We largely ignore the target type when preprocessing, but still
dereference options.target_type in a couple of places. E.g. when setting
info.target_type. We could add more checks for options.preprocess_only
to account for that, but it seems more robust to make sure
options.target_type is never NULL.
Since commit eaebef4265 we may receive
vsir generated from HLSL sources, which typically contains SSA values.
We could previously receive vsir with SSA values as well, but HLSL
sources would go through d3dbc/tpf as an intermediate step, making the
resulting vsir much less likely to contain SSA values.
This works around an issue introduced by commit
66cb2815f0. SV_PRIMITIVE_ID inputs in
geometry shaders use VKD3DSPR_PRIMID registers, and we create the
corresponding SPIR-V inputs using spirv_compiler_emit_io_register().
Unfortunately we also have an input signature element for the same
input, and simply creating another PrimitiveId input would run into
VUID-StandaloneSpirv-OpEntryPoint-09658.
Before the commit mentioned above, we'd use DCL_INPUT instructions to
emit input declarations, and these would help to distinguish whether
VKD3DSPR_INPUT or VKD3DSPR_PRIMID registers were used for primitive ID
inputs. Note that we can't simply ignore input signature element with
SIGNATURE_TARGET_LOCATION_UNUSED; the DXIL parser emits SV_SAMPLE_INDEX
inputs with that target location, but does require them to use a
VKD3DSPR_INPUT register.
Instead of storing the list inside struct hlsl_ctx. The source file
names in the list are used by the location information that the HLSL
frontend produces, and end up being referenced by the vsir program. If
we want the vsir program to be able to outlive the hlsl_ctx, its
location information can't reference data owned by the hlsl_ctx.
