[johnmaia]

What

Fixes #1497.

On Linux, GuestModuleBuilder renamed every import whose NID a guest module exports to its guest-suffixed name, whatever module the import declares: the executable's imports (GuestModuleBuilder.cpp:180 on main) and the guest modules' own (:218). A title that bundles a module exporting system NIDs then had those imports bound to it.

DOOM PPSA99666 declares snprintf, vsnprintf, strlen, memcpy, malloc and fopen from libSceLibcInternal (#G#H, module 7) and nothing from libc. The libc.prx loader shim it bundles (ps5-native-app-boilerplate) exports placeholder functions under the same NIDs, so 71 imports became <nid>#guest and returned 0: no notification text, 27 opens of uninitialized paths, no WAD. Windows already binds by the declared module (WindowsEntryStubBuilder) and was not affected.

An import now binds to a guest export only when the module it declares is a guest module:

• Executable imports: the declared module of each symbol comes from ImportModules through its relocation. A symbol declared from a module that is not a guest keeps its plain name and binds to the host library.
• Guest-module imports: the same rule with GuestSymbol::Library; such an import is not searched among guest exports, renamed, or counted as a guest dependency.
• Unchanged: imports without import-module metadata, imports declared from a guest module (a title's own libc.prx imported from libc still binds to it, as in 7103f39 and test_guest_host_libc.py), and Windows (declaresHost is false there).

test_guest_symbol_names gets two cases, using the qualified-import fixtures from test_windows_import_modules: an executable that declares one NID from libc (a guest) and one from libSceLibcInternal (a host), both exported by the guest libc.prx; and a guest module importing shared from libSceLibcInternal while another guest exports it. On main the first fails with {'EEEEEEEEEEE#guest', 'AAAAAAAAAAA#guest'}.

Overlap: #1462 edits the lines next to the executable import loop, so whichever merges second needs a small rebase. #1403 and #1249 merge cleanly.

Tested

• Linux (Ubuntu 24.04.5, GCC 13.3, RX 7600): full ctest 433/433, and the 152 guest/relinker tests again on the final revision; guest_symbol_names fails on main with the new cases and passes with this change.
• Windows 11 (WinLibs GCC 15.2.0 r7): 157 guest/relinker/windows tests pass, and the 19 relinker tests (including windows_import_modules) again on the final revision.
• DOOM PPSA99666 v1.1.1 on Linux, built from source (byte-identical to the release eboot.bin and libc.prx), relinked with its bundled libc.prx and no --exclude-sce-module, on this branch plus feat(libkernel): implement poll for guest sockets #1278 (poll), fix(libkernel): accept MSG_NOSIGNAL in send and sendto #1279 and fix(libkernel): translate scalar POSIX I/O errors #1307: 0 imports bound to the guest module (71 on main); the title opens /app0/wads/doom1.wad and stops at window creation on this headless machine, as with --exclude-sce-module libc.prx. Not run to E1M1 here (no display).

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 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: none (DOOM itself also needs feat(libkernel): implement poll for guest sockets #1278 for poll)
• AI-assisted: yes

🤖 Generated with Claude Code

[klosixgaming-arch]

Excellent fix for a subtle but important relinker issue. The problem of guest modules incorrectly binding to imports declared from host modules is well-documented and the solution is clean.

Strengths:

• Clear identification of the root cause: GuestModuleBuilder was renaming all imports matching exported NIDs regardless of declaring module
• Proper use of relocation table scanning to determine import module provenance
• Comprehensive test coverage including real-world DOOM example (PPSA99666)
• Good overlap analysis with other open PRs (#1462, #1403, #1249)
• Thorough testing on both Linux and Windows

Technical notes:

• The declaresHost lambda is a clean way to encapsulate the host vs. guest module distinction
• Building the executableModules mapping from relocation tables is the right approach for determining import provenance
• Test cases properly cover both executable imports and guest-module imports with qualified NIDs

This will significantly improve compatibility with titles that bundle libc.prx loader shims. Ready to merge.

[SleglJan]

Confirmed on Linux with a display and a discrete GPU, on top of current main.

main 4b005b2 + d06fafd (clean merge), CachyOS (kernel 7.2.9, GCC 16.2.1), NVIDIA RTX 3080 (driver 615.71.09), full build with BUILD_TESTING=ON and ANYPS5_ENABLE_SPIRV_TOOLS=ON.

• Full ctest 466/466; the 167 relinker and guest tests again on their own.
• DOOM PPSA99666 v1.1.1 (catalog release ZIP; eboot.bin and sce_module/libc.prx unwrapped from their SELF containers), relinked with the bundled libc.prx as a guest module, no --exclude-sce-module:
  • LD_DEBUG=bindings: 0 executable imports bind to libc.prx.guest.prx; 77 bind to the host libc.prx. The guest module's own 101 imports bind to host prx as before.
  • The title finds /app0/wads/doom1.wad, writes /download0/doom.log and shows the game list at 59.93 FPS (window-title counter, 1412 frames in 25 s). Run without input, so not played to E1M1.
  • As in Linux: a bundled libc.prx loader shim (ps5-native-app-boilerplate) binds libc imports to placeholder exports; DOOM PPSA99666 finds no WAD #1497 on llvmpipe and the Windows notes in docs: record DOOM homebrew integration compatibility #690, GPU presentation fails and the title switches to its CPU scaler; here after frame 426 rather than frame 1.
• Same input relinked with the main 4b005b2 relinker, same host libraries: 70 executable imports bind to the guest module, [notification] has no text, doom.log stays empty and the process exits with code 0 within a second (the failure in Linux: a bundled libc.prx loader shim (ps5-native-app-boilerplate) binds libc imports to placeholder exports; DOOM PPSA99666 finds no WAD #1497).

AI-assisted: yes (Claude Code); the runs and counts above were checked by hand.

[oneandonlydean]

This conflicts with main after today's merges; could you rebase? It's otherwise ready.

[oneandonlydean]

Merged with current main, guest_symbol_names fails at test_guest_symbol_names.py:175:

    assert importer == ["shared"], importer
AssertionError: ['shared', '__aps5_get_pending_module_args_nid_no_patch', '__aps5_set_module_init_result_nid_no_patch']

Since #1102 every guest module also imports those two __aps5_* symbols. Please rebase and check shared against the module's imports without requiring it to be the only one.

[MarcZX10]

I reproduced the reported current-main failure on b5f5362d: the implementation passes its focused behavior once the test stops requiring shared to be the only import. main injects __aps5_get_pending_module_args_nid_no_patch and __aps5_set_module_init_result_nid_no_patch into each guest module, so the useful assertion here is that shared is present and shared#guest is absent (or an equivalent membership check), rather than exact list equality.\n\nOn a detached worktree at b5f5362d, I applied the PR diff and that test-only assertion adjustment; the full Linux relinker-only build succeeded and CTest passed 97/97. Conventions and git diff --check also pass. The official PR checks still need to run on the updated head after you rebase.

[DavidSamir]

@johnmaia @oneandonlydean the rebase and the test change asked for above are on DavidSamir:fix/relinker-declared-host-imports (12cd7d17, on main b5f5362), in case it saves a round trip. Take it into this PR or carry it, whichever is easier.

• 8f5a66f9: this PR's commit, cherry-picked unchanged (author kept). It applied without conflicts.
• 8ce04ddd: the consumer check no longer requires shared to be the module's only symbol, since every guest module imports the two __aps5_* helpers after fix(libkernel): pass module args and argp to runtime-loaded modules on Linux #1102. It checks that shared is there and shared#guest is not.
• 12cd7d17: the declared-module case is converted with --macos as well. GuestModuleBuilder takes the non-Windows path for macOS, so the same binding happened there. The test reads the Mach-O bind names.

Tested

macOS 26.7.1 (Apple silicon M2), Apple clang 21.0.0, relinker-only Release build:

• ctest 90/90. With GuestModuleBuilder.cpp restored to main, guest_symbol_names fails and the other 89 pass.
• The new macOS check alone, against a relinker without the fix: _EEEEEEEEEEE#guest is bound. With the fix: _EEEEEEEEEEE.
• tools/check_conventions.py --base origin/main: passed.
• DOOM PPSA99666 v1.1.1 (catalog ZIP, unwrapped from SELF), executable imports bound to the bundled libc.prx: 70 of 170 on main b5f5362 and 0 with the branch for Linux output; 70 of 171 and 0 for --macos output.
• The same title run on macOS (macos-port 92d7098 libraries, relinker with and without this PR's change): without it, two empty notifications and exit after 7 s; with it, doom1.wad opens and the demo loop runs. Details in macOS support #1 (comment).

Not tested: Linux and Windows runs (no such host here).

AI-assisted: yes

[adriadam10]

Rebased on current main (one conflict: Io::Utf8Path in the rejectSharedImport calls) and added one commit on top, as asked by @oneandonlydean:

• test_guest_symbol_names.py: the guest-module consumer check now asserts shared is present and shared#guest is absent, instead of requiring shared to be the only import (every guest module also imports the two __aps5_* helpers since fix(libkernel): pass module args and argp to runtime-loaded modules on Linux #1102). Same change as proposed by @MarcZX100 / @DavidSamir.

Checked: relinker build + 89 relinker tests green on the rebased tree; guest_symbol_names fails with the GuestModuleBuilder.cpp change reverted. Confirmed on a display + RTX 3080 by @SleglJan. Merging on green CI.