`branch_deaths` was a relic from before I had a full understanding of
AIR's control flow structure, and so was unnecessary. This change
simplifies Liveness, fixes a bug exposed by #15235, and likely improves
performance (due to cloning hashmaps less often).
The same declaration can be added to the dependency set multiple
times, and in this case we still need to emit it once. By making
this list a hash map instead, we can do that quite easily.
This commit also introduces some additional debug logging regarding
decls.
Implements AIR lowering for is_null and is_non_null tags.
Additionally this cleans up and centralizes the logic to convert from 'direct'
representation to 'indirect' representation and vice-versa. The related functions,
as well as the functions that use it, are all moved near eachother so that the
conversion logic remains in a central place. Extracting/inserting fields
and loading/storing pointers should go through these functions.
The result-id and result-type-id of the OpPhi used to merge the
break values was not properly emitted, as some of the operands
were not written out. This caused an invalid spir-v module.
Implements code generation for the try air tag. This commit also adds
a utility `errorUnionLayout` function that helps keeping the layout
of a spir-v error union consistent.
resolve() is now able to emit instructions. If usage of this function is
interleaved with calls to emitRaw() and writeOperand(), then an
instruction may get inserted between operands, causing an invalid module.
The solution here is to just perform a temporary allocation.
The pointer to a slot in a hash map was fetched before a recursive call.
If the hash map's size changed during the recursive call, this would write
to an invalid pointer.
The solution is to use an index instead of a pointer. Note that care must be
taken that resolved types (from the type_cahce) must not be accessed, as they
might be incomplete during this operation.
Define the size of the c types according the OpenCL specification.
Note that OpenCL does not define the size of long double. Clang generates
fp128, even though there is no extension that allows such types. The
llvm-spirv translator simply crashes.
With the old logic, it was possible for a bunch of processes to queue up
to update a cache entry, and then each to do so one at a time. Now, it
rechecks whether there still a cache miss or another process has
completed the work in the interim.
When checking a cache entry with no input files for a hit, if
`createFile` returned `error.WouldBlock` we would forget about the fact
that the file has been created, and all future checks will assume that a
cache hit has happened, even though one never has or does, leading to
rare `FileNotFound` errors trying the access the protected files.
This fix works by writing an extra byte to the manifest file to
distinguish hits and misses when there no input files to write.
