#13227: Loss of full laziness in mapFB
-------------------------------------+-------------------------------------
Reporter: simonpj | Owner:
Type: bug | Status: new
Priority: normal | Milestone:
Component: Compiler | Version: 8.0.1
Keywords: | Operating System: Unknown/Multiple
Architecture: | Type of failure: None/Unknown
Unknown/Multiple |
Test Case: | Blocked By:
Blocking: | Related Tickets:
Differential Rev(s): | Wiki Page:
-------------------------------------+-------------------------------------
I've just discovered this
{{{
g4 x expensive ys = let h = \7 -> y + expensive x
in map h ys
}}}
Of course we'd expect the `(expensive x)` call to be floated out of the
`\y`-abstraction, so that it is only done once.
But it isn't! Here's the simplified core with -O:
{{{
g4 = \ (@ b_aPG)
(@ p_atr)
($dNum_aPP :: Num b)
(x_arW :: p)
(expensive_arX :: p -> b)
(ys_arY :: [b]) ->
map @ b @ b
(\ (y_as0 [OS=ProbOneShot] :: b) ->
+ @ b $dNum_aPP y_as0 (expensive_arX x_arW))
ys_arY
}}}
Yikes! What is happening?
Answer: look at that suspicious `ProbOneShot` on the `\y` above. Read
`Note [Computing one-shot info, and ProbOneShot]` in `Demand.hs`.
When `FloatOut` runs we have
{{{
g4 = \ (@ b_aPL)
(@ p_atw)
($dNum_aPU :: Num b)
(x_as1 :: p)
(expensive_as2 :: p -> b)
(ys_as3 :: [b]) ->
GHC.Base.build @ b
(\ (@ b1_aQs)
(c_aQt [OS=OneShot] :: b -> b1 -> b1)
(n_aQu [OS=OneShot] :: b1) ->
GHC.Base.foldr @ b @ b1
(GHC.Base.mapFB @ b @ b1 @ b
c_aQt
(\ (y_as5 [OS=ProbOneShot] :: b) ->
+ @ b $dNum_aPU y_as5 (expensive_as2 x_as1)))
n_aQu
ys_as3)
}}}
Why is the `\y` marked `ProbOneShot`? Because the occurrence analyser
marked it so, based on the cardinality info from `mapFB`, even though
`mapFB` was not saturated.
So `Demand.argsOneShots` makes a deliberate choice to play risky, and that
choice backfires badly for use of `map`. Not good!
The offending commit, which introduced this behaviour, is (I think)
{{{
commit 80989de947dc7edb55999456d1c1e8c337efc951
Author: Simon Peyton Jones