379 lines
15 KiB
Rust
Raw Normal View History

// Copyright 2012-2014 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.
//! An analysis to determine which locals require allocas and
//! which do not.
Merge indexed_set.rs into bitvec.rs, and rename it bit_set.rs. Currently we have two files implementing bitsets (and 2D bit matrices). This commit combines them into one, taking the best features from each. This involves renaming a lot of things. The high level changes are as follows. - bitvec.rs --> bit_set.rs - indexed_set.rs --> (removed) - BitArray + IdxSet --> BitSet (merged, see below) - BitVector --> GrowableBitSet - {,Sparse,Hybrid}IdxSet --> {,Sparse,Hybrid}BitSet - BitMatrix --> BitMatrix - SparseBitMatrix --> SparseBitMatrix The changes within the bitset types themselves are as follows. ``` OLD OLD NEW BitArray<C> IdxSet<T> BitSet<T> -------- ------ ------ grow - grow new - (remove) new_empty new_empty new_empty new_filled new_filled new_filled - to_hybrid to_hybrid clear clear clear set_up_to set_up_to set_up_to clear_above - clear_above count - count contains(T) contains(&T) contains(T) contains_all - superset is_empty - is_empty insert(T) add(&T) insert(T) insert_all - insert_all() remove(T) remove(&T) remove(T) words words words words_mut words_mut words_mut - overwrite overwrite merge union union - subtract subtract - intersect intersect iter iter iter ``` In general, when choosing names I went with: - names that are more obvious (e.g. `BitSet` over `IdxSet`). - names that are more like the Rust libraries (e.g. `T` over `C`, `insert` over `add`); - names that are more set-like (e.g. `union` over `merge`, `superset` over `contains_all`, `domain_size` over `num_bits`). Also, using `T` for index arguments seems more sensible than `&T` -- even though the latter is standard in Rust collection types -- because indices are always copyable. It also results in fewer `&` and `*` sigils in practice.
2018-09-14 15:07:25 +10:00
use rustc_data_structures::bit_set::BitSet;
2018-07-02 06:14:49 -04:00
use rustc_data_structures::graph::dominators::Dominators;
use rustc_data_structures::indexed_vec::{Idx, IndexVec};
use rustc::mir::{self, Location, TerminatorKind};
use rustc::mir::visit::{Visitor, PlaceContext, MutatingUseContext, NonMutatingUseContext};
use rustc::mir::traversal;
use rustc::ty;
use rustc::ty::layout::LayoutOf;
use type_of::LayoutLlvmExt;
use super::FunctionCx;
Merge indexed_set.rs into bitvec.rs, and rename it bit_set.rs. Currently we have two files implementing bitsets (and 2D bit matrices). This commit combines them into one, taking the best features from each. This involves renaming a lot of things. The high level changes are as follows. - bitvec.rs --> bit_set.rs - indexed_set.rs --> (removed) - BitArray + IdxSet --> BitSet (merged, see below) - BitVector --> GrowableBitSet - {,Sparse,Hybrid}IdxSet --> {,Sparse,Hybrid}BitSet - BitMatrix --> BitMatrix - SparseBitMatrix --> SparseBitMatrix The changes within the bitset types themselves are as follows. ``` OLD OLD NEW BitArray<C> IdxSet<T> BitSet<T> -------- ------ ------ grow - grow new - (remove) new_empty new_empty new_empty new_filled new_filled new_filled - to_hybrid to_hybrid clear clear clear set_up_to set_up_to set_up_to clear_above - clear_above count - count contains(T) contains(&T) contains(T) contains_all - superset is_empty - is_empty insert(T) add(&T) insert(T) insert_all - insert_all() remove(T) remove(&T) remove(T) words words words words_mut words_mut words_mut - overwrite overwrite merge union union - subtract subtract - intersect intersect iter iter iter ``` In general, when choosing names I went with: - names that are more obvious (e.g. `BitSet` over `IdxSet`). - names that are more like the Rust libraries (e.g. `T` over `C`, `insert` over `add`); - names that are more set-like (e.g. `union` over `merge`, `superset` over `contains_all`, `domain_size` over `num_bits`). Also, using `T` for index arguments seems more sensible than `&T` -- even though the latter is standard in Rust collection types -- because indices are always copyable. It also results in fewer `&` and `*` sigils in practice.
2018-09-14 15:07:25 +10:00
pub fn non_ssa_locals(fx: &FunctionCx<'a, 'll, 'tcx>) -> BitSet<mir::Local> {
let mir = fx.mir;
let mut analyzer = LocalAnalyzer::new(fx);
analyzer.visit_mir(mir);
for (index, ty) in mir.local_decls.iter().map(|l| l.ty).enumerate() {
let ty = fx.monomorphize(&ty);
debug!("local {} has type {:?}", index, ty);
let layout = fx.cx.layout_of(ty);
if layout.is_llvm_immediate() {
// These sorts of types are immediates that we can store
// in an Value without an alloca.
} else if layout.is_llvm_scalar_pair() {
// We allow pairs and uses of any of their 2 fields.
} else {
// These sorts of types require an alloca. Note that
// is_llvm_immediate() may *still* be true, particularly
// for newtypes, but we currently force some types
// (e.g. structs) into an alloca unconditionally, just so
// that we don't have to deal with having two pathways
2015-11-03 15:50:04 -05:00
// (gep vs extractvalue etc).
analyzer.not_ssa(mir::Local::new(index));
}
}
analyzer.non_ssa_locals
}
struct LocalAnalyzer<'mir, 'a: 'mir, 'll: 'a, 'tcx: 'll> {
fx: &'mir FunctionCx<'a, 'll, 'tcx>,
dominators: Dominators<mir::BasicBlock>,
Merge indexed_set.rs into bitvec.rs, and rename it bit_set.rs. Currently we have two files implementing bitsets (and 2D bit matrices). This commit combines them into one, taking the best features from each. This involves renaming a lot of things. The high level changes are as follows. - bitvec.rs --> bit_set.rs - indexed_set.rs --> (removed) - BitArray + IdxSet --> BitSet (merged, see below) - BitVector --> GrowableBitSet - {,Sparse,Hybrid}IdxSet --> {,Sparse,Hybrid}BitSet - BitMatrix --> BitMatrix - SparseBitMatrix --> SparseBitMatrix The changes within the bitset types themselves are as follows. ``` OLD OLD NEW BitArray<C> IdxSet<T> BitSet<T> -------- ------ ------ grow - grow new - (remove) new_empty new_empty new_empty new_filled new_filled new_filled - to_hybrid to_hybrid clear clear clear set_up_to set_up_to set_up_to clear_above - clear_above count - count contains(T) contains(&T) contains(T) contains_all - superset is_empty - is_empty insert(T) add(&T) insert(T) insert_all - insert_all() remove(T) remove(&T) remove(T) words words words words_mut words_mut words_mut - overwrite overwrite merge union union - subtract subtract - intersect intersect iter iter iter ``` In general, when choosing names I went with: - names that are more obvious (e.g. `BitSet` over `IdxSet`). - names that are more like the Rust libraries (e.g. `T` over `C`, `insert` over `add`); - names that are more set-like (e.g. `union` over `merge`, `superset` over `contains_all`, `domain_size` over `num_bits`). Also, using `T` for index arguments seems more sensible than `&T` -- even though the latter is standard in Rust collection types -- because indices are always copyable. It also results in fewer `&` and `*` sigils in practice.
2018-09-14 15:07:25 +10:00
non_ssa_locals: BitSet<mir::Local>,
// The location of the first visited direct assignment to each
// local, or an invalid location (out of bounds `block` index).
first_assignment: IndexVec<mir::Local, Location>
}
impl LocalAnalyzer<'mir, 'a, 'll, 'tcx> {
fn new(fx: &'mir FunctionCx<'a, 'll, 'tcx>) -> Self {
let invalid_location =
mir::BasicBlock::new(fx.mir.basic_blocks().len()).start_location();
let mut analyzer = LocalAnalyzer {
fx,
dominators: fx.mir.dominators(),
Merge indexed_set.rs into bitvec.rs, and rename it bit_set.rs. Currently we have two files implementing bitsets (and 2D bit matrices). This commit combines them into one, taking the best features from each. This involves renaming a lot of things. The high level changes are as follows. - bitvec.rs --> bit_set.rs - indexed_set.rs --> (removed) - BitArray + IdxSet --> BitSet (merged, see below) - BitVector --> GrowableBitSet - {,Sparse,Hybrid}IdxSet --> {,Sparse,Hybrid}BitSet - BitMatrix --> BitMatrix - SparseBitMatrix --> SparseBitMatrix The changes within the bitset types themselves are as follows. ``` OLD OLD NEW BitArray<C> IdxSet<T> BitSet<T> -------- ------ ------ grow - grow new - (remove) new_empty new_empty new_empty new_filled new_filled new_filled - to_hybrid to_hybrid clear clear clear set_up_to set_up_to set_up_to clear_above - clear_above count - count contains(T) contains(&T) contains(T) contains_all - superset is_empty - is_empty insert(T) add(&T) insert(T) insert_all - insert_all() remove(T) remove(&T) remove(T) words words words words_mut words_mut words_mut - overwrite overwrite merge union union - subtract subtract - intersect intersect iter iter iter ``` In general, when choosing names I went with: - names that are more obvious (e.g. `BitSet` over `IdxSet`). - names that are more like the Rust libraries (e.g. `T` over `C`, `insert` over `add`); - names that are more set-like (e.g. `union` over `merge`, `superset` over `contains_all`, `domain_size` over `num_bits`). Also, using `T` for index arguments seems more sensible than `&T` -- even though the latter is standard in Rust collection types -- because indices are always copyable. It also results in fewer `&` and `*` sigils in practice.
2018-09-14 15:07:25 +10:00
non_ssa_locals: BitSet::new_empty(fx.mir.local_decls.len()),
first_assignment: IndexVec::from_elem(invalid_location, &fx.mir.local_decls)
};
// Arguments get assigned to by means of the function being called
for arg in fx.mir.args_iter() {
analyzer.first_assignment[arg] = mir::START_BLOCK.start_location();
}
analyzer
}
fn first_assignment(&self, local: mir::Local) -> Option<Location> {
let location = self.first_assignment[local];
if location.block.index() < self.fx.mir.basic_blocks().len() {
Some(location)
} else {
None
}
}
fn not_ssa(&mut self, local: mir::Local) {
debug!("marking {:?} as non-SSA", local);
self.non_ssa_locals.insert(local);
}
fn assign(&mut self, local: mir::Local, location: Location) {
if self.first_assignment(local).is_some() {
self.not_ssa(local);
} else {
self.first_assignment[local] = location;
}
}
}
impl Visitor<'tcx> for LocalAnalyzer<'mir, 'a, 'll, 'tcx> {
fn visit_assign(&mut self,
block: mir::BasicBlock,
place: &mir::Place<'tcx>,
rvalue: &mir::Rvalue<'tcx>,
location: Location) {
debug!("visit_assign(block={:?}, place={:?}, rvalue={:?})", block, place, rvalue);
if let mir::Place::Local(index) = *place {
self.assign(index, location);
if !self.fx.rvalue_creates_operand(rvalue) {
self.not_ssa(index);
}
} else {
self.visit_place(
place,
PlaceContext::MutatingUse(MutatingUseContext::Store),
location
);
}
self.visit_rvalue(rvalue, location);
}
fn visit_terminator_kind(&mut self,
block: mir::BasicBlock,
kind: &mir::TerminatorKind<'tcx>,
location: Location) {
2018-01-16 09:24:38 +01:00
let check = match *kind {
mir::TerminatorKind::Call {
func: mir::Operand::Constant(ref c),
ref args, ..
} => match c.ty.sty {
ty::FnDef(did, _) => Some((did, args)),
2018-01-16 09:24:38 +01:00
_ => None,
},
2018-01-16 09:24:38 +01:00
_ => None,
};
if let Some((def_id, args)) = check {
if Some(def_id) == self.fx.cx.tcx.lang_items().box_free_fn() {
// box_free(x) shares with `drop x` the property that it
// is not guaranteed to be statically dominated by the
// definition of x, so x must always be in an alloca.
if let mir::Operand::Move(ref place) = args[0] {
self.visit_place(
place,
PlaceContext::MutatingUse(MutatingUseContext::Drop),
location
);
}
}
}
self.super_terminator_kind(block, kind, location);
}
fn visit_place(&mut self,
place: &mir::Place<'tcx>,
context: PlaceContext<'tcx>,
location: Location) {
debug!("visit_place(place={:?}, context={:?})", place, context);
let cx = self.fx.cx;
if let mir::Place::Projection(ref proj) = *place {
// Allow uses of projections that are ZSTs or from scalar fields.
let is_consume = match context {
PlaceContext::NonMutatingUse(NonMutatingUseContext::Copy) |
PlaceContext::NonMutatingUse(NonMutatingUseContext::Move) => true,
_ => false
};
if is_consume {
let base_ty = proj.base.ty(self.fx.mir, cx.tcx);
let base_ty = self.fx.monomorphize(&base_ty);
// ZSTs don't require any actual memory access.
let elem_ty = base_ty
.projection_ty(cx.tcx, &proj.elem)
.to_ty(cx.tcx);
let elem_ty = self.fx.monomorphize(&elem_ty);
2018-01-05 07:04:08 +02:00
if cx.layout_of(elem_ty).is_zst() {
return;
}
if let mir::ProjectionElem::Field(..) = proj.elem {
2018-01-05 07:04:08 +02:00
let layout = cx.layout_of(base_ty.to_ty(cx.tcx));
if layout.is_llvm_immediate() || layout.is_llvm_scalar_pair() {
// Recurse with the same context, instead of `Projection`,
// potentially stopping at non-operand projections,
// which would trigger `not_ssa` on locals.
self.visit_place(&proj.base, context, location);
return;
}
}
}
// A deref projection only reads the pointer, never needs the place.
if let mir::ProjectionElem::Deref = proj.elem {
return self.visit_place(
&proj.base,
PlaceContext::NonMutatingUse(NonMutatingUseContext::Copy),
location
);
}
}
self.super_place(place, context, location);
}
fn visit_local(&mut self,
&local: &mir::Local,
2017-12-01 14:31:47 +02:00
context: PlaceContext<'tcx>,
location: Location) {
match context {
PlaceContext::MutatingUse(MutatingUseContext::Call) => {
self.assign(local, location);
}
PlaceContext::NonUse(_) => {}
PlaceContext::NonMutatingUse(NonMutatingUseContext::Copy) |
PlaceContext::NonMutatingUse(NonMutatingUseContext::Move) => {
// Reads from uninitialized variables (e.g. in dead code, after
// optimizations) require locals to be in (uninitialized) memory.
// NB: there can be uninitialized reads of a local visited after
// an assignment to that local, if they happen on disjoint paths.
let ssa_read = match self.first_assignment(local) {
Some(assignment_location) => {
assignment_location.dominates(location, &self.dominators)
}
None => false
};
if !ssa_read {
self.not_ssa(local);
}
}
PlaceContext::NonMutatingUse(NonMutatingUseContext::Inspect) |
PlaceContext::MutatingUse(MutatingUseContext::Store) |
PlaceContext::MutatingUse(MutatingUseContext::AsmOutput) |
PlaceContext::MutatingUse(MutatingUseContext::Borrow(..)) |
PlaceContext::MutatingUse(MutatingUseContext::Projection) |
PlaceContext::NonMutatingUse(NonMutatingUseContext::SharedBorrow(..)) |
PlaceContext::NonMutatingUse(NonMutatingUseContext::UniqueBorrow(..)) |
PlaceContext::NonMutatingUse(NonMutatingUseContext::ShallowBorrow(..)) |
PlaceContext::NonMutatingUse(NonMutatingUseContext::Projection) => {
self.not_ssa(local);
}
PlaceContext::MutatingUse(MutatingUseContext::Drop) => {
let ty = mir::Place::Local(local).ty(self.fx.mir, self.fx.cx.tcx);
let ty = self.fx.monomorphize(&ty.to_ty(self.fx.cx.tcx));
// Only need the place if we're actually dropping it.
if self.fx.cx.type_needs_drop(ty) {
self.not_ssa(local);
}
}
}
}
}
#[derive(Copy, Clone, Debug, PartialEq, Eq)]
pub enum CleanupKind {
NotCleanup,
Funclet,
Internal { funclet: mir::BasicBlock }
}
impl CleanupKind {
pub fn funclet_bb(self, for_bb: mir::BasicBlock) -> Option<mir::BasicBlock> {
match self {
CleanupKind::NotCleanup => None,
CleanupKind::Funclet => Some(for_bb),
CleanupKind::Internal { funclet } => Some(funclet),
}
}
}
2016-12-17 19:54:32 -07:00
pub fn cleanup_kinds<'a, 'tcx>(mir: &mir::Mir<'tcx>) -> IndexVec<mir::BasicBlock, CleanupKind> {
fn discover_masters<'tcx>(result: &mut IndexVec<mir::BasicBlock, CleanupKind>,
mir: &mir::Mir<'tcx>) {
2016-06-07 21:20:50 +03:00
for (bb, data) in mir.basic_blocks().iter_enumerated() {
match data.terminator().kind {
TerminatorKind::Goto { .. } |
TerminatorKind::Resume |
TerminatorKind::Abort |
TerminatorKind::Return |
2016-12-26 14:34:03 +01:00
TerminatorKind::GeneratorDrop |
TerminatorKind::Unreachable |
2016-12-26 14:34:03 +01:00
TerminatorKind::SwitchInt { .. } |
TerminatorKind::Yield { .. } |
TerminatorKind::FalseEdges { .. } |
TerminatorKind::FalseUnwind { .. } => {
/* nothing to do */
}
TerminatorKind::Call { cleanup: unwind, .. } |
TerminatorKind::Assert { cleanup: unwind, .. } |
TerminatorKind::DropAndReplace { unwind, .. } |
TerminatorKind::Drop { unwind, .. } => {
if let Some(unwind) = unwind {
debug!("cleanup_kinds: {:?}/{:?} registering {:?} as funclet",
bb, data, unwind);
result[unwind] = CleanupKind::Funclet;
}
}
}
}
}
fn propagate<'tcx>(result: &mut IndexVec<mir::BasicBlock, CleanupKind>,
mir: &mir::Mir<'tcx>) {
2016-06-07 21:20:50 +03:00
let mut funclet_succs = IndexVec::from_elem(None, mir.basic_blocks());
let mut set_successor = |funclet: mir::BasicBlock, succ| {
match funclet_succs[funclet] {
ref mut s @ None => {
debug!("set_successor: updating successor of {:?} to {:?}",
funclet, succ);
*s = Some(succ);
},
Some(s) => if s != succ {
span_bug!(mir.span, "funclet {:?} has 2 parents - {:?} and {:?}",
funclet, s, succ);
}
}
};
for (bb, data) in traversal::reverse_postorder(mir) {
let funclet = match result[bb] {
CleanupKind::NotCleanup => continue,
CleanupKind::Funclet => bb,
CleanupKind::Internal { funclet } => funclet,
};
debug!("cleanup_kinds: {:?}/{:?}/{:?} propagating funclet {:?}",
bb, data, result[bb], funclet);
for &succ in data.terminator().successors() {
let kind = result[succ];
debug!("cleanup_kinds: propagating {:?} to {:?}/{:?}",
funclet, succ, kind);
match kind {
CleanupKind::NotCleanup => {
result[succ] = CleanupKind::Internal { funclet: funclet };
}
CleanupKind::Funclet => {
if funclet != succ {
set_successor(funclet, succ);
}
}
CleanupKind::Internal { funclet: succ_funclet } => {
if funclet != succ_funclet {
// `succ` has 2 different funclet going into it, so it must
// be a funclet by itself.
debug!("promoting {:?} to a funclet and updating {:?}", succ,
succ_funclet);
result[succ] = CleanupKind::Funclet;
set_successor(succ_funclet, succ);
set_successor(funclet, succ);
}
}
}
}
}
}
2016-06-07 21:20:50 +03:00
let mut result = IndexVec::from_elem(CleanupKind::NotCleanup, mir.basic_blocks());
discover_masters(&mut result, mir);
propagate(&mut result, mir);
debug!("cleanup_kinds: result={:?}", result);
result
}