1 //! Inlining pass for MIR functions
2 use crate::deref_separator::deref_finder;
3 use rustc_attr::InlineAttr;
4 use rustc_index::bit_set::BitSet;
5 use rustc_index::vec::Idx;
6 use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs};
7 use rustc_middle::mir::visit::*;
8 use rustc_middle::mir::*;
9 use rustc_middle::ty::{self, Instance, InstanceDef, ParamEnv, Ty, TyCtxt};
10 use rustc_session::config::OptLevel;
11 use rustc_span::def_id::DefId;
12 use rustc_span::{hygiene::ExpnKind, ExpnData, LocalExpnId, Span};
13 use rustc_target::abi::VariantIdx;
14 use rustc_target::spec::abi::Abi;
16 use crate::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 !util::is_subtype(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 !util::is_subtype(self.tcx, self.param_env, input_type, arg_ty) {
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 !util::is_subtype(self.tcx, self.param_env, input_type, arg_ty) {
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 // Two functions are compatible if the callee has no attribute (meaning
379 // that it's codegen agnostic), or sets an attribute that is identical
380 // to this function's attribute.
381 if callee_attrs.instruction_set.is_some()
382 && callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
384 return Err("incompatible instruction set");
387 for feature in &callee_attrs.target_features {
388 if !self.codegen_fn_attrs.target_features.contains(feature) {
389 return Err("incompatible target feature");
396 /// Returns inlining decision that is based on the examination of callee MIR body.
397 /// Assumes that codegen attributes have been checked for compatibility already.
398 #[instrument(level = "debug", skip(self, callee_body))]
401 callsite: &CallSite<'tcx>,
402 callee_body: &Body<'tcx>,
403 callee_attrs: &CodegenFnAttrs,
404 ) -> Result<(), &'static str> {
407 let mut threshold = if callee_attrs.requests_inline() {
408 self.tcx.sess.opts.unstable_opts.inline_mir_hint_threshold.unwrap_or(100)
410 self.tcx.sess.opts.unstable_opts.inline_mir_threshold.unwrap_or(50)
413 // Give a bonus functions with a small number of blocks,
414 // We normally have two or three blocks for even
415 // very small functions.
416 if callee_body.basic_blocks.len() <= 3 {
417 threshold += threshold / 4;
419 debug!(" final inline threshold = {}", threshold);
421 // FIXME: Give a bonus to functions with only a single caller
422 let diverges = matches!(
423 callee_body.basic_blocks[START_BLOCK].terminator().kind,
424 TerminatorKind::Unreachable | TerminatorKind::Call { target: None, .. }
426 if diverges && !matches!(callee_attrs.inline, InlineAttr::Always) {
427 return Err("callee diverges unconditionally");
430 let mut checker = CostChecker {
432 param_env: self.param_env,
433 instance: callsite.callee,
439 // Traverse the MIR manually so we can account for the effects of inlining on the CFG.
440 let mut work_list = vec![START_BLOCK];
441 let mut visited = BitSet::new_empty(callee_body.basic_blocks.len());
442 while let Some(bb) = work_list.pop() {
443 if !visited.insert(bb.index()) {
447 let blk = &callee_body.basic_blocks[bb];
448 checker.visit_basic_block_data(bb, blk);
450 let term = blk.terminator();
451 if let TerminatorKind::Drop { ref place, target, unwind }
452 | TerminatorKind::DropAndReplace { ref place, target, unwind, .. } = term.kind
454 work_list.push(target);
456 // If the place doesn't actually need dropping, treat it like a regular goto.
457 let ty = callsite.callee.subst_mir(self.tcx, &place.ty(callee_body, tcx).ty);
458 if ty.needs_drop(tcx, self.param_env) && let Some(unwind) = unwind {
459 work_list.push(unwind);
461 } else if callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
462 && matches!(term.kind, TerminatorKind::InlineAsm { .. })
464 // During the attribute checking stage we allow a callee with no
465 // instruction_set assigned to count as compatible with a function that does
466 // assign one. However, during this stage we require an exact match when any
467 // inline-asm is detected. LLVM will still possibly do an inline later on
468 // if the no-attribute function ends up with the same instruction set anyway.
469 return Err("Cannot move inline-asm across instruction sets");
471 work_list.extend(term.successors())
475 // Count up the cost of local variables and temps, if we know the size
476 // use that, otherwise we use a moderately-large dummy cost.
477 for v in callee_body.vars_and_temps_iter() {
478 checker.visit_local_decl(v, &callee_body.local_decls[v]);
481 // Abort if type validation found anything fishy.
484 let cost = checker.cost;
485 if let InlineAttr::Always = callee_attrs.inline {
486 debug!("INLINING {:?} because inline(always) [cost={}]", callsite, cost);
488 } else if cost <= threshold {
489 debug!("INLINING {:?} [cost={} <= threshold={}]", callsite, cost, threshold);
492 debug!("NOT inlining {:?} [cost={} > threshold={}]", callsite, cost, threshold);
493 Err("cost above threshold")
499 caller_body: &mut Body<'tcx>,
500 callsite: &CallSite<'tcx>,
501 mut callee_body: Body<'tcx>,
503 let terminator = caller_body[callsite.block].terminator.take().unwrap();
504 match terminator.kind {
505 TerminatorKind::Call { args, destination, cleanup, .. } => {
506 // If the call is something like `a[*i] = f(i)`, where
507 // `i : &mut usize`, then just duplicating the `a[*i]`
508 // Place could result in two different locations if `f`
509 // writes to `i`. To prevent this we need to create a temporary
510 // borrow of the place and pass the destination as `*temp` instead.
511 fn dest_needs_borrow(place: Place<'_>) -> bool {
512 for elem in place.projection.iter() {
514 ProjectionElem::Deref | ProjectionElem::Index(_) => return true,
522 let dest = if dest_needs_borrow(destination) {
523 trace!("creating temp for return destination");
524 let dest = Rvalue::Ref(
525 self.tcx.lifetimes.re_erased,
526 BorrowKind::Mut { allow_two_phase_borrow: false },
529 let dest_ty = dest.ty(caller_body, self.tcx);
530 let temp = Place::from(self.new_call_temp(caller_body, &callsite, dest_ty));
531 caller_body[callsite.block].statements.push(Statement {
532 source_info: callsite.source_info,
533 kind: StatementKind::Assign(Box::new((temp, dest))),
535 self.tcx.mk_place_deref(temp)
540 // Copy the arguments if needed.
541 let args: Vec<_> = self.make_call_args(args, &callsite, caller_body, &callee_body);
543 let mut expn_data = ExpnData::default(
545 callsite.source_info.span,
546 self.tcx.sess.edition(),
550 expn_data.def_site = callee_body.span;
552 self.tcx.with_stable_hashing_context(|hcx| LocalExpnId::fresh(expn_data, hcx));
553 let mut integrator = Integrator {
555 new_locals: Local::new(caller_body.local_decls.len())..,
556 new_scopes: SourceScope::new(caller_body.source_scopes.len())..,
557 new_blocks: BasicBlock::new(caller_body.basic_blocks.len())..,
559 callsite_scope: caller_body.source_scopes[callsite.source_info.scope].clone(),
561 cleanup_block: cleanup,
562 in_cleanup_block: false,
565 always_live_locals: BitSet::new_filled(callee_body.local_decls.len()),
568 // Map all `Local`s, `SourceScope`s and `BasicBlock`s to new ones
569 // (or existing ones, in a few special cases) in the caller.
570 integrator.visit_body(&mut callee_body);
572 // If there are any locals without storage markers, give them storage only for the
573 // duration of the call.
574 for local in callee_body.vars_and_temps_iter() {
575 if !callee_body.local_decls[local].internal
576 && integrator.always_live_locals.contains(local)
578 let new_local = integrator.map_local(local);
579 caller_body[callsite.block].statements.push(Statement {
580 source_info: callsite.source_info,
581 kind: StatementKind::StorageLive(new_local),
585 if let Some(block) = callsite.target {
586 // To avoid repeated O(n) insert, push any new statements to the end and rotate
589 for local in callee_body.vars_and_temps_iter().rev() {
590 if !callee_body.local_decls[local].internal
591 && integrator.always_live_locals.contains(local)
593 let new_local = integrator.map_local(local);
594 caller_body[block].statements.push(Statement {
595 source_info: callsite.source_info,
596 kind: StatementKind::StorageDead(new_local),
601 caller_body[block].statements.rotate_right(n);
604 // Insert all of the (mapped) parts of the callee body into the caller.
605 caller_body.local_decls.extend(callee_body.drain_vars_and_temps());
606 caller_body.source_scopes.extend(&mut callee_body.source_scopes.drain(..));
607 caller_body.var_debug_info.append(&mut callee_body.var_debug_info);
608 caller_body.basic_blocks_mut().extend(callee_body.basic_blocks_mut().drain(..));
610 caller_body[callsite.block].terminator = Some(Terminator {
611 source_info: callsite.source_info,
612 kind: TerminatorKind::Goto { target: integrator.map_block(START_BLOCK) },
615 // Copy only unevaluated constants from the callee_body into the caller_body.
616 // Although we are only pushing `ConstKind::Unevaluated` consts to
617 // `required_consts`, here we may not only have `ConstKind::Unevaluated`
618 // because we are calling `subst_and_normalize_erasing_regions`.
619 caller_body.required_consts.extend(
620 callee_body.required_consts.iter().copied().filter(|&ct| match ct.literal {
621 ConstantKind::Ty(_) => {
622 bug!("should never encounter ty::UnevaluatedConst in `required_consts`")
624 ConstantKind::Val(..) | ConstantKind::Unevaluated(..) => true,
628 kind => bug!("unexpected terminator kind {:?}", kind),
634 args: Vec<Operand<'tcx>>,
635 callsite: &CallSite<'tcx>,
636 caller_body: &mut Body<'tcx>,
637 callee_body: &Body<'tcx>,
641 // There is a bit of a mismatch between the *caller* of a closure and the *callee*.
642 // The caller provides the arguments wrapped up in a tuple:
644 // tuple_tmp = (a, b, c)
645 // Fn::call(closure_ref, tuple_tmp)
647 // meanwhile the closure body expects the arguments (here, `a`, `b`, and `c`)
648 // as distinct arguments. (This is the "rust-call" ABI hack.) Normally, codegen has
649 // the job of unpacking this tuple. But here, we are codegen. =) So we want to create
652 // [closure_ref, tuple_tmp.0, tuple_tmp.1, tuple_tmp.2]
654 // Except for one tiny wrinkle: we don't actually want `tuple_tmp.0`. It's more convenient
655 // if we "spill" that into *another* temporary, so that we can map the argument
656 // variable in the callee MIR directly to an argument variable on our side.
657 // So we introduce temporaries like:
659 // tmp0 = tuple_tmp.0
660 // tmp1 = tuple_tmp.1
661 // tmp2 = tuple_tmp.2
663 // and the vector is `[closure_ref, tmp0, tmp1, tmp2]`.
664 if callsite.fn_sig.abi() == Abi::RustCall && callee_body.spread_arg.is_none() {
665 let mut args = args.into_iter();
666 let self_ = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
667 let tuple = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
668 assert!(args.next().is_none());
670 let tuple = Place::from(tuple);
671 let ty::Tuple(tuple_tys) = tuple.ty(caller_body, tcx).ty.kind() else {
672 bug!("Closure arguments are not passed as a tuple");
675 // The `closure_ref` in our example above.
676 let closure_ref_arg = iter::once(self_);
678 // The `tmp0`, `tmp1`, and `tmp2` in our example above.
679 let tuple_tmp_args = tuple_tys.iter().enumerate().map(|(i, ty)| {
680 // This is e.g., `tuple_tmp.0` in our example above.
681 let tuple_field = Operand::Move(tcx.mk_place_field(tuple, Field::new(i), ty));
683 // Spill to a local to make e.g., `tmp0`.
684 self.create_temp_if_necessary(tuple_field, callsite, caller_body)
687 closure_ref_arg.chain(tuple_tmp_args).collect()
690 .map(|a| self.create_temp_if_necessary(a, callsite, caller_body))
695 /// If `arg` is already a temporary, returns it. Otherwise, introduces a fresh
696 /// temporary `T` and an instruction `T = arg`, and returns `T`.
697 fn create_temp_if_necessary(
700 callsite: &CallSite<'tcx>,
701 caller_body: &mut Body<'tcx>,
703 // Reuse the operand if it is a moved temporary.
704 if let Operand::Move(place) = &arg
705 && let Some(local) = place.as_local()
706 && caller_body.local_kind(local) == LocalKind::Temp
711 // Otherwise, create a temporary for the argument.
712 trace!("creating temp for argument {:?}", arg);
713 let arg_ty = arg.ty(caller_body, self.tcx);
714 let local = self.new_call_temp(caller_body, callsite, arg_ty);
715 caller_body[callsite.block].statements.push(Statement {
716 source_info: callsite.source_info,
717 kind: StatementKind::Assign(Box::new((Place::from(local), Rvalue::Use(arg)))),
722 /// Introduces a new temporary into the caller body that is live for the duration of the call.
725 caller_body: &mut Body<'tcx>,
726 callsite: &CallSite<'tcx>,
729 let local = caller_body.local_decls.push(LocalDecl::new(ty, callsite.source_info.span));
731 caller_body[callsite.block].statements.push(Statement {
732 source_info: callsite.source_info,
733 kind: StatementKind::StorageLive(local),
736 if let Some(block) = callsite.target {
737 caller_body[block].statements.insert(
740 source_info: callsite.source_info,
741 kind: StatementKind::StorageDead(local),
750 fn type_size_of<'tcx>(
752 param_env: ty::ParamEnv<'tcx>,
755 tcx.layout_of(param_env.and(ty)).ok().map(|layout| layout.size.bytes())
758 /// Verify that the callee body is compatible with the caller.
760 /// This visitor mostly computes the inlining cost,
761 /// but also needs to verify that types match because of normalization failure.
762 struct CostChecker<'b, 'tcx> {
764 param_env: ParamEnv<'tcx>,
766 callee_body: &'b Body<'tcx>,
767 instance: ty::Instance<'tcx>,
768 validation: Result<(), &'static str>,
771 impl<'tcx> Visitor<'tcx> for CostChecker<'_, 'tcx> {
772 fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) {
773 // Don't count StorageLive/StorageDead in the inlining cost.
774 match statement.kind {
775 StatementKind::StorageLive(_)
776 | StatementKind::StorageDead(_)
777 | StatementKind::Deinit(_)
778 | StatementKind::Nop => {}
779 _ => self.cost += INSTR_COST,
782 self.super_statement(statement, location);
785 fn visit_terminator(&mut self, terminator: &Terminator<'tcx>, location: Location) {
787 match terminator.kind {
788 TerminatorKind::Drop { ref place, unwind, .. }
789 | TerminatorKind::DropAndReplace { ref place, unwind, .. } => {
790 // If the place doesn't actually need dropping, treat it like a regular goto.
791 let ty = self.instance.subst_mir(tcx, &place.ty(self.callee_body, tcx).ty);
792 if ty.needs_drop(tcx, self.param_env) {
793 self.cost += CALL_PENALTY;
794 if unwind.is_some() {
795 self.cost += LANDINGPAD_PENALTY;
798 self.cost += INSTR_COST;
801 TerminatorKind::Call { func: Operand::Constant(ref f), cleanup, .. } => {
802 let fn_ty = self.instance.subst_mir(tcx, &f.literal.ty());
803 self.cost += if let ty::FnDef(def_id, _) = *fn_ty.kind() && tcx.is_intrinsic(def_id) {
804 // Don't give intrinsics the extra penalty for calls
809 if cleanup.is_some() {
810 self.cost += LANDINGPAD_PENALTY;
813 TerminatorKind::Assert { cleanup, .. } => {
814 self.cost += CALL_PENALTY;
815 if cleanup.is_some() {
816 self.cost += LANDINGPAD_PENALTY;
819 TerminatorKind::Resume => self.cost += RESUME_PENALTY,
820 TerminatorKind::InlineAsm { cleanup, .. } => {
821 self.cost += INSTR_COST;
822 if cleanup.is_some() {
823 self.cost += LANDINGPAD_PENALTY;
826 _ => self.cost += INSTR_COST,
829 self.super_terminator(terminator, location);
832 /// Count up the cost of local variables and temps, if we know the size
833 /// use that, otherwise we use a moderately-large dummy cost.
834 fn visit_local_decl(&mut self, local: Local, local_decl: &LocalDecl<'tcx>) {
836 let ptr_size = tcx.data_layout.pointer_size.bytes();
838 let ty = self.instance.subst_mir(tcx, &local_decl.ty);
839 // Cost of the var is the size in machine-words, if we know
841 if let Some(size) = type_size_of(tcx, self.param_env, ty) {
842 self.cost += ((size + ptr_size - 1) / ptr_size) as usize;
844 self.cost += UNKNOWN_SIZE_COST;
847 self.super_local_decl(local, local_decl)
850 /// This method duplicates code from MIR validation in an attempt to detect type mismatches due
851 /// to normalization failure.
852 fn visit_projection_elem(
855 proj_base: &[PlaceElem<'tcx>],
856 elem: PlaceElem<'tcx>,
857 context: PlaceContext,
860 if let ProjectionElem::Field(f, ty) = elem {
861 let parent = Place { local, projection: self.tcx.intern_place_elems(proj_base) };
862 let parent_ty = parent.ty(&self.callee_body.local_decls, self.tcx);
863 let check_equal = |this: &mut Self, f_ty| {
864 if !util::is_equal_up_to_subtyping(this.tcx, this.param_env, ty, f_ty) {
866 this.validation = Err("failed to normalize projection type");
871 let kind = match parent_ty.ty.kind() {
872 &ty::Opaque(def_id, substs) => {
873 self.tcx.bound_type_of(def_id).subst(self.tcx, substs).kind()
879 ty::Tuple(fields) => {
880 let Some(f_ty) = fields.get(f.as_usize()) else {
881 self.validation = Err("malformed MIR");
884 check_equal(self, *f_ty);
886 ty::Adt(adt_def, substs) => {
887 let var = parent_ty.variant_index.unwrap_or(VariantIdx::from_u32(0));
888 let Some(field) = adt_def.variant(var).fields.get(f.as_usize()) else {
889 self.validation = Err("malformed MIR");
892 check_equal(self, field.ty(self.tcx, substs));
894 ty::Closure(_, substs) => {
895 let substs = substs.as_closure();
896 let Some(f_ty) = substs.upvar_tys().nth(f.as_usize()) else {
897 self.validation = Err("malformed MIR");
900 check_equal(self, f_ty);
902 &ty::Generator(def_id, substs, _) => {
903 let f_ty = if let Some(var) = parent_ty.variant_index {
904 let gen_body = if def_id == self.callee_body.source.def_id() {
907 self.tcx.optimized_mir(def_id)
910 let Some(layout) = gen_body.generator_layout() else {
911 self.validation = Err("malformed MIR");
915 let Some(&local) = layout.variant_fields[var].get(f) else {
916 self.validation = Err("malformed MIR");
920 let Some(&f_ty) = layout.field_tys.get(local) else {
921 self.validation = Err("malformed MIR");
927 let Some(f_ty) = substs.as_generator().prefix_tys().nth(f.index()) else {
928 self.validation = Err("malformed MIR");
935 check_equal(self, f_ty);
937 _ => self.validation = Err("malformed MIR"),
941 self.super_projection_elem(local, proj_base, elem, context, location);
948 * Integrates blocks from the callee function into the calling function.
949 * Updates block indices, references to locals and other control flow
952 struct Integrator<'a, 'tcx> {
954 new_locals: RangeFrom<Local>,
955 new_scopes: RangeFrom<SourceScope>,
956 new_blocks: RangeFrom<BasicBlock>,
957 destination: Place<'tcx>,
958 callsite_scope: SourceScopeData<'tcx>,
959 callsite: &'a CallSite<'tcx>,
960 cleanup_block: Option<BasicBlock>,
961 in_cleanup_block: bool,
963 expn_data: LocalExpnId,
964 always_live_locals: BitSet<Local>,
967 impl Integrator<'_, '_> {
968 fn map_local(&self, local: Local) -> Local {
969 let new = if local == RETURN_PLACE {
970 self.destination.local
972 let idx = local.index() - 1;
973 if idx < self.args.len() {
976 Local::new(self.new_locals.start.index() + (idx - self.args.len()))
979 trace!("mapping local `{:?}` to `{:?}`", local, new);
983 fn map_scope(&self, scope: SourceScope) -> SourceScope {
984 let new = SourceScope::new(self.new_scopes.start.index() + scope.index());
985 trace!("mapping scope `{:?}` to `{:?}`", scope, new);
989 fn map_block(&self, block: BasicBlock) -> BasicBlock {
990 let new = BasicBlock::new(self.new_blocks.start.index() + block.index());
991 trace!("mapping block `{:?}` to `{:?}`", block, new);
995 fn map_unwind(&self, unwind: Option<BasicBlock>) -> Option<BasicBlock> {
996 if self.in_cleanup_block {
997 if unwind.is_some() {
998 bug!("cleanup on cleanup block");
1004 Some(target) => Some(self.map_block(target)),
1005 // Add an unwind edge to the original call's cleanup block
1006 None => self.cleanup_block,
1011 impl<'tcx> MutVisitor<'tcx> for Integrator<'_, 'tcx> {
1012 fn tcx(&self) -> TyCtxt<'tcx> {
1016 fn visit_local(&mut self, local: &mut Local, _ctxt: PlaceContext, _location: Location) {
1017 *local = self.map_local(*local);
1020 fn visit_source_scope_data(&mut self, scope_data: &mut SourceScopeData<'tcx>) {
1021 self.super_source_scope_data(scope_data);
1022 if scope_data.parent_scope.is_none() {
1023 // Attach the outermost callee scope as a child of the callsite
1024 // scope, via the `parent_scope` and `inlined_parent_scope` chains.
1025 scope_data.parent_scope = Some(self.callsite.source_info.scope);
1026 assert_eq!(scope_data.inlined_parent_scope, None);
1027 scope_data.inlined_parent_scope = if self.callsite_scope.inlined.is_some() {
1028 Some(self.callsite.source_info.scope)
1030 self.callsite_scope.inlined_parent_scope
1033 // Mark the outermost callee scope as an inlined one.
1034 assert_eq!(scope_data.inlined, None);
1035 scope_data.inlined = Some((self.callsite.callee, self.callsite.source_info.span));
1036 } else if scope_data.inlined_parent_scope.is_none() {
1037 // Make it easy to find the scope with `inlined` set above.
1038 scope_data.inlined_parent_scope = Some(self.map_scope(OUTERMOST_SOURCE_SCOPE));
1042 fn visit_source_scope(&mut self, scope: &mut SourceScope) {
1043 *scope = self.map_scope(*scope);
1046 fn visit_span(&mut self, span: &mut Span) {
1047 // Make sure that all spans track the fact that they were inlined.
1048 *span = span.fresh_expansion(self.expn_data);
1051 fn visit_place(&mut self, place: &mut Place<'tcx>, context: PlaceContext, location: Location) {
1052 for elem in place.projection {
1053 // FIXME: Make sure that return place is not used in an indexing projection, since it
1054 // won't be rebased as it is supposed to be.
1055 assert_ne!(ProjectionElem::Index(RETURN_PLACE), elem);
1058 // If this is the `RETURN_PLACE`, we need to rebase any projections onto it.
1059 let dest_proj_len = self.destination.projection.len();
1060 if place.local == RETURN_PLACE && dest_proj_len > 0 {
1061 let mut projs = Vec::with_capacity(dest_proj_len + place.projection.len());
1062 projs.extend(self.destination.projection);
1063 projs.extend(place.projection);
1065 place.projection = self.tcx.intern_place_elems(&*projs);
1067 // Handles integrating any locals that occur in the base
1069 self.super_place(place, context, location)
1072 fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
1073 self.in_cleanup_block = data.is_cleanup;
1074 self.super_basic_block_data(block, data);
1075 self.in_cleanup_block = false;
1078 fn visit_retag(&mut self, kind: &mut RetagKind, place: &mut Place<'tcx>, loc: Location) {
1079 self.super_retag(kind, place, loc);
1081 // We have to patch all inlined retags to be aware that they are no longer
1082 // happening on function entry.
1083 if *kind == RetagKind::FnEntry {
1084 *kind = RetagKind::Default;
1088 fn visit_statement(&mut self, statement: &mut Statement<'tcx>, location: Location) {
1089 if let StatementKind::StorageLive(local) | StatementKind::StorageDead(local) =
1092 self.always_live_locals.remove(local);
1094 self.super_statement(statement, location);
1097 fn visit_terminator(&mut self, terminator: &mut Terminator<'tcx>, loc: Location) {
1098 // Don't try to modify the implicit `_0` access on return (`return` terminators are
1099 // replaced down below anyways).
1100 if !matches!(terminator.kind, TerminatorKind::Return) {
1101 self.super_terminator(terminator, loc);
1104 match terminator.kind {
1105 TerminatorKind::GeneratorDrop | TerminatorKind::Yield { .. } => bug!(),
1106 TerminatorKind::Goto { ref mut target } => {
1107 *target = self.map_block(*target);
1109 TerminatorKind::SwitchInt { ref mut targets, .. } => {
1110 for tgt in targets.all_targets_mut() {
1111 *tgt = self.map_block(*tgt);
1114 TerminatorKind::Drop { ref mut target, ref mut unwind, .. }
1115 | TerminatorKind::DropAndReplace { ref mut target, ref mut unwind, .. } => {
1116 *target = self.map_block(*target);
1117 *unwind = self.map_unwind(*unwind);
1119 TerminatorKind::Call { ref mut target, ref mut cleanup, .. } => {
1120 if let Some(ref mut tgt) = *target {
1121 *tgt = self.map_block(*tgt);
1123 *cleanup = self.map_unwind(*cleanup);
1125 TerminatorKind::Assert { ref mut target, ref mut cleanup, .. } => {
1126 *target = self.map_block(*target);
1127 *cleanup = self.map_unwind(*cleanup);
1129 TerminatorKind::Return => {
1130 terminator.kind = if let Some(tgt) = self.callsite.target {
1131 TerminatorKind::Goto { target: tgt }
1133 TerminatorKind::Unreachable
1136 TerminatorKind::Resume => {
1137 if let Some(tgt) = self.cleanup_block {
1138 terminator.kind = TerminatorKind::Goto { target: tgt }
1141 TerminatorKind::Abort => {}
1142 TerminatorKind::Unreachable => {}
1143 TerminatorKind::FalseEdge { ref mut real_target, ref mut imaginary_target } => {
1144 *real_target = self.map_block(*real_target);
1145 *imaginary_target = self.map_block(*imaginary_target);
1147 TerminatorKind::FalseUnwind { real_target: _, unwind: _ } =>
1148 // see the ordering of passes in the optimized_mir query.
1150 bug!("False unwinds should have been removed before inlining")
1152 TerminatorKind::InlineAsm { ref mut destination, ref mut cleanup, .. } => {
1153 if let Some(ref mut tgt) = *destination {
1154 *tgt = self.map_block(*tgt);
1156 *cleanup = self.map_unwind(*cleanup);