Use -fspec-constr-keen GHC option

[harendra-kumar]

The terminating fold changes (PR #488) cause regression for several benchmarks. One of the benchmarks (copy/read/rawToNull) was found be improved using -fspec-constr-keen option. The following code:

import qualified Streamly.Prelude as S
import qualified Streamly.FileSystem.Handle as FH
import System.IO (openFile, IOMode(..))

main :: IO ()
main = do
    inh <- openFile "benchmark-tmpin-100MB.txt" ReadMode
    outh <- openFile "/dev/null" WriteMode
    S.fold (FH.write outh) (S.unfold FH.read inh)

generates a core where a W8 is getting boxed in a function, in which it is not inspected, but it is inspected in a function called by it. This caused the Word8 to not get unboxed in a tight loop processing bytes in an array.

                jump $wstep3_s9yD
                  sc10_s9Rn
                  sc9_s9Ro
                  sc8_s9Rp
                  sc7_s9Rq
                  sc6_s9Rr
                  (W8# ipv8_a8uK)
                  sc_s9Rx

unboxed in

                  exit1_X19 ww_s9yr
                            ww1_s9yw
                            ww2_s9yx
                            ww3_s9yy
                            ww4_s9yB
                            w_s9yh
                            w1_s9yi
                    = case w_s9yh of { W8# x_a8tC ->

-fspec-constr-keen removes this boxing. However, some benchmarks get worse due to this option. For example (allocations):

Benchmark                                                                                                     default(0)(MiB) default(1) - default(0)(%)
------------------------------------------------------------------------------------------------------------- --------------- --------------------------
FileSystem.Handle/o-1-space/copy/read/group-ungroup/S.interposeSuffix . S.splitOnSuffix(Array Word8) (1/10)             76.23                    +740.47
FileSystem.Handle/o-1-space/copy/read/group-ungroup/UA.unlines . UA.lines (Array Char) (1/10)                          225.39                    +250.78
FileSystem.Handle/o-1-space/reduce/read/S.splitOnSeq "\n" FL.drain                                                     101.44                     +37.00
FileSystem.Handle/o-1-space/reduce/read/S.splitOnSeq "a" FL.drain                                                      101.44                     +36.89
FileSystem.Handle/o-1-space/copy/read/group-ungroup/UA.unwords . UA.words (Array Char) (1/10)                         1574.39                     +22.09

Data.Fold/o-n-heap/serially/elimination/writeN                        0.00                  +Infinity
Data.Fold/o-n-heap/serially/elimination/lastN.Max               1602136.00                    +195.57

Memory.Array/o-1-space/elimination/toStreamRev                              0.00                  +Infinity
Memory.Array/o-1-space/elimination/length . IsList.toList                   0.00                  +Infinity
Memory.Array/o-1-space/elimination/min                                      0.00                  +Infinity
Memory.Array/o-1-space/elimination/<                                        0.00                  +Infinity
Memory.Array/o-1-space/elimination/id                                       0.00                  +Infinity
Memory.Array/o-1-space/generation/writeN . unfoldr                          0.00                  +Infinity
Memory.Array/o-1-space/generation/writeN . intFromTo                        0.00                  +Infinity
Memory.Array/o-1-space/generation/writeN . fromList                   6989224.00                     +29.78

Prelude.Serial/o-1-space/Applicative/(<*) (sqrt n x sqrt n)                     9401184.00                     +32.68
Prelude.Serial/o-n-stack/iterated/filterEven (n/10 x 10)                        4766320.00                     +22.07
Prelude.Serial/o-1-space/Monad/(>>) (sqrt n x sqrt n)                          10392416.00                     +10.23
Prelude.Serial/o-1-space/Applicative/(*>) (sqrt n x sqrt n)                    10392416.00                     +10.23
Prelude.Serial/o-n-heap/buffered/reverse                                        3129384.00                     +66.91
Prelude.Serial/o-n-space/Applicative/(<*) (n times)                            37705048.00                     +14.59

To enable this option by default we need to make sure that there are no significant regressions. Possibly tweak GHC or add some support in fusion-plugin. We could do the following:

1. Allow keen if the argument is getting unboxed in a child call
2. Is it possible to detect if spec constr actually helped in reducing allocations, and only then allow it?

[harendra-kumar]

The unboxing is happening inside the exit function created by the GHC exitification pass. If I use the -fno-exitification then also spec-constr is able to remove the boxing without using -fspec-constr-keen. Perhaps exitification is blocking unboxing by spec-constr. Is it possible to do exitification after spec-constr, can it help?

[harendra-kumar]

See this note. We would not have the spec constructor issue if we use the option D i.e. put exitification before the final simplifier pass.

We can possibly experiment with this using the fusion-plugin, pass -fno-exitification but force insert a pass just before the final simpl pass.

[harendra-kumar]

Placing exitification pass after spec-constr pass works pretty well. But it is not as good as doing it before spec-constr + using spec-constr-keen . It seems the important thing in the placement of exitification is its order wrt spec-constr, the placement wrt to other passes does not seem to have any effect.

Maybe we can teach spec-constr to be keen if a function itself is not examining a constructor but some exit join point inside it is examining it?