fix(recompiler): emit a GPU-selected formatted buffer access through one shared function - #2349

What

#1959 moves the switch over every buffer format into one function per resource when the descriptor is prepared. It keeps the inline switch for GPU-selected descriptors (memory.gpuDescriptor). For those, every buffer_load_format_* / buffer_store_format_* still emits all 77 cases inline, and each case inlines the address computation, bounds checks and BDA reads of formatted().

A title with about 300 such accesses in one compute program (Ghost of Yōtei, PPSA26344, 0x8000289300, 102 guest blocks) compiles to 37.1M SPIR-V words: 292 OpSwitches, about 700k blocks, 287 distinct block shapes. Each pipeline takes 40 s or more to create on NVIDIA, and each shader disk-cache entry is up to ~300 MB. This is probably also the 13.4M-word module reported on #1183.

• formattedGpuCall emits one DontInline SPIR-V function per component count, direction, d16, coherence, kind and stop mode, and runs formatSwitch inside it.
  • It takes the descriptor fields (base, records, stride, word 3, swizzle, valid), the index, offset, soffset and the access's PC. Each access is one call.
• When BDA faults stop invocations, the function returns {value, stopped} (just stopped for a store). Inside it, a stop returns that through bdaStopValue, as read_bda_span does. The caller calls StopBdaInvocationIf on the flag, so the pixel-shader interlock and OpKill handling stay in the entry function.
• BdaInstructionPc returns the instruction's PC, or the PC parameter inside the shared function (SpirvEmitterState::bdaPcOverride). The BDA read, write, store, atomic and overflow paths use it, so a fault records the PC of the access that faulted.

Tested

Windows 11, MinGW-w64 GCC 15.2.0, RTX 4090 (driver 617.42).

• agc_formatted_buffer_once gains a GPU-selected case: dword 3 of the V# comes from v_readfirstlane_b32 s11, v0.
  • It checks that two more loads add calls but no function, no switch and less than 10% words.
  • It fails on fix(recompiler): emit a formatted buffer access through one function per resource #1959 alone ("2 more format switches") and passes with this change.
• Full ctest: 506/506 passed.
• Ghost of Yōtei, same tree before and after:
  • 0x8000289300 goes from 37,110,613 to 1,447,831 words, and its recompile from 7.2 s to 0.25 s.
  • 0x800036e100 goes from 818,855 to 260,672 words; its pipeline now takes 2.4 s.
  • The average disk-cache entry drops below 1 MB.
  • The title reaches its first presented frames (the SIE intro card). Before, a 25-minute run was still creating the second and third 37M-word pipelines.

Checklist

• Based on current main; no other open PR implements the same functions (stacked on fix(recompiler): emit a formatted buffer access through one function per resource #1959)
• 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 in TechnicalDebt
• 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: fix(recompiler): emit a formatted buffer access through one function per resource #1959
• AI-assisted: yes

🤖 Generated with Claude Code