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