Check that none of the generated missing cases match explicitly given
clauses, which might happen if there's some overlaps in the clauses or
if there's some matching on implicits. Ignore erased argument positions
when checking the match.
If we postpone unification problems, it's not necessarily at the top
level of a term, but then if we retry and find we need to insert a
'Delay' or 'Force' then it had better be at the top level or we'll get
an incorrect term. So, keep track of this in postponed constraints.
I dont know how I thought that was supposed to work, but it needs to get
the S and Z cases separately then build the 'if' expression, with the
default case if either are missing.
I had thought this was unnecessary, because it treats erased patterns as
variables, but it turns out we do need to erase variables where they're
sometimes used in erased positions, and sometimes not, to ensure we use
the right one.
It's a bit more expensive than just using a lambda, because they cache
the result in a thread safe way, and the Idris Force/Delay doesn't
promise to cache the result anywya.
Need to make sure we've got the right real name for the function - this
had previously only been tested in parameters blocks where the name
doesn't change.
If there's a conversion failure in unification, there could still be a
way to rescue it if one of the terms is a lazy type, so postpone rather
than giving up. Fixes#225
For example, delayed elaborators. If it resolves one, and it's not a
complete solution, we might end up with unsolved holes. Or - possibly
worse - we might end up with an error in a program that is silently
fixed rather than reported!
Fixes#223
If there's a 'case' with multiple branches, some of which have holes, we
don't know what the effect on the usage should be, so leave it alone and
just update the hole types. It's get fully checked when all the holes
are resolved.
Most importantly: Case operators should not be inlined indiscriminately,
because that might duplicate work (e.g. via a variable pattern for the
scrutinee, where the variable is used more than once)
This is needed to get the time correctly on 32 bit systems (for example,
a Raspberry Pi!) otherwise it reports a negative number for the current
time...
This has two main advantages: firstly, it actually makes the build slightly
faster and less memory intensive, because Idris doesn't fork a new
process and eat all the memory! Secondly, it means we can build a
distribution with the C file, for building on machines which don't have
Idris 1 available.
Also includes a script for building such a distribution.
I expect someone who is better a writing Makefiles than me can tidy this
up. There are almost certainly problems on Windows too - please help if
you can!
There are new Make targets 'all-fromc' and 'install-fromc' which don't
require building from Idris source, so that we can build with no Idris 1
available.
A small consequence: make install-exec no longer builds the idris2
executable, just installs it.
Explaining how to write the most basic Idris bindings for readline,
taking account of how to allocate Strings in the completion callback.
Also adds the basic API as a sample, which can be used as a starting
point for other packages containing C bindings.
For an output file 'f', now generate a subdirectory in the executable
directory 'f_app', and put 'f.so' and 'f.ss' in there, along with any C
libraries we might have needed. Then generate a shell script 'f' to set
the library path to `f_app` and invoke `f.so`.
In :exec mode, f = _tmpchez, so we're using the executable directory for
the temporary files.
This means all the libraries and code needed to run the generated
program are copied to a single relocatable target directory.
This almost certainly won't work on Windows. Please help! Thanks :)
