1 // Copyright 2016 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 //! Inlining pass for MIR functions
14 use rustc::hir::TransFnAttrFlags;
15 use rustc::hir::def_id::DefId;
17 use rustc_data_structures::bitvec::BitVector;
18 use rustc_data_structures::indexed_vec::{Idx, IndexVec};
21 use rustc::mir::visit::*;
22 use rustc::ty::{self, Instance, Ty, TyCtxt};
23 use rustc::ty::subst::{Subst,Substs};
25 use std::collections::VecDeque;
27 use transform::{MirPass, MirSource};
28 use super::simplify::{remove_dead_blocks, CfgSimplifier};
31 use rustc_target::spec::abi::Abi;
33 const DEFAULT_THRESHOLD: usize = 50;
34 const HINT_THRESHOLD: usize = 100;
36 const INSTR_COST: usize = 5;
37 const CALL_PENALTY: usize = 25;
39 const UNKNOWN_SIZE_COST: usize = 10;
43 #[derive(Copy, Clone, Debug)]
44 struct CallSite<'tcx> {
46 substs: &'tcx Substs<'tcx>,
51 impl MirPass for Inline {
52 fn run_pass<'a, 'tcx>(&self,
53 tcx: TyCtxt<'a, 'tcx, 'tcx>,
55 mir: &mut Mir<'tcx>) {
56 if tcx.sess.opts.debugging_opts.mir_opt_level >= 2 {
57 Inliner { tcx, source }.run_pass(mir);
62 struct Inliner<'a, 'tcx: 'a> {
63 tcx: TyCtxt<'a, 'tcx, 'tcx>,
67 impl<'a, 'tcx> Inliner<'a, 'tcx> {
68 fn run_pass(&self, caller_mir: &mut Mir<'tcx>) {
69 // Keep a queue of callsites to try inlining on. We take
70 // advantage of the fact that queries detect cycles here to
71 // allow us to try and fetch the fully optimized MIR of a
72 // call; if it succeeds, we can inline it and we know that
73 // they do not call us. Otherwise, we just don't try to
76 // We use a queue so that we inline "broadly" before we inline
77 // in depth. It is unclear if this is the best heuristic,
78 // really, but that's true of all the heuristics in this
81 let mut callsites = VecDeque::new();
83 let param_env = self.tcx.param_env(self.source.def_id);
85 // Only do inlining into fn bodies.
86 let id = self.tcx.hir.as_local_node_id(self.source.def_id).unwrap();
87 let body_owner_kind = self.tcx.hir.body_owner_kind(id);
88 if let (hir::BodyOwnerKind::Fn, None) = (body_owner_kind, self.source.promoted) {
90 for (bb, bb_data) in caller_mir.basic_blocks().iter_enumerated() {
91 // Don't inline calls that are in cleanup blocks.
92 if bb_data.is_cleanup { continue; }
94 // Only consider direct calls to functions
95 let terminator = bb_data.terminator();
96 if let TerminatorKind::Call {
97 func: Operand::Constant(ref f), .. } = terminator.kind {
98 if let ty::TyFnDef(callee_def_id, substs) = f.ty.sty {
99 if let Some(instance) = Instance::resolve(self.tcx,
103 callsites.push_back(CallSite {
104 callee: instance.def_id(),
105 substs: instance.substs,
107 location: terminator.source_info
117 let mut local_change;
118 let mut changed = false;
121 local_change = false;
122 while let Some(callsite) = callsites.pop_front() {
123 debug!("checking whether to inline callsite {:?}", callsite);
124 if !self.tcx.is_mir_available(callsite.callee) {
125 debug!("checking whether to inline callsite {:?} - MIR unavailable", callsite);
129 let callee_mir = match self.tcx.try_get_query::<ty::queries::optimized_mir>(
130 callsite.location.span,
132 Ok(callee_mir) if self.should_inline(callsite, callee_mir) => {
133 self.tcx.subst_and_normalize_erasing_regions(
142 // FIXME(#43542) shouldn't have to cancel an error
148 let start = caller_mir.basic_blocks().len();
149 debug!("attempting to inline callsite {:?} - mir={:?}", callsite, callee_mir);
150 if !self.inline_call(callsite, caller_mir, callee_mir) {
151 debug!("attempting to inline callsite {:?} - failure", callsite);
154 debug!("attempting to inline callsite {:?} - success", callsite);
156 // Add callsites from inlined function
157 for (bb, bb_data) in caller_mir.basic_blocks().iter_enumerated().skip(start) {
158 // Only consider direct calls to functions
159 let terminator = bb_data.terminator();
160 if let TerminatorKind::Call {
161 func: Operand::Constant(ref f), .. } = terminator.kind {
162 if let ty::TyFnDef(callee_def_id, substs) = f.ty.sty {
163 // Don't inline the same function multiple times.
164 if callsite.callee != callee_def_id {
165 callsites.push_back(CallSite {
166 callee: callee_def_id,
169 location: terminator.source_info
185 // Simplify if we inlined anything.
187 debug!("Running simplify cfg on {:?}", self.source);
188 CfgSimplifier::new(caller_mir).simplify();
189 remove_dead_blocks(caller_mir);
193 fn should_inline(&self,
194 callsite: CallSite<'tcx>,
195 callee_mir: &Mir<'tcx>)
198 debug!("should_inline({:?})", callsite);
201 // Don't inline closures that have captures
202 // FIXME: Handle closures better
203 if callee_mir.upvar_decls.len() > 0 {
204 debug!(" upvar decls present - not inlining");
208 // Cannot inline generators which haven't been transformed yet
209 if callee_mir.yield_ty.is_some() {
210 debug!(" yield ty present - not inlining");
214 // Do not inline {u,i}128 lang items, trans const eval depends
215 // on detecting calls to these lang items and intercepting them
216 if tcx.is_binop_lang_item(callsite.callee).is_some() {
217 debug!(" not inlining 128bit integer lang item");
221 let trans_fn_attrs = tcx.trans_fn_attrs(callsite.callee);
223 let hinted = match trans_fn_attrs.inline {
224 // Just treat inline(always) as a hint for now,
225 // there are cases that prevent inlining that we
226 // need to check for first.
227 attr::InlineAttr::Always => true,
228 attr::InlineAttr::Never => {
229 debug!("#[inline(never)] present - not inlining");
232 attr::InlineAttr::Hint => true,
233 attr::InlineAttr::None => false,
236 // Only inline local functions if they would be eligible for cross-crate
237 // inlining. This is to ensure that the final crate doesn't have MIR that
238 // reference unexported symbols
239 if callsite.callee.is_local() {
240 if callsite.substs.types().count() == 0 && !hinted {
241 debug!(" callee is an exported function - not inlining");
246 let mut threshold = if hinted {
252 // Significantly lower the threshold for inlining cold functions
253 if trans_fn_attrs.flags.contains(TransFnAttrFlags::COLD) {
257 // Give a bonus functions with a small number of blocks,
258 // We normally have two or three blocks for even
259 // very small functions.
260 if callee_mir.basic_blocks().len() <= 3 {
261 threshold += threshold / 4;
263 debug!(" final inline threshold = {}", threshold);
265 // FIXME: Give a bonus to functions with only a single caller
267 let param_env = tcx.param_env(self.source.def_id);
269 let mut first_block = true;
272 // Traverse the MIR manually so we can account for the effects of
273 // inlining on the CFG.
274 let mut work_list = vec![START_BLOCK];
275 let mut visited = BitVector::new(callee_mir.basic_blocks().len());
276 while let Some(bb) = work_list.pop() {
277 if !visited.insert(bb.index()) { continue; }
278 let blk = &callee_mir.basic_blocks()[bb];
280 for stmt in &blk.statements {
281 // Don't count StorageLive/StorageDead in the inlining cost.
283 StatementKind::StorageLive(_) |
284 StatementKind::StorageDead(_) |
285 StatementKind::Nop => {}
286 _ => cost += INSTR_COST
289 let term = blk.terminator();
290 let mut is_drop = false;
292 TerminatorKind::Drop { ref location, target, unwind } |
293 TerminatorKind::DropAndReplace { ref location, target, unwind, .. } => {
295 work_list.push(target);
296 // If the location doesn't actually need dropping, treat it like
298 let ty = location.ty(callee_mir, tcx).subst(tcx, callsite.substs);
299 let ty = ty.to_ty(tcx);
300 if ty.needs_drop(tcx, param_env) {
301 cost += CALL_PENALTY;
302 if let Some(unwind) = unwind {
303 work_list.push(unwind);
310 TerminatorKind::Unreachable |
311 TerminatorKind::Call { destination: None, .. } if first_block => {
312 // If the function always diverges, don't inline
313 // unless the cost is zero
317 TerminatorKind::Call {func: Operand::Constant(ref f), .. } => {
318 if let ty::TyFnDef(def_id, _) = f.ty.sty {
319 // Don't give intrinsics the extra penalty for calls
320 let f = tcx.fn_sig(def_id);
321 if f.abi() == Abi::RustIntrinsic || f.abi() == Abi::PlatformIntrinsic {
324 cost += CALL_PENALTY;
328 TerminatorKind::Assert { .. } => cost += CALL_PENALTY,
329 _ => cost += INSTR_COST
333 for &succ in term.successors() {
334 work_list.push(succ);
341 // Count up the cost of local variables and temps, if we know the size
342 // use that, otherwise we use a moderately-large dummy cost.
344 let ptr_size = tcx.data_layout.pointer_size.bytes();
346 for v in callee_mir.vars_and_temps_iter() {
347 let v = &callee_mir.local_decls[v];
348 let ty = v.ty.subst(tcx, callsite.substs);
349 // Cost of the var is the size in machine-words, if we know
351 if let Some(size) = type_size_of(tcx, param_env.clone(), ty) {
352 cost += (size / ptr_size) as usize;
354 cost += UNKNOWN_SIZE_COST;
358 if let attr::InlineAttr::Always = trans_fn_attrs.inline {
359 debug!("INLINING {:?} because inline(always) [cost={}]", callsite, cost);
362 if cost <= threshold {
363 debug!("INLINING {:?} [cost={} <= threshold={}]", callsite, cost, threshold);
366 debug!("NOT inlining {:?} [cost={} > threshold={}]", callsite, cost, threshold);
372 fn inline_call(&self,
373 callsite: CallSite<'tcx>,
374 caller_mir: &mut Mir<'tcx>,
375 mut callee_mir: Mir<'tcx>) -> bool {
376 let terminator = caller_mir[callsite.bb].terminator.take().unwrap();
377 match terminator.kind {
378 // FIXME: Handle inlining of diverging calls
379 TerminatorKind::Call { args, destination: Some(destination), cleanup, .. } => {
380 debug!("Inlined {:?} into {:?}", callsite.callee, self.source);
382 let mut local_map = IndexVec::with_capacity(callee_mir.local_decls.len());
383 let mut scope_map = IndexVec::with_capacity(callee_mir.visibility_scopes.len());
384 let mut promoted_map = IndexVec::with_capacity(callee_mir.promoted.len());
386 for mut scope in callee_mir.visibility_scopes.iter().cloned() {
387 if scope.parent_scope.is_none() {
388 scope.parent_scope = Some(callsite.location.scope);
389 scope.span = callee_mir.span;
392 scope.span = callsite.location.span;
394 let idx = caller_mir.visibility_scopes.push(scope);
398 for loc in callee_mir.vars_and_temps_iter() {
399 let mut local = callee_mir.local_decls[loc].clone();
401 local.source_info.scope = scope_map[local.source_info.scope];
402 local.source_info.span = callsite.location.span;
404 let idx = caller_mir.local_decls.push(local);
408 for p in callee_mir.promoted.iter().cloned() {
409 let idx = caller_mir.promoted.push(p);
410 promoted_map.push(idx);
413 // If the call is something like `a[*i] = f(i)`, where
414 // `i : &mut usize`, then just duplicating the `a[*i]`
415 // Place could result in two different locations if `f`
416 // writes to `i`. To prevent this we need to create a temporary
417 // borrow of the place and pass the destination as `*temp` instead.
418 fn dest_needs_borrow(place: &Place) -> bool {
420 Place::Projection(ref p) => {
422 ProjectionElem::Deref |
423 ProjectionElem::Index(_) => true,
424 _ => dest_needs_borrow(&p.base)
427 // Static variables need a borrow because the callee
428 // might modify the same static.
429 Place::Static(_) => true,
434 let dest = if dest_needs_borrow(&destination.0) {
435 debug!("Creating temp for return destination");
436 let dest = Rvalue::Ref(
437 self.tcx.types.re_erased,
438 BorrowKind::Mut { allow_two_phase_borrow: false },
441 let ty = dest.ty(caller_mir, self.tcx);
443 let temp = LocalDecl::new_temp(ty, callsite.location.span);
445 let tmp = caller_mir.local_decls.push(temp);
446 let tmp = Place::Local(tmp);
448 let stmt = Statement {
449 source_info: callsite.location,
450 kind: StatementKind::Assign(tmp.clone(), dest)
452 caller_mir[callsite.bb]
453 .statements.push(stmt);
459 let return_block = destination.1;
461 // Copy the arguments if needed.
462 let args: Vec<_> = self.make_call_args(args, &callsite, caller_mir);
464 let bb_len = caller_mir.basic_blocks().len();
465 let mut integrator = Integrator {
474 cleanup_block: cleanup,
475 in_cleanup_block: false
479 for (bb, mut block) in callee_mir.basic_blocks_mut().drain_enumerated(..) {
480 integrator.visit_basic_block_data(bb, &mut block);
481 caller_mir.basic_blocks_mut().push(block);
484 let terminator = Terminator {
485 source_info: callsite.location,
486 kind: TerminatorKind::Goto { target: BasicBlock::new(bb_len) }
489 caller_mir[callsite.bb].terminator = Some(terminator);
494 caller_mir[callsite.bb].terminator = Some(Terminator {
495 source_info: terminator.source_info,
505 args: Vec<Operand<'tcx>>,
506 callsite: &CallSite<'tcx>,
507 caller_mir: &mut Mir<'tcx>,
511 // There is a bit of a mismatch between the *caller* of a closure and the *callee*.
512 // The caller provides the arguments wrapped up in a tuple:
514 // tuple_tmp = (a, b, c)
515 // Fn::call(closure_ref, tuple_tmp)
517 // meanwhile the closure body expects the arguments (here, `a`, `b`, and `c`)
518 // as distinct arguments. (This is the "rust-call" ABI hack.) Normally, trans has
519 // the job of unpacking this tuple. But here, we are trans. =) So we want to create
522 // [closure_ref, tuple_tmp.0, tuple_tmp.1, tuple_tmp.2]
524 // Except for one tiny wrinkle: we don't actually want `tuple_tmp.0`. It's more convenient
525 // if we "spill" that into *another* temporary, so that we can map the argument
526 // variable in the callee MIR directly to an argument variable on our side.
527 // So we introduce temporaries like:
529 // tmp0 = tuple_tmp.0
530 // tmp1 = tuple_tmp.1
531 // tmp2 = tuple_tmp.2
533 // and the vector is `[closure_ref, tmp0, tmp1, tmp2]`.
534 if tcx.is_closure(callsite.callee) {
535 let mut args = args.into_iter();
536 let self_ = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_mir);
537 let tuple = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_mir);
538 assert!(args.next().is_none());
540 let tuple = Place::Local(tuple);
541 let tuple_tys = if let ty::TyTuple(s) = tuple.ty(caller_mir, tcx).to_ty(tcx).sty {
544 bug!("Closure arguments are not passed as a tuple");
547 // The `closure_ref` in our example above.
548 let closure_ref_arg = iter::once(self_);
550 // The `tmp0`, `tmp1`, and `tmp2` in our example abonve.
552 tuple_tys.iter().enumerate().map(|(i, ty)| {
553 // This is e.g. `tuple_tmp.0` in our example above.
554 let tuple_field = Operand::Move(tuple.clone().field(Field::new(i), ty));
556 // Spill to a local to make e.g. `tmp0`.
557 self.create_temp_if_necessary(tuple_field, callsite, caller_mir)
560 closure_ref_arg.chain(tuple_tmp_args).collect()
563 .map(|a| self.create_temp_if_necessary(a, callsite, caller_mir))
568 /// If `arg` is already a temporary, returns it. Otherwise, introduces a fresh
569 /// temporary `T` and an instruction `T = arg`, and returns `T`.
570 fn create_temp_if_necessary(
573 callsite: &CallSite<'tcx>,
574 caller_mir: &mut Mir<'tcx>,
576 // FIXME: Analysis of the usage of the arguments to avoid
577 // unnecessary temporaries.
579 if let Operand::Move(Place::Local(local)) = arg {
580 if caller_mir.local_kind(local) == LocalKind::Temp {
581 // Reuse the operand if it's a temporary already
586 debug!("Creating temp for argument {:?}", arg);
587 // Otherwise, create a temporary for the arg
588 let arg = Rvalue::Use(arg);
590 let ty = arg.ty(caller_mir, self.tcx);
592 let arg_tmp = LocalDecl::new_temp(ty, callsite.location.span);
593 let arg_tmp = caller_mir.local_decls.push(arg_tmp);
595 let stmt = Statement {
596 source_info: callsite.location,
597 kind: StatementKind::Assign(Place::Local(arg_tmp), arg),
599 caller_mir[callsite.bb].statements.push(stmt);
604 fn type_size_of<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
605 param_env: ty::ParamEnv<'tcx>,
606 ty: Ty<'tcx>) -> Option<u64> {
607 tcx.layout_of(param_env.and(ty)).ok().map(|layout| layout.size.bytes())
613 * Integrates blocks from the callee function into the calling function.
614 * Updates block indices, references to locals and other control flow
617 struct Integrator<'a, 'tcx: 'a> {
620 local_map: IndexVec<Local, Local>,
621 scope_map: IndexVec<VisibilityScope, VisibilityScope>,
622 promoted_map: IndexVec<Promoted, Promoted>,
623 _callsite: CallSite<'tcx>,
624 destination: Place<'tcx>,
625 return_block: BasicBlock,
626 cleanup_block: Option<BasicBlock>,
627 in_cleanup_block: bool,
630 impl<'a, 'tcx> Integrator<'a, 'tcx> {
631 fn update_target(&self, tgt: BasicBlock) -> BasicBlock {
632 let new = BasicBlock::new(tgt.index() + self.block_idx);
633 debug!("Updating target `{:?}`, new: `{:?}`", tgt, new);
638 impl<'a, 'tcx> MutVisitor<'tcx> for Integrator<'a, 'tcx> {
639 fn visit_local(&mut self,
641 _ctxt: PlaceContext<'tcx>,
642 _location: Location) {
643 if *local == RETURN_PLACE {
644 match self.destination {
649 ref place => bug!("Return place is {:?}, not local", place)
652 let idx = local.index() - 1;
653 if idx < self.args.len() {
654 *local = self.args[idx];
657 *local = self.local_map[Local::new(idx - self.args.len())];
660 fn visit_place(&mut self,
661 place: &mut Place<'tcx>,
662 _ctxt: PlaceContext<'tcx>,
663 _location: Location) {
664 if let Place::Local(RETURN_PLACE) = *place {
665 // Return pointer; update the place itself
666 *place = self.destination.clone();
668 self.super_place(place, _ctxt, _location);
672 fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
673 self.in_cleanup_block = data.is_cleanup;
674 self.super_basic_block_data(block, data);
675 self.in_cleanup_block = false;
678 fn visit_terminator_kind(&mut self, block: BasicBlock,
679 kind: &mut TerminatorKind<'tcx>, loc: Location) {
680 self.super_terminator_kind(block, kind, loc);
683 TerminatorKind::GeneratorDrop |
684 TerminatorKind::Yield { .. } => bug!(),
685 TerminatorKind::Goto { ref mut target} => {
686 *target = self.update_target(*target);
688 TerminatorKind::SwitchInt { ref mut targets, .. } => {
690 *tgt = self.update_target(*tgt);
693 TerminatorKind::Drop { ref mut target, ref mut unwind, .. } |
694 TerminatorKind::DropAndReplace { ref mut target, ref mut unwind, .. } => {
695 *target = self.update_target(*target);
696 if let Some(tgt) = *unwind {
697 *unwind = Some(self.update_target(tgt));
698 } else if !self.in_cleanup_block {
699 // Unless this drop is in a cleanup block, add an unwind edge to
700 // the orignal call's cleanup block
701 *unwind = self.cleanup_block;
704 TerminatorKind::Call { ref mut destination, ref mut cleanup, .. } => {
705 if let Some((_, ref mut tgt)) = *destination {
706 *tgt = self.update_target(*tgt);
708 if let Some(tgt) = *cleanup {
709 *cleanup = Some(self.update_target(tgt));
710 } else if !self.in_cleanup_block {
711 // Unless this call is in a cleanup block, add an unwind edge to
712 // the orignal call's cleanup block
713 *cleanup = self.cleanup_block;
716 TerminatorKind::Assert { ref mut target, ref mut cleanup, .. } => {
717 *target = self.update_target(*target);
718 if let Some(tgt) = *cleanup {
719 *cleanup = Some(self.update_target(tgt));
720 } else if !self.in_cleanup_block {
721 // Unless this assert is in a cleanup block, add an unwind edge to
722 // the orignal call's cleanup block
723 *cleanup = self.cleanup_block;
726 TerminatorKind::Return => {
727 *kind = TerminatorKind::Goto { target: self.return_block };
729 TerminatorKind::Resume => {
730 if let Some(tgt) = self.cleanup_block {
731 *kind = TerminatorKind::Goto { target: tgt }
734 TerminatorKind::Abort => { }
735 TerminatorKind::Unreachable => { }
736 TerminatorKind::FalseEdges { ref mut real_target, ref mut imaginary_targets } => {
737 *real_target = self.update_target(*real_target);
738 for target in imaginary_targets {
739 *target = self.update_target(*target);
742 TerminatorKind::FalseUnwind { real_target: _ , unwind: _ } =>
743 // see the ordering of passes in the optimized_mir query.
744 bug!("False unwinds should have been removed before inlining")
748 fn visit_visibility_scope(&mut self, scope: &mut VisibilityScope) {
749 *scope = self.scope_map[*scope];
752 fn visit_literal(&mut self, literal: &mut Literal<'tcx>, loc: Location) {
753 if let Literal::Promoted { ref mut index } = *literal {
754 if let Some(p) = self.promoted_map.get(*index).cloned() {
758 self.super_literal(literal, loc);