Haddock commit c4a06a877be56d5cd790f8abb7928fa39458e1e4 or/and GHC commit 53c78be0aab76a3107c4dacbb1d177afacdd37fa is/are buggy

[awson]

Current ghc-head Haddock fails to produce language-c-quote-0.12 package documentation barfing something like unknown symbol languagezmczmquotezm0zi12zmU79M5Ik6y66BQup46vL86_LanguageziCziParserziMonad_emptyPState_closure.

Reverting 'Don't enable compilation for template haskell' commit fixes things for me.

Thus either this commit or relevant ghc commit or both is/are buggy. Perhaps, this should be cross-reported to GHC trac.

[alexbiehl]

Thanks @awson! @duog is the expert here, what do you say?

[duog]

It's more likely to be a ghc problem, trac is probably the place. It is probably replicatable by building with ghc --make -fno-code -fwrite-interface. I will look into this when I get a chance.

[awson]

Perhaps, but all the object code (both vanilla and profiled libraries and prelinked object file for ghci) was built by ghc with no problems.

[awson]

Well, doing ghc ... --make -fno-code -fwrite-interface ... I'm getting the same error.

But I don't quite understand how should we describe this bug. What exactly is buggy in ghc?

[alexbiehl]

IIRC this is the change at fault here: ghc/ghc@53c78be Ghc doesn't seem to compile your modules as though it should work this patch. awson <notifications@github.com> schrieb am Do., 22. Juni 2017, 19:07:

…

Well, doing ghc ... --make -fno-code -fwrite-interface ... I'm getting the same error. But I don't quite understand how should we describe this bug. What exactly is buggy in ghc? — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#640 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AByiiXk4Qfwqk0Vs3xiddQL_LaeVw6Ybks5sGp8hgaJpZM4OCXeM> .

[duog]

I made ghc -fno-code smarter: it enables object code generation for modules depended on by modules that use template haskell.

Haddock uses -fno-code, and used to enable object code generation for all modules in a project where template haskell is used at all. The commit you reverted disabled that behavior, because it should be unnecessary when ghc is already generating the necessary object code.

I think what is going wrong is that the quasi-qutoes in language-c-quote should also trigger object code generation.

if you pass -opt-ghc=-asm to haddock, that may work around the issue.

[duog]

trac

[awson]

Thanks.
Btw, this is exactly what I suspected too – template-haskell and quasi-quoting having different relevant code paths.

[alexbiehl]

@awson @duog is on fixing this in https://phabricator.haskell.org/D3677

[alexbiehl]

Alright, ghc/ghc@d55bea1 has all the magic to make it work again. Closing this for now. Feel free to reopen if you think your converns haven't been addressed.