broaden what Names can be reflected and refied
I did not add the Names I wasn't sure how to test but have put placeholders
that produce clearer error messages.
There's some missing flexibility in how foreign specifiers can be used with
scheme that is addressed here with minimal changes to how scheme specifiers
are read. This opens up uses for users that they otherwise would have had
to modify the compiler's support files to accomplish.
Auxiliary functions introduced in elaboration (e.g., through case splits and with clauses) now
have the same totality annotation as the function they're defined in.
Moved auxiliary function `findSetTotal` into `Context.idr` since it's
now used by `ProcessDef.idr` too.
Added a totality requirement argument to `checkClause` so that the
with-clause case could propagate it to the functions it generates in
elaboration.
Sandwhich the rhs elaboration in pattern matches with code that sets
the global, default, totality requirement to the current one, and
restores the previous default afterwards. It's a bit of a hacky way to
do it, but I don't think we have a better alternative with the current
design.
Until now namespaces were stored as (reversed) lists of strings.
It led to:
* confusing code where we work on the underlying representation of
namespaces rather than say what we mean (using `isSuffixOf` to mean
`isParentOf`)
* potentially introducing errors by not respecting the invariant cf.
bug report #616 (but also name generation in the scheme backend
although that did not lead to bugs as it was self-consistent AFAICT)
* ad-hoc code to circumvent overlapping interface implementation when
showing / pretty-printing namespaces
This PR introduces a `Namespace` newtype containing a list of strings.
Nested namespaces are still stored in reverse order but the exposed
interface aims to support programming by saying what we mean
(`isParentOf`, `isApproximationOf`, `X <.> Y` computes to `X.Y`, etc.)
irrespective of the underlying representation.
Until now namespaces were stored as (reversed) lists of strings.
It led to:
* confusing code where we work on the representation rather than say
what we mean (e.g. using `isSuffixOf` to mean `isParentOf`)
* potentially introducing errors by not respecting the invariant cf.
bug report #616 (but also name generation in the scheme backend
although that did not lead to bugs as it was self-consistent AFAICT)
* ad-hoc code to circumvent overlapping interface implementations when
showing / pretty-printing namespaces
This introduces a Namespace newtype containing non-empty lists of
strings. Nested namespaces are still stored in reverse order but the
exposed interface aims to support programming by saying what we mean
(`isParentOf`, `isApproximationOf`, `X <.> Y` computes to `X.Y`, etc.)
irrespective of the underlying representation.
All of these internal names are making the output fragile. This
cleanup should allow us to only have to update the golden file
when there is a genuinely interesting change.
- Added initial implementations for terms and values
- Error messages converted to pretty printer
- Colorization for error messages
- Color and console width option both as command line and repl command
As it was, it could break if the argument was repeated more than twice.
When checking dot patterns, we need to check that no further holes are
solved, and that the pattern variable doesn't unify with some other
pattern variable, but if it had already made progress (either for a good
or bad reason) we missed this. Fixes#536
As it was, there was significant backtracking for big expressions,
getting to the end, not finding a **, so having to try again for
application expressions. Fixes#532
This has involved quite a bit of reorganisation and some improvements in
resugaring so that the results look nice. In summary:
* Expression search now gives back a RawImp rather than a checked term,
which allows it to include case expressions
* Case with one pattern is resugared to a destructuring let
* Some name generation issues address in function generation
We look at intermediate results for local variables which are functions
that return a concrete type, or recursive calls that return a single
constructor type. In these cases, we:
* let bind the local variable/recursive call
* generate a new definition for the scope, as a 'case' function
When we recursively generate the definition, it's a bit more restricted
so as not to explode the search space. We only take the first result, we
only look one constructor deep, and we go right to left on variable
splitting so only deconstruct the name we've just added.
Sort by proportion of bound variables used, which is likely to get us
the right answer quicker. The results are generated in batches of 16 (a
completely arbitrary choice) then sorted.
This gives the number of implementations to reject before accepting one.
It's intended as a reasonably cheap way of giving multiple results from
interactive editing (e.g. the vim mode, which goes via the REPL and
--client rather than the IDE mode)
These continue the search from :ps and :gd next respectively, giving the
next search result until there are no more results.
Correspondingly, added ':proof-search-next' and ':generate-def-next' in
IDE mode, which continue the search from the previous ':proof-search'
and ':generate-def' respectively.
Rather than returning a complete list of results, return a pair of the
first result, and a continuation. The continuation explains how to
continue the search if the given result is deemed unacceptable (either
on encountering an error somewhere, or just if the caller wants the next
result).
This means we don't search needlessly if we're only looking for the
first result. Fixes#228
...until the definition is complete. This is necessary since sometimes
information outside the case block can help resolve interfaces, and in
the simplest case, we might just have delayed resolving a default
Integer. It turns out this was also an obscure bug waiting to happen
with coverage checking of nested case blocks (so there's a test update
there too).
Fixes#443
In a 'Bind', normalise the result of the first action, rather than
quoting the HNF. This improves performance since the HNF could be quite
big when quoted back.
Ideally, we wouldn't have to quote and unquote here, and we can probably
achieve this by tinkering with the evaluator.
This has an unfortunate effect on the reflection002 test, in that the
"typed template Idris" example now evaluates too much. But, I think the
overall performance is too important for the primary motivation
behind elaborator reflection. I will return to this!
This is partly to tidy things up, but also a good test for 'import as'.
Requires some internal changes since there are parts of reflection,
unelaboration and a compiler hack that rely on where things are in the
Prelude.
The Int represented the resolved name, but this isn't guaranteed to be
up to date after reloading and, worse, it doesn't display helpfully. I'm
bored of updating the tests which fail as a result!
This also fixes#407, which is about displaying the wrong name after
reloading the ttc.
There is an argument that, for import public, this should be automatic
(that is, the publicly imported things should be reexported with the
parent namespace). I decided not to do this, because another use of
import public which we do a lot in the Idris 2 code base is purely as a
convenience, where we still expect things to be in their original
namespace.
Also, where there's a choice between things being explicit and implicit,
I prefer to err on the side of explicit now.
So, if what you really want in an API is to reexport, you can do that,
but explicitly.
Instead of just the cursory name update that we used to do (which didn't
work properly anyway for a lot of reasons), now we add aliases for all
the names in the imported module.
So, like Idris 1, every global has a canonical name by which we can
refer to it, but it can also have aliases via "import ... as".
If we have a delayed thing, but we don't yet know the expected type,
don't commit to forcing because the expected type might turn out to be a
delay.
Fixes#395
'convert' doesn't solve holes, so might reject things that are solvable.
This can be an issue when resolving interfaces, because we were using
convert for arguments of the invertible holes that arise when trying to
resolve them. Fixes#66.
We need to make sure they are inferred again when elaborating methods,
so substitute in a _ in method types before substituting in the explicit
parameters.
In future, it might (probably will) also be useful to allow giving the
implicit parameters explicitly when defining implementations.
Fixes#374
It's worth delaying in case doing some more work (and some more
interface resolution) can make the type more concrete. This makes
test linear011 work more smoothly, and will help with this sort of
program in general.
A better way, later, would be to try elaborating and delay only if the
type is not yet known. We have the facilities, but it's too fiddly for
me to want to implement it right now...
We need to check below top level too, since there could be holes that
we're happy to resolve by searching. The linearity test added
illustrates a place where this is needed.
The namespace parser was not requiring a minimum indentation and instead
based its indentation on the following line, which meant that a line like:
namespace Foo
foodef : Int
placed foodef into namespace Foo instead of the module's top level.
And so made it unclear when a namespace ends.
Following a fairly detailed discussion on slack, the feeling is
generally that it's better to have a single interface. While precision
is nice, it doesn't appear to buy us anything here. If that turns out to
be wrong, or limiting somehow, we can revisit it later. Also:
- it's easier for backend authors if the type of IO operations is
slightly less restrictive. For example, if it's in HasIO, that limits
alternative implementations, which might be awkward for some
alternative back ends.
- it's one less extra detail to learn. This is minor, but there needs to
be a clear advantage if there's more detail to learn.
- It is difficult to think of an underlying type that can't have a Monad
instance (I have personally never encountered one - if they turns out
to exist, again, we can revisit!)
This addresses the case where you'd see an error of:
Ambiguous elaboration at:
r <- pure []
Possible correct results:
[]
[]
[]
By changing it to:
Possible correct results:
Main.Nil
PrimIO.Nil
Prelude.Nil
The output directory was previously called the executable directory, but I changed it because the output is not always an executable (depending on the code generator).
The code generators can now distinguish between where to place the (temporary) build files and the resulting output files.
unelabBinder was losing information for the purpose of displaying names
nicely. This is typically wanted but when working with elabortation it's
useful to have all the information you can get. Things like record field
names would be lost when querying with GetType but are now retained.
It takes a long time to run, but it is basically all
the testing we need of the API, if this works it's
unlikely to be broken. It requires a change in the build
system, so don't enable it for now.
I'm playing with some linear structures and finding these useful a lot,
so good to have a consistent syntax for it. '#' is chosen because it's
short, looks a bit like a cross if you look at it from the right angle
(!) and so as not to clash with '@@' in preorder reasoning syntax.
Without --no-index, git compares the files with their state in HEAD.
But we want to compare them to each other! This explain why we were
getting no output whatsoever.
With --word-diff=color, we can easily spot the small changes anywhere in
a line.
If a function is public export, the local definitions also need to be
public export for it to reduce properly. But if they're not, we don't
want them exported or they might affect the module hash and cause
unnecessary rebuilds.
When runing in interactive mode, the diff between expected and output is not being show,
Supposedly, Git-diff returns 0 on successful run and a non-zero for error.
However, this is not being picked up by the test suite.
We change git-diff's error reporting to be more diff like in which:
+ +ve indicates that there is a diff.
+ 0 indicates the files are the same.
+ -ve program error.
This should make diff reporting more robust.
This takes the responsibilty of finding the ipkg away from IDE mode,
which seems sensible given that we can do it ourselves. If there isn't
one, it'll load from the local directory as always.
This involves new primitives GCPtr and GCAnyPtr which are pointer types
that have finalisers attached. The finalisers are run when the
associated pointer goes out of scope.
In the test, I am assuming that the GC will only be called once, right
at the end. Otherwise, the output isn't guaranteed to be deterministic!
Let's see how this assumption holds...
This is currently Chez only. I think it'll be easy enough to add to
the Racket and Gambit back ends too.
Clash in a reflection test due to conflicting merges. I'd like to find a
way to make the internally generated numbers not matter here (and
elsewhere) but I don't see an obvious way.
Fixes#116. This is the solution Idris 1 took, and while it is a special
case, the syntax does make it explicit (in a way) that the result of the
case is unused - if you mean something other than (), you must now say
so!
helps with readability since these, especially named-IPi, come up a lot
didn't change everything that could need it like PiInfo or BindMode
PiInfo rarely has DefImplicit (so far) and BindMode hasn't come up a lot (so far)
reduced indentation for TTImp Show implementation
Another one from the "stop trying to be clever" files :). Instead of a
continuation for fallthrough in the evaluator, be explicit about whether
there's a result, no match, or evaluation is stuck.
Fixes#70
This is quite fiddly as it the blocks might be in different contexts so
we need to keep track of which variables correspond in the scrutinees of
the blocks. Once that's done, check the terms at the leaves convert,
then check the corresponding variables convert.
This may not be perfect yet, because we only look at case scrutinees to
find correspondence. It might also be a bit slower than it could be, but
at least these checks are quite rare.
Fixes#208 and maybe some others?
Allows quoting a term back to a TTImp. Test reflection007 shows one
possible use for this, building a reflected, type safe, representation
of an expression.
So the type of Elab now gives the expected type that's being elaborated
to, meaning that we can run 'check' in the middle of scripts and use the
result.
On the LHS, we want to match against the reflected thing, so FC and
implicits need to turn into match anything patterns, or we won't match
anything at all. This means we can put quoted terms on the LHS, with
pattern variables under ~().
Instead of merely generating a locally unique name, use the existing
code for generating a new unique variable name in the unifier, which is
therefore globally unique.
This invokes a script of type Elab (). %runElan in a term invokes a
script of type Elab TT. The elaborator now pushes in that type, so that
it'll report an appropriate error if you give it a script of the wrong
type.
A %macro must always be fully applied. Whenever the elaborator
encounters a %macro application (except in a function LHS) it evaluates
the application and sends the result to %runElab. So:
%macro
foo : args -> Elab TT
...
def = foo a b c
is equivalent to
foo : args -> Elab TT
...
def = %runElab foo a b c
Including appropriate casts, and Num/Eq/Ord/Show implementations.
Also includes new primitives in Data.Buffer, and calls to foreign
functions in C as 'unsigned'.
If available (sometimes, say a top level expression, it might need
inferring so there'll be no goal available). Also add the ability to log
the current goal, or indeed any term.
Add %runElab and start on scripts, although all they can do so far is
check a term. This does gives us, sort of, "template Idris" (as
demonstrated in test reflection002)
Can't export a type which refers to a private name. This has caught a
couple of visibility errors in the libraries, code and tests, so they've
been updated too.
Don't get too excited yet - I want this in so that it doesn't get too
out of sync, but I still have to think about exactly how it's going to
work in practice.
This means it abstracts over the value syntactically, rather than by
value, and can significantly speed up elaboration where large types are
involved, at a cost of being less general. Try it if "with" is slow.
There are more flags we want on with (well, at least one: "proof")
This was taking too long, and adding too many things, because it was
going too deep in the name of having everything accessible at the REPL
and for the compiler. So, it's done a bit differently now, only chasing
everything on a "full" load (i.e., final load at the REPL)
This has some effects:
+ As systems get bigger, load time gets better (on my machine, checking
Idris.Main now takes 52s from scratch, down from 76s)
+ You might find import errors that you didn't previously get, because
things were being imported that shouldn't have been. The new way is
correct!
An unfortunate effect is that sometimes you end up getting "undefined
name" errors even if you didn't explicitly use the name, because
sometimes a module uses a name from another module in a type, which then
gets exported, and eventually needs to be reduced. This mostly happens
because there is a compile time check that should be done which I
haven't implemented yet. That is, public export definitions should only
be allowed to use names that are also public export. I'll get to this
soon.
As it was in Idris1 being able to override some, or introduce new, options to an Idris IPKG is beneficial. For example, generate code for multiple codegens from a single source.
Overridable options are:
+ `--quiet`
+ `--verbose`
+ `--timing`
+ `--dumpcases <file>`
+ `--dumplifted <file>`
+ `--dumpvmcode <file>`
+ `--debug-elab-check`
This is mostly to make it easier to write linear function types without
having to invent names for everything, which might be noisy. Also it
improves the display of linear function types when the name isn't used
in the scope.
I don't know how complete this is, but it certainly detects some of the
most obvious cases which are most likely to be bugs.
While I'm at it, this is as good a time as any to add a general way of
reporting warnings, similar to the way of reporting errors.
Still a couple of things to resolve in coverage and totality checking
before we can switch on %default, so don't expect quite the right
behaviour just yet. More progress though!
Also working on this has caught a few totality errors in the Idris 2
code base that Idris 1 missed... so these are fixed on the way.
The hack (optimisation?) to normalise integer literals when below some
threshold is fine on the RHS, but on the LHS causes problems since we
need them in normal form for pattern matching. Fixes#112
This improves coverage checking, because it can now see that things like
(Z = S x) and (x = S x) are empty. Previously, it only checked that all
possible constructors had a disjoint index. Now, it looks for matches
and checks for disjointness in the matches, which catches a lot more
things especially with equality.
This is added to functions which are guaranteed to be productive. The
check is currently very conservative - just added when every clause is
constructor headed (or headed by an AllGuarded function), and there are
no other function applications.
If we never evaluate under Delay at all, we won't inline interface
methods, which means productive things defined in an interface can never
be today. So, make sure to set the tcinline flag before quoting the
Delayed closure.
We also need to separate building the runtests binary from running the
tests, because runtests refers to the boostrap libraries, and the tests
refer to the newly built libraries.
This worked locally, using inconsistent TTC versions for the bootstrap
version and new version, but let's see what it does on a clean machine
As in Idris 1 - if an interface declares a method as total or covering,
then all implementations have to satisfy that.
Temporarily turn off %default directive again, at least until totality
checking works as it should (this is probably better than removing it
from various places because I'll forget to put them back)
This means temporarily removing some '%default total' from the libraries
and tests, since the input for codata checking isn't right yet (it needs
appropriate use of inlining)
Should have done this ages ago, it was much easier than I expected...
this adds an explicit error clause before running the pattern match
compiler for runtime case trees, but only if the coverage checker finds
there are missing cases.
* Ignore build artifacts in 'tests' directory
* Remove unused variables in makefiles
* Add 'bootstrap-clean' rule to delete build artifacts from 'bootstrap' directory
* Add 'distclean' rule to delete all build artifacts from the source tree
