Fix #249: subset valid-shape inference with dynamic offsets

[TaoTao-real]

Summary

• Fix pto.subset return type inference so valid-shape inference does not depend on subset offsets.
• Add regression test test/basic/subset_infer_dynamic_offset.mlir to cover dynamic offset + static result type.

Motivation

• Fixes pto.subset dynamic offset causes inferred v_col=? vs declared static valid mismatch #249.
• With dynamic offsets, previous inference could degrade a static valid dim to dynamic or invalid values, causing type inference conflicts for valid code.

Design

• In SubsetOp::inferReturnTypes, derive result validShape only from:
  • static sizes
  • parent tile validShape
• Explicitly remove offset-dependent valid-shape derivation.
• Invariant: dynamic offsets must not force static valid dims to become dynamic.

Testing

• Local targeted check:
  • /tmp/PTOAS_upstream_compare_20260312_165206/build/tools/ptoas/ptoas test/basic/subset_infer_dynamic_offset.mlir (pass)
• Remote full-board validation on 101.245.68.6 (2026-03-12):
  • Upstream main baseline: OK=150 FAIL=3 SKIP=7
  • main + this fix: OK=150 FAIL=3 SKIP=7
  • FAIL set unchanged (plan_memory_scopes_independent, prelu, print), no new regressions.

Risk / Rollback

• Risk: low, change is localized to subset return type inference.
• Rollback: revert this commit.

Review Focus

• Verify SubsetOp::inferReturnTypes no longer uses offsets to derive valid dims.
• Verify dynamic-offset subset with static result type now passes.
• Verify no regressions in full-board run relative to upstream baseline.

[zhangstevenunity]

@codex

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e0733a11f2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve static-offset clipping in subset valid-shape inference

The new inference path ignores subset offsets entirely and always uses min(size, parentValid), which regresses constant-offset cases where the valid region should shrink. For example, with parent v_col=32, sizes=[...,32], and a static col offset of 16, this now infers v_col=32 instead of 16; that overstates validity and diverges from the lowering path (computeSubsetValidDim in PTOViewToMemref) that still applies offset-based clipping. This can make earlier IR passes treat padded elements as valid and produce inconsistent metadata across the pipeline.

Useful? React with 👍 / 👎.