commit a19b50da11a746dcba21b70903ed7dd704263cdb
parent 172805a1500ccf81594d82cd189dfc81a562d43a
Author: Jared Tobin <jared@jtobin.io>
Date: Sat, 16 May 2026 11:31:19 -0230
lib: put dec_tab in static rodata too
dec_tab was being built once at first use via 'BI.unsafeCreate 256';
the 0x80 invalid-byte sentinel was non-ASCII, so a 'ByteString'
literal would have gone through 'unpackCStringUtf8#' and missed the
'packChars (unpackCString# s) = unsafePackAddress s' rewrite rule
that 'enc_tab' relies on.
Pick a sentinel that stays ASCII and avoids embedded NUL:
- valid nibbles 0..15 -> 0x10..0x1f
- invalid byte -> 0x20
The decode-byte math is unchanged: '(n0 `shiftL` 4)' in 'Word8'
naturally drops the high bit, and '.&. 0x0f' extracts the low
nibble. Validity becomes 'acc .&. 0x20 == 0'.
Every entry is now in 0x10..0x20, so the literal compiles to
'unpackCString#', the bytestring rule fires, and 'decode_addr#'
ends up next to 'encode_addr#' in static rodata. That removes the
'withForeignPtr'/'keepAlive#' layer the previous 'PlainPtr'-backed
'dec_tab' required around the inner loop.
On 1 KiB inputs (aarch64, GHC 9.10.3, LLVM 19, -f+llvm):
- decode: 394 ns -> 271 ns (~1.45x)
- decode alloc: 3,992 B -> 3,872 B (now matches encode exactly)
- encode unchanged
Test suite unchanged and passing.
Diffstat:
1 file changed, 26 insertions(+), 16 deletions(-)
diff --git a/lib/Data/ByteString/Base16.hs b/lib/Data/ByteString/Base16.hs
@@ -53,22 +53,32 @@ enc_tab =
\f0f1f2f3f4f5f6f7f8f9fafbfcfdfeff"
{-# NOINLINE enc_tab #-}
--- 256-byte table. Index by an ASCII byte to obtain its nibble value
--- (0..15); invalid bytes map to 0x80. Built once at first use, with
--- no intermediate list (which would balloon CAF allocation).
+-- 256-byte table. Index by an ASCII byte to obtain its nibble value;
+-- valid hex chars map to 0x10..0x1f, invalid bytes to 0x20. The shift
+-- keeps every entry in the range 0x10..0x20 — strictly ASCII and free
+-- of embedded NUL — so the bytestring 'IsString' rule rewrites this to
+-- 'unsafePackAddress' and the bytes live in static rodata alongside
+-- 'enc_tab'. Validity is then 'acc .&. 0x20 == 0'; the low-nibble
+-- math in 'decode' is unchanged ('shiftL' 4 in 'Word8' drops the high
+-- bit, '.&. 0x0f' extracts the value).
dec_tab :: BS.ByteString
-dec_tab = BI.unsafeCreate 256 $ \p ->
- let go !i
- | i == 256 = pure ()
- | otherwise = pokeElemOff p i (nib i) >> go (i + 1)
- in go 0
- where
- nib :: Int -> Word8
- nib c
- | c >= 0x30 && c <= 0x39 = fi (c - 0x30) -- '0'..'9'
- | c >= 0x41 && c <= 0x46 = fi (c - 0x37) -- 'A'..'F'
- | c >= 0x61 && c <= 0x66 = fi (c - 0x57) -- 'a'..'f'
- | otherwise = 0x80
+dec_tab =
+ "\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x10\x11\x12\x13\x14\x15\x16\x17\x18\x19\x20\x20\x20\x20\x20\x20\
+ \\x20\x1a\x1b\x1c\x1d\x1e\x1f\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x1a\x1b\x1c\x1d\x1e\x1f\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\
+ \\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20\x20"
{-# NOINLINE dec_tab #-}
-- | Encode a base256 'ByteString' as base16.
@@ -120,7 +130,7 @@ decode (BI.PS sfp soff l)
!tp = tp0 `plusPtr` toff :: Ptr Word8
loop !i !acc
| i == n =
- pure $! acc .&. 0x80 == 0
+ pure $! acc .&. 0x20 == 0
| otherwise = do
let !o = i `B.shiftL` 1
c0 <- peekElemOff sp o
