- Rename GPU address spaces to match with SPIR-V spec.
- Emit `Block` Decoration for Uniform/PushConstant variables.
- Don't emit `OpTypeForwardPointer` for non-opencl targets.
(there's still a false-positive about recursive structs)
Signed-off-by: Ali Cheraghi <alichraghi@proton.me>
The old vectorization helper (WipElementWise) was clunky and a bit
annoying to use, and it wasn't really flexible enough.
This introduces a new vectorization helper, which uses Temporary and
Operation types to deduce a Vectorization to perform the operation
in a reasonably efficient manner. It removes the outer loop
required by WipElementWise so that implementations of AIR instructions
are cleaner. This helps with sanity when we start to introduce support
for composite integers.
airShift, convertToDirect, convertToIndirect, and normalize are initially
implemented using this new method.
This allows us to more sanely allocate a continuous
range of result-ids, and avoids a bunch of nasty
casting code in a few places. Its currently not used
very often, but will be useful in the future.
This reverts commit 9f0359d78f9facc38418e32b0e8c1bf6f99f0d26 in an attempt to
make the tests pass again. The CI failure from that merge should be unrelated
to this commit.
This reverts commit b822e841cda0adabe3fec260ff51c18508f7ee32, reversing
changes made to 0c99ba1eab63865592bb084feb271cd4e4b0357e.
This caused a CI failure when it landed in master branch.
The Khronos SPIRV-LLVM translator does not parse OpSource correctly. This
was causing tests to fail and other mysterious issues.
These are resolved by only generating a single OpSource instruction for now,
which does not have the source file locations also.
See https://github.com/KhronosGroup/SPIRV-LLVM-Translator/issues/2188
Most of this migration was performed automatically with `zig fmt`. There
were a few exceptions which I had to manually fix:
* `@alignCast` and `@addrSpaceCast` cannot be automatically rewritten
* `@truncate`'s fixup is incorrect for vectors
* Test cases are not formatted, and their error locations change
This is in preparation of removing indirect lowering again. Also
modifies constant() to accept a repr so that both direct as well
as indirect representations can be generated. Indirect is not yet
used, but will be used for globals.
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.
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.
It seems that some implementations may have problems with these right now,
like Intel and Rusticl. In theory, these attributes should be superficial
on the pointer type, as alignment guarantees are also added via the
alignment option of the OpLoad and OpStore instructions. Therefore, get rid
of them for now.
Entry points need to be attributed with a complete list of
global variables that they use. To that end, the global dependencies
mechanism is extended to also allow functions - when flushing the
module, the list of dependencies is examined to generate this
list of global variable result-ids.
SPIR-V cannot represent function pointers without extensions
that no vendor implements. For the time being, generate a test
kernel for each error, so that we can at least run SOME tests.
In the future we may be able to emulate function pointers in some
way, but that is not today.
SPIR-V globals must be emitted in order, so that any
declaration precedes usage. Zig, however, generates globals in
random order. To this end we keep for each global a list of
dependencies and perform a topological sort when flushing the
module.
Lowering constants is currently not really compatible with unions. In
this commit, constant lowering is drastically overhauled: instead of
playing nice and generating SPIR-V constant representations for everything
directly, we're just going to treat globals as an untyped bag of bytes (
or rather, SPIR-V 32-bit words), which we cast to the desired type at
usage. This is similar to how Rust generates constants in its LLVm backend.
Similar to function locals, taking the address of a global that does
not have an explicit address space assigned to it should result
in a generic pointer, not a global pointer. Also similar to function
locals, they cannot be generated into the generic storage class, and
so are generated into the global storage class and then cast to a
generic pointer, using OpSpecConstantOp. Note that using
OpSpecConstantOp results is only allowed by a hand full of other
OpSpecConstant instructions - which is why we generate constant
structs using OpSpecConstantComposite: These may use OpVariable
and OpSpecConstantOp results, while OpConstantComposite may not.
There are two main ways in which a value can be stored: "Direct", as it
will be operated on as an immediate value, and "indirect", as it is stored
in memory. Some types need a different representation here: Bools, for
example, are opaque in SPIR-V, and so these need to have a different
representation in memory. The bool operations are not easily interchangable
with integer operations, though, so they need to be OpTypeBool as
immediate value.
