← Research index

sra r0, 0 — shift-by-zero ISA design question (#513)

Date: 2026-06-05
Issue: #513
Status: Decision pending — awaiting input from Prof. Dos Reis and Charlie

What exists today

Both OG LCC and lccjs accept sra r0, 0 without error. They agree on the encoding.

Instruction Count field Machine word Behavior
sra r0 1 (default — no operand given) 0xa023 shifts right by 1
sra r0, 0 0 (explicit) 0xa003 different word — shifts right by 0
sra r0, 1 1 (explicit) 0xa023 same word as no-operand form

This is not a parity deviation — oracle and lccjs produce identical output for all three forms (confirmed in #502).

The open question

sra r0, 0 is legal today in both tools. The question is whether it should be:

Option What lccjs does going forward Oracle behavior Repo change needed
1 — Keep as-is Accepts sra r0, 0, encodes count=0 (distinct word) Same (accepts it) Nothing
2 — Reject it Errors: "shift count must be 1–15" Still accepts it (deliberate divergence, BY DESIGN) Tighten lower bound of evaluateImmediate in assembleSRA: 0 → 1; add test
3 — Normalize it Accepts it but silently encodes as count=1 (same word as sra r0) Still encodes count=0 (deliberate divergence) Add normalize path; add test

Options 2 and 3 would make lccjs intentionally stricter or smarter than OG LCC — both tools currently agree, so either change creates a deliberate BY DESIGN deviation. LCC.js already has precedent for this pattern (e.g. §15 .string unknown escapes, §18 name.nnn fail-fast).

Scope note

If option 2 or 3 is chosen, the same decision cascades to all five shift instructions — SRL, SLL, ROL, ROR — once #512 adds range-checking to those siblings. Currently only SRA has a range check at all; the others don't validate the count operand at the evaluateImmediate level.

Open questions for Prof. Dos Reis / Charlie

  1. ISA intent: Was count=0 for a shift instruction ever intended to be a valid encoding, or is it a silent gap in the original spec?
  2. Correct behavior: Should assemblers accept sra r0, 0 (Option 1), reject it (Option 2), or normalize it to 1 (Option 3)?
  3. All five siblings: If count=0 is invalid, should that apply uniformly to SRA, SRL, SLL, ROL, ROR?

References