Migrate runtime from AST post-order execution to bytecode IR + VM with execution-only benchmarks and VM stack optimization (#64)

Author: Copilot

README.md

@@ -6,7 +6,7 @@
 
 ## Introduction
 
-Expression engine is a library written in pure Rust which provides an engine to compile and execute expressions. An expression indicates a string-like sentence that can be executed with some contexts and return a value (mostly, but not limited to, boolean, string and number).
+Expression engine is a library written in pure Rust which parses expressions into AST, compiles them into bytecode, and executes them with context. An expression indicates a string-like sentence that can be executed with some contexts and return a value (mostly, but not limited to, boolean, string and number).
 
 Expression engine aims to provide an engine for users that can execute complex logics using configurations without recompiling. It's a proper alternative as the basis to build business rule engines.
 

benches/execute_expression.rs

@@ -1,5 +1,6 @@
-use criterion::{criterion_group, criterion_main, Criterion};
-use expression_engine::{create_context, execute, parse_expression, Value};
+use criterion::{black_box, criterion_group, criterion_main, Criterion};
+use expression_engine::{bytecode, create_context, execute, parse_expression, Value};
+use std::sync::Arc;
 
 fn bench_execute_expression(c: &mut Criterion) {
     let input = "c = 5+3; c+=10+f; c";
@@ -8,9 +9,9 @@ fn bench_execute_expression(c: &mut Criterion) {
             execute(
                 input,
                 create_context!(
-                    "d" => 2,
-                    "b" => true,
-                    "f" => Arc::new(|_| Ok(Value::from(3)))
+                "d" => 2,
+                "b" => true,
+                "f" => Arc::new(|_| Ok(Value::from(3)))
                 ),
             )
         })
@@ -22,5 +23,60 @@ fn bench_parse_expression(c: &mut Criterion) {
     c.bench_function("parse_expression", |b| b.iter(|| parse_expression(input)));
 }
 
-criterion_group!(benches, bench_execute_expression, bench_parse_expression);
+fn create_bench_context(with_func: bool) -> expression_engine::Context {
+    if with_func {
+        create_context!(
+            "d" => 2,
+            "b" => true,
+            "f" => Arc::new(|_| Ok(Value::from(3)))
+        )
+    } else {
+        create_context!(
+            "d" => 2,
+            "b" => true
+        )
+    }
+}
+
+fn bench_execution_only_ast_vs_bytecode(c: &mut Criterion) {
+    let scenarios = [
+        ("short_expression", "1+2*3-4", false),
+        (
+            "long_expression",
+            "2+3*5-2/2+6*(2+4)-20+1+2+3+4+5+6+7+8+9+10",
+            false,
+        ),
+        ("function_call", "f(3)+2*f(4)", true),
+        ("list_map_mix", "{'a':1+2, 'b':[1,2,3,4], 5: 6>2}", false),
+        ("ternary_chain", "d > 1 ? (2 < 3 ? 11 : 12) : 13", false),
+    ];
+
+    for (name, expr, with_func) in scenarios {
+        let ast = parse_expression(expr).unwrap();
+        let program = bytecode::compile_expression(&ast).unwrap();
+        let mut group = c.benchmark_group(format!("execution_only/{}", name));
+
+        group.bench_function("ast_exec", |b| {
+            b.iter(|| {
+                let mut ctx = create_bench_context(with_func);
+                black_box(ast.exec(&mut ctx).unwrap())
+            })
+        });
+
+        group.bench_function("bytecode_exec", |b| {
+            b.iter(|| {
+                let mut ctx = create_bench_context(with_func);
+                black_box(bytecode::execute_program(&program, &mut ctx).unwrap())
+            })
+        });
+        group.finish();
+    }
+}
+
+criterion_group!(
+    benches,
+    bench_execute_expression,
+    bench_parse_expression,
+    bench_execution_only_ast_vs_bytecode
+);
 criterion_main!(benches);

benchmarks_execution_only.md

@@ -0,0 +1,34 @@
+# Execution-only benchmark: AST vs Bytecode VM
+
+Command:
+
+```bash
+cargo bench --bench execute_expression
+```
+
+Method:
+
+- Parse each expression once to AST.
+- Compile the AST once to bytecode `Program`.
+- Benchmark **execution only**:
+  - `ast.exec(&mut ctx)`
+  - `bytecode::execute_program(&program, &mut ctx)`
+- Create a fresh context per iteration for both paths.
+
+## Results (Criterion, median)
+
+| Scenario | AST exec | Bytecode exec | Delta (Bytecode vs AST) |
+| --- | ---: | ---: | ---: |
+| short_expression | 597.55 ns | 594.13 ns | -0.57% |
+| long_expression | 2.5106 µs | 2.5698 µs | +2.36% |
+| function_call | 695.27 ns | 655.10 ns | -5.78% |
+| list_map_mix | 864.45 ns | 679.70 ns | -21.37% |
+| ternary_chain | 493.75 ns | 494.15 ns | +0.08% |
+
+## Takeaways
+
+- At pure execution level, bytecode is **not uniformly slower**.
+- Bytecode is close to parity for short-expression and ternary-chain cases in this run.
+- Bytecode is slower in the long arithmetic-heavy case in this run.
+- Bytecode is faster in function-call and list/map construction cases in this run.
+- The previous end-to-end slowdown mostly comes from the extra **compile stage per call** in `execute()` (parse + compile + run).