Fix p2p comm insert sync modeling

[zhangstevenunity]

Summary

• Model pto.comm.tput and pto.comm.tget for both InsertSync and GraphSyncSolver.
• In GSS, split synchronous p2p calls into MTE2 staging and MTE3 commit RW phases so scratch writes are attributed to MTE2 and scratch reads to MTE3.
• Skip only the internal phase pair for the same TPut/TGet op, while preserving loop-carried same-op dependencies.
• Add issue706 lit coverage for InsertSync, GraphSyncSolver, and the GSS staging-tile reuse case that requires PIPE_V -> PIPE_MTE2.

Tests

• cmake -G Ninja -S . -B build-wsl ... from .codex/worktrees/issue706-p2p-gss-phase
• ninja -C build-wsl tools/ptoas/ptoas
• build-wsl/tools/ptoas/ptoas --pto-arch=a3 --pto-level=level3 --enable-insert-sync test/lit/pto/issue706_comm_p2p_insert_sync.pto | /home/rdp/llvm-workspace/llvm-project/build-shared/bin/FileCheck test/lit/pto/issue706_comm_p2p_insert_sync.pto
• build-wsl/tools/ptoas/ptoas --pto-arch=a3 --pto-level=level3 --enable-graph-sync-solver --graph-sync-solver-event-id-max=64 test/lit/pto/issue706_comm_p2p_insert_sync_gss.pto | /home/rdp/llvm-workspace/llvm-project/build-shared/bin/FileCheck test/lit/pto/issue706_comm_p2p_insert_sync_gss.pto
• build-wsl/tools/ptoas/ptoas --pto-level=level3 --pto-arch=a5 --enable-graph-sync-solver --graph-sync-solver-event-id-max=64 test/lit/pto/issue664_mscatter_pipe_selection_gss.pto | /home/rdp/llvm-workspace/llvm-project/build-shared/bin/FileCheck test/lit/pto/issue664_mscatter_pipe_selection_gss.pto
• build-wsl/tools/ptoas/ptoas --pto-level=level3 --pto-arch a5 --enable-graph-sync-solver --graph-sync-solver-event-id-max=64 test/lit/pto/insert_sync_level3_enable_gss.pto | /home/rdp/llvm-workspace/llvm-project/build-shared/bin/FileCheck test/lit/pto/insert_sync_level3_enable_gss.pto
• build-wsl/tools/ptoas/ptoas --pto-level=level3 --pto-arch a5 --enable-graph-sync-solver --graph-sync-solver-event-id-max=8 --emit-pto-ir test/lit/pto/graph_sync_solver_basic.pto | /home/rdp/llvm-workspace/llvm-project/build-shared/bin/FileCheck test/lit/pto/graph_sync_solver_basic.pto
• build-wsl/tools/ptoas/ptoas --pto-arch=a3 test/lit/pto/comm_p2p_emitc.pto 2>&1 | /home/rdp/llvm-workspace/llvm-project/build-shared/bin/FileCheck test/lit/pto/comm_p2p_emitc.pto --check-prefix=A3

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[reedhecre]

Codex Review

该评论由 review 机器人自动更新。

• PR: Fix p2p comm insert sync modeling #707 Fix p2p comm insert sync modeling
• Author: zhangstevenunity
• Base/Head: main / codex/issue706-p2p-comm-sync
• Head SHA: a4afa97e07c9
• Trigger: PR 有新提交
• Generated At: 2026-05-27T09:13:34Z
• Previous Head SHA: da3694618847
• Status: completed

Summary

GraphSyncSolver 版修复仍会漏掉循环里的 p2p stage 复用同步，存在运行时 correctness 风险。

Findings

1. P1 GSS 会把同一 p2p op 的跨迭代 phase 冲突也一起跳过 lib/PTO/Transforms/GraphSyncSolver/SyncSolver.cpp:2325

这里按 rwOp1->op == rwOp2->op 无条件跳过 pto::TPutOp/pto::TGetOp 的双 phase 配对。GraphSyncSolver 的循环建模会把同一个 loop body 复制成两组 occurrence 来发现 loop-carried hazard，所以同一个 pto.comm.tput/tget 在 iter0 的 MTE3 phase 和 iter1 的 MTE2 phase 仍然共享同一个 MLIR Operation*，也会被这个条件直接过滤掉。结果是循环里复用同一个 ping/pong stage buffer 的 p2p 通信仍然可能缺少必须的 MTE3 -> MTE2 同步，前一轮还在消费 stage 时下一轮就能覆盖它，属于真实的运行时错误。当前新增用例都是直线代码，CI 覆盖不到这个场景。

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!