Pointers and GEP are untyped. SPIR-V required structured OpAccessChain.
This means the backend will have to determine a good way to retrieve the
structured access from an untyped GEP. This is not a trivial problem,
and needs to be addressed to have a robust compiler.
The issue is other workstreams relies on the access chain deduction to
work. So we have 2 options:
- pause all dependent work until we have a good chain deduction.
- submit this limited fix to we can work on both this and other features
in parallel.
Choice we want to make is #2: submitting this **knowing this is not a
good** fix. It only increase the number of patterns we can work with,
thus allowing others to continue working on other parts of the backend.
This patch as-is has many limitations:
- If cannot robustly determine the depth of the structured access from a
GEP. Fixing this would require looking ahead at the full GEP chain.
- It cannot always figure out the correct access indices, especially
with dynamic indices. This will require frontend collaboration.
Because we know this is a temporary hack, this patch only impacts the
logical SPIR-V target. Physical SPIR-V, which can rely on pointer cast
remains on the old method.
Related to #145002
Prior to this change spv_ptrcast (and OpBitcast) was never emitted for
GEP resulting pointers. While such SPIR-V was (mostly) accepted by the
NEO GPU driver, the generated SPIR-V was incorrect.
The newly added test (pointers/getelementptr-bitcast-load.ll) verifies
that a correct bitcast is added for more complex cases and passes
spirv-val. The test is based on an OpenCL CTS test (basic/prefetch).
This commit adds 2 new instructions in the selector:
- OpAccessChain
- OpInBoundsAccessChain.
The choice between the two relies on the `inbounds` marker.
Those instruction are not used for OpenCL, to maintain the same
behavior as previously. They are only added when building for logical
SPIR-V, as it doesn't support the pointer equivalent.
Because logical SPIR-V doesn't support pointer cast either, the
assign_ptr_type intrinsic need to be generated so OpAccessChain gets
lowered with the correct pointer type, instead of i8*.
Fixes#66107
---------
Signed-off-by: Nathan Gauër <brioche@google.com>
