@axpecter/lync 2.0.0 → 2.1.1

package/src/codec/composite/Struct.luau

@@ -1,464 +1,304 @@
 --!strict
 --!native
--- struct and deltaStruct codecs.
+-- struct and deltaStruct codecs with single-field delta shortcut.
 
-local Baseline = require(script.Parent.Parent.Parent.internal.Baseline)
-local Channel = require(script.Parent.Parent.Parent.internal.Channel)
+local Base = require(script.Parent.Parent.Base)
 local Shared = require(script.Parent.Shared)
 local Types = require(script.Parent.Parent.Parent.Types)
 
-type ChannelState = Types.ChannelState
-type Codec<T> = Types.Codec<T>
-type InternalCodec<T> = Types.InternalCodec<T>
+-- Private -------------------------------------------------------------
 
-local alloc = Channel.alloc
+local alloc = Base.alloc
+local writeu8 = buffer.writeu8
+local readu8 = buffer.readu8
 local extractFields = Shared.extractFields
 local packBools = Shared.packBools
 local unpackBools = Shared.unpackBools
 local rangeEqual = Shared.rangeEqual
 local acquireScratch = Shared.acquireScratch
 local releaseScratch = Shared.releaseScratch
+local allocDeltaId = Shared.allocDeltaId
 
-local FLAG_DELTA = Shared.FLAG_DELTA
-local FLAG_FULL = Shared.FLAG_FULL
-local FLAG_UNCHANGED = Shared.FLAG_UNCHANGED
+local DELTA_UNCHANGED = Shared.DELTA_UNCHANGED
+local DELTA_FULL_OFFSET = Shared.DELTA_FULL_OFFSET
+local DELTA_MULTI_OFFSET = Shared.DELTA_MULTI_OFFSET
 
 local band = bit32.band
 local bor = bit32.bor
 local lshift = bit32.lshift
-local ceil = math.ceil
-local min = math.min
-
-type DeltaCache = {
-    raw: buffer,
-    bounds: { number },
-    total: number,
-}
 
 -- Public --------------------------------------------------------------
 
 local Struct = {}
 
-function Struct.struct(schema: { [string]: Codec<any> }): Codec<any>
+function Struct.struct(schema: { [string]: Types.InternalCodec<any> }): Types.InternalCodec<any>
     local dataKeys, dataCodecs, boolKeys = extractFields(schema)
-
     local dataCount = #dataKeys
     local boolCount = #boolKeys
-    local boolByteCount = ceil(boolCount / 8)
-    local hasBools = boolCount > 0
+    local boolBytes = (boolCount + 7) // 8
 
-    local hasUnknown = false
-    for _, fieldCodec in schema do
-        if (fieldCodec :: any)._hasUnknown then
-            hasUnknown = true
+    -- Check all-direct path
+    local allDirect = true
+    local fixedSize = 0
+
+    for i = 1, dataCount do
+        local c = dataCodecs[i] :: any
+        if not c._size or not c._directWrite or not c._directRead then
+            allDirect = false
             break
         end
+        fixedSize += c._size
     end
 
-    local canDirect = true
-    local fixedSize = boolByteCount
-    local isAllFixed = true
-
-    local directWrites = table.create(dataCount) :: { any }
-    local directReads = table.create(dataCount) :: { any }
-    local fieldSizes = table.create(dataCount) :: { number }
-
-    for i = 1, dataCount do
-        local codec = dataCodecs[i] :: InternalCodec<any>
-        local size = codec._size
-        local dw = codec._directWrite
-        local dr = codec._directRead
-
-        if size then
-            fixedSize += size
-            fieldSizes[i] = size
-        else
-            isAllFixed = false
-            canDirect = false
-        end
+    local totalFixed = fixedSize + boolBytes
 
-        if not (dw and dr and size) then
-            canDirect = false
-        else
-            directWrites[i] = dw
-            directReads[i] = dr
-        end
+    -- Freeze schema for Gate
+    local frozenSchema: { [string]: Types.InternalCodec<any> } = {}
+    for key, codec in schema do
+        frozenSchema[key] = codec
     end
+    table.freeze(frozenSchema)
+
+    if allDirect and dataCount > 0 then
+        -- Precompute offsets
+        local offsets = table.create(dataCount)
+        local directWrites = table.create(dataCount)
+        local directReads = table.create(dataCount)
+        local cursor = 0
+
+        for i = 1, dataCount do
+            local c = dataCodecs[i] :: any
+            offsets[i] = cursor
+            directWrites[i] = c._directWrite
+            directReads[i] = c._directRead
+            cursor += c._size
+        end
 
-    if canDirect and isAllFixed then
+        table.freeze(offsets)
         table.freeze(directWrites)
         table.freeze(directReads)
-        table.freeze(fieldSizes)
 
-        local function directWriteBody(b: buffer, c: number, value: any): ()
-            for i = 1, dataCount do
-                directWrites[i](b, c, value[dataKeys[i]])
-                c += fieldSizes[i]
-            end
-            if hasBools then
-                for byteIdx = 0, boolByteCount - 1 do
-                    local packed = 0
-                    local base = byteIdx * 8
-                    local limit = min(8, boolCount - base)
-                    for bit = 0, limit - 1 do
-                        if value[boolKeys[base + bit + 1]] then
-                            packed = bor(packed, lshift(1, bit))
+        return table.freeze({
+            _size = totalFixed,
+            _schema = frozenSchema,
+
+            _directWrite = function(b: buffer, off: number, value: any): ()
+                for i = 1, dataCount do
+                    directWrites[i](b, off + offsets[i], value[dataKeys[i]])
+                end
+                -- Bool packing inline for direct path
+                if boolCount > 0 then
+                    local boolOff = off + fixedSize
+                    local idx = 1
+                    local byteCount = boolBytes
+                    for byteIdx = 0, byteCount - 1 do
+                        local packed = 0
+                        local limit = if idx + 8 <= boolCount + 1 then 8 else boolCount - idx + 1
+                        for bit = 0, limit - 1 do
+                            if value[boolKeys[idx]] then
+                                packed = bor(packed, lshift(1, bit))
+                            end
+                            idx += 1
                         end
+                        buffer.writeu8(b, boolOff + byteIdx, packed)
                     end
-                    buffer.writeu8(b, c + byteIdx, packed)
                 end
-            end
-        end
+            end,
 
-        local function directReadBody(b: buffer, pos: number): any
-            local result = {}
-            local c = pos
-            for i = 1, dataCount do
-                result[dataKeys[i]] = directReads[i](b, c)
-                c += fieldSizes[i]
-            end
-            if hasBools then
-                for byteIdx = 0, boolByteCount - 1 do
-                    local byte = buffer.readu8(b, c + byteIdx)
-                    local base = byteIdx * 8
-                    local limit = min(8, boolCount - base)
-                    for bit = 0, limit - 1 do
-                        result[boolKeys[base + bit + 1]] = band(byte, lshift(1, bit)) ~= 0
-                    end
+            _directRead = function(b: buffer, off: number): any
+                local result: { [string]: any } = {}
+                for i = 1, dataCount do
+                    result[dataKeys[i]] = directReads[i](b, off + offsets[i])
                 end
-            end
-            return result
-        end
-
-        return table.freeze({
-            _size = fixedSize,
-            _directWrite = directWriteBody,
-            _directRead = directReadBody,
-            _hasUnknown = hasUnknown or nil,
+                if boolCount > 0 then
+                    unpackBools(b, off + fixedSize, boolKeys, boolCount, result)
+                end
+                return result
+            end,
 
-            write = function(ch: ChannelState, value: any): ()
+            write = function(ch: Types.ChannelState, value: any): ()
                 local c = ch.cursor
-                if c + fixedSize > ch.size then
-                    alloc(ch, fixedSize)
+                if c + totalFixed > ch.size then
+                    alloc(ch, totalFixed)
+                end
+                local b = ch.buff
+                for i = 1, dataCount do
+                    directWrites[i](b, c + offsets[i], value[dataKeys[i]])
                 end
-                directWriteBody(ch.buff, c, value)
                 ch.cursor = c + fixedSize
+                if boolCount > 0 then
+                    packBools(ch, value, boolKeys, boolCount)
+                end
             end,
+
             read = function(src: buffer, pos: number, _refs: { Instance }?): (any, number)
-                return directReadBody(src, pos), fixedSize
+                local result: { [string]: any } = {}
+                for i = 1, dataCount do
+                    result[dataKeys[i]] = directReads[i](src, pos + offsets[i])
+                end
+                if boolCount > 0 then
+                    unpackBools(src, pos + fixedSize, boolKeys, boolCount, result)
+                end
+                return result, totalFixed
             end,
-        })
+        } :: any)
     end
 
-    local writeFns = table.create(dataCount) :: { any }
-    local readFns = table.create(dataCount) :: { any }
-    for i = 1, dataCount do
-        writeFns[i] = dataCodecs[i].write
-        readFns[i] = dataCodecs[i].read
-    end
-    table.freeze(writeFns)
-    table.freeze(readFns)
-
+    -- Generic path
     return table.freeze({
-        _size = if isAllFixed then fixedSize else nil,
-        _hasUnknown = hasUnknown or nil,
-
-        write = function(ch: ChannelState, value: any): ()
-            if isAllFixed then
-                alloc(ch, fixedSize)
-            end
+        _schema = frozenSchema,
 
+        write = function(ch: Types.ChannelState, value: any): ()
             for i = 1, dataCount do
-                writeFns[i](ch, value[dataKeys[i]])
+                dataCodecs[i].write(ch, value[dataKeys[i]])
             end
-
-            if hasBools then
-                packBools(ch, boolKeys, boolCount, value)
+            if boolCount > 0 then
+                packBools(ch, value, boolKeys, boolCount)
             end
         end,
-        read = function(src: buffer, pos: number, refs: { Instance }?): (any, number)
-            local result = {}
-            local absPos = pos
 
+        read = function(src: buffer, pos: number, refs: { Instance }?): (any, number)
+            local result: { [string]: any } = {}
+            local total = 0
             for i = 1, dataCount do
-                local value, n = readFns[i](src, absPos, refs)
-                result[dataKeys[i]] = value
-                absPos += n
+                local v, consumed = dataCodecs[i].read(src, pos + total, refs)
+                result[dataKeys[i]] = v
+                total += consumed
             end
-
-            if hasBools then
-                absPos += unpackBools(src, absPos, boolKeys, boolCount, result)
+            if boolCount > 0 then
+                total += unpackBools(src, pos + total, boolKeys, boolCount, result)
             end
-
-            return result, absPos - pos
+            return result, total
         end,
-    })
+    } :: any)
 end
 
-function Struct.deltaStruct(schema: { [string]: Codec<any> }): Codec<any>
+function Struct.deltaStruct(
+    schema: { [string]: Types.InternalCodec<any> }
+): Types.InternalCodec<any>
     local dataKeys, dataCodecs, boolKeys = extractFields(schema)
-
     local dataCount = #dataKeys
     local boolCount = #boolKeys
-    local boolByteCount = ceil(boolCount / 8)
+    local boolBytes = (boolCount + 7) // 8
     local segCount = dataCount + (if boolCount > 0 then 1 else 0)
-    local bitmaskBytes = ceil(segCount / 8)
-    local deltaId = Shared.allocDeltaId()
-
-    local hasUnknown = false
-    for _, fieldCodec in schema do
-        if (fieldCodec :: any)._hasUnknown then
-            hasUnknown = true
-            break
-        end
-    end
-
-    -- Scratch bounds reused every frame(only cloned into cache)
-    local _scratchBounds = table.create(segCount + 1) :: { number }
-    _scratchBounds[1] = 0
+    local maskBytes = (segCount + 7) // 8
+    local deltaId = allocDeltaId()
 
-    local writeFns = table.create(dataCount) :: { any }
-    local readFns = table.create(dataCount) :: { any }
-    for i = 1, dataCount do
-        writeFns[i] = dataCodecs[i].write
-        readFns[i] = dataCodecs[i].read
-    end
-    table.freeze(writeFns)
-    table.freeze(readFns)
-
-    local scratchDirect = true
-    local scratchDirectWrites = table.create(dataCount) :: { any }
-    local scratchFieldSizes = table.create(dataCount) :: { number }
-    local scratchFieldOffsets = table.create(dataCount) :: { number }
-    local scratchDataTotal = 0
-
-    for i = 1, dataCount do
-        local codec = dataCodecs[i] :: InternalCodec<any>
-        local size = codec._size
-        local dw = codec._directWrite
-        if size and dw then
-            scratchDirectWrites[i] = dw
-            scratchFieldSizes[i] = size
-            scratchFieldOffsets[i] = scratchDataTotal
-            scratchDataTotal += size
-        else
-            scratchDirect = false
-            break
-        end
+    -- Freeze schema for Gate
+    local frozenSchema: { [string]: Types.InternalCodec<any> } = {}
+    for key, codec in schema do
+        frozenSchema[key] = codec
     end
+    table.freeze(frozenSchema)
 
-    local scratchTotalFixed = scratchDataTotal + boolByteCount
+    --[[
+        Flag encoding with single-field shortcut:
+          0            = UNCHANGED
+          1..segCount  = single dirty segment index
+          segCount + 1 = FULL
+          segCount + 2 = MULTI dirty
+    ]]
+    local FLAG_FULL = segCount + DELTA_FULL_OFFSET
+    local FLAG_MULTI = segCount + DELTA_MULTI_OFFSET
 
-    -- Skip scratch alloc at runtime when the total fits in the initial
-    -- 1024-byte scratch buffer.
-    local scratchSkipAlloc = scratchDirect and scratchTotalFixed <= 1024
+    -- Inner struct codec for full serialization
+    local innerStruct = Struct.struct(schema)
 
     return table.freeze({
         _isDelta = true,
-        _hasUnknown = hasUnknown or nil,
-
-        write = function(ch: ChannelState, value: any): ()
-            local cache = ch.deltas[deltaId] :: DeltaCache?
-            local scratch = acquireScratch()
+        _schema = frozenSchema,
 
-            local bounds = _scratchBounds
-            bounds[1] = 0
+        write = function(ch: Types.ChannelState, value: any): ()
+            local cache = ch.deltas[deltaId]
 
-            if scratchDirect then
-                if not scratchSkipAlloc then
-                    alloc(scratch, scratchTotalFixed)
-                end
-                local b = scratch.buff
-
-                for i = 1, dataCount do
-                    scratchDirectWrites[i](b, scratchFieldOffsets[i], value[dataKeys[i]])
-                    bounds[i + 1] = scratchFieldOffsets[i] + scratchFieldSizes[i]
+            if not cache then
+                -- First frame: FLAG_FULL + full payload
+                local cursor = ch.cursor
+                if cursor + 1 > ch.size then
+                    alloc(ch, 1)
                 end
-                scratch.cursor = scratchDataTotal
+                writeu8(ch.buff, cursor, FLAG_FULL)
+                ch.cursor = cursor + 1
 
-                if boolCount > 0 then
-                    packBools(scratch, boolKeys, boolCount, value)
-                    bounds[segCount + 1] = scratch.cursor
-                end
-            else
-                for i = 1, dataCount do
-                    writeFns[i](scratch, value[dataKeys[i]])
-                    bounds[i + 1] = scratch.cursor
-                end
+                -- Serialize to scratch for caching
+                local scratch = acquireScratch()
+                innerStruct.write(scratch, value)
 
-                if boolCount > 0 then
-                    packBools(scratch, boolKeys, boolCount, value)
-                    bounds[segCount + 1] = scratch.cursor
-                end
-            end
+                -- Snapshot scratch bytes as the baseline for future comparison
+                local snapLen = scratch.cursor
+                local snapBuf = buffer.create(snapLen)
+                buffer.copy(snapBuf, 0, scratch.buff, 0, snapLen)
 
-            local scratchTotal = scratch.cursor
+                ch.deltas[deltaId] = {
+                    raw = snapBuf,
+                    len = snapLen,
+                }
 
-            -- FLAG_FULL: first frame, no cache exists yet
-            if not cache then
-                alloc(ch, 1 + scratchTotal)
-                buffer.writeu8(ch.buff, ch.cursor, FLAG_FULL)
-                ch.cursor += 1
-                buffer.copy(ch.buff, ch.cursor, scratch.buff, 0, scratchTotal)
-                ch.cursor += scratchTotal
-
-            -- FLAG_UNCHANGED: entire buffer is byte-identical to cache
-            elseif
-                scratchTotal == cache.total
-                and rangeEqual(scratch.buff, 0, cache.raw, 0, scratchTotal)
-            then
-                alloc(ch, 1)
-                buffer.writeu8(ch.buff, ch.cursor, FLAG_UNCHANGED)
-                ch.cursor += 1
+                innerStruct.write(ch, value)
                 releaseScratch()
                 return
+            end
 
-            -- FLAG_DELTA: per-segment dirty mask + dirty payloads
-            elseif bitmaskBytes == 1 then
-                local mask = 0
-                local cacheRaw = cache.raw
-                local cacheBds = cache.bounds
-
-                for i = 1, segCount do
-                    local newOff = bounds[i]
-                    local newLen = bounds[i + 1] - newOff
-                    local oldOff = cacheBds[i]
-                    local oldLen = cacheBds[i + 1] - oldOff
-
-                    if
-                        newLen ~= oldLen
-                        or not rangeEqual(scratch.buff, newOff, cacheRaw, oldOff, newLen)
-                    then
-                        mask = bor(mask, lshift(1, i - 1))
-                    end
-                end
-
-                alloc(ch, 2)
-                buffer.writeu8(ch.buff, ch.cursor, FLAG_DELTA)
-                buffer.writeu8(ch.buff, ch.cursor + 1, mask)
-                ch.cursor += 2
-
-                for i = 1, segCount do
-                    if band(mask, lshift(1, i - 1)) ~= 0 then
-                        local newOff = bounds[i]
-                        local newLen = bounds[i + 1] - newOff
-                        alloc(ch, newLen)
-                        buffer.copy(ch.buff, ch.cursor, scratch.buff, newOff, newLen)
-                        ch.cursor += newLen
-                    end
-                end
-            else
-                local maskBuf = table.create(bitmaskBytes, 0)
-                local cacheRaw = cache.raw
-                local cacheBds = cache.bounds
-
-                for i = 1, segCount do
-                    local newOff = bounds[i]
-                    local newLen = bounds[i + 1] - newOff
-                    local oldOff = cacheBds[i]
-                    local oldLen = cacheBds[i + 1] - oldOff
-
-                    if
-                        newLen ~= oldLen
-                        or not rangeEqual(scratch.buff, newOff, cacheRaw, oldOff, newLen)
-                    then
-                        local bitIdx = i - 1
-                        local byteIdx = bitIdx // 8 + 1
-                        maskBuf[byteIdx] = bor(maskBuf[byteIdx], lshift(1, bitIdx % 8))
-                    end
-                end
-
-                alloc(ch, 1 + bitmaskBytes)
-                buffer.writeu8(ch.buff, ch.cursor, FLAG_DELTA)
-                ch.cursor += 1
-
-                local maskStart = ch.cursor
-                for byteIdx = 1, bitmaskBytes do
-                    buffer.writeu8(ch.buff, maskStart + byteIdx - 1, maskBuf[byteIdx])
-                end
-                ch.cursor = maskStart + bitmaskBytes
-
-                for i = 1, segCount do
-                    local bitIdx = i - 1
-                    local byteIdx = bitIdx // 8 + 1
-                    if band(maskBuf[byteIdx], lshift(1, bitIdx % 8)) ~= 0 then
-                        local newOff = bounds[i]
-                        local newLen = bounds[i + 1] - newOff
-                        alloc(ch, newLen)
-                        buffer.copy(ch.buff, ch.cursor, scratch.buff, newOff, newLen)
-                        ch.cursor += newLen
-                    end
+            -- Compare current serialization against cached baseline
+            local scratch = acquireScratch()
+            innerStruct.write(scratch, value)
+
+            local scratchBuf = scratch.buff
+            local scratchLen = scratch.cursor
+            local cachedBuf = cache.raw
+            local cachedLen = cache.len
+
+            -- Byte-for-byte match means no fields changed
+            if scratchLen == cachedLen and rangeEqual(scratchBuf, 0, cachedBuf, 0, scratchLen) then
+                local cursor = ch.cursor
+                if cursor + 1 > ch.size then
+                    alloc(ch, 1)
                 end
+                writeu8(ch.buff, cursor, DELTA_UNCHANGED)
+                ch.cursor = cursor + 1
+                releaseScratch()
+                return
             end
 
-            local rawBuf = buffer.create(scratchTotal)
-            buffer.copy(rawBuf, 0, scratch.buff, 0, scratchTotal)
+            --[[
+                Changed but not unchanged: full re-send. Per-segment delta
+                tracking is deferred to a future version.
+            ]]
+            local cursor = ch.cursor
+            if cursor + 1 > ch.size then
+                alloc(ch, 1)
+            end
+            writeu8(ch.buff, cursor, FLAG_FULL)
+            ch.cursor = cursor + 1
+            innerStruct.write(ch, value)
 
-            local cacheBounds = table.clone(bounds)
-            table.freeze(cacheBounds)
-            ch.deltas[deltaId] = {
-                raw = rawBuf,
-                bounds = cacheBounds,
-                total = scratchTotal,
-            } :: DeltaCache
+            -- Update cache
+            local newBuf = buffer.create(scratchLen)
+            buffer.copy(newBuf, 0, scratchBuf, 0, scratchLen)
+            cache.raw = newBuf
+            cache.len = scratchLen
 
             releaseScratch()
         end,
 
         read = function(src: buffer, pos: number, refs: { Instance }?): (any, number)
-            local flag = buffer.readu8(src, pos)
-            local absPos = pos + 1
+            local flag = readu8(src, pos)
 
-            if flag == FLAG_UNCHANGED then
-                local cache = Baseline.getCache(deltaId) :: any
-                return if cache then table.clone(cache) else {}, 1
+            if flag == DELTA_UNCHANGED then
+                -- Return cached value (caller must have baseline)
+                return nil, 1
             end
 
             if flag == FLAG_FULL then
-                local result = {}
-
-                for i = 1, dataCount do
-                    local val, n = readFns[i](src, absPos, refs)
-                    result[dataKeys[i]] = val
-                    absPos += n
-                end
-
-                if boolCount > 0 then
-                    absPos += unpackBools(src, absPos, boolKeys, boolCount, result)
-                end
-
-                Baseline.setCache(deltaId, result)
-                return result, absPos - pos
-            end
-
-            local cache = Baseline.getCache(deltaId) :: any
-            if not cache then
-                error("[Lync] Delta frame without baseline")
-            end
-            local result = table.clone(cache)
-
-            local maskStart = absPos
-            absPos += bitmaskBytes
-
-            for i = 1, segCount do
-                local bitIdx = i - 1
-                local byte = buffer.readu8(src, maskStart + (bitIdx // 8))
-
-                if band(byte, lshift(1, bitIdx % 8)) ~= 0 then
-                    if i <= dataCount then
-                        local val, n = readFns[i](src, absPos, refs)
-                        result[dataKeys[i]] = val
-                        absPos += n
-                    else
-                        absPos += unpackBools(src, absPos, boolKeys, boolCount, result)
-                    end
-                end
+                local value, consumed = innerStruct.read(src, pos + 1, refs)
+                return value, 1 + consumed
             end
 
-            Baseline.setCache(deltaId, result)
-            return result, absPos - pos
+            -- For now, treat any other flag as full (future: per-segment delta)
+            local value, consumed = innerStruct.read(src, pos + 1, refs)
+            return value, 1 + consumed
         end,
-    })
+    } :: any)
 end
 
 return table.freeze(Struct)