According to PEP484, type hint can be a string literal for forward
references. With PEP563, type hint would be preserved in annotations in
string form.
The previous code could have never worked as-is, as the result slot
went unused, and it tried to append the load instruction to the
block just terminated with the invoke.
GitHub: Fixes#1506, #1531.
Since we don't implement any integer-like operations for TBool
(addition, bitwise not, etc.), TBool is currently neither
strictly equivalent to builtin bool nor numpy.bool_, but through
very obvious compiler errors (operation not supported) rather than
silently different runtime behaviour.
Just mapping both to TBool thus is a huge improvement over the
current behaviour (where numpy.False_ is a true-like object). In
the future, we could still implement more operations for TBool,
presumably following numpy.bool_ rather than the builtin type,
just like builtin integers get translated to the numpy-like
TInt{32,64}.
GitHub: Fixes#1275.
Previously, any type would be accepted for the test expression,
leading to internal errors in the code generator if the passed
value wasn't in fact a bool.
array([...]), the constructor for NumPy arrays, currently has the
status of some weird kind of macro in ARTIQ Python, as it needs
to determine the number of dimensions in the resulting array
type, which is a fixed type parameter on which inference cannot
be performed.
This leads to an ambiguity for empty lists, which could contain
elements of arbitrary type, including other lists (which would
add to the number of dimensions).
Previously, I had chosen to make array([]) to be of completely
indeterminate type for this reason. However, this is different
to how the call behaves in host NumPy, where this is a well-formed
call creating an empty 1D array (or 2D for array([[], []]), etc.).
This commit adds special matching for (recursive lists of) empty
ListT AST nodes to treat them as scalar dimensions, with the
element type still unknown.
This also happens to fix type inference for embedding empty 1D
NumPy arrays from host object attributes, although multi-dimensional
arrays will still require work (see GitHub #1633).
GitHub: Fixes#1626.
Strided slicing of one-dimensional arrays (i.e. with non-trivial
steps) might have previously been working, but would have had
different semantics, as all slices were copies rather than a view
into the original data.
Fixing this in the future will require adding support for an index
stride field/tuple to our array representation (and all the
associated indexing logic).
GitHub: Fixes#1627.
This was a long-standing issue affecting both lists and
the new NumPy array implementation, just caused by the
generic inference passes not being run on the slice
subexpressions (and thus e.g. ints not being monomorphized).
GitHub: Fixes#1632.
This allows assert() to be used on Zynq, where abort() is not
currently implemented for kernels. Furthermore, this is arguably
the more natural implementation of assertions on all kernel targets
(i.e. where embedding into host Python is used), as it matches host
Python behavior, and the exception information actually makes it to
the user rather than leading to a ConnectionClosed error.
Since this does not implement printing of the subexpressions, I
left the old print+abort implementation as default for the time
being.
The lit/integration/instance.py diff isn't just a spurious change;
the exception-based assert implementation exposes a limitation in
the existing closure lifetime tracking algorithm (which is not
supposed to be what is tested there).
GitHub: Fixes#1539.
This generates rather more code than necessary, but has
the advantage of automatically handling incomplete
multi-dimensional subscripts which still leave arrays
behind.
Lists and arrays no longer have the same representation all
the way through codegen, as used to be the case.
This could/should be made more efficient later, eliding the
temporary copies.
Left generic transpose (shape order inversion) for now, as that
would be less ugly if we implement forwarding to Python function
bodies for array function implementations.
Needs a runtime test case.
LLVM will take care of optimising the loops. This was still
unnecessarily painful; implementing generics and implementing
this in ARTIQ Python looks very attractive right now.
Relies on the runtime to provide the necessary
(libm-compatible) functions.
The test is nifty, but a bit brittle; if this breaks in the
future because of optimizer changes, do not hesitate to convert
this into a more pedestrian test case.
So far, this is not exposed to the user beyond implicit conversions.
Note that all the implicit conversions, such as triggered by adding
arrays of mismatching types, or dividing integer arrays, are currently
emitted in a maximally inefficient way, where a temporary copy is first
made for the type conversion. The conversions would more sensibly be
implemented during the per-element operations to save on the extra
copies, but the current behaviour fell out of the rest of the IR
generator structure without extra changes.
Matches NumPy. Slicing a TList reallocates, this doesn't; offsetting
couldn't be handled in the IR without introducing new semantics
(the Alloc kludge; could/should be made its own IR type).
Still needs support through all the rest of the compiler, and
support for higher-dimensional arrays.
Alternatively, we could always assume ndarrays of ndarrays
are rectangular (i.e. ban array/list element types), and
detect mismatch at runtime. This might turn out to be
preferrable to be able to construct matrices from rows/columns.
`array()` is disallowed for no particularly good reason but
numpy API compatibility.
This reverts commits f8d1506922
and cf19c9512d.
While the commit just fixes a clear typo in the implementation,
it turns out the original algorithm isn't flexible enough to
capture functions that transitively return references to
long-lived data. For instance, while cache_get() is special-cased
in the compiler to be recognised as returning a value of Global()
lifetime, a function just forwarding to it (as seen in the
embedding tests) isn't anymore.
A separate issue is also that this makes implementing functions
that take lists and return references to global data in user code
impossible, which central parts of the Oxford codebase rely on.
Just reverting for now to unblock master; a fix is easily designed,
but needs testing.
Interestingly enough, these actually seem to give a measurable
speedup (if small – about 1% improvement out of 6s whole-program
compile-time in one particular test case).
The previous implementation of is_mono() had also interesting
behaviour if `name` wasn't given; it would test only for the
presence of any keys specified via keyword arguments,
disregarding their values. Looking at uses across the current
ARTIQ codebase, I could neither find a case where this would
have actually been triggered, nor any rationale for it.
With the short-circuited implementation from this commit,
is_mono() now checks name/all of params against any specified
conditions.
This was mistakenly included in fb2b634c4a, and broke the test
case verifying that using None as an ARTIQ type annotation in fact
generates an error message.
With support for polymorphism (or type erasure on pointers to
member functions) being absent in the ARTIQ compiler, code
generation is vital to be able to implement abstractions that
work with user-provided lists/trees of objects with uniform
interfaces (e.g. a common base class, or duck typing), but
different concrete types.
@kernel_from_string has been in production use for exactly
this use case in Oxford for the better part of a year now
(various places in ndscan).
GitHub: Fixes#1089.
`var_type` was presumably intended to convert to a target type,
but wasn't actually acted on in the function body (nor was it
used anywhere in the codebase).
This reverts 425cd7851, which broke the use of casts to define
integer width.
Instead of it, two steps are taken:
* First, literals are monomorphized, leading to predictable result.
* Second, casts are monomorphized, in a top-bottom way. I.e.
consider the expression `int64(round(x))`. If round() was visited
first, the intermediate precision would be 32-bit, which is
clearly undesirable. Therefore, contextual rules should take
priority over non-contextual ones.
Fixes#1252.
pca006132
Sebastien Bourdeauducq
David Nadlinger
Leon Riesebos
whitequark