2015-11-10 14:38:36 -06:00
|
|
|
// Copyright 2015 The Rust Project Developers. See the COPYRIGHT
|
|
|
|
// file at the top-level directory of this distribution and at
|
|
|
|
// http://rust-lang.org/COPYRIGHT.
|
|
|
|
//
|
|
|
|
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
|
|
|
|
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
|
|
|
|
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
|
|
|
|
// option. This file may not be copied, modified, or distributed
|
|
|
|
// except according to those terms.
|
|
|
|
|
2016-08-15 20:59:50 -05:00
|
|
|
//! A number of passes which remove various redundancies in the CFG.
|
2016-06-06 14:58:28 -05:00
|
|
|
//!
|
2016-08-15 20:59:50 -05:00
|
|
|
//! The `SimplifyCfg` pass gets rid of unnecessary blocks in the CFG, whereas the `SimplifyLocals`
|
|
|
|
//! gets rid of all the unnecessary local variable declarations.
|
2016-06-06 14:58:28 -05:00
|
|
|
//!
|
2016-08-15 20:59:50 -05:00
|
|
|
//! The `SimplifyLocals` pass is kinda expensive and therefore not very suitable to be run often.
|
|
|
|
//! Most of the passes should not care or be impacted in meaningful ways due to extra locals
|
|
|
|
//! either, so running the pass once, right before translation, should suffice.
|
|
|
|
//!
|
|
|
|
//! On the other side of the spectrum, the `SimplifyCfg` pass is considerably cheap to run, thus
|
|
|
|
//! one should run it after every pass which may modify CFG in significant ways. This pass must
|
|
|
|
//! also be run before any analysis passes because it removes dead blocks, and some of these can be
|
|
|
|
//! ill-typed.
|
|
|
|
//!
|
|
|
|
//! The cause of this typing issue is typeck allowing most blocks whose end is not reachable have
|
|
|
|
//! an arbitrary return type, rather than having the usual () return type (as a note, typeck's
|
|
|
|
//! notion of reachability is in fact slightly weaker than MIR CFG reachability - see #31617). A
|
|
|
|
//! standard example of the situation is:
|
2016-06-06 14:58:28 -05:00
|
|
|
//!
|
|
|
|
//! ```rust
|
|
|
|
//! fn example() {
|
|
|
|
//! let _a: char = { return; };
|
|
|
|
//! }
|
|
|
|
//! ```
|
|
|
|
//!
|
2016-08-15 20:59:50 -05:00
|
|
|
//! Here the block (`{ return; }`) has the return type `char`, rather than `()`, but the MIR we
|
|
|
|
//! naively generate still contains the `_a = ()` write in the unreachable block "after" the
|
|
|
|
//! return.
|
2016-06-06 14:58:28 -05:00
|
|
|
|
2015-11-14 16:52:17 -06:00
|
|
|
use rustc_data_structures::bitvec::BitVector;
|
2016-06-07 09:28:36 -05:00
|
|
|
use rustc_data_structures::indexed_vec::{Idx, IndexVec};
|
2016-03-22 10:30:57 -05:00
|
|
|
use rustc::ty::TyCtxt;
|
2016-09-19 15:50:00 -05:00
|
|
|
use rustc::mir::*;
|
2017-04-25 17:23:33 -05:00
|
|
|
use rustc::mir::transform::{MirPass, MirSource};
|
2016-08-15 20:59:50 -05:00
|
|
|
use rustc::mir::visit::{MutVisitor, Visitor, LvalueContext};
|
2017-04-25 17:23:33 -05:00
|
|
|
use std::borrow::Cow;
|
2016-02-26 10:05:50 -06:00
|
|
|
|
2017-04-25 17:23:33 -05:00
|
|
|
pub struct SimplifyCfg { label: String }
|
2015-11-10 14:38:36 -06:00
|
|
|
|
2017-04-25 17:23:33 -05:00
|
|
|
impl SimplifyCfg {
|
|
|
|
pub fn new(label: &str) -> Self {
|
|
|
|
SimplifyCfg { label: format!("SimplifyCfg-{}", label) }
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2015-11-14 16:52:17 -06:00
|
|
|
}
|
|
|
|
|
2017-03-09 12:36:01 -06:00
|
|
|
pub fn simplify_cfg(mir: &mut Mir) {
|
|
|
|
CfgSimplifier::new(mir).simplify();
|
|
|
|
remove_dead_blocks(mir);
|
|
|
|
|
|
|
|
// FIXME: Should probably be moved into some kind of pass manager
|
|
|
|
mir.basic_blocks_mut().raw.shrink_to_fit();
|
|
|
|
}
|
|
|
|
|
2017-04-25 17:23:33 -05:00
|
|
|
impl MirPass for SimplifyCfg {
|
|
|
|
fn name<'a>(&'a self) -> Cow<'a, str> {
|
|
|
|
Cow::Borrowed(&self.label)
|
2015-11-14 16:52:17 -06:00
|
|
|
}
|
|
|
|
|
2017-04-25 17:23:33 -05:00
|
|
|
fn run_pass<'a, 'tcx>(&self,
|
|
|
|
_tcx: TyCtxt<'a, 'tcx, 'tcx>,
|
|
|
|
_src: MirSource,
|
|
|
|
mir: &mut Mir<'tcx>) {
|
|
|
|
debug!("SimplifyCfg({:?}) - simplifying {:?}", self.label, mir);
|
|
|
|
simplify_cfg(mir);
|
2016-06-08 13:03:06 -05:00
|
|
|
}
|
|
|
|
}
|
2015-11-14 16:52:17 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
pub struct CfgSimplifier<'a, 'tcx: 'a> {
|
|
|
|
basic_blocks: &'a mut IndexVec<BasicBlock, BasicBlockData<'tcx>>,
|
|
|
|
pred_count: IndexVec<BasicBlock, u32>
|
|
|
|
}
|
2015-11-10 14:38:36 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
impl<'a, 'tcx: 'a> CfgSimplifier<'a, 'tcx> {
|
2017-02-08 03:24:49 -06:00
|
|
|
pub fn new(mir: &'a mut Mir<'tcx>) -> Self {
|
2016-06-08 16:10:15 -05:00
|
|
|
let mut pred_count = IndexVec::from_elem(0u32, mir.basic_blocks());
|
2015-11-14 16:52:17 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
// we can't use mir.predecessors() here because that counts
|
|
|
|
// dead blocks, which we don't want to.
|
2016-10-04 08:23:01 -05:00
|
|
|
pred_count[START_BLOCK] = 1;
|
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
for (_, data) in traversal::preorder(mir) {
|
|
|
|
if let Some(ref term) = data.terminator {
|
|
|
|
for &tgt in term.successors().iter() {
|
|
|
|
pred_count[tgt] += 1;
|
2015-11-14 16:52:17 -06:00
|
|
|
}
|
|
|
|
}
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
let basic_blocks = mir.basic_blocks_mut();
|
|
|
|
|
|
|
|
CfgSimplifier {
|
|
|
|
basic_blocks: basic_blocks,
|
|
|
|
pred_count: pred_count
|
2015-11-14 16:52:17 -06:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-08 03:24:49 -06:00
|
|
|
pub fn simplify(mut self) {
|
2016-06-08 16:10:15 -05:00
|
|
|
loop {
|
|
|
|
let mut changed = false;
|
|
|
|
|
|
|
|
for bb in (0..self.basic_blocks.len()).map(BasicBlock::new) {
|
|
|
|
if self.pred_count[bb] == 0 {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
|
|
|
debug!("simplifying {:?}", bb);
|
|
|
|
|
|
|
|
let mut terminator = self.basic_blocks[bb].terminator.take()
|
|
|
|
.expect("invalid terminator state");
|
|
|
|
|
|
|
|
for successor in terminator.successors_mut() {
|
|
|
|
self.collapse_goto_chain(successor, &mut changed);
|
2015-11-14 16:52:17 -06:00
|
|
|
}
|
2016-06-08 16:10:15 -05:00
|
|
|
|
2017-04-19 17:20:50 -05:00
|
|
|
changed |= self.simplify_unwind(&mut terminator);
|
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
let mut new_stmts = vec![];
|
|
|
|
let mut inner_changed = true;
|
|
|
|
while inner_changed {
|
|
|
|
inner_changed = false;
|
|
|
|
inner_changed |= self.simplify_branch(&mut terminator);
|
|
|
|
inner_changed |= self.merge_successor(&mut new_stmts, &mut terminator);
|
|
|
|
changed |= inner_changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
self.basic_blocks[bb].statements.extend(new_stmts);
|
|
|
|
self.basic_blocks[bb].terminator = Some(terminator);
|
|
|
|
|
|
|
|
changed |= inner_changed;
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2016-06-08 16:10:15 -05:00
|
|
|
|
|
|
|
if !changed { break }
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2017-02-08 03:24:49 -06:00
|
|
|
|
|
|
|
self.strip_nops()
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
// Collapse a goto chain starting from `start`
|
|
|
|
fn collapse_goto_chain(&mut self, start: &mut BasicBlock, changed: &mut bool) {
|
|
|
|
let mut terminator = match self.basic_blocks[*start] {
|
|
|
|
BasicBlockData {
|
|
|
|
ref statements,
|
|
|
|
terminator: ref mut terminator @ Some(Terminator {
|
|
|
|
kind: TerminatorKind::Goto { .. }, ..
|
|
|
|
}), ..
|
|
|
|
} if statements.is_empty() => terminator.take(),
|
|
|
|
// if `terminator` is None, this means we are in a loop. In that
|
|
|
|
// case, let all the loop collapse to its entry.
|
|
|
|
_ => return
|
|
|
|
};
|
|
|
|
|
|
|
|
let target = match terminator {
|
|
|
|
Some(Terminator { kind: TerminatorKind::Goto { ref mut target }, .. }) => {
|
|
|
|
self.collapse_goto_chain(target, changed);
|
|
|
|
*target
|
|
|
|
}
|
|
|
|
_ => unreachable!()
|
|
|
|
};
|
|
|
|
self.basic_blocks[*start].terminator = terminator;
|
2015-11-14 16:52:17 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
debug!("collapsing goto chain from {:?} to {:?}", *start, target);
|
2015-11-10 14:38:36 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
*changed |= *start != target;
|
2016-10-04 08:27:27 -05:00
|
|
|
|
|
|
|
if self.pred_count[*start] == 1 {
|
|
|
|
// This is the last reference to *start, so the pred-count to
|
|
|
|
// to target is moved into the current block.
|
|
|
|
self.pred_count[*start] = 0;
|
|
|
|
} else {
|
|
|
|
self.pred_count[target] += 1;
|
|
|
|
self.pred_count[*start] -= 1;
|
|
|
|
}
|
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
*start = target;
|
|
|
|
}
|
2016-01-31 11:17:15 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
// merge a block with 1 `goto` predecessor to its parent
|
|
|
|
fn merge_successor(&mut self,
|
|
|
|
new_stmts: &mut Vec<Statement<'tcx>>,
|
|
|
|
terminator: &mut Terminator<'tcx>)
|
|
|
|
-> bool
|
|
|
|
{
|
|
|
|
let target = match terminator.kind {
|
|
|
|
TerminatorKind::Goto { target }
|
|
|
|
if self.pred_count[target] == 1
|
|
|
|
=> target,
|
|
|
|
_ => return false
|
|
|
|
};
|
|
|
|
|
|
|
|
debug!("merging block {:?} into {:?}", target, terminator);
|
|
|
|
*terminator = match self.basic_blocks[target].terminator.take() {
|
|
|
|
Some(terminator) => terminator,
|
|
|
|
None => {
|
|
|
|
// unreachable loop - this should not be possible, as we
|
|
|
|
// don't strand blocks, but handle it correctly.
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
};
|
|
|
|
new_stmts.extend(self.basic_blocks[target].statements.drain(..));
|
|
|
|
self.pred_count[target] = 0;
|
2016-01-31 11:17:15 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
true
|
|
|
|
}
|
2016-05-25 00:39:32 -05:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
// turn a branch with all successors identical to a goto
|
|
|
|
fn simplify_branch(&mut self, terminator: &mut Terminator<'tcx>) -> bool {
|
|
|
|
match terminator.kind {
|
|
|
|
TerminatorKind::SwitchInt { .. } => {},
|
|
|
|
_ => return false
|
|
|
|
};
|
|
|
|
|
|
|
|
let first_succ = {
|
|
|
|
let successors = terminator.successors();
|
|
|
|
if let Some(&first_succ) = terminator.successors().get(0) {
|
|
|
|
if successors.iter().all(|s| *s == first_succ) {
|
|
|
|
self.pred_count[first_succ] -= (successors.len()-1) as u32;
|
|
|
|
first_succ
|
|
|
|
} else {
|
|
|
|
return false
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2016-06-08 16:10:15 -05:00
|
|
|
} else {
|
|
|
|
return false
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2016-06-08 16:10:15 -05:00
|
|
|
};
|
2015-11-10 14:38:36 -06:00
|
|
|
|
2016-06-08 16:10:15 -05:00
|
|
|
debug!("simplifying branch {:?}", terminator);
|
|
|
|
terminator.kind = TerminatorKind::Goto { target: first_succ };
|
|
|
|
true
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2017-02-08 03:24:49 -06:00
|
|
|
|
2017-04-19 17:20:50 -05:00
|
|
|
// turn an unwind branch to a resume block into a None
|
|
|
|
fn simplify_unwind(&mut self, terminator: &mut Terminator<'tcx>) -> bool {
|
|
|
|
let unwind = match terminator.kind {
|
|
|
|
TerminatorKind::Drop { ref mut unwind, .. } |
|
|
|
|
TerminatorKind::DropAndReplace { ref mut unwind, .. } |
|
|
|
|
TerminatorKind::Call { cleanup: ref mut unwind, .. } |
|
|
|
|
TerminatorKind::Assert { cleanup: ref mut unwind, .. } =>
|
|
|
|
unwind,
|
|
|
|
_ => return false
|
|
|
|
};
|
|
|
|
|
|
|
|
if let &mut Some(unwind_block) = unwind {
|
|
|
|
let is_resume_block = match self.basic_blocks[unwind_block] {
|
|
|
|
BasicBlockData {
|
|
|
|
ref statements,
|
|
|
|
terminator: Some(Terminator {
|
|
|
|
kind: TerminatorKind::Resume, ..
|
|
|
|
}), ..
|
|
|
|
} if statements.is_empty() => true,
|
|
|
|
_ => false
|
|
|
|
};
|
|
|
|
if is_resume_block {
|
|
|
|
debug!("simplifying unwind to {:?} from {:?}",
|
|
|
|
unwind_block, terminator.source_info);
|
|
|
|
*unwind = None;
|
|
|
|
}
|
|
|
|
return is_resume_block;
|
|
|
|
}
|
|
|
|
|
|
|
|
false
|
|
|
|
}
|
|
|
|
|
2017-02-08 03:24:49 -06:00
|
|
|
fn strip_nops(&mut self) {
|
|
|
|
for blk in self.basic_blocks.iter_mut() {
|
|
|
|
blk.statements.retain(|stmt| if let StatementKind::Nop = stmt.kind {
|
|
|
|
false
|
|
|
|
} else {
|
|
|
|
true
|
|
|
|
})
|
|
|
|
}
|
|
|
|
}
|
2015-11-10 14:38:36 -06:00
|
|
|
}
|
2016-06-06 14:58:28 -05:00
|
|
|
|
2017-02-08 03:24:49 -06:00
|
|
|
pub fn remove_dead_blocks(mir: &mut Mir) {
|
2016-06-07 13:20:50 -05:00
|
|
|
let mut seen = BitVector::new(mir.basic_blocks().len());
|
2016-06-06 14:58:28 -05:00
|
|
|
for (bb, _) in traversal::preorder(mir) {
|
|
|
|
seen.insert(bb.index());
|
|
|
|
}
|
|
|
|
|
2016-06-07 13:20:50 -05:00
|
|
|
let basic_blocks = mir.basic_blocks_mut();
|
2016-06-06 14:58:28 -05:00
|
|
|
|
2016-06-07 13:20:50 -05:00
|
|
|
let num_blocks = basic_blocks.len();
|
2016-06-07 09:28:36 -05:00
|
|
|
let mut replacements : Vec<_> = (0..num_blocks).map(BasicBlock::new).collect();
|
2016-06-06 14:58:28 -05:00
|
|
|
let mut used_blocks = 0;
|
|
|
|
for alive_index in seen.iter() {
|
|
|
|
replacements[alive_index] = BasicBlock::new(used_blocks);
|
|
|
|
if alive_index != used_blocks {
|
|
|
|
// Swap the next alive block data with the current available slot. Since alive_index is
|
|
|
|
// non-decreasing this is a valid operation.
|
2016-06-07 13:20:50 -05:00
|
|
|
basic_blocks.raw.swap(alive_index, used_blocks);
|
2016-06-06 14:58:28 -05:00
|
|
|
}
|
|
|
|
used_blocks += 1;
|
|
|
|
}
|
2016-06-07 13:20:50 -05:00
|
|
|
basic_blocks.raw.truncate(used_blocks);
|
2016-06-06 14:58:28 -05:00
|
|
|
|
2016-06-07 13:20:50 -05:00
|
|
|
for block in basic_blocks {
|
|
|
|
for target in block.terminator_mut().successors_mut() {
|
2016-06-06 14:58:28 -05:00
|
|
|
*target = replacements[target.index()];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2016-08-15 20:59:50 -05:00
|
|
|
|
|
|
|
|
|
|
|
pub struct SimplifyLocals;
|
|
|
|
|
2017-04-25 17:23:33 -05:00
|
|
|
impl MirPass for SimplifyLocals {
|
|
|
|
fn run_pass<'a, 'tcx>(&self,
|
|
|
|
_: TyCtxt<'a, 'tcx, 'tcx>,
|
|
|
|
_: MirSource,
|
|
|
|
mir: &mut Mir<'tcx>) {
|
2016-08-15 20:59:50 -05:00
|
|
|
let mut marker = DeclMarker { locals: BitVector::new(mir.local_decls.len()) };
|
|
|
|
marker.visit_mir(mir);
|
|
|
|
// Return pointer and arguments are always live
|
|
|
|
marker.locals.insert(0);
|
|
|
|
for idx in mir.args_iter() {
|
|
|
|
marker.locals.insert(idx.index());
|
|
|
|
}
|
|
|
|
let map = make_local_map(&mut mir.local_decls, marker.locals);
|
|
|
|
// Update references to all vars and tmps now
|
|
|
|
LocalUpdater { map: map }.visit_mir(mir);
|
|
|
|
mir.local_decls.shrink_to_fit();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Construct the mapping while swapping out unused stuff out from the `vec`.
|
|
|
|
fn make_local_map<'tcx, I: Idx, V>(vec: &mut IndexVec<I, V>, mask: BitVector) -> Vec<usize> {
|
|
|
|
let mut map: Vec<usize> = ::std::iter::repeat(!0).take(vec.len()).collect();
|
|
|
|
let mut used = 0;
|
|
|
|
for alive_index in mask.iter() {
|
|
|
|
map[alive_index] = used;
|
|
|
|
if alive_index != used {
|
|
|
|
vec.swap(alive_index, used);
|
|
|
|
}
|
|
|
|
used += 1;
|
|
|
|
}
|
|
|
|
vec.truncate(used);
|
|
|
|
map
|
|
|
|
}
|
|
|
|
|
|
|
|
struct DeclMarker {
|
|
|
|
pub locals: BitVector,
|
|
|
|
}
|
|
|
|
|
|
|
|
impl<'tcx> Visitor<'tcx> for DeclMarker {
|
|
|
|
fn visit_lvalue(&mut self, lval: &Lvalue<'tcx>, ctx: LvalueContext<'tcx>, loc: Location) {
|
|
|
|
if ctx == LvalueContext::StorageLive || ctx == LvalueContext::StorageDead {
|
|
|
|
// ignore these altogether, they get removed along with their otherwise unused decls.
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if let Lvalue::Local(ref v) = *lval {
|
|
|
|
self.locals.insert(v.index());
|
|
|
|
}
|
|
|
|
self.super_lvalue(lval, ctx, loc);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
struct LocalUpdater {
|
|
|
|
map: Vec<usize>,
|
|
|
|
}
|
|
|
|
|
|
|
|
impl<'tcx> MutVisitor<'tcx> for LocalUpdater {
|
|
|
|
fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
|
|
|
|
// Remove unnecessary StorageLive and StorageDead annotations.
|
|
|
|
data.statements.retain(|stmt| {
|
|
|
|
match stmt.kind {
|
|
|
|
StatementKind::StorageLive(ref lval) | StatementKind::StorageDead(ref lval) => {
|
|
|
|
match *lval {
|
|
|
|
Lvalue::Local(l) => self.map[l.index()] != !0,
|
|
|
|
_ => true
|
|
|
|
}
|
|
|
|
}
|
|
|
|
_ => true
|
|
|
|
}
|
|
|
|
});
|
|
|
|
self.super_basic_block_data(block, data);
|
|
|
|
}
|
|
|
|
fn visit_lvalue(&mut self, lval: &mut Lvalue<'tcx>, ctx: LvalueContext<'tcx>, loc: Location) {
|
|
|
|
match *lval {
|
|
|
|
Lvalue::Local(ref mut l) => *l = Local::new(self.map[l.index()]),
|
|
|
|
_ => (),
|
|
|
|
};
|
|
|
|
self.super_lvalue(lval, ctx, loc);
|
|
|
|
}
|
|
|
|
}
|