Remove legalizer support from cranelift-codegen-meta

This commit is contained in:
bjorn3
2021-06-21 12:55:08 +02:00
parent d499933612
commit 18bd27e90b
20 changed files with 15 additions and 4630 deletions

View File

@@ -1,751 +0,0 @@
use crate::cdsl::instructions::{InstSpec, Instruction, InstructionPredicate};
use crate::cdsl::operands::{OperandKind, OperandKindFields};
use crate::cdsl::types::ValueType;
use crate::cdsl::typevar::{TypeSetBuilder, TypeVar};
use cranelift_entity::{entity_impl, PrimaryMap, SparseMap, SparseMapValue};
use std::fmt;
use std::iter::IntoIterator;
pub(crate) enum Expr {
Var(VarIndex),
Literal(Literal),
}
impl Expr {
pub fn maybe_literal(&self) -> Option<&Literal> {
match &self {
Expr::Literal(lit) => Some(lit),
_ => None,
}
}
pub fn maybe_var(&self) -> Option<VarIndex> {
if let Expr::Var(var) = &self {
Some(*var)
} else {
None
}
}
pub fn unwrap_var(&self) -> VarIndex {
self.maybe_var()
.expect("tried to unwrap a non-Var content in Expr::unwrap_var")
}
pub fn to_rust_code(&self, var_pool: &VarPool) -> String {
match self {
Expr::Var(var_index) => var_pool.get(*var_index).to_rust_code(),
Expr::Literal(literal) => literal.to_rust_code(),
}
}
}
/// An AST definition associates a set of variables with the values produced by an expression.
pub(crate) struct Def {
pub apply: Apply,
pub defined_vars: Vec<VarIndex>,
}
impl Def {
pub fn to_comment_string(&self, var_pool: &VarPool) -> String {
let results = self
.defined_vars
.iter()
.map(|&x| var_pool.get(x).name.as_str())
.collect::<Vec<_>>();
let results = if results.len() == 1 {
results[0].to_string()
} else {
format!("({})", results.join(", "))
};
format!("{} := {}", results, self.apply.to_comment_string(var_pool))
}
}
pub(crate) struct DefPool {
pool: PrimaryMap<DefIndex, Def>,
}
impl DefPool {
pub fn new() -> Self {
Self {
pool: PrimaryMap::new(),
}
}
pub fn get(&self, index: DefIndex) -> &Def {
self.pool.get(index).unwrap()
}
pub fn next_index(&self) -> DefIndex {
self.pool.next_key()
}
pub fn create_inst(&mut self, apply: Apply, defined_vars: Vec<VarIndex>) -> DefIndex {
self.pool.push(Def {
apply,
defined_vars,
})
}
}
#[derive(Copy, Clone, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub(crate) struct DefIndex(u32);
entity_impl!(DefIndex);
/// A definition which would lead to generate a block creation.
#[derive(Clone)]
pub(crate) struct Block {
/// Instruction index after which the block entry is set.
pub location: DefIndex,
/// Variable holding the new created block.
pub name: VarIndex,
}
pub(crate) struct BlockPool {
pool: SparseMap<DefIndex, Block>,
}
impl SparseMapValue<DefIndex> for Block {
fn key(&self) -> DefIndex {
self.location
}
}
impl BlockPool {
pub fn new() -> Self {
Self {
pool: SparseMap::new(),
}
}
pub fn get(&self, index: DefIndex) -> Option<&Block> {
self.pool.get(index)
}
pub fn create_block(&mut self, name: VarIndex, location: DefIndex) {
if self.pool.contains_key(location) {
panic!("Attempt to insert 2 blocks after the same instruction")
}
self.pool.insert(Block { location, name });
}
pub fn is_empty(&self) -> bool {
self.pool.is_empty()
}
}
// Implement IntoIterator such that we can iterate over blocks which are in the block pool.
impl<'a> IntoIterator for &'a BlockPool {
type Item = <&'a SparseMap<DefIndex, Block> as IntoIterator>::Item;
type IntoIter = <&'a SparseMap<DefIndex, Block> as IntoIterator>::IntoIter;
fn into_iter(self) -> Self::IntoIter {
self.pool.into_iter()
}
}
#[derive(Clone, Debug)]
pub(crate) enum Literal {
/// A value of an enumerated immediate operand.
///
/// Some immediate operand kinds like `intcc` and `floatcc` have an enumerated range of values
/// corresponding to a Rust enum type. An `Enumerator` object is an AST leaf node representing one
/// of the values.
Enumerator {
rust_type: &'static str,
value: &'static str,
},
/// A bitwise value of an immediate operand, used for bitwise exact floating point constants.
Bits { rust_type: &'static str, value: u64 },
/// A value of an integer immediate operand.
Int(i64),
/// A empty list of variable set of arguments.
EmptyVarArgs,
}
impl Literal {
pub fn enumerator_for(kind: &OperandKind, value: &'static str) -> Self {
let value = match &kind.fields {
OperandKindFields::ImmEnum(values) => values.get(value).unwrap_or_else(|| {
panic!(
"nonexistent value '{}' in enumeration '{}'",
value, kind.rust_type
)
}),
_ => panic!("enumerator is for enum values"),
};
Literal::Enumerator {
rust_type: kind.rust_type,
value,
}
}
pub fn bits(kind: &OperandKind, bits: u64) -> Self {
match kind.fields {
OperandKindFields::ImmValue => {}
_ => panic!("bits_of is for immediate scalar types"),
}
Literal::Bits {
rust_type: kind.rust_type,
value: bits,
}
}
pub fn constant(kind: &OperandKind, value: i64) -> Self {
match kind.fields {
OperandKindFields::ImmValue => {}
_ => panic!("constant is for immediate scalar types"),
}
Literal::Int(value)
}
pub fn empty_vararg() -> Self {
Literal::EmptyVarArgs
}
pub fn to_rust_code(&self) -> String {
match self {
Literal::Enumerator { rust_type, value } => format!("{}::{}", rust_type, value),
Literal::Bits { rust_type, value } => format!("{}::with_bits({:#x})", rust_type, value),
Literal::Int(val) => val.to_string(),
Literal::EmptyVarArgs => "&[]".into(),
}
}
}
#[derive(Clone, Copy, Debug)]
pub(crate) enum PatternPosition {
Source,
Destination,
}
/// A free variable.
///
/// When variables are used in `XForms` with source and destination patterns, they are classified
/// as follows:
///
/// Input values: Uses in the source pattern with no preceding def. These may appear as inputs in
/// the destination pattern too, but no new inputs can be introduced.
///
/// Output values: Variables that are defined in both the source and destination pattern. These
/// values may have uses outside the source pattern, and the destination pattern must compute the
/// same value.
///
/// Intermediate values: Values that are defined in the source pattern, but not in the destination
/// pattern. These may have uses outside the source pattern, so the defining instruction can't be
/// deleted immediately.
///
/// Temporary values are defined only in the destination pattern.
pub(crate) struct Var {
pub name: String,
/// The `Def` defining this variable in a source pattern.
pub src_def: Option<DefIndex>,
/// The `Def` defining this variable in a destination pattern.
pub dst_def: Option<DefIndex>,
/// TypeVar representing the type of this variable.
type_var: Option<TypeVar>,
/// Is this the original type variable, or has it be redefined with set_typevar?
is_original_type_var: bool,
}
impl Var {
fn new(name: String) -> Self {
Self {
name,
src_def: None,
dst_def: None,
type_var: None,
is_original_type_var: false,
}
}
/// Is this an input value to the src pattern?
pub fn is_input(&self) -> bool {
self.src_def.is_none() && self.dst_def.is_none()
}
/// Is this an output value, defined in both src and dst patterns?
pub fn is_output(&self) -> bool {
self.src_def.is_some() && self.dst_def.is_some()
}
/// Is this an intermediate value, defined only in the src pattern?
pub fn is_intermediate(&self) -> bool {
self.src_def.is_some() && self.dst_def.is_none()
}
/// Is this a temp value, defined only in the dst pattern?
pub fn is_temp(&self) -> bool {
self.src_def.is_none() && self.dst_def.is_some()
}
/// Get the def of this variable according to the position.
pub fn get_def(&self, position: PatternPosition) -> Option<DefIndex> {
match position {
PatternPosition::Source => self.src_def,
PatternPosition::Destination => self.dst_def,
}
}
pub fn set_def(&mut self, position: PatternPosition, def: DefIndex) {
assert!(
self.get_def(position).is_none(),
"redefinition of variable {}",
self.name
);
match position {
PatternPosition::Source => {
self.src_def = Some(def);
}
PatternPosition::Destination => {
self.dst_def = Some(def);
}
}
}
/// Get the type variable representing the type of this variable.
pub fn get_or_create_typevar(&mut self) -> TypeVar {
match &self.type_var {
Some(tv) => tv.clone(),
None => {
// Create a new type var in which we allow all types.
let tv = TypeVar::new(
format!("typeof_{}", self.name),
format!("Type of the pattern variable {:?}", self),
TypeSetBuilder::all(),
);
self.type_var = Some(tv.clone());
self.is_original_type_var = true;
tv
}
}
}
pub fn get_typevar(&self) -> Option<TypeVar> {
self.type_var.clone()
}
pub fn set_typevar(&mut self, tv: TypeVar) {
self.is_original_type_var = if let Some(previous_tv) = &self.type_var {
*previous_tv == tv
} else {
false
};
self.type_var = Some(tv);
}
/// Check if this variable has a free type variable. If not, the type of this variable is
/// computed from the type of another variable.
pub fn has_free_typevar(&self) -> bool {
match &self.type_var {
Some(tv) => tv.base.is_none() && self.is_original_type_var,
None => false,
}
}
pub fn to_rust_code(&self) -> String {
self.name.clone()
}
fn rust_type(&self) -> String {
self.type_var.as_ref().unwrap().to_rust_code()
}
}
impl fmt::Debug for Var {
fn fmt(&self, fmt: &mut fmt::Formatter) -> Result<(), fmt::Error> {
fmt.write_fmt(format_args!(
"Var({}{}{})",
self.name,
if self.src_def.is_some() { ", src" } else { "" },
if self.dst_def.is_some() { ", dst" } else { "" }
))
}
}
#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub(crate) struct VarIndex(u32);
entity_impl!(VarIndex);
pub(crate) struct VarPool {
pool: PrimaryMap<VarIndex, Var>,
}
impl VarPool {
pub fn new() -> Self {
Self {
pool: PrimaryMap::new(),
}
}
pub fn get(&self, index: VarIndex) -> &Var {
self.pool.get(index).unwrap()
}
pub fn get_mut(&mut self, index: VarIndex) -> &mut Var {
self.pool.get_mut(index).unwrap()
}
pub fn create(&mut self, name: impl Into<String>) -> VarIndex {
self.pool.push(Var::new(name.into()))
}
}
/// Contains constants created in the AST that must be inserted into the true [ConstantPool] when
/// the legalizer code is generated. The constant data is named in the order it is inserted;
/// inserting data using [insert] will avoid duplicates.
///
/// [ConstantPool]: ../../../cranelift_codegen/ir/constant/struct.ConstantPool.html
/// [insert]: ConstPool::insert
pub(crate) struct ConstPool {
pool: Vec<Vec<u8>>,
}
impl ConstPool {
/// Create an empty constant pool.
pub fn new() -> Self {
Self { pool: vec![] }
}
/// Create a name for a constant from its position in the pool.
fn create_name(position: usize) -> String {
format!("const{}", position)
}
/// Insert constant data into the pool, returning the name of the variable used to reference it.
/// This method will search for data that matches the new data and return the existing constant
/// name to avoid duplicates.
pub fn insert(&mut self, data: Vec<u8>) -> String {
let possible_position = self.pool.iter().position(|d| d == &data);
let position = if let Some(found_position) = possible_position {
found_position
} else {
let new_position = self.pool.len();
self.pool.push(data);
new_position
};
ConstPool::create_name(position)
}
/// Iterate over the name/value pairs in the pool.
pub fn iter(&self) -> impl Iterator<Item = (String, &Vec<u8>)> {
self.pool
.iter()
.enumerate()
.map(|(i, v)| (ConstPool::create_name(i), v))
}
}
/// Apply an instruction to arguments.
///
/// An `Apply` AST expression is created by using function call syntax on instructions. This
/// applies to both bound and unbound polymorphic instructions.
pub(crate) struct Apply {
pub inst: Instruction,
pub args: Vec<Expr>,
pub value_types: Vec<ValueType>,
}
impl Apply {
pub fn new(target: InstSpec, args: Vec<Expr>) -> Self {
let (inst, value_types) = match target {
InstSpec::Inst(inst) => (inst, Vec::new()),
InstSpec::Bound(bound_inst) => (bound_inst.inst, bound_inst.value_types),
};
// Apply should only operate on concrete value types, not "any".
let value_types = value_types
.into_iter()
.map(|vt| vt.expect())
.collect();
// Basic check on number of arguments.
assert!(
inst.operands_in.len() == args.len(),
"incorrect number of arguments in instruction {}",
inst.name
);
// Check that the kinds of Literals arguments match the expected operand.
for &imm_index in &inst.imm_opnums {
let arg = &args[imm_index];
if let Some(literal) = arg.maybe_literal() {
let op = &inst.operands_in[imm_index];
match &op.kind.fields {
OperandKindFields::ImmEnum(values) => {
if let Literal::Enumerator { value, .. } = literal {
assert!(
values.iter().any(|(_key, v)| v == value),
"Nonexistent enum value '{}' passed to field of kind '{}' -- \
did you use the right enum?",
value,
op.kind.rust_type
);
} else {
panic!(
"Passed non-enum field value {:?} to field of kind {}",
literal, op.kind.rust_type
);
}
}
OperandKindFields::ImmValue => match &literal {
Literal::Enumerator { value, .. } => panic!(
"Expected immediate value in immediate field of kind '{}', \
obtained enum value '{}'",
op.kind.rust_type, value
),
Literal::Bits { .. } | Literal::Int(_) | Literal::EmptyVarArgs => {}
},
_ => {
panic!(
"Literal passed to non-literal field of kind {}",
op.kind.rust_type
);
}
}
}
}
Self {
inst,
args,
value_types,
}
}
fn to_comment_string(&self, var_pool: &VarPool) -> String {
let args = self
.args
.iter()
.map(|arg| arg.to_rust_code(var_pool))
.collect::<Vec<_>>()
.join(", ");
let mut inst_and_bound_types = vec![self.inst.name.to_string()];
inst_and_bound_types.extend(self.value_types.iter().map(|vt| vt.to_string()));
let inst_name = inst_and_bound_types.join(".");
format!("{}({})", inst_name, args)
}
pub fn inst_predicate(&self, var_pool: &VarPool) -> InstructionPredicate {
let mut pred = InstructionPredicate::new();
for (format_field, &op_num) in self
.inst
.format
.imm_fields
.iter()
.zip(self.inst.imm_opnums.iter())
{
let arg = &self.args[op_num];
if arg.maybe_var().is_some() {
// Ignore free variables for now.
continue;
}
pred = pred.and(InstructionPredicate::new_is_field_equal_ast(
&*self.inst.format,
format_field,
arg.to_rust_code(var_pool),
));
}
// Add checks for any bound secondary type variables. We can't check the controlling type
// variable this way since it may not appear as the type of an operand.
if self.value_types.len() > 1 {
let poly = self
.inst
.polymorphic_info
.as_ref()
.expect("must have polymorphic info if it has bounded types");
for (bound_type, type_var) in
self.value_types[1..].iter().zip(poly.other_typevars.iter())
{
pred = pred.and(InstructionPredicate::new_typevar_check(
&self.inst, type_var, bound_type,
));
}
}
pred
}
/// Same as `inst_predicate()`, but also check the controlling type variable.
pub fn inst_predicate_with_ctrl_typevar(&self, var_pool: &VarPool) -> InstructionPredicate {
let mut pred = self.inst_predicate(var_pool);
if !self.value_types.is_empty() {
let bound_type = &self.value_types[0];
let poly = self.inst.polymorphic_info.as_ref().unwrap();
let type_check = if poly.use_typevar_operand {
InstructionPredicate::new_typevar_check(&self.inst, &poly.ctrl_typevar, bound_type)
} else {
InstructionPredicate::new_ctrl_typevar_check(&bound_type)
};
pred = pred.and(type_check);
}
pred
}
pub fn rust_builder(&self, defined_vars: &[VarIndex], var_pool: &VarPool) -> String {
let mut args = self
.args
.iter()
.map(|expr| expr.to_rust_code(var_pool))
.collect::<Vec<_>>()
.join(", ");
// Do we need to pass an explicit type argument?
if let Some(poly) = &self.inst.polymorphic_info {
if !poly.use_typevar_operand {
args = format!("{}, {}", var_pool.get(defined_vars[0]).rust_type(), args);
}
}
format!("{}({})", self.inst.snake_name(), args)
}
}
// Simple helpers for legalize actions construction.
pub(crate) enum DummyExpr {
Var(DummyVar),
Literal(Literal),
Constant(DummyConstant),
Apply(InstSpec, Vec<DummyExpr>),
Block(DummyVar),
}
#[derive(Clone)]
pub(crate) struct DummyVar {
pub name: String,
}
impl Into<DummyExpr> for DummyVar {
fn into(self) -> DummyExpr {
DummyExpr::Var(self)
}
}
impl Into<DummyExpr> for Literal {
fn into(self) -> DummyExpr {
DummyExpr::Literal(self)
}
}
#[derive(Clone)]
pub(crate) struct DummyConstant(pub(crate) Vec<u8>);
impl Into<DummyExpr> for DummyConstant {
fn into(self) -> DummyExpr {
DummyExpr::Constant(self)
}
}
pub(crate) fn var(name: &str) -> DummyVar {
DummyVar {
name: name.to_owned(),
}
}
pub(crate) struct DummyDef {
pub expr: DummyExpr,
pub defined_vars: Vec<DummyVar>,
}
pub(crate) struct ExprBuilder {
expr: DummyExpr,
}
impl ExprBuilder {
pub fn apply(inst: InstSpec, args: Vec<DummyExpr>) -> Self {
let expr = DummyExpr::Apply(inst, args);
Self { expr }
}
pub fn assign_to(self, defined_vars: Vec<DummyVar>) -> DummyDef {
DummyDef {
expr: self.expr,
defined_vars,
}
}
pub fn block(name: DummyVar) -> Self {
let expr = DummyExpr::Block(name);
Self { expr }
}
}
macro_rules! def_rhs {
// inst(a, b, c)
($inst:ident($($src:expr),*)) => {
ExprBuilder::apply($inst.into(), vec![$($src.clone().into()),*])
};
// inst.type(a, b, c)
($inst:ident.$type:ident($($src:expr),*)) => {
ExprBuilder::apply($inst.bind($type).into(), vec![$($src.clone().into()),*])
};
}
// Helper macro to define legalization recipes.
macro_rules! def {
// x = ...
($dest:ident = $($tt:tt)*) => {
def_rhs!($($tt)*).assign_to(vec![$dest.clone()])
};
// (x, y, ...) = ...
(($($dest:ident),*) = $($tt:tt)*) => {
def_rhs!($($tt)*).assign_to(vec![$($dest.clone()),*])
};
// An instruction with no results.
($($tt:tt)*) => {
def_rhs!($($tt)*).assign_to(Vec::new())
}
}
// Helper macro to define legalization recipes.
macro_rules! block {
// a basic block definition, splitting the current block in 2.
($block: ident) => {
ExprBuilder::block($block).assign_to(Vec::new())
};
}
#[cfg(test)]
mod tests {
use crate::cdsl::ast::ConstPool;
#[test]
fn const_pool_returns_var_names() {
let mut c = ConstPool::new();
assert_eq!(c.insert([0, 1, 2].to_vec()), "const0");
assert_eq!(c.insert([1, 2, 3].to_vec()), "const1");
}
#[test]
fn const_pool_avoids_duplicates() {
let data = [0, 1, 2].to_vec();
let mut c = ConstPool::new();
assert_eq!(c.pool.len(), 0);
assert_eq!(c.insert(data.clone()), "const0");
assert_eq!(c.pool.len(), 1);
assert_eq!(c.insert(data), "const0");
assert_eq!(c.pool.len(), 1);
}
#[test]
fn const_pool_iterates() {
let mut c = ConstPool::new();
c.insert([0, 1, 2].to_vec());
c.insert([3, 4, 5].to_vec());
let mut iter = c.iter();
assert_eq!(iter.next(), Some(("const0".to_owned(), &vec![0, 1, 2])));
assert_eq!(iter.next(), Some(("const1".to_owned(), &vec![3, 4, 5])));
assert_eq!(iter.next(), None);
}
}

View File

@@ -1,30 +0,0 @@
use std::collections::{HashMap, HashSet};
use std::iter::FromIterator;
use crate::cdsl::types::ValueType;
use crate::cdsl::xform::TransformGroupIndex;
pub(crate) struct CpuMode {
pub name: &'static str,
default_legalize: Option<TransformGroupIndex>,
monomorphic_legalize: Option<TransformGroupIndex>,
typed_legalize: HashMap<ValueType, TransformGroupIndex>,
}
impl CpuMode {
/// Returns a deterministically ordered, deduplicated list of TransformGroupIndex for the directly
/// reachable set of TransformGroup this TargetIsa uses.
pub fn direct_transform_groups(&self) -> Vec<TransformGroupIndex> {
let mut set = HashSet::new();
if let Some(i) = &self.default_legalize {
set.insert(*i);
}
if let Some(i) = &self.monomorphic_legalize {
set.insert(*i);
}
set.extend(self.typed_legalize.values().cloned());
let mut ret = Vec::from_iter(set);
ret.sort();
ret
}
}

View File

@@ -5,7 +5,7 @@ use std::fmt::{Display, Error, Formatter};
use std::rc::Rc;
use crate::cdsl::camel_case;
use crate::cdsl::formats::{FormatField, InstructionFormat};
use crate::cdsl::formats::InstructionFormat;
use crate::cdsl::operands::Operand;
use crate::cdsl::type_inference::Constraint;
use crate::cdsl::types::{LaneType, ReferenceType, ValueType};
@@ -21,46 +21,20 @@ pub(crate) type AllInstructions = PrimaryMap<OpcodeNumber, Instruction>;
pub(crate) struct InstructionGroupBuilder<'all_inst> {
all_instructions: &'all_inst mut AllInstructions,
own_instructions: Vec<Instruction>,
}
impl<'all_inst> InstructionGroupBuilder<'all_inst> {
pub fn new(all_instructions: &'all_inst mut AllInstructions) -> Self {
Self {
all_instructions,
own_instructions: Vec::new(),
}
}
pub fn push(&mut self, builder: InstructionBuilder) {
let opcode_number = OpcodeNumber(self.all_instructions.next_key().as_u32());
let inst = builder.build(opcode_number);
// Note this clone is cheap, since Instruction is a Rc<> wrapper for InstructionContent.
self.own_instructions.push(inst.clone());
self.all_instructions.push(inst);
}
pub fn build(self) -> InstructionGroup {
InstructionGroup {
instructions: self.own_instructions,
}
}
}
/// Every instruction must belong to exactly one instruction group. A given
/// target architecture can support instructions from multiple groups, and it
/// does not necessarily support all instructions in a group.
pub(crate) struct InstructionGroup {
instructions: Vec<Instruction>,
}
impl InstructionGroup {
pub fn by_name(&self, name: &'static str) -> &Instruction {
self.instructions
.iter()
.find(|inst| inst.name == name)
.unwrap_or_else(|| panic!("instruction with name '{}' does not exist", name))
}
}
/// Instructions can have parameters bound to them to specialize them for more specific encodings
@@ -143,17 +117,6 @@ impl InstructionContent {
&self.name
}
}
pub fn all_typevars(&self) -> Vec<&TypeVar> {
match &self.polymorphic_info {
Some(poly) => {
let mut result = vec![&poly.ctrl_typevar];
result.extend(&poly.other_typevars);
result
}
None => Vec::new(),
}
}
}
pub(crate) type Instruction = Rc<InstructionContent>;
@@ -375,20 +338,6 @@ impl InstructionBuilder {
}
}
/// A thin wrapper like Option<ValueType>, but with more precise semantics.
#[derive(Clone)]
pub(crate) enum ValueTypeOrAny {
ValueType(ValueType),
}
impl ValueTypeOrAny {
pub fn expect(self) -> ValueType {
match self {
ValueTypeOrAny::ValueType(vt) => vt,
}
}
}
/// An parameter used for binding instructions to specific types or values
pub(crate) enum BindParameter {
Lane(LaneType),
@@ -439,7 +388,7 @@ impl Display for Immediate {
#[derive(Clone)]
pub(crate) struct BoundInstruction {
pub inst: Instruction,
pub value_types: Vec<ValueTypeOrAny>,
pub value_types: Vec<ValueType>,
pub immediate_values: Vec<Immediate>,
}
@@ -502,11 +451,11 @@ impl Bindable for BoundInstruction {
match parameter.into() {
BindParameter::Lane(lane_type) => modified
.value_types
.push(ValueTypeOrAny::ValueType(lane_type.into())),
.push(lane_type.into()),
BindParameter::Reference(reference_type) => {
modified
.value_types
.push(ValueTypeOrAny::ValueType(reference_type.into()));
.push(reference_type.into());
}
}
modified.verify_bindings().unwrap();
@@ -719,206 +668,6 @@ fn is_ctrl_typevar_candidate(
Ok(other_typevars)
}
#[derive(Clone, Hash, PartialEq, Eq)]
pub(crate) enum FormatPredicateKind {
/// Is the field member equal to the expected value (stored here)?
IsEqual(String),
}
#[derive(Clone, Hash, PartialEq, Eq)]
pub(crate) struct FormatPredicateNode {
format_name: &'static str,
member_name: &'static str,
kind: FormatPredicateKind,
}
impl FormatPredicateNode {
fn new_raw(
format: &InstructionFormat,
member_name: &'static str,
kind: FormatPredicateKind,
) -> Self {
Self {
format_name: format.name,
member_name,
kind,
}
}
fn rust_predicate(&self) -> String {
match &self.kind {
FormatPredicateKind::IsEqual(arg) => {
format!("predicates::is_equal({}, {})", self.member_name, arg)
}
}
}
}
#[derive(Clone, Hash, PartialEq, Eq)]
pub(crate) enum TypePredicateNode {
/// Is the value argument (at the index designated by the first member) the same type as the
/// type name (second member)?
TypeVarCheck(usize, String),
/// Is the controlling type variable the same type as the one designated by the type name
/// (only member)?
CtrlTypeVarCheck(String),
}
impl TypePredicateNode {
fn rust_predicate(&self, func_str: &str) -> String {
match self {
TypePredicateNode::TypeVarCheck(index, value_type_name) => format!(
"{}.dfg.value_type(args[{}]) == {}",
func_str, index, value_type_name
),
TypePredicateNode::CtrlTypeVarCheck(value_type_name) => {
format!("{}.dfg.ctrl_typevar(inst) == {}", func_str, value_type_name)
}
}
}
}
/// A basic node in an instruction predicate: either an atom, or an AND of two conditions.
#[derive(Clone, Hash, PartialEq, Eq)]
pub(crate) enum InstructionPredicateNode {
FormatPredicate(FormatPredicateNode),
TypePredicate(TypePredicateNode),
/// An AND-combination of two or more other predicates.
And(Vec<InstructionPredicateNode>),
}
impl InstructionPredicateNode {
fn rust_predicate(&self, func_str: &str) -> String {
match self {
InstructionPredicateNode::FormatPredicate(node) => node.rust_predicate(),
InstructionPredicateNode::TypePredicate(node) => node.rust_predicate(func_str),
InstructionPredicateNode::And(nodes) => nodes
.iter()
.map(|x| x.rust_predicate(func_str))
.collect::<Vec<_>>()
.join(" && "),
}
}
}
#[derive(Clone, Hash, PartialEq, Eq)]
pub(crate) struct InstructionPredicate {
node: Option<InstructionPredicateNode>,
}
impl Into<InstructionPredicate> for InstructionPredicateNode {
fn into(self) -> InstructionPredicate {
InstructionPredicate { node: Some(self) }
}
}
impl InstructionPredicate {
pub fn new() -> Self {
Self { node: None }
}
pub fn new_typevar_check(
inst: &Instruction,
type_var: &TypeVar,
value_type: &ValueType,
) -> InstructionPredicateNode {
let index = inst
.value_opnums
.iter()
.enumerate()
.find(|(_, &op_num)| inst.operands_in[op_num].type_var().unwrap() == type_var)
.unwrap()
.0;
InstructionPredicateNode::TypePredicate(TypePredicateNode::TypeVarCheck(
index,
value_type.rust_name(),
))
}
pub fn new_ctrl_typevar_check(value_type: &ValueType) -> InstructionPredicateNode {
InstructionPredicateNode::TypePredicate(TypePredicateNode::CtrlTypeVarCheck(
value_type.rust_name(),
))
}
/// Used only for the AST module, which directly passes in the format field.
pub fn new_is_field_equal_ast(
format: &InstructionFormat,
field: &FormatField,
imm_value: String,
) -> InstructionPredicateNode {
InstructionPredicateNode::FormatPredicate(FormatPredicateNode::new_raw(
format,
field.member,
FormatPredicateKind::IsEqual(imm_value),
))
}
pub fn and(mut self, new_node: InstructionPredicateNode) -> Self {
let node = self.node;
let mut and_nodes = match node {
Some(node) => match node {
InstructionPredicateNode::And(nodes) => nodes,
_ => vec![node],
},
_ => Vec::new(),
};
and_nodes.push(new_node);
self.node = Some(InstructionPredicateNode::And(and_nodes));
self
}
pub fn rust_predicate(&self, func_str: &str) -> Option<String> {
self.node.as_ref().map(|root| root.rust_predicate(func_str))
}
}
#[derive(Copy, Clone, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub(crate) struct InstructionPredicateNumber(u32);
entity_impl!(InstructionPredicateNumber);
pub(crate) type InstructionPredicateMap =
PrimaryMap<InstructionPredicateNumber, InstructionPredicate>;
/// An instruction specification, containing an instruction that has bound types or not.
pub(crate) enum InstSpec {
Inst(Instruction),
Bound(BoundInstruction),
}
impl InstSpec {
pub fn inst(&self) -> &Instruction {
match &self {
InstSpec::Inst(inst) => inst,
InstSpec::Bound(bound_inst) => &bound_inst.inst,
}
}
}
impl Bindable for InstSpec {
fn bind(&self, parameter: impl Into<BindParameter>) -> BoundInstruction {
match self {
InstSpec::Inst(inst) => inst.bind(parameter.into()),
InstSpec::Bound(inst) => inst.bind(parameter.into()),
}
}
}
impl Into<InstSpec> for &Instruction {
fn into(self) -> InstSpec {
InstSpec::Inst(self.clone())
}
}
impl Into<InstSpec> for BoundInstruction {
fn into(self) -> InstSpec {
InstSpec::Bound(self)
}
}
#[cfg(test)]
mod test {
use super::*;

View File

@@ -1,89 +1,18 @@
use std::collections::HashSet;
use std::iter::FromIterator;
use crate::cdsl::cpu_modes::CpuMode;
use crate::cdsl::instructions::InstructionPredicateMap;
use crate::cdsl::recipes::Recipes;
use crate::cdsl::regs::IsaRegs;
use crate::cdsl::settings::SettingGroup;
use crate::cdsl::xform::{TransformGroupIndex, TransformGroups};
pub(crate) struct TargetIsa {
pub name: &'static str,
pub settings: SettingGroup,
pub regs: IsaRegs,
pub recipes: Recipes,
pub cpu_modes: Vec<CpuMode>,
pub encodings_predicates: InstructionPredicateMap,
/// TransformGroupIndex are global to all the ISAs, while we want to have indices into the
/// local array of transform groups that are directly used. We use this map to get this
/// information.
pub local_transform_groups: Vec<TransformGroupIndex>,
}
impl TargetIsa {
pub fn new(
name: &'static str,
settings: SettingGroup,
regs: IsaRegs,
recipes: Recipes,
cpu_modes: Vec<CpuMode>,
encodings_predicates: InstructionPredicateMap,
) -> Self {
// Compute the local TransformGroup index.
let mut local_transform_groups = Vec::new();
for cpu_mode in &cpu_modes {
let transform_groups = cpu_mode.direct_transform_groups();
for group_index in transform_groups {
// find() is fine here: the number of transform group is < 5 as of June 2019.
if local_transform_groups
.iter()
.find(|&val| group_index == *val)
.is_none()
{
local_transform_groups.push(group_index);
}
}
}
pub fn new(name: &'static str, settings: SettingGroup, regs: IsaRegs) -> Self {
Self {
name,
settings,
regs,
recipes,
cpu_modes,
encodings_predicates,
local_transform_groups,
}
}
/// Returns a deterministically ordered, deduplicated list of TransformGroupIndex for the
/// transitive set of TransformGroup this TargetIsa uses.
pub fn transitive_transform_groups(
&self,
all_groups: &TransformGroups,
) -> Vec<TransformGroupIndex> {
let mut set = HashSet::new();
for &root in self.local_transform_groups.iter() {
set.insert(root);
let mut base = root;
// Follow the chain of chain_with.
while let Some(chain_with) = &all_groups.get(base).chain_with {
set.insert(*chain_with);
base = *chain_with;
}
}
let mut vec = Vec::from_iter(set);
vec.sort();
vec
}
/// Returns a deterministically ordered, deduplicated list of TransformGroupIndex for the directly
/// reachable set of TransformGroup this TargetIsa uses.
pub fn direct_transform_groups(&self) -> &Vec<TransformGroupIndex> {
&self.local_transform_groups
}
}

View File

@@ -3,20 +3,15 @@
//! This module defines the classes that are used to define Cranelift
//! instructions and other entities.
#[macro_use]
pub mod ast;
pub mod cpu_modes;
pub mod formats;
pub mod instructions;
pub mod isa;
pub mod operands;
pub mod recipes;
pub mod regs;
pub mod settings;
pub mod type_inference;
pub mod types;
pub mod typevar;
pub mod xform;
/// A macro that converts boolean settings into predicates to look more natural.
#[macro_export]

View File

@@ -1,165 +0,0 @@
use std::rc::Rc;
use cranelift_entity::{entity_impl, PrimaryMap};
use crate::cdsl::formats::InstructionFormat;
use crate::cdsl::instructions::InstructionPredicate;
use crate::cdsl::regs::RegClassIndex;
use crate::cdsl::settings::SettingPredicateNumber;
/// A specific register in a register class.
///
/// A register is identified by the top-level register class it belongs to and
/// its first register unit.
///
/// Specific registers are used to describe constraints on instructions where
/// some operands must use a fixed register.
///
/// Register instances can be created with the constructor, or accessed as
/// attributes on the register class: `GPR.rcx`.
#[derive(Copy, Clone, Hash, PartialEq, Eq)]
pub(crate) struct Register {
pub regclass: RegClassIndex,
pub unit: u8,
}
/// An operand that must be in a stack slot.
///
/// A `Stack` object can be used to indicate an operand constraint for a value
/// operand that must live in a stack slot.
#[derive(Copy, Clone, Hash, PartialEq)]
pub(crate) struct Stack {
pub regclass: RegClassIndex,
}
#[derive(Clone, Hash, PartialEq)]
pub(crate) struct BranchRange {
pub inst_size: u64,
pub range: u64,
}
#[derive(Copy, Clone, Hash, PartialEq)]
pub(crate) enum OperandConstraint {
RegClass(RegClassIndex),
FixedReg(Register),
TiedInput(usize),
Stack(Stack),
}
impl Into<OperandConstraint> for RegClassIndex {
fn into(self) -> OperandConstraint {
OperandConstraint::RegClass(self)
}
}
impl Into<OperandConstraint> for Register {
fn into(self) -> OperandConstraint {
OperandConstraint::FixedReg(self)
}
}
impl Into<OperandConstraint> for usize {
fn into(self) -> OperandConstraint {
OperandConstraint::TiedInput(self)
}
}
impl Into<OperandConstraint> for Stack {
fn into(self) -> OperandConstraint {
OperandConstraint::Stack(self)
}
}
/// A recipe for encoding instructions with a given format.
///
/// Many different instructions can be encoded by the same recipe, but they
/// must all have the same instruction format.
///
/// The `operands_in` and `operands_out` arguments are tuples specifying the register
/// allocation constraints for the value operands and results respectively. The
/// possible constraints for an operand are:
///
/// - A `RegClass` specifying the set of allowed registers.
/// - A `Register` specifying a fixed-register operand.
/// - An integer indicating that this result is tied to a value operand, so
/// they must use the same register.
/// - A `Stack` specifying a value in a stack slot.
///
/// The `branch_range` argument must be provided for recipes that can encode
/// branch instructions. It is an `(origin, bits)` tuple describing the exact
/// range that can be encoded in a branch instruction.
#[derive(Clone)]
pub(crate) struct EncodingRecipe {
/// Short mnemonic name for this recipe.
pub name: String,
/// Associated instruction format.
pub format: Rc<InstructionFormat>,
/// Base number of bytes in the binary encoded instruction.
pub base_size: u64,
/// Tuple of register constraints for value operands.
pub operands_in: Vec<OperandConstraint>,
/// Tuple of register constraints for results.
pub operands_out: Vec<OperandConstraint>,
/// Function name to use when computing actual size.
pub compute_size: &'static str,
/// `(origin, bits)` range for branches.
pub branch_range: Option<BranchRange>,
/// This instruction clobbers `iflags` and `fflags`; true by default.
pub clobbers_flags: bool,
/// Instruction predicate.
pub inst_predicate: Option<InstructionPredicate>,
/// ISA predicate.
pub isa_predicate: Option<SettingPredicateNumber>,
/// Rust code for binary emission.
pub emit: Option<String>,
}
// Implement PartialEq ourselves: take all the fields into account but the name.
impl PartialEq for EncodingRecipe {
fn eq(&self, other: &Self) -> bool {
Rc::ptr_eq(&self.format, &other.format)
&& self.base_size == other.base_size
&& self.operands_in == other.operands_in
&& self.operands_out == other.operands_out
&& self.compute_size == other.compute_size
&& self.branch_range == other.branch_range
&& self.clobbers_flags == other.clobbers_flags
&& self.inst_predicate == other.inst_predicate
&& self.isa_predicate == other.isa_predicate
&& self.emit == other.emit
}
}
// To allow using it in a hashmap.
impl Eq for EncodingRecipe {}
#[derive(Copy, Clone, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub(crate) struct EncodingRecipeNumber(u32);
entity_impl!(EncodingRecipeNumber);
pub(crate) type Recipes = PrimaryMap<EncodingRecipeNumber, EncodingRecipe>;
#[derive(Clone)]
pub(crate) struct EncodingRecipeBuilder {
pub name: String,
format: Rc<InstructionFormat>,
pub base_size: u64,
pub operands_in: Option<Vec<OperandConstraint>>,
pub operands_out: Option<Vec<OperandConstraint>>,
pub compute_size: Option<&'static str>,
pub branch_range: Option<BranchRange>,
pub emit: Option<String>,
clobbers_flags: Option<bool>,
inst_predicate: Option<InstructionPredicate>,
isa_predicate: Option<SettingPredicateNumber>,
}

View File

@@ -1,8 +1,4 @@
use crate::cdsl::ast::{Def, DefIndex, DefPool, Var, VarIndex, VarPool};
use crate::cdsl::typevar::{DerivedFunc, TypeSet, TypeVar};
use std::collections::{HashMap, HashSet};
use std::iter::FromIterator;
use crate::cdsl::typevar::TypeVar;
#[derive(Debug, Hash, PartialEq, Eq)]
pub(crate) enum Constraint {
@@ -11,651 +7,4 @@ pub(crate) enum Constraint {
/// 1) They have the same number of lanes
/// 2) In a lane tv1 has at least as many bits as tv2.
WiderOrEq(TypeVar, TypeVar),
/// Constraint specifying that two derived type vars must have the same runtime type.
Eq(TypeVar, TypeVar),
/// Constraint specifying that a type var must belong to some typeset.
InTypeset(TypeVar, TypeSet),
}
impl Constraint {
fn translate_with<F: Fn(&TypeVar) -> TypeVar>(&self, func: F) -> Constraint {
match self {
Constraint::WiderOrEq(lhs, rhs) => {
let lhs = func(&lhs);
let rhs = func(&rhs);
Constraint::WiderOrEq(lhs, rhs)
}
Constraint::Eq(lhs, rhs) => {
let lhs = func(&lhs);
let rhs = func(&rhs);
Constraint::Eq(lhs, rhs)
}
Constraint::InTypeset(tv, ts) => {
let tv = func(&tv);
Constraint::InTypeset(tv, ts.clone())
}
}
}
/// Creates a new constraint by replacing type vars by their hashmap equivalent.
fn translate_with_map(
&self,
original_to_own_typevar: &HashMap<&TypeVar, TypeVar>,
) -> Constraint {
self.translate_with(|tv| substitute(original_to_own_typevar, tv))
}
/// Creates a new constraint by replacing type vars by their canonical equivalent.
fn translate_with_env(&self, type_env: &TypeEnvironment) -> Constraint {
self.translate_with(|tv| type_env.get_equivalent(tv))
}
fn is_trivial(&self) -> bool {
match self {
Constraint::WiderOrEq(lhs, rhs) => {
// Trivially true.
if lhs == rhs {
return true;
}
let ts1 = lhs.get_typeset();
let ts2 = rhs.get_typeset();
// Trivially true.
if ts1.is_wider_or_equal(&ts2) {
return true;
}
// Trivially false.
if ts1.is_narrower(&ts2) {
return true;
}
// Trivially false.
if (&ts1.lanes & &ts2.lanes).is_empty() {
return true;
}
self.is_concrete()
}
Constraint::Eq(lhs, rhs) => lhs == rhs || self.is_concrete(),
Constraint::InTypeset(_, _) => {
// The way InTypeset are made, they would always be trivial if we were applying the
// same logic as the Python code did, so ignore this.
self.is_concrete()
}
}
}
/// Returns true iff all the referenced type vars are singletons.
fn is_concrete(&self) -> bool {
match self {
Constraint::WiderOrEq(lhs, rhs) => {
lhs.singleton_type().is_some() && rhs.singleton_type().is_some()
}
Constraint::Eq(lhs, rhs) => {
lhs.singleton_type().is_some() && rhs.singleton_type().is_some()
}
Constraint::InTypeset(tv, _) => tv.singleton_type().is_some(),
}
}
fn typevar_args(&self) -> Vec<&TypeVar> {
match self {
Constraint::WiderOrEq(lhs, rhs) => vec![lhs, rhs],
Constraint::Eq(lhs, rhs) => vec![lhs, rhs],
Constraint::InTypeset(tv, _) => vec![tv],
}
}
}
#[derive(Clone, Copy)]
enum TypeEnvRank {
Singleton = 5,
Input = 4,
Intermediate = 3,
Output = 2,
Temp = 1,
Internal = 0,
}
/// Class encapsulating the necessary bookkeeping for type inference.
pub(crate) struct TypeEnvironment {
vars: HashSet<VarIndex>,
ranks: HashMap<TypeVar, TypeEnvRank>,
equivalency_map: HashMap<TypeVar, TypeVar>,
pub constraints: Vec<Constraint>,
}
impl TypeEnvironment {
fn new() -> Self {
TypeEnvironment {
vars: HashSet::new(),
ranks: HashMap::new(),
equivalency_map: HashMap::new(),
constraints: Vec::new(),
}
}
fn register(&mut self, var_index: VarIndex, var: &mut Var) {
self.vars.insert(var_index);
let rank = if var.is_input() {
TypeEnvRank::Input
} else if var.is_intermediate() {
TypeEnvRank::Intermediate
} else if var.is_output() {
TypeEnvRank::Output
} else {
assert!(var.is_temp());
TypeEnvRank::Temp
};
self.ranks.insert(var.get_or_create_typevar(), rank);
}
fn add_constraint(&mut self, constraint: Constraint) {
if self.constraints.iter().any(|item| *item == constraint) {
return;
}
// Check extra conditions for InTypeset constraints.
if let Constraint::InTypeset(tv, _) = &constraint {
assert!(
tv.base.is_none(),
"type variable is {:?}, while expecting none",
tv
);
assert!(
tv.name.starts_with("typeof_"),
"Name \"{}\" should start with \"typeof_\"",
tv.name
);
}
self.constraints.push(constraint);
}
/// Returns the canonical representative of the equivalency class of the given argument, or
/// duplicates it if it's not there yet.
pub fn get_equivalent(&self, tv: &TypeVar) -> TypeVar {
let mut tv = tv;
while let Some(found) = self.equivalency_map.get(tv) {
tv = found;
}
match &tv.base {
Some(parent) => self
.get_equivalent(&parent.type_var)
.derived(parent.derived_func),
None => tv.clone(),
}
}
/// Get the rank of tv in the partial order:
/// - TVs directly associated with a Var get their rank from the Var (see register()).
/// - Internally generated non-derived TVs implicitly get the lowest rank (0).
/// - Derived variables get their rank from their free typevar.
/// - Singletons have the highest rank.
/// - TVs associated with vars in a source pattern have a higher rank than TVs associated with
/// temporary vars.
fn rank(&self, tv: &TypeVar) -> u8 {
let actual_tv = match tv.base {
Some(_) => tv.free_typevar(),
None => Some(tv.clone()),
};
let rank = match actual_tv {
Some(actual_tv) => match self.ranks.get(&actual_tv) {
Some(rank) => Some(*rank),
None => {
assert!(
!actual_tv.name.starts_with("typeof_"),
"variable {} should be explicitly ranked",
actual_tv.name
);
None
}
},
None => None,
};
let rank = match rank {
Some(rank) => rank,
None => {
if tv.singleton_type().is_some() {
TypeEnvRank::Singleton
} else {
TypeEnvRank::Internal
}
}
};
rank as u8
}
/// Record the fact that the free tv1 is part of the same equivalence class as tv2. The
/// canonical representative of the merged class is tv2's canonical representative.
fn record_equivalent(&mut self, tv1: TypeVar, tv2: TypeVar) {
assert!(tv1.base.is_none());
assert!(self.get_equivalent(&tv1) == tv1);
if let Some(tv2_base) = &tv2.base {
// Ensure there are no cycles.
assert!(self.get_equivalent(&tv2_base.type_var) != tv1);
}
self.equivalency_map.insert(tv1, tv2);
}
/// Get the free typevars in the current type environment.
pub fn free_typevars(&self, var_pool: &mut VarPool) -> Vec<TypeVar> {
let mut typevars = Vec::new();
typevars.extend(self.equivalency_map.keys().cloned());
typevars.extend(
self.vars
.iter()
.map(|&var_index| var_pool.get_mut(var_index).get_or_create_typevar()),
);
let set: HashSet<TypeVar> = HashSet::from_iter(
typevars
.iter()
.map(|tv| self.get_equivalent(tv).free_typevar())
.filter(|opt_tv| {
// Filter out singleton types.
opt_tv.is_some()
})
.map(|tv| tv.unwrap()),
);
Vec::from_iter(set)
}
/// Normalize by collapsing any roots that don't correspond to a concrete type var AND have a
/// single type var derived from them or equivalent to them.
///
/// e.g. if we have a root of the tree that looks like:
///
/// typeof_a typeof_b
/// \\ /
/// typeof_x
/// |
/// half_width(1)
/// |
/// 1
///
/// we want to collapse the linear path between 1 and typeof_x. The resulting graph is:
///
/// typeof_a typeof_b
/// \\ /
/// typeof_x
fn normalize(&mut self, var_pool: &mut VarPool) {
let source_tvs: HashSet<TypeVar> = HashSet::from_iter(
self.vars
.iter()
.map(|&var_index| var_pool.get_mut(var_index).get_or_create_typevar()),
);
let mut children: HashMap<TypeVar, HashSet<TypeVar>> = HashMap::new();
// Insert all the parents found by the derivation relationship.
for type_var in self.equivalency_map.values() {
if type_var.base.is_none() {
continue;
}
let parent_tv = type_var.free_typevar();
if parent_tv.is_none() {
// Ignore this type variable, it's a singleton.
continue;
}
let parent_tv = parent_tv.unwrap();
children
.entry(parent_tv)
.or_insert_with(HashSet::new)
.insert(type_var.clone());
}
// Insert all the explicit equivalency links.
for (equivalent_tv, canon_tv) in self.equivalency_map.iter() {
children
.entry(canon_tv.clone())
.or_insert_with(HashSet::new)
.insert(equivalent_tv.clone());
}
// Remove links that are straight paths up to typevar of variables.
for free_root in self.free_typevars(var_pool) {
let mut root = &free_root;
while !source_tvs.contains(&root)
&& children.contains_key(&root)
&& children.get(&root).unwrap().len() == 1
{
let child = children.get(&root).unwrap().iter().next().unwrap();
assert_eq!(self.equivalency_map[child], root.clone());
self.equivalency_map.remove(child);
root = child;
}
}
}
/// Extract a clean type environment from self, that only mentions type vars associated with
/// real variables.
fn extract(self, var_pool: &mut VarPool) -> TypeEnvironment {
let vars_tv: HashSet<TypeVar> = HashSet::from_iter(
self.vars
.iter()
.map(|&var_index| var_pool.get_mut(var_index).get_or_create_typevar()),
);
let mut new_equivalency_map: HashMap<TypeVar, TypeVar> = HashMap::new();
for tv in &vars_tv {
let canon_tv = self.get_equivalent(tv);
if *tv != canon_tv {
new_equivalency_map.insert(tv.clone(), canon_tv.clone());
}
// Sanity check: the translated type map should only refer to real variables.
assert!(vars_tv.contains(tv));
let canon_free_tv = canon_tv.free_typevar();
assert!(canon_free_tv.is_none() || vars_tv.contains(&canon_free_tv.unwrap()));
}
let mut new_constraints: HashSet<Constraint> = HashSet::new();
for constraint in &self.constraints {
let constraint = constraint.translate_with_env(&self);
if constraint.is_trivial() || new_constraints.contains(&constraint) {
continue;
}
// Sanity check: translated constraints should refer only to real variables.
for arg in constraint.typevar_args() {
let arg_free_tv = arg.free_typevar();
assert!(arg_free_tv.is_none() || vars_tv.contains(&arg_free_tv.unwrap()));
}
new_constraints.insert(constraint);
}
TypeEnvironment {
vars: self.vars,
ranks: self.ranks,
equivalency_map: new_equivalency_map,
constraints: Vec::from_iter(new_constraints),
}
}
}
/// Replaces an external type variable according to the following rules:
/// - if a local copy is present in the map, return it.
/// - or if it's derived, create a local derived one that recursively substitutes the parent.
/// - or return itself.
fn substitute(map: &HashMap<&TypeVar, TypeVar>, external_type_var: &TypeVar) -> TypeVar {
match map.get(&external_type_var) {
Some(own_type_var) => own_type_var.clone(),
None => match &external_type_var.base {
Some(parent) => {
let parent_substitute = substitute(map, &parent.type_var);
TypeVar::derived(&parent_substitute, parent.derived_func)
}
None => external_type_var.clone(),
},
}
}
/// Normalize a (potentially derived) typevar using the following rules:
///
/// - vector and width derived functions commute
/// {HALF,DOUBLE}VECTOR({HALF,DOUBLE}WIDTH(base)) ->
/// {HALF,DOUBLE}WIDTH({HALF,DOUBLE}VECTOR(base))
///
/// - half/double pairs collapse
/// {HALF,DOUBLE}WIDTH({DOUBLE,HALF}WIDTH(base)) -> base
/// {HALF,DOUBLE}VECTOR({DOUBLE,HALF}VECTOR(base)) -> base
fn canonicalize_derivations(tv: TypeVar) -> TypeVar {
let base = match &tv.base {
Some(base) => base,
None => return tv,
};
let derived_func = base.derived_func;
if let Some(base_base) = &base.type_var.base {
let base_base_tv = &base_base.type_var;
match (derived_func, base_base.derived_func) {
(DerivedFunc::HalfWidth, DerivedFunc::DoubleWidth)
| (DerivedFunc::DoubleWidth, DerivedFunc::HalfWidth)
| (DerivedFunc::HalfVector, DerivedFunc::DoubleVector)
| (DerivedFunc::DoubleVector, DerivedFunc::HalfVector) => {
// Cancelling bijective transformations. This doesn't hide any overflow issues
// since derived type sets are checked upon derivaion, and base typesets are only
// allowed to shrink.
return canonicalize_derivations(base_base_tv.clone());
}
(DerivedFunc::HalfWidth, DerivedFunc::HalfVector)
| (DerivedFunc::HalfWidth, DerivedFunc::DoubleVector)
| (DerivedFunc::DoubleWidth, DerivedFunc::DoubleVector)
| (DerivedFunc::DoubleWidth, DerivedFunc::HalfVector) => {
// Arbitrarily put WIDTH derivations before VECTOR derivations, since they commute.
return canonicalize_derivations(
base_base_tv
.derived(derived_func)
.derived(base_base.derived_func),
);
}
_ => {}
};
}
canonicalize_derivations(base.type_var.clone()).derived(derived_func)
}
/// Given typevars tv1 and tv2 (which could be derived from one another), constrain their typesets
/// to be the same. When one is derived from the other, repeat the constrain process until
/// a fixed point is reached.
fn constrain_fixpoint(tv1: &TypeVar, tv2: &TypeVar) {
loop {
let old_tv1_ts = tv1.get_typeset().clone();
tv2.constrain_types(tv1.clone());
if tv1.get_typeset() == old_tv1_ts {
break;
}
}
let old_tv2_ts = tv2.get_typeset();
tv1.constrain_types(tv2.clone());
// The above loop should ensure that all reference cycles have been handled.
assert!(old_tv2_ts == tv2.get_typeset());
}
/// Unify tv1 and tv2 in the given type environment. tv1 must have a rank greater or equal to tv2's
/// one, modulo commutations.
fn unify(tv1: &TypeVar, tv2: &TypeVar, type_env: &mut TypeEnvironment) -> Result<(), String> {
let tv1 = canonicalize_derivations(type_env.get_equivalent(tv1));
let tv2 = canonicalize_derivations(type_env.get_equivalent(tv2));
if tv1 == tv2 {
// Already unified.
return Ok(());
}
if type_env.rank(&tv2) < type_env.rank(&tv1) {
// Make sure tv1 always has the smallest rank, since real variables have the higher rank
// and we want them to be the canonical representatives of their equivalency classes.
return unify(&tv2, &tv1, type_env);
}
constrain_fixpoint(&tv1, &tv2);
if tv1.get_typeset().size() == 0 || tv2.get_typeset().size() == 0 {
return Err(format!(
"Error: empty type created when unifying {} and {}",
tv1.name, tv2.name
));
}
let base = match &tv1.base {
Some(base) => base,
None => {
type_env.record_equivalent(tv1, tv2);
return Ok(());
}
};
if let Some(inverse) = base.derived_func.inverse() {
return unify(&base.type_var, &tv2.derived(inverse), type_env);
}
type_env.add_constraint(Constraint::Eq(tv1, tv2));
Ok(())
}
/// Perform type inference on one Def in the current type environment and return an updated type
/// environment or error.
///
/// At a high level this works by creating fresh copies of each formal type var in the Def's
/// instruction's signature, and unifying the formal typevar with the corresponding actual typevar.
fn infer_definition(
def: &Def,
var_pool: &mut VarPool,
type_env: TypeEnvironment,
last_type_index: &mut usize,
) -> Result<TypeEnvironment, String> {
let apply = &def.apply;
let inst = &apply.inst;
let mut type_env = type_env;
let free_formal_tvs = inst.all_typevars();
let mut original_to_own_typevar: HashMap<&TypeVar, TypeVar> = HashMap::new();
for &tv in &free_formal_tvs {
assert!(original_to_own_typevar
.insert(
tv,
TypeVar::copy_from(tv, format!("own_{}", last_type_index))
)
.is_none());
*last_type_index += 1;
}
// Update the mapping with any explicity bound type vars:
for (i, value_type) in apply.value_types.iter().enumerate() {
let singleton = TypeVar::new_singleton(value_type.clone());
assert!(original_to_own_typevar
.insert(free_formal_tvs[i], singleton)
.is_some());
}
// Get fresh copies for each typevar in the signature (both free and derived).
let mut formal_tvs = Vec::new();
formal_tvs.extend(inst.value_results.iter().map(|&i| {
substitute(
&original_to_own_typevar,
inst.operands_out[i].type_var().unwrap(),
)
}));
formal_tvs.extend(inst.value_opnums.iter().map(|&i| {
substitute(
&original_to_own_typevar,
inst.operands_in[i].type_var().unwrap(),
)
}));
// Get the list of actual vars.
let mut actual_vars = Vec::new();
actual_vars.extend(inst.value_results.iter().map(|&i| def.defined_vars[i]));
actual_vars.extend(
inst.value_opnums
.iter()
.map(|&i| apply.args[i].unwrap_var()),
);
// Get the list of the actual TypeVars.
let mut actual_tvs = Vec::new();
for var_index in actual_vars {
let var = var_pool.get_mut(var_index);
type_env.register(var_index, var);
actual_tvs.push(var.get_or_create_typevar());
}
// Make sure we start unifying with the control type variable first, by putting it at the
// front of both vectors.
if let Some(poly) = &inst.polymorphic_info {
let own_ctrl_tv = &original_to_own_typevar[&poly.ctrl_typevar];
let ctrl_index = formal_tvs.iter().position(|tv| tv == own_ctrl_tv).unwrap();
if ctrl_index != 0 {
formal_tvs.swap(0, ctrl_index);
actual_tvs.swap(0, ctrl_index);
}
}
// Unify each actual type variable with the corresponding formal type variable.
for (actual_tv, formal_tv) in actual_tvs.iter().zip(&formal_tvs) {
if let Err(msg) = unify(actual_tv, formal_tv, &mut type_env) {
return Err(format!(
"fail ti on {} <: {}: {}",
actual_tv.name, formal_tv.name, msg
));
}
}
// Add any instruction specific constraints.
for constraint in &inst.constraints {
type_env.add_constraint(constraint.translate_with_map(&original_to_own_typevar));
}
Ok(type_env)
}
/// Perform type inference on an transformation. Return an updated type environment or error.
pub(crate) fn infer_transform(
src: DefIndex,
dst: &[DefIndex],
def_pool: &DefPool,
var_pool: &mut VarPool,
) -> Result<TypeEnvironment, String> {
let mut type_env = TypeEnvironment::new();
let mut last_type_index = 0;
// Execute type inference on the source pattern.
type_env = infer_definition(def_pool.get(src), var_pool, type_env, &mut last_type_index)
.map_err(|err| format!("In src pattern: {}", err))?;
// Collect the type sets once after applying the source patterm; we'll compare the typesets
// after we've also considered the destination pattern, and will emit supplementary InTypeset
// checks if they don't match.
let src_typesets = type_env
.vars
.iter()
.map(|&var_index| {
let var = var_pool.get_mut(var_index);
let tv = type_env.get_equivalent(&var.get_or_create_typevar());
(var_index, tv.get_typeset())
})
.collect::<Vec<_>>();
// Execute type inference on the destination pattern.
for (i, &def_index) in dst.iter().enumerate() {
let def = def_pool.get(def_index);
type_env = infer_definition(def, var_pool, type_env, &mut last_type_index)
.map_err(|err| format!("line {}: {}", i, err))?;
}
for (var_index, src_typeset) in src_typesets {
let var = var_pool.get(var_index);
if !var.has_free_typevar() {
continue;
}
let tv = type_env.get_equivalent(&var.get_typevar().unwrap());
let new_typeset = tv.get_typeset();
assert!(
new_typeset.is_subset(&src_typeset),
"type sets can only get narrower"
);
if new_typeset != src_typeset {
type_env.add_constraint(Constraint::InTypeset(tv.clone(), new_typeset.clone()));
}
}
type_env.normalize(var_pool);
Ok(type_env.extract(var_pool))
}

View File

@@ -1,5 +1,5 @@
use std::cell::RefCell;
use std::collections::{BTreeSet, HashSet};
use std::collections::BTreeSet;
use std::fmt;
use std::hash;
use std::iter::FromIterator;
@@ -269,52 +269,6 @@ impl TypeVar {
pub fn merge_lanes(&self) -> TypeVar {
self.derived(DerivedFunc::MergeLanes)
}
/// Constrain the range of types this variable can assume to a subset of those in the typeset
/// ts.
/// May mutate itself if it's not derived, or its parent if it is.
pub fn constrain_types_by_ts(&self, type_set: TypeSet) {
match &self.base {
Some(base) => {
base.type_var
.constrain_types_by_ts(type_set.preimage(base.derived_func));
}
None => {
self.content
.borrow_mut()
.type_set
.inplace_intersect_with(&type_set);
}
}
}
/// Constrain the range of types this variable can assume to a subset of those `other` can
/// assume.
/// May mutate itself if it's not derived, or its parent if it is.
pub fn constrain_types(&self, other: TypeVar) {
if self == &other {
return;
}
self.constrain_types_by_ts(other.get_typeset());
}
/// Get a Rust expression that computes the type of this type variable.
pub fn to_rust_code(&self) -> String {
match &self.base {
Some(base) => format!(
"{}.{}().unwrap()",
base.type_var.to_rust_code(),
base.derived_func.name()
),
None => {
if let Some(singleton) = self.singleton_type() {
singleton.rust_name()
} else {
self.name.clone()
}
}
}
}
}
impl Into<TypeVar> for &TypeVar {
@@ -392,19 +346,6 @@ impl DerivedFunc {
DerivedFunc::MergeLanes => "merge_lanes",
}
}
/// Returns the inverse function of this one, if it is a bijection.
pub fn inverse(self) -> Option<DerivedFunc> {
match self {
DerivedFunc::HalfWidth => Some(DerivedFunc::DoubleWidth),
DerivedFunc::DoubleWidth => Some(DerivedFunc::HalfWidth),
DerivedFunc::HalfVector => Some(DerivedFunc::DoubleVector),
DerivedFunc::DoubleVector => Some(DerivedFunc::HalfVector),
DerivedFunc::MergeLanes => Some(DerivedFunc::SplitLanes),
DerivedFunc::SplitLanes => Some(DerivedFunc::MergeLanes),
_ => None,
}
}
}
#[derive(Debug, Hash)]
@@ -594,94 +535,6 @@ impl TypeSet {
assert_eq!(types.len(), 1);
types.remove(0)
}
/// Return the inverse image of self across the derived function func.
fn preimage(&self, func: DerivedFunc) -> TypeSet {
if self.size() == 0 {
// The inverse of the empty set is itself.
return self.clone();
}
match func {
DerivedFunc::LaneOf => {
let mut copy = self.clone();
copy.lanes =
NumSet::from_iter((0..=MAX_LANES.trailing_zeros()).map(|i| u16::pow(2, i)));
copy
}
DerivedFunc::AsBool => {
let mut copy = self.clone();
if self.bools.contains(&1) {
copy.ints = NumSet::from_iter(vec![8, 16, 32, 64, 128]);
copy.floats = NumSet::from_iter(vec![32, 64]);
} else {
copy.ints = &self.bools - &NumSet::from_iter(vec![1]);
copy.floats = &self.bools & &NumSet::from_iter(vec![32, 64]);
// If b1 is not in our typeset, than lanes=1 cannot be in the pre-image, as
// as_bool() of scalars is always b1.
copy.lanes = &self.lanes - &NumSet::from_iter(vec![1]);
}
copy
}
DerivedFunc::HalfWidth => self.double_width(),
DerivedFunc::DoubleWidth => self.half_width(),
DerivedFunc::HalfVector => self.double_vector(),
DerivedFunc::DoubleVector => self.half_vector(),
DerivedFunc::SplitLanes => self.double_width().half_vector(),
DerivedFunc::MergeLanes => self.half_width().double_vector(),
}
}
pub fn inplace_intersect_with(&mut self, other: &TypeSet) {
self.lanes = &self.lanes & &other.lanes;
self.ints = &self.ints & &other.ints;
self.floats = &self.floats & &other.floats;
self.bools = &self.bools & &other.bools;
self.refs = &self.refs & &other.refs;
let mut new_specials = Vec::new();
for spec in &self.specials {
if let Some(spec) = other.specials.iter().find(|&other_spec| other_spec == spec) {
new_specials.push(*spec);
}
}
self.specials = new_specials;
}
pub fn is_subset(&self, other: &TypeSet) -> bool {
self.lanes.is_subset(&other.lanes)
&& self.ints.is_subset(&other.ints)
&& self.floats.is_subset(&other.floats)
&& self.bools.is_subset(&other.bools)
&& self.refs.is_subset(&other.refs)
&& {
let specials: HashSet<SpecialType> = HashSet::from_iter(self.specials.clone());
let other_specials = HashSet::from_iter(other.specials.clone());
specials.is_subset(&other_specials)
}
}
pub fn is_wider_or_equal(&self, other: &TypeSet) -> bool {
set_wider_or_equal(&self.ints, &other.ints)
&& set_wider_or_equal(&self.floats, &other.floats)
&& set_wider_or_equal(&self.bools, &other.bools)
&& set_wider_or_equal(&self.refs, &other.refs)
}
pub fn is_narrower(&self, other: &TypeSet) -> bool {
set_narrower(&self.ints, &other.ints)
&& set_narrower(&self.floats, &other.floats)
&& set_narrower(&self.bools, &other.bools)
&& set_narrower(&self.refs, &other.refs)
}
}
fn set_wider_or_equal(s1: &NumSet, s2: &NumSet) -> bool {
!s1.is_empty() && !s2.is_empty() && s1.iter().min() >= s2.iter().max()
}
fn set_narrower(s1: &NumSet, s2: &NumSet) -> bool {
!s1.is_empty() && !s2.is_empty() && s1.iter().min() < s2.iter().max()
}
impl fmt::Debug for TypeSet {
@@ -806,18 +659,6 @@ impl TypeSetBuilder {
self.specials,
)
}
pub fn all() -> TypeSet {
TypeSetBuilder::new()
.ints(Interval::All)
.floats(Interval::All)
.bools(Interval::All)
.refs(Interval::All)
.simd_lanes(Interval::All)
.specials(ValueType::all_special_types().collect())
.includes_scalars(true)
.build()
}
}
#[derive(PartialEq)]
@@ -1054,135 +895,6 @@ fn test_forward_images() {
);
}
#[test]
fn test_backward_images() {
let empty_set = TypeSetBuilder::new().build();
// LaneOf.
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(1..1)
.ints(8..8)
.floats(32..32)
.build()
.preimage(DerivedFunc::LaneOf),
TypeSetBuilder::new()
.simd_lanes(Interval::All)
.ints(8..8)
.floats(32..32)
.build()
);
assert_eq!(empty_set.preimage(DerivedFunc::LaneOf), empty_set);
// AsBool.
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(1..4)
.bools(1..128)
.build()
.preimage(DerivedFunc::AsBool),
TypeSetBuilder::new()
.simd_lanes(1..4)
.ints(Interval::All)
.bools(Interval::All)
.floats(Interval::All)
.build()
);
// Double vector.
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(1..1)
.ints(8..8)
.build()
.preimage(DerivedFunc::DoubleVector)
.size(),
0
);
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(1..16)
.ints(8..16)
.floats(32..32)
.build()
.preimage(DerivedFunc::DoubleVector),
TypeSetBuilder::new()
.simd_lanes(1..8)
.ints(8..16)
.floats(32..32)
.build(),
);
// Half vector.
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(256..256)
.ints(8..8)
.build()
.preimage(DerivedFunc::HalfVector)
.size(),
0
);
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(64..128)
.bools(1..32)
.build()
.preimage(DerivedFunc::HalfVector),
TypeSetBuilder::new()
.simd_lanes(128..256)
.bools(1..32)
.build(),
);
// Half width.
assert_eq!(
TypeSetBuilder::new()
.ints(128..128)
.floats(64..64)
.bools(128..128)
.build()
.preimage(DerivedFunc::HalfWidth)
.size(),
0
);
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(64..256)
.bools(1..64)
.build()
.preimage(DerivedFunc::HalfWidth),
TypeSetBuilder::new()
.simd_lanes(64..256)
.bools(16..128)
.build(),
);
// Double width.
assert_eq!(
TypeSetBuilder::new()
.ints(8..8)
.floats(32..32)
.bools(1..8)
.build()
.preimage(DerivedFunc::DoubleWidth)
.size(),
0
);
assert_eq!(
TypeSetBuilder::new()
.simd_lanes(1..16)
.ints(8..16)
.floats(32..64)
.build()
.preimage(DerivedFunc::DoubleWidth),
TypeSetBuilder::new()
.simd_lanes(1..16)
.ints(8..8)
.floats(32..32)
.build()
);
}
#[test]
#[should_panic]

View File

@@ -1,478 +0,0 @@
use crate::cdsl::ast::{
Apply, BlockPool, ConstPool, DefIndex, DefPool, DummyDef, DummyExpr, Expr, PatternPosition,
VarIndex, VarPool,
};
use crate::cdsl::instructions::Instruction;
use crate::cdsl::type_inference::{infer_transform, TypeEnvironment};
use crate::cdsl::typevar::TypeVar;
use cranelift_entity::{entity_impl, PrimaryMap};
use std::collections::{HashMap, HashSet};
use std::iter::FromIterator;
/// An instruction transformation consists of a source and destination pattern.
///
/// Patterns are expressed in *register transfer language* as tuples of Def or Expr nodes. A
/// pattern may optionally have a sequence of TypeConstraints, that additionally limit the set of
/// cases when it applies.
///
/// The source pattern can contain only a single instruction.
pub(crate) struct Transform {
pub src: DefIndex,
pub dst: Vec<DefIndex>,
pub var_pool: VarPool,
pub def_pool: DefPool,
pub block_pool: BlockPool,
pub const_pool: ConstPool,
pub type_env: TypeEnvironment,
}
type SymbolTable = HashMap<String, VarIndex>;
impl Transform {
fn new(src: DummyDef, dst: Vec<DummyDef>) -> Self {
let mut var_pool = VarPool::new();
let mut def_pool = DefPool::new();
let mut block_pool = BlockPool::new();
let mut const_pool = ConstPool::new();
let mut input_vars: Vec<VarIndex> = Vec::new();
let mut defined_vars: Vec<VarIndex> = Vec::new();
// Maps variable names to our own Var copies.
let mut symbol_table: SymbolTable = SymbolTable::new();
// Rewrite variables in src and dst using our own copies.
let src = rewrite_def_list(
PatternPosition::Source,
vec![src],
&mut symbol_table,
&mut input_vars,
&mut defined_vars,
&mut var_pool,
&mut def_pool,
&mut block_pool,
&mut const_pool,
)[0];
let num_src_inputs = input_vars.len();
let dst = rewrite_def_list(
PatternPosition::Destination,
dst,
&mut symbol_table,
&mut input_vars,
&mut defined_vars,
&mut var_pool,
&mut def_pool,
&mut block_pool,
&mut const_pool,
);
// Sanity checks.
for &var_index in &input_vars {
assert!(
var_pool.get(var_index).is_input(),
"'{:?}' used as both input and def",
var_pool.get(var_index)
);
}
assert!(
input_vars.len() == num_src_inputs,
"extra input vars in dst pattern: {:?}",
input_vars
.iter()
.map(|&i| var_pool.get(i))
.skip(num_src_inputs)
.collect::<Vec<_>>()
);
// Perform type inference and cleanup.
let type_env = infer_transform(src, &dst, &def_pool, &mut var_pool).unwrap();
// Sanity check: the set of inferred free type variables should be a subset of the type
// variables corresponding to Vars appearing in the source pattern.
{
let free_typevars: HashSet<TypeVar> =
HashSet::from_iter(type_env.free_typevars(&mut var_pool));
let src_tvs = HashSet::from_iter(
input_vars
.clone()
.iter()
.chain(
defined_vars
.iter()
.filter(|&&var_index| !var_pool.get(var_index).is_temp()),
)
.map(|&var_index| var_pool.get(var_index).get_typevar())
.filter(|maybe_var| maybe_var.is_some())
.map(|var| var.unwrap()),
);
if !free_typevars.is_subset(&src_tvs) {
let missing_tvs = (&free_typevars - &src_tvs)
.iter()
.map(|tv| tv.name.clone())
.collect::<Vec<_>>()
.join(", ");
panic!("Some free vars don't appear in src: {}", missing_tvs);
}
}
for &var_index in input_vars.iter().chain(defined_vars.iter()) {
let var = var_pool.get_mut(var_index);
let canon_tv = type_env.get_equivalent(&var.get_or_create_typevar());
var.set_typevar(canon_tv);
}
Self {
src,
dst,
var_pool,
def_pool,
block_pool,
const_pool,
type_env,
}
}
fn verify_legalize(&self) {
let def = self.def_pool.get(self.src);
for &var_index in def.defined_vars.iter() {
let defined_var = self.var_pool.get(var_index);
assert!(
defined_var.is_output(),
"{:?} not defined in the destination pattern",
defined_var
);
}
}
}
/// Inserts, if not present, a name in the `symbol_table`. Then returns its index in the variable
/// pool `var_pool`. If the variable was not present in the symbol table, then add it to the list of
/// `defined_vars`.
fn var_index(
name: &str,
symbol_table: &mut SymbolTable,
defined_vars: &mut Vec<VarIndex>,
var_pool: &mut VarPool,
) -> VarIndex {
let name = name.to_string();
match symbol_table.get(&name) {
Some(&existing_var) => existing_var,
None => {
// Materialize the variable.
let new_var = var_pool.create(name.clone());
symbol_table.insert(name, new_var);
defined_vars.push(new_var);
new_var
}
}
}
/// Given a list of symbols defined in a Def, rewrite them to local symbols. Yield the new locals.
fn rewrite_defined_vars(
position: PatternPosition,
dummy_def: &DummyDef,
def_index: DefIndex,
symbol_table: &mut SymbolTable,
defined_vars: &mut Vec<VarIndex>,
var_pool: &mut VarPool,
) -> Vec<VarIndex> {
let mut new_defined_vars = Vec::new();
for var in &dummy_def.defined_vars {
let own_var = var_index(&var.name, symbol_table, defined_vars, var_pool);
var_pool.get_mut(own_var).set_def(position, def_index);
new_defined_vars.push(own_var);
}
new_defined_vars
}
/// Find all uses of variables in `expr` and replace them with our own local symbols.
fn rewrite_expr(
position: PatternPosition,
dummy_expr: DummyExpr,
symbol_table: &mut SymbolTable,
input_vars: &mut Vec<VarIndex>,
var_pool: &mut VarPool,
const_pool: &mut ConstPool,
) -> Apply {
let (apply_target, dummy_args) = if let DummyExpr::Apply(apply_target, dummy_args) = dummy_expr
{
(apply_target, dummy_args)
} else {
panic!("we only rewrite apply expressions");
};
assert_eq!(
apply_target.inst().operands_in.len(),
dummy_args.len(),
"number of arguments in instruction {} is incorrect\nexpected: {:?}",
apply_target.inst().name,
apply_target
.inst()
.operands_in
.iter()
.map(|operand| format!("{}: {}", operand.name, operand.kind.rust_type))
.collect::<Vec<_>>(),
);
let mut args = Vec::new();
for (i, arg) in dummy_args.into_iter().enumerate() {
match arg {
DummyExpr::Var(var) => {
let own_var = var_index(&var.name, symbol_table, input_vars, var_pool);
let var = var_pool.get(own_var);
assert!(
var.is_input() || var.get_def(position).is_some(),
"{:?} used as both input and def",
var
);
args.push(Expr::Var(own_var));
}
DummyExpr::Literal(literal) => {
assert!(!apply_target.inst().operands_in[i].is_value());
args.push(Expr::Literal(literal));
}
DummyExpr::Constant(constant) => {
let const_name = const_pool.insert(constant.0);
// Here we abuse var_index by passing an empty, immediately-dropped vector to
// `defined_vars`; the reason for this is that unlike the `Var` case above,
// constants will create a variable that is not an input variable (it is tracked
// instead by ConstPool).
let const_var = var_index(&const_name, symbol_table, &mut vec![], var_pool);
args.push(Expr::Var(const_var));
}
DummyExpr::Apply(..) => {
panic!("Recursive apply is not allowed.");
}
DummyExpr::Block(_block) => {
panic!("Blocks are not valid arguments.");
}
}
}
Apply::new(apply_target, args)
}
#[allow(clippy::too_many_arguments)]
fn rewrite_def_list(
position: PatternPosition,
dummy_defs: Vec<DummyDef>,
symbol_table: &mut SymbolTable,
input_vars: &mut Vec<VarIndex>,
defined_vars: &mut Vec<VarIndex>,
var_pool: &mut VarPool,
def_pool: &mut DefPool,
block_pool: &mut BlockPool,
const_pool: &mut ConstPool,
) -> Vec<DefIndex> {
let mut new_defs = Vec::new();
// Register variable names of new blocks first as a block name can be used to jump forward. Thus
// the name has to be registered first to avoid misinterpreting it as an input-var.
for dummy_def in dummy_defs.iter() {
if let DummyExpr::Block(ref var) = dummy_def.expr {
var_index(&var.name, symbol_table, defined_vars, var_pool);
}
}
// Iterate over the definitions and blocks, to map variables names to inputs or outputs.
for dummy_def in dummy_defs {
let def_index = def_pool.next_index();
let new_defined_vars = rewrite_defined_vars(
position,
&dummy_def,
def_index,
symbol_table,
defined_vars,
var_pool,
);
if let DummyExpr::Block(var) = dummy_def.expr {
let var_index = *symbol_table
.get(&var.name)
.or_else(|| {
panic!(
"Block {} was not registered during the first visit",
var.name
)
})
.unwrap();
var_pool.get_mut(var_index).set_def(position, def_index);
block_pool.create_block(var_index, def_index);
} else {
let new_apply = rewrite_expr(
position,
dummy_def.expr,
symbol_table,
input_vars,
var_pool,
const_pool,
);
assert!(
def_pool.next_index() == def_index,
"shouldn't have created new defs in the meanwhile"
);
assert_eq!(
new_apply.inst.value_results.len(),
new_defined_vars.len(),
"number of Var results in instruction is incorrect"
);
new_defs.push(def_pool.create_inst(new_apply, new_defined_vars));
}
}
new_defs
}
/// A group of related transformations.
pub(crate) struct TransformGroup {
pub name: &'static str,
pub doc: &'static str,
pub chain_with: Option<TransformGroupIndex>,
pub isa_name: Option<&'static str>,
pub id: TransformGroupIndex,
/// Maps Instruction camel_case names to custom legalization functions names.
pub custom_legalizes: HashMap<String, &'static str>,
pub transforms: Vec<Transform>,
}
impl TransformGroup {
pub fn rust_name(&self) -> String {
match self.isa_name {
Some(_) => {
// This is a function in the same module as the LEGALIZE_ACTIONS table referring to
// it.
self.name.to_string()
}
None => format!("crate::legalizer::{}", self.name),
}
}
}
#[derive(Copy, Clone, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub(crate) struct TransformGroupIndex(u32);
entity_impl!(TransformGroupIndex);
pub(crate) struct TransformGroupBuilder {
name: &'static str,
doc: &'static str,
chain_with: Option<TransformGroupIndex>,
isa_name: Option<&'static str>,
pub custom_legalizes: HashMap<String, &'static str>,
pub transforms: Vec<Transform>,
}
impl TransformGroupBuilder {
pub fn new(name: &'static str, doc: &'static str) -> Self {
Self {
name,
doc,
chain_with: None,
isa_name: None,
custom_legalizes: HashMap::new(),
transforms: Vec::new(),
}
}
pub fn chain_with(mut self, next_id: TransformGroupIndex) -> Self {
assert!(self.chain_with.is_none());
self.chain_with = Some(next_id);
self
}
/// Add a custom legalization action for `inst`.
///
/// The `func_name` parameter is the fully qualified name of a Rust function which takes the
/// same arguments as the `isa::Legalize` actions.
///
/// The custom function will be called to legalize `inst` and any return value is ignored.
pub fn custom_legalize(&mut self, inst: &Instruction, func_name: &'static str) {
assert!(
self.custom_legalizes
.insert(inst.camel_name.clone(), func_name)
.is_none(),
"custom legalization action for {} inserted twice",
inst.name
);
}
/// Add a legalization pattern to this group.
pub fn legalize(&mut self, src: DummyDef, dst: Vec<DummyDef>) {
let transform = Transform::new(src, dst);
transform.verify_legalize();
self.transforms.push(transform);
}
pub fn build_and_add_to(self, owner: &mut TransformGroups) -> TransformGroupIndex {
let next_id = owner.next_key();
owner.add(TransformGroup {
name: self.name,
doc: self.doc,
isa_name: self.isa_name,
id: next_id,
chain_with: self.chain_with,
custom_legalizes: self.custom_legalizes,
transforms: self.transforms,
})
}
}
pub(crate) struct TransformGroups {
groups: PrimaryMap<TransformGroupIndex, TransformGroup>,
}
impl TransformGroups {
pub fn new() -> Self {
Self {
groups: PrimaryMap::new(),
}
}
pub fn add(&mut self, new_group: TransformGroup) -> TransformGroupIndex {
for group in self.groups.values() {
assert!(
group.name != new_group.name,
"trying to insert {} for the second time",
new_group.name
);
}
self.groups.push(new_group)
}
pub fn get(&self, id: TransformGroupIndex) -> &TransformGroup {
&self.groups[id]
}
fn next_key(&self) -> TransformGroupIndex {
self.groups.next_key()
}
pub fn by_name(&self, name: &'static str) -> &TransformGroup {
for group in self.groups.values() {
if group.name == name {
return group;
}
}
panic!("transform group with name {} not found", name);
}
}
#[test]
#[should_panic]
fn test_double_custom_legalization() {
use crate::cdsl::formats::InstructionFormatBuilder;
use crate::cdsl::instructions::{AllInstructions, InstructionBuilder, InstructionGroupBuilder};
let nullary = InstructionFormatBuilder::new("nullary").build();
let mut dummy_all = AllInstructions::new();
let mut inst_group = InstructionGroupBuilder::new(&mut dummy_all);
inst_group.push(InstructionBuilder::new("dummy", "doc", &nullary));
let inst_group = inst_group.build();
let dummy_inst = inst_group.by_name("dummy");
let mut transform_group = TransformGroupBuilder::new("test", "doc");
transform_group.custom_legalize(&dummy_inst, "custom 1");
transform_group.custom_legalize(&dummy_inst, "custom 2");
}

View File

@@ -1,734 +0,0 @@
//! Generate transformations to legalize instructions without encodings.
use crate::cdsl::ast::{Def, DefPool, Expr, VarPool};
use crate::cdsl::isa::TargetIsa;
use crate::cdsl::operands::Operand;
use crate::cdsl::type_inference::Constraint;
use crate::cdsl::typevar::{TypeSet, TypeVar};
use crate::cdsl::xform::{Transform, TransformGroup, TransformGroups};
use crate::error;
use crate::gen_inst::gen_typesets_table;
use crate::srcgen::Formatter;
use crate::unique_table::UniqueTable;
use std::collections::{HashMap, HashSet};
use std::iter::FromIterator;
/// Given a `Def` node, emit code that extracts all the instruction fields from
/// `pos.func.dfg[iref]`.
///
/// Create local variables named after the `Var` instances in `node`.
///
/// Also create a local variable named `predicate` with the value of the evaluated instruction
/// predicate, or `true` if the node has no predicate.
fn unwrap_inst(transform: &Transform, fmt: &mut Formatter) -> bool {
let var_pool = &transform.var_pool;
let def_pool = &transform.def_pool;
let def = def_pool.get(transform.src);
let apply = &def.apply;
let inst = &apply.inst;
let iform = &inst.format;
fmt.comment(format!(
"Unwrap fields from instruction format {}",
def.to_comment_string(&transform.var_pool)
));
// Extract the Var arguments.
let arg_names = apply
.args
.iter()
.enumerate()
.filter(|(arg_num, _)| {
// Variable args are specially handled after extracting args.
!inst.operands_in[*arg_num].is_varargs()
})
.map(|(arg_num, arg)| match &arg {
Expr::Var(var_index) => var_pool.get(*var_index).name.as_ref(),
Expr::Literal(_) => {
let n = inst.imm_opnums.iter().position(|&i| i == arg_num).unwrap();
iform.imm_fields[n].member
}
})
.collect::<Vec<_>>()
.join(", ");
// May we need "args" in the values consumed by predicates?
let emit_args = iform.num_value_operands >= 1 || iform.has_value_list;
// We need a tuple:
// - if there's at least one value operand, then we emit a variable for the value, and the
// value list as args.
// - otherwise, if there's the count of immediate operands added to the presence of a value list exceeds one.
let need_tuple = if iform.num_value_operands >= 1 {
true
} else {
let mut imm_and_varargs = inst
.operands_in
.iter()
.filter(|op| op.is_immediate_or_entityref())
.count();
if iform.has_value_list {
imm_and_varargs += 1;
}
imm_and_varargs > 1
};
let maybe_args = if emit_args { ", args" } else { "" };
let defined_values = format!("{}{}", arg_names, maybe_args);
let tuple_or_value = if need_tuple {
format!("({})", defined_values)
} else {
defined_values
};
fmtln!(
fmt,
"let {} = if let ir::InstructionData::{} {{",
tuple_or_value,
iform.name
);
fmt.indent(|fmt| {
// Fields are encoded directly.
for field in &iform.imm_fields {
fmtln!(fmt, "{},", field.member);
}
if iform.has_value_list || iform.num_value_operands > 1 {
fmt.line("ref args,");
} else if iform.num_value_operands == 1 {
fmt.line("arg,");
}
fmt.line("..");
fmt.outdented_line("} = pos.func.dfg[inst] {");
if iform.has_value_list {
fmt.line("let args = args.as_slice(&pos.func.dfg.value_lists);");
} else if iform.num_value_operands == 1 {
fmt.line("let args = [arg];")
}
// Generate the values for the tuple.
let emit_one_value =
|fmt: &mut Formatter, needs_comma: bool, op_num: usize, op: &Operand| {
let comma = if needs_comma { "," } else { "" };
if op.is_immediate_or_entityref() {
let n = inst.imm_opnums.iter().position(|&i| i == op_num).unwrap();
fmtln!(fmt, "{}{}", iform.imm_fields[n].member, comma);
} else if op.is_value() {
let n = inst.value_opnums.iter().position(|&i| i == op_num).unwrap();
fmtln!(fmt, "pos.func.dfg.resolve_aliases(args[{}]),", n);
} else {
// This is a value list argument or a varargs.
assert!(iform.has_value_list || op.is_varargs());
}
};
if need_tuple {
fmt.line("(");
fmt.indent(|fmt| {
for (op_num, op) in inst.operands_in.iter().enumerate() {
let needs_comma = emit_args || op_num + 1 < inst.operands_in.len();
emit_one_value(fmt, needs_comma, op_num, op);
}
if emit_args {
fmt.line("args");
}
});
fmt.line(")");
} else {
// Only one of these can be true at the same time, otherwise we'd need a tuple.
emit_one_value(fmt, false, 0, &inst.operands_in[0]);
if emit_args {
fmt.line("args");
}
}
fmt.outdented_line("} else {");
fmt.line(r#"unreachable!("bad instruction format")"#);
});
fmtln!(fmt, "};");
fmt.empty_line();
assert_eq!(inst.operands_in.len(), apply.args.len());
for (i, op) in inst.operands_in.iter().enumerate() {
if op.is_varargs() {
let name = &var_pool
.get(apply.args[i].maybe_var().expect("vararg without name"))
.name;
let n = inst
.imm_opnums
.iter()
.chain(inst.value_opnums.iter())
.max()
.copied()
.unwrap_or(0);
fmtln!(fmt, "let {} = &Vec::from(&args[{}..]);", name, n);
}
}
for &op_num in &inst.value_opnums {
let arg = &apply.args[op_num];
if let Some(var_index) = arg.maybe_var() {
let var = var_pool.get(var_index);
if var.has_free_typevar() {
fmtln!(
fmt,
"let typeof_{} = pos.func.dfg.value_type({});",
var.name,
var.name
);
}
}
}
// If the definition creates results, detach the values and place them in locals.
let mut replace_inst = false;
if !def.defined_vars.is_empty() {
if def.defined_vars
== def_pool
.get(var_pool.get(def.defined_vars[0]).dst_def.unwrap())
.defined_vars
{
// Special case: The instruction replacing node defines the exact same values.
fmt.comment(format!(
"Results handled by {}.",
def_pool
.get(var_pool.get(def.defined_vars[0]).dst_def.unwrap())
.to_comment_string(var_pool)
));
fmt.line("let r = pos.func.dfg.inst_results(inst);");
for (i, &var_index) in def.defined_vars.iter().enumerate() {
let var = var_pool.get(var_index);
fmtln!(fmt, "let {} = &r[{}];", var.name, i);
fmtln!(
fmt,
"let typeof_{} = pos.func.dfg.value_type(*{});",
var.name,
var.name
);
}
replace_inst = true;
} else {
// Boring case: Detach the result values, capture them in locals.
for &var_index in &def.defined_vars {
fmtln!(fmt, "let {};", var_pool.get(var_index).name);
}
fmt.line("{");
fmt.indent(|fmt| {
fmt.line("let r = pos.func.dfg.inst_results(inst);");
for i in 0..def.defined_vars.len() {
let var = var_pool.get(def.defined_vars[i]);
fmtln!(fmt, "{} = r[{}];", var.name, i);
}
});
fmt.line("}");
for &var_index in &def.defined_vars {
let var = var_pool.get(var_index);
if var.has_free_typevar() {
fmtln!(
fmt,
"let typeof_{} = pos.func.dfg.value_type({});",
var.name,
var.name
);
}
}
}
}
replace_inst
}
fn build_derived_expr(tv: &TypeVar) -> String {
let base = match &tv.base {
Some(base) => base,
None => {
assert!(tv.name.starts_with("typeof_"));
return format!("Some({})", tv.name);
}
};
let base_expr = build_derived_expr(&base.type_var);
format!(
"{}.map(|t: crate::ir::Type| t.{}())",
base_expr,
base.derived_func.name()
)
}
/// Emit rust code for the given check.
///
/// The emitted code is a statement redefining the `predicate` variable like this:
/// let predicate = predicate && ...
fn emit_runtime_typecheck<'a>(
constraint: &'a Constraint,
type_sets: &mut UniqueTable<'a, TypeSet>,
fmt: &mut Formatter,
) {
match constraint {
Constraint::InTypeset(tv, ts) => {
let ts_index = type_sets.add(&ts);
fmt.comment(format!(
"{} must belong to {:?}",
tv.name,
type_sets.get(ts_index)
));
fmtln!(
fmt,
"let predicate = predicate && TYPE_SETS[{}].contains({});",
ts_index,
tv.name
);
}
Constraint::Eq(tv1, tv2) => {
fmtln!(
fmt,
"let predicate = predicate && match ({}, {}) {{",
build_derived_expr(tv1),
build_derived_expr(tv2)
);
fmt.indent(|fmt| {
fmt.line("(Some(a), Some(b)) => a == b,");
fmt.comment("On overflow, constraint doesn\'t apply");
fmt.line("_ => false,");
});
fmtln!(fmt, "};");
}
Constraint::WiderOrEq(tv1, tv2) => {
fmtln!(
fmt,
"let predicate = predicate && match ({}, {}) {{",
build_derived_expr(tv1),
build_derived_expr(tv2)
);
fmt.indent(|fmt| {
fmt.line("(Some(a), Some(b)) => a.wider_or_equal(b),");
fmt.comment("On overflow, constraint doesn\'t apply");
fmt.line("_ => false,");
});
fmtln!(fmt, "};");
}
}
}
/// Determine if `node` represents one of the value splitting instructions: `isplit` or `vsplit.
/// These instructions are lowered specially by the `legalize::split` module.
fn is_value_split(def: &Def) -> bool {
let name = &def.apply.inst.name;
name == "isplit" || name == "vsplit"
}
fn emit_dst_inst(def: &Def, def_pool: &DefPool, var_pool: &VarPool, fmt: &mut Formatter) {
let defined_vars = {
let vars = def
.defined_vars
.iter()
.map(|&var_index| var_pool.get(var_index).name.as_ref())
.collect::<Vec<&str>>();
if vars.len() == 1 {
vars[0].to_string()
} else {
format!("({})", vars.join(", "))
}
};
if is_value_split(def) {
// Split instructions are not emitted with the builder, but by calling special functions in
// the `legalizer::split` module. These functions will eliminate concat-split patterns.
fmt.line("let curpos = pos.position();");
fmt.line("let srcloc = pos.srcloc();");
fmtln!(
fmt,
"let {} = split::{}(pos.func, cfg, curpos, srcloc, {});",
defined_vars,
def.apply.inst.snake_name(),
def.apply.args[0].to_rust_code(var_pool)
);
return;
}
if def.defined_vars.is_empty() {
// This node doesn't define any values, so just insert the new instruction.
fmtln!(
fmt,
"pos.ins().{};",
def.apply.rust_builder(&def.defined_vars, var_pool)
);
return;
}
if let Some(src_def0) = var_pool.get(def.defined_vars[0]).src_def {
if def.defined_vars == def_pool.get(src_def0).defined_vars {
// The replacement instruction defines the exact same values as the source pattern.
// Unwrapping would have left the results intact. Replace the whole instruction.
fmtln!(
fmt,
"let {} = pos.func.dfg.replace(inst).{};",
defined_vars,
def.apply.rust_builder(&def.defined_vars, var_pool)
);
// We need to bump the cursor so following instructions are inserted *after* the
// replaced instruction.
fmt.line("if pos.current_inst() == Some(inst) {");
fmt.indent(|fmt| {
fmt.line("pos.next_inst();");
});
fmt.line("}");
return;
}
}
// Insert a new instruction.
let mut builder = format!("let {} = pos.ins()", defined_vars);
if def.defined_vars.len() == 1 && var_pool.get(def.defined_vars[0]).is_output() {
// Reuse the single source result value.
builder = format!(
"{}.with_result({})",
builder,
var_pool.get(def.defined_vars[0]).to_rust_code()
);
} else if def
.defined_vars
.iter()
.any(|&var_index| var_pool.get(var_index).is_output())
{
// There are more than one output values that can be reused.
let array = def
.defined_vars
.iter()
.map(|&var_index| {
let var = var_pool.get(var_index);
if var.is_output() {
format!("Some({})", var.name)
} else {
"None".into()
}
})
.collect::<Vec<_>>()
.join(", ");
builder = format!("{}.with_results([{}])", builder, array);
}
fmtln!(
fmt,
"{}.{};",
builder,
def.apply.rust_builder(&def.defined_vars, var_pool)
);
}
/// Emit code for `transform`, assuming that the opcode of transform's root instruction
/// has already been matched.
///
/// `inst: Inst` is the variable to be replaced. It is pointed to by `pos: Cursor`.
/// `dfg: DataFlowGraph` is available and mutable.
fn gen_transform<'a>(
replace_inst: bool,
transform: &'a Transform,
type_sets: &mut UniqueTable<'a, TypeSet>,
fmt: &mut Formatter,
) {
// Evaluate the instruction predicate if any.
let apply = &transform.def_pool.get(transform.src).apply;
let inst_predicate = apply
.inst_predicate_with_ctrl_typevar(&transform.var_pool)
.rust_predicate("pos.func");
let has_extra_constraints = !transform.type_env.constraints.is_empty();
if has_extra_constraints {
// Extra constraints rely on the predicate being a variable that we can rebind as we add
// more constraint predicates.
if let Some(pred) = &inst_predicate {
fmt.multi_line(&format!("let predicate = {};", pred));
} else {
fmt.line("let predicate = true;");
}
}
// Emit any runtime checks; these will rebind `predicate` emitted right above.
for constraint in &transform.type_env.constraints {
emit_runtime_typecheck(constraint, type_sets, fmt);
}
let do_expand = |fmt: &mut Formatter| {
// Emit any constants that must be created before use.
for (name, value) in transform.const_pool.iter() {
fmtln!(
fmt,
"let {} = pos.func.dfg.constants.insert(vec!{:?}.into());",
name,
value
);
}
// If we are adding some blocks, we need to recall the original block, such that we can
// recompute it.
if !transform.block_pool.is_empty() {
fmt.line("let orig_block = pos.current_block().unwrap();");
}
// If we're going to delete `inst`, we need to detach its results first so they can be
// reattached during pattern expansion.
if !replace_inst {
fmt.line("pos.func.dfg.clear_results(inst);");
}
// Emit new block creation.
for block in &transform.block_pool {
let var = transform.var_pool.get(block.name);
fmtln!(fmt, "let {} = pos.func.dfg.make_block();", var.name);
}
// Emit the destination pattern.
for &def_index in &transform.dst {
if let Some(block) = transform.block_pool.get(def_index) {
let var = transform.var_pool.get(block.name);
fmtln!(fmt, "pos.insert_block({});", var.name);
}
emit_dst_inst(
transform.def_pool.get(def_index),
&transform.def_pool,
&transform.var_pool,
fmt,
);
}
// Insert a new block after the last instruction, if needed.
let def_next_index = transform.def_pool.next_index();
if let Some(block) = transform.block_pool.get(def_next_index) {
let var = transform.var_pool.get(block.name);
fmtln!(fmt, "pos.insert_block({});", var.name);
}
// Delete the original instruction if we didn't have an opportunity to replace it.
if !replace_inst {
fmt.line("let removed = pos.remove_inst();");
fmt.line("debug_assert_eq!(removed, inst);");
}
if transform.block_pool.is_empty() {
if transform.def_pool.get(transform.src).apply.inst.is_branch {
// A branch might have been legalized into multiple branches, so we need to recompute
// the cfg.
fmt.line("cfg.recompute_block(pos.func, pos.current_block().unwrap());");
}
} else {
// Update CFG for the new blocks.
fmt.line("cfg.recompute_block(pos.func, orig_block);");
for block in &transform.block_pool {
let var = transform.var_pool.get(block.name);
fmtln!(fmt, "cfg.recompute_block(pos.func, {});", var.name);
}
}
fmt.line("return true;");
};
// Guard the actual expansion by `predicate`.
if has_extra_constraints {
fmt.line("if predicate {");
fmt.indent(|fmt| {
do_expand(fmt);
});
fmt.line("}");
} else if let Some(pred) = &inst_predicate {
fmt.multi_line(&format!("if {} {{", pred));
fmt.indent(|fmt| {
do_expand(fmt);
});
fmt.line("}");
} else {
// Unconditional transform (there was no predicate), just emit it.
do_expand(fmt);
}
}
fn gen_transform_group<'a>(
group: &'a TransformGroup,
transform_groups: &TransformGroups,
type_sets: &mut UniqueTable<'a, TypeSet>,
fmt: &mut Formatter,
) {
fmt.doc_comment(group.doc);
fmt.line("#[allow(unused_variables,unused_assignments,unused_imports,non_snake_case)]");
// Function arguments.
fmtln!(fmt, "pub fn {}(", group.name);
fmt.indent(|fmt| {
fmt.line("inst: crate::ir::Inst,");
fmt.line("func: &mut crate::ir::Function,");
fmt.line("cfg: &mut crate::flowgraph::ControlFlowGraph,");
fmt.line("isa: &dyn crate::isa::TargetIsa,");
});
fmtln!(fmt, ") -> bool {");
// Function body.
fmt.indent(|fmt| {
fmt.line("use crate::ir::InstBuilder;");
fmt.line("use crate::cursor::{Cursor, FuncCursor};");
fmt.line("let mut pos = FuncCursor::new(func).at_inst(inst);");
fmt.line("pos.use_srcloc(inst);");
// Group the transforms by opcode so we can generate a big switch.
// Preserve ordering.
let mut inst_to_transforms = HashMap::new();
for transform in &group.transforms {
let def_index = transform.src;
let inst = &transform.def_pool.get(def_index).apply.inst;
inst_to_transforms
.entry(inst.camel_name.clone())
.or_insert_with(Vec::new)
.push(transform);
}
let mut sorted_inst_names = Vec::from_iter(inst_to_transforms.keys());
sorted_inst_names.sort();
fmt.line("{");
fmt.indent(|fmt| {
fmt.line("match pos.func.dfg[inst].opcode() {");
fmt.indent(|fmt| {
for camel_name in sorted_inst_names {
fmtln!(fmt, "ir::Opcode::{} => {{", camel_name);
fmt.indent(|fmt| {
let transforms = inst_to_transforms.get(camel_name).unwrap();
// Unwrap the source instruction, create local variables for the input variables.
let replace_inst = unwrap_inst(&transforms[0], fmt);
fmt.empty_line();
for (i, transform) in transforms.iter().enumerate() {
if i > 0 {
fmt.empty_line();
}
gen_transform(replace_inst, transform, type_sets, fmt);
}
});
fmtln!(fmt, "}");
fmt.empty_line();
}
// Emit the custom transforms. The Rust compiler will complain about any overlap with
// the normal transforms.
let mut sorted_custom_legalizes = Vec::from_iter(&group.custom_legalizes);
sorted_custom_legalizes.sort();
for (inst_camel_name, func_name) in sorted_custom_legalizes {
fmtln!(fmt, "ir::Opcode::{} => {{", inst_camel_name);
fmt.indent(|fmt| {
fmtln!(fmt, "{}(inst, func, cfg, isa);", func_name);
fmt.line("return true;");
});
fmtln!(fmt, "}");
fmt.empty_line();
}
// We'll assume there are uncovered opcodes.
fmt.line("_ => {},");
});
fmt.line("}");
});
fmt.line("}");
// If we fall through, nothing was expanded; call the chain if any.
match &group.chain_with {
Some(group_id) => fmtln!(
fmt,
"{}(inst, func, cfg, isa)",
transform_groups.get(*group_id).rust_name()
),
None => fmt.line("false"),
};
});
fmtln!(fmt, "}");
fmt.empty_line();
}
/// Generate legalization functions for `isa` and add any shared `TransformGroup`s
/// encountered to `shared_groups`.
///
/// Generate `TYPE_SETS` and `LEGALIZE_ACTIONS` tables.
fn gen_isa(
isa: &TargetIsa,
transform_groups: &TransformGroups,
shared_group_names: &mut HashSet<&'static str>,
fmt: &mut Formatter,
) {
let mut type_sets = UniqueTable::new();
for group_index in isa.transitive_transform_groups(transform_groups) {
let group = transform_groups.get(group_index);
match group.isa_name {
Some(isa_name) => {
assert!(
isa_name == isa.name,
"ISA-specific legalizations must be used by the same ISA"
);
gen_transform_group(group, transform_groups, &mut type_sets, fmt);
}
None => {
shared_group_names.insert(group.name);
}
}
}
gen_typesets_table(&type_sets, fmt);
let direct_groups = isa.direct_transform_groups();
fmtln!(
fmt,
"pub static LEGALIZE_ACTIONS: [isa::Legalize; {}] = [",
direct_groups.len()
);
fmt.indent(|fmt| {
for &group_index in direct_groups {
fmtln!(fmt, "{},", transform_groups.get(group_index).rust_name());
}
});
fmtln!(fmt, "];");
}
/// Generate the legalizer files.
pub(crate) fn generate(
isas: &[TargetIsa],
transform_groups: &TransformGroups,
extra_legalization_groups: &[&'static str],
filename_prefix: &str,
out_dir: &str,
) -> Result<(), error::Error> {
let mut shared_group_names = HashSet::new();
for isa in isas {
let mut fmt = Formatter::new();
gen_isa(isa, transform_groups, &mut shared_group_names, &mut fmt);
fmt.update_file(format!("{}-{}.rs", filename_prefix, isa.name), out_dir)?;
}
// Add extra legalization groups that were explicitly requested.
for group in extra_legalization_groups {
shared_group_names.insert(group);
}
// Generate shared legalize groups.
let mut fmt = Formatter::new();
// Generate shared legalize groups.
let mut type_sets = UniqueTable::new();
let mut sorted_shared_group_names = Vec::from_iter(shared_group_names);
sorted_shared_group_names.sort();
for group_name in &sorted_shared_group_names {
let group = transform_groups.by_name(group_name);
gen_transform_group(group, transform_groups, &mut type_sets, &mut fmt);
}
gen_typesets_table(&type_sets, &mut fmt);
fmt.update_file(format!("{}r.rs", filename_prefix), out_dir)?;
Ok(())
}

View File

@@ -1,6 +1,4 @@
use crate::cdsl::instructions::InstructionPredicateMap;
use crate::cdsl::isa::TargetIsa;
use crate::cdsl::recipes::Recipes;
use crate::cdsl::regs::{IsaRegs, IsaRegsBuilder, RegBankBuilder, RegClassBuilder};
use crate::cdsl::settings::{SettingGroup, SettingGroupBuilder};
@@ -52,20 +50,5 @@ pub(crate) fn define(shared_defs: &mut SharedDefinitions) -> TargetIsa {
let settings = define_settings(&shared_defs.settings);
let regs = define_regs();
let cpu_modes = vec![];
// TODO implement arm32 recipes.
let recipes = Recipes::new();
// TODO implement arm32 encodings and predicates.
let encodings_predicates = InstructionPredicateMap::new();
TargetIsa::new(
"arm32",
settings,
regs,
recipes,
cpu_modes,
encodings_predicates,
)
TargetIsa::new("arm32", settings, regs)
}

View File

@@ -1,6 +1,4 @@
use crate::cdsl::instructions::InstructionPredicateMap;
use crate::cdsl::isa::TargetIsa;
use crate::cdsl::recipes::Recipes;
use crate::cdsl::regs::{IsaRegs, IsaRegsBuilder, RegBankBuilder, RegClassBuilder};
use crate::cdsl::settings::{SettingGroup, SettingGroupBuilder};
@@ -51,20 +49,5 @@ pub(crate) fn define(shared_defs: &mut SharedDefinitions) -> TargetIsa {
let settings = define_settings(&shared_defs.settings);
let regs = define_registers();
let cpu_modes = vec![];
// TODO implement arm64 recipes.
let recipes = Recipes::new();
// TODO implement arm64 encodings and predicates.
let encodings_predicates = InstructionPredicateMap::new();
TargetIsa::new(
"arm64",
settings,
regs,
recipes,
cpu_modes,
encodings_predicates,
)
TargetIsa::new("arm64", settings, regs)
}

View File

@@ -1,6 +1,4 @@
use crate::cdsl::instructions::InstructionPredicateMap;
use crate::cdsl::isa::TargetIsa;
use crate::cdsl::recipes::Recipes;
use crate::cdsl::regs::IsaRegsBuilder;
use crate::cdsl::settings::{SettingGroup, SettingGroupBuilder};
@@ -46,17 +44,6 @@ fn define_settings(_shared: &SettingGroup) -> SettingGroup {
pub(crate) fn define(shared_defs: &mut SharedDefinitions) -> TargetIsa {
let settings = define_settings(&shared_defs.settings);
let regs = IsaRegsBuilder::new().build();
let recipes = Recipes::new();
let encodings_predicates = InstructionPredicateMap::new();
let cpu_modes = vec![];
TargetIsa::new(
"s390x",
settings,
regs,
recipes,
cpu_modes,
encodings_predicates,
)
TargetIsa::new("s390x", settings, regs)
}

View File

@@ -1,6 +1,4 @@
use crate::cdsl::instructions::{InstructionGroupBuilder, InstructionPredicateMap};
use crate::cdsl::isa::TargetIsa;
use crate::cdsl::recipes::Recipes;
use crate::cdsl::regs::IsaRegsBuilder;
use crate::shared::Definitions as SharedDefinitions;
@@ -10,16 +8,5 @@ pub(crate) mod settings;
pub(crate) fn define(shared_defs: &mut SharedDefinitions) -> TargetIsa {
let settings = settings::define(&shared_defs.settings);
let inst_group = InstructionGroupBuilder::new(&mut shared_defs.all_instructions).build();
let cpu_modes = vec![];
TargetIsa::new(
"x86",
settings,
IsaRegsBuilder::new().build(),
Recipes::new(),
cpu_modes,
InstructionPredicateMap::new(),
)
TargetIsa::new("x86", settings, IsaRegsBuilder::new().build())
}

View File

@@ -8,7 +8,6 @@ pub mod error;
pub mod isa;
mod gen_inst;
mod gen_legalizer;
mod gen_registers;
mod gen_settings;
mod gen_types;
@@ -55,21 +54,6 @@ pub fn generate(
&out_dir,
)?;
let extra_legalization_groups: &[&'static str] = if !new_backend_isas.is_empty() {
// The new backend only requires the "expand" legalization group.
&["expand"]
} else {
&[]
};
gen_legalizer::generate(
&target_isas,
&shared_defs.transform_groups,
extra_legalization_groups,
"legalize",
&out_dir,
)?;
for isa in target_isas {
gen_registers::generate(&isa, &format!("registers-{}.rs", isa.name), &out_dir)?;

View File

@@ -1,7 +1,7 @@
#![allow(non_snake_case)]
use crate::cdsl::instructions::{
AllInstructions, InstructionBuilder as Inst, InstructionGroup, InstructionGroupBuilder,
AllInstructions, InstructionBuilder as Inst, InstructionGroupBuilder,
};
use crate::cdsl::operands::Operand;
use crate::cdsl::type_inference::Constraint::WiderOrEq;
@@ -767,7 +767,7 @@ pub(crate) fn define(
formats: &Formats,
imm: &Immediates,
entities: &EntityRefs,
) -> InstructionGroup {
) {
let mut ig = InstructionGroupBuilder::new(all_instructions);
define_control_flow(&mut ig, formats, imm, entities);
@@ -4647,6 +4647,4 @@ pub(crate) fn define(
)
.other_side_effects(true),
);
ig.build()
}

File diff suppressed because it is too large Load Diff

View File

@@ -4,14 +4,12 @@ pub mod entities;
pub mod formats;
pub mod immediates;
pub mod instructions;
pub mod legalize;
pub mod settings;
pub mod types;
use crate::cdsl::formats::{FormatStructure, InstructionFormat};
use crate::cdsl::instructions::{AllInstructions, InstructionGroup};
use crate::cdsl::instructions::{AllInstructions};
use crate::cdsl::settings::SettingGroup;
use crate::cdsl::xform::TransformGroups;
use crate::shared::entities::EntityRefs;
use crate::shared::formats::Formats;
@@ -24,10 +22,8 @@ use std::rc::Rc;
pub(crate) struct Definitions {
pub settings: SettingGroup,
pub all_instructions: AllInstructions,
pub instructions: InstructionGroup,
pub imm: Immediates,
pub formats: Formats,
pub transform_groups: TransformGroups,
pub entities: EntityRefs,
}
@@ -37,17 +33,13 @@ pub(crate) fn define() -> Definitions {
let immediates = Immediates::new();
let entities = EntityRefs::new();
let formats = Formats::new(&immediates, &entities);
let instructions =
instructions::define(&mut all_instructions, &formats, &immediates, &entities);
let transform_groups = legalize::define(&instructions, &immediates);
Definitions {
settings: settings::define(),
all_instructions,
instructions,
imm: immediates,
formats,
transform_groups,
entities,
}
}

View File

@@ -77,15 +77,6 @@ impl Formatter {
}
}
/// Get a string containing whitespace outdented one level. Used for
/// lines of code that are inside a single indented block.
fn get_outdent(&mut self) -> String {
self.indent_pop();
let s = self.get_indent();
self.indent_push();
s
}
/// Add an indented line.
pub fn line(&mut self, contents: impl AsRef<str>) {
let indented_line = format!("{}{}\n", self.get_indent(), contents.as_ref());
@@ -97,12 +88,6 @@ impl Formatter {
self.lines.push("\n".to_string());
}
/// Emit a line outdented one level.
pub fn outdented_line(&mut self, s: &str) {
let new_line = format!("{}{}\n", self.get_outdent(), s);
self.lines.push(new_line);
}
/// Write `self.lines` to a file.
pub fn update_file(
&self,

View File

@@ -32,9 +32,6 @@ impl<'entries, T: Eq + Hash> UniqueTable<'entries, T> {
pub fn len(&self) -> usize {
self.table.len()
}
pub fn get(&self, index: usize) -> &T {
self.table[index]
}
pub fn iter(&self) -> slice::Iter<&'entries T> {
self.table.iter()
}