fix(legacy): avoid undefined 32-bit shift

[youchengsong]

Summary

• avoid shifting a 32-bit dispatch transaction window by 32
• explicitly clear the window after all 32 pending slots become contiguous

Why

When window == 0xffffffffu, num_empty_slots is 32. Shifting a
32-bit value by its width is undefined behavior in C++, so the compiler is
not required to produce the intended zero value. If the bitmap is not
cleared, the dispatch transaction window can retain stale completion state.

The full-window case has no remaining pending bits, so assigning zero is the
direct representation of the intended state. Other shift counts retain the
existing behavior.

Fixes #366.

Validation

• standalone GCC 15.2 UBSan regression:
  • the old expression fails with shift exponent 32 is too large
  • the fixed expression returns 0u for the full-window case
  • shift counts 1 through 31 retain the previous result
• bash ./format.sh
• git diff --check upstream/main...HEAD

format.sh also reformats several pre-existing long lines in the same
legacy source file; the only behavioral change is the transaction-window
assignment.

[ds-review-bot]

🔵 suggestion: The MR mixes the functional fix with several formatting-only line-wrapping changes across the file (e.g., notify_dispatch, combine). These appear to come from ./format.sh and are harmless, but bundling unrelated reformatting with a bugfix increases review surface and makes bisecting harder. Consider isolating the functional change from the reformatting in future MRs.

_{🤖 v3}

[ds-review-bot]

🤖 ds-review-bot Code Review

v6

The substantive change correctly avoids shifting a 32-bit value by 32 while preserving the intended zero result when the transaction window is full. Other changes are formatting-only.

v4

⚠️ 未完成评审（no_result_file：模型未产出结果文件）

v3

This MR fixes an undefined-behavior bug in csrc/kernels/legacy/internode.cu where a 32-bit dispatch transaction window (window) could be shifted right by 32. When window == 0xffffffffu, num_empty_slots evaluates to 32; the previous window >>= num_empty_slots; shifts a 32-bit value by its full width, which is UB in C++/CUDA and may leave stale completion bits in the window bitmap. The fix window = num_empty_slots == 32 ? 0u : window >> num_empty_slots; explicitly assigns zero for the full-window case (the intended state: no remaining pending bits) and preserves the existing shift behavior for counts < 32. The boundary handling is correct and consistent with how num_empty_slots is computed ((~window) == 0 ? 32 : __ffs(~window) - 1): since window already ORs in the current offset bit before this block, num_empty_slots == 32 correctly corresponds to a fully contiguous/complete window. The remainder of the diff is formatting-only line-wrapping of long expressions (notify_dispatch, combine, etc.), consistent with running ./format.sh, and does not alter behavior. Overall the functional change is minimal, well-scoped, and correct.

Files reviewed: 1
Issues found: 🔵 1 suggestion
Inline comments posted: 1

⚠️ Parse warning: [v4] no_result_file：模型未产出结果文件