3.0 KiB
SPIR-V global narrow load/store regression
This fixture builds a small IR::Program, runs the real shader-info pass and
EmitSPIRV, then checks both the emitted instructions and the generated module
with spirv-val. It covers signed and unsigned halfword loads at naturally
aligned offsets, preservation of high guest-address bits, and a page-end access
at byte offset 0xffe whose helper word base is 0xffc. Narrow 8- and 16-bit
stores use an aligned-word compare-exchange loop and bit-field insert. A
mesh-stage load and store preserve the upstream regression cases. The existing
zero-result behavior without 64-bit integer support is checked too.
Eight single-storage-buffer scenarios trace the actual called U32 helper's
range branch. Its logical end must use the original descriptor address plus
size, then be capped by OpArrayLength. Concrete cbuf/descriptor inputs cover
a hit at 0x1004 with size 2 and alignment 16, a miss at 0x1008, an aligned-base
control, short and zero-length descriptor caps, signed/unsigned loads and a
high 64-bit base. These eight scenarios and the eleven upstream cases emit
nineteen modules. Two more vertex fixtures check that the 32-bit and
128-bit load helpers default to DontInline and that the runtime workaround
switches only the 128-bit helper to Inline.
The additional --cas-loops OUTPUT_DIRECTORY case emits one real compute IR
program with shared/storage byte writes, a storage/global float atomic add and
a global compare-and-swap. With RX_VK_FOLLOW_STORES=1, it checks seven emitted
retry loops across the SSBO and page-table routes. Each counter must be a leading
entry-block variable, reset before its loop, incremented once per failed or
successful exchange, and tested against 1024 alongside the actual exchange
success. The failed-attempt path must feed back the returned word and retry.
The two alternative single-exchange global CAS routes must return directly,
without being counted as retry loops. CI validates all twenty-two modules.
The fixture checks the emitted SPIR-V dependency chain; it does not execute a
shader or driver, and does not validate odd guest addresses or AMD hardware. It
tests the SPIR-V backend's lowering contract only.
The workflow checks out the pinned SPIR-V-Headers revision and uses the system
SPIR-V Tools validator with --uniform-buffer-standard-layout: the single-SSBO
fixture profiles use four-byte array strides in a uniform block. This validation
option describes those fixture profiles; it does not test device feature enablement.
Reproduce it with Clang in C++20 mode, as in the Linux workflow. The focused
target includes project compatibility headers with a function-local static
inside a constexpr assertion macro. GCC rejects that construct in C++20,
while the project's production Linux compiler is Clang in C++20 mode. The
fixture does not establish an unmodified MSVC build.
To include the CAS case when reproducing manually, run the built fixture again
with RX_VK_FOLLOW_STORES=1 and --cas-loops before validating the output modules.