Pin the rv_insn_t/opcode_fuse_t shared prefix

Closed
AL
7d ago
+27-0

Addresses the third item of #735.

try_fuse_sequence() populates the fuse array with
memcpy(ir->fuse + j, next_ir, sizeof(opcode_fuse_t)), reinterpreting the
leading bytes of rv_insn_t as an opcode_fuse_t. That is only valid while
both structures agree on the offset of every field the fused handlers read
back, but the existing _Static_asserts in decode.h pin the layout of
opcode_fuse_t alone. Nothing catches a change on the rv_insn_t side.

The issue mentions opcode, but since the memcpy copies the whole 12-byte
prefix, the contract covers imm, rd, rs1, rs2 and opcode, so this
pins all five plus a size relation.

Note that rv_insn_t gains members under EXT_C, EXT_V and EXT_F, so
the prefix layout is configuration-dependent; the asserts were checked with
EXT_V both enabled and disabled.

Verified by inserting a layout-shifting member ahead of opcode, which the
new assert rejects at compile time. Note that a uint8_t inserted there is
absorbed by existing alignment padding and does not shift opcode, so the
negative test needs a wider type.

make check passes; clang-format 20.1.7 clean.


Summary by cubic

Pins the shared prefix layout between rv_insn_t and opcode_fuse_t so the memcpy in try_fuse_sequence() can't silently corrupt fused sequences when the structs drift. Addresses the third item of #735.

  • Adds compile-time offset asserts for imm, rd, rs1, rs2, and opcode, plus a size check.