1
Fork 0

Auto merge of #111673 - cjgillot:dominator-preprocess, r=cjgillot,tmiasko

Preprocess and cache dominator tree

Preprocessing dominators has a very strong effect for https://github.com/rust-lang/rust/pull/111344.
That pass checks that assignments dominate their uses repeatedly. Using the unprocessed dominator tree caused a quadratic runtime (number of bbs x depth of the dominator tree).

This PR also caches the dominator tree and the pre-processed dominators in the MIR cfg cache.

Rebase of https://github.com/rust-lang/rust/pull/107157
cc `@tmiasko`
This commit is contained in:
bors 2023-05-24 16:18:21 +00:00
commit 97d328012b
9 changed files with 107 additions and 51 deletions

View file

@ -9,6 +9,7 @@ use rustc_index::{IndexSlice, IndexVec};
use rustc_middle::mir::coverage::*;
use rustc_middle::mir::{self, BasicBlock, BasicBlockData, Terminator, TerminatorKind};
use std::cmp::Ordering;
use std::ops::{Index, IndexMut};
const ID_SEPARATOR: &str = ",";
@ -212,8 +213,12 @@ impl CoverageGraph {
}
#[inline(always)]
pub fn dominators(&self) -> &Dominators<BasicCoverageBlock> {
self.dominators.as_ref().unwrap()
pub fn rank_partial_cmp(
&self,
a: BasicCoverageBlock,
b: BasicCoverageBlock,
) -> Option<Ordering> {
self.dominators.as_ref().unwrap().rank_partial_cmp(a, b)
}
}
@ -650,26 +655,6 @@ pub(super) fn find_loop_backedges(
let mut backedges = IndexVec::from_elem_n(Vec::<BasicCoverageBlock>::new(), num_bcbs);
// Identify loops by their backedges.
//
// The computational complexity is bounded by: n(s) x d where `n` is the number of
// `BasicCoverageBlock` nodes (the simplified/reduced representation of the CFG derived from the
// MIR); `s` is the average number of successors per node (which is most likely less than 2, and
// independent of the size of the function, so it can be treated as a constant);
// and `d` is the average number of dominators per node.
//
// The average number of dominators depends on the size and complexity of the function, and
// nodes near the start of the function's control flow graph typically have less dominators
// than nodes near the end of the CFG. Without doing a detailed mathematical analysis, I
// think the resulting complexity has the characteristics of O(n log n).
//
// The overall complexity appears to be comparable to many other MIR transform algorithms, and I
// don't expect that this function is creating a performance hot spot, but if this becomes an
// issue, there may be ways to optimize the `dominates` algorithm (as indicated by an
// existing `FIXME` comment in that code), or possibly ways to optimize it's usage here, perhaps
// by keeping track of results for visited `BasicCoverageBlock`s if they can be used to short
// circuit downstream `dominates` checks.
//
// For now, that kind of optimization seems unnecessarily complicated.
for (bcb, _) in basic_coverage_blocks.iter_enumerated() {
for &successor in &basic_coverage_blocks.successors[bcb] {
if basic_coverage_blocks.dominates(successor, bcb) {

View file

@ -344,7 +344,7 @@ impl<'a, 'tcx> CoverageSpans<'a, 'tcx> {
// before the dominated equal spans). When later comparing two spans in
// order, the first will either dominate the second, or they will have no
// dominator relationship.
self.basic_coverage_blocks.dominators().rank_partial_cmp(a.bcb, b.bcb)
self.basic_coverage_blocks.rank_partial_cmp(a.bcb, b.bcb)
}
} else {
// Sort hi() in reverse order so shorter spans are attempted after longer spans.