Recompute IR flags from efunc when the rewriter changes opcodes.

[maleadt]

The canonicalize! pass rewrites Core.Intrinsics.* to cuTile's Intrinsics.* (sub_int → subi, etc.). The rewrite driver previously cleared IR_FLAG_EFFECT_FREE on opcode change because the inferred flag described the OLD op. Without the flag, downstream CSE and LICM gate-fail on every rewritten call.

Plumb effect propagation through the rewriter itself: when the driver changes an opcode or inserts a new sub-call, recompute the IR flags from the new func's declared effects, mirroring inference's flags_for_effects. The single source of truth for per-intrinsic effects is the efunc(...) overrides — the same hooks that adjust inference results during method-instance compilation (e.g., RNG ops setting effect_free=ALWAYS_FALSE). No parallel impurity list to maintain; no post-hoc walk over the SCI.

The new inferred_flags(func) helper lives next to classify_memory_op in analysis/effects.jl. It's wired into the three sites that produce new instructions: the substitution and inplace branches of apply_rewrite!, the resolve_rhs insertion of synthetic sub-calls, and the imperative commute_arith_transparent rewriter.

Concrete IR impact on the matmul inner loop: 7 → 4 surviving subi(token_id, 1) after CSE, and LICM hoists invariant cmpi/bcast/ reshape out of the k-loop. matmul moves from 43.4 → 47.5 TFLOPS (+9.4% over cuTile Python).

[AntonOresten]

matmul moves from 43.4 → 47.5 TFLOPS (+9.4% over cuTile Python)

Based on the current README it seems it's already roughly +8-9% over cuTile Python, at 46.9 TFLOPS. Is this patching a regression since then, or timing noise?

[maleadt]

Is this patching a regression since then, or timing noise?

Yeah that's wrong; it doesn't by itself improve performance, but unlocks additional rewrites I'm going to push in a next PR. The goal is to enable vectorization for the MoE example, which is missing right now (and will improve performance of that example).

[AntonOresten]

Neat!