness = '239'
+    CompositorClientHang = '243'
+
+
+class CrashReportBase:
+    def __init__(self, metadata: Mapping, data: str, filename: str = None):
+        self.filename = filename
+        self._metadata = metadata
+        self._data = data
+        self._parse()
+
+    def _parse(self):
+        self._is_json = False
+        try:
+            modified_data = self._data
+            if '\n  \n' in modified_data:
+                modified_data, rest = modified_data.split('\n  \n', 1)
+                rest = '",' + rest.split('",', 1)[1]
+                modified_data += rest
+            self._data = json.loads(modified_data)
+            self._is_json = True
+        except json.decoder.JSONDecodeError:
+            pass
+
+    @cached_property
+    def bug_type(self) -> BugType:
+        return BugType(self.bug_type_str)
+
+    @cached_property
+    def bug_type_str(self) -> str:
+        return self._metadata['bug_type']
+
+    @cached_property
+    def incident_id(self):
+        return self._metadata.get('incident_id')
+
+    @cached_property
+    def timestamp(self) -> datetime:
+        timestamp = self._metadata.get('timestamp')
+        timestamp_without_timezone = timestamp.rsplit(' ', 1)[0]
+        return datetime.strptime(timestamp_without_timezone, '%Y-%m-%d %H:%M:%S.%f')
+
+    @cached_property
+    def name(self) -> str:
+        return self._metadata.get('name')
+
+    def __repr__(self) -> str:
+        filename = ''
+        if self.filename:
+            filename = f'FILENAME:{posixpath.basename(self.filename)} '
+        return f'<{self.__class__} {filename}TIMESTAMP:{self.timestamp}>'
+
+    def __str__(self) -> str:
+        filename = ''
+        if self.filename:
+            filename = self.filename
+
+        return cyan(f'{self.incident_id} {self.timestamp}\n{filename}\n\n')
+
+
+class UserModeCrashReport(CrashReportBase):
+    def _parse_field(self, name: str) -> str:
+        name += ':'
+        for line in self._data.split('\n'):
+            if line.startswith(name):
+                field = line.split(name, 1)[1]
+                field = field.strip()
+                return field
+
+    @cached_property
+    def faulting_thread(self) -> int:
+        if self._is_json:
+            return self._data['faultingThread']
+        else:
+            return int(self._parse_field('Triggered by Thread'))
+
+    @cached_property
+    def frames(self) -> List[Frame]:
+        result = []
+        if self._is_json:
+            thread_index = self.faulting_thread
+            images = self._data['usedImages']
+            for frame in self._data['threads'][thread_index]['frames']:
+                image = images[frame['imageIndex']]
+                result.append(
+                    Frame(image_name=image.get('path'), image_base=image.get('base'), symbol=frame.get('symbol'),
+                          image_offset=frame.get('imageOffset'), symbol_offset=frame.get('symbolLocation')))
+        else:
+            in_frames = False
+            for line in self._data.split('\n'):
+                if in_frames:
+                    splitted = line.split()
+
+                    if len(splitted) == 0:
+                        break
+
+                    assert splitted[-2] == '+'
+                    image_base = splitted[-3]
+                    if image_base.startswith('0x'):
+                        result.append(Frame(image_name=splitted[1], image_base=int(image_base, 16), symbol=None,
+                                            image_offset=int(splitted[-1]), symbol_offset=None))
+                    else:
+                        # symbolicated
+                        result.append(Frame(image_name=splitted[1], image_base=None, symbol=image_base,
+                                            image_offset=None, symbol_offset=int(splitted[-1])))
+
+                if line.startswith(f'Thread {self.faulting_thread} Crashed:'):
+                    in_frames = True
+
+        return result
+
+    @cached_property
+    def registers(self) -> List[Register]:
+        result = []
+        if self._is_json:
+            thread_index = self._data['faultingThread']
+            thread_state = self._data['threads'][thread_index]['threadState']
+
+            if 'x' in thread_state:
+                for i, reg_x in enumerate(thread_state['x']):
+                    result.append(Register(name=f'x{i}', value=reg_x['value']))
+
+            for i, (name, value) in enumerate(thread_state.items()):
+                if name == 'x':
+                    for j, reg_x in enumerate(value):
+                        result.append(Register(name=f'x{j}', value=reg_x['value']))
+                else:
+                    if isinstance(value, dict):
+                        result.append(Register(name=name, value=value['value']))
+        else:
+            in_frames = False
+            for line in self._data.split('\n'):
+                if in_frames:
+                    splitted = line.split()
+
+                    if len(splitted) == 0:
+                        break
+
+                    for i in range(0, len(splitted), 2):
+                        register_name = splitted[i]
+                        if not register_name.endswith(':'):
+                            break
+
+                        register_name = register_name[:-1]
+                        register_value = int(splitted[i + 1], 16)
+
+                        result.append(Register(name=register_name, value=register_value))
+
+                if line.startswith(f'Thread {self.faulting_thread} crashed with ARM Thread State'):
+                    in_frames = True
+
+        return result
+
+    @cached_property
+    def exception_type(self):
+        if self._is_json:
+            return self._data['exception'].get('type')
+        else:
+            return self._parse_field('Exception Type')
+
+    @cached_property
+    def exception_subtype(self) -> Optional[str]:
+        if self._is_json:
+            return self._data['exception'].get('subtype')
+        else:
+            return self._parse_field('Exception Subtype')
+
+    @cached_property
+    def application_specific_information(self) -> Optional[str]:
+        result = ''
+        if self._is_json:
+            asi = self._data.get('asi')
+            if asi is None:
+                return None
+            return asi
+        else:
+            in_frames = False
+            for line in self._data.split('\n'):
+                if in_frames:
+                    line = line.strip()
+                    if len(line) == 0:
+                        break
+
+                    result += line + '\n'
+
+                if line.startswith('Application Specific Information:'):
+                    in_frames = True
+
+        result = result.strip()
+        if not result:
+            return None
+        return result
+
+    def __str__(self) -> str:
+        result = super().__str__()
+        result += bold(f'Exception: {self.exception_type}\n')
+
+        if self.exception_subtype:
+            result += bold('Exception Subtype: ')
+            result += f'{self.exception_subtype}\n'
+
+        if self.application_specific_information:
+            result += bold('Application Specific Information: ')
+            result += str(self.application_specific_information)
+
+        result += '\n'
+
+        result += bold('Registers:')
+        for i, register in enumerate(self.registers):
+            if i % 4 == 0:
+                result += '\n'
+
+            result += f'{register.name} = 0x{register.value:016x} '.rjust(30)
+
+        result += '\n\n'
+
+        result += bold('Frames:\n')
+        for frame in self.frames:
+            image_base = '_HEADER'
+            if frame.image_base is not None:
+                image_base = f'0x{frame.image_base:x}'
+            result += f'\t[{frame.image_name}] {image_base}'
+            if frame.image_offset:
+                result += f' + 0x{frame.image_offset:x}'
+            if frame.symbol is not None:
+                result += f' ({frame.symbol} + 0x{frame.symbol_offset:x})'
+            result += '\n'
+
+        return result
+
+
+def get_crash_report_from_file(crash_report_file: IO) -> CrashReportBase:
+    metadata = json.loads(crash_report_file.readline())
+
+    try:
+        bug_type = BugType(metadata['bug_type'])
+    except ValueError:
+        return CrashReportBase(metadata, crash_report_file.read(), crash_report_file.name)
+
+    bug_type_parsers = {
+        BugType.Crash_109: UserModeCrashReport,
+        BugType.Crash_309: UserModeCrashReport,
+        BugType.ExcResourceThreads_327: UserModeCrashReport,
+        BugType.ExcResource_385: UserModeCrashReport,
+    }
+
+    parser = bug_type_parsers.get(bug_type)
+    if parser is None:
+        return CrashReportBase(metadata, crash_report_file.read(), crash_report_file.name)
+
+    return parser(metadata, crash_report_file.read(), crash_report_file.name)
+
+
+if __name__ == '__main__':
+    with open(sys.argv[-1]) as f:
+        print(get_crash_report_from_file(f))

.gitignore

@@ -4,6 +4,9 @@
 *.bin
 *_stub.pyi
 *_generated.go
+*_generated.s
+*_generated_test.go
+*_generated_test.s
 *_generated.h
 /.dmypy.json
 /dependencies
@@ -18,6 +21,7 @@ __pycache__/
 /glfw/wayland-*-client-protocol.[ch]
 /docs/_build/
 /docs/generated/
+/tools/simdstring/simdstring.test
 /.mypy_cache
 /.ruff_cache
 .DS_Store

3rdparty/base64/LICENSE

@@ -0,0 +1,28 @@
+Copyright (c) 2005-2007, Nick Galbreath
+Copyright (c) 2015-2018, Wojciech Muła
+Copyright (c) 2016-2017, Matthieu Darbois
+Copyright (c) 2013-2022, Alfred Klomp
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+- Redistributions of source code must retain the above copyright notice,
+  this list of conditions and the following disclaimer.
+
+- Redistributions in binary form must reproduce the above copyright
+  notice, this list of conditions and the following disclaimer in the
+  documentation and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

3rdparty/base64/README.md

@@ -0,0 +1,491 @@
+# Fast Base64 stream encoder/decoder
+
+[![Build Status](https://github.com/aklomp/base64/actions/workflows/test.yml/badge.svg)](https://github.com/aklomp/base64/actions/workflows/test.yml)
+
+This is an implementation of a base64 stream encoding/decoding library in C99
+with SIMD (AVX2, AVX512, NEON, AArch64/NEON, SSSE3, SSE4.1, SSE4.2, AVX) and
+[OpenMP](http://www.openmp.org) acceleration. It also contains wrapper functions
+to encode/decode simple length-delimited strings. This library aims to be:
+
+- FAST;
+- easy to use;
+- elegant.
+
+On x86, the library does runtime feature detection. The first time it's called,
+the library will determine the appropriate encoding/decoding routines for the
+machine. It then remembers them for the lifetime of the program. If your
+processor supports AVX2, SSSE3, SSE4.1, SSE4.2 or AVX instructions, the library
+will pick an optimized codec that lets it encode/decode 12 or 24 bytes at a
+time, which gives a speedup of four or more times compared to the "plain"
+bytewise codec.
+
+AVX512 support is only for encoding at present, utilizing the AVX512 VL and VBMI
+instructions. Decoding part reused AVX2 implementations. For CPUs later than
+Cannonlake (manufactured in 2018) supports these instructions.
+
+NEON support is hardcoded to on or off at compile time, because portable
+runtime feature detection is unavailable on ARM.
+
+Even if your processor does not support SIMD instructions, this is a very fast
+library. The fallback routine can process 32 or 64 bits of input in one round,
+depending on your processor's word width, which still makes it significantly
+faster than naive bytewise implementations. On some 64-bit machines, the 64-bit
+routines even outperform the SSSE3 ones.
+
+To the author's knowledge, at the time of original release, this was the only
+Base64 library to offer SIMD acceleration. The author wrote
+[an article](http://www.alfredklomp.com/programming/sse-base64) explaining one
+possible SIMD approach to encoding/decoding Base64. The article can help figure
+out what the code is doing, and why.
+
+Notable features:
+
+- Really fast on x86 and ARM systems by using SIMD vector processing;
+- Can use [OpenMP](http://www.openmp.org) for even more parallel speedups;
+- Really fast on other 32 or 64-bit platforms through optimized routines;
+- Reads/writes blocks of streaming data;
+- Does not dynamically allocate memory;
+- Valid C99 that compiles with pedantic options on;
+- Re-entrant and threadsafe;
+- Unit tested;
+- Uses Duff's Device.
+
+## Acknowledgements
+
+The original AVX2, NEON and Aarch64/NEON codecs were generously contributed by
+[Inkymail](https://github.com/inkymail/base64), who, in their fork, also
+implemented some additional features. Their work is slowly being backported
+into this project.
+
+The SSSE3 and AVX2 codecs were substantially improved by using some very clever
+optimizations described by Wojciech Muła in a
+[series](http://0x80.pl/notesen/2016-01-12-sse-base64-encoding.html) of
+[articles](http://0x80.pl/notesen/2016-01-17-sse-base64-decoding.html).
+His own code is [here](https://github.com/WojciechMula/toys/tree/master/base64).
+
+The AVX512 encoder is based on code from Wojciech Muła's
+[base64simd](https://github.com/WojciechMula/base64simd) library.
+
+The OpenMP implementation was added by Ferry Toth (@htot) from [Exalon Delft](http://www.exalondelft.nl).
+
+## Building
+
+The `lib` directory contains the code for the actual library.
+Typing `make` in the toplevel directory will build `lib/libbase64.o` and `bin/base64`.
+The first is a single, self-contained object file that you can link into your own project.
+The second is a standalone test binary that works similarly to the `base64` system utility.
+
+The matching header file needed to use this library is in `include/libbase64.h`.
+
+To compile just the "plain" library without SIMD codecs, type:
+
+```sh
+make lib/libbase64.o
+```
+
+Optional SIMD codecs can be included by specifying the `AVX2_CFLAGS`, `AVX512_CFLAGS`,
+`NEON32_CFLAGS`, `NEON64_CFLAGS`, `SSSE3_CFLAGS`, `SSE41_CFLAGS`, `SSE42_CFLAGS` and/or `AVX_CFLAGS` environment variables.
+A typical build invocation on x86 looks like this:
+
+```sh
+AVX2_CFLAGS=-mavx2 SSSE3_CFLAGS=-mssse3 SSE41_CFLAGS=-msse4.1 SSE42_CFLAGS=-msse4.2 AVX_CFLAGS=-mavx make lib/libbase64.o
+```
+
+### AVX2
+
+To build and include the AVX2 codec, set the `AVX2_CFLAGS` environment variable to a value that will turn on AVX2 support in your compiler, typically `-mavx2`.
+Example:
+
+```sh
+AVX2_CFLAGS=-mavx2 make
+```
+
+### AVX512
+
+To build and include the AVX512 codec, set the `AVX512_CFLAGS` environment variable to a value that will turn on AVX512 support in your compiler, typically `-mavx512vl -mavx512vbmi`.
+Example:
+
+```sh
+AVX512_CFLAGS="-mavx512vl -mavx512vbmi" make
+```
+
+The codec will only be used if runtime feature detection shows that the target machine supports AVX2.
+
+### SSSE3
+
+To build and include the SSSE3 codec, set the `SSSE3_CFLAGS` environment variable to a value that will turn on SSSE3 support in your compiler, typically `-mssse3`.
+Example:
+
+```sh
+SSSE3_CFLAGS=-mssse3 make
+```
+
+The codec will only be used if runtime feature detection shows that the target machine supports SSSE3.
+
+### NEON
+
+This library includes two NEON codecs: one for regular 32-bit ARM and one for the 64-bit AArch64 with NEON, which has double the amount of SIMD registers and can do full 64-byte table lookups.
+These codecs encode in 48-byte chunks and decode in massive 64-byte chunks, so they had to be augmented with an uint32/64 codec to stay fast on smaller inputs!
+
+Use LLVM/Clang for compiling the NEON codecs.
+The code generation of at least GCC 4.6 (the version shipped with Raspbian and used for testing) contains a bug when compiling `vstq4_u8()`, and the generated assembly code is of low quality.
+NEON intrinsics are a known weak area of GCC.
+Clang does a better job.
+
+NEON support can unfortunately not be portably detected at runtime from userland (the `mrc` instruction is privileged), so the default value for using the NEON codec is determined at compile-time.
+But you can do your own runtime detection.
+You can include the NEON codec and make it the default, then do a runtime check if the CPU has NEON support, and if not, force a downgrade to non-NEON with `BASE64_FORCE_PLAIN`.
+
+These are your options:
+
+1. Don't include NEON support;
+2. build NEON support and make it the default, but build all other code without NEON flags so that you can override the default at runtime with `BASE64_FORCE_PLAIN`;
+3. build everything with NEON support and make it the default;
+4. build everything with NEON support, but don't make it the default (which makes no sense).
+
+For option 1, simply don't specify any NEON-specific compiler flags at all, like so:
+
+```sh
+CC=clang CFLAGS="-march=armv6" make
+```
+
+For option 2, keep your `CFLAGS` plain, but set the `NEON32_CFLAGS` environment variable to a value that will build NEON support.
+The line below, for instance, will build all the code at ARMv6 level, except for the NEON codec, which is built at ARMv7.
+It will also make the NEON codec the default.
+For ARMv6 platforms, override that default at runtime with the `BASE64_FORCE_PLAIN` flag.
+No ARMv7/NEON code will then be touched.
+
+```sh
+CC=clang CFLAGS="-march=armv6" NEON32_CFLAGS="-march=armv7 -mfpu=neon" make
+```
+
+For option 3, put everything in your `CFLAGS` and use a stub, but non-empty, `NEON32_CFLAGS`.
+This example works for the Raspberry Pi 2B V1.1, which has NEON support:
+
+```sh
+CC=clang CFLAGS="-march=armv7 -mtune=cortex-a7" NEON32_CFLAGS="-mfpu=neon" make
+```
+
+To build and include the NEON64 codec, use `CFLAGS` as usual to define the platform and set `NEON64_CFLAGS` to a nonempty stub.
+(The AArch64 target has mandatory NEON64 support.)
+Example:
+
+```sh
+CC=clang CFLAGS="--target=aarch64-linux-gnu -march=armv8-a" NEON64_CFLAGS=" " make
+```
+
+### OpenMP
+
+To enable OpenMP on GCC you need to build with `-fopenmp`. This can be by setting the `OPENMP` environment variable to `1`.
+
+Example:
+
+```sh
+OPENMP=1 make
+```
+
+This will let the compiler define `_OPENMP`, which in turn will include the OpenMP optimized `lib_openmp.c` into `lib.c`.
+
+By default the number of parallel threads will be equal to the number of cores of the processor.
+On a quad core with hyperthreading eight cores will be detected, but hyperthreading will not increase the performance.
+
+To get verbose information about OpenMP start the program with `OMP_DISPLAY_ENV=VERBOSE`, for instance
+
+```sh
+OMP_DISPLAY_ENV=VERBOSE test/benchmark
+```
+
+To put a limit on the number of threads, start the program with `OMP_THREAD_LIMIT=n`, for instance
+
+```sh
+OMP_THREAD_LIMIT=2 test/benchmark
+```
+
+An example of running a benchmark with OpenMP, SSSE3 and AVX2 enabled:
+
+```sh
+make clean && OPENMP=1 SSSE3_CFLAGS=-mssse3 AVX2_CFLAGS=-mavx2 make && OPENMP=1 make -C test
+```
+
+## API reference
+
+Strings are represented as a pointer and a length; they are not
+zero-terminated. This was a conscious design decision. In the decoding step,
+relying on zero-termination would make no sense since the output could contain
+legitimate zero bytes. In the encoding step, returning the length saves the
+overhead of calling `strlen()` on the output. If you insist on the trailing
+zero, you can easily add it yourself at the given offset.
+
+### Flags
+
+Some API calls take a `flags` argument.
+That argument can be used to force the use of a specific codec, even if that codec is a no-op in the current build.
+Mainly there for testing purposes, this is also useful on ARM where the only way to do runtime NEON detection is to ask the OS if it's available.
+The following constants can be used:
+
+- `BASE64_FORCE_AVX2`
+- `BASE64_FORCE_AVX512`
+- `BASE64_FORCE_NEON32`
+- `BASE64_FORCE_NEON64`
+- `BASE64_FORCE_PLAIN`
+- `BASE64_FORCE_SSSE3`
+- `BASE64_FORCE_SSE41`
+- `BASE64_FORCE_SSE42`
+- `BASE64_FORCE_AVX`
+
+Set `flags` to `0` for the default behavior, which is runtime feature detection on x86, a compile-time fixed codec on ARM, and the plain codec on other platforms.
+
+### Encoding
+
+#### base64_encode
+
+```c
+void base64_encode
+    ( const char  *src
+    , size_t       srclen
+    , char        *out
+    , size_t      *outlen
+    , int          flags
+    ) ;
+```
+
+Wrapper function to encode a plain string of given length.
+Output is written to `out` without trailing zero.
+Output length in bytes is written to `outlen`.
+The buffer in `out` has been allocated by the caller and is at least 4/3 the size of the input.
+
+#### base64_stream_encode_init
+
+```c
+void base64_stream_encode_init
+    ( struct base64_state  *state
+    , int                   flags
+    ) ;
+```
+
+Call this before calling `base64_stream_encode()` to init the state.
+
+#### base64_stream_encode
+
+```c
+void base64_stream_encode
+    ( struct base64_state  *state
+    , const char           *src
+    , size_t                srclen
+    , char                 *out
+    , size_t               *outlen
+    ) ;
+```
+
+Encodes the block of data of given length at `src`, into the buffer at `out`.
+Caller is responsible for allocating a large enough out-buffer; it must be at least 4/3 the size of the in-buffer, but take some margin.
+Places the number of new bytes written into `outlen` (which is set to zero when the function starts).
+Does not zero-terminate or finalize the output.
+
+#### base64_stream_encode_final
+
+```c
+void base64_stream_encode_final
+    ( struct base64_state  *state
+    , char                 *out
+    , size_t               *outlen
+    ) ;
+```
+
+Finalizes the output begun by previous calls to `base64_stream_encode()`.
+Adds the required end-of-stream markers if appropriate.
+`outlen` is modified and will contain the number of new bytes written at `out` (which will quite often be zero).
+
+### Decoding
+
+#### base64_decode
+
+```c
+int base64_decode
+    ( const char  *src
+    , size_t       srclen
+    , char        *out
+    , size_t      *outlen
+    , int          flags
+    ) ;
+```
+
+Wrapper function to decode a plain string of given length.
+Output is written to `out` without trailing zero. Output length in bytes is written to `outlen`.
+The buffer in `out` has been allocated by the caller and is at least 3/4 the size of the input.
+Returns `1` for success, and `0` when a decode error has occured due to invalid input.
+Returns `-1` if the chosen codec is not included in the current build.
+
+#### base64_stream_decode_init
+
+```c
+void base64_stream_decode_init
+    ( struct base64_state  *state
+    , int                   flags
+    ) ;
+```
+
+Call this before calling `base64_stream_decode()` to init the state.
+
+#### base64_stream_decode
+
+```c
+int base64_stream_decode
+    ( struct base64_state  *state
+    , const char           *src
+    , size_t                srclen
+    , char                 *out
+    , size_t               *outlen
+    ) ;
+```
+
+Decodes the block of data of given length at `src`, into the buffer at `out`.
+Caller is responsible for allocating a large enough out-buffer; it must be at least 3/4 the size of the in-buffer, but take some margin.
+Places the number of new bytes written into `outlen` (which is set to zero when the function starts).
+Does not zero-terminate the output.
+Returns 1 if all is well, and 0 if a decoding error was found, such as an invalid character.
+Returns -1 if the chosen codec is not included in the current build.
+Used by the test harness to check whether a codec is available for testing.
+
+## Examples
+
+A simple example of encoding a static string to base64 and printing the output
+to stdout:
+
+```c
+#include <stdio.h>	/* fwrite */
+#include "libbase64.h"
+
+int main ()
+{
+	char src[] = "hello world";
+	char out[20];
+	size_t srclen = sizeof(src) - 1;
+	size_t outlen;
+
+	base64_encode(src, srclen, out, &outlen, 0);
+
+	fwrite(out, outlen, 1, stdout);
+
+	return 0;
+}
+```
+
+A simple example (no error checking, etc) of stream encoding standard input to
+standard output:
+
+```c
+#include <stdio.h>
+#include "libbase64.h"
+
+int main ()
+{
+	size_t nread, nout;
+	char buf[12000], out[16000];
+	struct base64_state state;
+
+	// Initialize stream encoder:
+	base64_stream_encode_init(&state, 0);
+
+	// Read contents of stdin into buffer:
+	while ((nread = fread(buf, 1, sizeof(buf), stdin)) > 0) {
+
+		// Encode buffer:
+		base64_stream_encode(&state, buf, nread, out, &nout);
+
+		// If there's output, print it to stdout:
+		if (nout) {
+			fwrite(out, nout, 1, stdout);
+		}
+
+		// If an error occurred, exit the loop:
+		if (feof(stdin)) {
+			break;
+		}
+	}
+
+	// Finalize encoding:
+	base64_stream_encode_final(&state, out, &nout);
+
+	// If the finalizing resulted in extra output bytes, print them:
+	if (nout) {
+		fwrite(out, nout, 1, stdout);
+	}
+
+	return 0;
+}
+```
+
+Also see `bin/base64.c` for a simple re-implementation of the `base64` utility.
+A file or standard input is fed through the encoder/decoder, and the output is
+written to standard output.
+
+## Tests
+
+See `tests/` for a small test suite. Testing is automated with
+[GitHub Actions](https://github.com/aklomp/base64/actions), which builds and
+tests the code across various architectures.
+
+## Benchmarks
+
+Benchmarks can be run with the built-in benchmark program as follows:
+
+```sh
+make -C test benchmark <buildflags> && test/benchmark
+```
+
+It will run an encoding and decoding benchmark for all of the compiled-in codecs.
+
+The tables below contain some results on random machines. All numbers measured with a 10MB buffer in MB/sec, rounded to the nearest integer.
+
+\*: Update needed
+
+x86 processors
+
+| Processor                                 | Plain enc | Plain dec | SSSE3 enc | SSSE3 dec | AVX enc | AVX dec | AVX2 enc | AVX2 dec |
+|-------------------------------------------|----------:|----------:|----------:|----------:|--------:|--------:|---------:|---------:|
+| i7-4771 @ 3.5 GHz                         | 833\*     | 1111\*    | 3333\*    | 4444\*    | TBD     | TBD     | 4999\*   | 6666\*   |
+| i7-4770 @ 3.4 GHz DDR1600                 | 1790\*    | 3038\*    | 4899\*    | 4043\*    | 4796\*  | 5709\*  | 4681\*   | 6386\*   |
+| i7-4770 @ 3.4 GHz DDR1600 OPENMP 1 thread | 1784\*    | 3041\*    | 4945\*    | 4035\*    | 4776\*  | 5719\*  | 4661\*   | 6294\*   |
+| i7-4770 @ 3.4 GHz DDR1600 OPENMP 2 thread | 3401\*    | 5729\*    | 5489\*    | 7444\*    | 5003\*  | 8624\*  | 5105\*   | 8558\*   |
+| i7-4770 @ 3.4 GHz DDR1600 OPENMP 4 thread | 4884\*    | 7099\*    | 4917\*    | 7057\*    | 4799\*  | 7143\*  | 4902\*   | 7219\*   |
+| i7-4770 @ 3.4 GHz DDR1600 OPENMP 8 thread | 5212\*    | 8849\*    | 5284\*    | 9099\*    | 5289\*  | 9220\*  | 4849\*   | 9200\*   |
+| i7-4870HQ @ 2.5 GHz                       | 1471\*    | 3066\*    | 6721\*    | 6962\*    | 7015\*  | 8267\*  | 8328\*   | 11576\*  |
+| i5-4590S @ 3.0 GHz                        | 3356      | 3197      | 4363      | 6104      | 4243\*  | 6233    | 4160\*   | 6344     |
+| Xeon X5570 @ 2.93 GHz                     | 2161      | 1508      | 3160      | 3915      | -       | -       | -        | -        |
+| Pentium4 @ 3.4 GHz                        | 896       | 740       | -         | -         | -       | -       | -        | -        |
+| Atom N270                                 | 243       | 266       | 508       | 387       | -       | -       | -        | -        |
+| AMD E-450                                 | 645       | 564       | 625       | 634       | -       | -       | -        | -        |
+| Intel Edison @ 500 MHz                    | 79\*      | 92\*      | 152\*     | 172\*     | -       | -       | -        | -        |
+| Intel Edison @ 500 MHz OPENMP 2 thread    | 158\*     | 184\*     | 300\*     | 343\*     | -       | -       | -        | -        |
+| Intel Edison @ 500 MHz (x86-64)           | 162       | 119       | 209       | 164       | -       | -       | -        | -        |
+| Intel Edison @ 500 MHz (x86-64) 2 thread  | 319       | 237       | 412       | 329       | -       | -       | -        | -        |
+
+ARM processors
+
+| Processor                                 | Plain enc | Plain dec | NEON32 enc | NEON32 dec | NEON64 enc | NEON64 dec |
+|-------------------------------------------|----------:|----------:|-----------:|-----------:|-----------:|-----------:|
+| Raspberry PI B+ V1.2                      | 46\*      | 40\*      | -          | -          | -          | -          |
+| Raspberry PI 2 B V1.1                     | 85        | 141       | 300        | 225        | -          | -          |
+| Apple iPhone SE armv7                     | 1056\*    | 895\*     | 2943\*     | 2618\*     | -          | -          |
+| Apple iPhone SE arm64                     | 1061\*    | 1239\*    | -          | -          | 4098\*     | 3983\*     |
+
+PowerPC processors
+
+| Processor                                 | Plain enc | Plain dec |
+|-------------------------------------------|----------:|----------:|
+| PowerPC E6500 @ 1.8GHz                    | 270\*     | 265\*     |
+
+
+Benchmarks on i7-4770 @ 3.4 GHz DDR1600 with varrying buffer sizes:
+![Benchmarks](base64-benchmarks.png)
+
+Note: optimal buffer size to take advantage of the cache is in the range of 100 kB to 1 MB, leading to 12x faster AVX encoding/decoding compared to Plain, or a throughput of 24/27GB/sec.
+Also note the performance degradation when the buffer size is less than 10 kB due to thread creation overhead.
+To prevent this from happening `lib_openmp.c` defines `OMP_THRESHOLD 20000`, requiring at least a 20000 byte buffer to enable multithreading.
+
+## License
+
+This repository is licensed under the
+[BSD 2-clause License](http://opensource.org/licenses/BSD-2-Clause). See the
+LICENSE file.

3rdparty/base64/include/libbase64.h

@@ -0,0 +1,146 @@
+#ifndef LIBBASE64_H
+#define LIBBASE64_H
+
+#include <stddef.h>	/* size_t */
+
+
+#if defined(_WIN32) || defined(__CYGWIN__)
+#define BASE64_SYMBOL_IMPORT __declspec(dllimport)
+#define BASE64_SYMBOL_EXPORT __declspec(dllexport)
+#define BASE64_SYMBOL_PRIVATE
+
+#elif __GNUC__ >= 4
+#define BASE64_SYMBOL_IMPORT   __attribute__ ((visibility ("default")))
+#define BASE64_SYMBOL_EXPORT   __attribute__ ((visibility ("default")))
+#define BASE64_SYMBOL_PRIVATE  __attribute__ ((visibility ("hidden")))
+
+#else
+#define BASE64_SYMBOL_IMPORT
+#define BASE64_SYMBOL_EXPORT
+#define BASE64_SYMBOL_PRIVATE
+#endif
+
+#if defined(BASE64_STATIC_DEFINE)
+#define BASE64_EXPORT
+#define BASE64_NO_EXPORT
+
+#else
+#if defined(BASE64_EXPORTS) // defined if we are building the shared library
+#define BASE64_EXPORT BASE64_SYMBOL_EXPORT
+
+#else
+#define BASE64_EXPORT BASE64_SYMBOL_IMPORT
+#endif
+
+#define BASE64_NO_EXPORT BASE64_SYMBOL_PRIVATE
+#endif
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* These are the flags that can be passed in the `flags` argument. The values
+ * below force the use of a given codec, even if that codec is a no-op in the
+ * current build. Used in testing. Set to 0 for the default behavior, which is
+ * runtime feature detection on x86, a compile-time fixed codec on ARM, and
+ * the plain codec on other platforms: */
+#define BASE64_FORCE_AVX2	(1 << 0)
+#define BASE64_FORCE_NEON32	(1 << 1)
+#define BASE64_FORCE_NEON64	(1 << 2)
+#define BASE64_FORCE_PLAIN	(1 << 3)
+#define BASE64_FORCE_SSSE3	(1 << 4)
+#define BASE64_FORCE_SSE41	(1 << 5)
+#define BASE64_FORCE_SSE42	(1 << 6)
+#define BASE64_FORCE_AVX	(1 << 7)
+#define BASE64_FORCE_AVX512	(1 << 8)
+
+struct base64_state {
+	int eof;
+	int bytes;
+	int flags;
+	unsigned char carry;
+};
+
+/* Wrapper function to encode a plain string of given length. Output is written
+ * to *out without trailing zero. Output length in bytes is written to *outlen.
+ * The buffer in `out` has been allocated by the caller and is at least 4/3 the
+ * size of the input. See above for `flags`; set to 0 for default operation: */
+void BASE64_EXPORT base64_encode
+	( const char		*src
+	, size_t		 srclen
+	, char			*out
+	, size_t		*outlen
+	, int			 flags
+	) ;
+
+/* Call this before calling base64_stream_encode() to init the state. See above
+ * for `flags`; set to 0 for default operation: */
+void BASE64_EXPORT base64_stream_encode_init
+	( struct base64_state	*state
+	, int			 flags
+	) ;
+
+/* Encodes the block of data of given length at `src`, into the buffer at
+ * `out`. Caller is responsible for allocating a large enough out-buffer; it
+ * must be at least 4/3 the size of the in-buffer, but take some margin. Places
+ * the number of new bytes written into `outlen` (which is set to zero when the
+ * function starts). Does not zero-terminate or finalize the output. */
+void BASE64_EXPORT base64_stream_encode
+	( struct base64_state	*state
+	, const char		*src
+	, size_t		 srclen
+	, char			*out
+	, size_t		*outlen
+	) ;
+
+/* Finalizes the output begun by previous calls to `base64_stream_encode()`.
+ * Adds the required end-of-stream markers if appropriate. `outlen` is modified
+ * and will contain the number of new bytes written at `out` (which will quite
+ * often be zero). */
+void BASE64_EXPORT base64_stream_encode_final
+	( struct base64_state	*state
+	, char			*out
+	, size_t		*outlen
+	) ;
+
+/* Wrapper function to decode a plain string of given length. Output is written
+ * to *out without trailing zero. Output length in bytes is written to *outlen.
+ * The buffer in `out` has been allocated by the caller and is at least 3/4 the
+ * size of the input. See above for `flags`, set to 0 for default operation: */
+int BASE64_EXPORT base64_decode
+	( const char		*src
+	, size_t		 srclen
+	, char			*out
+	, size_t		*outlen
+	, int			 flags
+	) ;
+
+/* Call this before calling base64_stream_decode() to init the state. See above
+ * for `flags`; set to 0 for default operation: */
+void BASE64_EXPORT base64_stream_decode_init
+	( struct base64_state	*state
+	, int			 flags
+	) ;
+
+/* Decodes the block of data of given length at `src`, into the buffer at
+ * `out`. Caller is responsible for allocating a large enough out-buffer; it
+ * must be at least 3/4 the size of the in-buffer, but take some margin. Places
+ * the number of new bytes written into `outlen` (which is set to zero when the
+ * function starts). Does not zero-terminate the output. Returns 1 if all is
+ * well, and 0 if a decoding error was found, such as an invalid character.
+ * Returns -1 if the chosen codec is not included in the current build. Used by
+ * the test harness to check whether a codec is available for testing. */
+int BASE64_EXPORT base64_stream_decode
+	( struct base64_state	*state
+	, const char		*src
+	, size_t		 srclen
+	, char			*out
+	, size_t		*outlen
+	) ;
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* LIBBASE64_H */

3rdparty/base64/lib/arch/avx/codec.c

@@ -0,0 +1,66 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdlib.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if HAVE_AVX
+#include <immintrin.h>
+
+// Only enable inline assembly on supported compilers and on 64-bit CPUs.
+#ifndef BASE64_AVX_USE_ASM
+# if (defined(__GNUC__) || defined(__clang__)) && BASE64_WORDSIZE == 64
+#  define BASE64_AVX_USE_ASM 1
+# else
+#  define BASE64_AVX_USE_ASM 0
+# endif
+#endif
+
+#include "../ssse3/dec_reshuffle.c"
+#include "../ssse3/dec_loop.c"
+
+#if BASE64_AVX_USE_ASM
+# include "enc_loop_asm.c"
+#else
+# include "../ssse3/enc_translate.c"
+# include "../ssse3/enc_reshuffle.c"
+# include "../ssse3/enc_loop.c"
+#endif
+
+#endif	// HAVE_AVX
+
+BASE64_ENC_FUNCTION(avx)
+{
+#if HAVE_AVX
+	#include "../generic/enc_head.c"
+
+	// For supported compilers, use a hand-optimized inline assembly
+	// encoder. Otherwise fall back on the SSSE3 encoder, but compiled with
+	// AVX flags to generate better optimized AVX code.
+
+#if BASE64_AVX_USE_ASM
+	enc_loop_avx(&s, &slen, &o, &olen);
+#else
+	enc_loop_ssse3(&s, &slen, &o, &olen);
+#endif
+
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(avx)
+{
+#if HAVE_AVX
+	#include "../generic/dec_head.c"
+	dec_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/avx/enc_loop_asm.c

@@ -0,0 +1,264 @@
+// Apologies in advance for combining the preprocessor with inline assembly,
+// two notoriously gnarly parts of C, but it was necessary to avoid a lot of
+// code repetition. The preprocessor is used to template large sections of
+// inline assembly that differ only in the registers used. If the code was
+// written out by hand, it would become very large and hard to audit.
+
+// Generate a block of inline assembly that loads register R0 from memory. The
+// offset at which the register is loaded is set by the given round.
+#define LOAD(R0, ROUND) \
+	"vlddqu ("#ROUND" * 12)(%[src]), %["R0"] \n\t"
+
+// Generate a block of inline assembly that deinterleaves and shuffles register
+// R0 using preloaded constants. Outputs in R0 and R1.
+#define SHUF(R0, R1, R2) \
+	"vpshufb  %[lut0], %["R0"], %["R1"] \n\t" \
+	"vpand    %["R1"], %[msk0], %["R2"] \n\t" \
+	"vpand    %["R1"], %[msk2], %["R1"] \n\t" \
+	"vpmulhuw %["R2"], %[msk1], %["R2"] \n\t" \
+	"vpmullw  %["R1"], %[msk3], %["R1"] \n\t" \
+	"vpor     %["R1"], %["R2"], %["R1"] \n\t"
+
+// Generate a block of inline assembly that takes R0 and R1 and translates
+// their contents to the base64 alphabet, using preloaded constants.
+#define TRAN(R0, R1, R2) \
+	"vpsubusb %[n51],  %["R1"], %["R0"] \n\t" \
+	"vpcmpgtb %[n25],  %["R1"], %["R2"] \n\t" \
+	"vpsubb   %["R2"], %["R0"], %["R0"] \n\t" \
+	"vpshufb  %["R0"], %[lut1], %["R2"] \n\t" \
+	"vpaddb   %["R1"], %["R2"], %["R0"] \n\t"
+
+// Generate a block of inline assembly that stores the given register R0 at an
+// offset set by the given round.
+#define STOR(R0, ROUND) \
+	"vmovdqu %["R0"], ("#ROUND" * 16)(%[dst]) \n\t"
+
+// Generate a block of inline assembly that generates a single self-contained
+// encoder round: fetch the data, process it, and store the result. Then update
+// the source and destination pointers.
+#define ROUND() \
+	LOAD("a", 0) \
+	SHUF("a", "b", "c") \
+	TRAN("a", "b", "c") \
+	STOR("a", 0) \
+	"add $12, %[src] \n\t" \
+	"add $16, %[dst] \n\t"
+
+// Define a macro that initiates a three-way interleaved encoding round by
+// preloading registers a, b and c from memory.
+// The register graph shows which registers are in use during each step, and
+// is a visual aid for choosing registers for that step. Symbol index:
+//
+//  +  indicates that a register is loaded by that step.
+//  |  indicates that a register is in use and must not be touched.
+//  -  indicates that a register is decommissioned by that step.
+//  x  indicates that a register is used as a temporary by that step.
+//  V  indicates that a register is an input or output to the macro.
+//
+#define ROUND_3_INIT() 			/*  a b c d e f  */ \
+	LOAD("a",   0)			/*  +            */ \
+	SHUF("a", "d", "e")		/*  |     + x    */ \
+	LOAD("b",   1)			/*  | +   |      */ \
+	TRAN("a", "d", "e")		/*  | |   - x    */ \
+	LOAD("c",   2)			/*  V V V        */
+
+// Define a macro that translates, shuffles and stores the input registers A, B
+// and C, and preloads registers D, E and F for the next round.
+// This macro can be arbitrarily daisy-chained by feeding output registers D, E
+// and F back into the next round as input registers A, B and C. The macro
+// carefully interleaves memory operations with data operations for optimal
+// pipelined performance.
+
+#define ROUND_3(ROUND, A,B,C,D,E,F) 	/*  A B C D E F  */ \
+	LOAD(D, (ROUND + 3))		/*  V V V +      */ \
+	SHUF(B, E, F)			/*  | | | | + x  */ \
+	STOR(A, (ROUND + 0))		/*  - | | | |    */ \
+	TRAN(B, E, F)			/*    | | | - x  */ \
+	LOAD(E, (ROUND + 4))		/*    | | | +    */ \
+	SHUF(C, A, F)			/*  + | | | | x  */ \
+	STOR(B, (ROUND + 1))		/*  | - | | |    */ \
+	TRAN(C, A, F)			/*  -   | | | x  */ \
+	LOAD(F, (ROUND + 5))		/*      | | | +  */ \
+	SHUF(D, A, B)			/*  + x | | | |  */ \
+	STOR(C, (ROUND + 2))		/*  |   - | | |  */ \
+	TRAN(D, A, B)			/*  - x   V V V  */
+
+// Define a macro that terminates a ROUND_3 macro by taking pre-loaded
+// registers D, E and F, and translating, shuffling and storing them.
+#define ROUND_3_END(ROUND, A,B,C,D,E,F)	/*  A B C D E F  */ \
+	SHUF(E, A, B)			/*  + x   V V V  */ \
+	STOR(D, (ROUND + 3))		/*  |     - | |  */ \
+	TRAN(E, A, B)			/*  - x     | |  */ \
+	SHUF(F, C, D)			/*      + x | |  */ \
+	STOR(E, (ROUND + 4))		/*      |   - |  */ \
+	TRAN(F, C, D)			/*      - x   |  */ \
+	STOR(F, (ROUND + 5))		/*            -  */
+
+// Define a type A round. Inputs are a, b, and c, outputs are d, e, and f.
+#define ROUND_3_A(ROUND) \
+	ROUND_3(ROUND, "a", "b", "c", "d", "e", "f")
+
+// Define a type B round. Inputs and outputs are swapped with regard to type A.
+#define ROUND_3_B(ROUND) \
+	ROUND_3(ROUND, "d", "e", "f", "a", "b", "c")
+
+// Terminating macro for a type A round.
+#define ROUND_3_A_LAST(ROUND) \
+	ROUND_3_A(ROUND) \
+	ROUND_3_END(ROUND, "a", "b", "c", "d", "e", "f")
+
+// Terminating macro for a type B round.
+#define ROUND_3_B_LAST(ROUND) \
+	ROUND_3_B(ROUND) \
+	ROUND_3_END(ROUND, "d", "e", "f", "a", "b", "c")
+
+// Suppress clang's warning that the literal string in the asm statement is
+// overlong (longer than the ISO-mandated minimum size of 4095 bytes for C99
+// compilers). It may be true, but the goal here is not C99 portability.
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+
+static inline void
+enc_loop_avx (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	// For a clearer explanation of the algorithm used by this function,
+	// please refer to the plain (not inline assembly) implementation. This
+	// function follows the same basic logic.
+
+	if (*slen < 16) {
+		return;
+	}
+
+	// Process blocks of 12 bytes at a time. Input is read in blocks of 16
+	// bytes, so "reserve" four bytes from the input buffer to ensure that
+	// we never read beyond the end of the input buffer.
+	size_t rounds = (*slen - 4) / 12;
+
+	*slen -= rounds * 12;   // 12 bytes consumed per round
+	*olen += rounds * 16;   // 16 bytes produced per round
+
+	// Number of times to go through the 36x loop.
+	size_t loops = rounds / 36;
+
+	// Number of rounds remaining after the 36x loop.
+	rounds %= 36;
+
+	// Lookup tables.
+	const __m128i lut0 = _mm_set_epi8(
+		10, 11,  9, 10,  7,  8,  6,  7,  4,  5,  3,  4,  1,  2,  0,  1);
+
+	const __m128i lut1 = _mm_setr_epi8(
+		65, 71, -4, -4, -4, -4, -4, -4, -4, -4, -4, -4, -19, -16, 0, 0);
+
+	// Temporary registers.
+	__m128i a, b, c, d, e, f;
+
+	__asm__ volatile (
+
+		// If there are 36 rounds or more, enter a 36x unrolled loop of
+		// interleaved encoding rounds. The rounds interleave memory
+		// operations (load/store) with data operations (table lookups,
+		// etc) to maximize pipeline throughput.
+		"    test %[loops], %[loops] \n\t"
+		"    jz   18f                \n\t"
+		"    jmp  36f                \n\t"
+		"                            \n\t"
+		".balign 64                  \n\t"
+		"36: " ROUND_3_INIT()
+		"    " ROUND_3_A( 0)
+		"    " ROUND_3_B( 3)
+		"    " ROUND_3_A( 6)
+		"    " ROUND_3_B( 9)
+		"    " ROUND_3_A(12)
+		"    " ROUND_3_B(15)
+		"    " ROUND_3_A(18)
+		"    " ROUND_3_B(21)
+		"    " ROUND_3_A(24)
+		"    " ROUND_3_B(27)
+		"    " ROUND_3_A_LAST(30)
+		"    add $(12 * 36), %[src] \n\t"
+		"    add $(16 * 36), %[dst] \n\t"
+		"    dec %[loops]           \n\t"
+		"    jnz 36b                \n\t"
+
+		// Enter an 18x unrolled loop for rounds of 18 or more.
+		"18: cmp $18, %[rounds] \n\t"
+		"    jl  9f             \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A(0)
+		"    " ROUND_3_B(3)
+		"    " ROUND_3_A(6)
+		"    " ROUND_3_B(9)
+		"    " ROUND_3_A_LAST(12)
+		"    sub $18,        %[rounds] \n\t"
+		"    add $(12 * 18), %[src]    \n\t"
+		"    add $(16 * 18), %[dst]    \n\t"
+
+		// Enter a 9x unrolled loop for rounds of 9 or more.
+		"9:  cmp $9, %[rounds] \n\t"
+		"    jl  6f            \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A(0)
+		"    " ROUND_3_B_LAST(3)
+		"    sub $9,        %[rounds] \n\t"
+		"    add $(12 * 9), %[src]    \n\t"
+		"    add $(16 * 9), %[dst]    \n\t"
+
+		// Enter a 6x unrolled loop for rounds of 6 or more.
+		"6:  cmp $6, %[rounds] \n\t"
+		"    jl  55f           \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A_LAST(0)
+		"    sub $6,        %[rounds] \n\t"
+		"    add $(12 * 6), %[src]    \n\t"
+		"    add $(16 * 6), %[dst]    \n\t"
+
+		// Dispatch the remaining rounds 0..5.
+		"55: cmp $3, %[rounds] \n\t"
+		"    jg  45f           \n\t"
+		"    je  3f            \n\t"
+		"    cmp $1, %[rounds] \n\t"
+		"    jg  2f            \n\t"
+		"    je  1f            \n\t"
+		"    jmp 0f            \n\t"
+
+		"45: cmp $4, %[rounds] \n\t"
+		"    je  4f            \n\t"
+
+		// Block of non-interlaced encoding rounds, which can each
+		// individually be jumped to. Rounds fall through to the next.
+		"5: " ROUND()
+		"4: " ROUND()
+		"3: " ROUND()
+		"2: " ROUND()
+		"1: " ROUND()
+		"0: \n\t"
+
+		// Outputs (modified).
+		: [rounds] "+r"  (rounds),
+		  [loops]  "+r"  (loops),
+		  [src]    "+r"  (*s),
+		  [dst]    "+r"  (*o),
+		  [a]      "=&x" (a),
+		  [b]      "=&x" (b),
+		  [c]      "=&x" (c),
+		  [d]      "=&x" (d),
+		  [e]      "=&x" (e),
+		  [f]      "=&x" (f)
+
+		// Inputs (not modified).
+		: [lut0] "x" (lut0),
+		  [lut1] "x" (lut1),
+		  [msk0] "x" (_mm_set1_epi32(0x0FC0FC00)),
+		  [msk1] "x" (_mm_set1_epi32(0x04000040)),
+		  [msk2] "x" (_mm_set1_epi32(0x003F03F0)),
+		  [msk3] "x" (_mm_set1_epi32(0x01000010)),
+		  [n51]  "x" (_mm_set1_epi8(51)),
+		  [n25]  "x" (_mm_set1_epi8(25))
+
+		// Clobbers.
+		: "cc", "memory"
+	);
+}
+
+#pragma GCC diagnostic pop

3rdparty/base64/lib/arch/avx2/codec.c

@@ -0,0 +1,56 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdlib.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if HAVE_AVX2
+#include <immintrin.h>
+
+// Only enable inline assembly on supported compilers and on 64-bit CPUs.
+#ifndef BASE64_AVX2_USE_ASM
+# if (defined(__GNUC__) || defined(__clang__)) && BASE64_WORDSIZE == 64
+#  define BASE64_AVX2_USE_ASM 1
+# else
+#  define BASE64_AVX2_USE_ASM 0
+# endif
+#endif
+
+#include "dec_reshuffle.c"
+#include "dec_loop.c"
+
+#if BASE64_AVX2_USE_ASM
+# include "enc_loop_asm.c"
+#else
+# include "enc_translate.c"
+# include "enc_reshuffle.c"
+# include "enc_loop.c"
+#endif
+
+#endif	// HAVE_AVX2
+
+BASE64_ENC_FUNCTION(avx2)
+{
+#if HAVE_AVX2
+	#include "../generic/enc_head.c"
+	enc_loop_avx2(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(avx2)
+{
+#if HAVE_AVX2
+	#include "../generic/dec_head.c"
+	dec_loop_avx2(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/avx2/dec_loop.c

@@ -0,0 +1,110 @@
+static inline int
+dec_loop_avx2_inner (const uint8_t **s, uint8_t **o, size_t *rounds)
+{
+	const __m256i lut_lo = _mm256_setr_epi8(
+		0x15, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11,
+		0x11, 0x11, 0x13, 0x1A, 0x1B, 0x1B, 0x1B, 0x1A,
+		0x15, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11,
+		0x11, 0x11, 0x13, 0x1A, 0x1B, 0x1B, 0x1B, 0x1A);
+
+	const __m256i lut_hi = _mm256_setr_epi8(
+		0x10, 0x10, 0x01, 0x02, 0x04, 0x08, 0x04, 0x08,
+		0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10,
+		0x10, 0x10, 0x01, 0x02, 0x04, 0x08, 0x04, 0x08,
+		0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10);
+
+	const __m256i lut_roll = _mm256_setr_epi8(
+		0,  16,  19,   4, -65, -65, -71, -71,
+		0,   0,   0,   0,   0,   0,   0,   0,
+		0,  16,  19,   4, -65, -65, -71, -71,
+		0,   0,   0,   0,   0,   0,   0,   0);
+
+	const __m256i mask_2F = _mm256_set1_epi8(0x2F);
+
+	// Load input:
+	__m256i str = _mm256_loadu_si256((__m256i *) *s);
+
+	// See the SSSE3 decoder for an explanation of the algorithm.
+	const __m256i hi_nibbles = _mm256_and_si256(_mm256_srli_epi32(str, 4), mask_2F);
+	const __m256i lo_nibbles = _mm256_and_si256(str, mask_2F);
+	const __m256i hi         = _mm256_shuffle_epi8(lut_hi, hi_nibbles);
+	const __m256i lo         = _mm256_shuffle_epi8(lut_lo, lo_nibbles);
+
+	if (!_mm256_testz_si256(lo, hi)) {
+		return 0;
+	}
+
+	const __m256i eq_2F = _mm256_cmpeq_epi8(str, mask_2F);
+	const __m256i roll  = _mm256_shuffle_epi8(lut_roll, _mm256_add_epi8(eq_2F, hi_nibbles));
+
+	// Now simply add the delta values to the input:
+	str = _mm256_add_epi8(str, roll);
+
+	// Reshuffle the input to packed 12-byte output format:
+	str = dec_reshuffle(str);
+
+	// Store the output:
+	_mm256_storeu_si256((__m256i *) *o, str);
+
+	*s += 32;
+	*o += 24;
+	*rounds -= 1;
+
+	return 1;
+}
+
+static inline void
+dec_loop_avx2 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 45) {
+		return;
+	}
+
+	// Process blocks of 32 bytes per round. Because 8 extra zero bytes are
+	// written after the output, ensure that there will be at least 13
+	// bytes of input data left to cover the gap. (11 data bytes and up to
+	// two end-of-string markers.)
+	size_t rounds = (*slen - 13) / 32;
+
+	*slen -= rounds * 32;	// 32 bytes consumed per round
+	*olen += rounds * 24;	// 24 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			if (dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 4) {
+			if (dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 2) {
+			if (dec_loop_avx2_inner(s, o, &rounds) &&
+			    dec_loop_avx2_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		dec_loop_avx2_inner(s, o, &rounds);
+		break;
+
+	} while (rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 32;
+	*olen -= rounds * 24;
+}

3rdparty/base64/lib/arch/avx2/dec_reshuffle.c

@@ -0,0 +1,34 @@
+static inline __m256i
+dec_reshuffle (const __m256i in)
+{
+	// in, lower lane, bits, upper case are most significant bits, lower
+	// case are least significant bits:
+	// 00llllll 00kkkkLL 00jjKKKK 00JJJJJJ
+	// 00iiiiii 00hhhhII 00ggHHHH 00GGGGGG
+	// 00ffffff 00eeeeFF 00ddEEEE 00DDDDDD
+	// 00cccccc 00bbbbCC 00aaBBBB 00AAAAAA
+
+	const __m256i merge_ab_and_bc = _mm256_maddubs_epi16(in, _mm256_set1_epi32(0x01400140));
+	// 0000kkkk LLllllll 0000JJJJ JJjjKKKK
+	// 0000hhhh IIiiiiii 0000GGGG GGggHHHH
+	// 0000eeee FFffffff 0000DDDD DDddEEEE
+	// 0000bbbb CCcccccc 0000AAAA AAaaBBBB
+
+	__m256i out = _mm256_madd_epi16(merge_ab_and_bc, _mm256_set1_epi32(0x00011000));
+	// 00000000 JJJJJJjj KKKKkkkk LLllllll
+	// 00000000 GGGGGGgg HHHHhhhh IIiiiiii
+	// 00000000 DDDDDDdd EEEEeeee FFffffff
+	// 00000000 AAAAAAaa BBBBbbbb CCcccccc
+
+	// Pack bytes together in each lane:
+	out = _mm256_shuffle_epi8(out, _mm256_setr_epi8(
+		2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1,
+		2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1));
+	// 00000000 00000000 00000000 00000000
+	// LLllllll KKKKkkkk JJJJJJjj IIiiiiii
+	// HHHHhhhh GGGGGGgg FFffffff EEEEeeee
+	// DDDDDDdd CCcccccc BBBBbbbb AAAAAAaa
+
+	// Pack lanes:
+	return _mm256_permutevar8x32_epi32(out, _mm256_setr_epi32(0, 1, 2, 4, 5, 6, -1, -1));
+}

3rdparty/base64/lib/arch/avx2/enc_loop.c

@@ -0,0 +1,89 @@
+static inline void
+enc_loop_avx2_inner_first (const uint8_t **s, uint8_t **o)
+{
+	// First load is done at s - 0 to not get a segfault:
+	__m256i src = _mm256_loadu_si256((__m256i *) *s);
+
+	// Shift by 4 bytes, as required by enc_reshuffle:
+	src = _mm256_permutevar8x32_epi32(src, _mm256_setr_epi32(0, 0, 1, 2, 3, 4, 5, 6));
+
+	// Reshuffle, translate, store:
+	src = enc_reshuffle(src);
+	src = enc_translate(src);
+	_mm256_storeu_si256((__m256i *) *o, src);
+
+	// Subsequent loads will be done at s - 4, set pointer for next round:
+	*s += 20;
+	*o += 32;
+}
+
+static inline void
+enc_loop_avx2_inner (const uint8_t **s, uint8_t **o)
+{
+	// Load input:
+	__m256i src = _mm256_loadu_si256((__m256i *) *s);
+
+	// Reshuffle, translate, store:
+	src = enc_reshuffle(src);
+	src = enc_translate(src);
+	_mm256_storeu_si256((__m256i *) *o, src);
+
+	*s += 24;
+	*o += 32;
+}
+
+static inline void
+enc_loop_avx2 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 32) {
+		return;
+	}
+
+	// Process blocks of 24 bytes at a time. Because blocks are loaded 32
+	// bytes at a time an offset of -4, ensure that there will be at least
+	// 4 remaining bytes after the last round, so that the final read will
+	// not pass beyond the bounds of the input buffer:
+	size_t rounds = (*slen - 4) / 24;
+
+	*slen -= rounds * 24;   // 24 bytes consumed per round
+	*olen += rounds * 32;   // 32 bytes produced per round
+
+	// The first loop iteration requires special handling to ensure that
+	// the read, which is done at an offset, does not underflow the buffer:
+	enc_loop_avx2_inner_first(s, o);
+	rounds--;
+
+	while (rounds > 0) {
+		if (rounds >= 8) {
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_avx2_inner(s, o);
+			enc_loop_avx2_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_avx2_inner(s, o);
+		break;
+	}
+
+	// Add the offset back:
+	*s += 4;
+}

3rdparty/base64/lib/arch/avx2/enc_loop_asm.c

@@ -0,0 +1,291 @@
+// Apologies in advance for combining the preprocessor with inline assembly,
+// two notoriously gnarly parts of C, but it was necessary to avoid a lot of
+// code repetition. The preprocessor is used to template large sections of
+// inline assembly that differ only in the registers used. If the code was
+// written out by hand, it would become very large and hard to audit.
+
+// Generate a block of inline assembly that loads register R0 from memory. The
+// offset at which the register is loaded is set by the given round and a
+// constant offset.
+#define LOAD(R0, ROUND, OFFSET) \
+	"vlddqu ("#ROUND" * 24 + "#OFFSET")(%[src]), %["R0"] \n\t"
+
+// Generate a block of inline assembly that deinterleaves and shuffles register
+// R0 using preloaded constants. Outputs in R0 and R1.
+#define SHUF(R0, R1, R2) \
+	"vpshufb  %[lut0], %["R0"], %["R1"] \n\t" \
+	"vpand    %["R1"], %[msk0], %["R2"] \n\t" \
+	"vpand    %["R1"], %[msk2], %["R1"] \n\t" \
+	"vpmulhuw %["R2"], %[msk1], %["R2"] \n\t" \
+	"vpmullw  %["R1"], %[msk3], %["R1"] \n\t" \
+	"vpor     %["R1"], %["R2"], %["R1"] \n\t"
+
+// Generate a block of inline assembly that takes R0 and R1 and translates
+// their contents to the base64 alphabet, using preloaded constants.
+#define TRAN(R0, R1, R2) \
+	"vpsubusb %[n51],  %["R1"], %["R0"] \n\t" \
+	"vpcmpgtb %[n25],  %["R1"], %["R2"] \n\t" \
+	"vpsubb   %["R2"], %["R0"], %["R0"] \n\t" \
+	"vpshufb  %["R0"], %[lut1], %["R2"] \n\t" \
+	"vpaddb   %["R1"], %["R2"], %["R0"] \n\t"
+
+// Generate a block of inline assembly that stores the given register R0 at an
+// offset set by the given round.
+#define STOR(R0, ROUND) \
+	"vmovdqu %["R0"], ("#ROUND" * 32)(%[dst]) \n\t"
+
+// Generate a block of inline assembly that generates a single self-contained
+// encoder round: fetch the data, process it, and store the result. Then update
+// the source and destination pointers.
+#define ROUND() \
+	LOAD("a", 0, -4) \
+	SHUF("a", "b", "c") \
+	TRAN("a", "b", "c") \
+	STOR("a", 0) \
+	"add $24, %[src] \n\t" \
+	"add $32, %[dst] \n\t"
+
+// Define a macro that initiates a three-way interleaved encoding round by
+// preloading registers a, b and c from memory.
+// The register graph shows which registers are in use during each step, and
+// is a visual aid for choosing registers for that step. Symbol index:
+//
+//  +  indicates that a register is loaded by that step.
+//  |  indicates that a register is in use and must not be touched.
+//  -  indicates that a register is decommissioned by that step.
+//  x  indicates that a register is used as a temporary by that step.
+//  V  indicates that a register is an input or output to the macro.
+//
+#define ROUND_3_INIT() 			/*  a b c d e f  */ \
+	LOAD("a",   0,  -4)		/*  +            */ \
+	SHUF("a", "d", "e")		/*  |     + x    */ \
+	LOAD("b",   1,  -4)		/*  | +   |      */ \
+	TRAN("a", "d", "e")		/*  | |   - x    */ \
+	LOAD("c",   2,  -4)		/*  V V V        */
+
+// Define a macro that translates, shuffles and stores the input registers A, B
+// and C, and preloads registers D, E and F for the next round.
+// This macro can be arbitrarily daisy-chained by feeding output registers D, E
+// and F back into the next round as input registers A, B and C. The macro
+// carefully interleaves memory operations with data operations for optimal
+// pipelined performance.
+
+#define ROUND_3(ROUND, A,B,C,D,E,F) 	/*  A B C D E F  */ \
+	LOAD(D, (ROUND + 3), -4)	/*  V V V +      */ \
+	SHUF(B, E, F)			/*  | | | | + x  */ \
+	STOR(A, (ROUND + 0))		/*  - | | | |    */ \
+	TRAN(B, E, F)			/*    | | | - x  */ \
+	LOAD(E, (ROUND + 4), -4)	/*    | | | +    */ \
+	SHUF(C, A, F)			/*  + | | | | x  */ \
+	STOR(B, (ROUND + 1))		/*  | - | | |    */ \
+	TRAN(C, A, F)			/*  -   | | | x  */ \
+	LOAD(F, (ROUND + 5), -4)	/*      | | | +  */ \
+	SHUF(D, A, B)			/*  + x | | | |  */ \
+	STOR(C, (ROUND + 2))		/*  |   - | | |  */ \
+	TRAN(D, A, B)			/*  - x   V V V  */
+
+// Define a macro that terminates a ROUND_3 macro by taking pre-loaded
+// registers D, E and F, and translating, shuffling and storing them.
+#define ROUND_3_END(ROUND, A,B,C,D,E,F)	/*  A B C D E F  */ \
+	SHUF(E, A, B)			/*  + x   V V V  */ \
+	STOR(D, (ROUND + 3))		/*  |     - | |  */ \
+	TRAN(E, A, B)			/*  - x     | |  */ \
+	SHUF(F, C, D)			/*      + x | |  */ \
+	STOR(E, (ROUND + 4))		/*      |   - |  */ \
+	TRAN(F, C, D)			/*      - x   |  */ \
+	STOR(F, (ROUND + 5))		/*            -  */
+
+// Define a type A round. Inputs are a, b, and c, outputs are d, e, and f.
+#define ROUND_3_A(ROUND) \
+	ROUND_3(ROUND, "a", "b", "c", "d", "e", "f")
+
+// Define a type B round. Inputs and outputs are swapped with regard to type A.
+#define ROUND_3_B(ROUND) \
+	ROUND_3(ROUND, "d", "e", "f", "a", "b", "c")
+
+// Terminating macro for a type A round.
+#define ROUND_3_A_LAST(ROUND) \
+	ROUND_3_A(ROUND) \
+	ROUND_3_END(ROUND, "a", "b", "c", "d", "e", "f")
+
+// Terminating macro for a type B round.
+#define ROUND_3_B_LAST(ROUND) \
+	ROUND_3_B(ROUND) \
+	ROUND_3_END(ROUND, "d", "e", "f", "a", "b", "c")
+
+// Suppress clang's warning that the literal string in the asm statement is
+// overlong (longer than the ISO-mandated minimum size of 4095 bytes for C99
+// compilers). It may be true, but the goal here is not C99 portability.
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+
+static inline void
+enc_loop_avx2 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	// For a clearer explanation of the algorithm used by this function,
+	// please refer to the plain (not inline assembly) implementation. This
+	// function follows the same basic logic.
+
+	if (*slen < 32) {
+		return;
+	}
+
+	// Process blocks of 24 bytes at a time. Because blocks are loaded 32
+	// bytes at a time an offset of -4, ensure that there will be at least
+	// 4 remaining bytes after the last round, so that the final read will
+	// not pass beyond the bounds of the input buffer.
+	size_t rounds = (*slen - 4) / 24;
+
+	*slen -= rounds * 24;   // 24 bytes consumed per round
+	*olen += rounds * 32;   // 32 bytes produced per round
+
+	// Pre-decrement the number of rounds to get the number of rounds
+	// *after* the first round, which is handled as a special case.
+	rounds--;
+
+	// Number of times to go through the 36x loop.
+	size_t loops = rounds / 36;
+
+	// Number of rounds remaining after the 36x loop.
+	rounds %= 36;
+
+	// Lookup tables.
+	const __m256i lut0 = _mm256_set_epi8(
+		10, 11,  9, 10,  7,  8,  6,  7,  4,  5,  3,  4,  1,  2,  0,  1,
+		14, 15, 13, 14, 11, 12, 10, 11,  8,  9,  7,  8,  5,  6,  4,  5);
+
+	const __m256i lut1 = _mm256_setr_epi8(
+		65, 71, -4, -4, -4, -4, -4, -4, -4, -4, -4, -4, -19, -16, 0, 0,
+		65, 71, -4, -4, -4, -4, -4, -4, -4, -4, -4, -4, -19, -16, 0, 0);
+
+	// Temporary registers.
+	__m256i a, b, c, d, e;
+
+	// Temporary register f doubles as the shift mask for the first round.
+	__m256i f = _mm256_setr_epi32(0, 0, 1, 2, 3, 4, 5, 6);
+
+	__asm__ volatile (
+
+		// The first loop iteration requires special handling to ensure
+		// that the read, which is normally done at an offset of -4,
+		// does not underflow the buffer. Load the buffer at an offset
+		// of 0 and permute the input to achieve the same effect.
+		LOAD("a", 0, 0)
+		"vpermd %[a], %[f], %[a] \n\t"
+
+		// Perform the standard shuffling and translation steps.
+		SHUF("a", "b", "c")
+		TRAN("a", "b", "c")
+
+		// Store the result and increment the source and dest pointers.
+		"vmovdqu %[a], (%[dst]) \n\t"
+		"add     $24,  %[src]   \n\t"
+		"add     $32,  %[dst]   \n\t"
+
+		// If there are 36 rounds or more, enter a 36x unrolled loop of
+		// interleaved encoding rounds. The rounds interleave memory
+		// operations (load/store) with data operations (table lookups,
+		// etc) to maximize pipeline throughput.
+		"    test %[loops], %[loops] \n\t"
+		"    jz   18f                \n\t"
+		"    jmp  36f                \n\t"
+		"                            \n\t"
+		".balign 64                  \n\t"
+		"36: " ROUND_3_INIT()
+		"    " ROUND_3_A( 0)
+		"    " ROUND_3_B( 3)
+		"    " ROUND_3_A( 6)
+		"    " ROUND_3_B( 9)
+		"    " ROUND_3_A(12)
+		"    " ROUND_3_B(15)
+		"    " ROUND_3_A(18)
+		"    " ROUND_3_B(21)
+		"    " ROUND_3_A(24)
+		"    " ROUND_3_B(27)
+		"    " ROUND_3_A_LAST(30)
+		"    add $(24 * 36), %[src] \n\t"
+		"    add $(32 * 36), %[dst] \n\t"
+		"    dec %[loops]           \n\t"
+		"    jnz 36b                \n\t"
+
+		// Enter an 18x unrolled loop for rounds of 18 or more.
+		"18: cmp $18, %[rounds] \n\t"
+		"    jl  9f             \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A(0)
+		"    " ROUND_3_B(3)
+		"    " ROUND_3_A(6)
+		"    " ROUND_3_B(9)
+		"    " ROUND_3_A_LAST(12)
+		"    sub $18,        %[rounds] \n\t"
+		"    add $(24 * 18), %[src]    \n\t"
+		"    add $(32 * 18), %[dst]    \n\t"
+
+		// Enter a 9x unrolled loop for rounds of 9 or more.
+		"9:  cmp $9, %[rounds] \n\t"
+		"    jl  6f            \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A(0)
+		"    " ROUND_3_B_LAST(3)
+		"    sub $9,        %[rounds] \n\t"
+		"    add $(24 * 9), %[src]    \n\t"
+		"    add $(32 * 9), %[dst]    \n\t"
+
+		// Enter a 6x unrolled loop for rounds of 6 or more.
+		"6:  cmp $6, %[rounds] \n\t"
+		"    jl  55f           \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A_LAST(0)
+		"    sub $6,        %[rounds] \n\t"
+		"    add $(24 * 6), %[src]    \n\t"
+		"    add $(32 * 6), %[dst]    \n\t"
+
+		// Dispatch the remaining rounds 0..5.
+		"55: cmp $3, %[rounds] \n\t"
+		"    jg  45f           \n\t"
+		"    je  3f            \n\t"
+		"    cmp $1, %[rounds] \n\t"
+		"    jg  2f            \n\t"
+		"    je  1f            \n\t"
+		"    jmp 0f            \n\t"
+
+		"45: cmp $4, %[rounds] \n\t"
+		"    je  4f            \n\t"
+
+		// Block of non-interlaced encoding rounds, which can each
+		// individually be jumped to. Rounds fall through to the next.
+		"5: " ROUND()
+		"4: " ROUND()
+		"3: " ROUND()
+		"2: " ROUND()
+		"1: " ROUND()
+		"0: \n\t"
+
+		// Outputs (modified).
+		: [rounds] "+r"  (rounds),
+		  [loops]  "+r"  (loops),
+		  [src]    "+r"  (*s),
+		  [dst]    "+r"  (*o),
+		  [a]      "=&x" (a),
+		  [b]      "=&x" (b),
+		  [c]      "=&x" (c),
+		  [d]      "=&x" (d),
+		  [e]      "=&x" (e),
+		  [f]      "+x"  (f)
+
+		// Inputs (not modified).
+		: [lut0] "x" (lut0),
+		  [lut1] "x" (lut1),
+		  [msk0] "x" (_mm256_set1_epi32(0x0FC0FC00)),
+		  [msk1] "x" (_mm256_set1_epi32(0x04000040)),
+		  [msk2] "x" (_mm256_set1_epi32(0x003F03F0)),
+		  [msk3] "x" (_mm256_set1_epi32(0x01000010)),
+		  [n51]  "x" (_mm256_set1_epi8(51)),
+		  [n25]  "x" (_mm256_set1_epi8(25))
+
+		// Clobbers.
+		: "cc", "memory"
+	);
+}
+
+#pragma GCC diagnostic pop

3rdparty/base64/lib/arch/avx2/enc_reshuffle.c

@@ -0,0 +1,83 @@
+static inline __m256i
+enc_reshuffle (const __m256i input)
+{
+	// Translation of the SSSE3 reshuffling algorithm to AVX2. This one
+	// works with shifted (4 bytes) input in order to be able to work
+	// efficiently in the two 128-bit lanes.
+
+	// Input, bytes MSB to LSB:
+	// 0 0 0 0 x w v u t s r q p o n m
+	// l k j i h g f e d c b a 0 0 0 0
+
+	const __m256i in = _mm256_shuffle_epi8(input, _mm256_set_epi8(
+		10, 11,  9, 10,
+		 7,  8,  6,  7,
+		 4,  5,  3,  4,
+		 1,  2,  0,  1,
+
+		14, 15, 13, 14,
+		11, 12, 10, 11,
+		 8,  9,  7,  8,
+		 5,  6,  4,  5));
+	// in, bytes MSB to LSB:
+	// w x v w
+	// t u s t
+	// q r p q
+	// n o m n
+	// k l j k
+	// h i g h
+	// e f d e
+	// b c a b
+
+	const __m256i t0 = _mm256_and_si256(in, _mm256_set1_epi32(0x0FC0FC00));
+	// bits, upper case are most significant bits, lower case are least
+	// significant bits.
+	// 0000wwww XX000000 VVVVVV00 00000000
+	// 0000tttt UU000000 SSSSSS00 00000000
+	// 0000qqqq RR000000 PPPPPP00 00000000
+	// 0000nnnn OO000000 MMMMMM00 00000000
+	// 0000kkkk LL000000 JJJJJJ00 00000000
+	// 0000hhhh II000000 GGGGGG00 00000000
+	// 0000eeee FF000000 DDDDDD00 00000000
+	// 0000bbbb CC000000 AAAAAA00 00000000
+
+	const __m256i t1 = _mm256_mulhi_epu16(t0, _mm256_set1_epi32(0x04000040));
+	// 00000000 00wwwwXX 00000000 00VVVVVV
+	// 00000000 00ttttUU 00000000 00SSSSSS
+	// 00000000 00qqqqRR 00000000 00PPPPPP
+	// 00000000 00nnnnOO 00000000 00MMMMMM
+	// 00000000 00kkkkLL 00000000 00JJJJJJ
+	// 00000000 00hhhhII 00000000 00GGGGGG
+	// 00000000 00eeeeFF 00000000 00DDDDDD
+	// 00000000 00bbbbCC 00000000 00AAAAAA
+
+	const __m256i t2 = _mm256_and_si256(in, _mm256_set1_epi32(0x003F03F0));
+	// 00000000 00xxxxxx 000000vv WWWW0000
+	// 00000000 00uuuuuu 000000ss TTTT0000
+	// 00000000 00rrrrrr 000000pp QQQQ0000
+	// 00000000 00oooooo 000000mm NNNN0000
+	// 00000000 00llllll 000000jj KKKK0000
+	// 00000000 00iiiiii 000000gg HHHH0000
+	// 00000000 00ffffff 000000dd EEEE0000
+	// 00000000 00cccccc 000000aa BBBB0000
+
+	const __m256i t3 = _mm256_mullo_epi16(t2, _mm256_set1_epi32(0x01000010));
+	// 00xxxxxx 00000000 00vvWWWW 00000000
+	// 00uuuuuu 00000000 00ssTTTT 00000000
+	// 00rrrrrr 00000000 00ppQQQQ 00000000
+	// 00oooooo 00000000 00mmNNNN 00000000
+	// 00llllll 00000000 00jjKKKK 00000000
+	// 00iiiiii 00000000 00ggHHHH 00000000
+	// 00ffffff 00000000 00ddEEEE 00000000
+	// 00cccccc 00000000 00aaBBBB 00000000
+
+	return _mm256_or_si256(t1, t3);
+	// 00xxxxxx 00wwwwXX 00vvWWWW 00VVVVVV
+	// 00uuuuuu 00ttttUU 00ssTTTT 00SSSSSS
+	// 00rrrrrr 00qqqqRR 00ppQQQQ 00PPPPPP
+	// 00oooooo 00nnnnOO 00mmNNNN 00MMMMMM
+	// 00llllll 00kkkkLL 00jjKKKK 00JJJJJJ
+	// 00iiiiii 00hhhhII 00ggHHHH 00GGGGGG
+	// 00ffffff 00eeeeFF 00ddEEEE 00DDDDDD
+	// 00cccccc 00bbbbCC 00aaBBBB 00AAAAAA
+}

3rdparty/base64/lib/arch/avx2/enc_translate.c

@@ -0,0 +1,30 @@
+static inline __m256i
+enc_translate (const __m256i in)
+{
+	// A lookup table containing the absolute offsets for all ranges:
+	const __m256i lut = _mm256_setr_epi8(
+		65, 71, -4, -4, -4, -4, -4, -4, -4, -4, -4, -4, -19, -16, 0, 0,
+		65, 71, -4, -4, -4, -4, -4, -4, -4, -4, -4, -4, -19, -16, 0, 0);
+
+	// Translate values 0..63 to the Base64 alphabet. There are five sets:
+	// #  From      To         Abs    Index  Characters
+	// 0  [0..25]   [65..90]   +65        0  ABCDEFGHIJKLMNOPQRSTUVWXYZ
+	// 1  [26..51]  [97..122]  +71        1  abcdefghijklmnopqrstuvwxyz
+	// 2  [52..61]  [48..57]    -4  [2..11]  0123456789
+	// 3  [62]      [43]       -19       12  +
+	// 4  [63]      [47]       -16       13  /
+
+	// Create LUT indices from the input. The index for range #0 is right,
+	// others are 1 less than expected:
+	__m256i indices = _mm256_subs_epu8(in, _mm256_set1_epi8(51));
+
+	// mask is 0xFF (-1) for range #[1..4] and 0x00 for range #0:
+	const __m256i mask = _mm256_cmpgt_epi8(in, _mm256_set1_epi8(25));
+
+	// Subtract -1, so add 1 to indices for range #[1..4]. All indices are
+	// now correct:
+	indices = _mm256_sub_epi8(indices, mask);
+
+	// Add offsets to input values:
+	return _mm256_add_epi8(in, _mm256_shuffle_epi8(lut, indices));
+}

3rdparty/base64/lib/arch/avx512/codec.c

@@ -0,0 +1,42 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdlib.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if HAVE_AVX512
+#include <immintrin.h>
+
+#include "../avx2/dec_reshuffle.c"
+#include "../avx2/dec_loop.c"
+#include "enc_reshuffle_translate.c"
+#include "enc_loop.c"
+
+#endif	// HAVE_AVX512
+
+BASE64_ENC_FUNCTION(avx512)
+{
+#if HAVE_AVX512
+	#include "../generic/enc_head.c"
+	enc_loop_avx512(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+// Reuse AVX2 decoding. Not supporting AVX512 at present
+BASE64_DEC_FUNCTION(avx512)
+{
+#if HAVE_AVX512
+	#include "../generic/dec_head.c"
+	dec_loop_avx2(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/avx512/enc_loop.c

@@ -0,0 +1,61 @@
+static inline void
+enc_loop_avx512_inner (const uint8_t **s, uint8_t **o)
+{
+	// Load input.
+	__m512i src = _mm512_loadu_si512((__m512i *) *s);
+
+	// Reshuffle, translate, store.
+	src = enc_reshuffle_translate(src);
+	_mm512_storeu_si512((__m512i *) *o, src);
+
+	*s += 48;
+	*o += 64;
+}
+
+static inline void
+enc_loop_avx512 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 64) {
+		return;
+	}
+
+	// Process blocks of 48 bytes at a time. Because blocks are loaded 64
+	// bytes at a time, ensure that there will be at least 24 remaining
+	// bytes after the last round, so that the final read will not pass
+	// beyond the bounds of the input buffer.
+	size_t rounds = (*slen - 24) / 48;
+
+	*slen -= rounds * 48;   // 48 bytes consumed per round
+	*olen += rounds * 64;   // 64 bytes produced per round
+
+	while (rounds > 0) {
+		if (rounds >= 8) {
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_avx512_inner(s, o);
+			enc_loop_avx512_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_avx512_inner(s, o);
+		break;
+	}
+}

3rdparty/base64/lib/arch/avx512/enc_reshuffle_translate.c

@@ -0,0 +1,50 @@
+// AVX512 algorithm is based on permutevar and multishift. The code is based on
+// https://github.com/WojciechMula/base64simd which is under BSD-2 license.
+
+static inline __m512i
+enc_reshuffle_translate (const __m512i input)
+{
+	// 32-bit input
+	// [ 0  0  0  0  0  0  0  0|c1 c0 d5 d4 d3 d2 d1 d0|
+	//  b3 b2 b1 b0 c5 c4 c3 c2|a5 a4 a3 a2 a1 a0 b5 b4]
+	// output order  [1, 2, 0, 1]
+	// [b3 b2 b1 b0 c5 c4 c3 c2|c1 c0 d5 d4 d3 d2 d1 d0|
+	//  a5 a4 a3 a2 a1 a0 b5 b4|b3 b2 b1 b0 c3 c2 c1 c0]
+
+	const __m512i shuffle_input = _mm512_setr_epi32(0x01020001,
+	                                                0x04050304,
+	                                                0x07080607,
+	                                                0x0a0b090a,
+	                                                0x0d0e0c0d,
+	                                                0x10110f10,
+	                                                0x13141213,
+	                                                0x16171516,
+	                                                0x191a1819,
+	                                                0x1c1d1b1c,
+	                                                0x1f201e1f,
+	                                                0x22232122,
+	                                                0x25262425,
+	                                                0x28292728,
+	                                                0x2b2c2a2b,
+	                                                0x2e2f2d2e);
+
+	// Reorder bytes
+	// [b3 b2 b1 b0 c5 c4 c3 c2|c1 c0 d5 d4 d3 d2 d1 d0|
+	//  a5 a4 a3 a2 a1 a0 b5 b4|b3 b2 b1 b0 c3 c2 c1 c0]
+	const __m512i in = _mm512_permutexvar_epi8(shuffle_input, input);
+
+	// After multishift a single 32-bit lane has following layout
+	// [c1 c0 d5 d4 d3 d2 d1 d0|b1 b0 c5 c4 c3 c2 c1 c0|
+	//  a1 a0 b5 b4 b3 b2 b1 b0|d1 d0 a5 a4 a3 a2 a1 a0]
+	// (a = [10:17], b = [4:11], c = [22:27], d = [16:21])
+
+	// 48, 54, 36, 42, 16, 22, 4, 10
+	const __m512i shifts = _mm512_set1_epi64(0x3036242a1016040alu);
+	__m512i shuffled_in = _mm512_multishift_epi64_epi8(shifts, in);
+
+	// Translate immediatedly after reshuffled.
+	const __m512i lookup = _mm512_loadu_si512(base64_table_enc_6bit);
+
+	// Translation 6-bit values to ASCII.
+	return _mm512_permutexvar_epi8(shuffled_in, lookup);
+}

3rdparty/base64/lib/arch/generic/32/dec_loop.c

@@ -0,0 +1,86 @@
+static inline int
+dec_loop_generic_32_inner (const uint8_t **s, uint8_t **o, size_t *rounds)
+{
+	const uint32_t str
+		= base64_table_dec_32bit_d0[(*s)[0]]
+		| base64_table_dec_32bit_d1[(*s)[1]]
+		| base64_table_dec_32bit_d2[(*s)[2]]
+		| base64_table_dec_32bit_d3[(*s)[3]];
+
+#if BASE64_LITTLE_ENDIAN
+
+	// LUTs for little-endian set MSB in case of invalid character:
+	if (str & UINT32_C(0x80000000)) {
+		return 0;
+	}
+#else
+	// LUTs for big-endian set LSB in case of invalid character:
+	if (str & UINT32_C(1)) {
+		return 0;
+	}
+#endif
+	// Store the output:
+	memcpy(*o, &str, sizeof (str));
+
+	*s += 4;
+	*o += 3;
+	*rounds -= 1;
+
+	return 1;
+}
+
+static inline void
+dec_loop_generic_32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 8) {
+		return;
+	}
+
+	// Process blocks of 4 bytes per round. Because one extra zero byte is
+	// written after the output, ensure that there will be at least 4 bytes
+	// of input data left to cover the gap. (Two data bytes and up to two
+	// end-of-string markers.)
+	size_t rounds = (*slen - 4) / 4;
+
+	*slen -= rounds * 4;	// 4 bytes consumed per round
+	*olen += rounds * 3;	// 3 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			if (dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 4) {
+			if (dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 2) {
+			if (dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		dec_loop_generic_32_inner(s, o, &rounds);
+		break;
+
+	} while (rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 4;
+	*olen -= rounds * 3;
+}

3rdparty/base64/lib/arch/generic/32/enc_loop.c

@@ -0,0 +1,73 @@
+static inline void
+enc_loop_generic_32_inner (const uint8_t **s, uint8_t **o)
+{
+	uint32_t src;
+
+	// Load input:
+	memcpy(&src, *s, sizeof (src));
+
+	// Reorder to 32-bit big-endian, if not already in that format. The
+	// workset must be in big-endian, otherwise the shifted bits do not
+	// carry over properly among adjacent bytes:
+	src = BASE64_HTOBE32(src);
+
+	// Two indices for the 12-bit lookup table:
+	const size_t index0 = (src >> 20) & 0xFFFU;
+	const size_t index1 = (src >>  8) & 0xFFFU;
+
+	// Table lookup and store:
+	memcpy(*o + 0, base64_table_enc_12bit + index0, 2);
+	memcpy(*o + 2, base64_table_enc_12bit + index1, 2);
+
+	*s += 3;
+	*o += 4;
+}
+
+static inline void
+enc_loop_generic_32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 4) {
+		return;
+	}
+
+	// Process blocks of 3 bytes at a time. Because blocks are loaded 4
+	// bytes at a time, ensure that there will be at least one remaining
+	// byte after the last round, so that the final read will not pass
+	// beyond the bounds of the input buffer:
+	size_t rounds = (*slen - 1) / 3;
+
+	*slen -= rounds * 3;	// 3 bytes consumed per round
+	*olen += rounds * 4;	// 4 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_generic_32_inner(s, o);
+		break;
+
+	} while (rounds > 0);
+}

3rdparty/base64/lib/arch/generic/64/enc_loop.c

@@ -0,0 +1,77 @@
+static inline void
+enc_loop_generic_64_inner (const uint8_t **s, uint8_t **o)
+{
+	uint64_t src;
+
+	// Load input:
+	memcpy(&src, *s, sizeof (src));
+
+	// Reorder to 64-bit big-endian, if not already in that format. The
+	// workset must be in big-endian, otherwise the shifted bits do not
+	// carry over properly among adjacent bytes:
+	src = BASE64_HTOBE64(src);
+
+	// Four indices for the 12-bit lookup table:
+	const size_t index0 = (src >> 52) & 0xFFFU;
+	const size_t index1 = (src >> 40) & 0xFFFU;
+	const size_t index2 = (src >> 28) & 0xFFFU;
+	const size_t index3 = (src >> 16) & 0xFFFU;
+
+	// Table lookup and store:
+	memcpy(*o + 0, base64_table_enc_12bit + index0, 2);
+	memcpy(*o + 2, base64_table_enc_12bit + index1, 2);
+	memcpy(*o + 4, base64_table_enc_12bit + index2, 2);
+	memcpy(*o + 6, base64_table_enc_12bit + index3, 2);
+
+	*s += 6;
+	*o += 8;
+}
+
+static inline void
+enc_loop_generic_64 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 8) {
+		return;
+	}
+
+	// Process blocks of 6 bytes at a time. Because blocks are loaded 8
+	// bytes at a time, ensure that there will be at least 2 remaining
+	// bytes after the last round, so that the final read will not pass
+	// beyond the bounds of the input buffer:
+	size_t rounds = (*slen - 2) / 6;
+
+	*slen -= rounds * 6;	// 6 bytes consumed per round
+	*olen += rounds * 8;	// 8 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_generic_64_inner(s, o);
+		break;
+
+	} while (rounds > 0);
+}

3rdparty/base64/lib/arch/generic/codec.c

@@ -0,0 +1,39 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <string.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if BASE64_WORDSIZE == 32
+#  include "32/enc_loop.c"
+#elif BASE64_WORDSIZE == 64
+#  include "64/enc_loop.c"
+#endif
+
+#if BASE64_WORDSIZE >= 32
+#  include "32/dec_loop.c"
+#endif
+
+BASE64_ENC_FUNCTION(plain)
+{
+	#include "enc_head.c"
+#if BASE64_WORDSIZE == 32
+	enc_loop_generic_32(&s, &slen, &o, &olen);
+#elif BASE64_WORDSIZE == 64
+	enc_loop_generic_64(&s, &slen, &o, &olen);
+#endif
+	#include "enc_tail.c"
+}
+
+BASE64_DEC_FUNCTION(plain)
+{
+	#include "dec_head.c"
+#if BASE64_WORDSIZE >= 32
+	dec_loop_generic_32(&s, &slen, &o, &olen);
+#endif
+	#include "dec_tail.c"
+}

3rdparty/base64/lib/arch/generic/dec_head.c

@@ -0,0 +1,37 @@
+int ret = 0;
+const uint8_t *s = (const uint8_t *) src;
+uint8_t *o = (uint8_t *) out;
+uint8_t q;
+
+// Use local temporaries to avoid cache thrashing:
+size_t olen = 0;
+size_t slen = srclen;
+struct base64_state st;
+st.eof = state->eof;
+st.bytes = state->bytes;
+st.carry = state->carry;
+
+// If we previously saw an EOF or an invalid character, bail out:
+if (st.eof) {
+	*outlen = 0;
+	ret = 0;
+	// If there was a trailing '=' to check, check it:
+	if (slen && (st.eof == BASE64_AEOF)) {
+		state->bytes = 0;
+		state->eof = BASE64_EOF;
+		ret = ((base64_table_dec_8bit[*s++] == 254) && (slen == 1)) ? 1 : 0;
+	}
+	return ret;
+}
+
+// Turn four 6-bit numbers into three bytes:
+// out[0] = 11111122
+// out[1] = 22223333
+// out[2] = 33444444
+
+// Duff's device again:
+switch (st.bytes)
+{
+	for (;;)
+	{
+	case 0:

3rdparty/base64/lib/arch/generic/dec_tail.c

@@ -0,0 +1,91 @@
+		if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.eof = BASE64_EOF;
+			// Treat character '=' as invalid for byte 0:
+			break;
+		}
+		st.carry = q << 2;
+		st.bytes++;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 1:	if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.eof = BASE64_EOF;
+			// Treat character '=' as invalid for byte 1:
+			break;
+		}
+		*o++ = st.carry | (q >> 4);
+		st.carry = q << 4;
+		st.bytes++;
+		olen++;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 2:	if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.bytes++;
+			// When q == 254, the input char is '='.
+			// Check if next byte is also '=':
+			if (q == 254) {
+				if (slen-- != 0) {
+					st.bytes = 0;
+					// EOF:
+					st.eof = BASE64_EOF;
+					q = base64_table_dec_8bit[*s++];
+					ret = ((q == 254) && (slen == 0)) ? 1 : 0;
+					break;
+				}
+				else {
+					// Almost EOF
+					st.eof = BASE64_AEOF;
+					ret = 1;
+					break;
+				}
+			}
+			// If we get here, there was an error:
+			break;
+		}
+		*o++ = st.carry | (q >> 2);
+		st.carry = q << 6;
+		st.bytes++;
+		olen++;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 3:	if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.bytes = 0;
+			st.eof = BASE64_EOF;
+			// When q == 254, the input char is '='. Return 1 and EOF.
+			// When q == 255, the input char is invalid. Return 0 and EOF.
+			ret = ((q == 254) && (slen == 0)) ? 1 : 0;
+			break;
+		}
+		*o++ = st.carry | q;
+		st.carry = 0;
+		st.bytes = 0;
+		olen++;
+	}
+}
+
+state->eof = st.eof;
+state->bytes = st.bytes;
+state->carry = st.carry;
+*outlen = olen;
+return ret;

3rdparty/base64/lib/arch/generic/enc_head.c

@@ -0,0 +1,24 @@
+// Assume that *out is large enough to contain the output.
+// Theoretically it should be 4/3 the length of src.
+const uint8_t *s = (const uint8_t *) src;
+uint8_t *o = (uint8_t *) out;
+
+// Use local temporaries to avoid cache thrashing:
+size_t olen = 0;
+size_t slen = srclen;
+struct base64_state st;
+st.bytes = state->bytes;
+st.carry = state->carry;
+
+// Turn three bytes into four 6-bit numbers:
+// in[0] = 00111111
+// in[1] = 00112222
+// in[2] = 00222233
+// in[3] = 00333333
+
+// Duff's device, a for() loop inside a switch() statement. Legal!
+switch (st.bytes)
+{
+	for (;;)
+	{
+	case 0:

3rdparty/base64/lib/arch/generic/enc_tail.c

@@ -0,0 +1,34 @@
+		if (slen-- == 0) {
+			break;
+		}
+		*o++ = base64_table_enc_6bit[*s >> 2];
+		st.carry = (*s++ << 4) & 0x30;
+		st.bytes++;
+		olen += 1;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 1:	if (slen-- == 0) {
+			break;
+		}
+		*o++ = base64_table_enc_6bit[st.carry | (*s >> 4)];
+		st.carry = (*s++ << 2) & 0x3C;
+		st.bytes++;
+		olen += 1;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 2:	if (slen-- == 0) {
+			break;
+		}
+		*o++ = base64_table_enc_6bit[st.carry | (*s >> 6)];
+		*o++ = base64_table_enc_6bit[*s++ & 0x3F];
+		st.bytes = 0;
+		olen += 2;
+	}
+}
+state->bytes = st.bytes;
+state->carry = st.carry;
+*outlen = olen;

3rdparty/base64/lib/arch/neon32/codec.c

@@ -0,0 +1,77 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <string.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#ifdef __arm__
+#  if (defined(__ARM_NEON__) || defined(__ARM_NEON)) && HAVE_NEON32
+#    define BASE64_USE_NEON32
+#  endif
+#endif
+
+#ifdef BASE64_USE_NEON32
+#include <arm_neon.h>
+
+// Only enable inline assembly on supported compilers.
+#if defined(__GNUC__) || defined(__clang__)
+#define BASE64_NEON32_USE_ASM
+#endif
+
+static inline uint8x16_t
+vqtbl1q_u8 (const uint8x16_t lut, const uint8x16_t indices)
+{
+	// NEON32 only supports 64-bit wide lookups in 128-bit tables. Emulate
+	// the NEON64 `vqtbl1q_u8` intrinsic to do 128-bit wide lookups.
+	uint8x8x2_t lut2;
+	uint8x8x2_t result;
+
+	lut2.val[0] = vget_low_u8(lut);
+	lut2.val[1] = vget_high_u8(lut);
+
+	result.val[0] = vtbl2_u8(lut2, vget_low_u8(indices));
+	result.val[1] = vtbl2_u8(lut2, vget_high_u8(indices));
+
+	return vcombine_u8(result.val[0], result.val[1]);
+}
+
+#include "../generic/32/dec_loop.c"
+#include "../generic/32/enc_loop.c"
+#include "dec_loop.c"
+#include "enc_reshuffle.c"
+#include "enc_translate.c"
+#include "enc_loop.c"
+
+#endif	// BASE64_USE_NEON32
+
+// Stride size is so large on these NEON 32-bit functions
+// (48 bytes encode, 32 bytes decode) that we inline the
+// uint32 codec to stay performant on smaller inputs.
+
+BASE64_ENC_FUNCTION(neon32)
+{
+#ifdef BASE64_USE_NEON32
+	#include "../generic/enc_head.c"
+	enc_loop_neon32(&s, &slen, &o, &olen);
+	enc_loop_generic_32(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(neon32)
+{
+#ifdef BASE64_USE_NEON32
+	#include "../generic/dec_head.c"
+	dec_loop_neon32(&s, &slen, &o, &olen);
+	dec_loop_generic_32(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/neon32/dec_loop.c

@@ -0,0 +1,106 @@
+static inline int
+is_nonzero (const uint8x16_t v)
+{
+	uint64_t u64;
+	const uint64x2_t v64 = vreinterpretq_u64_u8(v);
+	const uint32x2_t v32 = vqmovn_u64(v64);
+
+	vst1_u64(&u64, vreinterpret_u64_u32(v32));
+	return u64 != 0;
+}
+
+static inline uint8x16_t
+delta_lookup (const uint8x16_t v)
+{
+	const uint8x8_t lut = {
+		0, 16, 19, 4, (uint8_t) -65, (uint8_t) -65, (uint8_t) -71, (uint8_t) -71,
+	};
+
+	return vcombine_u8(
+		vtbl1_u8(lut, vget_low_u8(v)),
+		vtbl1_u8(lut, vget_high_u8(v)));
+}
+
+static inline uint8x16_t
+dec_loop_neon32_lane (uint8x16_t *lane)
+{
+	// See the SSSE3 decoder for an explanation of the algorithm.
+	const uint8x16_t lut_lo = {
+		0x15, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11,
+		0x11, 0x11, 0x13, 0x1A, 0x1B, 0x1B, 0x1B, 0x1A
+	};
+
+	const uint8x16_t lut_hi = {
+		0x10, 0x10, 0x01, 0x02, 0x04, 0x08, 0x04, 0x08,
+		0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10
+	};
+
+	const uint8x16_t mask_0F = vdupq_n_u8(0x0F);
+	const uint8x16_t mask_2F = vdupq_n_u8(0x2F);
+
+	const uint8x16_t hi_nibbles = vshrq_n_u8(*lane, 4);
+	const uint8x16_t lo_nibbles = vandq_u8(*lane, mask_0F);
+	const uint8x16_t eq_2F      = vceqq_u8(*lane, mask_2F);
+
+	const uint8x16_t hi = vqtbl1q_u8(lut_hi, hi_nibbles);
+	const uint8x16_t lo = vqtbl1q_u8(lut_lo, lo_nibbles);
+
+	// Now simply add the delta values to the input:
+	*lane = vaddq_u8(*lane, delta_lookup(vaddq_u8(eq_2F, hi_nibbles)));
+
+	// Return the validity mask:
+	return vandq_u8(lo, hi);
+}
+
+static inline void
+dec_loop_neon32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 64) {
+		return;
+	}
+
+	// Process blocks of 64 bytes per round. Unlike the SSE codecs, no
+	// extra trailing zero bytes are written, so it is not necessary to
+	// reserve extra input bytes:
+	size_t rounds = *slen / 64;
+
+	*slen -= rounds * 64;	// 64 bytes consumed per round
+	*olen += rounds * 48;	// 48 bytes produced per round
+
+	do {
+		uint8x16x3_t dec;
+
+		// Load 64 bytes and deinterleave:
+		uint8x16x4_t str = vld4q_u8(*s);
+
+		// Decode each lane, collect a mask of invalid inputs:
+		const uint8x16_t classified
+			= dec_loop_neon32_lane(&str.val[0])
+			| dec_loop_neon32_lane(&str.val[1])
+			| dec_loop_neon32_lane(&str.val[2])
+			| dec_loop_neon32_lane(&str.val[3]);
+
+		// Check for invalid input: if any of the delta values are
+		// zero, fall back on bytewise code to do error checking and
+		// reporting:
+		if (is_nonzero(classified)) {
+			break;
+		}
+
+		// Compress four bytes into three:
+		dec.val[0] = vorrq_u8(vshlq_n_u8(str.val[0], 2), vshrq_n_u8(str.val[1], 4));
+		dec.val[1] = vorrq_u8(vshlq_n_u8(str.val[1], 4), vshrq_n_u8(str.val[2], 2));
+		dec.val[2] = vorrq_u8(vshlq_n_u8(str.val[2], 6), str.val[3]);
+
+		// Interleave and store decoded result:
+		vst3q_u8(*o, dec);
+
+		*s += 64;
+		*o += 48;
+
+	} while (--rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 64;
+	*olen -= rounds * 48;
+}

3rdparty/base64/lib/arch/neon32/enc_loop.c

@@ -0,0 +1,170 @@
+#ifdef BASE64_NEON32_USE_ASM
+static inline void
+enc_loop_neon32_inner_asm (const uint8_t **s, uint8_t **o)
+{
+	// This function duplicates the functionality of enc_loop_neon32_inner,
+	// but entirely with inline assembly. This gives a significant speedup
+	// over using NEON intrinsics, which do not always generate very good
+	// code. The logic of the assembly is directly lifted from the
+	// intrinsics version, so it can be used as a guide to this code.
+
+	// Temporary registers, used as scratch space.
+	uint8x16_t tmp0, tmp1, tmp2, tmp3;
+	uint8x16_t mask0, mask1, mask2, mask3;
+
+	// A lookup table containing the absolute offsets for all ranges.
+	const uint8x16_t lut = {
+		  65U,  71U, 252U, 252U,
+		 252U, 252U, 252U, 252U,
+		 252U, 252U, 252U, 252U,
+		 237U, 240U,   0U,   0U
+	};
+
+	// Numeric constants.
+	const uint8x16_t n51 = vdupq_n_u8(51);
+	const uint8x16_t n25 = vdupq_n_u8(25);
+	const uint8x16_t n63 = vdupq_n_u8(63);
+
+	__asm__ (
+
+		// Load 48 bytes and deinterleave. The bytes are loaded to
+		// hard-coded registers q12, q13 and q14, to ensure that they
+		// are contiguous. Increment the source pointer.
+		"vld3.8 {d24, d26, d28}, [%[src]]! \n\t"
+		"vld3.8 {d25, d27, d29}, [%[src]]! \n\t"
+
+		// Reshuffle the bytes using temporaries.
+		"vshr.u8 %q[t0], q12,    #2      \n\t"
+		"vshr.u8 %q[t1], q13,    #4      \n\t"
+		"vshr.u8 %q[t2], q14,    #6      \n\t"
+		"vsli.8  %q[t1], q12,    #4      \n\t"
+		"vsli.8  %q[t2], q13,    #2      \n\t"
+		"vand.u8 %q[t1], %q[t1], %q[n63] \n\t"
+		"vand.u8 %q[t2], %q[t2], %q[n63] \n\t"
+		"vand.u8 %q[t3], q14,    %q[n63] \n\t"
+
+		// t0..t3 are the reshuffled inputs. Create LUT indices.
+		"vqsub.u8 q12, %q[t0], %q[n51] \n\t"
+		"vqsub.u8 q13, %q[t1], %q[n51] \n\t"
+		"vqsub.u8 q14, %q[t2], %q[n51] \n\t"
+		"vqsub.u8 q15, %q[t3], %q[n51] \n\t"
+
+		// Create the mask for range #0.
+		"vcgt.u8 %q[m0], %q[t0], %q[n25] \n\t"
+		"vcgt.u8 %q[m1], %q[t1], %q[n25] \n\t"
+		"vcgt.u8 %q[m2], %q[t2], %q[n25] \n\t"
+		"vcgt.u8 %q[m3], %q[t3], %q[n25] \n\t"
+
+		// Subtract -1 to correct the LUT indices.
+		"vsub.u8 q12, %q[m0] \n\t"
+		"vsub.u8 q13, %q[m1] \n\t"
+		"vsub.u8 q14, %q[m2] \n\t"
+		"vsub.u8 q15, %q[m3] \n\t"
+
+		// Lookup the delta values.
+		"vtbl.u8 d24, {%q[lut]}, d24 \n\t"
+		"vtbl.u8 d25, {%q[lut]}, d25 \n\t"
+		"vtbl.u8 d26, {%q[lut]}, d26 \n\t"
+		"vtbl.u8 d27, {%q[lut]}, d27 \n\t"
+		"vtbl.u8 d28, {%q[lut]}, d28 \n\t"
+		"vtbl.u8 d29, {%q[lut]}, d29 \n\t"
+		"vtbl.u8 d30, {%q[lut]}, d30 \n\t"
+		"vtbl.u8 d31, {%q[lut]}, d31 \n\t"
+
+		// Add the delta values.
+		"vadd.u8 q12, %q[t0] \n\t"
+		"vadd.u8 q13, %q[t1] \n\t"
+		"vadd.u8 q14, %q[t2] \n\t"
+		"vadd.u8 q15, %q[t3] \n\t"
+
+		// Store 64 bytes and interleave. Increment the dest pointer.
+		"vst4.8 {d24, d26, d28, d30}, [%[dst]]! \n\t"
+		"vst4.8 {d25, d27, d29, d31}, [%[dst]]! \n\t"
+
+		// Outputs (modified).
+		: [src] "+r"  (*s),
+		  [dst] "+r"  (*o),
+		  [t0]  "=&w" (tmp0),
+		  [t1]  "=&w" (tmp1),
+		  [t2]  "=&w" (tmp2),
+		  [t3]  "=&w" (tmp3),
+		  [m0]  "=&w" (mask0),
+		  [m1]  "=&w" (mask1),
+		  [m2]  "=&w" (mask2),
+		  [m3]  "=&w" (mask3)
+
+		// Inputs (not modified).
+		: [lut] "w" (lut),
+		  [n25] "w" (n25),
+		  [n51] "w" (n51),
+		  [n63] "w" (n63)
+
+		// Clobbers.
+		: "d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31",
+		  "cc", "memory"
+	);
+}
+#endif
+
+static inline void
+enc_loop_neon32_inner (const uint8_t **s, uint8_t **o)
+{
+#ifdef BASE64_NEON32_USE_ASM
+	enc_loop_neon32_inner_asm(s, o);
+#else
+	// Load 48 bytes and deinterleave:
+	uint8x16x3_t src = vld3q_u8(*s);
+
+	// Reshuffle:
+	uint8x16x4_t out = enc_reshuffle(src);
+
+	// Translate reshuffled bytes to the Base64 alphabet:
+	out = enc_translate(out);
+
+	// Interleave and store output:
+	vst4q_u8(*o, out);
+
+	*s += 48;
+	*o += 64;
+#endif
+}
+
+static inline void
+enc_loop_neon32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	size_t rounds = *slen / 48;
+
+	*slen -= rounds * 48;	// 48 bytes consumed per round
+	*olen += rounds * 64;	// 64 bytes produced per round
+
+	while (rounds > 0) {
+		if (rounds >= 8) {
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_neon32_inner(s, o);
+		break;
+	}
+}

3rdparty/base64/lib/arch/neon32/enc_reshuffle.c

@@ -0,0 +1,31 @@
+static inline uint8x16x4_t
+enc_reshuffle (uint8x16x3_t in)
+{
+	uint8x16x4_t out;
+
+	// Input:
+	// in[0]  = a7 a6 a5 a4 a3 a2 a1 a0
+	// in[1]  = b7 b6 b5 b4 b3 b2 b1 b0
+	// in[2]  = c7 c6 c5 c4 c3 c2 c1 c0
+
+	// Output:
+	// out[0] = 00 00 a7 a6 a5 a4 a3 a2
+	// out[1] = 00 00 a1 a0 b7 b6 b5 b4
+	// out[2] = 00 00 b3 b2 b1 b0 c7 c6
+	// out[3] = 00 00 c5 c4 c3 c2 c1 c0
+
+	// Move the input bits to where they need to be in the outputs. Except
+	// for the first output, the high two bits are not cleared.
+	out.val[0] = vshrq_n_u8(in.val[0], 2);
+	out.val[1] = vshrq_n_u8(in.val[1], 4);
+	out.val[2] = vshrq_n_u8(in.val[2], 6);
+	out.val[1] = vsliq_n_u8(out.val[1], in.val[0], 4);
+	out.val[2] = vsliq_n_u8(out.val[2], in.val[1], 2);
+
+	// Clear the high two bits in the second, third and fourth output.
+	out.val[1] = vandq_u8(out.val[1], vdupq_n_u8(0x3F));
+	out.val[2] = vandq_u8(out.val[2], vdupq_n_u8(0x3F));
+	out.val[3] = vandq_u8(in.val[2],  vdupq_n_u8(0x3F));
+
+	return out;
+}

3rdparty/base64/lib/arch/neon32/enc_translate.c

@@ -0,0 +1,57 @@
+static inline uint8x16x4_t
+enc_translate (const uint8x16x4_t in)
+{
+	// A lookup table containing the absolute offsets for all ranges:
+	const uint8x16_t lut = {
+		 65U,  71U, 252U, 252U,
+		252U, 252U, 252U, 252U,
+		252U, 252U, 252U, 252U,
+		237U, 240U,   0U,   0U
+	};
+
+	const uint8x16_t offset = vdupq_n_u8(51);
+
+	uint8x16x4_t indices, mask, delta, out;
+
+	// Translate values 0..63 to the Base64 alphabet. There are five sets:
+	// #  From      To         Abs    Index  Characters
+	// 0  [0..25]   [65..90]   +65        0  ABCDEFGHIJKLMNOPQRSTUVWXYZ
+	// 1  [26..51]  [97..122]  +71        1  abcdefghijklmnopqrstuvwxyz
+	// 2  [52..61]  [48..57]    -4  [2..11]  0123456789
+	// 3  [62]      [43]       -19       12  +
+	// 4  [63]      [47]       -16       13  /
+
+	// Create LUT indices from input:
+	// the index for range #0 is right, others are 1 less than expected:
+	indices.val[0] = vqsubq_u8(in.val[0], offset);
+	indices.val[1] = vqsubq_u8(in.val[1], offset);
+	indices.val[2] = vqsubq_u8(in.val[2], offset);
+	indices.val[3] = vqsubq_u8(in.val[3], offset);
+
+	// mask is 0xFF (-1) for range #[1..4] and 0x00 for range #0:
+	mask.val[0] = vcgtq_u8(in.val[0], vdupq_n_u8(25));
+	mask.val[1] = vcgtq_u8(in.val[1], vdupq_n_u8(25));
+	mask.val[2] = vcgtq_u8(in.val[2], vdupq_n_u8(25));
+	mask.val[3] = vcgtq_u8(in.val[3], vdupq_n_u8(25));
+
+	// Subtract -1, so add 1 to indices for range #[1..4], All indices are
+	// now correct:
+	indices.val[0] = vsubq_u8(indices.val[0], mask.val[0]);
+	indices.val[1] = vsubq_u8(indices.val[1], mask.val[1]);
+	indices.val[2] = vsubq_u8(indices.val[2], mask.val[2]);
+	indices.val[3] = vsubq_u8(indices.val[3], mask.val[3]);
+
+	// Lookup delta values:
+	delta.val[0] = vqtbl1q_u8(lut, indices.val[0]);
+	delta.val[1] = vqtbl1q_u8(lut, indices.val[1]);
+	delta.val[2] = vqtbl1q_u8(lut, indices.val[2]);
+	delta.val[3] = vqtbl1q_u8(lut, indices.val[3]);
+
+	// Add delta values:
+	out.val[0] = vaddq_u8(in.val[0], delta.val[0]);
+	out.val[1] = vaddq_u8(in.val[1], delta.val[1]);
+	out.val[2] = vaddq_u8(in.val[2], delta.val[2]);
+	out.val[3] = vaddq_u8(in.val[3], delta.val[3]);
+
+	return out;
+}

3rdparty/base64/lib/arch/neon64/codec.c

@@ -0,0 +1,97 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <string.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#ifdef __aarch64__
+#  if (defined(__ARM_NEON__) || defined(__ARM_NEON)) && HAVE_NEON64
+#    define BASE64_USE_NEON64
+#  endif
+#endif
+
+#ifdef BASE64_USE_NEON64
+#include <arm_neon.h>
+
+// Only enable inline assembly on supported compilers.
+#if defined(__GNUC__) || defined(__clang__)
+#define BASE64_NEON64_USE_ASM
+#endif
+
+static inline uint8x16x4_t
+load_64byte_table (const uint8_t *p)
+{
+#ifdef BASE64_NEON64_USE_ASM
+
+	// Force the table to be loaded into contiguous registers. GCC will not
+	// normally allocate contiguous registers for a `uint8x16x4_t'. These
+	// registers are chosen to not conflict with the ones in the enc loop.
+	register uint8x16_t t0 __asm__ ("v8");
+	register uint8x16_t t1 __asm__ ("v9");
+	register uint8x16_t t2 __asm__ ("v10");
+	register uint8x16_t t3 __asm__ ("v11");
+
+	__asm__ (
+		"ld1 {%[t0].16b, %[t1].16b, %[t2].16b, %[t3].16b}, [%[src]], #64 \n\t"
+		: [src] "+r" (p),
+		  [t0]  "=w" (t0),
+		  [t1]  "=w" (t1),
+		  [t2]  "=w" (t2),
+		  [t3]  "=w" (t3)
+	);
+
+	return (uint8x16x4_t) {
+		.val[0] = t0,
+		.val[1] = t1,
+		.val[2] = t2,
+		.val[3] = t3,
+	};
+#else
+	return vld1q_u8_x4(p);
+#endif
+}
+
+#include "../generic/32/dec_loop.c"
+#include "../generic/64/enc_loop.c"
+#include "dec_loop.c"
+
+#ifdef BASE64_NEON64_USE_ASM
+# include "enc_loop_asm.c"
+#else
+# include "enc_reshuffle.c"
+# include "enc_loop.c"
+#endif
+
+#endif	// BASE64_USE_NEON64
+
+// Stride size is so large on these NEON 64-bit functions
+// (48 bytes encode, 64 bytes decode) that we inline the
+// uint64 codec to stay performant on smaller inputs.
+
+BASE64_ENC_FUNCTION(neon64)
+{
+#ifdef BASE64_USE_NEON64
+	#include "../generic/enc_head.c"
+	enc_loop_neon64(&s, &slen, &o, &olen);
+	enc_loop_generic_64(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(neon64)
+{
+#ifdef BASE64_USE_NEON64
+	#include "../generic/dec_head.c"
+	dec_loop_neon64(&s, &slen, &o, &olen);
+	dec_loop_generic_32(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/neon64/dec_loop.c

@@ -0,0 +1,129 @@
+// The input consists of five valid character sets in the Base64 alphabet,
+// which we need to map back to the 6-bit values they represent.
+// There are three ranges, two singles, and then there's the rest.
+//
+//   #  From       To        LUT  Characters
+//   1  [0..42]    [255]      #1  invalid input
+//   2  [43]       [62]       #1  +
+//   3  [44..46]   [255]      #1  invalid input
+//   4  [47]       [63]       #1  /
+//   5  [48..57]   [52..61]   #1  0..9
+//   6  [58..63]   [255]      #1  invalid input
+//   7  [64]       [255]      #2  invalid input
+//   8  [65..90]   [0..25]    #2  A..Z
+//   9  [91..96]   [255]      #2  invalid input
+//  10  [97..122]  [26..51]   #2  a..z
+//  11  [123..126] [255]      #2  invalid input
+// (12) Everything else => invalid input
+
+// The first LUT will use the VTBL instruction (out of range indices are set to
+// 0 in destination).
+static const uint8_t dec_lut1[] = {
+	255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U,
+	255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U,
+	255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U, 255U,  62U, 255U, 255U, 255U,  63U,
+	 52U,  53U,  54U,  55U,  56U,  57U,  58U,  59U,  60U,  61U, 255U, 255U, 255U, 255U, 255U, 255U,
+};
+
+// The second LUT will use the VTBX instruction (out of range indices will be
+// unchanged in destination). Input [64..126] will be mapped to index [1..63]
+// in this LUT. Index 0 means that value comes from LUT #1.
+static const uint8_t dec_lut2[] = {
+	  0U, 255U,   0U,   1U,   2U,   3U,   4U,   5U,   6U,   7U,   8U,   9U,  10U,  11U,  12U,  13U,
+	 14U,  15U,  16U,  17U,  18U,  19U,  20U,  21U,  22U,  23U,  24U,  25U, 255U, 255U, 255U, 255U,
+	255U, 255U,  26U,  27U,  28U,  29U,  30U,  31U,  32U,  33U,  34U,  35U,  36U,  37U,  38U,  39U,
+	 40U,  41U,  42U,  43U,  44U,  45U,  46U,  47U,  48U,  49U,  50U,  51U, 255U, 255U, 255U, 255U,
+};
+
+// All input values in range for the first look-up will be 0U in the second
+// look-up result. All input values out of range for the first look-up will be
+// 0U in the first look-up result. Thus, the two results can be ORed without
+// conflicts.
+//
+// Invalid characters that are in the valid range for either look-up will be
+// set to 255U in the combined result. Other invalid characters will just be
+// passed through with the second look-up result (using the VTBX instruction).
+// Since the second LUT is 64 bytes, those passed-through values are guaranteed
+// to have a value greater than 63U. Therefore, valid characters will be mapped
+// to the valid [0..63] range and all invalid characters will be mapped to
+// values greater than 63.
+
+static inline void
+dec_loop_neon64 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 64) {
+		return;
+	}
+
+	// Process blocks of 64 bytes per round. Unlike the SSE codecs, no
+	// extra trailing zero bytes are written, so it is not necessary to
+	// reserve extra input bytes:
+	size_t rounds = *slen / 64;
+
+	*slen -= rounds * 64;	// 64 bytes consumed per round
+	*olen += rounds * 48;	// 48 bytes produced per round
+
+	const uint8x16x4_t tbl_dec1 = load_64byte_table(dec_lut1);
+	const uint8x16x4_t tbl_dec2 = load_64byte_table(dec_lut2);
+
+	do {
+		const uint8x16_t offset = vdupq_n_u8(63U);
+		uint8x16x4_t dec1, dec2;
+		uint8x16x3_t dec;
+
+		// Load 64 bytes and deinterleave:
+		uint8x16x4_t str = vld4q_u8((uint8_t *) *s);
+
+		// Get indices for second LUT:
+		dec2.val[0] = vqsubq_u8(str.val[0], offset);
+		dec2.val[1] = vqsubq_u8(str.val[1], offset);
+		dec2.val[2] = vqsubq_u8(str.val[2], offset);
+		dec2.val[3] = vqsubq_u8(str.val[3], offset);
+
+		// Get values from first LUT:
+		dec1.val[0] = vqtbl4q_u8(tbl_dec1, str.val[0]);
+		dec1.val[1] = vqtbl4q_u8(tbl_dec1, str.val[1]);
+		dec1.val[2] = vqtbl4q_u8(tbl_dec1, str.val[2]);
+		dec1.val[3] = vqtbl4q_u8(tbl_dec1, str.val[3]);
+
+		// Get values from second LUT:
+		dec2.val[0] = vqtbx4q_u8(dec2.val[0], tbl_dec2, dec2.val[0]);
+		dec2.val[1] = vqtbx4q_u8(dec2.val[1], tbl_dec2, dec2.val[1]);
+		dec2.val[2] = vqtbx4q_u8(dec2.val[2], tbl_dec2, dec2.val[2]);
+		dec2.val[3] = vqtbx4q_u8(dec2.val[3], tbl_dec2, dec2.val[3]);
+
+		// Get final values:
+		str.val[0] = vorrq_u8(dec1.val[0], dec2.val[0]);
+		str.val[1] = vorrq_u8(dec1.val[1], dec2.val[1]);
+		str.val[2] = vorrq_u8(dec1.val[2], dec2.val[2]);
+		str.val[3] = vorrq_u8(dec1.val[3], dec2.val[3]);
+
+		// Check for invalid input, any value larger than 63:
+		const uint8x16_t classified
+			= vcgtq_u8(str.val[0], vdupq_n_u8(63))
+			| vcgtq_u8(str.val[1], vdupq_n_u8(63))
+			| vcgtq_u8(str.val[2], vdupq_n_u8(63))
+			| vcgtq_u8(str.val[3], vdupq_n_u8(63));
+
+		// Check that all bits are zero:
+		if (vmaxvq_u8(classified) != 0U) {
+			break;
+		}
+
+		// Compress four bytes into three:
+		dec.val[0] = vshlq_n_u8(str.val[0], 2) | vshrq_n_u8(str.val[1], 4);
+		dec.val[1] = vshlq_n_u8(str.val[1], 4) | vshrq_n_u8(str.val[2], 2);
+		dec.val[2] = vshlq_n_u8(str.val[2], 6) | str.val[3];
+
+		// Interleave and store decoded result:
+		vst3q_u8((uint8_t *) *o, dec);
+
+		*s += 64;
+		*o += 48;
+
+	} while (--rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 64;
+	*olen -= rounds * 48;
+}

3rdparty/base64/lib/arch/neon64/enc_loop.c

@@ -0,0 +1,66 @@
+static inline void
+enc_loop_neon64_inner (const uint8_t **s, uint8_t **o, const uint8x16x4_t tbl_enc)
+{
+	// Load 48 bytes and deinterleave:
+	uint8x16x3_t src = vld3q_u8(*s);
+
+	// Divide bits of three input bytes over four output bytes:
+	uint8x16x4_t out = enc_reshuffle(src);
+
+	// The bits have now been shifted to the right locations;
+	// translate their values 0..63 to the Base64 alphabet.
+	// Use a 64-byte table lookup:
+	out.val[0] = vqtbl4q_u8(tbl_enc, out.val[0]);
+	out.val[1] = vqtbl4q_u8(tbl_enc, out.val[1]);
+	out.val[2] = vqtbl4q_u8(tbl_enc, out.val[2]);
+	out.val[3] = vqtbl4q_u8(tbl_enc, out.val[3]);
+
+	// Interleave and store output:
+	vst4q_u8(*o, out);
+
+	*s += 48;
+	*o += 64;
+}
+
+static inline void
+enc_loop_neon64 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	size_t rounds = *slen / 48;
+
+	*slen -= rounds * 48;	// 48 bytes consumed per round
+	*olen += rounds * 64;	// 64 bytes produced per round
+
+	// Load the encoding table:
+	const uint8x16x4_t tbl_enc = load_64byte_table(base64_table_enc_6bit);
+
+	while (rounds > 0) {
+		if (rounds >= 8) {
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			enc_loop_neon64_inner(s, o, tbl_enc);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_neon64_inner(s, o, tbl_enc);
+		break;
+	}
+}

3rdparty/base64/lib/arch/neon64/enc_loop_asm.c

@@ -0,0 +1,168 @@
+// Apologies in advance for combining the preprocessor with inline assembly,
+// two notoriously gnarly parts of C, but it was necessary to avoid a lot of
+// code repetition. The preprocessor is used to template large sections of
+// inline assembly that differ only in the registers used. If the code was
+// written out by hand, it would become very large and hard to audit.
+
+// Generate a block of inline assembly that loads three user-defined registers
+// A, B, C from memory and deinterleaves them, post-incrementing the src
+// pointer. The register set should be sequential.
+#define LOAD(A, B, C) \
+	"ld3 {"A".16b, "B".16b, "C".16b}, [%[src]], #48 \n\t"
+
+// Generate a block of inline assembly that takes three deinterleaved registers
+// and shuffles the bytes. The output is in temporary registers t0..t3.
+#define SHUF(A, B, C) \
+	"ushr %[t0].16b, "A".16b,   #2         \n\t" \
+	"ushr %[t1].16b, "B".16b,   #4         \n\t" \
+	"ushr %[t2].16b, "C".16b,   #6         \n\t" \
+	"sli  %[t1].16b, "A".16b,   #4         \n\t" \
+	"sli  %[t2].16b, "B".16b,   #2         \n\t" \
+	"and  %[t1].16b, %[t1].16b, %[n63].16b \n\t" \
+	"and  %[t2].16b, %[t2].16b, %[n63].16b \n\t" \
+	"and  %[t3].16b, "C".16b,   %[n63].16b \n\t"
+
+// Generate a block of inline assembly that takes temporary registers t0..t3
+// and translates them to the base64 alphabet, using a table loaded into
+// v8..v11. The output is in user-defined registers A..D.
+#define TRAN(A, B, C, D) \
+	"tbl "A".16b, {v8.16b-v11.16b}, %[t0].16b \n\t" \
+	"tbl "B".16b, {v8.16b-v11.16b}, %[t1].16b \n\t" \
+	"tbl "C".16b, {v8.16b-v11.16b}, %[t2].16b \n\t" \
+	"tbl "D".16b, {v8.16b-v11.16b}, %[t3].16b \n\t"
+
+// Generate a block of inline assembly that interleaves four registers and
+// stores them, post-incrementing the destination pointer.
+#define STOR(A, B, C, D) \
+	"st4 {"A".16b, "B".16b, "C".16b, "D".16b}, [%[dst]], #64 \n\t"
+
+// Generate a block of inline assembly that generates a single self-contained
+// encoder round: fetch the data, process it, and store the result.
+#define ROUND() \
+	LOAD("v12", "v13", "v14") \
+	SHUF("v12", "v13", "v14") \
+	TRAN("v12", "v13", "v14", "v15") \
+	STOR("v12", "v13", "v14", "v15")
+
+// Generate a block of assembly that generates a type A interleaved encoder
+// round. It uses registers that were loaded by the previous type B round, and
+// in turn loads registers for the next type B round.
+#define ROUND_A() \
+	SHUF("v2",  "v3",  "v4") \
+	LOAD("v12", "v13", "v14") \
+	TRAN("v2",  "v3",  "v4", "v5") \
+	STOR("v2",  "v3",  "v4", "v5")
+
+// Type B interleaved encoder round. Same as type A, but register sets swapped.
+#define ROUND_B() \
+	SHUF("v12", "v13", "v14") \
+	LOAD("v2",  "v3",  "v4") \
+	TRAN("v12", "v13", "v14", "v15") \
+	STOR("v12", "v13", "v14", "v15")
+
+// The first type A round needs to load its own registers.
+#define ROUND_A_FIRST() \
+	LOAD("v2", "v3", "v4") \
+	ROUND_A()
+
+// The last type B round omits the load for the next step.
+#define ROUND_B_LAST() \
+	SHUF("v12", "v13", "v14") \
+	TRAN("v12", "v13", "v14", "v15") \
+	STOR("v12", "v13", "v14", "v15")
+
+// Suppress clang's warning that the literal string in the asm statement is
+// overlong (longer than the ISO-mandated minimum size of 4095 bytes for C99
+// compilers). It may be true, but the goal here is not C99 portability.
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+
+static inline void
+enc_loop_neon64 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	size_t rounds = *slen / 48;
+
+	if (rounds == 0) {
+		return;
+	}
+
+	*slen -= rounds * 48;	// 48 bytes consumed per round.
+	*olen += rounds * 64;	// 64 bytes produced per round.
+
+	// Number of times to go through the 8x loop.
+	size_t loops = rounds / 8;
+
+	// Number of rounds remaining after the 8x loop.
+	rounds %= 8;
+
+	// Temporary registers, used as scratch space.
+	uint8x16_t tmp0, tmp1, tmp2, tmp3;
+
+	__asm__ volatile (
+
+		// Load the encoding table into v8..v11.
+		"    ld1 {v8.16b-v11.16b}, [%[tbl]] \n\t"
+
+		// If there are eight rounds or more, enter an 8x unrolled loop
+		// of interleaved encoding rounds. The rounds interleave memory
+		// operations (load/store) with data operations to maximize
+		// pipeline throughput.
+		"    cbz %[loops], 4f \n\t"
+
+		// The SIMD instructions do not touch the flags.
+		"88: subs %[loops], %[loops], #1 \n\t"
+		"    " ROUND_A_FIRST()
+		"    " ROUND_B()
+		"    " ROUND_A()
+		"    " ROUND_B()
+		"    " ROUND_A()
+		"    " ROUND_B()
+		"    " ROUND_A()
+		"    " ROUND_B_LAST()
+		"    b.ne 88b \n\t"
+
+		// Enter a 4x unrolled loop for rounds of 4 or more.
+		"4:  cmp  %[rounds], #4 \n\t"
+		"    b.lt 30f           \n\t"
+		"    " ROUND_A_FIRST()
+		"    " ROUND_B()
+		"    " ROUND_A()
+		"    " ROUND_B_LAST()
+		"    sub %[rounds], %[rounds], #4 \n\t"
+
+		// Dispatch the remaining rounds 0..3.
+		"30: cbz  %[rounds], 0f \n\t"
+		"    cmp  %[rounds], #2 \n\t"
+		"    b.eq 2f            \n\t"
+		"    b.lt 1f            \n\t"
+
+		// Block of non-interlaced encoding rounds, which can each
+		// individually be jumped to. Rounds fall through to the next.
+		"3:  " ROUND()
+		"2:  " ROUND()
+		"1:  " ROUND()
+		"0:  \n\t"
+
+		// Outputs (modified).
+		: [loops] "+r"  (loops),
+		  [src]   "+r"  (*s),
+		  [dst]   "+r"  (*o),
+		  [t0]    "=&w" (tmp0),
+		  [t1]    "=&w" (tmp1),
+		  [t2]    "=&w" (tmp2),
+		  [t3]    "=&w" (tmp3)
+
+		// Inputs (not modified).
+		: [rounds] "r" (rounds),
+		  [tbl]    "r" (base64_table_enc_6bit),
+		  [n63]    "w" (vdupq_n_u8(63))
+
+		// Clobbers.
+		: "v2",  "v3",  "v4",  "v5",
+		  "v8",  "v9",  "v10", "v11",
+		  "v12", "v13", "v14", "v15",
+		  "cc", "memory"
+	);
+}
+
+#pragma GCC diagnostic pop

3rdparty/base64/lib/arch/neon64/enc_reshuffle.c

@@ -0,0 +1,31 @@
+static inline uint8x16x4_t
+enc_reshuffle (const uint8x16x3_t in)
+{
+	uint8x16x4_t out;
+
+	// Input:
+	// in[0]  = a7 a6 a5 a4 a3 a2 a1 a0
+	// in[1]  = b7 b6 b5 b4 b3 b2 b1 b0
+	// in[2]  = c7 c6 c5 c4 c3 c2 c1 c0
+
+	// Output:
+	// out[0] = 00 00 a7 a6 a5 a4 a3 a2
+	// out[1] = 00 00 a1 a0 b7 b6 b5 b4
+	// out[2] = 00 00 b3 b2 b1 b0 c7 c6
+	// out[3] = 00 00 c5 c4 c3 c2 c1 c0
+
+	// Move the input bits to where they need to be in the outputs. Except
+	// for the first output, the high two bits are not cleared.
+	out.val[0] = vshrq_n_u8(in.val[0], 2);
+	out.val[1] = vshrq_n_u8(in.val[1], 4);
+	out.val[2] = vshrq_n_u8(in.val[2], 6);
+	out.val[1] = vsliq_n_u8(out.val[1], in.val[0], 4);
+	out.val[2] = vsliq_n_u8(out.val[2], in.val[1], 2);
+
+	// Clear the high two bits in the second, third and fourth output.
+	out.val[1] = vandq_u8(out.val[1], vdupq_n_u8(0x3F));
+	out.val[2] = vandq_u8(out.val[2], vdupq_n_u8(0x3F));
+	out.val[3] = vandq_u8(in.val[2],  vdupq_n_u8(0x3F));
+
+	return out;
+}

3rdparty/base64/lib/arch/sse41/codec.c

@@ -0,0 +1,56 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdlib.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if HAVE_SSE41
+#include <smmintrin.h>
+
+// Only enable inline assembly on supported compilers and on 64-bit CPUs.
+#ifndef BASE64_SSE41_USE_ASM
+# if (defined(__GNUC__) || defined(__clang__)) && BASE64_WORDSIZE == 64
+#  define BASE64_SSE41_USE_ASM 1
+# else
+#  define BASE64_SSE41_USE_ASM 0
+# endif
+#endif
+
+#include "../ssse3/dec_reshuffle.c"
+#include "../ssse3/dec_loop.c"
+
+#if BASE64_SSE41_USE_ASM
+# include "../ssse3/enc_loop_asm.c"
+#else
+# include "../ssse3/enc_translate.c"
+# include "../ssse3/enc_reshuffle.c"
+# include "../ssse3/enc_loop.c"
+#endif
+
+#endif	// HAVE_SSE41
+
+BASE64_ENC_FUNCTION(sse41)
+{
+#if HAVE_SSE41
+	#include "../generic/enc_head.c"
+	enc_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(sse41)
+{
+#if HAVE_SSE41
+	#include "../generic/dec_head.c"
+	dec_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/sse42/codec.c

@@ -0,0 +1,56 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdlib.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if HAVE_SSE42
+#include <nmmintrin.h>
+
+// Only enable inline assembly on supported compilers and on 64-bit CPUs.
+#ifndef BASE64_SSE42_USE_ASM
+# if (defined(__GNUC__) || defined(__clang__)) && BASE64_WORDSIZE == 64
+#  define BASE64_SSE42_USE_ASM 1
+# else
+#  define BASE64_SSE42_USE_ASM 0
+# endif
+#endif
+
+#include "../ssse3/dec_reshuffle.c"
+#include "../ssse3/dec_loop.c"
+
+#if BASE64_SSE42_USE_ASM
+# include "../ssse3/enc_loop_asm.c"
+#else
+# include "../ssse3/enc_translate.c"
+# include "../ssse3/enc_reshuffle.c"
+# include "../ssse3/enc_loop.c"
+#endif
+
+#endif	// HAVE_SSE42
+
+BASE64_ENC_FUNCTION(sse42)
+{
+#if HAVE_SSE42
+	#include "../generic/enc_head.c"
+	enc_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(sse42)
+{
+#if HAVE_SSE42
+	#include "../generic/dec_head.c"
+	dec_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/ssse3/codec.c

@@ -0,0 +1,58 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdlib.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if HAVE_SSSE3
+#include <tmmintrin.h>
+
+// Only enable inline assembly on supported compilers and on 64-bit CPUs.
+// 32-bit CPUs with SSSE3 support, such as low-end Atoms, only have eight XMM
+// registers, which is not enough to run the inline assembly.
+#ifndef BASE64_SSSE3_USE_ASM
+# if (defined(__GNUC__) || defined(__clang__)) && BASE64_WORDSIZE == 64
+#  define BASE64_SSSE3_USE_ASM 1
+# else
+#  define BASE64_SSSE3_USE_ASM 0
+# endif
+#endif
+
+#include "dec_reshuffle.c"
+#include "dec_loop.c"
+
+#if BASE64_SSSE3_USE_ASM
+# include "enc_loop_asm.c"
+#else
+# include "enc_reshuffle.c"
+# include "enc_translate.c"
+# include "enc_loop.c"
+#endif
+
+#endif	// HAVE_SSSE3
+
+BASE64_ENC_FUNCTION(ssse3)
+{
+#if HAVE_SSSE3
+	#include "../generic/enc_head.c"
+	enc_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(ssse3)
+{
+#if HAVE_SSSE3
+	#include "../generic/dec_head.c"
+	dec_loop_ssse3(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

3rdparty/base64/lib/arch/ssse3/dec_loop.c

@@ -0,0 +1,173 @@
+// The input consists of six character sets in the Base64 alphabet, which we
+// need to map back to the 6-bit values they represent. There are three ranges,
+// two singles, and then there's the rest.
+//
+//  #  From       To        Add  Characters
+//  1  [43]       [62]      +19  +
+//  2  [47]       [63]      +16  /
+//  3  [48..57]   [52..61]   +4  0..9
+//  4  [65..90]   [0..25]   -65  A..Z
+//  5  [97..122]  [26..51]  -71  a..z
+// (6) Everything else => invalid input
+//
+// We will use lookup tables for character validation and offset computation.
+// Remember that 0x2X and 0x0X are the same index for _mm_shuffle_epi8, this
+// allows to mask with 0x2F instead of 0x0F and thus save one constant
+// declaration (register and/or memory access).
+//
+// For offsets:
+// Perfect hash for lut = ((src >> 4) & 0x2F) + ((src == 0x2F) ? 0xFF : 0x00)
+// 0000 = garbage
+// 0001 = /
+// 0010 = +
+// 0011 = 0-9
+// 0100 = A-Z
+// 0101 = A-Z
+// 0110 = a-z
+// 0111 = a-z
+// 1000 >= garbage
+//
+// For validation, here's the table.
+// A character is valid if and only if the AND of the 2 lookups equals 0:
+//
+// hi \ lo              0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
+//      LUT             0x15 0x11 0x11 0x11 0x11 0x11 0x11 0x11 0x11 0x11 0x13 0x1A 0x1B 0x1B 0x1B 0x1A
+//
+// 0000 0x10 char        NUL  SOH  STX  ETX  EOT  ENQ  ACK  BEL   BS   HT   LF   VT   FF   CR   SO   SI
+//           andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+//
+// 0001 0x10 char        DLE  DC1  DC2  DC3  DC4  NAK  SYN  ETB  CAN   EM  SUB  ESC   FS   GS   RS   US
+//           andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+//
+// 0010 0x01 char               !    "    #    $    %    &    '    (    )    *    +    ,    -    .    /
+//           andlut     0x01 0x01 0x01 0x01 0x01 0x01 0x01 0x01 0x01 0x01 0x01 0x00 0x01 0x01 0x01 0x00
+//
+// 0011 0x02 char          0    1    2    3    4    5    6    7    8    9    :    ;    <    =    >    ?
+//           andlut     0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x02 0x02 0x02 0x02 0x02 0x02
+//
+// 0100 0x04 char          @    A    B    C    D    E    F    G    H    I    J    K    L    M    N    O
+//           andlut     0x04 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00
+//
+// 0101 0x08 char          P    Q    R    S    T    U    V    W    X    Y    Z    [    \    ]    ^    _
+//           andlut     0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x08 0x08 0x08 0x08 0x08
+//
+// 0110 0x04 char          `    a    b    c    d    e    f    g    h    i    j    k    l    m    n    o
+//           andlut     0x04 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00
+// 0111 0x08 char          p    q    r    s    t    u    v    w    x    y    z    {    |    }    ~
+//           andlut     0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x08 0x08 0x08 0x08 0x08
+//
+// 1000 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1001 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1010 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1011 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1100 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1101 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1110 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+// 1111 0x10 andlut     0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10 0x10
+
+static inline int
+dec_loop_ssse3_inner (const uint8_t **s, uint8_t **o, size_t *rounds)
+{
+	const __m128i lut_lo = _mm_setr_epi8(
+		0x15, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11,
+		0x11, 0x11, 0x13, 0x1A, 0x1B, 0x1B, 0x1B, 0x1A);
+
+	const __m128i lut_hi = _mm_setr_epi8(
+		0x10, 0x10, 0x01, 0x02, 0x04, 0x08, 0x04, 0x08,
+		0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10);
+
+	const __m128i lut_roll = _mm_setr_epi8(
+		0,  16,  19,   4, -65, -65, -71, -71,
+		0,   0,   0,   0,   0,   0,   0,   0);
+
+	const __m128i mask_2F = _mm_set1_epi8(0x2F);
+
+	// Load input:
+	__m128i str = _mm_loadu_si128((__m128i *) *s);
+
+	// Table lookups:
+	const __m128i hi_nibbles = _mm_and_si128(_mm_srli_epi32(str, 4), mask_2F);
+	const __m128i lo_nibbles = _mm_and_si128(str, mask_2F);
+	const __m128i hi         = _mm_shuffle_epi8(lut_hi, hi_nibbles);
+	const __m128i lo         = _mm_shuffle_epi8(lut_lo, lo_nibbles);
+
+	// Check for invalid input: if any "and" values from lo and hi are not
+	// zero, fall back on bytewise code to do error checking and reporting:
+	if (_mm_movemask_epi8(_mm_cmpgt_epi8(_mm_and_si128(lo, hi), _mm_setzero_si128())) != 0) {
+		return 0;
+	}
+
+	const __m128i eq_2F = _mm_cmpeq_epi8(str, mask_2F);
+	const __m128i roll  = _mm_shuffle_epi8(lut_roll, _mm_add_epi8(eq_2F, hi_nibbles));
+
+	// Now simply add the delta values to the input:
+	str = _mm_add_epi8(str, roll);
+
+	// Reshuffle the input to packed 12-byte output format:
+	str = dec_reshuffle(str);
+
+	// Store the output:
+	_mm_storeu_si128((__m128i *) *o, str);
+
+	*s += 16;
+	*o += 12;
+	*rounds -= 1;
+
+	return 1;
+}
+
+static inline void
+dec_loop_ssse3 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 24) {
+		return;
+	}
+
+	// Process blocks of 16 bytes per round. Because 4 extra zero bytes are
+	// written after the output, ensure that there will be at least 8 bytes
+	// of input data left to cover the gap. (6 data bytes and up to two
+	// end-of-string markers.)
+	size_t rounds = (*slen - 8) / 16;
+
+	*slen -= rounds * 16;	// 16 bytes consumed per round
+	*olen += rounds * 12;	// 12 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			if (dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 4) {
+			if (dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 2) {
+			if (dec_loop_ssse3_inner(s, o, &rounds) &&
+			    dec_loop_ssse3_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		dec_loop_ssse3_inner(s, o, &rounds);
+		break;
+
+	} while (rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 16;
+	*olen -= rounds * 12;
+}

3rdparty/base64/lib/arch/ssse3/dec_reshuffle.c

@@ -0,0 +1,33 @@
+static inline __m128i
+dec_reshuffle (const __m128i in)
+{
+	// in, bits, upper case are most significant bits, lower case are least significant bits
+	// 00llllll 00kkkkLL 00jjKKKK 00JJJJJJ
+	// 00iiiiii 00hhhhII 00ggHHHH 00GGGGGG
+	// 00ffffff 00eeeeFF 00ddEEEE 00DDDDDD
+	// 00cccccc 00bbbbCC 00aaBBBB 00AAAAAA
+
+	const __m128i merge_ab_and_bc = _mm_maddubs_epi16(in, _mm_set1_epi32(0x01400140));
+	// 0000kkkk LLllllll 0000JJJJ JJjjKKKK
+	// 0000hhhh IIiiiiii 0000GGGG GGggHHHH
+	// 0000eeee FFffffff 0000DDDD DDddEEEE
+	// 0000bbbb CCcccccc 0000AAAA AAaaBBBB
+
+	const __m128i out = _mm_madd_epi16(merge_ab_and_bc, _mm_set1_epi32(0x00011000));
+	// 00000000 JJJJJJjj KKKKkkkk LLllllll
+	// 00000000 GGGGGGgg HHHHhhhh IIiiiiii
+	// 00000000 DDDDDDdd EEEEeeee FFffffff
+	// 00000000 AAAAAAaa BBBBbbbb CCcccccc
+
+	// Pack bytes together:
+	return  _mm_shuffle_epi8(out, _mm_setr_epi8(
+		 2,  1,  0,
+		 6,  5,  4,
+		10,  9,  8,
+		14, 13, 12,
+		-1, -1, -1, -1));
+	// 00000000 00000000 00000000 00000000
+	// LLllllll KKKKkkkk JJJJJJjj IIiiiiii
+	// HHHHhhhh GGGGGGgg FFffffff EEEEeeee
+	// DDDDDDdd CCcccccc BBBBbbbb AAAAAAaa
+}

3rdparty/base64/lib/arch/ssse3/enc_loop.c

@@ -0,0 +1,67 @@
+static inline void
+enc_loop_ssse3_inner (const uint8_t **s, uint8_t **o)
+{
+	// Load input:
+	__m128i str = _mm_loadu_si128((__m128i *) *s);
+
+	// Reshuffle:
+	str = enc_reshuffle(str);
+
+	// Translate reshuffled bytes to the Base64 alphabet:
+	str = enc_translate(str);
+
+	// Store:
+	_mm_storeu_si128((__m128i *) *o, str);
+
+	*s += 12;
+	*o += 16;
+}
+
+static inline void
+enc_loop_ssse3 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 16) {
+		return;
+	}
+
+	// Process blocks of 12 bytes at a time. Because blocks are loaded 16
+	// bytes at a time, ensure that there will be at least 4 remaining
+	// bytes after the last round, so that the final read will not pass
+	// beyond the bounds of the input buffer:
+	size_t rounds = (*slen - 4) / 12;
+
+	*slen -= rounds * 12;	// 12 bytes consumed per round
+	*olen += rounds * 16;	// 16 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_ssse3_inner(s, o);
+		break;
+
+	} while (rounds > 0);
+}

3rdparty/base64/lib/arch/ssse3/enc_loop_asm.c

@@ -0,0 +1,268 @@
+// Apologies in advance for combining the preprocessor with inline assembly,
+// two notoriously gnarly parts of C, but it was necessary to avoid a lot of
+// code repetition. The preprocessor is used to template large sections of
+// inline assembly that differ only in the registers used. If the code was
+// written out by hand, it would become very large and hard to audit.
+
+// Generate a block of inline assembly that loads register R0 from memory. The
+// offset at which the register is loaded is set by the given round.
+#define LOAD(R0, ROUND) \
+	"lddqu ("#ROUND" * 12)(%[src]), %["R0"] \n\t"
+
+// Generate a block of inline assembly that deinterleaves and shuffles register
+// R0 using preloaded constants. Outputs in R0 and R1.
+#define SHUF(R0, R1) \
+	"pshufb  %[lut0], %["R0"] \n\t" \
+	"movdqa  %["R0"], %["R1"] \n\t" \
+	"pand    %[msk0], %["R0"] \n\t" \
+	"pand    %[msk2], %["R1"] \n\t" \
+	"pmulhuw %[msk1], %["R0"] \n\t" \
+	"pmullw  %[msk3], %["R1"] \n\t" \
+	"por     %["R1"], %["R0"] \n\t"
+
+// Generate a block of inline assembly that takes R0 and R1 and translates
+// their contents to the base64 alphabet, using preloaded constants.
+#define TRAN(R0, R1, R2) \
+	"movdqa  %["R0"], %["R1"] \n\t" \
+	"movdqa  %["R0"], %["R2"] \n\t" \
+	"psubusb %[n51],  %["R1"] \n\t" \
+	"pcmpgtb %[n25],  %["R2"] \n\t" \
+	"psubb   %["R2"], %["R1"] \n\t" \
+	"movdqa  %[lut1], %["R2"] \n\t" \
+	"pshufb  %["R1"], %["R2"] \n\t" \
+	"paddb   %["R2"], %["R0"] \n\t"
+
+// Generate a block of inline assembly that stores the given register R0 at an
+// offset set by the given round.
+#define STOR(R0, ROUND) \
+	"movdqu %["R0"], ("#ROUND" * 16)(%[dst]) \n\t"
+
+// Generate a block of inline assembly that generates a single self-contained
+// encoder round: fetch the data, process it, and store the result. Then update
+// the source and destination pointers.
+#define ROUND() \
+	LOAD("a", 0) \
+	SHUF("a", "b") \
+	TRAN("a", "b", "c") \
+	STOR("a", 0) \
+	"add $12, %[src] \n\t" \
+	"add $16, %[dst] \n\t"
+
+// Define a macro that initiates a three-way interleaved encoding round by
+// preloading registers a, b and c from memory.
+// The register graph shows which registers are in use during each step, and
+// is a visual aid for choosing registers for that step. Symbol index:
+//
+//  +  indicates that a register is loaded by that step.
+//  |  indicates that a register is in use and must not be touched.
+//  -  indicates that a register is decommissioned by that step.
+//  x  indicates that a register is used as a temporary by that step.
+//  V  indicates that a register is an input or output to the macro.
+//
+#define ROUND_3_INIT() 			/*  a b c d e f  */ \
+	LOAD("a", 0)			/*  +            */ \
+	SHUF("a", "d")			/*  |     +      */ \
+	LOAD("b", 1)			/*  | +   |      */ \
+	TRAN("a", "d", "e")		/*  | |   - x    */ \
+	LOAD("c", 2)			/*  V V V        */
+
+// Define a macro that translates, shuffles and stores the input registers A, B
+// and C, and preloads registers D, E and F for the next round.
+// This macro can be arbitrarily daisy-chained by feeding output registers D, E
+// and F back into the next round as input registers A, B and C. The macro
+// carefully interleaves memory operations with data operations for optimal
+// pipelined performance.
+
+#define ROUND_3(ROUND, A,B,C,D,E,F) 	/*  A B C D E F  */ \
+	LOAD(D, (ROUND + 3))		/*  V V V +      */ \
+	SHUF(B, E)			/*  | | | | +    */ \
+	STOR(A, (ROUND + 0))		/*  - | | | |    */ \
+	TRAN(B, E, F)			/*    | | | - x  */ \
+	LOAD(E, (ROUND + 4))		/*    | | | +    */ \
+	SHUF(C, A)			/*  + | | | |    */ \
+	STOR(B, (ROUND + 1))		/*  | - | | |    */ \
+	TRAN(C, A, F)			/*  -   | | | x  */ \
+	LOAD(F, (ROUND + 5))		/*      | | | +  */ \
+	SHUF(D, A)			/*  +   | | | |  */ \
+	STOR(C, (ROUND + 2))		/*  |   - | | |  */ \
+	TRAN(D, A, B)			/*  - x   V V V  */
+
+// Define a macro that terminates a ROUND_3 macro by taking pre-loaded
+// registers D, E and F, and translating, shuffling and storing them.
+#define ROUND_3_END(ROUND, A,B,C,D,E,F)	/*  A B C D E F  */ \
+	SHUF(E, A)			/*  +     V V V  */ \
+	STOR(D, (ROUND + 3))		/*  |     - | |  */ \
+	TRAN(E, A, B)			/*  - x     | |  */ \
+	SHUF(F, C)			/*      +   | |  */ \
+	STOR(E, (ROUND + 4))		/*      |   - |  */ \
+	TRAN(F, C, D)			/*      - x   |  */ \
+	STOR(F, (ROUND + 5))		/*            -  */
+
+// Define a type A round. Inputs are a, b, and c, outputs are d, e, and f.
+#define ROUND_3_A(ROUND) \
+	ROUND_3(ROUND, "a", "b", "c", "d", "e", "f")
+
+// Define a type B round. Inputs and outputs are swapped with regard to type A.
+#define ROUND_3_B(ROUND) \
+	ROUND_3(ROUND, "d", "e", "f", "a", "b", "c")
+
+// Terminating macro for a type A round.
+#define ROUND_3_A_LAST(ROUND) \
+	ROUND_3_A(ROUND) \
+	ROUND_3_END(ROUND, "a", "b", "c", "d", "e", "f")
+
+// Terminating macro for a type B round.
+#define ROUND_3_B_LAST(ROUND) \
+	ROUND_3_B(ROUND) \
+	ROUND_3_END(ROUND, "d", "e", "f", "a", "b", "c")
+
+// Suppress clang's warning that the literal string in the asm statement is
+// overlong (longer than the ISO-mandated minimum size of 4095 bytes for C99
+// compilers). It may be true, but the goal here is not C99 portability.
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+
+static inline void
+enc_loop_ssse3 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	// For a clearer explanation of the algorithm used by this function,
+	// please refer to the plain (not inline assembly) implementation. This
+	// function follows the same basic logic.
+
+	if (*slen < 16) {
+		return;
+	}
+
+	// Process blocks of 12 bytes at a time. Input is read in blocks of 16
+	// bytes, so "reserve" four bytes from the input buffer to ensure that
+	// we never read beyond the end of the input buffer.
+	size_t rounds = (*slen - 4) / 12;
+
+	*slen -= rounds * 12;   // 12 bytes consumed per round
+	*olen += rounds * 16;   // 16 bytes produced per round
+
+	// Number of times to go through the 36x loop.
+	size_t loops = rounds / 36;
+
+	// Number of rounds remaining after the 36x loop.
+	rounds %= 36;
+
+	// Lookup tables.
+	const __m128i lut0 = _mm_set_epi8(
+		10, 11,  9, 10,  7,  8,  6,  7,  4,  5,  3,  4,  1,  2,  0,  1);
+
+	const __m128i lut1 = _mm_setr_epi8(
+		65, 71, -4, -4, -4, -4, -4, -4, -4, -4, -4, -4, -19, -16, 0, 0);
+
+	// Temporary registers.
+	__m128i a, b, c, d, e, f;
+
+	__asm__ volatile (
+
+		// If there are 36 rounds or more, enter a 36x unrolled loop of
+		// interleaved encoding rounds. The rounds interleave memory
+		// operations (load/store) with data operations (table lookups,
+		// etc) to maximize pipeline throughput.
+		"    test %[loops], %[loops] \n\t"
+		"    jz   18f                \n\t"
+		"    jmp  36f                \n\t"
+		"                            \n\t"
+		".balign 64                  \n\t"
+		"36: " ROUND_3_INIT()
+		"    " ROUND_3_A( 0)
+		"    " ROUND_3_B( 3)
+		"    " ROUND_3_A( 6)
+		"    " ROUND_3_B( 9)
+		"    " ROUND_3_A(12)
+		"    " ROUND_3_B(15)
+		"    " ROUND_3_A(18)
+		"    " ROUND_3_B(21)
+		"    " ROUND_3_A(24)
+		"    " ROUND_3_B(27)
+		"    " ROUND_3_A_LAST(30)
+		"    add $(12 * 36), %[src] \n\t"
+		"    add $(16 * 36), %[dst] \n\t"
+		"    dec %[loops]           \n\t"
+		"    jnz 36b                \n\t"
+
+		// Enter an 18x unrolled loop for rounds of 18 or more.
+		"18: cmp $18, %[rounds] \n\t"
+		"    jl  9f             \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A(0)
+		"    " ROUND_3_B(3)
+		"    " ROUND_3_A(6)
+		"    " ROUND_3_B(9)
+		"    " ROUND_3_A_LAST(12)
+		"    sub $18,        %[rounds] \n\t"
+		"    add $(12 * 18), %[src]    \n\t"
+		"    add $(16 * 18), %[dst]    \n\t"
+
+		// Enter a 9x unrolled loop for rounds of 9 or more.
+		"9:  cmp $9, %[rounds] \n\t"
+		"    jl  6f            \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A(0)
+		"    " ROUND_3_B_LAST(3)
+		"    sub $9,        %[rounds] \n\t"
+		"    add $(12 * 9), %[src]    \n\t"
+		"    add $(16 * 9), %[dst]    \n\t"
+
+		// Enter a 6x unrolled loop for rounds of 6 or more.
+		"6:  cmp $6, %[rounds] \n\t"
+		"    jl  55f           \n\t"
+		"    " ROUND_3_INIT()
+		"    " ROUND_3_A_LAST(0)
+		"    sub $6,        %[rounds] \n\t"
+		"    add $(12 * 6), %[src]    \n\t"
+		"    add $(16 * 6), %[dst]    \n\t"
+
+		// Dispatch the remaining rounds 0..5.
+		"55: cmp $3, %[rounds] \n\t"
+		"    jg  45f           \n\t"
+		"    je  3f            \n\t"
+		"    cmp $1, %[rounds] \n\t"
+		"    jg  2f            \n\t"
+		"    je  1f            \n\t"
+		"    jmp 0f            \n\t"
+
+		"45: cmp $4, %[rounds] \n\t"
+		"    je  4f            \n\t"
+
+		// Block of non-interlaced encoding rounds, which can each
+		// individually be jumped to. Rounds fall through to the next.
+		"5: " ROUND()
+		"4: " ROUND()
+		"3: " ROUND()
+		"2: " ROUND()
+		"1: " ROUND()
+		"0: \n\t"
+
+		// Outputs (modified).
+		: [rounds] "+r"  (rounds),
+		  [loops]  "+r"  (loops),
+		  [src]    "+r"  (*s),
+		  [dst]    "+r"  (*o),
+		  [a]      "=&x" (a),
+		  [b]      "=&x" (b),
+		  [c]      "=&x" (c),
+		  [d]      "=&x" (d),
+		  [e]      "=&x" (e),
+		  [f]      "=&x" (f)
+
+		// Inputs (not modified).
+		: [lut0] "x" (lut0),
+		  [lut1] "x" (lut1),
+		  [msk0] "x" (_mm_set1_epi32(0x0FC0FC00)),
+		  [msk1] "x" (_mm_set1_epi32(0x04000040)),
+		  [msk2] "x" (_mm_set1_epi32(0x003F03F0)),
+		  [msk3] "x" (_mm_set1_epi32(0x01000010)),
+		  [n51]  "x" (_mm_set1_epi8(51)),
+		  [n25]  "x" (_mm_set1_epi8(25))
+
+		// Clobbers.
+		: "cc", "memory"
+	);
+}
+
+#pragma GCC diagnostic pop

3rdparty/base64/lib/arch/ssse3/enc_reshuffle.c

@@ -0,0 +1,48 @@
+static inline __m128i
+enc_reshuffle (__m128i in)
+{
+	// Input, bytes MSB to LSB:
+	// 0 0 0 0 l k j i h g f e d c b a
+
+	in = _mm_shuffle_epi8(in, _mm_set_epi8(
+		10, 11,  9, 10,
+		 7,  8,  6,  7,
+		 4,  5,  3,  4,
+		 1,  2,  0,  1));
+	// in, bytes MSB to LSB:
+	// k l j k
+	// h i g h
+	// e f d e
+	// b c a b
+
+	const __m128i t0 = _mm_and_si128(in, _mm_set1_epi32(0x0FC0FC00));
+	// bits, upper case are most significant bits, lower case are least significant bits
+	// 0000kkkk LL000000 JJJJJJ00 00000000
+	// 0000hhhh II000000 GGGGGG00 00000000
+	// 0000eeee FF000000 DDDDDD00 00000000
+	// 0000bbbb CC000000 AAAAAA00 00000000
+
+	const __m128i t1 = _mm_mulhi_epu16(t0, _mm_set1_epi32(0x04000040));
+	// 00000000 00kkkkLL 00000000 00JJJJJJ
+	// 00000000 00hhhhII 00000000 00GGGGGG
+	// 00000000 00eeeeFF 00000000 00DDDDDD
+	// 00000000 00bbbbCC 00000000 00AAAAAA
+
+	const __m128i t2 = _mm_and_si128(in, _mm_set1_epi32(0x003F03F0));
+	// 00000000 00llllll 000000jj KKKK0000
+	// 00000000 00iiiiii 000000gg HHHH0000
+	// 00000000 00ffffff 000000dd EEEE0000
+	// 00000000 00cccccc 000000aa BBBB0000
+
+	const __m128i t3 = _mm_mullo_epi16(t2, _mm_set1_epi32(0x01000010));
+	// 00llllll 00000000 00jjKKKK 00000000
+	// 00iiiiii 00000000 00ggHHHH 00000000
+	// 00ffffff 00000000 00ddEEEE 00000000
+	// 00cccccc 00000000 00aaBBBB 00000000
+
+	return _mm_or_si128(t1, t3);
+	// 00llllll 00kkkkLL 00jjKKKK 00JJJJJJ
+	// 00iiiiii 00hhhhII 00ggHHHH 00GGGGGG
+	// 00ffffff 00eeeeFF 00ddEEEE 00DDDDDD
+	// 00cccccc 00bbbbCC 00aaBBBB 00AAAAAA
+}

3rdparty/base64/lib/arch/ssse3/enc_translate.c

@@ -0,0 +1,33 @@
+static inline __m128i
+enc_translate (const __m128i in)
+{
+	// A lookup table containing the absolute offsets for all ranges:
+	const __m128i lut = _mm_setr_epi8(
+		 65,  71, -4, -4,
+		 -4,  -4, -4, -4,
+		 -4,  -4, -4, -4,
+		-19, -16,  0,  0
+	);
+
+	// Translate values 0..63 to the Base64 alphabet. There are five sets:
+	// #  From      To         Abs    Index  Characters
+	// 0  [0..25]   [65..90]   +65        0  ABCDEFGHIJKLMNOPQRSTUVWXYZ
+	// 1  [26..51]  [97..122]  +71        1  abcdefghijklmnopqrstuvwxyz
+	// 2  [52..61]  [48..57]    -4  [2..11]  0123456789
+	// 3  [62]      [43]       -19       12  +
+	// 4  [63]      [47]       -16       13  /
+
+	// Create LUT indices from the input. The index for range #0 is right,
+	// others are 1 less than expected:
+	__m128i indices = _mm_subs_epu8(in, _mm_set1_epi8(51));
+
+	// mask is 0xFF (-1) for range #[1..4] and 0x00 for range #0:
+	__m128i mask = _mm_cmpgt_epi8(in, _mm_set1_epi8(25));
+
+	// Subtract -1, so add 1 to indices for range #[1..4]. All indices are
+	// now correct:
+	indices = _mm_sub_epi8(indices, mask);
+
+	// Add offsets to input values:
+	return _mm_add_epi8(in, _mm_shuffle_epi8(lut, indices));
+}

3rdparty/base64/lib/codec_choose.c

@@ -0,0 +1,305 @@
+#include <stdbool.h>
+#include <stdint.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+
+#include "../include/libbase64.h"
+#include "codecs.h"
+#include "config.h"
+#include "env.h"
+
+#if (__x86_64__ || __i386__ || _M_X86 || _M_X64)
+  #define BASE64_X86
+  #if (HAVE_SSSE3 || HAVE_SSE41 || HAVE_SSE42 || HAVE_AVX || HAVE_AVX2 || HAVE_AVX512)
+    #define BASE64_X86_SIMD
+  #endif
+#endif
+
+#ifdef BASE64_X86
+#ifdef _MSC_VER
+	#include <intrin.h>
+	#define __cpuid_count(__level, __count, __eax, __ebx, __ecx, __edx) \
+	{						\
+		int info[4];				\
+		__cpuidex(info, __level, __count);	\
+		__eax = info[0];			\
+		__ebx = info[1];			\
+		__ecx = info[2];			\
+		__edx = info[3];			\
+	}
+	#define __cpuid(__level, __eax, __ebx, __ecx, __edx) \
+		__cpuid_count(__level, 0, __eax, __ebx, __ecx, __edx)
+#else
+	#include <cpuid.h>
+	#if HAVE_AVX512 || HAVE_AVX2 || HAVE_AVX
+		#if ((__GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ >= 2) || (__clang_major__ >= 3))
+			static inline uint64_t _xgetbv (uint32_t index)
+			{
+				uint32_t eax, edx;
+				__asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
+				return ((uint64_t)edx << 32) | eax;
+			}
+		#else
+			#error "Platform not supported"
+		#endif
+	#endif
+#endif
+
+#ifndef bit_AVX512vl
+#define bit_AVX512vl (1 << 31)
+#endif
+#ifndef bit_AVX512vbmi
+#define bit_AVX512vbmi (1 << 1)
+#endif
+#ifndef bit_AVX2
+#define bit_AVX2 (1 << 5)
+#endif
+#ifndef bit_SSSE3
+#define bit_SSSE3 (1 << 9)
+#endif
+#ifndef bit_SSE41
+#define bit_SSE41 (1 << 19)
+#endif
+#ifndef bit_SSE42
+#define bit_SSE42 (1 << 20)
+#endif
+#ifndef bit_AVX
+#define bit_AVX (1 << 28)
+#endif
+
+#define bit_XSAVE_XRSTORE (1 << 27)
+
+#ifndef _XCR_XFEATURE_ENABLED_MASK
+#define _XCR_XFEATURE_ENABLED_MASK 0
+#endif
+
+#define _XCR_XMM_AND_YMM_STATE_ENABLED_BY_OS 0x6
+#endif
+
+// Function declarations:
+#define BASE64_CODEC_FUNCS(arch)	\
+	BASE64_ENC_FUNCTION(arch);	\
+	BASE64_DEC_FUNCTION(arch);	\
+
+BASE64_CODEC_FUNCS(avx512)
+BASE64_CODEC_FUNCS(avx2)
+BASE64_CODEC_FUNCS(neon32)
+BASE64_CODEC_FUNCS(neon64)
+BASE64_CODEC_FUNCS(plain)
+BASE64_CODEC_FUNCS(ssse3)
+BASE64_CODEC_FUNCS(sse41)
+BASE64_CODEC_FUNCS(sse42)
+BASE64_CODEC_FUNCS(avx)
+
+static bool
+codec_choose_forced (struct codec *codec, int flags)
+{
+	// If the user wants to use a certain codec,
+	// always allow it, even if the codec is a no-op.
+	// For testing purposes.
+
+	if (!(flags & 0xFFFF)) {
+		return false;
+	}
+
+	if (flags & BASE64_FORCE_AVX2) {
+		codec->enc = base64_stream_encode_avx2;
+		codec->dec = base64_stream_decode_avx2;
+		return true;
+	}
+	if (flags & BASE64_FORCE_NEON32) {
+		codec->enc = base64_stream_encode_neon32;
+		codec->dec = base64_stream_decode_neon32;
+		return true;
+	}
+	if (flags & BASE64_FORCE_NEON64) {
+		codec->enc = base64_stream_encode_neon64;
+		codec->dec = base64_stream_decode_neon64;
+		return true;
+	}
+	if (flags & BASE64_FORCE_PLAIN) {
+		codec->enc = base64_stream_encode_plain;
+		codec->dec = base64_stream_decode_plain;
+		return true;
+	}
+	if (flags & BASE64_FORCE_SSSE3) {
+		codec->enc = base64_stream_encode_ssse3;
+		codec->dec = base64_stream_decode_ssse3;
+		return true;
+	}
+	if (flags & BASE64_FORCE_SSE41) {
+		codec->enc = base64_stream_encode_sse41;
+		codec->dec = base64_stream_decode_sse41;
+		return true;
+	}
+	if (flags & BASE64_FORCE_SSE42) {
+		codec->enc = base64_stream_encode_sse42;
+		codec->dec = base64_stream_decode_sse42;
+		return true;
+	}
+	if (flags & BASE64_FORCE_AVX) {
+		codec->enc = base64_stream_encode_avx;
+		codec->dec = base64_stream_decode_avx;
+		return true;
+	}
+	if (flags & BASE64_FORCE_AVX512) {
+		codec->enc = base64_stream_encode_avx512;
+		codec->dec = base64_stream_decode_avx512;
+		return true;
+	}
+	return false;
+}
+
+static bool
+codec_choose_arm (struct codec *codec)
+{
+#if (defined(__ARM_NEON__) || defined(__ARM_NEON)) && ((defined(__aarch64__) && HAVE_NEON64) || HAVE_NEON32)
+
+	// Unfortunately there is no portable way to check for NEON
+	// support at runtime from userland in the same way that x86
+	// has cpuid, so just stick to the compile-time configuration:
+
+	#if defined(__aarch64__) && HAVE_NEON64
+	codec->enc = base64_stream_encode_neon64;
+	codec->dec = base64_stream_decode_neon64;
+	#else
+	codec->enc = base64_stream_encode_neon32;
+	codec->dec = base64_stream_decode_neon32;
+	#endif
+
+	return true;
+
+#else
+	(void)codec;
+	return false;
+#endif
+}
+
+static bool
+codec_choose_x86 (struct codec *codec)
+{
+#ifdef BASE64_X86_SIMD
+
+	unsigned int eax, ebx = 0, ecx = 0, edx;
+	unsigned int max_level;
+
+	#ifdef _MSC_VER
+	int info[4];
+	__cpuidex(info, 0, 0);
+	max_level = info[0];
+	#else
+	max_level = __get_cpuid_max(0, NULL);
+	#endif
+
+	#if HAVE_AVX512 || HAVE_AVX2 || HAVE_AVX
+	// Check for AVX/AVX2/AVX512 support:
+	// Checking for AVX requires 3 things:
+	// 1) CPUID indicates that the OS uses XSAVE and XRSTORE instructions
+	//    (allowing saving YMM registers on context switch)
+	// 2) CPUID indicates support for AVX
+	// 3) XGETBV indicates the AVX registers will be saved and restored on
+	//    context switch
+	//
+	// Note that XGETBV is only available on 686 or later CPUs, so the
+	// instruction needs to be conditionally run.
+	if (max_level >= 1) {
+		__cpuid_count(1, 0, eax, ebx, ecx, edx);
+		if (ecx & bit_XSAVE_XRSTORE) {
+			uint64_t xcr_mask;
+			xcr_mask = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
+			if ((xcr_mask & _XCR_XMM_AND_YMM_STATE_ENABLED_BY_OS) == _XCR_XMM_AND_YMM_STATE_ENABLED_BY_OS) { // check multiple bits at once
+				#if HAVE_AVX512
+				if (max_level >= 7) {
+					__cpuid_count(7, 0, eax, ebx, ecx, edx);
+					if ((ebx & bit_AVX512vl) && (ecx & bit_AVX512vbmi)) {
+						codec->enc = base64_stream_encode_avx512;
+						codec->dec = base64_stream_decode_avx512;
+						return true;
+					}
+				}
+				#endif
+				#if HAVE_AVX2
+				if (max_level >= 7) {
+					__cpuid_count(7, 0, eax, ebx, ecx, edx);
+					if (ebx & bit_AVX2) {
+						codec->enc = base64_stream_encode_avx2;
+						codec->dec = base64_stream_decode_avx2;
+						return true;
+					}
+				}
+				#endif
+				#if HAVE_AVX
+				__cpuid_count(1, 0, eax, ebx, ecx, edx);
+				if (ecx & bit_AVX) {
+					codec->enc = base64_stream_encode_avx;
+					codec->dec = base64_stream_decode_avx;
+					return true;
+				}
+				#endif
+			}
+		}
+	}
+	#endif
+
+	#if HAVE_SSE42
+	// Check for SSE42 support:
+	if (max_level >= 1) {
+		__cpuid(1, eax, ebx, ecx, edx);
+		if (ecx & bit_SSE42) {
+			codec->enc = base64_stream_encode_sse42;
+			codec->dec = base64_stream_decode_sse42;
+			return true;
+		}
+	}
+	#endif
+
+	#if HAVE_SSE41
+	// Check for SSE41 support:
+	if (max_level >= 1) {
+		__cpuid(1, eax, ebx, ecx, edx);
+		if (ecx & bit_SSE41) {
+			codec->enc = base64_stream_encode_sse41;
+			codec->dec = base64_stream_decode_sse41;
+			return true;
+		}
+	}
+	#endif
+
+	#if HAVE_SSSE3
+	// Check for SSSE3 support:
+	if (max_level >= 1) {
+		__cpuid(1, eax, ebx, ecx, edx);
+		if (ecx & bit_SSSE3) {
+			codec->enc = base64_stream_encode_ssse3;
+			codec->dec = base64_stream_decode_ssse3;
+			return true;
+		}
+	}
+	#endif
+
+#else
+	(void)codec;
+#endif
+
+	return false;
+}
+
+void
+codec_choose (struct codec *codec, int flags)
+{
+	// User forced a codec:
+	if (codec_choose_forced(codec, flags)) {
+		return;
+	}
+
+	// Runtime feature detection:
+	if (codec_choose_arm(codec)) {
+		return;
+	}
+	if (codec_choose_x86(codec)) {
+		return;
+	}
+	codec->enc = base64_stream_encode_plain;
+	codec->dec = base64_stream_decode_plain;
+}

3rdparty/base64/lib/codecs.h

@@ -0,0 +1,65 @@
+#include <stdint.h>
+#include <stddef.h>
+
+#include "../include/libbase64.h"
+#include "config.h"
+
+// Function parameters for encoding functions:
+#define BASE64_ENC_PARAMS			\
+	( struct base64_state	*state		\
+	, const char		*src		\
+	, size_t		 srclen		\
+	, char			*out		\
+	, size_t		*outlen		\
+	)
+
+// Function parameters for decoding functions:
+#define BASE64_DEC_PARAMS			\
+	( struct base64_state	*state		\
+	, const char		*src		\
+	, size_t		 srclen		\
+	, char			*out		\
+	, size_t		*outlen		\
+	)
+
+// Function signature for encoding functions:
+#define BASE64_ENC_FUNCTION(arch)		\
+	void					\
+	base64_stream_encode_ ## arch		\
+	BASE64_ENC_PARAMS
+
+// Function signature for decoding functions:
+#define BASE64_DEC_FUNCTION(arch)		\
+	int					\
+	base64_stream_decode_ ## arch		\
+	BASE64_DEC_PARAMS
+
+// Cast away unused variable, silence compiler:
+#define UNUSED(x)		((void)(x))
+
+// Stub function when encoder arch unsupported:
+#define BASE64_ENC_STUB				\
+	UNUSED(state);				\
+	UNUSED(src);				\
+	UNUSED(srclen);				\
+	UNUSED(out);				\
+						\
+	*outlen = 0;
+
+// Stub function when decoder arch unsupported:
+#define BASE64_DEC_STUB				\
+	UNUSED(state);				\
+	UNUSED(src);				\
+	UNUSED(srclen);				\
+	UNUSED(out);				\
+	UNUSED(outlen);				\
+						\
+	return -1;
+
+struct codec
+{
+	void (* enc) BASE64_ENC_PARAMS;
+	int  (* dec) BASE64_DEC_PARAMS;
+};
+
+extern void codec_choose (struct codec *, int flags);

3rdparty/base64/lib/env.h

@@ -0,0 +1,74 @@
+#ifndef BASE64_ENV_H
+#define BASE64_ENV_H
+
+// This header file contains macro definitions that describe certain aspects of
+// the compile-time environment. Compatibility and portability macros go here.
+
+// Define machine endianness. This is for GCC:
+#if (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#  define BASE64_LITTLE_ENDIAN 1
+#else
+#  define BASE64_LITTLE_ENDIAN 0
+#endif
+
+// This is for Clang:
+#ifdef __LITTLE_ENDIAN__
+#  define BASE64_LITTLE_ENDIAN 1
+#endif
+
+#ifdef __BIG_ENDIAN__
+#  define BASE64_LITTLE_ENDIAN 0
+#endif
+
+// MSVC++ needs intrin.h for _byteswap_uint64 (issue #68):
+#if BASE64_LITTLE_ENDIAN && defined(_MSC_VER)
+#  include <intrin.h>
+#endif
+
+// Endian conversion functions:
+#if BASE64_LITTLE_ENDIAN
+#  ifdef _MSC_VER
+//   Microsoft Visual C++:
+#    define BASE64_HTOBE32(x)	_byteswap_ulong(x)
+#    define BASE64_HTOBE64(x)	_byteswap_uint64(x)
+#  else
+//   GCC and Clang:
+#    define BASE64_HTOBE32(x)	__builtin_bswap32(x)
+#    define BASE64_HTOBE64(x)	__builtin_bswap64(x)
+#  endif
+#else
+// No conversion needed:
+#  define BASE64_HTOBE32(x)	(x)
+#  define BASE64_HTOBE64(x)	(x)
+#endif
+
+// Detect word size:
+#if defined (__x86_64__)
+// This also works for the x32 ABI, which has a 64-bit word size.
+#  define BASE64_WORDSIZE 64
+#elif defined (_INTEGRAL_MAX_BITS)
+#  define BASE64_WORDSIZE _INTEGRAL_MAX_BITS
+#elif defined (__WORDSIZE)
+#  define BASE64_WORDSIZE __WORDSIZE
+#elif defined (__SIZE_WIDTH__)
+#  define BASE64_WORDSIZE __SIZE_WIDTH__
+#else
+#  error BASE64_WORDSIZE_NOT_DEFINED
+#endif
+
+// End-of-file definitions.
+// Almost end-of-file when waiting for the last '=' character:
+#define BASE64_AEOF 1
+// End-of-file when stream end has been reached or invalid input provided:
+#define BASE64_EOF 2
+
+// GCC 7 defaults to issuing a warning for fallthrough in switch statements,
+// unless the fallthrough cases are marked with an attribute. As we use
+// fallthrough deliberately, define an alias for the attribute:
+#if __GNUC__ >= 7
+#  define BASE64_FALLTHROUGH  __attribute__((fallthrough));
+#else
+#  define BASE64_FALLTHROUGH
+#endif
+
+#endif	// BASE64_ENV_H

3rdparty/base64/lib/exports.txt

@@ -0,0 +1,7 @@
+base64_encode
+base64_stream_encode
+base64_stream_encode_init
+base64_stream_encode_final
+base64_decode
+base64_stream_decode
+base64_stream_decode_init

3rdparty/base64/lib/lib.c

@@ -0,0 +1,164 @@
+#include <stdint.h>
+#include <stddef.h>
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+#include "../include/libbase64.h"
+#include "tables/tables.h"
+#include "codecs.h"
+#include "env.h"
+
+// These static function pointers are initialized once when the library is
+// first used, and remain in use for the remaining lifetime of the program.
+// The idea being that CPU features don't change at runtime.
+static struct codec codec = { NULL, NULL };
+
+void
+base64_stream_encode_init (struct base64_state *state, int flags)
+{
+	// If any of the codec flags are set, redo choice:
+	if (codec.enc == NULL || flags & 0xFF) {
+		codec_choose(&codec, flags);
+	}
+	state->eof = 0;
+	state->bytes = 0;
+	state->carry = 0;
+	state->flags = flags;
+}
+
+void
+base64_stream_encode
+	( struct base64_state	*state
+	, const char		*src
+	, size_t		 srclen
+	, char			*out
+	, size_t		*outlen
+	)
+{
+	codec.enc(state, src, srclen, out, outlen);
+}
+
+void
+base64_stream_encode_final
+	( struct base64_state	*state
+	, char			*out
+	, size_t		*outlen
+	)
+{
+	uint8_t *o = (uint8_t *)out;
+
+	if (state->bytes == 1) {
+		*o++ = base64_table_enc_6bit[state->carry];
+		*o++ = '=';
+		*o++ = '=';
+		*outlen = 3;
+		return;
+	}
+	if (state->bytes == 2) {
+		*o++ = base64_table_enc_6bit[state->carry];
+		*o++ = '=';
+		*outlen = 2;
+		return;
+	}
+	*outlen = 0;
+}
+
+void
+base64_stream_decode_init (struct base64_state *state, int flags)
+{
+	// If any of the codec flags are set, redo choice:
+	if (codec.dec == NULL || flags & 0xFFFF) {
+		codec_choose(&codec, flags);
+	}
+	state->eof = 0;
+	state->bytes = 0;
+	state->carry = 0;
+	state->flags = flags;
+}
+
+int
+base64_stream_decode
+	( struct base64_state	*state
+	, const char		*src
+	, size_t		 srclen
+	, char			*out
+	, size_t		*outlen
+	)
+{
+	return codec.dec(state, src, srclen, out, outlen);
+}
+
+#ifdef _OPENMP
+
+	// Due to the overhead of initializing OpenMP and creating a team of
+	// threads, we require the data length to be larger than a threshold:
+	#define OMP_THRESHOLD 20000
+
+	// Conditionally include OpenMP-accelerated codec implementations:
+	#include "lib_openmp.c"
+#endif
+
+void
+base64_encode
+	( const char	*src
+	, size_t	 srclen
+	, char		*out
+	, size_t	*outlen
+	, int		 flags
+	)
+{
+	size_t s;
+	size_t t;
+	struct base64_state state;
+
+	#ifdef _OPENMP
+	if (srclen >= OMP_THRESHOLD) {
+		base64_encode_openmp(src, srclen, out, outlen, flags);
+		return;
+	}
+	#endif
+
+	// Init the stream reader:
+	base64_stream_encode_init(&state, flags);
+
+	// Feed the whole string to the stream reader:
+	base64_stream_encode(&state, src, srclen, out, &s);
+
+	// Finalize the stream by writing trailer if any:
+	base64_stream_encode_final(&state, out + s, &t);
+
+	// Final output length is stream length plus tail:
+	*outlen = s + t;
+}
+
+int
+base64_decode
+	( const char	*src
+	, size_t	 srclen
+	, char		*out
+	, size_t	*outlen
+	, int		 flags
+	)
+{
+	int ret;
+	struct base64_state state;
+
+	#ifdef _OPENMP
+	if (srclen >= OMP_THRESHOLD) {
+		return base64_decode_openmp(src, srclen, out, outlen, flags);
+	}
+	#endif
+
+	// Init the stream reader:
+	base64_stream_decode_init(&state, flags);
+
+	// Feed the whole string to the stream reader:
+	ret = base64_stream_decode(&state, src, srclen, out, outlen);
+
+	// If when decoding a whole block, we're still waiting for input then fail:
+	if (ret && (state.bytes == 0)) {
+		return ret;
+	}
+	return 0;
+}

3rdparty/base64/lib/lib_openmp.c

@@ -0,0 +1,149 @@
+// This code makes some assumptions on the implementation of
+// base64_stream_encode_init(), base64_stream_encode() and base64_stream_decode().
+// Basically these assumptions boil down to that when breaking the src into
+// parts, out parts can be written without side effects.
+// This is met when:
+// 1) base64_stream_encode() and base64_stream_decode() don't use globals;
+// 2) the shared variables src and out are not read or written outside of the
+//    bounds of their parts, i.e.  when base64_stream_encode() reads a multiple
+//    of 3 bytes, it must write no more then a multiple of 4 bytes, not even
+//    temporarily;
+// 3) the state flag can be discarded after base64_stream_encode() and
+//    base64_stream_decode() on the parts.
+
+static inline void
+base64_encode_openmp
+	( const char	*src
+	, size_t	 srclen
+	, char		*out
+	, size_t	*outlen
+	, int		 flags
+	)
+{
+	size_t s;
+	size_t t;
+	size_t sum = 0, len, last_len;
+	struct base64_state state, initial_state;
+	int num_threads, i;
+
+	// Request a number of threads but not necessarily get them:
+	#pragma omp parallel
+	{
+		// Get the number of threads used from one thread only,
+		// as num_threads is a shared var:
+		#pragma omp single
+		{
+			num_threads = omp_get_num_threads();
+
+			// Split the input string into num_threads parts, each
+			// part a multiple of 3 bytes. The remaining bytes will
+			// be done later:
+			len = srclen / (num_threads * 3);
+			len *= 3;
+			last_len = srclen - num_threads * len;
+
+			// Init the stream reader:
+			base64_stream_encode_init(&state, flags);
+			initial_state = state;
+		}
+
+		// Single has an implicit barrier for all threads to wait here
+		// for the above to complete:
+		#pragma omp for firstprivate(state) private(s) reduction(+:sum) schedule(static,1)
+		for (i = 0; i < num_threads; i++)
+		{
+			// Feed each part of the string to the stream reader:
+			base64_stream_encode(&state, src + i * len, len, out + i * len * 4 / 3, &s);
+			sum += s;
+		}
+	}
+
+	// As encoding should never fail and we encode an exact multiple
+	// of 3 bytes, we can discard state:
+	state = initial_state;
+
+	// Encode the remaining bytes:
+	base64_stream_encode(&state, src + num_threads * len, last_len, out + num_threads * len * 4 / 3, &s);
+
+	// Finalize the stream by writing trailer if any:
+	base64_stream_encode_final(&state, out + num_threads * len * 4 / 3 + s, &t);
+
+	// Final output length is stream length plus tail:
+	sum += s + t;
+	*outlen = sum;
+}
+
+static inline int
+base64_decode_openmp
+	( const char	*src
+	, size_t	 srclen
+	, char		*out
+	, size_t	*outlen
+	, int		 flags
+	)
+{
+	int num_threads, result = 0, i;
+	size_t sum = 0, len, last_len, s;
+	struct base64_state state, initial_state;
+
+	// Request a number of threads but not necessarily get them:
+	#pragma omp parallel
+	{
+		// Get the number of threads used from one thread only,
+		// as num_threads is a shared var:
+		#pragma omp single
+		{
+			num_threads = omp_get_num_threads();
+
+			// Split the input string into num_threads parts, each
+			// part a multiple of 4 bytes. The remaining bytes will
+			// be done later:
+			len = srclen / (num_threads * 4);
+			len *= 4;
+			last_len = srclen - num_threads * len;
+
+			// Init the stream reader:
+			base64_stream_decode_init(&state, flags);
+
+			initial_state = state;
+		}
+
+		// Single has an implicit barrier to wait here for the above to
+		// complete:
+		#pragma omp for firstprivate(state) private(s) reduction(+:sum, result) schedule(static,1)
+		for (i = 0; i < num_threads; i++)
+		{
+			int this_result;
+
+			// Feed each part of the string to the stream reader:
+			this_result = base64_stream_decode(&state, src + i * len, len, out + i * len * 3 / 4, &s);
+			sum += s;
+			result += this_result;
+		}
+	}
+
+	// If `result' equals `-num_threads', then all threads returned -1,
+	// indicating that the requested codec is not available:
+	if (result == -num_threads) {
+		return -1;
+	}
+
+	// If `result' does not equal `num_threads', then at least one of the
+	// threads hit a decode error:
+	if (result != num_threads) {
+		return 0;
+	}
+
+	// So far so good, now decode whatever remains in the buffer. Reuse the
+	// initial state, since we are at a 4-byte boundary:
+	state = initial_state;
+	result = base64_stream_decode(&state, src + num_threads * len, last_len, out + num_threads * len * 3 / 4, &s);
+	sum += s;
+	*outlen = sum;
+
+	// If when decoding a whole block, we're still waiting for input then fail:
+	if (result && (state.bytes == 0)) {
+		return result;
+	}
+	return 0;
+}

3rdparty/base64/lib/tables/.gitignore

@@ -0,0 +1 @@
+table_generator

3rdparty/base64/lib/tables/Makefile

@@ -0,0 +1,17 @@
+.PHONY: all clean
+
+TARGETS := table_dec_32bit.h table_enc_12bit.h table_generator
+
+all: $(TARGETS)
+
+clean:
+	$(RM) $(TARGETS)
+
+table_dec_32bit.h: table_generator
+	./$^ > $@
+
+table_enc_12bit.h: table_enc_12bit.py
+	./$^ > $@
+
+table_generator: table_generator.c
+	$(CC) $(CFLAGS) -o $@ $^

3rdparty/base64/lib/tables/table_dec_32bit.h

@@ -0,0 +1,393 @@
+#include <stdint.h>
+#define CHAR62 '+'
+#define CHAR63 '/'
+#define CHARPAD '='
+
+
+#if BASE64_LITTLE_ENDIAN
+
+
+/* SPECIAL DECODE TABLES FOR LITTLE ENDIAN (INTEL) CPUS */
+
+const uint32_t base64_table_dec_32bit_d0[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x000000f8, 0xffffffff, 0xffffffff, 0xffffffff, 0x000000fc,
+0x000000d0, 0x000000d4, 0x000000d8, 0x000000dc, 0x000000e0, 0x000000e4,
+0x000000e8, 0x000000ec, 0x000000f0, 0x000000f4, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00000004, 0x00000008, 0x0000000c, 0x00000010, 0x00000014, 0x00000018,
+0x0000001c, 0x00000020, 0x00000024, 0x00000028, 0x0000002c, 0x00000030,
+0x00000034, 0x00000038, 0x0000003c, 0x00000040, 0x00000044, 0x00000048,
+0x0000004c, 0x00000050, 0x00000054, 0x00000058, 0x0000005c, 0x00000060,
+0x00000064, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x00000068, 0x0000006c, 0x00000070, 0x00000074, 0x00000078,
+0x0000007c, 0x00000080, 0x00000084, 0x00000088, 0x0000008c, 0x00000090,
+0x00000094, 0x00000098, 0x0000009c, 0x000000a0, 0x000000a4, 0x000000a8,
+0x000000ac, 0x000000b0, 0x000000b4, 0x000000b8, 0x000000bc, 0x000000c0,
+0x000000c4, 0x000000c8, 0x000000cc, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+const uint32_t base64_table_dec_32bit_d1[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x0000e003, 0xffffffff, 0xffffffff, 0xffffffff, 0x0000f003,
+0x00004003, 0x00005003, 0x00006003, 0x00007003, 0x00008003, 0x00009003,
+0x0000a003, 0x0000b003, 0x0000c003, 0x0000d003, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00001000, 0x00002000, 0x00003000, 0x00004000, 0x00005000, 0x00006000,
+0x00007000, 0x00008000, 0x00009000, 0x0000a000, 0x0000b000, 0x0000c000,
+0x0000d000, 0x0000e000, 0x0000f000, 0x00000001, 0x00001001, 0x00002001,
+0x00003001, 0x00004001, 0x00005001, 0x00006001, 0x00007001, 0x00008001,
+0x00009001, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x0000a001, 0x0000b001, 0x0000c001, 0x0000d001, 0x0000e001,
+0x0000f001, 0x00000002, 0x00001002, 0x00002002, 0x00003002, 0x00004002,
+0x00005002, 0x00006002, 0x00007002, 0x00008002, 0x00009002, 0x0000a002,
+0x0000b002, 0x0000c002, 0x0000d002, 0x0000e002, 0x0000f002, 0x00000003,
+0x00001003, 0x00002003, 0x00003003, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+const uint32_t base64_table_dec_32bit_d2[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x00800f00, 0xffffffff, 0xffffffff, 0xffffffff, 0x00c00f00,
+0x00000d00, 0x00400d00, 0x00800d00, 0x00c00d00, 0x00000e00, 0x00400e00,
+0x00800e00, 0x00c00e00, 0x00000f00, 0x00400f00, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00400000, 0x00800000, 0x00c00000, 0x00000100, 0x00400100, 0x00800100,
+0x00c00100, 0x00000200, 0x00400200, 0x00800200, 0x00c00200, 0x00000300,
+0x00400300, 0x00800300, 0x00c00300, 0x00000400, 0x00400400, 0x00800400,
+0x00c00400, 0x00000500, 0x00400500, 0x00800500, 0x00c00500, 0x00000600,
+0x00400600, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x00800600, 0x00c00600, 0x00000700, 0x00400700, 0x00800700,
+0x00c00700, 0x00000800, 0x00400800, 0x00800800, 0x00c00800, 0x00000900,
+0x00400900, 0x00800900, 0x00c00900, 0x00000a00, 0x00400a00, 0x00800a00,
+0x00c00a00, 0x00000b00, 0x00400b00, 0x00800b00, 0x00c00b00, 0x00000c00,
+0x00400c00, 0x00800c00, 0x00c00c00, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+const uint32_t base64_table_dec_32bit_d3[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x003e0000, 0xffffffff, 0xffffffff, 0xffffffff, 0x003f0000,
+0x00340000, 0x00350000, 0x00360000, 0x00370000, 0x00380000, 0x00390000,
+0x003a0000, 0x003b0000, 0x003c0000, 0x003d0000, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00010000, 0x00020000, 0x00030000, 0x00040000, 0x00050000, 0x00060000,
+0x00070000, 0x00080000, 0x00090000, 0x000a0000, 0x000b0000, 0x000c0000,
+0x000d0000, 0x000e0000, 0x000f0000, 0x00100000, 0x00110000, 0x00120000,
+0x00130000, 0x00140000, 0x00150000, 0x00160000, 0x00170000, 0x00180000,
+0x00190000, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x001a0000, 0x001b0000, 0x001c0000, 0x001d0000, 0x001e0000,
+0x001f0000, 0x00200000, 0x00210000, 0x00220000, 0x00230000, 0x00240000,
+0x00250000, 0x00260000, 0x00270000, 0x00280000, 0x00290000, 0x002a0000,
+0x002b0000, 0x002c0000, 0x002d0000, 0x002e0000, 0x002f0000, 0x00300000,
+0x00310000, 0x00320000, 0x00330000, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+#else
+
+
+/* SPECIAL DECODE TABLES FOR BIG ENDIAN (IBM/MOTOROLA/SUN) CPUS */
+
+const uint32_t base64_table_dec_32bit_d0[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xf8000000, 0xffffffff, 0xffffffff, 0xffffffff, 0xfc000000,
+0xd0000000, 0xd4000000, 0xd8000000, 0xdc000000, 0xe0000000, 0xe4000000,
+0xe8000000, 0xec000000, 0xf0000000, 0xf4000000, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x04000000, 0x08000000, 0x0c000000, 0x10000000, 0x14000000, 0x18000000,
+0x1c000000, 0x20000000, 0x24000000, 0x28000000, 0x2c000000, 0x30000000,
+0x34000000, 0x38000000, 0x3c000000, 0x40000000, 0x44000000, 0x48000000,
+0x4c000000, 0x50000000, 0x54000000, 0x58000000, 0x5c000000, 0x60000000,
+0x64000000, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x68000000, 0x6c000000, 0x70000000, 0x74000000, 0x78000000,
+0x7c000000, 0x80000000, 0x84000000, 0x88000000, 0x8c000000, 0x90000000,
+0x94000000, 0x98000000, 0x9c000000, 0xa0000000, 0xa4000000, 0xa8000000,
+0xac000000, 0xb0000000, 0xb4000000, 0xb8000000, 0xbc000000, 0xc0000000,
+0xc4000000, 0xc8000000, 0xcc000000, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+const uint32_t base64_table_dec_32bit_d1[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x03e00000, 0xffffffff, 0xffffffff, 0xffffffff, 0x03f00000,
+0x03400000, 0x03500000, 0x03600000, 0x03700000, 0x03800000, 0x03900000,
+0x03a00000, 0x03b00000, 0x03c00000, 0x03d00000, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00100000, 0x00200000, 0x00300000, 0x00400000, 0x00500000, 0x00600000,
+0x00700000, 0x00800000, 0x00900000, 0x00a00000, 0x00b00000, 0x00c00000,
+0x00d00000, 0x00e00000, 0x00f00000, 0x01000000, 0x01100000, 0x01200000,
+0x01300000, 0x01400000, 0x01500000, 0x01600000, 0x01700000, 0x01800000,
+0x01900000, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x01a00000, 0x01b00000, 0x01c00000, 0x01d00000, 0x01e00000,
+0x01f00000, 0x02000000, 0x02100000, 0x02200000, 0x02300000, 0x02400000,
+0x02500000, 0x02600000, 0x02700000, 0x02800000, 0x02900000, 0x02a00000,
+0x02b00000, 0x02c00000, 0x02d00000, 0x02e00000, 0x02f00000, 0x03000000,
+0x03100000, 0x03200000, 0x03300000, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+const uint32_t base64_table_dec_32bit_d2[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x000f8000, 0xffffffff, 0xffffffff, 0xffffffff, 0x000fc000,
+0x000d0000, 0x000d4000, 0x000d8000, 0x000dc000, 0x000e0000, 0x000e4000,
+0x000e8000, 0x000ec000, 0x000f0000, 0x000f4000, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00004000, 0x00008000, 0x0000c000, 0x00010000, 0x00014000, 0x00018000,
+0x0001c000, 0x00020000, 0x00024000, 0x00028000, 0x0002c000, 0x00030000,
+0x00034000, 0x00038000, 0x0003c000, 0x00040000, 0x00044000, 0x00048000,
+0x0004c000, 0x00050000, 0x00054000, 0x00058000, 0x0005c000, 0x00060000,
+0x00064000, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x00068000, 0x0006c000, 0x00070000, 0x00074000, 0x00078000,
+0x0007c000, 0x00080000, 0x00084000, 0x00088000, 0x0008c000, 0x00090000,
+0x00094000, 0x00098000, 0x0009c000, 0x000a0000, 0x000a4000, 0x000a8000,
+0x000ac000, 0x000b0000, 0x000b4000, 0x000b8000, 0x000bc000, 0x000c0000,
+0x000c4000, 0x000c8000, 0x000cc000, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+const uint32_t base64_table_dec_32bit_d3[256] = {
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x00003e00, 0xffffffff, 0xffffffff, 0xffffffff, 0x00003f00,
+0x00003400, 0x00003500, 0x00003600, 0x00003700, 0x00003800, 0x00003900,
+0x00003a00, 0x00003b00, 0x00003c00, 0x00003d00, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0x00000000,
+0x00000100, 0x00000200, 0x00000300, 0x00000400, 0x00000500, 0x00000600,
+0x00000700, 0x00000800, 0x00000900, 0x00000a00, 0x00000b00, 0x00000c00,
+0x00000d00, 0x00000e00, 0x00000f00, 0x00001000, 0x00001100, 0x00001200,
+0x00001300, 0x00001400, 0x00001500, 0x00001600, 0x00001700, 0x00001800,
+0x00001900, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0x00001a00, 0x00001b00, 0x00001c00, 0x00001d00, 0x00001e00,
+0x00001f00, 0x00002000, 0x00002100, 0x00002200, 0x00002300, 0x00002400,
+0x00002500, 0x00002600, 0x00002700, 0x00002800, 0x00002900, 0x00002a00,
+0x00002b00, 0x00002c00, 0x00002d00, 0x00002e00, 0x00002f00, 0x00003000,
+0x00003100, 0x00003200, 0x00003300, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff,
+0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+
+#endif

3rdparty/base64/lib/tables/table_enc_12bit.py

@@ -0,0 +1,45 @@
+#!/usr/bin/python3
+
+def tr(x):
+    """Translate a 6-bit value to the Base64 alphabet."""
+    s = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ' \
+      + 'abcdefghijklmnopqrstuvwxyz' \
+      + '0123456789' \
+      + '+/'
+    return ord(s[x])
+
+def table(fn):
+    """Generate a 12-bit lookup table."""
+    ret = []
+    for n in range(0, 2**12):
+        pre = "\n\t" if n % 8 == 0 else " "
+        pre = "\t" if n == 0 else pre
+        ret.append("{}0x{:04X}U,".format(pre, fn(n)))
+    return "".join(ret)
+
+def table_be():
+    """Generate a 12-bit big-endian lookup table."""
+    return table(lambda n: (tr(n & 0x3F) << 0) | (tr(n >> 6) << 8))
+
+def table_le():
+    """Generate a 12-bit little-endian lookup table."""
+    return table(lambda n: (tr(n >> 6) << 0) | (tr(n & 0x3F) << 8))
+
+def main():
+    """Entry point."""
+    lines = [
+        "#include <stdint.h>",
+        "",
+        "const uint16_t base64_table_enc_12bit[] = {",
+        "#if BASE64_LITTLE_ENDIAN",
+        table_le(),
+        "#else",
+        table_be(),
+        "#endif",
+        "};"
+    ]
+    for line in lines:
+        print(line)
+
+if __name__ == "__main__":
+    main()

3rdparty/base64/lib/tables/table_generator.c

@@ -0,0 +1,184 @@
+/**
+ *
+ * Copyright 2005, 2006 Nick Galbreath -- nickg [at] modp [dot] com
+ * Copyright 2017 Matthieu Darbois
+ * All rights reserved.
+ *
+ * http://modp.com/release/base64
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ * - Redistributions of source code must retain the above copyright notice,
+ *  this list of conditions and the following disclaimer.
+ *
+ * - Redistributions in binary form must reproduce the above copyright
+ *   notice, this list of conditions and the following disclaimer in the
+ *   documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+ * IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+ * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+ * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+/****************************/
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <inttypes.h>
+
+static uint8_t b64chars[64] = {
+	'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M',
+	'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z',
+	'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
+	'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z',
+	'0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '+', '/'
+};
+
+static uint8_t padchar = '=';
+
+static void printStart(void)
+{
+	printf("#include <stdint.h>\n");
+	printf("#define CHAR62 '%c'\n", b64chars[62]);
+	printf("#define CHAR63 '%c'\n", b64chars[63]);
+	printf("#define CHARPAD '%c'\n", padchar);
+}
+
+static void clearDecodeTable(uint32_t* ary)
+{
+	int i = 0;
+	for (i = 0; i < 256; ++i) {
+		ary[i] = 0xFFFFFFFF;
+	}
+}
+
+/* dump uint32_t as hex digits */
+void uint32_array_to_c_hex(const uint32_t* ary, size_t sz, const char* name)
+{
+	size_t i = 0;
+
+	printf("const uint32_t %s[%d] = {\n", name, (int)sz);
+	for (;;) {
+		printf("0x%08" PRIx32, ary[i]);
+		++i;
+		if (i == sz)
+			break;
+		if (i % 6 == 0) {
+			printf(",\n");
+		} else {
+			printf(", ");
+		}
+	}
+	printf("\n};\n");
+}
+
+int main(int argc, char** argv)
+{
+	uint32_t x;
+	uint32_t i = 0;
+	uint32_t ary[256];
+
+	/*  over-ride standard alphabet */
+	if (argc == 2) {
+		uint8_t* replacements = (uint8_t*)argv[1];
+		if (strlen((char*)replacements) != 3) {
+			fprintf(stderr, "input must be a string of 3 characters '-', '.' or '_'\n");
+			exit(1);
+		}
+		fprintf(stderr, "fusing '%s' as replacements in base64 encoding\n", replacements);
+		b64chars[62] = replacements[0];
+		b64chars[63] = replacements[1];
+		padchar = replacements[2];
+	}
+
+	printStart();
+
+	printf("\n\n#if BASE64_LITTLE_ENDIAN\n");
+
+	printf("\n\n/* SPECIAL DECODE TABLES FOR LITTLE ENDIAN (INTEL) CPUS */\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = i << 2;
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d0");
+	printf("\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = ((i & 0x30) >> 4) | ((i & 0x0F) << 12);
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d1");
+	printf("\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = ((i & 0x03) << 22) | ((i & 0x3c) << 6);
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d2");
+	printf("\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = i << 16;
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d3");
+	printf("\n\n");
+
+	printf("#else\n");
+
+	printf("\n\n/* SPECIAL DECODE TABLES FOR BIG ENDIAN (IBM/MOTOROLA/SUN) CPUS */\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = i << 26;
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d0");
+	printf("\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = i << 20;
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d1");
+	printf("\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = i << 14;
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d2");
+	printf("\n\n");
+
+	clearDecodeTable(ary);
+	for (i = 0; i < 64; ++i) {
+		x = b64chars[i];
+		ary[x] = i << 8;
+	}
+	uint32_array_to_c_hex(ary, sizeof(ary) / sizeof(uint32_t), "base64_table_dec_32bit_d3");
+	printf("\n\n");
+
+	printf("#endif\n");
+
+	return 0;
+}

3rdparty/base64/lib/tables/tables.c

@@ -0,0 +1,40 @@
+#include "tables.h"
+
+const uint8_t
+base64_table_enc_6bit[] =
+	"ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+	"abcdefghijklmnopqrstuvwxyz"
+	"0123456789"
+	"+/";
+
+// In the lookup table below, note that the value for '=' (character 61) is
+// 254, not 255. This character is used for in-band signaling of the end of
+// the datastream, and we will use that later. The characters A-Z, a-z, 0-9
+// and + / are mapped to their "decoded" values. The other bytes all map to
+// the value 255, which flags them as "invalid input".
+
+const uint8_t
+base64_table_dec_8bit[] =
+{
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,		//   0..15
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,		//  16..31
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,  62, 255, 255, 255,  63,		//  32..47
+	 52,  53,  54,  55,  56,  57,  58,  59,  60,  61, 255, 255, 255, 254, 255, 255,		//  48..63
+	255,   0,   1,   2,   3,   4,   5,   6,   7,   8,   9,  10,  11,  12,  13,  14,		//  64..79
+	 15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25, 255, 255, 255, 255, 255,		//  80..95
+	255,  26,  27,  28,  29,  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,		//  96..111
+	 41,  42,  43,  44,  45,  46,  47,  48,  49,  50,  51, 255, 255, 255, 255, 255,		// 112..127
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,		// 128..143
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+};
+
+#if BASE64_WORDSIZE >= 32
+#  include "table_dec_32bit.h"
+#  include "table_enc_12bit.h"
+#endif

3rdparty/base64/lib/tables/tables.h

@@ -0,0 +1,23 @@
+#ifndef BASE64_TABLES_H
+#define BASE64_TABLES_H
+
+#include <stdint.h>
+
+#include "../env.h"
+
+// These tables are used by all codecs for fallback plain encoding/decoding:
+extern const uint8_t base64_table_enc_6bit[];
+extern const uint8_t base64_table_dec_8bit[];
+
+// These tables are used for the 32-bit and 64-bit generic decoders:
+#if BASE64_WORDSIZE >= 32
+extern const uint32_t base64_table_dec_32bit_d0[];
+extern const uint32_t base64_table_dec_32bit_d1[];
+extern const uint32_t base64_table_dec_32bit_d2[];
+extern const uint32_t base64_table_dec_32bit_d3[];
+
+// This table is used by the 32 and 64-bit generic encoders:
+extern const uint16_t base64_table_enc_12bit[];
+#endif
+
+#endif	// BASE64_TABLES_H

3rdparty/uthash.h

@@ -1,5 +1,5 @@
 /*
-Copyright (c) 2003-2021, Troy D. Hanson     http://troydhanson.github.io/uthash/
+Copyright (c) 2003-2022, Troy D. Hanson  https://troydhanson.github.io/uthash/
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
@@ -51,6 +51,8 @@ typedef unsigned char uint8_t;
 #else                   /* VS2008 or older (or VS2010 in C mode) */
 #define NO_DECLTYPE
 #endif
+#elif defined(__MCST__)  /* Elbrus C Compiler */
+#define DECLTYPE(x) (__typeof(x))
 #elif defined(__BORLANDC__) || defined(__ICCARM__) || defined(__LCC__) || defined(__WATCOMC__)
 #define NO_DECLTYPE
 #else                   /* GNU, Sun and other compilers */
@@ -450,7 +452,7 @@ do {
 
 #define HASH_DELETE_HH(hh,head,delptrhh)                                         \
 do {                                                                             \
-  struct UT_hash_handle *_hd_hh_del = (delptrhh);                                \
+  const struct UT_hash_handle *_hd_hh_del = (delptrhh);                          \
   if ((_hd_hh_del->prev == NULL) && (_hd_hh_del->next == NULL)) {                \
     HASH_BLOOM_FREE((head)->hh.tbl);                                             \
     uthash_free((head)->hh.tbl->buckets,                                         \
@@ -593,7 +595,9 @@ do {
 
 
 /* SAX/FNV/OAT/JEN hash functions are macro variants of those listed at
- * http://eternallyconfuzzled.com/tuts/algorithms/jsw_tut_hashing.aspx */
+ * http://eternallyconfuzzled.com/tuts/algorithms/jsw_tut_hashing.aspx
+ * (archive link: https://archive.is/Ivcan )
+ */
 #define HASH_SAX(key,keylen,hashv)                                               \
 do {                                                                             \
   unsigned _sx_i;                                                                \

build-terminfo

@@ -27,6 +27,7 @@ def compile_terminfo(base):
         os.makedirs(odir, exist_ok=True)
         ofile = os.path.join(odir, xterm_kitty)
         shutil.move(tfile, ofile)
+    return ofile
 
 
 def generate_terminfo():
@@ -46,7 +47,14 @@ def generate_terminfo():
     with open('terminfo/kitty.termcap', 'w') as f:
         f.write(tcap)
 
-    compile_terminfo(os.path.join(base, 'terminfo'))
+    dbfile = compile_terminfo(os.path.join(base, 'terminfo'))
+    with open(dbfile, 'rb') as f:
+        data = f.read()
+    with open('kitty/terminfo.h', 'w') as f:
+        print(f'static const uint8_t terminfo_data[{len(data)}] = ''{', file=f)
+        for b in data:
+            print(b, end=', ', file=f)
+        print('};', file=f)
 
 
 if __name__ == '__main__':

bypy/devenv.go

@@ -242,7 +242,9 @@ func dependencies(args []string) {
 	chdir_to_base()
 	nf := flag.NewFlagSet("deps", flag.ExitOnError)
 	docsptr := nf.Bool("for-docs", false, "download the dependencies needed to build the documentation")
-	nf.Parse(args)
+	if err := nf.Parse(args); err != nil {
+		exit(err)
+	}
 	if *docsptr {
 		dependencies_for_docs()
 		fmt.Println("Dependencies needed to generate documentation have been installed. Build docs with ./dev.sh docs")
@@ -323,7 +325,7 @@ func dependencies(args []string) {
 	}); err != nil {
 		exit(err)
 	}
-	fmt.Println(`Dependencies downloaded. Now build kitty with: make develop`)
+	fmt.Println(`Dependencies downloaded. Now build kitty with: ./dev.sh build`)
 }
 
 // }}}
@@ -384,7 +386,9 @@ func docs(args []string) {
 	nf := flag.NewFlagSet("deps", flag.ExitOnError)
 	livereload := nf.Bool("live-reload", false, "build the docs and make them available via s local server with live reloading for ease of development")
 	failwarn := nf.Bool("fail-warn", false, "make warnings fatal when building the docs")
-	nf.Parse(args)
+	if err := nf.Parse(args); err != nil {
+		exit(err)
+	}
 	exe := filepath.Join(root_dir(), "bin", "sphinx-build")
 	aexe := filepath.Join(root_dir(), "bin", "sphinx-autobuild")
 	target := "docs"

bypy/init_env.py

@@ -46,7 +46,7 @@ def run(*args, **extra_env):
     return subprocess.call(list(args), env=env, cwd=cwd)
 
 
-SETUP_CMD = [PYTHON, 'setup.py', '--build-universal-binary']
+SETUP_CMD = [PYTHON, 'setup.py']
 
 
 def build_frozen_launcher(extra_include_dirs):

bypy/sources.json

@@ -250,6 +250,15 @@
         }
     },
 
+    {
+        "name": "simde",
+        "unix": {
+            "filename": "simde-amalgamated-0.7.6.tar.xz",
+            "hash": "sha256:703eac1f2af7de1f7e4aea2286130b98e1addcc0559426e78304c92e2b4eb5e1",
+            "urls": ["https://github.com/simd-everywhere/simde/releases/download/v0.7.6/{filename}"]
+        }
+    },
+
     {
         "name": "wayland",
         "os": "linux",

count-lines-of-code

@@ -5,12 +5,13 @@ import subprocess
 ls_files = subprocess.check_output([ 'git', 'ls-files']).decode('utf-8')
 all_files = set(ls_files.splitlines())
 all_files.discard('')
-cp = subprocess.run(['git', 'check-attr', 'linguist-generated', '--stdin'],
-                    check=True, stdout=subprocess.PIPE, input='\n'.join(all_files).encode('utf-8'))
-for line in cp.stdout.decode().splitlines():
-    if line.endswith(' true'):
-        fname = line.split(':', 1)[0]
-        all_files.discard(fname)
+for attr in ('linguist-generated', 'linguist-vendored'):
+    cp = subprocess.run(['git', 'check-attr', attr, '--stdin'],
+                        check=True, stdout=subprocess.PIPE, input='\n'.join(all_files).encode('utf-8'))
+    for line in cp.stdout.decode().splitlines():
+        if line.endswith(' true'):
+            fname = line.split(':', 1)[0]
+            all_files.discard(fname)
 
 all_files -= {'gen/nerd-fonts-glyphs.txt', 'gen/rowcolumn-diacritics.txt'}
 cp = subprocess.run(['cloc', '--list-file', '-'], input='\n'.join(all_files).encode())

docs/basic.rst

@@ -104,6 +104,8 @@ or another OS window::
     map ctrl+f3 detach_window tab-prev
     # moves the window into the tab at the left of the active tab
     map ctrl+f3 detach_window tab-left
+    # moves the window into a new tab created to the left of the active tab
+    map ctrl+f3 detach_window new-tab-left
     # asks which tab to move the window into
     map ctrl+f4 detach_window ask
 

docs/binary.rst

@@ -54,14 +54,14 @@ particular desktop, but it should work for most major desktop environments.
     cp ~/.local/kitty.app/share/applications/kitty.desktop ~/.local/share/applications/
     # If you want to open text files and images in kitty via your file manager also add the kitty-open.desktop file
     cp ~/.local/kitty.app/share/applications/kitty-open.desktop ~/.local/share/applications/
-    # Update the paths to the kitty and its icon in the kitty.desktop file(s)
+    # Update the paths to the kitty and its icon in the kitty desktop file(s)
     sed -i "s|Icon=kitty|Icon=/home/$USER/.local/kitty.app/share/icons/hicolor/256x256/apps/kitty.png|g" ~/.local/share/applications/kitty*.desktop
     sed -i "s|Exec=kitty|Exec=/home/$USER/.local/kitty.app/bin/kitty|g" ~/.local/share/applications/kitty*.desktop
 
 .. note::
     In :file:`kitty-open.desktop`, kitty is registered to handle some supported
     MIME types. This will cause kitty to take precedence on some systems where
-    the default apps are not explicitly set. For example, you expect to use
+    the default apps are not explicitly set. For example, if you expect to use
     other GUI file managers to open dir paths when using commands such as
     :program:`xdg-open`, you should configure the default opener for the MIME
     type ``inode/directory``::

docs/build.rst

@@ -30,7 +30,7 @@ to build kitty with your changes.
 
 .. note::
    If you plan to run kitty from source long-term, there are a couple of
-   caveats to be aware of. You should occassionally run ``./dev.sh deps``
+   caveats to be aware of. You should occasionally run ``./dev.sh deps``
    to have the dependencies re-downloaded as they are updated periodically.
    Also, the built kitty executable assumes it will find source in whatever
    directory you first ran :code:`./dev.sh build` in. If you move/rename the
@@ -96,6 +96,7 @@ Run-time dependencies:
 Build-time dependencies:
 
 * ``gcc`` or ``clang``
+* ``simde``
 * ``go`` >= _build_go_version (see :file:`go.mod` for go packages used during building)
 * ``pkg-config``
 * For building on Linux in addition to the above dependencies you might also
@@ -115,6 +116,7 @@ Build-time dependencies:
   - ``libssl-dev``
   - ``libpython3-dev``
   - ``libxxhash-dev``
+  - ``libsmide-dev``
 
 
 Build and run from source with Nix

docs/changelog.rst

@@ -9,40 +9,200 @@ To update |kitty|, :doc:`follow the instructions <binary>`.
 Recent major new features
 ---------------------------
 
-File transfer over the tty device
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Wayland goodies [0.34]
+~~~~~~~~~~~~~~~~~~~~~~~
 
-Transfer files to and from remote computers over the ``TTY`` device itself.
-This means that file transfer works over nested SSH sessions, serial links,
-etc. Anywhere you have a terminal device, you can transfer files.
+Wayland users should rejoice as kitty now comes with major Wayland
+quality-of-life improvements:
 
-Simply ssh into a remote computer using the :doc:`ssh kitten </kittens/ssh>`
-and run the :doc:`transfer kitten </kittens/transfer>` (which the ssh kitten
-makes available for you on the remote computer automatically). For example, to
-copy a file from a remote computer::
+* Draw GPU accelerated :doc:`desktop panels and background </kittens/panel>`
+  running arbitrary terminal programs. For example, run `btop
+  <https://github.com/aristocratos/btop/>`__ as your desktop background
 
-    <local computer>  $ kitten ssh my-remote-computer
-    <remote computer> $ kitten transfer some-file /path/on/local/computer
+* Background blur for transparent windows is now supported under KDE
+  using a custom KDE specific protocol
 
-The kitten can transfer files to and from the remote computer. It supports
-recursive transfer of directories, symlinks and hardlinks. It can even use the
-rsync algorithm to speed up repeated transfers of large files.
+* The kitty window decorations in GNOME are now fully functional with buttons
+  and they follow system dark/light mode automatically
 
-Truly convenient SSH
-~~~~~~~~~~~~~~~~~~~~~~~~
+* kitty now supports fractional scaling in Wayland which means pixel perfect
+  rendering when you use a fractional scale with no wasted performance on
+  resizing an overdrawn pixmap in the compositor
 
-The :doc:`ssh kitten <kittens/ssh>` is redesigned with powerful new features:
+With this release kitty's Wayland support is now on par with X11, provided
+you use a decent Wayland compositor.
 
-* Automatic :ref:`shell_integration` on remote machines
-* Easily :ref:`clone local shell/editor config <real_world_ssh_kitten_config>` on remote machines
-* Easily :ref:`edit files in your local editor <edit_file>` on remote machines
-* Automatic :opt:`re-use of existing connections <kitten-ssh.share_connections>` to avoid connection setup latency
+Cheetah speed 🐆 [0.33]
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
+kitty has grown up and become a cheetah. It now parses data it receives in
+parallel :iss:`using SIMD vector CPU instructions <7005>` for a 2x speedup in
+benchmarks and a 10%-50% real world speedup depending on workload. There is a
+new benchmarking kitten ``kitten __benchmark__`` that can be used to measure
+terminal throughput. There is also :ref:`a table <throughput>` showing kitty is
+much faster than other terminal emulators based on the benchmark kitten. While
+kitty was already so fast that its performance was never a bottleneck, this
+improvement makes it even faster and more importantly reduces the energy
+consumption to do the same tasks.
 
 .. }}}
 
 Detailed list of changes
 -------------------------------------
 
+0.34.0 [2024-04-15]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+- Wayland: :doc:`panel kitten <kittens/panel>`: Add support for drawing desktop background and bars
+  using the panel kitten for all compositors that support the `requisite Wayland
+  protocol <https://wayland.app/protocols/wlr-layer-shell-unstable-v1>`__ which is practically speaking all of them but GNOME (:pull:`2590`)
+
+- Show a small :opt:`scrollback indicator <scrollback_indicator_opacity>` along the right window edge when viewing
+  the scrollback to keep track of scroll position (:iss:`2502`)
+
+- Wayland: Support fractional scales so that there is no wasted drawing at larger scale followed by resizing in the compositor
+
+- Wayland KDE: Support :opt:`background_blur`
+
+- Wayland GNOME: The window titlebar now has buttons to minimize/maximize/close the window
+
+- Wayland GNOME: The window titlebar color now follows the system light/dark color scheme preference, see :opt:`wayland_titlebar_color`
+
+- Wayland KDE: Fix mouse cursor hiding not working in Plasma 6 (:iss:`7265`)
+
+- Wayland IME: Fix a bug with handling synthetic keypresses generated by ZMK keyboard + fcitx (:pull:`7283`)
+
+- A new option :opt:`terminfo_type` to allow passing the terminfo database embedded into the :envvar:`TERMINFO` env var directly instead of via a file
+
+- Mouse reporting: Fix drag release event outside the window not being reported in legacy mouse reporting modes (:iss:`7244`)
+
+- macOS: Fix a regression in the previous release that broke rendering of some symbols on some systems (:iss:`7249`)
+
+- Fix handling of tab character when cursor is at end of line and wrapping is enabled (:iss:`7250`)
+
+- Splits layout: Fix :ac:`move_window_forward` not working (:iss:`7264`)
+
+- macOS: Fix an abort due to an assertion when a program tries to set an invalid window title (:iss:`7271`)
+
+- fish shell integration: Fix clicking at the prompt causing autosuggestions to be accepted, needs fish >= 3.8.0 (:iss:`7168`)
+
+- Linux: Fix for a regression in 0.32.0 that caused some CJK fonts to not render glyphs (:iss:`7263`)
+
+- Wayland: Support preferred integer scales
+
+- Wayland: A new option :opt:`wayland_enable_ime` to turn off Input Method Extensions which add latency and create bugs
+
+- Wayland: Fix :opt:`hide_window_decorations` not working on non GNOME desktops
+
+- When asking for quit confirmation because of a running program, mention the program name (:iss:`7331`)
+
+- Fix flickering of prompt during window resize (:iss:`7324`)
+
+0.33.1 [2024-03-21]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+- Fix a regression in the previous release that caused requesting data from the clipboard via OSC 52 to instead return data from the primary selection (:iss:`7213`)
+
+- Splits layout: Allow resizing until one of the halves in a split is minimally sized (:iss:`7220`)
+
+- macOS: Fix text rendered with fallback fonts not respecting bold/italic styling (:disc:`7241`)
+
+- macOS: When CoreText fails to find a fallback font for a character in the first Private Use Unicode Area, preferentially use the NERD font, if available, for it (:iss:`6043`)
+
+
+0.33.0 [2024-03-12]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+- :ref:`Cheetah speed <throughput>` with a redesigned render loop and a 2x faster escape code
+  parser that uses SIMD CPU vector instruction to parse data in parallel
+  (:iss:`7005`)
+
+- A new benchmark kitten (``kitten __benchmark__``) to measure terminal
+  throughput performance
+
+- Graphics protocol: Add a new delete mode for deleting images whose ids fall within a range. Useful for bulk deletion (:iss:`7080`)
+
+- Keyboard protocol: Fix the :kbd:`Enter`, :kbd:`Tab` and :kbd:`Backspace` keys
+  generating spurious release events even when report all keys as escape codes
+  is not set (:iss:`7136`)
+
+- macOS: The command line args from :file:`macos-launch-services-cmdline` are now
+  prefixed to any args from ``open --args`` rather than overwriting them (:iss:`7135`)
+
+- Allow specifying where the new tab is created for :ac:`detach_window` (:pull:`7134`)
+
+- hints kitten: The option to set the text color for hints now allows arbitrary
+  colors (:pull:`7150`)
+
+- icat kitten: Add a command line argument to override terminal window size detection (:iss:`7165`)
+
+- A new action :ac:`toggle_tab` to easily switch to and back from a tab with a single shortcut (:iss:`7203`)
+
+- When :ac:`clearing terminal <clear_terminal>` add a new type ``to_cursor_scroll`` which can be
+  used to clear to prompt while moving cleared lines into the scrollback
+
+- Fix a performance bottleneck when dealing with thousands of small images
+  (:iss:`7080`)
+
+- kitten @ ls: Return the timestamp at which the window was created (:iss:`7178`)
+
+- hints kitten: Use default editor rather than hardcoding vim to open file at specific line (:iss:`7186`)
+
+- Remote control: Fix ``--match`` argument not working for @ls, @send-key,
+  @set-background-image (:iss:`7192`)
+
+- Keyboard protocol: Do not deliver a fake key release events on OS window focus out for engaged modifiers (:iss:`7196`)
+
+- Ignore :opt:`startup_session` when kitty is invoked with command line options specifying a command to run (:pull:`7198`)
+
+- Box drawing: Specialize rendering for the Fira Code progress bar/spinner glyphs
+
+0.32.2 [2024-02-12]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+- kitten @ load-config: Allow (re)loading kitty.conf via remote control
+
+- Remote control: Allow running mappable actions via remote control (`kitten @ action`)
+
+- kitten @ send-text: Add a new option to automatically wrap the sent text in
+  bracketed paste escape codes if the program in the destination window has
+  turned on bracketed paste.
+
+- Fix a single key mapping not overriding a previously defined multi-key mapping
+
+- macOS: Fix :code:`kitten @ select-window` leaving the keyboard in a partially functional state (:iss:`7074`)
+
+- Graphics protocol: Improve display of images using Unicode placeholders or
+  row/column boxes by resizing them using linear instead of nearest neighbor
+  interpolation on the GPU (:iss:`7070`)
+
+- When matching URLs use the definition of legal characters in URLs from the
+  `WHATWG spec <https://url.spec.whatwg.org/#url-code-points>`__ rather than older standards (:iss:`7095`)
+
+- hints kitten: Respect the kitty :opt:`url_excluded_characters` option
+  (:iss:`7075`)
+
+- macOS: Fix an abort when changing OS window chrome for a full screen window via remote control or the themes kitten (:iss:`7106`)
+
+- Special case rendering of some more box drawing characters using shades from the block of symbols for legacy computing (:iss:`7110`)
+
+- A new action :ac:`close_other_os_windows` to close non active OS windows (:disc:`7113`)
+
+0.32.1 [2024-01-26]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+- macOS: Fix a regression in the previous release that broke overriding keyboard shortcuts for actions present in the global menu bar (:iss:`7016`)
+
+- Fix a regression in the previous release that caused multi-key sequences to not abort when pressing an unknown key (:iss:`7022`)
+
+- Fix a regression in the previous release that caused `kitten @ launch --cwd=current` to fail over SSH (:iss:`7028`)
+
+- Fix a regression in the previous release that caused `kitten @ send-text` with a match tab parameter to send text twice to the active window (:iss:`7027`)
+
+- Fix a regression in the previous release that caused overriding of existing multi-key mappings to fail (:iss:`7044`, :iss:`7058`)
+
+- Wayland+NVIDIA: Do not request an sRGB output buffer as a bug in Wayland causes kitty to not start (:iss:`7021`)
+
 0.32.0 [2024-01-19]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

docs/conf.py

@@ -211,16 +211,16 @@ def commit_role(
     ' Link to a github commit '
     try:
         commit_id = subprocess.check_output(
-            f'git rev-list --max-count=1 --skip=# {text}'.split()).decode('utf-8').strip()
+            f'git rev-list --max-count=1 {text}'.split()).decode('utf-8').strip()
     except Exception:
         msg = inliner.reporter.error(
-            f'GitHub commit id "{text}" not recognized.', line=lineno)
+            f'git commit id "{text}" not recognized.', line=lineno)
         prb = inliner.problematic(rawtext, rawtext, msg)
         return [prb], [msg]
     url = f'https://github.com/kovidgoyal/kitty/commit/{commit_id}'
     set_classes(options)
     short_id = subprocess.check_output(
-        f'git rev-list --max-count=1 --abbrev-commit --skip=# {commit_id}'.split()).decode('utf-8').strip()
+        f'git rev-list --max-count=1 --abbrev-commit {commit_id}'.split()).decode('utf-8').strip()
     node = nodes.reference(rawtext, f'commit: {short_id}', refuri=url, **options)
     return [node], []
 # }}}

docs/conf.rst

@@ -16,10 +16,10 @@ frames-per-second. See below for an overview of all customization possibilities.
 You can open the config file within kitty by pressing :sc:`edit_config_file`
 (:kbd:`⌘+,` on macOS). A :file:`kitty.conf` with commented default
 configurations and descriptions will be created if the file does not exist.
-You can reload the config file within kitty by pressing :sc:`reload_config_file`
-(:kbd:`⌃+⌘+,` on macOS) or sending kitty the ``SIGUSR1`` signal.
-You can also display the current configuration by pressing :sc:`debug_config`
-(:kbd:`⌥+⌘+,` on macOS).
+You can reload the config file within kitty by pressing
+:sc:`reload_config_file` (:kbd:`⌃+⌘+,` on macOS) or sending kitty the
+``SIGUSR1`` signal with ``kill -SIGUSR1 $KITTY_PID``. You can also display the
+current configuration by pressing :sc:`debug_config` (:kbd:`⌥+⌘+,` on macOS).
 
 .. _confloc:
 

docs/faq.rst

@@ -263,11 +263,12 @@ fonts to be freely resizable, so it does not support bitmapped fonts.
    symbols from it automatically, and you can tell it to do so explicitly in
    case it doesn't with the :opt:`symbol_map` directive::
 
-        # Nerd Fonts v2.3.3
+        # Nerd Fonts v3.1.0
 
-        symbol_map U+23FB-U+23FE,U+2665,U+26A1,U+2B58,U+E000-U+E00A,U+E0A0-U+E0A3,U+E0B0-U+E0D4,U+E200-U+E2A9,U+E300-U+E3E3,U+E5FA-U+E6AA,U+E700-U+E7C5,U+EA60-U+EBEB,U+F000-U+F2E0,U+F300-U+F32F,U+F400-U+F4A9,U+F500-U+F8FF,U+F0001-U+F1AF0 Symbols Nerd Font Mono
+        symbol_map U+e000-U+e00a,U+ea60-U+ebeb,U+e0a0-U+e0c8,U+e0ca,U+e0cc-U+e0d4,U+e200-U+e2a9,U+e300-U+e3e3,U+e5fa-U+e6b1,U+e700-U+e7c5,U+f000-U+f2e0,U+f300-U+f372,U+f400-U+f532,U+f0001-U+f1af0 Symbols Nerd Font Mono
 
-   Those Unicode symbols beyond the ``E000-F8FF`` Unicode private use area are
+   Those Unicode symbols not in the `Unicode private use areas
+   <https://en.wikipedia.org/wiki/Private_Use_Areas>`__ are
    not included.
 
 If your font is not listed in ``kitty +list-fonts`` it means that it is not
@@ -386,7 +387,7 @@ You can also change the icon manually by following the steps:
 How do I map key presses in kitty to different keys in the terminal program?
 --------------------------------------------------------------------------------------
 
-This is accomplished by using ``map`` with :sc:`send_text <send_text>` in :file:`kitty.conf`.
+This is accomplished by using ``map`` with :ac:`send_key` in :file:`kitty.conf`.
 For example::
 
     map alt+s send_key ctrl+s
@@ -396,7 +397,8 @@ you press the :kbd:`alt+s` key. To see this in action, run::
 
     kitten show-key -m kitty
 
-Which will print out what key events it receives.
+Which will print out what key events it receives. To send arbitrary text rather
+than a key press, see :sc:`send_text <send_text>` instead.
 
 
 How do I open a new window or tab with the same working directory as the current window?
@@ -447,9 +449,6 @@ do not use them, if at all possible. kitty contains features that do all of what
 tmux does, but better, with the exception of remote persistence (:iss:`391`).
 If you still want to use tmux, read on.
 
-Image display will not work, see `tmux issue
-<https://github.com/tmux/tmux/issues/1391>`__.
-
 Using ancient versions of tmux such as 1.8 will cause gibberish on screen when
 pressing keys (:iss:`3541`).
 
@@ -458,11 +457,17 @@ and then switch to another and these terminals have different :envvar:`TERM`
 variables, tmux will break. You will need to restart it as tmux does not support
 multiple terminfo definitions.
 
+Displaying images while inside programs such as nvim or ranger may not work
+depending on whether those programs have adopted support for the :ref:`unicode
+placeholders <graphics_unicode_placeholders>` workaround that kitty created
+for tmux refusing to support images.
+
 If you use any of the advanced features that kitty has innovated, such as
 :doc:`styled underlines </underlines>`, :doc:`desktop notifications
 </desktop-notifications>`, :doc:`extended keyboard support
-</keyboard-protocol>`, etc. they may or may not work, depending on the whims of
-tmux's maintainer, your version of tmux, etc.
+</keyboard-protocol>`, :doc:`file transfer </kittens/transfer>`,  etc.
+they may or may not work, depending on the whims of tmux's maintainer,
+your version of tmux, etc.
 
 
 I opened and closed a lot of windows/tabs and top shows kitty's memory usage is very high?

docs/glossary.rst

@@ -164,7 +164,8 @@ Variables that kitty sets when running child programs
 
 .. envvar:: TERMINFO
 
-   Path to a directory containing the kitty terminfo database.
+   Path to a directory containing the kitty terminfo database. Or the terminfo
+   database itself encoded in base64. See :opt:`terminfo_type`.
 
 .. envvar:: KITTY_INSTALLATION_DIR
 
@@ -230,3 +231,11 @@ Variables that kitty sets when running child programs
 
    Set to ``1`` when kitty is running a shell because of the ``--hold`` flag. Can
    be used to specialize shell behavior in the shell rc files as desired.
+
+.. envvar:: KITTY_SIMD
+
+   Set it to ``128`` to use 128 bit vector registers, ``256`` to use 256 bit
+   vector registers or any other value to prevent kitty from using SIMD CPU
+   vector instructions. Warning, this overrides CPU capability detection so
+   will cause kitty to crash with SIGILL if your CPU does not support the
+   necessary SIMD extensions.

docs/graphics-protocol.rst

@@ -460,7 +460,10 @@ When you specify a placement id, it will be added to the acknowledgement code
 above. Every placement is uniquely identified by the pair of the ``image id``
 and the ``placement id``. If you specify a placement id for an image that does
 not have an id (i.e. has id=0), it will be ignored. In particular this means
-there can exist multiple images with ``image id=0, placement id=0``.
+there can exist multiple images with ``image id=0, placement id=0``. Not
+specifying a placement id or using ``p=0`` for multiple put commands (``a=p``)
+with the same non-zero image id results in multiple placements the image.
+
 An example response::
 
     <ESC>_Gi=<image id>,p=<placement id>;OK<ESC>\
@@ -634,7 +637,7 @@ terminal may apply other heuristics (but it doesn't have to).
 It is important to distinguish between virtual image placements and real images
 displayed on top of Unicode placeholders. Virtual placements are invisible and only play
 the role of prototypes for real images. Virtual placements can be deleted by a
-deletion command only when the `d` key is equal to ``i``, ``I``, ``n`` or ``N``.
+deletion command only when the `d` key is equal to ``i``, ``I``, ``r``, ``R``, ``n`` or ``N``.
 The key values ``a``, ``c``, ``p``, ``q``, ``x``, ``y``, ``z`` and their capital
 variants never affect virtual placements because they do not have a physical
 location on the screen.
@@ -726,13 +729,14 @@ scrollback buffer. The values of the ``x`` and ``y`` keys are the same as cursor
 Value of ``d``       Meaning
 =================    ============
 ``a`` or ``A``       Delete all placements visible on screen
-``i`` or ``I``       Delete all images with the specified id, specified using the ``i`` key. If you specify a ``p`` key for the placement                      id as well, then only the placement with the specified image id and placement id will be deleted.
+``i`` or ``I``       Delete all images with the specified id, specified using the ``i`` key. If you specify a ``p`` key for the placement                          id as well, then only the placement with the specified image id and placement id will be deleted.
 ``n`` or ``N``       Delete newest image with the specified number, specified using the ``I`` key. If you specify a ``p`` key for the
                      placement id as well, then only the placement with the specified number and placement id will be deleted.
 ``c`` or ``C``       Delete all placements that intersect with the current cursor position.
 ``f`` or ``F``       Delete animation frames.
 ``p`` or ``P``       Delete all placements that intersect a specific cell, the cell is specified using the ``x`` and ``y`` keys
 ``q`` or ``Q``       Delete all placements that intersect a specific cell having a specific z-index. The cell and z-index is specified using the ``x``, ``y`` and ``z`` keys.
+``r`` or ``R``       Delete all images whose id is greater than or equal to the value of the ``x`` key and less than or equal to the value of the ``y`` (added in kitty version 0.33.0).
 ``x`` or ``X``       Delete all placements that intersect the specified column, specified using the ``x`` key.
 ``y`` or ``Y``       Delete all placements that intersect the specified row, specified using the ``y`` key.
 ``z`` or ``Z``       Delete all placements that have the specified z-index, specified using the ``z`` key.
@@ -1080,9 +1084,11 @@ Key      Value                 Default    Description
 **Keys for deleting images**
 -----------------------------------------------------------
 ``d``    Single character.     ``a``      What to delete.
-         ``(a, A, c, C, n, N,
-         i, I, p, P, q, Q, x,
-         X, y, Y, z, Z)``.
+         ``(
+         a, A, c, C, n, N,
+         i, I, p, P, q, Q, r,
+         R, x, X, y, Y, z, Z
+         )``.
 =======  ====================  =========  =================
 
 

docs/integrations.rst

@@ -164,6 +164,12 @@ Add this to bashrc and then to plot a function, simply do:
 
     iplot 'sin(x*3)*exp(x*.2)'
 
+.. tool_tgutui:
+
+`tgutui <https://github.com/tgu-ltd/tgutui>`_
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+A Terminal Operating Test hardware equipment
+
 .. tool_onefetch:
 
 `onefetch <https://github.com/o2sh/onefetch>`_
@@ -189,13 +195,6 @@ A tool to display weather information in your terminal with curl
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 View and manage the system clipboard under Wayland in your kitty terminal
 
-.. tool_dmenu_term:
-
-`dmenu-term <https://github.com/maximbaz/dmenu-term>`_
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-Run applications on your system with fuzzy find inside a kitty window
-
-
 Editor integration
 -----------------------
 

docs/keyboard-protocol.rst

@@ -49,8 +49,10 @@ In addition to kitty, this protocol is also implemented in:
 * The `dte text editor <https://gitlab.com/craigbarnes/dte/-/issues/138>`__
 * The `Helix text editor <https://github.com/helix-editor/helix/pull/4939>`__
 * The `far2l file manager <https://github.com/elfmz/far2l/commit/e1f2ee0ef2b8332e5fa3ad7f2e4afefe7c96fc3b>`__
+* The `yazi file manager <https://github.com/sxyazi/yazi>`__
 * The `awrit web browser <https://github.com/chase/awrit>`__
 * The `nushell shell <https://github.com/nushell/nushell/pull/10540>`__
+* The `fish shell <https://github.com/fish-shell/fish-shell/commit/8bf8b10f685d964101f491b9cc3da04117a308b4>`__
 
 .. versionadded:: 0.20.0
 
@@ -83,7 +85,7 @@ text (``CSI`` is the bytes ``0x1b 0x5b``)::
 The ``number`` in the first form above will be either the Unicode codepoint for a
 key, such as ``97`` for the :kbd:`a` key, or one of the numbers from the
 :ref:`functional` table below. The ``modifiers`` optional parameter encodes any
-modifiers pressed for the key event. The encoding is described in the
+modifiers active for the key event. The encoding is described in the
 :ref:`modifiers` section.
 
 The second form is used for a few functional keys, such as the :kbd:`Home`,
@@ -105,9 +107,7 @@ do not. When a key event produces text, the text is sent directly as UTF-8
 encoded bytes. This is safe as UTF-8 contains no C0 control codes.
 When the key event does not have text, the key event is encoded as an escape code. In
 legacy compatibility mode (the default) this uses legacy escape codes, so old terminal
-applications continue to work. Key events that could not be represented in
-legacy mode are encoded using a ``CSI u`` escape code, that most terminal
-programs should just ignore. For more advanced features, such as release/repeat
+applications continue to work. For more advanced features, such as release/repeat
 reporting etc., applications can tell the terminal they want this information by
 sending an escape code to :ref:`progressively enhance <progressive_enhancement>` the data reported for
 key events.
@@ -181,10 +181,12 @@ bit field with::
     num_lock  0b10000000  (128)
 
 In the escape code, the modifier value is encoded as a decimal number which is
-``1 + actual modifiers``. So to represent :kbd:`shift` only, the value would be ``1 +
-1 = 2``, to represent :kbd:`ctrl+shift` the value would be ``1 + 0b101 = 6``
-and so on. If the modifier field is not present in the escape code, its default
-value is ``1`` which means no modifiers.
+``1 + actual modifiers``. So to represent :kbd:`shift` only, the value would be
+``1 + 1 = 2``, to represent :kbd:`ctrl+shift` the value would be ``1 + 0b101 =
+6`` and so on. If the modifier field is not present in the escape code, its
+default value is ``1`` which means no modifiers. If a modifier is *active* when
+the key event occurs, i.e. if the key is pressed or the lock (for caps lock/num
+lock) is enabled, the key event must have the bit for that modifier set.
 
 When the key event is related to an actual modifier key, the corresponding
 modifier's bit must be set to the modifier state including the effect for the
@@ -229,8 +231,10 @@ enhancement <progressive_enhancement>` mechanism described below. Some examples:
     shift+a -> CSI 97 ; 2 ; 65 u  # The text 'A' is reported as 65
     option+a -> CSI 97 ; ; 229 u  # The text 'å' is reported as 229
 
-If multiple code points are present, they must be separated by colons.
-If no known key is associated with the text the key number ``0`` must be used.
+If multiple code points are present, they must be separated by colons.  If no
+known key is associated with the text the key number ``0`` must be used. The
+associated text must not contain control codes (control codes are code points
+below U+0020 and codepoints in the C0 and C1 blocks).
 
 
 Non-Unicode keys
@@ -336,7 +340,9 @@ much easier to integrate into the application event loop. The only exceptions
 are the :kbd:`Enter`, :kbd:`Tab` and :kbd:`Backspace` keys which still generate the same
 bytes as in legacy mode this is to allow the user to type and execute commands
 in the shell such as ``reset`` after a program that sets this mode crashes
-without clearing it.
+without clearing it. Note that the Lock modifiers are not reported for text
+producing keys, to keep them useable in legacy programs. To get lock modifiers
+for all keys use the :ref:`report_all_keys` enhancement.
 
 .. _report_events:
 
@@ -348,6 +354,13 @@ and key release events. Normally only key press events are reported and key
 repeat events are treated as key press events. See :ref:`event_types` for
 details on how these are reported.
 
+.. note::
+
+   The :kbd:`Enter`, :kbd:`Tab` and :kbd:`Backspace` keys will not have release
+   events unless :ref:`report_all_keys` is also set, so that the user can still
+   type reset at a shell prompt when a program that sets this mode ends without
+   resetting it.
+
 .. _report_alternates:
 
 Report alternate keys
@@ -482,6 +495,12 @@ must correspond to the :kbd:`Backspace` key.
 All keypad keys are reported as their equivalent non-keypad keys. To
 distinguish these, use the :ref:`disambiguate <disambiguate>` flag.
 
+Terminals may choose what they want to do about functional keys that have no
+legacy encoding. kitty chooses to encode these using ``CSI u`` encoding even in
+legacy mode, so that they become usable even in programs that do not
+understand the full kitty keyboard protocol. However, terminals may instead choose to
+ignore such keys in legacy mode instead, or have an option to control this behavior.
+
 .. _legacy_text:
 
 Legacy text keys

docs/kittens/hints.rst

@@ -26,8 +26,10 @@ and adding them to the command line for the next command.
 
 You can also press :sc:`goto_file_line` to select anything that looks like a
 path or filename followed by a colon and a line number and open the file in
-:program:`vim` at the specified line number. The patterns and editor to be used
-can be modified using options passed to the kitten. For example::
+your default editor at the specified line number (opening at line number will
+work only if your editor supports the +linenum command line syntax or is a
+"known" editor). The patterns and editor to be used can be modified using
+options passed to the kitten. For example::
 
     map ctrl+g kitten hints --type=linenum --linenum-action=tab nvim +{line} {path}
 

docs/kittens/icat.rst

@@ -45,15 +45,29 @@ from inside other programs to display images. In particular, :option:`--place`,
 :option:`--detect-support` and :option:`--print-window-size`.
 
 If you are trying to integrate icat into a complex program like a file manager
-or editor, there are a few things to keep in mind. icat works by communicating
+or editor, there are a few things to keep in mind. icat normally works by communicating
 over the TTY device, it both writes to and reads from the TTY. So it is
 imperative that while it is running the host program does not do any TTY I/O.
 Any key presses or other input from the user on the TTY device will be
-discarded. At a minimum, you should use the :option:`--transfer-mode`
-command line arguments. To be really robust you should
-consider writing proper support for the :doc:`kitty graphics protocol
-</graphics-protocol>` in the program instead. Nowadays there are many libraries
-that have support for it.
+discarded. If you would instead like to use it just as a backend to generate
+the escape codes for image display, you need to pass it options to tell it the
+window dimensions, where to place the image in the window and the transfer mode
+to use. If you do that, it will not try to communicate with the TTY device at
+all. The requisite options are: :option:`--use-window-size`, :option:`--place`
+and :option:`--transfer-mode`, :option:`--stdin=no`.
+For example, to demonstrate usage without access to the TTY:
+
+.. code:: sh
+
+   zsh -c 'setsid kitten icat --stdin=no --use-window-size $COLUMNS,$LINES,3000,2000 --transfer-mode=file myimage.png'
+
+Here, ``setsid`` ensures icat has no access to the TTY device.
+The values, 3000, 2000 are made up. They are the window width and height in
+pixels, to obtain which access to the TTY is needed.
+
+To be really robust you should consider writing proper support for the
+:doc:`kitty graphics protocol </graphics-protocol>` in the program instead.
+Nowadays there are many libraries that have support for it.
 
 
 .. include:: /generated/cli-kitten-icat.rst

docs/kittens/panel.rst

@@ -10,12 +10,12 @@ Draw a GPU accelerated dock panel on your desktop
 
 
 You can use this kitten to draw a GPU accelerated panel on the edge of your
-screen, that shows the output from an arbitrary terminal program.
+screen or as the desktop wallpaper, that shows the output from an arbitrary
+terminal program.
 
 It is useful for showing status information or notifications on your desktop
 using terminal programs instead of GUI toolkits.
 
-
 .. figure:: ../screenshots/panel.png
    :alt: Screenshot, showing a sample panel
    :align: center
@@ -28,18 +28,32 @@ The screenshot above shows a sample panel that displays the current desktop and
 window title as well as miscellaneous system information such as network
 activity, CPU load, date/time, etc.
 
+.. versionadded:: 0.34.0
+   Support for Wayland
+
 .. note::
 
-    This kitten currently only works on X11 desktops
+    This kitten currently only works on X11 desktops and Wayland compositors
+    that support the `wlr layer shell protocol
+    <https://wayland.app/protocols/wlr-layer-shell-unstable-v1#compositor-support>`__
+    (which is almost all of them except the, as usual, crippled GNOME).
 
 Using this kitten is simple, for example::
 
     kitty +kitten panel sh -c 'printf "\n\n\nHello, world."; sleep 5s'
 
 This will show ``Hello, world.`` at the top edge of your screen for five
-seconds. Here the terminal program we are running is :program:`sh` with a script
+seconds. Here, the terminal program we are running is :program:`sh` with a script
 to print out ``Hello, world!``. You can make the terminal program as complex as
 you like, as demonstrated in the screenshot above.
 
+If you are on Wayland, you can, for instance run::
+
+    kitty +kitten panel --edge=background htop
+
+to display htop as your desktop background. Remember this works in everything
+but GNOME and also, in sway, you have to disable the background wallpaper as
+sway renders that over the panel kitten surface.
+
 
 .. include:: ../generated/cli-kitten-panel.rst

docs/layouts.rst

@@ -141,7 +141,7 @@ The Splits Layout
 --------------------
 
 This is the most flexible layout. You can create any arrangement of windows
-by splitting exiting windows repeatedly. To best use this layout you should
+by splitting existing windows repeatedly. To best use this layout you should
 define a few extra key bindings in :file:`kitty.conf`::
 
     # Create a new window splitting the space used by the existing one so that

docs/mapping.rst

@@ -81,7 +81,7 @@ control scripts. To run a kitten on a key press::
 
     map f1 kitten mykitten.py
 
-Many of kitty;s features are themselves implemented as kittens, for example,
+Many of kitty's features are themselves implemented as kittens, for example,
 :doc:`/kittens/unicode_input`, :doc:`/kittens/hints` and
 :doc:`/kittens/themes`. To learn about writing your own kittens, see
 :doc:`/kittens/custom`.
@@ -189,16 +189,46 @@ has :code:`keyboard protocol` in its title. Run the show-key kitten as::
 Press :kbd:`ctrl+shift+t` and instead of a new tab opening, you will
 see the key press being reported by the kitten. :code:`--when-focus-on` can test
 the focused window using very powerful criteria, see :ref:`search_syntax` for
-details. A more practical example unmaps the key when the focused window is running vim::
+details. A more practical example unmaps the key when the focused window is
+running an editor::
 
-    map --when-focus-on var:in_editor
+    map --when-focus-on var:in_editor kitty_mod+c
 
-In order to make this work, you need the following lines in your :file:`.vimrc`::
+In order to make this work, you need to configure your editor as show below:
 
-    let &t_ti = &t_ti . "\\033]1337;SetUserVar=in_editor=MQo\\007"
-    let &t_te = &t_te . "\\033]1337;SetUserVar=in_editor\\007"
+.. tab:: vim
 
-These cause vim to set the :code:`in_editor` variable in kitty and unset it when leaving vim.
+   In :file:`~/.vimrc` add:
+    .. code-block:: vim
+
+        let &t_ti = &t_ti . "\\033]1337;SetUserVar=in_editor=MQo\\007"
+        let &t_te = &t_te . "\\033]1337;SetUserVar=in_editor\\007"
+
+.. tab:: neovim
+
+   In :file:`~/.config/nvim/init.lua` add:
+
+    .. code-block:: lua
+
+        vim.api.nvim_create_autocmd({ "VimEnter", "VimResume" }, {
+            group = vim.api.nvim_create_augroup("KittySetVarVimEnter", { clear = true }),
+            callback = function()
+                io.stdout:write("\x1b]1337;SetUserVar=in_editor=MQo\007")
+            end,
+        })
+
+        vim.api.nvim_create_autocmd({ "VimLeave", "VimSuspend" }, {
+            group = vim.api.nvim_create_augroup("KittyUnsetVarVimLeave", { clear = true }),
+            callback = function()
+                io.stdout:write("\x1b]1337;SetUserVar=in_editor\007")
+            end,
+        })
+
+These cause the editor to set the :code:`in_editor` variable in kitty and unset it when exiting.
+As a result, the :kbd:`ctrl+shift+c` key will be passed to the editor instead of
+copying to clipboard. In the editor, you can map it to copy to the clipboard,
+thereby allowing use of a common shortcut both inside and outside the editor
+for copying to clipboard.
 
 Sending arbitrary text or keys to the program running in kitty
 --------------------------------------------------------------------------------

docs/overview.rst

@@ -251,8 +251,11 @@ The scrollback buffer
 -----------------------
 
 |kitty| supports scrolling back to view history, just like most terminals. You
-can use either keyboard shortcuts or the mouse scroll wheel to do so. However,
-|kitty| has an extra, neat feature. Sometimes you need to explore the scrollback
+can use either keyboard shortcuts or the mouse scroll wheel to do so. While
+you are browsing the scrollback a :opt:`small indicator <scrollback_indicator_opacity>`
+is displayed along the right edge of the window to show how far back you are.
+
+However, |kitty| has an extra, neat feature. Sometimes you need to explore the scrollback
 buffer in more detail, maybe search for some text or refer to it side-by-side
 while typing in a follow-up command. |kitty| allows you to do this by pressing
 the :sc:`show_scrollback` shortcut, which will open the scrollback buffer in

docs/performance.rst

@@ -3,10 +3,14 @@ Performance
 
 The main goals for |kitty| performance are user perceived latency while typing
 and "smoothness" while scrolling as well as CPU usage. |kitty| tries hard to
-find an optimum balance for these. To that end it keeps a cache of each rendered
-glyph in video RAM so that font rendering is not a bottleneck. Interaction with
-child programs takes place in a separate thread from rendering, to improve
-smoothness.
+find an optimum balance for these. To that end it keeps a cache of each
+rendered glyph in video RAM so that font rendering is not a bottleneck.
+Interaction with child programs takes place in a separate thread from
+rendering, to improve smoothness. Parsing of the byte stream is done using
+`vector CPU instructions
+<https://en.wikipedia.org/wiki/Single_instruction,_multiple_data>`__ for
+maximum performance. Updates to the screen typically require sending just a few
+bytes to the GPU.
 
 There are two config options you can tune to adjust the performance,
 :opt:`repaint_delay` and :opt:`input_delay`. These control the artificial delays
@@ -15,19 +19,110 @@ introduced into the render loop to reduce CPU usage. See
 option to further decrease latency at the cost of some `screen tearing
 <https://en.wikipedia.org/wiki/Screen_tearing>`__ while scrolling.
 
-You can generate detailed per-function performance data using
-`gperftools <https://github.com/gperftools/gperftools>`__. Build |kitty| with
-``make profile``. Run kitty and perform the task you want to analyse, for
-example, scrolling a large file with :program:`less`. After you quit, function
-call statistics will be printed to STDOUT and you can use tools like
-*KCachegrind* for more detailed analysis.
+Benchmarks
+-------------
 
-Here are some CPU usage numbers for the task of scrolling a file continuously in
-:program:`less`. The CPU usage is for the terminal process and X together and is
-measured using :program:`htop`. The measurements are taken at the same font and
-window size for all terminals on a ``Intel(R) Core(TM) i7-4820K CPU @ 3.70GHz``
-CPU with a ``Advanced Micro Devices, Inc. [AMD/ATI] Cape Verde XT [Radeon HD
-7770/8760 / R7 250X]`` GPU.
+Measuring terminal emulator performance is fairly subtle, there are three main
+axes on which performance is measured: Energy usage for typical tasks,
+Keyboard to screen latency, and throughput (processing large amounts of data).
+
+Keyboard to screen latency
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This is measured either with dedicated hardware, or software such as `Typometer
+<https://pavelfatin.com/typometer/>`__. Third party measurements comparing
+kitty with other terminal emulators on various systems show kitty has best in
+class keyboard to screen latency.
+
+Note that to minimize latency at the expense of more energy usage, use the
+following settings in kitty.conf::
+
+    input_delay 0
+    repaint_delay 2
+    sync_to_monitor no
+    wayland_enable_ime no
+
+`Hardware based measurement on macOS
+<https://thume.ca/2020/05/20/making-a-latency-tester/>`__ show that kitty and
+Apple's Terminal.app share the crown for best latency. These
+measurements were done with :opt:`input_delay` at its default value of ``3 ms``
+which means kitty's actual numbers would be even lower.
+
+`Typometer based measurements on Linux
+<https://github.com/kovidgoyal/kitty/issues/2701#issuecomment-911089374>`__
+show that kitty has far and away the best latency of the terminals tested.
+
+.. _throughput:
+
+Throughput
+^^^^^^^^^^^^^^^^
+
+kitty has a builtin kitten to measure throughput, it works by dumping large
+amounts of data of different types into the tty device and measuring how fast
+the terminal parses and responds to it. The measurements below were taken with
+the same font, font size and window size for all terminals, and default
+settings, on the same computer. They clearly show kitty has the fastest
+throughput. To run the tests yourself, run ``kitten __benchmark__`` in the
+terminal emulator you want to test, where the kitten binary is part of the
+kitty install.
+
+The numbers are megabytes per second of data that the terminal
+processes. Measurements were taken under Linux/X11 with an ``AMD Ryzen 7 PRO
+5850U``. Entries are in order of decreasing performance. kitty is twice
+as fast as the next best.
+
+================   ======  ======= ===== ====== =======
+Terminal           ASCII   Unicode CSI   Images Average
+================   ======  ======= ===== ====== =======
+kitty 0.33         121.8   105.0   59.8  251.6  134.55
+gnometerm 3.50.1   33.4    55.0    16.1  142.8  61.83
+alacritty 0.13.1   43.1    46.5    32.5  94.1   54.05
+wezterm 20230712   16.4    26.0    11.1  140.5  48.5
+xterm 389          47.7    18.3    0.6   56.3   30.72
+konsole 23.08.04   25.2    37.7    23.6  23.4   27.48
+alacritty+tmux     30.3    7.8     14.7  46.1   24.73
+================   ======  ======= ===== ====== =======
+
+In this table, each column represents different types of data. The CSI column
+is for data consisting of a mix of typical formatting escape codes and some
+ASCII only text.
+
+.. note::
+
+   By default, the benchmark kitten suppresses actual rendering, to better
+   focus on parser speed, you can pass it the ``--render`` flag to not suppress
+   rendering. However, modern terminals typically render asynchronously,
+   therefore the numbers are not really useful for comparison, as it is just a
+   game about how much input to *batch* before rendering the next frame.
+   However, even with rendering enabled kitty is still faster than all the
+   rest. For brevity those numbers are not included.
+
+.. note::
+
+   foot, iterm2 and Terminal.app are left out as they do not run under X11.
+   Alacritty+tmux is included just to show the effect of putting a terminal
+   multiplexer into the mix (halving throughput) and because alacritty isnt
+   remotely comparable to any of the other terminals feature wise without tmux.
+
+.. note::
+
+   konsole, gnome-terminal and xterm do not support the `Synchronized update
+   <https://gitlab.com/gnachman/iterm2/-/wikis/synchronized-updates-spec>`__
+   escape code used to suppress rendering, if and when they gain support for it
+   their numbers are likely to improve by ``20 - 50%``, depending on how well they
+   implement it.
+
+
+Energy usage
+^^^^^^^^^^^^^^^^^
+
+Sadly, I do not have the infrastructure to measure actual energy usage so CPU
+usage will have to stand in for it. Here are some CPU usage numbers for the
+task of scrolling a file continuously in :program:`less`. The CPU usage is for
+the terminal process and X together and is measured using :program:`htop`. The
+measurements are taken at the same font and window size for all terminals on a
+``Intel(R) Core(TM) i7-4820K CPU @ 3.70GHz`` CPU with a ``Advanced Micro
+Devices, Inc. [AMD/ATI] Cape Verde XT [Radeon HD 7770/8760 / R7 250X]`` GPU.
 
 ==============   =========================
 Terminal         CPU usage (X + terminal)
@@ -40,21 +135,16 @@ gnome-terminal   15 - 17%
 konsole          29 - 31%
 ==============   =========================
 
-
 As you can see, |kitty| uses much less CPU than all terminals, except xterm, but
 its scrolling "smoothness" is much better than that of xterm (at least to my,
 admittedly biased, eyes).
 
+Instrumenting kitty
+-----------------------
 
-.. _perf-cat:
-
-.. note::
-
-    Some people have asked why kitty does not perform better than terminal XXX
-    in the test of sinking large amounts of data, such as catting a large text
-    file. The answer is because this is not a goal for kitty. kitty deliberately
-    throttles input parsing and output rendering to minimize resource usage
-    while still being able to sink output faster than any real world program can
-    produce it. Reducing CPU usage, and hence battery drain while achieving
-    instant response times and smooth scrolling to a human eye is a far more
-    important goal.
+You can generate detailed per-function performance data using
+`gperftools <https://github.com/gperftools/gperftools>`__. Build |kitty| with
+``make profile``. Run kitty and perform the task you want to analyse, for
+example, scrolling a large file with :program:`less`. After you quit, function
+call statistics will be displayed in *KCachegrind*. Hence, profiling is best done
+on Linux which has these tools easily available.