1 //! Inlining pass for MIR functions
2 use crate::deref_separator
::deref_finder
;
3 use rustc_attr
::InlineAttr
;
4 use rustc_hir
::def_id
::DefId
;
5 use rustc_index
::bit_set
::BitSet
;
6 use rustc_index
::vec
::Idx
;
7 use rustc_middle
::middle
::codegen_fn_attrs
::{CodegenFnAttrFlags, CodegenFnAttrs}
;
8 use rustc_middle
::mir
::visit
::*;
9 use rustc_middle
::mir
::*;
10 use rustc_middle
::ty
::{self, Instance, InstanceDef, ParamEnv, Ty, TyCtxt}
;
11 use rustc_session
::config
::OptLevel
;
12 use rustc_span
::{hygiene::ExpnKind, ExpnData, LocalExpnId, Span}
;
13 use rustc_target
::abi
::VariantIdx
;
14 use rustc_target
::spec
::abi
::Abi
;
16 use crate::simplify
::{remove_dead_blocks, CfgSimplifier}
;
20 use std
::ops
::{Range, RangeFrom}
;
24 const INSTR_COST
: usize = 5;
25 const CALL_PENALTY
: usize = 25;
26 const LANDINGPAD_PENALTY
: usize = 50;
27 const RESUME_PENALTY
: usize = 45;
29 const UNKNOWN_SIZE_COST
: usize = 10;
31 const TOP_DOWN_DEPTH_LIMIT
: usize = 5;
35 #[derive(Copy, Clone, Debug)]
36 struct CallSite
<'tcx
> {
37 callee
: Instance
<'tcx
>,
38 fn_sig
: ty
::PolyFnSig
<'tcx
>,
40 target
: Option
<BasicBlock
>,
41 source_info
: SourceInfo
,
44 impl<'tcx
> MirPass
<'tcx
> for Inline
{
45 fn is_enabled(&self, sess
: &rustc_session
::Session
) -> bool
{
46 if let Some(enabled
) = sess
.opts
.unstable_opts
.inline_mir
{
50 match sess
.mir_opt_level() {
53 (sess
.opts
.optimize
== OptLevel
::Default
54 || sess
.opts
.optimize
== OptLevel
::Aggressive
)
55 && sess
.opts
.incremental
== None
61 fn run_pass(&self, tcx
: TyCtxt
<'tcx
>, body
: &mut Body
<'tcx
>) {
62 let span
= trace_span
!("inline", body
= %tcx
.def_path_str(body
.source
.def_id()));
63 let _guard
= span
.enter();
64 if inline(tcx
, body
) {
65 debug
!("running simplify cfg on {:?}", body
.source
);
66 CfgSimplifier
::new(body
).simplify();
67 remove_dead_blocks(tcx
, body
);
68 deref_finder(tcx
, body
);
73 fn inline
<'tcx
>(tcx
: TyCtxt
<'tcx
>, body
: &mut Body
<'tcx
>) -> bool
{
74 let def_id
= body
.source
.def_id().expect_local();
76 // Only do inlining into fn bodies.
77 if !tcx
.hir().body_owner_kind(def_id
).is_fn_or_closure() {
80 if body
.source
.promoted
.is_some() {
83 // Avoid inlining into generators, since their `optimized_mir` is used for layout computation,
84 // which can create a cycle, even when no attempt is made to inline the function in the other
86 if body
.generator
.is_some() {
90 let param_env
= tcx
.param_env_reveal_all_normalized(def_id
);
92 let mut this
= Inliner
{
95 codegen_fn_attrs
: tcx
.codegen_fn_attrs(def_id
),
99 let blocks
= BasicBlock
::new(0)..body
.basic_blocks
.next_index();
100 this
.process_blocks(body
, blocks
);
104 struct Inliner
<'tcx
> {
106 param_env
: ParamEnv
<'tcx
>,
107 /// Caller codegen attributes.
108 codegen_fn_attrs
: &'tcx CodegenFnAttrs
,
109 /// Stack of inlined instances.
110 /// We only check the `DefId` and not the substs because we want to
111 /// avoid inlining cases of polymorphic recursion.
112 /// The number of `DefId`s is finite, so checking history is enough
113 /// to ensure that we do not loop endlessly while inlining.
115 /// Indicates that the caller body has been modified.
119 impl<'tcx
> Inliner
<'tcx
> {
120 fn process_blocks(&mut self, caller_body
: &mut Body
<'tcx
>, blocks
: Range
<BasicBlock
>) {
121 // How many callsites in this body are we allowed to inline? We need to limit this in order
122 // to prevent super-linear growth in MIR size
123 let inline_limit
= match self.history
.len() {
125 1..=TOP_DOWN_DEPTH_LIMIT
=> 1,
128 let mut inlined_count
= 0;
130 let bb_data
= &caller_body
[bb
];
131 if bb_data
.is_cleanup
{
135 let Some(callsite
) = self.resolve_callsite(caller_body
, bb
, bb_data
) else {
139 let span
= trace_span
!("process_blocks", %callsite
.callee
, ?bb
);
140 let _guard
= span
.enter();
142 match self.try_inlining(caller_body
, &callsite
) {
144 debug
!("not-inlined {} [{}]", callsite
.callee
, reason
);
148 debug
!("inlined {}", callsite
.callee
);
151 if inlined_count
== inline_limit
{
154 self.history
.push(callsite
.callee
.def_id());
155 self.process_blocks(caller_body
, new_blocks
);
162 /// Attempts to inline a callsite into the caller body. When successful returns basic blocks
163 /// containing the inlined body. Otherwise returns an error describing why inlining didn't take
167 caller_body
: &mut Body
<'tcx
>,
168 callsite
: &CallSite
<'tcx
>,
169 ) -> Result
<std
::ops
::Range
<BasicBlock
>, &'
static str> {
170 let callee_attrs
= self.tcx
.codegen_fn_attrs(callsite
.callee
.def_id());
171 self.check_codegen_attributes(callsite
, callee_attrs
)?
;
172 self.check_mir_is_available(caller_body
, &callsite
.callee
)?
;
173 let callee_body
= self.tcx
.instance_mir(callsite
.callee
.def
);
174 self.check_mir_body(callsite
, callee_body
, callee_attrs
)?
;
176 if !self.tcx
.consider_optimizing(|| {
177 format
!("Inline {:?} into {:?}", callsite
.callee
, caller_body
.source
)
179 return Err("optimization fuel exhausted");
182 let Ok(callee_body
) = callsite
.callee
.try_subst_mir_and_normalize_erasing_regions(
187 return Err("failed to normalize callee body");
190 // Check call signature compatibility.
191 // Normally, this shouldn't be required, but trait normalization failure can create a
193 let terminator
= caller_body
[callsite
.block
].terminator
.as_ref().unwrap();
194 let TerminatorKind
::Call { args, destination, .. }
= &terminator
.kind
else { bug!() }
;
195 let destination_ty
= destination
.ty(&caller_body
.local_decls
, self.tcx
).ty
;
196 let output_type
= callee_body
.return_ty();
197 if !util
::is_subtype(self.tcx
, self.param_env
, output_type
, destination_ty
) {
198 trace
!(?output_type
, ?destination_ty
);
199 return Err("failed to normalize return type");
201 if callsite
.fn_sig
.abi() == Abi
::RustCall
{
202 let (arg_tuple
, skipped_args
) = match &args
[..] {
203 [arg_tuple
] => (arg_tuple
, 0),
204 [_
, arg_tuple
] => (arg_tuple
, 1),
205 _
=> bug
!("Expected `rust-call` to have 1 or 2 args"),
208 let arg_tuple_ty
= arg_tuple
.ty(&caller_body
.local_decls
, self.tcx
);
209 let ty
::Tuple(arg_tuple_tys
) = arg_tuple_ty
.kind() else {
210 bug
!("Closure arguments are not passed as a tuple");
213 for (arg_ty
, input
) in
214 arg_tuple_tys
.iter().zip(callee_body
.args_iter().skip(skipped_args
))
216 let input_type
= callee_body
.local_decls
[input
].ty
;
217 if !util
::is_subtype(self.tcx
, self.param_env
, input_type
, arg_ty
) {
218 trace
!(?arg_ty
, ?input_type
);
219 return Err("failed to normalize tuple argument type");
223 for (arg
, input
) in args
.iter().zip(callee_body
.args_iter()) {
224 let input_type
= callee_body
.local_decls
[input
].ty
;
225 let arg_ty
= arg
.ty(&caller_body
.local_decls
, self.tcx
);
226 if !util
::is_subtype(self.tcx
, self.param_env
, input_type
, arg_ty
) {
227 trace
!(?arg_ty
, ?input_type
);
228 return Err("failed to normalize argument type");
233 let old_blocks
= caller_body
.basic_blocks
.next_index();
234 self.inline_call(caller_body
, &callsite
, callee_body
);
235 let new_blocks
= old_blocks
..caller_body
.basic_blocks
.next_index();
240 fn check_mir_is_available(
242 caller_body
: &Body
<'tcx
>,
243 callee
: &Instance
<'tcx
>,
244 ) -> Result
<(), &'
static str> {
245 let caller_def_id
= caller_body
.source
.def_id();
246 let callee_def_id
= callee
.def_id();
247 if callee_def_id
== caller_def_id
{
248 return Err("self-recursion");
252 InstanceDef
::Item(_
) => {
253 // If there is no MIR available (either because it was not in metadata or
254 // because it has no MIR because it's an extern function), then the inliner
255 // won't cause cycles on this.
256 if !self.tcx
.is_mir_available(callee_def_id
) {
257 return Err("item MIR unavailable");
260 // These have no own callable MIR.
261 InstanceDef
::Intrinsic(_
) | InstanceDef
::Virtual(..) => {
262 return Err("instance without MIR (intrinsic / virtual)");
264 // This cannot result in an immediate cycle since the callee MIR is a shim, which does
265 // not get any optimizations run on it. Any subsequent inlining may cause cycles, but we
266 // do not need to catch this here, we can wait until the inliner decides to continue
267 // inlining a second time.
268 InstanceDef
::VTableShim(_
)
269 | InstanceDef
::ReifyShim(_
)
270 | InstanceDef
::FnPtrShim(..)
271 | InstanceDef
::ClosureOnceShim { .. }
272 | InstanceDef
::DropGlue(..)
273 | InstanceDef
::CloneShim(..) => return Ok(()),
276 if self.tcx
.is_constructor(callee_def_id
) {
277 trace
!("constructors always have MIR");
278 // Constructor functions cannot cause a query cycle.
282 if callee_def_id
.is_local() {
283 // Avoid a cycle here by only using `instance_mir` only if we have
284 // a lower `DefPathHash` than the callee. This ensures that the callee will
285 // not inline us. This trick even works with incremental compilation,
286 // since `DefPathHash` is stable.
287 if self.tcx
.def_path_hash(caller_def_id
).local_hash()
288 < self.tcx
.def_path_hash(callee_def_id
).local_hash()
293 // If we know for sure that the function we're calling will itself try to
294 // call us, then we avoid inlining that function.
295 if self.tcx
.mir_callgraph_reachable((*callee
, caller_def_id
.expect_local())) {
296 return Err("caller might be reachable from callee (query cycle avoidance)");
301 // This cannot result in an immediate cycle since the callee MIR is from another crate
302 // and is already optimized. Any subsequent inlining may cause cycles, but we do
303 // not need to catch this here, we can wait until the inliner decides to continue
304 // inlining a second time.
305 trace
!("functions from other crates always have MIR");
312 caller_body
: &Body
<'tcx
>,
314 bb_data
: &BasicBlockData
<'tcx
>,
315 ) -> Option
<CallSite
<'tcx
>> {
316 // Only consider direct calls to functions
317 let terminator
= bb_data
.terminator();
318 if let TerminatorKind
::Call { ref func, target, fn_span, .. }
= terminator
.kind
{
319 let func_ty
= func
.ty(caller_body
, self.tcx
);
320 if let ty
::FnDef(def_id
, substs
) = *func_ty
.kind() {
321 // To resolve an instance its substs have to be fully normalized.
322 let substs
= self.tcx
.try_normalize_erasing_regions(self.param_env
, substs
).ok()?
;
324 Instance
::resolve(self.tcx
, self.param_env
, def_id
, substs
).ok().flatten()?
;
326 if let InstanceDef
::Virtual(..) | InstanceDef
::Intrinsic(_
) = callee
.def
{
330 if self.history
.contains(&callee
.def_id()) {
334 let fn_sig
= self.tcx
.bound_fn_sig(def_id
).subst(self.tcx
, substs
);
335 let source_info
= SourceInfo { span: fn_span, ..terminator.source_info }
;
337 return Some(CallSite { callee, fn_sig, block: bb, target, source_info }
);
344 /// Returns an error if inlining is not possible based on codegen attributes alone. A success
345 /// indicates that inlining decision should be based on other criteria.
346 fn check_codegen_attributes(
348 callsite
: &CallSite
<'tcx
>,
349 callee_attrs
: &CodegenFnAttrs
,
350 ) -> Result
<(), &'
static str> {
351 match callee_attrs
.inline
{
352 InlineAttr
::Never
=> return Err("never inline hint"),
353 InlineAttr
::Always
| InlineAttr
::Hint
=> {}
354 InlineAttr
::None
=> {
355 if self.tcx
.sess
.mir_opt_level() <= 2 {
356 return Err("at mir-opt-level=2, only #[inline] is inlined");
361 // Only inline local functions if they would be eligible for cross-crate
362 // inlining. This is to ensure that the final crate doesn't have MIR that
363 // reference unexported symbols
364 if callsite
.callee
.def_id().is_local() {
365 let is_generic
= callsite
.callee
.substs
.non_erasable_generics().next().is_some();
366 if !is_generic
&& !callee_attrs
.requests_inline() {
367 return Err("not exported");
371 if callsite
.fn_sig
.c_variadic() {
372 return Err("C variadic");
375 if callee_attrs
.flags
.contains(CodegenFnAttrFlags
::COLD
) {
379 if callee_attrs
.no_sanitize
!= self.codegen_fn_attrs
.no_sanitize
{
380 return Err("incompatible sanitizer set");
383 // Two functions are compatible if the callee has no attribute (meaning
384 // that it's codegen agnostic), or sets an attribute that is identical
385 // to this function's attribute.
386 if callee_attrs
.instruction_set
.is_some()
387 && callee_attrs
.instruction_set
!= self.codegen_fn_attrs
.instruction_set
389 return Err("incompatible instruction set");
392 for feature
in &callee_attrs
.target_features
{
393 if !self.codegen_fn_attrs
.target_features
.contains(feature
) {
394 return Err("incompatible target feature");
401 /// Returns inlining decision that is based on the examination of callee MIR body.
402 /// Assumes that codegen attributes have been checked for compatibility already.
403 #[instrument(level = "debug", skip(self, callee_body))]
406 callsite
: &CallSite
<'tcx
>,
407 callee_body
: &Body
<'tcx
>,
408 callee_attrs
: &CodegenFnAttrs
,
409 ) -> Result
<(), &'
static str> {
412 let mut threshold
= if callee_attrs
.requests_inline() {
413 self.tcx
.sess
.opts
.unstable_opts
.inline_mir_hint_threshold
.unwrap_or(100)
415 self.tcx
.sess
.opts
.unstable_opts
.inline_mir_threshold
.unwrap_or(50)
418 // Give a bonus functions with a small number of blocks,
419 // We normally have two or three blocks for even
420 // very small functions.
421 if callee_body
.basic_blocks
.len() <= 3 {
422 threshold
+= threshold
/ 4;
424 debug
!(" final inline threshold = {}", threshold
);
426 // FIXME: Give a bonus to functions with only a single caller
427 let diverges
= matches
!(
428 callee_body
.basic_blocks
[START_BLOCK
].terminator().kind
,
429 TerminatorKind
::Unreachable
| TerminatorKind
::Call { target: None, .. }
431 if diverges
&& !matches
!(callee_attrs
.inline
, InlineAttr
::Always
) {
432 return Err("callee diverges unconditionally");
435 let mut checker
= CostChecker
{
437 param_env
: self.param_env
,
438 instance
: callsite
.callee
,
444 // Traverse the MIR manually so we can account for the effects of inlining on the CFG.
445 let mut work_list
= vec
![START_BLOCK
];
446 let mut visited
= BitSet
::new_empty(callee_body
.basic_blocks
.len());
447 while let Some(bb
) = work_list
.pop() {
448 if !visited
.insert(bb
.index()) {
452 let blk
= &callee_body
.basic_blocks
[bb
];
453 checker
.visit_basic_block_data(bb
, blk
);
455 let term
= blk
.terminator();
456 if let TerminatorKind
::Drop { ref place, target, unwind }
457 | TerminatorKind
::DropAndReplace { ref place, target, unwind, .. }
= term
.kind
459 work_list
.push(target
);
461 // If the place doesn't actually need dropping, treat it like a regular goto.
462 let ty
= callsite
.callee
.subst_mir(self.tcx
, &place
.ty(callee_body
, tcx
).ty
);
463 if ty
.needs_drop(tcx
, self.param_env
) && let Some(unwind
) = unwind
{
464 work_list
.push(unwind
);
466 } else if callee_attrs
.instruction_set
!= self.codegen_fn_attrs
.instruction_set
467 && matches
!(term
.kind
, TerminatorKind
::InlineAsm { .. }
)
469 // During the attribute checking stage we allow a callee with no
470 // instruction_set assigned to count as compatible with a function that does
471 // assign one. However, during this stage we require an exact match when any
472 // inline-asm is detected. LLVM will still possibly do an inline later on
473 // if the no-attribute function ends up with the same instruction set anyway.
474 return Err("Cannot move inline-asm across instruction sets");
476 work_list
.extend(term
.successors())
480 // Count up the cost of local variables and temps, if we know the size
481 // use that, otherwise we use a moderately-large dummy cost.
482 for v
in callee_body
.vars_and_temps_iter() {
483 checker
.visit_local_decl(v
, &callee_body
.local_decls
[v
]);
486 // Abort if type validation found anything fishy.
489 let cost
= checker
.cost
;
490 if let InlineAttr
::Always
= callee_attrs
.inline
{
491 debug
!("INLINING {:?} because inline(always) [cost={}]", callsite
, cost
);
493 } else if cost
<= threshold
{
494 debug
!("INLINING {:?} [cost={} <= threshold={}]", callsite
, cost
, threshold
);
497 debug
!("NOT inlining {:?} [cost={} > threshold={}]", callsite
, cost
, threshold
);
498 Err("cost above threshold")
504 caller_body
: &mut Body
<'tcx
>,
505 callsite
: &CallSite
<'tcx
>,
506 mut callee_body
: Body
<'tcx
>,
508 let terminator
= caller_body
[callsite
.block
].terminator
.take().unwrap();
509 match terminator
.kind
{
510 TerminatorKind
::Call { args, destination, cleanup, .. }
=> {
511 // If the call is something like `a[*i] = f(i)`, where
512 // `i : &mut usize`, then just duplicating the `a[*i]`
513 // Place could result in two different locations if `f`
514 // writes to `i`. To prevent this we need to create a temporary
515 // borrow of the place and pass the destination as `*temp` instead.
516 fn dest_needs_borrow(place
: Place
<'_
>) -> bool
{
517 for elem
in place
.projection
.iter() {
519 ProjectionElem
::Deref
| ProjectionElem
::Index(_
) => return true,
527 let dest
= if dest_needs_borrow(destination
) {
528 trace
!("creating temp for return destination");
529 let dest
= Rvalue
::Ref(
530 self.tcx
.lifetimes
.re_erased
,
531 BorrowKind
::Mut { allow_two_phase_borrow: false }
,
534 let dest_ty
= dest
.ty(caller_body
, self.tcx
);
535 let temp
= Place
::from(self.new_call_temp(caller_body
, &callsite
, dest_ty
));
536 caller_body
[callsite
.block
].statements
.push(Statement
{
537 source_info
: callsite
.source_info
,
538 kind
: StatementKind
::Assign(Box
::new((temp
, dest
))),
540 self.tcx
.mk_place_deref(temp
)
545 // Always create a local to hold the destination, as `RETURN_PLACE` may appear
546 // where a full `Place` is not allowed.
547 let (remap_destination
, destination_local
) = if let Some(d
) = dest
.as_local() {
555 destination
.ty(caller_body
, self.tcx
).ty
,
560 // Copy the arguments if needed.
561 let args
: Vec
<_
> = self.make_call_args(args
, &callsite
, caller_body
, &callee_body
);
563 let mut expn_data
= ExpnData
::default(
565 callsite
.source_info
.span
,
566 self.tcx
.sess
.edition(),
570 expn_data
.def_site
= callee_body
.span
;
572 self.tcx
.with_stable_hashing_context(|hcx
| LocalExpnId
::fresh(expn_data
, hcx
));
573 let mut integrator
= Integrator
{
575 new_locals
: Local
::new(caller_body
.local_decls
.len())..,
576 new_scopes
: SourceScope
::new(caller_body
.source_scopes
.len())..,
577 new_blocks
: BasicBlock
::new(caller_body
.basic_blocks
.len())..,
578 destination
: destination_local
,
579 callsite_scope
: caller_body
.source_scopes
[callsite
.source_info
.scope
].clone(),
581 cleanup_block
: cleanup
,
582 in_cleanup_block
: false,
585 always_live_locals
: BitSet
::new_filled(callee_body
.local_decls
.len()),
588 // Map all `Local`s, `SourceScope`s and `BasicBlock`s to new ones
589 // (or existing ones, in a few special cases) in the caller.
590 integrator
.visit_body(&mut callee_body
);
592 // If there are any locals without storage markers, give them storage only for the
593 // duration of the call.
594 for local
in callee_body
.vars_and_temps_iter() {
595 if !callee_body
.local_decls
[local
].internal
596 && integrator
.always_live_locals
.contains(local
)
598 let new_local
= integrator
.map_local(local
);
599 caller_body
[callsite
.block
].statements
.push(Statement
{
600 source_info
: callsite
.source_info
,
601 kind
: StatementKind
::StorageLive(new_local
),
605 if let Some(block
) = callsite
.target
{
606 // To avoid repeated O(n) insert, push any new statements to the end and rotate
609 if remap_destination
{
610 caller_body
[block
].statements
.push(Statement
{
611 source_info
: callsite
.source_info
,
612 kind
: StatementKind
::Assign(Box
::new((
614 Rvalue
::Use(Operand
::Move(destination_local
.into())),
619 for local
in callee_body
.vars_and_temps_iter().rev() {
620 if !callee_body
.local_decls
[local
].internal
621 && integrator
.always_live_locals
.contains(local
)
623 let new_local
= integrator
.map_local(local
);
624 caller_body
[block
].statements
.push(Statement
{
625 source_info
: callsite
.source_info
,
626 kind
: StatementKind
::StorageDead(new_local
),
631 caller_body
[block
].statements
.rotate_right(n
);
634 // Insert all of the (mapped) parts of the callee body into the caller.
635 caller_body
.local_decls
.extend(callee_body
.drain_vars_and_temps());
636 caller_body
.source_scopes
.extend(&mut callee_body
.source_scopes
.drain(..));
637 caller_body
.var_debug_info
.append(&mut callee_body
.var_debug_info
);
638 caller_body
.basic_blocks_mut().extend(callee_body
.basic_blocks_mut().drain(..));
640 caller_body
[callsite
.block
].terminator
= Some(Terminator
{
641 source_info
: callsite
.source_info
,
642 kind
: TerminatorKind
::Goto { target: integrator.map_block(START_BLOCK) }
,
645 // Copy only unevaluated constants from the callee_body into the caller_body.
646 // Although we are only pushing `ConstKind::Unevaluated` consts to
647 // `required_consts`, here we may not only have `ConstKind::Unevaluated`
648 // because we are calling `subst_and_normalize_erasing_regions`.
649 caller_body
.required_consts
.extend(
650 callee_body
.required_consts
.iter().copied().filter(|&ct
| match ct
.literal
{
651 ConstantKind
::Ty(_
) => {
652 bug
!("should never encounter ty::UnevaluatedConst in `required_consts`")
654 ConstantKind
::Val(..) | ConstantKind
::Unevaluated(..) => true,
658 kind
=> bug
!("unexpected terminator kind {:?}", kind
),
664 args
: Vec
<Operand
<'tcx
>>,
665 callsite
: &CallSite
<'tcx
>,
666 caller_body
: &mut Body
<'tcx
>,
667 callee_body
: &Body
<'tcx
>,
671 // There is a bit of a mismatch between the *caller* of a closure and the *callee*.
672 // The caller provides the arguments wrapped up in a tuple:
674 // tuple_tmp = (a, b, c)
675 // Fn::call(closure_ref, tuple_tmp)
677 // meanwhile the closure body expects the arguments (here, `a`, `b`, and `c`)
678 // as distinct arguments. (This is the "rust-call" ABI hack.) Normally, codegen has
679 // the job of unpacking this tuple. But here, we are codegen. =) So we want to create
682 // [closure_ref, tuple_tmp.0, tuple_tmp.1, tuple_tmp.2]
684 // Except for one tiny wrinkle: we don't actually want `tuple_tmp.0`. It's more convenient
685 // if we "spill" that into *another* temporary, so that we can map the argument
686 // variable in the callee MIR directly to an argument variable on our side.
687 // So we introduce temporaries like:
689 // tmp0 = tuple_tmp.0
690 // tmp1 = tuple_tmp.1
691 // tmp2 = tuple_tmp.2
693 // and the vector is `[closure_ref, tmp0, tmp1, tmp2]`.
694 if callsite
.fn_sig
.abi() == Abi
::RustCall
&& callee_body
.spread_arg
.is_none() {
695 let mut args
= args
.into_iter();
696 let self_
= self.create_temp_if_necessary(args
.next().unwrap(), callsite
, caller_body
);
697 let tuple
= self.create_temp_if_necessary(args
.next().unwrap(), callsite
, caller_body
);
698 assert
!(args
.next().is_none());
700 let tuple
= Place
::from(tuple
);
701 let ty
::Tuple(tuple_tys
) = tuple
.ty(caller_body
, tcx
).ty
.kind() else {
702 bug
!("Closure arguments are not passed as a tuple");
705 // The `closure_ref` in our example above.
706 let closure_ref_arg
= iter
::once(self_
);
708 // The `tmp0`, `tmp1`, and `tmp2` in our example above.
709 let tuple_tmp_args
= tuple_tys
.iter().enumerate().map(|(i
, ty
)| {
710 // This is e.g., `tuple_tmp.0` in our example above.
711 let tuple_field
= Operand
::Move(tcx
.mk_place_field(tuple
, Field
::new(i
), ty
));
713 // Spill to a local to make e.g., `tmp0`.
714 self.create_temp_if_necessary(tuple_field
, callsite
, caller_body
)
717 closure_ref_arg
.chain(tuple_tmp_args
).collect()
720 .map(|a
| self.create_temp_if_necessary(a
, callsite
, caller_body
))
725 /// If `arg` is already a temporary, returns it. Otherwise, introduces a fresh
726 /// temporary `T` and an instruction `T = arg`, and returns `T`.
727 fn create_temp_if_necessary(
730 callsite
: &CallSite
<'tcx
>,
731 caller_body
: &mut Body
<'tcx
>,
733 // Reuse the operand if it is a moved temporary.
734 if let Operand
::Move(place
) = &arg
735 && let Some(local
) = place
.as_local()
736 && caller_body
.local_kind(local
) == LocalKind
::Temp
741 // Otherwise, create a temporary for the argument.
742 trace
!("creating temp for argument {:?}", arg
);
743 let arg_ty
= arg
.ty(caller_body
, self.tcx
);
744 let local
= self.new_call_temp(caller_body
, callsite
, arg_ty
);
745 caller_body
[callsite
.block
].statements
.push(Statement
{
746 source_info
: callsite
.source_info
,
747 kind
: StatementKind
::Assign(Box
::new((Place
::from(local
), Rvalue
::Use(arg
)))),
752 /// Introduces a new temporary into the caller body that is live for the duration of the call.
755 caller_body
: &mut Body
<'tcx
>,
756 callsite
: &CallSite
<'tcx
>,
759 let local
= caller_body
.local_decls
.push(LocalDecl
::new(ty
, callsite
.source_info
.span
));
761 caller_body
[callsite
.block
].statements
.push(Statement
{
762 source_info
: callsite
.source_info
,
763 kind
: StatementKind
::StorageLive(local
),
766 if let Some(block
) = callsite
.target
{
767 caller_body
[block
].statements
.insert(
770 source_info
: callsite
.source_info
,
771 kind
: StatementKind
::StorageDead(local
),
780 fn type_size_of
<'tcx
>(
782 param_env
: ty
::ParamEnv
<'tcx
>,
785 tcx
.layout_of(param_env
.and(ty
)).ok().map(|layout
| layout
.size
.bytes())
788 /// Verify that the callee body is compatible with the caller.
790 /// This visitor mostly computes the inlining cost,
791 /// but also needs to verify that types match because of normalization failure.
792 struct CostChecker
<'b
, 'tcx
> {
794 param_env
: ParamEnv
<'tcx
>,
796 callee_body
: &'b Body
<'tcx
>,
797 instance
: ty
::Instance
<'tcx
>,
798 validation
: Result
<(), &'
static str>,
801 impl<'tcx
> Visitor
<'tcx
> for CostChecker
<'_
, 'tcx
> {
802 fn visit_statement(&mut self, statement
: &Statement
<'tcx
>, location
: Location
) {
803 // Don't count StorageLive/StorageDead in the inlining cost.
804 match statement
.kind
{
805 StatementKind
::StorageLive(_
)
806 | StatementKind
::StorageDead(_
)
807 | StatementKind
::Deinit(_
)
808 | StatementKind
::Nop
=> {}
809 _
=> self.cost
+= INSTR_COST
,
812 self.super_statement(statement
, location
);
815 fn visit_terminator(&mut self, terminator
: &Terminator
<'tcx
>, location
: Location
) {
817 match terminator
.kind
{
818 TerminatorKind
::Drop { ref place, unwind, .. }
819 | TerminatorKind
::DropAndReplace { ref place, unwind, .. }
=> {
820 // If the place doesn't actually need dropping, treat it like a regular goto.
821 let ty
= self.instance
.subst_mir(tcx
, &place
.ty(self.callee_body
, tcx
).ty
);
822 if ty
.needs_drop(tcx
, self.param_env
) {
823 self.cost
+= CALL_PENALTY
;
824 if unwind
.is_some() {
825 self.cost
+= LANDINGPAD_PENALTY
;
828 self.cost
+= INSTR_COST
;
831 TerminatorKind
::Call { func: Operand::Constant(ref f), cleanup, .. }
=> {
832 let fn_ty
= self.instance
.subst_mir(tcx
, &f
.literal
.ty());
833 self.cost
+= if let ty
::FnDef(def_id
, _
) = *fn_ty
.kind() && tcx
.is_intrinsic(def_id
) {
834 // Don't give intrinsics the extra penalty for calls
839 if cleanup
.is_some() {
840 self.cost
+= LANDINGPAD_PENALTY
;
843 TerminatorKind
::Assert { cleanup, .. }
=> {
844 self.cost
+= CALL_PENALTY
;
845 if cleanup
.is_some() {
846 self.cost
+= LANDINGPAD_PENALTY
;
849 TerminatorKind
::Resume
=> self.cost
+= RESUME_PENALTY
,
850 TerminatorKind
::InlineAsm { cleanup, .. }
=> {
851 self.cost
+= INSTR_COST
;
852 if cleanup
.is_some() {
853 self.cost
+= LANDINGPAD_PENALTY
;
856 _
=> self.cost
+= INSTR_COST
,
859 self.super_terminator(terminator
, location
);
862 /// Count up the cost of local variables and temps, if we know the size
863 /// use that, otherwise we use a moderately-large dummy cost.
864 fn visit_local_decl(&mut self, local
: Local
, local_decl
: &LocalDecl
<'tcx
>) {
866 let ptr_size
= tcx
.data_layout
.pointer_size
.bytes();
868 let ty
= self.instance
.subst_mir(tcx
, &local_decl
.ty
);
869 // Cost of the var is the size in machine-words, if we know
871 if let Some(size
) = type_size_of(tcx
, self.param_env
, ty
) {
872 self.cost
+= ((size
+ ptr_size
- 1) / ptr_size
) as usize;
874 self.cost
+= UNKNOWN_SIZE_COST
;
877 self.super_local_decl(local
, local_decl
)
880 /// This method duplicates code from MIR validation in an attempt to detect type mismatches due
881 /// to normalization failure.
882 fn visit_projection_elem(
885 proj_base
: &[PlaceElem
<'tcx
>],
886 elem
: PlaceElem
<'tcx
>,
887 context
: PlaceContext
,
890 if let ProjectionElem
::Field(f
, ty
) = elem
{
891 let parent
= Place { local, projection: self.tcx.intern_place_elems(proj_base) }
;
892 let parent_ty
= parent
.ty(&self.callee_body
.local_decls
, self.tcx
);
893 let check_equal
= |this
: &mut Self, f_ty
| {
894 if !util
::is_equal_up_to_subtyping(this
.tcx
, this
.param_env
, ty
, f_ty
) {
896 this
.validation
= Err("failed to normalize projection type");
901 let kind
= match parent_ty
.ty
.kind() {
902 &ty
::Alias(ty
::Opaque
, ty
::AliasTy { def_id, substs, .. }
) => {
903 self.tcx
.bound_type_of(def_id
).subst(self.tcx
, substs
).kind()
909 ty
::Tuple(fields
) => {
910 let Some(f_ty
) = fields
.get(f
.as_usize()) else {
911 self.validation
= Err("malformed MIR");
914 check_equal(self, *f_ty
);
916 ty
::Adt(adt_def
, substs
) => {
917 let var
= parent_ty
.variant_index
.unwrap_or(VariantIdx
::from_u32(0));
918 let Some(field
) = adt_def
.variant(var
).fields
.get(f
.as_usize()) else {
919 self.validation
= Err("malformed MIR");
922 check_equal(self, field
.ty(self.tcx
, substs
));
924 ty
::Closure(_
, substs
) => {
925 let substs
= substs
.as_closure();
926 let Some(f_ty
) = substs
.upvar_tys().nth(f
.as_usize()) else {
927 self.validation
= Err("malformed MIR");
930 check_equal(self, f_ty
);
932 &ty
::Generator(def_id
, substs
, _
) => {
933 let f_ty
= if let Some(var
) = parent_ty
.variant_index
{
934 let gen_body
= if def_id
== self.callee_body
.source
.def_id() {
937 self.tcx
.optimized_mir(def_id
)
940 let Some(layout
) = gen_body
.generator_layout() else {
941 self.validation
= Err("malformed MIR");
945 let Some(&local
) = layout
.variant_fields
[var
].get(f
) else {
946 self.validation
= Err("malformed MIR");
950 let Some(&f_ty
) = layout
.field_tys
.get(local
) else {
951 self.validation
= Err("malformed MIR");
957 let Some(f_ty
) = substs
.as_generator().prefix_tys().nth(f
.index()) else {
958 self.validation
= Err("malformed MIR");
965 check_equal(self, f_ty
);
967 _
=> self.validation
= Err("malformed MIR"),
971 self.super_projection_elem(local
, proj_base
, elem
, context
, location
);
978 * Integrates blocks from the callee function into the calling function.
979 * Updates block indices, references to locals and other control flow
982 struct Integrator
<'a
, 'tcx
> {
984 new_locals
: RangeFrom
<Local
>,
985 new_scopes
: RangeFrom
<SourceScope
>,
986 new_blocks
: RangeFrom
<BasicBlock
>,
988 callsite_scope
: SourceScopeData
<'tcx
>,
989 callsite
: &'a CallSite
<'tcx
>,
990 cleanup_block
: Option
<BasicBlock
>,
991 in_cleanup_block
: bool
,
993 expn_data
: LocalExpnId
,
994 always_live_locals
: BitSet
<Local
>,
997 impl Integrator
<'_
, '_
> {
998 fn map_local(&self, local
: Local
) -> Local
{
999 let new
= if local
== RETURN_PLACE
{
1002 let idx
= local
.index() - 1;
1003 if idx
< self.args
.len() {
1006 Local
::new(self.new_locals
.start
.index() + (idx
- self.args
.len()))
1009 trace
!("mapping local `{:?}` to `{:?}`", local
, new
);
1013 fn map_scope(&self, scope
: SourceScope
) -> SourceScope
{
1014 let new
= SourceScope
::new(self.new_scopes
.start
.index() + scope
.index());
1015 trace
!("mapping scope `{:?}` to `{:?}`", scope
, new
);
1019 fn map_block(&self, block
: BasicBlock
) -> BasicBlock
{
1020 let new
= BasicBlock
::new(self.new_blocks
.start
.index() + block
.index());
1021 trace
!("mapping block `{:?}` to `{:?}`", block
, new
);
1025 fn map_unwind(&self, unwind
: Option
<BasicBlock
>) -> Option
<BasicBlock
> {
1026 if self.in_cleanup_block
{
1027 if unwind
.is_some() {
1028 bug
!("cleanup on cleanup block");
1034 Some(target
) => Some(self.map_block(target
)),
1035 // Add an unwind edge to the original call's cleanup block
1036 None
=> self.cleanup_block
,
1041 impl<'tcx
> MutVisitor
<'tcx
> for Integrator
<'_
, 'tcx
> {
1042 fn tcx(&self) -> TyCtxt
<'tcx
> {
1046 fn visit_local(&mut self, local
: &mut Local
, _ctxt
: PlaceContext
, _location
: Location
) {
1047 *local
= self.map_local(*local
);
1050 fn visit_source_scope_data(&mut self, scope_data
: &mut SourceScopeData
<'tcx
>) {
1051 self.super_source_scope_data(scope_data
);
1052 if scope_data
.parent_scope
.is_none() {
1053 // Attach the outermost callee scope as a child of the callsite
1054 // scope, via the `parent_scope` and `inlined_parent_scope` chains.
1055 scope_data
.parent_scope
= Some(self.callsite
.source_info
.scope
);
1056 assert_eq
!(scope_data
.inlined_parent_scope
, None
);
1057 scope_data
.inlined_parent_scope
= if self.callsite_scope
.inlined
.is_some() {
1058 Some(self.callsite
.source_info
.scope
)
1060 self.callsite_scope
.inlined_parent_scope
1063 // Mark the outermost callee scope as an inlined one.
1064 assert_eq
!(scope_data
.inlined
, None
);
1065 scope_data
.inlined
= Some((self.callsite
.callee
, self.callsite
.source_info
.span
));
1066 } else if scope_data
.inlined_parent_scope
.is_none() {
1067 // Make it easy to find the scope with `inlined` set above.
1068 scope_data
.inlined_parent_scope
= Some(self.map_scope(OUTERMOST_SOURCE_SCOPE
));
1072 fn visit_source_scope(&mut self, scope
: &mut SourceScope
) {
1073 *scope
= self.map_scope(*scope
);
1076 fn visit_span(&mut self, span
: &mut Span
) {
1077 // Make sure that all spans track the fact that they were inlined.
1078 *span
= span
.fresh_expansion(self.expn_data
);
1081 fn visit_basic_block_data(&mut self, block
: BasicBlock
, data
: &mut BasicBlockData
<'tcx
>) {
1082 self.in_cleanup_block
= data
.is_cleanup
;
1083 self.super_basic_block_data(block
, data
);
1084 self.in_cleanup_block
= false;
1087 fn visit_retag(&mut self, kind
: &mut RetagKind
, place
: &mut Place
<'tcx
>, loc
: Location
) {
1088 self.super_retag(kind
, place
, loc
);
1090 // We have to patch all inlined retags to be aware that they are no longer
1091 // happening on function entry.
1092 if *kind
== RetagKind
::FnEntry
{
1093 *kind
= RetagKind
::Default
;
1097 fn visit_statement(&mut self, statement
: &mut Statement
<'tcx
>, location
: Location
) {
1098 if let StatementKind
::StorageLive(local
) | StatementKind
::StorageDead(local
) =
1101 self.always_live_locals
.remove(local
);
1103 self.super_statement(statement
, location
);
1106 fn visit_terminator(&mut self, terminator
: &mut Terminator
<'tcx
>, loc
: Location
) {
1107 // Don't try to modify the implicit `_0` access on return (`return` terminators are
1108 // replaced down below anyways).
1109 if !matches
!(terminator
.kind
, TerminatorKind
::Return
) {
1110 self.super_terminator(terminator
, loc
);
1113 match terminator
.kind
{
1114 TerminatorKind
::GeneratorDrop
| TerminatorKind
::Yield { .. }
=> bug
!(),
1115 TerminatorKind
::Goto { ref mut target }
=> {
1116 *target
= self.map_block(*target
);
1118 TerminatorKind
::SwitchInt { ref mut targets, .. }
=> {
1119 for tgt
in targets
.all_targets_mut() {
1120 *tgt
= self.map_block(*tgt
);
1123 TerminatorKind
::Drop { ref mut target, ref mut unwind, .. }
1124 | TerminatorKind
::DropAndReplace { ref mut target, ref mut unwind, .. }
=> {
1125 *target
= self.map_block(*target
);
1126 *unwind
= self.map_unwind(*unwind
);
1128 TerminatorKind
::Call { ref mut target, ref mut cleanup, .. }
=> {
1129 if let Some(ref mut tgt
) = *target
{
1130 *tgt
= self.map_block(*tgt
);
1132 *cleanup
= self.map_unwind(*cleanup
);
1134 TerminatorKind
::Assert { ref mut target, ref mut cleanup, .. }
=> {
1135 *target
= self.map_block(*target
);
1136 *cleanup
= self.map_unwind(*cleanup
);
1138 TerminatorKind
::Return
=> {
1139 terminator
.kind
= if let Some(tgt
) = self.callsite
.target
{
1140 TerminatorKind
::Goto { target: tgt }
1142 TerminatorKind
::Unreachable
1145 TerminatorKind
::Resume
=> {
1146 if let Some(tgt
) = self.cleanup_block
{
1147 terminator
.kind
= TerminatorKind
::Goto { target: tgt }
1150 TerminatorKind
::Abort
=> {}
1151 TerminatorKind
::Unreachable
=> {}
1152 TerminatorKind
::FalseEdge { ref mut real_target, ref mut imaginary_target }
=> {
1153 *real_target
= self.map_block(*real_target
);
1154 *imaginary_target
= self.map_block(*imaginary_target
);
1156 TerminatorKind
::FalseUnwind { real_target: _, unwind: _ }
=>
1157 // see the ordering of passes in the optimized_mir query.
1159 bug
!("False unwinds should have been removed before inlining")
1161 TerminatorKind
::InlineAsm { ref mut destination, ref mut cleanup, .. }
=> {
1162 if let Some(ref mut tgt
) = *destination
{
1163 *tgt
= self.map_block(*tgt
);
1165 *cleanup
= self.map_unwind(*cleanup
);