This module provides Builder primitives, which are lower level building blocks for constructing Builders. You don't need to go down to this level but it can be slightly faster.

Morally, builder primitives are like functions a -> Builder, that is they take a value and encode it as a sequence of bytes, represented as a Builder. Of course their implementation is a bit more specialised.

Builder primitives come in two forms: fixed-size and bounded-size.

• Fixed(-size) primitives are builder primitives that always result in a sequence of bytes of a fixed length. That is, the length is independent of the value that is encoded. An example of a fixed size primitive is the big-endian encoding of a Word64, which always results in exactly 8 bytes.
• Bounded(-size) primitives are builder primitives that always result in a sequence of bytes that is no larger than a predetermined bound. That is, the bound is independent of the value that is encoded but the actual length will depend on the value. An example for a bounded primitive is the UTF-8 encoding of a Char, which can be 1,2,3 or 4 bytes long, so the bound is 4 bytes.

Note that fixed primitives can be considered as a special case of bounded primitives, and we can lift from fixed to bounded.

Because bounded primitives are the more general case, in this documentation we only refer to fixed size primitives where it matters that the resulting sequence of bytes is of a fixed length. Otherwise, we just refer to bounded size primitives.

The purpose of using builder primitives is to improve the performance of Builders. These improvements stem from making the two most common steps performed by a Builder more efficient. We explain these two steps in turn.

The first most common step is the concatenation of two Builders. Internally, concatenation corresponds to function composition. (Note that Builders can be seen as difference-lists of buffer-filling functions; cf. http://hackage.haskell.org/cgi-bin/hackage-scripts/package/dlist. ) Function composition is a fast O(1) operation. However, we can use bounded primitives to remove some of these function compositions altogether, which is more efficient.

The second most common step performed by a Builder is to fill a buffer using a bounded primitives, which works as follows. The Builder checks whether there is enough space left to execute the bounded primitive. If there is, then the Builder executes the bounded primitive and calls the next Builder with the updated buffer. Otherwise, the Builder signals its driver that it requires a new buffer. This buffer must be at least as large as the bound of the primitive. We can use bounded primitives to reduce the number of buffer-free checks by fusing the buffer-free checks of consecutive Builders. We can also use bounded primitives to simplify the control flow for signalling that a buffer is full by ensuring that we check first that there is enough space left and only then decide on how to encode a given value.

Let us illustrate these improvements on the CSV-table rendering example from Data.ByteString.Builder. Its "hot code" is the rendering of a table's cells, which we implement as follows using only the functions from the Builder API.

import Data.ByteString.Builder as B

renderCell :: Cell -> Builder
renderCell (StringC cs) = renderString cs
renderCell (IntC i)     = B.intDec i

renderString :: String -> Builder
renderString cs = B.charUtf8 '"' <> foldMap escape cs <> B.charUtf8 '"'
  where
    escape '\\' = B.charUtf8 '\\' <> B.charUtf8 '\\'
    escape '\"' = B.charUtf8 '\\' <> B.charUtf8 '\"'
    escape c    = B.charUtf8 c

Efficient encoding of Ints as decimal numbers is performed by intDec. Optimization potential exists for the escaping of Strings. The above implementation has two optimization opportunities. First, the buffer-free checks of the Builders for escaping double quotes and backslashes can be fused. Second, the concatenations performed by foldMap can be eliminated. The following implementation exploits these optimizations.

import qualified Data.ByteString.Builder.Prim  as P
import           Data.ByteString.Builder.Prim
                 ( condB, liftFixedToBounded, (>*<), (>$<) )

renderString :: String -> Builder
renderString cs =
    B.charUtf8 '"' <> primMapListBounded escape cs <> B.charUtf8 '"'
  where
    escape :: BoundedPrim Char
    escape =
      condB (== '\\') (fixed2 ('\\', '\\')) $
      condB (== '\"') (fixed2 ('\\', '\"')) $
      charUtf8
     
    {-# INLINE fixed2 #-}
    fixed2 x = liftFixedToBounded $ const x >$< char7 >*< char7

The code should be mostly self-explanatory. The slightly awkward syntax is because the combinators are written such that the size-bound of the resulting BoundedPrim can be computed at compile time. We also explicitly inline the fixed2 primitive, which encodes a fixed tuple of characters, to ensure that the bound computation happens at compile time. When encoding the following list of Strings, the optimized implementation of renderString is two times faster.

maxiStrings :: [String]
maxiStrings = take 1000 $ cycle ["hello", "\"1\"", "λ-wörld"]

Most of the performance gain stems from using primMapListBounded, which encodes a list of values from left-to-right with a BoundedPrim. It exploits the Builder internals to avoid unnecessary function compositions (i.e., concatenations). In the future, we might expect the compiler to perform the optimizations implemented in primMapListBounded. However, it seems that the code is currently to complicated for the compiler to see through. Therefore, we provide the BoundedPrim escape hatch, which allows data structures to provide very efficient encoding traversals, like primMapListBounded for lists.

Note that BoundedPrims are a bit verbose, but quite versatile. Here is an example of a BoundedPrim for combined HTML escaping and UTF-8 encoding. It exploits that the escaped character with the maximal Unicode codepoint is '>'.

{-# INLINE charUtf8HtmlEscaped #-}
charUtf8HtmlEscaped :: BoundedPrim Char
charUtf8HtmlEscaped =
    condB (>  '>' ) charUtf8 $
    condB (== '<' ) (fixed4 ('&',('l',('t',';')))) $        -- &lt;
    condB (== '>' ) (fixed4 ('&',('g',('t',';')))) $        -- &gt;
    condB (== '&' ) (fixed5 ('&',('a',('m',('p',';'))))) $  -- &amp;
    condB (== '"' ) (fixed5 ('&',('#',('3',('4',';'))))) $  -- &#34;
    condB (== '\'') (fixed5 ('&',('#',('3',('9',';'))))) $  -- &#39;
    (liftFixedToBounded char7)         -- fallback for Chars smaller than '>'
  where
    {-# INLINE fixed4 #-}
    fixed4 x = liftFixedToBounded $ const x >$<
      char7 >*< char7 >*< char7 >*< char7
     
    {-# INLINE fixed5 #-}
    fixed5 x = liftFixedToBounded $ const x >$<
      char7 >*< char7 >*< char7 >*< char7 >*< char7

This module currently does not expose functions that require the special properties of fixed-size primitives. They are useful for prefixing Builders with their size or for implementing chunked encodings. We will expose the corresponding functions in future releases of this library.

The combinators for BoundedPrims are implemented such that the size of the resulting BoundedPrim can be computed at compile time.

The combinators for FixedPrims are implemented such that the size of the resulting FixedPrim is computed at compile time.

The (>*<) and (>$<) pairing and mapping operators can be used with FixedPrim.

In terms of expressivity, the function fixedPrim would be sufficient for constructing Builders from FixedPrims. The fused variants of this function are provided because they allow for more efficient implementations. Our compilers are just not smart enough yet; and for some of the employed optimizations (see the code of primMapByteStringFixed) they will very likely never be.

Note that functions marked with "Heavy inlining." are forced to be inlined because they must be specialized for concrete encodings, but are rather heavy in terms of code size. We recommend to define a top-level function for every concrete instantiation of such a function in order to share its code. A typical example is the function byteStringHex from Data.ByteString.Builder.ASCII, which is implemented as follows.

byteStringHex :: S.StrictByteString -> Builder
byteStringHex = primMapByteStringFixed word8HexFixed

The ISO/IEC 8859-1 encoding is an 8-bit encoding often known as Latin-1. The Char8 encoding implemented here works by truncating the Unicode codepoint to 8-bits and encoding them as a single byte. For the codepoints 0-255 this corresponds to the ISO/IEC 8859-1 encoding. Note that the Char8 encoding is equivalent to the ASCII encoding on the Unicode codepoints 0-127. Hence, functions such as intDec can also be used for encoding Ints as a decimal number with Char8 encoded characters.

The UTF-8 encoding can encode all Unicode codepoints. It is equivalent to the ASCII encoding on the Unicode codepoints 0-127. Hence, functions such as intDec can also be used for encoding Ints as a decimal number with UTF-8 encoded characters.