This commit is contained in:
Chris Fallin 2022-11-22 21:43:50 -08:00
parent d954fa9fe6
commit abc46f1d14
3 changed files with 14 additions and 257 deletions

View file

@ -2,31 +2,24 @@
use crate::entity::{EntityRef, PerEntity};
use crate::ir::{Block, FunctionBody, Value, ValueDef};
use crate::passes::rpo::{RPOIndex, RPO};
use anyhow::Result;
pub mod stackify;
use stackify::{Mark, Marks, RPOIndex, RPO};
pub mod treeify;
use treeify::Trees;
pub struct WasmBackend<'a> {
body: &'a FunctionBody,
rpo: RPO,
marks: Marks,
trees: Trees,
}
impl<'a> WasmBackend<'a> {
pub fn new(body: &'a FunctionBody) -> Result<WasmBackend<'a>> {
let rpo = RPO::compute(body);
let marks = Marks::compute(body, &rpo)?;
let trees = Trees::compute(body);
Ok(WasmBackend {
body,
rpo,
marks,
trees,
})
Ok(WasmBackend { body, rpo, trees })
}
pub fn compile(&self) -> Result<Vec<u8>> {

View file

@ -1,251 +1,15 @@
//! Stackify implementation to produce structured control flow from an
//! arbitrary CFG.
//!
//! Note on algorithm:
//! See the paper
//!
//! - We sort in RPO, then mark loops, then place blocks within loops
//! or at top level to give forward edges appropriate targets.
//! - Norman Ramsey. Beyond Relooper: recursive translation of
//! unstructured control flow to structured control flow. In ICFP
//! 2022 (Functional Pearl). https://dl.acm.org/doi/10.1145/3547621
//!
//! - The RPO sort order we choose is quite special: we need loop
//! bodies to be placed contiguously, without blocks that do not
//! belong to the loop in the middle. Otherwise we may not be able
//! to properly nest a block to allow a forward edge.
//!
//! Consider the following CFG:
//!
//! ```plain
//! 1
//! |
//! 2 <-.
//! / | |
//! | 3 --'
//! | |
//! `> 4
//! |
//! 5
//! ```
//!
//! A normal RPO sort may produce 1, 2, 4, 5, 3 or 1, 2, 3, 4, 5
//! depending on which child order it chooses from block 2. (If it
//! visits 3 first, it will emit it first in postorder hence it comes
//! last.)
//!
//! One way of ensuring we get the right order would be to compute the
//! loop nest and make note of loops when choosing children to visit,
//! but we really would rather not do that, since we don't otherwise
//! have the infrastructure to compute that or the need for it.
//!
//! Instead, we keep a "pending" list: as we have nodes on the stack
//! during postorder traversal, we keep a list of other children that
//! we will visit once we get back to a given level. If another node
//! is pending, and is a successor we are considering, we visit it
//! *first* in postorder, so it is last in RPO. This is a way to
//! ensure that (e.g.) block 4 above is visited first when considering
//! successors of block 2.
//! for more details on how this algorithm works.
use crate::declare_entity;
use crate::entity::{EntityRef, EntityVec, PerEntity};
use crate::cfg::CFGInfo;
use crate::ir::{Block, FunctionBody};
use crate::passes::rpo::{RPOIndex, RPO};
use std::collections::{HashMap, HashSet};
declare_entity!(RPOIndex, "rpo");
impl RPOIndex {
pub fn prev(self) -> RPOIndex {
RPOIndex(self.0.checked_sub(1).unwrap())
}
}
pub struct RPO {
pub order: EntityVec<RPOIndex, Block>,
pub rev: PerEntity<Block, RPOIndex>,
}
impl RPO {
pub fn compute(body: &FunctionBody) -> RPO {
let mut postorder = vec![];
let mut visited = HashSet::new();
let mut pending = vec![];
let mut pending_idx = HashMap::new();
visited.insert(body.entry);
Self::visit(
body,
body.entry,
&mut visited,
&mut pending,
&mut pending_idx,
&mut postorder,
);
postorder.reverse();
let mut rev = PerEntity::default();
for (i, block) in postorder.iter().copied().enumerate() {
rev[block] = RPOIndex(i as u32);
}
RPO {
order: EntityVec::from(postorder),
rev,
}
}
fn visit(
body: &FunctionBody,
block: Block,
visited: &mut HashSet<Block>,
pending: &mut Vec<Block>,
pending_idx: &mut HashMap<Block, usize>,
postorder: &mut Vec<Block>,
) {
// `pending` is a Vec, not a Set; we prioritize based on
// position (first in pending go first in postorder -> last in
// RPO). A case with nested loops to show why this matters:
//
// TODO example
let pending_top = pending.len();
pending.extend(body.blocks[block].succs.clone());
// Sort new entries in `pending` by index at which they appear
// earlier. Those that don't appear in `pending` at all should
// be visited last (to appear in RPO first), so we want `None`
// values to sort first here (hence the "unwrap or MAX"
// idiom). Then those that appear earlier in `pending` should
// be visited earlier here to appear later in RPO, so they
// sort later.
pending[pending_top..]
.sort_by_key(|entry| pending_idx.get(entry).copied().unwrap_or(usize::MAX));
// Above we placed items in order they are to be visited;
// below we pop off the end, so we reverse here.
pending[pending_top..].reverse();
// Now update indices in `pending_idx`: insert entries for
// those blocks not yet present.
for i in pending_top..pending.len() {
pending_idx.entry(pending[i]).or_insert(i);
}
for _ in 0..(pending.len() - pending_top) {
let succ = pending.pop().unwrap();
if pending_idx.get(&succ) == Some(&pending.len()) {
pending_idx.remove(&succ);
}
if visited.insert(succ) {
Self::visit(body, succ, visited, pending, pending_idx, postorder);
}
}
postorder.push(block);
}
}
/// Start and end marks for loops.
#[derive(Debug)]
pub struct Marks(HashMap<RPOIndex, Vec<Mark>>);
// Sorting-order note: Loop comes second, so Blocks sort first with
// smaller regions first. Thus, *reverse* sort order places loops
// outermost then larger blocks before smaller blocks.
#[derive(Debug, PartialEq, Eq, PartialOrd, Ord)]
pub enum Mark {
Block { last_inclusive: RPOIndex },
Loop { last_inclusive: RPOIndex },
}
impl Marks {
pub fn compute(body: &FunctionBody, rpo: &RPO) -> anyhow::Result<Marks> {
let mut marks = HashMap::new();
// Pass 1: Place loop markers.
let mut loop_end: HashMap<RPOIndex, RPOIndex> = HashMap::new();
for (rpo_block, &block) in rpo.order.entries() {
for &succ in &body.blocks[block].succs {
let rpo_succ = rpo.rev[succ];
assert!(rpo_succ.is_valid());
if rpo_succ <= rpo_block {
let end = loop_end.entry(rpo_succ).or_insert(RPOIndex::invalid());
if end.is_invalid() {
*end = rpo_block;
} else {
// Already-existing loop header. Adjust `end`.
*end = std::cmp::max(*end, rpo_block);
}
}
}
}
// Pass 2: properly nest loops by extending the reach of outer
// loops to fully contain inner loops.
for rpo_block in rpo.order.iter().rev() {
if let Some(rpo_loop_end) = loop_end.get(&rpo_block).copied() {
let mut updated_end = rpo_loop_end;
for body_block in rpo_block.index()..=rpo_loop_end.index() {
let body_block = RPOIndex::new(body_block);
if let Some(inner_end) = loop_end.get(&body_block).copied() {
updated_end = std::cmp::max(updated_end, inner_end);
}
}
if updated_end != rpo_loop_end {
loop_end.insert(rpo_block, updated_end);
}
}
}
// Pass 3: compute location of innermost loop for each
// block.
let mut innermost_loop: PerEntity<RPOIndex, Option<RPOIndex>> = PerEntity::default();
let mut loop_stack: Vec<(RPOIndex, RPOIndex)> = vec![];
for rpo_block in rpo.order.iter() {
while let Some(innermost) = loop_stack.last() {
if innermost.1 >= rpo_block {
break;
}
loop_stack.pop();
}
if let Some(rpo_loop_end) = loop_end.get(&rpo_block).copied() {
loop_stack.push((rpo_block, rpo_loop_end));
}
innermost_loop[rpo_block] = loop_stack.last().map(|lp| lp.0);
}
// Copy loop-start markers over.
for (lp, end) in loop_end {
marks.insert(
lp,
vec![Mark::Loop {
last_inclusive: end,
}],
);
}
// Pass 4: place block markers.
for (rpo_block, &block) in rpo.order.entries() {
for &succ in &body.blocks[block].succs {
let rpo_succ = rpo.rev[succ];
assert!(rpo_succ.is_valid());
if rpo_succ > rpo_block {
// Determine the innermost loop for the target,
// and add the block just inside the loop.
let block_start = innermost_loop[rpo_succ].unwrap_or(RPOIndex(0));
let start_marks = marks.entry(block_start).or_insert_with(|| vec![]);
let mark = Mark::Block {
last_inclusive: rpo_succ.prev(),
};
start_marks.push(mark);
}
}
}
// Sort markers at each block.
for marklist in marks.values_mut() {
marklist.sort();
marklist.dedup();
marklist.reverse();
}
Ok(Marks(marks))
}
}

View file

@ -44,19 +44,19 @@ use std::collections::{HashMap, HashSet};
declare_entity!(RPOIndex, "rpo");
impl RPOIndex {
fn prev(self) -> RPOIndex {
pub fn prev(self) -> RPOIndex {
RPOIndex::from(self.0.checked_sub(1).unwrap())
}
}
#[derive(Clone, Debug, Default)]
struct RPO {
order: EntityVec<RPOIndex, Block>,
rev: PerEntity<Block, Option<RPOIndex>>,
pub struct RPO {
pub order: EntityVec<RPOIndex, Block>,
pub rev: PerEntity<Block, Option<RPOIndex>>,
}
impl RPO {
fn compute(body: &FunctionBody) -> RPO {
pub fn compute(body: &FunctionBody) -> RPO {
let mut postorder = vec![];
let mut visited = HashSet::new();
let mut pending = vec![];