I did look into this a little. There are several things going on * GHC doesn't really expect you to use INLINE and SPECIALISE together. INLINE says to inline a copy of the function at every call site, which is the best possible form of specialisation, so SPECIALISE is a bit redundant if you are happy to do that. (Incidentally, INLINE is ignored for recursive functions, whereas SPECIALISE works fine.) * If you have both INLINE and SPECIALISE, GHC makes a specialised copy of the function, and the specialised copy takes precedence. So a call at the specialised type turns into a call to the specialised function. (A call at any other type would not use the specialisation, and so the un-specialised function would be inlined.) Moreover, the specialised copies are themselves not marked INLINE. * GHC's HEAD in CVS (i.e 6.5) has a new feature {#- SPECIALISE INLINE f :: <type> #-} which tells GHC to specialise f and inline the specialised copy (only). Ordinary, un-specialised calls will not be inlined. * When you delete the pragmas, GHC uses a size-based heuristic to decide when to inline a function at a call site. Oddly, this heuristic decides (just) to inline the un-specialised function, but when it's specialised GHC thinks it's bigger and doesn't inline it. You can increase the threshold for inlining with -funfolding-use-threshold=20, say. http://www.haskell.org/ghc/docs/latest/html/users_guide/options-optimise .html#options-f This really is a bit odd, but I don't want to fiddle with GHC's inlining heuristics without doing lots of benchmark runs etc. Actually, it'd be quite an easy thing for others to try fiddling with, and I bet there are performance gains to be had. Simon | -----Original Message----- | From: Joel Reymont [mailto:joelr1@gmail.com] | Sent: 22 December 2005 14:09 | To: Haskell-Cafe Cafe | Cc: Simon Peyton-Jones; Simon Marlow | Subject: GHC optimization issue | | Folks, | | I have been trying to improve my byte swapping routines as part of my | effort to speed up serialization. I then tried to look at the core | output from GHC to see what it was converting my code into. Brandon | (skew on #haskell) helped me code a TH version but then I went with a | regular CPP version instead. | | The point of contention is that the logical approach to optimization | does not produce expected results. The only difference between the | following two core outputs is the presence or absence of specialize | and inline pragmas. The code is the same. | | The version without pragmas inlines foo1 into | | Main.foo1 :: GHC.Word.Word16 | [GlobalId] | [Str: DmdType] | Main.foo1 = case GHC.Word.Word16 GHC.Word.$wshift3 __word 3855 (-8) | of ww1_a2eo { __DEFAULT -> | case GHC.Word.Word16 GHC.Word.$wshift3 __word 15 8 of | ww_a2f4 { __DEFAULT -> | GHC.Word.W16# (GHC.Prim.narrow16Word# | (GHC.Prim.plusWord# ww1_a2eo ww_a2f4)) | } | } | | Whereas the version _with_ pragmas produces a function call: | | Main.foo1 :: GHC.Word.Word16 | [GlobalId] | [Str: DmdType] | Main.foo1 = case GHC.Word.Word16 Swap1.$w$sswap16 __word 3855 | of ww1_s1Ia { __DEFAULT -> | GHC.Word.W16# ww1_s1Ia | } | | Is there a reasonable explanation? | | Both versions compiled thusly: ghc --make Foo.hs -O -ddump-simpl > foo | | Swap: | | {-# OPTIONS_GHC -fglasgow-exts -cpp #-} | module Swap1 | ( | swap16 | ) | where | | import Data.Word | import Data.Int | import Data.Bits | | #define BIG_ENDIAN 1 | | {-# SPECIALIZE swap16 :: Word16 -> Word16 #-} | {-# SPECIALIZE swap16 :: Int16 -> Int16 #-} | {-# INLINE swap16 #-} | | swap16 :: Bits a => a -> a | #ifdef BIG_ENDIAN | swap16 v = (v `shiftR` 8) + ((v .&. 0xFF) `shiftL` 8) | #else | swap16 v = v | #endif | | Foo: | | module Main where | | import Data.Word | import Swap1 | | foo1 :: Word16 | foo1 = swap16 0x0f0f | | main = putStrLn $ show foo1 | | Core output for the version WITH pragmas, see the version w/o pragmas | way below | | ==================== Tidy Core Rules ==================== | "SPEC Swap1.swap16" __forall {$dBits_X1Co :: {Data.Bits.Bits | GHC.Int.Int16}} | Swap1.swap16 @ GHC.Int.Int16 $dBits_X1Co | = Swap1.$sswap16 ; | "SPEC Swap1.swap16" __forall {$dBits_X1Cv :: {Data.Bits.Bits | GHC.Word.Word16}} | Swap1.swap16 @ GHC.Word.Word16 $dBits_X1Cv | = Swap1.$sswap161 ; | | | ==================== Tidy Core ==================== | Main.foo1 :: GHC.Word.Word16 | [GlobalId] | [Str: DmdType] | Main.foo1 = case GHC.Word.Word16 Swap1.$w$sswap16 __word 3855 | of ww1_s1Ia { __DEFAULT -> | GHC.Word.W16# ww1_s1Ia | } | | ==================== Tidy Core ==================== | Swap1.$w$sswap16 :: GHC.Prim.Word# -> GHC.Prim.Word# | [GlobalId] | [Arity 1 | NoCafRefs | Str: DmdType L] | Swap1.$w$sswap16 = \ (ww_s1I5 :: GHC.Prim.Word#) -> | case GHC.Prim.Word# GHC.Word.$wshift3 ww_s1I5 | (-8) of ww1_a1Gx { __DEFAULT -> | case GHC.Prim.Word# GHC.Word.$wshift3 | (GHC.Prim.and# ww_s1I5 __word 255) 8 | of ww11_a1Hd { __DEFAULT -> | GHC.Prim.narrow16Word# (GHC.Prim.plusWord# | ww1_a1Gx ww11_a1Hd) | } | } | | Swap1.$w$sswap161 :: GHC.Prim.Int# -> GHC.Prim.Int# | [GlobalId] | [Arity 1 | NoCafRefs | Str: DmdType L] | Swap1.$w$sswap161 = \ (ww_s1HV :: GHC.Prim.Int#) -> | case GHC.Prim.Int# GHC.Int.$wshift2 ww_s1HV | (-8) of ww1_a1Fk { __DEFAULT -> | case GHC.Prim.Int# GHC.Int.$wshift2 | (GHC.Prim.word2Int# | (GHC.Prim.and# (GHC.Prim.int2Word# ww_s1HV) __word 255)) | 8 | of ww11_a1G5 { __DEFAULT -> | GHC.Prim.narrow16Int# (GHC.Prim.+# ww1_a1Fk | ww11_a1G5) | } | } | | Swap1.$sswap16 :: GHC.Int.Int16 -> GHC.Int.Int16 | [GlobalId] | [Arity 1 | Worker Swap1.$w$sswap161 | NoCafRefs | Str: DmdType U(L)m] | Swap1.$sswap16 = __inline_me (\ (w_s1HT :: GHC.Int.Int16) -> | case GHC.Int.Int16 w_s1HT of w1_X1I5 | { GHC.Int.I16# ww_s1HV -> | case GHC.Int.Int16 Swap1.$w$sswap161 | ww_s1HV of ww1_s1I0 { __DEFAULT -> | GHC.Int.I16# ww1_s1I0 | } | }) | | Swap1.$sswap161 :: GHC.Word.Word16 -> GHC.Word.Word16 | [GlobalId] | [Arity 1 | Worker Swap1.$w$sswap16 | NoCafRefs | Str: DmdType U(L)m] | Swap1.$sswap161 = __inline_me (\ (w_s1I3 :: GHC.Word.Word16) -> | case GHC.Word.Word16 w_s1I3 of | w1_X1Ih { GHC.Word.W16# ww_s1I5 -> | case GHC.Word.Word16 Swap1.$w | $sswap16 ww_s1I5 of ww1_s1Ia { __DEFAULT -> | GHC.Word.W16# ww1_s1Ia | } | }) | | Swap1.swap16 :: forall a_a1aQ. (Data.Bits.Bits a_a1aQ) => a_a1aQ -> | a_a1aQ | [GlobalId] | [Arity 1 | NoCafRefs | Str: DmdType L] | Swap1.swap16 = __inline_me (\ (@ a_a1dB) | ($dBits_a1ip :: {Data.Bits.Bits | a_a1dB}) -> | let { | $dNum_s1BO :: {GHC.Num.Num a_a1dB} | [Str: DmdType {a1ip->U | (SAAAAAAAAAAAAAAAAA)}] | $dNum_s1BO = Data.Bits.$p1Bits @ | a_a1dB $dBits_a1ip } in | let { | lit_s1BN :: a_a1dB | [Str: DmdType {a1ip->U | (SAAAAAAAAAAAAAAAAA) s1BO->U(AAAAAAAAS)}] | lit_s1BN = GHC.Num.fromInteger @ | a_a1dB $dNum_s1BO (GHC.Num.S# 255) | } in | \ (v_a1aW :: a_a1dB) -> | GHC.Num.+ | @ a_a1dB | $dNum_s1BO | (Data.Bits.shiftR @ a_a1dB | $dBits_a1ip v_a1aW (GHC.Base.I# 8)) | (Data.Bits.shiftL | @ a_a1dB | $dBits_a1ip | (Data.Bits..&. @ a_a1dB | $dBits_a1ip v_a1aW lit_s1BN) | (GHC.Base.I# 8))) | | | Core output without pragmas: | | Main.foo1 :: GHC.Word.Word16 | [GlobalId] | [Str: DmdType] | Main.foo1 = case GHC.Word.Word16 GHC.Word.$wshift3 __word 3855 (-8) | of ww1_a2eo { __DEFAULT -> | case GHC.Word.Word16 GHC.Word.$wshift3 __word 15 8 of | ww_a2f4 { __DEFAULT -> | GHC.Word.W16# (GHC.Prim.narrow16Word# | (GHC.Prim.plusWord# ww1_a2eo ww_a2f4)) | } | } | | --- | Swap1.swap16 :: forall a_a1aQ. (Data.Bits.Bits a_a1aQ) => a_a1aQ -> | a_a1aQ | [GlobalId] | [Arity 1 | NoCafRefs | Str: DmdType L] | Swap1.swap16 = \ (@ a_a1dB) ($dBits_a1ip :: {Data.Bits.Bits a_a1dB}) -> | let { | $dNum_s1BK :: {GHC.Num.Num a_a1dB} | [Str: DmdType] | $dNum_s1BK = case {GHC.Num.Num a_a1dB} $dBits_a1ip | of tpl_B1 { Data.Bits.:DBits tpl1_B2 | tpl2_B3 tpl3_B4 tpl4_B5 tpl5_B6 tpl6_B7 tpl7_B8 tpl8_B9 tpl9_Ba | tpl10_Bb tpl11_Bc tpl12_Bd tpl13_Be tpl14_Bf tpl15_Bg tpl16_Bh | tpl17_Bi tpl18_Bj -> | tpl1_B2 | } } in | let { | lit_s1BM :: a_a1dB | [Str: DmdType] | lit_s1BM = case a_a1dB $dNum_s1BK | of tpl_B1 { GHC.Num.:DNum tpl1_B2 | tpl2_B3 tpl3_B4 tpl4_B5 tpl5_B6 tpl6_B7 tpl7_B8 tpl8_B9 tpl9_Ba -> | tpl9_Ba Swap1.lvl1 | } | } in | \ (v_a1aW :: a_a1dB) -> | tpl3_B4 | (case a_a1dB $dBits_a1ip | of tpl10_Xy { Data.Bits.:DBits tpl11_XB | tpl12_XD tpl13_XF tpl14_XH tpl15_XJ tpl16_XL tpl17_XN tpl18_XP | tpl19_XR tpl20_Bb tpl21_Bc tpl22_Bd tpl23_Be tpl24_Bf tpl25_Bg | tpl26_Bh tpl27_Bi tpl28_Bj -> | tpl26_Bh v_a1aW Swap1.lvl | }) | (case a_a1dB $dBits_a1ip | of tpl10_Xy { Data.Bits.:DBits tpl11_XB | tpl12_XD tpl13_XF tpl14_XH tpl15_XJ tpl16_XL tpl17_XN tpl18_XP | tpl19_XR tpl20_Bb tpl21_Bc tpl22_Bd tpl23_Be tpl24_Bf tpl25_Bg | tpl26_Bh tpl27_Bi tpl28_Bj -> | tpl25_Bg (tpl12_XD v_a1aW lit_s1BM) Swap1.lvl | }) | | -- | http://wagerlabs.com/ | | | |