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::def_id::DefId;
  12 use rustc_span::{hygiene::ExpnKind, ExpnData, LocalExpnId, Span};
  13 use rustc_target::abi::VariantIdx;
  14 use rustc_target::spec::abi::Abi;
  15
  16 use crate::simplify::{remove_dead_blocks, CfgSimplifier};
  17 use crate::util;
  18 use crate::MirPass;
  19 use std::iter;
  20 use std::ops::{Range, RangeFrom};
  21
  22 pub(crate) mod cycle;
  23
  24 const INSTR_COST: usize = 5;
  25 const CALL_PENALTY: usize = 25;
  26 const LANDINGPAD_PENALTY: usize = 50;
  27 const RESUME_PENALTY: usize = 45;
  28
  29 const UNKNOWN_SIZE_COST: usize = 10;
  30
  31 pub struct Inline;
  32
  33 #[derive(Copy, Clone, Debug)]
  34 struct CallSite<'tcx> {
  35     callee: Instance<'tcx>,
  36     fn_sig: ty::PolyFnSig<'tcx>,
  37     block: BasicBlock,
  38     target: Option<BasicBlock>,
  39     source_info: SourceInfo,
  40 }
  41
  42 impl<'tcx> MirPass<'tcx> for Inline {
  43     fn is_enabled(&self, sess: &rustc_session::Session) -> bool {
  44         if let Some(enabled) = sess.opts.unstable_opts.inline_mir {
  45             return enabled;
  46         }
  47
  48         match sess.mir_opt_level() {
  49             0 | 1 => false,
  50             2 => {
  51                 (sess.opts.optimize == OptLevel::Default
  52                     || sess.opts.optimize == OptLevel::Aggressive)
  53                     && sess.opts.incremental == None
  54             }
  55             _ => true,
  56         }
  57     }
  58
  59     fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
  60         let span = trace_span!("inline", body = %tcx.def_path_str(body.source.def_id()));
  61         let _guard = span.enter();
  62         if inline(tcx, body) {
  63             debug!("running simplify cfg on {:?}", body.source);
  64             CfgSimplifier::new(body).simplify();
  65             remove_dead_blocks(tcx, body);
  66             deref_finder(tcx, body);
  67         }
  68     }
  69 }
  70
  71 fn inline<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) -> bool {
  72     let def_id = body.source.def_id().expect_local();
  73
  74     // Only do inlining into fn bodies.
  75     if !tcx.hir().body_owner_kind(def_id).is_fn_or_closure() {
  76         return false;
  77     }
  78     if body.source.promoted.is_some() {
  79         return false;
  80     }
  81     // Avoid inlining into generators, since their `optimized_mir` is used for layout computation,
  82     // which can create a cycle, even when no attempt is made to inline the function in the other
  83     // direction.
  84     if body.generator.is_some() {
  85         return false;
  86     }
  87
  88     let param_env = tcx.param_env_reveal_all_normalized(def_id);
  89
  90     let mut this = Inliner {
  91         tcx,
  92         param_env,
  93         codegen_fn_attrs: tcx.codegen_fn_attrs(def_id),
  94         history: Vec::new(),
  95         changed: false,
  96     };
  97     let blocks = BasicBlock::new(0)..body.basic_blocks.next_index();
  98     this.process_blocks(body, blocks);
  99     this.changed
 100 }
 101
 102 struct Inliner<'tcx> {
 103     tcx: TyCtxt<'tcx>,
 104     param_env: ParamEnv<'tcx>,
 105     /// Caller codegen attributes.
 106     codegen_fn_attrs: &'tcx CodegenFnAttrs,
 107     /// Stack of inlined instances.
 108     /// We only check the `DefId` and not the substs because we want to
 109     /// avoid inlining cases of polymorphic recursion.
 110     /// The number of `DefId`s is finite, so checking history is enough
 111     /// to ensure that we do not loop endlessly while inlining.
 112     history: Vec<DefId>,
 113     /// Indicates that the caller body has been modified.
 114     changed: bool,
 115 }
 116
 117 impl<'tcx> Inliner<'tcx> {
 118     fn process_blocks(&mut self, caller_body: &mut Body<'tcx>, blocks: Range<BasicBlock>) {
 119         for bb in blocks {
 120             let bb_data = &caller_body[bb];
 121             if bb_data.is_cleanup {
 122                 continue;
 123             }
 124
 125             let Some(callsite) = self.resolve_callsite(caller_body, bb, bb_data) else {
 126                 continue;
 127             };
 128
 129             let span = trace_span!("process_blocks", %callsite.callee, ?bb);
 130             let _guard = span.enter();
 131
 132             match self.try_inlining(caller_body, &callsite) {
 133                 Err(reason) => {
 134                     debug!("not-inlined {} [{}]", callsite.callee, reason);
 135                     continue;
 136                 }
 137                 Ok(new_blocks) => {
 138                     debug!("inlined {}", callsite.callee);
 139                     self.changed = true;
 140                     self.history.push(callsite.callee.def_id());
 141                     self.process_blocks(caller_body, new_blocks);
 142                     self.history.pop();
 143                 }
 144             }
 145         }
 146     }
 147
 148     /// Attempts to inline a callsite into the caller body. When successful returns basic blocks
 149     /// containing the inlined body. Otherwise returns an error describing why inlining didn't take
 150     /// place.
 151     fn try_inlining(
 152         &self,
 153         caller_body: &mut Body<'tcx>,
 154         callsite: &CallSite<'tcx>,
 155     ) -> Result<std::ops::Range<BasicBlock>, &'static str> {
 156         let callee_attrs = self.tcx.codegen_fn_attrs(callsite.callee.def_id());
 157         self.check_codegen_attributes(callsite, callee_attrs)?;
 158         self.check_mir_is_available(caller_body, &callsite.callee)?;
 159         let callee_body = self.tcx.instance_mir(callsite.callee.def);
 160         self.check_mir_body(callsite, callee_body, callee_attrs)?;
 161
 162         if !self.tcx.consider_optimizing(|| {
 163             format!("Inline {:?} into {:?}", callsite.callee, caller_body.source)
 164         }) {
 165             return Err("optimization fuel exhausted");
 166         }
 167
 168         let Ok(callee_body) = callsite.callee.try_subst_mir_and_normalize_erasing_regions(
 169             self.tcx,
 170             self.param_env,
 171             callee_body.clone(),
 172         ) else {
 173             return Err("failed to normalize callee body");
 174         };
 175
 176         // Check call signature compatibility.
 177         // Normally, this shouldn't be required, but trait normalization failure can create a
 178         // validation ICE.
 179         let terminator = caller_body[callsite.block].terminator.as_ref().unwrap();
 180         let TerminatorKind::Call { args, destination, .. } = &terminator.kind else { bug!() };
 181         let destination_ty = destination.ty(&caller_body.local_decls, self.tcx).ty;
 182         let output_type = callee_body.return_ty();
 183         if !util::is_subtype(self.tcx, self.param_env, output_type, destination_ty) {
 184             trace!(?output_type, ?destination_ty);
 185             return Err("failed to normalize return type");
 186         }
 187         if callsite.fn_sig.abi() == Abi::RustCall {
 188             let (arg_tuple, skipped_args) = match &args[..] {
 189                 [arg_tuple] => (arg_tuple, 0),
 190                 [_, arg_tuple] => (arg_tuple, 1),
 191                 _ => bug!("Expected `rust-call` to have 1 or 2 args"),
 192             };
 193
 194             let arg_tuple_ty = arg_tuple.ty(&caller_body.local_decls, self.tcx);
 195             let ty::Tuple(arg_tuple_tys) = arg_tuple_ty.kind() else {
 196                 bug!("Closure arguments are not passed as a tuple");
 197             };
 198
 199             for (arg_ty, input) in
 200                 arg_tuple_tys.iter().zip(callee_body.args_iter().skip(skipped_args))
 201             {
 202                 let input_type = callee_body.local_decls[input].ty;
 203                 if !util::is_subtype(self.tcx, self.param_env, input_type, arg_ty) {
 204                     trace!(?arg_ty, ?input_type);
 205                     return Err("failed to normalize tuple argument type");
 206                 }
 207             }
 208         } else {
 209             for (arg, input) in args.iter().zip(callee_body.args_iter()) {
 210                 let input_type = callee_body.local_decls[input].ty;
 211                 let arg_ty = arg.ty(&caller_body.local_decls, self.tcx);
 212                 if !util::is_subtype(self.tcx, self.param_env, input_type, arg_ty) {
 213                     trace!(?arg_ty, ?input_type);
 214                     return Err("failed to normalize argument type");
 215                 }
 216             }
 217         }
 218
 219         let old_blocks = caller_body.basic_blocks.next_index();
 220         self.inline_call(caller_body, &callsite, callee_body);
 221         let new_blocks = old_blocks..caller_body.basic_blocks.next_index();
 222
 223         Ok(new_blocks)
 224     }
 225
 226     fn check_mir_is_available(
 227         &self,
 228         caller_body: &Body<'tcx>,
 229         callee: &Instance<'tcx>,
 230     ) -> Result<(), &'static str> {
 231         let caller_def_id = caller_body.source.def_id();
 232         let callee_def_id = callee.def_id();
 233         if callee_def_id == caller_def_id {
 234             return Err("self-recursion");
 235         }
 236
 237         match callee.def {
 238             InstanceDef::Item(_) => {
 239                 // If there is no MIR available (either because it was not in metadata or
 240                 // because it has no MIR because it's an extern function), then the inliner
 241                 // won't cause cycles on this.
 242                 if !self.tcx.is_mir_available(callee_def_id) {
 243                     return Err("item MIR unavailable");
 244                 }
 245             }
 246             // These have no own callable MIR.
 247             InstanceDef::Intrinsic(_) | InstanceDef::Virtual(..) => {
 248                 return Err("instance without MIR (intrinsic / virtual)");
 249             }
 250             // This cannot result in an immediate cycle since the callee MIR is a shim, which does
 251             // not get any optimizations run on it. Any subsequent inlining may cause cycles, but we
 252             // do not need to catch this here, we can wait until the inliner decides to continue
 253             // inlining a second time.
 254             InstanceDef::VTableShim(_)
 255             | InstanceDef::ReifyShim(_)
 256             | InstanceDef::FnPtrShim(..)
 257             | InstanceDef::ClosureOnceShim { .. }
 258             | InstanceDef::DropGlue(..)
 259             | InstanceDef::CloneShim(..) => return Ok(()),
 260         }
 261
 262         if self.tcx.is_constructor(callee_def_id) {
 263             trace!("constructors always have MIR");
 264             // Constructor functions cannot cause a query cycle.
 265             return Ok(());
 266         }
 267
 268         if callee_def_id.is_local() {
 269             // Avoid a cycle here by only using `instance_mir` only if we have
 270             // a lower `DefPathHash` than the callee. This ensures that the callee will
 271             // not inline us. This trick even works with incremental compilation,
 272             // since `DefPathHash` is stable.
 273             if self.tcx.def_path_hash(caller_def_id).local_hash()
 274                 < self.tcx.def_path_hash(callee_def_id).local_hash()
 275             {
 276                 return Ok(());
 277             }
 278
 279             // If we know for sure that the function we're calling will itself try to
 280             // call us, then we avoid inlining that function.
 281             if self.tcx.mir_callgraph_reachable((*callee, caller_def_id.expect_local())) {
 282                 return Err("caller might be reachable from callee (query cycle avoidance)");
 283             }
 284
 285             Ok(())
 286         } else {
 287             // This cannot result in an immediate cycle since the callee MIR is from another crate
 288             // and is already optimized. Any subsequent inlining may cause cycles, but we do
 289             // not need to catch this here, we can wait until the inliner decides to continue
 290             // inlining a second time.
 291             trace!("functions from other crates always have MIR");
 292             Ok(())
 293         }
 294     }
 295
 296     fn resolve_callsite(
 297         &self,
 298         caller_body: &Body<'tcx>,
 299         bb: BasicBlock,
 300         bb_data: &BasicBlockData<'tcx>,
 301     ) -> Option<CallSite<'tcx>> {
 302         // Only consider direct calls to functions
 303         let terminator = bb_data.terminator();
 304         if let TerminatorKind::Call { ref func, target, .. } = terminator.kind {
 305             let func_ty = func.ty(caller_body, self.tcx);
 306             if let ty::FnDef(def_id, substs) = *func_ty.kind() {
 307                 // To resolve an instance its substs have to be fully normalized.
 308                 let substs = self.tcx.try_normalize_erasing_regions(self.param_env, substs).ok()?;
 309                 let callee =
 310                     Instance::resolve(self.tcx, self.param_env, def_id, substs).ok().flatten()?;
 311
 312                 if let InstanceDef::Virtual(..) | InstanceDef::Intrinsic(_) = callee.def {
 313                     return None;
 314                 }
 315
 316                 if self.history.contains(&callee.def_id()) {
 317                     return None;
 318                 }
 319
 320                 let fn_sig = self.tcx.bound_fn_sig(def_id).subst(self.tcx, substs);
 321
 322                 return Some(CallSite {
 323                     callee,
 324                     fn_sig,
 325                     block: bb,
 326                     target,
 327                     source_info: terminator.source_info,
 328                 });
 329             }
 330         }
 331
 332         None
 333     }
 334
 335     /// Returns an error if inlining is not possible based on codegen attributes alone. A success
 336     /// indicates that inlining decision should be based on other criteria.
 337     fn check_codegen_attributes(
 338         &self,
 339         callsite: &CallSite<'tcx>,
 340         callee_attrs: &CodegenFnAttrs,
 341     ) -> Result<(), &'static str> {
 342         match callee_attrs.inline {
 343             InlineAttr::Never => return Err("never inline hint"),
 344             InlineAttr::Always | InlineAttr::Hint => {}
 345             InlineAttr::None => {
 346                 if self.tcx.sess.mir_opt_level() <= 2 {
 347                     return Err("at mir-opt-level=2, only #[inline] is inlined");
 348                 }
 349             }
 350         }
 351
 352         // Only inline local functions if they would be eligible for cross-crate
 353         // inlining. This is to ensure that the final crate doesn't have MIR that
 354         // reference unexported symbols
 355         if callsite.callee.def_id().is_local() {
 356             let is_generic = callsite.callee.substs.non_erasable_generics().next().is_some();
 357             if !is_generic && !callee_attrs.requests_inline() {
 358                 return Err("not exported");
 359             }
 360         }
 361
 362         if callsite.fn_sig.c_variadic() {
 363             return Err("C variadic");
 364         }
 365
 366         if callee_attrs.flags.contains(CodegenFnAttrFlags::COLD) {
 367             return Err("cold");
 368         }
 369
 370         if callee_attrs.no_sanitize != self.codegen_fn_attrs.no_sanitize {
 371             return Err("incompatible sanitizer set");
 372         }
 373
 374         // Two functions are compatible if the callee has no attribute (meaning
 375         // that it's codegen agnostic), or sets an attribute that is identical
 376         // to this function's attribute.
 377         if callee_attrs.instruction_set.is_some()
 378             && callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
 379         {
 380             return Err("incompatible instruction set");
 381         }
 382
 383         for feature in &callee_attrs.target_features {
 384             if !self.codegen_fn_attrs.target_features.contains(feature) {
 385                 return Err("incompatible target feature");
 386             }
 387         }
 388
 389         Ok(())
 390     }
 391
 392     /// Returns inlining decision that is based on the examination of callee MIR body.
 393     /// Assumes that codegen attributes have been checked for compatibility already.
 394     #[instrument(level = "debug", skip(self, callee_body))]
 395     fn check_mir_body(
 396         &self,
 397         callsite: &CallSite<'tcx>,
 398         callee_body: &Body<'tcx>,
 399         callee_attrs: &CodegenFnAttrs,
 400     ) -> Result<(), &'static str> {
 401         let tcx = self.tcx;
 402
 403         let mut threshold = if callee_attrs.requests_inline() {
 404             self.tcx.sess.opts.unstable_opts.inline_mir_hint_threshold.unwrap_or(100)
 405         } else {
 406             self.tcx.sess.opts.unstable_opts.inline_mir_threshold.unwrap_or(50)
 407         };
 408
 409         // Give a bonus functions with a small number of blocks,
 410         // We normally have two or three blocks for even
 411         // very small functions.
 412         if callee_body.basic_blocks.len() <= 3 {
 413             threshold += threshold / 4;
 414         }
 415         debug!("    final inline threshold = {}", threshold);
 416
 417         // FIXME: Give a bonus to functions with only a single caller
 418         let diverges = matches!(
 419             callee_body.basic_blocks[START_BLOCK].terminator().kind,
 420             TerminatorKind::Unreachable | TerminatorKind::Call { target: None, .. }
 421         );
 422         if diverges && !matches!(callee_attrs.inline, InlineAttr::Always) {
 423             return Err("callee diverges unconditionally");
 424         }
 425
 426         let mut checker = CostChecker {
 427             tcx: self.tcx,
 428             param_env: self.param_env,
 429             instance: callsite.callee,
 430             callee_body,
 431             cost: 0,
 432             validation: Ok(()),
 433         };
 434
 435         // Traverse the MIR manually so we can account for the effects of inlining on the CFG.
 436         let mut work_list = vec![START_BLOCK];
 437         let mut visited = BitSet::new_empty(callee_body.basic_blocks.len());
 438         while let Some(bb) = work_list.pop() {
 439             if !visited.insert(bb.index()) {
 440                 continue;
 441             }
 442
 443             let blk = &callee_body.basic_blocks[bb];
 444             checker.visit_basic_block_data(bb, blk);
 445
 446             let term = blk.terminator();
 447             if let TerminatorKind::Drop { ref place, target, unwind }
 448             | TerminatorKind::DropAndReplace { ref place, target, unwind, .. } = term.kind
 449             {
 450                 work_list.push(target);
 451
 452                 // If the place doesn't actually need dropping, treat it like a regular goto.
 453                 let ty = callsite.callee.subst_mir(self.tcx, &place.ty(callee_body, tcx).ty);
 454                 if ty.needs_drop(tcx, self.param_env) && let Some(unwind) = unwind {
 455                         work_list.push(unwind);
 456                     }
 457             } else if callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
 458                 && matches!(term.kind, TerminatorKind::InlineAsm { .. })
 459             {
 460                 // During the attribute checking stage we allow a callee with no
 461                 // instruction_set assigned to count as compatible with a function that does
 462                 // assign one. However, during this stage we require an exact match when any
 463                 // inline-asm is detected. LLVM will still possibly do an inline later on
 464                 // if the no-attribute function ends up with the same instruction set anyway.
 465                 return Err("Cannot move inline-asm across instruction sets");
 466             } else {
 467                 work_list.extend(term.successors())
 468             }
 469         }
 470
 471         // Count up the cost of local variables and temps, if we know the size
 472         // use that, otherwise we use a moderately-large dummy cost.
 473         for v in callee_body.vars_and_temps_iter() {
 474             checker.visit_local_decl(v, &callee_body.local_decls[v]);
 475         }
 476
 477         // Abort if type validation found anything fishy.
 478         checker.validation?;
 479
 480         let cost = checker.cost;
 481         if let InlineAttr::Always = callee_attrs.inline {
 482             debug!("INLINING {:?} because inline(always) [cost={}]", callsite, cost);
 483             Ok(())
 484         } else if cost <= threshold {
 485             debug!("INLINING {:?} [cost={} <= threshold={}]", callsite, cost, threshold);
 486             Ok(())
 487         } else {
 488             debug!("NOT inlining {:?} [cost={} > threshold={}]", callsite, cost, threshold);
 489             Err("cost above threshold")
 490         }
 491     }
 492
 493     fn inline_call(
 494         &self,
 495         caller_body: &mut Body<'tcx>,
 496         callsite: &CallSite<'tcx>,
 497         mut callee_body: Body<'tcx>,
 498     ) {
 499         let terminator = caller_body[callsite.block].terminator.take().unwrap();
 500         match terminator.kind {
 501             TerminatorKind::Call { args, destination, cleanup, .. } => {
 502                 // If the call is something like `a[*i] = f(i)`, where
 503                 // `i : &mut usize`, then just duplicating the `a[*i]`
 504                 // Place could result in two different locations if `f`
 505                 // writes to `i`. To prevent this we need to create a temporary
 506                 // borrow of the place and pass the destination as `*temp` instead.
 507                 fn dest_needs_borrow(place: Place<'_>) -> bool {
 508                     for elem in place.projection.iter() {
 509                         match elem {
 510                             ProjectionElem::Deref | ProjectionElem::Index(_) => return true,
 511                             _ => {}
 512                         }
 513                     }
 514
 515                     false
 516                 }
 517
 518                 let dest = if dest_needs_borrow(destination) {
 519                     trace!("creating temp for return destination");
 520                     let dest = Rvalue::Ref(
 521                         self.tcx.lifetimes.re_erased,
 522                         BorrowKind::Mut { allow_two_phase_borrow: false },
 523                         destination,
 524                     );
 525                     let dest_ty = dest.ty(caller_body, self.tcx);
 526                     let temp = Place::from(self.new_call_temp(caller_body, &callsite, dest_ty));
 527                     caller_body[callsite.block].statements.push(Statement {
 528                         source_info: callsite.source_info,
 529                         kind: StatementKind::Assign(Box::new((temp, dest))),
 530                     });
 531                     self.tcx.mk_place_deref(temp)
 532                 } else {
 533                     destination
 534                 };
 535
 536                 // Copy the arguments if needed.
 537                 let args: Vec<_> = self.make_call_args(args, &callsite, caller_body, &callee_body);
 538
 539                 let mut expn_data = ExpnData::default(
 540                     ExpnKind::Inlined,
 541                     callsite.source_info.span,
 542                     self.tcx.sess.edition(),
 543                     None,
 544                     None,
 545                 );
 546                 expn_data.def_site = callee_body.span;
 547                 let expn_data =
 548                     self.tcx.with_stable_hashing_context(|hcx| LocalExpnId::fresh(expn_data, hcx));
 549                 let mut integrator = Integrator {
 550                     args: &args,
 551                     new_locals: Local::new(caller_body.local_decls.len())..,
 552                     new_scopes: SourceScope::new(caller_body.source_scopes.len())..,
 553                     new_blocks: BasicBlock::new(caller_body.basic_blocks.len())..,
 554                     destination: dest,
 555                     callsite_scope: caller_body.source_scopes[callsite.source_info.scope].clone(),
 556                     callsite,
 557                     cleanup_block: cleanup,
 558                     in_cleanup_block: false,
 559                     tcx: self.tcx,
 560                     expn_data,
 561                     always_live_locals: BitSet::new_filled(callee_body.local_decls.len()),
 562                 };
 563
 564                 // Map all `Local`s, `SourceScope`s and `BasicBlock`s to new ones
 565                 // (or existing ones, in a few special cases) in the caller.
 566                 integrator.visit_body(&mut callee_body);
 567
 568                 // If there are any locals without storage markers, give them storage only for the
 569                 // duration of the call.
 570                 for local in callee_body.vars_and_temps_iter() {
 571                     if !callee_body.local_decls[local].internal
 572                         && integrator.always_live_locals.contains(local)
 573                     {
 574                         let new_local = integrator.map_local(local);
 575                         caller_body[callsite.block].statements.push(Statement {
 576                             source_info: callsite.source_info,
 577                             kind: StatementKind::StorageLive(new_local),
 578                         });
 579                     }
 580                 }
 581                 if let Some(block) = callsite.target {
 582                     // To avoid repeated O(n) insert, push any new statements to the end and rotate
 583                     // the slice once.
 584                     let mut n = 0;
 585                     for local in callee_body.vars_and_temps_iter().rev() {
 586                         if !callee_body.local_decls[local].internal
 587                             && integrator.always_live_locals.contains(local)
 588                         {
 589                             let new_local = integrator.map_local(local);
 590                             caller_body[block].statements.push(Statement {
 591                                 source_info: callsite.source_info,
 592                                 kind: StatementKind::StorageDead(new_local),
 593                             });
 594                             n += 1;
 595                         }
 596                     }
 597                     caller_body[block].statements.rotate_right(n);
 598                 }
 599
 600                 // Insert all of the (mapped) parts of the callee body into the caller.
 601                 caller_body.local_decls.extend(callee_body.drain_vars_and_temps());
 602                 caller_body.source_scopes.extend(&mut callee_body.source_scopes.drain(..));
 603                 caller_body.var_debug_info.append(&mut callee_body.var_debug_info);
 604                 caller_body.basic_blocks_mut().extend(callee_body.basic_blocks_mut().drain(..));
 605
 606                 caller_body[callsite.block].terminator = Some(Terminator {
 607                     source_info: callsite.source_info,
 608                     kind: TerminatorKind::Goto { target: integrator.map_block(START_BLOCK) },
 609                 });
 610
 611                 // Copy only unevaluated constants from the callee_body into the caller_body.
 612                 // Although we are only pushing `ConstKind::Unevaluated` consts to
 613                 // `required_consts`, here we may not only have `ConstKind::Unevaluated`
 614                 // because we are calling `subst_and_normalize_erasing_regions`.
 615                 caller_body.required_consts.extend(
 616                     callee_body.required_consts.iter().copied().filter(|&ct| match ct.literal {
 617                         ConstantKind::Ty(_) => {
 618                             bug!("should never encounter ty::UnevaluatedConst in `required_consts`")
 619                         }
 620                         ConstantKind::Val(..) | ConstantKind::Unevaluated(..) => true,
 621                     }),
 622                 );
 623             }
 624             kind => bug!("unexpected terminator kind {:?}", kind),
 625         }
 626     }
 627
 628     fn make_call_args(
 629         &self,
 630         args: Vec<Operand<'tcx>>,
 631         callsite: &CallSite<'tcx>,
 632         caller_body: &mut Body<'tcx>,
 633         callee_body: &Body<'tcx>,
 634     ) -> Vec<Local> {
 635         let tcx = self.tcx;
 636
 637         // There is a bit of a mismatch between the *caller* of a closure and the *callee*.
 638         // The caller provides the arguments wrapped up in a tuple:
 639         //
 640         //     tuple_tmp = (a, b, c)
 641         //     Fn::call(closure_ref, tuple_tmp)
 642         //
 643         // meanwhile the closure body expects the arguments (here, `a`, `b`, and `c`)
 644         // as distinct arguments. (This is the "rust-call" ABI hack.) Normally, codegen has
 645         // the job of unpacking this tuple. But here, we are codegen. =) So we want to create
 646         // a vector like
 647         //
 648         //     [closure_ref, tuple_tmp.0, tuple_tmp.1, tuple_tmp.2]
 649         //
 650         // Except for one tiny wrinkle: we don't actually want `tuple_tmp.0`. It's more convenient
 651         // if we "spill" that into *another* temporary, so that we can map the argument
 652         // variable in the callee MIR directly to an argument variable on our side.
 653         // So we introduce temporaries like:
 654         //
 655         //     tmp0 = tuple_tmp.0
 656         //     tmp1 = tuple_tmp.1
 657         //     tmp2 = tuple_tmp.2
 658         //
 659         // and the vector is `[closure_ref, tmp0, tmp1, tmp2]`.
 660         if callsite.fn_sig.abi() == Abi::RustCall && callee_body.spread_arg.is_none() {
 661             let mut args = args.into_iter();
 662             let self_ = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
 663             let tuple = self.create_temp_if_necessary(args.next().unwrap(), callsite, caller_body);
 664             assert!(args.next().is_none());
 665
 666             let tuple = Place::from(tuple);
 667             let ty::Tuple(tuple_tys) = tuple.ty(caller_body, tcx).ty.kind() else {
 668                 bug!("Closure arguments are not passed as a tuple");
 669             };
 670
 671             // The `closure_ref` in our example above.
 672             let closure_ref_arg = iter::once(self_);
 673
 674             // The `tmp0`, `tmp1`, and `tmp2` in our example above.
 675             let tuple_tmp_args = tuple_tys.iter().enumerate().map(|(i, ty)| {
 676                 // This is e.g., `tuple_tmp.0` in our example above.
 677                 let tuple_field = Operand::Move(tcx.mk_place_field(tuple, Field::new(i), ty));
 678
 679                 // Spill to a local to make e.g., `tmp0`.
 680                 self.create_temp_if_necessary(tuple_field, callsite, caller_body)
 681             });
 682
 683             closure_ref_arg.chain(tuple_tmp_args).collect()
 684         } else {
 685             args.into_iter()
 686                 .map(|a| self.create_temp_if_necessary(a, callsite, caller_body))
 687                 .collect()
 688         }
 689     }
 690
 691     /// If `arg` is already a temporary, returns it. Otherwise, introduces a fresh
 692     /// temporary `T` and an instruction `T = arg`, and returns `T`.
 693     fn create_temp_if_necessary(
 694         &self,
 695         arg: Operand<'tcx>,
 696         callsite: &CallSite<'tcx>,
 697         caller_body: &mut Body<'tcx>,
 698     ) -> Local {
 699         // Reuse the operand if it is a moved temporary.
 700         if let Operand::Move(place) = &arg
 701             && let Some(local) = place.as_local()
 702             && caller_body.local_kind(local) == LocalKind::Temp
 703         {
 704             return local;
 705         }
 706
 707         // Otherwise, create a temporary for the argument.
 708         trace!("creating temp for argument {:?}", arg);
 709         let arg_ty = arg.ty(caller_body, self.tcx);
 710         let local = self.new_call_temp(caller_body, callsite, arg_ty);
 711         caller_body[callsite.block].statements.push(Statement {
 712             source_info: callsite.source_info,
 713             kind: StatementKind::Assign(Box::new((Place::from(local), Rvalue::Use(arg)))),
 714         });
 715         local
 716     }
 717
 718     /// Introduces a new temporary into the caller body that is live for the duration of the call.
 719     fn new_call_temp(
 720         &self,
 721         caller_body: &mut Body<'tcx>,
 722         callsite: &CallSite<'tcx>,
 723         ty: Ty<'tcx>,
 724     ) -> Local {
 725         let local = caller_body.local_decls.push(LocalDecl::new(ty, callsite.source_info.span));
 726
 727         caller_body[callsite.block].statements.push(Statement {
 728             source_info: callsite.source_info,
 729             kind: StatementKind::StorageLive(local),
 730         });
 731
 732         if let Some(block) = callsite.target {
 733             caller_body[block].statements.insert(
 734                 0,
 735                 Statement {
 736                     source_info: callsite.source_info,
 737                     kind: StatementKind::StorageDead(local),
 738                 },
 739             );
 740         }
 741
 742         local
 743     }
 744 }
 745
 746 fn type_size_of<'tcx>(
 747     tcx: TyCtxt<'tcx>,
 748     param_env: ty::ParamEnv<'tcx>,
 749     ty: Ty<'tcx>,
 750 ) -> Option<u64> {
 751     tcx.layout_of(param_env.and(ty)).ok().map(|layout| layout.size.bytes())
 752 }
 753
 754 /// Verify that the callee body is compatible with the caller.
 755 ///
 756 /// This visitor mostly computes the inlining cost,
 757 /// but also needs to verify that types match because of normalization failure.
 758 struct CostChecker<'b, 'tcx> {
 759     tcx: TyCtxt<'tcx>,
 760     param_env: ParamEnv<'tcx>,
 761     cost: usize,
 762     callee_body: &'b Body<'tcx>,
 763     instance: ty::Instance<'tcx>,
 764     validation: Result<(), &'static str>,
 765 }
 766
 767 impl<'tcx> Visitor<'tcx> for CostChecker<'_, 'tcx> {
 768     fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) {
 769         // Don't count StorageLive/StorageDead in the inlining cost.
 770         match statement.kind {
 771             StatementKind::StorageLive(_)
 772             | StatementKind::StorageDead(_)
 773             | StatementKind::Deinit(_)
 774             | StatementKind::Nop => {}
 775             _ => self.cost += INSTR_COST,
 776         }
 777
 778         self.super_statement(statement, location);
 779     }
 780
 781     fn visit_terminator(&mut self, terminator: &Terminator<'tcx>, location: Location) {
 782         let tcx = self.tcx;
 783         match terminator.kind {
 784             TerminatorKind::Drop { ref place, unwind, .. }
 785             | TerminatorKind::DropAndReplace { ref place, unwind, .. } => {
 786                 // If the place doesn't actually need dropping, treat it like a regular goto.
 787                 let ty = self.instance.subst_mir(tcx, &place.ty(self.callee_body, tcx).ty);
 788                 if ty.needs_drop(tcx, self.param_env) {
 789                     self.cost += CALL_PENALTY;
 790                     if unwind.is_some() {
 791                         self.cost += LANDINGPAD_PENALTY;
 792                     }
 793                 } else {
 794                     self.cost += INSTR_COST;
 795                 }
 796             }
 797             TerminatorKind::Call { func: Operand::Constant(ref f), cleanup, .. } => {
 798                 let fn_ty = self.instance.subst_mir(tcx, &f.literal.ty());
 799                 self.cost += if let ty::FnDef(def_id, _) = *fn_ty.kind() && tcx.is_intrinsic(def_id) {
 800                     // Don't give intrinsics the extra penalty for calls
 801                     INSTR_COST
 802                 } else {
 803                     CALL_PENALTY
 804                 };
 805                 if cleanup.is_some() {
 806                     self.cost += LANDINGPAD_PENALTY;
 807                 }
 808             }
 809             TerminatorKind::Assert { cleanup, .. } => {
 810                 self.cost += CALL_PENALTY;
 811                 if cleanup.is_some() {
 812                     self.cost += LANDINGPAD_PENALTY;
 813                 }
 814             }
 815             TerminatorKind::Resume => self.cost += RESUME_PENALTY,
 816             TerminatorKind::InlineAsm { cleanup, .. } => {
 817                 self.cost += INSTR_COST;
 818                 if cleanup.is_some() {
 819                     self.cost += LANDINGPAD_PENALTY;
 820                 }
 821             }
 822             _ => self.cost += INSTR_COST,
 823         }
 824
 825         self.super_terminator(terminator, location);
 826     }
 827
 828     /// Count up the cost of local variables and temps, if we know the size
 829     /// use that, otherwise we use a moderately-large dummy cost.
 830     fn visit_local_decl(&mut self, local: Local, local_decl: &LocalDecl<'tcx>) {
 831         let tcx = self.tcx;
 832         let ptr_size = tcx.data_layout.pointer_size.bytes();
 833
 834         let ty = self.instance.subst_mir(tcx, &local_decl.ty);
 835         // Cost of the var is the size in machine-words, if we know
 836         // it.
 837         if let Some(size) = type_size_of(tcx, self.param_env, ty) {
 838             self.cost += ((size + ptr_size - 1) / ptr_size) as usize;
 839         } else {
 840             self.cost += UNKNOWN_SIZE_COST;
 841         }
 842
 843         self.super_local_decl(local, local_decl)
 844     }
 845
 846     /// This method duplicates code from MIR validation in an attempt to detect type mismatches due
 847     /// to normalization failure.
 848     fn visit_projection_elem(
 849         &mut self,
 850         local: Local,
 851         proj_base: &[PlaceElem<'tcx>],
 852         elem: PlaceElem<'tcx>,
 853         context: PlaceContext,
 854         location: Location,
 855     ) {
 856         if let ProjectionElem::Field(f, ty) = elem {
 857             let parent = Place { local, projection: self.tcx.intern_place_elems(proj_base) };
 858             let parent_ty = parent.ty(&self.callee_body.local_decls, self.tcx);
 859             let check_equal = |this: &mut Self, f_ty| {
 860                 if !util::is_equal_up_to_subtyping(this.tcx, this.param_env, ty, f_ty) {
 861                     trace!(?ty, ?f_ty);
 862                     this.validation = Err("failed to normalize projection type");
 863                     return;
 864                 }
 865             };
 866
 867             let kind = match parent_ty.ty.kind() {
 868                 &ty::Opaque(def_id, substs) => {
 869                     self.tcx.bound_type_of(def_id).subst(self.tcx, substs).kind()
 870                 }
 871                 kind => kind,
 872             };
 873
 874             match kind {
 875                 ty::Tuple(fields) => {
 876                     let Some(f_ty) = fields.get(f.as_usize()) else {
 877                         self.validation = Err("malformed MIR");
 878                         return;
 879                     };
 880                     check_equal(self, *f_ty);
 881                 }
 882                 ty::Adt(adt_def, substs) => {
 883                     let var = parent_ty.variant_index.unwrap_or(VariantIdx::from_u32(0));
 884                     let Some(field) = adt_def.variant(var).fields.get(f.as_usize()) else {
 885                         self.validation = Err("malformed MIR");
 886                         return;
 887                     };
 888                     check_equal(self, field.ty(self.tcx, substs));
 889                 }
 890                 ty::Closure(_, substs) => {
 891                     let substs = substs.as_closure();
 892                     let Some(f_ty) = substs.upvar_tys().nth(f.as_usize()) else {
 893                         self.validation = Err("malformed MIR");
 894                         return;
 895                     };
 896                     check_equal(self, f_ty);
 897                 }
 898                 &ty::Generator(def_id, substs, _) => {
 899                     let f_ty = if let Some(var) = parent_ty.variant_index {
 900                         let gen_body = if def_id == self.callee_body.source.def_id() {
 901                             self.callee_body
 902                         } else {
 903                             self.tcx.optimized_mir(def_id)
 904                         };
 905
 906                         let Some(layout) = gen_body.generator_layout() else {
 907                             self.validation = Err("malformed MIR");
 908                             return;
 909                         };
 910
 911                         let Some(&local) = layout.variant_fields[var].get(f) else {
 912                             self.validation = Err("malformed MIR");
 913                             return;
 914                         };
 915
 916                         let Some(&f_ty) = layout.field_tys.get(local) else {
 917                             self.validation = Err("malformed MIR");
 918                             return;
 919                         };
 920
 921                         f_ty
 922                     } else {
 923                         let Some(f_ty) = substs.as_generator().prefix_tys().nth(f.index()) else {
 924                             self.validation = Err("malformed MIR");
 925                             return;
 926                         };
 927
 928                         f_ty
 929                     };
 930
 931                     check_equal(self, f_ty);
 932                 }
 933                 _ => self.validation = Err("malformed MIR"),
 934             }
 935         }
 936
 937         self.super_projection_elem(local, proj_base, elem, context, location);
 938     }
 939 }
 940
 941 /**
 942  * Integrator.
 943  *
 944  * Integrates blocks from the callee function into the calling function.
 945  * Updates block indices, references to locals and other control flow
 946  * stuff.
 947 */
 948 struct Integrator<'a, 'tcx> {
 949     args: &'a [Local],
 950     new_locals: RangeFrom<Local>,
 951     new_scopes: RangeFrom<SourceScope>,
 952     new_blocks: RangeFrom<BasicBlock>,
 953     destination: Place<'tcx>,
 954     callsite_scope: SourceScopeData<'tcx>,
 955     callsite: &'a CallSite<'tcx>,
 956     cleanup_block: Option<BasicBlock>,
 957     in_cleanup_block: bool,
 958     tcx: TyCtxt<'tcx>,
 959     expn_data: LocalExpnId,
 960     always_live_locals: BitSet<Local>,
 961 }
 962
 963 impl Integrator<'_, '_> {
 964     fn map_local(&self, local: Local) -> Local {
 965         let new = if local == RETURN_PLACE {
 966             self.destination.local
 967         } else {
 968             let idx = local.index() - 1;
 969             if idx < self.args.len() {
 970                 self.args[idx]
 971             } else {
 972                 Local::new(self.new_locals.start.index() + (idx - self.args.len()))
 973             }
 974         };
 975         trace!("mapping local `{:?}` to `{:?}`", local, new);
 976         new
 977     }
 978
 979     fn map_scope(&self, scope: SourceScope) -> SourceScope {
 980         let new = SourceScope::new(self.new_scopes.start.index() + scope.index());
 981         trace!("mapping scope `{:?}` to `{:?}`", scope, new);
 982         new
 983     }
 984
 985     fn map_block(&self, block: BasicBlock) -> BasicBlock {
 986         let new = BasicBlock::new(self.new_blocks.start.index() + block.index());
 987         trace!("mapping block `{:?}` to `{:?}`", block, new);
 988         new
 989     }
 990
 991     fn map_unwind(&self, unwind: Option<BasicBlock>) -> Option<BasicBlock> {
 992         if self.in_cleanup_block {
 993             if unwind.is_some() {
 994                 bug!("cleanup on cleanup block");
 995             }
 996             return unwind;
 997         }
 998
 999         match unwind {
1000             Some(target) => Some(self.map_block(target)),
1001             // Add an unwind edge to the original call's cleanup block
1002             None => self.cleanup_block,
1003         }
1004     }
1005 }
1006
1007 impl<'tcx> MutVisitor<'tcx> for Integrator<'_, 'tcx> {
1008     fn tcx(&self) -> TyCtxt<'tcx> {
1009         self.tcx
1010     }
1011
1012     fn visit_local(&mut self, local: &mut Local, _ctxt: PlaceContext, _location: Location) {
1013         *local = self.map_local(*local);
1014     }
1015
1016     fn visit_source_scope_data(&mut self, scope_data: &mut SourceScopeData<'tcx>) {
1017         self.super_source_scope_data(scope_data);
1018         if scope_data.parent_scope.is_none() {
1019             // Attach the outermost callee scope as a child of the callsite
1020             // scope, via the `parent_scope` and `inlined_parent_scope` chains.
1021             scope_data.parent_scope = Some(self.callsite.source_info.scope);
1022             assert_eq!(scope_data.inlined_parent_scope, None);
1023             scope_data.inlined_parent_scope = if self.callsite_scope.inlined.is_some() {
1024                 Some(self.callsite.source_info.scope)
1025             } else {
1026                 self.callsite_scope.inlined_parent_scope
1027             };
1028
1029             // Mark the outermost callee scope as an inlined one.
1030             assert_eq!(scope_data.inlined, None);
1031             scope_data.inlined = Some((self.callsite.callee, self.callsite.source_info.span));
1032         } else if scope_data.inlined_parent_scope.is_none() {
1033             // Make it easy to find the scope with `inlined` set above.
1034             scope_data.inlined_parent_scope = Some(self.map_scope(OUTERMOST_SOURCE_SCOPE));
1035         }
1036     }
1037
1038     fn visit_source_scope(&mut self, scope: &mut SourceScope) {
1039         *scope = self.map_scope(*scope);
1040     }
1041
1042     fn visit_span(&mut self, span: &mut Span) {
1043         // Make sure that all spans track the fact that they were inlined.
1044         *span = span.fresh_expansion(self.expn_data);
1045     }
1046
1047     fn visit_place(&mut self, place: &mut Place<'tcx>, context: PlaceContext, location: Location) {
1048         for elem in place.projection {
1049             // FIXME: Make sure that return place is not used in an indexing projection, since it
1050             // won't be rebased as it is supposed to be.
1051             assert_ne!(ProjectionElem::Index(RETURN_PLACE), elem);
1052         }
1053
1054         // If this is the `RETURN_PLACE`, we need to rebase any projections onto it.
1055         let dest_proj_len = self.destination.projection.len();
1056         if place.local == RETURN_PLACE && dest_proj_len > 0 {
1057             let mut projs = Vec::with_capacity(dest_proj_len + place.projection.len());
1058             projs.extend(self.destination.projection);
1059             projs.extend(place.projection);
1060
1061             place.projection = self.tcx.intern_place_elems(&*projs);
1062         }
1063         // Handles integrating any locals that occur in the base
1064         // or projections
1065         self.super_place(place, context, location)
1066     }
1067
1068     fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
1069         self.in_cleanup_block = data.is_cleanup;
1070         self.super_basic_block_data(block, data);
1071         self.in_cleanup_block = false;
1072     }
1073
1074     fn visit_retag(&mut self, kind: &mut RetagKind, place: &mut Place<'tcx>, loc: Location) {
1075         self.super_retag(kind, place, loc);
1076
1077         // We have to patch all inlined retags to be aware that they are no longer
1078         // happening on function entry.
1079         if *kind == RetagKind::FnEntry {
1080             *kind = RetagKind::Default;
1081         }
1082     }
1083
1084     fn visit_statement(&mut self, statement: &mut Statement<'tcx>, location: Location) {
1085         if let StatementKind::StorageLive(local) | StatementKind::StorageDead(local) =
1086             statement.kind
1087         {
1088             self.always_live_locals.remove(local);
1089         }
1090         self.super_statement(statement, location);
1091     }
1092
1093     fn visit_terminator(&mut self, terminator: &mut Terminator<'tcx>, loc: Location) {
1094         // Don't try to modify the implicit `_0` access on return (`return` terminators are
1095         // replaced down below anyways).
1096         if !matches!(terminator.kind, TerminatorKind::Return) {
1097             self.super_terminator(terminator, loc);
1098         }
1099
1100         match terminator.kind {
1101             TerminatorKind::GeneratorDrop | TerminatorKind::Yield { .. } => bug!(),
1102             TerminatorKind::Goto { ref mut target } => {
1103                 *target = self.map_block(*target);
1104             }
1105             TerminatorKind::SwitchInt { ref mut targets, .. } => {
1106                 for tgt in targets.all_targets_mut() {
1107                     *tgt = self.map_block(*tgt);
1108                 }
1109             }
1110             TerminatorKind::Drop { ref mut target, ref mut unwind, .. }
1111             | TerminatorKind::DropAndReplace { ref mut target, ref mut unwind, .. } => {
1112                 *target = self.map_block(*target);
1113                 *unwind = self.map_unwind(*unwind);
1114             }
1115             TerminatorKind::Call { ref mut target, ref mut cleanup, .. } => {
1116                 if let Some(ref mut tgt) = *target {
1117                     *tgt = self.map_block(*tgt);
1118                 }
1119                 *cleanup = self.map_unwind(*cleanup);
1120             }
1121             TerminatorKind::Assert { ref mut target, ref mut cleanup, .. } => {
1122                 *target = self.map_block(*target);
1123                 *cleanup = self.map_unwind(*cleanup);
1124             }
1125             TerminatorKind::Return => {
1126                 terminator.kind = if let Some(tgt) = self.callsite.target {
1127                     TerminatorKind::Goto { target: tgt }
1128                 } else {
1129                     TerminatorKind::Unreachable
1130                 }
1131             }
1132             TerminatorKind::Resume => {
1133                 if let Some(tgt) = self.cleanup_block {
1134                     terminator.kind = TerminatorKind::Goto { target: tgt }
1135                 }
1136             }
1137             TerminatorKind::Abort => {}
1138             TerminatorKind::Unreachable => {}
1139             TerminatorKind::FalseEdge { ref mut real_target, ref mut imaginary_target } => {
1140                 *real_target = self.map_block(*real_target);
1141                 *imaginary_target = self.map_block(*imaginary_target);
1142             }
1143             TerminatorKind::FalseUnwind { real_target: _, unwind: _ } =>
1144             // see the ordering of passes in the optimized_mir query.
1145             {
1146                 bug!("False unwinds should have been removed before inlining")
1147             }
1148             TerminatorKind::InlineAsm { ref mut destination, ref mut cleanup, .. } => {
1149                 if let Some(ref mut tgt) = *destination {
1150                     *tgt = self.map_block(*tgt);
1151                 }
1152                 *cleanup = self.map_unwind(*cleanup);
1153             }
1154         }
1155     }
1156 }