program with GADTs segfaults when compiled with -prof and -auto-all

[guest]

Consider the following modules Main.hs and M.hs:

module Main ( main )

where

import M

main :: IO ()
main = do let f = T [undefined, undefined]
          [a] <- c f (\t -> return [r (UR undefined) t])
          a

{-# OPTIONS_GHC -fglasgow-exts -fno-warn-incomplete-patterns #-}
module M ( R(..), U, r, T(..), c )

where

data T f where T :: [T f] -> T ()

data R t where
    UR  :: (T f -> ()) -> R (U f)
    R :: R () 

data U f

c ::  T f -> (T () -> a) -> a
c p@(T _) f = f p

r :: R (U f) -> T f -> IO ()
r a f = (fu a f) `seq` return ()

fu :: R (U f) -> T f -> ()
fu (UR f) = f

I get:

$ ghc --make -prof -auto-all Main
$ ./Main +RTS -p
Segmentation fault
$./Main
Segmentation fault

This is a boiled-down example taken from a large project. A couple of observations:

• I haven't been able to reproduce it using only one module

• The R constructor from type R is not used, but if removed, the segfault no longer happens

• If -auto-all is not used, the segfault does not happen

• I've been able to reproduce it both on linux and os x, using ghc 6.6.1

[sorear]

Thanks for the report.

I can reproduce this with 6.6.1 and 6.7.20070223, but not with 6.7.20070323 or any later version (fail with error: Prelude.undefined).

I don't see any patches containing the word 'crash' or 'prof' in that span, though.

[simonpj]

I've just tried this again and it seems to work fine with the 6.8 branch and with the HEAD. Rather than chase further, I'm going to downgrade this to 'low', with the intention of not looking at it further unless it crops up again. It'd be nicer to chase down exactly where the seg-fault occurs, but life is too short!

It might be worth revisiting if we ever released 6.6.2 (which we don't currently plan to).

Simon

[simonmar]

We won't be doing 6.6.2.