jlebar wrote:
I pushed the AI's changes to your branch. Feel free to revert them or
whatever, this just seemed like a good way to share them. The text of the
review is below.
The "one thing worth flagging" at the bottom seems significant.
----------
**Written by Fable**
1. Unsourced semantic changes in the assume-operand-bundle docs (LangRef.md
~3214–3270, ~1613). The conversion, billed as a pure rst→md rewrite, silently
imported normative text from a newer upstream revision: a never-reviewed
grammar block (with a malformed "<tag>"([ <arguments>] ]) line), "attributes"
changed to "operand bundles", and new dereferenceable-bundle semantics ("at the
point of the assumption… Only %size > 0 implies dereferenceable") plus a
matching forward-reference added to the dereferenceable(<n>) attribute. Fixed
by restoring the faithful conversion of the source rst at the merge base in all
four spots.
2. Empty ```text fences in the llvm.experimental.vp.strided.load/store examples
(~24029, ~24085). Each fence opened and immediately closed, with the example IR
orphaned outside as tab-indented text — rendering an empty code box plus a
detached block. Fixed by moving the IR inside the fences.
Cleanups — all applied:
3. Leftover RST :: markers in three vector-function-abi-variant prose lines →
single :.
4. Escaped seq\_cst (×3) and va\_arg (×2) in prose → `seq_cst`/`va_arg` code
spans, matching the file's own convention.
5. The lone 4-space-indented module-flags example in Stack Alignment Metadata →
fenced ```llvm block (now gets IR highlighting like its sibling examples).
6. Redundant (vp)= anchor (added with no rst counterpart, zero references
in-tree, and shadowed by the MyST auto-anchor) → removed.
7. Four cross-doc links pinned to .rst source paths (OpaquePointers.rst ×2,
PCSectionsMetadata.rst, ExtendingLLVM.rst) → {doc}` ` role, which survives if
those docs are themselves converted to Markdown later.
8. Five anchor-free .md path links (GarbageCollection.md ×3, PointerAuth.md,
AssignmentTracking.md) → {doc}` ` role, consolidating on the mechanism already
used ~20 times in the file. Anchor-bearing links were left alone, as those
legitimately need the path form.
Skipped: nothing — no findings were refuted or judged false positives.
Verification: sphinx-build over llvm/docs completes with zero LangRef warnings,
and I checked the rendered LangRef.html directly: the strided-load/store IR now
sits inside its text code box, the :: artifacts and backslash escapes are gone,
all nine converted links resolve to the right .html targets, and the restored
assume/dereferenceable wording matches the source rst. Net diff:
llvm/docs/LangRef.md, +45/−72 lines, uncommitted in the working tree.
One thing worth flagging to the PR author: the imported newer-upstream text in
finding 1 suggests the conversion may have been done against a different LLVM
revision than the branch point — worth a skim for other sections that drifted
the same way, though the review's finders didn't surface any others.
https://github.com/llvm/llvm-project/pull/201975
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
