Commit 6f990c54 authored by Ben Gamari's avatar Ben Gamari Committed by Ben Gamari
Browse files

cmm/CBE: Fix comparison between blocks of different lengths

Previously CBE computed equality by taking the lists of middle nodes of
the blocks being compared and zipping them together. It would then map
over this list with the equality relation, and accumulate the result.

However, this is completely wrong: Consider what will happen when we
compare a block with no middle nodes with one with one or more. The
result of `zip` will be empty and consequently the pass may conclude
that the two are indeed equivalent (if their last nodes also match).
This is very bad and the cause of #14361.

The solution I chose was just to write out an explicit recursion, like I
distinctly recall considering doing when I first wrote this code.
Unfortunately I was feeling clever at the time.

Unfortunately this case was just rare enough not to be triggered by the
testsuite. I still need to find a testcase that doesn't have external
dependencies.

Test Plan: Need to find a more minimal testcase

Reviewers: austin, simonmar, michalt

Reviewed By: michalt

Subscribers: michalt, rwbarton, thomie, hvr

GHC Trac Issues: #14361

Differential Revision: https://phabricator.haskell.org/D4152
parent 4dfb790c
......@@ -371,11 +371,15 @@ eqBlockBodyWith dflags eqBid block block'
(_,m',l') = blockSplit block'
nodes' = filter (not . dont_care) (blockToList m')
(env_mid, eqs_mid) =
List.mapAccumL (\acc (a,b) -> eqMiddleWith dflags eqBid acc a b)
emptyUFM
(List.zip nodes nodes')
equal = and eqs_mid && eqLastWith eqBid env_mid l l'
eqMids :: LocalRegMapping -> [CmmNode O O] -> [CmmNode O O] -> Bool
eqMids env (a:as) (b:bs)
| eq = eqMids env' as bs
where
(env', eq) = eqMiddleWith dflags eqBid env a b
eqMids env [] [] = eqLastWith eqBid env l l'
eqMids _ _ _ = False
equal = eqMids emptyUFM nodes nodes'
eqLastWith :: (BlockId -> BlockId -> Bool) -> LocalRegMapping
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment