perf(shader): reuse packed f16 operand lanes

[MarcZX10]

What

packedFloat16 emits the packed f16 operand selection and lane extraction once to form a native f16x2 add/multiply. When it emits the exact NaN/overflow repair path, it recomputes the same selected low/high halves for those operands. Cache each half lazily during the translation and reuse it for the result and repair logic. This avoids duplicate IR for extraction, OP_SEL, absolute value, and negate handling. The instruction result and existing repair rules are unchanged.

This is the smaller follow-up suggested in the review of closed #5057 after the guarded repair path showed no Palworld gain and grew the SPIR-V module. It makes no title-level performance claim. It depends on #4072, which changes the same packed-f16 translation function.

Tested

• The changed translation unit compiled with the repository's Release C++20 compiler flags (GCC 12.2).
• git diff --check passed.
• Focused execution tests were not run: the available stale CMake build expanded the test target into a broad rebuild of unrelated dependencies, so I stopped it before running out-of-scope work.

Checklist

• Based on current main; no other open PR implements this reuse
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: perf(shader): compute v_pk_fma_f16 and v_pk_fmac_f16 with OpFmaKHR #4072
• Unblocks: none
• GPUs tested: none; this is an IR reuse change
• AI-assisted: yes

[adriadam10]

Checked the diff: sourceLaneBits only memoizes laneBits per (source, half) within one packedFloat16 call. The emitted operations and bit rules are unchanged, and straight-line IR means the cached value dominates every use. Ran build + ctest for the AGC driver (including PackedFloat16Native, PackedAlu, Float16MinMaxBits, Fp16Overflow, Float16FmaRounding, Dx10Clamp); all pass. The only failure is agc_driver_tessellation, which also fails on main. The diff applies cleanly on main, so the #4072 dependency isn't needed. Merging once the full build is green.