Files
wasmtime/cranelift/src/bugpoint.rs

823 lines
28 KiB
Rust

//! CLI tool to reduce Cranelift IR files crashing during compilation.
use crate::disasm::{PrintRelocs, PrintStackmaps, PrintTraps};
use crate::utils::{parse_sets_and_triple, read_to_string};
use cranelift_codegen::cursor::{Cursor, FuncCursor};
use cranelift_codegen::ir::types::{F32, F64};
use cranelift_codegen::ir::{
self, Ebb, FuncRef, Function, GlobalValueData, Inst, InstBuilder, InstructionData, StackSlots,
TrapCode,
};
use cranelift_codegen::isa::TargetIsa;
use cranelift_codegen::Context;
use cranelift_entity::PrimaryMap;
use cranelift_reader::{parse_test, ParseOptions};
use std::collections::HashMap;
use std::path::Path;
use indicatif::{ProgressBar, ProgressDrawTarget, ProgressStyle};
pub fn run(
filename: &str,
flag_set: &[String],
flag_isa: &str,
verbose: bool,
) -> Result<(), String> {
let parsed = parse_sets_and_triple(flag_set, flag_isa)?;
let fisa = parsed.as_fisa();
let path = Path::new(&filename).to_path_buf();
let buffer = read_to_string(&path).map_err(|e| format!("{}: {}", filename, e))?;
let test_file =
parse_test(&buffer, ParseOptions::default()).map_err(|e| format!("{}: {}", filename, e))?;
// If we have an isa from the command-line, use that. Otherwise if the
// file contains a unique isa, use that.
let isa = if let Some(isa) = fisa.isa {
isa
} else if let Some(isa) = test_file.isa_spec.unique_isa() {
isa
} else {
return Err(String::from("compilation requires a target isa"));
};
std::env::set_var("RUST_BACKTRACE", "0"); // Disable backtraces to reduce verbosity
for (func, _) in test_file.functions {
let (orig_ebb_count, orig_inst_count) = (ebb_count(&func), inst_count(&func));
match reduce(isa, func, verbose) {
Ok((func, crash_msg)) => {
println!("Crash message: {}", crash_msg);
println!("\n{}", func);
println!(
"{} ebbs {} insts -> {} ebbs {} insts",
orig_ebb_count,
orig_inst_count,
ebb_count(&func),
inst_count(&func)
);
}
Err(err) => println!("Warning: {}", err),
}
}
Ok(())
}
enum ProgressStatus {
/// The mutation raised or reduced the amount of instructions or ebbs.
ExpandedOrShrinked,
/// The mutation only changed an instruction. Performing another round of mutations may only
/// reduce the test case if another mutation shrank the test case.
Changed,
/// No need to re-test if the program crashes, because the mutation had no effect, but we want
/// to keep on iterating.
Skip,
}
trait Mutator {
fn name(&self) -> &'static str;
fn mutation_count(&self, func: &Function) -> usize;
fn mutate(&mut self, func: Function) -> Option<(Function, String, ProgressStatus)>;
}
/// Try to remove instructions.
struct RemoveInst {
ebb: Ebb,
inst: Inst,
}
impl RemoveInst {
fn new(func: &Function) -> Self {
let first_ebb = func.layout.entry_block().unwrap();
let first_inst = func.layout.first_inst(first_ebb).unwrap();
Self {
ebb: first_ebb,
inst: first_inst,
}
}
}
impl Mutator for RemoveInst {
fn name(&self) -> &'static str {
"remove inst"
}
fn mutation_count(&self, func: &Function) -> usize {
inst_count(func)
}
fn mutate(&mut self, mut func: Function) -> Option<(Function, String, ProgressStatus)> {
next_inst_ret_prev(&func, &mut self.ebb, &mut self.inst).map(|(prev_ebb, prev_inst)| {
func.layout.remove_inst(prev_inst);
let msg = if func.layout.ebb_insts(prev_ebb).next().is_none() {
// Make sure empty ebbs are removed, as `next_inst_ret_prev` depends on non empty ebbs
func.layout.remove_ebb(prev_ebb);
format!("Remove inst {} and empty ebb {}", prev_inst, prev_ebb)
} else {
format!("Remove inst {}", prev_inst)
};
(func, msg, ProgressStatus::ExpandedOrShrinked)
})
}
}
/// Try to replace instructions with `iconst` or `fconst`.
struct ReplaceInstWithConst {
ebb: Ebb,
inst: Inst,
}
impl ReplaceInstWithConst {
fn new(func: &Function) -> Self {
let first_ebb = func.layout.entry_block().unwrap();
let first_inst = func.layout.first_inst(first_ebb).unwrap();
Self {
ebb: first_ebb,
inst: first_inst,
}
}
fn const_for_type<'f, T: InstBuilder<'f>>(builder: T, ty: ir::Type) -> &'static str {
// Try to keep the result type consistent, and default to an integer type
// otherwise: this will cover all the cases for f32/f64 and integer types, or
// create verifier errors otherwise.
if ty == F32 {
builder.f32const(0.0);
"f32const"
} else if ty == F64 {
builder.f64const(0.0);
"f64const"
} else {
builder.iconst(ty, 0);
"iconst"
}
}
}
impl Mutator for ReplaceInstWithConst {
fn name(&self) -> &'static str {
"replace inst with const"
}
fn mutation_count(&self, func: &Function) -> usize {
inst_count(func)
}
fn mutate(&mut self, mut func: Function) -> Option<(Function, String, ProgressStatus)> {
next_inst_ret_prev(&func, &mut self.ebb, &mut self.inst).map(|(_prev_ebb, prev_inst)| {
let num_results = func.dfg.inst_results(prev_inst).len();
let opcode = func.dfg[prev_inst].opcode();
if num_results == 0
|| opcode == ir::Opcode::Iconst
|| opcode == ir::Opcode::F32const
|| opcode == ir::Opcode::F64const
{
return (func, format!(""), ProgressStatus::Skip);
}
if num_results == 1 {
let ty = func.dfg.value_type(func.dfg.first_result(prev_inst));
let new_inst_name = Self::const_for_type(func.dfg.replace(prev_inst), ty);
return (
func,
format!("Replace inst {} with {}.", prev_inst, new_inst_name),
ProgressStatus::Changed,
);
}
// At least 2 results. Replace each instruction with as many const instructions as
// there are results.
let mut pos = FuncCursor::new(&mut func).at_inst(prev_inst);
// Copy result SSA names into our own vector; otherwise we couldn't mutably borrow pos
// in the loop below.
let results = pos
.func
.dfg
.inst_results(prev_inst)
.iter()
.cloned()
.collect::<Vec<_>>();
// Detach results from the previous instruction, since we're going to reuse them.
pos.func.dfg.clear_results(prev_inst);
let mut inst_names = Vec::new();
for r in results {
let ty = pos.func.dfg.value_type(r);
let builder = pos.ins().with_results([Some(r)]);
let new_inst_name = Self::const_for_type(builder, ty);
inst_names.push(new_inst_name);
}
// Remove the instruction.
assert_eq!(pos.remove_inst(), prev_inst);
(
func,
format!("Replace inst {} with {}", prev_inst, inst_names.join(" / ")),
ProgressStatus::ExpandedOrShrinked,
)
})
}
}
/// Try to replace instructions with `trap`.
struct ReplaceInstWithTrap {
ebb: Ebb,
inst: Inst,
}
impl ReplaceInstWithTrap {
fn new(func: &Function) -> Self {
let first_ebb = func.layout.entry_block().unwrap();
let first_inst = func.layout.first_inst(first_ebb).unwrap();
Self {
ebb: first_ebb,
inst: first_inst,
}
}
}
impl Mutator for ReplaceInstWithTrap {
fn name(&self) -> &'static str {
"replace inst with trap"
}
fn mutation_count(&self, func: &Function) -> usize {
inst_count(func)
}
fn mutate(&mut self, mut func: Function) -> Option<(Function, String, ProgressStatus)> {
next_inst_ret_prev(&func, &mut self.ebb, &mut self.inst).map(|(_prev_ebb, prev_inst)| {
let status = if func.dfg[prev_inst].opcode() == ir::Opcode::Trap {
ProgressStatus::Skip
} else {
func.dfg.replace(prev_inst).trap(TrapCode::User(0));
ProgressStatus::Changed
};
(
func,
format!("Replace inst {} with trap", prev_inst),
status,
)
})
}
}
/// Try to remove an ebb.
struct RemoveEbb {
ebb: Ebb,
}
impl RemoveEbb {
fn new(func: &Function) -> Self {
Self {
ebb: func.layout.entry_block().unwrap(),
}
}
}
impl Mutator for RemoveEbb {
fn name(&self) -> &'static str {
"remove ebb"
}
fn mutation_count(&self, func: &Function) -> usize {
ebb_count(func)
}
fn mutate(&mut self, mut func: Function) -> Option<(Function, String, ProgressStatus)> {
func.layout.next_ebb(self.ebb).map(|next_ebb| {
self.ebb = next_ebb;
while let Some(inst) = func.layout.last_inst(self.ebb) {
func.layout.remove_inst(inst);
}
func.layout.remove_ebb(self.ebb);
(
func,
format!("Remove ebb {}", next_ebb),
ProgressStatus::ExpandedOrShrinked,
)
})
}
}
/// Try to remove unused entities.
struct RemoveUnusedEntities {
kind: u32,
}
impl RemoveUnusedEntities {
fn new() -> Self {
Self { kind: 0 }
}
}
impl Mutator for RemoveUnusedEntities {
fn name(&self) -> &'static str {
"remove unused entities"
}
fn mutation_count(&self, _func: &Function) -> usize {
4
}
fn mutate(&mut self, mut func: Function) -> Option<(Function, String, ProgressStatus)> {
let name = match self.kind {
0 => {
let mut ext_func_usage_map = HashMap::new();
for ebb in func.layout.ebbs() {
for inst in func.layout.ebb_insts(ebb) {
match func.dfg[inst] {
// Add new cases when there are new instruction formats taking a `FuncRef`.
InstructionData::Call { func_ref, .. }
| InstructionData::FuncAddr { func_ref, .. } => {
ext_func_usage_map
.entry(func_ref)
.or_insert_with(Vec::new)
.push(inst);
}
_ => {}
}
}
}
let mut ext_funcs = PrimaryMap::new();
for (func_ref, ext_func_data) in func.dfg.ext_funcs.clone().into_iter() {
if let Some(func_ref_usage) = ext_func_usage_map.get(&func_ref) {
let new_func_ref = ext_funcs.push(ext_func_data.clone());
for &inst in func_ref_usage {
match func.dfg[inst] {
// Keep in sync with the above match.
InstructionData::Call {
ref mut func_ref, ..
}
| InstructionData::FuncAddr {
ref mut func_ref, ..
} => {
*func_ref = new_func_ref;
}
_ => unreachable!(),
}
}
}
}
func.dfg.ext_funcs = ext_funcs;
"Remove unused ext funcs"
}
1 => {
#[derive(Copy, Clone)]
enum SigRefUser {
Instruction(Inst),
ExtFunc(FuncRef),
}
let mut signatures_usage_map = HashMap::new();
for ebb in func.layout.ebbs() {
for inst in func.layout.ebb_insts(ebb) {
match func.dfg[inst] {
// Add new cases when there are new instruction formats taking a `SigRef`.
InstructionData::CallIndirect { sig_ref, .. } => {
signatures_usage_map
.entry(sig_ref)
.or_insert_with(Vec::new)
.push(SigRefUser::Instruction(inst));
}
_ => {}
}
}
}
for (func_ref, ext_func_data) in func.dfg.ext_funcs.iter() {
signatures_usage_map
.entry(ext_func_data.signature)
.or_insert_with(Vec::new)
.push(SigRefUser::ExtFunc(func_ref));
}
let mut signatures = PrimaryMap::new();
for (sig_ref, sig_data) in func.dfg.signatures.clone().into_iter() {
if let Some(sig_ref_usage) = signatures_usage_map.get(&sig_ref) {
let new_sig_ref = signatures.push(sig_data.clone());
for &sig_ref_user in sig_ref_usage {
match sig_ref_user {
SigRefUser::Instruction(inst) => match func.dfg[inst] {
// Keep in sync with the above match.
InstructionData::CallIndirect {
ref mut sig_ref, ..
} => {
*sig_ref = new_sig_ref;
}
_ => unreachable!(),
},
SigRefUser::ExtFunc(func_ref) => {
func.dfg.ext_funcs[func_ref].signature = new_sig_ref;
}
}
}
}
}
func.dfg.signatures = signatures;
"Remove unused signatures"
}
2 => {
let mut stack_slot_usage_map = HashMap::new();
for ebb in func.layout.ebbs() {
for inst in func.layout.ebb_insts(ebb) {
match func.dfg[inst] {
// Add new cases when there are new instruction formats taking a `StackSlot`.
InstructionData::StackLoad { stack_slot, .. }
| InstructionData::StackStore { stack_slot, .. } => {
stack_slot_usage_map
.entry(stack_slot)
.or_insert_with(Vec::new)
.push(inst);
}
InstructionData::RegSpill { dst, .. } => {
stack_slot_usage_map
.entry(dst)
.or_insert_with(Vec::new)
.push(inst);
}
InstructionData::RegFill { src, .. } => {
stack_slot_usage_map
.entry(src)
.or_insert_with(Vec::new)
.push(inst);
}
_ => {}
}
}
}
let mut stack_slots = StackSlots::new();
for (stack_slot, stack_slot_data) in func.stack_slots.clone().iter() {
if let Some(stack_slot_usage) = stack_slot_usage_map.get(&stack_slot) {
let new_stack_slot = stack_slots.push(stack_slot_data.clone());
for &inst in stack_slot_usage {
match &mut func.dfg[inst] {
// Keep in sync with the above match.
InstructionData::StackLoad { stack_slot, .. }
| InstructionData::StackStore { stack_slot, .. } => {
*stack_slot = new_stack_slot;
}
InstructionData::RegSpill { dst, .. } => {
*dst = new_stack_slot;
}
InstructionData::RegFill { src, .. } => {
*src = new_stack_slot;
}
_ => unreachable!(),
}
}
}
}
func.stack_slots = stack_slots;
"Remove unused stack slots"
}
3 => {
let mut global_value_usage_map = HashMap::new();
for ebb in func.layout.ebbs() {
for inst in func.layout.ebb_insts(ebb) {
match func.dfg[inst] {
// Add new cases when there are new instruction formats taking a `GlobalValue`.
InstructionData::UnaryGlobalValue { global_value, .. } => {
global_value_usage_map
.entry(global_value)
.or_insert_with(Vec::new)
.push(inst);
}
_ => {}
}
}
}
for (_global_value, global_value_data) in func.global_values.iter() {
match *global_value_data {
GlobalValueData::VMContext | GlobalValueData::Symbol { .. } => {}
// These can create cyclic references, which cause complications. Just skip
// the global value removal for now.
// FIXME Handle them in a better way.
GlobalValueData::Load { base: _, .. }
| GlobalValueData::IAddImm { base: _, .. } => return None,
}
}
let mut global_values = PrimaryMap::new();
for (global_value, global_value_data) in func.global_values.clone().into_iter() {
if let Some(global_value_usage) = global_value_usage_map.get(&global_value) {
let new_global_value = global_values.push(global_value_data.clone());
for &inst in global_value_usage {
match &mut func.dfg[inst] {
// Keep in sync with the above match.
InstructionData::UnaryGlobalValue { global_value, .. } => {
*global_value = new_global_value;
}
_ => unreachable!(),
}
}
}
}
func.global_values = global_values;
"Remove unused global values"
}
_ => return None,
};
self.kind += 1;
Some((func, name.to_owned(), ProgressStatus::Changed))
}
}
fn next_inst_ret_prev(func: &Function, ebb: &mut Ebb, inst: &mut Inst) -> Option<(Ebb, Inst)> {
let prev = (*ebb, *inst);
if let Some(next_inst) = func.layout.next_inst(*inst) {
*inst = next_inst;
return Some(prev);
}
if let Some(next_ebb) = func.layout.next_ebb(*ebb) {
*ebb = next_ebb;
*inst = func.layout.first_inst(*ebb).expect("no inst");
return Some(prev);
}
None
}
fn ebb_count(func: &Function) -> usize {
func.layout.ebbs().count()
}
fn inst_count(func: &Function) -> usize {
func.layout
.ebbs()
.map(|ebb| func.layout.ebb_insts(ebb).count())
.sum()
}
fn resolve_aliases(func: &mut Function) {
for ebb in func.layout.ebbs() {
for inst in func.layout.ebb_insts(ebb) {
func.dfg.resolve_aliases_in_arguments(inst);
}
}
}
fn reduce(
isa: &dyn TargetIsa,
mut func: Function,
verbose: bool,
) -> Result<(Function, String), String> {
let mut context = CrashCheckContext::new(isa);
match context.check_for_crash(&func) {
CheckResult::Succeed => {
return Err(format!(
"Given function compiled successfully or gave a verifier error."
));
}
CheckResult::Crash(_) => {}
}
resolve_aliases(&mut func);
let progress_bar = ProgressBar::with_draw_target(0, ProgressDrawTarget::stdout());
progress_bar.set_style(
ProgressStyle::default_bar().template("{bar:60} {prefix:40} {pos:>4}/{len:>4} {msg}"),
);
for pass_idx in 0..100 {
let mut should_keep_reducing = false;
let mut phase = 0;
loop {
let mut mutator: Box<dyn Mutator> = match phase {
0 => Box::new(RemoveInst::new(&func)),
1 => Box::new(ReplaceInstWithConst::new(&func)),
2 => Box::new(ReplaceInstWithTrap::new(&func)),
3 => Box::new(RemoveEbb::new(&func)),
4 => Box::new(RemoveUnusedEntities::new()),
_ => break,
};
progress_bar.set_prefix(&format!("pass {} phase {}", pass_idx, mutator.name()));
progress_bar.set_length(mutator.mutation_count(&func) as u64);
// Reset progress bar.
progress_bar.set_position(0);
progress_bar.set_draw_delta(0);
for _ in 0..10000 {
progress_bar.inc(1);
let (mutated_func, msg, mutation_kind) = match mutator.mutate(func.clone()) {
Some(res) => res,
None => {
break;
}
};
if let ProgressStatus::Skip = mutation_kind {
// The mutator didn't change anything, but we want to try more mutator
// iterations.
continue;
}
progress_bar.set_message(&msg);
match context.check_for_crash(&mutated_func) {
CheckResult::Succeed => {
// Shrinking didn't hit the problem anymore, discard changes.
continue;
}
CheckResult::Crash(_) => {
// Panic remained while shrinking, make changes definitive.
func = mutated_func;
let verb = match mutation_kind {
ProgressStatus::ExpandedOrShrinked => {
should_keep_reducing = true;
"shrink"
}
ProgressStatus::Changed => "changed",
ProgressStatus::Skip => unreachable!(),
};
if verbose {
progress_bar.println(format!("{}: {}", msg, verb));
}
}
}
}
phase += 1;
}
progress_bar.println(format!(
"After pass {}, remaining insts/ebbs: {}/{}",
pass_idx,
inst_count(&func),
ebb_count(&func)
));
if !should_keep_reducing {
// No new shrinking opportunities have been found this pass. This means none will ever
// be found. Skip the rest of the passes over the function.
break;
}
}
progress_bar.finish();
let crash_msg = match context.check_for_crash(&func) {
CheckResult::Succeed => unreachable!("Used to crash, but doesn't anymore???"),
CheckResult::Crash(crash_msg) => crash_msg,
};
Ok((func, crash_msg))
}
struct CrashCheckContext<'a> {
/// Cached `Context`, to prevent repeated allocation.
context: Context,
/// Cached code memory, to prevent repeated allocation.
code_memory: Vec<u8>,
/// The target isa to compile for.
isa: &'a dyn TargetIsa,
}
fn get_panic_string(panic: Box<dyn std::any::Any>) -> String {
let panic = match panic.downcast::<&'static str>() {
Ok(panic_msg) => {
return panic_msg.to_string();
}
Err(panic) => panic,
};
match panic.downcast::<String>() {
Ok(panic_msg) => *panic_msg,
Err(_) => "Box<Any>".to_string(),
}
}
enum CheckResult {
/// The function compiled fine, or the verifier noticed an error.
Succeed,
/// The compilation of the function panicked.
Crash(String),
}
impl<'a> CrashCheckContext<'a> {
fn new(isa: &'a dyn TargetIsa) -> Self {
CrashCheckContext {
context: Context::new(),
code_memory: Vec::new(),
isa,
}
}
#[cfg_attr(test, allow(unreachable_code))]
fn check_for_crash(&mut self, func: &Function) -> CheckResult {
self.context.clear();
self.code_memory.clear();
self.context.func = func.clone();
use std::io::Write;
std::io::stdout().flush().unwrap(); // Flush stdout to sync with panic messages on stderr
match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
cranelift_codegen::verifier::verify_function(&func, self.isa).err()
})) {
Ok(Some(_)) => return CheckResult::Succeed,
Ok(None) => {}
// The verifier panicked. Compiling it will probably give the same panic.
// We treat it as succeeding to make it possible to reduce for the actual error.
// FIXME prevent verifier panic on removing ebb0.
Err(_) => return CheckResult::Succeed,
}
#[cfg(test)]
{
// For testing purposes we emulate a panic caused by the existence of
// a `call` instruction.
let contains_call = func.layout.ebbs().any(|ebb| {
func.layout.ebb_insts(ebb).any(|inst| match func.dfg[inst] {
InstructionData::Call { .. } => true,
_ => false,
})
});
if contains_call {
return CheckResult::Crash("test crash".to_string());
} else {
return CheckResult::Succeed;
}
}
let old_panic_hook = std::panic::take_hook();
std::panic::set_hook(Box::new(|_| {})); // silence panics
let res = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
let mut relocs = PrintRelocs::new(false);
let mut traps = PrintTraps::new(false);
let mut stackmaps = PrintStackmaps::new(false);
let _ = self.context.compile_and_emit(
self.isa,
&mut self.code_memory,
&mut relocs,
&mut traps,
&mut stackmaps,
);
})) {
Ok(()) => CheckResult::Succeed,
Err(err) => CheckResult::Crash(get_panic_string(err)),
};
std::panic::set_hook(old_panic_hook);
res
}
}
#[cfg(test)]
mod tests {
use super::*;
use cranelift_reader::ParseOptions;
#[test]
fn test_reduce() {
const TEST: &'static str = include_str!("../tests/bugpoint_test.clif");
const EXPECTED: &'static str = include_str!("../tests/bugpoint_test_expected.clif");
let test_file = parse_test(TEST, ParseOptions::default()).unwrap();
// If we have an isa from the command-line, use that. Otherwise if the
// file contains a unique isa, use that.
let isa = test_file.isa_spec.unique_isa().expect("Unknown isa");
for (func, _) in test_file.functions {
let (func, crash_msg) = reduce(isa, func, false).expect("Couldn't reduce test case");
assert_eq!(crash_msg, "test crash");
assert_eq!(format!("{}", func), EXPECTED.replace("\r\n", "\n"));
}
}
}