S.D.Mechveliani writes:
> Here is an -O -ish bug in ghc-3.02-i386-unknown...
>
....
> remove -Onot and run `make all'
> It reports
> -----------------------------------------------------------------
> several successful compilations, then
>
> Vec0_.hs:22: Warning:
> Warning: failed to find (optional) interface decl for
> `AddSemigroup_sc1'
> desired at
> Semigr_.hi:311
> tcUnfolding failed with: Semigr_.zeroS{-riN-}
> export/Semigr_.hi:311:
> Warning: no binding for Semigr_.AddSemigroup_sc1
>
> - and such things.
>
> Then
>
> tcUnfolding failed with: Semigr_.zeroS{-rh4-}
> export/Semigr_.hi:311:
> Warning: no binding for Semigr_.AddSemigroup_sc1{-r2IX-}
> In an interface-file signature for Semigr_.zeroS{-rh4-}
>
> panic! (the `impossible' happened):
> lookupBindC:no info!
> for: restore_smf5
> (probably: data dependencies broken by an optimisation pass)
> static binds for:
> DPair.$g2{-g2-}
> ...
> -------------------------------------------------------------------------
>
Hi,
thanks for a fine report. Your sources showed up a couple of problems
with 3.02:
- putting {-# INLINE #-} pragmas on overloaded functions doesn't
work properly. (the compiler assumes that all super class selectors
are unfolded away, but this wasn't happening on the RHS of a
function marked as {-# INLINE #-} .)
- percolation of specialisation information was done in a cheap and
cheerful manner for 3.02, something that came back to haunt us when
compiling your sources (witness the lookupBindC panic above :-(
Both of these problems have now been fixed, but in the meantime I
suggest working around the problems as follows with 3.02:
- leave out {-# INLINE #-} pragmas on overloaded functions.
- compile with -fno-specialise the modules that elicits a panic from
lookupBindC (there's two of them in DoCon, I think.)
--Sigbjorn
BTW, I had to change the DoCon Makefile to generate dependencies and
noticed that you're being bitten by a rather bad (but semi-known)
mkdependHS bug - it does not understand Haskell block comments:
--source/auxil/Common_.hs
...
module Common_ ( AlgTypeCons(..) )
where
...
{- OLD ***********************************************************
--
-- The Algebraic type construction is certain expression
-- describing a type.
-- It is either BasicType <mode> <typeName> or
-- <constructor> <operands-parameters>
--
-- <mode> = "f" means a finitely enumerated type, as, say, Char,
-- or the type given by data .. = v1 | ... | vn,
-- "" - arbitrary.
import Group_ ( Properties_subgroup )
import Ring0_ ( Properties_subring, Properties_ideal )
import Ring_ ( PIRChineseIdeal(..) )
...
-}
i.e., the script will here generate dependencies on the above imports,
leading to all sorts of weird loops.
