diff --git a/examples/50-smoke.sx b/examples/50-smoke.sx
index 3695cc9..e6c4d8a 100644
--- a/examples/50-smoke.sx
+++ b/examples/50-smoke.sx
@@ -38,6 +38,8 @@ Perms :: enum flags { read; write; execute; }
 
 Status :: enum u8 { ok; err; timeout; }
 
+WindowFlags :: enum flags u32 { vsync :: 64; resizable :: 4; hidden :: 128; }
+
 // --- Top-level functions ---
 
 add :: (a: s32, b: s32) -> s32 { a + b; }
@@ -72,13 +74,22 @@ early_return :: (x: s32) -> s32 {
     x;
 }
 
-// #run compile-time constant
-CT_VAL :: #run add(10, 15);
+vec3 :: (x: f32, y: f32, z: f32) -> Vector(3, f32) {
+    .[x, y, z];
+}
 
-// #insert helper
+// #run compile-time constants
+CT_VAL :: #run add(10, 15);
+CT_MUL :: #run mul(6, 7);
+CT_CHAIN :: #run add(CT_VAL, 5);
+
+// #insert helpers
 gen_code :: () -> string {
     return "print(\"insert-ok\\n\");";
 }
+gen_val :: () -> string {
+    return "print(\"insert-gen: {}\\n\", 42);";
+}
 
 // ============================================================
 main :: {
@@ -128,6 +139,18 @@ END;
     c : Color = .green;
     print("enum-lit: {}\n", c);
 
+    // Null pointer
+    np : *s32 = null;
+    print("null-ptr: {}\n", np);
+
+    // String .len
+    slen := "hello";
+    print("string-len: {}\n", slen.len);
+
+    // Empty string .len
+    es := "";
+    print("empty-string: {}\n", es.len);
+
     // ========================================================
     // 2. OPERATORS & PRECEDENCE
     // ========================================================
@@ -155,16 +178,37 @@ END;
     print("chain-gt: {}\n", 100 > v > 0);
     print("chain-mixed: {}\n", 100 > v >= 0);
 
+    // Equality chains
+    print("eq-chain: {}\n", 5 == 5 == 5);
+    print("eq-chain-f: {}\n", 5 == 5 == 6);
+
     // Bitwise
     print("band: {}\n", 0xFF & 0x0F);
     print("bor: {}\n", 1 | 2 | 4);
 
+    // Bitwise on variables
+    bv1 := 0xFF;
+    bv2 := 0x0F;
+    print("band-var: {}\n", bv1 & bv2);
+    bv3 := 1;
+    bv4 := 6;
+    print("bor-var: {}\n", bv3 | bv4);
+
+    // Modulo on variables
+    mv1 := 17;
+    mv2 := 5;
+    print("mod-var: {}\n", mv1 % mv2);
+
     // Logical (short-circuit)
     print("and: {}\n", true and true);
     print("and-false: {}\n", true and false);
     print("or: {}\n", false or true);
     print("or-false: {}\n", false or false);
 
+    // Short-circuit verification
+    print("short-and: {}\n", false and true);
+    print("short-or: {}\n", true or false);
+
     // Compound assignment
     ca := 10;
     ca += 5;
@@ -185,6 +229,36 @@ END;
     small : u8 = xx big2;
     print("xx-cast: {}\n", small);
 
+    // Implicit widening conversions
+    wu : u8 = 200;
+    ws : s64 = wu;
+    print("widen-u8-s64: {}\n", ws);
+
+    wi3 : s32 = 42;
+    wf : f64 = wi3;
+    print("widen-s32-f64: {}\n", wf);
+
+    wf32 : f32 = 1.5;
+    wf64 : f64 = wf32;
+    print("widen-f32-f64: {}\n", wf64);
+
+    wu2 : u8 = 100;
+    ws2 : s16 = wu2;
+    print("widen-u8-s16: {}\n", ws2);
+
+    // More xx narrowing
+    xl : s64 = 12345;
+    xs : s32 = xx xl;
+    print("xx-s64-s32: {}\n", xs);
+
+    xd : f64 = 1.5;
+    xf : f32 = xx xd;
+    print("xx-f64-f32: {}\n", xf);
+
+    xdf : f64 = 7.9;
+    xdi : s32 = xx xdf;
+    print("xx-f64-s32: {}\n", xdi);
+
     // ========================================================
     // 3. TYPE SYSTEM
     // ========================================================
@@ -241,6 +315,13 @@ END;
     ec : Color = .red;
     print("enum: {}\n", ec);
 
+    // Enum comparison
+    ce1 : Color = .red;
+    ce2 : Color = .red;
+    ce3 : Color = .blue;
+    print("enum-eq: {}\n", ce1 == ce2);
+    print("enum-neq: {}\n", ce1 != ce3);
+
     // Backing type
     st : Status = .err;
     print("backing: {}\n", st);
@@ -257,6 +338,16 @@ END;
     sh = .none;
     print("void-variant: {}\n", sh);
 
+    // Variant reassignment
+    sh = .circle(1.0);
+    print("reassign: {}\n", sh);
+    sh = .rect(.{ 5, 3 });
+    print("reassign2: {}\n", sh);
+
+    // Type-prefix construction
+    tp := Shape.circle(2.5);
+    print("enum-prefix: {}\n", tp);
+
     // Pattern matching
     sh2 : Shape = .rect(.{ 5, 3 });
     if sh2 == {
@@ -274,6 +365,15 @@ END;
     }
     print("match-expr: {}\n", ms);
 
+    // Match expression with else
+    me_val := 42;
+    me_res := if me_val == {
+        case 1: 10;
+        case 2: 20;
+        else: 99;
+    }
+    print("match-expr-else: {}\n", me_res);
+
     // Payload capture (block form)
     sh4 : Shape = .circle(9.5);
     if sh4 == {
@@ -282,6 +382,14 @@ END;
         case .none: print("capture: none\n");
     }
 
+    // Payload capture (arrow form)
+    sh_ca : Shape = .circle(7.5);
+    if sh_ca == {
+        case .circle: (r) => print("capture-arrow: {}\n", r);
+        case .rect: (sz) => print("capture-arrow: rect\n");
+        case .none: print("capture-arrow: none\n");
+    }
+
     // else arm in match
     num := 42;
     if num == {
@@ -298,6 +406,26 @@ END;
         case 3: print("int-match: three\n");
     }
 
+    // Integer match with else
+    im_code := 99;
+    if im_code == {
+        case 1: print("int-match-else: one\n");
+        case 2: print("int-match-else: two\n");
+        else: print("int-match-else: unknown\n");
+    }
+
+    // Bool pattern matching
+    bm := true;
+    if bm == {
+        case true: print("bool-match-t: yes\n");
+        case false: print("bool-match-t: no\n");
+    }
+    bm2 := false;
+    if bm2 == {
+        case true: print("bool-match-f: yes\n");
+        case false: print("bool-match-f: no\n");
+    }
+
     // Bool conditional
     flag := true;
     if flag { print("bool: true\n"); }
@@ -321,11 +449,21 @@ END;
     print("arr[2]: {}\n", arr[2]);
     print("arr.len: {}\n", arr.len);
 
+    // Array element assignment
+    aa : [3]s32 = .[1, 2, 3];
+    aa[1] = 99;
+    print("arr-assign: {}\n", aa);
+
     // --- Slices ---
     sl : []s32 = .[1, 2, 3, 4, 5];
     print("sl[0]: {}\n", sl[0]);
     print("sl.len: {}\n", sl.len);
 
+    // Slice element write
+    sla : []s32 = .[10, 20, 30];
+    sla[1] = 55;
+    print("sl-assign: {}\n", sla);
+
     // Subslicing
     sub := arr[1..4];
     print("sub: {}\n", sub);
@@ -334,9 +472,17 @@ END;
     tail := arr[2..];
     print("tail: {}\n", tail);
 
+    // Slice of slice
+    sos : []s32 = .[10, 20, 30, 40, 50];
+    mid := sos[1..4];
+    inner := mid[0..2];
+    print("slice-of-slice: {}\n", inner);
+
     // String subslicing
     msg := "hello world";
     print("strsub: {}\n", msg[6..11]);
+    print("str-prefix: {}\n", msg[..5]);
+    print("str-suffix: {}\n", msg[6..]);
 
     // --- Pointers ---
     pv := Point.{ 10, 20 };
@@ -351,6 +497,32 @@ END;
     print("mp[0]: {}\n", mp[0]);
     print("mp[3]: {}\n", mp[3]);
 
+    // Many-pointer write
+    mpw : [5]s32 = .[10, 20, 30, 40, 50];
+    mpw_ptr : [*]s32 = @mpw[0];
+    mpw_ptr[2] = 99;
+    print("mp-write: {}\n", mpw[2]);
+
+    // --- Vectors ---
+    vc := vec3(1, 3, 2);
+    print("vec-construct: {}\n", vc);
+
+    va := vec3(1, 2, 3);
+    vb := vec3(4, 5, 6);
+    print("vec-add: {}\n", va + vb);
+    print("vec-sub: {}\n", vec3(5, 5, 5) - vec3(1, 2, 3));
+    print("vec-mul: {}\n", vec3(2, 3, 4) * vec3(1, 2, 3));
+    print("vec-div: {}\n", vec3(10, 9, 8) / vec3(2, 3, 4));
+
+    print("vec-scalar: {}\n", vec3(1, 3, 2) * 2.0);
+    print("vec-neg: {}\n", -vec3(1, 3, 2));
+
+    ve := vec3(10, 20, 30);
+    print("vec-x: {}\n", ve.x);
+    print("vec-y: {}\n", ve.y);
+    print("vec-z: {}\n", ve.z);
+    print("vec-idx: {}\n", ve[1]);
+
     // ========================================================
     // 4. CONTROL FLOW
     // ========================================================
@@ -360,11 +532,38 @@ END;
     ite := if true then 1 else 2;
     print("ite: {}\n", ite);
 
+    // If-then-else both branches
+    ie_a := if true then 10 else 20;
+    ie_b := if false then 10 else 20;
+    print("ite-both: {} {}\n", ie_a, ie_b);
+
     // If block
     if 1 < 2 {
         print("if-block: yes\n");
     }
 
+    // If without else (statement)
+    if false { print("should-not-print\n"); }
+    print("if-no-else: after\n");
+
+    // Nested if
+    nx := 10;
+    if nx > 5 {
+        if nx > 8 {
+            print("nested-if: deep\n");
+        }
+    }
+
+    // If-else-if chain
+    eiv := 2;
+    if eiv == 1 {
+        print("if-else-if: first\n");
+    } else if eiv == 2 {
+        print("if-else-if: second\n");
+    } else {
+        print("if-else-if: other\n");
+    }
+
     // If block as expression
     ibe := 10 + if true { 5; } else { 0; };
     print("if-block-expr: {}\n", ibe);
@@ -374,6 +573,10 @@ END;
     while wi < 5 { wi += 1; }
     print("while: {}\n", wi);
 
+    // While with false condition (never executes)
+    while false { print("should-not-print\n"); }
+    print("while-false: skipped\n");
+
     // While with break
     wb := 0;
     while wb < 100 {
@@ -392,7 +595,44 @@ END;
     }
     print("while-continue: {}\n", wsum);
 
-    // For loop basic (using write like example 19)
+    // While sum 1..10
+    wsum2 := 0;
+    wi2 := 1;
+    while wi2 <= 10 {
+        wsum2 += wi2;
+        wi2 += 1;
+    }
+    print("while-sum: {}\n", wsum2);
+
+    // Nested while
+    nw_outer := 0;
+    nw_count := 0;
+    while nw_outer < 3 {
+        nw_inner := 0;
+        while nw_inner < 3 {
+            nw_count += 1;
+            nw_inner += 1;
+        }
+        nw_outer += 1;
+    }
+    print("nested-while: {}\n", nw_count);
+
+    // Nested while with break in inner
+    nb_outer := 0;
+    nb_icount := 0;
+    while nb_outer < 5 {
+        nb_i := 0;
+        while nb_i < 5 {
+            if nb_i == 1 { break; }
+            nb_i += 1;
+        }
+        nb_icount += nb_i;
+        nb_outer += 1;
+        if nb_outer == 2 { break; }
+    }
+    print("nested-break: {} {}\n", nb_outer, nb_icount);
+
+    // For loop basic
     farr : [4]s32 = .[10, 20, 30, 40];
     write("for:");
     for farr {
@@ -439,6 +679,44 @@ END;
     }
     write("\n");
 
+    // For on slice
+    fsl : []s32 = .[10, 20, 30];
+    write("for-slice:");
+    for fsl {
+        print(" {}", it);
+    }
+    write("\n");
+
+    // For on slice with it_index
+    write("for-slice-idx:");
+    for fsl {
+        print(" {}:{}", it_index, it);
+    }
+    write("\n");
+
+    // Nested for
+    nf_a : [2]s32 = .[0, 1];
+    nf_b : [2]s32 = .[0, 1];
+    write("for-nested:");
+    for nf_a {
+        oa := it;
+        for nf_b {
+            print(" ({},{})", oa, it);
+        }
+    }
+    write("\n");
+
+    // For with break preserving it_index
+    fbi : [5]s32 = .[10, 20, 30, 40, 50];
+    fbi_idx := 0;
+    for fbi {
+        if it == 30 { fbi_idx = it_index; break; }
+    }
+    print("for-break-idx: {}\n", fbi_idx);
+
+    // Multiple print placeholders
+    print("multi: {} {} {}\n", 1, 2, 3);
+
     // ========================================================
     // 5. FUNCTIONS & DECLARATIONS
     // ========================================================
@@ -470,6 +748,7 @@ END;
     // Generic — single param
     print("generic-s32: {}\n", identity(42));
     print("generic-f32: {}\n", identity(1.5));
+    print("generic-bool: {}\n", identity(true));
 
     // Generic — multiple params
     print("generic-multi: {}\n", pair_add(10, 20));
@@ -482,6 +761,13 @@ END;
     halve :: (x: f32) -> f32 => x / 2.0;
     print("lambda-ret: {}\n", halve(10.0));
 
+    // Local function (non-lambda)
+    local_add :: (a: s32, b: s32) -> s32 { a + b; }
+    print("local-fn: {}\n", local_add(3, 4));
+
+    // Nested function calls
+    print("fn-nested: {}\n", add(mul(2, 3), mul(4, 5)));
+
     // Variadic (typed)
     print("varargs: {}\n", typed_sum(1, 2, 3, 4, 5));
 
@@ -509,6 +795,14 @@ END;
     }
     print("outer: {}\n", sv);
 
+    // Shadow with different type
+    st_v := 42;
+    print("shadow-type: {}\n", st_v);
+    {
+        st_v := 3.14;
+        print("shadow-type: {}\n", st_v);
+    }
+
     // Nested scopes (3 levels)
     nv := 1;
     {
@@ -521,6 +815,15 @@ END;
     }
     print("nest1: {}\n", nv);
 
+    // Scope isolation
+    { iso := 100; print("scope-isolate: {}\n", iso); }
+
+    // Reuse name after scope exit
+    sr := 1;
+    print("scope-reuse: {}\n", sr);
+    { sr := 2; print("scope-reuse: {}\n", sr); }
+    print("scope-reuse: {}\n", sr);
+
     // Multiple defers (LIFO order)
     {
         defer print("defer-c\n");
@@ -528,6 +831,14 @@ END;
         defer print("defer-a\n");
     }
 
+    // Four defers
+    {
+        defer print("d1\n");
+        defer print("d2\n");
+        defer print("d3\n");
+        defer print("d4\n");
+    }
+
     // Defer in nested scopes
     {
         defer print("outer-defer\n");
@@ -536,6 +847,12 @@ END;
         }
     }
 
+    // Defer in if block
+    if true {
+        defer print("defer-in-if: deferred\n");
+        print("defer-in-if: body\n");
+    }
+
     // ========================================================
     // 7. BUILT-IN FUNCTIONS
     // ========================================================
@@ -546,10 +863,12 @@ END;
 
     // sqrt
     print("sqrt: {}\n", sqrt(9.0));
+    print("sqrt-f64: {}\n", sqrt(16.0));
 
     // size_of
     print("sizeof-s32: {}\n", size_of(s32));
     print("sizeof-f64: {}\n", size_of(f64));
+    print("sizeof-struct: {}\n", size_of(Point));
 
     // type_of + category matching
     tv := 42;
@@ -560,16 +879,58 @@ END;
         else: print("typeof: other\n");
     }
 
+    // type_of — float
+    tf := 3.14;
+    if type_of(tf) == {
+        case float: print("typeof-float: float\n");
+        else: print("typeof-float: other\n");
+    }
+
+    // type_of — string
+    ts := "hello";
+    if type_of(ts) == {
+        case string: print("typeof-string: string\n");
+        else: print("typeof-string: other\n");
+    }
+
+    // type_of — bool
+    tb := true;
+    if type_of(tb) == {
+        case bool: print("typeof-bool: bool\n");
+        else: print("typeof-bool: other\n");
+    }
+
+    // type_of — struct
+    tst := Point.{ 1, 2 };
+    if type_of(tst) == {
+        case struct: print("typeof-struct: struct\n");
+        else: print("typeof-struct: other\n");
+    }
+
+    // type_of — enum
+    ten : Color = .red;
+    if type_of(ten) == {
+        case enum: print("typeof-enum: enum\n");
+        else: print("typeof-enum: other\n");
+    }
+
     // type_name
     print("typename: {}\n", type_name(Point));
 
-    // field_count
+    // field_count on struct
     print("fieldcount: {}\n", field_count(Point));
 
-    // field_name
+    // field_count on enum
+    print("fieldcount-enum: {}\n", field_count(Color));
+
+    // field_name on struct
     print("fieldname0: {}\n", field_name(Point, 0));
     print("fieldname1: {}\n", field_name(Point, 1));
 
+    // field_name on enum
+    print("fieldname-enum0: {}\n", field_name(Color, 0));
+    print("fieldname-enum2: {}\n", field_name(Color, 2));
+
     // field_value (use any_to_string to avoid sext-on-Any bug)
     fv_pt := Point.{ 11, 22 };
     write("fieldval0: ");
@@ -579,13 +940,21 @@ END;
     write(any_to_string(field_value(fv_pt, 1)));
     write("\n");
 
-    // field_index on enum
+    // field_index on plain enum
     fi_c : Color = .green;
     print("fieldidx: {}\n", field_index(Color, fi_c));
 
+    // field_index on tagged enum
+    fi_sh : Shape = .circle(1.0);
+    print("fieldidx-tagged: {}\n", field_index(Shape, fi_sh));
+    fi_sh2 : Shape = .none;
+    print("fieldidx-tagged2: {}\n", field_index(Shape, fi_sh2));
+
     // cast
     cval : f64 = 3.7;
     print("cast: {}\n", cast(s32) cval);
+    cv2 : s32 = 42;
+    print("cast-int-f64: {}\n", cast(f64) cv2);
 
     // ========================================================
     // 8. COMPILE-TIME
@@ -595,9 +964,18 @@ END;
     // #run constant
     print("run-const: {}\n", CT_VAL);
 
+    // #run with expression
+    print("run-expr: {}\n", CT_MUL);
+
+    // #run chained dependency
+    print("run-chain: {}\n", CT_CHAIN);
+
     // #insert with function
     #insert gen_code();
 
+    // #insert additional
+    #insert gen_val();
+
     // ========================================================
     // 9. FLAGS
     // ========================================================
@@ -611,8 +989,29 @@ END;
     if perm & .read { print("has-read: yes\n"); }
     if perm & .execute { print("has-exec: yes\n"); }
 
+    // Test flag negative
+    pt : Perms = .write;
+    if pt & .read {
+        print("flags-neg: has-read\n");
+    } else {
+        print("flags-neg: no-read\n");
+    }
+
+    // Single flag
+    ps : Perms = .execute;
+    print("flags-single: {}\n", ps);
+
+    // All flags
+    pall : Perms = .read | .write | .execute;
+    print("flags-all: {}\n", pall);
+
     // Cast to int
     print("flags-raw: {}\n", cast(s64) perm);
 
+    // Flags with explicit values
+    wf : WindowFlags = .vsync | .resizable;
+    print("flags-explicit: {}\n", wf);
+    print("flags-explicit-raw: {}\n", cast(s64) wf);
+
     print("=== DONE ===\n");
 }
diff --git a/src/codegen.zig b/src/codegen.zig
index eae543d..2b95dc8 100644
--- a/src/codegen.zig
+++ b/src/codegen.zig
@@ -104,6 +104,12 @@ pub const CodeGen = struct {
     module: c.LLVMModuleRef,
     builder: c.LLVMBuilderRef,
     allocator: std.mem.Allocator,
+    // ORC ThreadSafeContext — wraps the LLVMContext for JIT compatibility
+    ts_context: c.LLVMOrcThreadSafeContextRef = null,
+    // Whether we still own the module (false after JIT takes ownership)
+    module_owned: bool = true,
+    // Cached target machine (created in init, reused by emitToFile)
+    target_machine: c.LLVMTargetMachineRef = null,
 
     // Symbol table: maps variable names to their alloca pointers
     named_values: std.StringHashMap(NamedValue),
@@ -188,6 +194,16 @@ pub const CodeGen = struct {
     function_return_types: std.StringHashMap(Type),
     // Target configuration (triple, cpu, opt level, lib paths, linker)
     target_config: TargetConfig = .{},
+    // Cached primitive LLVM types (initialized once in init(), avoids repeated FFI calls)
+    cached_i1: c.LLVMTypeRef = null,
+    cached_i8: c.LLVMTypeRef = null,
+    cached_i16: c.LLVMTypeRef = null,
+    cached_i32: c.LLVMTypeRef = null,
+    cached_i64: c.LLVMTypeRef = null,
+    cached_f32: c.LLVMTypeRef = null,
+    cached_f64: c.LLVMTypeRef = null,
+    cached_ptr: c.LLVMTypeRef = null,
+    cached_void: c.LLVMTypeRef = null,
 
     const DeferredFn = struct {
         fd: ast.FnDecl,
@@ -311,12 +327,18 @@ pub const CodeGen = struct {
     };
 
     pub fn init(allocator: std.mem.Allocator, module_name: [*:0]const u8, target_config: TargetConfig) CodeGen {
-        const ctx = c.LLVMContextCreate();
+        // Create context via ORC ThreadSafeContext for JIT compatibility
+        const ts_ctx = c.LLVMOrcCreateNewThreadSafeContext();
+        const ctx = c.LLVMOrcThreadSafeContextGetContext(ts_ctx);
         const module = c.LLVMModuleCreateWithNameInContext(module_name, ctx);
         const builder = c.LLVMCreateBuilderInContext(ctx);
 
-        // Initialize LLVM targets and set data layout early so alignment queries work
-        llvm.initAllTargets();
+        // Initialize LLVM targets — native-only when targeting host, all for cross-compilation
+        if (target_config.triple == null) {
+            llvm.initNativeTarget();
+        } else {
+            llvm.initAllTargets();
+        }
 
         const triple_owned = target_config.triple == null;
         const triple = target_config.triple orelse c.LLVMGetDefaultTargetTriple();
@@ -326,8 +348,9 @@ pub const CodeGen = struct {
 
         var target: c.LLVMTargetRef = null;
         var err_msg: [*c]u8 = null;
+        var tm: c.LLVMTargetMachineRef = null;
         if (c.LLVMGetTargetFromTriple(triple, &target, &err_msg) == 0) {
-            const tm = c.LLVMCreateTargetMachine(
+            tm = c.LLVMCreateTargetMachine(
                 target,
                 triple,
                 target_config.getCpu(),
@@ -339,7 +362,6 @@ pub const CodeGen = struct {
             const dl = c.LLVMCreateTargetDataLayout(tm);
             c.LLVMSetModuleDataLayout(module, dl);
             c.LLVMDisposeTargetData(dl);
-            c.LLVMDisposeTargetMachine(tm);
         } else {
             if (err_msg != null) c.LLVMDisposeMessage(err_msg);
         }
@@ -348,6 +370,8 @@ pub const CodeGen = struct {
             .module = module,
             .builder = builder,
             .allocator = allocator,
+            .ts_context = ts_ctx,
+            .target_machine = tm,
             .named_values = std.StringHashMap(NamedValue).init(allocator),
             .type_registry = std.StringHashMap(TypeRegistryEntry).init(allocator),
             .flags_enum_types = std.StringHashMap(void).init(allocator),
@@ -375,6 +399,15 @@ pub const CodeGen = struct {
             .global_mutable_vars = std.StringHashMap(NamedValue).init(allocator),
             .function_return_types = std.StringHashMap(Type).init(allocator),
             .target_config = target_config,
+            .cached_i1 = c.LLVMInt1TypeInContext(ctx),
+            .cached_i8 = c.LLVMInt8TypeInContext(ctx),
+            .cached_i16 = c.LLVMInt16TypeInContext(ctx),
+            .cached_i32 = c.LLVMInt32TypeInContext(ctx),
+            .cached_i64 = c.LLVMInt64TypeInContext(ctx),
+            .cached_f32 = c.LLVMFloatTypeInContext(ctx),
+            .cached_f64 = c.LLVMDoubleTypeInContext(ctx),
+            .cached_ptr = c.LLVMPointerTypeInContext(ctx, 0),
+            .cached_void = c.LLVMVoidTypeInContext(ctx),
         };
     }
 
@@ -396,8 +429,15 @@ pub const CodeGen = struct {
         self.foreign_libraries.deinit(self.allocator);
         self.foreign_fns.deinit();
         c.LLVMDisposeBuilder(self.builder);
-        c.LLVMDisposeModule(self.module);
-        c.LLVMContextDispose(self.context);
+        if (self.target_machine) |tm| c.LLVMDisposeTargetMachine(tm);
+        if (self.module_owned) {
+            c.LLVMDisposeModule(self.module);
+        }
+        if (self.ts_context) |ts_ctx| {
+            c.LLVMOrcDisposeThreadSafeContext(ts_ctx);
+        } else {
+            c.LLVMContextDispose(self.context);
+        }
     }
 
     fn getStructInfo(self: *CodeGen, name: []const u8) !StructInfo {
@@ -510,8 +550,14 @@ pub const CodeGen = struct {
 
     pub fn typeToLLVM(self: *CodeGen, ty: Type) c.LLVMTypeRef {
         return switch (ty) {
-            .signed => |w| c.LLVMIntTypeInContext(self.context, w),
-            .unsigned => |w| c.LLVMIntTypeInContext(self.context, w),
+            .signed, .unsigned => |w| switch (w) {
+                1 => self.cached_i1.?,
+                8 => self.cached_i8.?,
+                16 => self.cached_i16.?,
+                32 => self.cached_i32.?,
+                64 => self.cached_i64.?,
+                else => c.LLVMIntTypeInContext(self.context, w),
+            },
             .f32 => self.f32Type(),
             .f64 => self.f64Type(),
             .void_type => self.voidType(),
@@ -736,15 +782,15 @@ pub const CodeGen = struct {
         return self.buildFatPointer(self.getStringStructType(), ptr, len_val);
     }
 
-    // LLVM type shortcuts
-    fn i1Type(self: *CodeGen) c.LLVMTypeRef { return c.LLVMInt1TypeInContext(self.context); }
-    fn i8Type(self: *CodeGen) c.LLVMTypeRef { return c.LLVMInt8TypeInContext(self.context); }
-    fn i32Type(self: *CodeGen) c.LLVMTypeRef { return c.LLVMInt32TypeInContext(self.context); }
-    fn i64Type(self: *CodeGen) c.LLVMTypeRef { return c.LLVMInt64TypeInContext(self.context); }
-    fn f32Type(self: *CodeGen) c.LLVMTypeRef { return c.LLVMFloatTypeInContext(self.context); }
-    fn f64Type(self: *CodeGen) c.LLVMTypeRef { return c.LLVMDoubleTypeInContext(self.context); }
-    fn ptrType(self: *CodeGen) c.LLVMTypeRef { return c.LLVMPointerTypeInContext(self.context, 0); }
-    fn voidType(self: *CodeGen) c.LLVMTypeRef { return c.LLVMVoidTypeInContext(self.context); }
+    // LLVM type shortcuts (cached — no FFI call)
+    fn i1Type(self: *CodeGen) c.LLVMTypeRef { return self.cached_i1.?; }
+    fn i8Type(self: *CodeGen) c.LLVMTypeRef { return self.cached_i8.?; }
+    fn i32Type(self: *CodeGen) c.LLVMTypeRef { return self.cached_i32.?; }
+    fn i64Type(self: *CodeGen) c.LLVMTypeRef { return self.cached_i64.?; }
+    fn f32Type(self: *CodeGen) c.LLVMTypeRef { return self.cached_f32.?; }
+    fn f64Type(self: *CodeGen) c.LLVMTypeRef { return self.cached_f64.?; }
+    fn ptrType(self: *CodeGen) c.LLVMTypeRef { return self.cached_ptr.?; }
+    fn voidType(self: *CodeGen) c.LLVMTypeRef { return self.cached_void.?; }
 
     fn gepArrayElement(self: *CodeGen, arr_ty: c.LLVMTypeRef, arr_ptr: c.LLVMValueRef, idx: c.LLVMValueRef, name: [*c]const u8) c.LLVMValueRef {
         var indices = [_]c.LLVMValueRef{ self.constInt32(0), idx };
@@ -6206,13 +6252,16 @@ pub const CodeGen = struct {
         // Merge block
         self.positionAt(merge_bb);
 
-        // PHI node if both branches produced values
+        // PHI node if both branches produced values (skip for void type)
         if (then_val != null and else_val != null) {
-            const phi = c.LLVMBuildPhi(self.builder, c.LLVMTypeOf(then_val), "iftmp");
-            var vals = [2]c.LLVMValueRef{ then_val, else_val };
-            var blocks = [2]c.LLVMBasicBlockRef{ then_bb, else_bb };
-            c.LLVMAddIncoming(phi, &vals, &blocks, 2);
-            return phi;
+            const ty = c.LLVMTypeOf(then_val);
+            if (c.LLVMGetTypeKind(ty) != c.LLVMVoidTypeKind) {
+                const phi = c.LLVMBuildPhi(self.builder, ty, "iftmp");
+                var vals = [2]c.LLVMValueRef{ then_val, else_val };
+                var blocks = [2]c.LLVMBasicBlockRef{ then_bb, else_bb };
+                c.LLVMAddIncoming(phi, &vals, &blocks, 2);
+                return phi;
+            }
         }
 
         return null;
@@ -7176,39 +7225,12 @@ pub const CodeGen = struct {
     }
 
     fn emitToFile(self: *CodeGen, output_path: [*:0]const u8, file_type: c.LLVMCodeGenFileType) !void {
-        llvm.initAllTargets();
+        const tm = self.target_machine orelse return self.emitError("no target machine available");
 
-        const cfg = self.target_config;
-        const triple_owned = cfg.triple == null;
-        const triple = cfg.triple orelse c.LLVMGetDefaultTargetTriple();
-        defer if (triple_owned) c.LLVMDisposeMessage(@constCast(triple));
-
-        var target: c.LLVMTargetRef = null;
         var err_msg: [*c]u8 = null;
-
-        if (c.LLVMGetTargetFromTriple(triple, &target, &err_msg) != 0) {
+        if (c.LLVMTargetMachineEmitToFile(tm, self.module, output_path, file_type, &err_msg) != 0) {
             defer c.LLVMDisposeMessage(err_msg);
             const msg = std.mem.span(err_msg);
-            return self.emitErrorFmt("failed to get target: {s}", .{msg});
-        }
-
-        const tm = c.LLVMCreateTargetMachine(
-            target,
-            triple,
-            cfg.getCpu(),
-            cfg.getFeatures(),
-            cfg.opt_level.toLLVM(),
-            c.LLVMRelocPIC,
-            c.LLVMCodeModelDefault,
-        );
-        defer c.LLVMDisposeTargetMachine(tm);
-
-        c.LLVMSetTarget(self.module, triple);
-
-        var err_msg2: [*c]u8 = null;
-        if (c.LLVMTargetMachineEmitToFile(tm, self.module, output_path, file_type, &err_msg2) != 0) {
-            defer c.LLVMDisposeMessage(err_msg2);
-            const msg = std.mem.span(err_msg2);
             return self.emitErrorFmt("failed to emit file: {s}", .{msg});
         }
     }
@@ -7221,6 +7243,75 @@ pub const CodeGen = struct {
         return self.emitToFile(output_path, c.LLVMAssemblyFile);
     }
 
+    /// Emit the module as an object file to a memory buffer.
+    /// Caller owns the returned buffer and must dispose or pass to JIT.
+    pub fn emitObjectToMemory(self: *CodeGen) !c.LLVMMemoryBufferRef {
+        const tm = self.target_machine orelse return self.emitError("no target machine available");
+        var err_msg: [*c]u8 = null;
+        var buf: c.LLVMMemoryBufferRef = null;
+        if (c.LLVMTargetMachineEmitToMemoryBuffer(tm, self.module, c.LLVMObjectFile, &err_msg, &buf) != 0) {
+            if (err_msg != null) {
+                defer c.LLVMDisposeMessage(err_msg);
+                const msg = std.mem.span(err_msg);
+                return self.emitErrorFmt("failed to emit object to memory: {s}", .{msg});
+            }
+            return error.CompileError;
+        }
+        return buf;
+    }
+
+    /// Execute a precompiled object file in-process using LLVM's ORC JIT.
+    /// Takes ownership of obj_buf. Returns the exit code from main().
+    pub fn runJITFromObject(obj_buf: c.LLVMMemoryBufferRef) !u8 {
+        // Create LLJIT with default builder (no custom TM needed — .o is precompiled)
+        var jit: c.LLVMOrcLLJITRef = null;
+        var err = c.LLVMOrcCreateLLJIT(&jit, null);
+        if (err != null) {
+            const msg = c.LLVMGetErrorMessage(err);
+            defer c.LLVMDisposeErrorMessage(msg);
+            std.debug.print("JIT error: {s}\n", .{std.mem.span(msg)});
+            return error.CompileError;
+        }
+        defer _ = c.LLVMOrcDisposeLLJIT(jit);
+
+        // Add process symbols so JIT can find libc (printf, write, etc.)
+        const jd = c.LLVMOrcLLJITGetMainJITDylib(jit);
+        const prefix = c.LLVMOrcLLJITGetGlobalPrefix(jit);
+        var gen: c.LLVMOrcDefinitionGeneratorRef = null;
+        err = c.LLVMOrcCreateDynamicLibrarySearchGeneratorForProcess(&gen, prefix, null, null);
+        if (err != null) {
+            const msg = c.LLVMGetErrorMessage(err);
+            defer c.LLVMDisposeErrorMessage(msg);
+            std.debug.print("JIT symbol gen error: {s}\n", .{std.mem.span(msg)});
+            return error.CompileError;
+        }
+        c.LLVMOrcJITDylibAddGenerator(jd, gen);
+
+        // Add precompiled object file (transfers ownership of obj_buf)
+        err = c.LLVMOrcLLJITAddObjectFile(jit, jd, obj_buf);
+        if (err != null) {
+            const msg = c.LLVMGetErrorMessage(err);
+            defer c.LLVMDisposeErrorMessage(msg);
+            std.debug.print("JIT add object error: {s}\n", .{std.mem.span(msg)});
+            return error.CompileError;
+        }
+
+        // Look up the "main" function
+        var main_addr: c.LLVMOrcExecutorAddress = 0;
+        err = c.LLVMOrcLLJITLookup(jit, &main_addr, "main");
+        if (err != null) {
+            const msg = c.LLVMGetErrorMessage(err);
+            defer c.LLVMDisposeErrorMessage(msg);
+            std.debug.print("JIT lookup error: {s}\n", .{std.mem.span(msg)});
+            return error.CompileError;
+        }
+
+        // Cast to function pointer and call
+        const main_fn: *const fn () callconv(.c) i32 = @ptrFromInt(main_addr);
+        const result = main_fn();
+        return if (result >= 0 and result <= 255) @intCast(result) else 1;
+    }
+
     pub fn link(allocator: std.mem.Allocator, io: std.Io, output_obj: []const u8, output_bin: []const u8, libraries: []const []const u8, target_config: TargetConfig) !void {
         var argv = std.ArrayList([]const u8).empty;
 
diff --git a/src/llvm_api.zig b/src/llvm_api.zig
index 6a2c701..078b04d 100644
--- a/src/llvm_api.zig
+++ b/src/llvm_api.zig
@@ -1,7 +1,7 @@
 pub const c = @cImport({
     @cInclude("llvm-c/Core.h");
     @cInclude("llvm-c/Analysis.h");
-    @cInclude("llvm-c/BitWriter.h");
+
     @cInclude("llvm-c/Target.h");
     @cInclude("llvm-c/TargetMachine.h");
     @cInclude("llvm-c/LLJIT.h");
diff --git a/src/main.zig b/src/main.zig
index 20460eb..0567668 100644
--- a/src/main.zig
+++ b/src/main.zig
@@ -23,6 +23,9 @@ pub fn main(init: std.process.Init) !void {
     var input_path: ?[]const u8 = null;
     var target_config = sx.codegen.TargetConfig{};
     var lib_paths = std.ArrayList([]const u8).empty;
+    var show_timing: bool = false;
+    var explicit_opt: bool = false;
+    var no_cache: bool = false;
 
     var i: usize = 2;
     while (i < args.len) : (i += 1) {
@@ -42,6 +45,7 @@ pub fn main(init: std.process.Init) !void {
                 std.debug.print("error: invalid --opt value '{s}' (expected: none/0, less/1, default/2, aggressive/3)\n", .{args[i]});
                 return;
             };
+            explicit_opt = true;
         } else if (std.mem.eql(u8, arg, "-o")) {
             i += 1;
             if (i >= args.len) { std.debug.print("error: -o requires a value\n", .{}); return; }
@@ -54,6 +58,10 @@ pub fn main(init: std.process.Init) !void {
             i += 1;
             if (i >= args.len) { std.debug.print("error: --sysroot requires a value\n", .{}); return; }
             target_config.sysroot = args[i];
+        } else if (std.mem.eql(u8, arg, "--time")) {
+            show_timing = true;
+        } else if (std.mem.eql(u8, arg, "--no-cache")) {
+            no_cache = true;
         } else if (std.mem.startsWith(u8, arg, "-L")) {
             if (arg.len > 2) {
                 try lib_paths.append(allocator, arg[2..]);
@@ -79,33 +87,84 @@ pub fn main(init: std.process.Init) !void {
 
     if (std.mem.eql(u8, command, "build")) {
         const output_name = target_config.output_path orelse deriveOutputName(path);
-        compile(allocator, io, path, output_name, target_config) catch return;
+        compile(allocator, io, path, output_name, target_config, show_timing, no_cache) catch return;
         std.debug.print("compiled: {s}\n", .{output_name});
     } else if (std.mem.eql(u8, command, "ir")) {
         emitIR(allocator, io, path, target_config) catch return;
     } else if (std.mem.eql(u8, command, "asm")) {
         emitAsm(allocator, io, path, target_config) catch return;
     } else if (std.mem.eql(u8, command, "run")) {
-        const tmp_bin = if (comptime @import("builtin").os.tag == .windows) "sx_run_tmp.exe" else "/tmp/sx_run_tmp";
-        compile(allocator, io, path, tmp_bin, target_config) catch return;
-        defer {
-            std.Io.Dir.deleteFile(.cwd(), io, tmp_bin) catch {};
-        }
-        var child = std.process.spawn(io, .{
-            .argv = &.{tmp_bin},
-        }) catch {
-            std.debug.print("error: failed to run program\n", .{});
+        // Default to -O0 for run (faster compile) unless user explicitly set --opt
+        if (!explicit_opt) target_config.opt_level = .none;
+        var timer = Timing.init(show_timing);
+
+        // Phase A: read + parse + resolveImports (for cache key)
+        timer.mark();
+        const source = readSource(allocator, io, path) catch return;
+        timer.record("read");
+
+        var comp = sx.core.Compilation.init(allocator, io, path, source, target_config);
+        defer comp.deinit();
+
+        timer.mark();
+        comp.parse() catch { comp.renderErrors(); return; };
+        timer.record("parse");
+
+        timer.mark();
+        comp.resolveImports() catch { comp.renderErrors(); return; };
+        timer.record("imports");
+
+        // Cache check — use .o files (precompiled object, skip IR compilation in JIT)
+        // Disable caching for files with top-level #run (side effects lost on cache hit)
+        const root = comp.resolved_root orelse comp.root orelse return;
+        const use_cache = !no_cache and !hasTopLevelRun(root);
+        const key = computeCacheKey(source, &comp.import_sources, target_config);
+        const cache_obj = cachePath(allocator, key, "o") catch return;
+
+        timer.mark();
+        const obj_buf: sx.llvm_api.c.LLVMMemoryBufferRef = blk: {
+            if (use_cache) {
+                // Try loading cached .o from disk
+                var buf: sx.llvm_api.c.LLVMMemoryBufferRef = null;
+                var err_msg: [*c]u8 = null;
+                if (sx.llvm_api.c.LLVMCreateMemoryBufferWithContentsOfFile(cache_obj.ptr, &buf, &err_msg) == 0) {
+                    timer.record("cache");
+                    break :blk buf;
+                }
+                if (err_msg != null) sx.llvm_api.c.LLVMDisposeMessage(err_msg);
+            }
+
+            // Cache MISS — codegen + emit .o to memory (verify skipped: JIT catches errors)
+            comp.generateCode() catch { comp.renderErrors(); return; };
+            timer.record("codegen");
+
+            timer.mark();
+            var cg = &comp.cg.?;
+            const buf = cg.emitObjectToMemory() catch { comp.renderErrors(); return; };
+            timer.record("emit");
+
+            // Save .o to cache (extract data before JIT takes ownership)
+            if (use_cache) {
+                saveObjectToCache(buf, io, cache_obj);
+            }
+
+            break :blk buf;
+        };
+
+        // JIT from precompiled object (relocation only, no IR compilation)
+        sx.llvm_api.initNativeTarget();
+        timer.mark();
+        const exit_code = sx.codegen.CodeGen.runJITFromObject(obj_buf) catch {
+            // JIT failed — fall back to AOT
+            timer.record("jit-fail");
+            runAOT(allocator, io, path, target_config, &timer, no_cache) catch return;
+            timer.printAll();
             return;
         };
-        const term = child.wait(io) catch {
-            std.debug.print("error: program execution failed\n", .{});
-            return;
-        };
-        switch (term) {
-            .exited => |code| if (code != 0) std.process.exit(code),
-            .signal => std.process.exit(1),
-            .stopped, .unknown => std.process.exit(1),
-        }
+        timer.record("jit");
+        timer.printAll();
+
+        if (exit_code != 0) std.process.exit(exit_code);
     } else {
         printUsage();
     }
@@ -138,6 +197,8 @@ fn printUsage() void {
         \\  -L <path>           Library search path (repeatable)
         \\  --linker <cmd>      Linker command (default: cc)
         \\  --sysroot <path>    Sysroot for cross-compilation
+        \\  --no-cache          Disable build caching
+        \\  --time              Show compilation timing breakdown
         \\
     , .{});
 }
@@ -191,30 +252,44 @@ fn readSource(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8)
     return try allocator.dupeZ(u8, source_bytes);
 }
 
-fn compilePipeline(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, target_config: sx.codegen.TargetConfig) !sx.core.Compilation {
+fn compilePipeline(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, target_config: sx.codegen.TargetConfig, timer: *Timing) !sx.core.Compilation {
+    timer.mark();
     const source = try readSource(allocator, io, input_path);
+    timer.record("read");
 
     var comp = sx.core.Compilation.init(allocator, io, input_path, source, target_config);
     errdefer comp.deinit();
 
+    timer.mark();
     comp.parse() catch { comp.renderErrors(); return error.CompileError; };
-    comp.resolveImports() catch { comp.renderErrors(); return error.CompileError; };
-    comp.generateCode() catch { comp.renderErrors(); return error.CompileError; };
+    timer.record("parse");
 
+    timer.mark();
+    comp.resolveImports() catch { comp.renderErrors(); return error.CompileError; };
+    timer.record("imports");
+
+    timer.mark();
+    comp.generateCode() catch { comp.renderErrors(); return error.CompileError; };
+    timer.record("codegen");
+
+    timer.mark();
     var cg = &comp.cg.?;
     cg.verify() catch { comp.renderErrors(); return error.CompileError; };
+    timer.record("verify");
 
     return comp;
 }
 
 fn emitIR(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, target_config: sx.codegen.TargetConfig) !void {
-    var comp = try compilePipeline(allocator, io, input_path, target_config);
+    var timer = Timing.init(false);
+    var comp = try compilePipeline(allocator, io, input_path, target_config, &timer);
     defer comp.deinit();
     comp.cg.?.printIR();
 }
 
 fn emitAsm(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, target_config: sx.codegen.TargetConfig) !void {
-    var comp = try compilePipeline(allocator, io, input_path, target_config);
+    var timer = Timing.init(false);
+    var comp = try compilePipeline(allocator, io, input_path, target_config, &timer);
     defer comp.deinit();
     const asm_path = target_config.output_path orelse blk: {
         const name = deriveOutputName(input_path);
@@ -225,22 +300,236 @@ fn emitAsm(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, tar
     std.debug.print("emitted: {s}\n", .{asm_path});
 }
 
-fn compile(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, output_path: []const u8, target_config: sx.codegen.TargetConfig) !void {
-    var comp = try compilePipeline(allocator, io, input_path, target_config);
+fn compile(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, output_path: []const u8, target_config: sx.codegen.TargetConfig, show_timing: bool, no_cache: bool) !void {
+    var timer = Timing.init(show_timing);
+    try compileWithTimer(allocator, io, input_path, output_path, target_config, &timer, no_cache);
+    timer.printAll();
+}
+
+fn compileWithTimer(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, output_path: []const u8, target_config: sx.codegen.TargetConfig, timer: *Timing, no_cache: bool) !void {
+    // Phase A: read + parse + resolveImports (fast: ~0.5ms)
+    timer.mark();
+    const source = try readSource(allocator, io, input_path);
+    timer.record("read");
+
+    var comp = sx.core.Compilation.init(allocator, io, input_path, source, target_config);
+    errdefer comp.deinit();
     defer comp.deinit();
 
-    var cg = &comp.cg.?;
+    timer.mark();
+    comp.parse() catch { comp.renderErrors(); return error.CompileError; };
+    timer.record("parse");
+
+    timer.mark();
+    comp.resolveImports() catch { comp.renderErrors(); return error.CompileError; };
+    timer.record("imports");
+
+    // Extract library names from AST (needed for linking regardless of cache)
+    const root = comp.resolved_root orelse comp.root orelse return error.CompileError;
+    const libs = try extractLibraries(allocator, root);
 
-    // Emit object file
     const obj_path = try std.fmt.allocPrintSentinel(allocator, "{s}.o", .{output_path}, 0);
-    cg.emitObject(obj_path.ptr) catch { comp.renderErrors(); return error.CompileError; };
+
+    // Cache: compute key and check for cached binary/.o
+    const key = computeCacheKey(source, &comp.import_sources, target_config);
+    const cache_obj = try cachePath(allocator, key, "o");
+    const cache_bin = try cachePath(allocator, key, "bin");
+
+    // Level 1: Try cached binary (skip everything — no codegen, no link)
+    if (!no_cache) bin_cache: {
+        std.Io.Dir.copyFile(.cwd(), cache_bin, .cwd(), output_path, io, .{}) catch break :bin_cache;
+        timer.record("cache");
+        return;
+    }
+
+    // Level 2: Try cached .o (skip codegen+emit, still need link)
+    const used_obj_cache = blk: {
+        if (no_cache) break :blk false;
+        std.Io.Dir.copyFile(.cwd(), cache_obj, .cwd(), obj_path, io, .{}) catch break :blk false;
+        break :blk true;
+    };
+
+    if (used_obj_cache) {
+        timer.record("cache");
+    } else {
+        // Cache MISS — full codegen + emit
+        timer.mark();
+        comp.generateCode() catch { comp.renderErrors(); return error.CompileError; };
+        timer.record("codegen");
+
+        timer.mark();
+        var cg = &comp.cg.?;
+        cg.verify() catch { comp.renderErrors(); return error.CompileError; };
+        timer.record("verify");
+
+        timer.mark();
+        cg.emitObject(obj_path.ptr) catch { comp.renderErrors(); return error.CompileError; };
+        timer.record("emit");
+
+        // Save .o to cache
+        if (!no_cache) {
+            std.Io.Dir.copyFile(.cwd(), obj_path, .cwd(), cache_obj, io, .{ .make_path = true }) catch {};
+        }
+    }
 
     // Link
-    sx.codegen.CodeGen.link(allocator, io, obj_path, output_path, cg.foreign_libraries.items, target_config) catch {
+    timer.mark();
+    sx.codegen.CodeGen.link(allocator, io, obj_path, output_path, libs, target_config) catch {
         std.debug.print("error: linking failed\n", .{});
         return error.CompileError;
     };
+    timer.record("link");
+
+    // Save linked binary to cache
+    if (!no_cache) {
+        std.Io.Dir.copyFile(.cwd(), output_path, .cwd(), cache_bin, io, .{ .make_path = true }) catch {};
+    }
 
     // Clean up object file
     std.Io.Dir.deleteFile(.cwd(), io, obj_path) catch {};
 }
+
+fn runAOT(allocator: std.mem.Allocator, io: std.Io, input_path: []const u8, target_config: sx.codegen.TargetConfig, timer: *Timing, no_cache: bool) !void {
+    const tmp_bin = if (comptime @import("builtin").os.tag == .windows) "sx_run_tmp.exe" else "/tmp/sx_run_tmp";
+    try compileWithTimer(allocator, io, input_path, tmp_bin, target_config, timer, no_cache);
+    defer {
+        std.Io.Dir.deleteFile(.cwd(), io, tmp_bin) catch {};
+    }
+
+    timer.mark();
+    var child = std.process.spawn(io, .{
+        .argv = &.{tmp_bin},
+    }) catch {
+        std.debug.print("error: failed to run program\n", .{});
+        return error.CompileError;
+    };
+    const term = child.wait(io) catch {
+        std.debug.print("error: program execution failed\n", .{});
+        return error.CompileError;
+    };
+    timer.record("exec");
+
+    switch (term) {
+        .exited => |code| if (code != 0) std.process.exit(code),
+        .signal => std.process.exit(1),
+        .stopped, .unknown => std.process.exit(1),
+    }
+}
+
+// --- Cache helpers ---
+
+fn computeCacheKey(source: [:0]const u8, import_sources: *const std.StringHashMap([:0]const u8), target_config: sx.codegen.TargetConfig) u64 {
+    const Wyhash = std.hash.Wyhash;
+    var key = Wyhash.hash(0, source);
+
+    // XOR import hashes for order independence (HashMap iteration is non-deterministic)
+    var import_hash: u64 = 0;
+    var it = import_sources.iterator();
+    while (it.next()) |entry| {
+        var h = Wyhash.hash(0, entry.key_ptr.*);
+        h = Wyhash.hash(h, entry.value_ptr.*);
+        import_hash ^= h;
+    }
+    key = Wyhash.hash(key, std.mem.asBytes(&import_hash));
+
+    // Hash target config fields that affect codegen
+    if (target_config.triple) |t| key = Wyhash.hash(key, std.mem.span(t));
+    if (target_config.cpu) |cp| key = Wyhash.hash(key, std.mem.span(cp));
+    if (target_config.features) |f| key = Wyhash.hash(key, std.mem.span(f));
+    key = Wyhash.hash(key, std.mem.asBytes(&target_config.opt_level));
+
+    return key;
+}
+
+fn cachePath(allocator: std.mem.Allocator, key: u64, ext: []const u8) ![:0]const u8 {
+    return try std.fmt.allocPrintSentinel(allocator, ".sx-cache/{x:0>16}.{s}", .{ key, ext }, 0);
+}
+
+fn saveObjectToCache(obj_buf: sx.llvm_api.c.LLVMMemoryBufferRef, io: std.Io, cache_path: [:0]const u8) void {
+    const c_api = sx.llvm_api.c;
+    const start = c_api.LLVMGetBufferStart(obj_buf);
+    const size = c_api.LLVMGetBufferSize(obj_buf);
+    if (start == null or size == 0) return;
+    const data = @as([*]const u8, @ptrCast(start))[0..size];
+    // Write to temp file, then copy to cache (make_path creates .sx-cache/ if needed)
+    std.Io.Dir.writeFile(.cwd(), io, .{ .sub_path = ".sx-cache-tmp", .data = data }) catch return;
+    std.Io.Dir.copyFile(.cwd(), ".sx-cache-tmp", .cwd(), cache_path, io, .{ .make_path = true }) catch {};
+    std.Io.Dir.deleteFile(.cwd(), io, ".sx-cache-tmp") catch {};
+}
+
+fn hasTopLevelRun(root: *const sx.ast.Node) bool {
+    for (root.data.root.decls) |decl| {
+        if (decl.data == .comptime_expr) return true;
+    }
+    return false;
+}
+
+fn extractLibraries(allocator: std.mem.Allocator, root: *const sx.ast.Node) ![]const []const u8 {
+    var libs = std.ArrayList([]const u8).empty;
+    for (root.data.root.decls) |decl| {
+        switch (decl.data) {
+            .library_decl => |ld| try libs.append(allocator, ld.lib_name),
+            .namespace_decl => |ns| {
+                for (ns.decls) |nd| {
+                    switch (nd.data) {
+                        .library_decl => |ld| try libs.append(allocator, ld.lib_name),
+                        else => {},
+                    }
+                }
+            },
+            else => {},
+        }
+    }
+    return try libs.toOwnedSlice(allocator);
+}
+
+// Simple timing helper — records stage durations and prints a summary table.
+const Timing = struct {
+    const max_entries = 16;
+
+    enabled: bool,
+    names: [max_entries][]const u8,
+    durations_ns: [max_entries]u64,
+    count: usize,
+    last: ?std.time.Instant,
+
+    fn init(enabled: bool) Timing {
+        return .{
+            .enabled = enabled,
+            .names = undefined,
+            .durations_ns = undefined,
+            .count = 0,
+            .last = if (enabled) (std.time.Instant.now() catch null) else null,
+        };
+    }
+
+    fn mark(self: *Timing) void {
+        if (self.enabled) self.last = std.time.Instant.now() catch null;
+    }
+
+    fn record(self: *Timing, name: []const u8) void {
+        if (!self.enabled) return;
+        const now = std.time.Instant.now() catch null;
+        const elapsed_ns: u64 = if (self.last != null and now != null) now.?.since(self.last.?) else 0;
+        if (self.count < max_entries) {
+            self.names[self.count] = name;
+            self.durations_ns[self.count] = elapsed_ns;
+            self.count += 1;
+        }
+        self.last = now;
+    }
+
+    fn printAll(self: *const Timing) void {
+        if (!self.enabled or self.count == 0) return;
+        var total_ns: u64 = 0;
+        for (self.durations_ns[0..self.count]) |d| total_ns += d;
+
+        std.debug.print("\n--- timing ---\n", .{});
+        for (0..self.count) |idx| {
+            const ms = @as(f64, @floatFromInt(self.durations_ns[idx])) / 1_000_000.0;
+            std.debug.print("  {s:<10} {d:>7.1} ms\n", .{ self.names[idx], ms });
+        }
+        const total_ms = @as(f64, @floatFromInt(total_ns)) / 1_000_000.0;
+        std.debug.print("  {s:<10} {d:>7.1} ms\n", .{ "total", total_ms });
+    }
+};
diff --git a/tests/expected/50-smoke.txt b/tests/expected/50-smoke.txt
index 5230f77..c3308cc 100644
--- a/tests/expected/50-smoke.txt
+++ b/tests/expected/50-smoke.txt
@@ -13,6 +13,9 @@ heredoc: raw heredoc
 
 undef-then-set: 77
 enum-lit: .green
+null-ptr: null
+string-len: 5
+empty-string: 0
 === 2. Operators ===
 add: 7
 sub: 7
@@ -29,12 +32,19 @@ ge: true
 chain: true
 chain-gt: true
 chain-mixed: true
+eq-chain: true
+eq-chain-f: false
 band: 15
 bor: 7
+band-var: 15
+bor-var: 7
+mod-var: 2
 and: true
 and-false: false
 or: true
 or-false: false
+short-and: false
+short-or: true
 ca+=: 15
 ca-=: 12
 ca*=: 24
@@ -42,6 +52,13 @@ ca/=: 4
 prec1: 14
 prec2: 20
 xx-cast: 200
+widen-u8-s64: 200
+widen-s32-f64: 42.000000
+widen-f32-f64: 1.500000
+widen-u8-s16: 100
+xx-s64-s32: 12345
+xx-f64-f32: 1.500000
+xx-f64-s32: 7
 === 3. Types ===
 s8: 127
 s16: 32000
@@ -57,14 +74,22 @@ struct-shorthand: Point{x: 5, y: 6}
 defaults: a=0 b=99
 field-assign: Point{x: 42, y: 99}
 enum: .red
+enum-eq: true
+enum-neq: true
 backing: .err
 tagged: .circle(3.140000)
 payload: 3.140000
 void-variant: .none
+reassign: .circle(1.000000)
+reassign2: .rect(Shape.rect{w: 5.000000, h: 3.000000})
+enum-prefix: .circle(2.500000)
 match: rect
 match-expr: 10
+match-expr-else: 99
 capture: 9.500000
+capture-arrow: 7.500000
 else-match: other
+int-match-else: unknown
 bool: true
 union-f: 3.140000
 union-i: 1078523331
@@ -72,29 +97,59 @@ promoted-x: 1.000000
 promoted-data0: 1.000000
 arr[2]: 30
 arr.len: 5
+arr-assign: [1, 99, 0]
 sl[0]: 1
 sl.len: 5
+sl-assign: [10, 55, 0]
 sub: [20, 30, 40]
 head: [10, 20, 30]
 tail: [30, 40, 50]
+slice-of-slice: [20, 30]
 strsub: world
+str-prefix: hello
+str-suffix: world
 deref: Point{x: 10, y: 20}
 auto-deref: 10
 mp[0]: 10
 mp[3]: 40
+mp-write: 99
+vec-construct: [1.000000, 3.000000, 2.000000]
+vec-add: [5.000000, 7.000000, 9.000000]
+vec-sub: [4.000000, 3.000000, 2.000000]
+vec-mul: [2.000000, 6.000000, 12.000000]
+vec-div: [5.000000, 3.000000, 2.000000]
+vec-scalar: [2.000000, 6.000000, 4.000000]
+vec-neg: [-1.000000, -3.000000, -2.000000]
+vec-x: 10.000000
+vec-y: 20.000000
+vec-z: 30.000000
+vec-idx: 20.000000
 === 4. Control Flow ===
 ite: 1
+ite-both: 10 20
 if-block: yes
+if-no-else: after
+nested-if: deep
+if-else-if: second
 if-block-expr: 15
 while: 5
+while-false: skipped
 while-break: 7
 while-continue: 25
+while-sum: 55
+nested-while: 9
+nested-break: 2 2
 for: 10 20 30 40
 for-print: 10 20 30 40
 for-idx: 0 1 2 3
 for-2arg: 10@0 20@1 30@2 40@3
 for-break: 10 20
 for-continue: 10 30 40
+for-slice: 10 20 30
+for-slice-idx: 0:10 1:20 2:30
+for-nested: (0,0) (0,1) (1,0) (1,1)
+for-break-idx: 2
+multi: 1 2 3
 === 5. Functions ===
 const: 42
 typed-const: 3.140000
@@ -105,9 +160,12 @@ early-ret2: 99
 void-return: ok
 generic-s32: 42
 generic-f32: 1.500000
+generic-bool: true
 generic-multi: 30
 lambda: 14
 lambda-ret: 5.000000
+local-fn: 7
+fn-nested: 26
 varargs: 15
 spread: 60
 fp: 7
@@ -116,33 +174,66 @@ fp-apply: 30
 === 6. Scoping ===
 inner: 200
 outer: 100
+shadow-type: 42
+shadow-type: 3.140000
 nest3: 3
 nest2: 2
 nest1: 1
+scope-isolate: 100
+scope-reuse: 1
+scope-reuse: 2
+scope-reuse: 1
 defer-a
 defer-b
 defer-c
+d4
+d3
+d2
+d1
 inner-defer
 outer-defer
+defer-in-if: body
+defer-in-if: deferred
 === 7. Builtins ===
 write-ok
 sqrt: 3.000000
+sqrt-f64: 4.000000
 sizeof-s32: 4
 sizeof-f64: 8
+sizeof-struct: 8
 typeof: int
+typeof-float: float
+typeof-string: string
+typeof-bool: bool
+typeof-struct: struct
+typeof-enum: enum
 typename: Point
 fieldcount: 2
+fieldcount-enum: 3
 fieldname0: x
 fieldname1: y
+fieldname-enum0: red
+fieldname-enum2: blue
 fieldval0: 11
 fieldval1: 22
 fieldidx: 1
+fieldidx-tagged: 0
+fieldidx-tagged2: 2
 cast: 3
+cast-int-f64: 42.000000
 === 8. Comptime ===
 run-const: 25
+run-expr: 42
+run-chain: 30
 insert-ok
+insert-gen: 42
 === 9. Flags ===
 flags: .read | .write
 has-read: yes
+flags-neg: no-read
+flags-single: .execute
+flags-all: .read | .write | .execute
 flags-raw: 3
+flags-explicit: .vsync | .resizable
+flags-explicit-raw: 68
 === DONE ===