* damlc: Avoid immediately invoked closures
We add rules to the simplifier which rewrite expressions of the form
```
(\x1 ... xn -> E) A1 ... An
```
into
```
let x1 = A1 in ... let xn = An in E
```
provided that `xi` is not free in `Aj` for any `i < j`.
In cases where `E` is _not_ a lambda, this rewriting is beneficial
since it removes a costly closure allocation and the immediate entering
of the closure. If `E` is a closure itself, this rewriting is not
detrimental since the only difference it makes is that the values for
`A1, ..., An` end up in the list of captured variables rather than in
the list of already applied variables.
Even though users might not write expressions like the one we're
simplifying here themselves, various desugarings produce them
nevertheless. A common pattern is to define multiple auxiliary
functions that a only used once in a where cluase. For instance, with
the help these new rewriting rules the function
```haskell
f: Int -> Int -> Int
f x y = g (h x) y
where
g x y = x+y
h x = 2*x
```
gets translated to
```
def f : Int64 -> Int64 -> Int64 =
\(x : Int64) (y : Int64).
let x2 : Int64 =
let x2 : Int64 = x
in MUL_INT64 2 x2
y2 : Int64 = y
in ADD_INT64 x2 y2
```
Without the simplification, `g` and `h` would each allocate a closure:
```
def f : Int64 -> Int64 -> Int64 =
\(x : Int64) (y : Int64).
(\(x2 : Int64) (y2 : Int64). ADD_INT64 x2 y2)
((\(x2 : Int64). MUL_INT64 2 x2) x)
y
```
The `collect-authority` benchmarck is sped up by 1.03x by this change.
This is not a huge improvement but the change is simple enough to
merge it nevertheless.
CHANGELOG_BEGIN
CHANGELOG_END
* Let code speak not comments
CHANGELOG_BEGIN
CHANGELOG_END
* Improve explanations
CHANGELOG_BEGIN
CHANGELOG_END
* Inline typeclasses and apply projections.
This draft PR:
* adds a Subst module for substitution within LF expressions.
This implementation piggybacks on the existing type
substition. But there are not enough tests yet.
* passes World data into the simplifier, in order to have
have the ability to inline functions (selectively)
* inlines typeclass dictionaries and projection functions
* beta reduces type lambdas (which are just ignored by
speedy afaik)
* beta reduces lambdas that are passed units or dictionaries
* runs the simplifier twice so it has a change to
perform the inlining, reduction, AND projection
-- this is probably avoidable by sequencing the
simplifier steps in a specific order, but running
the simplifier twice is simple enough for now.
* together, these fix#5748 (see result on a toy module)
TODO:
* add lots of tests for Subst
* run this on a benchmark to see how big of a difference it makes.
* reduce the jankiness of running the simplifier twice.
Results:
DAML input:
```
module Main where
hello : Int
hello = 10 + 30
```
Original DAML-LF output:
```
module Main where
@location(7:0-7:5)
def hello : Int64 =
a284919a95c4a515cd1efac0d89be302d0e9d61e692a2176128c871ad8067e36:GHC.Num:+
@Int64
a284919a95c4a515cd1efac0d89be302d0e9d61e692a2176128c871ad8067e36:GHC.Num:$fAdditiveInt
10
30
```
Current DAML-LF output:
```
module Main where
@location(7:0-7:5)
def hello : Int64 = ADD_INT64 10 30
```
changelog_begin
changelog_end
* Update copyright header
* Lint
* More direct typeclass simplification
* extendWorldSelf + getTypeClassDictionary
* Fix visual
* Fix visual comment
* Disable cross-module inlining for incremental builds
* Cleanup subst/freevars.
* copyright header
* Alpha equivalence for LF expressions.
* copyright header
* lots of tests
* fix comment
* Apply review suggestions
I've completely removed the possibility to call `daml codegen ts`. I'm
happy to add in back in a followup PR once I've seen that all our tests
pass without it existing.
CHANGELOG_BEGIN
CHANGELOG_END
* Make name collision check more strict
This PR extends the name collision check to catch collisions between
A:B (type B in module A) and module A.B.C. For now this is just a
warning and not an error. Once we turn it into an error, we also need
to add this to the Scala collision checker.
There is a fair bit of plumbing required to make warnings work but on
the plus side we get multiple errors at once now instead of erroring
out on the first one.
changelog_begin
- [DAML Compiler] The name collision check has been extended to also
count the case as a collision where you have a type B in module A and a module
A.B.C (but no module A.B). This is a warning in this SDK release but
will become an error in a future release. The typescript codegen is
not usable on packages that don’t uphold this restriction.
changelog_end
* Address review comments
* Use a safer safeToReexport
This is much safer than the approximation from last time. The only downside is
introducing a dependency between data dependencies and our type checker,
but that seems safer than having two versions of `expandTypeSynonyms`
floating around (and perhaps this dependency is something we would end
up adding anyway).
changelog_begin
changelog_end
* Add own package to extPackages.
* Use mkTForalls
* simplify mkTForalls
* CHANGELOG_BEGIN
Type-check type synonyms.
CHANGELOG_END
* placate HLint
* comments
* Add an example that requires the check in kindOf
* check types containing syn-apps are well formed even when there is no expression of that type
* show type mismatch error after synonyms are expanded
* typeOf calls expandTypeSynonyms; track vars bound by TForall during expansion
* test interaction of syn-expansion and free-vars; add one bigger testcase
* extend bigger example with pointed typeclass, having functor as a super class
Co-authored-by: Moritz Kiefer <moritz.kiefer@purelyfunctional.org>
* Full package name collision check
* Handle type synonyms appropriately
* Better comment
* Make isAscendant case-insensitive
* Document isAscendant and explain case-insensitivity
* Add a package-wide name collision test.
* daml-lf: rename Map to TextMap in archive proto
+ in Scala/haskell AST
* a bit more renamming
* Update compiler/daml-lf-tools/src/DA/Daml/LF/TypeChecker/Serializability.hs
Co-Authored-By: associahedron <231829+associahedron@users.noreply.github.com>
* fix test
* Apply suggestions from code review
Co-Authored-By: associahedron <231829+associahedron@users.noreply.github.com>
* Update compiler/daml-lf-ast/src/DA/Daml/LF/Ast/Base.hs
Co-Authored-By: associahedron <231829+associahedron@users.noreply.github.com>
* Generalize AnyTemplate type to Any in DAML-LF
See #3131 for the motivation for this. The tl;dr is that we need
something like AnyTemplate for choice types as well.
Since the protobuf was already more general in anticipation of such a
change, this change only changes the internal AST on the Haskell and
Scala side.
Since AnyTemplate change has never made it out of 1.dev, I updated the
changelog in the LF spec instead of adding a new entry.
* Update daml-lf/spec/daml-lf-1.rst
Co-Authored-By: Andreas Herrmann <42969706+aherrmann-da@users.noreply.github.com>
* windows debugging
* more windows debugging
* clean expunge
* don’t cat the config file
* remove comment on type equality
* windows …
* gnah
* foobar
* foobar
* does anything ever work?
* reenable caching
* Do not build daml-lf-ast separately
* Remove polymorphic literals from library
* Add BECastNumeric and BEShiftNumeric in DAML-LF AST.
* Update rest
* Copyright header
* Remove polymorphic literals from LF conversion.
* Update tests
* Add test for forbidden polymorphic literals
* Add newline in DA.Numeric
* use MulNumeric instead of CastNumeric in DA.Internal.Prelude
* Add Any type and to_any/from_any primitives to protobuf
Following a suggestion by Rémy, the protobuf representation is more
general and is associated with an arbitrary type instead of a
typecon. This allows us to easily extend this later to a full Any
type.
I’ve still called the type in the protobuf Any instead of Haskell’s
Dynamic since I find AnyTemplate more clear than DynamicTemplate and
having AnyTemplate and Dynamic seems confusing.
Right now, the decoder enforces that the type is a TypeCon.
* Fix some mistakes in the spec
* Update daml-lf/spec/daml-lf-1.rst
Co-Authored-By: Remy <remy.haemmerle@daml.com>
* Update daml-lf/spec/daml-lf-1.rst
Co-Authored-By: Remy <remy.haemmerle@daml.com>
* Update daml-lf/spec/daml-lf-1.rst
Co-Authored-By: Remy <remy.haemmerle@daml.com>
* Update daml-lf/spec/daml-lf-1.rst
Co-Authored-By: Remy <remy.haemmerle@daml.com>
* Add evaluation rule for to_any_template
* Update daml-lf/spec/daml-lf-1.rst
Co-Authored-By: Remy <remy.haemmerle@daml.com>
* Add missing alpha-equivalence case.
* Type-level nat is not serializable.
* Convert decimal primitives as numeric primitives if feature is available.
* Convert decimal literals to numeric if available.
* Better approach to decimal primitves.
* Fixing BEDecimalFromText
* Add issue number
* Set Numeric feature at v1.7
Currently, we use `Safe 1` as the safety of the `BEMapEmpty` primitive. This
is clearly wrong since `BEMapEmtpy` is not supposed to be applied to a value
(but only to a type). Thus, it must be `Safe 0`. This has not caused any
problems in the past because the type checker would have caught any
application of `BEMapEmpty` to a value.
Previously we rewrote references to `PRSelf` to `PRImport` references
every time we call `initWorld`. However, we call `initWorld` quite
often (e.g. every time we run a scenario) while the packages that need
to be rewritten stay constant. Since rewriting the package references
requires traversing the whole package, this can be quite expensive.
This PR moves the rewriting of package references out of `initWorld`
and caches it as part of `GeneratePackageMap`.
On a large project where I tested this, this caused a drop in the
runtime of `daml test` from 250s to 200s. In the IDE this can also
make a pretty big difference since we call this everytime we run a
scenario which we do on every file change.
* Allow shadowing of type variables in DAML and DAML-LF
We relax the DAML-LF type checker to allow for shadowing of type variables.
This does not need big changes since the substitution we use already avoids
name capture. The test for alpha equivalence converts to de Bruijn indices
on the fly and is hence not a problem either.
We inline type synonyms during the conversion from GHC Core to DAML-LF. This
requires alpha renaming. The cheapest way to get this, is to use the unique
names of type variables instead of their surface names.
This fixes#1915.
* Fix Scala test
* Relax memory constraints for bond-trading test
My primary goal is to avoid confusion when starting to support DAML-LF's
enum types. I also think these types were never particularly useful,
although I introduced them myself...
