[Thomasb81]

Summary

This PR resolves the constrained randomization performance bottlenecks discussed in #876:

1. Affine and inside bounds narrowing in scope foreach solves (inline_constraints):

   • Added inline_foreach_affine_ref to recognize affine element offsets like (chunk[i] + start) and arr[idx] +/- C.
   • Narrowed bounds for inside ranges so (chunk[i] + start) inside [0:255] restricts $hi$ to $255 - \text{start} = 252$.
   • Preserved weighted dist draws when checking unweighted inside constraints if the current value already satisfies the range.
   • Enforced relational interval repair in solve_inline_foreach_elem regardless of prior pins to prevent drawing values that violate < or <= bounds.
   • Prevents spurious failures in chunked array randomization (tb_rand_grind_01) that triggered 10-attempt rejection cycles and fell back to a monolithic joint CSP solve over 8,192 variables.

2. Compiled foreach check plans (constraint_targets):

   • In check_coll_foreach: compiles element-independent foreach constraint bodies into a CompiledForeachCheck plan once before the loop, pre-evaluating loop-invariant expressions, resolving struct member offsets, and checking elements using direct bit-shift and range comparisons.
   • Replaces recursive AST expression evaluation and string/hash lookups on every element with a tight loop over raw element bits.

3. Complete factorized array solve with top-level pins (rand_csp):

   • Runs solve_forced across all constraint items during post-size factorized solving rather than only items with item_has_foreach, binding top-level element equality pins (arr[0].member == val) in inline with blocks directly on the fast path.

4. Foreach frame reuse and fallback gating (constraint_domains, constraint_targets):

   • Added body_needs_generate_and_test to skip expensive item_holds sweeps when constraints are purely structural.
   • Added body_has_bare_inside to skip re-evaluating body_inside_ranges when no bare array inside constraints exist.
   • Fast u64 bitwise splice in splice_elem_member for scalar element widths $\le 64$.
   • Added packed_struct_member_memo in declared_packed_access to memoize struct member bit offsets.

5. Formal argument fix in tests (uvm_config_db_tests):

   • Passed null as the parent argument to uvm_component::new in testbench snippets to comply with IEEE 1800-2017 §13.5.

---

Benchmarks (Release Build)

• Test Case 2 (tb_rand_grind_01.sv, 100,000 elements across 13 chunks of 8,192):

  • Baseline (0.11.4): 921.27 s (~15.3 min)
  • Reference simulator: 1.59 s user time (target: $\le 6.0$ s)
  • This PR: 4.61 s user time (~200x speedup, within $2.9\times$ of the reference simulator)

• Test Case 1 (tb_rand_grind_00.sv, 2,900 randomize() calls):

  • Baseline (0.11.4): 184.21 s
  • This PR: 50.71 s (~3.6x speedup)
  • Phase E: 27.6 s (down from 123.6 s, ~4.5x faster)
  • Phase C: 8.3 s (down from 39.3 s, ~4.7x faster)

---

Verification

• Output verified byte-for-byte against reference simulators on assertion lines for both benchmark testcases.
• Full regression suite passing: 6,186 passed, 0 failed across all test targets (classes, hierarchy, collections, misc, types, scheduling, strings, gates, lrm, perf, upf).
• Added regression test tests/classes/randomize_scope_chunked_dist_bounds.rs.

[aionhw]

Integrated by cherry-pick with @Thomasb81's authorship preserved: b19450cd, the tests 939be3dd, and the uvm_component::new fix to the config_db tests as its own commit, ac86e74b. Output on the #876 benchmarks is unchanged, with 2.2× and 125× fewer instructions.

Follow-up e0a0531e, e539bbac:

• the push and pop of this go through one helper, so an early return cannot leave an extra handle;
• the compiled check compares by operand signedness and width (§11.8.1), with mixed-sign cases falling back to the general check;
• the member memo is keyed by declaration instead of by source byte offset.

The rewritten foreach shapes can pick different solutions than before for the same seed; seeded runs remain reproducible.

Published in xezim 603c9bf3 with xezim-core 212dde0.
Validation: the full xezim suite passes (6610 tests, 0 failed, 1 ignored), with xezim-core 514 and parser 233 tests, the JIT smoke and DPI fallback tests, and a 1.92 MSRV check.