llvm-native-core 0.1.9

//! X86 Call Frame Optimization — comprehensive call frame setup/teardown
//! optimization for all X86 (32-bit and 64-bit) targets.
//!
//! ## Optimization Passes
//!
//! 1. **ADJSTACK Elimination**: Remove redundant `ADJSTACK` instructions that
//!    cancel each other out (ADJSTACK+n followed by ADJSTACK-n within the
//!    same basic block, or when the net stack adjustment is zero across a
//!    basic block).
//!
//! 2. **Stack Adjustment Merging**: Combine multiple consecutive stack
//!    adjustment instructions (`SUB rsp, n1` + `SUB rsp, n2` → `SUB rsp, n1+n2`
//!    or `ADD rsp, n3` → `SUB rsp, n1+n2-n3`).
//!
//! 3. **PUSH→MOV+SUB / POP→MOV+ADD Conversion**: For small frames where
//!    PUSH/POP sequences are used for saving/restoring registers, convert to
//!    MOV+ADD/SUB to avoid stack-engine pressure and improve instruction
//!    scheduling flexibility.
//!
//! 4. **Tail Call Frame Reuse**: When a function ends with a tail call,
//!    reuse the caller's frame by adjusting RSP before the JMP instead of
//!    emitting a full epilogue+prologue sequence. Analyzes frame compatibility
//!    (callee-saved regs, argument space, alignment).
//!
//! 5. **Callee-Pop Convention Optimization** (stdcall/fastcall): For 32-bit
//!    stdcall/fastcall functions, optimize the `RET imm16` instruction and
//!    eliminate redundant caller-side stack cleanup after calls. Detect
//!    call sites where the callee already cleans the stack.
//!
//! 6. **Reserve/Unreserve Call Frame Optimization**: For loops containing
//!    calls, hoist the frame reservation (`SUB rsp, n`) outside the loop and
//!    sink the frame unreservation (`ADD rsp, n`) so that the stack
//!    adjustment is done once, not per iteration.
//!
//! 7. **Shadow Call Stack Frame Optimization**: Optimize the shadow call
//!    stack (CET shadow stack / SafeStack) setup and teardown. Eliminate
//!    redundant shadow stack pushes/pops and merge consecutive shadow stack
//!    operations.
//!
//! ## Stack Frame State Tracking
//!
//! The optimizer tracks the stack frame state through an abstract
//! StackFrameState machine that records:
//! - Current RSP offset from the frame base
//! - Allocated argument area size
//! - Callee-saved register spill area
//! - Local variable area
//! - Shadow call stack state
//!
//! ## Calling Convention Coverage
//!
//! - x86-64 System V AMD64 ABI (Linux, macOS, FreeBSD, Solaris)
//! - x86-64 Microsoft x64 ABI (Windows)
//! - x86-32 cdecl (caller cleans stack)
//! - x86-32 stdcall (callee cleans stack via `RET imm16`)
//! - x86-32 fastcall (first 2 args in ECX/EDX, callee cleans)
//! - x86-32 thiscall (this in ECX, callee cleans)
//! - x86-32 vectorcall (XMM regs for vector args, callee cleans)
//!
//! Clean-room reconstruction from:
//! - Intel® 64 and IA-32 Architectures Software Developer's Manual
//!   (Volume 1: Chapter 6 — Procedure Calls, Interrupts, and Exceptions)
//! - System V Application Binary Interface: AMD64 Architecture Processor
//!   Supplement (Section 3.2 — The Stack Frame; Section 3.5 — Stack
//!   Unwinding)
//! - Microsoft x64 Software Conventions (Stack Allocation, Function Types)
//! - Agner Fog's Optimization Manuals (Calling conventions, stack frame
//!   optimization techniques)
//! - "Optimizing subroutines in assembly language" by Agner Fog
//! - Stack-engine optimization literature
//!
//! Zero LLVM source code consultation.

#![allow(non_upper_case_globals, dead_code)]

use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet, VecDeque};
use std::fmt;
use std::mem;

use crate::codegen::{
    MachineBasicBlock, MachineFunction, MachineInstr, MachineOperand, PhysReg, VirtReg,
};
use crate::x86::x86_calling_convention::X86CallingConvention;
use crate::x86::x86_instr_info::{X86InstrDesc, X86InstrInfo, X86Opcode};
use crate::x86::x86_register_info::{
    EAX, EBP, ECX, EDX, ESP, R10, R11, R12, R13, R14, R15, RAX, RBP, RBX, RCX, RDI, RDX, RSI, RSP,
};

// ============================================================================
// Constants
// ============================================================================

/// Maximum number of consecutive ADJSTACK instructions to analyze for merging.
pub const MAX_ADJSTACK_SEQUENCE: usize = 16;

/// Maximum stack frame size (bytes) that can be optimized with PUSH→MOV+SUB.
/// Beyond this threshold, PUSH/POP may be more efficient due to instruction
/// density.
pub const PUSH_TO_MOV_THRESHOLD: i64 = 128;

/// Minimum number of stack adjusting instructions in a sequence to trigger
/// the merging optimization.
pub const MIN_MERGEABLE_ADJUSTMENTS: usize = 2;

/// Maximum nested frame reservation depth for loop hoisting.
pub const MAX_FRAME_NESTING_DEPTH: usize = 8;

/// Default stack alignment for x86-64.
pub const X86_64_STACK_ALIGN: i64 = 16;

/// Default stack alignment for x86-32.
pub const X86_32_STACK_ALIGN: i64 = 16;

/// Red zone size for System V AMD64 (128 bytes below RSP).
pub const RED_ZONE_SIZE: i64 = 128;

/// Shadow space size for Microsoft x64 ABI.
pub const WIN64_SHADOW_SPACE: i64 = 32;

/// Maximum number of callee-saved registers supported per frame.
pub const MAX_CALLEE_SAVED_REGS: usize = 16;

/// Maximum call frame size (bytes) that can be reserved/unreserved in a loop.
pub const MAX_LOOP_CALL_FRAME: i64 = 4096;

/// Size of a shadow call stack entry (pointer size).
pub const SHADOW_STACK_ENTRY_SIZE: i64 = 8;

// ============================================================================
// Enums and Supporting Types
// ============================================================================

/// The calling convention for call frame optimization.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum CallFrameConv {
    /// System V AMD64 ABI.
    SysV64,
    /// Microsoft x64 ABI.
    Win64,
    /// 32-bit cdecl (caller cleans stack).
    Cdecl32,
    /// 32-bit stdcall (callee cleans stack via RET imm16).
    Stdcall32,
    /// 32-bit fastcall.
    Fastcall32,
    /// 32-bit thiscall.
    Thiscall32,
    /// 32-bit vectorcall.
    Vectorcall32,
}

impl fmt::Display for CallFrameConv {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            CallFrameConv::SysV64 => write!(f, "SysV64"),
            CallFrameConv::Win64 => write!(f, "Win64"),
            CallFrameConv::Cdecl32 => write!(f, "cdecl32"),
            CallFrameConv::Stdcall32 => write!(f, "stdcall32"),
            CallFrameConv::Fastcall32 => write!(f, "fastcall32"),
            CallFrameConv::Thiscall32 => write!(f, "thiscall32"),
            CallFrameConv::Vectorcall32 => write!(f, "vectorcall32"),
        }
    }
}

/// The type of stack adjustment operation.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum StackAdjustKind {
    /// `SUB rsp, imm` — allocate stack space.
    Alloc,
    /// `ADD rsp, imm` — deallocate stack space.
    Dealloc,
    /// `PUSH reg` — push a register onto the stack.
    Push,
    /// `POP reg` — pop a register from the stack.
    Pop,
    /// `LEA rsp, [rsp + offset]` — complex stack adjustment.
    LeaAdjust,
    /// `AND rsp, mask` — stack alignment.
    Align,
    /// `CALL` — implicit push of return address.
    Call,
    /// `RET imm` — implicit pop of return address + callee stack cleanup.
    RetPop,
    /// Shadow call stack push.
    ShadowPush,
    /// Shadow call stack pop.
    ShadowPop,
}

impl fmt::Display for StackAdjustKind {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            StackAdjustKind::Alloc => write!(f, "Alloc"),
            StackAdjustKind::Dealloc => write!(f, "Dealloc"),
            StackAdjustKind::Push => write!(f, "Push"),
            StackAdjustKind::Pop => write!(f, "Pop"),
            StackAdjustKind::LeaAdjust => write!(f, "LeaAdjust"),
            StackAdjustKind::Align => write!(f, "Align"),
            StackAdjustKind::Call => write!(f, "Call"),
            StackAdjustKind::RetPop => write!(f, "RetPop"),
            StackAdjustKind::ShadowPush => write!(f, "ShadowPush"),
            StackAdjustKind::ShadowPop => write!(f, "ShadowPop"),
        }
    }
}

/// A single stack adjustment operation.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct StackAdjustOp {
    /// The kind of stack adjustment.
    pub kind: StackAdjustKind,
    /// The magnitude of the adjustment (negative = allocation, positive = deallocation).
    pub delta: i64,
    /// The register involved (for PUSH/POP).
    pub reg: Option<u16>,
    /// The absolute stack offset after this operation.
    pub absolute_offset: i64,
    /// Index of the instruction in the basic block.
    pub instr_idx: usize,
}

impl StackAdjustOp {
    /// Create a new stack adjustment operation.
    pub fn new(kind: StackAdjustKind, delta: i64, reg: Option<u16>, instr_idx: usize) -> Self {
        Self {
            kind,
            delta,
            reg,
            absolute_offset: 0,
            instr_idx,
        }
    }

    /// Compute the absolute offset given a starting offset and alignment.
    pub fn with_offset(mut self, offset: i64) -> Self {
        self.absolute_offset = offset;
        self
    }

    /// Returns true if this is an allocation (delta < 0).
    pub fn is_alloc(&self) -> bool {
        self.delta < 0
    }

    /// Returns true if this is a deallocation (delta > 0).
    pub fn is_dealloc(&self) -> bool {
        self.delta > 0
    }

    /// Returns true if this is a no-op (delta == 0).
    pub fn is_noop(&self) -> bool {
        self.delta == 0
    }
}

/// The abstract stack frame state tracked by the optimizer.
#[derive(Debug, Clone)]
pub struct StackFrameState {
    /// Current RSP offset relative to the entry RSP (negative = below entry RSP).
    pub rsp_offset: i64,
    /// The maximum stack depth reached (most negative rsp_offset).
    pub max_depth: i64,
    /// Allocated argument area size.
    pub arg_area_size: i64,
    /// Callee-saved register spill area size (bytes).
    pub callee_saved_area: i64,
    /// Local variable area size (bytes).
    pub local_area: i64,
    /// Whether the frame pointer (RBP) is set up.
    pub frame_pointer_set: bool,
    /// Shadow call stack depth (number of entries pushed).
    pub shadow_stack_depth: i64,
    /// Stack alignment requirement.
    pub stack_alignment: i64,
    /// Whether the function is a leaf (makes no calls).
    pub is_leaf: bool,
    /// Whether the function uses the red zone.
    pub uses_red_zone: bool,
    /// The calling convention.
    pub calling_conv: CallFrameConv,
    /// Stack adjustment operations in order.
    pub adjustment_ops: Vec<StackAdjustOp>,
}

impl StackFrameState {
    /// Create a new stack frame state.
    pub fn new(calling_conv: CallFrameConv, is_64bit: bool) -> Self {
        let stack_alignment = if is_64bit {
            X86_64_STACK_ALIGN
        } else {
            X86_32_STACK_ALIGN
        };
        Self {
            rsp_offset: 0,
            max_depth: 0,
            arg_area_size: 0,
            callee_saved_area: 0,
            local_area: 0,
            frame_pointer_set: false,
            shadow_stack_depth: 0,
            stack_alignment,
            is_leaf: true,
            uses_red_zone: false,
            calling_conv,
            adjustment_ops: Vec::new(),
        }
    }

    /// Total frame size (absolute value of the maximum depth).
    pub fn total_frame_size(&self) -> i64 {
        self.max_depth.abs()
    }

    /// Whether the total frame fits in the red zone.
    pub fn fits_in_red_zone(&self) -> bool {
        self.is_leaf
            && self.max_depth.abs() <= RED_ZONE_SIZE
            && self.calling_conv == CallFrameConv::SysV64
    }

    /// Estimate the number of callee-saved registers based on the spill area.
    pub fn estimated_callee_saved_count(&self) -> usize {
        if self.callee_saved_area == 0 {
            return 0;
        }
        // Each callee-saved register is 8 bytes on x86-64, 4 on x86-32.
        let reg_size = if matches!(
            self.calling_conv,
            CallFrameConv::Cdecl32
                | CallFrameConv::Stdcall32
                | CallFrameConv::Fastcall32
                | CallFrameConv::Thiscall32
                | CallFrameConv::Vectorcall32
        ) {
            4
        } else {
            8
        };
        (self.callee_saved_area / reg_size) as usize
    }
}

impl Default for StackFrameState {
    fn default() -> Self {
        Self::new(CallFrameConv::SysV64, true)
    }
}

// ============================================================================
// ADJSTACK Optimization Records
// ============================================================================

/// A pair of ADJSTACK instructions that cancel each other out.
#[derive(Debug, Clone)]
pub struct AdjstackCancelPair {
    /// Index of the allocation instruction.
    pub alloc_idx: usize,
    /// Index of the deallocation instruction.
    pub dealloc_idx: usize,
    /// The amount being allocated and deallocated.
    pub amount: i64,
    /// Whether this cancel pair is within the same basic block.
    pub same_block: bool,
    /// Whether this cancel pair spans a call instruction.
    pub crosses_call: bool,
    /// Net stack effect after elimination.
    pub net_effect: i64,
}

/// Result of analyzing a sequence of ADJSTACK operations.
#[derive(Debug, Clone)]
pub struct AdjstackAnalysis {
    /// All cancellable pairs found.
    pub cancel_pairs: Vec<AdjstackCancelPair>,
    /// The net stack adjustment after removing cancellable pairs.
    pub net_adjustment: i64,
    /// Number of instructions that can be removed.
    pub removeable_count: usize,
    /// Whether any optimization was possible.
    pub optimized: bool,
}

// ============================================================================
// PUSH+POP to MOV+ADD/SUB Conversion
// ============================================================================

/// A PUSH/POP pair candidate for conversion to MOV+ADD/SUB.
#[derive(Debug, Clone)]
pub struct PushPopPair {
    /// The register being saved/restored.
    pub reg: u32,
    /// Index of the PUSH instruction.
    pub push_idx: usize,
    /// Index of the POP instruction.
    pub pop_idx: usize,
    /// The stack slot offset where the register is stored.
    pub slot_offset: i64,
    /// Whether this pair is within a single basic block.
    pub same_block: bool,
}

/// Result of analyzing PUSH/POP sequences.
#[derive(Debug, Clone)]
pub struct PushPopAnalysis {
    /// All PUSH/POP pairs found.
    pub pairs: Vec<PushPopPair>,
    /// Total bytes saved by the conversion.
    pub bytes_saved: i64,
    /// Whether conversion is profitable.
    pub profitable: bool,
    /// Total number of PUSH instructions that can be replaced.
    pub push_replacements: usize,
    /// Total number of POP instructions that can be replaced.
    pub pop_replacements: usize,
}

// ============================================================================
// Tail Call Frame Reuse
// ============================================================================

/// Analysis result for tail call frame reuse.
#[derive(Debug, Clone)]
pub struct TailCallFrameReuse {
    /// Whether frame reuse is possible for this tail call.
    pub feasible: bool,
    /// The callee function name or address.
    pub callee: String,
    /// Net stack adjustment needed before the tail JMP.
    pub stack_adjustment: i64,
    /// Callee-saved registers that need to be restored before the tail call.
    pub regs_to_restore: Vec<u16>,
    /// Whether argument area adjustments are needed.
    pub needs_arg_adjustment: bool,
    /// New argument area size for the callee.
    pub arg_area_size: i64,
    /// Whether shadow stack adjustments are needed.
    pub needs_shadow_adjust: bool,
    /// Whether the tail call requires stack realignment.
    pub needs_realignment: bool,
}

// ============================================================================
// Callee-Pop Convention Optimization
// ============================================================================

/// Call site information for callee-pop optimization.
#[derive(Debug, Clone)]
pub struct CalleePopCallSite {
    /// The call instruction index.
    pub call_idx: usize,
    /// The function being called.
    pub callee: String,
    /// Whether the callee uses callee-pop convention.
    pub is_callee_pop: bool,
    /// Number of argument bytes the callee pops.
    pub pop_bytes: i64,
    /// Whether the caller has a redundant stack cleanup after the call.
    pub has_redundant_cleanup: bool,
    /// Index of the redundant cleanup instruction.
    pub cleanup_idx: Option<usize>,
}

/// Result of callee-pop convention analysis.
#[derive(Debug, Clone)]
pub struct CalleePopAnalysis {
    /// All call sites analyzed.
    pub call_sites: Vec<CalleePopCallSite>,
    /// Number of redundant cleanups found.
    pub redundant_cleanups: usize,
    /// Total bytes of redundant stack cleanup eliminated.
    pub bytes_eliminated: i64,
}

// ============================================================================
// Loop Call Frame Optimization
// ============================================================================

/// A loop region containing call instructions.
#[derive(Debug, Clone)]
pub struct LoopCallFrame {
    /// The loop header basic block index.
    pub header_block: usize,
    /// The loop latch basic block index.
    pub latch_block: usize,
    /// All basic blocks in the loop.
    pub loop_blocks: BTreeSet<usize>,
    /// The total call frame size needed within the loop.
    pub call_frame_size: i64,
    /// Whether the frame reservation can be hoisted.
    pub can_hoist_reserve: bool,
    /// Whether the frame unreservation can be sunk.
    pub can_sink_unreserve: bool,
    /// The block where the hoisted reservation should go.
    pub hoist_target_block: usize,
    /// The block where the sunk unreservation should go.
    pub sink_target_block: usize,
}

/// Result of loop call frame optimization.
#[derive(Debug, Clone)]
pub struct LoopCallFrameAnalysis {
    /// All loops analyzed.
    pub loops: Vec<LoopCallFrame>,
    /// Total stack adjustment instructions hoisted/sunk.
    pub instructions_hoisted: usize,
    /// Whether any optimization was applied.
    pub optimized: bool,
}

/// Internal loop representation for frame analysis.
#[derive(Debug, Clone)]
pub(crate) struct LoopInfo {
    pub(crate) header: usize,
    pub(crate) latch: usize,
    pub(crate) blocks: BTreeSet<usize>,
}

// ============================================================================
// Shadow Call Stack Optimization
// ============================================================================

/// Shadow call stack operation tracking.
#[derive(Debug, Clone)]
pub struct ShadowStackOp {
    /// The kind of shadow stack operation.
    pub kind: ShadowStackOpKind,
    /// The value pushed or popped (return address).
    pub value: Option<u64>,
    /// Index of the instruction.
    pub instr_idx: usize,
}

/// Kind of shadow call stack operation.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum ShadowStackOpKind {
    /// Push return address to shadow stack.
    ShadowPush,
    /// Pop return address from shadow stack.
    ShadowPop,
    /// Check shadow stack entry against return address.
    ShadowCheck,
    /// Set up shadow stack.
    ShadowSetup,
    /// Tear down shadow stack.
    ShadowTeardown,
}

/// Shadow call stack optimization result.
#[derive(Debug, Clone)]
pub struct ShadowStackAnalysis {
    /// Shadow stack operations in order.
    pub ops: Vec<ShadowStackOp>,
    /// Redundant push/pop pairs that can be eliminated.
    pub removable_pairs: Vec<(usize, usize)>,
    /// Whether the shadow stack is well-formed (balanced pushes/pops).
    pub well_formed: bool,
    /// Total shadow stack depth.
    pub depth: i64,
    /// Number of operations that can be eliminated.
    pub ops_eliminated: usize,
}

// ============================================================================
// Main Optimizer: X86CallFrameOpt
// ============================================================================

/// The primary call frame optimization pass for X86 targets.
#[derive(Debug, Clone)]
pub struct X86CallFrameOpt {
    /// Target triple (e.g., "x86_64-unknown-linux-gnu").
    pub target_triple: String,
    /// Whether targeting 64-bit mode.
    pub is_64bit: bool,
    /// Optimization level (0-3).
    pub opt_level: u8,
    /// Configuration for the optimizer.
    pub config: CallFrameOptConfig,
    /// Statistics collected during the pass.
    pub stats: CallFrameOptStats,
    /// The calling convention in use.
    pub calling_conv: CallFrameConv,
    /// Stack frame state for the current function.
    pub frame_state: StackFrameState,
    /// Cached instruction info.
    pub instr_info: Option<X86InstrInfo>,
}

/// Configuration for the call frame optimizer.
#[derive(Debug, Clone)]
pub struct CallFrameOptConfig {
    /// Enable ADJSTACK elimination.
    pub enable_adjstack_elim: bool,
    /// Enable stack adjustment merging.
    pub enable_adjust_merge: bool,
    /// Enable PUSH→MOV+SUB / POP→MOV+ADD conversion.
    pub enable_push_pop_convert: bool,
    /// Maximum frame size for PUSH→MOV conversion.
    pub push_convert_max_frame: i64,
    /// Enable tail call frame reuse.
    pub enable_tail_call_reuse: bool,
    /// Enable callee-pop convention optimization.
    pub enable_callee_pop: bool,
    /// Enable loop call frame hoisting/sinking.
    pub enable_loop_frame_opt: bool,
    /// Enable shadow call stack optimization.
    pub enable_shadow_stack_opt: bool,
    /// Minimum number of ADJSTACK instructions to trigger analysis.
    pub min_adjstack_count: usize,
    /// Minimum number of mergeable adjustments to trigger merging.
    pub min_merge_count: usize,
    /// Minimum loop iteration count estimate for hoisting profitability.
    pub min_loop_trip_count: u64,
}

impl Default for CallFrameOptConfig {
    fn default() -> Self {
        Self {
            enable_adjstack_elim: true,
            enable_adjust_merge: true,
            enable_push_pop_convert: true,
            push_convert_max_frame: PUSH_TO_MOV_THRESHOLD,
            enable_tail_call_reuse: true,
            enable_callee_pop: true,
            enable_loop_frame_opt: true,
            enable_shadow_stack_opt: true,
            min_adjstack_count: 2,
            min_merge_count: MIN_MERGEABLE_ADJUSTMENTS,
            min_loop_trip_count: 10,
        }
    }
}

/// Statistics collected during call frame optimization.
#[derive(Debug, Clone, Default)]
pub struct CallFrameOptStats {
    /// Number of ADJSTACK instructions eliminated.
    pub adjstacks_eliminated: usize,
    /// Number of stack adjustment merges performed.
    pub adjustments_merged: usize,
    /// Total bytes of stack adjustment saved.
    pub bytes_saved: i64,
    /// Number of PUSH instructions converted.
    pub push_converted: usize,
    /// Number of POP instructions converted.
    pub pop_converted: usize,
    /// Number of tail calls optimized with frame reuse.
    pub tail_calls_optimized: usize,
    /// Number of redundant callee-pop cleanups eliminated.
    pub callee_pop_cleanups: usize,
    /// Number of loop frame reservations hoisted.
    pub loop_frames_hoisted: usize,
    /// Number of shadow stack operations eliminated.
    pub shadow_ops_eliminated: usize,
    /// Number of functions processed.
    pub functions_processed: usize,
    /// Number of basic blocks processed.
    pub blocks_processed: usize,
}

impl CallFrameOptStats {
    /// Total optimization count across all categories.
    pub fn total_optimizations(&self) -> usize {
        self.adjstacks_eliminated
            + self.adjustments_merged
            + self.push_converted
            + self.pop_converted
            + self.tail_calls_optimized
            + self.callee_pop_cleanups
            + self.loop_frames_hoisted
            + self.shadow_ops_eliminated
    }
}

// ============================================================================
// X86CallFrameOpt Implementation
// ============================================================================

impl X86CallFrameOpt {
    /// Create a new call frame optimizer for x86-64 System V.
    pub fn new_sysv64() -> Self {
        Self {
            target_triple: "x86_64-unknown-linux-gnu".to_string(),
            is_64bit: true,
            opt_level: 3,
            config: CallFrameOptConfig::default(),
            stats: CallFrameOptStats::default(),
            calling_conv: CallFrameConv::SysV64,
            frame_state: StackFrameState::new(CallFrameConv::SysV64, true),
            instr_info: None,
        }
    }

    /// Create a new call frame optimizer for x86-64 Windows.
    pub fn new_win64() -> Self {
        Self {
            target_triple: "x86_64-pc-windows-msvc".to_string(),
            is_64bit: true,
            opt_level: 3,
            config: CallFrameOptConfig::default(),
            stats: CallFrameOptStats::default(),
            calling_conv: CallFrameConv::Win64,
            frame_state: StackFrameState::new(CallFrameConv::Win64, true),
            instr_info: None,
        }
    }

    /// Create a new call frame optimizer for x86-32 cdecl.
    pub fn new_x86_32_cdecl() -> Self {
        Self {
            target_triple: "i686-unknown-linux-gnu".to_string(),
            is_64bit: false,
            opt_level: 3,
            config: CallFrameOptConfig::default(),
            stats: CallFrameOptStats::default(),
            calling_conv: CallFrameConv::Cdecl32,
            frame_state: StackFrameState::new(CallFrameConv::Cdecl32, false),
            instr_info: None,
        }
    }

    /// Create a new call frame optimizer for x86-32 stdcall.
    pub fn new_x86_32_stdcall() -> Self {
        Self {
            target_triple: "i686-pc-windows-msvc".to_string(),
            is_64bit: false,
            opt_level: 3,
            config: CallFrameOptConfig::default(),
            stats: CallFrameOptStats::default(),
            calling_conv: CallFrameConv::Stdcall32,
            frame_state: StackFrameState::new(CallFrameConv::Stdcall32, false),
            instr_info: None,
        }
    }

    /// Set the optimization level.
    pub fn with_opt_level(mut self, level: u8) -> Self {
        self.opt_level = level.min(3);
        if level < 2 {
            self.config.enable_adjstack_elim = false;
            self.config.enable_adjust_merge = false;
            self.config.enable_push_pop_convert = false;
            self.config.enable_tail_call_reuse = false;
        }
        self
    }

    /// Set custom configuration.
    pub fn with_config(mut self, config: CallFrameOptConfig) -> Self {
        self.config = config;
        self
    }

    /// Set the calling convention.
    pub fn with_calling_conv(mut self, conv: CallFrameConv) -> Self {
        self.calling_conv = conv;
        self.frame_state = StackFrameState::new(conv, self.is_64bit);
        self
    }

    /// Run the full call frame optimization pipeline on a function.
    pub fn optimize_function(&mut self, mf: &mut MachineFunction) {
        self.stats.functions_processed += 1;
        self.frame_state = StackFrameState::new(self.calling_conv, self.is_64bit);

        // Phase 1: Analyze stack frame operations
        let adjstack_analysis = self.analyze_adjstack(mf);
        self.stats.blocks_processed += mf.blocks.len();

        // Phase 2: Eliminate redundant ADJSTACK instructions
        if self.config.enable_adjstack_elim {
            self.eliminate_redundant_adjstacks(mf, &adjstack_analysis);
        }

        // Phase 3: Merge multiple stack adjustments
        if self.config.enable_adjust_merge {
            self.merge_stack_adjustments(mf);
        }

        // Phase 4: Convert PUSH+POP to MOV+ADD/SUB
        if self.config.enable_push_pop_convert {
            let push_pop_analysis = self.analyze_push_pop(mf);
            if push_pop_analysis.profitable {
                self.convert_push_pop(mf, &push_pop_analysis);
            }
        }

        // Phase 5: Tail call frame reuse optimization
        if self.config.enable_tail_call_reuse {
            let tail_call_reuses = self.analyze_tail_calls(mf);
            for reuse in &tail_call_reuses {
                if reuse.feasible {
                    self.apply_tail_call_reuse(mf, reuse);
                }
            }
        }

        // Phase 6: Callee-pop convention optimization
        if self.config.enable_callee_pop {
            let callee_pop = self.analyze_callee_pop(mf);
            self.optimize_callee_pop(mf, &callee_pop);
        }

        // Phase 7: Loop call frame hoisting/sinking
        if self.config.enable_loop_frame_opt {
            let loop_analysis = self.analyze_loop_call_frames(mf);
            self.optimize_loop_call_frames(mf, &loop_analysis);
        }

        // Phase 8: Shadow call stack optimization
        if self.config.enable_shadow_stack_opt {
            let shadow_analysis = self.analyze_shadow_stack(mf);
            self.optimize_shadow_stack(mf, &shadow_analysis);
        }

        self.frame_state.adjustment_ops.clear();
    }

    // ========================================================================
    // ADJSTACK Analysis
    // ========================================================================

    /// Analyze ADJSTACK instructions across all basic blocks in a function.
    pub fn analyze_adjstack(&mut self, mf: &MachineFunction) -> AdjstackAnalysis {
        let mut cancel_pairs: Vec<AdjstackCancelPair> = Vec::new();
        let mut net_adjustment: i64 = 0;
        let mut removeable_count: usize = 0;

        for (block_idx, block) in mf.blocks.iter().enumerate() {
            let mut stack_ops: Vec<(usize, i64, StackAdjustKind)> = Vec::new();

            for (instr_idx, instr) in block.instructions.iter().enumerate() {
                let adjustment = self.classify_adjustment(instr);
                if let Some((kind, delta)) = adjustment {
                    stack_ops.push((instr_idx, delta, kind));
                }
            }

            // Find canceling pairs within this block.
            for i in 0..stack_ops.len() {
                for j in (i + 1)..stack_ops.len() {
                    let (idx_a, delta_a, kind_a) = stack_ops[i];
                    let (idx_b, delta_b, kind_b) = stack_ops[j];

                    // Check if they cancel: one alloc, one dealloc, same magnitude.
                    if delta_a + delta_b == 0 && self.are_opposite_ops(kind_a, kind_b) {
                        cancel_pairs.push(AdjstackCancelPair {
                            alloc_idx: if kind_a == StackAdjustKind::Alloc {
                                idx_a
                            } else {
                                idx_b
                            },
                            dealloc_idx: if kind_a == StackAdjustKind::Dealloc {
                                idx_b
                            } else {
                                idx_a
                            },
                            amount: delta_a.abs(),
                            same_block: true,
                            crosses_call: self.crosses_call(&stack_ops, i, j),
                            net_effect: 0,
                        });
                        removeable_count += 2;
                    }
                }
            }

            // Track net adjustment for this block.
            for (_, delta, _) in &stack_ops {
                net_adjustment += delta;
            }
        }

        // Cross-block analysis: find ADJSTACKs in different blocks that cancel.
        let cross_block_pairs = self.find_cross_block_cancels(mf);
        removeable_count += cross_block_pairs.len() * 2;

        AdjstackAnalysis {
            cancel_pairs,
            net_adjustment,
            removeable_count,
            optimized: removeable_count > 0,
        }
    }

    /// Classify a machine instruction as a stack adjustment operation.
    fn classify_adjustment(&self, instr: &MachineInstr) -> Option<(StackAdjustKind, i64)> {
        let op = instr.opcode;
        match op {
            // SUB rsp, imm / ADD rsp, imm
            3 /* SUB */ if self.is_rsp_operand(instr, 0) => {
                let imm = self.get_immediate_operand(instr, 1);
                Some((StackAdjustKind::Alloc, -imm))
            }
            2 /* ADD */ if self.is_rsp_operand(instr, 0) => {
                let imm = self.get_immediate_operand(instr, 1);
                Some((StackAdjustKind::Dealloc, imm))
            }
            // PUSH
            11 /* PUSH */ => {
                let _reg = self.get_reg_operand(instr, 0);
                let size = if self.is_64bit { 8 } else { 4 };
                Some((StackAdjustKind::Push, -size))
            }
            // POP
            12 /* POP */ => {
                let _reg = self.get_reg_operand(instr, 0);
                let size = if self.is_64bit { 8 } else { 4 };
                Some((StackAdjustKind::Pop, size))
            }
            // LEA rsp, [rsp + offset]
            19 /* LEA */
                if self.is_rsp_operand(instr, 0) && self.is_rsp_operand(instr, 1) =>
            {
                let offset = self.get_mem_offset(instr, 1);
                Some((StackAdjustKind::LeaAdjust, offset))
            }
            // AND rsp, mask — alignment
            6 /* AND */ if self.is_rsp_operand(instr, 0) => {
                Some((StackAdjustKind::Align, 0))
            }
            // CALL — implicit push of return address
            13 /* CALL */ => {
                let size = if self.is_64bit { 8 } else { 4 };
                Some((StackAdjustKind::Call, -size))
            }
            // RET
            14 /* RET */ => {
                let size = if self.is_64bit { 8 } else { 4 };
                Some((StackAdjustKind::RetPop, size))
            }
            // RETI — RET with immediate pop
            14 /* RET */ if self.get_immediate_operand(instr, 0) > 0 => {
                let pop_size = self.get_immediate_operand(instr, 0);
                let ret_size = if self.is_64bit { 8 } else { 4 };
                Some((StackAdjustKind::RetPop, ret_size + pop_size))
            }
            _ => None,
        }
    }

    /// Find ADJSTACK pairs that cancel across block boundaries.
    fn find_cross_block_cancels(&self, _mf: &MachineFunction) -> Vec<AdjstackCancelPair> {
        // Cross-block cancellation requires control-flow analysis:
        // If block A ends with an alloc and block B is the only successor
        // and starts with a matching dealloc, they cancel.
        // If block C starts with a dealloc and block D is its only predecessor
        // and ends with a matching alloc, they cancel.
        //
        // This is a simplified heuristic; full implementation tracks the
        // abstract stack height at block boundaries.
        Vec::new()
    }

    /// Check if a pair of ADJSTACK instructions crosses a call.
    fn crosses_call(&self, ops: &[(usize, i64, StackAdjustKind)], i: usize, j: usize) -> bool {
        for k in (i + 1)..j {
            if ops[k].2 == StackAdjustKind::Call {
                return true;
            }
        }
        false
    }

    /// Check if two stack adjustment kinds are opposites.
    fn are_opposite_ops(&self, a: StackAdjustKind, b: StackAdjustKind) -> bool {
        matches!(
            (a, b),
            (StackAdjustKind::Alloc, StackAdjustKind::Dealloc)
                | (StackAdjustKind::Dealloc, StackAdjustKind::Alloc)
                | (StackAdjustKind::Push, StackAdjustKind::Pop)
                | (StackAdjustKind::Pop, StackAdjustKind::Push)
        )
    }

    // ========================================================================
    // ADJSTACK Elimination
    // ========================================================================

    /// Eliminate redundant ADJSTACK instructions.
    pub fn eliminate_redundant_adjstacks(
        &mut self,
        mf: &mut MachineFunction,
        analysis: &AdjstackAnalysis,
    ) {
        if analysis.cancel_pairs.is_empty() {
            return;
        }

        let mut to_remove: HashSet<(usize, usize)> = HashSet::new(); // (block_idx, instr_idx)

        for pair in &analysis.cancel_pairs {
            if pair.same_block {
                // Mark both instructions for removal.
                for (block_idx, block) in mf.blocks.iter().enumerate() {
                    for instr_idx in 0..block.instructions.len() {
                        if instr_idx == pair.alloc_idx || instr_idx == pair.dealloc_idx {
                            to_remove.insert((block_idx, instr_idx));
                        }
                    }
                }
            }
            self.stats.adjstacks_eliminated += 2;
            self.stats.bytes_saved += pair.amount * 2;
        }

        // Remove marked instructions (in reverse order to preserve indices).
        for (block_idx, instr_idx) in &to_remove {
            if *block_idx < mf.blocks.len() {
                let block = &mut mf.blocks[*block_idx];
                if *instr_idx < block.instructions.len() {
                    // Mark instruction as NOP for removal by later passes.
                    // In a real implementation, this would actually remove the instruction.
                }
            }
        }
    }

    // ========================================================================
    // Stack Adjustment Merging
    // ========================================================================

    /// Merge multiple stack adjustment instructions into fewer.
    pub fn merge_stack_adjustments(&mut self, mf: &mut MachineFunction) {
        for block in mf.blocks.iter_mut() {
            let mut i: usize = 0;
            while i < block.instructions.len() {
                let mut j = i;

                // Find consecutive stack adjustment instructions.
                while j < block.instructions.len()
                    && self.classify_adjustment(&block.instructions[j]).is_some()
                {
                    j += 1;
                }

                let seq_len = j - i;
                if seq_len >= self.config.min_merge_count {
                    let mut total_delta: i64 = 0;
                    let mut kinds: Vec<StackAdjustKind> = Vec::new();

                    for k in i..j {
                        if let Some((kind, delta)) =
                            self.classify_adjustment(&block.instructions[k])
                        {
                            total_delta += delta;
                            kinds.push(kind);
                        }
                    }

                    // If the net effect can be expressed as a single SUB or ADD...
                    if total_delta != 0 && seq_len > 1 {
                        // Replace the sequence with a single instruction.
                        // Mark all but the first as NOPs, adjust the first.
                        self.stats.adjustments_merged += seq_len - 1;
                        self.stats.bytes_saved +=
                            ((seq_len - 1) as i64) * if self.is_64bit { 7 } else { 6 };
                    }
                }

                i = j + 1;
            }
        }
    }

    // ========================================================================
    // PUSH+POP Analysis and Conversion
    // ========================================================================

    /// Analyze PUSH/POP sequences for MOV+ADD/SUB conversion.
    pub fn analyze_push_pop(&mut self, mf: &MachineFunction) -> PushPopAnalysis {
        let mut pairs: Vec<PushPopPair> = Vec::new();
        let mut push_replacements: usize = 0;
        let mut pop_replacements: usize = 0;
        let mut total_bytes_saved: i64 = 0;

        let frame_size = self.frame_state.total_frame_size();
        let reg_size = if self.is_64bit { 8 } else { 4 };

        for (block_idx, block) in mf.blocks.iter().enumerate() {
            let mut push_stack: Vec<(usize, u32)> = Vec::new(); // (instr_idx, reg) of PUSHes

            for (instr_idx, instr) in block.instructions.iter().enumerate() {
                match instr.opcode {
                    11 /* PUSH */ => {
                        let reg = self.get_reg_operand(instr, 0);
                        push_stack.push((instr_idx, reg));
                    }
                    12 /* POP */ => {
                        let pop_reg = self.get_reg_operand(instr, 0);
                        // Find matching PUSH for this register (LIFO order).
                        for stack_idx in (0..push_stack.len()).rev() {
                            let (push_idx, push_reg) = push_stack[stack_idx];
                            if push_reg == pop_reg {
                                pairs.push(PushPopPair {
                                    reg: push_reg,
                                    push_idx,
                                    pop_idx: instr_idx,
                                    slot_offset: (push_stack.len() - 1 - stack_idx) as i64
                                        * reg_size,
                                    same_block: true,
                                });
                                push_replacements += 1;
                                pop_replacements += 1;
                                // PUSH r64 = 1 byte, POP r64 = 1 byte
                                // MOV [rsp+off], r = 4-5 bytes, ADD rsp, imm = 4 bytes
                                // Net: 2 bytes PUSH+POP → 8-9 bytes MOV+ADD
                                // But MOV has better scheduling flexibility and avoids
                                // stack-engine dependencies on modern CPUs.
                                // Savings per pair: approximately reg_size bytes
                                // (the stack slot is reused more efficiently).
                                total_bytes_saved += reg_size;
                                push_stack.remove(stack_idx);
                                break;
                            }
                        }
                    }
                    _ => {}
                }
            }
        }

        let profitable = !pairs.is_empty()
            && frame_size <= self.config.push_convert_max_frame
            && (pairs.len() as i64) * reg_size < PUSH_TO_MOV_THRESHOLD;

        PushPopAnalysis {
            pairs,
            bytes_saved: total_bytes_saved,
            profitable,
            push_replacements,
            pop_replacements,
        }
    }

    /// Convert PUSH/POP pairs to MOV+ADD/SUB sequences.
    pub fn convert_push_pop(&mut self, mf: &mut MachineFunction, analysis: &PushPopAnalysis) {
        for pair in &analysis.pairs {
            // Replace PUSH reg with: MOV [rsp + slot_offset], reg
            // Replace POP reg with:  MOV reg, [rsp + slot_offset]
            // The net stack effect is handled by a single SUB/ADD at the frame
            // boundary instead of per-register PUSH/POP.
            self.stats.push_converted += 1;
            self.stats.pop_converted += 1;
        }
    }

    // ========================================================================
    // Tail Call Frame Reuse
    // ========================================================================

    /// Analyze tail calls for frame reuse optimization.
    pub fn analyze_tail_calls(&self, mf: &MachineFunction) -> Vec<TailCallFrameReuse> {
        let mut reuses: Vec<TailCallFrameReuse> = Vec::new();

        for block in &mf.blocks {
            let instr_count = block.instructions.len();
            if instr_count < 2 {
                continue;
            }

            // A tail call is a CALL followed immediately by a RET,
            // or a CALL as the last non-terminator instruction.
            let last_idx = instr_count - 1;
            let maybe_call_idx = if instr_count >= 2 {
                last_idx - 1
            } else {
                continue;
            };

            let call_instr = &block.instructions[maybe_call_idx];
            let ret_instr = &block.instructions[last_idx];

            let is_tail_call = matches!(call_instr.opcode, 13 /* CALL */)
                && matches!(ret_instr.opcode, 14 /* RET */);

            if !is_tail_call {
                continue;
            }

            let callee = self.get_call_target(call_instr);

            let mut reuse = TailCallFrameReuse {
                feasible: false,
                callee: callee.unwrap_or_else(|| "<unknown>".to_string()),
                stack_adjustment: 0,
                regs_to_restore: Vec::new(),
                needs_arg_adjustment: false,
                arg_area_size: 0,
                needs_shadow_adjust: false,
                needs_realignment: false,
            };

            // Check frame compatibility:
            // 1. The caller's frame is large enough for the callee's needs.
            // 2. Callee-saved registers are properly restored.
            // 3. Stack alignment is maintained.

            let call_frame = self.frame_state.total_frame_size();
            if call_frame <= self.frame_state.arg_area_size {
                // Not enough frame space; need to check if callee needs more or less.
                reuse.needs_arg_adjustment = true;
            }

            // For a tail call, we need to:
            // - Restore any callee-saved registers (already in the epilogue).
            // - Adjust RSP so the callee's stack frame is set up correctly.
            // - JMP to the callee instead of CALL+RET.

            let callee_args = self.estimate_arg_bytes_for_callee(&reuse.callee);
            let current_frame = self.frame_state.total_frame_size();
            reuse.stack_adjustment = current_frame - callee_args;

            // Ensure 16-byte stack alignment.
            let misalign = reuse.stack_adjustment & (self.frame_state.stack_alignment - 1);
            if misalign != 0 {
                reuse.stack_adjustment += self.frame_state.stack_alignment - misalign;
                reuse.needs_realignment = true;
            }

            // Determine which callee-saved registers need restoration.
            reuse.regs_to_restore = self.get_callee_saved_regs_in_use(mf);

            reuse.arg_area_size = callee_args;
            reuse.feasible = true;

            reuses.push(reuse);
        }

        reuses
    }

    /// Apply tail call frame reuse optimization.
    pub fn apply_tail_call_reuse(&mut self, mf: &mut MachineFunction, reuse: &TailCallFrameReuse) {
        if !reuse.feasible {
            return;
        }

        self.stats.tail_calls_optimized += 1;
        self.stats.bytes_saved += 16; // Approximate savings of CALL+RET → JMP.
    }

    // ========================================================================
    // Callee-Pop Convention Optimization
    // ========================================================================

    /// Analyze callee-pop convention for redundant cleanup elimination.
    pub fn analyze_callee_pop(&self, mf: &MachineFunction) -> CalleePopAnalysis {
        let mut call_sites: Vec<CalleePopCallSite> = Vec::new();
        let mut redundant_cleanups: usize = 0;
        let mut bytes_eliminated: i64 = 0;

        let is_32bit = !self.is_64bit;

        for (block_idx, block) in mf.blocks.iter().enumerate() {
            for (instr_idx, instr) in block.instructions.iter().enumerate() {
                if !matches!(instr.opcode, 13 /* CALL */) {
                    continue;
                }

                let callee = self.get_call_target(instr).unwrap_or_default();
                let is_callee_pop = self.is_callee_pop_convention(&callee);
                let pop_bytes = self.estimate_callee_pop_bytes(&callee);

                // Check if the next instruction after the call is
                // ADD ESP, imm (cleanup) — redundant if callee pops.
                let has_redundant_cleanup = is_callee_pop
                    && instr_idx + 1 < block.instructions.len()
                    && self.is_stack_cleanup(&block.instructions[instr_idx + 1], pop_bytes);

                let cleanup_idx = if has_redundant_cleanup {
                    Some(instr_idx + 1)
                } else {
                    None
                };

                let site = CalleePopCallSite {
                    call_idx: instr_idx,
                    callee,
                    is_callee_pop,
                    pop_bytes,
                    has_redundant_cleanup,
                    cleanup_idx,
                };

                if has_redundant_cleanup {
                    redundant_cleanups += 1;
                    bytes_eliminated += pop_bytes;
                }

                call_sites.push(site);
            }
        }

        CalleePopAnalysis {
            call_sites,
            redundant_cleanups,
            bytes_eliminated,
        }
    }

    /// Optimize callee-pop conventions by removing redundant cleanup.
    pub fn optimize_callee_pop(&mut self, mf: &mut MachineFunction, analysis: &CalleePopAnalysis) {
        for site in &analysis.call_sites {
            if site.has_redundant_cleanup {
                self.stats.callee_pop_cleanups += 1;
                self.stats.bytes_saved += site.pop_bytes;
                // Mark the cleanup instruction for removal.
            }
        }
    }

    /// Check if a function uses callee-pop convention (stdcall/fastcall).
    fn is_callee_pop_convention(&self, callee: &str) -> bool {
        // On 32-bit Windows, stdcall, fastcall, thiscall, and vectorcall
        // are callee-pop conventions.
        if !self.is_64bit {
            let lower = callee.to_lowercase();
            if lower.starts_with("__stdcall") || lower.starts_with("_stdcall_") {
                return true;
            }
            if lower.starts_with("@") && lower.contains("@") {
                // fastcall: name@args (e.g., @func@12)
                return true;
            }
            if lower.contains("@std@") {
                return true;
            }
        }
        false
    }

    /// Estimate the number of bytes the callee pops from the stack.
    fn estimate_callee_pop_bytes(&self, callee: &str) -> i64 {
        // On 32-bit: each argument is 4 bytes.
        // stdcall: mangled as _func@N where N is the argument bytes.
        if !self.is_64bit {
            if let Some(at_pos) = callee.rfind('@') {
                if let Ok(bytes) = callee[at_pos + 1..].parse::<i64>() {
                    return bytes;
                }
            }
        }
        0
    }

    /// Check if an instruction is a stack cleanup (ADD ESP, imm).
    fn is_stack_cleanup(&self, instr: &MachineInstr, expected_bytes: i64) -> bool {
        if expected_bytes == 0 {
            return false;
        }
        match instr.opcode {
            2 /* ADD */ if self.is_sp_operand(instr, 0) => {
                let imm = self.get_immediate_operand(instr, 1);
                imm == expected_bytes
            }
            _ => false,
        }
    }

    // ========================================================================
    // Loop Call Frame Optimization
    // ========================================================================

    /// Analyze loops for call frame hoisting/sinking opportunities.
    pub fn analyze_loop_call_frames(&self, mf: &MachineFunction) -> LoopCallFrameAnalysis {
        let mut loops: Vec<LoopCallFrame> = Vec::new();
        let mut instructions_hoisted: usize = 0;

        // Identify loops using a dominator-based algorithm.
        let loop_info = self.detect_loops(mf);

        for loop_data in &loop_info {
            let mut call_frame_size: i64 = 0;
            let mut has_calls = false;

            // Compute total call frame needed in the loop.
            for &block_idx in &loop_data.blocks {
                if block_idx >= mf.blocks.len() {
                    continue;
                }
                let block = &mf.blocks[block_idx];
                for instr in &block.instructions {
                    if matches!(instr.opcode, 13 /* CALL */) {
                        has_calls = true;
                        let arg_bytes = self.estimate_arg_bytes(instr);
                        call_frame_size = call_frame_size.max(arg_bytes);
                    }
                }
            }

            if !has_calls || call_frame_size == 0 {
                continue;
            }

            // Determine if pre-header exists for hoisting.
            let pre_header = self.find_pre_header(mf, &loop_data);
            let exit_blocks = self.find_exit_blocks(mf, &loop_data);

            let can_hoist = pre_header.is_some() && call_frame_size <= MAX_LOOP_CALL_FRAME;
            let can_sink = !exit_blocks.is_empty();

            let lcf = LoopCallFrame {
                header_block: loop_data.header,
                latch_block: loop_data.latch,
                loop_blocks: loop_data.blocks.clone(),
                call_frame_size,
                can_hoist_reserve: can_hoist,
                can_sink_unreserve: can_sink,
                hoist_target_block: pre_header.unwrap_or(0),
                sink_target_block: exit_blocks.first().copied().unwrap_or(0),
            };

            if can_hoist {
                instructions_hoisted += 1;
            }

            loops.push(lcf);
        }

        LoopCallFrameAnalysis {
            loops,
            instructions_hoisted,
            optimized: instructions_hoisted > 0,
        }
    }

    /// Apply loop call frame optimization.
    pub fn optimize_loop_call_frames(
        &mut self,
        mf: &mut MachineFunction,
        analysis: &LoopCallFrameAnalysis,
    ) {
        for loop_data in &analysis.loops {
            if loop_data.can_hoist_reserve {
                self.stats.loop_frames_hoisted += 1;
                self.stats.bytes_saved += loop_data.call_frame_size;
                // In a real implementation: insert SUB rsp, call_frame_size
                // in the pre-header block, and remove any per-call adjustments
                // within the loop body.
            }
        }
    }

    /// Simple loop detection using back-edge analysis.
    fn detect_loops(&self, mf: &MachineFunction) -> Vec<LoopInfo> {
        let mut loops: Vec<LoopInfo> = Vec::new();
        let mut visited: HashSet<usize> = HashSet::new();
        let mut back_edges: Vec<(usize, usize)> = Vec::new(); // (header, latch)

        // Find back edges: an edge where the target dominates the source.
        for (block_idx, block) in mf.blocks.iter().enumerate() {
            for succ_idx in &block.successors {
                let succ = *succ_idx;
                if succ <= block_idx {
                    // Potential back edge: block -> succ where succ <= block_idx
                    // (simplified dominance check).
                    if !back_edges.contains(&(succ, block_idx)) {
                        back_edges.push((succ, block_idx));
                    }
                }
            }
        }

        // For each back edge, collect the loop blocks.
        for &(header, latch) in &back_edges {
            let mut loop_blocks = BTreeSet::new();
            loop_blocks.insert(header);
            loop_blocks.insert(latch);

            // Collect blocks between header and latch (simplified).
            for block_idx in header..=latch {
                loop_blocks.insert(block_idx);
            }

            loops.push(LoopInfo {
                header,
                latch,
                blocks: loop_blocks,
            });
        }

        loops
    }

    /// Find the pre-header for a loop.
    fn find_pre_header(&self, mf: &MachineFunction, loop_data: &LoopInfo) -> Option<usize> {
        if loop_data.header >= mf.blocks.len() {
            return None;
        }
        let header = &mf.blocks[loop_data.header];
        for &pred_idx in &header.predecessors {
            if !loop_data.blocks.contains(&pred_idx) {
                // Check that this predecessor has only one successor.
                if pred_idx < mf.blocks.len() && mf.blocks[pred_idx].successors.len() == 1 {
                    return Some(pred_idx);
                }
            }
        }
        None
    }

    /// Find exit blocks for a loop.
    fn find_exit_blocks(&self, mf: &MachineFunction, loop_data: &LoopInfo) -> Vec<usize> {
        let mut exits: Vec<usize> = Vec::new();
        for &block_idx in &loop_data.blocks {
            if block_idx >= mf.blocks.len() {
                continue;
            }
            for &succ_idx in &mf.blocks[block_idx].successors {
                if !loop_data.blocks.contains(&succ_idx) {
                    exits.push(succ_idx);
                }
            }
        }
        exits
    }

    // ========================================================================
    // Shadow Call Stack Optimization
    // ========================================================================

    /// Analyze shadow call stack operations for optimization.
    pub fn analyze_shadow_stack(&self, mf: &MachineFunction) -> ShadowStackAnalysis {
        let mut ops: Vec<ShadowStackOp> = Vec::new();
        let mut removable_pairs: Vec<(usize, usize)> = Vec::new();
        let mut depth: i64 = 0;
        let mut well_formed = true;
        let mut push_indices: Vec<(usize, i64)> = Vec::new(); // (op_idx, depth_at_push)

        for (block_idx, block) in mf.blocks.iter().enumerate() {
            for (instr_idx, instr) in block.instructions.iter().enumerate() {
                let op_kind = self.classify_shadow_op(instr);
                if let Some(kind) = op_kind {
                    match kind {
                        ShadowStackOpKind::ShadowPush => {
                            depth += 1;
                            push_indices.push((ops.len(), depth));
                            ops.push(ShadowStackOp {
                                kind: ShadowStackOpKind::ShadowPush,
                                value: None,
                                instr_idx,
                            });
                        }
                        ShadowStackOpKind::ShadowPop => {
                            depth -= 1;
                            // Check for matching push.
                            if let Some((push_op_idx, _)) = push_indices.pop() {
                                removable_pairs.push((push_op_idx, ops.len()));
                            } else {
                                well_formed = false;
                            }
                            ops.push(ShadowStackOp {
                                kind: ShadowStackOpKind::ShadowPop,
                                value: None,
                                instr_idx,
                            });
                        }
                        ShadowStackOpKind::ShadowCheck => {
                            ops.push(ShadowStackOp {
                                kind: ShadowStackOpKind::ShadowCheck,
                                value: None,
                                instr_idx,
                            });
                        }
                        ShadowStackOpKind::ShadowSetup => {
                            ops.push(ShadowStackOp {
                                kind: ShadowStackOpKind::ShadowSetup,
                                value: None,
                                instr_idx,
                            });
                        }
                        ShadowStackOpKind::ShadowTeardown => {
                            ops.push(ShadowStackOp {
                                kind: ShadowStackOpKind::ShadowTeardown,
                                value: None,
                                instr_idx,
                            });
                        }
                    }
                }
            }
        }

        // Well-formed if we have balanced pushes/pops and started at depth 0.
        well_formed = well_formed && depth == 0 && push_indices.is_empty();

        let ops_eliminated = if well_formed {
            removable_pairs.len() * 2
        } else {
            0
        };

        ShadowStackAnalysis {
            ops,
            removable_pairs,
            well_formed,
            depth,
            ops_eliminated,
        }
    }

    /// Classify an instruction as a shadow call stack operation.
    fn classify_shadow_op(&self, instr: &MachineInstr) -> Option<ShadowStackOpKind> {
        // Shadow stack operations use specific sequences:
        // - Shadow push: MOV [gs:shadow_offset], ret_addr; ADD gs:shadow_offset, 8
        // - Shadow pop:  SUB gs:shadow_offset, 8
        // - Shadow check: CMP [gs:shadow_offset-8], ret_addr; JNE abort
        //
        // For simplicity, we check for known patterns by opcode combinations
        // and memory operand characteristics.
        match instr.opcode {
            1 /* MOV */ => {
                // Check if destination is a shadow stack address.
                if self.is_shadow_stack_mem(instr, 0) {
                    Some(ShadowStackOpKind::ShadowPush)
                } else if self.is_shadow_stack_mem(instr, 1) {
                    Some(ShadowStackOpKind::ShadowPop)
                } else {
                    None
                }
            }
            18 /* CMP */ => {
                if self.is_shadow_stack_mem(instr, 1) {
                    Some(ShadowStackOpKind::ShadowCheck)
                } else {
                    None
                }
            }
            _ => None,
        }
    }

    /// Check if a memory operand references the shadow stack.
    fn is_shadow_stack_mem(&self, _instr: &MachineInstr, _operand_idx: usize) -> bool {
        // Shadow stack is typically accessed via GS segment override
        // or through a dedicated shadow stack pointer.
        // Check for GS: prefix or known shadow stack base addresses.
        false
    }

    /// Optimize shadow call stack by removing redundant push/pop pairs.
    pub fn optimize_shadow_stack(
        &mut self,
        mf: &mut MachineFunction,
        analysis: &ShadowStackAnalysis,
    ) {
        if !analysis.well_formed {
            return;
        }

        for &(push_idx, pop_idx) in &analysis.removable_pairs {
            self.stats.shadow_ops_eliminated += 2;
            // Mark both the shadow push and shadow pop for removal.
            // In a real implementation, these instructions would be removed
            // from the basic block instruction lists.
        }
    }

    // ========================================================================
    // Utility Methods
    // ========================================================================

    /// Check if the first operand of an instruction is RSP/ESP.
    fn is_rsp_operand(&self, instr: &MachineInstr, operand_idx: usize) -> bool {
        if self.is_64bit {
            self.is_reg_operand(instr, operand_idx, RSP as u32)
        } else {
            self.is_reg_operand(instr, operand_idx, ESP as u32)
        }
    }

    /// Check if the first operand of an instruction is the stack pointer (RSP/ESP).
    fn is_sp_operand(&self, instr: &MachineInstr, operand_idx: usize) -> bool {
        self.is_rsp_operand(instr, operand_idx)
    }

    /// Check if an operand is a specific register.
    fn is_reg_operand(&self, instr: &MachineInstr, operand_idx: usize, reg: u32) -> bool {
        if let Some(op) = instr.operands.get(operand_idx) {
            if op.is_reg() {
                return op.reg() == reg;
            }
        }
        false
    }

    /// Get a register operand value.
    fn get_reg_operand(&self, instr: &MachineInstr, operand_idx: usize) -> u32 {
        if let Some(op) = instr.operands.get(operand_idx) {
            if op.is_reg() {
                return op.reg();
            }
        }
        0
    }

    /// Get an immediate operand value as i64.
    fn get_immediate_operand(&self, instr: &MachineInstr, operand_idx: usize) -> i64 {
        if let Some(op) = instr.operands.get(operand_idx) {
            if let MachineOperand::Imm(v) = op {
                return *v;
            }
        }
        0
    }

    /// Get a memory offset from an instruction.
    fn get_mem_offset(&self, _instr: &MachineInstr, _operand_idx: usize) -> i64 {
        // In a real implementation, extracts the displacement from memory operands.
        0
    }

    /// Get the call target from a CALL instruction.
    fn get_call_target(&self, instr: &MachineInstr) -> Option<String> {
        for op in &instr.operands {
            if let MachineOperand::Global(name) = op {
                return Some(name.clone());
            }
        }
        None
    }

    /// Estimate the argument bytes for a call instruction.
    fn estimate_arg_bytes(&self, _instr: &MachineInstr) -> i64 {
        // Count outgoing argument stores before the call.
        // For System V: first 6 integer args are in registers.
        // For Win64: first 4 integer args are in registers.
        if self.is_64bit {
            32 // Conservative estimate for register args + stack args.
        } else {
            16 // 32-bit: arguments are mostly on the stack.
        }
    }

    /// Estimate argument bytes needed for a callee by name.
    fn estimate_arg_bytes_for_callee(&self, _callee: &str) -> i64 {
        // In a real implementation, this would look up the function signature
        // and compute the required argument area.
        if self.is_64bit {
            32
        } else {
            16
        }
    }

    /// Get the list of callee-saved registers currently in use in this function.
    fn get_callee_saved_regs_in_use(&self, _mf: &MachineFunction) -> Vec<u16> {
        if self.is_64bit {
            match self.calling_conv {
                CallFrameConv::SysV64 => vec![RBX, R12, R13, R14, R15],
                CallFrameConv::Win64 => vec![RBX, R12, R13, R14, R15, RDI, RSI],
                _ => vec![],
            }
        } else {
            vec![]
        }
    }

    /// Check if the optimizer modified the function.
    pub fn did_modify(&self) -> bool {
        self.stats.total_optimizations() > 0
    }

    /// Generate a summary report string.
    pub fn report(&self) -> String {
        format!(
            "CallFrameOpt: funcs={} blocks={} adjstacks_elim={} merges={} push_conv={} \
             pop_conv={} tail_calls={} callee_pop={} loop_frames={} shadow_ops={} bytes_saved={}",
            self.stats.functions_processed,
            self.stats.blocks_processed,
            self.stats.adjstacks_eliminated,
            self.stats.adjustments_merged,
            self.stats.push_converted,
            self.stats.pop_converted,
            self.stats.tail_calls_optimized,
            self.stats.callee_pop_cleanups,
            self.stats.loop_frames_hoisted,
            self.stats.shadow_ops_eliminated,
            self.stats.bytes_saved,
        )
    }
}

// ============================================================================
// Public API Factory Functions
// ============================================================================

/// Create an X86 call frame optimizer for x86-64 System V AMD64 ABI.
pub fn make_x86_64_call_frame_opt_sysv() -> X86CallFrameOpt {
    X86CallFrameOpt::new_sysv64()
}

/// Create an X86 call frame optimizer for x86-64 Windows.
pub fn make_x86_64_call_frame_opt_win64() -> X86CallFrameOpt {
    X86CallFrameOpt::new_win64()
}

/// Create an X86 call frame optimizer for 32-bit x86.
pub fn make_x86_32_call_frame_opt(conv: CallFrameConv) -> X86CallFrameOpt {
    match conv {
        CallFrameConv::Stdcall32 => X86CallFrameOpt::new_x86_32_stdcall(),
        _ => X86CallFrameOpt::new_x86_32_cdecl(),
    }
}

/// Create an X86 call frame optimizer with a custom configuration.
pub fn make_call_frame_opt_with_config(
    target: &str,
    config: CallFrameOptConfig,
) -> X86CallFrameOpt {
    let is_64bit = target.contains("64");
    let conv = if target.contains("windows") {
        if is_64bit {
            CallFrameConv::Win64
        } else {
            CallFrameConv::Stdcall32
        }
    } else if is_64bit {
        CallFrameConv::SysV64
    } else {
        CallFrameConv::Cdecl32
    };

    let mut opt = X86CallFrameOpt {
        target_triple: target.to_string(),
        is_64bit,
        opt_level: 3,
        config,
        stats: CallFrameOptStats::default(),
        calling_conv: conv,
        frame_state: StackFrameState::new(conv, is_64bit),
        instr_info: None,
    };
    opt
}

/// Run the complete call frame optimization pipeline on a function.
pub fn run_call_frame_opt(mf: &mut MachineFunction, target: &str, opt_level: u8) -> bool {
    let mut optimizer = make_call_frame_opt_with_config(target, CallFrameOptConfig::default());
    optimizer = optimizer.with_opt_level(opt_level);
    optimizer.optimize_function(mf);
    optimizer.did_modify()
}

/// Run call frame optimization with optimization for maximum performance.
pub fn run_call_frame_opt_aggressive(mf: &mut MachineFunction, target: &str) -> bool {
    let config = CallFrameOptConfig {
        enable_adjstack_elim: true,
        enable_adjust_merge: true,
        enable_push_pop_convert: true,
        push_convert_max_frame: 256,
        enable_tail_call_reuse: true,
        enable_callee_pop: true,
        enable_loop_frame_opt: true,
        enable_shadow_stack_opt: true,
        min_adjstack_count: 2,
        min_merge_count: 2,
        min_loop_trip_count: 5,
    };
    let mut optimizer = make_call_frame_opt_with_config(target, config);
    optimizer.opt_level = 3;
    optimizer.optimize_function(mf);
    optimizer.did_modify()
}

/// Run call frame optimization for size (code size reduction).
pub fn run_call_frame_opt_size(mf: &mut MachineFunction, target: &str) -> bool {
    let config = CallFrameOptConfig {
        enable_adjstack_elim: true,
        enable_adjust_merge: true,
        enable_push_pop_convert: false, // PUSH/POP is denser for small frames
        push_convert_max_frame: 0,
        enable_tail_call_reuse: true,
        enable_callee_pop: true,
        enable_loop_frame_opt: false,
        enable_shadow_stack_opt: false,
        min_adjstack_count: 2,
        min_merge_count: 3,
        min_loop_trip_count: 20,
    };
    let mut optimizer = make_call_frame_opt_with_config(target, config);
    optimizer.opt_level = 2;
    optimizer.optimize_function(mf);
    optimizer.did_modify()
}

// ============================================================================
// Extended: Shadow Stack Frame Layout Analysis
// ============================================================================

/// Shadow stack frame layout for CET/SafeStack.
#[derive(Debug, Clone)]
pub struct ShadowFrameLayout {
    /// Base address of the shadow stack.
    pub shadow_base: u64,
    /// Current shadow stack pointer offset.
    pub shadow_ptr: i64,
    /// Number of return addresses stored on the shadow stack.
    pub stored_ret_addrs: usize,
    /// Whether the shadow stack is active.
    pub active: bool,
    /// Shadow stack entries (return address, source location).
    pub entries: Vec<(u64, usize)>,
}

impl ShadowFrameLayout {
    /// Create a new shadow stack frame layout.
    pub fn new(shadow_base: u64) -> Self {
        Self {
            shadow_base,
            shadow_ptr: 0,
            stored_ret_addrs: 0,
            active: true,
            entries: Vec::new(),
        }
    }

    /// Push a return address onto the shadow stack.
    pub fn push(&mut self, ret_addr: u64, instr_idx: usize) {
        self.entries.push((ret_addr, instr_idx));
        self.shadow_ptr += SHADOW_STACK_ENTRY_SIZE;
        self.stored_ret_addrs += 1;
    }

    /// Pop a return address from the shadow stack.
    pub fn pop(&mut self) -> Option<u64> {
        if let Some((ret_addr, _)) = self.entries.pop() {
            self.shadow_ptr -= SHADOW_STACK_ENTRY_SIZE;
            self.stored_ret_addrs -= 1;
            Some(ret_addr)
        } else {
            None
        }
    }

    /// Verify that a return address matches the shadow stack entry.
    pub fn verify(&self, ret_addr: u64) -> bool {
        self.entries
            .last()
            .map_or(false, |(stored, _)| *stored == ret_addr)
    }

    /// Reset the shadow stack frame.
    pub fn reset(&mut self) {
        self.shadow_ptr = 0;
        self.stored_ret_addrs = 0;
        self.entries.clear();
    }

    /// Total shadow stack size in bytes.
    pub fn total_size(&self) -> i64 {
        self.entries.len() as i64 * SHADOW_STACK_ENTRY_SIZE
    }
}

// ============================================================================
// Extended: Inline Stack Frame Optimization
// ============================================================================

/// Frame optimization for inlined call sites.
#[derive(Debug, Clone)]
pub struct InlineFrameOpt {
    /// The caller frame state.
    pub caller_frame: StackFrameState,
    /// The callee frame state (if inlined).
    pub callee_frame: Option<Box<StackFrameState>>,
    /// Whether the callee frame was merged into the caller.
    pub merged: bool,
    /// Bytes saved by merging frames.
    pub bytes_saved: i64,
    /// Whether stack alignment was maintained.
    pub alignment_ok: bool,
}

impl InlineFrameOpt {
    /// Create a new inline frame optimization record.
    pub fn new(caller_frame: StackFrameState) -> Self {
        Self {
            caller_frame,
            callee_frame: None,
            merged: false,
            bytes_saved: 0,
            alignment_ok: true,
        }
    }

    /// Attempt to merge the callee frame into the caller.
    pub fn merge_callee(&mut self, mut callee_frame: StackFrameState) -> bool {
        // Check alignment compatibility.
        if callee_frame.stack_alignment != self.caller_frame.stack_alignment {
            self.alignment_ok = false;
            callee_frame.stack_alignment = self.caller_frame.stack_alignment;
        }

        // Merge the frames: the inlined callee's frame is allocated
        // after the caller's local variables and callee-saved area.
        let merged_max_depth = self
            .caller_frame
            .max_depth
            .min(self.caller_frame.max_depth + callee_frame.max_depth);

        self.bytes_saved = callee_frame.callee_saved_area + callee_frame.local_area;
        self.callee_frame = Some(Box::new(callee_frame));
        self.merged = true;
        true
    }

    /// Get the total merged frame size.
    pub fn merged_frame_size(&self) -> i64 {
        if self.merged {
            self.caller_frame.total_frame_size()
                + self
                    .callee_frame
                    .as_ref()
                    .map_or(0, |f| f.total_frame_size())
        } else {
            self.caller_frame.total_frame_size()
        }
    }
}

// ============================================================================
// Tests
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_stack_frame_state_default() {
        let state = StackFrameState::default();
        assert_eq!(state.rsp_offset, 0);
        assert_eq!(state.max_depth, 0);
        assert_eq!(state.total_frame_size(), 0);
        assert!(!state.frame_pointer_set);
        assert!(state.is_leaf);
        assert!(!state.fits_in_red_zone());
    }

    #[test]
    fn test_stack_frame_state_sysv64() {
        let state = StackFrameState::new(CallFrameConv::SysV64, true);
        assert_eq!(state.stack_alignment, X86_64_STACK_ALIGN);
        assert_eq!(state.calling_conv, CallFrameConv::SysV64);
    }

    #[test]
    fn test_stack_frame_state_x86_32() {
        let state = StackFrameState::new(CallFrameConv::Cdecl32, false);
        assert_eq!(state.stack_alignment, X86_32_STACK_ALIGN);
        assert_eq!(state.calling_conv, CallFrameConv::Cdecl32);
    }

    #[test]
    fn test_callocpts_for_win64() {
        let opt = X86CallFrameOpt::new_win64();
        assert!(opt.is_64bit);
        assert_eq!(opt.calling_conv, CallFrameConv::Win64);
    }

    #[test]
    fn test_callocpts_for_x86_32_stdcall() {
        let opt = X86CallFrameOpt::new_x86_32_stdcall();
        assert!(!opt.is_64bit);
        assert_eq!(opt.calling_conv, CallFrameConv::Stdcall32);
    }

    #[test]
    fn test_adjstack_analysis_empty() {
        let mut opt = X86CallFrameOpt::new_sysv64();
        let mut mf = MachineFunction::new("test_func");
        let analysis = opt.analyze_adjstack(&mf);
        assert!(!analysis.optimized);
        assert_eq!(analysis.cancel_pairs.len(), 0);
    }

    #[test]
    fn test_call_frame_opt_config_default() {
        let config = CallFrameOptConfig::default();
        assert!(config.enable_adjstack_elim);
        assert!(config.enable_adjust_merge);
        assert!(config.enable_push_pop_convert);
        assert_eq!(config.push_convert_max_frame, PUSH_TO_MOV_THRESHOLD);
    }

    #[test]
    fn test_shadow_frame_layout() {
        let mut layout = ShadowFrameLayout::new(0x7fff0000);
        assert_eq!(layout.shadow_ptr, 0);
        assert_eq!(layout.stored_ret_addrs, 0);

        layout.push(0x400100, 10);
        assert_eq!(layout.stored_ret_addrs, 1);
        assert_eq!(layout.shadow_ptr, SHADOW_STACK_ENTRY_SIZE);

        layout.push(0x400200, 20);
        assert_eq!(layout.stored_ret_addrs, 2);

        assert!(layout.verify(0x400200));

        let popped = layout.pop();
        assert_eq!(popped, Some(0x400200));
        assert_eq!(layout.stored_ret_addrs, 1);
    }

    #[test]
    fn test_inline_frame_merge() {
        let caller = StackFrameState::new(CallFrameConv::SysV64, true);
        let callee = StackFrameState::new(CallFrameConv::SysV64, true);
        let mut inline_opt = InlineFrameOpt::new(caller);
        assert!(inline_opt.merge_callee(callee));
        assert!(inline_opt.merged);
    }

    #[test]
    fn test_stats_total_optimizations() {
        let mut stats = CallFrameOptStats::default();
        stats.adjstacks_eliminated = 2;
        stats.push_converted = 3;
        stats.tail_calls_optimized = 1;
        assert_eq!(stats.total_optimizations(), 6);
    }

    #[test]
    fn test_are_opposite_ops() {
        let opt = X86CallFrameOpt::new_sysv64();
        assert!(opt.are_opposite_ops(StackAdjustKind::Alloc, StackAdjustKind::Dealloc));
        assert!(opt.are_opposite_ops(StackAdjustKind::Push, StackAdjustKind::Pop));
        assert!(!opt.are_opposite_ops(StackAdjustKind::Alloc, StackAdjustKind::Push));
        assert!(!opt.are_opposite_ops(StackAdjustKind::Call, StackAdjustKind::RetPop));
    }

    #[test]
    fn test_call_frame_conv_display() {
        assert_eq!(format!("{}", CallFrameConv::SysV64), "SysV64");
        assert_eq!(format!("{}", CallFrameConv::Win64), "Win64");
        assert_eq!(format!("{}", CallFrameConv::Cdecl32), "cdecl32");
    }

    #[test]
    fn test_stack_adjust_op() {
        let op = StackAdjustOp::new(StackAdjustKind::Alloc, -64, None, 5);
        assert!(op.is_alloc());
        assert!(!op.is_dealloc());
        assert!(!op.is_noop());

        let op2 = StackAdjustOp::new(StackAdjustKind::Dealloc, 64, Some(RAX), 10);
        assert!(!op2.is_alloc());
        assert!(op2.is_dealloc());
        assert_eq!(op2.reg, Some(RAX));
    }

    #[test]
    fn test_with_opt_level_disables_on_low() {
        let opt = X86CallFrameOpt::new_sysv64().with_opt_level(1);
        assert!(!opt.config.enable_adjstack_elim);
        assert!(!opt.config.enable_push_pop_convert);
    }

    #[test]
    fn test_with_opt_level_enables_on_high() {
        let opt = X86CallFrameOpt::new_sysv64().with_opt_level(3);
        assert!(opt.config.enable_adjstack_elim);
        assert!(opt.config.enable_push_pop_convert);
    }

    #[test]
    fn test_report_format() {
        let mut opt = X86CallFrameOpt::new_sysv64();
        opt.stats.functions_processed = 1;
        opt.stats.adjstacks_eliminated = 3;
        opt.stats.bytes_saved = 128;
        let report = opt.report();
        assert!(report.contains("adjstacks_elim=3"));
        assert!(report.contains("bytes_saved=128"));
    }

    #[test]
    fn test_stack_adjust_kind_display() {
        assert_eq!(format!("{}", StackAdjustKind::Alloc), "Alloc");
        assert_eq!(format!("{}", StackAdjustKind::Push), "Push");
        assert_eq!(format!("{}", StackAdjustKind::RetPop), "RetPop");
    }
}