The work done by the partial evaluator is now divied in two stages:
- A static "term traversal" stage that happens only once per term and uses
only statically known information. In particular, the values of lambda bound
variables are unknown during this stage. Some tables are transformed to
reduce the cost of pattern matching.
- A dynamic "function application" stage, where function bodies can be
evaluated repeatedly with different arguments, without the term traversal
overhead and without recomputing statically known information.
Also the treatment of predefined functions has been reworked to take advantage
of the staging and better handle partial applications.
* Evaluate operators once, not every time they are looked up
* Remember the list of parameter values instead of recomputing it from the
pattern type every time a table selection is made.
* Quick fix for partial application of some predefined functions.
GF.Compile.Compute.ConcreteNew + two new modules contain a new
partial evaluator intended to solve some performance problems with the old
partial evalutator in GF.Compile.Compute.ConcreteLazy. It has been around for
a while, but is now complete enough to compile the RGL and the Phrasebook.
The old partial evaluator is still used by default. The new one can be activated
in two ways:
- by using the command line option -new-comp when invoking GF.
- by using cabal configure -fnew-comp to make -new-comp the default. In this
case you can also use the command line option -old-comp to revert to the old
partial evaluator.
In the GF shell, the cc command uses the old evaluator regardless of -new-comp
for now, but you can use "cc -new ..." to invoke the new evaluator.
With -new-comp, computations happen in GF.Compile.GeneratePMCFG instead of
GF.Compile.Optimize. This is implemented by testing the flag optNewComp in
both modules, to omit calls to the old partial evaluator from GF.Compile.Optimize
and add calls to the new partial evaluator in GF.Compile.GeneratePMCFG.
This also means that -new-comp effectively implies -noexpand.
In GF.Compile.CheckGrammar, there is a check that restricted inheritance is used
correctly. However, when -noexpand is used, this check causes unexpected errors,
so it has been converted to generate warnings, for now.
-new-comp no longer enables the new type checker in
GF.Compile.Typeckeck.ConcreteNew.
The GF version number has been bumped to 3.3.10-darcs
There was 55 lines of rather repetitive code with calls to 6 compiler passes.
They have been replaced with 19 lines that call the 6 compiler passes
plus 26 lines of helper functions.
Introduced the function
parallelCheck :: [Check a] -> Check [a]
that runs independent checks in parallel, potentially allowing faster grammar
compilation on multi-core computers, if you run gf with +RTS -N.
However, on my dual core laptop, this seems to slow down compilation somewhat
even though CPU utilization goes up as high as 170% at times.
(This is with GF compiled with GHC 7.0.4.)
In GF.Compile.CheckGrammar, use a new topological sorting function that
groups independent judgements, allowing them all to be checked before
continuing or reporting errors.
Using accumulated errors in the Check monad.
TODO: some errors are still not accumulated, but thanks to checkMapRecover
at least one error per judgement is reported.
In addition to warnings, the Check monad in GF.Infra.CheckM can now accumulate
errors. There are two new functions
checkAccumError: Message -> Check ()
accumulateError :: (a -> Check a) -> a -> Check a
The former (with the same type as checkWarn) is used to report an accumulated
(nonfatal) error. The latter converts fatal errors into accumulated errors.
Accumulated errors are reported as regular errors by runCheck.
Also, the Check monad type has been made abstract.
An apparent bug in ghc-7.2.2 causes the type Value to be exported from PGF.Data.
Workaround: restrict the imports from PGF.Data in GF.Command.Abstract and
GF.Compile.GeneratePMCFG to avoid the clash with locally defined type Value.
(ghc-7.0.4 and ghc-7.4.1 appear to be free from this bug.)
