1 //! Inlining pass for MIR functions
2 use crate::deref_separator::deref_finder;
3 use rustc_attr::InlineAttr;
4 use rustc_const_eval::transform::validate::equal_up_to_regions;
5 use rustc_index::bit_set::BitSet;
6 use rustc_index::vec::Idx;
7 use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs};
8 use rustc_middle::mir::visit::*;
9 use rustc_middle::mir::*;
10 use rustc_middle::ty::{self, Instance, InstanceDef, ParamEnv, Ty, TyCtxt};
11 use rustc_session::config::OptLevel;
12 use rustc_span::def_id::DefId;
13 use rustc_span::{hygiene::ExpnKind, ExpnData, LocalExpnId, Span};
14 use rustc_target::abi::VariantIdx;
15 use rustc_target::spec::abi::Abi;
17 use super::simplify::{remove_dead_blocks, CfgSimplifier};
20 use std::ops::{Range, RangeFrom};
24 const INSTR_COST: usize = 5;
25 const CALL_PENALTY: usize = 25;
26 const LANDINGPAD_PENALTY: usize = 50;
27 const RESUME_PENALTY: usize = 45;
29 const UNKNOWN_SIZE_COST: usize = 10;
33 #[derive(Copy, Clone, Debug)]
34 struct CallSite<'tcx> {
35 callee: Instance<'tcx>,
36 fn_sig: ty::PolyFnSig<'tcx>,
38 target: Option<BasicBlock>,
39 source_info: SourceInfo,
42 impl<'tcx> MirPass<'tcx> for Inline {
43 fn is_enabled(&self, sess: &rustc_session::Session) -> bool {
44 if let Some(enabled) = sess.opts.unstable_opts.inline_mir {
48 match sess.mir_opt_level() {
51 (sess.opts.optimize == OptLevel::Default
52 || sess.opts.optimize == OptLevel::Aggressive)
53 && sess.opts.incremental == None
59 fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
60 let span = trace_span!("inline", body = %tcx.def_path_str(body.source.def_id()));
61 let _guard = span.enter();
62 if inline(tcx, body) {
63 debug!("running simplify cfg on {:?}", body.source);
64 CfgSimplifier::new(body).simplify();
65 remove_dead_blocks(tcx, body);
66 deref_finder(tcx, body);
71 fn inline<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) -> bool {
72 let def_id = body.source.def_id().expect_local();
74 // Only do inlining into fn bodies.
75 if !tcx.hir().body_owner_kind(def_id).is_fn_or_closure() {
78 if body.source.promoted.is_some() {
81 // Avoid inlining into generators, since their `optimized_mir` is used for layout computation,
82 // which can create a cycle, even when no attempt is made to inline the function in the other
84 if body.generator.is_some() {
88 let param_env = tcx.param_env_reveal_all_normalized(def_id);
90 let mut this = Inliner {
93 codegen_fn_attrs: tcx.codegen_fn_attrs(def_id),
97 let blocks = BasicBlock::new(0)..body.basic_blocks.next_index();
98 this.process_blocks(body, blocks);
102 struct Inliner<'tcx> {
104 param_env: ParamEnv<'tcx>,
105 /// Caller codegen attributes.
106 codegen_fn_attrs: &'tcx CodegenFnAttrs,
107 /// Stack of inlined instances.
108 /// We only check the `DefId` and not the substs because we want to
109 /// avoid inlining cases of polymorphic recursion.
110 /// The number of `DefId`s is finite, so checking history is enough
111 /// to ensure that we do not loop endlessly while inlining.
113 /// Indicates that the caller body has been modified.
117 impl<'tcx> Inliner<'tcx> {
118 fn process_blocks(&mut self, caller_body: &mut Body<'tcx>, blocks: Range<BasicBlock>) {
120 let bb_data = &caller_body[bb];
121 if bb_data.is_cleanup {
125 let Some(callsite) = self.resolve_callsite(caller_body, bb, bb_data) else {
129 let span = trace_span!("process_blocks", %callsite.callee, ?bb);
130 let _guard = span.enter();
132 match self.try_inlining(caller_body, &callsite) {
134 debug!("not-inlined {} [{}]", callsite.callee, reason);
138 debug!("inlined {}", callsite.callee);
140 self.history.push(callsite.callee.def_id());
141 self.process_blocks(caller_body, new_blocks);
148 /// Attempts to inline a callsite into the caller body. When successful returns basic blocks
149 /// containing the inlined body. Otherwise returns an error describing why inlining didn't take
153 caller_body: &mut Body<'tcx>,
154 callsite: &CallSite<'tcx>,
155 ) -> Result<std::ops::Range<BasicBlock>, &'static str> {
156 let callee_attrs = self.tcx.codegen_fn_attrs(callsite.callee.def_id());
157 self.check_codegen_attributes(callsite, callee_attrs)?;
158 self.check_mir_is_available(caller_body, &callsite.callee)?;
159 let callee_body = self.tcx.instance_mir(callsite.callee.def);
160 self.check_mir_body(callsite, callee_body, callee_attrs)?;
162 if !self.tcx.consider_optimizing(|| {
163 format!("Inline {:?} into {:?}", callsite.callee, caller_body.source)
165 return Err("optimization fuel exhausted");
168 let Ok(callee_body) = callsite.callee.try_subst_mir_and_normalize_erasing_regions(
173 return Err("failed to normalize callee body");
176 // Check call signature compatibility.
177 // Normally, this shouldn't be required, but trait normalization failure can create a
179 let terminator = caller_body[callsite.block].terminator.as_ref().unwrap();
180 let TerminatorKind::Call { args, destination, .. } = &terminator.kind else { bug!() };
181 let destination_ty = destination.ty(&caller_body.local_decls, self.tcx).ty;
182 let output_type = callee_body.return_ty();
183 if !equal_up_to_regions(self.tcx, self.param_env, output_type, destination_ty) {
184 trace!(?output_type, ?destination_ty);
185 return Err("failed to normalize return type");
187 if callsite.fn_sig.abi() == Abi::RustCall {
188 let (arg_tuple, skipped_args) = match &args[..] {
189 [arg_tuple] => (arg_tuple, 0),
190 [_, arg_tuple] => (arg_tuple, 1),
191 _ => bug!("Expected `rust-call` to have 1 or 2 args"),
194 let arg_tuple_ty = arg_tuple.ty(&caller_body.local_decls, self.tcx);
195 let ty::Tuple(arg_tuple_tys) = arg_tuple_ty.kind() else {
196 bug!("Closure arguments are not passed as a tuple");
199 for (arg_ty, input) in
200 arg_tuple_tys.iter().zip(callee_body.args_iter().skip(skipped_args))
202 let input_type = callee_body.local_decls[input].ty;
203 if !equal_up_to_regions(self.tcx, self.param_env, arg_ty, input_type) {
204 trace!(?arg_ty, ?input_type);
205 return Err("failed to normalize tuple argument type");
209 for (arg, input) in args.iter().zip(callee_body.args_iter()) {
210 let input_type = callee_body.local_decls[input].ty;
211 let arg_ty = arg.ty(&caller_body.local_decls, self.tcx);
212 if !equal_up_to_regions(self.tcx, self.param_env, arg_ty, input_type) {
213 trace!(?arg_ty, ?input_type);
214 return Err("failed to normalize argument type");
219 let old_blocks = caller_body.basic_blocks.next_index();
220 self.inline_call(caller_body, &callsite, callee_body);
221 let new_blocks = old_blocks..caller_body.basic_blocks.next_index();
226 fn check_mir_is_available(
228 caller_body: &Body<'tcx>,
229 callee: &Instance<'tcx>,
230 ) -> Result<(), &'static str> {
231 let caller_def_id = caller_body.source.def_id();
232 let callee_def_id = callee.def_id();
233 if callee_def_id == caller_def_id {
234 return Err("self-recursion");
238 InstanceDef::Item(_) => {
239 // If there is no MIR available (either because it was not in metadata or
240 // because it has no MIR because it's an extern function), then the inliner
241 // won't cause cycles on this.
242 if !self.tcx.is_mir_available(callee_def_id) {
243 return Err("item MIR unavailable");
246 // These have no own callable MIR.
247 InstanceDef::Intrinsic(_) | InstanceDef::Virtual(..) => {
248 return Err("instance without MIR (intrinsic / virtual)");
250 // This cannot result in an immediate cycle since the callee MIR is a shim, which does
251 // not get any optimizations run on it. Any subsequent inlining may cause cycles, but we
252 // do not need to catch this here, we can wait until the inliner decides to continue
253 // inlining a second time.
254 InstanceDef::VTableShim(_)
255 | InstanceDef::ReifyShim(_)
256 | InstanceDef::FnPtrShim(..)
257 | InstanceDef::ClosureOnceShim { .. }
258 | InstanceDef::DropGlue(..)
259 | InstanceDef::CloneShim(..) => return Ok(()),
262 if self.tcx.is_constructor(callee_def_id) {
263 trace!("constructors always have MIR");
264 // Constructor functions cannot cause a query cycle.
268 if callee_def_id.is_local() {
269 // Avoid a cycle here by only using `instance_mir` only if we have
270 // a lower `DefPathHash` than the callee. This ensures that the callee will
271 // not inline us. This trick even works with incremental compilation,
272 // since `DefPathHash` is stable.
273 if self.tcx.def_path_hash(caller_def_id).local_hash()
274 < self.tcx.def_path_hash(callee_def_id).local_hash()
279 // If we know for sure that the function we're calling will itself try to
280 // call us, then we avoid inlining that function.
281 if self.tcx.mir_callgraph_reachable((*callee, caller_def_id.expect_local())) {
282 return Err("caller might be reachable from callee (query cycle avoidance)");
287 // This cannot result in an immediate cycle since the callee MIR is from another crate
288 // and is already optimized. Any subsequent inlining may cause cycles, but we do
289 // not need to catch this here, we can wait until the inliner decides to continue
290 // inlining a second time.
291 trace!("functions from other crates always have MIR");
298 caller_body: &Body<'tcx>,
300 bb_data: &BasicBlockData<'tcx>,
301 ) -> Option<CallSite<'tcx>> {
302 // Only consider direct calls to functions
303 let terminator = bb_data.terminator();
304 if let TerminatorKind::Call { ref func, target, .. } = terminator.kind {
305 let func_ty = func.ty(caller_body, self.tcx);
306 if let ty::FnDef(def_id, substs) = *func_ty.kind() {
307 // To resolve an instance its substs have to be fully normalized.
308 let substs = self.tcx.try_normalize_erasing_regions(self.param_env, substs).ok()?;
310 Instance::resolve(self.tcx, self.param_env, def_id, substs).ok().flatten()?;
312 if let InstanceDef::Virtual(..) | InstanceDef::Intrinsic(_) = callee.def {
316 if self.history.contains(&callee.def_id()) {
320 let fn_sig = self.tcx.bound_fn_sig(def_id).subst(self.tcx, substs);
322 return Some(CallSite {
327 source_info: terminator.source_info,
335 /// Returns an error if inlining is not possible based on codegen attributes alone. A success
336 /// indicates that inlining decision should be based on other criteria.
337 fn check_codegen_attributes(
339 callsite: &CallSite<'tcx>,
340 callee_attrs: &CodegenFnAttrs,
341 ) -> Result<(), &'static str> {
342 match callee_attrs.inline {
343 InlineAttr::Never => return Err("never inline hint"),
344 InlineAttr::Always | InlineAttr::Hint => {}
345 InlineAttr::None => {
346 if self.tcx.sess.mir_opt_level() <= 2 {
347 return Err("at mir-opt-level=2, only #[inline] is inlined");
352 // Only inline local functions if they would be eligible for cross-crate
353 // inlining. This is to ensure that the final crate doesn't have MIR that
354 // reference unexported symbols
355 if callsite.callee.def_id().is_local() {
356 let is_generic = callsite.callee.substs.non_erasable_generics().next().is_some();
357 if !is_generic && !callee_attrs.requests_inline() {
358 return Err("not exported");
362 if callsite.fn_sig.c_variadic() {
363 return Err("C variadic");
366 if callee_attrs.flags.contains(CodegenFnAttrFlags::NAKED) {
370 if callee_attrs.flags.contains(CodegenFnAttrFlags::COLD) {
374 if callee_attrs.no_sanitize != self.codegen_fn_attrs.no_sanitize {
375 return Err("incompatible sanitizer set");
378 if callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set {
379 return Err("incompatible instruction set");
382 for feature in &callee_attrs.target_features {
383 if !self.codegen_fn_attrs.target_features.contains(feature) {
384 return Err("incompatible target feature");
391 /// Returns inlining decision that is based on the examination of callee MIR body.
392 /// Assumes that codegen attributes have been checked for compatibility already.
393 #[instrument(level = "debug", skip(self, callee_body))]
396 callsite: &CallSite<'tcx>,
397 callee_body: &Body<'tcx>,
398 callee_attrs: &CodegenFnAttrs,
399 ) -> Result<(), &'static str> {
402 let mut threshold = if callee_attrs.requests_inline() {
403 self.tcx.sess.opts.unstable_opts.inline_mir_hint_threshold.unwrap_or(100)
405 self.tcx.sess.opts.unstable_opts.inline_mir_threshold.unwrap_or(50)
408 // Give a bonus functions with a small number of blocks,
409 // We normally have two or three blocks for even
410 // very small functions.
411 if callee_body.basic_blocks.len() <= 3 {
412 threshold += threshold / 4;
414 debug!(" final inline threshold = {}", threshold);
416 // FIXME: Give a bonus to functions with only a single caller
417 let diverges = matches!(
418 callee_body.basic_blocks[START_BLOCK].terminator().kind,
419 TerminatorKind::Unreachable | TerminatorKind::Call { target: None, .. }
421 if diverges && !matches!(callee_attrs.inline, InlineAttr::Always) {
422 return Err("callee diverges unconditionally");
425 let mut checker = CostChecker {
427 param_env: self.param_env,
428 instance: callsite.callee,
434 // Traverse the MIR manually so we can account for the effects of inlining on the CFG.
435 let mut work_list = vec![START_BLOCK];
436 let mut visited = BitSet::new_empty(callee_body.basic_blocks.len());
437 while let Some(bb) = work_list.pop() {
438 if !visited.insert(bb.index()) {
442 let blk = &callee_body.basic_blocks[bb];
443 checker.visit_basic_block_data(bb, blk);
445 let term = blk.terminator();
446 if let TerminatorKind::Drop { ref place, target, unwind }
447 | TerminatorKind::DropAndReplace { ref place, target, unwind, .. } = term.kind
449 work_list.push(target);
451 // If the place doesn't actually need dropping, treat it like a regular goto.
452 let ty = callsite.callee.subst_mir(self.tcx, &place.ty(callee_body, tcx).ty);
453 if ty.needs_drop(tcx, self.param_env) && let Some(unwind) = unwind {
454 work_list.push(unwind);
457 work_list.extend(term.successors())
461 // Count up the cost of local variables and temps, if we know the size
462 // use that, otherwise we use a moderately-large dummy cost.
463 for v in callee_body.vars_and_temps_iter() {
464 checker.visit_local_decl(v, &callee_body.local_decls[v]);
467 // Abort if type validation found anything fishy.
470 let cost = checker.cost;
471 if let InlineAttr::Always = callee_attrs.inline {
472 debug!("INLINING {:?} because inline(always) [cost={}]", callsite, cost);
474 } else if cost <= threshold {
475 debug!("INLINING {:?} [cost={} <= threshold={}]", callsite, cost, threshold);
478 debug!("NOT inlining {:?} [cost={} > threshold={}]", callsite, cost, threshold);
479 Err("cost above threshold")
485 caller_body: &mut Body<'tcx>,
486 callsite: &CallSite<'tcx>,
487 mut callee_body: Body<'tcx>,
489 let terminator = caller_body[callsite.block].terminator.take().unwrap();
490 match terminator.kind {
491 TerminatorKind::Call { args, destination, cleanup, .. } => {
492 // If the call is something like `a[*i] = f(i)`, where
493 // `i : &mut usize`, then just duplicating the `a[*i]`
494 // Place could result in two different locations if `f`
495 // writes to `i`. To prevent this we need to create a temporary
496 // borrow of the place and pass the destination as `*temp` instead.
497 fn dest_needs_borrow(place: Place<'_>) -> bool {
498 for elem in place.projection.iter() {
500 ProjectionElem::Deref | ProjectionElem::Index(_) => return true,
508 let dest = if dest_needs_borrow(destination) {
509 trace!("creating temp for return destination");
510 let dest = Rvalue::Ref(
511 self.tcx.lifetimes.re_erased,
512 BorrowKind::Mut { allow_two_phase_borrow: false },
515 let dest_ty = dest.ty(caller_body, self.tcx);
516 let temp = Place::from(self.new_call_temp(caller_body, &callsite, dest_ty));
517 caller_body[callsite.block].statements.push(Statement {
518 source_info: callsite.source_info,
519 kind: StatementKind::Assign(Box::new((temp, dest))),
521 self.tcx.mk_place_deref(temp)
526 // Copy the arguments if needed.
527 let args: Vec<_> = self.make_call_args(args, &callsite, caller_body, &callee_body);
529 let mut expn_data = ExpnData::default(
531 callsite.source_info.span,
532 self.tcx.sess.edition(),
536 expn_data.def_site = callee_body.span;
538 self.tcx.with_stable_hashing_context(|hcx| LocalExpnId::fresh(expn_data, hcx));
539 let mut integrator = Integrator {
541 new_locals: Local::new(caller_body.local_decls.len())..,
542 new_scopes: SourceScope::new(caller_body.source_scopes.len())..,
543 new_blocks: BasicBlock::new(caller_body.basic_blocks.len())..,
545 callsite_scope: caller_body.source_scopes[callsite.source_info.scope].clone(),
547 cleanup_block: cleanup,
548 in_cleanup_block: false,
551 always_live_locals: BitSet::new_filled(callee_body.local_decls.len()),
554 // Map all `Local`s, `SourceScope`s and `BasicBlock`s to new ones
555 // (or existing ones, in a few special cases) in the caller.
556 integrator.visit_body(&mut callee_body);
558 // If there are any locals without storage markers, give them storage only for the
559 // duration of the call.
560 for local in callee_body.vars_and_temps_iter() {
561 if !callee_body.local_decls[local].internal
562 && integrator.always_live_locals.contains(local)
564 let new_local = integrator.map_local(local);
565 caller_body[callsite.block].statements.push(Statement {
566 source_info: callsite.source_info,
567 kind: StatementKind::StorageLive(new_local),
571 if let Some(block) = callsite.target {
572 // To avoid repeated O(n) insert, push any new statements to the end and rotate
575 for local in callee_body.vars_and_temps_iter().rev() {
576 if !callee_body.local_decls[local].internal
577 && integrator.always_live_locals.contains(local)
579 let new_local = integrator.map_local(local);
580 caller_body[block].statements.push(Statement {
581 source_info: callsite.source_info,
582 kind: StatementKind::StorageDead(new_local),
587 caller_body[block].statements.rotate_right(n);
590 // Insert all of the (mapped) parts of the callee body into the caller.
591 caller_body.local_decls.extend(callee_body.drain_vars_and_temps());
592 caller_body.source_scopes.extend(&mut callee_body.source_scopes.drain(..));
593 caller_body.var_debug_info.append(&mut callee_body.var_debug_info);
594 caller_body.basic_blocks_mut().extend(callee_body.basic_blocks_mut().drain(..));
596 caller_body[callsite.block].terminator = Some(Terminator {
597 source_info: callsite.source_info,
598 kind: TerminatorKind::Goto { target: integrator.map_block(START_BLOCK) },
601 // Copy only unevaluated constants from the callee_body into the caller_body.
602 // Although we are only pushing `ConstKind::Unevaluated` consts to
603 // `required_consts`, here we may not only have `ConstKind::Unevaluated`
604 // because we are calling `subst_and_normalize_erasing_regions`.
605 caller_body.required_consts.extend(
606 callee_body.required_consts.iter().copied().filter(|&ct| match ct.literal {
607 ConstantKind::Ty(_) => {
608 bug!("should never encounter ty::UnevaluatedConst in `required_consts`")
610 ConstantKind::Val(..) | ConstantKind::Unevaluated(..) => true,
614 kind => bug!("unexpected terminator kind {:?}", kind),
620 args: Vec<Operand<'tcx>>,
621 callsite: &CallSite<'tcx>,
622 caller_body: &mut Body<'tcx>,
623 callee_body: &Body<'tcx>,
627 // There is a bit of a mismatch between the *caller* of a closure and the *callee*.
628 // The caller provides the arguments wrapped up in a tuple:
630 // tuple_tmp = (a, b, c)
631 // Fn::call(closure_ref, tuple_tmp)
633 // meanwhile the closure body expects the arguments (here, `a`, `b`, and `c`)
634 // as distinct arguments. (This is the "rust-call" ABI hack.) Normally, codegen has
635 // the job of unpacking this tuple. But here, we are codegen. =) So we want to create
638 // [closure_ref, tuple_tmp.0, tuple_tmp.1, tuple_tmp.2]
640 // Except for one tiny wrinkle: we don't actually want `tuple_tmp.0`. It's more convenient
641 // if we "spill" that into *another* temporary, so that we can map the argument
642 // variable in the callee MIR directly to an argument variable on our side.
643 // So we introduce temporaries like:
645 // tmp0 = tuple_tmp.0
646 // tmp1 = tuple_tmp.1
647 // tmp2 = tuple_tmp.2
649 // and the vector is `[closure_ref, tmp0, tmp1, tmp2]`.
650 if callsite.fn_sig.abi() == Abi::RustCall && callee_body.spread_arg.is_none() {
651 let mut args = args.into_iter();
652 let self_ = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
653 let tuple = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
654 assert!(args.next().is_none());
656 let tuple = Place::from(tuple);
657 let ty::Tuple(tuple_tys) = tuple.ty(caller_body, tcx).ty.kind() else {
658 bug!("Closure arguments are not passed as a tuple");
661 // The `closure_ref` in our example above.
662 let closure_ref_arg = iter::once(self_);
664 // The `tmp0`, `tmp1`, and `tmp2` in our example above.
665 let tuple_tmp_args = tuple_tys.iter().enumerate().map(|(i, ty)| {
666 // This is e.g., `tuple_tmp.0` in our example above.
667 let tuple_field = Operand::Move(tcx.mk_place_field(tuple, Field::new(i), ty));
669 // Spill to a local to make e.g., `tmp0`.
670 self.create_temp_if_necessary(tuple_field, callsite, caller_body)
673 closure_ref_arg.chain(tuple_tmp_args).collect()
676 .map(|a| self.create_temp_if_necessary(a, callsite, caller_body))
681 /// If `arg` is already a temporary, returns it. Otherwise, introduces a fresh
682 /// temporary `T` and an instruction `T = arg`, and returns `T`.
683 fn create_temp_if_necessary(
686 callsite: &CallSite<'tcx>,
687 caller_body: &mut Body<'tcx>,
689 // Reuse the operand if it is a moved temporary.
690 if let Operand::Move(place) = &arg
691 && let Some(local) = place.as_local()
692 && caller_body.local_kind(local) == LocalKind::Temp
697 // Otherwise, create a temporary for the argument.
698 trace!("creating temp for argument {:?}", arg);
699 let arg_ty = arg.ty(caller_body, self.tcx);
700 let local = self.new_call_temp(caller_body, callsite, arg_ty);
701 caller_body[callsite.block].statements.push(Statement {
702 source_info: callsite.source_info,
703 kind: StatementKind::Assign(Box::new((Place::from(local), Rvalue::Use(arg)))),
708 /// Introduces a new temporary into the caller body that is live for the duration of the call.
711 caller_body: &mut Body<'tcx>,
712 callsite: &CallSite<'tcx>,
715 let local = caller_body.local_decls.push(LocalDecl::new(ty, callsite.source_info.span));
717 caller_body[callsite.block].statements.push(Statement {
718 source_info: callsite.source_info,
719 kind: StatementKind::StorageLive(local),
722 if let Some(block) = callsite.target {
723 caller_body[block].statements.insert(
726 source_info: callsite.source_info,
727 kind: StatementKind::StorageDead(local),
736 fn type_size_of<'tcx>(
738 param_env: ty::ParamEnv<'tcx>,
741 tcx.layout_of(param_env.and(ty)).ok().map(|layout| layout.size.bytes())
744 /// Verify that the callee body is compatible with the caller.
746 /// This visitor mostly computes the inlining cost,
747 /// but also needs to verify that types match because of normalization failure.
748 struct CostChecker<'b, 'tcx> {
750 param_env: ParamEnv<'tcx>,
752 callee_body: &'b Body<'tcx>,
753 instance: ty::Instance<'tcx>,
754 validation: Result<(), &'static str>,
757 impl<'tcx> Visitor<'tcx> for CostChecker<'_, 'tcx> {
758 fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) {
759 // Don't count StorageLive/StorageDead in the inlining cost.
760 match statement.kind {
761 StatementKind::StorageLive(_)
762 | StatementKind::StorageDead(_)
763 | StatementKind::Deinit(_)
764 | StatementKind::Nop => {}
765 _ => self.cost += INSTR_COST,
768 self.super_statement(statement, location);
771 fn visit_terminator(&mut self, terminator: &Terminator<'tcx>, location: Location) {
773 match terminator.kind {
774 TerminatorKind::Drop { ref place, unwind, .. }
775 | TerminatorKind::DropAndReplace { ref place, unwind, .. } => {
776 // If the place doesn't actually need dropping, treat it like a regular goto.
777 let ty = self.instance.subst_mir(tcx, &place.ty(self.callee_body, tcx).ty);
778 if ty.needs_drop(tcx, self.param_env) {
779 self.cost += CALL_PENALTY;
780 if unwind.is_some() {
781 self.cost += LANDINGPAD_PENALTY;
784 self.cost += INSTR_COST;
787 TerminatorKind::Call { func: Operand::Constant(ref f), cleanup, .. } => {
788 let fn_ty = self.instance.subst_mir(tcx, &f.literal.ty());
789 self.cost += if let ty::FnDef(def_id, _) = *fn_ty.kind() && tcx.is_intrinsic(def_id) {
790 // Don't give intrinsics the extra penalty for calls
795 if cleanup.is_some() {
796 self.cost += LANDINGPAD_PENALTY;
799 TerminatorKind::Assert { cleanup, .. } => {
800 self.cost += CALL_PENALTY;
801 if cleanup.is_some() {
802 self.cost += LANDINGPAD_PENALTY;
805 TerminatorKind::Resume => self.cost += RESUME_PENALTY,
806 TerminatorKind::InlineAsm { cleanup, .. } => {
807 self.cost += INSTR_COST;
808 if cleanup.is_some() {
809 self.cost += LANDINGPAD_PENALTY;
812 _ => self.cost += INSTR_COST,
815 self.super_terminator(terminator, location);
818 /// Count up the cost of local variables and temps, if we know the size
819 /// use that, otherwise we use a moderately-large dummy cost.
820 fn visit_local_decl(&mut self, local: Local, local_decl: &LocalDecl<'tcx>) {
822 let ptr_size = tcx.data_layout.pointer_size.bytes();
824 let ty = self.instance.subst_mir(tcx, &local_decl.ty);
825 // Cost of the var is the size in machine-words, if we know
827 if let Some(size) = type_size_of(tcx, self.param_env, ty) {
828 self.cost += ((size + ptr_size - 1) / ptr_size) as usize;
830 self.cost += UNKNOWN_SIZE_COST;
833 self.super_local_decl(local, local_decl)
836 /// This method duplicates code from MIR validation in an attempt to detect type mismatches due
837 /// to normalization failure.
838 fn visit_projection_elem(
841 proj_base: &[PlaceElem<'tcx>],
842 elem: PlaceElem<'tcx>,
843 context: PlaceContext,
846 if let ProjectionElem::Field(f, ty) = elem {
847 let parent = Place { local, projection: self.tcx.intern_place_elems(proj_base) };
848 let parent_ty = parent.ty(&self.callee_body.local_decls, self.tcx);
849 let check_equal = |this: &mut Self, f_ty| {
850 if !equal_up_to_regions(this.tcx, this.param_env, ty, f_ty) {
852 this.validation = Err("failed to normalize projection type");
857 let kind = match parent_ty.ty.kind() {
858 &ty::Opaque(def_id, substs) => {
859 self.tcx.bound_type_of(def_id).subst(self.tcx, substs).kind()
865 ty::Tuple(fields) => {
866 let Some(f_ty) = fields.get(f.as_usize()) else {
867 self.validation = Err("malformed MIR");
870 check_equal(self, *f_ty);
872 ty::Adt(adt_def, substs) => {
873 let var = parent_ty.variant_index.unwrap_or(VariantIdx::from_u32(0));
874 let Some(field) = adt_def.variant(var).fields.get(f.as_usize()) else {
875 self.validation = Err("malformed MIR");
878 check_equal(self, field.ty(self.tcx, substs));
880 ty::Closure(_, substs) => {
881 let substs = substs.as_closure();
882 let Some(f_ty) = substs.upvar_tys().nth(f.as_usize()) else {
883 self.validation = Err("malformed MIR");
886 check_equal(self, f_ty);
888 &ty::Generator(def_id, substs, _) => {
889 let f_ty = if let Some(var) = parent_ty.variant_index {
890 let gen_body = if def_id == self.callee_body.source.def_id() {
893 self.tcx.optimized_mir(def_id)
896 let Some(layout) = gen_body.generator_layout() else {
897 self.validation = Err("malformed MIR");
901 let Some(&local) = layout.variant_fields[var].get(f) else {
902 self.validation = Err("malformed MIR");
906 let Some(&f_ty) = layout.field_tys.get(local) else {
907 self.validation = Err("malformed MIR");
913 let Some(f_ty) = substs.as_generator().prefix_tys().nth(f.index()) else {
914 self.validation = Err("malformed MIR");
921 check_equal(self, f_ty);
923 _ => self.validation = Err("malformed MIR"),
927 self.super_projection_elem(local, proj_base, elem, context, location);
934 * Integrates blocks from the callee function into the calling function.
935 * Updates block indices, references to locals and other control flow
938 struct Integrator<'a, 'tcx> {
940 new_locals: RangeFrom<Local>,
941 new_scopes: RangeFrom<SourceScope>,
942 new_blocks: RangeFrom<BasicBlock>,
943 destination: Place<'tcx>,
944 callsite_scope: SourceScopeData<'tcx>,
945 callsite: &'a CallSite<'tcx>,
946 cleanup_block: Option<BasicBlock>,
947 in_cleanup_block: bool,
949 expn_data: LocalExpnId,
950 always_live_locals: BitSet<Local>,
953 impl Integrator<'_, '_> {
954 fn map_local(&self, local: Local) -> Local {
955 let new = if local == RETURN_PLACE {
956 self.destination.local
958 let idx = local.index() - 1;
959 if idx < self.args.len() {
962 Local::new(self.new_locals.start.index() + (idx - self.args.len()))
965 trace!("mapping local `{:?}` to `{:?}`", local, new);
969 fn map_scope(&self, scope: SourceScope) -> SourceScope {
970 let new = SourceScope::new(self.new_scopes.start.index() + scope.index());
971 trace!("mapping scope `{:?}` to `{:?}`", scope, new);
975 fn map_block(&self, block: BasicBlock) -> BasicBlock {
976 let new = BasicBlock::new(self.new_blocks.start.index() + block.index());
977 trace!("mapping block `{:?}` to `{:?}`", block, new);
981 fn map_unwind(&self, unwind: Option<BasicBlock>) -> Option<BasicBlock> {
982 if self.in_cleanup_block {
983 if unwind.is_some() {
984 bug!("cleanup on cleanup block");
990 Some(target) => Some(self.map_block(target)),
991 // Add an unwind edge to the original call's cleanup block
992 None => self.cleanup_block,
997 impl<'tcx> MutVisitor<'tcx> for Integrator<'_, 'tcx> {
998 fn tcx(&self) -> TyCtxt<'tcx> {
1002 fn visit_local(&mut self, local: &mut Local, _ctxt: PlaceContext, _location: Location) {
1003 *local = self.map_local(*local);
1006 fn visit_source_scope_data(&mut self, scope_data: &mut SourceScopeData<'tcx>) {
1007 self.super_source_scope_data(scope_data);
1008 if scope_data.parent_scope.is_none() {
1009 // Attach the outermost callee scope as a child of the callsite
1010 // scope, via the `parent_scope` and `inlined_parent_scope` chains.
1011 scope_data.parent_scope = Some(self.callsite.source_info.scope);
1012 assert_eq!(scope_data.inlined_parent_scope, None);
1013 scope_data.inlined_parent_scope = if self.callsite_scope.inlined.is_some() {
1014 Some(self.callsite.source_info.scope)
1016 self.callsite_scope.inlined_parent_scope
1019 // Mark the outermost callee scope as an inlined one.
1020 assert_eq!(scope_data.inlined, None);
1021 scope_data.inlined = Some((self.callsite.callee, self.callsite.source_info.span));
1022 } else if scope_data.inlined_parent_scope.is_none() {
1023 // Make it easy to find the scope with `inlined` set above.
1024 scope_data.inlined_parent_scope = Some(self.map_scope(OUTERMOST_SOURCE_SCOPE));
1028 fn visit_source_scope(&mut self, scope: &mut SourceScope) {
1029 *scope = self.map_scope(*scope);
1032 fn visit_span(&mut self, span: &mut Span) {
1033 // Make sure that all spans track the fact that they were inlined.
1034 *span = span.fresh_expansion(self.expn_data);
1037 fn visit_place(&mut self, place: &mut Place<'tcx>, context: PlaceContext, location: Location) {
1038 for elem in place.projection {
1039 // FIXME: Make sure that return place is not used in an indexing projection, since it
1040 // won't be rebased as it is supposed to be.
1041 assert_ne!(ProjectionElem::Index(RETURN_PLACE), elem);
1044 // If this is the `RETURN_PLACE`, we need to rebase any projections onto it.
1045 let dest_proj_len = self.destination.projection.len();
1046 if place.local == RETURN_PLACE && dest_proj_len > 0 {
1047 let mut projs = Vec::with_capacity(dest_proj_len + place.projection.len());
1048 projs.extend(self.destination.projection);
1049 projs.extend(place.projection);
1051 place.projection = self.tcx.intern_place_elems(&*projs);
1053 // Handles integrating any locals that occur in the base
1055 self.super_place(place, context, location)
1058 fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
1059 self.in_cleanup_block = data.is_cleanup;
1060 self.super_basic_block_data(block, data);
1061 self.in_cleanup_block = false;
1064 fn visit_retag(&mut self, kind: &mut RetagKind, place: &mut Place<'tcx>, loc: Location) {
1065 self.super_retag(kind, place, loc);
1067 // We have to patch all inlined retags to be aware that they are no longer
1068 // happening on function entry.
1069 if *kind == RetagKind::FnEntry {
1070 *kind = RetagKind::Default;
1074 fn visit_statement(&mut self, statement: &mut Statement<'tcx>, location: Location) {
1075 if let StatementKind::StorageLive(local) | StatementKind::StorageDead(local) =
1078 self.always_live_locals.remove(local);
1080 self.super_statement(statement, location);
1083 fn visit_terminator(&mut self, terminator: &mut Terminator<'tcx>, loc: Location) {
1084 // Don't try to modify the implicit `_0` access on return (`return` terminators are
1085 // replaced down below anyways).
1086 if !matches!(terminator.kind, TerminatorKind::Return) {
1087 self.super_terminator(terminator, loc);
1090 match terminator.kind {
1091 TerminatorKind::GeneratorDrop | TerminatorKind::Yield { .. } => bug!(),
1092 TerminatorKind::Goto { ref mut target } => {
1093 *target = self.map_block(*target);
1095 TerminatorKind::SwitchInt { ref mut targets, .. } => {
1096 for tgt in targets.all_targets_mut() {
1097 *tgt = self.map_block(*tgt);
1100 TerminatorKind::Drop { ref mut target, ref mut unwind, .. }
1101 | TerminatorKind::DropAndReplace { ref mut target, ref mut unwind, .. } => {
1102 *target = self.map_block(*target);
1103 *unwind = self.map_unwind(*unwind);
1105 TerminatorKind::Call { ref mut target, ref mut cleanup, .. } => {
1106 if let Some(ref mut tgt) = *target {
1107 *tgt = self.map_block(*tgt);
1109 *cleanup = self.map_unwind(*cleanup);
1111 TerminatorKind::Assert { ref mut target, ref mut cleanup, .. } => {
1112 *target = self.map_block(*target);
1113 *cleanup = self.map_unwind(*cleanup);
1115 TerminatorKind::Return => {
1116 terminator.kind = if let Some(tgt) = self.callsite.target {
1117 TerminatorKind::Goto { target: tgt }
1119 TerminatorKind::Unreachable
1122 TerminatorKind::Resume => {
1123 if let Some(tgt) = self.cleanup_block {
1124 terminator.kind = TerminatorKind::Goto { target: tgt }
1127 TerminatorKind::Abort => {}
1128 TerminatorKind::Unreachable => {}
1129 TerminatorKind::FalseEdge { ref mut real_target, ref mut imaginary_target } => {
1130 *real_target = self.map_block(*real_target);
1131 *imaginary_target = self.map_block(*imaginary_target);
1133 TerminatorKind::FalseUnwind { real_target: _, unwind: _ } =>
1134 // see the ordering of passes in the optimized_mir query.
1136 bug!("False unwinds should have been removed before inlining")
1138 TerminatorKind::InlineAsm { ref mut destination, ref mut cleanup, .. } => {
1139 if let Some(ref mut tgt) = *destination {
1140 *tgt = self.map_block(*tgt);
1142 *cleanup = self.map_unwind(*cleanup);