wasmer-compiler-cranelift 7.4.2

//! Helpers for generating DWARF LSDA data for Cranelift-compiled functions.
//!
//! The structures and encoding implemented here mirror what LLVM produces for
//! Wasm exception handling so that Wasmer's libunwind personalities can parse
//! the tables without any runtime changes.

use cranelift_codegen::{
    ExceptionContextLoc, FinalizedMachCallSite, FinalizedMachExceptionHandler,
    isa::unwind::UnwindInst,
};
use cranelift_entity::EntityRef;
use itertools::Itertools;
use std::collections::hash_map::Entry;
use std::collections::{HashMap, HashSet};
use std::convert::TryFrom;
use std::io::{Cursor, Write};

use wasmer_compiler::types::{
    relocation::{Relocation, RelocationKind, RelocationTarget},
    section::{CustomSection, CustomSectionProtection, SectionBody, SectionIndex},
};
use wasmer_types::{LibCall, LocalFunctionIndex};

/// Relocation information for an LSDA entry that references a tag constant.
#[derive(Debug, Clone)]
pub struct TagRelocation {
    /// Offset within the LSDA blob where the relocation should be applied.
    pub offset: u32,
    /// The module-local exception tag value.
    pub tag: u32,
}

/// Fully encoded LSDA bytes for a single function, together with pending tag
/// relocations that will be resolved once the global tag section is built.
#[derive(Debug, Clone)]
pub struct FunctionLsdaData {
    pub bytes: Vec<u8>,
    pub relocations: Vec<TagRelocation>,
}

/// Build the LSDA for a single function given the finalized Cranelift
/// call-site metadata.
pub fn build_function_lsda<'a>(
    call_sites: impl Iterator<Item = FinalizedMachCallSite<'a>>,
    function_length: usize,
    pointer_bytes: u8,
    pcrel_type_table: bool,
) -> Option<FunctionLsdaData> {
    let mut sites = Vec::new();

    for site in call_sites {
        let mut catches = Vec::new();
        let mut landing_pad = None;

        // Our landing pads handle all the tags considered for a call instruction, thus
        // we use the latest landing pad.
        for handler in site.exception_handlers {
            match handler {
                FinalizedMachExceptionHandler::Tag(tag, offset) => {
                    landing_pad = Some(landing_pad.unwrap_or(*offset));
                    catches.push(ExceptionType::Tag {
                        tag: u32::try_from(tag.index()).expect("tag index fits in u32"),
                    });
                }
                FinalizedMachExceptionHandler::Default(offset) => {
                    landing_pad = Some(landing_pad.unwrap_or(*offset));
                    catches.push(ExceptionType::CatchAll);
                }
                FinalizedMachExceptionHandler::Context(context) => {
                    // Context records are used by Cranelift to thread VMContext
                    // information through the landing pad. We emit the LSDA
                    // regardless of whether we see them; nothing to do here.
                    match context {
                        ExceptionContextLoc::SPOffset(_) | ExceptionContextLoc::GPR(_) => {}
                    }
                }
            }
        }

        if catches.is_empty() {
            continue;
        }

        let landing_pad = landing_pad.expect("landing pad offset set when catches exist");
        let cs_start = site.ret_addr.saturating_sub(1);

        sites.push(CallSiteDesc {
            start: cs_start,
            len: 1,
            landing_pad,
            actions: catches,
        });
    }

    if sites.is_empty() {
        return None;
    }

    // Ensure all instructions in the function are covered by filling gaps with
    // default unwinding behavior (no catch actions).
    let mut current_pos = 0u32;
    let mut filled_sites = Vec::new();

    for site in sites {
        if site.start > current_pos {
            // Gap found: add a default site that covers instructions with no handlers
            filled_sites.push(CallSiteDesc {
                start: current_pos,
                len: site.start - current_pos,
                landing_pad: 0,
                actions: Vec::new(),
            });
        }
        current_pos = site.start + site.len;
        filled_sites.push(site);
    }

    // Cover any remaining instructions at the end of the function
    if current_pos < function_length as u32 {
        filled_sites.push(CallSiteDesc {
            start: current_pos,
            len: function_length as u32 - current_pos,
            landing_pad: 0,
            actions: Vec::new(),
        });
    }

    let sites = filled_sites;

    let mut type_entries = TypeTable::new();
    let mut callsite_actions = Vec::with_capacity(sites.len());

    for site in &sites {
        #[cfg(debug_assertions)]
        {
            // CatchAll must always be the last item in the action list; otherwise, the tags that follow
            // it will be ignored.
            let catch_all_positions = site
                .actions
                .iter()
                .positions(|a| matches!(a, ExceptionType::CatchAll))
                .collect_vec();
            assert!(catch_all_positions.iter().at_most_one().is_ok());
            if let Some(&i) = catch_all_positions.first() {
                assert!(i == site.actions.len() - 1);
            }
        }

        let action_indices = site
            .actions
            .iter()
            // Reverse actions to ensure CatchAll is always last in the chain, since the action table
            // encoding uses back references and relies on this ordering.
            .rev()
            .map(|action| type_entries.get_or_insert(*action) as i32)
            .collect_vec();
        callsite_actions.push(action_indices);
    }

    let action_table = encode_action_table(&callsite_actions);
    let call_site_table = encode_call_site_table(&sites, &action_table);
    let (type_table_bytes, type_table_relocs) = if pcrel_type_table {
        type_entries.encode_relocated()
    } else {
        type_entries.encode(pointer_bytes)
    };

    let call_site_table_len = call_site_table.len() as u64;
    let mut writer = Cursor::new(Vec::new());
    writer
        .write_all(&cranelift_codegen::gimli::DW_EH_PE_omit.0.to_le_bytes())
        .unwrap(); // lpstart encoding omitted (relative to function start)

    if type_entries.is_empty() {
        writer
            .write_all(&cranelift_codegen::gimli::DW_EH_PE_omit.0.to_le_bytes())
            .unwrap();
    } else if pcrel_type_table {
        // PC-relative, 4-byte entries. This keeps the `.gcc_except_table`
        // section relocations position-independent (e.g. `R_X86_64_PC32`).
        writer
            .write_all(
                &(cranelift_codegen::gimli::DW_EH_PE_pcrel
                    | cranelift_codegen::gimli::DW_EH_PE_sdata4)
                    .0
                    .to_le_bytes(),
            )
            .unwrap();
    } else {
        writer
            .write_all(&cranelift_codegen::gimli::DW_EH_PE_absptr.0.to_le_bytes())
            .unwrap();
    }

    if !type_entries.is_empty() {
        let ttype_table_end = 1 // call-site encoding byte
            + uleb128_len(call_site_table_len)
            + call_site_table.len()
            + action_table.bytes.len()
            + type_table_bytes.len();
        leb128::write::unsigned(&mut writer, ttype_table_end as u64).unwrap();
    }

    writer
        .write_all(&cranelift_codegen::gimli::DW_EH_PE_udata4.0.to_le_bytes())
        .unwrap();
    leb128::write::unsigned(&mut writer, call_site_table_len).unwrap();
    writer.write_all(&call_site_table).unwrap();
    writer.write_all(&action_table.bytes).unwrap();

    let type_table_offset = writer.position() as u32;
    writer.write_all(&type_table_bytes).unwrap();

    let mut relocations = Vec::new();
    for reloc in type_table_relocs {
        relocations.push(TagRelocation {
            offset: type_table_offset + reloc.offset,
            tag: reloc.tag,
        });
    }

    Some(FunctionLsdaData {
        bytes: writer.into_inner(),
        relocations,
    })
}

/// Build the global tag section and a tag->offset map.
pub fn build_tag_section(
    lsda_data: &[Option<FunctionLsdaData>],
) -> Option<(CustomSection, HashMap<u32, u32>)> {
    let mut unique_tags = HashSet::new();
    for data in lsda_data.iter().flatten() {
        for reloc in &data.relocations {
            unique_tags.insert(reloc.tag);
        }
    }

    if unique_tags.is_empty() {
        return None;
    }

    let mut tags: Vec<u32> = unique_tags.into_iter().collect();
    tags.sort_unstable();

    let mut bytes = Vec::with_capacity(tags.len() * std::mem::size_of::<u32>());
    let mut offsets = HashMap::new();
    for tag in tags {
        let offset = bytes.len() as u32;
        bytes.extend_from_slice(&tag.to_ne_bytes());
        offsets.insert(tag, offset);
    }

    let section = CustomSection {
        protection: CustomSectionProtection::Read,
        alignment: None,
        bytes: SectionBody::new_with_vec(bytes),
        relocations: Vec::new(),
    };

    Some((section, offsets))
}

/// Build the LSDA custom section and record the offset for each function.
///
/// Returns the section (if any) and a vector mapping each function index to
/// its LSDA offset inside the section. Even when utilizing the same landing pad for exception tags,
/// Cranelift generates separate landing pad locations.
/// These locations are essentially small trampolines that redirect to the basic block we established (the EH dispatch block).
///
/// The section can be dumped using the elfutils' readelf tool:
/// ```shell
/// objcopy -I binary -O elf64-x86-64 --rename-section .data=.gcc_except_table,alloc,contents lsda.bin object.o && eu-readelf -w object.o
/// ```
pub fn build_lsda_section(
    lsda_data: Vec<Option<FunctionLsdaData>>,
    pointer_bytes: u8,
    tag_offsets: &HashMap<u32, u32>,
    tag_section_index: Option<SectionIndex>,
) -> (Option<CustomSection>, Vec<Option<u32>>) {
    let mut bytes = Vec::new();
    let mut relocations = Vec::new();
    let mut offsets_per_function = Vec::with_capacity(lsda_data.len());

    let pointer_kind = match pointer_bytes {
        4 => RelocationKind::Abs4,
        8 => RelocationKind::Abs8,
        other => panic!("unsupported pointer size {other} for LSDA generation"),
    };

    for data in lsda_data.into_iter() {
        if let Some(data) = data {
            let base = bytes.len() as u32;
            bytes.extend_from_slice(&data.bytes);

            for reloc in &data.relocations {
                let target_offset = tag_offsets
                    .get(&reloc.tag)
                    .copied()
                    .expect("missing tag offset for relocation");
                relocations.push(Relocation {
                    kind: pointer_kind,
                    reloc_target: RelocationTarget::CustomSection(
                        tag_section_index
                            .expect("tag section index must exist when relocations are present"),
                    ),
                    offset: base + reloc.offset,
                    addend: target_offset as i64,
                });
            }

            offsets_per_function.push(Some(base));
        } else {
            offsets_per_function.push(None);
        }
    }

    if bytes.is_empty() {
        (None, offsets_per_function)
    } else {
        (
            Some(CustomSection {
                protection: CustomSectionProtection::Read,
                alignment: None,
                bytes: SectionBody::new_with_vec(bytes),
                relocations,
            }),
            offsets_per_function,
        )
    }
}

#[derive(Debug, Clone)]
pub struct CompactUnwindEntryData {
    pub function: LocalFunctionIndex,
    pub function_length: u32,
    pub compact_encoding: u32,
    pub lsda_offset: Option<u32>,
}

/// Build the 64-bit Mach-O `__compact_unwind` section consumed by the
/// runtime compact-unwind publisher.
pub fn build_compact_unwind_section(
    entries: impl IntoIterator<Item = CompactUnwindEntryData>,
    lsda_section_index: Option<SectionIndex>,
) -> Option<CustomSection> {
    const ENTRY_SIZE: usize = 32;
    const FUNCTION_ADDR_OFFSET: u32 = 0;
    const PERSONALITY_ADDR_OFFSET: u32 = 16;
    const LSDA_ADDR_OFFSET: u32 = 24;

    let entries = entries.into_iter().collect::<Vec<_>>();
    if entries.is_empty() {
        return None;
    }

    let mut bytes = Vec::with_capacity(entries.len() * ENTRY_SIZE);
    let mut relocations = Vec::new();

    for entry in entries {
        let base = bytes.len() as u32;

        bytes.extend_from_slice(&0u64.to_le_bytes());
        bytes.extend_from_slice(&entry.function_length.to_le_bytes());
        bytes.extend_from_slice(&entry.compact_encoding.to_le_bytes());
        bytes.extend_from_slice(&0u64.to_le_bytes());
        bytes.extend_from_slice(&0u64.to_le_bytes());

        relocations.push(Relocation {
            kind: RelocationKind::Abs8,
            reloc_target: RelocationTarget::LocalFunc(entry.function),
            offset: base + FUNCTION_ADDR_OFFSET,
            addend: 0,
        });
        relocations.push(Relocation {
            kind: RelocationKind::Abs8,
            reloc_target: RelocationTarget::LibCall(LibCall::EHPersonality),
            offset: base + PERSONALITY_ADDR_OFFSET,
            addend: 0,
        });

        if let Some(lsda_offset) = entry.lsda_offset {
            relocations.push(Relocation {
                kind: RelocationKind::Abs8,
                reloc_target: RelocationTarget::CustomSection(
                    lsda_section_index.expect("LSDA section index required for LSDA relocation"),
                ),
                offset: base + LSDA_ADDR_OFFSET,
                addend: lsda_offset as i64,
            });
        }
    }

    Some(CustomSection {
        protection: CustomSectionProtection::Read,
        alignment: Some(8),
        bytes: SectionBody::new_with_vec(bytes),
        relocations,
    })
}

// Constants are defined in compact_unwind_encoding.h file.
const UNWIND_ARM64_MODE_FRAMELESS: u32 = 0x02000000;
const UNWIND_ARM64_MODE_FRAME: u32 = 0x04000000;

const UNWIND_ARM64_FRAMELESS_STACK_SIZE_SHIFT: u32 = 12;
const UNWIND_ARM64_FRAME_X19_X20_PAIR: u32 = 0x00000001;
const UNWIND_ARM64_FRAME_X21_X22_PAIR: u32 = 0x00000002;
const UNWIND_ARM64_FRAME_X23_X24_PAIR: u32 = 0x00000004;
const UNWIND_ARM64_FRAME_X25_X26_PAIR: u32 = 0x00000008;
const UNWIND_ARM64_FRAME_X27_X28_PAIR: u32 = 0x00000010;
const UNWIND_ARM64_FRAME_D8_D9_PAIR: u32 = 0x00000100;
const UNWIND_ARM64_FRAME_D10_D11_PAIR: u32 = 0x00000200;
const UNWIND_ARM64_FRAME_D12_D13_PAIR: u32 = 0x00000400;
const UNWIND_ARM64_FRAME_D14_D15_PAIR: u32 = 0x00000800;

const STACK_SIZE_UNIT: u32 = 16;

pub fn compact_unwind_encoding_aarch64(unwind_info: &[(u32, UnwindInst)]) -> Result<u32, String> {
    let mut has_frame = false;
    let mut stack_size = 0u32;
    let mut saved_int = HashSet::new();
    let mut saved_float = HashSet::new();

    for (_, inst) in unwind_info {
        match inst {
            UnwindInst::PushFrameRegs { .. } | UnwindInst::DefineNewFrame { .. } => {
                has_frame = true;
            }
            UnwindInst::StackAlloc { size } => {
                stack_size = stack_size
                    .checked_add(*size)
                    .ok_or_else(|| "aarch64 compact-unwind stack size overflow".to_string())?;
            }
            UnwindInst::SaveReg { reg, .. } => match reg.class() {
                regalloc2::RegClass::Int => {
                    saved_int.insert(reg.hw_enc());
                }
                regalloc2::RegClass::Float => {
                    saved_float.insert(reg.hw_enc());
                }
                regalloc2::RegClass::Vector => {
                    return Err(
                        "aarch64 compact-unwind cannot encode vector register saves".to_owned()
                    );
                }
            },
            UnwindInst::RegStackOffset { .. } => {
                return Err("aarch64 compact-unwind cannot encode RegStackOffset".to_owned());
            }
            UnwindInst::Aarch64SetPointerAuth { .. } => {}
        }
    }

    if !has_frame {
        if !saved_int.is_empty() || !saved_float.is_empty() {
            return Err("aarch64 frameless compact-unwind cannot encode saved registers".into());
        }
        if !stack_size.is_multiple_of(STACK_SIZE_UNIT) {
            return Err("aarch64 compact-unwind stack size must be 16-byte aligned".into());
        }
        let stack_units = stack_size / STACK_SIZE_UNIT;
        if stack_units > 0x0fff {
            return Err("aarch64 compact-unwind stack size is too large".into());
        }
        return Ok(
            UNWIND_ARM64_MODE_FRAMELESS | (stack_units << UNWIND_ARM64_FRAMELESS_STACK_SIZE_SHIFT)
        );
    }

    let encode_saved_pair = |saved: &mut HashSet<_>, lo, hi, bit, class_name| match (
        saved.remove(&lo),
        saved.remove(&hi),
    ) {
        (false, false) => Ok(0),
        (true, true) => Ok(bit),
        _ => Err(format!(
            "aarch64 compact-unwind cannot encode unpaired {class_name}{lo}/{class_name}{hi} save"
        )),
    };

    let mut encoding = UNWIND_ARM64_MODE_FRAME;
    for (lo, hi, bit) in [
        (19, 20, UNWIND_ARM64_FRAME_X19_X20_PAIR),
        (21, 22, UNWIND_ARM64_FRAME_X21_X22_PAIR),
        (23, 24, UNWIND_ARM64_FRAME_X23_X24_PAIR),
        (25, 26, UNWIND_ARM64_FRAME_X25_X26_PAIR),
        (27, 28, UNWIND_ARM64_FRAME_X27_X28_PAIR),
    ] {
        encoding |= encode_saved_pair(&mut saved_int, lo, hi, bit, "x")?;
    }
    for (lo, hi, bit) in [
        (8, 9, UNWIND_ARM64_FRAME_D8_D9_PAIR),
        (10, 11, UNWIND_ARM64_FRAME_D10_D11_PAIR),
        (12, 13, UNWIND_ARM64_FRAME_D12_D13_PAIR),
        (14, 15, UNWIND_ARM64_FRAME_D14_D15_PAIR),
    ] {
        encoding |= encode_saved_pair(&mut saved_float, lo, hi, bit, "d")?;
    }

    if !saved_int.is_empty() || !saved_float.is_empty() {
        return Err("aarch64 compact-unwind encountered unsupported saved register".to_owned());
    }

    Ok(encoding)
}

#[derive(Debug)]
struct CallSiteDesc {
    start: u32,
    len: u32,
    landing_pad: u32,
    actions: Vec<ExceptionType>,
}

#[derive(Debug, Hash, PartialEq, Eq, Clone, Copy)]
enum ExceptionType {
    Tag { tag: u32 },
    CatchAll,
}

#[derive(Debug)]
struct TypeTable {
    entries: indexmap::IndexSet<ExceptionType>,
}

impl TypeTable {
    fn new() -> Self {
        Self {
            entries: indexmap::IndexSet::new(),
        }
    }

    fn is_empty(&self) -> bool {
        self.entries.is_empty()
    }

    fn get_or_insert(&mut self, exception: ExceptionType) -> usize {
        self.entries.insert(exception);

        // The indices are one-based!
        self.entries
            .get_index_of(&exception)
            .expect("must be already inserted")
            + 1
    }

    fn encode(&self, pointer_bytes: u8) -> (Vec<u8>, Vec<TagRelocation>) {
        let mut bytes = Vec::with_capacity(self.entries.len() * pointer_bytes as usize);
        let mut relocations = Vec::new();

        // Note the exception types must be streamed in the reverse order!
        for entry in self.entries.iter().rev() {
            let offset = bytes.len() as u32;
            match entry {
                ExceptionType::Tag { tag } => {
                    bytes.extend(std::iter::repeat_n(0, pointer_bytes as usize));
                    relocations.push(TagRelocation { offset, tag: *tag });
                }
                ExceptionType::CatchAll => {
                    bytes.extend(std::iter::repeat_n(0, pointer_bytes as usize));
                }
            }
        }

        (bytes, relocations)
    }

    /// Encode the type table as PC-relative, 4-byte slots resolved through
    /// relocations against the per-object tag section. Used by the ELF
    /// artifact format, where each function's LSDA lives in its own object.
    fn encode_relocated(&self) -> (Vec<u8>, Vec<TagRelocation>) {
        const ENTRY_SIZE: usize = 4;
        let mut bytes = Vec::with_capacity(self.entries.len() * ENTRY_SIZE);
        let mut relocations = Vec::new();

        // Note the exception types must be streamed in the reverse order!
        for entry in self.entries.iter().rev() {
            let offset = bytes.len() as u32;
            match entry {
                ExceptionType::Tag { tag } => {
                    bytes.extend(std::iter::repeat_n(0, ENTRY_SIZE));
                    relocations.push(TagRelocation { offset, tag: *tag });
                }
                ExceptionType::CatchAll => {
                    bytes.extend(std::iter::repeat_n(0, ENTRY_SIZE));
                }
            }
        }

        (bytes, relocations)
    }
}

struct ActionTable {
    bytes: Vec<u8>,
    first_action_offsets: Vec<Option<u32>>,
}

fn encode_action_table(callsite_actions: &[Vec<i32>]) -> ActionTable {
    let mut writer = Cursor::new(Vec::new());
    let mut first_action_offsets = Vec::new();

    let mut cache = HashMap::new();

    for actions in callsite_actions {
        if actions.is_empty() {
            first_action_offsets.push(None);
        } else {
            match cache.entry(actions.clone()) {
                Entry::Occupied(entry) => {
                    first_action_offsets.push(Some(*entry.get()));
                }
                Entry::Vacant(entry) => {
                    let mut last_action_start = 0;
                    for (i, &ttype_index) in actions.iter().enumerate() {
                        let next_action_start = writer.position();
                        leb128::write::signed(&mut writer, ttype_index as i64)
                            .expect("leb128 write failed");

                        if i != 0 {
                            // Make a linked list to the previous action
                            let displacement = last_action_start - writer.position() as i64;
                            leb128::write::signed(&mut writer, displacement)
                                .expect("leb128 write failed");
                        } else {
                            leb128::write::signed(&mut writer, 0).expect("leb128 write failed");
                        }
                        last_action_start = next_action_start as i64;
                    }
                    let last_action_start = last_action_start as u32;
                    entry.insert(last_action_start);
                    first_action_offsets.push(Some(last_action_start));
                }
            }
        }
    }

    ActionTable {
        bytes: writer.into_inner(),
        first_action_offsets,
    }
}

fn encode_call_site_table(callsites: &[CallSiteDesc], action_table: &ActionTable) -> Vec<u8> {
    let mut writer = Cursor::new(Vec::new());
    for (idx, site) in callsites.iter().enumerate() {
        write_encoded_offset(site.start, &mut writer);
        write_encoded_offset(site.len, &mut writer);
        write_encoded_offset(site.landing_pad, &mut writer);

        let action = match action_table.first_action_offsets[idx] {
            Some(offset) => offset as u64 + 1,
            None => 0,
        };
        leb128::write::unsigned(&mut writer, action).expect("leb128 write failed");
    }
    writer.into_inner()
}

fn write_encoded_offset(val: u32, out: &mut impl Write) {
    // We use DW_EH_PE_udata4 for all offsets.
    out.write_all(&val.to_le_bytes())
        .expect("write to buffer failed")
}

fn uleb128_len(value: u64) -> usize {
    let mut cursor = Cursor::new([0u8; 10]);
    leb128::write::unsigned(&mut cursor, value).unwrap()
}