There was a log message that we are creating typedef void* for empty
structs, but we have missed that piece of code. This patch adds that
logic.
In the case of unions, we completely ignored empty ones, and also
there was not handling if the type was generated by ordinal (as
in the case of structs). This patch also fixes it.
InstCombine was disabled in favor of some more basic constant
propagation, but it turns that under certain circumstances this leads to
direct branches appearing as indirect branches until we run InstCombine
during finalization, which triggers an assertion, since we do not expect
any direct jump there.
If this, affects translation performance, we can think to run it less
often.
`TupleTreeReference`s use the Root type as a pointer only so they can
operate with forward declarations. Before this commit there were static
asserts that prevented this use case, and thus required the headers
defining the root type to be included before this one to work correctly.
We drop those `static_assert`s to make sure that all headers are parsable
on their own.
This change introduces some duplication but ensures an important
property of `tuple_tree_generate`d: data structures: all the leaves are
scalars. Previously, yield::Function was using efa::BasicBlock, making
things more difficult under certain conditions.
Specifically, we can rely on the fact that, when generating a visit to
the TupleTree, we know everything about all non-scalars.
This commit reduces build times by introducing a type-erasure layer when
performing a visit on TupleTrees. Basically instead of propagating the
type of the visitor along all of the template castle, we wrap the
visitor into a virtual class with one method for each possible type in
the TupleTree.
This enables a single instatiation of visit algorithm.
This commit ensures that `target_tuple_tree_generator` can handle being
invoke multiple times on the same target.
This avoids the need to create a dedicated library as a workaround.
We used to assign an uninitialized field to `false`. However, what we
really wanted to is to construct the optional by invoking the default
constructor of the `value_type` of the optional.
Djordje Todorovic