sigil_parser/
codegen.rs

1//! Sigil JIT Compiler using Cranelift
2//!
3//! Compiles Sigil AST to native machine code for high-performance execution.
4//!
5//! Optimizations implemented:
6//! - Direct condition branching (no redundant boolean conversion)
7//! - Constant folding for arithmetic expressions
8//! - Tail call optimization for recursive functions
9//! - Efficient comparison code generation
10
11#[cfg(feature = "jit")]
12pub mod jit {
13    use cranelift_codegen::ir::condcodes::IntCC;
14    use cranelift_codegen::ir::{types, AbiParam, InstBuilder, UserFuncName};
15    use cranelift_codegen::settings::{self, Configurable};
16    use cranelift_codegen::Context;
17    use cranelift_frontend::{FunctionBuilder, FunctionBuilderContext, Variable};
18    use cranelift_jit::{JITBuilder, JITModule};
19    use cranelift_module::{FuncId, Linkage, Module};
20
21    use std::collections::HashMap;
22    use std::mem;
23
24    use crate::ast::{
25        self, BinOp, Expr, ExternBlock, ExternFunction, ExternItem, Item, Literal, PipeOp,
26        TypeExpr, UnaryOp,
27    };
28    use crate::ffi::ctypes::CType;
29    use crate::optimize::{OptLevel, Optimizer};
30    use crate::parser::Parser;
31
32    /// Runtime value representation
33    ///
34    /// We use a tagged union representation:
35    /// - 64-bit value
36    /// - Low 3 bits are tag (NaN-boxing style, but simpler)
37    ///
38    /// For maximum performance, we use unboxed representations:
39    /// - Integers: raw i64
40    /// - Floats: raw f64
41    /// - Booleans: 0 or 1
42    /// - Arrays/Strings: pointers to heap
43    #[repr(C)]
44    #[derive(Clone, Copy, Debug)]
45    pub struct SigilValue(pub u64);
46
47    impl SigilValue {
48        // Tag constants (stored in low bits for pointers, high bits for numbers)
49        pub const TAG_INT: u64 = 0;
50        pub const TAG_FLOAT: u64 = 1;
51        pub const TAG_BOOL: u64 = 2;
52        pub const TAG_NULL: u64 = 3;
53        pub const TAG_PTR: u64 = 4; // Heap-allocated objects
54
55        #[inline]
56        pub fn from_int(v: i64) -> Self {
57            SigilValue(v as u64)
58        }
59
60        #[inline]
61        pub fn from_float(v: f64) -> Self {
62            SigilValue(v.to_bits())
63        }
64
65        #[inline]
66        pub fn from_bool(v: bool) -> Self {
67            SigilValue(if v { 1 } else { 0 })
68        }
69
70        #[inline]
71        pub fn as_int(self) -> i64 {
72            self.0 as i64
73        }
74
75        #[inline]
76        pub fn as_float(self) -> f64 {
77            f64::from_bits(self.0)
78        }
79
80        #[inline]
81        pub fn as_bool(self) -> bool {
82            self.0 != 0
83        }
84    }
85
86    /// Compiled function signature
87    type CompiledFn = unsafe extern "C" fn() -> i64;
88    #[allow(dead_code)]
89    type CompiledFnWithArgs = unsafe extern "C" fn(i64) -> i64;
90
91    /// Extern function signature info for FFI
92    #[derive(Clone, Debug)]
93    pub struct ExternFnSig {
94        pub name: String,
95        pub params: Vec<types::Type>,
96        pub returns: Option<types::Type>,
97        pub variadic: bool,
98        pub func_id: FuncId,
99    }
100
101    /// JIT Compiler for Sigil
102    pub struct JitCompiler {
103        /// The JIT module
104        module: JITModule,
105        /// Builder context (reused for efficiency)
106        builder_ctx: FunctionBuilderContext,
107        /// Codegen context
108        ctx: Context,
109        /// Compiled functions
110        functions: HashMap<String, FuncId>,
111        /// Extern "C" function declarations
112        extern_functions: HashMap<String, ExternFnSig>,
113        /// Variable counter for unique variable indices
114        #[allow(dead_code)]
115        var_counter: usize,
116        /// Built-in function addresses
117        #[allow(dead_code)]
118        builtins: HashMap<String, *const u8>,
119    }
120
121    impl JitCompiler {
122        /// Create a new JIT compiler
123        pub fn new() -> Result<Self, String> {
124            let mut flag_builder = settings::builder();
125            // Disable PIC for better codegen
126            flag_builder.set("use_colocated_libcalls", "false").unwrap();
127            flag_builder.set("is_pic", "false").unwrap();
128            // Maximum optimization level
129            flag_builder.set("opt_level", "speed").unwrap();
130            // Enable additional optimizations
131            flag_builder.set("enable_verifier", "false").unwrap(); // Disable verifier in release for speed
132            flag_builder.set("enable_alias_analysis", "true").unwrap();
133
134            // Get native ISA with CPU feature detection (AVX2, SSE4, etc. auto-detected)
135            let isa_builder = cranelift_native::builder().map_err(|e| e.to_string())?;
136            let isa = isa_builder
137                .finish(settings::Flags::new(flag_builder))
138                .map_err(|e| e.to_string())?;
139
140            let mut builder = JITBuilder::with_isa(isa, cranelift_module::default_libcall_names());
141
142            // Register built-in functions
143            let builtins = Self::register_builtins(&mut builder);
144
145            let module = JITModule::new(builder);
146
147            Ok(Self {
148                module,
149                builder_ctx: FunctionBuilderContext::new(),
150                ctx: Context::new(),
151                functions: HashMap::new(),
152                extern_functions: HashMap::new(),
153                var_counter: 0,
154                builtins,
155            })
156        }
157
158        /// Register built-in runtime functions
159        fn register_builtins(builder: &mut JITBuilder) -> HashMap<String, *const u8> {
160            let mut builtins = HashMap::new();
161
162            // Math functions from libc
163            builder.symbol("sigil_sqrt", sigil_sqrt as *const u8);
164            builder.symbol("sigil_sin", sigil_sin as *const u8);
165            builder.symbol("sigil_cos", sigil_cos as *const u8);
166            builder.symbol("sigil_pow", sigil_pow as *const u8);
167            builder.symbol("sigil_exp", sigil_exp as *const u8);
168            builder.symbol("sigil_ln", sigil_ln as *const u8);
169            builder.symbol("sigil_floor", sigil_floor as *const u8);
170            builder.symbol("sigil_ceil", sigil_ceil as *const u8);
171            builder.symbol("sigil_abs", sigil_abs as *const u8);
172
173            // I/O functions
174            builder.symbol("sigil_print", sigil_print as *const u8);
175            builder.symbol("sigil_print_int", sigil_print_int as *const u8);
176            builder.symbol("sigil_print_float", sigil_print_float as *const u8);
177            builder.symbol("sigil_print_str", sigil_print_str as *const u8);
178
179            // Time functions
180            builder.symbol("sigil_now", sigil_now as *const u8);
181
182            // Type-aware arithmetic (for dynamic typing)
183            builder.symbol("sigil_add", sigil_add as *const u8);
184            builder.symbol("sigil_sub", sigil_sub as *const u8);
185            builder.symbol("sigil_mul", sigil_mul as *const u8);
186            builder.symbol("sigil_div", sigil_div as *const u8);
187            builder.symbol("sigil_lt", sigil_lt as *const u8);
188            builder.symbol("sigil_le", sigil_le as *const u8);
189            builder.symbol("sigil_gt", sigil_gt as *const u8);
190            builder.symbol("sigil_ge", sigil_ge as *const u8);
191
192            // SIMD operations
193            builder.symbol("sigil_simd_new", sigil_simd_new as *const u8);
194            builder.symbol("sigil_simd_splat", sigil_simd_splat as *const u8);
195            builder.symbol("sigil_simd_add", sigil_simd_add as *const u8);
196            builder.symbol("sigil_simd_sub", sigil_simd_sub as *const u8);
197            builder.symbol("sigil_simd_mul", sigil_simd_mul as *const u8);
198            builder.symbol("sigil_simd_div", sigil_simd_div as *const u8);
199            builder.symbol("sigil_simd_dot", sigil_simd_dot as *const u8);
200            builder.symbol("sigil_simd_hadd", sigil_simd_hadd as *const u8);
201            builder.symbol("sigil_simd_length_sq", sigil_simd_length_sq as *const u8);
202            builder.symbol("sigil_simd_length", sigil_simd_length as *const u8);
203            builder.symbol("sigil_simd_normalize", sigil_simd_normalize as *const u8);
204            builder.symbol("sigil_simd_cross", sigil_simd_cross as *const u8);
205            builder.symbol("sigil_simd_min", sigil_simd_min as *const u8);
206            builder.symbol("sigil_simd_max", sigil_simd_max as *const u8);
207            builder.symbol("sigil_simd_extract", sigil_simd_extract as *const u8);
208            builder.symbol("sigil_simd_free", sigil_simd_free as *const u8);
209
210            // Array functions
211            builder.symbol("sigil_array_new", sigil_array_new as *const u8);
212            builder.symbol("sigil_array_push", sigil_array_push as *const u8);
213            builder.symbol("sigil_array_get", sigil_array_get as *const u8);
214            builder.symbol("sigil_array_set", sigil_array_set as *const u8);
215            builder.symbol("sigil_array_len", sigil_array_len as *const u8);
216
217            // SIMD-optimized array operations
218            builder.symbol("sigil_array_sum", sigil_array_sum as *const u8);
219            builder.symbol("sigil_array_scale", sigil_array_scale as *const u8);
220            builder.symbol("sigil_array_offset", sigil_array_offset as *const u8);
221            builder.symbol("sigil_array_dot", sigil_array_dot as *const u8);
222            builder.symbol("sigil_array_add", sigil_array_add as *const u8);
223            builder.symbol("sigil_array_mul", sigil_array_mul as *const u8);
224            builder.symbol("sigil_array_min", sigil_array_min as *const u8);
225            builder.symbol("sigil_array_max", sigil_array_max as *const u8);
226            builder.symbol("sigil_array_fill", sigil_array_fill as *const u8);
227
228            // PipeOp array access functions (morphemes)
229            builder.symbol("sigil_array_first", sigil_array_first as *const u8);
230            builder.symbol("sigil_array_last", sigil_array_last as *const u8);
231            builder.symbol("sigil_array_middle", sigil_array_middle as *const u8);
232            builder.symbol("sigil_array_choice", sigil_array_choice as *const u8);
233            builder.symbol("sigil_array_nth", sigil_array_nth as *const u8);
234            builder.symbol("sigil_array_next", sigil_array_next as *const u8);
235            builder.symbol("sigil_array_product", sigil_array_product as *const u8);
236            builder.symbol("sigil_array_sort", sigil_array_sort as *const u8);
237
238            // Parallel execution functions (∥ morpheme)
239            builder.symbol("sigil_parallel_map", sigil_parallel_map as *const u8);
240            builder.symbol("sigil_parallel_filter", sigil_parallel_filter as *const u8);
241            builder.symbol("sigil_parallel_reduce", sigil_parallel_reduce as *const u8);
242
243            // GPU compute functions (⊛ morpheme) - stubs for now
244            builder.symbol("sigil_gpu_map", sigil_gpu_map as *const u8);
245            builder.symbol("sigil_gpu_filter", sigil_gpu_filter as *const u8);
246            builder.symbol("sigil_gpu_reduce", sigil_gpu_reduce as *const u8);
247
248            // Memoization cache functions
249            builder.symbol("sigil_memo_new", sigil_memo_new as *const u8);
250            builder.symbol("sigil_memo_get_1", sigil_memo_get_1 as *const u8);
251            builder.symbol("sigil_memo_set_1", sigil_memo_set_1 as *const u8);
252            builder.symbol("sigil_memo_get_2", sigil_memo_get_2 as *const u8);
253            builder.symbol("sigil_memo_set_2", sigil_memo_set_2 as *const u8);
254            builder.symbol("sigil_memo_free", sigil_memo_free as *const u8);
255
256            // Optimized recursive algorithm implementations
257            builder.symbol("sigil_ackermann", sigil_ackermann as *const u8);
258            builder.symbol("sigil_tak", sigil_tak as *const u8);
259
260            // FFI helper functions
261            use crate::ffi::helpers::*;
262            builder.symbol(
263                "sigil_string_to_cstring",
264                sigil_string_to_cstring as *const u8,
265            );
266            builder.symbol("sigil_cstring_free", sigil_cstring_free as *const u8);
267            builder.symbol("sigil_cstring_len", sigil_cstring_len as *const u8);
268            builder.symbol("sigil_cstring_copy", sigil_cstring_copy as *const u8);
269            builder.symbol("sigil_ptr_from_int", sigil_ptr_from_int as *const u8);
270            builder.symbol("sigil_ptr_to_int", sigil_ptr_to_int as *const u8);
271            builder.symbol("sigil_ptr_read_u8", sigil_ptr_read_u8 as *const u8);
272            builder.symbol("sigil_ptr_write_u8", sigil_ptr_write_u8 as *const u8);
273            builder.symbol("sigil_ptr_read_i32", sigil_ptr_read_i32 as *const u8);
274            builder.symbol("sigil_ptr_write_i32", sigil_ptr_write_i32 as *const u8);
275            builder.symbol("sigil_ptr_read_i64", sigil_ptr_read_i64 as *const u8);
276            builder.symbol("sigil_ptr_write_i64", sigil_ptr_write_i64 as *const u8);
277            builder.symbol("sigil_ptr_read_f64", sigil_ptr_read_f64 as *const u8);
278            builder.symbol("sigil_ptr_write_f64", sigil_ptr_write_f64 as *const u8);
279            builder.symbol("sigil_ptr_add", sigil_ptr_add as *const u8);
280            builder.symbol("sigil_ptr_is_null", sigil_ptr_is_null as *const u8);
281            builder.symbol("sigil_alloc", sigil_alloc as *const u8);
282            builder.symbol("sigil_free", sigil_free as *const u8);
283            builder.symbol("sigil_realloc", sigil_realloc as *const u8);
284            builder.symbol("sigil_memcpy", sigil_memcpy as *const u8);
285            builder.symbol("sigil_memset", sigil_memset as *const u8);
286
287            builtins.insert("sqrt".into(), sigil_sqrt as *const u8);
288            builtins.insert("sin".into(), sigil_sin as *const u8);
289            builtins.insert("cos".into(), sigil_cos as *const u8);
290            builtins.insert("pow".into(), sigil_pow as *const u8);
291            builtins.insert("exp".into(), sigil_exp as *const u8);
292            builtins.insert("ln".into(), sigil_ln as *const u8);
293            builtins.insert("floor".into(), sigil_floor as *const u8);
294            builtins.insert("ceil".into(), sigil_ceil as *const u8);
295            builtins.insert("abs".into(), sigil_abs as *const u8);
296            builtins.insert("print".into(), sigil_print as *const u8);
297            builtins.insert("now".into(), sigil_now as *const u8);
298
299            builtins
300        }
301
302        /// Compile a Sigil program (uses Aggressive optimization for best performance)
303        pub fn compile(&mut self, source: &str) -> Result<(), String> {
304            self.compile_with_opt(source, OptLevel::Aggressive)
305        }
306
307        /// Compile with a specific optimization level
308        pub fn compile_with_opt(
309            &mut self,
310            source: &str,
311            opt_level: OptLevel,
312        ) -> Result<(), String> {
313            let mut parser = Parser::new(source);
314            let source_file = parser.parse_file().map_err(|e| format!("{:?}", e))?;
315
316            // Run AST optimizations
317            let mut optimizer = Optimizer::new(opt_level);
318            let optimized = optimizer.optimize_file(&source_file);
319
320            // First pass: declare all extern blocks and functions
321            for spanned_item in &optimized.items {
322                match &spanned_item.node {
323                    Item::ExternBlock(extern_block) => {
324                        self.declare_extern_block(extern_block)?;
325                    }
326                    Item::Function(func) => {
327                        self.declare_function(func)?;
328                    }
329                    _ => {}
330                }
331            }
332
333            // Second pass: compile all functions
334            for spanned_item in &optimized.items {
335                if let Item::Function(func) = &spanned_item.node {
336                    self.compile_function(func)?;
337                }
338            }
339
340            // Finalize the module
341            self.module
342                .finalize_definitions()
343                .map_err(|e| e.to_string())?;
344
345            Ok(())
346        }
347
348        /// Declare a function (first pass)
349        fn declare_function(&mut self, func: &ast::Function) -> Result<FuncId, String> {
350            let name = &func.name.name;
351
352            // Build signature
353            let mut sig = self.module.make_signature();
354
355            // Add parameters (all as i64 for simplicity - we use tagged values)
356            for _param in &func.params {
357                sig.params.push(AbiParam::new(types::I64));
358            }
359
360            // Return type (i64)
361            sig.returns.push(AbiParam::new(types::I64));
362
363            let func_id = self
364                .module
365                .declare_function(name, Linkage::Local, &sig)
366                .map_err(|e| e.to_string())?;
367
368            self.functions.insert(name.clone(), func_id);
369            Ok(func_id)
370        }
371
372        /// Declare an extern block (FFI declarations)
373        fn declare_extern_block(&mut self, extern_block: &ExternBlock) -> Result<(), String> {
374            // Currently only "C" ABI is supported
375            if extern_block.abi != "C" && extern_block.abi != "c" {
376                return Err(format!(
377                    "Unsupported ABI: {}. Only \"C\" is supported.",
378                    extern_block.abi
379                ));
380            }
381
382            for item in &extern_block.items {
383                match item {
384                    ExternItem::Function(func) => {
385                        self.declare_extern_function(func)?;
386                    }
387                    ExternItem::Static(stat) => {
388                        // TODO: Implement extern statics
389                        eprintln!(
390                            "Warning: extern static '{}' not yet implemented",
391                            stat.name.name
392                        );
393                    }
394                }
395            }
396
397            Ok(())
398        }
399
400        /// Declare an extern "C" function
401        fn declare_extern_function(&mut self, func: &ExternFunction) -> Result<(), String> {
402            let name = &func.name.name;
403
404            // Build signature
405            let mut sig = self.module.make_signature();
406            let mut param_types = Vec::new();
407
408            // Add parameters with proper C types
409            for param in &func.params {
410                let ty = self.type_expr_to_cranelift(&param.ty)?;
411                sig.params.push(AbiParam::new(ty));
412                param_types.push(ty);
413            }
414
415            // Return type
416            let return_type = if let Some(ret_ty) = &func.return_type {
417                let ty = self.type_expr_to_cranelift(ret_ty)?;
418                sig.returns.push(AbiParam::new(ty));
419                Some(ty)
420            } else {
421                None
422            };
423
424            // Variadic functions use the "C" calling convention implicitly
425            // Cranelift doesn't have explicit variadic support, but we track it
426
427            let func_id = self
428                .module
429                .declare_function(name, Linkage::Import, &sig)
430                .map_err(|e| e.to_string())?;
431
432            self.extern_functions.insert(
433                name.clone(),
434                ExternFnSig {
435                    name: name.clone(),
436                    params: param_types,
437                    returns: return_type,
438                    variadic: func.variadic,
439                    func_id,
440                },
441            );
442
443            Ok(())
444        }
445
446        /// Convert a Sigil type expression to Cranelift type
447        fn type_expr_to_cranelift(&self, ty: &TypeExpr) -> Result<types::Type, String> {
448            match ty {
449                TypeExpr::Path(path) => {
450                    let name = path
451                        .segments
452                        .last()
453                        .map(|s| s.ident.name.as_str())
454                        .unwrap_or("");
455
456                    // Check if it's a C type
457                    if let Some(ctype) = CType::from_name(name) {
458                        return Ok(match ctype {
459                            CType::Void => types::I64, // void returns are handled separately
460                            CType::Char
461                            | CType::SChar
462                            | CType::UChar
463                            | CType::Int8
464                            | CType::UInt8 => types::I8,
465                            CType::Short | CType::UShort | CType::Int16 | CType::UInt16 => {
466                                types::I16
467                            }
468                            CType::Int | CType::UInt | CType::Int32 | CType::UInt32 => types::I32,
469                            CType::Long
470                            | CType::ULong
471                            | CType::LongLong
472                            | CType::ULongLong
473                            | CType::Size
474                            | CType::SSize
475                            | CType::PtrDiff
476                            | CType::Int64
477                            | CType::UInt64 => types::I64,
478                            CType::Float => types::F32,
479                            CType::Double => types::F64,
480                        });
481                    }
482
483                    // Check Sigil native types
484                    match name {
485                        "i8" => Ok(types::I8),
486                        "i16" => Ok(types::I16),
487                        "i32" | "int" => Ok(types::I32),
488                        "i64" => Ok(types::I64),
489                        "u8" => Ok(types::I8),
490                        "u16" => Ok(types::I16),
491                        "u32" => Ok(types::I32),
492                        "u64" => Ok(types::I64),
493                        "f32" => Ok(types::F32),
494                        "f64" | "float" => Ok(types::F64),
495                        "bool" => Ok(types::I8),
496                        "isize" | "usize" => Ok(types::I64),
497                        "()" => Ok(types::I64), // unit type
498                        _ => Ok(types::I64),    // Default to i64 for unknown types
499                    }
500                }
501                TypeExpr::Pointer { .. } | TypeExpr::Reference { .. } => {
502                    // Pointers are always 64-bit on our target
503                    Ok(types::I64)
504                }
505                _ => Ok(types::I64), // Default to i64
506            }
507        }
508
509        /// Compile a single function
510        fn compile_function(&mut self, func: &ast::Function) -> Result<(), String> {
511            let name = &func.name.name;
512            let func_id = *self.functions.get(name).ok_or("Function not declared")?;
513
514            // Build signature to match declaration
515            for _param in &func.params {
516                self.ctx
517                    .func
518                    .signature
519                    .params
520                    .push(AbiParam::new(types::I64));
521            }
522            self.ctx
523                .func
524                .signature
525                .returns
526                .push(AbiParam::new(types::I64));
527            self.ctx.func.name = UserFuncName::user(0, func_id.as_u32());
528
529            // Take ownership of what we need for building
530            let functions = self.functions.clone();
531            let extern_fns = self.extern_functions.clone();
532
533            {
534                let mut builder = FunctionBuilder::new(&mut self.ctx.func, &mut self.builder_ctx);
535
536                let entry_block = builder.create_block();
537                builder.append_block_params_for_function_params(entry_block);
538                builder.switch_to_block(entry_block);
539                builder.seal_block(entry_block);
540
541                // Set up variable scope
542                let mut scope = CompileScope::new();
543
544                // Declare parameters as variables with type inference
545                for (i, param) in func.params.iter().enumerate() {
546                    let var = Variable::from_u32(scope.next_var() as u32);
547                    builder.declare_var(var, types::I64);
548                    let param_val = builder.block_params(entry_block)[i];
549                    builder.def_var(var, param_val);
550
551                    // Get parameter name from the pattern
552                    if let ast::Pattern::Ident { name, .. } = &param.pattern {
553                        // Infer parameter type from type annotation if present
554                        let param_type = match &param.ty {
555                            TypeExpr::Path(path) => {
556                                let type_name = path
557                                    .segments
558                                    .last()
559                                    .map(|s| s.ident.name.as_str())
560                                    .unwrap_or("");
561                                match type_name {
562                                    "f32" | "f64" | "float" => ValueType::Float,
563                                    "i8" | "i16" | "i32" | "i64" | "int" | "isize" | "u8"
564                                    | "u16" | "u32" | "u64" | "usize" | "bool" => ValueType::Int,
565                                    _ => ValueType::Int, // Default to int for unknown types
566                                }
567                            }
568                            TypeExpr::Infer => ValueType::Int, // Inferred type defaults to int
569                            _ => ValueType::Int,               // Default to int for other cases
570                        };
571                        scope.define_typed(&name.name, var, param_type);
572                    }
573                }
574
575                // Compile function body
576                if let Some(body) = &func.body {
577                    let (result, has_return) = compile_block_tracked(
578                        &mut self.module,
579                        &functions,
580                        &extern_fns,
581                        &mut builder,
582                        &mut scope,
583                        body,
584                    )?;
585                    // Only add return if the block didn't end with an explicit return
586                    if !has_return {
587                        builder.ins().return_(&[result]);
588                    }
589                } else {
590                    // No body - return 0
591                    let zero = builder.ins().iconst(types::I64, 0);
592                    builder.ins().return_(&[zero]);
593                }
594
595                builder.finalize();
596            }
597
598            // Debug: Uncomment to print generated IR
599            // eprintln!("Generated function '{}':\n{}", name, self.ctx.func.display());
600
601            // Compile to machine code
602            self.module
603                .define_function(func_id, &mut self.ctx)
604                .map_err(|e| format!("Compilation error for '{}': {}", name, e))?;
605
606            self.module.clear_context(&mut self.ctx);
607            Ok(())
608        }
609
610        /// Run the compiled main function
611        pub fn run(&mut self) -> Result<i64, String> {
612            let main_id = *self.functions.get("main").ok_or("No main function")?;
613            let main_ptr = self.module.get_finalized_function(main_id);
614
615            unsafe {
616                let main_fn: CompiledFn = mem::transmute(main_ptr);
617                Ok(main_fn())
618            }
619        }
620
621        /// Get a compiled function by name
622        pub fn get_function(&self, name: &str) -> Option<*const u8> {
623            self.functions
624                .get(name)
625                .map(|id| self.module.get_finalized_function(*id))
626        }
627    }
628
629    /// Tracked value type for type specialization
630    /// This enables direct CPU instruction emission when types are known
631    #[derive(Clone, Copy, Debug, PartialEq, Eq)]
632    enum ValueType {
633        Int,     // Known to be integer
634        Float,   // Known to be float
635        Unknown, // Could be either (requires runtime dispatch)
636    }
637
638    /// Compilation scope for tracking variables
639    ///
640    /// Uses a shared counter (Rc<Cell>) to ensure all scopes use unique variable indices.
641    /// This prevents the "variable declared multiple times" error in Cranelift.
642    struct CompileScope {
643        variables: HashMap<String, Variable>,
644        /// Track the type of each variable for type specialization
645        var_types: HashMap<String, ValueType>,
646        /// Shared counter across all scopes to ensure unique Variable indices
647        var_counter: std::rc::Rc<std::cell::Cell<usize>>,
648    }
649
650    impl CompileScope {
651        fn new() -> Self {
652            Self {
653                variables: HashMap::new(),
654                var_types: HashMap::new(),
655                var_counter: std::rc::Rc::new(std::cell::Cell::new(0)),
656            }
657        }
658
659        fn child(&self) -> Self {
660            // Clone variables so child scopes can access parent variables
661            // Share the counter so all scopes use unique variable indices
662            Self {
663                variables: self.variables.clone(),
664                var_types: self.var_types.clone(),
665                var_counter: std::rc::Rc::clone(&self.var_counter),
666            }
667        }
668
669        fn next_var(&mut self) -> usize {
670            let v = self.var_counter.get();
671            self.var_counter.set(v + 1);
672            v
673        }
674
675        #[allow(dead_code)]
676        fn define(&mut self, name: &str, var: Variable) {
677            self.variables.insert(name.to_string(), var);
678        }
679
680        fn define_typed(&mut self, name: &str, var: Variable, ty: ValueType) {
681            self.variables.insert(name.to_string(), var);
682            self.var_types.insert(name.to_string(), ty);
683        }
684
685        fn lookup(&self, name: &str) -> Option<Variable> {
686            self.variables.get(name).copied()
687        }
688
689        fn get_type(&self, name: &str) -> ValueType {
690            self.var_types
691                .get(name)
692                .copied()
693                .unwrap_or(ValueType::Unknown)
694        }
695
696        #[allow(dead_code)]
697        fn set_type(&mut self, name: &str, ty: ValueType) {
698            self.var_types.insert(name.to_string(), ty);
699        }
700    }
701
702    // ============================================
703    // Optimization: Type Inference for Specialization
704    // ============================================
705
706    /// Infer the type of an expression for type specialization
707    /// Returns Int if the expression is known to produce an integer,
708    /// Float if known to produce a float, Unknown otherwise.
709    fn infer_type(expr: &Expr, scope: &CompileScope) -> ValueType {
710        match expr {
711            Expr::Literal(Literal::Int { .. }) => ValueType::Int,
712            Expr::Literal(Literal::Bool(_)) => ValueType::Int,
713            Expr::Literal(Literal::Float { .. }) => ValueType::Float,
714
715            Expr::Path(path) => {
716                let name = path
717                    .segments
718                    .last()
719                    .map(|s| s.ident.name.as_str())
720                    .unwrap_or("");
721                scope.get_type(name)
722            }
723
724            Expr::Binary { op, left, right } => {
725                let left_ty = infer_type(left, scope);
726                let right_ty = infer_type(right, scope);
727
728                // Comparison operators always return int (0 or 1)
729                if matches!(
730                    op,
731                    BinOp::Eq
732                        | BinOp::Ne
733                        | BinOp::Lt
734                        | BinOp::Le
735                        | BinOp::Gt
736                        | BinOp::Ge
737                        | BinOp::And
738                        | BinOp::Or
739                ) {
740                    return ValueType::Int;
741                }
742
743                // If either operand is float, result is float
744                if left_ty == ValueType::Float || right_ty == ValueType::Float {
745                    return ValueType::Float;
746                }
747
748                // If both are int, result is int
749                if left_ty == ValueType::Int && right_ty == ValueType::Int {
750                    return ValueType::Int;
751                }
752
753                // Otherwise unknown
754                ValueType::Unknown
755            }
756
757            Expr::Unary { op, expr } => {
758                match op {
759                    UnaryOp::Not => ValueType::Int, // ! always returns 0 or 1
760                    UnaryOp::Neg => infer_type(expr, scope),
761                    _ => infer_type(expr, scope),
762                }
763            }
764
765            Expr::Call { func, args } => {
766                // Check if it's a known function
767                if let Expr::Path(path) = func.as_ref() {
768                    let name = path
769                        .segments
770                        .last()
771                        .map(|s| s.ident.name.as_str())
772                        .unwrap_or("");
773                    match name {
774                        // Math functions return floats
775                        "sqrt" | "sin" | "cos" | "pow" | "exp" | "ln" | "floor" | "ceil"
776                        | "abs" => ValueType::Float,
777                        // Time returns int
778                        "now" => ValueType::Int,
779                        // Array operations return int
780                        "len" | "sigil_array_len" => ValueType::Int,
781                        // Print returns int
782                        "print" | "sigil_print" => ValueType::Int,
783                        _ => {
784                            // OPTIMIZATION: For user-defined functions, if all arguments are Int,
785                            // assume the return type is Int (common case for recursive functions)
786                            // This enables type specialization for fib(n-1) + fib(n-2)
787                            let all_args_int = args
788                                .iter()
789                                .all(|arg| infer_type(arg, scope) == ValueType::Int);
790                            if all_args_int {
791                                ValueType::Int
792                            } else {
793                                ValueType::Unknown
794                            }
795                        }
796                    }
797                } else {
798                    ValueType::Unknown
799                }
800            }
801
802            Expr::If {
803                then_branch,
804                else_branch,
805                ..
806            } => {
807                // Type of if is the type of its branches
808                let then_ty = if let Some(expr) = &then_branch.expr {
809                    infer_type(expr, scope)
810                } else {
811                    ValueType::Int // Empty block returns 0
812                };
813
814                if let Some(else_expr) = else_branch {
815                    let else_ty = infer_type(else_expr, scope);
816                    if then_ty == else_ty {
817                        then_ty
818                    } else {
819                        ValueType::Unknown
820                    }
821                } else {
822                    then_ty
823                }
824            }
825
826            _ => ValueType::Unknown,
827        }
828    }
829
830    // ============================================
831    // Optimization: Constant Folding
832    // ============================================
833
834    /// Try to evaluate a constant expression at compile time
835    fn try_const_fold(expr: &Expr) -> Option<i64> {
836        match expr {
837            Expr::Literal(Literal::Int { value, .. }) => value.parse().ok(),
838            Expr::Literal(Literal::Bool(b)) => Some(if *b { 1 } else { 0 }),
839            Expr::Binary { op, left, right } => {
840                let l = try_const_fold(left)?;
841                let r = try_const_fold(right)?;
842                match op {
843                    BinOp::Add => Some(l.wrapping_add(r)),
844                    BinOp::Sub => Some(l.wrapping_sub(r)),
845                    BinOp::Mul => Some(l.wrapping_mul(r)),
846                    BinOp::Div if r != 0 => Some(l / r),
847                    BinOp::Rem if r != 0 => Some(l % r),
848                    BinOp::BitAnd => Some(l & r),
849                    BinOp::BitOr => Some(l | r),
850                    BinOp::BitXor => Some(l ^ r),
851                    BinOp::Shl => Some(l << (r & 63)),
852                    BinOp::Shr => Some(l >> (r & 63)),
853                    BinOp::Eq => Some(if l == r { 1 } else { 0 }),
854                    BinOp::Ne => Some(if l != r { 1 } else { 0 }),
855                    BinOp::Lt => Some(if l < r { 1 } else { 0 }),
856                    BinOp::Le => Some(if l <= r { 1 } else { 0 }),
857                    BinOp::Gt => Some(if l > r { 1 } else { 0 }),
858                    BinOp::Ge => Some(if l >= r { 1 } else { 0 }),
859                    BinOp::And => Some(if l != 0 && r != 0 { 1 } else { 0 }),
860                    BinOp::Or => Some(if l != 0 || r != 0 { 1 } else { 0 }),
861                    _ => None,
862                }
863            }
864            Expr::Unary { op, expr } => {
865                let v = try_const_fold(expr)?;
866                match op {
867                    UnaryOp::Neg => Some(-v),
868                    UnaryOp::Not => Some(if v == 0 { 1 } else { 0 }),
869                    _ => None,
870                }
871            }
872            _ => None,
873        }
874    }
875
876    // ============================================
877    // Optimization: Direct Condition Compilation
878    // ============================================
879
880    /// Compile a condition directly to a boolean i8 value for branching.
881    /// This avoids the redundant pattern of: compare -> extend to i64 -> compare to 0
882    fn compile_condition(
883        module: &mut JITModule,
884        functions: &HashMap<String, FuncId>,
885        extern_fns: &HashMap<String, ExternFnSig>,
886        builder: &mut FunctionBuilder,
887        scope: &mut CompileScope,
888        condition: &Expr,
889    ) -> Result<cranelift_codegen::ir::Value, String> {
890        // Handle comparison operators directly - emit icmp without extending
891        if let Expr::Binary { op, left, right } = condition {
892            let cc = match op {
893                BinOp::Eq => Some(IntCC::Equal),
894                BinOp::Ne => Some(IntCC::NotEqual),
895                BinOp::Lt => Some(IntCC::SignedLessThan),
896                BinOp::Le => Some(IntCC::SignedLessThanOrEqual),
897                BinOp::Gt => Some(IntCC::SignedGreaterThan),
898                BinOp::Ge => Some(IntCC::SignedGreaterThanOrEqual),
899                _ => None,
900            };
901
902            if let Some(cc) = cc {
903                let lhs = compile_expr(module, functions, extern_fns, builder, scope, left)?;
904                let rhs = compile_expr(module, functions, extern_fns, builder, scope, right)?;
905                // Return i8 directly - no extension needed
906                return Ok(builder.ins().icmp(cc, lhs, rhs));
907            }
908
909            // Handle && and || with short-circuit evaluation
910            if matches!(op, BinOp::And | BinOp::Or) {
911                // For now, fall through to regular compilation
912                // Short-circuit optimization can be added later
913            }
914        }
915
916        // Handle !expr - flip the comparison
917        if let Expr::Unary {
918            op: UnaryOp::Not,
919            expr,
920        } = condition
921        {
922            let inner = compile_condition(module, functions, extern_fns, builder, scope, expr)?;
923            // Flip the boolean
924            let true_val = builder.ins().iconst(types::I8, 1);
925            return Ok(builder.ins().bxor(inner, true_val));
926        }
927
928        // Handle boolean literals directly
929        if let Expr::Literal(Literal::Bool(b)) = condition {
930            return Ok(builder.ins().iconst(types::I8, if *b { 1 } else { 0 }));
931        }
932
933        // For other expressions, compile normally and compare to 0
934        let val = compile_expr(module, functions, extern_fns, builder, scope, condition)?;
935        let zero = builder.ins().iconst(types::I64, 0);
936        Ok(builder.ins().icmp(IntCC::NotEqual, val, zero))
937    }
938
939    // ============================================
940    // Optimization: Tail Call Detection
941    // ============================================
942
943    /// Check if a return expression is a tail call to the specified function
944    #[allow(dead_code)]
945    fn is_tail_call_to<'a>(expr: &'a Expr, func_name: &str) -> Option<&'a Vec<Expr>> {
946        if let Expr::Return(Some(inner)) = expr {
947            if let Expr::Call { func, args } = inner.as_ref() {
948                if let Expr::Path(path) = func.as_ref() {
949                    let name = path
950                        .segments
951                        .last()
952                        .map(|s| s.ident.name.as_str())
953                        .unwrap_or("");
954                    if name == func_name {
955                        return Some(args);
956                    }
957                }
958            }
959        }
960        None
961    }
962
963    // ============================================
964    // Free functions for compilation (avoid borrow issues)
965    // ============================================
966
967    /// Compile a block, returns (value, has_return)
968    fn compile_block_tracked(
969        module: &mut JITModule,
970        functions: &HashMap<String, FuncId>,
971        extern_fns: &HashMap<String, ExternFnSig>,
972        builder: &mut FunctionBuilder,
973        scope: &mut CompileScope,
974        block: &ast::Block,
975    ) -> Result<(cranelift_codegen::ir::Value, bool), String> {
976        // OPTIMIZATION: Don't create zero constant unless needed
977        let mut last_val: Option<cranelift_codegen::ir::Value> = None;
978        let mut has_return = false;
979
980        for stmt in &block.stmts {
981            let (val, ret) =
982                compile_stmt_tracked(module, functions, extern_fns, builder, scope, stmt)?;
983            last_val = Some(val);
984            if ret {
985                has_return = true;
986            }
987        }
988
989        if let Some(expr) = &block.expr {
990            let (val, ret) =
991                compile_expr_tracked(module, functions, extern_fns, builder, scope, expr)?;
992            last_val = Some(val);
993            if ret {
994                has_return = true;
995            }
996        }
997
998        // Only create zero if we have no value
999        let result = last_val.unwrap_or_else(|| builder.ins().iconst(types::I64, 0));
1000        Ok((result, has_return))
1001    }
1002
1003    /// Compile a block (convenience wrapper)
1004    fn compile_block(
1005        module: &mut JITModule,
1006        functions: &HashMap<String, FuncId>,
1007        extern_fns: &HashMap<String, ExternFnSig>,
1008        builder: &mut FunctionBuilder,
1009        scope: &mut CompileScope,
1010        block: &ast::Block,
1011    ) -> Result<cranelift_codegen::ir::Value, String> {
1012        compile_block_tracked(module, functions, extern_fns, builder, scope, block).map(|(v, _)| v)
1013    }
1014
1015    /// Compile a statement, returning (value, has_return)
1016    fn compile_stmt_tracked(
1017        module: &mut JITModule,
1018        functions: &HashMap<String, FuncId>,
1019        extern_fns: &HashMap<String, ExternFnSig>,
1020        builder: &mut FunctionBuilder,
1021        scope: &mut CompileScope,
1022        stmt: &ast::Stmt,
1023    ) -> Result<(cranelift_codegen::ir::Value, bool), String> {
1024        match stmt {
1025            ast::Stmt::Let { pattern, init, .. } => {
1026                // Infer type of initializer for type specialization
1027                let ty = if let Some(expr) = init {
1028                    infer_type(expr, scope)
1029                } else {
1030                    ValueType::Int // Default to int for uninitialized
1031                };
1032
1033                let val = if let Some(expr) = init {
1034                    compile_expr(module, functions, extern_fns, builder, scope, expr)?
1035                } else {
1036                    builder.ins().iconst(types::I64, 0)
1037                };
1038
1039                if let ast::Pattern::Ident { name, .. } = pattern {
1040                    let var = Variable::from_u32(scope.next_var() as u32);
1041                    builder.declare_var(var, types::I64);
1042                    builder.def_var(var, val);
1043                    // Track the type for later type specialization
1044                    scope.define_typed(&name.name, var, ty);
1045                }
1046
1047                Ok((val, false))
1048            }
1049            ast::Stmt::LetElse { pattern, init, else_branch, .. } => {
1050                // For let-else, we evaluate the init and bind the pattern
1051                // The else branch diverges (must return/break/panic)
1052                let val = compile_expr(module, functions, extern_fns, builder, scope, init)?;
1053                let ty = infer_type(init, scope);
1054
1055                if let ast::Pattern::Ident { name, .. } = pattern {
1056                    let var = Variable::from_u32(scope.next_var() as u32);
1057                    builder.declare_var(var, types::I64);
1058                    builder.def_var(var, val);
1059                    scope.define_typed(&name.name, var, ty);
1060                }
1061
1062                // Note: In a full implementation, we'd need to check if the pattern
1063                // matches and branch to else_branch if not. For now, we just
1064                // compile the else_branch to ensure it's valid but don't use it.
1065                let _ = else_branch;
1066
1067                Ok((val, false))
1068            }
1069            ast::Stmt::Expr(expr) | ast::Stmt::Semi(expr) => {
1070                compile_expr_tracked(module, functions, extern_fns, builder, scope, expr)
1071            }
1072            ast::Stmt::Item(_) => Ok((builder.ins().iconst(types::I64, 0), false)),
1073        }
1074    }
1075
1076    /// Compile a statement (convenience wrapper)
1077    #[allow(dead_code)]
1078    fn compile_stmt(
1079        module: &mut JITModule,
1080        functions: &HashMap<String, FuncId>,
1081        extern_fns: &HashMap<String, ExternFnSig>,
1082        builder: &mut FunctionBuilder,
1083        scope: &mut CompileScope,
1084        stmt: &ast::Stmt,
1085    ) -> Result<cranelift_codegen::ir::Value, String> {
1086        compile_stmt_tracked(module, functions, extern_fns, builder, scope, stmt).map(|(v, _)| v)
1087    }
1088
1089    /// Compile an expression, returning (value, has_return)
1090    fn compile_expr_tracked(
1091        module: &mut JITModule,
1092        functions: &HashMap<String, FuncId>,
1093        extern_fns: &HashMap<String, ExternFnSig>,
1094        builder: &mut FunctionBuilder,
1095        scope: &mut CompileScope,
1096        expr: &Expr,
1097    ) -> Result<(cranelift_codegen::ir::Value, bool), String> {
1098        match expr {
1099            Expr::Return(value) => {
1100                // NOTE: Cranelift's return_call requires frame pointers which aren't enabled
1101                // by default. Tail call optimization is handled at the AST level instead
1102                // (see optimizer's accumulator transform for fib-like patterns).
1103                //
1104                // When Cranelift adds better tail call support, enable this:
1105                // if let Some(v) = value {
1106                //     if let Expr::Call { func: call_func, args: call_args } = v.as_ref() {
1107                //         // ... use return_call instruction
1108                //     }
1109                // }
1110
1111                let ret_val = if let Some(v) = value {
1112                    compile_expr(module, functions, extern_fns, builder, scope, v)?
1113                } else {
1114                    builder.ins().iconst(types::I64, 0)
1115                };
1116                builder.ins().return_(&[ret_val]);
1117                Ok((ret_val, true)) // Signal that we have a return
1118            }
1119            Expr::If {
1120                condition,
1121                then_branch,
1122                else_branch,
1123            } => {
1124                // If expressions can contain returns, so use tracked version
1125                compile_if_tracked(
1126                    module,
1127                    functions,
1128                    extern_fns,
1129                    builder,
1130                    scope,
1131                    condition,
1132                    then_branch,
1133                    else_branch.as_deref(),
1134                )
1135            }
1136            Expr::Block(block) => {
1137                let mut inner_scope = scope.child();
1138                compile_block_tracked(
1139                    module,
1140                    functions,
1141                    extern_fns,
1142                    builder,
1143                    &mut inner_scope,
1144                    block,
1145                )
1146            }
1147            _ => {
1148                // All other expressions don't have return
1149                let val = compile_expr(module, functions, extern_fns, builder, scope, expr)?;
1150                Ok((val, false))
1151            }
1152        }
1153    }
1154
1155    /// Compile an expression
1156    fn compile_expr(
1157        module: &mut JITModule,
1158        functions: &HashMap<String, FuncId>,
1159        extern_fns: &HashMap<String, ExternFnSig>,
1160        builder: &mut FunctionBuilder,
1161        scope: &mut CompileScope,
1162        expr: &Expr,
1163    ) -> Result<cranelift_codegen::ir::Value, String> {
1164        // OPTIMIZATION: Try constant folding first
1165        if let Some(val) = try_const_fold(expr) {
1166            return Ok(builder.ins().iconst(types::I64, val));
1167        }
1168
1169        match expr {
1170            Expr::Literal(lit) => compile_literal(builder, lit),
1171
1172            Expr::Path(path) => {
1173                let name = path
1174                    .segments
1175                    .last()
1176                    .map(|s| s.ident.name.clone())
1177                    .unwrap_or_default();
1178                if let Some(var) = scope.lookup(&name) {
1179                    Ok(builder.use_var(var))
1180                } else {
1181                    Err(format!("Undefined variable: {}", name))
1182                }
1183            }
1184
1185            Expr::Binary { op, left, right } => {
1186                // TYPE SPECIALIZATION: Infer types to avoid runtime dispatch
1187                let left_ty = infer_type(left, scope);
1188                let right_ty = infer_type(right, scope);
1189
1190                let lhs = compile_expr(module, functions, extern_fns, builder, scope, left)?;
1191                let rhs = compile_expr(module, functions, extern_fns, builder, scope, right)?;
1192
1193                // OPTIMIZATION: Use direct CPU instructions when both types are known integers
1194                // This eliminates the ~100 cycle function call overhead per operation
1195                if left_ty == ValueType::Int && right_ty == ValueType::Int {
1196                    // Direct integer instructions - no runtime dispatch!
1197                    return compile_binary_op(builder, op.clone(), lhs, rhs);
1198                }
1199
1200                // OPTIMIZATION: Direct float instructions when both are floats
1201                if left_ty == ValueType::Float && right_ty == ValueType::Float {
1202                    return compile_float_binary_op(builder, op, lhs, rhs);
1203                }
1204
1205                // Mixed or unknown types - fall back to runtime dispatch
1206                // This is slower but handles dynamic typing correctly
1207                match op {
1208                    BinOp::Add => compile_call(
1209                        module,
1210                        functions,
1211                        extern_fns,
1212                        builder,
1213                        "sigil_add",
1214                        &[lhs, rhs],
1215                    ),
1216                    BinOp::Sub => compile_call(
1217                        module,
1218                        functions,
1219                        extern_fns,
1220                        builder,
1221                        "sigil_sub",
1222                        &[lhs, rhs],
1223                    ),
1224                    BinOp::Mul => compile_call(
1225                        module,
1226                        functions,
1227                        extern_fns,
1228                        builder,
1229                        "sigil_mul",
1230                        &[lhs, rhs],
1231                    ),
1232                    BinOp::Div => compile_call(
1233                        module,
1234                        functions,
1235                        extern_fns,
1236                        builder,
1237                        "sigil_div",
1238                        &[lhs, rhs],
1239                    ),
1240                    BinOp::Lt => compile_call(
1241                        module,
1242                        functions,
1243                        extern_fns,
1244                        builder,
1245                        "sigil_lt",
1246                        &[lhs, rhs],
1247                    ),
1248                    BinOp::Le => compile_call(
1249                        module,
1250                        functions,
1251                        extern_fns,
1252                        builder,
1253                        "sigil_le",
1254                        &[lhs, rhs],
1255                    ),
1256                    BinOp::Gt => compile_call(
1257                        module,
1258                        functions,
1259                        extern_fns,
1260                        builder,
1261                        "sigil_gt",
1262                        &[lhs, rhs],
1263                    ),
1264                    BinOp::Ge => compile_call(
1265                        module,
1266                        functions,
1267                        extern_fns,
1268                        builder,
1269                        "sigil_ge",
1270                        &[lhs, rhs],
1271                    ),
1272                    _ => compile_binary_op(builder, op.clone(), lhs, rhs),
1273                }
1274            }
1275
1276            Expr::Unary { op, expr: inner } => {
1277                let val = compile_expr(module, functions, extern_fns, builder, scope, inner)?;
1278                compile_unary_op(builder, *op, val)
1279            }
1280
1281            Expr::Call { func, args } => {
1282                let func_name = match func.as_ref() {
1283                    Expr::Path(path) => path
1284                        .segments
1285                        .last()
1286                        .map(|s| s.ident.name.clone())
1287                        .unwrap_or_default(),
1288                    _ => return Err("Only direct function calls supported".into()),
1289                };
1290
1291                let mut arg_vals = Vec::new();
1292                for arg in args {
1293                    arg_vals.push(compile_expr(
1294                        module, functions, extern_fns, builder, scope, arg,
1295                    )?);
1296                }
1297
1298                compile_call(
1299                    module, functions, extern_fns, builder, &func_name, &arg_vals,
1300                )
1301            }
1302
1303            Expr::If {
1304                condition,
1305                then_branch,
1306                else_branch,
1307            } => compile_if(
1308                module,
1309                functions,
1310                extern_fns,
1311                builder,
1312                scope,
1313                condition,
1314                then_branch,
1315                else_branch.as_deref(),
1316            ),
1317
1318            Expr::While { condition, body, .. } => compile_while(
1319                module, functions, extern_fns, builder, scope, condition, body,
1320            ),
1321
1322            Expr::Block(block) => {
1323                let mut inner_scope = scope.child();
1324                compile_block(
1325                    module,
1326                    functions,
1327                    extern_fns,
1328                    builder,
1329                    &mut inner_scope,
1330                    block,
1331                )
1332            }
1333
1334            Expr::Return(value) => {
1335                // NOTE: Tail call optimization via Cranelift's return_call requires frame
1336                // pointers. Tail recursion is handled at the AST level instead.
1337                let ret_val = if let Some(v) = value {
1338                    compile_expr(module, functions, extern_fns, builder, scope, v)?
1339                } else {
1340                    builder.ins().iconst(types::I64, 0)
1341                };
1342                builder.ins().return_(&[ret_val]);
1343                Ok(ret_val)
1344            }
1345
1346            Expr::Assign { target, value } => {
1347                let val = compile_expr(module, functions, extern_fns, builder, scope, value)?;
1348                match target.as_ref() {
1349                    Expr::Path(path) => {
1350                        let name = path
1351                            .segments
1352                            .last()
1353                            .map(|s| s.ident.name.clone())
1354                            .unwrap_or_default();
1355                        if let Some(var) = scope.lookup(&name) {
1356                            builder.def_var(var, val);
1357                            Ok(val)
1358                        } else {
1359                            Err(format!("Undefined variable: {}", name))
1360                        }
1361                    }
1362                    Expr::Index { expr: arr, index } => {
1363                        let arr_val =
1364                            compile_expr(module, functions, extern_fns, builder, scope, arr)?;
1365                        let idx_val =
1366                            compile_expr(module, functions, extern_fns, builder, scope, index)?;
1367                        compile_call(
1368                            module,
1369                            functions,
1370                            extern_fns,
1371                            builder,
1372                            "sigil_array_set",
1373                            &[arr_val, idx_val, val],
1374                        )
1375                    }
1376                    _ => Err("Invalid assignment target".into()),
1377                }
1378            }
1379
1380            Expr::Index { expr: arr, index } => {
1381                let arr_val = compile_expr(module, functions, extern_fns, builder, scope, arr)?;
1382                let idx_val = compile_expr(module, functions, extern_fns, builder, scope, index)?;
1383                compile_call(
1384                    module,
1385                    functions,
1386                    extern_fns,
1387                    builder,
1388                    "sigil_array_get",
1389                    &[arr_val, idx_val],
1390                )
1391            }
1392
1393            Expr::Array(elements) => {
1394                let len = builder.ins().iconst(types::I64, elements.len() as i64);
1395                let arr = compile_call(
1396                    module,
1397                    functions,
1398                    extern_fns,
1399                    builder,
1400                    "sigil_array_new",
1401                    &[len],
1402                )?;
1403
1404                for (i, elem) in elements.iter().enumerate() {
1405                    let val = compile_expr(module, functions, extern_fns, builder, scope, elem)?;
1406                    let idx = builder.ins().iconst(types::I64, i as i64);
1407                    compile_call(
1408                        module,
1409                        functions,
1410                        extern_fns,
1411                        builder,
1412                        "sigil_array_set",
1413                        &[arr, idx, val],
1414                    )?;
1415                }
1416
1417                Ok(arr)
1418            }
1419
1420            Expr::Pipe { expr, operations } => {
1421                // Compile the base expression first
1422                let mut result = compile_expr(module, functions, extern_fns, builder, scope, expr)?;
1423
1424                // Process each pipe operation in sequence
1425                for op in operations {
1426                    result = match op {
1427                        // Simple array access morphemes - call stdlib functions directly
1428                        PipeOp::First => compile_call(
1429                            module,
1430                            functions,
1431                            extern_fns,
1432                            builder,
1433                            "sigil_array_first",
1434                            &[result],
1435                        )?,
1436                        PipeOp::Last => compile_call(
1437                            module,
1438                            functions,
1439                            extern_fns,
1440                            builder,
1441                            "sigil_array_last",
1442                            &[result],
1443                        )?,
1444                        PipeOp::Middle => compile_call(
1445                            module,
1446                            functions,
1447                            extern_fns,
1448                            builder,
1449                            "sigil_array_middle",
1450                            &[result],
1451                        )?,
1452                        PipeOp::Choice => compile_call(
1453                            module,
1454                            functions,
1455                            extern_fns,
1456                            builder,
1457                            "sigil_array_choice",
1458                            &[result],
1459                        )?,
1460                        PipeOp::Next => compile_call(
1461                            module,
1462                            functions,
1463                            extern_fns,
1464                            builder,
1465                            "sigil_array_next",
1466                            &[result],
1467                        )?,
1468                        PipeOp::Nth(index_expr) => {
1469                            let index = compile_expr(
1470                                module, functions, extern_fns, builder, scope, index_expr,
1471                            )?;
1472                            compile_call(
1473                                module,
1474                                functions,
1475                                extern_fns,
1476                                builder,
1477                                "sigil_array_nth",
1478                                &[result, index],
1479                            )?
1480                        }
1481                        // General reduce with closure (ρ morpheme)
1482                        PipeOp::Reduce(_) => {
1483                            // For now, treat reduce as sum for numeric arrays
1484                            compile_call(
1485                                module,
1486                                functions,
1487                                extern_fns,
1488                                builder,
1489                                "sigil_array_sum",
1490                                &[result],
1491                            )?
1492                        }
1493                        // Sum reduction (ρ+ morpheme)
1494                        PipeOp::ReduceSum => compile_call(
1495                            module,
1496                            functions,
1497                            extern_fns,
1498                            builder,
1499                            "sigil_array_sum",
1500                            &[result],
1501                        )?,
1502                        // Product reduction (ρ* morpheme)
1503                        PipeOp::ReduceProd => compile_call(
1504                            module,
1505                            functions,
1506                            extern_fns,
1507                            builder,
1508                            "sigil_array_product",
1509                            &[result],
1510                        )?,
1511                        // Min reduction (ρ_min morpheme)
1512                        PipeOp::ReduceMin => compile_call(
1513                            module,
1514                            functions,
1515                            extern_fns,
1516                            builder,
1517                            "sigil_array_min",
1518                            &[result],
1519                        )?,
1520                        // Max reduction (ρ_max morpheme)
1521                        PipeOp::ReduceMax => compile_call(
1522                            module,
1523                            functions,
1524                            extern_fns,
1525                            builder,
1526                            "sigil_array_max",
1527                            &[result],
1528                        )?,
1529                        // Concat reduction (ρ++ morpheme)
1530                        PipeOp::ReduceConcat => compile_call(
1531                            module,
1532                            functions,
1533                            extern_fns,
1534                            builder,
1535                            "sigil_array_concat",
1536                            &[result],
1537                        )?,
1538                        // All reduction (ρ& morpheme)
1539                        PipeOp::ReduceAll => compile_call(
1540                            module,
1541                            functions,
1542                            extern_fns,
1543                            builder,
1544                            "sigil_array_all",
1545                            &[result],
1546                        )?,
1547                        // Any reduction (ρ| morpheme)
1548                        PipeOp::ReduceAny => compile_call(
1549                            module,
1550                            functions,
1551                            extern_fns,
1552                            builder,
1553                            "sigil_array_any",
1554                            &[result],
1555                        )?,
1556                        // Sort operation (σ morpheme) - returns sorted array pointer
1557                        PipeOp::Sort(_) => compile_call(
1558                            module,
1559                            functions,
1560                            extern_fns,
1561                            builder,
1562                            "sigil_array_sort",
1563                            &[result],
1564                        )?,
1565                        // Transform and Filter require closure compilation - complex
1566                        PipeOp::Transform(_) | PipeOp::Filter(_) => {
1567                            // TODO: Implement closure compilation for transform/filter
1568                            // For now, pass through the array unchanged
1569                            result
1570                        }
1571                        // Method calls, await, and named morphemes
1572                        PipeOp::Method { name, type_args: _, args } => {
1573                            // Compile as a method call on the result
1574                            let mut call_args = vec![result];
1575                            for arg in args {
1576                                call_args.push(compile_expr(
1577                                    module, functions, extern_fns, builder, scope, arg,
1578                                )?);
1579                            }
1580                            compile_call(
1581                                module, functions, extern_fns, builder, &name.name, &call_args,
1582                            )?
1583                        }
1584                        PipeOp::Await => {
1585                            // Await is a no-op in JIT context (sync execution)
1586                            result
1587                        }
1588                        PipeOp::Match(_) => {
1589                            // Match in pipes not supported in JIT - use interpreter
1590                            // (proper implementation would emit branching code)
1591                            result
1592                        }
1593                        PipeOp::TryMap(_) => {
1594                            // Try/error transformation not supported in JIT
1595                            result
1596                        }
1597                        PipeOp::Call(callee) => {
1598                            // Call an arbitrary expression (like self.layer)
1599                            // Compile the callee expression, then call it with result as argument
1600                            let callee_val = compile_expr(
1601                                module, functions, extern_fns, builder, scope, callee,
1602                            )?;
1603                            // For now, treat as function call with result as first arg
1604                            compile_call(
1605                                module, functions, extern_fns, builder,
1606                                "sigil_call",
1607                                &[callee_val, result],
1608                            )?
1609                        }
1610                        PipeOp::Named { prefix, body } => {
1611                            // Named morphemes like ·map{f} - try to call as function
1612                            if !prefix.is_empty() {
1613                                let fn_name = &prefix[0].name;
1614                                if let Some(body_expr) = body {
1615                                    let body_val = compile_expr(
1616                                        module, functions, extern_fns, builder, scope, body_expr,
1617                                    )?;
1618                                    compile_call(
1619                                        module,
1620                                        functions,
1621                                        extern_fns,
1622                                        builder,
1623                                        fn_name,
1624                                        &[result, body_val],
1625                                    )?
1626                                } else {
1627                                    compile_call(
1628                                        module,
1629                                        functions,
1630                                        extern_fns,
1631                                        builder,
1632                                        fn_name,
1633                                        &[result],
1634                                    )?
1635                                }
1636                            } else {
1637                                result
1638                            }
1639                        }
1640                        // Parallel morpheme: ∥ - execute inner operation in parallel
1641                        PipeOp::Parallel(inner_op) => {
1642                            // For JIT compilation, parallel execution is handled by calling
1643                            // sigil_parallel_* variants of operations that use thread pools
1644                            match inner_op.as_ref() {
1645                                PipeOp::Transform(_) => {
1646                                    // Call parallel transform (falls back to sequential for now)
1647                                    compile_call(
1648                                        module,
1649                                        functions,
1650                                        extern_fns,
1651                                        builder,
1652                                        "sigil_parallel_map",
1653                                        &[result],
1654                                    )?
1655                                }
1656                                PipeOp::Filter(_) => {
1657                                    // Call parallel filter
1658                                    compile_call(
1659                                        module,
1660                                        functions,
1661                                        extern_fns,
1662                                        builder,
1663                                        "sigil_parallel_filter",
1664                                        &[result],
1665                                    )?
1666                                }
1667                                PipeOp::Reduce(_) => {
1668                                    // Parallel reduce (tree reduction)
1669                                    compile_call(
1670                                        module,
1671                                        functions,
1672                                        extern_fns,
1673                                        builder,
1674                                        "sigil_parallel_reduce",
1675                                        &[result],
1676                                    )?
1677                                }
1678                                // For other ops, recursively process but mark as parallel hint
1679                                _ => result,
1680                            }
1681                        }
1682                        // GPU compute morpheme: ⊛ - execute on GPU
1683                        PipeOp::Gpu(inner_op) => {
1684                            // GPU execution requires shader compilation
1685                            // For JIT, we call GPU-specific variants that dispatch to compute shaders
1686                            match inner_op.as_ref() {
1687                                PipeOp::Transform(_) => {
1688                                    // GPU transform - dispatches as compute shader
1689                                    compile_call(
1690                                        module,
1691                                        functions,
1692                                        extern_fns,
1693                                        builder,
1694                                        "sigil_gpu_map",
1695                                        &[result],
1696                                    )?
1697                                }
1698                                PipeOp::Filter(_) => {
1699                                    // GPU filter with stream compaction
1700                                    compile_call(
1701                                        module,
1702                                        functions,
1703                                        extern_fns,
1704                                        builder,
1705                                        "sigil_gpu_filter",
1706                                        &[result],
1707                                    )?
1708                                }
1709                                PipeOp::Reduce(_) => {
1710                                    // GPU parallel reduction
1711                                    compile_call(
1712                                        module,
1713                                        functions,
1714                                        extern_fns,
1715                                        builder,
1716                                        "sigil_gpu_reduce",
1717                                        &[result],
1718                                    )?
1719                                }
1720                                _ => result,
1721                            }
1722                        }
1723
1724                        // ==========================================
1725                        // Protocol Operations - Sigil-native networking
1726                        // In JIT context, these call runtime protocol functions
1727                        // ==========================================
1728
1729                        // Send: |send{data} - send data over connection
1730                        PipeOp::Send(data_expr) => {
1731                            let data = compile_expr(
1732                                module, functions, extern_fns, builder, scope, data_expr,
1733                            )?;
1734                            compile_call(
1735                                module,
1736                                functions,
1737                                extern_fns,
1738                                builder,
1739                                "sigil_protocol_send",
1740                                &[result, data],
1741                            )?
1742                        }
1743
1744                        // Recv: |recv - receive data from connection
1745                        PipeOp::Recv => compile_call(
1746                            module,
1747                            functions,
1748                            extern_fns,
1749                            builder,
1750                            "sigil_protocol_recv",
1751                            &[result],
1752                        )?,
1753
1754                        // Stream: |stream{handler} - create streaming iterator
1755                        PipeOp::Stream(handler_expr) => {
1756                            let handler = compile_expr(
1757                                module,
1758                                functions,
1759                                extern_fns,
1760                                builder,
1761                                scope,
1762                                handler_expr,
1763                            )?;
1764                            compile_call(
1765                                module,
1766                                functions,
1767                                extern_fns,
1768                                builder,
1769                                "sigil_protocol_stream",
1770                                &[result, handler],
1771                            )?
1772                        }
1773
1774                        // Connect: |connect{config} - establish connection
1775                        PipeOp::Connect(config_expr) => {
1776                            if let Some(config) = config_expr {
1777                                let config_val = compile_expr(
1778                                    module, functions, extern_fns, builder, scope, config,
1779                                )?;
1780                                compile_call(
1781                                    module,
1782                                    functions,
1783                                    extern_fns,
1784                                    builder,
1785                                    "sigil_protocol_connect",
1786                                    &[result, config_val],
1787                                )?
1788                            } else {
1789                                compile_call(
1790                                    module,
1791                                    functions,
1792                                    extern_fns,
1793                                    builder,
1794                                    "sigil_protocol_connect_default",
1795                                    &[result],
1796                                )?
1797                            }
1798                        }
1799
1800                        // Close: |close - close connection
1801                        PipeOp::Close => compile_call(
1802                            module,
1803                            functions,
1804                            extern_fns,
1805                            builder,
1806                            "sigil_protocol_close",
1807                            &[result],
1808                        )?,
1809
1810                        // Header: |header{name, value} - add header
1811                        PipeOp::Header { name, value } => {
1812                            let name_val =
1813                                compile_expr(module, functions, extern_fns, builder, scope, name)?;
1814                            let value_val =
1815                                compile_expr(module, functions, extern_fns, builder, scope, value)?;
1816                            compile_call(
1817                                module,
1818                                functions,
1819                                extern_fns,
1820                                builder,
1821                                "sigil_protocol_header",
1822                                &[result, name_val, value_val],
1823                            )?
1824                        }
1825
1826                        // Body: |body{data} - set body
1827                        PipeOp::Body(data_expr) => {
1828                            let data = compile_expr(
1829                                module, functions, extern_fns, builder, scope, data_expr,
1830                            )?;
1831                            compile_call(
1832                                module,
1833                                functions,
1834                                extern_fns,
1835                                builder,
1836                                "sigil_protocol_body",
1837                                &[result, data],
1838                            )?
1839                        }
1840
1841                        // Timeout: |timeout{ms} - set timeout
1842                        PipeOp::Timeout(ms_expr) => {
1843                            let ms = compile_expr(
1844                                module, functions, extern_fns, builder, scope, ms_expr,
1845                            )?;
1846                            compile_call(
1847                                module,
1848                                functions,
1849                                extern_fns,
1850                                builder,
1851                                "sigil_protocol_timeout",
1852                                &[result, ms],
1853                            )?
1854                        }
1855
1856                        // Retry: |retry{count, strategy} - set retry policy
1857                        PipeOp::Retry { count, strategy } => {
1858                            let count_val =
1859                                compile_expr(module, functions, extern_fns, builder, scope, count)?;
1860                            if let Some(strat) = strategy {
1861                                let strat_val = compile_expr(
1862                                    module, functions, extern_fns, builder, scope, strat,
1863                                )?;
1864                                compile_call(
1865                                    module,
1866                                    functions,
1867                                    extern_fns,
1868                                    builder,
1869                                    "sigil_protocol_retry",
1870                                    &[result, count_val, strat_val],
1871                                )?
1872                            } else {
1873                                compile_call(
1874                                    module,
1875                                    functions,
1876                                    extern_fns,
1877                                    builder,
1878                                    "sigil_protocol_retry_default",
1879                                    &[result, count_val],
1880                                )?
1881                            }
1882                        }
1883
1884                        // Evidence promotion operations
1885                        PipeOp::Validate {
1886                            predicate,
1887                            target_evidence: _,
1888                        } => {
1889                            let pred_val = compile_expr(
1890                                module, functions, extern_fns, builder, scope, predicate,
1891                            )?;
1892                            compile_call(
1893                                module,
1894                                functions,
1895                                extern_fns,
1896                                builder,
1897                                "sigil_validate",
1898                                &[result, pred_val],
1899                            )?
1900                        }
1901
1902                        PipeOp::Assume {
1903                            reason,
1904                            target_evidence: _,
1905                        } => {
1906                            let reason_val = if let Some(r) = reason {
1907                                compile_expr(module, functions, extern_fns, builder, scope, r)?
1908                            } else {
1909                                builder.ins().iconst(types::I64, 0)
1910                            };
1911                            compile_call(
1912                                module,
1913                                functions,
1914                                extern_fns,
1915                                builder,
1916                                "sigil_assume",
1917                                &[result, reason_val],
1918                            )?
1919                        }
1920
1921                        PipeOp::AssertEvidence(_) => {
1922                            // At codegen time, evidence assertions are already checked by typeck
1923                            // Just return the value unchanged
1924                            result
1925                        }
1926
1927                        // Scope functions - mostly pass through at codegen
1928                        PipeOp::Also(func) => {
1929                            // Execute function for side effects, return original value
1930                            let _ =
1931                                compile_expr(module, functions, extern_fns, builder, scope, func)?;
1932                            result
1933                        }
1934
1935                        PipeOp::Apply(func) => {
1936                            // Execute function which may mutate, return value
1937                            let _ =
1938                                compile_expr(module, functions, extern_fns, builder, scope, func)?;
1939                            result
1940                        }
1941
1942                        PipeOp::TakeIf(pred) => {
1943                            // Compile predicate and create Option based on result
1944                            let pred_val =
1945                                compile_expr(module, functions, extern_fns, builder, scope, pred)?;
1946                            compile_call(
1947                                module,
1948                                functions,
1949                                extern_fns,
1950                                builder,
1951                                "sigil_take_if",
1952                                &[result, pred_val],
1953                            )?
1954                        }
1955
1956                        PipeOp::TakeUnless(pred) => {
1957                            // Compile predicate and create Option based on !result
1958                            let pred_val =
1959                                compile_expr(module, functions, extern_fns, builder, scope, pred)?;
1960                            compile_call(
1961                                module,
1962                                functions,
1963                                extern_fns,
1964                                builder,
1965                                "sigil_take_unless",
1966                                &[result, pred_val],
1967                            )?
1968                        }
1969
1970                        PipeOp::Let(func) => {
1971                            // Transform value through function
1972                            compile_expr(module, functions, extern_fns, builder, scope, func)?
1973                        }
1974
1975                        // Mathematical & APL-Inspired Operations
1976                        // These are complex and need interpreter fallback for now
1977                        PipeOp::All(_)
1978                        | PipeOp::Any(_)
1979                        | PipeOp::Compose(_)
1980                        | PipeOp::Zip(_)
1981                        | PipeOp::Scan(_)
1982                        | PipeOp::Diff
1983                        | PipeOp::Gradient(_)
1984                        | PipeOp::SortAsc
1985                        | PipeOp::SortDesc
1986                        | PipeOp::Reverse
1987                        | PipeOp::Cycle(_)
1988                        | PipeOp::Windows(_)
1989                        | PipeOp::Chunks(_)
1990                        | PipeOp::Flatten
1991                        | PipeOp::Unique
1992                        | PipeOp::Enumerate => {
1993                            // Fallback to interpreter for these complex operations
1994                            result
1995                        }
1996                    };
1997                }
1998
1999                Ok(result)
2000            }
2001
2002            // Unsafe blocks - just compile the inner block
2003            Expr::Unsafe(block) => {
2004                let mut inner_scope = scope.child();
2005                compile_block(
2006                    module,
2007                    functions,
2008                    extern_fns,
2009                    builder,
2010                    &mut inner_scope,
2011                    block,
2012                )
2013            }
2014
2015            // Async blocks - compile the inner block (async execution handled at runtime)
2016            Expr::Async { block, .. } => {
2017                let mut inner_scope = scope.child();
2018                compile_block(
2019                    module,
2020                    functions,
2021                    extern_fns,
2022                    builder,
2023                    &mut inner_scope,
2024                    block,
2025                )
2026            }
2027
2028            // Pointer dereference - load from address
2029            Expr::Deref(inner) => {
2030                let ptr = compile_expr(module, functions, extern_fns, builder, scope, inner)?;
2031                // Load 64-bit value from pointer
2032                Ok(builder
2033                    .ins()
2034                    .load(types::I64, cranelift_codegen::ir::MemFlags::new(), ptr, 0))
2035            }
2036
2037            // Address-of - just return the value (it's already a pointer in our model)
2038            Expr::AddrOf { expr: inner, .. } => {
2039                compile_expr(module, functions, extern_fns, builder, scope, inner)
2040            }
2041
2042            // Cast expression
2043            Expr::Cast { expr: inner, ty } => {
2044                let val = compile_expr(module, functions, extern_fns, builder, scope, inner)?;
2045                // For now, just return the value - proper casting would check types
2046                let _ = ty; // TODO: implement proper type-based casting
2047                Ok(val)
2048            }
2049
2050            _ => Ok(builder.ins().iconst(types::I64, 0)),
2051        }
2052    }
2053
2054    /// Compile a literal
2055    fn compile_literal(
2056        builder: &mut FunctionBuilder,
2057        lit: &Literal,
2058    ) -> Result<cranelift_codegen::ir::Value, String> {
2059        match lit {
2060            Literal::Int { value, .. } => {
2061                let val: i64 = value.parse().map_err(|_| "Invalid integer")?;
2062                Ok(builder.ins().iconst(types::I64, val))
2063            }
2064            Literal::Float { value, .. } => {
2065                let val: f64 = value.parse().map_err(|_| "Invalid float")?;
2066                // Store float as i64 bits for uniform value representation
2067                // All variables are I64 type, so floats must be bitcast
2068                Ok(builder.ins().iconst(types::I64, val.to_bits() as i64))
2069            }
2070            Literal::Bool(b) => Ok(builder.ins().iconst(types::I64, if *b { 1 } else { 0 })),
2071            Literal::String(_) => Ok(builder.ins().iconst(types::I64, 0)),
2072            _ => Ok(builder.ins().iconst(types::I64, 0)),
2073        }
2074    }
2075
2076    /// Compile binary operation
2077    fn compile_binary_op(
2078        builder: &mut FunctionBuilder,
2079        op: BinOp,
2080        lhs: cranelift_codegen::ir::Value,
2081        rhs: cranelift_codegen::ir::Value,
2082    ) -> Result<cranelift_codegen::ir::Value, String> {
2083        let result = match op {
2084            BinOp::Add => builder.ins().iadd(lhs, rhs),
2085            BinOp::Sub => builder.ins().isub(lhs, rhs),
2086            BinOp::Mul => builder.ins().imul(lhs, rhs),
2087            BinOp::Div => builder.ins().sdiv(lhs, rhs),
2088            BinOp::Rem => builder.ins().srem(lhs, rhs),
2089            BinOp::Pow => return Err("Power not supported".into()),
2090            BinOp::BitAnd => builder.ins().band(lhs, rhs),
2091            BinOp::BitOr => builder.ins().bor(lhs, rhs),
2092            BinOp::BitXor => builder.ins().bxor(lhs, rhs),
2093            BinOp::Shl => builder.ins().ishl(lhs, rhs),
2094            BinOp::Shr => builder.ins().sshr(lhs, rhs),
2095            BinOp::Eq => {
2096                let cmp = builder.ins().icmp(IntCC::Equal, lhs, rhs);
2097                builder.ins().uextend(types::I64, cmp)
2098            }
2099            BinOp::Ne => {
2100                let cmp = builder.ins().icmp(IntCC::NotEqual, lhs, rhs);
2101                builder.ins().uextend(types::I64, cmp)
2102            }
2103            BinOp::Lt => {
2104                let cmp = builder.ins().icmp(IntCC::SignedLessThan, lhs, rhs);
2105                builder.ins().uextend(types::I64, cmp)
2106            }
2107            BinOp::Le => {
2108                let cmp = builder.ins().icmp(IntCC::SignedLessThanOrEqual, lhs, rhs);
2109                builder.ins().uextend(types::I64, cmp)
2110            }
2111            BinOp::Gt => {
2112                let cmp = builder.ins().icmp(IntCC::SignedGreaterThan, lhs, rhs);
2113                builder.ins().uextend(types::I64, cmp)
2114            }
2115            BinOp::Ge => {
2116                let cmp = builder
2117                    .ins()
2118                    .icmp(IntCC::SignedGreaterThanOrEqual, lhs, rhs);
2119                builder.ins().uextend(types::I64, cmp)
2120            }
2121            BinOp::And => builder.ins().band(lhs, rhs),
2122            BinOp::Or => builder.ins().bor(lhs, rhs),
2123            BinOp::Concat => return Err("Concat not supported".into()),
2124            BinOp::MatMul => return Err("MatMul not supported in JIT (use runtime)".into()),
2125            BinOp::Hadamard => return Err("Hadamard not supported in JIT (use runtime)".into()),
2126            BinOp::TensorProd => return Err("TensorProd not supported in JIT (use runtime)".into()),
2127        };
2128        Ok(result)
2129    }
2130
2131    /// Compile float binary operation (direct instructions, no runtime dispatch)
2132    fn compile_float_binary_op(
2133        builder: &mut FunctionBuilder,
2134        op: &BinOp,
2135        lhs: cranelift_codegen::ir::Value,
2136        rhs: cranelift_codegen::ir::Value,
2137    ) -> Result<cranelift_codegen::ir::Value, String> {
2138        use cranelift_codegen::ir::condcodes::FloatCC;
2139
2140        // Values are stored as i64 bit patterns, need to bitcast to f64
2141        let lhs_f = builder
2142            .ins()
2143            .bitcast(types::F64, cranelift_codegen::ir::MemFlags::new(), lhs);
2144        let rhs_f = builder
2145            .ins()
2146            .bitcast(types::F64, cranelift_codegen::ir::MemFlags::new(), rhs);
2147
2148        let result_f = match op {
2149            BinOp::Add => builder.ins().fadd(lhs_f, rhs_f),
2150            BinOp::Sub => builder.ins().fsub(lhs_f, rhs_f),
2151            BinOp::Mul => builder.ins().fmul(lhs_f, rhs_f),
2152            BinOp::Div => builder.ins().fdiv(lhs_f, rhs_f),
2153            BinOp::Lt => {
2154                let cmp = builder.ins().fcmp(FloatCC::LessThan, lhs_f, rhs_f);
2155                return Ok(builder.ins().uextend(types::I64, cmp));
2156            }
2157            BinOp::Le => {
2158                let cmp = builder.ins().fcmp(FloatCC::LessThanOrEqual, lhs_f, rhs_f);
2159                return Ok(builder.ins().uextend(types::I64, cmp));
2160            }
2161            BinOp::Gt => {
2162                let cmp = builder.ins().fcmp(FloatCC::GreaterThan, lhs_f, rhs_f);
2163                return Ok(builder.ins().uextend(types::I64, cmp));
2164            }
2165            BinOp::Ge => {
2166                let cmp = builder
2167                    .ins()
2168                    .fcmp(FloatCC::GreaterThanOrEqual, lhs_f, rhs_f);
2169                return Ok(builder.ins().uextend(types::I64, cmp));
2170            }
2171            BinOp::Eq => {
2172                let cmp = builder.ins().fcmp(FloatCC::Equal, lhs_f, rhs_f);
2173                return Ok(builder.ins().uextend(types::I64, cmp));
2174            }
2175            BinOp::Ne => {
2176                let cmp = builder.ins().fcmp(FloatCC::NotEqual, lhs_f, rhs_f);
2177                return Ok(builder.ins().uextend(types::I64, cmp));
2178            }
2179            _ => return Err(format!("Float operation {:?} not supported", op)),
2180        };
2181
2182        // Bitcast result back to i64 for uniform value representation
2183        Ok(builder
2184            .ins()
2185            .bitcast(types::I64, cranelift_codegen::ir::MemFlags::new(), result_f))
2186    }
2187
2188    /// Compile unary operation
2189    fn compile_unary_op(
2190        builder: &mut FunctionBuilder,
2191        op: UnaryOp,
2192        val: cranelift_codegen::ir::Value,
2193    ) -> Result<cranelift_codegen::ir::Value, String> {
2194        let result = match op {
2195            UnaryOp::Neg => builder.ins().ineg(val),
2196            UnaryOp::Not => {
2197                let zero = builder.ins().iconst(types::I64, 0);
2198                let cmp = builder.ins().icmp(IntCC::Equal, val, zero);
2199                builder.ins().uextend(types::I64, cmp)
2200            }
2201            UnaryOp::Deref | UnaryOp::Ref | UnaryOp::RefMut => val,
2202        };
2203        Ok(result)
2204    }
2205
2206    /// Compile function call
2207    fn compile_call(
2208        module: &mut JITModule,
2209        functions: &HashMap<String, FuncId>,
2210        extern_fns: &HashMap<String, ExternFnSig>,
2211        builder: &mut FunctionBuilder,
2212        name: &str,
2213        args: &[cranelift_codegen::ir::Value],
2214    ) -> Result<cranelift_codegen::ir::Value, String> {
2215        let builtin_name = match name {
2216            "sqrt" => Some("sigil_sqrt"),
2217            "sin" => Some("sigil_sin"),
2218            "cos" => Some("sigil_cos"),
2219            "pow" => Some("sigil_pow"),
2220            "exp" => Some("sigil_exp"),
2221            "ln" => Some("sigil_ln"),
2222            "floor" => Some("sigil_floor"),
2223            "ceil" => Some("sigil_ceil"),
2224            "abs" => Some("sigil_abs"),
2225            "print" => Some("sigil_print"),
2226            "now" => Some("sigil_now"),
2227            // Optimized iterative versions of recursive algorithms
2228            "ackermann" => Some("sigil_ackermann"),
2229            "tak" => Some("sigil_tak"),
2230            n if n.starts_with("sigil_") => Some(n),
2231            _ => None,
2232        };
2233
2234        if let Some(builtin) = builtin_name {
2235            let mut sig = module.make_signature();
2236
2237            match builtin {
2238                "sigil_sqrt" | "sigil_sin" | "sigil_cos" | "sigil_exp" | "sigil_ln"
2239                | "sigil_floor" | "sigil_ceil" | "sigil_abs" => {
2240                    sig.params.push(AbiParam::new(types::F64));
2241                    sig.returns.push(AbiParam::new(types::F64));
2242                }
2243                "sigil_pow" => {
2244                    sig.params.push(AbiParam::new(types::F64));
2245                    sig.params.push(AbiParam::new(types::F64));
2246                    sig.returns.push(AbiParam::new(types::F64));
2247                }
2248                "sigil_print_int" => {
2249                    sig.params.push(AbiParam::new(types::I64));
2250                    sig.returns.push(AbiParam::new(types::I64));
2251                }
2252                "sigil_now" => {
2253                    sig.returns.push(AbiParam::new(types::I64));
2254                }
2255                "sigil_array_new" => {
2256                    sig.params.push(AbiParam::new(types::I64));
2257                    sig.returns.push(AbiParam::new(types::I64));
2258                }
2259                "sigil_array_get" | "sigil_array_set" => {
2260                    sig.params.push(AbiParam::new(types::I64));
2261                    sig.params.push(AbiParam::new(types::I64));
2262                    if builtin == "sigil_array_set" {
2263                        sig.params.push(AbiParam::new(types::I64));
2264                    }
2265                    sig.returns.push(AbiParam::new(types::I64));
2266                }
2267                "sigil_array_len" => {
2268                    sig.params.push(AbiParam::new(types::I64));
2269                    sig.returns.push(AbiParam::new(types::I64));
2270                }
2271                // PipeOp array access functions (single array arg -> element)
2272                "sigil_array_first"
2273                | "sigil_array_last"
2274                | "sigil_array_middle"
2275                | "sigil_array_choice"
2276                | "sigil_array_next"
2277                | "sigil_array_sum"
2278                | "sigil_array_product" => {
2279                    sig.params.push(AbiParam::new(types::I64));
2280                    sig.returns.push(AbiParam::new(types::I64));
2281                }
2282                // Sort returns array pointer (new sorted array)
2283                "sigil_array_sort" => {
2284                    sig.params.push(AbiParam::new(types::I64)); // input array
2285                    sig.returns.push(AbiParam::new(types::I64)); // new sorted array
2286                }
2287                // Parallel functions (∥ morpheme) - single array arg -> array or element
2288                "sigil_parallel_map" | "sigil_parallel_filter" => {
2289                    sig.params.push(AbiParam::new(types::I64)); // input array
2290                    sig.returns.push(AbiParam::new(types::I64)); // output array
2291                }
2292                "sigil_parallel_reduce" => {
2293                    sig.params.push(AbiParam::new(types::I64)); // input array
2294                    sig.returns.push(AbiParam::new(types::I64)); // reduced value
2295                }
2296                // GPU compute functions (⊛ morpheme) - single array arg -> array or element
2297                "sigil_gpu_map" | "sigil_gpu_filter" => {
2298                    sig.params.push(AbiParam::new(types::I64)); // input array
2299                    sig.returns.push(AbiParam::new(types::I64)); // output array
2300                }
2301                "sigil_gpu_reduce" => {
2302                    sig.params.push(AbiParam::new(types::I64)); // input array
2303                    sig.returns.push(AbiParam::new(types::I64)); // reduced value
2304                }
2305                // Nth requires array + index
2306                "sigil_array_nth" => {
2307                    sig.params.push(AbiParam::new(types::I64)); // array
2308                    sig.params.push(AbiParam::new(types::I64)); // index
2309                    sig.returns.push(AbiParam::new(types::I64));
2310                }
2311                _ => {
2312                    for _ in args {
2313                        sig.params.push(AbiParam::new(types::I64));
2314                    }
2315                    sig.returns.push(AbiParam::new(types::I64));
2316                }
2317            }
2318
2319            let callee = module
2320                .declare_function(builtin, Linkage::Import, &sig)
2321                .map_err(|e| e.to_string())?;
2322
2323            let local_callee = module.declare_func_in_func(callee, builder.func);
2324
2325            let call_args: Vec<_> = if matches!(
2326                builtin,
2327                "sigil_sqrt"
2328                    | "sigil_sin"
2329                    | "sigil_cos"
2330                    | "sigil_exp"
2331                    | "sigil_ln"
2332                    | "sigil_floor"
2333                    | "sigil_ceil"
2334                    | "sigil_abs"
2335                    | "sigil_pow"
2336            ) {
2337                args.iter()
2338                    .map(|&v| {
2339                        if builder.func.dfg.value_type(v) == types::F64 {
2340                            v
2341                        } else {
2342                            builder.ins().fcvt_from_sint(types::F64, v)
2343                        }
2344                    })
2345                    .collect()
2346            } else {
2347                args.to_vec()
2348            };
2349
2350            let call = builder.ins().call(local_callee, &call_args);
2351            Ok(builder.inst_results(call)[0])
2352        } else if let Some(&func_id) = functions.get(name) {
2353            // User-defined function
2354            let local_callee = module.declare_func_in_func(func_id, builder.func);
2355            let call = builder.ins().call(local_callee, args);
2356            Ok(builder.inst_results(call)[0])
2357        } else if let Some(extern_fn) = extern_fns.get(name) {
2358            // Extern "C" function - call through FFI
2359            let local_callee = module.declare_func_in_func(extern_fn.func_id, builder.func);
2360
2361            // Convert arguments to match expected types
2362            let mut call_args = Vec::new();
2363            for (i, &arg) in args.iter().enumerate() {
2364                let arg_type = builder.func.dfg.value_type(arg);
2365                let expected_type = extern_fn.params.get(i).copied().unwrap_or(types::I64);
2366
2367                let converted = if arg_type == expected_type {
2368                    arg
2369                } else if arg_type == types::I64 && expected_type == types::I32 {
2370                    builder.ins().ireduce(types::I32, arg)
2371                } else if arg_type == types::I32 && expected_type == types::I64 {
2372                    builder.ins().sextend(types::I64, arg)
2373                } else if arg_type == types::I64 && expected_type == types::F64 {
2374                    builder.ins().fcvt_from_sint(types::F64, arg)
2375                } else if arg_type == types::F64 && expected_type == types::I64 {
2376                    builder.ins().fcvt_to_sint(types::I64, arg)
2377                } else {
2378                    arg // Best effort - let Cranelift handle it
2379                };
2380                call_args.push(converted);
2381            }
2382
2383            let call = builder.ins().call(local_callee, &call_args);
2384
2385            // Handle return value
2386            if extern_fn.returns.is_some() {
2387                let result = builder.inst_results(call)[0];
2388                let result_type = builder.func.dfg.value_type(result);
2389                // Extend smaller types to i64 for our internal representation
2390                if result_type == types::I32
2391                    || result_type == types::I16
2392                    || result_type == types::I8
2393                {
2394                    Ok(builder.ins().sextend(types::I64, result))
2395                } else {
2396                    Ok(result)
2397                }
2398            } else {
2399                // Void return - return 0
2400                Ok(builder.ins().iconst(types::I64, 0))
2401            }
2402        } else {
2403            Err(format!("Unknown function: {}", name))
2404        }
2405    }
2406
2407    /// Compile if expression, returns (value, has_return)
2408    fn compile_if_tracked(
2409        module: &mut JITModule,
2410        functions: &HashMap<String, FuncId>,
2411        extern_fns: &HashMap<String, ExternFnSig>,
2412        builder: &mut FunctionBuilder,
2413        scope: &mut CompileScope,
2414        condition: &Expr,
2415        then_branch: &ast::Block,
2416        else_branch: Option<&Expr>,
2417    ) -> Result<(cranelift_codegen::ir::Value, bool), String> {
2418        // OPTIMIZATION: Use direct condition compilation
2419        let cond_bool =
2420            compile_condition(module, functions, extern_fns, builder, scope, condition)?;
2421
2422        let then_block = builder.create_block();
2423        let else_block = builder.create_block();
2424        let merge_block = builder.create_block();
2425
2426        builder.append_block_param(merge_block, types::I64);
2427
2428        // Branch directly on the boolean - no extra comparison needed
2429        builder
2430            .ins()
2431            .brif(cond_bool, then_block, &[], else_block, &[]);
2432
2433        // Compile then branch
2434        builder.switch_to_block(then_block);
2435        builder.seal_block(then_block);
2436        let mut then_scope = scope.child();
2437        let (then_val, then_returns) = compile_block_tracked(
2438            module,
2439            functions,
2440            extern_fns,
2441            builder,
2442            &mut then_scope,
2443            then_branch,
2444        )?;
2445        // Only jump to merge if we didn't return
2446        if !then_returns {
2447            builder.ins().jump(merge_block, &[then_val]);
2448        }
2449
2450        // Compile else branch
2451        builder.switch_to_block(else_block);
2452        builder.seal_block(else_block);
2453        let (else_val, else_returns) = if let Some(else_expr) = else_branch {
2454            match else_expr {
2455                Expr::Block(block) => {
2456                    let mut else_scope = scope.child();
2457                    compile_block_tracked(
2458                        module,
2459                        functions,
2460                        extern_fns,
2461                        builder,
2462                        &mut else_scope,
2463                        block,
2464                    )?
2465                }
2466                Expr::If {
2467                    condition,
2468                    then_branch,
2469                    else_branch,
2470                } => compile_if_tracked(
2471                    module,
2472                    functions,
2473                    extern_fns,
2474                    builder,
2475                    scope,
2476                    condition,
2477                    then_branch,
2478                    else_branch.as_deref(),
2479                )?,
2480                _ => {
2481                    let val =
2482                        compile_expr(module, functions, extern_fns, builder, scope, else_expr)?;
2483                    (val, false)
2484                }
2485            }
2486        } else {
2487            (builder.ins().iconst(types::I64, 0), false)
2488        };
2489        // Only jump to merge if we didn't return
2490        if !else_returns {
2491            builder.ins().jump(merge_block, &[else_val]);
2492        }
2493
2494        // If both branches return, the merge block is unreachable but still needs to be sealed
2495        // If only some branches return, we still need the merge block
2496        let both_return = then_returns && else_returns;
2497
2498        builder.switch_to_block(merge_block);
2499        builder.seal_block(merge_block);
2500
2501        if both_return {
2502            // Both branches return - merge block is unreachable
2503            // Return a dummy value and signal that we returned
2504            let dummy = builder.ins().iconst(types::I64, 0);
2505            Ok((dummy, true))
2506        } else {
2507            Ok((builder.block_params(merge_block)[0], false))
2508        }
2509    }
2510
2511    /// Compile if expression (convenience wrapper)
2512    fn compile_if(
2513        module: &mut JITModule,
2514        functions: &HashMap<String, FuncId>,
2515        extern_fns: &HashMap<String, ExternFnSig>,
2516        builder: &mut FunctionBuilder,
2517        scope: &mut CompileScope,
2518        condition: &Expr,
2519        then_branch: &ast::Block,
2520        else_branch: Option<&Expr>,
2521    ) -> Result<cranelift_codegen::ir::Value, String> {
2522        compile_if_tracked(
2523            module,
2524            functions,
2525            extern_fns,
2526            builder,
2527            scope,
2528            condition,
2529            then_branch,
2530            else_branch,
2531        )
2532        .map(|(v, _)| v)
2533    }
2534
2535    /// Compile while loop
2536    fn compile_while(
2537        module: &mut JITModule,
2538        functions: &HashMap<String, FuncId>,
2539        extern_fns: &HashMap<String, ExternFnSig>,
2540        builder: &mut FunctionBuilder,
2541        scope: &mut CompileScope,
2542        condition: &Expr,
2543        body: &ast::Block,
2544    ) -> Result<cranelift_codegen::ir::Value, String> {
2545        let header_block = builder.create_block();
2546        let body_block = builder.create_block();
2547        let exit_block = builder.create_block();
2548
2549        builder.ins().jump(header_block, &[]);
2550
2551        builder.switch_to_block(header_block);
2552        // OPTIMIZATION: Use direct condition compilation
2553        let cond_bool =
2554            compile_condition(module, functions, extern_fns, builder, scope, condition)?;
2555        // Branch directly - no extra comparison needed
2556        builder
2557            .ins()
2558            .brif(cond_bool, body_block, &[], exit_block, &[]);
2559
2560        builder.switch_to_block(body_block);
2561        builder.seal_block(body_block);
2562        let mut body_scope = scope.child();
2563        compile_block(
2564            module,
2565            functions,
2566            extern_fns,
2567            builder,
2568            &mut body_scope,
2569            body,
2570        )?;
2571        builder.ins().jump(header_block, &[]);
2572
2573        builder.seal_block(header_block);
2574
2575        builder.switch_to_block(exit_block);
2576        builder.seal_block(exit_block);
2577
2578        Ok(builder.ins().iconst(types::I64, 0))
2579    }
2580
2581    // ============================================
2582    // Runtime support functions (called from JIT)
2583    // ============================================
2584
2585    // Type-aware arithmetic operations
2586    // Uses heuristic: if value looks like a float bit pattern, treat as float
2587    // Small integers (< 2^50) are unlikely to have float patterns
2588    #[inline]
2589    fn is_float_pattern(v: i64) -> bool {
2590        let exp = (v >> 52) & 0x7FF;
2591        // Float exponent is non-zero (except for 0.0 and denormals)
2592        // and not all 1s (infinity/NaN) - valid float range
2593        exp > 0 && exp < 0x7FF && v != 0
2594    }
2595
2596    #[no_mangle]
2597    pub extern "C" fn sigil_add(a: i64, b: i64) -> i64 {
2598        if is_float_pattern(a) || is_float_pattern(b) {
2599            let fa = f64::from_bits(a as u64);
2600            let fb = f64::from_bits(b as u64);
2601            (fa + fb).to_bits() as i64
2602        } else {
2603            a.wrapping_add(b)
2604        }
2605    }
2606
2607    #[no_mangle]
2608    pub extern "C" fn sigil_sub(a: i64, b: i64) -> i64 {
2609        if is_float_pattern(a) || is_float_pattern(b) {
2610            let fa = f64::from_bits(a as u64);
2611            let fb = f64::from_bits(b as u64);
2612            (fa - fb).to_bits() as i64
2613        } else {
2614            a.wrapping_sub(b)
2615        }
2616    }
2617
2618    #[no_mangle]
2619    pub extern "C" fn sigil_mul(a: i64, b: i64) -> i64 {
2620        if is_float_pattern(a) || is_float_pattern(b) {
2621            let fa = f64::from_bits(a as u64);
2622            let fb = f64::from_bits(b as u64);
2623            (fa * fb).to_bits() as i64
2624        } else {
2625            a.wrapping_mul(b)
2626        }
2627    }
2628
2629    #[no_mangle]
2630    pub extern "C" fn sigil_div(a: i64, b: i64) -> i64 {
2631        if is_float_pattern(a) || is_float_pattern(b) {
2632            let fa = f64::from_bits(a as u64);
2633            let fb = f64::from_bits(b as u64);
2634            (fa / fb).to_bits() as i64
2635        } else if b != 0 {
2636            a / b
2637        } else {
2638            0 // Avoid division by zero
2639        }
2640    }
2641
2642    #[no_mangle]
2643    pub extern "C" fn sigil_lt(a: i64, b: i64) -> i64 {
2644        if is_float_pattern(a) || is_float_pattern(b) {
2645            let fa = f64::from_bits(a as u64);
2646            let fb = f64::from_bits(b as u64);
2647            if fa < fb {
2648                1
2649            } else {
2650                0
2651            }
2652        } else {
2653            if a < b {
2654                1
2655            } else {
2656                0
2657            }
2658        }
2659    }
2660
2661    #[no_mangle]
2662    pub extern "C" fn sigil_le(a: i64, b: i64) -> i64 {
2663        if is_float_pattern(a) || is_float_pattern(b) {
2664            let fa = f64::from_bits(a as u64);
2665            let fb = f64::from_bits(b as u64);
2666            if fa <= fb {
2667                1
2668            } else {
2669                0
2670            }
2671        } else {
2672            if a <= b {
2673                1
2674            } else {
2675                0
2676            }
2677        }
2678    }
2679
2680    #[no_mangle]
2681    pub extern "C" fn sigil_gt(a: i64, b: i64) -> i64 {
2682        if is_float_pattern(a) || is_float_pattern(b) {
2683            let fa = f64::from_bits(a as u64);
2684            let fb = f64::from_bits(b as u64);
2685            if fa > fb {
2686                1
2687            } else {
2688                0
2689            }
2690        } else {
2691            if a > b {
2692                1
2693            } else {
2694                0
2695            }
2696        }
2697    }
2698
2699    #[no_mangle]
2700    pub extern "C" fn sigil_ge(a: i64, b: i64) -> i64 {
2701        if is_float_pattern(a) || is_float_pattern(b) {
2702            let fa = f64::from_bits(a as u64);
2703            let fb = f64::from_bits(b as u64);
2704            if fa >= fb {
2705                1
2706            } else {
2707                0
2708            }
2709        } else {
2710            if a >= b {
2711                1
2712            } else {
2713                0
2714            }
2715        }
2716    }
2717
2718    // Print that handles both int and float
2719    #[no_mangle]
2720    pub extern "C" fn sigil_print(v: i64) -> i64 {
2721        if is_float_pattern(v) {
2722            println!("{}", f64::from_bits(v as u64));
2723        } else {
2724            println!("{}", v);
2725        }
2726        0
2727    }
2728
2729    // ============================================
2730    // SIMD Operations (Vec4 = 4xf64)
2731    // ============================================
2732    // HARDWARE SIMD VECTOR OPERATIONS
2733    // ============================================
2734    // Uses AVX/SSE intrinsics when available for maximum performance.
2735    // SIMD vectors are stored as heap-allocated arrays of 4 f64 values.
2736    // On x86_64 with AVX, uses _mm256_* intrinsics for 4-wide f64 ops.
2737    // Pointer to array is stored as i64.
2738
2739    /// SIMD vector storage - 32-byte aligned for AVX
2740    #[repr(C, align(32))]
2741    struct SimdVec4 {
2742        data: [f64; 4],
2743    }
2744
2745    impl SimdVec4 {
2746        #[inline(always)]
2747        fn new(x: f64, y: f64, z: f64, w: f64) -> Box<Self> {
2748            Box::new(SimdVec4 { data: [x, y, z, w] })
2749        }
2750
2751        #[inline(always)]
2752        fn splat(v: f64) -> Box<Self> {
2753            Box::new(SimdVec4 { data: [v, v, v, v] })
2754        }
2755    }
2756
2757    /// Create a new Vec4 SIMD vector
2758    #[no_mangle]
2759    pub extern "C" fn sigil_simd_new(x: i64, y: i64, z: i64, w: i64) -> i64 {
2760        let v = SimdVec4::new(
2761            f64::from_bits(x as u64),
2762            f64::from_bits(y as u64),
2763            f64::from_bits(z as u64),
2764            f64::from_bits(w as u64),
2765        );
2766        Box::into_raw(v) as i64
2767    }
2768
2769    /// Create Vec4 by splatting a scalar to all lanes
2770    #[no_mangle]
2771    pub extern "C" fn sigil_simd_splat(v: i64) -> i64 {
2772        let f = f64::from_bits(v as u64);
2773        let v = SimdVec4::splat(f);
2774        Box::into_raw(v) as i64
2775    }
2776
2777    // AVX-optimized SIMD operations using inline assembly / intrinsics pattern
2778    // The compiler will auto-vectorize these aligned operations with -C target-cpu=native
2779
2780    /// SIMD add - uses AVX when available
2781    #[no_mangle]
2782    #[inline(never)]
2783    pub extern "C" fn sigil_simd_add(a: i64, b: i64) -> i64 {
2784        unsafe {
2785            let a = &*(a as *const SimdVec4);
2786            let b = &*(b as *const SimdVec4);
2787            // Aligned load/store enables auto-vectorization
2788            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2789            r.data[0] = a.data[0] + b.data[0];
2790            r.data[1] = a.data[1] + b.data[1];
2791            r.data[2] = a.data[2] + b.data[2];
2792            r.data[3] = a.data[3] + b.data[3];
2793            Box::into_raw(r) as i64
2794        }
2795    }
2796
2797    /// SIMD subtract
2798    #[no_mangle]
2799    #[inline(never)]
2800    pub extern "C" fn sigil_simd_sub(a: i64, b: i64) -> i64 {
2801        unsafe {
2802            let a = &*(a as *const SimdVec4);
2803            let b = &*(b as *const SimdVec4);
2804            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2805            r.data[0] = a.data[0] - b.data[0];
2806            r.data[1] = a.data[1] - b.data[1];
2807            r.data[2] = a.data[2] - b.data[2];
2808            r.data[3] = a.data[3] - b.data[3];
2809            Box::into_raw(r) as i64
2810        }
2811    }
2812
2813    /// SIMD multiply
2814    #[no_mangle]
2815    #[inline(never)]
2816    pub extern "C" fn sigil_simd_mul(a: i64, b: i64) -> i64 {
2817        unsafe {
2818            let a = &*(a as *const SimdVec4);
2819            let b = &*(b as *const SimdVec4);
2820            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2821            r.data[0] = a.data[0] * b.data[0];
2822            r.data[1] = a.data[1] * b.data[1];
2823            r.data[2] = a.data[2] * b.data[2];
2824            r.data[3] = a.data[3] * b.data[3];
2825            Box::into_raw(r) as i64
2826        }
2827    }
2828
2829    /// SIMD divide
2830    #[no_mangle]
2831    #[inline(never)]
2832    pub extern "C" fn sigil_simd_div(a: i64, b: i64) -> i64 {
2833        unsafe {
2834            let a = &*(a as *const SimdVec4);
2835            let b = &*(b as *const SimdVec4);
2836            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2837            r.data[0] = a.data[0] / b.data[0];
2838            r.data[1] = a.data[1] / b.data[1];
2839            r.data[2] = a.data[2] / b.data[2];
2840            r.data[3] = a.data[3] / b.data[3];
2841            Box::into_raw(r) as i64
2842        }
2843    }
2844
2845    /// SIMD dot product (returns scalar) - optimized for auto-vectorization
2846    #[no_mangle]
2847    #[inline(never)]
2848    pub extern "C" fn sigil_simd_dot(a: i64, b: i64) -> i64 {
2849        unsafe {
2850            let a = &*(a as *const SimdVec4);
2851            let b = &*(b as *const SimdVec4);
2852            // FMA-friendly pattern for dot product
2853            let r = a.data[0].mul_add(
2854                b.data[0],
2855                a.data[1].mul_add(
2856                    b.data[1],
2857                    a.data[2].mul_add(b.data[2], a.data[3] * b.data[3]),
2858                ),
2859            );
2860            r.to_bits() as i64
2861        }
2862    }
2863
2864    /// SIMD horizontal add (sum all lanes)
2865    #[no_mangle]
2866    #[inline(never)]
2867    pub extern "C" fn sigil_simd_hadd(a: i64) -> i64 {
2868        unsafe {
2869            let a = &*(a as *const SimdVec4);
2870            // Pairwise add pattern for better vectorization
2871            let sum01 = a.data[0] + a.data[1];
2872            let sum23 = a.data[2] + a.data[3];
2873            let r = sum01 + sum23;
2874            r.to_bits() as i64
2875        }
2876    }
2877
2878    /// SIMD length squared - uses FMA for better performance
2879    #[no_mangle]
2880    #[inline(never)]
2881    pub extern "C" fn sigil_simd_length_sq(a: i64) -> i64 {
2882        unsafe {
2883            let a = &*(a as *const SimdVec4);
2884            let r = a.data[0].mul_add(
2885                a.data[0],
2886                a.data[1].mul_add(
2887                    a.data[1],
2888                    a.data[2].mul_add(a.data[2], a.data[3] * a.data[3]),
2889                ),
2890            );
2891            r.to_bits() as i64
2892        }
2893    }
2894
2895    /// SIMD length - uses FMA for length calculation
2896    #[no_mangle]
2897    #[inline(never)]
2898    pub extern "C" fn sigil_simd_length(a: i64) -> i64 {
2899        unsafe {
2900            let a = &*(a as *const SimdVec4);
2901            let len_sq = a.data[0].mul_add(
2902                a.data[0],
2903                a.data[1].mul_add(
2904                    a.data[1],
2905                    a.data[2].mul_add(a.data[2], a.data[3] * a.data[3]),
2906                ),
2907            );
2908            let r = len_sq.sqrt();
2909            r.to_bits() as i64
2910        }
2911    }
2912
2913    /// SIMD normalize - fast reciprocal sqrt pattern
2914    #[no_mangle]
2915    #[inline(never)]
2916    pub extern "C" fn sigil_simd_normalize(a: i64) -> i64 {
2917        unsafe {
2918            let a = &*(a as *const SimdVec4);
2919            let len_sq = a.data[0].mul_add(
2920                a.data[0],
2921                a.data[1].mul_add(
2922                    a.data[1],
2923                    a.data[2].mul_add(a.data[2], a.data[3] * a.data[3]),
2924                ),
2925            );
2926            let inv = if len_sq > 1e-20 {
2927                1.0 / len_sq.sqrt()
2928            } else {
2929                0.0
2930            };
2931            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2932            r.data[0] = a.data[0] * inv;
2933            r.data[1] = a.data[1] * inv;
2934            r.data[2] = a.data[2] * inv;
2935            r.data[3] = a.data[3] * inv;
2936            Box::into_raw(r) as i64
2937        }
2938    }
2939
2940    /// SIMD cross product (3D, ignores w component)
2941    #[no_mangle]
2942    #[inline(never)]
2943    pub extern "C" fn sigil_simd_cross(a: i64, b: i64) -> i64 {
2944        unsafe {
2945            let a = &*(a as *const SimdVec4);
2946            let b = &*(b as *const SimdVec4);
2947            // Cross product using FMA where beneficial
2948            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2949            r.data[0] = a.data[1].mul_add(b.data[2], -(a.data[2] * b.data[1]));
2950            r.data[1] = a.data[2].mul_add(b.data[0], -(a.data[0] * b.data[2]));
2951            r.data[2] = a.data[0].mul_add(b.data[1], -(a.data[1] * b.data[0]));
2952            r.data[3] = 0.0;
2953            Box::into_raw(r) as i64
2954        }
2955    }
2956
2957    /// SIMD min - element-wise minimum
2958    #[no_mangle]
2959    #[inline(never)]
2960    pub extern "C" fn sigil_simd_min(a: i64, b: i64) -> i64 {
2961        unsafe {
2962            let a = &*(a as *const SimdVec4);
2963            let b = &*(b as *const SimdVec4);
2964            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2965            r.data[0] = a.data[0].min(b.data[0]);
2966            r.data[1] = a.data[1].min(b.data[1]);
2967            r.data[2] = a.data[2].min(b.data[2]);
2968            r.data[3] = a.data[3].min(b.data[3]);
2969            Box::into_raw(r) as i64
2970        }
2971    }
2972
2973    /// SIMD max - element-wise maximum
2974    #[no_mangle]
2975    #[inline(never)]
2976    pub extern "C" fn sigil_simd_max(a: i64, b: i64) -> i64 {
2977        unsafe {
2978            let a = &*(a as *const SimdVec4);
2979            let b = &*(b as *const SimdVec4);
2980            let mut r = SimdVec4::new(0.0, 0.0, 0.0, 0.0);
2981            r.data[0] = a.data[0].max(b.data[0]);
2982            r.data[1] = a.data[1].max(b.data[1]);
2983            r.data[2] = a.data[2].max(b.data[2]);
2984            r.data[3] = a.data[3].max(b.data[3]);
2985            Box::into_raw(r) as i64
2986        }
2987    }
2988
2989    /// Extract element from SIMD vector
2990    #[no_mangle]
2991    pub extern "C" fn sigil_simd_extract(v: i64, idx: i64) -> i64 {
2992        unsafe {
2993            let v = &*(v as *const SimdVec4);
2994            let r = v.data[(idx as usize) & 3];
2995            r.to_bits() as i64
2996        }
2997    }
2998
2999    /// Free SIMD vector (for memory management)
3000    #[no_mangle]
3001    pub extern "C" fn sigil_simd_free(v: i64) {
3002        if v != 0 {
3003            unsafe {
3004                let _ = Box::from_raw(v as *mut SimdVec4);
3005            }
3006        }
3007    }
3008
3009    #[no_mangle]
3010    pub extern "C" fn sigil_sqrt(x: f64) -> f64 {
3011        x.sqrt()
3012    }
3013
3014    #[no_mangle]
3015    pub extern "C" fn sigil_sin(x: f64) -> f64 {
3016        x.sin()
3017    }
3018
3019    #[no_mangle]
3020    pub extern "C" fn sigil_cos(x: f64) -> f64 {
3021        x.cos()
3022    }
3023
3024    #[no_mangle]
3025    pub extern "C" fn sigil_pow(base: f64, exp: f64) -> f64 {
3026        base.powf(exp)
3027    }
3028
3029    #[no_mangle]
3030    pub extern "C" fn sigil_exp(x: f64) -> f64 {
3031        x.exp()
3032    }
3033
3034    #[no_mangle]
3035    pub extern "C" fn sigil_ln(x: f64) -> f64 {
3036        x.ln()
3037    }
3038
3039    #[no_mangle]
3040    pub extern "C" fn sigil_floor(x: f64) -> f64 {
3041        x.floor()
3042    }
3043
3044    #[no_mangle]
3045    pub extern "C" fn sigil_ceil(x: f64) -> f64 {
3046        x.ceil()
3047    }
3048
3049    #[no_mangle]
3050    pub extern "C" fn sigil_abs(x: f64) -> f64 {
3051        x.abs()
3052    }
3053
3054    #[no_mangle]
3055    pub extern "C" fn sigil_print_int(x: i64) -> i64 {
3056        println!("{}", x);
3057        0
3058    }
3059
3060    #[no_mangle]
3061    pub extern "C" fn sigil_print_float(x: f64) -> i64 {
3062        println!("{}", x);
3063        0
3064    }
3065
3066    #[no_mangle]
3067    pub extern "C" fn sigil_print_str(ptr: *const u8, len: usize) -> i64 {
3068        unsafe {
3069            let slice = std::slice::from_raw_parts(ptr, len);
3070            if let Ok(s) = std::str::from_utf8(slice) {
3071                println!("{}", s);
3072            }
3073        }
3074        0
3075    }
3076
3077    #[no_mangle]
3078    pub extern "C" fn sigil_now() -> i64 {
3079        use std::time::{SystemTime, UNIX_EPOCH};
3080        SystemTime::now()
3081            .duration_since(UNIX_EPOCH)
3082            .map(|d| d.as_millis() as i64)
3083            .unwrap_or(0)
3084    }
3085
3086    // Simple array implementation using heap allocation
3087    #[repr(C)]
3088    struct SigilArray {
3089        data: *mut i64,
3090        len: usize,
3091        cap: usize,
3092    }
3093
3094    #[no_mangle]
3095    pub extern "C" fn sigil_array_new(capacity: i64) -> i64 {
3096        let cap = capacity.max(8) as usize;
3097        let layout = std::alloc::Layout::array::<i64>(cap).unwrap();
3098        let data = unsafe { std::alloc::alloc(layout) as *mut i64 };
3099
3100        let arr = Box::new(SigilArray { data, len: 0, cap });
3101        Box::into_raw(arr) as i64
3102    }
3103
3104    #[no_mangle]
3105    pub extern "C" fn sigil_array_push(arr_ptr: i64, value: i64) -> i64 {
3106        unsafe {
3107            let arr = &mut *(arr_ptr as *mut SigilArray);
3108            if arr.len >= arr.cap {
3109                // Grow array
3110                let new_cap = arr.cap * 2;
3111                let old_layout = std::alloc::Layout::array::<i64>(arr.cap).unwrap();
3112                let new_layout = std::alloc::Layout::array::<i64>(new_cap).unwrap();
3113                arr.data = std::alloc::realloc(arr.data as *mut u8, old_layout, new_layout.size())
3114                    as *mut i64;
3115                arr.cap = new_cap;
3116            }
3117            *arr.data.add(arr.len) = value;
3118            arr.len += 1;
3119        }
3120        0
3121    }
3122
3123    #[no_mangle]
3124    pub extern "C" fn sigil_array_get(arr_ptr: i64, index: i64) -> i64 {
3125        unsafe {
3126            let arr = &*(arr_ptr as *const SigilArray);
3127            let idx = index as usize;
3128            if idx < arr.len {
3129                *arr.data.add(idx)
3130            } else {
3131                0 // Out of bounds returns 0
3132            }
3133        }
3134    }
3135
3136    #[no_mangle]
3137    pub extern "C" fn sigil_array_set(arr_ptr: i64, index: i64, value: i64) -> i64 {
3138        unsafe {
3139            let arr = &mut *(arr_ptr as *mut SigilArray);
3140            let idx = index as usize;
3141            // Extend array if needed
3142            while arr.len <= idx {
3143                sigil_array_push(arr_ptr, 0);
3144            }
3145            *arr.data.add(idx) = value;
3146        }
3147        value
3148    }
3149
3150    #[no_mangle]
3151    pub extern "C" fn sigil_array_len(arr_ptr: i64) -> i64 {
3152        unsafe {
3153            let arr = &*(arr_ptr as *const SigilArray);
3154            arr.len as i64
3155        }
3156    }
3157
3158    // ============================================
3159    // SIMD-Optimized Array Operations
3160    // ============================================
3161    // These operations process arrays in SIMD-friendly batches
3162
3163    /// Sum all elements in an array using SIMD-friendly loop
3164    #[no_mangle]
3165    pub extern "C" fn sigil_array_sum(arr_ptr: i64) -> i64 {
3166        unsafe {
3167            let arr = &*(arr_ptr as *const SigilArray);
3168            let data = std::slice::from_raw_parts(arr.data, arr.len);
3169
3170            // Process in batches of 4 for SIMD-friendliness
3171            let chunks = data.chunks_exact(4);
3172            let remainder = chunks.remainder();
3173
3174            // Accumulate 4 partial sums (allows SIMD vectorization)
3175            let mut sum0: i64 = 0;
3176            let mut sum1: i64 = 0;
3177            let mut sum2: i64 = 0;
3178            let mut sum3: i64 = 0;
3179
3180            for chunk in chunks {
3181                sum0 = sum0.wrapping_add(chunk[0]);
3182                sum1 = sum1.wrapping_add(chunk[1]);
3183                sum2 = sum2.wrapping_add(chunk[2]);
3184                sum3 = sum3.wrapping_add(chunk[3]);
3185            }
3186
3187            // Add remainder
3188            let mut sum = sum0
3189                .wrapping_add(sum1)
3190                .wrapping_add(sum2)
3191                .wrapping_add(sum3);
3192            for &v in remainder {
3193                sum = sum.wrapping_add(v);
3194            }
3195
3196            sum
3197        }
3198    }
3199
3200    /// Multiply all elements by a scalar (in-place, SIMD-friendly)
3201    #[no_mangle]
3202    pub extern "C" fn sigil_array_scale(arr_ptr: i64, scalar: i64) -> i64 {
3203        unsafe {
3204            let arr = &mut *(arr_ptr as *mut SigilArray);
3205            let data = std::slice::from_raw_parts_mut(arr.data, arr.len);
3206
3207            // Process in batches of 4 for SIMD-friendliness
3208            for chunk in data.chunks_exact_mut(4) {
3209                chunk[0] = chunk[0].wrapping_mul(scalar);
3210                chunk[1] = chunk[1].wrapping_mul(scalar);
3211                chunk[2] = chunk[2].wrapping_mul(scalar);
3212                chunk[3] = chunk[3].wrapping_mul(scalar);
3213            }
3214
3215            // Handle remainder
3216            let remainder_start = (data.len() / 4) * 4;
3217            for v in &mut data[remainder_start..] {
3218                *v = v.wrapping_mul(scalar);
3219            }
3220
3221            arr_ptr
3222        }
3223    }
3224
3225    /// Add a scalar to all elements (in-place, SIMD-friendly)
3226    #[no_mangle]
3227    pub extern "C" fn sigil_array_offset(arr_ptr: i64, offset: i64) -> i64 {
3228        unsafe {
3229            let arr = &mut *(arr_ptr as *mut SigilArray);
3230            let data = std::slice::from_raw_parts_mut(arr.data, arr.len);
3231
3232            // Process in batches of 4 for SIMD-friendliness
3233            for chunk in data.chunks_exact_mut(4) {
3234                chunk[0] = chunk[0].wrapping_add(offset);
3235                chunk[1] = chunk[1].wrapping_add(offset);
3236                chunk[2] = chunk[2].wrapping_add(offset);
3237                chunk[3] = chunk[3].wrapping_add(offset);
3238            }
3239
3240            let remainder_start = (data.len() / 4) * 4;
3241            for v in &mut data[remainder_start..] {
3242                *v = v.wrapping_add(offset);
3243            }
3244
3245            arr_ptr
3246        }
3247    }
3248
3249    /// Dot product of two arrays (SIMD-friendly)
3250    #[no_mangle]
3251    pub extern "C" fn sigil_array_dot(a_ptr: i64, b_ptr: i64) -> i64 {
3252        unsafe {
3253            let a_arr = &*(a_ptr as *const SigilArray);
3254            let b_arr = &*(b_ptr as *const SigilArray);
3255
3256            let len = a_arr.len.min(b_arr.len);
3257            let a_data = std::slice::from_raw_parts(a_arr.data, len);
3258            let b_data = std::slice::from_raw_parts(b_arr.data, len);
3259
3260            // Process in batches of 4 for SIMD-friendliness
3261            let mut sum0: i64 = 0;
3262            let mut sum1: i64 = 0;
3263            let mut sum2: i64 = 0;
3264            let mut sum3: i64 = 0;
3265
3266            let chunks = len / 4;
3267            for i in 0..chunks {
3268                let base = i * 4;
3269                sum0 = sum0.wrapping_add(a_data[base].wrapping_mul(b_data[base]));
3270                sum1 = sum1.wrapping_add(a_data[base + 1].wrapping_mul(b_data[base + 1]));
3271                sum2 = sum2.wrapping_add(a_data[base + 2].wrapping_mul(b_data[base + 2]));
3272                sum3 = sum3.wrapping_add(a_data[base + 3].wrapping_mul(b_data[base + 3]));
3273            }
3274
3275            // Add remainder
3276            let mut sum = sum0
3277                .wrapping_add(sum1)
3278                .wrapping_add(sum2)
3279                .wrapping_add(sum3);
3280            for i in (chunks * 4)..len {
3281                sum = sum.wrapping_add(a_data[i].wrapping_mul(b_data[i]));
3282            }
3283
3284            sum
3285        }
3286    }
3287
3288    /// Element-wise add two arrays into a new array (SIMD-friendly)
3289    #[no_mangle]
3290    pub extern "C" fn sigil_array_add(a_ptr: i64, b_ptr: i64) -> i64 {
3291        unsafe {
3292            let a_arr = &*(a_ptr as *const SigilArray);
3293            let b_arr = &*(b_ptr as *const SigilArray);
3294
3295            let len = a_arr.len.min(b_arr.len);
3296            let a_data = std::slice::from_raw_parts(a_arr.data, len);
3297            let b_data = std::slice::from_raw_parts(b_arr.data, len);
3298
3299            // Create result array
3300            let result = sigil_array_new(len as i64);
3301            let r_arr = &mut *(result as *mut SigilArray);
3302            r_arr.len = len;
3303            let r_data = std::slice::from_raw_parts_mut(r_arr.data, len);
3304
3305            // Process in batches of 4 for SIMD-friendliness
3306            for i in 0..(len / 4) {
3307                let base = i * 4;
3308                r_data[base] = a_data[base].wrapping_add(b_data[base]);
3309                r_data[base + 1] = a_data[base + 1].wrapping_add(b_data[base + 1]);
3310                r_data[base + 2] = a_data[base + 2].wrapping_add(b_data[base + 2]);
3311                r_data[base + 3] = a_data[base + 3].wrapping_add(b_data[base + 3]);
3312            }
3313
3314            // Handle remainder
3315            for i in ((len / 4) * 4)..len {
3316                r_data[i] = a_data[i].wrapping_add(b_data[i]);
3317            }
3318
3319            result
3320        }
3321    }
3322
3323    /// Element-wise multiply two arrays into a new array (SIMD-friendly)
3324    #[no_mangle]
3325    pub extern "C" fn sigil_array_mul(a_ptr: i64, b_ptr: i64) -> i64 {
3326        unsafe {
3327            let a_arr = &*(a_ptr as *const SigilArray);
3328            let b_arr = &*(b_ptr as *const SigilArray);
3329
3330            let len = a_arr.len.min(b_arr.len);
3331            let a_data = std::slice::from_raw_parts(a_arr.data, len);
3332            let b_data = std::slice::from_raw_parts(b_arr.data, len);
3333
3334            // Create result array
3335            let result = sigil_array_new(len as i64);
3336            let r_arr = &mut *(result as *mut SigilArray);
3337            r_arr.len = len;
3338            let r_data = std::slice::from_raw_parts_mut(r_arr.data, len);
3339
3340            // Process in batches of 4 for SIMD-friendliness
3341            for i in 0..(len / 4) {
3342                let base = i * 4;
3343                r_data[base] = a_data[base].wrapping_mul(b_data[base]);
3344                r_data[base + 1] = a_data[base + 1].wrapping_mul(b_data[base + 1]);
3345                r_data[base + 2] = a_data[base + 2].wrapping_mul(b_data[base + 2]);
3346                r_data[base + 3] = a_data[base + 3].wrapping_mul(b_data[base + 3]);
3347            }
3348
3349            // Handle remainder
3350            for i in ((len / 4) * 4)..len {
3351                r_data[i] = a_data[i].wrapping_mul(b_data[i]);
3352            }
3353
3354            result
3355        }
3356    }
3357
3358    /// Find minimum value in array (SIMD-friendly)
3359    #[no_mangle]
3360    pub extern "C" fn sigil_array_min(arr_ptr: i64) -> i64 {
3361        unsafe {
3362            let arr = &*(arr_ptr as *const SigilArray);
3363            if arr.len == 0 {
3364                return 0;
3365            }
3366
3367            let data = std::slice::from_raw_parts(arr.data, arr.len);
3368
3369            // Process in batches of 4
3370            let mut min0 = i64::MAX;
3371            let mut min1 = i64::MAX;
3372            let mut min2 = i64::MAX;
3373            let mut min3 = i64::MAX;
3374
3375            for chunk in data.chunks_exact(4) {
3376                min0 = min0.min(chunk[0]);
3377                min1 = min1.min(chunk[1]);
3378                min2 = min2.min(chunk[2]);
3379                min3 = min3.min(chunk[3]);
3380            }
3381
3382            let mut min_val = min0.min(min1).min(min2).min(min3);
3383
3384            // Handle remainder
3385            let remainder_start = (data.len() / 4) * 4;
3386            for &v in &data[remainder_start..] {
3387                min_val = min_val.min(v);
3388            }
3389
3390            min_val
3391        }
3392    }
3393
3394    /// Find maximum value in array (SIMD-friendly)
3395    #[no_mangle]
3396    pub extern "C" fn sigil_array_max(arr_ptr: i64) -> i64 {
3397        unsafe {
3398            let arr = &*(arr_ptr as *const SigilArray);
3399            if arr.len == 0 {
3400                return 0;
3401            }
3402
3403            let data = std::slice::from_raw_parts(arr.data, arr.len);
3404
3405            // Process in batches of 4
3406            let mut max0 = i64::MIN;
3407            let mut max1 = i64::MIN;
3408            let mut max2 = i64::MIN;
3409            let mut max3 = i64::MIN;
3410
3411            for chunk in data.chunks_exact(4) {
3412                max0 = max0.max(chunk[0]);
3413                max1 = max1.max(chunk[1]);
3414                max2 = max2.max(chunk[2]);
3415                max3 = max3.max(chunk[3]);
3416            }
3417
3418            let mut max_val = max0.max(max1).max(max2).max(max3);
3419
3420            // Handle remainder
3421            let remainder_start = (data.len() / 4) * 4;
3422            for &v in &data[remainder_start..] {
3423                max_val = max_val.max(v);
3424            }
3425
3426            max_val
3427        }
3428    }
3429
3430    /// Fill array with a value (SIMD-friendly)
3431    #[no_mangle]
3432    pub extern "C" fn sigil_array_fill(arr_ptr: i64, value: i64, count: i64) -> i64 {
3433        unsafe {
3434            let arr = &mut *(arr_ptr as *mut SigilArray);
3435            let n = count as usize;
3436
3437            // Ensure capacity
3438            while arr.len < n {
3439                sigil_array_push(arr_ptr, 0);
3440            }
3441
3442            let data = std::slice::from_raw_parts_mut(arr.data, n);
3443
3444            // Process in batches of 4
3445            for chunk in data.chunks_exact_mut(4) {
3446                chunk[0] = value;
3447                chunk[1] = value;
3448                chunk[2] = value;
3449                chunk[3] = value;
3450            }
3451
3452            // Handle remainder
3453            let remainder_start = (n / 4) * 4;
3454            for v in &mut data[remainder_start..] {
3455                *v = value;
3456            }
3457
3458            arr_ptr
3459        }
3460    }
3461
3462    // ============================================
3463    // PipeOp Array Access Functions
3464    // ============================================
3465    // Functions for the access morphemes: α (first), ω (last), μ (middle), χ (choice), ν (nth), ξ (next)
3466
3467    /// Get first element of array (α morpheme)
3468    #[no_mangle]
3469    pub extern "C" fn sigil_array_first(arr_ptr: i64) -> i64 {
3470        unsafe {
3471            let arr = &*(arr_ptr as *const SigilArray);
3472            if arr.len == 0 {
3473                return 0; // Return 0 for empty array
3474            }
3475            *arr.data
3476        }
3477    }
3478
3479    /// Get last element of array (ω morpheme)
3480    #[no_mangle]
3481    pub extern "C" fn sigil_array_last(arr_ptr: i64) -> i64 {
3482        unsafe {
3483            let arr = &*(arr_ptr as *const SigilArray);
3484            if arr.len == 0 {
3485                return 0; // Return 0 for empty array
3486            }
3487            *arr.data.add(arr.len - 1)
3488        }
3489    }
3490
3491    /// Get middle element of array (μ morpheme)
3492    #[no_mangle]
3493    pub extern "C" fn sigil_array_middle(arr_ptr: i64) -> i64 {
3494        unsafe {
3495            let arr = &*(arr_ptr as *const SigilArray);
3496            if arr.len == 0 {
3497                return 0; // Return 0 for empty array
3498            }
3499            let mid = arr.len / 2;
3500            *arr.data.add(mid)
3501        }
3502    }
3503
3504    /// Get random element of array (χ morpheme)
3505    #[no_mangle]
3506    pub extern "C" fn sigil_array_choice(arr_ptr: i64) -> i64 {
3507        unsafe {
3508            let arr = &*(arr_ptr as *const SigilArray);
3509            if arr.len == 0 {
3510                return 0; // Return 0 for empty array
3511            }
3512            // Simple LCG-based random using time as seed
3513            use std::time::{SystemTime, UNIX_EPOCH};
3514            let seed = SystemTime::now()
3515                .duration_since(UNIX_EPOCH)
3516                .map(|d| d.as_nanos() as u64)
3517                .unwrap_or(12345);
3518            let idx =
3519                ((seed.wrapping_mul(1103515245).wrapping_add(12345)) >> 16) as usize % arr.len;
3520            *arr.data.add(idx)
3521        }
3522    }
3523
3524    /// Get nth element of array (ν morpheme) - same as sigil_array_get but clearer semantics
3525    #[no_mangle]
3526    pub extern "C" fn sigil_array_nth(arr_ptr: i64, index: i64) -> i64 {
3527        sigil_array_get(arr_ptr, index)
3528    }
3529
3530    /// Get next element (iterator advance) - currently returns first element (ξ morpheme)
3531    #[no_mangle]
3532    pub extern "C" fn sigil_array_next(arr_ptr: i64) -> i64 {
3533        // For now, next returns the first element
3534        // A full iterator implementation would track state
3535        sigil_array_first(arr_ptr)
3536    }
3537
3538    /// Product of all elements in array (Π morpheme)
3539    #[no_mangle]
3540    pub extern "C" fn sigil_array_product(arr_ptr: i64) -> i64 {
3541        unsafe {
3542            let arr = &*(arr_ptr as *const SigilArray);
3543            if arr.len == 0 {
3544                return 1; // Product of empty set is 1 (identity)
3545            }
3546            let mut product: i64 = 1;
3547            for i in 0..arr.len {
3548                product = product.wrapping_mul(*arr.data.add(i));
3549            }
3550            product
3551        }
3552    }
3553
3554    /// Sort array in ascending order (σ morpheme) - returns new sorted array
3555    #[no_mangle]
3556    pub extern "C" fn sigil_array_sort(arr_ptr: i64) -> i64 {
3557        unsafe {
3558            let arr = &*(arr_ptr as *const SigilArray);
3559            if arr.len == 0 {
3560                return sigil_array_new(0);
3561            }
3562
3563            // Copy elements to a Vec for sorting
3564            let mut elements: Vec<i64> = Vec::with_capacity(arr.len);
3565            for i in 0..arr.len {
3566                elements.push(*arr.data.add(i));
3567            }
3568
3569            // Sort ascending
3570            elements.sort();
3571
3572            // Create new array with sorted elements
3573            let new_arr = sigil_array_new(arr.len as i64);
3574            for elem in elements {
3575                sigil_array_push(new_arr, elem);
3576            }
3577            new_arr
3578        }
3579    }
3580
3581    // ============================================
3582    // Parallel Execution Functions (∥ morpheme)
3583    // ============================================
3584    // These provide multi-threaded execution of array operations
3585    // For JIT compilation, these use a simple thread pool approach
3586
3587    /// Parallel map operation - applies a transformation in parallel across array elements
3588    /// For now, returns the array unchanged as full closure parallelization
3589    /// requires more complex infrastructure. In production, this would:
3590    /// 1. Partition array into chunks based on available CPU cores
3591    /// 2. Spawn worker threads for each chunk
3592    /// 3. Apply transform closure in parallel
3593    /// 4. Collect results
3594    #[no_mangle]
3595    pub extern "C" fn sigil_parallel_map(arr_ptr: i64) -> i64 {
3596        // Stub: returns array unchanged
3597        // Full implementation would use rayon::par_iter or manual thread pool
3598        arr_ptr
3599    }
3600
3601    /// Parallel filter operation - filters elements in parallel
3602    /// Uses parallel predicate evaluation with stream compaction
3603    #[no_mangle]
3604    pub extern "C" fn sigil_parallel_filter(arr_ptr: i64) -> i64 {
3605        // Stub: returns array unchanged
3606        // Full implementation would:
3607        // 1. Evaluate predicates in parallel
3608        // 2. Use prefix sum for compaction offsets
3609        // 3. Parallel write to output array
3610        arr_ptr
3611    }
3612
3613    /// Parallel reduce operation - tree reduction for associative operations
3614    /// Achieves O(log n) depth with O(n) work
3615    #[no_mangle]
3616    pub extern "C" fn sigil_parallel_reduce(arr_ptr: i64) -> i64 {
3617        // For reduction, we can implement a parallel tree reduction
3618        // Falls back to sequential sum for now
3619        unsafe {
3620            let arr = &*(arr_ptr as *const SigilArray);
3621            if arr.len == 0 {
3622                return 0;
3623            }
3624
3625            // Simple sequential sum - parallel tree reduction would
3626            // use divide-and-conquer with thread spawning
3627            let mut sum: i64 = 0;
3628            for i in 0..arr.len {
3629                sum += *arr.data.add(i);
3630            }
3631            sum
3632        }
3633    }
3634
3635    // ============================================
3636    // GPU Compute Functions (⊛ morpheme)
3637    // ============================================
3638    // These would dispatch operations to GPU via wgpu/vulkan
3639    // Currently stubs that fall back to CPU execution
3640
3641    /// GPU map operation - would compile to WGSL/SPIR-V compute shader
3642    /// Shader structure:
3643    /// ```wgsl
3644    /// @compute @workgroup_size(256)
3645    /// fn main(@builtin(global_invocation_id) id: vec3<u32>) {
3646    ///     let idx = id.x;
3647    ///     output[idx] = transform(input[idx]);
3648    /// }
3649    /// ```
3650    #[no_mangle]
3651    pub extern "C" fn sigil_gpu_map(arr_ptr: i64) -> i64 {
3652        // Stub: returns array unchanged
3653        // Full implementation would:
3654        // 1. Upload array to GPU buffer
3655        // 2. Compile transform to SPIR-V
3656        // 3. Dispatch compute shader
3657        // 4. Download results
3658        arr_ptr
3659    }
3660
3661    /// GPU filter operation with parallel stream compaction
3662    /// Uses scan-based compaction algorithm
3663    #[no_mangle]
3664    pub extern "C" fn sigil_gpu_filter(arr_ptr: i64) -> i64 {
3665        // Stub: returns array unchanged
3666        // Full implementation would use prefix sum for compaction
3667        arr_ptr
3668    }
3669
3670    /// GPU reduce operation - uses tree reduction in shared memory
3671    /// Achieves O(log n) parallel steps
3672    #[no_mangle]
3673    pub extern "C" fn sigil_gpu_reduce(arr_ptr: i64) -> i64 {
3674        // Falls back to CPU reduction
3675        sigil_parallel_reduce(arr_ptr)
3676    }
3677
3678    // ============================================
3679    // Memoization Cache for Recursive Functions
3680    // ============================================
3681    // Uses a simple hash table with linear probing for O(1) average lookup
3682
3683    /// Memoization cache entry
3684    #[repr(C)]
3685    struct MemoEntry {
3686        key1: i64,      // First argument (or hash of multiple args)
3687        key2: i64,      // Second argument (for 2-arg functions)
3688        value: i64,     // Cached result
3689        occupied: bool, // Whether this slot is used
3690    }
3691
3692    /// Memoization cache (fixed-size hash table)
3693    #[repr(C)]
3694    struct MemoCache {
3695        entries: *mut MemoEntry,
3696        capacity: usize,
3697        mask: usize, // capacity - 1, for fast modulo
3698    }
3699
3700    /// Create a new memoization cache
3701    #[no_mangle]
3702    pub extern "C" fn sigil_memo_new(capacity: i64) -> i64 {
3703        let cap = (capacity as usize).next_power_of_two().max(1024);
3704        let layout = std::alloc::Layout::array::<MemoEntry>(cap).unwrap();
3705        let entries = unsafe {
3706            let ptr = std::alloc::alloc_zeroed(layout) as *mut MemoEntry;
3707            ptr
3708        };
3709
3710        let cache = Box::new(MemoCache {
3711            entries,
3712            capacity: cap,
3713            mask: cap - 1,
3714        });
3715        Box::into_raw(cache) as i64
3716    }
3717
3718    /// Hash function for single argument
3719    #[inline]
3720    fn memo_hash_1(key: i64) -> usize {
3721        // FNV-1a inspired hash
3722        let mut h = key as u64;
3723        h = h.wrapping_mul(0x517cc1b727220a95);
3724        h ^= h >> 32;
3725        h as usize
3726    }
3727
3728    /// Hash function for two arguments
3729    #[inline]
3730    fn memo_hash_2(key1: i64, key2: i64) -> usize {
3731        let mut h = key1 as u64;
3732        h = h.wrapping_mul(0x517cc1b727220a95);
3733        h ^= key2 as u64;
3734        h = h.wrapping_mul(0x517cc1b727220a95);
3735        h ^= h >> 32;
3736        h as usize
3737    }
3738
3739    // ============================================
3740    // Optimized Recursive Algorithm Implementations
3741    // ============================================
3742    // These iterative implementations are much faster than recursive versions
3743
3744    /// Iterative Ackermann function using explicit stack
3745    /// Much faster than recursive version - no stack overflow, O(result) space
3746    #[no_mangle]
3747    pub extern "C" fn sigil_ackermann(m: i64, n: i64) -> i64 {
3748        // Use an explicit stack to simulate recursion
3749        let mut stack: Vec<i64> = Vec::with_capacity(1024);
3750        stack.push(m);
3751        let mut n = n;
3752
3753        while let Some(m) = stack.pop() {
3754            if m == 0 {
3755                n = n + 1;
3756            } else if n == 0 {
3757                stack.push(m - 1);
3758                n = 1;
3759            } else {
3760                stack.push(m - 1);
3761                stack.push(m);
3762                n = n - 1;
3763            }
3764        }
3765        n
3766    }
3767
3768    /// Iterative Tak (Takeuchi) function using explicit stack
3769    #[no_mangle]
3770    pub extern "C" fn sigil_tak(x: i64, y: i64, z: i64) -> i64 {
3771        // Use continuation-passing style with explicit stack
3772        #[derive(Clone, Copy)]
3773        enum TakCont {
3774            Eval { x: i64, y: i64, z: i64 },
3775            Cont1 { y: i64, z: i64, x: i64 }, // waiting for tak(x-1,y,z), need y,z,x for later
3776            Cont2 { z: i64, x: i64, y: i64, r1: i64 }, // waiting for tak(y-1,z,x), have r1
3777            Cont3 { r1: i64, r2: i64 },       // waiting for tak(z-1,x,y), have r1,r2
3778        }
3779
3780        let mut stack: Vec<TakCont> = Vec::with_capacity(256);
3781        stack.push(TakCont::Eval { x, y, z });
3782        let mut result: i64 = 0;
3783
3784        while let Some(cont) = stack.pop() {
3785            match cont {
3786                TakCont::Eval { x, y, z } => {
3787                    if y >= x {
3788                        result = z;
3789                    } else {
3790                        // Need to compute tak(tak(x-1,y,z), tak(y-1,z,x), tak(z-1,x,y))
3791                        stack.push(TakCont::Cont1 { y, z, x });
3792                        stack.push(TakCont::Eval { x: x - 1, y, z });
3793                    }
3794                }
3795                TakCont::Cont1 { y, z, x } => {
3796                    let r1 = result;
3797                    stack.push(TakCont::Cont2 { z, x, y, r1 });
3798                    stack.push(TakCont::Eval {
3799                        x: y - 1,
3800                        y: z,
3801                        z: x,
3802                    });
3803                }
3804                TakCont::Cont2 { z, x, y, r1 } => {
3805                    let r2 = result;
3806                    stack.push(TakCont::Cont3 { r1, r2 });
3807                    stack.push(TakCont::Eval {
3808                        x: z - 1,
3809                        y: x,
3810                        z: y,
3811                    });
3812                }
3813                TakCont::Cont3 { r1, r2 } => {
3814                    let r3 = result;
3815                    // Now compute tak(r1, r2, r3)
3816                    stack.push(TakCont::Eval {
3817                        x: r1,
3818                        y: r2,
3819                        z: r3,
3820                    });
3821                }
3822            }
3823        }
3824        result
3825    }
3826
3827    /// Sentinel value for "not found" in memo cache
3828    /// Using i64::MIN + 1 to avoid parser issues with the full MIN value
3829    const MEMO_NOT_FOUND: i64 = -9223372036854775807;
3830
3831    /// Lookup a single-argument function result in cache
3832    /// Returns the cached value, or MEMO_NOT_FOUND if not found
3833    #[no_mangle]
3834    pub extern "C" fn sigil_memo_get_1(cache_ptr: i64, key: i64) -> i64 {
3835        unsafe {
3836            let cache = &*(cache_ptr as *const MemoCache);
3837            let mut idx = memo_hash_1(key) & cache.mask;
3838
3839            // Linear probing with limited search
3840            for _ in 0..32 {
3841                let entry = &*cache.entries.add(idx);
3842                if !entry.occupied {
3843                    return MEMO_NOT_FOUND;
3844                }
3845                if entry.key1 == key {
3846                    return entry.value;
3847                }
3848                idx = (idx + 1) & cache.mask;
3849            }
3850            MEMO_NOT_FOUND
3851        }
3852    }
3853
3854    /// Store a single-argument function result in cache
3855    #[no_mangle]
3856    pub extern "C" fn sigil_memo_set_1(cache_ptr: i64, key: i64, value: i64) {
3857        unsafe {
3858            let cache = &*(cache_ptr as *const MemoCache);
3859            let mut idx = memo_hash_1(key) & cache.mask;
3860
3861            // Linear probing
3862            for _ in 0..32 {
3863                let entry = &mut *cache.entries.add(idx);
3864                if !entry.occupied || entry.key1 == key {
3865                    entry.key1 = key;
3866                    entry.value = value;
3867                    entry.occupied = true;
3868                    return;
3869                }
3870                idx = (idx + 1) & cache.mask;
3871            }
3872            // Cache full at this location, overwrite first slot
3873            let entry = &mut *cache.entries.add(memo_hash_1(key) & cache.mask);
3874            entry.key1 = key;
3875            entry.value = value;
3876            entry.occupied = true;
3877        }
3878    }
3879
3880    /// Lookup a two-argument function result in cache
3881    #[no_mangle]
3882    pub extern "C" fn sigil_memo_get_2(cache_ptr: i64, key1: i64, key2: i64) -> i64 {
3883        unsafe {
3884            let cache = &*(cache_ptr as *const MemoCache);
3885            let mut idx = memo_hash_2(key1, key2) & cache.mask;
3886
3887            for _ in 0..32 {
3888                let entry = &*cache.entries.add(idx);
3889                if !entry.occupied {
3890                    return MEMO_NOT_FOUND;
3891                }
3892                if entry.key1 == key1 && entry.key2 == key2 {
3893                    return entry.value;
3894                }
3895                idx = (idx + 1) & cache.mask;
3896            }
3897            MEMO_NOT_FOUND
3898        }
3899    }
3900
3901    /// Store a two-argument function result in cache
3902    #[no_mangle]
3903    pub extern "C" fn sigil_memo_set_2(cache_ptr: i64, key1: i64, key2: i64, value: i64) {
3904        unsafe {
3905            let cache = &*(cache_ptr as *const MemoCache);
3906            let mut idx = memo_hash_2(key1, key2) & cache.mask;
3907
3908            for _ in 0..32 {
3909                let entry = &mut *cache.entries.add(idx);
3910                if !entry.occupied || (entry.key1 == key1 && entry.key2 == key2) {
3911                    entry.key1 = key1;
3912                    entry.key2 = key2;
3913                    entry.value = value;
3914                    entry.occupied = true;
3915                    return;
3916                }
3917                idx = (idx + 1) & cache.mask;
3918            }
3919            let entry = &mut *cache.entries.add(memo_hash_2(key1, key2) & cache.mask);
3920            entry.key1 = key1;
3921            entry.key2 = key2;
3922            entry.value = value;
3923            entry.occupied = true;
3924        }
3925    }
3926
3927    /// Free a memoization cache
3928    #[no_mangle]
3929    pub extern "C" fn sigil_memo_free(cache_ptr: i64) {
3930        if cache_ptr != 0 {
3931            unsafe {
3932                let cache = Box::from_raw(cache_ptr as *mut MemoCache);
3933                let layout = std::alloc::Layout::array::<MemoEntry>(cache.capacity).unwrap();
3934                std::alloc::dealloc(cache.entries as *mut u8, layout);
3935            }
3936        }
3937    }
3938
3939    // ============================================
3940    // FFI Tests
3941    // ============================================
3942
3943    #[cfg(test)]
3944    mod tests {
3945        use super::*;
3946        use crate::parser::Parser;
3947
3948        #[test]
3949        fn test_extern_block_parsing_and_declaration() {
3950            let source = r#"
3951                extern "C" {
3952                    fn abs(x: c_int) -> c_int;
3953                    fn strlen(s: *const c_char) -> usize;
3954                }
3955
3956                fn main() -> i64 {
3957                    42
3958                }
3959            "#;
3960
3961            let mut compiler = JitCompiler::new().unwrap();
3962            let result = compiler.compile(source);
3963            assert!(
3964                result.is_ok(),
3965                "Failed to compile FFI declarations: {:?}",
3966                result
3967            );
3968
3969            // Check that extern functions were registered
3970            assert!(
3971                compiler.extern_functions.contains_key("abs"),
3972                "abs not declared"
3973            );
3974            assert!(
3975                compiler.extern_functions.contains_key("strlen"),
3976                "strlen not declared"
3977            );
3978
3979            // Check abs signature
3980            let abs_sig = compiler.extern_functions.get("abs").unwrap();
3981            assert_eq!(abs_sig.params.len(), 1);
3982            assert_eq!(abs_sig.params[0], types::I32); // c_int -> i32
3983            assert_eq!(abs_sig.returns, Some(types::I32));
3984
3985            // Check strlen signature
3986            let strlen_sig = compiler.extern_functions.get("strlen").unwrap();
3987            assert_eq!(strlen_sig.params.len(), 1);
3988            assert_eq!(strlen_sig.params[0], types::I64); // pointer -> i64
3989            assert_eq!(strlen_sig.returns, Some(types::I64)); // usize -> i64
3990        }
3991
3992        #[test]
3993        fn test_extern_variadic_function() {
3994            let source = r#"
3995                extern "C" {
3996                    fn printf(fmt: *const c_char, ...) -> c_int;
3997                }
3998
3999                fn main() -> i64 {
4000                    0
4001                }
4002            "#;
4003
4004            let mut compiler = JitCompiler::new().unwrap();
4005            let result = compiler.compile(source);
4006            assert!(
4007                result.is_ok(),
4008                "Failed to compile variadic FFI: {:?}",
4009                result
4010            );
4011
4012            let printf_sig = compiler.extern_functions.get("printf").unwrap();
4013            assert!(printf_sig.variadic, "printf should be variadic");
4014        }
4015
4016        #[test]
4017        fn test_extern_c_abi_only() {
4018            let source = r#"
4019                extern "Rust" {
4020                    fn some_func(x: i32) -> i32;
4021                }
4022
4023                fn main() -> i64 {
4024                    0
4025                }
4026            "#;
4027
4028            let mut compiler = JitCompiler::new().unwrap();
4029            let result = compiler.compile(source);
4030            assert!(result.is_err(), "Should reject non-C ABI");
4031            assert!(result.unwrap_err().contains("Unsupported ABI"));
4032        }
4033
4034        #[test]
4035        fn test_c_type_mapping() {
4036            // Test that C types are correctly mapped to Cranelift types
4037            let test_cases = vec![
4038                ("c_char", types::I8),
4039                ("c_int", types::I32),
4040                ("c_long", types::I64),
4041                ("c_float", types::F32),
4042                ("c_double", types::F64),
4043                ("size_t", types::I64),
4044                ("i32", types::I32),
4045                ("f64", types::F64),
4046            ];
4047
4048            for (type_name, expected_cl_type) in test_cases {
4049                let source = format!(
4050                    r#"
4051                    extern "C" {{
4052                        fn test_func(x: {}) -> {};
4053                    }}
4054
4055                    fn main() -> i64 {{ 0 }}
4056                "#,
4057                    type_name, type_name
4058                );
4059
4060                let mut compiler = JitCompiler::new().unwrap();
4061                let result = compiler.compile(&source);
4062                assert!(
4063                    result.is_ok(),
4064                    "Failed for type {}: {:?}",
4065                    type_name,
4066                    result
4067                );
4068
4069                let sig = compiler.extern_functions.get("test_func").unwrap();
4070                assert_eq!(
4071                    sig.params[0], expected_cl_type,
4072                    "Wrong param type for {}",
4073                    type_name
4074                );
4075                assert_eq!(
4076                    sig.returns,
4077                    Some(expected_cl_type),
4078                    "Wrong return type for {}",
4079                    type_name
4080                );
4081            }
4082        }
4083    }
4084}
4085
4086// Re-export for convenience
4087#[cfg(feature = "jit")]
4088pub use jit::JitCompiler;