Reject invalid JALR and SYSTEM encodings

[thc1006]

decode_word() accept every opcode 0x67 word as JALR and every opcode 0x73 word as SYSTEM. This allows 0x00009067, whose JALR funct3 is reserved and to be lowered as a direct jump. It also allow 0x00001073, a CSR instruction, to become SYS_EXIT when a7 is 93.

JALR requires funct3 == 0, while ECALL and EBREAK use exact instruction words. Reject the other encodings, check the exact ECALL word in analyze_syscalls(), and add hand-encoded regression cases.

make check-all passes.

---

Summary by cubic

Reject invalid JALR and SYSTEM encodings to stop reserved forms from being lowered as jumps or syscalls. JALR now requires funct3==000, and SYSTEM matches only exact ECALL/EBREAK words.

• Bug Fixes
  • Decoder: JALR only with funct3 000; SYSTEM only when word is 0x00000073 (ECALL) or 0x00100073 (EBREAK); otherwise mark illegal.
  • Syscall analysis: Match the exact ECALL word; treat EBREAK and non-ECALL as bad syscalls.
  • Tests: Add hand-encoded cases in the Makefile to reject bad encodings and accept valid ones, using python3 to pack words.
  • Docs: Note decode-time checks so reserved forms are rejected as unsupported ops.

^{Written for commit 4c45bf0. Summary will update on new commits.}

[jserv]

ebreak is grouped with the decode-gate rejections, but unlike the other RVOPT_BAD_ENC cases it decodes successfully as K_SYSTEM and is rejected later as SYS_BAD in analyze_syscalls() — not by the JALR/SYSTEM decode gate. The test still passes (rc=1 + "unsupported op"), so this is only a grouping nuance: it exercises the SYS_BAD emission path rather than the decode gate the neighboring cases test. Might be worth a one-line comment saying so.

[thc1006]

Thx prof's advice. ebreak pass decode as K_SYSTEM and is reject later as SYS_BAD. Add a comment to distinguish it from the decode-gate rejection cases.

[jserv]

Thank @thc1006 for contributing!