Skip to content

pr-description.md

Latest commit

 

History

History
69 lines (54 loc) · 4.49 KB

pr-description.md

File metadata and controls

69 lines (54 loc) · 4.49 KB

Summary

  • What changed: Completed decode support for all remaining standard RISC-V instructions across the I/M/A/F/D/C extensions, fixed multiple long-standing immediate-extraction bugs in the compressed-instruction decoder, and added comprehensive test coverage.
  • Why: The decoder previously returned unimplemented_instruction for RV64C integer loads/stores (c.ld, c.sd, c.ldsp, c.sdsp) and had no decode paths for any C.FP instructions (c.flw, c.fsw, c.fld, c.fsd, and their stack-pointer variants). Privileged system instructions (mret, sret, uret, wfi, sfence.vma) were misclassified as CSR operations, and fcvt.s.d was missing from the RVF decoder.

Detailed Changes

Phase 1 — Privileged & Conversion Instructions

  • RVI: decode_system() now correctly decodes mret, sret, uret, wfi, and sfence.vma as system instructions instead of falling through to CSR logic.
  • Decoder: infer_groups() tags the above mnemonics with the "system" group.
  • RVF: Added fcvt.s.d decode path for funct5=01000, fmt=00, rs2=00001.

Phase 2 — RV64C Integer Compressed Instructions

  • Implemented c.ld, c.sd, c.ldsp, c.sdsp with correct CL/CS/CI/CSS operand formatting and Capstone aliases (ld / sd).
  • Added correct immediate extractions in decoder.rs:
    • uimm_cld for c.ld / c.sd
    • uimm_cldsp for c.ldsp
    • uimm_sdsp for c.sdsp

Phase 3 — C.FP (Compressed Floating-Point) Instructions

  • Implemented all 8 C.FP instructions:
    • c.fld, c.fsd, c.fldsp, c.fsdsp (gated on D extension)
    • c.flw, c.fsw, c.flwsp, c.fswsp (gated on F extension, RV32 only)
  • Extended InstructionExtension::try_decode_compressed signature with extensions: &Extensions so the C extension can check F/D availability at decode time.
  • Updated all extension implementations (RVI, RVA, RVM, RVF, RVD, T-Head CMov) to match the new trait signature.

Immediate Extraction Fixes

Fixed several bugs in decoder.rs that produced incorrect displacements for compressed load/store instructions:

  • uimm_cl / uimm_cs (used by c.lw / c.sw)
  • uimm_css (used by c.swsp)
  • uimm_fldsp / uimm_cldsp (used by c.fldsp / c.ldsp)

Phase 4 — Tests & Cleanup

  • Added 10+ new unit tests in decoder.rs covering:
    • fcvt.s.d
    • System privileged instructions
    • RV64C (c.ld, c.sd, c.ldsp, c.sdsp)
    • C.FP on RV32 and RV64
    • Extension gating (C.FP rejected without F/D)
  • Added E2E CLI test test_riscv_new_compressed_instructions_decode.
  • Updated existing tests whose expectations were invalidated by the new decode paths:
    • test_rv64_only_compressed_families_report_unsupported_mode_on_rv32
    • test_invalid_compressed_encoding_reports_failure
    • test_json_formatter_emits_data_pseudo_for_unsupported_compressed
  • Removed dead code: decode_c_unimplemented() and the unimplemented_instruction error helper (no longer called anywhere).

Compatibility Checklist

  • This change affects CLI compatibility promises.
  • This change affects shared IR semantics or formatter behavior. New mnemonics (c.ld, c.sd, c.flw, c.fsw, c.fld, c.fsd, and stack variants) are now emitted instead of errors for previously-unhandled encodings.
  • This change affects public Rust API behavior.
  • This change updates benchmark baselines, known differences, or support-matrix claims.

Verification

  • cargo test --workspace --all-features — 222 tests pass, 0 failures

ISA / Compatibility Guardrails

  • If this change adds or changes ISA/compatibility behavior, I added or updated parity coverage under test/architectures/<arch>/test_cases.txt.
  • If this change adds or changes ISA/compatibility behavior, I added or updated Rust tests/golden/property coverage in the affected crate.
  • If parity is intentionally incomplete, I updated tests/differential/known-differences.toml with owner/expiry metadata instead of leaving the divergence undocumented.

Notes

  • Follow-up work or accepted limitations:
    • The existing uimm_cl / uimm_cs / uimm_css immediate extractions contained latent bugs that were fixed as part of this PR. Existing test cases happened to use encodings where the bugs did not manifest (e.g. inst[6]=0 and inst[11]=0), so no golden fixtures needed updating.
    • Capstone parity coverage for the new instructions is not yet present in test_cases.txt because cstool is not available in the current environment. Parity data should be regenerated once the Capstone binary is built.