[borisbat]

Why. The nightly linux lane failed llvm_jit_global_lookup.das on #4475: the hoisted lookups sat in iter.check, a run-once block the x86 vectorizer adds before the loop, and the test counted every block but the entry and .preheader ones as "inside the loop".

What changes.

• A lookup counts as inside the loop from the first loop header on - daslang's for_start/while_start or the vectorizer's vector.ph/vector.body, a preheader of either excluded - instead of from any block that is not a preheader.

Observable behavior.

• a lookup under a branch in a loop leaves the loop on the linux nightly lane: expected: 0, got: 2 -> passes with the same IR.
• With speculatable dropped from jit_get_global_mnh the cell still fails (got: 2), so the test keeps catching an un-hoisted lookup.

Where to look. is_loop_header and lookups_inside_loop in modules/dasLLVM/tests/llvm_jit_global_lookup.das.

Validation, claims, ledger

Validation

• The negative control above was run on this box: the test fails without the attribute and passes restored.
• The failing IR from the lane (extended_checks (linux, all) on goto: a labelled block resolves only its own labels; a computed goto checks its target #4475, run 42370157567) puts both lookups in iter.check, before vector.main.loop.iter.check and vector.ph, which the new rule reads as run-once.

Not done

• The test still reads block layout from the dump text; a lookup hoisted into a block after the loop would count, and no fixture produces one.

#nightly - the lane that runs this test is nightly.

[Copilot]

Note

Copilot was unable to complete this review at the selected tier due to model policy or plan restrictions.

🟢 Approval recommended

The update refines the regression test’s loop-header detection to avoid false failures from pre-loop blocks, and the only reviewer feedback is a non-blocking suggestion about naming/exit handling rather than an unresolved issue.

No open issues 🎉

What changed in this PR

This pull request updates the dasLLVM JIT IR-parsing regression test so that “lookups inside a loop” are counted starting from the first detected loop-header basic block (daslang loop headers or LLVM vectorizer loop headers), while excluding preheaders. This prevents false failures when LLVM introduces run-once pre-loop blocks (e.g., iter.check) that may contain hoisted lookups.

Changes:

• Add is_loop_header to recognize loop-header block labels (and exclude .preheader blocks).
• Update lookups_inside_loop to begin counting only after the first detected loop header, rather than treating most non-preheader blocks as “in-loop”.
• Update/clarify the helper behavior to match the new counting boundary.

Reviewer note (moderate): lookups_inside_loop now effectively counts lookups in every block after the first loop header because in_loop never resets, which can mislead (name/expectations) and may count post-loop blocks; consider renaming to match semantics or adding loop-exit detection to make it strictly loop-body-only.

File	Summary
modules/​dasLLVM/​tests/​llvm_jit_global_lookup.das	Refines IR block classification so loop-lookup counting starts at the first real loop header and ignores run-once pre-loop blocks (fixing the nightly false-positive).

This pull request requires changes.

_{Review effort: Balanced}