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 ) -> CallInfo<T> { 714 self.vcode 715 .abi() 716 .gen_call_info(self.vcode.sigs(), sig, dest, uses, defs, try_call_info) 717 } 718 719 /// Has this instruction been sunk to a use-site (i.e., away from its 720 /// original location)? 721 fn is_inst_sunk(&self, inst: Inst) -> bool { 722 self.inst_sunk.contains(&inst) 723 } 724 725 // Is any result of this instruction needed? 726 fn is_any_inst_result_needed(&self, inst: Inst) -> bool { 727 self.f 728 .dfg 729 .inst_results(inst) 730 .iter() 731 .any(|&result| self.value_lowered_uses[result] > 0) 732 } 733 734 fn lower_clif_block<B: LowerBackend<MInst = I>>( 735 &mut self, 736 backend: &B, 737 block: Block, 738 ctrl_plane: &mut ControlPlane, 739 ) -> CodegenResult<()> { 740 self.cur_scan_entry_color = Some(self.block_end_colors[block]); 741 // Lowering loop: 742 // - For each non-branch instruction, in reverse order: 743 // - If side-effecting (load, store, branch/call/return, 744 // possible trap), or if used outside of this block, or if 745 // demanded by another inst, then lower. 746 // 747 // That's it! Lowering of side-effecting ops will force all *needed* 748 // (live) non-side-effecting ops to be lowered at the right places, via 749 // the `use_input_reg()` callback on the `Lower` (that's us). That's 750 // because `use_input_reg()` sets the eager/demand bit for any insts 751 // whose result registers are used. 752 // 753 // We set the VCodeBuilder to "backward" mode, so we emit 754 // blocks in reverse order wrt the BlockIndex sequence, and 755 // emit instructions in reverse order within blocks. Because 756 // the machine backend calls `ctx.emit()` in forward order, we 757 // collect per-IR-inst lowered instructions in `ir_insts`, 758 // then reverse these and append to the VCode at the end of 759 // each IR instruction. 760 for inst in self.f.layout.block_insts(block).rev() { 761 let data = &self.f.dfg.insts[inst]; 762 let has_side_effect = has_lowering_side_effect(self.f, inst); 763 // If inst has been sunk to another location, skip it. 764 if self.is_inst_sunk(inst) { 765 continue; 766 } 767 // Are any outputs used at least once? 768 let value_needed = self.is_any_inst_result_needed(inst); 769 trace!( 770 "lower_clif_block: block {} inst {} ({:?}) is_branch {} side_effect {} value_needed {}", 771 block, 772 inst, 773 data, 774 data.opcode().is_branch(), 775 has_side_effect, 776 value_needed, 777 ); 778 779 // Update scan state to color prior to this inst (as we are scanning 780 // backward). 781 self.cur_inst = Some(inst); 782 if has_side_effect { 783 let entry_color = *self 784 .side_effect_inst_entry_colors 785 .get(&inst) 786 .expect("every side-effecting inst should have a color-map entry"); 787 self.cur_scan_entry_color = Some(entry_color); 788 } 789 790 // Skip lowering branches; these are handled separately 791 // (see `lower_clif_branches()` below). 792 if self.f.dfg.insts[inst].opcode().is_branch() { 793 continue; 794 } 795 796 // Value defined by "inst" becomes live after it in normal 797 // order, and therefore **before** in reversed order. 798 self.emit_value_label_live_range_start_for_inst(inst); 799 800 // Normal instruction: codegen if the instruction is side-effecting 801 // or any of its outputs is used. 802 if has_side_effect || value_needed { 803 trace!("lowering: inst {}: {}", inst, self.f.dfg.display_inst(inst)); 804 let temp_regs = match backend.lower(self, inst) { 805 Some(regs) => regs, 806 None => { 807 let ty = if self.num_outputs(inst) > 0 { 808 Some(self.output_ty(inst, 0)) 809 } else { 810 None 811 }; 812 return Err(CodegenError::Unsupported(format!( 813 "should be implemented in ISLE: inst = `{}`, type = `{:?}`", 814 self.f.dfg.display_inst(inst), 815 ty 816 ))); 817 } 818 }; 819 820 // The ISLE generated code emits its own registers to define the 821 // instruction's lowered values in. However, other instructions 822 // that use this SSA value will be lowered assuming that the value 823 // is generated into a pre-assigned, different, register. 824 // 825 // To connect the two, we set up "aliases" in the VCodeBuilder 826 // that apply when it is building the Operand table for the 827 // regalloc to use. These aliases effectively rewrite any use of 828 // the pre-assigned register to the register that was returned by 829 // the ISLE lowering logic. 830 let results = self.f.dfg.inst_results(inst); 831 debug_assert_eq!(temp_regs.len(), results.len()); 832 for (regs, &result) in temp_regs.iter().zip(results) { 833 let dsts = self.value_regs[result]; 834 let mut regs = regs.regs().iter(); 835 for &dst in dsts.regs().iter() { 836 let temp = regs.next().copied().unwrap_or(Reg::invalid_sentinel()); 837 trace!("set vreg alias: {result:?} = {dst:?}, lowering = {temp:?}"); 838 self.vregs.set_vreg_alias(dst, temp); 839 } 840 } 841 } 842 843 let start = self.vcode.vcode.num_insts(); 844 let loc = self.srcloc(inst); 845 self.finish_ir_inst(loc); 846 847 // If the instruction had a user stack map, forward it from the CLIF 848 // to the vcode. 849 if let Some(entries) = self.f.dfg.user_stack_map_entries(inst) { 850 let end = self.vcode.vcode.num_insts(); 851 debug_assert!(end > start); 852 debug_assert_eq!( 853 (start..end) 854 .filter(|i| self.vcode.vcode[InsnIndex::new(*i)].is_safepoint()) 855 .count(), 856 1 857 ); 858 for i in start..end { 859 let iix = InsnIndex::new(i); 860 if self.vcode.vcode[iix].is_safepoint() { 861 trace!( 862 "Adding user stack map from clif\n\n\ 863 {inst:?} `{}`\n\n\ 864 to vcode\n\n\ 865 {iix:?} `{}`", 866 self.f.dfg.display_inst(inst), 867 &self.vcode.vcode[iix].pretty_print_inst(&mut Default::default()), 868 ); 869 self.vcode 870 .add_user_stack_map(BackwardsInsnIndex::new(iix.index()), entries); 871 break; 872 } 873 } 874 } 875 876 // maybe insert random instruction 877 if ctrl_plane.get_decision() { 878 if ctrl_plane.get_decision() { 879 let imm: u64 = ctrl_plane.get_arbitrary(); 880 let reg = self.alloc_tmp(crate::ir::types::I64).regs()[0]; 881 I::gen_imm_u64(imm, reg).map(|inst| self.emit(inst)); 882 } else { 883 let imm: f64 = ctrl_plane.get_arbitrary(); 884 let tmp = self.alloc_tmp(crate::ir::types::I64).regs()[0]; 885 let reg = self.alloc_tmp(crate::ir::types::F64).regs()[0]; 886 for inst in I::gen_imm_f64(imm, tmp, reg) { 887 self.emit(inst); 888 } 889 } 890 } 891 } 892 893 // Add the block params to this block. 894 self.add_block_params(block)?; 895 896 self.cur_scan_entry_color = None; 897 Ok(()) 898 } 899 900 fn add_block_params(&mut self, block: Block) -> CodegenResult<()> { 901 for ¶m in self.f.dfg.block_params(block) { 902 for ® in self.value_regs[param].regs() { 903 let vreg = reg.to_virtual_reg().unwrap(); 904 self.vcode.add_block_param(vreg); 905 } 906 } 907 Ok(()) 908 } 909 910 fn get_value_labels<'a>(&'a self, val: Value, depth: usize) -> Option<&'a [ValueLabelStart]> { 911 if let Some(ref values_labels) = self.f.dfg.values_labels { 912 debug_assert!(self.f.dfg.value_is_real(val)); 913 trace!( 914 "get_value_labels: val {} -> {:?}", 915 val, 916 values_labels.get(&val) 917 ); 918 match values_labels.get(&val) { 919 Some(&ValueLabelAssignments::Starts(ref list)) => Some(&list[..]), 920 Some(&ValueLabelAssignments::Alias { value, .. }) if depth < 10 => { 921 self.get_value_labels(value, depth + 1) 922 } 923 _ => None, 924 } 925 } else { 926 None 927 } 928 } 929 930 fn emit_value_label_marks_for_value(&mut self, val: Value) { 931 let regs = self.value_regs[val]; 932 if regs.len() > 1 { 933 return; 934 } 935 let reg = regs.only_reg().unwrap(); 936 937 if let Some(label_starts) = self.get_value_labels(val, 0) { 938 let labels = label_starts 939 .iter() 940 .map(|&ValueLabelStart { label, .. }| label) 941 .collect::<FxHashSet<_>>(); 942 for label in labels { 943 trace!( 944 "value labeling: defines val {:?} -> reg {:?} -> label {:?}", 945 val, reg, label, 946 ); 947 self.vcode.add_value_label(reg, label); 948 } 949 } 950 } 951 952 fn emit_value_label_live_range_start_for_inst(&mut self, inst: Inst) { 953 if self.f.dfg.values_labels.is_none() { 954 return; 955 } 956 957 trace!( 958 "value labeling: srcloc {}: inst {}", 959 self.srcloc(inst), 960 inst 961 ); 962 for &val in self.f.dfg.inst_results(inst) { 963 self.emit_value_label_marks_for_value(val); 964 } 965 } 966 967 fn emit_value_label_live_range_start_for_block_args(&mut self, block: Block) { 968 if self.f.dfg.values_labels.is_none() { 969 return; 970 } 971 972 trace!("value labeling: block {}", block); 973 for &arg in self.f.dfg.block_params(block) { 974 self.emit_value_label_marks_for_value(arg); 975 } 976 self.finish_ir_inst(Default::default()); 977 } 978 979 fn finish_ir_inst(&mut self, loc: RelSourceLoc) { 980 // The VCodeBuilder builds in reverse order (and reverses at 981 // the end), but `ir_insts` is in forward order, so reverse 982 // it. 983 for inst in self.ir_insts.drain(..).rev() { 984 self.vcode.push(inst, loc); 985 } 986 } 987 988 fn finish_bb(&mut self) { 989 self.vcode.end_bb(); 990 } 991 992 fn lower_clif_branch<B: LowerBackend<MInst = I>>( 993 &mut self, 994 backend: &B, 995 // Lowered block index: 996 bindex: BlockIndex, 997 // Original CLIF block: 998 block: Block, 999 branch: Inst, 1000 targets: &[MachLabel], 1001 ) -> CodegenResult<()> { 1002 trace!( 1003 "lower_clif_branch: block {} branch {:?} targets {:?}", 1004 block, branch, targets, 1005 ); 1006 // When considering code-motion opportunities, consider the current 1007 // program point to be this branch. 1008 self.cur_inst = Some(branch); 1009 1010 // Lower the branch in ISLE. 1011 backend 1012 .lower_branch(self, branch, targets) 1013 .unwrap_or_else(|| { 1014 panic!( 1015 "should be implemented in ISLE: branch = `{}`", 1016 self.f.dfg.display_inst(branch), 1017 ) 1018 }); 1019 let loc = self.srcloc(branch); 1020 self.finish_ir_inst(loc); 1021 // Add block param outputs for current block. 1022 self.lower_branch_blockparam_args(bindex); 1023 Ok(()) 1024 } 1025 1026 fn lower_branch_blockparam_args(&mut self, block: BlockIndex) { 1027 let mut branch_arg_vregs: SmallVec<[Reg; 16]> = smallvec![]; 1028 1029 // TODO: why not make `block_order` public? 1030 for succ_idx in 0..self.vcode.block_order().succ_indices(block).1.len() { 1031 branch_arg_vregs.clear(); 1032 let (succ, args) = self.collect_block_call(block, succ_idx, &mut branch_arg_vregs); 1033 self.vcode.add_succ(succ, args); 1034 } 1035 } 1036 1037 fn collect_branch_and_targets( 1038 &self, 1039 bindex: BlockIndex, 1040 _bb: Block, 1041 targets: &mut SmallVec<[MachLabel; 2]>, 1042 ) -> Option<Inst> { 1043 targets.clear(); 1044 let (opt_inst, succs) = self.vcode.block_order().succ_indices(bindex); 1045 targets.extend(succs.iter().map(|succ| MachLabel::from_block(*succ))); 1046 opt_inst 1047 } 1048 1049 /// Collect the outgoing block-call arguments for a given edge out 1050 /// of a lowered block. 1051 fn collect_block_call<'a>( 1052 &mut self, 1053 block: BlockIndex, 1054 succ_idx: usize, 1055 buffer: &'a mut SmallVec<[Reg; 16]>, 1056 ) -> (BlockIndex, &'a [Reg]) { 1057 let block_order = self.vcode.block_order(); 1058 let (_, succs) = block_order.succ_indices(block); 1059 let succ = succs[succ_idx]; 1060 let this_lb = block_order.lowered_order()[block.index()]; 1061 let succ_lb = block_order.lowered_order()[succ.index()]; 1062 1063 let (branch_inst, succ_idx) = match (this_lb, succ_lb) { 1064 (_, LoweredBlock::CriticalEdge { .. }) => { 1065 // The successor is a split-critical-edge block. In this 1066 // case, this block-call has no arguments, and the 1067 // arguments go on the critical edge block's unconditional 1068 // branch instead. 1069 return (succ, &[]); 1070 } 1071 (LoweredBlock::CriticalEdge { pred, succ_idx, .. }, _) => { 1072 // This is a split-critical-edge block. In this case, our 1073 // block-call has the arguments that in the CLIF appear in 1074 // the predecessor's branch to this edge. 1075 let branch_inst = self.f.layout.last_inst(pred).unwrap(); 1076 (branch_inst, succ_idx as usize) 1077 } 1078 1079 (this, _) => { 1080 let block = this.orig_block().unwrap(); 1081 // Ordinary block, with an ordinary block as 1082 // successor. Take the arguments from the branch. 1083 let branch_inst = self.f.layout.last_inst(block).unwrap(); 1084 (branch_inst, succ_idx) 1085 } 1086 }; 1087 1088 let block_call = self.f.dfg.insts[branch_inst] 1089 .branch_destination(&self.f.dfg.jump_tables, &self.f.dfg.exception_tables)[succ_idx]; 1090 for arg in block_call.args(&self.f.dfg.value_lists) { 1091 match arg { 1092 BlockArg::Value(arg) => { 1093 debug_assert!(self.f.dfg.value_is_real(arg)); 1094 let regs = self.put_value_in_regs(arg); 1095 buffer.extend_from_slice(regs.regs()); 1096 } 1097 BlockArg::TryCallRet(i) => { 1098 let regs = self.try_call_rets.get(&branch_inst).unwrap()[i as usize] 1099 .map(|r| r.to_reg()); 1100 buffer.extend_from_slice(regs.regs()); 1101 } 1102 BlockArg::TryCallExn(i) => { 1103 let reg = 1104 self.try_call_payloads.get(&branch_inst).unwrap()[i as usize].to_reg(); 1105 buffer.push(reg); 1106 } 1107 } 1108 } 1109 (succ, &buffer[..]) 1110 } 1111 1112 /// Lower the function. 1113 pub fn lower<B: LowerBackend<MInst = I>>( 1114 mut self, 1115 backend: &B, 1116 ctrl_plane: &mut ControlPlane, 1117 ) -> CodegenResult<VCode<I>> { 1118 trace!("about to lower function: {:?}", self.f); 1119 1120 self.vcode.init_retval_area(&mut self.vregs)?; 1121 1122 // Get the pinned reg here (we only parameterize this function on `B`, 1123 // not the whole `Lower` impl). 1124 self.pinned_reg = backend.maybe_pinned_reg(); 1125 1126 self.vcode.set_entry(BlockIndex::new(0)); 1127 1128 // Reused vectors for branch lowering. 1129 let mut targets: SmallVec<[MachLabel; 2]> = SmallVec::new(); 1130 1131 // get a copy of the lowered order; we hold this separately because we 1132 // need a mut ref to the vcode to mutate it below. 1133 let lowered_order: SmallVec<[LoweredBlock; 64]> = self 1134 .vcode 1135 .block_order() 1136 .lowered_order() 1137 .iter() 1138 .cloned() 1139 .collect(); 1140 1141 // Main lowering loop over lowered blocks. 1142 for (bindex, lb) in lowered_order.iter().enumerate().rev() { 1143 let bindex = BlockIndex::new(bindex); 1144 1145 // Lower the block body in reverse order (see comment in 1146 // `lower_clif_block()` for rationale). 1147 1148 // End branch. 1149 if let Some(bb) = lb.orig_block() { 1150 if let Some(branch) = self.collect_branch_and_targets(bindex, bb, &mut targets) { 1151 self.lower_clif_branch(backend, bindex, bb, branch, &targets)?; 1152 self.finish_ir_inst(self.srcloc(branch)); 1153 } 1154 } else { 1155 // If no orig block, this must be a pure edge block; 1156 // get the successor and emit a jump. This block has 1157 // no block params; and this jump's block-call args 1158 // will be filled in by 1159 // `lower_branch_blockparam_args`. 1160 let succ = self.vcode.block_order().succ_indices(bindex).1[0]; 1161 self.emit(I::gen_jump(MachLabel::from_block(succ))); 1162 self.finish_ir_inst(Default::default()); 1163 self.lower_branch_blockparam_args(bindex); 1164 } 1165 1166 // Original block body. 1167 if let Some(bb) = lb.orig_block() { 1168 self.lower_clif_block(backend, bb, ctrl_plane)?; 1169 self.emit_value_label_live_range_start_for_block_args(bb); 1170 } 1171 1172 if bindex.index() == 0 { 1173 // Set up the function with arg vreg inits. 1174 self.gen_arg_setup(); 1175 self.finish_ir_inst(Default::default()); 1176 } 1177 1178 self.finish_bb(); 1179 1180 // Check for any deferred vreg-temp allocation errors, and 1181 // bubble one up at this time if it exists. 1182 if let Some(e) = self.vregs.take_deferred_error() { 1183 return Err(e); 1184 } 1185 } 1186 1187 // Now that we've emitted all instructions into the 1188 // VCodeBuilder, let's build the VCode. 1189 trace!( 1190 "built vcode:\n{:?}Backwards {:?}", 1191 &self.vregs, &self.vcode.vcode 1192 ); 1193 let vcode = self.vcode.build(self.vregs); 1194 1195 Ok(vcode) 1196 } 1197 1198 pub fn value_is_unused(&self, val: Value) -> bool { 1199 match self.value_ir_uses[val] { 1200 ValueUseState::Unused => true, 1201 _ => false, 1202 } 1203 } 1204 } 1205 1206 /// Pre-analysis: compute `value_ir_uses`. See comment on 1207 /// `ValueUseState` for a description of what this analysis 1208 /// computes. 1209 fn compute_use_states( 1210 f: &Function, 1211 sret_param: Option<Value>, 1212 ) -> SecondaryMap<Value, ValueUseState> { 1213 // We perform the analysis without recursion, so we don't 1214 // overflow the stack on long chains of ops in the input. 1215 // 1216 // This is sort of a hybrid of a "shallow use-count" pass and 1217 // a DFS. We iterate over all instructions and mark their args 1218 // as used. However when we increment a use-count to 1219 // "Multiple" we push its args onto the stack and do a DFS, 1220 // immediately marking the whole dependency tree as 1221 // Multiple. Doing both (shallow use-counting over all insts, 1222 // and deep Multiple propagation) lets us trim both 1223 // traversals, stopping recursion when a node is already at 1224 // the appropriate state. 1225 // 1226 // In particular, note that the *coarsening* into {Unused, 1227 // Once, Multiple} is part of what makes this pass more 1228 // efficient than a full indirect-use-counting pass. 1229 1230 let mut value_ir_uses = SecondaryMap::with_default(ValueUseState::Unused); 1231 1232 if let Some(sret_param) = sret_param { 1233 // There's an implicit use of the struct-return parameter in each 1234 // copy of the function epilogue, which we count here. 1235 value_ir_uses[sret_param] = ValueUseState::Multiple; 1236 } 1237 1238 // Stack of iterators over Values as we do DFS to mark 1239 // Multiple-state subtrees. The iterator type is whatever is 1240 // returned by `uses` below. 1241 let mut stack: SmallVec<[_; 16]> = smallvec![]; 1242 1243 // Find the args for the inst corresponding to the given value. 1244 // 1245 // Note that "root" instructions are skipped here. This means that multiple 1246 // uses of any result of a multi-result instruction are not considered 1247 // multiple uses of the operands of a multi-result instruction. This 1248 // requires tight coupling with `get_value_as_source_or_const` above which 1249 // is the consumer of the map that this function is producing. 1250 let uses = |value| { 1251 trace!(" -> pushing args for {} onto stack", value); 1252 if let ValueDef::Result(src_inst, _) = f.dfg.value_def(value) { 1253 if is_value_use_root(f, src_inst) { 1254 None 1255 } else { 1256 Some(f.dfg.inst_values(src_inst)) 1257 } 1258 } else { 1259 None 1260 } 1261 }; 1262 1263 // Do a DFS through `value_ir_uses` to mark a subtree as 1264 // Multiple. 1265 for inst in f 1266 .layout 1267 .blocks() 1268 .flat_map(|block| f.layout.block_insts(block)) 1269 { 1270 // Iterate over all values used by all instructions, noting an 1271 // additional use on each operand. 1272 for arg in f.dfg.inst_values(inst) { 1273 debug_assert!(f.dfg.value_is_real(arg)); 1274 let old = value_ir_uses[arg]; 1275 value_ir_uses[arg].inc(); 1276 let new = value_ir_uses[arg]; 1277 trace!("arg {} used, old state {:?}, new {:?}", arg, old, new); 1278 1279 // On transition to Multiple, do DFS. 1280 if old == ValueUseState::Multiple || new != ValueUseState::Multiple { 1281 continue; 1282 } 1283 if let Some(iter) = uses(arg) { 1284 stack.push(iter); 1285 } 1286 while let Some(iter) = stack.last_mut() { 1287 if let Some(value) = iter.next() { 1288 debug_assert!(f.dfg.value_is_real(value)); 1289 trace!(" -> DFS reaches {}", value); 1290 if value_ir_uses[value] == ValueUseState::Multiple { 1291 // Truncate DFS here: no need to go further, 1292 // as whole subtree must already be Multiple. 1293 // With debug asserts, check one level of 1294 // that invariant at least. 1295 debug_assert!(uses(value).into_iter().flatten().all(|arg| { 1296 debug_assert!(f.dfg.value_is_real(arg)); 1297 value_ir_uses[arg] == ValueUseState::Multiple 1298 })); 1299 continue; 1300 } 1301 value_ir_uses[value] = ValueUseState::Multiple; 1302 trace!(" -> became Multiple"); 1303 if let Some(iter) = uses(value) { 1304 stack.push(iter); 1305 } 1306 } else { 1307 // Empty iterator, discard. 1308 stack.pop(); 1309 } 1310 } 1311 } 1312 } 1313 1314 value_ir_uses 1315 } 1316 1317 /// Definition of a "root" instruction for the calculation of `ValueUseState`. 1318 /// 1319 /// This function calculates whether `inst` is considered a "root" for value-use 1320 /// information. This concept is used to forcibly prevent looking-through the 1321 /// instruction during `get_value_as_source_or_const` as it additionally 1322 /// prevents propagating `Multiple`-used results of the `inst` here to the 1323 /// operands of the instruction. 1324 /// 1325 /// Currently this is defined as multi-result instructions. That means that 1326 /// lowerings are never allowed to look through a multi-result instruction to 1327 /// generate patterns. Note that this isn't possible in ISLE today anyway so 1328 /// this isn't currently much of a loss. 1329 /// 1330 /// The main purpose of this function is to prevent the operands of a 1331 /// multi-result instruction from being forcibly considered `Multiple`-used 1332 /// regardless of circumstances. 1333 fn is_value_use_root(f: &Function, inst: Inst) -> bool { 1334 f.dfg.inst_results(inst).len() > 1 1335 } 1336 1337 /// Function-level queries. 1338 impl<'func, I: VCodeInst> Lower<'func, I> { 1339 pub fn dfg(&self) -> &DataFlowGraph { 1340 &self.f.dfg 1341 } 1342 1343 /// Get the `Callee`. 1344 pub fn abi(&self) -> &Callee<I::ABIMachineSpec> { 1345 self.vcode.abi() 1346 } 1347 1348 /// Get the `Callee`. 1349 pub fn abi_mut(&mut self) -> &mut Callee<I::ABIMachineSpec> { 1350 self.vcode.abi_mut() 1351 } 1352 } 1353 1354 /// Instruction input/output queries. 1355 impl<'func, I: VCodeInst> Lower<'func, I> { 1356 /// Get the instdata for a given IR instruction. 1357 pub fn data(&self, ir_inst: Inst) -> &InstructionData { 1358 &self.f.dfg.insts[ir_inst] 1359 } 1360 1361 /// Likewise, but starting with a GlobalValue identifier. 1362 pub fn symbol_value_data<'b>( 1363 &'b self, 1364 global_value: GlobalValue, 1365 ) -> Option<(&'b ExternalName, RelocDistance, i64)> { 1366 let gvdata = &self.f.global_values[global_value]; 1367 match gvdata { 1368 &GlobalValueData::Symbol { 1369 ref name, 1370 ref offset, 1371 colocated, 1372 .. 1373 } => { 1374 let offset = offset.bits(); 1375 let dist = if colocated { 1376 RelocDistance::Near 1377 } else { 1378 RelocDistance::Far 1379 }; 1380 Some((name, dist, offset)) 1381 } 1382 _ => None, 1383 } 1384 } 1385 1386 /// Returns the memory flags of a given memory access. 1387 pub fn memflags(&self, ir_inst: Inst) -> Option<MemFlags> { 1388 match &self.f.dfg.insts[ir_inst] { 1389 &InstructionData::AtomicCas { flags, .. } => Some(flags), 1390 &InstructionData::AtomicRmw { flags, .. } => Some(flags), 1391 &InstructionData::Load { flags, .. } 1392 | &InstructionData::LoadNoOffset { flags, .. } 1393 | &InstructionData::Store { flags, .. } => Some(flags), 1394 &InstructionData::StoreNoOffset { flags, .. } => Some(flags), 1395 _ => None, 1396 } 1397 } 1398 1399 /// Get the source location for a given instruction. 1400 pub fn srcloc(&self, ir_inst: Inst) -> RelSourceLoc { 1401 self.f.rel_srclocs()[ir_inst] 1402 } 1403 1404 /// Get the number of inputs to the given IR instruction. This is a count only of the Value 1405 /// arguments to the instruction: block arguments will not be included in this count. 1406 pub fn num_inputs(&self, ir_inst: Inst) -> usize { 1407 self.f.dfg.inst_args(ir_inst).len() 1408 } 1409 1410 /// Get the number of outputs to the given IR instruction. 1411 pub fn num_outputs(&self, ir_inst: Inst) -> usize { 1412 self.f.dfg.inst_results(ir_inst).len() 1413 } 1414 1415 /// Get the type for an instruction's input. 1416 pub fn input_ty(&self, ir_inst: Inst, idx: usize) -> Type { 1417 self.value_ty(self.input_as_value(ir_inst, idx)) 1418 } 1419 1420 /// Get the type for a value. 1421 pub fn value_ty(&self, val: Value) -> Type { 1422 self.f.dfg.value_type(val) 1423 } 1424 1425 /// Get the type for an instruction's output. 1426 pub fn output_ty(&self, ir_inst: Inst, idx: usize) -> Type { 1427 self.f.dfg.value_type(self.f.dfg.inst_results(ir_inst)[idx]) 1428 } 1429 1430 /// Get the value of a constant instruction (`iconst`, etc.) as a 64-bit 1431 /// value, if possible. 1432 pub fn get_constant(&self, ir_inst: Inst) -> Option<u64> { 1433 self.inst_constants.get(&ir_inst).map(|&c| { 1434 // The upper bits must be zero, enforced during legalization and by 1435 // the CLIF verifier. 1436 debug_assert_eq!(c, { 1437 let input_size = self.output_ty(ir_inst, 0).bits() as u64; 1438 let shift = 64 - input_size; 1439 (c << shift) >> shift 1440 }); 1441 c 1442 }) 1443 } 1444 1445 /// Get the input as one of two options other than a direct register: 1446 /// 1447 /// - An instruction, given that it is effect-free or able to sink its 1448 /// effect to the current instruction being lowered, and given it has only 1449 /// one output, and if effect-ful, given that this is the only use; 1450 /// - A constant, if the value is a constant. 1451 /// 1452 /// The instruction input may be available in either of these forms. It may 1453 /// be available in neither form, if the conditions are not met; if so, use 1454 /// `put_input_in_regs()` instead to get it in a register. 1455 /// 1456 /// If the backend merges the effect of a side-effecting instruction, it 1457 /// must call `sink_inst()`. When this is called, it indicates that the 1458 /// effect has been sunk to the current scan location. The sunk 1459 /// instruction's result(s) must have *no* uses remaining, because it will 1460 /// not be codegen'd (it has been integrated into the current instruction). 1461 pub fn input_as_value(&self, ir_inst: Inst, idx: usize) -> Value { 1462 let val = self.f.dfg.inst_args(ir_inst)[idx]; 1463 debug_assert!(self.f.dfg.value_is_real(val)); 1464 val 1465 } 1466 1467 /// Resolves a particular input of an instruction to the `Value` that it is 1468 /// represented with. 1469 /// 1470 /// For more information see [`Lower::get_value_as_source_or_const`]. 1471 pub fn get_input_as_source_or_const(&self, ir_inst: Inst, idx: usize) -> NonRegInput { 1472 let val = self.input_as_value(ir_inst, idx); 1473 self.get_value_as_source_or_const(val) 1474 } 1475 1476 /// Resolves a `Value` definition to the source instruction it came from 1477 /// plus whether it's a unique-use of that instruction. 1478 /// 1479 /// This function is the workhorse of pattern-matching in ISLE which enables 1480 /// combining multiple instructions together. This is used implicitly in 1481 /// patterns such as `(iadd x (iconst y))` where this function is used to 1482 /// extract the `(iconst y)` operand. 1483 /// 1484 /// At its core this function is a wrapper around 1485 /// [`DataFlowGraph::value_def`]. This function applies a filter on top of 1486 /// that, however, to determine when it is actually safe to "look through" 1487 /// the `val` definition here and view the underlying instruction. This 1488 /// protects against duplicating side effects, such as loads, for example. 1489 /// 1490 /// Internally this uses the data computed from `compute_use_states` along 1491 /// with other instruction properties to know what to return. 1492 pub fn get_value_as_source_or_const(&self, val: Value) -> NonRegInput { 1493 trace!( 1494 "get_input_for_val: val {} at cur_inst {:?} cur_scan_entry_color {:?}", 1495 val, self.cur_inst, self.cur_scan_entry_color, 1496 ); 1497 let inst = match self.f.dfg.value_def(val) { 1498 // OK to merge source instruction if we have a source 1499 // instruction, and one of these two conditions hold: 1500 // 1501 // - It has no side-effects and this instruction is not a "value-use 1502 // root" instruction. Instructions which are considered "roots" 1503 // for value-use calculations do not have accurate information 1504 // known about the `ValueUseState` of their operands. This is 1505 // currently done for multi-result instructions to prevent a use 1506 // of each result from forcing all operands of the multi-result 1507 // instruction to also be `Multiple`. This in turn means that the 1508 // `ValueUseState` for operands of a "root" instruction to be a 1509 // lie if pattern matching were to look through the multi-result 1510 // instruction. As a result the "look through this instruction" 1511 // logic only succeeds if it's not a root instruction. 1512 // 1513 // - It has a side-effect, has one output value, that one 1514 // output has only one use, directly or indirectly (so 1515 // cannot be duplicated -- see comment on 1516 // `ValueUseState`), and the instruction's color is *one 1517 // less than* the current scan color. 1518 // 1519 // This latter set of conditions is testing whether a 1520 // side-effecting instruction can sink to the current scan 1521 // location; this is possible if the in-color of this inst is 1522 // equal to the out-color of the producing inst, so no other 1523 // side-effecting ops occur between them (which will only be true 1524 // if they are in the same BB, because color increments at each BB 1525 // start). 1526 // 1527 // If it is actually sunk, then in `merge_inst()`, we update the 1528 // scan color so that as we scan over the range past which the 1529 // instruction was sunk, we allow other instructions (that came 1530 // prior to the sunk instruction) to sink. 1531 ValueDef::Result(src_inst, result_idx) => { 1532 let src_side_effect = has_lowering_side_effect(self.f, src_inst); 1533 trace!(" -> src inst {}", self.f.dfg.display_inst(src_inst)); 1534 trace!(" -> has lowering side effect: {}", src_side_effect); 1535 if is_value_use_root(self.f, src_inst) { 1536 // If this instruction is a "root instruction" then it's 1537 // required that we can't look through it to see the 1538 // definition. This means that the `ValueUseState` for the 1539 // operands of this result assume that this instruction is 1540 // generated exactly once which might get violated were we 1541 // to allow looking through it. 1542 trace!(" -> is a root instruction"); 1543 InputSourceInst::None 1544 } else if !src_side_effect { 1545 // Otherwise if this instruction has no side effects and the 1546 // value is used only once then we can look through it with 1547 // a "unique" tag. A non-unique `Use` can be shown for other 1548 // values ensuring consumers know how it's computed but that 1549 // it's not available to omit. 1550 if self.value_ir_uses[val] == ValueUseState::Once { 1551 InputSourceInst::UniqueUse(src_inst, result_idx) 1552 } else { 1553 InputSourceInst::Use(src_inst, result_idx) 1554 } 1555 } else { 1556 // Side-effect: test whether this is the only use of the 1557 // only result of the instruction, and whether colors allow 1558 // the code-motion. 1559 trace!( 1560 " -> side-effecting op {} for val {}: use state {:?}", 1561 src_inst, val, self.value_ir_uses[val] 1562 ); 1563 if self.cur_scan_entry_color.is_some() 1564 && self.value_ir_uses[val] == ValueUseState::Once 1565 && self.num_outputs(src_inst) == 1 1566 && self 1567 .side_effect_inst_entry_colors 1568 .get(&src_inst) 1569 .unwrap() 1570 .get() 1571 + 1 1572 == self.cur_scan_entry_color.unwrap().get() 1573 { 1574 InputSourceInst::UniqueUse(src_inst, 0) 1575 } else { 1576 InputSourceInst::None 1577 } 1578 } 1579 } 1580 _ => InputSourceInst::None, 1581 }; 1582 let constant = inst.as_inst().and_then(|(inst, _)| self.get_constant(inst)); 1583 1584 NonRegInput { inst, constant } 1585 } 1586 1587 /// Increment the reference count for the Value, ensuring that it gets lowered. 1588 pub fn increment_lowered_uses(&mut self, val: Value) { 1589 self.value_lowered_uses[val] += 1 1590 } 1591 1592 /// Put the `idx`th input into register(s) and return the assigned register. 1593 pub fn put_input_in_regs(&mut self, ir_inst: Inst, idx: usize) -> ValueRegs<Reg> { 1594 let val = self.f.dfg.inst_args(ir_inst)[idx]; 1595 self.put_value_in_regs(val) 1596 } 1597 1598 /// Put the given value into register(s) and return the assigned register. 1599 pub fn put_value_in_regs(&mut self, val: Value) -> ValueRegs<Reg> { 1600 debug_assert!(self.f.dfg.value_is_real(val)); 1601 trace!("put_value_in_regs: val {}", val); 1602 1603 if let Some(inst) = self.f.dfg.value_def(val).inst() { 1604 assert!(!self.inst_sunk.contains(&inst)); 1605 } 1606 1607 let regs = self.value_regs[val]; 1608 trace!(" -> regs {:?}", regs); 1609 assert!(regs.is_valid()); 1610 1611 self.value_lowered_uses[val] += 1; 1612 1613 regs 1614 } 1615 1616 /// Get the ValueRegs for the edge-defined values for special 1617 /// try-call-return block arguments. 1618 pub fn try_call_return_defs(&mut self, ir_inst: Inst) -> &[ValueRegs<Writable<Reg>>] { 1619 &self.try_call_rets.get(&ir_inst).unwrap()[..] 1620 } 1621 1622 /// Get the Regs for the edge-defined values for special 1623 /// try-call-return exception payload arguments. 1624 pub fn try_call_exception_defs(&mut self, ir_inst: Inst) -> &[Writable<Reg>] { 1625 &self.try_call_payloads.get(&ir_inst).unwrap()[..] 1626 } 1627 } 1628 1629 /// Codegen primitives: allocate temps, emit instructions, set result registers, 1630 /// ask for an input to be gen'd into a register. 1631 impl<'func, I: VCodeInst> Lower<'func, I> { 1632 /// Get a new temp. 1633 pub fn alloc_tmp(&mut self, ty: Type) -> ValueRegs<Writable<Reg>> { 1634 writable_value_regs(self.vregs.alloc_with_deferred_error(ty)) 1635 } 1636 1637 /// Get the current root instruction that we are lowering. 1638 pub fn cur_inst(&self) -> Inst { 1639 self.cur_inst.unwrap() 1640 } 1641 1642 /// Emit a machine instruction. 1643 pub fn emit(&mut self, mach_inst: I) { 1644 trace!("emit: {:?}", mach_inst); 1645 self.ir_insts.push(mach_inst); 1646 } 1647 1648 /// Indicate that the side-effect of an instruction has been sunk to the 1649 /// current scan location. This should only be done with the instruction's 1650 /// original results are not used (i.e., `put_input_in_regs` is not invoked 1651 /// for the input produced by the sunk instruction), otherwise the 1652 /// side-effect will occur twice. 1653 pub fn sink_inst(&mut self, ir_inst: Inst) { 1654 assert!(has_lowering_side_effect(self.f, ir_inst)); 1655 assert!(self.cur_scan_entry_color.is_some()); 1656 1657 for result in self.dfg().inst_results(ir_inst) { 1658 assert!(self.value_lowered_uses[*result] == 0); 1659 } 1660 1661 let sunk_inst_entry_color = self 1662 .side_effect_inst_entry_colors 1663 .get(&ir_inst) 1664 .cloned() 1665 .unwrap(); 1666 let sunk_inst_exit_color = InstColor::new(sunk_inst_entry_color.get() + 1); 1667 assert!(sunk_inst_exit_color == self.cur_scan_entry_color.unwrap()); 1668 self.cur_scan_entry_color = Some(sunk_inst_entry_color); 1669 self.inst_sunk.insert(ir_inst); 1670 } 1671 1672 /// Retrieve immediate data given a handle. 1673 pub fn get_immediate_data(&self, imm: Immediate) -> &ConstantData { 1674 self.f.dfg.immediates.get(imm).unwrap() 1675 } 1676 1677 /// Retrieve constant data given a handle. 1678 pub fn get_constant_data(&self, constant_handle: Constant) -> &ConstantData { 1679 self.f.dfg.constants.get(constant_handle) 1680 } 1681 1682 /// Indicate that a constant should be emitted. 1683 pub fn use_constant(&mut self, constant: VCodeConstantData) -> VCodeConstant { 1684 self.vcode.constants().insert(constant) 1685 } 1686 1687 /// Cause the value in `reg` to be in a virtual reg, by copying it into a 1688 /// new virtual reg if `reg` is a real reg. `ty` describes the type of the 1689 /// value in `reg`. 1690 pub fn ensure_in_vreg(&mut self, reg: Reg, ty: Type) -> Reg { 1691 if reg.to_virtual_reg().is_some() { 1692 reg 1693 } else { 1694 let new_reg = self.alloc_tmp(ty).only_reg().unwrap(); 1695 self.emit(I::gen_move(new_reg, reg, ty)); 1696 new_reg.to_reg() 1697 } 1698 } 1699 1700 /// Add a range fact to a register, if no other fact is present. 1701 pub fn add_range_fact(&mut self, reg: Reg, bit_width: u16, min: u64, max: u64) { 1702 if self.flags.enable_pcc() { 1703 self.vregs.set_fact_if_missing( 1704 reg.to_virtual_reg().unwrap(), 1705 Fact::Range { 1706 bit_width, 1707 min, 1708 max, 1709 }, 1710 ); 1711 } 1712 } 1713 } 1714 1715 #[cfg(test)] 1716 mod tests { 1717 use super::ValueUseState; 1718 use crate::cursor::{Cursor, FuncCursor}; 1719 use crate::ir::types; 1720 use crate::ir::{Function, InstBuilder}; 1721 1722 #[test] 1723 fn multi_result_use_once() { 1724 let mut func = Function::new(); 1725 let block0 = func.dfg.make_block(); 1726 let mut pos = FuncCursor::new(&mut func); 1727 pos.insert_block(block0); 1728 let v1 = pos.ins().iconst(types::I64, 0); 1729 let v2 = pos.ins().iconst(types::I64, 1); 1730 let v3 = pos.ins().iconcat(v1, v2); 1731 let (v4, v5) = pos.ins().isplit(v3); 1732 pos.ins().return_(&[v4, v5]); 1733 let func = pos.func; 1734 1735 let uses = super::compute_use_states(&func, None); 1736 assert_eq!(uses[v1], ValueUseState::Once); 1737 assert_eq!(uses[v2], ValueUseState::Once); 1738 assert_eq!(uses[v3], ValueUseState::Once); 1739 assert_eq!(uses[v4], ValueUseState::Once); 1740 assert_eq!(uses[v5], ValueUseState::Once); 1741 } 1742 1743 #[test] 1744 fn results_used_twice_but_not_operands() { 1745 let mut func = Function::new(); 1746 let block0 = func.dfg.make_block(); 1747 let mut pos = FuncCursor::new(&mut func); 1748 pos.insert_block(block0); 1749 let v1 = pos.ins().iconst(types::I64, 0); 1750 let v2 = pos.ins().iconst(types::I64, 1); 1751 let v3 = pos.ins().iconcat(v1, v2); 1752 let (v4, v5) = pos.ins().isplit(v3); 1753 pos.ins().return_(&[v4, v4]); 1754 let func = pos.func; 1755 1756 let uses = super::compute_use_states(&func, None); 1757 assert_eq!(uses[v1], ValueUseState::Once); 1758 assert_eq!(uses[v2], ValueUseState::Once); 1759 assert_eq!(uses[v3], ValueUseState::Once); 1760 assert_eq!(uses[v4], ValueUseState::Multiple); 1761 assert_eq!(uses[v5], ValueUseState::Unused); 1762 } 1763 } 1764