Skip to content

Commit

Permalink
Auto merge of #85646 - Moxinilian:separate-const-switch, r=cjgillot
Browse files Browse the repository at this point in the history 
MIR opt: separate constant predecessors of a switch

For each block S ending with a switch, this pass copies S for each of S's predecessors that seem to assign the value being switched over as a const. This is done using a somewhat simple heuristic to determine what seems to be a const transitively.

More precisely, this is what the pass does:
- find a block that ends in a switch
- track if there is an unique place set before the current basic block that determines the result of the switch (this is the part that resolves switching over discriminants)
- if there is, iterate over the parents that have a reasonable terminator and find if the found determining place is likely to be (transitively) set from a const within that parent block
- if so, add the corresponding edge to a vector of edges to duplicate
- once this is done, iterate over the found edges: copy the target block and replace the reference to the target block in the origin block with the new block

This pass is not optimal and could probably duplicate in more cases, but the intention was mostly to address cases like in #85133 or #85365, to avoid creating new enums that get destroyed immediately afterwards (notably making the new try v2 `?` desugar zero-cost).

A benefit of this pass working the way it does is that it is easy to ensure its correctness: the worst that can happen is for it to needlessly copy a basic block, which is likely to be destroyed by cleanup passes afterwards. The complex parts where aliasing matters are only heuristics and the hard work is left to further passes like ConstProp.

# LLVM blocker

Unfortunately, I believe it would be unwise to enable this optimization by default for now. Indeed, currently switch lowering passes like SimplifyCFG in LLVM lose the information on the set of possible variant values, which means it tends to actually generate worse code with this optimization enabled. A fix would have to be done in LLVM itself. This is something I also want to look into. I have opened [a bug report at the LLVM bug tracker](https://bugs.llvm.org/show_bug.cgi?id=50455).

When this is done, I hope we can enable this pass by default. It should be fairly fast and I think it is beneficial in many cases. Notably, it should be a sound alternative to simplify-arm-identity. By the way, ConstProp only seems to pick up the optimization in functions that are not generic. This is however most likely an issue in ConstProp that I will look into afterwards.

This is my first contribution to rustc, and I would like to thank everyone on the Zulip mir-opt chat for the help and support, and especially `@scottmcm` for the guidance.
  • Loading branch information
@bors
bors committed Jul 25, 2021
2 parents 6489ee1 + a77e2ad commit 70f7471
Show file tree
Hide file tree
Showing 9 changed files with 1,055 additions and 0 deletions.
2 changes: 2 additions & 0 deletions compiler/rustc_mir/src/transform/mod.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,7 @@ pub mod remove_unneeded_drops;
pub mod remove_zsts;
pub mod required_consts;
pub mod rustc_peek;
pub mod separate_const_switch;
pub mod simplify;
pub mod simplify_branches;
pub mod simplify_comparison_integral;
Expand Down Expand Up @@ -501,6 +502,7 @@ fn run_optimization_passes<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
// inst combine is after MatchBranchSimplification to clean up Ne(_1, false)
&multiple_return_terminators::MultipleReturnTerminators,
&instcombine::InstCombine,
&separate_const_switch::SeparateConstSwitch,
&const_prop::ConstProp,
&simplify_branches::SimplifyBranches::new("after-const-prop"),
&early_otherwise_branch::EarlyOtherwiseBranch,
Expand Down
343 changes: 343 additions & 0 deletions compiler/rustc_mir/src/transform/separate_const_switch.rs
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,343 @@
//! A pass that duplicates switch-terminated blocks
//! into a new copy for each predecessor, provided
//! the predecessor sets the value being switched
//! over to a constant.
//!
//! The purpose of this pass is to help constant
//! propagation passes to simplify the switch terminator
//! of the copied blocks into gotos when some predecessors
//! statically determine the output of switches.
//!
//! ```text
//! x = 12 --- ---> something
//! \ / 12
//! --> switch x
//! / \ otherwise
//! x = y --- ---> something else
//! ```
//! becomes
//! ```text
//! x = 12 ---> switch x ------> something
//! \ / 12
//! X
//! / \ otherwise
//! x = y ---> switch x ------> something else
//! ```
//! so it can hopefully later be turned by another pass into
//! ```text
//! x = 12 --------------------> something
//! / 12
//! /
//! / otherwise
//! x = y ---- switch x ------> something else
//! ```
//!
//! This optimization is meant to cover simple cases
//! like `?` desugaring. For now, it thus focuses on
//! simplicity rather than completeness (it notably
//! sometimes duplicates abusively).

use crate::transform::MirPass;
use rustc_middle::mir::*;
use rustc_middle::ty::TyCtxt;
use smallvec::SmallVec;

pub struct SeparateConstSwitch;

impl<'tcx> MirPass<'tcx> for SeparateConstSwitch {
fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
if tcx.sess.mir_opt_level() < 4 {
return;
}

// If execution did something, applying a simplification layer
// helps later passes optimize the copy away.
if separate_const_switch(body) > 0 {
super::simplify::simplify_cfg(tcx, body);
}
}
}

/// Returns the amount of blocks that were duplicated
pub fn separate_const_switch<'tcx>(body: &mut Body<'tcx>) -> usize {
let mut new_blocks: SmallVec<[(BasicBlock, BasicBlock); 6]> = SmallVec::new();
let predecessors = body.predecessors();
'block_iter: for (block_id, block) in body.basic_blocks().iter_enumerated() {
if let TerminatorKind::SwitchInt {
discr: Operand::Copy(switch_place) | Operand::Move(switch_place),
..
} = block.terminator().kind
{
// If the block is on an unwind path, do not
// apply the optimization as unwind paths
// rely on a unique parent invariant
if block.is_cleanup {
continue 'block_iter;
}

// If the block has fewer than 2 predecessors, ignore it
// we could maybe chain blocks that have exactly one
// predecessor, but for now we ignore that
if predecessors[block_id].len() < 2 {
continue 'block_iter;
}

// First, let's find a non-const place
// that determines the result of the switch
if let Some(switch_place) = find_determining_place(switch_place, block) {
// We now have an input place for which it would
// be interesting if predecessors assigned it from a const

let mut predecessors_left = predecessors[block_id].len();
'predec_iter: for predecessor_id in predecessors[block_id].iter().copied() {
let predecessor = &body.basic_blocks()[predecessor_id];

// First we make sure the predecessor jumps
// in a reasonable way
match &predecessor.terminator().kind {
// The following terminators are
// unconditionally valid
TerminatorKind::Goto { .. } | TerminatorKind::SwitchInt { .. } => {}

TerminatorKind::FalseEdge { real_target, .. } => {
if *real_target != block_id {
continue 'predec_iter;
}
}

// The following terminators are not allowed
TerminatorKind::Resume
| TerminatorKind::Drop { .. }
| TerminatorKind::DropAndReplace { .. }
| TerminatorKind::Call { .. }
| TerminatorKind::Assert { .. }
| TerminatorKind::FalseUnwind { .. }
| TerminatorKind::Yield { .. }
| TerminatorKind::Abort
| TerminatorKind::Return
| TerminatorKind::Unreachable
| TerminatorKind::InlineAsm { .. }
| TerminatorKind::GeneratorDrop => {
continue 'predec_iter;
}
}

if is_likely_const(switch_place, predecessor) {
new_blocks.push((predecessor_id, block_id));
predecessors_left -= 1;
if predecessors_left < 2 {
// If the original block only has one predecessor left,
// we have nothing left to do
break 'predec_iter;
}
}
}
}
}
}

// Once the analysis is done, perform the duplication
let body_span = body.span;
let copied_blocks = new_blocks.len();
let blocks = body.basic_blocks_mut();
for (pred_id, target_id) in new_blocks {
let new_block = blocks[target_id].clone();
let new_block_id = blocks.push(new_block);
let terminator = blocks[pred_id].terminator_mut();

match terminator.kind {
TerminatorKind::Goto { ref mut target } => {
*target = new_block_id;
}

TerminatorKind::FalseEdge { ref mut real_target, .. } => {
if *real_target == target_id {
*real_target = new_block_id;
}
}

TerminatorKind::SwitchInt { ref mut targets, .. } => {
targets.all_targets_mut().iter_mut().for_each(|x| {
if *x == target_id {
*x = new_block_id;
}
});
}

TerminatorKind::Resume
| TerminatorKind::Abort
| TerminatorKind::Return
| TerminatorKind::Unreachable
| TerminatorKind::GeneratorDrop
| TerminatorKind::Assert { .. }
| TerminatorKind::DropAndReplace { .. }
| TerminatorKind::FalseUnwind { .. }
| TerminatorKind::Drop { .. }
| TerminatorKind::Call { .. }
| TerminatorKind::InlineAsm { .. }
| TerminatorKind::Yield { .. } => {
span_bug!(
body_span,
"basic block terminator had unexpected kind {:?}",
&terminator.kind
)
}
}
}

copied_blocks
}

/// This function describes a rough heuristic guessing
/// whether a place is last set with a const within the block.
/// Notably, it will be overly pessimistic in cases that are already
/// not handled by `separate_const_switch`.
fn is_likely_const<'tcx>(mut tracked_place: Place<'tcx>, block: &BasicBlockData<'tcx>) -> bool {
for statement in block.statements.iter().rev() {
match &statement.kind {
StatementKind::Assign(assign) => {
if assign.0 == tracked_place {
match assign.1 {
// These rvalues are definitely constant
Rvalue::Use(Operand::Constant(_))
| Rvalue::Ref(_, _, _)
| Rvalue::AddressOf(_, _)
| Rvalue::Cast(_, Operand::Constant(_), _)
| Rvalue::NullaryOp(_, _)
| Rvalue::UnaryOp(_, Operand::Constant(_)) => return true,

// These rvalues make things ambiguous
Rvalue::Repeat(_, _)
| Rvalue::ThreadLocalRef(_)
| Rvalue::Len(_)
| Rvalue::BinaryOp(_, _)
| Rvalue::CheckedBinaryOp(_, _)
| Rvalue::Aggregate(_, _) => return false,

// These rvalues move the place to track
Rvalue::Cast(_, Operand::Copy(place) | Operand::Move(place), _)
| Rvalue::Use(Operand::Copy(place) | Operand::Move(place))
| Rvalue::UnaryOp(_, Operand::Copy(place) | Operand::Move(place))
| Rvalue::Discriminant(place) => tracked_place = place,
}
}
}

// If the discriminant is set, it is always set
// as a constant, so the job is done.
// As we are **ignoring projections**, if the place
// we are tracking sees its discriminant be set,
// that means we had to be tracking the discriminant
// specifically (as it is impossible to switch over
// an enum directly, and if we were switching over
// its content, we would have had to at least cast it to
// some variant first)
StatementKind::SetDiscriminant { place, .. } => {
if **place == tracked_place {
return true;
}
}

// If inline assembly is found, we probably should
// not try to analyze the code
StatementKind::LlvmInlineAsm(_) => return false,

// These statements have no influence on the place
// we are interested in
StatementKind::FakeRead(_)
| StatementKind::StorageLive(_)
| StatementKind::Retag(_, _)
| StatementKind::AscribeUserType(_, _)
| StatementKind::Coverage(_)
| StatementKind::StorageDead(_)
| StatementKind::CopyNonOverlapping(_)
| StatementKind::Nop => {}
}
}

// If no good reason for the place to be const is found,
// give up. We could maybe go up predecessors, but in
// most cases giving up now should be sufficient.
false
}

/// Finds a unique place that entirely determines the value
/// of `switch_place`, if it exists. This is only a heuristic.
/// Ideally we would like to track multiple determining places
/// for some edge cases, but one is enough for a lot of situations.
fn find_determining_place<'tcx>(
mut switch_place: Place<'tcx>,
block: &BasicBlockData<'tcx>,
) -> Option<Place<'tcx>> {
for statement in block.statements.iter().rev() {
match &statement.kind {
StatementKind::Assign(op) => {
if op.0 != switch_place {
continue;
}

match op.1 {
// The following rvalues move the place
// that may be const in the predecessor
Rvalue::Use(Operand::Move(new) | Operand::Copy(new))
| Rvalue::UnaryOp(_, Operand::Copy(new) | Operand::Move(new))
| Rvalue::Cast(_, Operand::Move(new) | Operand::Copy(new), _)
| Rvalue::Repeat(Operand::Move(new) | Operand::Copy(new), _)
| Rvalue::Discriminant(new)
=> switch_place = new,

// The following rvalues might still make the block
// be valid but for now we reject them
Rvalue::Len(_)
| Rvalue::Ref(_, _, _)
| Rvalue::BinaryOp(_, _)
| Rvalue::CheckedBinaryOp(_, _)
| Rvalue::Aggregate(_, _)

// The following rvalues definitely mean we cannot
// or should not apply this optimization
| Rvalue::Use(Operand::Constant(_))
| Rvalue::Repeat(Operand::Constant(_), _)
| Rvalue::ThreadLocalRef(_)
| Rvalue::AddressOf(_, _)
| Rvalue::NullaryOp(_, _)
| Rvalue::UnaryOp(_, Operand::Constant(_))
| Rvalue::Cast(_, Operand::Constant(_), _)
=> return None,
}
}

// These statements have no influence on the place
// we are interested in
StatementKind::FakeRead(_)
| StatementKind::StorageLive(_)
| StatementKind::StorageDead(_)
| StatementKind::Retag(_, _)
| StatementKind::AscribeUserType(_, _)
| StatementKind::Coverage(_)
| StatementKind::CopyNonOverlapping(_)
| StatementKind::Nop => {}

// If inline assembly is found, we probably should
// not try to analyze the code
StatementKind::LlvmInlineAsm(_) => return None,

// If the discriminant is set, it is always set
// as a constant, so the job is already done.
// As we are **ignoring projections**, if the place
// we are tracking sees its discriminant be set,
// that means we had to be tracking the discriminant
// specifically (as it is impossible to switch over
// an enum directly, and if we were switching over
// its content, we would have had to at least cast it to
// some variant first)
StatementKind::SetDiscriminant { place, .. } => {
if **place == switch_place {
return None;
}
}
}
}

Some(switch_place)
}
Loading

0 comments on commit 70f7471

Please sign in to comment.