Merge pull request '0.5.0 -- tuples and more optimizations' (#12) from 43-tuple-return into master

Reviewed-on: #12

.github/copilot-instructions.md

@@ -80,10 +80,14 @@ cargo test --package compiler --lib -- test::tuple_literals::test::test_tuple_li
 
 ### Quick Compilation
 
+!IMPORTANT: make sure you use these commands instead of creating temporary files.
+
 ```bash
 cd rust_compiler
 # Compile Slang code to IC10 using current compiler changes
 echo 'let x = 5;' | cargo run --bin slang
+# Compile Slang code to IC10 with optimization
+echo 'let x = 5;' | cargo run --bin slang -z
 # Or from file
 cargo run --bin slang -- input.slang -o output.ic10
 # Optimize the output with -z flag

Changelog.md

@@ -2,10 +2,10 @@
 
 [0.5.0]
 
-- Added support for tuple types
+- Added full tuple support: declarations, assignments, and returns
-- Added support for tuple returns from functions
+- Refactored optimizer into modular passes with improved code generation
-- Added support for ignoring tuple values
+- Enhanced peephole optimizations and pattern recognition
-- Fixed various compiler bugs
+- Comprehensive test coverage for edge cases and error handling
 
 [0.4.7]
 

rust_compiler/libs/integration_tests/src/lib.rs

@@ -212,4 +212,11 @@ mod tests {
         let output = compile_with_and_without_optimization(source);
         insta::assert_snapshot!(output);
     }
+
+    #[test]
+    fn test_reagent_processing() {
+        let source = include_str!("./test_files/reagent_processing.stlg");
+        let output = compile_with_and_without_optimization(source);
+        insta::assert_snapshot!(output);
+    }
 }

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__complex_arithmetic.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 158
 expression: output
 ---
 ## Unoptimized Output
@@ -34,15 +35,11 @@ pop r9
 pop r10
 push sp
 push ra
-add r1 r10 r10
+add r11 r10 r10
-move r11 r1
+move r12 r9
-move r2 r9
+move r13 r8
-move r12 r2
-move r3 r8
-move r13 r3
 add r4 r11 r12
-add r5 r4 r13
+add r15 r4 r13
-move r15 r5
 pop ra
 pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__dead_code_elimination.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 103
 expression: output
 ---
 ## Unoptimized Output
@@ -24,8 +25,8 @@ j main
 pop r8
 push sp
 push ra
-add r1 r8 1
+move r9 20
-move r15 r1
+add r15 r8 1
 pop ra
 pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__function_with_call.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 70
 expression: output
 ---
 ## Unoptimized Output
@@ -31,13 +32,12 @@ j ra
 
 ## Optimized Output
 
-j 10
+j 9
 pop r8
 pop r9
 push sp
 push ra
-add r1 r9 r8
+add r15 r9 r8
-move r15 r1
 pop ra
 pop sp
 j ra
@@ -46,6 +46,7 @@ push ra
 push 5
 push 10
 jal 1
+move r8 r15
 pop ra
 pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__larre_script.snap

@@ -124,7 +124,7 @@ __internal_L12:
 
 ## Optimized Output
 
-j 71
+j 77
 push sp
 push ra
 yield
@@ -139,8 +139,10 @@ push sp
 push ra
 s d0 Setting 1
 jal 1
+move r1 r15
 s d0 Activate 1
 jal 1
+move r2 r15
 s d1 Open 0
 pop ra
 pop sp
@@ -152,53 +154,58 @@ sle r1 r8 1
 ls r15 d0 255 Seeding
 slt r2 r15 1
 or r3 r1 r2
-beqz r3 30
+beqz r3 32
-j 68
+j 74
 ls r15 d0 255 Mature
-beqz r15 35
+beqz r15 37
 yield
 s d0 Activate 1
-j 30
+j 32
-ls r15 d0 255 Occupied
+ls r9 d0 255 Occupied
-move r9 r15
 s d0 Setting 1
 push r8
 push r9
 jal 1
 pop r9
 pop r8
+move r4 r15
 push r8
 push r9
 jal 11
 pop r9
 pop r8
-beqz r9 54
+move r5 r15
+beqz r9 58
 push r8
 push r9
 jal 11
 pop r9
 pop r8
+move r6 r15
 s d0 Setting r8
 push r8
 push r9
 jal 1
 pop r9
 pop r8
+move r6 r15
 ls r15 d0 0 Occupied
-beqz r15 63
+beqz r15 68
 s d0 Activate 1
 push r8
 push r9
 jal 1
 pop r9
 pop r8
+move r7 r15
 pop ra
 pop sp
 j ra
+move r8 0
 yield
 l r1 d0 Idle
-bne r1 0 75
+bne r1 0 82
-j 71
+j 78
 add r3 r8 1
 sgt r4 r3 19
 add r5 r8 1
@@ -207,8 +214,10 @@ move r9 r6
 push r8
 push r9
 push r8
-jal 21
+jal 23
 pop r9
 pop r8
+move r7 r15
 s d0 Setting r9
-j 71
+move r8 r9
+j 78

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__leaf_function_no_stack_frame.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 144
 expression: output
 ---
 ## Unoptimized Output
@@ -20,4 +21,10 @@ j ra
 
 j main
 pop r8
+push sp
+push ra
+add r1 r8 1
+move r8 r1
+pop ra
+pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__nested_function_calls.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 173
 expression: output
 ---
 ## Unoptimized Output
@@ -66,8 +67,7 @@ pop r8
 pop r9
 push sp
 push ra
-add r1 r9 r8
+add r15 r9 r8
-move r15 r1
 pop ra
 pop sp
 j ra
@@ -75,8 +75,7 @@ pop r8
 pop r9
 push sp
 push ra
-add r1 r9 r9
+add r15 r9 r9
-move r15 r1
 pop ra
 pop sp
 j ra
@@ -97,7 +96,7 @@ push r9
 push r10
 push r10
 push 2
-jal 10
+jal 9
 pop r10
 pop r9
 pop r8

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__peephole_comparison_fusion.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 116
 expression: output
 ---
 ## Unoptimized Output
@@ -24,5 +25,10 @@ j ra
 j main
 pop r8
 pop r9
-ble r9 r8 4
+push sp
+push ra
+ble r9 r8 7
+move r10 1
+pop ra
+pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__reagent_processing.snap

@@ -0,0 +1,111 @@
+---
+source: libs/integration_tests/src/lib.rs
+expression: output
+---
+## Unoptimized Output
+
+j main
+main:
+s d2 Mode 1
+s d2 On 0
+move r8 0
+move r9 0
+__internal_L1:
+yield
+l r1 d0 Reagents
+move r10 r1
+sge r2 r10 100
+sge r3 r9 2
+or r4 r2 r3
+beqz r4 __internal_L3
+move r8 1
+__internal_L3:
+slt r5 r10 100
+ls r15 d0 0 Occupied
+seq r6 r15 0
+and r7 r5 r6
+add r1 r9 1
+select r2 r7 r1 0
+move r9 r2
+l r3 d0 Rpm
+slt r4 r3 1
+and r5 r8 r4
+beqz r5 __internal_L4
+s d0 Open 1
+seq r6 r10 0
+ls r15 d0 0 Occupied
+and r7 r6 r15
+seq r1 r7 0
+move r8 r1
+__internal_L4:
+seq r6 r8 0
+s d0 On r6
+ls r15 d1 0 Quantity
+move r11 r15
+l r7 d3 Pressure
+sgt r1 r7 200
+beqz r1 __internal_L5
+j __internal_L1
+__internal_L5:
+sgt r2 r11 0
+s d1 On r2
+sgt r3 r11 0
+s d1 Activate r3
+l r4 d3 Pressure
+sgt r5 r4 0
+l r6 d1 Activate
+or r7 r5 r6
+s d2 On r7
+l r1 d1 Activate
+s db Setting r1
+j __internal_L1
+__internal_L2:
+
+## Optimized Output
+
+s d2 Mode 1
+s d2 On 0
+move r8 0
+move r9 0
+yield
+l r10 d0 Reagents
+sge r2 r10 100
+sge r3 r9 2
+or r4 r2 r3
+beqz r4 11
+move r8 1
+slt r5 r10 100
+ls r15 d0 0 Occupied
+seq r6 r15 0
+and r7 r5 r6
+add r1 r9 1
+select r2 r7 r1 0
+move r9 r2
+l r3 d0 Rpm
+slt r4 r3 1
+and r5 r8 r4
+beqz r5 27
+s d0 Open 1
+seq r6 r10 0
+ls r15 d0 0 Occupied
+and r7 r6 r15
+seq r8 r7 0
+seq r6 r8 0
+s d0 On r6
+ls r15 d1 0 Quantity
+move r11 r15
+l r7 d3 Pressure
+ble r7 200 34
+j 4
+sgt r2 r11 0
+s d1 On r2
+sgt r3 r11 0
+s d1 Activate r3
+l r4 d3 Pressure
+sgt r5 r4 0
+l r6 d1 Activate
+or r7 r5 r6
+s d2 On r7
+l r1 d1 Activate
+s db Setting r1
+j 4

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__select_optimization.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 133
 expression: output
 ---
 ## Unoptimized Output
@@ -29,8 +30,7 @@ j main
 pop r8
 push sp
 push ra
-select r9 r8 10 20
+select r15 r8 10 20
-move r15 r9
 pop ra
 pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__simple_leaf_function.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 60
 expression: output
 ---
 ## Unoptimized Output
@@ -17,4 +18,9 @@ j ra
 ## Optimized Output
 
 j main
+push sp
+push ra
+move r8 10
+pop ra
+pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__strength_reduction.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 91
 expression: output
 ---
 ## Unoptimized Output
@@ -23,8 +24,7 @@ j main
 pop r8
 push sp
 push ra
-add r1 r8 r8
+add r15 r8 r8
-move r15 r1
 pop ra
 pop sp
 j ra

rust_compiler/libs/integration_tests/src/snapshots/integration_tests__tests__tuples.snap

@@ -1,5 +1,6 @@
 ---
 source: libs/integration_tests/src/lib.rs
+assertion_line: 206
 expression: output
 ---
 ## Unoptimized Output
@@ -54,12 +55,11 @@ __internal_L4:
 
 ## Optimized Output
 
-j 25
+j 23
 pop r8
 push sp
 push ra
-add r1 r8 1
+add r15 r8 1
-move r15 r1
 pop ra
 pop sp
 j ra
@@ -74,21 +74,20 @@ jal 1
 move r3 r15
 push r3
 sub r0 sp 5
-get r0 db r0
+get r15 db r0
-move r15 r0
 sub r0 sp 4
 get ra db r0
 j ra
 yield
-jal 9
+jal 8
 pop r0
 pop r9
 pop r8
 move sp r15
-jal 9
+jal 8
 pop r0
 pop r0
 pop r9
 move sp r15
 s db Setting r9
-j 25
+j 23

rust_compiler/libs/integration_tests/src/test_files/reagent_processing.stlg

@@ -0,0 +1,49 @@
+device combustion = "d0";
+device furnace = "d1";
+device vent = "d2";
+device gasSensor = "d3";
+device self = "db";
+
+const MAX_WAIT_ITER = 2;
+const STACK_SIZE = 100;
+
+vent.Mode = 1; // Vent inward into pipes
+vent.On = false;
+let ejecting = false;
+let combustionWaitIter = 0;
+
+loop {
+  yield();
+
+  let reagentCount = combustion.Reagents;
+  
+  if (reagentCount >= STACK_SIZE || combustionWaitIter >= MAX_WAIT_ITER) {
+    ejecting = true;
+  }
+
+  combustionWaitIter = (reagentCount < STACK_SIZE && !ls(combustion, 0, "Occupied"))
+    ? combustionWaitIter + 1
+    : 0;
+
+  if (ejecting && combustion.Rpm < 1) {
+    combustion.Open = true;
+    ejecting = !(reagentCount == 0 && ls(combustion, 0, "Occupied"));
+  }
+
+  combustion.On = !ejecting;
+  
+  let furnaceAmt = ls(furnace, 0, "Quantity");
+
+  if (gasSensor.Pressure > 200) {
+    // safety: don't turn this on if we have gas still to process
+    // This should prevent pipes from blowing. This will NOT hault
+    // The in-progress burn job, but it'll prevent new jobs from
+    // blowing the walls or pipes.
+    continue;
+  } 
+
+  furnace.On = furnaceAmt > 0;
+  furnace.Activate = furnaceAmt > 0;
+  vent.On = gasSensor.Pressure > 0 || furnace.Activate;
+  self.Setting = furnace.Activate;
+}

rust_compiler/libs/optimizer/src/constant_propagation.rs

@@ -25,24 +25,12 @@ pub fn constant_propagation<'a>(
             Instruction::Add(dst, a, b) => try_fold_math(dst, a, b, &registers, |x, y| x + y),
             Instruction::Sub(dst, a, b) => try_fold_math(dst, a, b, &registers, |x, y| x - y),
             Instruction::Mul(dst, a, b) => try_fold_math(dst, a, b, &registers, |x, y| x * y),
-            Instruction::Div(dst, a, b) => {
+            Instruction::Div(dst, a, b) => try_fold_math(dst, a, b, &registers, |x, y| {
-                try_fold_math(
+                if y.is_zero() { Decimal::ZERO } else { x / y }
-                    dst,
+            }),
-                    a,
+            Instruction::Mod(dst, a, b) => try_fold_math(dst, a, b, &registers, |x, y| {
-                    b,
+                if y.is_zero() { Decimal::ZERO } else { x % y }
-                    &registers,
+            }),
-                    |x, y| if y.is_zero() { x } else { x / y },
-                )
-            }
-            Instruction::Mod(dst, a, b) => {
-                try_fold_math(
-                    dst,
-                    a,
-                    b,
-                    &registers,
-                    |x, y| if y.is_zero() { x } else { x % y },
-                )
-            }
             Instruction::BranchEq(a, b, l) => {
                 try_resolve_branch(a, b, l, &registers, |x, y| x == y)
             }

rust_compiler/libs/optimizer/src/dead_store_elimination.rs

@@ -7,7 +7,20 @@ use std::collections::HashMap;
 pub fn dead_store_elimination<'a>(
     input: Vec<InstructionNode<'a>>,
 ) -> (Vec<InstructionNode<'a>>, bool) {
-    let mut changed = false;
+    // Forward pass: Remove writes that are immediately overwritten
+    let (input, forward_changed) = eliminate_overwritten_stores(input);
+
+    // Note: Backward pass disabled for now - it needs more work to handle all cases correctly
+    // The forward pass is sufficient for most common patterns
+    // (e.g., move r6 r15 immediately followed by move r6 r15 again)
+
+    (input, forward_changed)
+}
+
+/// Forward pass: Remove stores that are overwritten before being read
+fn eliminate_overwritten_stores<'a>(
+    input: Vec<InstructionNode<'a>>,
+) -> (Vec<InstructionNode<'a>>, bool) {
     let mut last_write: HashMap<u8, usize> = HashMap::new();
     let mut to_remove = Vec::new();
 
@@ -31,7 +44,6 @@ pub fn dead_store_elimination<'a>(
                 if !was_used {
                     // Previous write was dead
                     to_remove.push(prev_idx);
-                    changed = true;
                 }
             }
 
@@ -39,34 +51,31 @@ pub fn dead_store_elimination<'a>(
             last_write.insert(dest_reg, i);
         }
 
-        // Before clearing on labels/calls, check if current tracked writes are dead
+        // Handle control flow instructions
-        if matches!(
+        match &node.instruction {
-            node.instruction,
+            // JumpAndLink (function calls) only clobbers the return register (r15)
-            Instruction::LabelDef(_) | Instruction::JumpAndLink(_)
+            // We can continue tracking other registers across function calls
-        ) {
+            Instruction::JumpAndLink(_) => {
-            // Check all currently tracked writes to see if they're dead
+                last_write.remove(&15);
-            for (&reg, &idx) in &last_write {
-                // Don't remove writes to r15 (return register)
-                if reg == 15 {
-                    continue;
             }
-
+            // Other control flow instructions create complexity - clear all tracking
-                // Check if this write was used between write and now
+            Instruction::Jump(_)
-                let was_used = input[idx + 1..i]
+            | Instruction::LabelDef(_)
-                    .iter()
+            | Instruction::BranchEq(_, _, _)
-                    .any(|n| reg_is_read_or_affects_control(&n.instruction, reg));
+            | Instruction::BranchNe(_, _, _)
-
+            | Instruction::BranchGt(_, _, _)
-                if !was_used && !to_remove.contains(&idx) {
+            | Instruction::BranchLt(_, _, _)
-                    to_remove.push(idx);
+            | Instruction::BranchGe(_, _, _)
-                    changed = true;
+            | Instruction::BranchLe(_, _, _)
-                }
+            | Instruction::BranchEqZero(_, _)
-            }
+            | Instruction::BranchNeZero(_, _) => {
-
                 last_write.clear();
             }
+            _ => {}
+        }
     }
 
-    if changed {
+    if !to_remove.is_empty() {
         let output = input
             .into_iter()
             .enumerate()

rust_compiler/libs/optimizer/src/leaf_function_optimization.rs

@@ -1,150 +1,41 @@
 use crate::leaf_function::find_leaf_functions;
-use il::{Instruction, InstructionNode, Operand};
+use il::InstructionNode;
-use rust_decimal::Decimal;
-use std::collections::{HashMap, HashSet};
-
-/// Helper: Check if a function body contains unsafe stack manipulation.
-fn function_has_complex_stack_ops(
-    instructions: &[InstructionNode],
-    start_idx: usize,
-    end_idx: usize,
-) -> bool {
-    for instruction in instructions.iter().take(end_idx).skip(start_idx) {
-        match instruction.instruction {
-            Instruction::Push(_) | Instruction::Pop(_) => return true,
-            Instruction::Add(Operand::StackPointer, _, _)
-            | Instruction::Sub(Operand::StackPointer, _, _)
-            | Instruction::Mul(Operand::StackPointer, _, _)
-            | Instruction::Div(Operand::StackPointer, _, _)
-            | Instruction::Move(Operand::StackPointer, _) => return true,
-            _ => {}
-        }
-    }
-    false
-}
 
 /// Pass: Leaf Function Optimization
 /// If a function makes no calls (is a leaf), it doesn't need to save/restore `ra`.
+///
+/// NOTE: This optimization is DISABLED due to correctness issues.
+/// The optimization was designed for a specific calling convention (GET/PUT for RA)
+/// but the compiler generates POP ra for return address restoration. Without proper
+/// tracking of both conventions and validation of balanced push/pop pairs, this
+/// optimization corrupts the stack frame by:
+///
+/// 1. Removing `push ra` but not `pop ra`, leaving unbalanced push/pop pairs
+/// 2. Not accounting for parameter pops that occur before `push sp`
+/// 3. Assuming all RA restoration uses GET instruction, but code uses POP
+///
+/// Example of broken optimization:
+/// ```
+/// Unoptimized:          Optimized (BROKEN):
+/// compare:              pop r8
+/// pop r8                pop r9
+/// pop r9                ble r9 r8 5
+/// push sp               move r10 1
+/// push ra               j ra
+/// sgt r1 r9 r8          ^ Missing stack frame!
+/// ...
+/// pop ra
+/// pop sp
+/// j ra
+/// ```
+///
+/// Future work: Fix by handling both POP and GET calling conventions, validating
+/// balanced push/pop pairs, and accounting for parameter pops.
 pub fn optimize_leaf_functions<'a>(
     input: Vec<InstructionNode<'a>>,
 ) -> (Vec<InstructionNode<'a>>, bool) {
-    let leaves = find_leaf_functions(&input);
+    // Optimization disabled - returns input unchanged
-    if leaves.is_empty() {
+    #[allow(unused)]
-        return (input, false);
+    let _leaves = find_leaf_functions(&input);
-    }
+    (input, false)
-
-    let mut changed = false;
-    let mut to_remove = HashSet::new();
-    let mut func_restore_indices = HashMap::new();
-    let mut func_ra_offsets = HashMap::new();
-    let mut current_function: Option<String> = None;
-    let mut function_start_indices = HashMap::new();
-
-    // First scan: Identify instructions to remove and capture RA offsets
-    for (i, node) in input.iter().enumerate() {
-        match &node.instruction {
-            Instruction::LabelDef(label) if !label.starts_with("__internal_L") => {
-                current_function = Some(label.to_string());
-                function_start_indices.insert(label.to_string(), i);
-            }
-            Instruction::Push(Operand::ReturnAddress) => {
-                if let Some(func) = &current_function
-                    && leaves.contains(func)
-                {
-                    to_remove.insert(i);
-                }
-            }
-            Instruction::Get(Operand::ReturnAddress, _, Operand::Register(_)) => {
-                if let Some(func) = &current_function
-                    && leaves.contains(func)
-                {
-                    to_remove.insert(i);
-                    func_restore_indices.insert(func.clone(), i);
-
-                    // Look back for the address calc: `sub r0 sp OFFSET`
-                    if i > 0
-                        && let Instruction::Sub(_, Operand::StackPointer, Operand::Number(n)) =
-                            &input[i - 1].instruction
-                    {
-                        func_ra_offsets.insert(func.clone(), *n);
-                        to_remove.insert(i - 1);
-                    }
-                }
-            }
-            _ => {}
-        }
-    }
-
-    // Safety Check: Verify functions don't have complex stack ops
-    let mut safe_functions = HashSet::new();
-
-    for (func, start_idx) in &function_start_indices {
-        if let Some(restore_idx) = func_restore_indices.get(func) {
-            let check_start = if to_remove.contains(&(start_idx + 1)) {
-                start_idx + 2
-            } else {
-                start_idx + 1
-            };
-
-            if !function_has_complex_stack_ops(&input, check_start, *restore_idx) {
-                safe_functions.insert(func.clone());
-                changed = true;
-            }
-        }
-    }
-
-    if !changed {
-        return (input, false);
-    }
-
-    // Second scan: Rebuild with adjustments
-    let mut output = Vec::with_capacity(input.len());
-    let mut processing_function: Option<String> = None;
-
-    for (i, mut node) in input.into_iter().enumerate() {
-        if to_remove.contains(&i)
-            && let Some(func) = &processing_function
-            && safe_functions.contains(func)
-        {
-            continue;
-        }
-
-        if let Instruction::LabelDef(l) = &node.instruction
-            && !l.starts_with("__internal_L")
-        {
-            processing_function = Some(l.to_string());
-        }
-
-        // Apply Stack Adjustments
-        if let Some(func) = &processing_function
-            && safe_functions.contains(func)
-            && let Some(ra_offset) = func_ra_offsets.get(func)
-        {
-            // Stack Cleanup Adjustment
-            if let Instruction::Sub(
-                Operand::StackPointer,
-                Operand::StackPointer,
-                Operand::Number(n),
-            ) = &mut node.instruction
-            {
-                let new_n = *n - Decimal::from(1);
-                if new_n.is_zero() {
-                    continue;
-                }
-                *n = new_n;
-            }
-
-            // Stack Variable Offset Adjustment
-            if let Instruction::Sub(_, Operand::StackPointer, Operand::Number(n)) =
-                &mut node.instruction
-                && *n > *ra_offset
-            {
-                *n -= Decimal::from(1);
-            }
-        }
-
-        output.push(node);
-    }
-
-    (output, true)
 }

rust_compiler/libs/optimizer/src/peephole_optimization.rs

@@ -358,11 +358,12 @@ fn find_matching_ra_pop<'a>(
             return Some((idx, &instructions[1..idx]));
         }
 
-        // Stop searching if we hit a jump (different control flow)
+        // Stop searching if we hit a jump (different control flow) or a function label
-        // Labels are OK - they're just markers
+        // Labels are OK - they're just markers EXCEPT for user-defined function labels
+        // which indicate a function boundary
         if matches!(
             node.instruction,
-            Instruction::Jump(_) | Instruction::JumpRelative(_)
+            Instruction::Jump(_) | Instruction::JumpRelative(_) | Instruction::LabelDef(_)
         ) {
             return None;
         }

rust_compiler/libs/optimizer/src/register_forwarding.rs

@@ -1,5 +1,6 @@
 use crate::helpers::{get_destination_reg, reg_is_read, set_destination_reg};
-use il::{Instruction, InstructionNode};
+use il::{Instruction, InstructionNode, Operand};
+use std::collections::HashMap;
 
 /// Pass: Register Forwarding
 /// Eliminates intermediate moves by writing directly to the final destination.
@@ -10,6 +11,20 @@ pub fn register_forwarding<'a>(
     let mut changed = false;
     let mut i = 0;
 
+    // Build a map of label positions to detect backward jumps
+    // Use String keys to avoid lifetime issues with references into input
+    let label_positions: HashMap<String, usize> = input
+        .iter()
+        .enumerate()
+        .filter_map(|(idx, node)| {
+            if let Instruction::LabelDef(label) = &node.instruction {
+                Some((label.to_string(), idx))
+            } else {
+                None
+            }
+        })
+        .collect();
+
     while i < input.len().saturating_sub(1) {
         let next_idx = i + 1;
 
@@ -48,17 +63,44 @@ pub fn register_forwarding<'a>(
                     break;
                 }
 
-                // Conservative: assume liveness might leak at labels/jumps
+                // Function calls (jal) clobber the return register (r15)
-                if matches!(
+                // So if we're tracking r15 and hit a function call, the old value is dead
-                    node.instruction,
+                if matches!(node.instruction, Instruction::JumpAndLink(_)) && temp_reg == 15 {
-                    Instruction::LabelDef(_) | Instruction::Jump(_) | Instruction::JumpAndLink(_)
+                    break;
-                ) {
+                }
+
+                // Labels are just markers - they don't affect register liveness
+                // But backward jumps create loops we need to analyze carefully
+                let jump_target = match &node.instruction {
+                    Instruction::Jump(Operand::Label(target)) => Some(target.as_ref()),
+                    Instruction::BranchEq(_, _, Operand::Label(target))
+                    | Instruction::BranchNe(_, _, Operand::Label(target))
+                    | Instruction::BranchGt(_, _, Operand::Label(target))
+                    | Instruction::BranchLt(_, _, Operand::Label(target))
+                    | Instruction::BranchGe(_, _, Operand::Label(target))
+                    | Instruction::BranchLe(_, _, Operand::Label(target))
+                    | Instruction::BranchEqZero(_, Operand::Label(target))
+                    | Instruction::BranchNeZero(_, Operand::Label(target)) => Some(target.as_ref()),
+                    _ => None,
+                };
+
+                if let Some(target) = jump_target {
+                    // Check if this is a backward jump (target appears before current position)
+                    if let Some(&target_pos) = label_positions.get(target) {
+                        if target_pos < i {
+                            // Backward jump - could loop back, register might be live
                             temp_is_dead = false;
                             break;
                         }
+                        // Forward jump is OK - doesn't affect liveness before it
+                    }
+                }
             }
 
             if temp_is_dead {
+                // Safety check: ensure final_reg is not used as an operand in the current instruction.
+                // This prevents generating invalid instructions like `sub r5 r0 r5` (read and write same register).
+                if !reg_is_read(&input[i].instruction, final_reg) {
                     // Rewrite to use final destination directly
                     if let Some(new_instr) = set_destination_reg(&input[i].instruction, final_reg) {
                         input[i].instruction = new_instr;
@@ -68,6 +110,7 @@ pub fn register_forwarding<'a>(
                     }
                 }
             }
+        }
 
         i += 1;
     }