compiler/rustc_mir_transform/src/inline.rs

   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::{hygiene::ExpnKind, ExpnData, LocalExpnId, Span};
  12 use rustc_target::abi::VariantIdx;
  13 use rustc_target::spec::abi::Abi;
  14
  15 use crate::simplify::{remove_dead_blocks, CfgSimplifier};
  16 use crate::util;
  17 use crate::MirPass;
  18 use std::iter;
  19 use std::ops::{Range, RangeFrom};
  20
  21 pub(crate) mod cycle;
  22
  23 const INSTR_COST: usize = 5;
  24 const CALL_PENALTY: usize = 25;
  25 const LANDINGPAD_PENALTY: usize = 50;
  26 const RESUME_PENALTY: usize = 45;
  27
  28 const UNKNOWN_SIZE_COST: usize = 10;
  29
  30 pub struct Inline;
  31
  32 #[derive(Copy, Clone, Debug)]
  33 struct CallSite<'tcx> {
  34     callee: Instance<'tcx>,
  35     fn_sig: ty::PolyFnSig<'tcx>,
  36     block: BasicBlock,
  37     target: Option<BasicBlock>,
  38     source_info: SourceInfo,
  39 }
  40
  41 impl<'tcx> MirPass<'tcx> for Inline {
  42     fn is_enabled(&self, sess: &rustc_session::Session) -> bool {
  43         if let Some(enabled) = sess.opts.unstable_opts.inline_mir {
  44             return enabled;
  45         }
  46
  47         match sess.mir_opt_level() {
  48             0 | 1 => false,
  49             2 => {
  50                 (sess.opts.optimize == OptLevel::Default
  51                     || sess.opts.optimize == OptLevel::Aggressive)
  52                     && sess.opts.incremental == None
  53             }
  54             _ => true,
  55         }
  56     }
  57
  58     fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
  59         let span = trace_span!("inline", body = %tcx.def_path_str(body.source.def_id()));
  60         let _guard = span.enter();
  61         if inline(tcx, body) {
  62             debug!("running simplify cfg on {:?}", body.source);
  63             CfgSimplifier::new(body).simplify();
  64             remove_dead_blocks(tcx, body);
  65             deref_finder(tcx, body);
  66         }
  67     }
  68 }
  69
  70 fn inline<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) -> bool {
  71     let def_id = body.source.def_id().expect_local();
  72
  73     // Only do inlining into fn bodies.
  74     if !tcx.hir().body_owner_kind(def_id).is_fn_or_closure() {
  75         return false;
  76     }
  77     if body.source.promoted.is_some() {
  78         return false;
  79     }
  80     // Avoid inlining into generators, since their `optimized_mir` is used for layout computation,
  81     // which can create a cycle, even when no attempt is made to inline the function in the other
  82     // direction.
  83     if body.generator.is_some() {
  84         return false;
  85     }
  86
  87     let param_env = tcx.param_env_reveal_all_normalized(def_id);
  88
  89     let mut this =
  90         Inliner { tcx, param_env, codegen_fn_attrs: tcx.codegen_fn_attrs(def_id), changed: false };
  91     let blocks = BasicBlock::new(0)..body.basic_blocks.next_index();
  92     this.process_blocks(body, blocks);
  93     this.changed
  94 }
  95
  96 struct Inliner<'tcx> {
  97     tcx: TyCtxt<'tcx>,
  98     param_env: ParamEnv<'tcx>,
  99     /// Caller codegen attributes.
 100     codegen_fn_attrs: &'tcx CodegenFnAttrs,
 101     /// Indicates that the caller body has been modified.
 102     changed: bool,
 103 }
 104
 105 impl<'tcx> Inliner<'tcx> {
 106     fn process_blocks(&mut self, caller_body: &mut Body<'tcx>, blocks: Range<BasicBlock>) {
 107         for bb in blocks {
 108             let bb_data = &caller_body[bb];
 109             if bb_data.is_cleanup {
 110                 continue;
 111             }
 112
 113             let Some(callsite) = self.resolve_callsite(caller_body, bb, bb_data) else {
 114                 continue;
 115             };
 116
 117             let span = trace_span!("process_blocks", %callsite.callee, ?bb);
 118             let _guard = span.enter();
 119
 120             match self.try_inlining(caller_body, &callsite) {
 121                 Err(reason) => {
 122                     debug!("not-inlined {} [{}]", callsite.callee, reason);
 123                     continue;
 124                 }
 125                 Ok(_) => {
 126                     debug!("inlined {}", callsite.callee);
 127                     self.changed = true;
 128                     // We could process the blocks returned by `try_inlining` here. However, that
 129                     // leads to exponential compile times due to the top-down nature of this kind
 130                     // of inlining.
 131                 }
 132             }
 133         }
 134     }
 135
 136     /// Attempts to inline a callsite into the caller body. When successful returns basic blocks
 137     /// containing the inlined body. Otherwise returns an error describing why inlining didn't take
 138     /// place.
 139     fn try_inlining(
 140         &self,
 141         caller_body: &mut Body<'tcx>,
 142         callsite: &CallSite<'tcx>,
 143     ) -> Result<std::ops::Range<BasicBlock>, &'static str> {
 144         let callee_attrs = self.tcx.codegen_fn_attrs(callsite.callee.def_id());
 145         self.check_codegen_attributes(callsite, callee_attrs)?;
 146         self.check_mir_is_available(caller_body, &callsite.callee)?;
 147         let callee_body = self.tcx.instance_mir(callsite.callee.def);
 148         self.check_mir_body(callsite, callee_body, callee_attrs)?;
 149
 150         if !self.tcx.consider_optimizing(|| {
 151             format!("Inline {:?} into {:?}", callsite.callee, caller_body.source)
 152         }) {
 153             return Err("optimization fuel exhausted");
 154         }
 155
 156         let Ok(callee_body) = callsite.callee.try_subst_mir_and_normalize_erasing_regions(
 157             self.tcx,
 158             self.param_env,
 159             callee_body.clone(),
 160         ) else {
 161             return Err("failed to normalize callee body");
 162         };
 163
 164         // Check call signature compatibility.
 165         // Normally, this shouldn't be required, but trait normalization failure can create a
 166         // validation ICE.
 167         let terminator = caller_body[callsite.block].terminator.as_ref().unwrap();
 168         let TerminatorKind::Call { args, destination, .. } = &terminator.kind else { bug!() };
 169         let destination_ty = destination.ty(&caller_body.local_decls, self.tcx).ty;
 170         let output_type = callee_body.return_ty();
 171         if !util::is_subtype(self.tcx, self.param_env, output_type, destination_ty) {
 172             trace!(?output_type, ?destination_ty);
 173             return Err("failed to normalize return type");
 174         }
 175         if callsite.fn_sig.abi() == Abi::RustCall {
 176             let (arg_tuple, skipped_args) = match &args[..] {
 177                 [arg_tuple] => (arg_tuple, 0),
 178                 [_, arg_tuple] => (arg_tuple, 1),
 179                 _ => bug!("Expected `rust-call` to have 1 or 2 args"),
 180             };
 181
 182             let arg_tuple_ty = arg_tuple.ty(&caller_body.local_decls, self.tcx);
 183             let ty::Tuple(arg_tuple_tys) = arg_tuple_ty.kind() else {
 184                 bug!("Closure arguments are not passed as a tuple");
 185             };
 186
 187             for (arg_ty, input) in
 188                 arg_tuple_tys.iter().zip(callee_body.args_iter().skip(skipped_args))
 189             {
 190                 let input_type = callee_body.local_decls[input].ty;
 191                 if !util::is_subtype(self.tcx, self.param_env, input_type, arg_ty) {
 192                     trace!(?arg_ty, ?input_type);
 193                     return Err("failed to normalize tuple argument type");
 194                 }
 195             }
 196         } else {
 197             for (arg, input) in args.iter().zip(callee_body.args_iter()) {
 198                 let input_type = callee_body.local_decls[input].ty;
 199                 let arg_ty = arg.ty(&caller_body.local_decls, self.tcx);
 200                 if !util::is_subtype(self.tcx, self.param_env, input_type, arg_ty) {
 201                     trace!(?arg_ty, ?input_type);
 202                     return Err("failed to normalize argument type");
 203                 }
 204             }
 205         }
 206
 207         let old_blocks = caller_body.basic_blocks.next_index();
 208         self.inline_call(caller_body, &callsite, callee_body);
 209         let new_blocks = old_blocks..caller_body.basic_blocks.next_index();
 210
 211         Ok(new_blocks)
 212     }
 213
 214     fn check_mir_is_available(
 215         &self,
 216         caller_body: &Body<'tcx>,
 217         callee: &Instance<'tcx>,
 218     ) -> Result<(), &'static str> {
 219         let caller_def_id = caller_body.source.def_id();
 220         let callee_def_id = callee.def_id();
 221         if callee_def_id == caller_def_id {
 222             return Err("self-recursion");
 223         }
 224
 225         match callee.def {
 226             InstanceDef::Item(_) => {
 227                 // If there is no MIR available (either because it was not in metadata or
 228                 // because it has no MIR because it's an extern function), then the inliner
 229                 // won't cause cycles on this.
 230                 if !self.tcx.is_mir_available(callee_def_id) {
 231                     return Err("item MIR unavailable");
 232                 }
 233             }
 234             // These have no own callable MIR.
 235             InstanceDef::Intrinsic(_) | InstanceDef::Virtual(..) => {
 236                 return Err("instance without MIR (intrinsic / virtual)");
 237             }
 238             // This cannot result in an immediate cycle since the callee MIR is a shim, which does
 239             // not get any optimizations run on it. Any subsequent inlining may cause cycles, but we
 240             // do not need to catch this here, we can wait until the inliner decides to continue
 241             // inlining a second time.
 242             InstanceDef::VTableShim(_)
 243             | InstanceDef::ReifyShim(_)
 244             | InstanceDef::FnPtrShim(..)
 245             | InstanceDef::ClosureOnceShim { .. }
 246             | InstanceDef::DropGlue(..)
 247             | InstanceDef::CloneShim(..) => return Ok(()),
 248         }
 249
 250         if self.tcx.is_constructor(callee_def_id) {
 251             trace!("constructors always have MIR");
 252             // Constructor functions cannot cause a query cycle.
 253             return Ok(());
 254         }
 255
 256         if callee_def_id.is_local() {
 257             // Avoid a cycle here by only using `instance_mir` only if we have
 258             // a lower `DefPathHash` than the callee. This ensures that the callee will
 259             // not inline us. This trick even works with incremental compilation,
 260             // since `DefPathHash` is stable.
 261             if self.tcx.def_path_hash(caller_def_id).local_hash()
 262                 < self.tcx.def_path_hash(callee_def_id).local_hash()
 263             {
 264                 return Ok(());
 265             }
 266
 267             // If we know for sure that the function we're calling will itself try to
 268             // call us, then we avoid inlining that function.
 269             if self.tcx.mir_callgraph_reachable((*callee, caller_def_id.expect_local())) {
 270                 return Err("caller might be reachable from callee (query cycle avoidance)");
 271             }
 272
 273             Ok(())
 274         } else {
 275             // This cannot result in an immediate cycle since the callee MIR is from another crate
 276             // and is already optimized. Any subsequent inlining may cause cycles, but we do
 277             // not need to catch this here, we can wait until the inliner decides to continue
 278             // inlining a second time.
 279             trace!("functions from other crates always have MIR");
 280             Ok(())
 281         }
 282     }
 283
 284     fn resolve_callsite(
 285         &self,
 286         caller_body: &Body<'tcx>,
 287         bb: BasicBlock,
 288         bb_data: &BasicBlockData<'tcx>,
 289     ) -> Option<CallSite<'tcx>> {
 290         // Only consider direct calls to functions
 291         let terminator = bb_data.terminator();
 292         if let TerminatorKind::Call { ref func, target, .. } = terminator.kind {
 293             let func_ty = func.ty(caller_body, self.tcx);
 294             if let ty::FnDef(def_id, substs) = *func_ty.kind() {
 295                 // To resolve an instance its substs have to be fully normalized.
 296                 let substs = self.tcx.try_normalize_erasing_regions(self.param_env, substs).ok()?;
 297                 let callee =
 298                     Instance::resolve(self.tcx, self.param_env, def_id, substs).ok().flatten()?;
 299
 300                 if let InstanceDef::Virtual(..) | InstanceDef::Intrinsic(_) = callee.def {
 301                     return None;
 302                 }
 303
 304                 let fn_sig = self.tcx.bound_fn_sig(def_id).subst(self.tcx, substs);
 305
 306                 return Some(CallSite {
 307                     callee,
 308                     fn_sig,
 309                     block: bb,
 310                     target,
 311                     source_info: terminator.source_info,
 312                 });
 313             }
 314         }
 315
 316         None
 317     }
 318
 319     /// Returns an error if inlining is not possible based on codegen attributes alone. A success
 320     /// indicates that inlining decision should be based on other criteria.
 321     fn check_codegen_attributes(
 322         &self,
 323         callsite: &CallSite<'tcx>,
 324         callee_attrs: &CodegenFnAttrs,
 325     ) -> Result<(), &'static str> {
 326         match callee_attrs.inline {
 327             InlineAttr::Never => return Err("never inline hint"),
 328             InlineAttr::Always | InlineAttr::Hint => {}
 329             InlineAttr::None => {
 330                 if self.tcx.sess.mir_opt_level() <= 2 {
 331                     return Err("at mir-opt-level=2, only #[inline] is inlined");
 332                 }
 333             }
 334         }
 335
 336         // Only inline local functions if they would be eligible for cross-crate
 337         // inlining. This is to ensure that the final crate doesn't have MIR that
 338         // reference unexported symbols
 339         if callsite.callee.def_id().is_local() {
 340             let is_generic = callsite.callee.substs.non_erasable_generics().next().is_some();
 341             if !is_generic && !callee_attrs.requests_inline() {
 342                 return Err("not exported");
 343             }
 344         }
 345
 346         if callsite.fn_sig.c_variadic() {
 347             return Err("C variadic");
 348         }
 349
 350         if callee_attrs.flags.contains(CodegenFnAttrFlags::COLD) {
 351             return Err("cold");
 352         }
 353
 354         if callee_attrs.no_sanitize != self.codegen_fn_attrs.no_sanitize {
 355             return Err("incompatible sanitizer set");
 356         }
 357
 358         // Two functions are compatible if the callee has no attribute (meaning
 359         // that it's codegen agnostic), or sets an attribute that is identical
 360         // to this function's attribute.
 361         if callee_attrs.instruction_set.is_some()
 362             && callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
 363         {
 364             return Err("incompatible instruction set");
 365         }
 366
 367         for feature in &callee_attrs.target_features {
 368             if !self.codegen_fn_attrs.target_features.contains(feature) {
 369                 return Err("incompatible target feature");
 370             }
 371         }
 372
 373         Ok(())
 374     }
 375
 376     /// Returns inlining decision that is based on the examination of callee MIR body.
 377     /// Assumes that codegen attributes have been checked for compatibility already.
 378     #[instrument(level = "debug", skip(self, callee_body))]
 379     fn check_mir_body(
 380         &self,
 381         callsite: &CallSite<'tcx>,
 382         callee_body: &Body<'tcx>,
 383         callee_attrs: &CodegenFnAttrs,
 384     ) -> Result<(), &'static str> {
 385         let tcx = self.tcx;
 386
 387         let mut threshold = if callee_attrs.requests_inline() {
 388             self.tcx.sess.opts.unstable_opts.inline_mir_hint_threshold.unwrap_or(100)
 389         } else {
 390             self.tcx.sess.opts.unstable_opts.inline_mir_threshold.unwrap_or(50)
 391         };
 392
 393         // Give a bonus functions with a small number of blocks,
 394         // We normally have two or three blocks for even
 395         // very small functions.
 396         if callee_body.basic_blocks.len() <= 3 {
 397             threshold += threshold / 4;
 398         }
 399         debug!("    final inline threshold = {}", threshold);
 400
 401         // FIXME: Give a bonus to functions with only a single caller
 402         let diverges = matches!(
 403             callee_body.basic_blocks[START_BLOCK].terminator().kind,
 404             TerminatorKind::Unreachable | TerminatorKind::Call { target: None, .. }
 405         );
 406         if diverges && !matches!(callee_attrs.inline, InlineAttr::Always) {
 407             return Err("callee diverges unconditionally");
 408         }
 409
 410         let mut checker = CostChecker {
 411             tcx: self.tcx,
 412             param_env: self.param_env,
 413             instance: callsite.callee,
 414             callee_body,
 415             cost: 0,
 416             validation: Ok(()),
 417         };
 418
 419         // Traverse the MIR manually so we can account for the effects of inlining on the CFG.
 420         let mut work_list = vec![START_BLOCK];
 421         let mut visited = BitSet::new_empty(callee_body.basic_blocks.len());
 422         while let Some(bb) = work_list.pop() {
 423             if !visited.insert(bb.index()) {
 424                 continue;
 425             }
 426
 427             let blk = &callee_body.basic_blocks[bb];
 428             checker.visit_basic_block_data(bb, blk);
 429
 430             let term = blk.terminator();
 431             if let TerminatorKind::Drop { ref place, target, unwind }
 432             | TerminatorKind::DropAndReplace { ref place, target, unwind, .. } = term.kind
 433             {
 434                 work_list.push(target);
 435
 436                 // If the place doesn't actually need dropping, treat it like a regular goto.
 437                 let ty = callsite.callee.subst_mir(self.tcx, &place.ty(callee_body, tcx).ty);
 438                 if ty.needs_drop(tcx, self.param_env) && let Some(unwind) = unwind {
 439                         work_list.push(unwind);
 440                     }
 441             } else if callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
 442                 && matches!(term.kind, TerminatorKind::InlineAsm { .. })
 443             {
 444                 // During the attribute checking stage we allow a callee with no
 445                 // instruction_set assigned to count as compatible with a function that does
 446                 // assign one. However, during this stage we require an exact match when any
 447                 // inline-asm is detected. LLVM will still possibly do an inline later on
 448                 // if the no-attribute function ends up with the same instruction set anyway.
 449                 return Err("Cannot move inline-asm across instruction sets");
 450             } else {
 451                 work_list.extend(term.successors())
 452             }
 453         }
 454
 455         // Count up the cost of local variables and temps, if we know the size
 456         // use that, otherwise we use a moderately-large dummy cost.
 457         for v in callee_body.vars_and_temps_iter() {
 458             checker.visit_local_decl(v, &callee_body.local_decls[v]);
 459         }
 460
 461         // Abort if type validation found anything fishy.
 462         checker.validation?;
 463
 464         let cost = checker.cost;
 465         if let InlineAttr::Always = callee_attrs.inline {
 466             debug!("INLINING {:?} because inline(always) [cost={}]", callsite, cost);
 467             Ok(())
 468         } else if cost <= threshold {
 469             debug!("INLINING {:?} [cost={} <= threshold={}]", callsite, cost, threshold);
 470             Ok(())
 471         } else {
 472             debug!("NOT inlining {:?} [cost={} > threshold={}]", callsite, cost, threshold);
 473             Err("cost above threshold")
 474         }
 475     }
 476
 477     fn inline_call(
 478         &self,
 479         caller_body: &mut Body<'tcx>,
 480         callsite: &CallSite<'tcx>,
 481         mut callee_body: Body<'tcx>,
 482     ) {
 483         let terminator = caller_body[callsite.block].terminator.take().unwrap();
 484         match terminator.kind {
 485             TerminatorKind::Call { args, destination, cleanup, .. } => {
 486                 // If the call is something like `a[*i] = f(i)`, where
 487                 // `i : &mut usize`, then just duplicating the `a[*i]`
 488                 // Place could result in two different locations if `f`
 489                 // writes to `i`. To prevent this we need to create a temporary
 490                 // borrow of the place and pass the destination as `*temp` instead.
 491                 fn dest_needs_borrow(place: Place<'_>) -> bool {
 492                     for elem in place.projection.iter() {
 493                         match elem {
 494                             ProjectionElem::Deref | ProjectionElem::Index(_) => return true,
 495                             _ => {}
 496                         }
 497                     }
 498
 499                     false
 500                 }
 501
 502                 let dest = if dest_needs_borrow(destination) {
 503                     trace!("creating temp for return destination");
 504                     let dest = Rvalue::Ref(
 505                         self.tcx.lifetimes.re_erased,
 506                         BorrowKind::Mut { allow_two_phase_borrow: false },
 507                         destination,
 508                     );
 509                     let dest_ty = dest.ty(caller_body, self.tcx);
 510                     let temp = Place::from(self.new_call_temp(caller_body, &callsite, dest_ty));
 511                     caller_body[callsite.block].statements.push(Statement {
 512                         source_info: callsite.source_info,
 513                         kind: StatementKind::Assign(Box::new((temp, dest))),
 514                     });
 515                     self.tcx.mk_place_deref(temp)
 516                 } else {
 517                     destination
 518                 };
 519
 520                 // Copy the arguments if needed.
 521                 let args: Vec<_> = self.make_call_args(args, &callsite, caller_body, &callee_body);
 522
 523                 let mut expn_data = ExpnData::default(
 524                     ExpnKind::Inlined,
 525                     callsite.source_info.span,
 526                     self.tcx.sess.edition(),
 527                     None,
 528                     None,
 529                 );
 530                 expn_data.def_site = callee_body.span;
 531                 let expn_data =
 532                     self.tcx.with_stable_hashing_context(|hcx| LocalExpnId::fresh(expn_data, hcx));
 533                 let mut integrator = Integrator {
 534                     args: &args,
 535                     new_locals: Local::new(caller_body.local_decls.len())..,
 536                     new_scopes: SourceScope::new(caller_body.source_scopes.len())..,
 537                     new_blocks: BasicBlock::new(caller_body.basic_blocks.len())..,
 538                     destination: dest,
 539                     callsite_scope: caller_body.source_scopes[callsite.source_info.scope].clone(),
 540                     callsite,
 541                     cleanup_block: cleanup,
 542                     in_cleanup_block: false,
 543                     tcx: self.tcx,
 544                     expn_data,
 545                     always_live_locals: BitSet::new_filled(callee_body.local_decls.len()),
 546                 };
 547
 548                 // Map all `Local`s, `SourceScope`s and `BasicBlock`s to new ones
 549                 // (or existing ones, in a few special cases) in the caller.
 550                 integrator.visit_body(&mut callee_body);
 551
 552                 // If there are any locals without storage markers, give them storage only for the
 553                 // duration of the call.
 554                 for local in callee_body.vars_and_temps_iter() {
 555                     if !callee_body.local_decls[local].internal
 556                         && integrator.always_live_locals.contains(local)
 557                     {
 558                         let new_local = integrator.map_local(local);
 559                         caller_body[callsite.block].statements.push(Statement {
 560                             source_info: callsite.source_info,
 561                             kind: StatementKind::StorageLive(new_local),
 562                         });
 563                     }
 564                 }
 565                 if let Some(block) = callsite.target {
 566                     // To avoid repeated O(n) insert, push any new statements to the end and rotate
 567                     // the slice once.
 568                     let mut n = 0;
 569                     for local in callee_body.vars_and_temps_iter().rev() {
 570                         if !callee_body.local_decls[local].internal
 571                             && integrator.always_live_locals.contains(local)
 572                         {
 573                             let new_local = integrator.map_local(local);
 574                             caller_body[block].statements.push(Statement {
 575                                 source_info: callsite.source_info,
 576                                 kind: StatementKind::StorageDead(new_local),
 577                             });
 578                             n += 1;
 579                         }
 580                     }
 581                     caller_body[block].statements.rotate_right(n);
 582                 }
 583
 584                 // Insert all of the (mapped) parts of the callee body into the caller.
 585                 caller_body.local_decls.extend(callee_body.drain_vars_and_temps());
 586                 caller_body.source_scopes.extend(&mut callee_body.source_scopes.drain(..));
 587                 caller_body.var_debug_info.append(&mut callee_body.var_debug_info);
 588                 caller_body.basic_blocks_mut().extend(callee_body.basic_blocks_mut().drain(..));
 589
 590                 caller_body[callsite.block].terminator = Some(Terminator {
 591                     source_info: callsite.source_info,
 592                     kind: TerminatorKind::Goto { target: integrator.map_block(START_BLOCK) },
 593                 });
 594
 595                 // Copy only unevaluated constants from the callee_body into the caller_body.
 596                 // Although we are only pushing `ConstKind::Unevaluated` consts to
 597                 // `required_consts`, here we may not only have `ConstKind::Unevaluated`
 598                 // because we are calling `subst_and_normalize_erasing_regions`.
 599                 caller_body.required_consts.extend(
 600                     callee_body.required_consts.iter().copied().filter(|&ct| match ct.literal {
 601                         ConstantKind::Ty(_) => {
 602                             bug!("should never encounter ty::UnevaluatedConst in `required_consts`")
 603                         }
 604                         ConstantKind::Val(..) | ConstantKind::Unevaluated(..) => true,
 605                     }),
 606                 );
 607             }
 608             kind => bug!("unexpected terminator kind {:?}", kind),
 609         }
 610     }
 611
 612     fn make_call_args(
 613         &self,
 614         args: Vec<Operand<'tcx>>,
 615         callsite: &CallSite<'tcx>,
 616         caller_body: &mut Body<'tcx>,
 617         callee_body: &Body<'tcx>,
 618     ) -> Vec<Local> {
 619         let tcx = self.tcx;
 620
 621         // There is a bit of a mismatch between the *caller* of a closure and the *callee*.
 622         // The caller provides the arguments wrapped up in a tuple:
 623         //
 624         //     tuple_tmp = (a, b, c)
 625         //     Fn::call(closure_ref, tuple_tmp)
 626         //
 627         // meanwhile the closure body expects the arguments (here, `a`, `b`, and `c`)
 628         // as distinct arguments. (This is the "rust-call" ABI hack.) Normally, codegen has
 629         // the job of unpacking this tuple. But here, we are codegen. =) So we want to create
 630         // a vector like
 631         //
 632         //     [closure_ref, tuple_tmp.0, tuple_tmp.1, tuple_tmp.2]
 633         //
 634         // Except for one tiny wrinkle: we don't actually want `tuple_tmp.0`. It's more convenient
 635         // if we "spill" that into *another* temporary, so that we can map the argument
 636         // variable in the callee MIR directly to an argument variable on our side.
 637         // So we introduce temporaries like:
 638         //
 639         //     tmp0 = tuple_tmp.0
 640         //     tmp1 = tuple_tmp.1
 641         //     tmp2 = tuple_tmp.2
 642         //
 643         // and the vector is `[closure_ref, tmp0, tmp1, tmp2]`.
 644         if callsite.fn_sig.abi() == Abi::RustCall && callee_body.spread_arg.is_none() {
 645             let mut args = args.into_iter();
 646             let self_ = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
 647             let tuple = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
 648             assert!(args.next().is_none());
 649
 650             let tuple = Place::from(tuple);
 651             let ty::Tuple(tuple_tys) = tuple.ty(caller_body, tcx).ty.kind() else {
 652                 bug!("Closure arguments are not passed as a tuple");
 653             };
 654
 655             // The `closure_ref` in our example above.
 656             let closure_ref_arg = iter::once(self_);
 657
 658             // The `tmp0`, `tmp1`, and `tmp2` in our example above.
 659             let tuple_tmp_args = tuple_tys.iter().enumerate().map(|(i, ty)| {
 660                 // This is e.g., `tuple_tmp.0` in our example above.
 661                 let tuple_field = Operand::Move(tcx.mk_place_field(tuple, Field::new(i), ty));
 662
 663                 // Spill to a local to make e.g., `tmp0`.
 664                 self.create_temp_if_necessary(tuple_field, callsite, caller_body)
 665             });
 666
 667             closure_ref_arg.chain(tuple_tmp_args).collect()
 668         } else {
 669             args.into_iter()
 670                 .map(|a| self.create_temp_if_necessary(a, callsite, caller_body))
 671                 .collect()
 672         }
 673     }
 674
 675     /// If `arg` is already a temporary, returns it. Otherwise, introduces a fresh
 676     /// temporary `T` and an instruction `T = arg`, and returns `T`.
 677     fn create_temp_if_necessary(
 678         &self,
 679         arg: Operand<'tcx>,
 680         callsite: &CallSite<'tcx>,
 681         caller_body: &mut Body<'tcx>,
 682     ) -> Local {
 683         // Reuse the operand if it is a moved temporary.
 684         if let Operand::Move(place) = &arg
 685             && let Some(local) = place.as_local()
 686             && caller_body.local_kind(local) == LocalKind::Temp
 687         {
 688             return local;
 689         }
 690
 691         // Otherwise, create a temporary for the argument.
 692         trace!("creating temp for argument {:?}", arg);
 693         let arg_ty = arg.ty(caller_body, self.tcx);
 694         let local = self.new_call_temp(caller_body, callsite, arg_ty);
 695         caller_body[callsite.block].statements.push(Statement {
 696             source_info: callsite.source_info,
 697             kind: StatementKind::Assign(Box::new((Place::from(local), Rvalue::Use(arg)))),
 698         });
 699         local
 700     }
 701
 702     /// Introduces a new temporary into the caller body that is live for the duration of the call.
 703     fn new_call_temp(
 704         &self,
 705         caller_body: &mut Body<'tcx>,
 706         callsite: &CallSite<'tcx>,
 707         ty: Ty<'tcx>,
 708     ) -> Local {
 709         let local = caller_body.local_decls.push(LocalDecl::new(ty, callsite.source_info.span));
 710
 711         caller_body[callsite.block].statements.push(Statement {
 712             source_info: callsite.source_info,
 713             kind: StatementKind::StorageLive(local),
 714         });
 715
 716         if let Some(block) = callsite.target {
 717             caller_body[block].statements.insert(
 718                 0,
 719                 Statement {
 720                     source_info: callsite.source_info,
 721                     kind: StatementKind::StorageDead(local),
 722                 },
 723             );
 724         }
 725
 726         local
 727     }
 728 }
 729
 730 fn type_size_of<'tcx>(
 731     tcx: TyCtxt<'tcx>,
 732     param_env: ty::ParamEnv<'tcx>,
 733     ty: Ty<'tcx>,
 734 ) -> Option<u64> {
 735     tcx.layout_of(param_env.and(ty)).ok().map(|layout| layout.size.bytes())
 736 }
 737
 738 /// Verify that the callee body is compatible with the caller.
 739 ///
 740 /// This visitor mostly computes the inlining cost,
 741 /// but also needs to verify that types match because of normalization failure.
 742 struct CostChecker<'b, 'tcx> {
 743     tcx: TyCtxt<'tcx>,
 744     param_env: ParamEnv<'tcx>,
 745     cost: usize,
 746     callee_body: &'b Body<'tcx>,
 747     instance: ty::Instance<'tcx>,
 748     validation: Result<(), &'static str>,
 749 }
 750
 751 impl<'tcx> Visitor<'tcx> for CostChecker<'_, 'tcx> {
 752     fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) {
 753         // Don't count StorageLive/StorageDead in the inlining cost.
 754         match statement.kind {
 755             StatementKind::StorageLive(_)
 756             | StatementKind::StorageDead(_)
 757             | StatementKind::Deinit(_)
 758             | StatementKind::Nop => {}
 759             _ => self.cost += INSTR_COST,
 760         }
 761
 762         self.super_statement(statement, location);
 763     }
 764
 765     fn visit_terminator(&mut self, terminator: &Terminator<'tcx>, location: Location) {
 766         let tcx = self.tcx;
 767         match terminator.kind {
 768             TerminatorKind::Drop { ref place, unwind, .. }
 769             | TerminatorKind::DropAndReplace { ref place, unwind, .. } => {
 770                 // If the place doesn't actually need dropping, treat it like a regular goto.
 771                 let ty = self.instance.subst_mir(tcx, &place.ty(self.callee_body, tcx).ty);
 772                 if ty.needs_drop(tcx, self.param_env) {
 773                     self.cost += CALL_PENALTY;
 774                     if unwind.is_some() {
 775                         self.cost += LANDINGPAD_PENALTY;
 776                     }
 777                 } else {
 778                     self.cost += INSTR_COST;
 779                 }
 780             }
 781             TerminatorKind::Call { func: Operand::Constant(ref f), cleanup, .. } => {
 782                 let fn_ty = self.instance.subst_mir(tcx, &f.literal.ty());
 783                 self.cost += if let ty::FnDef(def_id, _) = *fn_ty.kind() && tcx.is_intrinsic(def_id) {
 784                     // Don't give intrinsics the extra penalty for calls
 785                     INSTR_COST
 786                 } else {
 787                     CALL_PENALTY
 788                 };
 789                 if cleanup.is_some() {
 790                     self.cost += LANDINGPAD_PENALTY;
 791                 }
 792             }
 793             TerminatorKind::Assert { cleanup, .. } => {
 794                 self.cost += CALL_PENALTY;
 795                 if cleanup.is_some() {
 796                     self.cost += LANDINGPAD_PENALTY;
 797                 }
 798             }
 799             TerminatorKind::Resume => self.cost += RESUME_PENALTY,
 800             TerminatorKind::InlineAsm { cleanup, .. } => {
 801                 self.cost += INSTR_COST;
 802                 if cleanup.is_some() {
 803                     self.cost += LANDINGPAD_PENALTY;
 804                 }
 805             }
 806             _ => self.cost += INSTR_COST,
 807         }
 808
 809         self.super_terminator(terminator, location);
 810     }
 811
 812     /// Count up the cost of local variables and temps, if we know the size
 813     /// use that, otherwise we use a moderately-large dummy cost.
 814     fn visit_local_decl(&mut self, local: Local, local_decl: &LocalDecl<'tcx>) {
 815         let tcx = self.tcx;
 816         let ptr_size = tcx.data_layout.pointer_size.bytes();
 817
 818         let ty = self.instance.subst_mir(tcx, &local_decl.ty);
 819         // Cost of the var is the size in machine-words, if we know
 820         // it.
 821         if let Some(size) = type_size_of(tcx, self.param_env, ty) {
 822             self.cost += ((size + ptr_size - 1) / ptr_size) as usize;
 823         } else {
 824             self.cost += UNKNOWN_SIZE_COST;
 825         }
 826
 827         self.super_local_decl(local, local_decl)
 828     }
 829
 830     /// This method duplicates code from MIR validation in an attempt to detect type mismatches due
 831     /// to normalization failure.
 832     fn visit_projection_elem(
 833         &mut self,
 834         local: Local,
 835         proj_base: &[PlaceElem<'tcx>],
 836         elem: PlaceElem<'tcx>,
 837         context: PlaceContext,
 838         location: Location,
 839     ) {
 840         if let ProjectionElem::Field(f, ty) = elem {
 841             let parent = Place { local, projection: self.tcx.intern_place_elems(proj_base) };
 842             let parent_ty = parent.ty(&self.callee_body.local_decls, self.tcx);
 843             let check_equal = |this: &mut Self, f_ty| {
 844                 if !util::is_equal_up_to_subtyping(this.tcx, this.param_env, ty, f_ty) {
 845                     trace!(?ty, ?f_ty);
 846                     this.validation = Err("failed to normalize projection type");
 847                     return;
 848                 }
 849             };
 850
 851             let kind = match parent_ty.ty.kind() {
 852                 &ty::Opaque(def_id, substs) => {
 853                     self.tcx.bound_type_of(def_id).subst(self.tcx, substs).kind()
 854                 }
 855                 kind => kind,
 856             };
 857
 858             match kind {
 859                 ty::Tuple(fields) => {
 860                     let Some(f_ty) = fields.get(f.as_usize()) else {
 861                         self.validation = Err("malformed MIR");
 862                         return;
 863                     };
 864                     check_equal(self, *f_ty);
 865                 }
 866                 ty::Adt(adt_def, substs) => {
 867                     let var = parent_ty.variant_index.unwrap_or(VariantIdx::from_u32(0));
 868                     let Some(field) = adt_def.variant(var).fields.get(f.as_usize()) else {
 869                         self.validation = Err("malformed MIR");
 870                         return;
 871                     };
 872                     check_equal(self, field.ty(self.tcx, substs));
 873                 }
 874                 ty::Closure(_, substs) => {
 875                     let substs = substs.as_closure();
 876                     let Some(f_ty) = substs.upvar_tys().nth(f.as_usize()) else {
 877                         self.validation = Err("malformed MIR");
 878                         return;
 879                     };
 880                     check_equal(self, f_ty);
 881                 }
 882                 &ty::Generator(def_id, substs, _) => {
 883                     let f_ty = if let Some(var) = parent_ty.variant_index {
 884                         let gen_body = if def_id == self.callee_body.source.def_id() {
 885                             self.callee_body
 886                         } else {
 887                             self.tcx.optimized_mir(def_id)
 888                         };
 889
 890                         let Some(layout) = gen_body.generator_layout() else {
 891                             self.validation = Err("malformed MIR");
 892                             return;
 893                         };
 894
 895                         let Some(&local) = layout.variant_fields[var].get(f) else {
 896                             self.validation = Err("malformed MIR");
 897                             return;
 898                         };
 899
 900                         let Some(&f_ty) = layout.field_tys.get(local) else {
 901                             self.validation = Err("malformed MIR");
 902                             return;
 903                         };
 904
 905                         f_ty
 906                     } else {
 907                         let Some(f_ty) = substs.as_generator().prefix_tys().nth(f.index()) else {
 908                             self.validation = Err("malformed MIR");
 909                             return;
 910                         };
 911
 912                         f_ty
 913                     };
 914
 915                     check_equal(self, f_ty);
 916                 }
 917                 _ => self.validation = Err("malformed MIR"),
 918             }
 919         }
 920
 921         self.super_projection_elem(local, proj_base, elem, context, location);
 922     }
 923 }
 924
 925 /**
 926  * Integrator.
 927  *
 928  * Integrates blocks from the callee function into the calling function.
 929  * Updates block indices, references to locals and other control flow
 930  * stuff.
 931 */
 932 struct Integrator<'a, 'tcx> {
 933     args: &'a [Local],
 934     new_locals: RangeFrom<Local>,
 935     new_scopes: RangeFrom<SourceScope>,
 936     new_blocks: RangeFrom<BasicBlock>,
 937     destination: Place<'tcx>,
 938     callsite_scope: SourceScopeData<'tcx>,
 939     callsite: &'a CallSite<'tcx>,
 940     cleanup_block: Option<BasicBlock>,
 941     in_cleanup_block: bool,
 942     tcx: TyCtxt<'tcx>,
 943     expn_data: LocalExpnId,
 944     always_live_locals: BitSet<Local>,
 945 }
 946
 947 impl Integrator<'_, '_> {
 948     fn map_local(&self, local: Local) -> Local {
 949         let new = if local == RETURN_PLACE {
 950             self.destination.local
 951         } else {
 952             let idx = local.index() - 1;
 953             if idx < self.args.len() {
 954                 self.args[idx]
 955             } else {
 956                 Local::new(self.new_locals.start.index() + (idx - self.args.len()))
 957             }
 958         };
 959         trace!("mapping local `{:?}` to `{:?}`", local, new);
 960         new
 961     }
 962
 963     fn map_scope(&self, scope: SourceScope) -> SourceScope {
 964         let new = SourceScope::new(self.new_scopes.start.index() + scope.index());
 965         trace!("mapping scope `{:?}` to `{:?}`", scope, new);
 966         new
 967     }
 968
 969     fn map_block(&self, block: BasicBlock) -> BasicBlock {
 970         let new = BasicBlock::new(self.new_blocks.start.index() + block.index());
 971         trace!("mapping block `{:?}` to `{:?}`", block, new);
 972         new
 973     }
 974
 975     fn map_unwind(&self, unwind: Option<BasicBlock>) -> Option<BasicBlock> {
 976         if self.in_cleanup_block {
 977             if unwind.is_some() {
 978                 bug!("cleanup on cleanup block");
 979             }
 980             return unwind;
 981         }
 982
 983         match unwind {
 984             Some(target) => Some(self.map_block(target)),
 985             // Add an unwind edge to the original call's cleanup block
 986             None => self.cleanup_block,
 987         }
 988     }
 989 }
 990
 991 impl<'tcx> MutVisitor<'tcx> for Integrator<'_, 'tcx> {
 992     fn tcx(&self) -> TyCtxt<'tcx> {
 993         self.tcx
 994     }
 995
 996     fn visit_local(&mut self, local: &mut Local, _ctxt: PlaceContext, _location: Location) {
 997         *local = self.map_local(*local);
 998     }
 999
1000     fn visit_source_scope_data(&mut self, scope_data: &mut SourceScopeData<'tcx>) {
1001         self.super_source_scope_data(scope_data);
1002         if scope_data.parent_scope.is_none() {
1003             // Attach the outermost callee scope as a child of the callsite
1004             // scope, via the `parent_scope` and `inlined_parent_scope` chains.
1005             scope_data.parent_scope = Some(self.callsite.source_info.scope);
1006             assert_eq!(scope_data.inlined_parent_scope, None);
1007             scope_data.inlined_parent_scope = if self.callsite_scope.inlined.is_some() {
1008                 Some(self.callsite.source_info.scope)
1009             } else {
1010                 self.callsite_scope.inlined_parent_scope
1011             };
1012
1013             // Mark the outermost callee scope as an inlined one.
1014             assert_eq!(scope_data.inlined, None);
1015             scope_data.inlined = Some((self.callsite.callee, self.callsite.source_info.span));
1016         } else if scope_data.inlined_parent_scope.is_none() {
1017             // Make it easy to find the scope with `inlined` set above.
1018             scope_data.inlined_parent_scope = Some(self.map_scope(OUTERMOST_SOURCE_SCOPE));
1019         }
1020     }
1021
1022     fn visit_source_scope(&mut self, scope: &mut SourceScope) {
1023         *scope = self.map_scope(*scope);
1024     }
1025
1026     fn visit_span(&mut self, span: &mut Span) {
1027         // Make sure that all spans track the fact that they were inlined.
1028         *span = span.fresh_expansion(self.expn_data);
1029     }
1030
1031     fn visit_place(&mut self, place: &mut Place<'tcx>, context: PlaceContext, location: Location) {
1032         for elem in place.projection {
1033             // FIXME: Make sure that return place is not used in an indexing projection, since it
1034             // won't be rebased as it is supposed to be.
1035             assert_ne!(ProjectionElem::Index(RETURN_PLACE), elem);
1036         }
1037
1038         // If this is the `RETURN_PLACE`, we need to rebase any projections onto it.
1039         let dest_proj_len = self.destination.projection.len();
1040         if place.local == RETURN_PLACE && dest_proj_len > 0 {
1041             let mut projs = Vec::with_capacity(dest_proj_len + place.projection.len());
1042             projs.extend(self.destination.projection);
1043             projs.extend(place.projection);
1044
1045             place.projection = self.tcx.intern_place_elems(&*projs);
1046         }
1047         // Handles integrating any locals that occur in the base
1048         // or projections
1049         self.super_place(place, context, location)
1050     }
1051
1052     fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
1053         self.in_cleanup_block = data.is_cleanup;
1054         self.super_basic_block_data(block, data);
1055         self.in_cleanup_block = false;
1056     }
1057
1058     fn visit_retag(&mut self, kind: &mut RetagKind, place: &mut Place<'tcx>, loc: Location) {
1059         self.super_retag(kind, place, loc);
1060
1061         // We have to patch all inlined retags to be aware that they are no longer
1062         // happening on function entry.
1063         if *kind == RetagKind::FnEntry {
1064             *kind = RetagKind::Default;
1065         }
1066     }
1067
1068     fn visit_statement(&mut self, statement: &mut Statement<'tcx>, location: Location) {
1069         if let StatementKind::StorageLive(local) | StatementKind::StorageDead(local) =
1070             statement.kind
1071         {
1072             self.always_live_locals.remove(local);
1073         }
1074         self.super_statement(statement, location);
1075     }
1076
1077     fn visit_terminator(&mut self, terminator: &mut Terminator<'tcx>, loc: Location) {
1078         // Don't try to modify the implicit `_0` access on return (`return` terminators are
1079         // replaced down below anyways).
1080         if !matches!(terminator.kind, TerminatorKind::Return) {
1081             self.super_terminator(terminator, loc);
1082         }
1083
1084         match terminator.kind {
1085             TerminatorKind::GeneratorDrop | TerminatorKind::Yield { .. } => bug!(),
1086             TerminatorKind::Goto { ref mut target } => {
1087                 *target = self.map_block(*target);
1088             }
1089             TerminatorKind::SwitchInt { ref mut targets, .. } => {
1090                 for tgt in targets.all_targets_mut() {
1091                     *tgt = self.map_block(*tgt);
1092                 }
1093             }
1094             TerminatorKind::Drop { ref mut target, ref mut unwind, .. }
1095             | TerminatorKind::DropAndReplace { ref mut target, ref mut unwind, .. } => {
1096                 *target = self.map_block(*target);
1097                 *unwind = self.map_unwind(*unwind);
1098             }
1099             TerminatorKind::Call { ref mut target, ref mut cleanup, .. } => {
1100                 if let Some(ref mut tgt) = *target {
1101                     *tgt = self.map_block(*tgt);
1102                 }
1103                 *cleanup = self.map_unwind(*cleanup);
1104             }
1105             TerminatorKind::Assert { ref mut target, ref mut cleanup, .. } => {
1106                 *target = self.map_block(*target);
1107                 *cleanup = self.map_unwind(*cleanup);
1108             }
1109             TerminatorKind::Return => {
1110                 terminator.kind = if let Some(tgt) = self.callsite.target {
1111                     TerminatorKind::Goto { target: tgt }
1112                 } else {
1113                     TerminatorKind::Unreachable
1114                 }
1115             }
1116             TerminatorKind::Resume => {
1117                 if let Some(tgt) = self.cleanup_block {
1118                     terminator.kind = TerminatorKind::Goto { target: tgt }
1119                 }
1120             }
1121             TerminatorKind::Abort => {}
1122             TerminatorKind::Unreachable => {}
1123             TerminatorKind::FalseEdge { ref mut real_target, ref mut imaginary_target } => {
1124                 *real_target = self.map_block(*real_target);
1125                 *imaginary_target = self.map_block(*imaginary_target);
1126             }
1127             TerminatorKind::FalseUnwind { real_target: _, unwind: _ } =>
1128             // see the ordering of passes in the optimized_mir query.
1129             {
1130                 bug!("False unwinds should have been removed before inlining")
1131             }
1132             TerminatorKind::InlineAsm { ref mut destination, ref mut cleanup, .. } => {
1133                 if let Some(ref mut tgt) = *destination {
1134                     *tgt = self.map_block(*tgt);
1135                 }
1136                 *cleanup = self.map_unwind(*cleanup);
1137             }
1138         }
1139     }
1140 }