Changes between Version 20 and Version 21 of Commentary/Compiler/StrictnessAnalysis/KirstenNotes


Ignore:
Timestamp:
Nov 17, 2006 6:33:58 PM (7 years ago)
Author:
kirsten
Comment:

explaining demand transformers

Legend:

Unmodified
Added
Removed
Modified
  • Commentary/Compiler/StrictnessAnalysis/KirstenNotes

    v20 v21  
    7474When we run into an expression like {{{(Cast e co)}}} that we're placing demand {{{d}}} on, we analyze {{{e}}} to get {{{dmd_ty}}}, then check whether the depth of {{{e}}} is equal to the depth of {{{dmd_ty}}} or not. This is necessary because we might be casting a function to a non-function type. So, if {{{d}}} and {{{dmd_ty}}} have equal depth, we return {{{dmd_ty}}} as is; if {{{d}}}'s arity is less, we drop the appropriate number of args from {{{dmd_ty}}}; if {{{dmd_ty}}}'s arity is less, we add the appropriate number of dummy argument demands to it. 
    7575 
    76 = WARN: arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs) = 
     76= <nowiki>WARN: arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs)</nowiki> = 
    7777 
    7878This warning was happening for (at least) two reasons: 
    7979- lambdas with a strict non-call demand placed on them were being handled wrong (see the first two examples in [[Commentary/Compiler/StrictnessAnalysis/Examples]]) 
    8080- coercions were being handled wrong, resulting in a demand type with depth 0 being assigned to an rhs consisting of a cast from/to a function type 
     81 
     82= Explaining demand transformers = 
     83 
     84For those who, like me, are a little slow, this example might go in section 5.1 of the paper: 
     85 
     86(a): 
     87{{{ 
     88f = \ x ->  
     89      case x of 
     90        ... -> \ y -> ... 
     91        ... -> \ y -> ... 
     92}}} 
     93(b): 
     94{{{ 
     95f = \ x -> 
     96     \ y -> 
     97        case x of 
     98           ... 
     99}}} 
     100 
     101In both (a) and (b), {{{f}}}'s rhs places a strict demand on {{{x}}}. So if we see: 
     102{{{ 
     103(f x) 
     104}}} 
     105with a strict demand placed on it, it wouldn't be sound to look at {{{f}}}'s demand signature and say that {{{(f x)}}} places a strict demand on {{{x}}} under {{{f}}} -- because we don't know whether {{{f}}} is like (a) or like (b). This is why when we see a partial application of {{{f}}}, we discard all of the argument information in {{{f}}}'s demand type.