[GHC] #4900: DEPENDS pragma

[GHC]

#4900: DEPENDS pragma
-------------------------------+--------------------------------------------
  Reporter:  cdsmith           |          Owner:                  
      Type:  feature request   |         Status:  new             
  Priority:  normal            |      Milestone:  7.6.1           
 Component:  Compiler          |        Version:                  
Resolution:                    |       Keywords:                  
        Os:  Unknown/Multiple  |   Architecture:  Unknown/Multiple
   Failure:  None/Unknown      |     Difficulty:  Unknown         
  Testcase:  TH_Depends        |      Blockedby:                  
  Blocking:                    |        Related:                  
-------------------------------+--------------------------------------------

Comment(by parcs):

 Replying to [comment:53 simonmar]:
 > I looked at the new patch, and maybe I'm being stupid here, but it looks
 like my previous concerns still apply.  Looking at the HPT during the
 summarise stage rings alarm bells - the `hsc_env` should only be used for
 its `DynFlags` and the finder cache.  So I'm still not sure how to reason
 that this patch works correctly.

 Regarding your previous concerns, the important part of this new patch is
 that `ms_uf_date` is not read at all during the summarize stage; it is
 only used to propagate the latest usage file date computed during the
 summarize stage to the stability check. Instead of comparing the previous
 `ms_uf_date` to the latest `ms_uf_date`, as the last patch did, the
 `usg_mtime`s are compared and calculated piecewise from the module's
 `HMI`. This is better because these mtimes get calculated during the
 `upsweep`, meaning that they will be as accurate as they can be for the
 next `load` cycle. (Compare the current `isOkUsageFileDate` with the
 previous patch's `getUsageFileDate`).

 I _believe_ that reading the `HPT` in the way that is done in the patch is
 safe -- it is only read if an old summary is available. Is it guaranteed
 that if an old summary is available, then a corresponding `HMI` exists
 within the `HPT`? It seems so, but I am not sure.

 Also, I should mention that it is necessary to compute and check the usage
 file dates before or during the summarise stage, because that's when a
 file is ran through the preprocessor, and a usage file may be a CPP
 include. Without the `HPT`, the only real information you have at that
 point is an old `ModSummary`. Therefore, if the `HPT` should not be read
 during the summarise stage, then it seems that the only alternative would
 be to stuff the `UsageFile` data from the `HMI` to a field of the
 corresponding `ModSummary`. An `HMI` is created during the `upsweep`
 stage, so the `hsc_mod_graph` would have to be updated after the `upsweep`
 occurs, which is currently not done and doing so would probably introduce
 its own subtleties. (Currently, the mod graph is only updated right after
 the summarise stage). Does this seem more reasonable to you?

-- 
Ticket URL: <http://hackage.haskell.org/trac/ghc/ticket/4900#comment:55>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler