[maleadt]

SPIR-V on the CPU device has been disabled on 32-bit hosts since pocl#1883, as a stopgap for pocl#1880 ("until someone finishes it"). The crash in that issue, a double free when a 32-bit device rejected the 64-bit spirv/printf module, has since been fixed. What remained were two real gaps, both fixed here for i386 System V hosts (e.g. i686 Linux):

• The SPIR builtin wrapper. 32-bit hosts used the checked-in spir_wrapper32.ll from 2019, which only has scalar builtins: no vectors, no atomics with a scope, no work_group_barrier(flags, scope), and 38 wrappers with broken pointer parameters. generate_spir_wrapper.py now takes --address-bits 32, which types and mangles size_t parameters as 32-bit and selects the i386 ABI: clang passes and returns every OpenCL vector directly there, except that <8 x i8>, <4 x i16> and <2 x i32> arguments are coerced to i64. That holds for every distro kernel library variant (i386 through AVX-512), so the wrapper is generated once per variant. Comparing every wrapper declaration with the kernel library definitions gives 0 mismatches on all 13 i686 variants, and the same missing builtins as the x86-64 wrappers. The output for 64-bit targets is unchanged.

• Struct layouts. Before linking a SPIR module with the kernel library, PoCL replaces its datalayout with the host's. SPIR struct types aren't padded explicitly, because the SPIR datalayout already aligns long and double to 8 bytes, as OpenCL C requires. i386 aligns them to 4, so the kernels silently accessed fields, struct arguments and arrays of structs at the wrong offsets, while sizeof/offsetof (folded by the SPIR compiler) still reported the OpenCL layout. A new step before the datalayout swap replaces every struct type whose layout differs, wherever it describes memory, by a packed struct with explicit padding that keeps the SPIR offsets and size. That covers GEPs (with renumbered field indices), allocas, globals and their initializers, and the byval/sret/byref/elementtype attributes. Aggregate SSA values have no layout and keep their type; aggregate loads and stores are split into per-element accesses at the SPIR offsets. Clang does the equivalent for OpenCL C on i386 (explicitly padded struct types). Where the layouts agree, which is the case on all 64-bit hosts, the step changes nothing: on x86-64 every SPIR-V-built program.bc of the test suite is byte-identical with and without it.

ARM32 (AAPCS lowers vectors differently and needs its own wrapper ABI table), RISC-V32 and 32-bit Windows stay disabled.

Tests:

• regression/test_spirv_struct_layout: struct buffers, by-value struct arguments, nested structs, tail padding, private arrays of structs, constant initializers, and aggregate loads, stores and SSA values (from a hand-written .ll). It has 38 wrong fields on i686 without the layout fix.
• regression/test_spirv_builtins: the builtins whose wrappers depend on the address bits or the calling convention, plus scoped atomics, work_group_barrier and sub-groups.
• Both use 32-bit and 64-bit SPIR-V built from the same sources, with the commands in the sources.
• The regression tests with 64-bit SPIR-V inputs now skip on 32-bit devices, the compiler unit tests with .spirv inputs are only registered on 64-bit hosts, and spirv/printf (the test from test spirv/printf segfaults on 32-bit architectures pocl/pocl#1880) and example1 get 32-bit twins.
• A new GitHub-hosted CI job builds and tests PoCL in an i386 Debian testing container with LLVM 21. It builds with -msse2 because of Build fails on i686: error: inlining failed in call to ‘always_inline’ ‘_mm_getcsr’: target specific option mismatch pocl/pocl#2097, and excludes the tests that fail on i686 regardless of SPIR-V (listed below). LLVM 21 is needed because before it, JITLink can't handle R_386_GOT32X. With Debian trixie's LLVM 19, the CPU JIT fails on i386 with "Unsupported i386 relocation:43", which breaks many OpenCL C tests too. That is independent of this PR and probably calls for gating the JIT on LLVM 21 for i386.

Validation:

• Native i686 build (LLVM 22, -DKERNELLIB_HOST_CPU_VARIANTS=distro): 346/357 pass. The 11 failures are identical with -DENABLE_SPIRV=OFF, so they're unrelated:
  • kernel/test_printf (%ld of LONG_MIN);
  • regression/test_signal_handlers* (the SIGFPE avoidance is x86-64 only);
  • runtime/test_image_normalized_conv;
  • segfaults in runtime/test_event_free and examples/vecbuiltin*.
• The CI recipe in the Debian testing i386 container (LLVM 21): 344/344 pass with the exclusions. Without them, the only other failures are the first three items above and regression/test_issue_2005 (allocations beyond the 32-bit address space).
• x86-64 (LLVM 23): 374/374 pass.
• A struct-heavy kernel runs equally fast on i686 from OpenCL C source (Clang's padded structs) and from SPIR-V (the packed replacements).