I think another way to go about this problem is to figure out an alternative to baking in DynFlags to SDocContext (which I feel is the core problem here). The only use of those DynFlags is via sdocWithDynFlags and 94 call sites use them. - In the frontend, it's used to check for the presence of the flag (like suppress module prefixes, etc.) - In the code generator, it's used to get the word size, endianness, and other platform specific stuff for platform-specific printing. - Backpack-related stuff needed to get the package state - Used in exactly 2 cases: Outputable instances for ComponentId and InstalledUnitId

From what I observed with the majority of use cases, sdocWithDynFlags is used to obviate the need for passing dflags to various ppr* functions, which is a good idea since without it, we'd probably have to pass around DynFlags to a whole lot more pure functions throughout the codebase.

So as others have said, Show instances are just not practical because printing many of the GHC types is highly dependent on the platform and what flags GHC was invoked with. There are three solutions here: 1.) Figure out a subset of DynFlags (flags, platform details, package state) and only allow those inside of SDocContext and extend SDocContext as new use cases come up. This is probably not practical as it would require sweeping changes. 2.) Provide a stock set of DynFlags for the purpose of printing with Outputable. It's easy to do for flags and platform details, but tricky to do for package state. This seems to be the most reasonable solution if some sane substitute for package state can be used. Syd, can you tell us what kind of things you were trying to print out? You can try to pass in unsafeGlobalDynFlags but it may not be what you want. It gets written to on the initialisation of the GHC monad and after the command line options are parsed (so everything will be properly initialised except for package state). Hope that helps, Rahul On Sun, Mar 19, 2017 at 1:14 AM, Edward Z. Yang wrote:

We can't add Show instances for these types because many types below them, e.g., Type, are cyclic, and would result in infinite output.

Perhaps we can add a new type class which a) faithfully represents the Haskell syntax, but b) can deal with cyclic data. I think that's something people would like (extra compilation time not withstanding). But it sounds annoying to do since the deriving mechanism is not going to help you.

Edward

Excerpts from Tom Sydney Kerckhove's message of 2017-03-18 14:03:48 +0100:

...

Dear GHC Devs,

I am trying to use GHC as a library but I'm having a lot of trouble with understanding what everything means. Up to now, I have been able to figure out what to do by reading the sources, but it ocured to me that much of my struggles could have been mitigated if the relevant types had Show instances.

I am specifically talking about the types concerning type checking. TypecheckedModule and everything below that. I am aware that most of the types have an Outputable instance, but there are two problems with that:

- 'Outputting' a value requires DynFlags. (yes, I know about pprTrace) - These instances are not intended to show the internal structure of a value, but rather a 'human readable' representation of a value.

My questions for you:

- Is there a reason that there are no derived 'Show' instances for most types? - Would you accept a diff that adds these?

Thank you for your time.

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

-- Rahul Muttineni

Hi Rahul, This ticket may be of interest: https://ghc.haskell.org/trac/ghc/ticket/10143 Edward Excerpts from Rahul Muttineni's message of 2017-03-19 02:08:56 +0530:

I think another way to go about this problem is to figure out an alternative to baking in DynFlags to SDocContext (which I feel is the core problem here). The only use of those DynFlags is via sdocWithDynFlags and 94 call sites use them.

- In the frontend, it's used to check for the presence of the flag (like suppress module prefixes, etc.) - In the code generator, it's used to get the word size, endianness, and other platform specific stuff for platform-specific printing. - Backpack-related stuff needed to get the package state - Used in exactly 2 cases: Outputable instances for ComponentId and InstalledUnitId

From what I observed with the majority of use cases, sdocWithDynFlags is used to obviate the need for passing dflags to various ppr* functions, which is a good idea since without it, we'd probably have to pass around DynFlags to a whole lot more pure functions throughout the codebase.

So as others have said, Show instances are just not practical because printing many of the GHC types is highly dependent on the platform and what flags GHC was invoked with.

There are three solutions here: 1.) Figure out a subset of DynFlags (flags, platform details, package state) and only allow those inside of SDocContext and extend SDocContext as new use cases come up. This is probably not practical as it would require sweeping changes. 2.) Provide a stock set of DynFlags for the purpose of printing with Outputable. It's easy to do for flags and platform details, but tricky to do for package state. This seems to be the most reasonable solution if some sane substitute for package state can be used.

Syd, can you tell us what kind of things you were trying to print out? You can try to pass in unsafeGlobalDynFlags but it may not be what you want. It gets written to on the initialisation of the GHC monad and after the command line options are parsed (so everything will be properly initialised except for package state).

Hope that helps, Rahul

On Sun, Mar 19, 2017 at 1:14 AM, Edward Z. Yang wrote:

...

We can't add Show instances for these types because many types below them, e.g., Type, are cyclic, and would result in infinite output.

Perhaps we can add a new type class which a) faithfully represents the Haskell syntax, but b) can deal with cyclic data. I think that's something people would like (extra compilation time not withstanding). But it sounds annoying to do since the deriving mechanism is not going to help you.

Edward

Excerpts from Tom Sydney Kerckhove's message of 2017-03-18 14:03:48 +0100:

...

Dear GHC Devs,

I am trying to use GHC as a library but I'm having a lot of trouble with understanding what everything means. Up to now, I have been able to figure out what to do by reading the sources, but it ocured to me that much of my struggles could have been mitigated if the relevant types had Show instances.

I am specifically talking about the types concerning type checking. TypecheckedModule and everything below that. I am aware that most of the types have an Outputable instance, but there are two problems with that:

- 'Outputting' a value requires DynFlags. (yes, I know about pprTrace) - These instances are not intended to show the internal structure of a value, but rather a 'human readable' representation of a value.

My questions for you:

- Is there a reason that there are no derived 'Show' instances for most types? - Would you accept a diff that adds these?

Thank you for your time.

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

Tom Sydney Kerckhove writes:

On 19-03-17 02:08:56, Rahul Muttineni wrote:

...

Syd, can you tell us what kind of things you were trying to print out?

Maybe I wasn't very clear. I'm trying to visualise the internal structure of some of the typechecker's output. I specifically do NOT need to see the output of Outputable's functions. They show the human-readibly version and not the internal structure.

Indeed I am sympathetic to this request. In my time working on GHC I have written raw several variants of `Type -> SDoc`, each exposing various levels of detail. These are handy and can be a good way to gain insight into the AST, but I feel like it is hard to come up with something that is generally applicable; I find each time I need to expose slightly different details about the internal structure of the representation. Cheers, - Ben

On 18-03-17 16:13:52, Ben Gamari wrote:

On March 18, 2017 9:03:48 AM EDT, Tom Sydney Kerckhove wrote:

Snip.

...

My questions for you:

- Is there a reason that there are no derived 'Show' instances for most types?

As Richard mentioned, we don't derive Show due to code size and compilation time concerns.

Okay.

Show in particular is rather expensive to derive and seeing as we already have Outputable I don't it would make sense to derive it by default.

Show and Outputable have very different goals though.

I would really like to avoid introducing more CPP into the code base for this particular problem.

Fair enough.

One alternative which will work in many cases is to simply derive Show yourself using StandaloneDeriving. Does this help?

That doesn't work if some type doesn't have the constructors exposed. I tried this already, and it would be a good solution if all constructors were exposed, ...

Cheers,

- Ben

-- Sent from my Android device with K-9 Mail. Please excuse my brevity.

-- Tom Sydney Kerckhove

Hi, Am Montag, den 20.03.2017, 12:51 +0000 schrieb Simon Peyton Jones via ghc-devs:

Show and Outputable have very different goals though.

Really? What's wrong with using Outputable, plus, as Joachim says, 

showSDocUnsafe . ppr :: Outputable a => a -> String

Maybe you want to really really see the precise data constructors used.  But for the most part the Outputable instance tells you that, but much more legibly.

and if using Outputable is a bit annoying sometimes (e.g. I found it hard to get the IdInfo of an Id that is not part of a binder, if I recall correctly, because the default instance does not include that info, and the pretty printer for “id with idinfo” is not exported), well, we can always improve that by adding a few more helpful functions to Outputable. Greetings, Joachim -- Joachim “nomeata” Breitner   mail@joachim-breitner.de • https://www.joachim-breitner.de/   XMPP: nomeata@joachim-breitner.de • OpenPGP-Key: 0xF0FBF51F   Debian Developer: nomeata@debian.org