Fix loop-tail synthetic wait ordering

[TaoTao-real]

Summary

• Keep loop-tail synthetic waits ahead of post-loop local sets on the same LOOP_END anchor.
• Add an issue622 reproducer and adjust related lit checks to match the repaired ordering.

Motivation

• Fixes issue [Bug] QK_PRELOAD=4 FlashAttention DSL kernel builds but deadlocks under auto-sync #622.
• Before this change, UpdateBackwardMatchSync appended the synthetic tail wait to LOOP_END.pipeAfter. When the same anchor also carried a sunk post-loop local set reusing the same (srcPipe, dstPipe, eventId), the generated C++ could emit:
  • set(E0)
  • set(E0)
  • wait(E0)
  • wait(E0)
• That creates overlapping same-key use on the loop-tail anchor. The intended order is serialized reuse:
  • set(E0)
  • wait(E0)
  • set(E0)
  • wait(E0)

Design

• In SyncEventIdAllocation::UpdateBackwardMatchSync, insert the synthetic tail wait with push_front instead of push_back on LOOP_END.pipeAfter.
• This preserves the intended order on the shared loop-tail anchor:
  • consume the carried event first
  • then emit the new post-loop local set
• The loop-head synthetic set remains unchanged (push_back on LOOP_BEGIN.pipeBefore), which is already the conservative order for PRE syncs.
• Because this changes legal generated sync order, the affected lit checks are updated to match the new stable shape.

Testing

• Local targeted checks:
  • ptoas --pto-arch=a3 --enable-insert-sync test/lit/pto/issue428_cube_sync_regression.pto
  • ptoas --pto-arch=a3 --pto-level=level3 --enable-insert-sync test/lit/pto/issue564_k_loop_mte1_mte2_wait_regression.pto
  • ptoas --pto-arch=a3 --pto-level=level3 --enable-insert-sync test/lit/pto/issue622_v_mte2_eventid_overlap_reproducer.pto
• Verified issue622 changes from overlapping same-key use to serialized reuse at loop tail.

Risk / Rollback

• Risk: loop-tail sync order changes for synthetic backward-match waits that share a LOOP_END anchor with sunk local sets.
• Rollback: revert this PR.

[gemini-code-assist]

Code Review

This pull request addresses synchronization issues where event IDs could be incorrectly reused across outstanding sync pairs, specifically targeting issue #622. The primary change in SyncEventIdAllocation.cpp ensures that synthetic tail waits are inserted at the beginning of the pipeAfter list, allowing them to be emitted before post-loop sets. Additionally, the PR updates several regression tests and introduces a new test case to verify the fix for event ID overlap in mixed-width pipelines. I have no feedback to provide.

[reedhecre]

Codex Review

该评论由 review 机器人自动更新。

• PR: Fix loop-tail synthetic wait ordering #645 Fix loop-tail synthetic wait ordering
• Author: TaoTao-real
• Base/Head: main / codex/issue622-loop-tail-ordering
• Head SHA: d0ee3a47b8d7
• Trigger: 检测到新的 open PR
• Generated At: 2026-05-08T09:12:22Z
• Status: completed

Summary

发现 1 个 P2 问题：loop-tail synthetic wait 改为前插后，会在 ChangeNoEventIdSyncToPipeAll() 扫描同一 pipeAfter 时提前触发自修改，可能漏掉后续空 event-id 的 set，并在 codegen 阶段崩溃或生成错误同步序列。

Findings

1. P2 将 tail wait 前插后，会在 no-event-id 降级扫描里提前自修改同一 `pipeAfter` deque lib/PTO/Transforms/InsertSync/SyncEventIdAllocation.cpp:402

把 reallocated loop-tail synthetic wait 改成 push_front 后，这个 WAIT 会成为 pipeAfter 的第一个元素。后面的 ChangeNoEventIdSyncToPipeAll() 仍然用 for (auto &sync : e->pipeAfter) 顺序扫描同一个 deque；一旦先看到这个 WAIT，就会进入 MoveOutBackwardMatchSync()，而无冲突分支会再次调用 UpdateBackwardMatchSync()，继续往同一个 pipeAfter 里插入新的 tail wait。这样现在是在“还没扫描完后续 loop-end SET_EVENT”之前就自修改正在遍历的容器，后面的 empty-event-id set 可能被跳过，无法走 568-573 行的 PIPE_ALL 降级。那类遗漏的 set 到 codegen 阶段会直接取 eventIds[0]（SyncCodegen.cpp 340/353），会变成编译崩溃或坏的同步序列。旧实现把 synthetic wait 放在末尾时，至少不会在剩余 set 尚未处理前触发这条路径。