1 //! This module implements lowering (instruction selection) from Cranelift IR 2 //! to machine instructions with virtual registers. This is *almost* the final 3 //! machine code, except for register allocation. 4 5 // TODO: separate the IR-query core of `Lower` from the lowering logic built on 6 // top of it, e.g. the side-effect/coloring analysis and the scan support. 7 8 use crate::entity::SecondaryMap; 9 use crate::inst_predicates::{has_lowering_side_effect, is_constant_64bit}; 10 use crate::ir::pcc::{Fact, FactContext, PccError, PccResult}; 11 use crate::ir::{ 12 ArgumentPurpose, Block, BlockArg, Constant, ConstantData, DataFlowGraph, ExternalName, 13 Function, GlobalValue, GlobalValueData, Immediate, Inst, InstructionData, MemFlags, 14 RelSourceLoc, SigRef, Signature, Type, Value, ValueDef, ValueLabelAssignments, ValueLabelStart, 15 }; 16 use crate::machinst::valueregs::InvalidSentinel; 17 use crate::machinst::{ 18 ABIMachineSpec, BackwardsInsnIndex, BlockIndex, BlockLoweringOrder, CallArgList, CallInfo, 19 CallRetList, Callee, InsnIndex, LoweredBlock, MachLabel, Reg, Sig, SigSet, TryCallInfo, VCode, 20 VCodeBuilder, VCodeConstant, VCodeConstantData, VCodeConstants, VCodeInst, ValueRegs, Writable, 21 writable_value_regs, 22 }; 23 use crate::settings::Flags; 24 use crate::{CodegenError, CodegenResult, trace}; 25 use alloc::vec::Vec; 26 use cranelift_control::ControlPlane; 27 use rustc_hash::{FxHashMap, FxHashSet}; 28 use smallvec::{SmallVec, smallvec}; 29 use std::fmt::Debug; 30 31 use super::{VCodeBuildDirection, VRegAllocator}; 32 33 /// A vector of ValueRegs, used to represent the outputs of an instruction. 34 pub type InstOutput = SmallVec<[ValueRegs<Reg>; 2]>; 35 36 /// An "instruction color" partitions CLIF instructions by side-effecting ops. 37 /// All instructions with the same "color" are guaranteed not to be separated by 38 /// any side-effecting op (for this purpose, loads are also considered 39 /// side-effecting, to avoid subtle questions w.r.t. the memory model), and 40 /// furthermore, it is guaranteed that for any two instructions A and B such 41 /// that color(A) == color(B), either A dominates B and B postdominates A, or 42 /// vice-versa. (For now, in practice, only ops in the same basic block can ever 43 /// have the same color, trivially providing the second condition.) Intuitively, 44 /// this means that the ops of the same color must always execute "together", as 45 /// part of one atomic contiguous section of the dynamic execution trace, and 46 /// they can be freely permuted (modulo true dataflow dependencies) without 47 /// affecting program behavior. 48 #[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)] 49 struct InstColor(u32); 50 impl InstColor { 51 fn new(n: u32) -> InstColor { 52 InstColor(n) 53 } 54 55 /// Get an arbitrary index representing this color. The index is unique 56 /// *within a single function compilation*, but indices may be reused across 57 /// functions. 58 pub fn get(self) -> u32 { 59 self.0 60 } 61 } 62 63 /// A representation of all of the ways in which a value is available, aside 64 /// from as a direct register. 65 /// 66 /// - An instruction, if it would be allowed to occur at the current location 67 /// instead (see [Lower::get_input_as_source_or_const()] for more details). 68 /// 69 /// - A constant, if the value is known to be a constant. 70 #[derive(Clone, Copy, Debug)] 71 pub struct NonRegInput { 72 /// An instruction produces this value (as the given output), and its 73 /// computation (and side-effect if applicable) could occur at the 74 /// current instruction's location instead. 75 /// 76 /// If this instruction's operation is merged into the current instruction, 77 /// the backend must call [Lower::sink_inst()]. 78 /// 79 /// This enum indicates whether this use of the source instruction 80 /// is unique or not. 81 pub inst: InputSourceInst, 82 /// The value is a known constant. 83 pub constant: Option<u64>, 84 } 85 86 /// When examining an input to an instruction, this enum provides one 87 /// of several options: there is or isn't a single instruction (that 88 /// we can see and merge with) that produces that input's value, and 89 /// we are or aren't the single user of that instruction. 90 #[derive(Clone, Copy, Debug)] 91 pub enum InputSourceInst { 92 /// The input in question is the single, unique use of the given 93 /// instruction and output index, and it can be sunk to the 94 /// location of this input. 95 UniqueUse(Inst, usize), 96 /// The input in question is one of multiple uses of the given 97 /// instruction. It can still be sunk to the location of this 98 /// input. 99 Use(Inst, usize), 100 /// We cannot determine which instruction produced the input, or 101 /// it is one of several instructions (e.g., due to a control-flow 102 /// merge and blockparam), or the source instruction cannot be 103 /// allowed to sink to the current location due to side-effects. 104 None, 105 } 106 107 impl InputSourceInst { 108 /// Get the instruction and output index for this source, whether 109 /// we are its single or one of many users. 110 pub fn as_inst(&self) -> Option<(Inst, usize)> { 111 match self { 112 &InputSourceInst::UniqueUse(inst, output_idx) 113 | &InputSourceInst::Use(inst, output_idx) => Some((inst, output_idx)), 114 &InputSourceInst::None => None, 115 } 116 } 117 } 118 119 /// A machine backend. 120 pub trait LowerBackend { 121 /// The machine instruction type. 122 type MInst: VCodeInst; 123 124 /// Lower a single instruction. 125 /// 126 /// For a branch, this function should not generate the actual branch 127 /// instruction. However, it must force any values it needs for the branch 128 /// edge (block-param actuals) into registers, because the actual branch 129 /// generation (`lower_branch()`) happens *after* any possible merged 130 /// out-edge. 131 /// 132 /// Returns `None` if no lowering for the instruction was found. 133 fn lower(&self, ctx: &mut Lower<Self::MInst>, inst: Inst) -> Option<InstOutput>; 134 135 /// Lower a block-terminating group of branches (which together can be seen 136 /// as one N-way branch), given a vcode MachLabel for each target. 137 /// 138 /// Returns `None` if no lowering for the branch was found. 139 fn lower_branch( 140 &self, 141 ctx: &mut Lower<Self::MInst>, 142 inst: Inst, 143 targets: &[MachLabel], 144 ) -> Option<()>; 145 146 /// A bit of a hack: give a fixed register that always holds the result of a 147 /// `get_pinned_reg` instruction, if known. This allows elision of moves 148 /// into the associated vreg, instead using the real reg directly. 149 fn maybe_pinned_reg(&self) -> Option<Reg> { 150 None 151 } 152 153 /// The type of state carried between `check_fact` invocations. 154 type FactFlowState: Default + Clone + Debug; 155 156 /// Check any facts about an instruction, given VCode with facts 157 /// on VRegs. Takes mutable `VCode` so that it can propagate some 158 /// kinds of facts automatically. 159 fn check_fact( 160 &self, 161 _ctx: &FactContext<'_>, 162 _vcode: &mut VCode<Self::MInst>, 163 _inst: InsnIndex, 164 _state: &mut Self::FactFlowState, 165 ) -> PccResult<()> { 166 Err(PccError::UnimplementedBackend) 167 } 168 } 169 170 /// Machine-independent lowering driver / machine-instruction container. Maintains a correspondence 171 /// from original Inst to MachInsts. 172 pub struct Lower<'func, I: VCodeInst> { 173 /// The function to lower. 174 pub(crate) f: &'func Function, 175 176 /// Lowered machine instructions. 177 vcode: VCodeBuilder<I>, 178 179 /// VReg allocation context, given to the vcode field at build time to finalize the vcode. 180 vregs: VRegAllocator<I>, 181 182 /// Mapping from `Value` (SSA value in IR) to virtual register. 183 value_regs: SecondaryMap<Value, ValueRegs<Reg>>, 184 185 /// sret registers, if needed. 186 sret_reg: Option<ValueRegs<Reg>>, 187 188 /// Instruction colors at block exits. From this map, we can recover all 189 /// instruction colors by scanning backward from the block end and 190 /// decrementing on any color-changing (side-effecting) instruction. 191 block_end_colors: SecondaryMap<Block, InstColor>, 192 193 /// Instruction colors at side-effecting ops. This is the *entry* color, 194 /// i.e., the version of global state that exists before an instruction 195 /// executes. For each side-effecting instruction, the *exit* color is its 196 /// entry color plus one. 197 side_effect_inst_entry_colors: FxHashMap<Inst, InstColor>, 198 199 /// Current color as we scan during lowering. While we are lowering an 200 /// instruction, this is equal to the color *at entry to* the instruction. 201 cur_scan_entry_color: Option<InstColor>, 202 203 /// Current instruction as we scan during lowering. 204 cur_inst: Option<Inst>, 205 206 /// Instruction constant values, if known. 207 inst_constants: FxHashMap<Inst, u64>, 208 209 /// Use-counts per SSA value, as counted in the input IR. These 210 /// are "coarsened", in the abstract-interpretation sense: we only 211 /// care about "0, 1, many" states, as this is all we need and 212 /// this lets us do an efficient fixpoint analysis. 213 /// 214 /// See doc comment on `ValueUseState` for more details. 215 value_ir_uses: SecondaryMap<Value, ValueUseState>, 216 217 /// Actual uses of each SSA value so far, incremented while lowering. 218 value_lowered_uses: SecondaryMap<Value, u32>, 219 220 /// Effectful instructions that have been sunk; they are not codegen'd at 221 /// their original locations. 222 inst_sunk: FxHashSet<Inst>, 223 224 /// Instructions collected for the CLIF inst in progress, in forward order. 225 ir_insts: Vec<I>, 226 227 /// Try-call block arg normal-return values, indexed by instruction. 228 try_call_rets: FxHashMap<Inst, SmallVec<[ValueRegs<Writable<Reg>>; 2]>>, 229 230 /// Try-call block arg exceptional-return payloads, indexed by 231 /// instruction. Payloads are carried in registers per the ABI and 232 /// can only be one register each. 233 try_call_payloads: FxHashMap<Inst, SmallVec<[Writable<Reg>; 2]>>, 234 235 /// The register to use for GetPinnedReg, if any, on this architecture. 236 pinned_reg: Option<Reg>, 237 238 /// Compilation flags. 239 flags: Flags, 240 } 241 242 /// How is a value used in the IR? 243 /// 244 /// This can be seen as a coarsening of an integer count. We only need 245 /// distinct states for zero, one, or many. 246 /// 247 /// This analysis deserves further explanation. The basic idea is that 248 /// we want to allow instruction lowering to know whether a value that 249 /// an instruction references is *only* referenced by that one use, or 250 /// by others as well. This is necessary to know when we might want to 251 /// move a side-effect: we cannot, for example, duplicate a load, so 252 /// we cannot let instruction lowering match a load as part of a 253 /// subpattern and potentially incorporate it. 254 /// 255 /// Note that a lot of subtlety comes into play once we have 256 /// *indirect* uses. The classical example of this in our development 257 /// history was the x86 compare instruction, which is incorporated 258 /// into flags users (e.g. `selectif`, `trueif`, branches) and can 259 /// subsequently incorporate loads, or at least we would like it 260 /// to. However, danger awaits: the compare might be the only user of 261 /// a load, so we might think we can just move the load (and nothing 262 /// is duplicated -- success!), except that the compare itself is 263 /// codegen'd in multiple places, where it is incorporated as a 264 /// subpattern itself. 265 /// 266 /// So we really want a notion of "unique all the way along the 267 /// matching path". Rust's `&T` and `&mut T` offer a partial analogy 268 /// to the semantics that we want here: we want to know when we've 269 /// matched a unique use of an instruction, and that instruction's 270 /// unique use of another instruction, etc, just as `&mut T` can only 271 /// be obtained by going through a chain of `&mut T`. If one has a 272 /// `&T` to a struct containing `&mut T` (one of several uses of an 273 /// instruction that itself has a unique use of an instruction), one 274 /// can only get a `&T` (one can only get a "I am one of several users 275 /// of this instruction" result). 276 /// 277 /// We could track these paths, either dynamically as one "looks up the operand 278 /// tree" or precomputed. But the former requires state and means that the 279 /// `Lower` API carries that state implicitly, which we'd like to avoid if we 280 /// can. And the latter implies O(n^2) storage: it is an all-pairs property (is 281 /// inst `i` unique from the point of view of `j`). 282 /// 283 /// To make matters even a little more complex still, a value that is 284 /// not uniquely used when initially viewing the IR can *become* 285 /// uniquely used, at least as a root allowing further unique uses of 286 /// e.g. loads to merge, if no other instruction actually merges 287 /// it. To be more concrete, if we have `v1 := load; v2 := op v1; v3 288 /// := op v2; v4 := op v2` then `v2` is non-uniquely used, so from the 289 /// point of view of lowering `v4` or `v3`, we cannot merge the load 290 /// at `v1`. But if we decide just to use the assigned register for 291 /// `v2` at both `v3` and `v4`, then we only actually codegen `v2` 292 /// once, so it *is* a unique root at that point and we *can* merge 293 /// the load. 294 /// 295 /// Note also that the color scheme is not sufficient to give us this 296 /// information, for various reasons: reasoning about side-effects 297 /// does not tell us about potential duplication of uses through pure 298 /// ops. 299 /// 300 /// To keep things simple and avoid error-prone lowering APIs that 301 /// would extract more information about whether instruction merging 302 /// happens or not (we don't have that info now, and it would be 303 /// difficult to refactor to get it and make that refactor 100% 304 /// correct), we give up on the above "can become unique if not 305 /// actually merged" point. Instead, we compute a 306 /// transitive-uniqueness. That is what this enum represents. 307 /// 308 /// There is one final caveat as well to the result of this analysis. Notably, 309 /// we define some instructions to be "root" instructions, which means that we 310 /// assume they will always be codegen'd at the root of a matching tree, and not 311 /// matched. (This comes with the caveat that we actually enforce this property 312 /// by making them "opaque" to subtree matching in 313 /// `get_value_as_source_or_const`). Because they will always be codegen'd once, 314 /// they in some sense "reset" multiplicity: these root instructions can be used 315 /// many times, but because their result(s) are only computed once, they only 316 /// use their inputs once. 317 /// 318 /// We currently define all multi-result instructions to be "root" instructions, 319 /// because it is too complex to reason about matching through them, and they 320 /// cause too-coarse-grained approximation of multiplicity otherwise: the 321 /// analysis would have to assume (as it used to!) that they are always 322 /// multiply-used, simply because they have multiple outputs even if those 323 /// outputs are used only once. 324 /// 325 /// In the future we could define other instructions to be "root" instructions 326 /// as well, if we make the corresponding change to get_value_as_source_or_const 327 /// as well. 328 /// 329 /// To define `ValueUseState` more plainly: a value is `Unused` if no references 330 /// exist to it; `Once` if only one other op refers to it, *and* that other op 331 /// is `Unused` or `Once`; and `Multiple` otherwise. In other words, `Multiple` 332 /// is contagious (except through root instructions): even if an op's result 333 /// value is directly used only once in the CLIF, that value is `Multiple` if 334 /// the op that uses it is itself used multiple times (hence could be codegen'd 335 /// multiple times). In brief, this analysis tells us whether, if every op 336 /// merged all of its operand tree, a given op could be codegen'd in more than 337 /// one place. 338 /// 339 /// To compute this, we first consider direct uses. At this point 340 /// `Unused` answers are correct, `Multiple` answers are correct, but 341 /// some `Once`s may change to `Multiple`s. Then we propagate 342 /// `Multiple` transitively using a workqueue/fixpoint algorithm. 343 #[derive(Clone, Copy, Debug, PartialEq, Eq)] 344 enum ValueUseState { 345 /// Not used at all. 346 Unused, 347 /// Used exactly once. 348 Once, 349 /// Used multiple times. 350 Multiple, 351 } 352 353 impl ValueUseState { 354 /// Add one use. 355 fn inc(&mut self) { 356 let new = match self { 357 Self::Unused => Self::Once, 358 Self::Once | Self::Multiple => Self::Multiple, 359 }; 360 *self = new; 361 } 362 } 363 364 /// Notion of "relocation distance". This gives an estimate of how far away a symbol will be from a 365 /// reference. 366 #[derive(Clone, Copy, Debug, PartialEq, Eq)] 367 pub enum RelocDistance { 368 /// Target of relocation is "nearby". The threshold for this is fuzzy but should be interpreted 369 /// as approximately "within the compiled output of one module"; e.g., within AArch64's +/- 370 /// 128MB offset. If unsure, use `Far` instead. 371 Near, 372 /// Target of relocation could be anywhere in the address space. 373 Far, 374 } 375 376 impl<'func, I: VCodeInst> Lower<'func, I> { 377 /// Prepare a new lowering context for the given IR function. 378 pub fn new( 379 f: &'func Function, 380 abi: Callee<I::ABIMachineSpec>, 381 emit_info: I::Info, 382 block_order: BlockLoweringOrder, 383 sigs: SigSet, 384 flags: Flags, 385 ) -> CodegenResult<Self> { 386 let constants = VCodeConstants::with_capacity(f.dfg.constants.len()); 387 let vcode = VCodeBuilder::new( 388 sigs, 389 abi, 390 emit_info, 391 block_order, 392 constants, 393 VCodeBuildDirection::Backward, 394 flags.log2_min_function_alignment(), 395 ); 396 397 // We usually need two VRegs per instruction result, plus extras for 398 // various temporaries, but two per Value is a good starting point. 399 let mut vregs = VRegAllocator::with_capacity(f.dfg.num_values() * 2); 400 401 let mut value_regs = SecondaryMap::with_default(ValueRegs::invalid()); 402 let mut try_call_rets = FxHashMap::default(); 403 let mut try_call_payloads = FxHashMap::default(); 404 405 // Assign a vreg to each block param, each inst result, and 406 // each edge-defined block-call arg. 407 for bb in f.layout.blocks() { 408 for ¶m in f.dfg.block_params(bb) { 409 let ty = f.dfg.value_type(param); 410 if value_regs[param].is_invalid() { 411 let regs = vregs.alloc_with_maybe_fact(ty, f.dfg.facts[param].clone())?; 412 value_regs[param] = regs; 413 trace!("bb {} param {}: regs {:?}", bb, param, regs); 414 } 415 } 416 for inst in f.layout.block_insts(bb) { 417 for &result in f.dfg.inst_results(inst) { 418 let ty = f.dfg.value_type(result); 419 if value_regs[result].is_invalid() && !ty.is_invalid() { 420 let regs = vregs.alloc_with_maybe_fact(ty, f.dfg.facts[result].clone())?; 421 value_regs[result] = regs; 422 trace!( 423 "bb {} inst {} ({:?}): result {} regs {:?}", 424 bb, inst, f.dfg.insts[inst], result, regs, 425 ); 426 } 427 } 428 429 if let Some(et) = f.dfg.insts[inst].exception_table() { 430 let exdata = &f.dfg.exception_tables[et]; 431 let sig = &f.dfg.signatures[exdata.signature()]; 432 433 let mut rets = smallvec![]; 434 for ty in sig.returns.iter().map(|ret| ret.value_type) { 435 rets.push(vregs.alloc(ty)?.map(|r| Writable::from_reg(r))); 436 } 437 try_call_rets.insert(inst, rets); 438 439 let mut payloads = smallvec![]; 440 // Note that this is intentionally using the calling 441 // convention of the callee to determine what payload types 442 // are available. The callee defines that, not the calling 443 // convention of the caller. 444 for &ty in sig 445 .call_conv 446 .exception_payload_types(I::ABIMachineSpec::word_type()) 447 { 448 payloads.push(Writable::from_reg(vregs.alloc(ty)?.only_reg().unwrap())); 449 } 450 try_call_payloads.insert(inst, payloads); 451 } 452 } 453 } 454 455 // Find the sret register, if it's used. 456 let mut sret_param = None; 457 for ret in vcode.abi().signature().returns.iter() { 458 if ret.purpose == ArgumentPurpose::StructReturn { 459 let entry_bb = f.stencil.layout.entry_block().unwrap(); 460 for (¶m, sig_param) in f 461 .dfg 462 .block_params(entry_bb) 463 .iter() 464 .zip(vcode.abi().signature().params.iter()) 465 { 466 if sig_param.purpose == ArgumentPurpose::StructReturn { 467 assert!(sret_param.is_none()); 468 sret_param = Some(param); 469 } 470 } 471 472 assert!(sret_param.is_some()); 473 } 474 } 475 476 let sret_reg = sret_param.map(|param| { 477 let regs = value_regs[param]; 478 assert!(regs.len() == 1); 479 regs 480 }); 481 482 // Compute instruction colors, find constant instructions, and find instructions with 483 // side-effects, in one combined pass. 484 let mut cur_color = 0; 485 let mut block_end_colors = SecondaryMap::with_default(InstColor::new(0)); 486 let mut side_effect_inst_entry_colors = FxHashMap::default(); 487 let mut inst_constants = FxHashMap::default(); 488 for bb in f.layout.blocks() { 489 cur_color += 1; 490 for inst in f.layout.block_insts(bb) { 491 let side_effect = has_lowering_side_effect(f, inst); 492 493 trace!("bb {} inst {} has color {}", bb, inst, cur_color); 494 if side_effect { 495 side_effect_inst_entry_colors.insert(inst, InstColor::new(cur_color)); 496 trace!(" -> side-effecting; incrementing color for next inst"); 497 cur_color += 1; 498 } 499 500 // Determine if this is a constant; if so, add to the table. 501 if let Some(c) = is_constant_64bit(f, inst) { 502 trace!(" -> constant: {}", c); 503 inst_constants.insert(inst, c); 504 } 505 } 506 507 block_end_colors[bb] = InstColor::new(cur_color); 508 } 509 510 let value_ir_uses = compute_use_states(f, sret_param); 511 512 Ok(Lower { 513 f, 514 vcode, 515 vregs, 516 value_regs, 517 sret_reg, 518 block_end_colors, 519 side_effect_inst_entry_colors, 520 inst_constants, 521 value_ir_uses, 522 value_lowered_uses: SecondaryMap::default(), 523 inst_sunk: FxHashSet::default(), 524 cur_scan_entry_color: None, 525 cur_inst: None, 526 ir_insts: vec![], 527 try_call_rets, 528 try_call_payloads, 529 pinned_reg: None, 530 flags, 531 }) 532 } 533 534 pub fn sigs(&self) -> &SigSet { 535 self.vcode.sigs() 536 } 537 538 pub fn sigs_mut(&mut self) -> &mut SigSet { 539 self.vcode.sigs_mut() 540 } 541 542 pub fn vregs_mut(&mut self) -> &mut VRegAllocator<I> { 543 &mut self.vregs 544 } 545 546 fn gen_arg_setup(&mut self) { 547 if let Some(entry_bb) = self.f.layout.entry_block() { 548 trace!( 549 "gen_arg_setup: entry BB {} args are:\n{:?}", 550 entry_bb, 551 self.f.dfg.block_params(entry_bb) 552 ); 553 554 for (i, param) in self.f.dfg.block_params(entry_bb).iter().enumerate() { 555 if self.value_ir_uses[*param] == ValueUseState::Unused { 556 continue; 557 } 558 let regs = writable_value_regs(self.value_regs[*param]); 559 for insn in self 560 .vcode 561 .vcode 562 .abi 563 .gen_copy_arg_to_regs(&self.vcode.vcode.sigs, i, regs, &mut self.vregs) 564 .into_iter() 565 { 566 self.emit(insn); 567 } 568 } 569 if let Some(insn) = self 570 .vcode 571 .vcode 572 .abi 573 .gen_retval_area_setup(&self.vcode.vcode.sigs, &mut self.vregs) 574 { 575 self.emit(insn); 576 } 577 578 // The `args` instruction below must come first. Finish 579 // the current "IR inst" (with a default source location, 580 // as for other special instructions inserted during 581 // lowering) and continue the scan backward. 582 self.finish_ir_inst(Default::default()); 583 584 if let Some(insn) = self.vcode.vcode.abi.take_args() { 585 self.emit(insn); 586 } 587 } 588 } 589 590 /// Generate the return instruction. 591 pub fn gen_return(&mut self, rets: &[ValueRegs<Reg>]) { 592 let mut out_rets = vec![]; 593 594 let mut rets = rets.into_iter(); 595 for (i, ret) in self 596 .abi() 597 .signature() 598 .returns 599 .clone() 600 .into_iter() 601 .enumerate() 602 { 603 let regs = if ret.purpose == ArgumentPurpose::StructReturn { 604 self.sret_reg.unwrap() 605 } else { 606 *rets.next().unwrap() 607 }; 608 609 let (regs, insns) = self.vcode.abi().gen_copy_regs_to_retval( 610 self.vcode.sigs(), 611 i, 612 regs, 613 &mut self.vregs, 614 ); 615 out_rets.extend(regs); 616 for insn in insns { 617 self.emit(insn); 618 } 619 } 620 621 // Hack: generate a virtual instruction that uses vmctx in 622 // order to keep it alive for the duration of the function, 623 // for the benefit of debuginfo. 624 if self.f.dfg.values_labels.is_some() { 625 if let Some(vmctx_val) = self.f.special_param(ArgumentPurpose::VMContext) { 626 if self.value_ir_uses[vmctx_val] != ValueUseState::Unused { 627 let vmctx_reg = self.value_regs[vmctx_val].only_reg().unwrap(); 628 self.emit(I::gen_dummy_use(vmctx_reg)); 629 } 630 } 631 } 632 633 let inst = self.abi().gen_rets(out_rets); 634 self.emit(inst); 635 } 636 637 /// Generate list of registers to hold the output of a call with 638 /// signature `sig`. 639 pub fn gen_call_output(&mut self, sig: &Signature) -> InstOutput { 640 let mut rets = smallvec![]; 641 for ty in sig.returns.iter().map(|ret| ret.value_type) { 642 rets.push(self.vregs.alloc_with_deferred_error(ty)); 643 } 644 rets 645 } 646 647 /// Likewise, but for a `SigRef` instead. 648 pub fn gen_call_output_from_sig_ref(&mut self, sig_ref: SigRef) -> InstOutput { 649 self.gen_call_output(&self.f.dfg.signatures[sig_ref]) 650 } 651 652 /// Set up arguments values `args` for a call with signature `sig`. 653 pub fn gen_call_args(&mut self, sig: Sig, args: &[ValueRegs<Reg>]) -> CallArgList { 654 let (uses, insts) = self.vcode.abi().gen_call_args( 655 self.vcode.sigs(), 656 sig, 657 args, 658 /* is_tail_call */ false, 659 &self.flags, 660 &mut self.vregs, 661 ); 662 for insn in insts { 663 self.emit(insn); 664 } 665 uses 666 } 667 668 /// Likewise, but for a `return_call`. 669 pub fn gen_return_call_args(&mut self, sig: Sig, args: &[ValueRegs<Reg>]) -> CallArgList { 670 let (uses, insts) = self.vcode.abi().gen_call_args( 671 self.vcode.sigs(), 672 sig, 673 args, 674 /* is_tail_call */ true, 675 &self.flags, 676 &mut self.vregs, 677 ); 678 for insn in insts { 679 self.emit(insn); 680 } 681 uses 682 } 683 684 /// Set up return values `outputs` for a call with signature `sig`. 685 pub fn gen_call_rets(&mut self, sig: Sig, outputs: &[ValueRegs<Reg>]) -> CallRetList { 686 self.vcode 687 .abi() 688 .gen_call_rets(self.vcode.sigs(), sig, outputs, None, &mut self.vregs) 689 } 690 691 /// Likewise, but for a `try_call`. 692 pub fn gen_try_call_rets(&mut self, sig: Sig) -> CallRetList { 693 let ir_inst = self.cur_inst.unwrap(); 694 let mut outputs: SmallVec<[ValueRegs<Reg>; 2]> = smallvec![]; 695 for return_def in self.try_call_rets.get(&ir_inst).unwrap() { 696 outputs.push(return_def.map(|r| r.to_reg())); 697 } 698 let payloads = Some(&self.try_call_payloads.get(&ir_inst).unwrap()[..]); 699 700 self.vcode 701 .abi() 702 .gen_call_rets(self.vcode.sigs(), sig, &outputs, payloads, &mut self.vregs) 703 } 704 705 /// Populate a `CallInfo` for a call with signature `sig`. 706 pub fn gen_call_info<T>( 707 &mut self, 708 sig: Sig, 709 dest: T, 710 uses: CallArgList, 711 defs: CallRetList, 712 try_call_info: Option<TryCallInfo>, 713 patchable: bool, 714 ) -> CallInfo<T> { 715 self.vcode.abi().gen_call_info( 716 self.vcode.sigs(), 717 sig, 718 dest, 719 uses, 720 defs, 721 try_call_info, 722 patchable, 723 ) 724 } 725 726 /// Has this instruction been sunk to a use-site (i.e., away from its 727 /// original location)? 728 fn is_inst_sunk(&self, inst: Inst) -> bool { 729 self.inst_sunk.contains(&inst) 730 } 731 732 // Is any result of this instruction needed? 733 fn is_any_inst_result_needed(&self, inst: Inst) -> bool { 734 self.f 735 .dfg 736 .inst_results(inst) 737 .iter() 738 .any(|&result| self.value_lowered_uses[result] > 0) 739 } 740 741 fn lower_clif_block<B: LowerBackend<MInst = I>>( 742 &mut self, 743 backend: &B, 744 block: Block, 745 ctrl_plane: &mut ControlPlane, 746 ) -> CodegenResult<()> { 747 self.cur_scan_entry_color = Some(self.block_end_colors[block]); 748 // Lowering loop: 749 // - For each non-branch instruction, in reverse order: 750 // - If side-effecting (load, store, branch/call/return, 751 // possible trap), or if used outside of this block, or if 752 // demanded by another inst, then lower. 753 // 754 // That's it! Lowering of side-effecting ops will force all *needed* 755 // (live) non-side-effecting ops to be lowered at the right places, via 756 // the `use_input_reg()` callback on the `Lower` (that's us). That's 757 // because `use_input_reg()` sets the eager/demand bit for any insts 758 // whose result registers are used. 759 // 760 // We set the VCodeBuilder to "backward" mode, so we emit 761 // blocks in reverse order wrt the BlockIndex sequence, and 762 // emit instructions in reverse order within blocks. Because 763 // the machine backend calls `ctx.emit()` in forward order, we 764 // collect per-IR-inst lowered instructions in `ir_insts`, 765 // then reverse these and append to the VCode at the end of 766 // each IR instruction. 767 for inst in self.f.layout.block_insts(block).rev() { 768 let data = &self.f.dfg.insts[inst]; 769 let has_side_effect = has_lowering_side_effect(self.f, inst); 770 // If inst has been sunk to another location, skip it. 771 if self.is_inst_sunk(inst) { 772 continue; 773 } 774 // Are any outputs used at least once? 775 let value_needed = self.is_any_inst_result_needed(inst); 776 trace!( 777 "lower_clif_block: block {} inst {} ({:?}) is_branch {} side_effect {} value_needed {}", 778 block, 779 inst, 780 data, 781 data.opcode().is_branch(), 782 has_side_effect, 783 value_needed, 784 ); 785 786 // Update scan state to color prior to this inst (as we are scanning 787 // backward). 788 self.cur_inst = Some(inst); 789 if has_side_effect { 790 let entry_color = *self 791 .side_effect_inst_entry_colors 792 .get(&inst) 793 .expect("every side-effecting inst should have a color-map entry"); 794 self.cur_scan_entry_color = Some(entry_color); 795 } 796 797 // Skip lowering branches; these are handled separately 798 // (see `lower_clif_branches()` below). 799 if self.f.dfg.insts[inst].opcode().is_branch() { 800 continue; 801 } 802 803 // Value defined by "inst" becomes live after it in normal 804 // order, and therefore **before** in reversed order. 805 self.emit_value_label_live_range_start_for_inst(inst); 806 807 // Normal instruction: codegen if the instruction is side-effecting 808 // or any of its outputs is used. 809 if has_side_effect || value_needed { 810 trace!("lowering: inst {}: {}", inst, self.f.dfg.display_inst(inst)); 811 let temp_regs = match backend.lower(self, inst) { 812 Some(regs) => regs, 813 None => { 814 let ty = if self.num_outputs(inst) > 0 { 815 Some(self.output_ty(inst, 0)) 816 } else { 817 None 818 }; 819 return Err(CodegenError::Unsupported(format!( 820 "should be implemented in ISLE: inst = `{}`, type = `{:?}`", 821 self.f.dfg.display_inst(inst), 822 ty 823 ))); 824 } 825 }; 826 827 // The ISLE generated code emits its own registers to define the 828 // instruction's lowered values in. However, other instructions 829 // that use this SSA value will be lowered assuming that the value 830 // is generated into a pre-assigned, different, register. 831 // 832 // To connect the two, we set up "aliases" in the VCodeBuilder 833 // that apply when it is building the Operand table for the 834 // regalloc to use. These aliases effectively rewrite any use of 835 // the pre-assigned register to the register that was returned by 836 // the ISLE lowering logic. 837 let results = self.f.dfg.inst_results(inst); 838 debug_assert_eq!(temp_regs.len(), results.len()); 839 for (regs, &result) in temp_regs.iter().zip(results) { 840 let dsts = self.value_regs[result]; 841 let mut regs = regs.regs().iter(); 842 for &dst in dsts.regs().iter() { 843 let temp = regs.next().copied().unwrap_or(Reg::invalid_sentinel()); 844 trace!("set vreg alias: {result:?} = {dst:?}, lowering = {temp:?}"); 845 self.vregs.set_vreg_alias(dst, temp); 846 } 847 } 848 } 849 850 let start = self.vcode.vcode.num_insts(); 851 let loc = self.srcloc(inst); 852 self.finish_ir_inst(loc); 853 854 // If the instruction had a user stack map, forward it from the CLIF 855 // to the vcode. 856 if let Some(entries) = self.f.dfg.user_stack_map_entries(inst) { 857 let end = self.vcode.vcode.num_insts(); 858 debug_assert!(end > start); 859 debug_assert_eq!( 860 (start..end) 861 .filter(|i| self.vcode.vcode[InsnIndex::new(*i)].is_safepoint()) 862 .count(), 863 1 864 ); 865 for i in start..end { 866 let iix = InsnIndex::new(i); 867 if self.vcode.vcode[iix].is_safepoint() { 868 trace!( 869 "Adding user stack map from clif\n\n\ 870 {inst:?} `{}`\n\n\ 871 to vcode\n\n\ 872 {iix:?} `{}`", 873 self.f.dfg.display_inst(inst), 874 &self.vcode.vcode[iix].pretty_print_inst(&mut Default::default()), 875 ); 876 self.vcode 877 .add_user_stack_map(BackwardsInsnIndex::new(iix.index()), entries); 878 break; 879 } 880 } 881 } 882 883 // If the CLIF instruction had debug tags, copy them to 884 // the VCode. Place on all VCode instructions lowered from 885 // this CLIF instruction. 886 let debug_tags = self.f.debug_tags.get(inst); 887 if !debug_tags.is_empty() && self.vcode.vcode.num_insts() > 0 { 888 let end = self.vcode.vcode.num_insts(); 889 for i in start..end { 890 let backwards_index = BackwardsInsnIndex::new(i); 891 log::trace!( 892 "debug tags on {inst}; associating {debug_tags:?} with {backwards_index:?}" 893 ); 894 self.vcode.add_debug_tags(backwards_index, debug_tags); 895 } 896 } 897 898 // maybe insert random instruction 899 if ctrl_plane.get_decision() { 900 if ctrl_plane.get_decision() { 901 let imm: u64 = ctrl_plane.get_arbitrary(); 902 let reg = self.alloc_tmp(crate::ir::types::I64).regs()[0]; 903 I::gen_imm_u64(imm, reg).map(|inst| self.emit(inst)); 904 } else { 905 let imm: f64 = ctrl_plane.get_arbitrary(); 906 let tmp = self.alloc_tmp(crate::ir::types::I64).regs()[0]; 907 let reg = self.alloc_tmp(crate::ir::types::F64).regs()[0]; 908 for inst in I::gen_imm_f64(imm, tmp, reg) { 909 self.emit(inst); 910 } 911 } 912 } 913 } 914 915 // Add the block params to this block. 916 self.add_block_params(block)?; 917 918 self.cur_scan_entry_color = None; 919 Ok(()) 920 } 921 922 fn add_block_params(&mut self, block: Block) -> CodegenResult<()> { 923 for ¶m in self.f.dfg.block_params(block) { 924 for ® in self.value_regs[param].regs() { 925 let vreg = reg.to_virtual_reg().unwrap(); 926 self.vcode.add_block_param(vreg); 927 } 928 } 929 Ok(()) 930 } 931 932 fn get_value_labels<'a>(&'a self, val: Value, depth: usize) -> Option<&'a [ValueLabelStart]> { 933 if let Some(ref values_labels) = self.f.dfg.values_labels { 934 debug_assert!(self.f.dfg.value_is_real(val)); 935 trace!( 936 "get_value_labels: val {} -> {:?}", 937 val, 938 values_labels.get(&val) 939 ); 940 match values_labels.get(&val) { 941 Some(&ValueLabelAssignments::Starts(ref list)) => Some(&list[..]), 942 Some(&ValueLabelAssignments::Alias { value, .. }) if depth < 10 => { 943 self.get_value_labels(value, depth + 1) 944 } 945 _ => None, 946 } 947 } else { 948 None 949 } 950 } 951 952 fn emit_value_label_marks_for_value(&mut self, val: Value) { 953 let regs = self.value_regs[val]; 954 if regs.len() > 1 { 955 return; 956 } 957 let reg = regs.only_reg().unwrap(); 958 959 if let Some(label_starts) = self.get_value_labels(val, 0) { 960 let labels = label_starts 961 .iter() 962 .map(|&ValueLabelStart { label, .. }| label) 963 .collect::<FxHashSet<_>>(); 964 for label in labels { 965 trace!( 966 "value labeling: defines val {:?} -> reg {:?} -> label {:?}", 967 val, reg, label, 968 ); 969 self.vcode.add_value_label(reg, label); 970 } 971 } 972 } 973 974 fn emit_value_label_live_range_start_for_inst(&mut self, inst: Inst) { 975 if self.f.dfg.values_labels.is_none() { 976 return; 977 } 978 979 trace!( 980 "value labeling: srcloc {}: inst {}", 981 self.srcloc(inst), 982 inst 983 ); 984 for &val in self.f.dfg.inst_results(inst) { 985 self.emit_value_label_marks_for_value(val); 986 } 987 } 988 989 fn emit_value_label_live_range_start_for_block_args(&mut self, block: Block) { 990 if self.f.dfg.values_labels.is_none() { 991 return; 992 } 993 994 trace!("value labeling: block {}", block); 995 for &arg in self.f.dfg.block_params(block) { 996 self.emit_value_label_marks_for_value(arg); 997 } 998 self.finish_ir_inst(Default::default()); 999 } 1000 1001 fn finish_ir_inst(&mut self, loc: RelSourceLoc) { 1002 // The VCodeBuilder builds in reverse order (and reverses at 1003 // the end), but `ir_insts` is in forward order, so reverse 1004 // it. 1005 for inst in self.ir_insts.drain(..).rev() { 1006 self.vcode.push(inst, loc); 1007 } 1008 } 1009 1010 fn finish_bb(&mut self) { 1011 self.vcode.end_bb(); 1012 } 1013 1014 fn lower_clif_branch<B: LowerBackend<MInst = I>>( 1015 &mut self, 1016 backend: &B, 1017 // Lowered block index: 1018 bindex: BlockIndex, 1019 // Original CLIF block: 1020 block: Block, 1021 branch: Inst, 1022 targets: &[MachLabel], 1023 ) -> CodegenResult<()> { 1024 trace!( 1025 "lower_clif_branch: block {} branch {:?} targets {:?}", 1026 block, branch, targets, 1027 ); 1028 // When considering code-motion opportunities, consider the current 1029 // program point to be this branch. 1030 self.cur_inst = Some(branch); 1031 1032 // Lower the branch in ISLE. 1033 backend 1034 .lower_branch(self, branch, targets) 1035 .unwrap_or_else(|| { 1036 panic!( 1037 "should be implemented in ISLE: branch = `{}`", 1038 self.f.dfg.display_inst(branch), 1039 ) 1040 }); 1041 let loc = self.srcloc(branch); 1042 self.finish_ir_inst(loc); 1043 // Add block param outputs for current block. 1044 self.lower_branch_blockparam_args(bindex); 1045 Ok(()) 1046 } 1047 1048 fn lower_branch_blockparam_args(&mut self, block: BlockIndex) { 1049 let mut branch_arg_vregs: SmallVec<[Reg; 16]> = smallvec![]; 1050 1051 // TODO: why not make `block_order` public? 1052 for succ_idx in 0..self.vcode.block_order().succ_indices(block).1.len() { 1053 branch_arg_vregs.clear(); 1054 let (succ, args) = self.collect_block_call(block, succ_idx, &mut branch_arg_vregs); 1055 self.vcode.add_succ(succ, args); 1056 } 1057 } 1058 1059 fn collect_branch_and_targets( 1060 &self, 1061 bindex: BlockIndex, 1062 _bb: Block, 1063 targets: &mut SmallVec<[MachLabel; 2]>, 1064 ) -> Option<Inst> { 1065 targets.clear(); 1066 let (opt_inst, succs) = self.vcode.block_order().succ_indices(bindex); 1067 targets.extend(succs.iter().map(|succ| MachLabel::from_block(*succ))); 1068 opt_inst 1069 } 1070 1071 /// Collect the outgoing block-call arguments for a given edge out 1072 /// of a lowered block. 1073 fn collect_block_call<'a>( 1074 &mut self, 1075 block: BlockIndex, 1076 succ_idx: usize, 1077 buffer: &'a mut SmallVec<[Reg; 16]>, 1078 ) -> (BlockIndex, &'a [Reg]) { 1079 let block_order = self.vcode.block_order(); 1080 let (_, succs) = block_order.succ_indices(block); 1081 let succ = succs[succ_idx]; 1082 let this_lb = block_order.lowered_order()[block.index()]; 1083 let succ_lb = block_order.lowered_order()[succ.index()]; 1084 1085 let (branch_inst, succ_idx) = match (this_lb, succ_lb) { 1086 (_, LoweredBlock::CriticalEdge { .. }) => { 1087 // The successor is a split-critical-edge block. In this 1088 // case, this block-call has no arguments, and the 1089 // arguments go on the critical edge block's unconditional 1090 // branch instead. 1091 return (succ, &[]); 1092 } 1093 (LoweredBlock::CriticalEdge { pred, succ_idx, .. }, _) => { 1094 // This is a split-critical-edge block. In this case, our 1095 // block-call has the arguments that in the CLIF appear in 1096 // the predecessor's branch to this edge. 1097 let branch_inst = self.f.layout.last_inst(pred).unwrap(); 1098 (branch_inst, succ_idx as usize) 1099 } 1100 1101 (this, _) => { 1102 let block = this.orig_block().unwrap(); 1103 // Ordinary block, with an ordinary block as 1104 // successor. Take the arguments from the branch. 1105 let branch_inst = self.f.layout.last_inst(block).unwrap(); 1106 (branch_inst, succ_idx) 1107 } 1108 }; 1109 1110 let block_call = self.f.dfg.insts[branch_inst] 1111 .branch_destination(&self.f.dfg.jump_tables, &self.f.dfg.exception_tables)[succ_idx]; 1112 for arg in block_call.args(&self.f.dfg.value_lists) { 1113 match arg { 1114 BlockArg::Value(arg) => { 1115 debug_assert!(self.f.dfg.value_is_real(arg)); 1116 let regs = self.put_value_in_regs(arg); 1117 buffer.extend_from_slice(regs.regs()); 1118 } 1119 BlockArg::TryCallRet(i) => { 1120 let regs = self.try_call_rets.get(&branch_inst).unwrap()[i as usize] 1121 .map(|r| r.to_reg()); 1122 buffer.extend_from_slice(regs.regs()); 1123 } 1124 BlockArg::TryCallExn(i) => { 1125 let reg = 1126 self.try_call_payloads.get(&branch_inst).unwrap()[i as usize].to_reg(); 1127 buffer.push(reg); 1128 } 1129 } 1130 } 1131 (succ, &buffer[..]) 1132 } 1133 1134 /// Lower the function. 1135 pub fn lower<B: LowerBackend<MInst = I>>( 1136 mut self, 1137 backend: &B, 1138 ctrl_plane: &mut ControlPlane, 1139 ) -> CodegenResult<VCode<I>> { 1140 trace!("about to lower function: {:?}", self.f); 1141 1142 self.vcode.init_retval_area(&mut self.vregs)?; 1143 1144 // Get the pinned reg here (we only parameterize this function on `B`, 1145 // not the whole `Lower` impl). 1146 self.pinned_reg = backend.maybe_pinned_reg(); 1147 1148 self.vcode.set_entry(BlockIndex::new(0)); 1149 1150 // Reused vectors for branch lowering. 1151 let mut targets: SmallVec<[MachLabel; 2]> = SmallVec::new(); 1152 1153 // get a copy of the lowered order; we hold this separately because we 1154 // need a mut ref to the vcode to mutate it below. 1155 let lowered_order: SmallVec<[LoweredBlock; 64]> = self 1156 .vcode 1157 .block_order() 1158 .lowered_order() 1159 .iter() 1160 .cloned() 1161 .collect(); 1162 1163 // Main lowering loop over lowered blocks. 1164 for (bindex, lb) in lowered_order.iter().enumerate().rev() { 1165 let bindex = BlockIndex::new(bindex); 1166 1167 // Lower the block body in reverse order (see comment in 1168 // `lower_clif_block()` for rationale). 1169 1170 // End branch. 1171 if let Some(bb) = lb.orig_block() { 1172 if let Some(branch) = self.collect_branch_and_targets(bindex, bb, &mut targets) { 1173 self.lower_clif_branch(backend, bindex, bb, branch, &targets)?; 1174 self.finish_ir_inst(self.srcloc(branch)); 1175 } 1176 } else { 1177 // If no orig block, this must be a pure edge block; 1178 // get the successor and emit a jump. This block has 1179 // no block params; and this jump's block-call args 1180 // will be filled in by 1181 // `lower_branch_blockparam_args`. 1182 let succ = self.vcode.block_order().succ_indices(bindex).1[0]; 1183 self.emit(I::gen_jump(MachLabel::from_block(succ))); 1184 self.finish_ir_inst(Default::default()); 1185 self.lower_branch_blockparam_args(bindex); 1186 } 1187 1188 // Original block body. 1189 if let Some(bb) = lb.orig_block() { 1190 self.lower_clif_block(backend, bb, ctrl_plane)?; 1191 self.emit_value_label_live_range_start_for_block_args(bb); 1192 } 1193 1194 if bindex.index() == 0 { 1195 // Set up the function with arg vreg inits. 1196 self.gen_arg_setup(); 1197 self.finish_ir_inst(Default::default()); 1198 } 1199 1200 self.finish_bb(); 1201 1202 // Check for any deferred vreg-temp allocation errors, and 1203 // bubble one up at this time if it exists. 1204 if let Some(e) = self.vregs.take_deferred_error() { 1205 return Err(e); 1206 } 1207 } 1208 1209 // Now that we've emitted all instructions into the 1210 // VCodeBuilder, let's build the VCode. 1211 trace!( 1212 "built vcode:\n{:?}Backwards {:?}", 1213 &self.vregs, &self.vcode.vcode 1214 ); 1215 let vcode = self.vcode.build(self.vregs); 1216 1217 Ok(vcode) 1218 } 1219 1220 pub fn value_is_unused(&self, val: Value) -> bool { 1221 match self.value_ir_uses[val] { 1222 ValueUseState::Unused => true, 1223 _ => false, 1224 } 1225 } 1226 1227 pub fn block_successor_label(&self, block: Block, succ: usize) -> MachLabel { 1228 trace!("block_successor_label: block {block} succ {succ}"); 1229 let lowered = self 1230 .vcode 1231 .block_order() 1232 .lowered_index_for_block(block) 1233 .expect("Unreachable block"); 1234 trace!(" -> lowered block {lowered:?}"); 1235 let (_, succs) = self.vcode.block_order().succ_indices(lowered); 1236 trace!(" -> succs {succs:?}"); 1237 let succ_block = *succs.get(succ).expect("Successor index out of range"); 1238 MachLabel::from_block(succ_block) 1239 } 1240 } 1241 1242 /// Pre-analysis: compute `value_ir_uses`. See comment on 1243 /// `ValueUseState` for a description of what this analysis 1244 /// computes. 1245 fn compute_use_states( 1246 f: &Function, 1247 sret_param: Option<Value>, 1248 ) -> SecondaryMap<Value, ValueUseState> { 1249 // We perform the analysis without recursion, so we don't 1250 // overflow the stack on long chains of ops in the input. 1251 // 1252 // This is sort of a hybrid of a "shallow use-count" pass and 1253 // a DFS. We iterate over all instructions and mark their args 1254 // as used. However when we increment a use-count to 1255 // "Multiple" we push its args onto the stack and do a DFS, 1256 // immediately marking the whole dependency tree as 1257 // Multiple. Doing both (shallow use-counting over all insts, 1258 // and deep Multiple propagation) lets us trim both 1259 // traversals, stopping recursion when a node is already at 1260 // the appropriate state. 1261 // 1262 // In particular, note that the *coarsening* into {Unused, 1263 // Once, Multiple} is part of what makes this pass more 1264 // efficient than a full indirect-use-counting pass. 1265 1266 let mut value_ir_uses = SecondaryMap::with_default(ValueUseState::Unused); 1267 1268 if let Some(sret_param) = sret_param { 1269 // There's an implicit use of the struct-return parameter in each 1270 // copy of the function epilogue, which we count here. 1271 value_ir_uses[sret_param] = ValueUseState::Multiple; 1272 } 1273 1274 // Stack of iterators over Values as we do DFS to mark 1275 // Multiple-state subtrees. The iterator type is whatever is 1276 // returned by `uses` below. 1277 let mut stack: SmallVec<[_; 16]> = smallvec![]; 1278 1279 // Find the args for the inst corresponding to the given value. 1280 // 1281 // Note that "root" instructions are skipped here. This means that multiple 1282 // uses of any result of a multi-result instruction are not considered 1283 // multiple uses of the operands of a multi-result instruction. This 1284 // requires tight coupling with `get_value_as_source_or_const` above which 1285 // is the consumer of the map that this function is producing. 1286 let uses = |value| { 1287 trace!(" -> pushing args for {} onto stack", value); 1288 if let ValueDef::Result(src_inst, _) = f.dfg.value_def(value) { 1289 if is_value_use_root(f, src_inst) { 1290 None 1291 } else { 1292 Some(f.dfg.inst_values(src_inst)) 1293 } 1294 } else { 1295 None 1296 } 1297 }; 1298 1299 // Do a DFS through `value_ir_uses` to mark a subtree as 1300 // Multiple. 1301 for inst in f 1302 .layout 1303 .blocks() 1304 .flat_map(|block| f.layout.block_insts(block)) 1305 { 1306 // Iterate over all values used by all instructions, noting an 1307 // additional use on each operand. 1308 for arg in f.dfg.inst_values(inst) { 1309 debug_assert!(f.dfg.value_is_real(arg)); 1310 let old = value_ir_uses[arg]; 1311 value_ir_uses[arg].inc(); 1312 let new = value_ir_uses[arg]; 1313 trace!("arg {} used, old state {:?}, new {:?}", arg, old, new); 1314 1315 // On transition to Multiple, do DFS. 1316 if old == ValueUseState::Multiple || new != ValueUseState::Multiple { 1317 continue; 1318 } 1319 if let Some(iter) = uses(arg) { 1320 stack.push(iter); 1321 } 1322 while let Some(iter) = stack.last_mut() { 1323 if let Some(value) = iter.next() { 1324 debug_assert!(f.dfg.value_is_real(value)); 1325 trace!(" -> DFS reaches {}", value); 1326 if value_ir_uses[value] == ValueUseState::Multiple { 1327 // Truncate DFS here: no need to go further, 1328 // as whole subtree must already be Multiple. 1329 // With debug asserts, check one level of 1330 // that invariant at least. 1331 debug_assert!(uses(value).into_iter().flatten().all(|arg| { 1332 debug_assert!(f.dfg.value_is_real(arg)); 1333 value_ir_uses[arg] == ValueUseState::Multiple 1334 })); 1335 continue; 1336 } 1337 value_ir_uses[value] = ValueUseState::Multiple; 1338 trace!(" -> became Multiple"); 1339 if let Some(iter) = uses(value) { 1340 stack.push(iter); 1341 } 1342 } else { 1343 // Empty iterator, discard. 1344 stack.pop(); 1345 } 1346 } 1347 } 1348 } 1349 1350 value_ir_uses 1351 } 1352 1353 /// Definition of a "root" instruction for the calculation of `ValueUseState`. 1354 /// 1355 /// This function calculates whether `inst` is considered a "root" for value-use 1356 /// information. This concept is used to forcibly prevent looking-through the 1357 /// instruction during `get_value_as_source_or_const` as it additionally 1358 /// prevents propagating `Multiple`-used results of the `inst` here to the 1359 /// operands of the instruction. 1360 /// 1361 /// Currently this is defined as multi-result instructions. That means that 1362 /// lowerings are never allowed to look through a multi-result instruction to 1363 /// generate patterns. Note that this isn't possible in ISLE today anyway so 1364 /// this isn't currently much of a loss. 1365 /// 1366 /// The main purpose of this function is to prevent the operands of a 1367 /// multi-result instruction from being forcibly considered `Multiple`-used 1368 /// regardless of circumstances. 1369 fn is_value_use_root(f: &Function, inst: Inst) -> bool { 1370 f.dfg.inst_results(inst).len() > 1 1371 } 1372 1373 /// Function-level queries. 1374 impl<'func, I: VCodeInst> Lower<'func, I> { 1375 pub fn dfg(&self) -> &DataFlowGraph { 1376 &self.f.dfg 1377 } 1378 1379 /// Get the `Callee`. 1380 pub fn abi(&self) -> &Callee<I::ABIMachineSpec> { 1381 self.vcode.abi() 1382 } 1383 1384 /// Get the `Callee`. 1385 pub fn abi_mut(&mut self) -> &mut Callee<I::ABIMachineSpec> { 1386 self.vcode.abi_mut() 1387 } 1388 } 1389 1390 /// Instruction input/output queries. 1391 impl<'func, I: VCodeInst> Lower<'func, I> { 1392 /// Get the instdata for a given IR instruction. 1393 pub fn data(&self, ir_inst: Inst) -> &InstructionData { 1394 &self.f.dfg.insts[ir_inst] 1395 } 1396 1397 /// Likewise, but starting with a GlobalValue identifier. 1398 pub fn symbol_value_data<'b>( 1399 &'b self, 1400 global_value: GlobalValue, 1401 ) -> Option<(&'b ExternalName, RelocDistance, i64)> { 1402 let gvdata = &self.f.global_values[global_value]; 1403 match gvdata { 1404 &GlobalValueData::Symbol { 1405 ref name, 1406 ref offset, 1407 colocated, 1408 .. 1409 } => { 1410 let offset = offset.bits(); 1411 let dist = if colocated { 1412 RelocDistance::Near 1413 } else { 1414 RelocDistance::Far 1415 }; 1416 Some((name, dist, offset)) 1417 } 1418 _ => None, 1419 } 1420 } 1421 1422 /// Returns the memory flags of a given memory access. 1423 pub fn memflags(&self, ir_inst: Inst) -> Option<MemFlags> { 1424 match &self.f.dfg.insts[ir_inst] { 1425 &InstructionData::AtomicCas { flags, .. } => Some(flags), 1426 &InstructionData::AtomicRmw { flags, .. } => Some(flags), 1427 &InstructionData::Load { flags, .. } 1428 | &InstructionData::LoadNoOffset { flags, .. } 1429 | &InstructionData::Store { flags, .. } => Some(flags), 1430 &InstructionData::StoreNoOffset { flags, .. } => Some(flags), 1431 _ => None, 1432 } 1433 } 1434 1435 /// Get the source location for a given instruction. 1436 pub fn srcloc(&self, ir_inst: Inst) -> RelSourceLoc { 1437 self.f.rel_srclocs()[ir_inst] 1438 } 1439 1440 /// Get the number of inputs to the given IR instruction. This is a count only of the Value 1441 /// arguments to the instruction: block arguments will not be included in this count. 1442 pub fn num_inputs(&self, ir_inst: Inst) -> usize { 1443 self.f.dfg.inst_args(ir_inst).len() 1444 } 1445 1446 /// Get the number of outputs to the given IR instruction. 1447 pub fn num_outputs(&self, ir_inst: Inst) -> usize { 1448 self.f.dfg.inst_results(ir_inst).len() 1449 } 1450 1451 /// Get the type for an instruction's input. 1452 pub fn input_ty(&self, ir_inst: Inst, idx: usize) -> Type { 1453 self.value_ty(self.input_as_value(ir_inst, idx)) 1454 } 1455 1456 /// Get the type for a value. 1457 pub fn value_ty(&self, val: Value) -> Type { 1458 self.f.dfg.value_type(val) 1459 } 1460 1461 /// Get the type for an instruction's output. 1462 pub fn output_ty(&self, ir_inst: Inst, idx: usize) -> Type { 1463 self.f.dfg.value_type(self.f.dfg.inst_results(ir_inst)[idx]) 1464 } 1465 1466 /// Get the value of a constant instruction (`iconst`, etc.) as a 64-bit 1467 /// value, if possible. 1468 pub fn get_constant(&self, ir_inst: Inst) -> Option<u64> { 1469 self.inst_constants.get(&ir_inst).map(|&c| { 1470 // The upper bits must be zero, enforced during legalization and by 1471 // the CLIF verifier. 1472 debug_assert_eq!(c, { 1473 let input_size = self.output_ty(ir_inst, 0).bits() as u64; 1474 let shift = 64 - input_size; 1475 (c << shift) >> shift 1476 }); 1477 c 1478 }) 1479 } 1480 1481 /// Get the input as one of two options other than a direct register: 1482 /// 1483 /// - An instruction, given that it is effect-free or able to sink its 1484 /// effect to the current instruction being lowered, and given it has only 1485 /// one output, and if effect-ful, given that this is the only use; 1486 /// - A constant, if the value is a constant. 1487 /// 1488 /// The instruction input may be available in either of these forms. It may 1489 /// be available in neither form, if the conditions are not met; if so, use 1490 /// `put_input_in_regs()` instead to get it in a register. 1491 /// 1492 /// If the backend merges the effect of a side-effecting instruction, it 1493 /// must call `sink_inst()`. When this is called, it indicates that the 1494 /// effect has been sunk to the current scan location. The sunk 1495 /// instruction's result(s) must have *no* uses remaining, because it will 1496 /// not be codegen'd (it has been integrated into the current instruction). 1497 pub fn input_as_value(&self, ir_inst: Inst, idx: usize) -> Value { 1498 let val = self.f.dfg.inst_args(ir_inst)[idx]; 1499 debug_assert!(self.f.dfg.value_is_real(val)); 1500 val 1501 } 1502 1503 /// Resolves a particular input of an instruction to the `Value` that it is 1504 /// represented with. 1505 /// 1506 /// For more information see [`Lower::get_value_as_source_or_const`]. 1507 pub fn get_input_as_source_or_const(&self, ir_inst: Inst, idx: usize) -> NonRegInput { 1508 let val = self.input_as_value(ir_inst, idx); 1509 self.get_value_as_source_or_const(val) 1510 } 1511 1512 /// Resolves a `Value` definition to the source instruction it came from 1513 /// plus whether it's a unique-use of that instruction. 1514 /// 1515 /// This function is the workhorse of pattern-matching in ISLE which enables 1516 /// combining multiple instructions together. This is used implicitly in 1517 /// patterns such as `(iadd x (iconst y))` where this function is used to 1518 /// extract the `(iconst y)` operand. 1519 /// 1520 /// At its core this function is a wrapper around 1521 /// [`DataFlowGraph::value_def`]. This function applies a filter on top of 1522 /// that, however, to determine when it is actually safe to "look through" 1523 /// the `val` definition here and view the underlying instruction. This 1524 /// protects against duplicating side effects, such as loads, for example. 1525 /// 1526 /// Internally this uses the data computed from `compute_use_states` along 1527 /// with other instruction properties to know what to return. 1528 pub fn get_value_as_source_or_const(&self, val: Value) -> NonRegInput { 1529 trace!( 1530 "get_input_for_val: val {} at cur_inst {:?} cur_scan_entry_color {:?}", 1531 val, self.cur_inst, self.cur_scan_entry_color, 1532 ); 1533 let inst = match self.f.dfg.value_def(val) { 1534 // OK to merge source instruction if we have a source 1535 // instruction, and one of these two conditions hold: 1536 // 1537 // - It has no side-effects and this instruction is not a "value-use 1538 // root" instruction. Instructions which are considered "roots" 1539 // for value-use calculations do not have accurate information 1540 // known about the `ValueUseState` of their operands. This is 1541 // currently done for multi-result instructions to prevent a use 1542 // of each result from forcing all operands of the multi-result 1543 // instruction to also be `Multiple`. This in turn means that the 1544 // `ValueUseState` for operands of a "root" instruction to be a 1545 // lie if pattern matching were to look through the multi-result 1546 // instruction. As a result the "look through this instruction" 1547 // logic only succeeds if it's not a root instruction. 1548 // 1549 // - It has a side-effect, has one output value, that one 1550 // output has only one use, directly or indirectly (so 1551 // cannot be duplicated -- see comment on 1552 // `ValueUseState`), and the instruction's color is *one 1553 // less than* the current scan color. 1554 // 1555 // This latter set of conditions is testing whether a 1556 // side-effecting instruction can sink to the current scan 1557 // location; this is possible if the in-color of this inst is 1558 // equal to the out-color of the producing inst, so no other 1559 // side-effecting ops occur between them (which will only be true 1560 // if they are in the same BB, because color increments at each BB 1561 // start). 1562 // 1563 // If it is actually sunk, then in `merge_inst()`, we update the 1564 // scan color so that as we scan over the range past which the 1565 // instruction was sunk, we allow other instructions (that came 1566 // prior to the sunk instruction) to sink. 1567 ValueDef::Result(src_inst, result_idx) => { 1568 let src_side_effect = has_lowering_side_effect(self.f, src_inst); 1569 trace!(" -> src inst {}", self.f.dfg.display_inst(src_inst)); 1570 trace!(" -> has lowering side effect: {}", src_side_effect); 1571 if is_value_use_root(self.f, src_inst) { 1572 // If this instruction is a "root instruction" then it's 1573 // required that we can't look through it to see the 1574 // definition. This means that the `ValueUseState` for the 1575 // operands of this result assume that this instruction is 1576 // generated exactly once which might get violated were we 1577 // to allow looking through it. 1578 trace!(" -> is a root instruction"); 1579 InputSourceInst::None 1580 } else if !src_side_effect { 1581 // Otherwise if this instruction has no side effects and the 1582 // value is used only once then we can look through it with 1583 // a "unique" tag. A non-unique `Use` can be shown for other 1584 // values ensuring consumers know how it's computed but that 1585 // it's not available to omit. 1586 if self.value_ir_uses[val] == ValueUseState::Once { 1587 InputSourceInst::UniqueUse(src_inst, result_idx) 1588 } else { 1589 InputSourceInst::Use(src_inst, result_idx) 1590 } 1591 } else { 1592 // Side-effect: test whether this is the only use of the 1593 // only result of the instruction, and whether colors allow 1594 // the code-motion. 1595 trace!( 1596 " -> side-effecting op {} for val {}: use state {:?}", 1597 src_inst, val, self.value_ir_uses[val] 1598 ); 1599 if self.cur_scan_entry_color.is_some() 1600 && self.value_ir_uses[val] == ValueUseState::Once 1601 && self.num_outputs(src_inst) == 1 1602 && self 1603 .side_effect_inst_entry_colors 1604 .get(&src_inst) 1605 .unwrap() 1606 .get() 1607 + 1 1608 == self.cur_scan_entry_color.unwrap().get() 1609 { 1610 InputSourceInst::UniqueUse(src_inst, 0) 1611 } else { 1612 InputSourceInst::None 1613 } 1614 } 1615 } 1616 _ => InputSourceInst::None, 1617 }; 1618 let constant = inst.as_inst().and_then(|(inst, _)| self.get_constant(inst)); 1619 1620 NonRegInput { inst, constant } 1621 } 1622 1623 /// Increment the reference count for the Value, ensuring that it gets lowered. 1624 pub fn increment_lowered_uses(&mut self, val: Value) { 1625 self.value_lowered_uses[val] += 1 1626 } 1627 1628 /// Put the `idx`th input into register(s) and return the assigned register. 1629 pub fn put_input_in_regs(&mut self, ir_inst: Inst, idx: usize) -> ValueRegs<Reg> { 1630 let val = self.f.dfg.inst_args(ir_inst)[idx]; 1631 self.put_value_in_regs(val) 1632 } 1633 1634 /// Put the given value into register(s) and return the assigned register. 1635 pub fn put_value_in_regs(&mut self, val: Value) -> ValueRegs<Reg> { 1636 debug_assert!(self.f.dfg.value_is_real(val)); 1637 trace!("put_value_in_regs: val {}", val); 1638 1639 if let Some(inst) = self.f.dfg.value_def(val).inst() { 1640 assert!(!self.inst_sunk.contains(&inst)); 1641 } 1642 1643 let regs = self.value_regs[val]; 1644 trace!(" -> regs {:?}", regs); 1645 assert!(regs.is_valid()); 1646 1647 self.value_lowered_uses[val] += 1; 1648 1649 regs 1650 } 1651 1652 /// Get the ValueRegs for the edge-defined values for special 1653 /// try-call-return block arguments. 1654 pub fn try_call_return_defs(&mut self, ir_inst: Inst) -> &[ValueRegs<Writable<Reg>>] { 1655 &self.try_call_rets.get(&ir_inst).unwrap()[..] 1656 } 1657 1658 /// Get the Regs for the edge-defined values for special 1659 /// try-call-return exception payload arguments. 1660 pub fn try_call_exception_defs(&mut self, ir_inst: Inst) -> &[Writable<Reg>] { 1661 &self.try_call_payloads.get(&ir_inst).unwrap()[..] 1662 } 1663 } 1664 1665 /// Codegen primitives: allocate temps, emit instructions, set result registers, 1666 /// ask for an input to be gen'd into a register. 1667 impl<'func, I: VCodeInst> Lower<'func, I> { 1668 /// Get a new temp. 1669 pub fn alloc_tmp(&mut self, ty: Type) -> ValueRegs<Writable<Reg>> { 1670 writable_value_regs(self.vregs.alloc_with_deferred_error(ty)) 1671 } 1672 1673 /// Get the current root instruction that we are lowering. 1674 pub fn cur_inst(&self) -> Inst { 1675 self.cur_inst.unwrap() 1676 } 1677 1678 /// Emit a machine instruction. 1679 pub fn emit(&mut self, mach_inst: I) { 1680 trace!("emit: {:?}", mach_inst); 1681 self.ir_insts.push(mach_inst); 1682 } 1683 1684 /// Indicate that the side-effect of an instruction has been sunk to the 1685 /// current scan location. This should only be done with the instruction's 1686 /// original results are not used (i.e., `put_input_in_regs` is not invoked 1687 /// for the input produced by the sunk instruction), otherwise the 1688 /// side-effect will occur twice. 1689 pub fn sink_inst(&mut self, ir_inst: Inst) { 1690 assert!(has_lowering_side_effect(self.f, ir_inst)); 1691 assert!(self.cur_scan_entry_color.is_some()); 1692 1693 for result in self.dfg().inst_results(ir_inst) { 1694 assert!(self.value_lowered_uses[*result] == 0); 1695 } 1696 1697 let sunk_inst_entry_color = self 1698 .side_effect_inst_entry_colors 1699 .get(&ir_inst) 1700 .cloned() 1701 .unwrap(); 1702 let sunk_inst_exit_color = InstColor::new(sunk_inst_entry_color.get() + 1); 1703 assert!(sunk_inst_exit_color == self.cur_scan_entry_color.unwrap()); 1704 self.cur_scan_entry_color = Some(sunk_inst_entry_color); 1705 self.inst_sunk.insert(ir_inst); 1706 } 1707 1708 /// Retrieve immediate data given a handle. 1709 pub fn get_immediate_data(&self, imm: Immediate) -> &ConstantData { 1710 self.f.dfg.immediates.get(imm).unwrap() 1711 } 1712 1713 /// Retrieve constant data given a handle. 1714 pub fn get_constant_data(&self, constant_handle: Constant) -> &ConstantData { 1715 self.f.dfg.constants.get(constant_handle) 1716 } 1717 1718 /// Indicate that a constant should be emitted. 1719 pub fn use_constant(&mut self, constant: VCodeConstantData) -> VCodeConstant { 1720 self.vcode.constants().insert(constant) 1721 } 1722 1723 /// Cause the value in `reg` to be in a virtual reg, by copying it into a 1724 /// new virtual reg if `reg` is a real reg. `ty` describes the type of the 1725 /// value in `reg`. 1726 pub fn ensure_in_vreg(&mut self, reg: Reg, ty: Type) -> Reg { 1727 if reg.to_virtual_reg().is_some() { 1728 reg 1729 } else { 1730 let new_reg = self.alloc_tmp(ty).only_reg().unwrap(); 1731 self.emit(I::gen_move(new_reg, reg, ty)); 1732 new_reg.to_reg() 1733 } 1734 } 1735 1736 /// Add a range fact to a register, if no other fact is present. 1737 pub fn add_range_fact(&mut self, reg: Reg, bit_width: u16, min: u64, max: u64) { 1738 if self.flags.enable_pcc() { 1739 self.vregs.set_fact_if_missing( 1740 reg.to_virtual_reg().unwrap(), 1741 Fact::Range { 1742 bit_width, 1743 min, 1744 max, 1745 }, 1746 ); 1747 } 1748 } 1749 } 1750 1751 #[cfg(test)] 1752 mod tests { 1753 use super::ValueUseState; 1754 use crate::cursor::{Cursor, FuncCursor}; 1755 use crate::ir::types; 1756 use crate::ir::{Function, InstBuilder}; 1757 1758 #[test] 1759 fn multi_result_use_once() { 1760 let mut func = Function::new(); 1761 let block0 = func.dfg.make_block(); 1762 let mut pos = FuncCursor::new(&mut func); 1763 pos.insert_block(block0); 1764 let v1 = pos.ins().iconst(types::I64, 0); 1765 let v2 = pos.ins().iconst(types::I64, 1); 1766 let v3 = pos.ins().iconcat(v1, v2); 1767 let (v4, v5) = pos.ins().isplit(v3); 1768 pos.ins().return_(&[v4, v5]); 1769 let func = pos.func; 1770 1771 let uses = super::compute_use_states(&func, None); 1772 assert_eq!(uses[v1], ValueUseState::Once); 1773 assert_eq!(uses[v2], ValueUseState::Once); 1774 assert_eq!(uses[v3], ValueUseState::Once); 1775 assert_eq!(uses[v4], ValueUseState::Once); 1776 assert_eq!(uses[v5], ValueUseState::Once); 1777 } 1778 1779 #[test] 1780 fn results_used_twice_but_not_operands() { 1781 let mut func = Function::new(); 1782 let block0 = func.dfg.make_block(); 1783 let mut pos = FuncCursor::new(&mut func); 1784 pos.insert_block(block0); 1785 let v1 = pos.ins().iconst(types::I64, 0); 1786 let v2 = pos.ins().iconst(types::I64, 1); 1787 let v3 = pos.ins().iconcat(v1, v2); 1788 let (v4, v5) = pos.ins().isplit(v3); 1789 pos.ins().return_(&[v4, v4]); 1790 let func = pos.func; 1791 1792 let uses = super::compute_use_states(&func, None); 1793 assert_eq!(uses[v1], ValueUseState::Once); 1794 assert_eq!(uses[v2], ValueUseState::Once); 1795 assert_eq!(uses[v3], ValueUseState::Once); 1796 assert_eq!(uses[v4], ValueUseState::Multiple); 1797 assert_eq!(uses[v5], ValueUseState::Unused); 1798 } 1799 } 1800