fix(codegen): hoist loop-body allocas to the function entry block

src/codegen/mod.rs

@@ -262,6 +262,36 @@ impl<'ctx> CodeGenerator<'ctx> {
         }
     }
 
+    /// Allocate a stack slot in the function's entry block, regardless of where
+    /// the builder is currently positioned.
+    ///
+    /// `alloca` re-executes every time control reaches it, so an alloca emitted
+    /// inside a loop body allocates fresh stack on each iteration and is never
+    /// reclaimed until the function returns — the stack grows without bound.
+    /// mem2reg/SROA only seed promotion from the entry block, so such allocas
+    /// also survive optimization. Hoisting every alloca to the entry block (the
+    /// standard "static alloca" frontend pattern) makes it a single fixed slot
+    /// that promotes reliably at every opt level. Uses a throwaway builder so
+    /// the caller's insertion point is left untouched.
+    pub(crate) fn entry_block_alloca(
+        &self,
+        function: FunctionValue<'ctx>,
+        llvm_ty: BasicTypeEnum<'ctx>,
+        name: &str,
+    ) -> crate::error::Result<PointerValue<'ctx>> {
+        let builder = self.context.create_builder();
+        let entry = function.get_first_basic_block().ok_or_else(|| {
+            CompileError::codegen_error("function has no entry block for alloca".to_string())
+        })?;
+        match entry.get_first_instruction() {
+            Some(first) => builder.position_before(&first),
+            None => builder.position_at_end(entry),
+        }
+        builder
+            .build_alloca(llvm_ty, name)
+            .map_err(|e| CompileError::codegen_error(e.to_string()))
+    }
+
     pub(crate) fn type_alignment(ty: &Type) -> u32 {
         match ty {
             Type::Bool | Type::I8 | Type::U8 => 1,

src/codegen/statements.rs

@@ -96,10 +96,10 @@ impl<'ctx> CodeGenerator<'ctx> {
                 let declared = Self::resolve_annotation(ty);
                 self.validate_type_for_target(&declared)?;
                 let llvm_ty = self.llvm_type(&declared);
-                let alloca = self
-                    .builder
-                    .build_alloca(llvm_ty, name)
-                    .map_err(|e| CompileError::codegen_error(e.to_string()))?;
+                // Hoist to the entry block so a `let` inside a loop body does
+                // not allocate fresh stack on every iteration. See
+                // CodeGenerator::entry_block_alloca.
+                let alloca = self.entry_block_alloca(function, llvm_ty, name)?;
                 let val = self.compile_expr_typed(value, Some(&declared), function)?;
                 self.builder
                     .build_store(alloca, val)
@@ -287,21 +287,8 @@ impl<'ctx> CodeGenerator<'ctx> {
                 // Alloca in function entry block (not loop block) to avoid
                 // stack growth at O0 where mem2reg does not run.
                 let i32_type = self.context.i32_type();
-                let fn_entry = function.get_first_basic_block().unwrap();
-                let alloca = if let Some(first_instr) = fn_entry.get_first_instruction() {
-                    self.builder.position_before(&first_instr);
-                    let a = self
-                        .builder
-                        .build_alloca(i32_type, var)
-                        .map_err(|e| CompileError::codegen_error(e.to_string()))?;
-                    self.builder.position_at_end(cond_bb);
-                    a
-                } else {
-                    self.builder.position_at_end(cond_bb);
-                    self.builder
-                        .build_alloca(i32_type, var)
-                        .map_err(|e| CompileError::codegen_error(e.to_string()))?
-                };
+                let alloca = self.entry_block_alloca(function, i32_type.into(), var)?;
+                self.builder.position_at_end(cond_bb);
 
                 // Condition block with phi node
                 let phi = self

src/codegen/structs.rs

@@ -20,10 +20,9 @@ impl<'ctx> CodeGenerator<'ctx> {
         let field_map = self.struct_fields.get(name).cloned().ok_or_else(|| {
             CompileError::codegen_error(format!("unknown struct fields for '{name}'"))
         })?;
-        let alloca = self
-            .builder
-            .build_alloca(struct_type, "struct_tmp")
-            .map_err(|e| CompileError::codegen_error(e.to_string()))?;
+        // Entry-block alloca so a struct literal inside a loop body does not
+        // leak stack per iteration. See CodeGenerator::entry_block_alloca.
+        let alloca = self.entry_block_alloca(function, struct_type.into(), "struct_tmp")?;
 
         for (field_name, field_expr) in fields {
             let (idx, field_type) = field_map

tests/loop_alloca_hoisting.rs

@@ -0,0 +1,94 @@
+// Regression: `let` bindings inside a loop body must alloca in the function
+// entry block, not the loop body. An alloca in a loop body is a dynamic stack
+// allocation that re-executes every iteration and is never reclaimed until the
+// function returns, so the stack grows per iteration. mem2reg/SROA only seed
+// promotion from the entry block, so loop-body allocas also survive past -O0.
+//
+// This is the root cause of the eacompute stack-overflow found via Olorin's
+// log_level_scan.ea: a SIMD loop body with ~50 u8x16 `let` bindings overflowed
+// the 8 MB main-thread stack on >=~1 MB inputs (~8 bytes of stack per input
+// byte). The frontend emits unoptimized IR here, so the assertion is exact and
+// opt-level independent.
+
+#[cfg(feature = "llvm")]
+mod common;
+
+#[cfg(feature = "llvm")]
+mod tests {
+    use super::common::*;
+
+    /// Assert that every `alloca` in `ir` lives in a basic block named `entry`.
+    /// Tracks the current block label while scanning the IR text.
+    fn assert_all_allocas_in_entry(ir: &str) {
+        let mut current_block = String::new();
+        for line in ir.lines() {
+            let trimmed = line.trim();
+            // Basic-block label lines look like `while_body71:` or
+            // `while_body71:    ; preds = ...`. Strip any `; ...` comment first
+            // (the `preds = ...` comment contains `=`), then match `<label>:`
+            // with nothing but whitespace before the colon.
+            let code = trimmed.split(';').next().unwrap_or("").trim();
+            if let Some(label) = code.strip_suffix(':') {
+                let is_label = !label.is_empty()
+                    && label
+                        .chars()
+                        .all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '.');
+                if is_label {
+                    current_block = label.to_string();
+                    continue;
+                }
+            }
+            if trimmed.contains("= alloca") {
+                assert_eq!(
+                    current_block, "entry",
+                    "alloca emitted in block `{current_block}`, expected `entry`:\n  {trimmed}\n\nFull IR:\n{ir}"
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn let_bindings_in_while_body_alloca_in_entry() {
+        // Mirrors the log_level_scan.ea pattern: many vector `let`s per chunk.
+        let source = r#"
+            export func k(text: *u8, n: i32, out: *mut i32) {
+                let mut i: i32 = 0
+                while i + 16 <= n {
+                    let a: u8x16 = load(text, i)
+                    let b: u8x16 = load(text, i + 1)
+                    let c: u8x16 = a .& b
+                    let d: u8x16 = a .| b
+                    let e: u8x16 = select(a .== b, c, d)
+                    let f: u8x16 = c .& d
+                    out[0] = out[0] + to_i32(reduce_add(e))
+                    out[1] = out[1] + to_i32(reduce_add(f))
+                    i = i + 16
+                }
+            }
+        "#;
+        let ir = compile_to_ir(source);
+        assert_all_allocas_in_entry(&ir);
+    }
+
+    #[test]
+    fn nested_while_let_bindings_alloca_in_entry() {
+        // A `let` inside a nested loop is the worst case: per-iteration growth
+        // multiplied across both loop trip counts.
+        let source = r#"
+            export func k(text: *u8, n: i32, out: *mut i32) {
+                let mut i: i32 = 0
+                while i + 16 <= n {
+                    let mut j: i32 = 0
+                    while j < 4 {
+                        let v: u8x16 = load(text, i)
+                        out[0] = out[0] + to_i32(reduce_add(v))
+                        j = j + 1
+                    }
+                    i = i + 16
+                }
+            }
+        "#;
+        let ir = compile_to_ir(source);
+        assert_all_allocas_in_entry(&ir);
+    }
+}