Fix DSP 40-bit MAC accumulator (pink-noise / long IIR)

[JoeMatt]

Summary

The Jaguar DSP MAC is a 40-bit signed accumulator. The core kept dsp_acc in a 64-bit variable and only added products with dsp_acc += (int64_t)prod without folding to 40 bits. After many IMACN steps (typical of IIR / pink-noise style synthesis on the DSP), the emulated value diverged from hardware, with RESMAC and the high-byte read path (dsp_acc >> 32 in DSPReadLong) no longer matching a real 40-bit register (and bits 63..40 should stay zero for those reads).

This change introduces src/dsp_acc40.h with wrap/sign-extend helpers, uses them from dsp_opcode_imacn / DSP_imacn and dsp_opcode_imultn / DSP_imultn, and adds test/test_dsp_mac40.c plus a CI step (next to the blitter SIMD tests).

Context (240p / pink noise)

Long IIR MAC chains are sensitive to accumulator width. A prior hypothesis involved host __muldi3 on 32-bit libretro builds; this fix addresses emulation accuracy of the DSP MAC itself. If any platform still shows issues after this, the next place to check is linking libgcc / compiler-rt for 64-bit soft-mul on that target.

Test plan

• cc -I./src -o test/test_dsp_mac40 test/test_dsp_mac40.c && ./test/test_dsp_mac40
• Full make on macOS
• CI: new step runs with existing native SIMD test job

Made with Cursor

[JoeMatt]

verified with updated 240P test rom pink noise test now renders.

[Copilot]

Pull request overview

This PR fixes Jaguar DSP emulation accuracy by enforcing the correct 40-bit signed semantics for the DSP MAC accumulator, preventing long IMACN chains (e.g., IIR/pink-noise workloads) from diverging due to unbounded 64-bit growth.

Changes:

• Add a small header (src/dsp_acc40.h) providing 40-bit wrap + sign-extension helpers.
• Update DSP IMACN/IMULTN paths to use the helpers so dsp_acc stays in a 40-bit domain.
• Add a C regression/unit test and run it in CI alongside the existing blitter SIMD tests.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/dsp.c	Uses new helper functions for IMACN/IMULTN accumulator updates.
src/dsp_acc40.h	Introduces 40-bit accumulator wrap/sign-extend utilities.
test/test_dsp_mac40.c	Adds a unit/regression test for 40-bit accumulator behavior.
.github/workflows/c-cpp.yml	Runs the new DSP accumulator test in CI.
CLAUDE.md	Documents the new test harness entry.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

dsp_acc is now kept with bits 63..40 cleared, but dsp_opcode_sat32s / DSP_sat32s elsewhere in this file do int32_t temp = dsp_acc >> 32; and rely on the upper bits being sign-extended (so temp becomes -1/0 for in-range values). With the new 40-bit storage, negative accumulator values will produce temp in the 0..255 range and saturate incorrectly. Update those SAT32S implementations to explicitly sign-extend from the 40-bit high byte (e.g., cast via int8_t) or derive temp from dsp_acc_i40_signed().

[Copilot]

Now that the code intends to keep only the low 40 bits in dsp_acc (bits 63..40 must be zero), consider normalizing dsp_acc after savestate loads as well. STATE_LOAD_VAR(buf, dsp_acc) can restore older states where the accumulator was stored sign-extended in the upper bits; masking to 40 bits (or using the new helper) would keep the new invariant consistent immediately after load.

[Copilot]

The build hint in the file header uses -I../src, but this test is invoked from the repo root in CI (and matches other tests) with -I src. Please update the comment to reflect the correct working directory/include path (or mention both) to avoid confusing local builds.

Suggested change

	* Build: cc -O2 -Wall -I../src -o test_dsp_mac40 test/test_dsp_mac40.c
	* Build from repo root (CI): cc -O2 -Wall -I src -o test_dsp_mac40 test/test_dsp_mac40.c
	* Build from test/: cc -O2 -Wall -I../src -o test_dsp_mac40 test_dsp_mac40.c