[lucifer1004]

Summary

Add .shared::cta.u32 address forms for the classic ldmatrix APIs, so callers can convert a generic shared pointer once and keep byte-offset/swizzle arithmetic in the 32-bit PTX state-space address domain.

This is the first implementation slice of #1235. It does not add the counted register-MMA carrier described there.

Changes

• add cvta_generic_to_shared_u32(*const u8) -> u32 alongside the existing u64 descriptor-address conversion
• add x1/x2/x4 normal and transposed ldmatrix_*_shared_u32(u32) source forms
• declare those source aliases in the ldmatrix catalog, so importer dispatch, collector recognition, ABI markers, references, and generated tests remain one projection
• allow the existing nvvm.ldmatrix operation to carry either a compatible pointer or a .shared::cta.u32 state-space address
• pass u32 directly to the libNVVM inline-PTX route; materialize an address-space-3 pointer only at the LLVM-NVPTX intrinsic boundary
• extend the existing generated-ldmatrix GPU oracle and lowering tests

The large generated-file count is the repository's normal catalog-digest propagation; the semantic additions remain in the ldmatrix overlay/generator and handwritten device stubs.

Testing

• just check passes (all formatting, Clippy, unit tests, and structural guards passed; the final cargo deny --locked check could not run because this environment has no cargo-deny subcommand)
• cargo oxide run generated_ldmatrix --arch sm_120a
• cargo oxide run generated_ldmatrix --emit-nvvm-ir --arch sm_120a
• New example added (not applicable; extended the existing generated_ldmatrix oracle)

Both GPU runs passed all 128 x4 fragment checks and both 32-lane dynamic-index checks on sm_120 after the API was renamed to the explicit _shared_u32 spelling.

Additional targeted release tests passed for cuda-intrinsics-gen, cuda-device, dialect-nvvm, mir-importer, and mir-lower. Targeted Clippy passed with -D warnings, including the example in the CUDA 13.3 environment.

Checklist

• All commits signed off (git commit -s)
• SPDX headers on new source files (no new handwritten source file; generated files retain generated SPDX headers)

[nihalpasham]

Reviewed and validated:

• .shared::cta.u32 ldmatrix forms plus cvta_generic_to_shared_u32; the u32 rides to the intrinsic boundary on the NVPTX path and straight through on the libNVVM path
• slice 1 of Keep counted register-MMA shared addressing in the PTX u32 domain #1235, the counted register-MMA carrier stays open

Layered three commits:

• book catalog stamp to 3df52944; the six new _shared_u32 compatibility paths moved the hash
• rustfmt on generated_ldmatrix/src/main.rs
• merge of main after fix(mir-importer): preserve call destination evaluation order #1229 landed; it rewrote the intrinsic destination-write path the cvta lowering shares, resolved by keeping the prepare step and the widened result type, generated intrinsics verified current

Verified here: 128 fragments + 2x32 dynamic selections on sm_120, default and --arch sm_120a, full smoketest 226/226 on the reconciled head, fork CI green.

Merging.