[Mips] Restore the return address first in the R3000A epilogue (PS2-184) #8

Merged

techwritescode merged 1 commit from r3000a-load-delay-fill into r5900-ee 2026-10-13 19:19:25 +00:00

Conversation 0 Commits 1 Files changed 6 +72 -2

techwritescode commented 2026-10-13 18:33:18 +00:00

Owner

Copy link

MIPS I has no load interlock, so a load directly before its use needs a nop. The return reads $ra, and restoring it last put a nop before the jump in nearly every non-leaf function (1,501 of the 6,337 load-delay nops in RiptOPL's IOP code, against 194 for GCC).

A new ra-restore-first subtarget feature, implied by r3000a only, restores $ra first in the epilogue so the other restores fill its delay. It is implemented as a restoreCalleeSavedRegisters override in MipsSEFrameLowering; other CPUs keep the default order unless they ask for the feature. Test: CodeGen/Mips/r3000a-epilogue.ll (first on r3000a and with +ra-restore-first, last on mips1 and with -ra-restore-first).

Results (-Os, RiptOPL's 313 IOP objects): load-delay nops 6,900 -> 5,907 (GCC 3,761); the 46 IRX files total 755,046 -> 750,950 bytes (GCC 650,850, so +16.0% -> +15.4%). The effect is smaller than the epilogue count suggests because many functions already had another instruction between the restore and the jump.

Tested

• Mips CodeGen and MC lit suites: 1,515 passed, 23 expected failures, 0 failed.
• sdk-core regression tests (8) and GCC against LLVM diff runs (15) in PCSX2 with the new compiler.
• RiptOPL builds completely (892 steps, 46 IRX).
• On hardware (Deckard): IOP register-access test, iop-module, the sdsq/sdhd/reboot modules and audsrv all pass.
• Not tested: RiptOPL itself on hardware.

Remaining load-delay nops (5,907 against GCC's 3,761) are mostly a load directly followed by its use inside one block; a pass that moves an earlier independent instruction into the slot removed only about 22 nops on the sdk-core modules, so it is not included.

🤖 Generated with Claude Code

MIPS I has no load interlock, so a load directly before its use needs a nop. The return reads `$ra`, and restoring it last put a nop before the jump in nearly every non-leaf function (1,501 of the 6,337 load-delay nops in RiptOPL's IOP code, against 194 for GCC). A new `ra-restore-first` subtarget feature, implied by `r3000a` only, restores `$ra` first in the epilogue so the other restores fill its delay. It is implemented as a `restoreCalleeSavedRegisters` override in `MipsSEFrameLowering`; other CPUs keep the default order unless they ask for the feature. Test: `CodeGen/Mips/r3000a-epilogue.ll` (first on `r3000a` and with `+ra-restore-first`, last on `mips1` and with `-ra-restore-first`). **Results** (-Os, RiptOPL's 313 IOP objects): load-delay nops 6,900 -> 5,907 (GCC 3,761); the 46 IRX files total 755,046 -> 750,950 bytes (GCC 650,850, so +16.0% -> +15.4%). The effect is smaller than the epilogue count suggests because many functions already had another instruction between the restore and the jump. **Tested** - Mips CodeGen and MC lit suites: 1,515 passed, 23 expected failures, 0 failed. - sdk-core regression tests (8) and GCC against LLVM diff runs (15) in PCSX2 with the new compiler. - RiptOPL builds completely (892 steps, 46 IRX). - On hardware (Deckard): IOP register-access test, iop-module, the sdsq/sdhd/reboot modules and audsrv all pass. - Not tested: RiptOPL itself on hardware. Remaining load-delay nops (5,907 against GCC's 3,761) are mostly a load directly followed by its use inside one block; a pass that moves an earlier independent instruction into the slot removed only about 22 nops on the sdk-core modules, so it is not included. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

techwritescode added 1 commit 2026-10-13 18:33:18 +00:00

[Mips] Restore the return address first in the R3000A epilogue ... 2b74f6a346

MIPS I has no load interlock, so a load directly before its use needs a nop.
The return reads $ra, and restoring it last put a nop before the jump in
nearly every non-leaf function. Add an ra-restore-first subtarget feature,
implied by r3000a, that restores $ra first so the other restores fill the
delay.

techwritescode referenced this pull request 2026-10-13 18:55:45 +00:00

[Mips] Share the address of a global between repeated accesses on the R3000A (PS2-184) #9

techwritescode referenced this pull request 2026-10-13 19:13:58 +00:00

[Mips] Permit tail calls on the R3000A (PS2-184) #10

techwritescode merged commit 2b74f6a346 into r5900-ee 2026-10-13 19:19:25 +00:00

techwritescode deleted branch r3000a-load-delay-fill 2026-10-13 19:19:27 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

ps2/llvm!8

No description provided.