Semantically-equivalent expressions evaluating at different values with -O1

The following code shows inconsistent results when compiled with -O1:

import Data.Bits ((.&.))

flags :: Int -> [Int]
flags x = filter (\y -> x .&. y > 0) [1, 128, 129, 255]

main :: IO ()
main = do
  putStrLn $ show $ flags $ 255   -- These lines should display
  putStrLn  (show  (flags   255)) -- the same value, right?

-- It appears that this is related to something that started
-- with GHC 7.10, but was also changed in GHC 8.0, as we can
-- see from the following tests:
--
-- Stack snapshot lts-0.7 (GHC 7.8.3), with -O2
-- [1,128,129,255]
-- [1,128,129,255]
--
-- Stack snapshot lts-2.22 (GHC 7.8.4) with -O2
-- [1,128,129,255]
-- [1,128,129,255]
--
-- Stack snapshot lts-3.22 (GHC 7.10.2) with -O2
-- [1]
-- [1]
--
-- Stack snapshot lts-6.30 (GHC 7.10.3) with -O2
-- [1]
-- [1]
--
-- Stack snapshot lts-7.20 (GHC 8.0.1) with -O2
-- [1]
-- [1,128,129,255]
--
-- Stack snapshot lts-8.5 (GHC 8.0.2) with -O2
-- [1]
-- [1,128,129,255]
--
-- In GHC 8.0.2, compiling with any of the following...
--   -O0
--   -O2 -fno-cmm-sink
--   -O2 -fno-enable-rewrite-rules
--
-- ...we get the expected results:
-- [1,128,129,255]
-- [1,128,129,255]
--
-- Using only -O1, we get the abnormal behavior again:
-- [1]
-- [1,128,129,255]

Trac metadata

Trac field	Value
Version	8.0.2
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture

changed weight to 5

added Tbug Trac import labels

That is deeply strange. I have established

• It happens with HEAD.
• The combination of -ffull-laziness and -fspec-constr makes it produce the wrong answer [1]. Omitting either makes it give the right answer [1,128,129,255].
• Here are my command lines. I switched off everything I could. Failure case:

simonpj@cam-05-unx:~/tmp$ ~/5builds/HEAD-5/inplace/bin/ghc-stage1 -dcore-lint -O2 -o T13425 T13425.hs -fforce-recomp -fno-enable-rewrite-rules -fno-liberate-case -fno-strictness -fno-cse -fno-call-arity  -fno-specialise -fno-float-in -fno-late-dmd-anal
[1 of 1] Compiling Main             ( T13425.hs, T13425.o )
Linking T13425 ...
simonpj@cam-05-unx:~/tmp$ ./T13425
[1]

Switch off full laziness (works)

simonpj@cam-05-unx:~/tmp$ ~/5builds/HEAD-5/inplace/bin/ghc-stage1 -dcore-lint -O -o T13425 T13425.hs -fforce-recomp -fno-enable-rewrite-rules -fno-liberate-case -fno-strictness -fno-cse -fno-call-arity  -fno-specialise -fno-float-in -fno-late-dmd-anal -fno-full-laziness
[1 of 1] Compiling Main             ( T13425.hs, T13425.o )
Linking T13425 ...
simonpj@cam-05-unx:~/tmp$ ./T13425
[1,128,129,255]

Switch off spec-constr (works)

simonpj@cam-05-unx:~/tmp$ ~/5builds/HEAD-5/inplace/bin/ghc-stage1 -dcore-lint -O -o T13425 T13425.hs -fforce-recomp -fno-enable-rewrite-rules -fno-liberate-case -fno-strictness -fno-cse -fno-call-arity  -fno-specialise -fno-float-in -fno-late-dmd-anal -fno-spec-constr
[1 of 1] Compiling Main             ( T13425.hs, T13425.o )
Linking T13425 ...
simonpj@cam-05-unx:~/tmp$ ./T13425
[1,128,129,255]

I'm out the rest of today. This is a bad bug; might someone look?

I'll take a peek today.

Trac metadata

Trac field	Value
CC	- → dfeuer

Reid thinks he's found the ultimate source, in an optimization he made intended to help with tag bit checks. Here's a greatly reduced reproduction case:

import Data.Bits ((.&.))

flags :: Int -> Int
flags x
  | x .&. 128 > 0 = 12
  | otherwise = 13
{-# NOINLINE flags #-}

main :: IO ()
main = print (flags 255)

changed milestone to %8.2.1

Yep, sorry about that!

Trac metadata

Trac field	Value
Differential revisions	- → D3359

mentioned in commit caf94b06

closed

changed weight to 7

Merged to ghc-8.2 as bdb0c43c.

Trac metadata

Trac field	Value
Priority	normal → high
Resolution	Unresolved → ResolvedFixed

added incorrect runtime result label

added Phigh label