feat: Add syncall and tprefetch_async PTO ops

[FangRui0]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces the pto.tprefetch_async and pto.syncall operations, providing support for asynchronous global memory prefetching and cross-core synchronization. The changes include ODS definitions, IR verification logic, EmitC lowering patterns, and comprehensive documentation, alongside a new guide for adding PTO operations. However, the reviewer identified two critical issues where the implementations for the static helper functions isByteIntegerType and verifyAsyncFlatContiguous1DGMViewLike are missing from lib/PTO/IR/PTO.cpp, which will cause compilation failures.

[gemini-code-assist]

The implementation of the static helper function isByteIntegerType is missing from this file, which will lead to a compilation error. Please provide the implementation, for example:

static bool isByteIntegerType(Type ty) {
  auto intTy = dyn_cast<IntegerType>(ty);
  return intTy && intTy.getWidth() == 8;
}

[gemini-code-assist]

The implementation of the static helper function verifyAsyncFlatContiguous1DGMViewLike is missing from this file. This will cause a compilation error as it is called in TPrefetchAsyncOp::verify(). Please add the implementation to the file.

[reedhecre]

Codex Review

该评论由 review 机器人自动更新。

• PR: feat: Add syncall and tprefetch_async PTO ops #688 feat: Add syncall and tprefetch_async PTO ops
• Author: FangRui0
• Base/Head: main / add_newop
• Head SHA: f88ce4076967
• Trigger: PR 有新提交
• Generated At: 2026-05-20T07:35:23Z
• Previous Head SHA: 0f748d469860
• Status: completed

Summary

发现 3 个会影响新接口可用性的缺陷：tprefetch_async 有一类 memref 在验证通过后会于 EmitC 降低阶段失败，syncall 的 rank-1 tile_buf 形式无法降低，以及 tprefetch_async 文档声明的 tensor_view 输入实际无法构造。

Findings

1. P1 `tprefetch_async` 会放行动态-layout memref，但 EmitC 降低无法处理 lib/PTO/IR/PTO.cpp:2974

verifyAsyncFlatContiguous1DGMMemRef() 在检测到动态 stride/offset 时直接 return success()，所以带动态 layout 的 1D GM memref 会通过校验；但 PTOTPrefetchAsyncToEmitC 对 memref 路径总是调用 buildGlobalTensorFromMemref()，而这个 helper 只接受静态 shape + 静态 layout，失败后会在 PTOToEmitC 阶段报 failed to build GlobalTensor src。也就是说，一类当前被 verifier 明确放行的 tprefetch_async 输入会在代码生成阶段稳定失败，属于前端可见的编译阻断。

2. P2 `syncall` 声称支持 rank-1 `tile_buf` 工作区，但该形式无法降低到 C++ lib/PTO/IR/PTO.cpp:12138

verifySyncAllTileWorkspace() 明确接受 rank-1 或 rank-2 的 tile_buf/memref 工作区；但 EmitC 类型转换只支持 rank-2 TileBufType（getEmitCTileTypeString() 对非 2 维 tile 直接返回 nullopt），buildSyncAllWorkspaceTileValue() 也只对 memref 提供 1D -> 2D 的兜底路径。结果是 !pto.tile_buf<vec, Nxi32> / !pto.tile_buf<mat, Nxi32> 这种当前 verifier 认可的 syncall 形式，到了 lowering 会失败。

3. P2 `tprefetch_async` 实际并不接受文档承诺的 `!pto.tensor_view` 输入 include/PTO/IR/PTOOps.td:603

新文档和 verifyAsyncFlatContiguous1DGMViewLike() 都把 !pto.tensor_view 列为 src 的合法输入，但 TPrefetchAsyncOp 的 ODS 仍然把 src 声明成 PTODpsType，而 PTODpsType 并不包含 TensorViewType。因此前端实际上无法构造文档承诺的 tprefetch_async(%tensor_view, %ctx) 形式，这是新的 API/文档契约不一致。