Performance: reduce boxing churn in arithmetic and comparison hot paths

[bertysentry]

Motivation

Same baseline as #536: jawk is ~4.4x slower than gawk on BEGIN { s=0; for(i=0;i<5000000;i++) s=s+i; print s } (~6.5 s vs ~1.5 s). Beyond the operand stack itself (#536), the arithmetic opcodes allocate on every operation.

Problem

Each binary arithmetic opcode (ADD, SUBTRACT, MULTIPLY, ...) does:

Object o2 = pop();
Object o1 = pop();
double d1 = JRT.toDouble(o1);
double d2 = JRT.toDouble(o2);
double ans = d1 + d2;
push(ans);   // autoboxes a new Double on every single operation

So a 5M-iteration loop allocates at minimum 10M+ boxed Doubles (the add, the increment, the comparison result), plus the garbage-collection pressure that goes with it. INC/DEC similarly rebox on every iteration.

Possible directions

1. Keep exact integral results as Long where both operands are integral and the result fits — avoids Long -> Double churn and improves output fidelity (relates to the known quirk where printf "%s" of an incremented counter prints 1.0 where gawk prints 1).
2. Box cache: route results through Long.valueOf/a small value cache so common small integers do not allocate (Java caches -128..127 automatically once values stay long).
3. Longer term: a tagged accumulator or dual-stack design (separate double stack) so arithmetic chains stay unboxed between variable loads and stores. Larger change; only worth it after 1-2 and Performance: replace the boxed ArrayDeque operand stack with a flat Object[] stack #536 are measured.

Benchmarks should be re-run after #536, since the two overheads overlap on the same opcodes.

Measurement

Use the project's existing JMH harness rather than wall-clock CLI timings:

mvn package -Pbenchmark -DskipTests
java -jar target/jawk-*-benchmarks.jar

src/jmh/java/io/jawk/backend/AVMExpressionBenchmark.java already exercises the interpreter's expression evaluation and is the right place to anchor before/after numbers (add a tight arithmetic-loop case there if not already covered). AwkScriptBenchmark gives the end-to-end view. The gawk-vs-jawk CLI numbers above are motivation only — JMH results are the acceptance criterion.

🤖 Generated with Claude Code