Optimize temporary local variable allocation

[fglock]

Summary

Reduce over-allocation of temporary local variables that was causing significant bytecode bloat.

Changes

Before:

int preInitTempLocalsCount = Math.max(128, tempCountVisitor.getMaxTempCount() + 64);

• Always allocated minimum 128 slots
• Added 64-slot buffer on top of actual count
• Total minimum: 192 slots regardless of actual need

After:

int preInitTempLocalsCount = tempCountVisitor.getMaxTempCount() + 32;

• Use visitor's estimate with modest 32-slot buffer
• No artificial minimum
• Scales with actual code complexity

Results

Bytecode Reduction for japh.pl

• Total ASTORE instructions: 1019 → 255 (75% reduction)
• Null initializations: ~888 → 124 (86% reduction)
• Total bytecode lines: 3345 → 1817 (46% reduction)
• Savings: 1528 bytecode lines

Testing

• ✅ All 152 unit tests pass (100% pass rate)
• ✅ Tested with perl test runner (7010/7010 tests pass)
• ✅ No VerifyErrors or other issues

Why 32-slot buffer?

The TempLocalCountVisitor only counts 3 specific cases:

• Logical operators (&&, ||, //)
• For loops
• local() operators

But there are ~90 places in codegen that dynamically allocate temp variables (method calls, array/hash operations, regex, etc.). The visitor underestimates significantly.

Testing results:

• No buffer: Broke perl5_t tests (uni/variables.t, uni/fold.t, opbasic/cmp.t, uni/lower.t)
• Buffer=32: All tests pass, massive bytecode reduction

The 32-slot buffer provides safety margin for complex expressions without the extreme waste of the previous min-128 + 64-buffer approach.

Performance Impact

Smaller bytecode means:

• Faster class loading
• Better JIT compiler performance
• Reduced memory footprint
• Improved code cache utilization

🤖 Generated with Claude Code

[fglock]

Merged directly to master after confirming no regressions. The issue was incorrect test execution method - using jperl directly instead of perl_test_runner.pl gave false positives. All tests pass with the optimization.