Changes between Version 3 and Version 4 of Plugins/Phases


Ignore:
Timestamp:
Jun 15, 2008 12:49:12 AM (7 years ago)
Author:
guest
Comment:

--

Legend:

Unmodified
Added
Removed
Modified
  • Plugins/Phases

    v3 v4  
    55We need to be able to specify compiler phases to:
    66
    7  * Be able to specify where user phases should fit
    8  * Control when rules fire
     7 * Be able to specify where user phases from plugins should fit
     8 * Control when rules and inlinings will fire
    99
    10 == Possible Solution ==
     10The current solution where we have phases 0, 1 and 2 is too restrictive to let us do this well, hence this page proposes a generalization that should solve our problems.
     11
     12I believe the proposed system represents a significant improvement in modularity and usability of phase control within GHC.
     13
     14== Solution ==
    1115
    1216{{{
     
    2125
    2226This code has a PHASE declaration which brings a new phase into being. Later rules then use that phase name to control their firing, in contrast to the current system of controlling firing with a limited set of phase numbers.
     27
     28Phase names have the same name format as data constructors. There is no technical reason for this, as there is never any ambiguity as to whether a name is that of a phase: the choice is purely asthetic and could be changed.
    2329
    2430Phase names are exported, so:
     
    5258}}}
    5359
    54 This module explicitly exports its local phase C, which is defined to occur before the SpecConstr phase. However the programmer is totally free to remove it from the exports list and hence prevent other modules from referring to it.
     60This module explicitly exports its local phase C, which is defined to occur before the SpecConstr phase. However the programmer is totally free to remove it from the exports list and hence prevent other modules from referring to it. Likewise, you can selectively import phases:
     61
     62{{{
     63
     64module Baz where
     65
     66import Spqr({-# PHASE C #-})
     67
     68{-# PHASE D > C #-}
     69
     70{-# INLINE [~D] foo #-}
     71
     72foo = ..
     73
     74}}}
    5575
    5676== Expressing Dependence ==
     
    7595
    7696The square brackets are meant to be evocative of optionality in Backus-Naur form, but I'm not yet sure if that is too easily confused with Haskell list syntax.
    77  
     97
     98== Compatability Concerns ==
     99
     100There are two principal concerns:
     101 * Code that assumes the current phase control mechanism where we have phases 0, 1 and 2 should still work in this new system
     102 * Compilers that are unable to parse the PHASE pragma should still be able to deal with source code that uses it
     103
     104To handle these concerns, first we must provide three "wired in" phase names that support the old usage:
     105
     106{{{
     107
     108module Buzz where
     109
     110{-# PHASE E < 1, > 0 #-}
     111
     112{-# INLINE [~0] bar #-}
     113{-# INLINE [E] sqpr #-}
     114
     115bar = ...
     116spqr = ...
     117
     118}}}
     119
     120Note that actually the old syntax allowed arbitrary positive integers to be used, not just the set 0-2. However, supporting an infinite set of wired in names is a bit of a headache and I believe that the higher phase numbers were sufficiently rarely used that supporting them is not a major concern. You currently have to supply an additional flag to the compiler (to change the number of simplifier iterations) to even make the higher phases behave differently than phase 2.
     121
     122The PHASEs 0, 1 and 2 will be implicitly and irrevocably imported into every program GHC compiles. A possible alternate design choice is to have them live in the Prelude, so e.g. you can get rid of them by e.g. explicitly importing the Prelude with an empty import list. This reduces backwards compatability however, and is a little trickier to implement.
     123
     124Supporting compilers that do not understand the pragma is mostly easy, with the subtelty that we must not require commas between PHASE pragmas that appear in import/export lists. In my opinion we should not even accept such commas on the basis that by doing so would allow users to inadvertently write programs that do not compile on non-GHC and old-GHC compilers.
     125
     126An example of how it would look is:
     127
     128{{{
     129
     130module Qux({-# PHASE F #-} {-# PHASE G #-} {-# PHASE H #-} where
     131
     132import Quux({-# PHASE I #-} {-# PHASE J #-})
     133
     134... PHASE declarations and uses ...
     135
     136}}}
     137
     138It turns out that we have exactly the required code for this already in the parser to deal with Haddock pragmas.