1 //! This implements the VCode container: a CFG of Insts that have been lowered. 2 //! 3 //! VCode is virtual-register code. An instruction in VCode is almost a machine 4 //! instruction; however, its register slots can refer to virtual registers in 5 //! addition to real machine registers. 6 //! 7 //! VCode is structured with traditional basic blocks, and 8 //! each block must be terminated by an unconditional branch (one target), a 9 //! conditional branch (two targets), or a return (no targets). Note that this 10 //! slightly differs from the machine code of most ISAs: in most ISAs, a 11 //! conditional branch has one target (and the not-taken case falls through). 12 //! However, we expect that machine backends will elide branches to the following 13 //! block (i.e., zero-offset jumps), and will be able to codegen a branch-cond / 14 //! branch-uncond pair if *both* targets are not fallthrough. This allows us to 15 //! play with layout prior to final binary emission, as well, if we want. 16 //! 17 //! See the main module comment in `mod.rs` for more details on the VCode-based 18 //! backend pipeline. 19 20 use crate::ir::pcc::*; 21 use crate::ir::{self, types, Constant, ConstantData, ValueLabel}; 22 use crate::ranges::Ranges; 23 use crate::timing; 24 use crate::trace; 25 use crate::CodegenError; 26 use crate::{machinst::*, trace_log_enabled}; 27 use crate::{LabelValueLoc, ValueLocRange}; 28 use regalloc2::{ 29 Edit, Function as RegallocFunction, InstOrEdit, InstRange, MachineEnv, Operand, 30 OperandConstraint, OperandKind, PRegSet, RegClass, 31 }; 32 use rustc_hash::FxHashMap; 33 34 use core::cmp::Ordering; 35 use core::fmt::{self, Write}; 36 use core::mem::take; 37 use cranelift_entity::{entity_impl, Keys}; 38 use std::collections::hash_map::Entry; 39 use std::collections::HashMap; 40 41 /// Index referring to an instruction in VCode. 42 pub type InsnIndex = regalloc2::Inst; 43 44 /// Extension trait for `InsnIndex` to allow conversion to a 45 /// `BackwardsInsnIndex`. 46 trait ToBackwardsInsnIndex { 47 fn to_backwards_insn_index(&self, num_insts: usize) -> BackwardsInsnIndex; 48 } 49 50 impl ToBackwardsInsnIndex for InsnIndex { 51 fn to_backwards_insn_index(&self, num_insts: usize) -> BackwardsInsnIndex { 52 BackwardsInsnIndex::new(num_insts - self.index() - 1) 53 } 54 } 55 56 /// An index referring to an instruction in the VCode when it is backwards, 57 /// during VCode construction. 58 #[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)] 59 #[cfg_attr( 60 feature = "enable-serde", 61 derive(::serde::Serialize, ::serde::Deserialize) 62 )] 63 pub struct BackwardsInsnIndex(InsnIndex); 64 65 impl BackwardsInsnIndex { 66 pub fn new(i: usize) -> Self { 67 BackwardsInsnIndex(InsnIndex::new(i)) 68 } 69 } 70 71 /// Index referring to a basic block in VCode. 72 pub type BlockIndex = regalloc2::Block; 73 74 /// VCodeInst wraps all requirements for a MachInst to be in VCode: it must be 75 /// a `MachInst` and it must be able to emit itself at least to a `SizeCodeSink`. 76 pub trait VCodeInst: MachInst + MachInstEmit {} 77 impl<I: MachInst + MachInstEmit> VCodeInst for I {} 78 79 /// A function in "VCode" (virtualized-register code) form, after 80 /// lowering. This is essentially a standard CFG of basic blocks, 81 /// where each basic block consists of lowered instructions produced 82 /// by the machine-specific backend. 83 /// 84 /// Note that the VCode is immutable once produced, and is not 85 /// modified by register allocation in particular. Rather, register 86 /// allocation on the `VCode` produces a separate `regalloc2::Output` 87 /// struct, and this can be passed to `emit`. `emit` in turn does not 88 /// modify the vcode, but produces an `EmitResult`, which contains the 89 /// machine code itself, and the associated disassembly and/or 90 /// metadata as requested. 91 pub struct VCode<I: VCodeInst> { 92 /// VReg IR-level types. 93 vreg_types: Vec<Type>, 94 95 /// Lowered machine instructions in order corresponding to the original IR. 96 insts: Vec<I>, 97 98 /// A map from backwards instruction index to the user stack map for that 99 /// instruction. 100 /// 101 /// This is a sparse side table that only has entries for instructions that 102 /// are safepoints, and only for a subset of those that have an associated 103 /// user stack map. 104 user_stack_maps: FxHashMap<BackwardsInsnIndex, ir::UserStackMap>, 105 106 /// Operands: pre-regalloc references to virtual registers with 107 /// constraints, in one flattened array. This allows the regalloc 108 /// to efficiently access all operands without requiring expensive 109 /// matches or method invocations on insts. 110 operands: Vec<Operand>, 111 112 /// Operand index ranges: for each instruction in `insts`, there 113 /// is a tuple here providing the range in `operands` for that 114 /// instruction's operands. 115 operand_ranges: Ranges, 116 117 /// Clobbers: a sparse map from instruction indices to clobber masks. 118 clobbers: FxHashMap<InsnIndex, PRegSet>, 119 120 /// Source locations for each instruction. (`SourceLoc` is a `u32`, so it is 121 /// reasonable to keep one of these per instruction.) 122 srclocs: Vec<RelSourceLoc>, 123 124 /// Entry block. 125 entry: BlockIndex, 126 127 /// Block instruction indices. 128 block_ranges: Ranges, 129 130 /// Block successors: index range in the `block_succs` list. 131 block_succ_range: Ranges, 132 133 /// Block successor lists, concatenated into one vec. The 134 /// `block_succ_range` list of tuples above gives (start, end) 135 /// ranges within this list that correspond to each basic block's 136 /// successors. 137 block_succs: Vec<regalloc2::Block>, 138 139 /// Block predecessors: index range in the `block_preds` list. 140 block_pred_range: Ranges, 141 142 /// Block predecessor lists, concatenated into one vec. The 143 /// `block_pred_range` list of tuples above gives (start, end) 144 /// ranges within this list that correspond to each basic block's 145 /// predecessors. 146 block_preds: Vec<regalloc2::Block>, 147 148 /// Block parameters: index range in `block_params` below. 149 block_params_range: Ranges, 150 151 /// Block parameter lists, concatenated into one vec. The 152 /// `block_params_range` list of tuples above gives (start, end) 153 /// ranges within this list that correspond to each basic block's 154 /// blockparam vregs. 155 block_params: Vec<regalloc2::VReg>, 156 157 /// Outgoing block arguments on branch instructions, concatenated 158 /// into one list. 159 /// 160 /// Note that this is conceptually a 3D array: we have a VReg list 161 /// per block, per successor. We flatten those three dimensions 162 /// into this 1D vec, then store index ranges in two levels of 163 /// indirection. 164 /// 165 /// Indexed by the indices in `branch_block_arg_succ_range`. 166 branch_block_args: Vec<regalloc2::VReg>, 167 168 /// Array of sequences of (start, end) tuples in 169 /// `branch_block_args`, one for each successor; these sequences 170 /// for each block are concatenated. 171 /// 172 /// Indexed by the indices in `branch_block_arg_succ_range`. 173 branch_block_arg_range: Ranges, 174 175 /// For a given block, indices in `branch_block_arg_range` 176 /// corresponding to all of its successors. 177 branch_block_arg_succ_range: Ranges, 178 179 /// Block-order information. 180 block_order: BlockLoweringOrder, 181 182 /// ABI object. 183 pub(crate) abi: Callee<I::ABIMachineSpec>, 184 185 /// Constant information used during code emission. This should be 186 /// immutable across function compilations within the same module. 187 emit_info: I::Info, 188 189 /// Constants. 190 pub(crate) constants: VCodeConstants, 191 192 /// Value labels for debuginfo attached to vregs. 193 debug_value_labels: Vec<(VReg, InsnIndex, InsnIndex, u32)>, 194 195 pub(crate) sigs: SigSet, 196 197 /// Facts on VRegs, for proof-carrying code verification. 198 facts: Vec<Option<Fact>>, 199 200 log2_min_function_alignment: u8, 201 } 202 203 /// The result of `VCode::emit`. Contains all information computed 204 /// during emission: actual machine code, optionally a disassembly, 205 /// and optionally metadata about the code layout. 206 pub struct EmitResult { 207 /// The MachBuffer containing the machine code. 208 pub buffer: MachBufferFinalized<Stencil>, 209 210 /// Offset of each basic block, recorded during emission. Computed 211 /// only if `debug_value_labels` is non-empty. 212 pub bb_offsets: Vec<CodeOffset>, 213 214 /// Final basic-block edges, in terms of code offsets of 215 /// bb-starts. Computed only if `debug_value_labels` is non-empty. 216 pub bb_edges: Vec<(CodeOffset, CodeOffset)>, 217 218 /// Final length of function body. 219 pub func_body_len: CodeOffset, 220 221 /// The pretty-printed disassembly, if any. This uses the same 222 /// pretty-printing for MachInsts as the pre-regalloc VCode Debug 223 /// implementation, but additionally includes the prologue and 224 /// epilogue(s), and makes use of the regalloc results. 225 pub disasm: Option<String>, 226 227 /// Offsets of sized stackslots. 228 pub sized_stackslot_offsets: PrimaryMap<StackSlot, u32>, 229 230 /// Offsets of dynamic stackslots. 231 pub dynamic_stackslot_offsets: PrimaryMap<DynamicStackSlot, u32>, 232 233 /// Value-labels information (debug metadata). 234 pub value_labels_ranges: ValueLabelsRanges, 235 236 /// Stack frame size. 237 pub frame_size: u32, 238 } 239 240 /// A builder for a VCode function body. 241 /// 242 /// This builder has the ability to accept instructions in either 243 /// forward or reverse order, depending on the pass direction that 244 /// produces the VCode. The lowering from CLIF to VCode<MachInst> 245 /// ordinarily occurs in reverse order (in order to allow instructions 246 /// to be lowered only if used, and not merged) so a reversal will 247 /// occur at the end of lowering to ensure the VCode is in machine 248 /// order. 249 /// 250 /// If built in reverse, block and instruction indices used once the 251 /// VCode is built are relative to the final (reversed) order, not the 252 /// order of construction. Note that this means we do not know the 253 /// final block or instruction indices when building, so we do not 254 /// hand them out. (The user is assumed to know them when appending 255 /// terminator instructions with successor blocks.) 256 pub struct VCodeBuilder<I: VCodeInst> { 257 /// In-progress VCode. 258 pub(crate) vcode: VCode<I>, 259 260 /// In what direction is the build occurring? 261 direction: VCodeBuildDirection, 262 263 /// Debug-value label in-progress map, keyed by label. For each 264 /// label, we keep disjoint ranges mapping to vregs. We'll flatten 265 /// this into (vreg, range, label) tuples when done. 266 debug_info: FxHashMap<ValueLabel, Vec<(InsnIndex, InsnIndex, VReg)>>, 267 } 268 269 /// Direction in which a VCodeBuilder builds VCode. 270 #[derive(Clone, Copy, Debug, PartialEq, Eq)] 271 pub enum VCodeBuildDirection { 272 // TODO: add `Forward` once we need it and can test it adequately. 273 /// Backward-build pass: we expect the producer to call `emit()` 274 /// with instructions in reverse program order within each block. 275 Backward, 276 } 277 278 impl<I: VCodeInst> VCodeBuilder<I> { 279 /// Create a new VCodeBuilder. 280 pub fn new( 281 sigs: SigSet, 282 abi: Callee<I::ABIMachineSpec>, 283 emit_info: I::Info, 284 block_order: BlockLoweringOrder, 285 constants: VCodeConstants, 286 direction: VCodeBuildDirection, 287 log2_min_function_alignment: u8, 288 ) -> Self { 289 let vcode = VCode::new( 290 sigs, 291 abi, 292 emit_info, 293 block_order, 294 constants, 295 log2_min_function_alignment, 296 ); 297 298 VCodeBuilder { 299 vcode, 300 direction, 301 debug_info: FxHashMap::default(), 302 } 303 } 304 305 pub fn init_retval_area(&mut self, vregs: &mut VRegAllocator<I>) -> CodegenResult<()> { 306 self.vcode.abi.init_retval_area(&self.vcode.sigs, vregs) 307 } 308 309 /// Access the ABI object. 310 pub fn abi(&self) -> &Callee<I::ABIMachineSpec> { 311 &self.vcode.abi 312 } 313 314 /// Access the ABI object. 315 pub fn abi_mut(&mut self) -> &mut Callee<I::ABIMachineSpec> { 316 &mut self.vcode.abi 317 } 318 319 pub fn sigs(&self) -> &SigSet { 320 &self.vcode.sigs 321 } 322 323 pub fn sigs_mut(&mut self) -> &mut SigSet { 324 &mut self.vcode.sigs 325 } 326 327 /// Access to the BlockLoweringOrder object. 328 pub fn block_order(&self) -> &BlockLoweringOrder { 329 &self.vcode.block_order 330 } 331 332 /// Set the current block as the entry block. 333 pub fn set_entry(&mut self, block: BlockIndex) { 334 self.vcode.entry = block; 335 } 336 337 /// End the current basic block. Must be called after emitting vcode insts 338 /// for IR insts and prior to ending the function (building the VCode). 339 pub fn end_bb(&mut self) { 340 let end_idx = self.vcode.insts.len(); 341 // Add the instruction index range to the list of blocks. 342 self.vcode.block_ranges.push_end(end_idx); 343 // End the successors list. 344 let succ_end = self.vcode.block_succs.len(); 345 self.vcode.block_succ_range.push_end(succ_end); 346 // End the blockparams list. 347 let block_params_end = self.vcode.block_params.len(); 348 self.vcode.block_params_range.push_end(block_params_end); 349 // End the branch blockparam args list. 350 let branch_block_arg_succ_end = self.vcode.branch_block_arg_range.len(); 351 self.vcode 352 .branch_block_arg_succ_range 353 .push_end(branch_block_arg_succ_end); 354 } 355 356 pub fn add_block_param(&mut self, param: VirtualReg) { 357 self.vcode.block_params.push(param.into()); 358 } 359 360 fn add_branch_args_for_succ(&mut self, args: &[Reg]) { 361 self.vcode 362 .branch_block_args 363 .extend(args.iter().map(|&arg| VReg::from(arg))); 364 let end = self.vcode.branch_block_args.len(); 365 self.vcode.branch_block_arg_range.push_end(end); 366 } 367 368 /// Push an instruction for the current BB and current IR inst 369 /// within the BB. 370 pub fn push(&mut self, insn: I, loc: RelSourceLoc) { 371 assert!(!insn.is_low_level_branch()); // These are not meant to be in VCode. 372 self.vcode.insts.push(insn); 373 self.vcode.srclocs.push(loc); 374 } 375 376 /// Add a successor block with branch args. 377 pub fn add_succ(&mut self, block: BlockIndex, args: &[Reg]) { 378 self.vcode.block_succs.push(block); 379 self.add_branch_args_for_succ(args); 380 } 381 382 /// Add a debug value label to a register. 383 pub fn add_value_label(&mut self, reg: Reg, label: ValueLabel) { 384 // We'll fix up labels in reverse(). Because we're generating 385 // code bottom-to-top, the liverange of the label goes *from* 386 // the last index at which was defined (or 0, which is the end 387 // of the eventual function) *to* just this instruction, and 388 // no further. 389 let inst = InsnIndex::new(self.vcode.insts.len()); 390 let labels = self.debug_info.entry(label).or_insert_with(|| vec![]); 391 let last = labels 392 .last() 393 .map(|(_start, end, _vreg)| *end) 394 .unwrap_or(InsnIndex::new(0)); 395 labels.push((last, inst, reg.into())); 396 } 397 398 /// Access the constants. 399 pub fn constants(&mut self) -> &mut VCodeConstants { 400 &mut self.vcode.constants 401 } 402 403 fn compute_preds_from_succs(&mut self) { 404 // Do a linear-time counting sort: first determine how many 405 // times each block appears as a successor. 406 let mut starts = vec![0u32; self.vcode.num_blocks()]; 407 for succ in &self.vcode.block_succs { 408 starts[succ.index()] += 1; 409 } 410 411 // Determine for each block the starting index where that 412 // block's predecessors should go. This is equivalent to the 413 // ranges we need to store in block_pred_range. 414 self.vcode.block_pred_range.reserve(starts.len()); 415 let mut end = 0; 416 for count in starts.iter_mut() { 417 let start = end; 418 end += *count; 419 *count = start; 420 self.vcode.block_pred_range.push_end(end as usize); 421 } 422 let end = end as usize; 423 debug_assert_eq!(end, self.vcode.block_succs.len()); 424 425 // Walk over the successors again, this time grouped by 426 // predecessor, and push the predecessor at the current 427 // starting position of each of its successors. We build 428 // each group of predecessors in whatever order Ranges::iter 429 // returns them; regalloc2 doesn't care. 430 self.vcode.block_preds.resize(end, BlockIndex::invalid()); 431 for (pred, range) in self.vcode.block_succ_range.iter() { 432 let pred = BlockIndex::new(pred); 433 for succ in &self.vcode.block_succs[range] { 434 let pos = &mut starts[succ.index()]; 435 self.vcode.block_preds[*pos as usize] = pred; 436 *pos += 1; 437 } 438 } 439 debug_assert!(self.vcode.block_preds.iter().all(|pred| pred.is_valid())); 440 } 441 442 /// Called once, when a build in Backward order is complete, to 443 /// perform the overall reversal (into final forward order) and 444 /// finalize metadata accordingly. 445 fn reverse_and_finalize(&mut self, vregs: &VRegAllocator<I>) { 446 let n_insts = self.vcode.insts.len(); 447 if n_insts == 0 { 448 return; 449 } 450 451 // Reverse the per-block and per-inst sequences. 452 self.vcode.block_ranges.reverse_index(); 453 self.vcode.block_ranges.reverse_target(n_insts); 454 // block_params_range is indexed by block (and blocks were 455 // traversed in reverse) so we reverse it; but block-param 456 // sequences in the concatenated vec can remain in reverse 457 // order (it is effectively an arena of arbitrarily-placed 458 // referenced sequences). 459 self.vcode.block_params_range.reverse_index(); 460 // Likewise, we reverse block_succ_range, but the block_succ 461 // concatenated array can remain as-is. 462 self.vcode.block_succ_range.reverse_index(); 463 self.vcode.insts.reverse(); 464 self.vcode.srclocs.reverse(); 465 // Likewise, branch_block_arg_succ_range is indexed by block 466 // so must be reversed. 467 self.vcode.branch_block_arg_succ_range.reverse_index(); 468 469 // To translate an instruction index *endpoint* in reversed 470 // order to forward order, compute `n_insts - i`. 471 // 472 // Why not `n_insts - 1 - i`? That would be correct to 473 // translate an individual instruction index (for ten insts 0 474 // to 9 inclusive, inst 0 becomes 9, and inst 9 becomes 475 // 0). But for the usual inclusive-start, exclusive-end range 476 // idiom, inclusive starts become exclusive ends and 477 // vice-versa, so e.g. an (inclusive) start of 0 becomes an 478 // (exclusive) end of 10. 479 let translate = |inst: InsnIndex| InsnIndex::new(n_insts - inst.index()); 480 481 // Generate debug-value labels based on per-label maps. 482 for (label, tuples) in &self.debug_info { 483 for &(start, end, vreg) in tuples { 484 let vreg = vregs.resolve_vreg_alias(vreg); 485 let fwd_start = translate(end); 486 let fwd_end = translate(start); 487 self.vcode 488 .debug_value_labels 489 .push((vreg, fwd_start, fwd_end, label.as_u32())); 490 } 491 } 492 493 // Now sort debug value labels by VReg, as required 494 // by regalloc2. 495 self.vcode 496 .debug_value_labels 497 .sort_unstable_by_key(|(vreg, _, _, _)| *vreg); 498 } 499 500 fn collect_operands(&mut self, vregs: &VRegAllocator<I>) { 501 let allocatable = PRegSet::from(self.vcode.machine_env()); 502 for (i, insn) in self.vcode.insts.iter_mut().enumerate() { 503 // Push operands from the instruction onto the operand list. 504 // 505 // We rename through the vreg alias table as we collect 506 // the operands. This is better than a separate post-pass 507 // over operands, because it has more cache locality: 508 // operands only need to pass through L1 once. This is 509 // also better than renaming instructions' 510 // operands/registers while lowering, because here we only 511 // need to do the `match` over the instruction to visit 512 // its register fields (which is slow, branchy code) once. 513 514 let mut op_collector = 515 OperandCollector::new(&mut self.vcode.operands, allocatable, |vreg| { 516 vregs.resolve_vreg_alias(vreg) 517 }); 518 insn.get_operands(&mut op_collector); 519 let (ops, clobbers) = op_collector.finish(); 520 self.vcode.operand_ranges.push_end(ops); 521 522 if clobbers != PRegSet::default() { 523 self.vcode.clobbers.insert(InsnIndex::new(i), clobbers); 524 } 525 526 if let Some((dst, src)) = insn.is_move() { 527 // We should never see non-virtual registers present in move 528 // instructions. 529 assert!( 530 src.is_virtual(), 531 "the real register {src:?} was used as the source of a move instruction" 532 ); 533 assert!( 534 dst.to_reg().is_virtual(), 535 "the real register {:?} was used as the destination of a move instruction", 536 dst.to_reg() 537 ); 538 } 539 } 540 541 // Translate blockparam args via the vreg aliases table as well. 542 for arg in &mut self.vcode.branch_block_args { 543 let new_arg = vregs.resolve_vreg_alias(*arg); 544 trace!("operandcollector: block arg {:?} -> {:?}", arg, new_arg); 545 *arg = new_arg; 546 } 547 } 548 549 /// Build the final VCode. 550 pub fn build(mut self, mut vregs: VRegAllocator<I>) -> VCode<I> { 551 self.vcode.vreg_types = take(&mut vregs.vreg_types); 552 self.vcode.facts = take(&mut vregs.facts); 553 554 if self.direction == VCodeBuildDirection::Backward { 555 self.reverse_and_finalize(&vregs); 556 } 557 self.collect_operands(&vregs); 558 559 self.compute_preds_from_succs(); 560 self.vcode.debug_value_labels.sort_unstable(); 561 562 // At this point, nothing in the vcode should mention any 563 // VReg which has been aliased. All the appropriate rewriting 564 // should have happened above. Just to be sure, let's 565 // double-check each field which has vregs. 566 // Note: can't easily check vcode.insts, resolved in collect_operands. 567 // Operands are resolved in collect_operands. 568 vregs.debug_assert_no_vreg_aliases(self.vcode.operands.iter().map(|op| op.vreg())); 569 // Currently block params are never aliased to another vreg. 570 vregs.debug_assert_no_vreg_aliases(self.vcode.block_params.iter().copied()); 571 // Branch block args are resolved in collect_operands. 572 vregs.debug_assert_no_vreg_aliases(self.vcode.branch_block_args.iter().copied()); 573 // Debug value labels are resolved in reverse_and_finalize. 574 vregs.debug_assert_no_vreg_aliases( 575 self.vcode.debug_value_labels.iter().map(|&(vreg, ..)| vreg), 576 ); 577 // Facts are resolved eagerly during set_vreg_alias. 578 vregs.debug_assert_no_vreg_aliases( 579 self.vcode 580 .facts 581 .iter() 582 .zip(&vregs.vreg_types) 583 .enumerate() 584 .filter(|(_, (fact, _))| fact.is_some()) 585 .map(|(vreg, (_, &ty))| { 586 let (regclasses, _) = I::rc_for_type(ty).unwrap(); 587 VReg::new(vreg, regclasses[0]) 588 }), 589 ); 590 591 self.vcode 592 } 593 594 /// Add a user stack map for the associated instruction. 595 pub fn add_user_stack_map( 596 &mut self, 597 inst: BackwardsInsnIndex, 598 entries: &[ir::UserStackMapEntry], 599 ) { 600 let stack_map = ir::UserStackMap::new(entries, self.vcode.abi.sized_stackslot_offsets()); 601 let old_entry = self.vcode.user_stack_maps.insert(inst, stack_map); 602 debug_assert!(old_entry.is_none()); 603 } 604 } 605 606 const NO_INST_OFFSET: CodeOffset = u32::MAX; 607 608 impl<I: VCodeInst> VCode<I> { 609 /// New empty VCode. 610 fn new( 611 sigs: SigSet, 612 abi: Callee<I::ABIMachineSpec>, 613 emit_info: I::Info, 614 block_order: BlockLoweringOrder, 615 constants: VCodeConstants, 616 log2_min_function_alignment: u8, 617 ) -> Self { 618 let n_blocks = block_order.lowered_order().len(); 619 VCode { 620 sigs, 621 vreg_types: vec![], 622 insts: Vec::with_capacity(10 * n_blocks), 623 user_stack_maps: FxHashMap::default(), 624 operands: Vec::with_capacity(30 * n_blocks), 625 operand_ranges: Ranges::with_capacity(10 * n_blocks), 626 clobbers: FxHashMap::default(), 627 srclocs: Vec::with_capacity(10 * n_blocks), 628 entry: BlockIndex::new(0), 629 block_ranges: Ranges::with_capacity(n_blocks), 630 block_succ_range: Ranges::with_capacity(n_blocks), 631 block_succs: Vec::with_capacity(n_blocks), 632 block_pred_range: Ranges::default(), 633 block_preds: Vec::new(), 634 block_params_range: Ranges::with_capacity(n_blocks), 635 block_params: Vec::with_capacity(5 * n_blocks), 636 branch_block_args: Vec::with_capacity(10 * n_blocks), 637 branch_block_arg_range: Ranges::with_capacity(2 * n_blocks), 638 branch_block_arg_succ_range: Ranges::with_capacity(n_blocks), 639 block_order, 640 abi, 641 emit_info, 642 constants, 643 debug_value_labels: vec![], 644 facts: vec![], 645 log2_min_function_alignment, 646 } 647 } 648 649 /// Get the ABI-dependent MachineEnv for managing register allocation. 650 pub fn machine_env(&self) -> &MachineEnv { 651 self.abi.machine_env(&self.sigs) 652 } 653 654 /// Get the number of blocks. Block indices will be in the range `0 .. 655 /// (self.num_blocks() - 1)`. 656 pub fn num_blocks(&self) -> usize { 657 self.block_ranges.len() 658 } 659 660 /// The number of lowered instructions. 661 pub fn num_insts(&self) -> usize { 662 self.insts.len() 663 } 664 665 fn compute_clobbers(&self, regalloc: ®alloc2::Output) -> Vec<Writable<RealReg>> { 666 let mut clobbered = PRegSet::default(); 667 668 // All moves are included in clobbers. 669 for (_, Edit::Move { to, .. }) in ®alloc.edits { 670 if let Some(preg) = to.as_reg() { 671 clobbered.add(preg); 672 } 673 } 674 675 for (i, range) in self.operand_ranges.iter() { 676 let operands = &self.operands[range.clone()]; 677 let allocs = ®alloc.allocs[range]; 678 for (operand, alloc) in operands.iter().zip(allocs.iter()) { 679 if operand.kind() == OperandKind::Def { 680 if let Some(preg) = alloc.as_reg() { 681 clobbered.add(preg); 682 } 683 } 684 } 685 686 // Also add explicitly-clobbered registers. 687 // 688 // Skip merging this instruction's clobber list if not 689 // "included in clobbers" as per the MachInst. (Some 690 // backends use this to implement ABI specifics; e.g., 691 // excluding calls of the same ABI as the current function 692 // from clobbers, because by definition everything 693 // clobbered by the call can be clobbered by this function 694 // without saving as well. 695 // 696 // This is important for a particular optimization: when 697 // some registers are "half-clobbered", e.g. vector/float 698 // registers on aarch64, we want them to be seen as 699 // clobbered by regalloc so it avoids carrying values 700 // across calls in these registers but not seen as 701 // clobbered by prologue generation here (because the 702 // actual half-clobber implied by the clobber list fits 703 // within the clobbers that we allow without 704 // clobber-saves). 705 if self.insts[i].is_included_in_clobbers() { 706 if let Some(&inst_clobbered) = self.clobbers.get(&InsnIndex::new(i)) { 707 clobbered.union_from(inst_clobbered); 708 } 709 } 710 } 711 712 clobbered 713 .into_iter() 714 .map(|preg| Writable::from_reg(RealReg::from(preg))) 715 .collect() 716 } 717 718 /// Emit the instructions to a `MachBuffer`, containing fixed-up 719 /// code and external reloc/trap/etc. records ready for use. Takes 720 /// the regalloc results as well. 721 /// 722 /// Returns the machine code itself, and optionally metadata 723 /// and/or a disassembly, as an `EmitResult`. The `VCode` itself 724 /// is consumed by the emission process. 725 pub fn emit( 726 mut self, 727 regalloc: ®alloc2::Output, 728 want_disasm: bool, 729 flags: &settings::Flags, 730 ctrl_plane: &mut ControlPlane, 731 ) -> EmitResult 732 where 733 I: VCodeInst, 734 { 735 let _tt = timing::vcode_emit(); 736 let mut buffer = MachBuffer::new(); 737 buffer.set_log2_min_function_alignment(self.log2_min_function_alignment); 738 let mut bb_starts: Vec<Option<CodeOffset>> = vec![]; 739 740 // The first M MachLabels are reserved for block indices. 741 buffer.reserve_labels_for_blocks(self.num_blocks()); 742 743 // Register all allocated constants with the `MachBuffer` to ensure that 744 // any references to the constants during instructions can be handled 745 // correctly. 746 buffer.register_constants(&self.constants); 747 748 // Construct the final order we emit code in: cold blocks at the end. 749 let mut final_order: SmallVec<[BlockIndex; 16]> = smallvec![]; 750 let mut cold_blocks: SmallVec<[BlockIndex; 16]> = smallvec![]; 751 for block in 0..self.num_blocks() { 752 let block = BlockIndex::new(block); 753 if self.block_order.is_cold(block) { 754 cold_blocks.push(block); 755 } else { 756 final_order.push(block); 757 } 758 } 759 final_order.extend(cold_blocks.clone()); 760 761 // Compute/save info we need for the prologue: clobbers and 762 // number of spillslots. 763 // 764 // We clone `abi` here because we will mutate it as we 765 // generate the prologue and set other info, but we can't 766 // mutate `VCode`. The info it usually carries prior to 767 // setting clobbers is fairly minimal so this should be 768 // relatively cheap. 769 let clobbers = self.compute_clobbers(regalloc); 770 self.abi 771 .compute_frame_layout(&self.sigs, regalloc.num_spillslots, clobbers); 772 773 // Emit blocks. 774 let mut cur_srcloc = None; 775 let mut last_offset = None; 776 let mut inst_offsets = vec![]; 777 let mut state = I::State::new(&self.abi, std::mem::take(ctrl_plane)); 778 779 let mut disasm = String::new(); 780 781 if !self.debug_value_labels.is_empty() { 782 inst_offsets.resize(self.insts.len(), NO_INST_OFFSET); 783 } 784 785 // Count edits per block ahead of time; this is needed for 786 // lookahead island emission. (We could derive it per-block 787 // with binary search in the edit list, but it's more 788 // efficient to do it in one pass here.) 789 let mut ra_edits_per_block: SmallVec<[u32; 64]> = smallvec![]; 790 let mut edit_idx = 0; 791 for block in 0..self.num_blocks() { 792 let end_inst = InsnIndex::new(self.block_ranges.get(block).end); 793 let start_edit_idx = edit_idx; 794 while edit_idx < regalloc.edits.len() && regalloc.edits[edit_idx].0.inst() < end_inst { 795 edit_idx += 1; 796 } 797 let end_edit_idx = edit_idx; 798 ra_edits_per_block.push((end_edit_idx - start_edit_idx) as u32); 799 } 800 801 let is_forward_edge_cfi_enabled = self.abi.is_forward_edge_cfi_enabled(); 802 let mut bb_padding = match flags.bb_padding_log2_minus_one() { 803 0 => Vec::new(), 804 n => vec![0; 1 << (n - 1)], 805 }; 806 let mut total_bb_padding = 0; 807 808 for (block_order_idx, &block) in final_order.iter().enumerate() { 809 trace!("emitting block {:?}", block); 810 811 // Call the new block hook for state 812 state.on_new_block(); 813 814 // Emit NOPs to align the block. 815 let new_offset = I::align_basic_block(buffer.cur_offset()); 816 while new_offset > buffer.cur_offset() { 817 // Pad with NOPs up to the aligned block offset. 818 let nop = I::gen_nop((new_offset - buffer.cur_offset()) as usize); 819 nop.emit(&mut buffer, &self.emit_info, &mut Default::default()); 820 } 821 assert_eq!(buffer.cur_offset(), new_offset); 822 823 let do_emit = |inst: &I, 824 disasm: &mut String, 825 buffer: &mut MachBuffer<I>, 826 state: &mut I::State| { 827 if want_disasm && !inst.is_args() { 828 let mut s = state.clone(); 829 writeln!(disasm, " {}", inst.pretty_print_inst(&mut s)).unwrap(); 830 } 831 inst.emit(buffer, &self.emit_info, state); 832 }; 833 834 // Is this the first block? Emit the prologue directly if so. 835 if block == self.entry { 836 trace!(" -> entry block"); 837 buffer.start_srcloc(Default::default()); 838 for inst in &self.abi.gen_prologue() { 839 do_emit(&inst, &mut disasm, &mut buffer, &mut state); 840 } 841 buffer.end_srcloc(); 842 } 843 844 // Now emit the regular block body. 845 846 buffer.bind_label(MachLabel::from_block(block), state.ctrl_plane_mut()); 847 848 if want_disasm { 849 writeln!(&mut disasm, "block{}:", block.index()).unwrap(); 850 } 851 852 if flags.machine_code_cfg_info() { 853 // Track BB starts. If we have backed up due to MachBuffer 854 // branch opts, note that the removed blocks were removed. 855 let cur_offset = buffer.cur_offset(); 856 if last_offset.is_some() && cur_offset <= last_offset.unwrap() { 857 for i in (0..bb_starts.len()).rev() { 858 if bb_starts[i].is_some() && cur_offset > bb_starts[i].unwrap() { 859 break; 860 } 861 bb_starts[i] = None; 862 } 863 } 864 bb_starts.push(Some(cur_offset)); 865 last_offset = Some(cur_offset); 866 } 867 868 if let Some(block_start) = I::gen_block_start( 869 self.block_order.is_indirect_branch_target(block), 870 is_forward_edge_cfi_enabled, 871 ) { 872 do_emit(&block_start, &mut disasm, &mut buffer, &mut state); 873 } 874 875 for inst_or_edit in regalloc.block_insts_and_edits(&self, block) { 876 match inst_or_edit { 877 InstOrEdit::Inst(iix) => { 878 if !self.debug_value_labels.is_empty() { 879 // If we need to produce debug info, 880 // record the offset of each instruction 881 // so that we can translate value-label 882 // ranges to machine-code offsets. 883 884 // Cold blocks violate monotonicity 885 // assumptions elsewhere (that 886 // instructions in inst-index order are in 887 // order in machine code), so we omit 888 // their offsets here. Value-label range 889 // generation below will skip empty ranges 890 // and ranges with to-offsets of zero. 891 if !self.block_order.is_cold(block) { 892 inst_offsets[iix.index()] = buffer.cur_offset(); 893 } 894 } 895 896 // Update the srcloc at this point in the buffer. 897 let srcloc = self.srclocs[iix.index()]; 898 if cur_srcloc != Some(srcloc) { 899 if cur_srcloc.is_some() { 900 buffer.end_srcloc(); 901 } 902 buffer.start_srcloc(srcloc); 903 cur_srcloc = Some(srcloc); 904 } 905 906 // If this is a safepoint, compute a stack map 907 // and pass it to the emit state. 908 let stack_map_disasm = if self.insts[iix.index()].is_safepoint() { 909 let (user_stack_map, user_stack_map_disasm) = { 910 // The `user_stack_maps` is keyed by reverse 911 // instruction index, so we must flip the 912 // index. We can't put this into a helper method 913 // due to borrowck issues because parts of 914 // `self` are borrowed mutably elsewhere in this 915 // function. 916 let index = iix.to_backwards_insn_index(self.num_insts()); 917 let user_stack_map = self.user_stack_maps.remove(&index); 918 let user_stack_map_disasm = 919 user_stack_map.as_ref().map(|m| format!(" ; {m:?}")); 920 (user_stack_map, user_stack_map_disasm) 921 }; 922 923 state.pre_safepoint(user_stack_map); 924 925 user_stack_map_disasm 926 } else { 927 None 928 }; 929 930 // If the instruction we are about to emit is 931 // a return, place an epilogue at this point 932 // (and don't emit the return; the actual 933 // epilogue will contain it). 934 if self.insts[iix.index()].is_term() == MachTerminator::Ret { 935 for inst in self.abi.gen_epilogue() { 936 do_emit(&inst, &mut disasm, &mut buffer, &mut state); 937 } 938 } else { 939 // Update the operands for this inst using the 940 // allocations from the regalloc result. 941 let mut allocs = regalloc.inst_allocs(iix).iter(); 942 self.insts[iix.index()].get_operands( 943 &mut |reg: &mut Reg, constraint, _kind, _pos| { 944 let alloc = 945 allocs.next().expect("enough allocations for all operands"); 946 947 if let Some(alloc) = alloc.as_reg() { 948 let alloc: Reg = alloc.into(); 949 if let OperandConstraint::FixedReg(rreg) = constraint { 950 debug_assert_eq!(Reg::from(rreg), alloc); 951 } 952 *reg = alloc; 953 } else if let Some(alloc) = alloc.as_stack() { 954 let alloc: Reg = alloc.into(); 955 *reg = alloc; 956 } 957 }, 958 ); 959 debug_assert!(allocs.next().is_none()); 960 961 // Emit the instruction! 962 do_emit( 963 &self.insts[iix.index()], 964 &mut disasm, 965 &mut buffer, 966 &mut state, 967 ); 968 if let Some(stack_map_disasm) = stack_map_disasm { 969 disasm.push_str(&stack_map_disasm); 970 disasm.push('\n'); 971 } 972 } 973 } 974 975 InstOrEdit::Edit(Edit::Move { from, to }) => { 976 // Create a move/spill/reload instruction and 977 // immediately emit it. 978 match (from.as_reg(), to.as_reg()) { 979 (Some(from), Some(to)) => { 980 // Reg-to-reg move. 981 let from_rreg = Reg::from(from); 982 let to_rreg = Writable::from_reg(Reg::from(to)); 983 debug_assert_eq!(from.class(), to.class()); 984 let ty = I::canonical_type_for_rc(from.class()); 985 let mv = I::gen_move(to_rreg, from_rreg, ty); 986 do_emit(&mv, &mut disasm, &mut buffer, &mut state); 987 } 988 (Some(from), None) => { 989 // Spill from register to spillslot. 990 let to = to.as_stack().unwrap(); 991 let from_rreg = RealReg::from(from); 992 let spill = self.abi.gen_spill(to, from_rreg); 993 do_emit(&spill, &mut disasm, &mut buffer, &mut state); 994 } 995 (None, Some(to)) => { 996 // Load from spillslot to register. 997 let from = from.as_stack().unwrap(); 998 let to_rreg = Writable::from_reg(RealReg::from(to)); 999 let reload = self.abi.gen_reload(to_rreg, from); 1000 do_emit(&reload, &mut disasm, &mut buffer, &mut state); 1001 } 1002 (None, None) => { 1003 panic!("regalloc2 should have eliminated stack-to-stack moves!"); 1004 } 1005 } 1006 } 1007 } 1008 } 1009 1010 if cur_srcloc.is_some() { 1011 buffer.end_srcloc(); 1012 cur_srcloc = None; 1013 } 1014 1015 // Do we need an island? Get the worst-case size of the next BB, add 1016 // it to the optional padding behind the block, and pass this to the 1017 // `MachBuffer` to determine if an island is necessary. 1018 let worst_case_next_bb = if block_order_idx < final_order.len() - 1 { 1019 let next_block = final_order[block_order_idx + 1]; 1020 let next_block_range = self.block_ranges.get(next_block.index()); 1021 let next_block_size = next_block_range.len() as u32; 1022 let next_block_ra_insertions = ra_edits_per_block[next_block.index()]; 1023 I::worst_case_size() * (next_block_size + next_block_ra_insertions) 1024 } else { 1025 0 1026 }; 1027 let padding = if bb_padding.is_empty() { 1028 0 1029 } else { 1030 bb_padding.len() as u32 + I::LabelUse::ALIGN - 1 1031 }; 1032 if buffer.island_needed(padding + worst_case_next_bb) { 1033 buffer.emit_island(padding + worst_case_next_bb, ctrl_plane); 1034 } 1035 1036 // Insert padding, if configured, to stress the `MachBuffer`'s 1037 // relocation and island calculations. 1038 // 1039 // Padding can get quite large during fuzzing though so place a 1040 // total cap on it where when a per-function threshold is exceeded 1041 // the padding is turned back down to zero. This avoids a small-ish 1042 // test case generating a GB+ memory footprint in Cranelift for 1043 // example. 1044 if !bb_padding.is_empty() { 1045 buffer.put_data(&bb_padding); 1046 buffer.align_to(I::LabelUse::ALIGN); 1047 total_bb_padding += bb_padding.len(); 1048 if total_bb_padding > (150 << 20) { 1049 bb_padding = Vec::new(); 1050 } 1051 } 1052 } 1053 1054 debug_assert!( 1055 self.user_stack_maps.is_empty(), 1056 "any stack maps should have been consumed by instruction emission, still have: {:#?}", 1057 self.user_stack_maps, 1058 ); 1059 1060 // Do any optimizations on branches at tail of buffer, as if we had 1061 // bound one last label. 1062 buffer.optimize_branches(ctrl_plane); 1063 1064 // emission state is not needed anymore, move control plane back out 1065 *ctrl_plane = state.take_ctrl_plane(); 1066 1067 let func_body_len = buffer.cur_offset(); 1068 1069 // Create `bb_edges` and final (filtered) `bb_starts`. 1070 let mut bb_edges = vec![]; 1071 let mut bb_offsets = vec![]; 1072 if flags.machine_code_cfg_info() { 1073 for block in 0..self.num_blocks() { 1074 if bb_starts[block].is_none() { 1075 // Block was deleted by MachBuffer; skip. 1076 continue; 1077 } 1078 let from = bb_starts[block].unwrap(); 1079 1080 bb_offsets.push(from); 1081 // Resolve each `succ` label and add edges. 1082 let succs = self.block_succs(BlockIndex::new(block)); 1083 for &succ in succs.iter() { 1084 let to = buffer.resolve_label_offset(MachLabel::from_block(succ)); 1085 bb_edges.push((from, to)); 1086 } 1087 } 1088 } 1089 1090 self.monotonize_inst_offsets(&mut inst_offsets[..], func_body_len); 1091 let value_labels_ranges = 1092 self.compute_value_labels_ranges(regalloc, &inst_offsets[..], func_body_len); 1093 let frame_size = self.abi.frame_size(); 1094 1095 EmitResult { 1096 buffer: buffer.finish(&self.constants, ctrl_plane), 1097 bb_offsets, 1098 bb_edges, 1099 func_body_len, 1100 disasm: if want_disasm { Some(disasm) } else { None }, 1101 sized_stackslot_offsets: self.abi.sized_stackslot_offsets().clone(), 1102 dynamic_stackslot_offsets: self.abi.dynamic_stackslot_offsets().clone(), 1103 value_labels_ranges, 1104 frame_size, 1105 } 1106 } 1107 1108 fn monotonize_inst_offsets(&self, inst_offsets: &mut [CodeOffset], func_body_len: u32) { 1109 if self.debug_value_labels.is_empty() { 1110 return; 1111 } 1112 1113 // During emission, branch removal can make offsets of instructions incorrect. 1114 // Consider the following sequence: [insi][jmp0][jmp1][jmp2][insj] 1115 // It will be recorded as (say): [30] [34] [38] [42] [<would be 46>] 1116 // When the jumps get removed we are left with (in "inst_offsets"): 1117 // [insi][jmp0][jmp1][jmp2][insj][...] 1118 // [30] [34] [38] [42] [34] 1119 // Which violates the monotonicity invariant. This method sets offsets of these 1120 // removed instructions such as to make them appear zero-sized: 1121 // [insi][jmp0][jmp1][jmp2][insj][...] 1122 // [30] [34] [34] [34] [34] 1123 // 1124 let mut next_offset = func_body_len; 1125 for inst_index in (0..(inst_offsets.len() - 1)).rev() { 1126 let inst_offset = inst_offsets[inst_index]; 1127 1128 // Not all instructions get their offsets recorded. 1129 if inst_offset == NO_INST_OFFSET { 1130 continue; 1131 } 1132 1133 if inst_offset > next_offset { 1134 trace!( 1135 "Fixing code offset of the removed Inst {}: {} -> {}", 1136 inst_index, 1137 inst_offset, 1138 next_offset 1139 ); 1140 inst_offsets[inst_index] = next_offset; 1141 continue; 1142 } 1143 1144 next_offset = inst_offset; 1145 } 1146 } 1147 1148 fn compute_value_labels_ranges( 1149 &self, 1150 regalloc: ®alloc2::Output, 1151 inst_offsets: &[CodeOffset], 1152 func_body_len: u32, 1153 ) -> ValueLabelsRanges { 1154 if self.debug_value_labels.is_empty() { 1155 return ValueLabelsRanges::default(); 1156 } 1157 1158 if trace_log_enabled!() { 1159 self.log_value_labels_ranges(regalloc, inst_offsets); 1160 } 1161 1162 let mut value_labels_ranges: ValueLabelsRanges = HashMap::new(); 1163 for &(label, from, to, alloc) in ®alloc.debug_locations { 1164 let label = ValueLabel::from_u32(label); 1165 let ranges = value_labels_ranges.entry(label).or_insert_with(|| vec![]); 1166 let from_offset = inst_offsets[from.inst().index()]; 1167 let to_offset = if to.inst().index() == inst_offsets.len() { 1168 func_body_len 1169 } else { 1170 inst_offsets[to.inst().index()] 1171 }; 1172 1173 // Empty ranges or unavailable offsets can happen 1174 // due to cold blocks and branch removal (see above). 1175 if from_offset == NO_INST_OFFSET 1176 || to_offset == NO_INST_OFFSET 1177 || from_offset == to_offset 1178 { 1179 continue; 1180 } 1181 1182 let loc = if let Some(preg) = alloc.as_reg() { 1183 LabelValueLoc::Reg(Reg::from(preg)) 1184 } else { 1185 let slot = alloc.as_stack().unwrap(); 1186 let slot_offset = self.abi.get_spillslot_offset(slot); 1187 let slot_base_to_caller_sp_offset = self.abi.slot_base_to_caller_sp_offset(); 1188 let caller_sp_to_cfa_offset = 1189 crate::isa::unwind::systemv::caller_sp_to_cfa_offset(); 1190 // NOTE: this is a negative offset because it's relative to the caller's SP 1191 let cfa_to_sp_offset = 1192 -((slot_base_to_caller_sp_offset + caller_sp_to_cfa_offset) as i64); 1193 LabelValueLoc::CFAOffset(cfa_to_sp_offset + slot_offset) 1194 }; 1195 1196 // ValueLocRanges are recorded by *instruction-end 1197 // offset*. `from_offset` is the *start* of the 1198 // instruction; that is the same as the end of another 1199 // instruction, so we only want to begin coverage once 1200 // we are past the previous instruction's end. 1201 let start = from_offset + 1; 1202 1203 // Likewise, `end` is exclusive, but we want to 1204 // *include* the end of the last 1205 // instruction. `to_offset` is the start of the 1206 // `to`-instruction, which is the exclusive end, i.e., 1207 // the first instruction not covered. That 1208 // instruction's start is the same as the end of the 1209 // last instruction that is included, so we go one 1210 // byte further to be sure to include it. 1211 let end = to_offset + 1; 1212 1213 // Coalesce adjacent ranges that for the same location 1214 // to minimize output size here and for the consumers. 1215 if let Some(last_loc_range) = ranges.last_mut() { 1216 if last_loc_range.loc == loc && last_loc_range.end == start { 1217 trace!( 1218 "Extending debug range for {:?} in {:?} to {}", 1219 label, 1220 loc, 1221 end 1222 ); 1223 last_loc_range.end = end; 1224 continue; 1225 } 1226 } 1227 1228 trace!( 1229 "Recording debug range for {:?} in {:?}: [Inst {}..Inst {}) [{}..{})", 1230 label, 1231 loc, 1232 from.inst().index(), 1233 to.inst().index(), 1234 start, 1235 end 1236 ); 1237 1238 ranges.push(ValueLocRange { loc, start, end }); 1239 } 1240 1241 value_labels_ranges 1242 } 1243 1244 fn log_value_labels_ranges(&self, regalloc: ®alloc2::Output, inst_offsets: &[CodeOffset]) { 1245 debug_assert!(trace_log_enabled!()); 1246 1247 // What debug labels do we have? Note we'll skip those that have not been 1248 // allocated any location at all. They will show up as numeric gaps in the table. 1249 let mut labels = vec![]; 1250 for &(label, _, _, _) in ®alloc.debug_locations { 1251 if Some(&label) == labels.last() { 1252 continue; 1253 } 1254 labels.push(label); 1255 } 1256 1257 // Reformat the data on what VRegs were the VLs assigned to by lowering, since 1258 // the array we have is sorted by VReg, and we want it sorted by VL for easy 1259 // access in the loop below. 1260 let mut vregs = vec![]; 1261 for &(vreg, start, end, label) in &self.debug_value_labels { 1262 if matches!(labels.binary_search(&label), Ok(_)) { 1263 vregs.push((label, start, end, vreg)); 1264 } 1265 } 1266 vregs.sort_unstable_by( 1267 |(l_label, l_start, _, _), (r_label, r_start, _, _)| match l_label.cmp(r_label) { 1268 Ordering::Equal => l_start.cmp(r_start), 1269 cmp => cmp, 1270 }, 1271 ); 1272 1273 #[derive(PartialEq)] 1274 enum Mode { 1275 Measure, 1276 Emit, 1277 } 1278 #[derive(PartialEq)] 1279 enum Row { 1280 Head, 1281 Line, 1282 Inst(usize), 1283 } 1284 1285 let mut widths = vec![0; 2 + 2 * labels.len()]; 1286 let mut row = String::new(); 1287 let mut output_row = |row_kind: Row, mode: Mode| { 1288 let mut column_index = 0; 1289 row.clear(); 1290 1291 macro_rules! output_cell_impl { 1292 ($fill:literal, $span:literal, $($cell_fmt:tt)*) => { 1293 let column_start = row.len(); 1294 { 1295 row.push('|'); 1296 write!(row, $($cell_fmt)*).unwrap(); 1297 } 1298 1299 let next_column_index = column_index + $span; 1300 let expected_width: usize = widths[column_index..next_column_index].iter().sum(); 1301 if mode == Mode::Measure { 1302 let actual_width = row.len() - column_start; 1303 if actual_width > expected_width { 1304 widths[next_column_index - 1] += actual_width - expected_width; 1305 } 1306 } else { 1307 let column_end = column_start + expected_width; 1308 while row.len() != column_end { 1309 row.push($fill); 1310 } 1311 } 1312 column_index = next_column_index; 1313 }; 1314 } 1315 macro_rules! output_cell { 1316 ($($cell_fmt:tt)*) => { 1317 output_cell_impl!(' ', 1, $($cell_fmt)*); 1318 }; 1319 } 1320 1321 match row_kind { 1322 Row::Head => { 1323 output_cell!("Inst"); 1324 output_cell!("IP"); 1325 for label in &labels { 1326 output_cell_impl!(' ', 2, "{:?}", ValueLabel::from_u32(*label)); 1327 } 1328 } 1329 Row::Line => { 1330 debug_assert!(mode == Mode::Emit); 1331 output_cell_impl!('-', 1, ""); 1332 output_cell_impl!('-', 1, ""); 1333 for _ in &labels { 1334 output_cell_impl!('-', 2, ""); 1335 } 1336 } 1337 Row::Inst(inst_index) => { 1338 output_cell!("Inst {inst_index} "); 1339 output_cell!("{} ", inst_offsets[inst_index]); 1340 1341 // The ranges which we query below operate on the logic of 1342 // "IP(inst) == IP after inst", while the rows of our table 1343 // represent IPs 'before' "inst", so we need to convert "inst" 1344 // into these "IP after" coordinates. 1345 let inst_ip_index = inst_index.wrapping_sub(1); 1346 for label in &labels { 1347 // First, the VReg. 1348 use regalloc2::Inst; 1349 let vreg_cmp = |inst: usize, 1350 vreg_label: &u32, 1351 range_start: &Inst, 1352 range_end: &Inst| { 1353 match vreg_label.cmp(&label) { 1354 Ordering::Equal => { 1355 if range_end.index() <= inst { 1356 Ordering::Less 1357 } else if range_start.index() > inst { 1358 Ordering::Greater 1359 } else { 1360 Ordering::Equal 1361 } 1362 } 1363 cmp => cmp, 1364 } 1365 }; 1366 let vreg_index = 1367 vregs.binary_search_by(|(l, s, e, _)| vreg_cmp(inst_ip_index, l, s, e)); 1368 if let Ok(vreg_index) = vreg_index { 1369 let mut prev_vreg = None; 1370 if inst_ip_index > 0 { 1371 let prev_vreg_index = vregs.binary_search_by(|(l, s, e, _)| { 1372 vreg_cmp(inst_ip_index - 1, l, s, e) 1373 }); 1374 if let Ok(prev_vreg_index) = prev_vreg_index { 1375 prev_vreg = Some(vregs[prev_vreg_index].3); 1376 } 1377 } 1378 1379 let vreg = vregs[vreg_index].3; 1380 if Some(vreg) == prev_vreg { 1381 output_cell!("*"); 1382 } else { 1383 output_cell!("{}", vreg); 1384 } 1385 } else { 1386 output_cell!(""); 1387 } 1388 1389 // Second, the allocated location. 1390 let range_index = regalloc.debug_locations.binary_search_by( 1391 |(range_label, range_start, range_end, _)| match range_label.cmp(label) 1392 { 1393 Ordering::Equal => { 1394 if range_end.inst().index() <= inst_ip_index { 1395 Ordering::Less 1396 } else if range_start.inst().index() > inst_ip_index { 1397 Ordering::Greater 1398 } else { 1399 Ordering::Equal 1400 } 1401 } 1402 cmp => cmp, 1403 }, 1404 ); 1405 if let Ok(range_index) = range_index { 1406 // Live at this instruction, print the location. 1407 if let Some(reg) = regalloc.debug_locations[range_index].3.as_reg() { 1408 output_cell!("{:?}", Reg::from(reg)); 1409 } else { 1410 output_cell!("Stk"); 1411 } 1412 } else { 1413 // Not live at this instruction. 1414 output_cell!(""); 1415 } 1416 } 1417 } 1418 } 1419 row.push('|'); 1420 1421 if mode == Mode::Emit { 1422 trace!("{}", row.as_str()); 1423 } 1424 }; 1425 1426 for inst_index in 0..inst_offsets.len() { 1427 output_row(Row::Inst(inst_index), Mode::Measure); 1428 } 1429 output_row(Row::Head, Mode::Measure); 1430 1431 output_row(Row::Head, Mode::Emit); 1432 output_row(Row::Line, Mode::Emit); 1433 for inst_index in 0..inst_offsets.len() { 1434 output_row(Row::Inst(inst_index), Mode::Emit); 1435 } 1436 } 1437 1438 /// Get the IR block for a BlockIndex, if one exists. 1439 pub fn bindex_to_bb(&self, block: BlockIndex) -> Option<ir::Block> { 1440 self.block_order.lowered_order()[block.index()].orig_block() 1441 } 1442 1443 /// Get the type of a VReg. 1444 pub fn vreg_type(&self, vreg: VReg) -> Type { 1445 self.vreg_types[vreg.vreg()] 1446 } 1447 1448 /// Get the fact, if any, for a given VReg. 1449 pub fn vreg_fact(&self, vreg: VReg) -> Option<&Fact> { 1450 self.facts[vreg.vreg()].as_ref() 1451 } 1452 1453 /// Set the fact for a given VReg. 1454 pub fn set_vreg_fact(&mut self, vreg: VReg, fact: Fact) { 1455 trace!("set fact on {}: {:?}", vreg, fact); 1456 self.facts[vreg.vreg()] = Some(fact); 1457 } 1458 1459 /// Does a given instruction define any facts? 1460 pub fn inst_defines_facts(&self, inst: InsnIndex) -> bool { 1461 self.inst_operands(inst) 1462 .iter() 1463 .filter(|o| o.kind() == OperandKind::Def) 1464 .map(|o| o.vreg()) 1465 .any(|vreg| self.facts[vreg.vreg()].is_some()) 1466 } 1467 1468 /// Get the user stack map associated with the given forward instruction index. 1469 pub fn get_user_stack_map(&self, inst: InsnIndex) -> Option<&ir::UserStackMap> { 1470 let index = inst.to_backwards_insn_index(self.num_insts()); 1471 self.user_stack_maps.get(&index) 1472 } 1473 } 1474 1475 impl<I: VCodeInst> std::ops::Index<InsnIndex> for VCode<I> { 1476 type Output = I; 1477 fn index(&self, idx: InsnIndex) -> &Self::Output { 1478 &self.insts[idx.index()] 1479 } 1480 } 1481 1482 impl<I: VCodeInst> RegallocFunction for VCode<I> { 1483 fn num_insts(&self) -> usize { 1484 self.insts.len() 1485 } 1486 1487 fn num_blocks(&self) -> usize { 1488 self.block_ranges.len() 1489 } 1490 1491 fn entry_block(&self) -> BlockIndex { 1492 self.entry 1493 } 1494 1495 fn block_insns(&self, block: BlockIndex) -> InstRange { 1496 let range = self.block_ranges.get(block.index()); 1497 InstRange::new(InsnIndex::new(range.start), InsnIndex::new(range.end)) 1498 } 1499 1500 fn block_succs(&self, block: BlockIndex) -> &[BlockIndex] { 1501 let range = self.block_succ_range.get(block.index()); 1502 &self.block_succs[range] 1503 } 1504 1505 fn block_preds(&self, block: BlockIndex) -> &[BlockIndex] { 1506 let range = self.block_pred_range.get(block.index()); 1507 &self.block_preds[range] 1508 } 1509 1510 fn block_params(&self, block: BlockIndex) -> &[VReg] { 1511 // As a special case we don't return block params for the entry block, as all the arguments 1512 // will be defined by the `Inst::Args` instruction. 1513 if block == self.entry { 1514 return &[]; 1515 } 1516 1517 let range = self.block_params_range.get(block.index()); 1518 &self.block_params[range] 1519 } 1520 1521 fn branch_blockparams(&self, block: BlockIndex, _insn: InsnIndex, succ_idx: usize) -> &[VReg] { 1522 let succ_range = self.branch_block_arg_succ_range.get(block.index()); 1523 debug_assert!(succ_idx < succ_range.len()); 1524 let branch_block_args = self.branch_block_arg_range.get(succ_range.start + succ_idx); 1525 &self.branch_block_args[branch_block_args] 1526 } 1527 1528 fn is_ret(&self, insn: InsnIndex) -> bool { 1529 match self.insts[insn.index()].is_term() { 1530 // We treat blocks terminated by an unconditional trap like a return for regalloc. 1531 MachTerminator::None => self.insts[insn.index()].is_trap(), 1532 MachTerminator::Ret | MachTerminator::RetCall => true, 1533 MachTerminator::Branch => false, 1534 } 1535 } 1536 1537 fn is_branch(&self, insn: InsnIndex) -> bool { 1538 match self.insts[insn.index()].is_term() { 1539 MachTerminator::Branch => true, 1540 _ => false, 1541 } 1542 } 1543 1544 fn inst_operands(&self, insn: InsnIndex) -> &[Operand] { 1545 let range = self.operand_ranges.get(insn.index()); 1546 &self.operands[range] 1547 } 1548 1549 fn inst_clobbers(&self, insn: InsnIndex) -> PRegSet { 1550 self.clobbers.get(&insn).cloned().unwrap_or_default() 1551 } 1552 1553 fn num_vregs(&self) -> usize { 1554 self.vreg_types.len() 1555 } 1556 1557 fn debug_value_labels(&self) -> &[(VReg, InsnIndex, InsnIndex, u32)] { 1558 &self.debug_value_labels 1559 } 1560 1561 fn spillslot_size(&self, regclass: RegClass) -> usize { 1562 self.abi.get_spillslot_size(regclass) as usize 1563 } 1564 1565 fn allow_multiple_vreg_defs(&self) -> bool { 1566 // At least the s390x backend requires this, because the 1567 // `Loop` pseudo-instruction aggregates all Operands so pinned 1568 // vregs (RealRegs) may occur more than once. 1569 true 1570 } 1571 } 1572 1573 impl<I: VCodeInst> Debug for VRegAllocator<I> { 1574 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { 1575 writeln!(f, "VRegAllocator {{")?; 1576 1577 let mut alias_keys = self.vreg_aliases.keys().cloned().collect::<Vec<_>>(); 1578 alias_keys.sort_unstable(); 1579 for key in alias_keys { 1580 let dest = self.vreg_aliases.get(&key).unwrap(); 1581 writeln!(f, " {:?} := {:?}", Reg::from(key), Reg::from(*dest))?; 1582 } 1583 1584 for (vreg, fact) in self.facts.iter().enumerate() { 1585 if let Some(fact) = fact { 1586 writeln!(f, " v{vreg} ! {fact}")?; 1587 } 1588 } 1589 1590 writeln!(f, "}}") 1591 } 1592 } 1593 1594 impl<I: VCodeInst> fmt::Debug for VCode<I> { 1595 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { 1596 writeln!(f, "VCode {{")?; 1597 writeln!(f, " Entry block: {}", self.entry.index())?; 1598 1599 let mut state = Default::default(); 1600 1601 for block in 0..self.num_blocks() { 1602 let block = BlockIndex::new(block); 1603 writeln!( 1604 f, 1605 "Block {}({:?}):", 1606 block.index(), 1607 self.block_params(block) 1608 )?; 1609 if let Some(bb) = self.bindex_to_bb(block) { 1610 writeln!(f, " (original IR block: {bb})")?; 1611 } 1612 for (succ_idx, succ) in self.block_succs(block).iter().enumerate() { 1613 writeln!( 1614 f, 1615 " (successor: Block {}({:?}))", 1616 succ.index(), 1617 self.branch_blockparams(block, InsnIndex::new(0) /* dummy */, succ_idx) 1618 )?; 1619 } 1620 for inst in self.block_ranges.get(block.index()) { 1621 writeln!( 1622 f, 1623 " Inst {}: {}", 1624 inst, 1625 self.insts[inst].pretty_print_inst(&mut state) 1626 )?; 1627 if !self.operands.is_empty() { 1628 for operand in self.inst_operands(InsnIndex::new(inst)) { 1629 if operand.kind() == OperandKind::Def { 1630 if let Some(fact) = &self.facts[operand.vreg().vreg()] { 1631 writeln!(f, " v{} ! {}", operand.vreg().vreg(), fact)?; 1632 } 1633 } 1634 } 1635 } 1636 if let Some(user_stack_map) = self.get_user_stack_map(InsnIndex::new(inst)) { 1637 writeln!(f, " {user_stack_map:?}")?; 1638 } 1639 } 1640 } 1641 1642 writeln!(f, "}}")?; 1643 Ok(()) 1644 } 1645 } 1646 1647 /// This structure manages VReg allocation during the lifetime of the VCodeBuilder. 1648 pub struct VRegAllocator<I> { 1649 /// VReg IR-level types. 1650 vreg_types: Vec<Type>, 1651 1652 /// VReg aliases. When the final VCode is built we rewrite all 1653 /// uses of the keys in this table to their replacement values. 1654 /// 1655 /// We use these aliases to rename an instruction's expected 1656 /// result vregs to the returned vregs from lowering, which are 1657 /// usually freshly-allocated temps. 1658 vreg_aliases: FxHashMap<regalloc2::VReg, regalloc2::VReg>, 1659 1660 /// A deferred error, to be bubbled up to the top level of the 1661 /// lowering algorithm. We take this approach because we cannot 1662 /// currently propagate a `Result` upward through ISLE code (the 1663 /// lowering rules) or some ABI code. 1664 deferred_error: Option<CodegenError>, 1665 1666 /// Facts on VRegs, for proof-carrying code. 1667 facts: Vec<Option<Fact>>, 1668 1669 /// The type of instruction that this allocator makes registers for. 1670 _inst: core::marker::PhantomData<I>, 1671 } 1672 1673 impl<I: VCodeInst> VRegAllocator<I> { 1674 /// Make a new VRegAllocator. 1675 pub fn with_capacity(capacity: usize) -> Self { 1676 let capacity = first_user_vreg_index() + capacity; 1677 let mut vreg_types = Vec::with_capacity(capacity); 1678 vreg_types.resize(first_user_vreg_index(), types::INVALID); 1679 Self { 1680 vreg_types, 1681 vreg_aliases: FxHashMap::with_capacity_and_hasher(capacity, Default::default()), 1682 deferred_error: None, 1683 facts: Vec::with_capacity(capacity), 1684 _inst: core::marker::PhantomData::default(), 1685 } 1686 } 1687 1688 /// Allocate a fresh ValueRegs. 1689 pub fn alloc(&mut self, ty: Type) -> CodegenResult<ValueRegs<Reg>> { 1690 if self.deferred_error.is_some() { 1691 return Err(CodegenError::CodeTooLarge); 1692 } 1693 let v = self.vreg_types.len(); 1694 let (regclasses, tys) = I::rc_for_type(ty)?; 1695 if v + regclasses.len() >= VReg::MAX { 1696 return Err(CodegenError::CodeTooLarge); 1697 } 1698 1699 let regs: ValueRegs<Reg> = match regclasses { 1700 &[rc0] => ValueRegs::one(VReg::new(v, rc0).into()), 1701 &[rc0, rc1] => ValueRegs::two(VReg::new(v, rc0).into(), VReg::new(v + 1, rc1).into()), 1702 // We can extend this if/when we support 32-bit targets; e.g., 1703 // an i128 on a 32-bit machine will need up to four machine regs 1704 // for a `Value`. 1705 _ => panic!("Value must reside in 1 or 2 registers"), 1706 }; 1707 for (®_ty, ®) in tys.iter().zip(regs.regs().iter()) { 1708 let vreg = reg.to_virtual_reg().unwrap(); 1709 debug_assert_eq!(self.vreg_types.len(), vreg.index()); 1710 self.vreg_types.push(reg_ty); 1711 } 1712 1713 // Create empty facts for each allocated vreg. 1714 self.facts.resize(self.vreg_types.len(), None); 1715 1716 Ok(regs) 1717 } 1718 1719 /// Allocate a fresh ValueRegs, deferring any out-of-vregs 1720 /// errors. This is useful in places where we cannot bubble a 1721 /// `CodegenResult` upward easily, and which are known to be 1722 /// invoked from within the lowering loop that checks the deferred 1723 /// error status below. 1724 pub fn alloc_with_deferred_error(&mut self, ty: Type) -> ValueRegs<Reg> { 1725 match self.alloc(ty) { 1726 Ok(x) => x, 1727 Err(e) => { 1728 self.deferred_error = Some(e); 1729 self.bogus_for_deferred_error(ty) 1730 } 1731 } 1732 } 1733 1734 /// Take any deferred error that was accumulated by `alloc_with_deferred_error`. 1735 pub fn take_deferred_error(&mut self) -> Option<CodegenError> { 1736 self.deferred_error.take() 1737 } 1738 1739 /// Produce an bogus VReg placeholder with the proper number of 1740 /// registers for the given type. This is meant to be used with 1741 /// deferred allocation errors (see `Lower::alloc_tmp()`). 1742 fn bogus_for_deferred_error(&self, ty: Type) -> ValueRegs<Reg> { 1743 let (regclasses, _tys) = I::rc_for_type(ty).expect("must have valid type"); 1744 match regclasses { 1745 &[rc0] => ValueRegs::one(VReg::new(0, rc0).into()), 1746 &[rc0, rc1] => ValueRegs::two(VReg::new(0, rc0).into(), VReg::new(1, rc1).into()), 1747 _ => panic!("Value must reside in 1 or 2 registers"), 1748 } 1749 } 1750 1751 /// Rewrite any mention of `from` into `to`. 1752 pub fn set_vreg_alias(&mut self, from: Reg, to: Reg) { 1753 let from = from.into(); 1754 let resolved_to = self.resolve_vreg_alias(to.into()); 1755 // Disallow cycles (see below). 1756 assert_ne!(resolved_to, from); 1757 1758 // Maintain the invariant that PCC facts only exist on vregs 1759 // which aren't aliases. We want to preserve whatever was 1760 // stated about the vreg before its producer was lowered. 1761 if let Some(fact) = self.facts[from.vreg()].take() { 1762 self.set_fact(resolved_to, fact); 1763 } 1764 1765 let old_alias = self.vreg_aliases.insert(from, resolved_to); 1766 debug_assert_eq!(old_alias, None); 1767 } 1768 1769 fn resolve_vreg_alias(&self, mut vreg: regalloc2::VReg) -> regalloc2::VReg { 1770 // We prevent cycles from existing by resolving targets of 1771 // aliases eagerly before setting them. If the target resolves 1772 // to the origin of the alias, then a cycle would be created 1773 // and the alias is disallowed. Because of the structure of 1774 // SSA code (one instruction can refer to another's defs but 1775 // not vice-versa, except indirectly through 1776 // phis/blockparams), cycles should not occur as we use 1777 // aliases to redirect vregs to the temps that actually define 1778 // them. 1779 while let Some(to) = self.vreg_aliases.get(&vreg) { 1780 vreg = *to; 1781 } 1782 vreg 1783 } 1784 1785 #[inline] 1786 fn debug_assert_no_vreg_aliases(&self, mut list: impl Iterator<Item = VReg>) { 1787 debug_assert!(list.all(|vreg| !self.vreg_aliases.contains_key(&vreg))); 1788 } 1789 1790 /// Set the proof-carrying code fact on a given virtual register. 1791 /// 1792 /// Returns the old fact, if any (only one fact can be stored). 1793 fn set_fact(&mut self, vreg: regalloc2::VReg, fact: Fact) -> Option<Fact> { 1794 trace!("vreg {:?} has fact: {:?}", vreg, fact); 1795 debug_assert!(!self.vreg_aliases.contains_key(&vreg)); 1796 self.facts[vreg.vreg()].replace(fact) 1797 } 1798 1799 /// Set a fact only if one doesn't already exist. 1800 pub fn set_fact_if_missing(&mut self, vreg: VirtualReg, fact: Fact) { 1801 let vreg = self.resolve_vreg_alias(vreg.into()); 1802 if self.facts[vreg.vreg()].is_none() { 1803 self.set_fact(vreg, fact); 1804 } 1805 } 1806 1807 /// Allocate a fresh ValueRegs, with a given fact to apply if 1808 /// the value fits in one VReg. 1809 pub fn alloc_with_maybe_fact( 1810 &mut self, 1811 ty: Type, 1812 fact: Option<Fact>, 1813 ) -> CodegenResult<ValueRegs<Reg>> { 1814 let result = self.alloc(ty)?; 1815 1816 // Ensure that we don't lose a fact on a value that splits 1817 // into multiple VRegs. 1818 assert!(result.len() == 1 || fact.is_none()); 1819 if let Some(fact) = fact { 1820 self.set_fact(result.regs()[0].into(), fact); 1821 } 1822 1823 Ok(result) 1824 } 1825 } 1826 1827 /// This structure tracks the large constants used in VCode that will be emitted separately by the 1828 /// [MachBuffer]. 1829 /// 1830 /// First, during the lowering phase, constants are inserted using 1831 /// [VCodeConstants.insert]; an intermediate handle, `VCodeConstant`, tracks what constants are 1832 /// used in this phase. Some deduplication is performed, when possible, as constant 1833 /// values are inserted. 1834 /// 1835 /// Secondly, during the emission phase, the [MachBuffer] assigns [MachLabel]s for each of the 1836 /// constants so that instructions can refer to the value's memory location. The [MachBuffer] 1837 /// then writes the constant values to the buffer. 1838 #[derive(Default)] 1839 pub struct VCodeConstants { 1840 constants: PrimaryMap<VCodeConstant, VCodeConstantData>, 1841 pool_uses: HashMap<Constant, VCodeConstant>, 1842 well_known_uses: HashMap<*const [u8], VCodeConstant>, 1843 u64s: HashMap<[u8; 8], VCodeConstant>, 1844 } 1845 impl VCodeConstants { 1846 /// Initialize the structure with the expected number of constants. 1847 pub fn with_capacity(expected_num_constants: usize) -> Self { 1848 Self { 1849 constants: PrimaryMap::with_capacity(expected_num_constants), 1850 pool_uses: HashMap::with_capacity(expected_num_constants), 1851 well_known_uses: HashMap::new(), 1852 u64s: HashMap::new(), 1853 } 1854 } 1855 1856 /// Insert a constant; using this method indicates that a constant value will be used and thus 1857 /// will be emitted to the `MachBuffer`. The current implementation can deduplicate constants 1858 /// that are [VCodeConstantData::Pool] or [VCodeConstantData::WellKnown] but not 1859 /// [VCodeConstantData::Generated]. 1860 pub fn insert(&mut self, data: VCodeConstantData) -> VCodeConstant { 1861 match data { 1862 VCodeConstantData::Generated(_) => self.constants.push(data), 1863 VCodeConstantData::Pool(constant, _) => match self.pool_uses.get(&constant) { 1864 None => { 1865 let vcode_constant = self.constants.push(data); 1866 self.pool_uses.insert(constant, vcode_constant); 1867 vcode_constant 1868 } 1869 Some(&vcode_constant) => vcode_constant, 1870 }, 1871 VCodeConstantData::WellKnown(data_ref) => { 1872 match self.well_known_uses.entry(data_ref as *const [u8]) { 1873 Entry::Vacant(v) => { 1874 let vcode_constant = self.constants.push(data); 1875 v.insert(vcode_constant); 1876 vcode_constant 1877 } 1878 Entry::Occupied(o) => *o.get(), 1879 } 1880 } 1881 VCodeConstantData::U64(value) => match self.u64s.entry(value) { 1882 Entry::Vacant(v) => { 1883 let vcode_constant = self.constants.push(data); 1884 v.insert(vcode_constant); 1885 vcode_constant 1886 } 1887 Entry::Occupied(o) => *o.get(), 1888 }, 1889 } 1890 } 1891 1892 /// Return the number of constants inserted. 1893 pub fn len(&self) -> usize { 1894 self.constants.len() 1895 } 1896 1897 /// Iterate over the `VCodeConstant` keys inserted in this structure. 1898 pub fn keys(&self) -> Keys<VCodeConstant> { 1899 self.constants.keys() 1900 } 1901 1902 /// Iterate over the `VCodeConstant` keys and the data (as a byte slice) inserted in this 1903 /// structure. 1904 pub fn iter(&self) -> impl Iterator<Item = (VCodeConstant, &VCodeConstantData)> { 1905 self.constants.iter() 1906 } 1907 1908 /// Returns the data associated with the specified constant. 1909 pub fn get(&self, c: VCodeConstant) -> &VCodeConstantData { 1910 &self.constants[c] 1911 } 1912 1913 /// Checks if the given [VCodeConstantData] is registered as 1914 /// used by the pool. 1915 pub fn pool_uses(&self, constant: &VCodeConstantData) -> bool { 1916 match constant { 1917 VCodeConstantData::Pool(c, _) => self.pool_uses.contains_key(c), 1918 _ => false, 1919 } 1920 } 1921 } 1922 1923 /// A use of a constant by one or more VCode instructions; see [VCodeConstants]. 1924 #[derive(Clone, Copy, Debug, PartialEq, Eq)] 1925 pub struct VCodeConstant(u32); 1926 entity_impl!(VCodeConstant); 1927 1928 /// Identify the different types of constant that can be inserted into [VCodeConstants]. Tracking 1929 /// these separately instead of as raw byte buffers allows us to avoid some duplication. 1930 pub enum VCodeConstantData { 1931 /// A constant already present in the Cranelift IR 1932 /// [ConstantPool](crate::ir::constant::ConstantPool). 1933 Pool(Constant, ConstantData), 1934 /// A reference to a well-known constant value that is statically encoded within the compiler. 1935 WellKnown(&'static [u8]), 1936 /// A constant value generated during lowering; the value may depend on the instruction context 1937 /// which makes it difficult to de-duplicate--if possible, use other variants. 1938 Generated(ConstantData), 1939 /// A constant of at most 64 bits. These are deduplicated as 1940 /// well. Stored as a fixed-size array of `u8` so that we do not 1941 /// encounter endianness problems when cross-compiling. 1942 U64([u8; 8]), 1943 } 1944 impl VCodeConstantData { 1945 /// Retrieve the constant data as a byte slice. 1946 pub fn as_slice(&self) -> &[u8] { 1947 match self { 1948 VCodeConstantData::Pool(_, d) | VCodeConstantData::Generated(d) => d.as_slice(), 1949 VCodeConstantData::WellKnown(d) => d, 1950 VCodeConstantData::U64(value) => &value[..], 1951 } 1952 } 1953 1954 /// Calculate the alignment of the constant data. 1955 pub fn alignment(&self) -> u32 { 1956 if self.as_slice().len() <= 8 { 1957 8 1958 } else { 1959 16 1960 } 1961 } 1962 } 1963 1964 #[cfg(test)] 1965 mod test { 1966 use super::*; 1967 use std::mem::size_of; 1968 1969 #[test] 1970 fn size_of_constant_structs() { 1971 assert_eq!(size_of::<Constant>(), 4); 1972 assert_eq!(size_of::<VCodeConstant>(), 4); 1973 assert_eq!(size_of::<ConstantData>(), 3 * size_of::<usize>()); 1974 assert_eq!(size_of::<VCodeConstantData>(), 4 * size_of::<usize>()); 1975 assert_eq!( 1976 size_of::<PrimaryMap<VCodeConstant, VCodeConstantData>>(), 1977 3 * size_of::<usize>() 1978 ); 1979 // TODO The VCodeConstants structure's memory size could be further optimized. 1980 // With certain versions of Rust, each `HashMap` in `VCodeConstants` occupied at 1981 // least 48 bytes, making an empty `VCodeConstants` cost 120 bytes. 1982 } 1983 } 1984