feat(tensilelite): Add TDMLoadWaveSync

[chengchingwen]

JIRA ID: AIHPBLAS-4146

Motivation

Insert a workgroup barrier between two groups of tensor_load_to_lds — an urgent, waited-for group and a deferrable prefetch-ahead group. This PR automates that barrier insertion in StinkyTofu.

Technical Details

Adds TDMLoadWaveSyncPass (gfx1250), gated by a new off-by-default TensileLite solution param TDMLoadWaveSync. It runs after tensorcnt insertion and a kernel-scope CFGBuilder, and inserts a barrier (s_barrier_signal -1 / s_barrier_wait -1) between the two TDM groups.

• Trigger / classification. Each s_wait_tensorcnt is a trigger; its drained token set S (read from MemTokenData, now attached to emitted waits by StinkyWaitCntInsertionPass) splits preceding loads into urgent (token ∈ S) vs deferrable (token ∉ S). Wait-group identity == memtoken.
• Backward monotone dataflow. A per-trigger worklist scans backward, carrying a ScanState = {frontier of open deferrable groups, workingS}. workingS is the still-urgent subset of S and shrinks by a wait's drained tokens when the scan crosses it (an already-drained token is no longer urgent on that path); the flow ends when workingS empties or an urgent load is reached. State is unioned at CFG joins and iterated to a fixpoint — O(blocks × waits × lattice-height), no path enumeration. This covers prologue→loop, steady-state in-loop, diamond, and nested diamond-in-loop control flow.
• No-redundant / tagging. A pre-existing barrier already at a split suppresses insertion; such barriers are annotated in the emitted asm (a TDM wait-group barrier comment with the wait's drained set) so they are distinguishable from freshly-inserted and plain barriers.
• Supporting change. StinkyWaitCntInsertionPass tags each emitted s_wait_tensorcnt with the exact drained-token union (WaitDataflow::drainedTensorTokens).
• Naming. TDMLoadWaveSync is added to the kernel-name parameter set, so two solutions differing only by it do not dedup to one code object.
• Debug aid. DumpMemTokenIRStructurePass (registered in stinkytofu-opt as --DumpMemTokenIRStructurePass) dumps a reduced IR view — branch structure + memtoken-candidate instructions + comments — for inspecting wait-group flow.

TensileLite wiring: ValidParameters/GlobalParameters (default [False]), Solution.py (reject unless ScheduleIterAlg=4; disable when TDM is off or NumWaves<=1), KernelWriter.py (thread into the StinkyTofu module options), and a benchmark section in tdm_split_gfx1250.yaml.

Test Plan

• tdm_load_wave_sync.stir: 22 FileCheck cases covering flat/idempotent, cross-BB prologue→loop, steady-state in-loop, diamonds, nested diamond-in-loop, per-arm token-exact barriers, multi-token drained sets with partial-drain arms, and the no-memtoken / intervening-drain negatives.
• tdm_load_wave_sync_single_wave.stir: NumWaves==1 no-op (pass inserts nothing; matches the Solution.py guard that disables TDMLoadWaveSync when NumWaves<=1).
• Producer-side memtoken checks in waitcnt_insertion_tensor_anchor_ds_read_num_waves_1.stir.
• TensileLite: tdm_split_gfx1250.yaml exercises TDMLoadWaveSync: [False, True]
  (MXF8SS Block32, ScheduleIterAlg=4, TDMSplit, NumWaves>1) via the tox kernel-gen tests.

Test Result

Tests passed

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[cycheng]

Have had a discussion with Peter offline, refactoring (Analysis+Transform or DataFlow) will be implemented in the follow up PR.