Measure Dreamcast frame rate for the first time, and plan for full speed

[awest813]

What this is

The port had never had a frame rate recorded. The Dreamcast build reports timing through SDL_WM_SetCaption, which does nothing on target, so every performance discussion so far has been hypothesis. This PR records a baseline, audits where the time goes, and orders the work to fix it.

Baseline

Flycast 2.7, stock 16 MB, 200 MHz. The Dreamcast build presents exactly one flip per emulated GBA frame, so the host emulator's frame counter reads GBA speed directly.

Workload	fps	Share of the 59.73 target
Homebrew title, fully resident	53.9	90%
Super Puzzle Fighter II, mean over 65 s	23.2	39%
Super Puzzle Fighter II, worst sample	8.5	14%

The homebrew control carries the argument. The disc, video and flip path sustain near-full speed, so the deficit on commercial code is recompiler throughput rather than presentation.

Flycast's SH-4 timing is approximate rather than cycle-exact, so these are first-order estimates of Dreamcast speed, not hardware measurements. Real GD-ROM seek latency is not reproduced at all, so large-ROM paging cost does not appear here.

The audit

DC_PERFORMANCE_PLAN_2026-09-10.md records thirteen findings and a five-phase plan. Emission costs were counted, not estimated: dc/sh4_emit.h compiles on the host against the shim tests/sh4_emit_encoding_test.c already uses, and each emitter macro was run against a counting buffer.

Translated instruction	SH-4 instructions emitted
ARM ADD r0,r1,r2	15
Hand-written equivalent	4

Leading findings:

• F1 The entire used region of the translation cache is flushed on every newly translated block, so the cost grows as the cache fills. Cheapest large win, and not previously tracked.
• F2 Every data-processing instruction compiles to a call to a one-line C helper, costing twelve instructions of address arithmetic before the branch.
• F3 32-bit constants are built a byte at a time, 9.0 instructions on average, against one for a PC-relative literal pool load.
• F4 Constant shifts emit one instruction per bit, up to 31.
• F5 The flag slots sit just past the load/store displacement window, so every flag access costs three instructions instead of one.
• F12 Automatic frameskip is the default but exists only in the PSP build, so the Dreamcast never drops a frame to catch up. Not previously tracked.
• F13 Grand Theft Auto Advance boots to its title sequence and then stops with a recompiler bad-jump fatal error.

The plan supersedes roadmap rows B8 and B9.

Known defect, not fixed here

Grand Theft Auto Advance (16 MB) halts with bad jump 1a3019f8 (3000154) (0). An indirect branch produced a target outside any valid GBA region while executing in internal RAM. Two smaller titles do not reproduce it, and one of those is also demand-paged from disc, so ROM paging alone does not explain it. This is a correctness defect rather than a performance one, but it blocks benchmarking the most demanding title available. Isolating it needs the register dump that Phase 0 instrumentation would add.

Reviewer notes

This PR carries two separate bodies of work, in two commits.

1. a0ab5b6 is the dynarec audit 4 work that was already sitting uncommitted in the worktree: the ADC/SBC/RSC flag fixes, register shift masking, masked CPSR writes, exception-return alignment, plus the new host test coverage and the save/build hardening. It is committed here so the measurement work has a reviewable base and so the plan document's links to DYNAREC_AUDIT_2026-09-10.md resolve. Review it on its own terms; it was not authored in the session that produced the second commit.
2. bda07fd is this session's work: the plan document and the recorded measurements. Documentation only.

Testing. The Dreamcast cross-build was verified with the CI container. The host suite was not run locally, because no host compiler is available in the KOS container and Git Bash has none either; CI runs make -C tests test on Ubuntu and should be treated as the gate for commit 1. The second commit changes only Markdown.

Reproducing the run. scripts/smoke-test-results.md and section 4 of the plan carry the disc recipe. Three details each cost a failed boot to find: the boot binary must be scrambled, the ISO session offset must be 11702 to match where cdi4dc places the data track, and gba_bios.bin goes at the disc root rather than under gbaDC/. mkdcdisc, which dc/dc.sh expects, is not in the CI container; scramble, mkisofs and cdi4dc are.

🤖 Generated with Claude Code