Module inline 

Let-Inlining pass for TLExpr trees.

This module provides a let-inlining optimization pass that substitutes Let-bound variables into their usage sites, reducing the number of explicit bindings and enabling downstream passes to work with smaller, simpler trees.

Inlining Strategy

Three independent criteria control whether a binding is inlined:

1. Single-use inlining (inline_single_use): If the bound variable appears free exactly once in the body, inlining is always safe — it does not duplicate work.

2. Constant inlining (inline_constants): If the bound value is a Constant(f64), it is cheap to duplicate so we always inline regardless of use count.

3. Variable-alias inlining (inline_vars): If the bound value is a zero-argument Pred (which serves as a variable reference in TLExpr), we inline it unconditionally because the binding is just a rename.

A max_inline_depth guard prevents inlining of deeply nested sub-trees to keep code-size growth bounded.

Correctness

Substitution respects capture: when descending into a binder that re-uses the same variable name, substitution stops at that binder boundary.

Example

use tensorlogic_compiler::inline::{LetInliner, InlineConfig};
use tensorlogic_ir::TLExpr;

let inliner = LetInliner::with_default();
// Let x = 5.0 in Add(x, x)  →  Add(5.0, 5.0)
let expr = TLExpr::let_binding(
    "x",
    TLExpr::Constant(5.0),
    TLExpr::add(
        TLExpr::pred("x", vec![]),
        TLExpr::pred("x", vec![]),
    ),
);
let (result, stats) = inliner.run(expr);
assert_eq!(stats.constant_inlines, 1);

Re-exports

pub use config::InlineConfig;

pub use config::InlineStats;

pub use traversal::LetInliner;

Modules

config

helpers

substitute

traversal