Speed up changesUponFlattening

[sjakobi]

Since the first Union alternative was produced by the previous
call to to changesUponFlattening, it seems unnecessary to
scrutinize it again.

This speeds up some prettyprinter-heavy applications in dhall
by 20 to 42%.

This assumes some kind of idempotence in changesUponFlattening.

Fixes #99.

[sjakobi]

This assumes some kind of idempotence in changesUponFlattening.

I believe the property should be

group x == group (group x)

Does this look right @quchen? I could add a property test for this.

[sjakobi]

Ping @quchen.

This could help alleviate some performance issues in dhall, so I'm very interested to hear what you think about this! :)

[sjakobi]

Another idea:

Suggested change

	Union x _ -> Just x
	Union x _ -> Nothing

[sjakobi]

Ah, no. This causes the fusion tests to fail with the classic

»SFail« must not appear in a rendered »SimpleDocStream«.

I'm not quite sure why though. Proper shrinking and diagnostic output would help… :/

[sjakobi]

This assumes some kind of idempotence in changesUponFlattening.

I believe the property should be

group x == group (group x)

Does this look right @quchen? I could add a property test for this.

I have written a test for this on top of #96 at master...sjakobi:99-group-tests.

At least for the Docs generated there, render . group == render . group . group seems to hold.

[sjakobi]

While looking at some Docs with the new Diag, I noticed that even with this patch, repeated groups build up trees of Unions, just a bit faster than before:

> diag $ group . group . group $ hardline 
Union Fail (Union Fail (Union Fail Line))

In the layouter, this will still be slow since any overly wide Doc will be checked repeatedly until we finally decide to go with the innermost second alternative.

This patch would fix that:

@@ -523,6 +523,7 @@ hardline = Line
 -- use of it.
 group :: Doc ann -> Doc ann
 -- See note [Group: special flattening]
+group x@Union{} = x
 group x = case changesUponFlattening x of
     Nothing -> x
     Just x' -> Union x' x

And maybe we can even do something smarter and look whether there's anything flattenable before we reach a first Union…

[sjakobi]

Superseded by #116.