+ Instead of including lists of parameter values generated by GF, generate
code to enumerate parameter values (in the same order as GF). This seems
to give a factor of 2-3 code size reduction in the Phrasebook (e.g.
from 84MB to 25MB for Hin, from 338MB to 154MB for Fre).
+ Deduplicate table entries, i.e. convert "table [..,E,..,E,..,E,..]" into
"let x = E in table [..,x,..,x,..,x,..]". This gives even more significant
code size reduction in some cases, e.g. from 569MB to 15MB for
PhrasebookFin.
All phrasebook languages can now be converted to compilable Haskell code,
except PhrasebookPes, which still has the name clash problem.
Many Phrasebook languages can now be converted to compilable Haskell code.
Some languages (Fre, Hin, Snd, Urd) generate too much Haskell code to be
practically useful (e.g. 338MB for Fre). One language (Fin) took too long
to convert to Haskell. One language (Pes) has problems with name clashes in
the generated Haskell code.
STILL TODO:
- variants
- pre { ... }
- reduce code duplication for large tables
- generate qualified names to avoid name clashes
The translation is currently good enough to translate all concrete syntaxes
of the Foods and Letter grammars, and some concrete syntaxes of the Phrasebook
grammar (e.g. PhrasebookEng & PhrasebookSpa works, but there are problems with
e.g. PhrasebookSwe and PhrasebookChi)
This functionality is enabled by running
gf -make -output-format=haskell -haskell=concrete ...
TODO:
- variants
- pre { ... }
- eta expansion of linearization functions
- record subtyping can still cause type errors in the Haskell code
in some cases
- reduce code large tables
It was used only in cases where a lock field needed to be added to a
run-time variable, like e.g. in examples/phrasebook/SentencesTha.gf:
lin
PGreetingMale g = mkText (lin Text g) (lin Text (ss "ครับ")) | g ;
PGreetingFemale g = mkText (lin Text g) (lin Text (ss "ค่ะ")) | g ;
But lock fields are only meaningful during type checking and can safely be
ignored in later passes.
* The following modules are no longer used and have been removed completely:
GF.Compile.Compute.ConcreteLazy
GF.Compile.Compute.ConcreteStrict
GF.Compile.Refresh
* The STM monad has been commented out. It was only used in
GF.Compile.SubExpOpt, where could be replaced with a plain State monad,
since no error handling was needed. One of the functions was hardwired to
the Err monad, but did in fact not use error handling, so it was turned
into a pure function.
* The function errVal has been renamed to fromErr (since it is analogous to
fromMaybe).
* Replaced 'fail' with 'raise' and 'return ()' with 'done' in a few places.
* Some additional old code that was already commented out has been removed.
On my laptop these changes speed up the full build of the RGL and example
grammars with 'cabal build' from ~95s to ~43s and the zero build from ~18s
to ~5s.
The main change is the introduction of the module GF.CompileInParallel that
replaces GF.Compile and the function GF.Compile.ReadFiles.getAllFiles. At
present, it is activated with the new -j flag, and it is only used when
combined with --make or --batch. In addition, to get parallel computations,
you need to add GHC run-time flags, e.g., +RTS -N -A20M -RTS, to the command
line.
The Setup.hs script has been modified to pass the appropriate flags to GF
for parallel compilation when compiling the RGL and example grammars, but you
need a recent version of Cabal for this to work (probably >=1.20).
Some additonal refactoring were made during this work. A new monad is used to
avoid warnings/error messages from different modules to be intertwined when
compiling in parallel, so some functios that were hardiwred to the IO or IOE
monads have been lifted to work in arbitrary monads that are instances in
the appropriate classes.
In particular, the function compileOne has been moved to the new module
GF.CompileOne and its type has been changed from
compileOne :: ... -> CompileEnv -> FilePath -> IOE CompileEnv
to
compileOne :: ... -> SourceGrammar -> FilePath -> IOE OneCompiledModule
making it more suitable for use in a parallel compiler.
The def rules are now compiled to byte code by the compiler and then to
native code by the JIT compiler in the runtime. Not all constructions
are implemented yet. The partial implementation is now in the repository
but it is not activated by default since this requires changes in the
PGF format. I will enable it only after it is complete.
GF.Text.Pretty provides the class Pretty and overloaded versions of the pretty
printing combinators in Text.PrettyPrint, allowing pretty printable values to
be used directly instead of first having to convert them to Doc with functions
like text, int, char and ppIdent. Some modules have been converted to use
GF.Text.Pretty, but not all. Precedences could be added to simplify the pretty
printers for terms and patterns.
GF.Infra.Location contains the types Location and L, factored out from
GF.Grammar.Grammar, and the class HasSourcePath. This allowed the import
of GF.Grammar.Grammar to be removed from GF.Infra.CheckM, making it more
like a pure library module.
PGF exports the public, stable API.
PGF.Internal exports additional things needed in the GF compiler & shell,
including the nonstardard version of Data.Binary.
(table { p_i => t_i } ! x).l ==> table { p_i => t_i.l } ! x
This was used in the old partial evaluator and can significantly reduce term
sizes in some cases.
Eta expansion is applied between partial evaluation and PMCFG generation.
The buggy version generated type incorrect terms, but PMCFG generation
apparently worked anyway.
