[fdfawef324]

What

mergeBesideReturns (ControlFlow/src/Structurizer.cpp) finds the merge of a selection whose arms do not meet. It computes post-dominators with the edges into blocks that only lead to a return left out. On main a block counted as one that only leads to a return when a straight run of single-predecessor blocks starts at it and ends in a return (hasLinearPathToTerminal, linearTail). An arm whose return path branches again, such as an early return inside a nested if/else, still counted as a block of its own. The arms then had no common post-dominator, and the structurizer gave the whole program to route variables ("conditional block 0 has no structured merge").

• A block now only leads to a return when it has one predecessor, the header dominates it and every successor only leads to a return (a block with no successor returns). The walk uses an explicit stack and keeps one result per block; a block on a cycle does not count.

Why it matters: in the route-variable form every original block becomes a guarded region, and the register phis at the region joins also carry values from paths that never reach a later use. In Resident Evil 4 (PPSA07411) pixel program 0x20762add00, the T# of an image_load became Phi(0, load through Phi(user data, compare mask)), and resource analysis refused it. The title stopped there, in sceAgcCreateInterpolantMapping_0100 -> Driver::ResolveShaderAbi on a game thread:

terminate called after throwing an instance of 'std::runtime_error'
  what():  shader 0x20762add00 stage 5 (shader_20762add00.req): ShaderRecompiler::Recompile: GetImageResource dword 0 is not a valid runtime value; chain: Phi(Immediate, LoadAddressU32@pc=0x00000214(GetAddressResource(Phi(GetUserData(Void), CompositeExtractU32x4(Ballot(LogicalAnd(IEqual32(GetBuiltin, Immediate), ULessThan32(Immediate, SelectU32))), Immediate)), Phi(GetUserData(Void), ...

Structured, the T# is the load through the user-data pointer, and the program prepares.

The title gets to this program only after other stops: main stops it at Value::setTypeMismatchHandler (#6126) and then at its NGG geometry states (#5884). The run above was a private build of main 8cd937e with #6126 and local changes that get it past the geometry states. The request the driver saved for the program (APS5_DUMP_SHADERS=1) shows the refusal on current main without the title, see Tested.

Overlap: #6025 (dominator sets as bitsets) also changes Structurizer.cpp; git merge-tree merges the two cleanly. No open pull request changes mergeBesideReturns.

Tested

Windows Server 2025 (10.0.26100), AMD Ryzen 9 5900X, NVIDIA GeForce RTX 3060 (driver 591.86), WinLibs MinGW-w64 GCC 15.2.0, Ninja, Release, BUILD_TESTING=ON.

• agc_structurizer_tests: a new graph (14 blocks; both arms of the header reach a return through nested branches and share one block) must keep the same executions as the original graph and must not route through a goto variable. On main (a5e2a80, with only the test added) it fails: an arm whose returns branch once a tail shared across arms is cloned: block 0 routes through a goto variable instead of a structured selection. With the change it passes.
• agc_shader_replay --prepare of the request the driver saved for the title's program:
  • main a5e2a80: resource analysis failed: GetImageResource dword 0 is not a valid runtime value; chain: Phi(Immediate, LoadAddressU32@pc=0x00000214(GetAddressResource(Phi(GetUserData(Void), ...
  • with the change (on 0e4ea13): prepared: 81214 SPIR-V words.
• ctest -R "agc|shader|video_out" -j4 on the RTX 3060, on this branch (main 0e4ea13): 621 of 622 pass, 14 skip. The failure is agc_driver_store_over_untracked (store over untracked: the first result was not stored), Windows: StoreOverUntracked fails at "the first result was not stored" #6347; it fails on main here without the change too (5 of 5 runs at a5e2a80). agc_structurizer, its _intersection_off, _iteration_selection_off and _forced_fallback modes, agc_shader_control_flow and its two modes pass, and so does control_flow_graph. Before a rebase, on main 71bd287 with the change: 617 of 618, the same failure.
• tools/check_conventions.py from main with this title and description: passes.
• Resident Evil 4 (PPSA07411), with this change on a private build (main 8cd937e + feat(libSceJson2): implement Value::setTypeMismatchHandler and store the element access failure and special float handlers #6126 + feat(agc): draw line-strip output of NGG geometry programs as mesh lines #5884 + a read-lane change like fix(recompiler): read lane spills back through relative-addressing window registers #6277 + feat(agc): rasterize a multisampled draw into one-sample targets at sample 0's locations #5907 and a follow-up to it): the title no longer stops on this program. It reached the brightness calibration screen (1,485 s run) and the "About the Settings for Sending Play Data" screen (2,096 s run); both runs ended on the harness's timed dump, with no driver failure in the log.

Checklist

• Based on current main; no other open PR implements the same functions
• 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: none
• Unblocks: PPSA07411 (no Game bring-up discussion yet; also needs feat(libSceJson2): implement Value::setTypeMismatchHandler and store the element access failure and special float handlers #6126 and feat(agc): draw line-strip output of NGG geometry programs as mesh lines #5884 to get here)
• GPUs tested: NVIDIA GeForce RTX 3060, driver 591.86, Windows Server 2025
• AI-assisted: yes (Claude Code)

🤖 Generated with Claude Code

[Dextroid17]

I ran the test case this PR adds, on Linux.

Setup: head a2f497a merged into main at ff23657 (no conflicts; the PR's base is 25 commits behind), Linux, default CMake build (CMAKE_BUILD_TYPE empty, unoptimized), clean environment (env -i with only HOME, PATH and VK_ICD_FILENAMES set). ctest -R '^agc_structurizer$', 3 runs each.

• with this PR: 3 of 3 runs pass.
• with core/shader/recompiler/ControlFlow/src/Structurizer.cpp put back to main's version (the PR's test changes kept): 3 of 3 runs fail, at "an arm whose returns branch once a tail shared across arms is cloned: block 0 routes through a goto variable instead of a structured selection".

The same-executions check on that graph passes on main too, so main gives the same paths, only through a goto variable. I only ran the structurizer test, not a title.
I did not check Windows or macOS.

AI-assisted: yes