Compilation of my ghc tree with ghc-6.12.2 -dcore-lint fails.

[milan]

In the middle of the work on GHC backend I triggered a bug in GHC 6.12.2. When using 6.12.2 to compile compiler/cmm/CmmSpillReload.hs we get:

*** Core Lint Errors: in result of Simplifier mode gentle, iteration 1 out of 4 ***
	<no location info>:
	    In the expression: ic_a1xa @ $co$_a1wz
	    Kinds don't match in type application:
	    Type variable: co_wild_B1 :: e_a1tx ~ Compiler.Hoopl.Graph.C
	    Arg type: $co$_a1wz :: e_a1zK ~ Compiler.Hoopl.Graph.C

After discussing with simonpj, I have snapshotted the source tree, as I could not reduce the test case.

When compiling with ghc-6.12.2 -dcore-lint, the compilation fails in the stage 1. It is enough to invoke 'make' in the tree to get to the compilation error. I attach the dump of -dcore-lint output.

The file causing the bug is compiler/cmm/CmmSpillReload.hs. On line 131, there is a description of a change that makes the source compile.

[milan]

The dump of ghc -dcore-lint error message.

[milan]

The tree is at /64playpen/t-mistra/ghc-bug-4044. Anyone in GHC group can modify everything.

[BenMoseley]

We're seeing what seems to be this bug too (GHC 6.12.1). (Were getting segfaults in the finished program which compiled fine without "-dcore-lint", the segfaults disappear with "-O0" and appear with "-O1" and "-O2").

Building GenBarriers?

ghc.exe -odir ../Build\obj -hidir ../Build\obj -Wall -fno-warn-missing-signatures -fno-warn-name-shadowing -fno-warn-type-defaults -fno-warn-unused-do-bind -package conf=C:\ws\fpfdev\depot\QA\EDG\EDG_priv\FPF_Dev.br\ThirdParty?/haskell_packages\fpf.package.conf flags.c -O2 -Werror -i.. -XTemplateHaskell --make BackendBarriers/GenBarriers?.hs -main-is BackendBarriers?.GenBarriers?.main -o ..\Out\EXEs\5\GenBarriers?.exe 
*** Core Lint Errors: in result of Simplifier mode gentle, iteration 1 out of 4 *** <no location info>:

In the expression: ic_a700 @ $co$_a6ZE Kinds don't match in type application:

Type variable: co_wild_B1

a_a6I4 ~ Utils.RangeAnalysis?.FC.FCQty

Arg type: $co$_a6ZE

a_a7k7 ~ Utils.RangeAnalysis?.FC.FCQty 
*** Offending Program ***

[simonpj]

Hmm. Milan saw this when compiling (a variant of) GHC itself, and that was pretty hard to reproduce. Maybe your test case is smaller?

Is it possible for you to make a reproducible test case? It's embarassing that 6.12 doesn't work properly, even though I'm busy rewriting that very part of the type checker!

Simon

[BenMoseley]

Sadly it's in one of our more complex bits of code so it would take some effort to try and reduce it to a small example. Unfortunately I don't have the time to investigate further right now.

[simonpj]

OK fair enough. I'm not rushing to fix this if it's not getting in anyone's way, because the new type checker is very different.

[BenMoseley]

No problem, right now it's not a major issue for us. If it becomes so I'll investigate further, try to boil it down to a small reproducible case and post that here.

[simonpj]

This bug was never reproducible and seems to have gone away, so I'm closing it.