1 //! In-memory representation of compiled machine code, with labels and fixups to 2 //! refer to those labels. Handles constant-pool island insertion and also 3 //! veneer insertion for out-of-range jumps. 4 //! 5 //! This code exists to solve three problems: 6 //! 7 //! - Branch targets for forward branches are not known until later, when we 8 //! emit code in a single pass through the instruction structs. 9 //! 10 //! - On many architectures, address references or offsets have limited range. 11 //! For example, on AArch64, conditional branches can only target code +/- 1MB 12 //! from the branch itself. 13 //! 14 //! - The lowering of control flow from the CFG-with-edges produced by 15 //! [BlockLoweringOrder](super::BlockLoweringOrder), combined with many empty 16 //! edge blocks when the register allocator does not need to insert any 17 //! spills/reloads/moves in edge blocks, results in many suboptimal branch 18 //! patterns. The lowering also pays no attention to block order, and so 19 //! two-target conditional forms (cond-br followed by uncond-br) can often by 20 //! avoided because one of the targets is the fallthrough. There are several 21 //! cases here where we can simplify to use fewer branches. 22 //! 23 //! This "buffer" implements a single-pass code emission strategy (with a later 24 //! "fixup" pass, but only through recorded fixups, not all instructions). The 25 //! basic idea is: 26 //! 27 //! - Emit branches as they are, including two-target (cond/uncond) compound 28 //! forms, but with zero offsets and optimistically assuming the target will be 29 //! in range. Record the "fixup" for later. Targets are denoted instead by 30 //! symbolic "labels" that are then bound to certain offsets in the buffer as 31 //! we emit code. (Nominally, there is a label at the start of every basic 32 //! block.) 33 //! 34 //! - As we do this, track the offset in the buffer at which the first label 35 //! reference "goes out of range". We call this the "deadline". If we reach the 36 //! deadline and we still have not bound the label to which an unresolved branch 37 //! refers, we have a problem! 38 //! 39 //! - To solve this problem, we emit "islands" full of "veneers". An island is 40 //! simply a chunk of code inserted in the middle of the code actually produced 41 //! by the emitter (e.g., vcode iterating over instruction structs). The emitter 42 //! has some awareness of this: it either asks for an island between blocks, so 43 //! it is not accidentally executed, or else it emits a branch around the island 44 //! when all other options fail (see `Inst::EmitIsland` meta-instruction). 45 //! 46 //! - A "veneer" is an instruction (or sequence of instructions) in an "island" 47 //! that implements a longer-range reference to a label. The idea is that, for 48 //! example, a branch with a limited range can branch to a "veneer" instead, 49 //! which is simply a branch in a form that can use a longer-range reference. On 50 //! AArch64, for example, conditionals have a +/- 1 MB range, but a conditional 51 //! can branch to an unconditional branch which has a +/- 128 MB range. Hence, a 52 //! conditional branch's label reference can be fixed up with a "veneer" to 53 //! achieve a longer range. 54 //! 55 //! - To implement all of this, we require the backend to provide a `LabelUse` 56 //! type that implements a trait. This is nominally an enum that records one of 57 //! several kinds of references to an offset in code -- basically, a relocation 58 //! type -- and will usually correspond to different instruction formats. The 59 //! `LabelUse` implementation specifies the maximum range, how to patch in the 60 //! actual label location when known, and how to generate a veneer to extend the 61 //! range. 62 //! 63 //! That satisfies label references, but we still may have suboptimal branch 64 //! patterns. To clean up the branches, we do a simple "peephole"-style 65 //! optimization on the fly. To do so, the emitter (e.g., `Inst::emit()`) 66 //! informs the buffer of branches in the code and, in the case of conditionals, 67 //! the code that would have been emitted to invert this branch's condition. We 68 //! track the "latest branches": these are branches that are contiguous up to 69 //! the current offset. (If any code is emitted after a branch, that branch or 70 //! run of contiguous branches is no longer "latest".) The latest branches are 71 //! those that we can edit by simply truncating the buffer and doing something 72 //! else instead. 73 //! 74 //! To optimize branches, we implement several simple rules, and try to apply 75 //! them to the "latest branches" when possible: 76 //! 77 //! - A branch with a label target, when that label is bound to the ending 78 //! offset of the branch (the fallthrough location), can be removed altogether, 79 //! because the branch would have no effect). 80 //! 81 //! - An unconditional branch that starts at a label location, and branches to 82 //! another label, results in a "label alias": all references to the label bound 83 //! *to* this branch instruction are instead resolved to the *target* of the 84 //! branch instruction. This effectively removes empty blocks that just 85 //! unconditionally branch to the next block. We call this "branch threading". 86 //! 87 //! - A conditional followed by an unconditional, when the conditional branches 88 //! to the unconditional's fallthrough, results in (i) the truncation of the 89 //! unconditional, (ii) the inversion of the condition's condition, and (iii) 90 //! replacement of the conditional's target (using the original target of the 91 //! unconditional). This is a fancy way of saying "we can flip a two-target 92 //! conditional branch's taken/not-taken targets if it works better with our 93 //! fallthrough". To make this work, the emitter actually gives the buffer 94 //! *both* forms of every conditional branch: the true form is emitted into the 95 //! buffer, and the "inverted" machine-code bytes are provided as part of the 96 //! branch-fixup metadata. 97 //! 98 //! - An unconditional B preceded by another unconditional P, when B's label(s) have 99 //! been redirected to target(B), can be removed entirely. This is an extension 100 //! of the branch-threading optimization, and is valid because if we know there 101 //! will be no fallthrough into this branch instruction (the prior instruction 102 //! is an unconditional jump), and if we know we have successfully redirected 103 //! all labels, then this branch instruction is unreachable. Note that this 104 //! works because the redirection happens before the label is ever resolved 105 //! (fixups happen at island emission time, at which point latest-branches are 106 //! cleared, or at the end of emission), so we are sure to catch and redirect 107 //! all possible paths to this instruction. 108 //! 109 //! # Branch-optimization Correctness 110 //! 111 //! The branch-optimization mechanism depends on a few data structures with 112 //! invariants, which are always held outside the scope of top-level public 113 //! methods: 114 //! 115 //! - The latest-branches list. Each entry describes a span of the buffer 116 //! (start/end offsets), the label target, the corresponding fixup-list entry 117 //! index, and the bytes (must be the same length) for the inverted form, if 118 //! conditional. The list of labels that are bound to the start-offset of this 119 //! branch is *complete* (if any label has a resolved offset equal to `start` 120 //! and is not an alias, it must appear in this list) and *precise* (no label 121 //! in this list can be bound to another offset). No label in this list should 122 //! be an alias. No two branch ranges can overlap, and branches are in 123 //! ascending-offset order. 124 //! 125 //! - The labels-at-tail list. This contains all MachLabels that have been bound 126 //! to (whose resolved offsets are equal to) the tail offset of the buffer. 127 //! No label in this list should be an alias. 128 //! 129 //! - The label_offsets array, containing the bound offset of a label or 130 //! UNKNOWN. No label can be bound at an offset greater than the current 131 //! buffer tail. 132 //! 133 //! - The label_aliases array, containing another label to which a label is 134 //! bound or UNKNOWN. A label's resolved offset is the resolved offset 135 //! of the label it is aliased to, if this is set. 136 //! 137 //! We argue below, at each method, how the invariants in these data structures 138 //! are maintained (grep for "Post-invariant"). 139 //! 140 //! Given these invariants, we argue why each optimization preserves execution 141 //! semantics below (grep for "Preserves execution semantics"). 142 143 use crate::binemit::{Addend, CodeOffset, Reloc, StackMap}; 144 use crate::ir::function::FunctionParameters; 145 use crate::ir::{ExternalName, Opcode, RelSourceLoc, SourceLoc, TrapCode}; 146 use crate::isa::unwind::UnwindInst; 147 use crate::machinst::{ 148 BlockIndex, MachInstLabelUse, TextSectionBuilder, VCodeConstant, VCodeConstants, VCodeInst, 149 }; 150 use crate::timing; 151 use crate::trace; 152 use cranelift_entity::{entity_impl, SecondaryMap}; 153 use smallvec::SmallVec; 154 use std::convert::TryFrom; 155 use std::mem; 156 use std::string::String; 157 use std::vec::Vec; 158 159 #[cfg(feature = "enable-serde")] 160 use serde::{Deserialize, Serialize}; 161 162 #[cfg(feature = "enable-serde")] 163 pub trait CompilePhase { 164 type MachSrcLocType: for<'a> Deserialize<'a> + Serialize + core::fmt::Debug + PartialEq + Clone; 165 type SourceLocType: for<'a> Deserialize<'a> + Serialize + core::fmt::Debug + PartialEq + Clone; 166 } 167 168 #[cfg(not(feature = "enable-serde"))] 169 pub trait CompilePhase { 170 type MachSrcLocType: core::fmt::Debug + PartialEq + Clone; 171 type SourceLocType: core::fmt::Debug + PartialEq + Clone; 172 } 173 174 /// Status of a compiled artifact that needs patching before being used. 175 /// 176 /// Only used internally. 177 #[derive(Clone, Debug, PartialEq)] 178 #[cfg_attr(feature = "enable-serde", derive(Serialize, Deserialize))] 179 pub struct Stencil; 180 181 /// Status of a compiled artifact ready to use. 182 #[derive(Clone, Debug, PartialEq)] 183 pub struct Final; 184 185 impl CompilePhase for Stencil { 186 type MachSrcLocType = MachSrcLoc<Stencil>; 187 type SourceLocType = RelSourceLoc; 188 } 189 190 impl CompilePhase for Final { 191 type MachSrcLocType = MachSrcLoc<Final>; 192 type SourceLocType = SourceLoc; 193 } 194 195 /// A buffer of output to be produced, fixed up, and then emitted to a CodeSink 196 /// in bulk. 197 /// 198 /// This struct uses `SmallVec`s to support small-ish function bodies without 199 /// any heap allocation. As such, it will be several kilobytes large. This is 200 /// likely fine as long as it is stack-allocated for function emission then 201 /// thrown away; but beware if many buffer objects are retained persistently. 202 pub struct MachBuffer<I: VCodeInst> { 203 /// The buffer contents, as raw bytes. 204 data: SmallVec<[u8; 1024]>, 205 /// Any relocations referring to this code. Note that only *external* 206 /// relocations are tracked here; references to labels within the buffer are 207 /// resolved before emission. 208 relocs: SmallVec<[MachReloc; 16]>, 209 /// Any trap records referring to this code. 210 traps: SmallVec<[MachTrap; 16]>, 211 /// Any call site records referring to this code. 212 call_sites: SmallVec<[MachCallSite; 16]>, 213 /// Any source location mappings referring to this code. 214 srclocs: SmallVec<[MachSrcLoc<Stencil>; 64]>, 215 /// Any stack maps referring to this code. 216 stack_maps: SmallVec<[MachStackMap; 8]>, 217 /// Any unwind info at a given location. 218 unwind_info: SmallVec<[(CodeOffset, UnwindInst); 8]>, 219 /// The current source location in progress (after `start_srcloc()` and 220 /// before `end_srcloc()`). This is a (start_offset, src_loc) tuple. 221 cur_srcloc: Option<(CodeOffset, RelSourceLoc)>, 222 /// Known label offsets; `UNKNOWN_LABEL_OFFSET` if unknown. 223 label_offsets: SmallVec<[CodeOffset; 16]>, 224 /// Label aliases: when one label points to an unconditional jump, and that 225 /// jump points to another label, we can redirect references to the first 226 /// label immediately to the second. 227 /// 228 /// Invariant: we don't have label-alias cycles. We ensure this by, 229 /// before setting label A to alias label B, resolving B's alias 230 /// target (iteratively until a non-aliased label); if B is already 231 /// aliased to A, then we cannot alias A back to B. 232 label_aliases: SmallVec<[MachLabel; 16]>, 233 /// Constants that must be emitted at some point. 234 pending_constants: SmallVec<[MachLabelConstant; 16]>, 235 /// Fixups that must be performed after all code is emitted. 236 fixup_records: SmallVec<[MachLabelFixup<I>; 16]>, 237 /// Current deadline at which all constants are flushed and all code labels 238 /// are extended by emitting long-range jumps in an island. This flush 239 /// should be rare (e.g., on AArch64, the shortest-range PC-rel references 240 /// are +/- 1MB for conditional jumps and load-literal instructions), so 241 /// it's acceptable to track a minimum and flush-all rather than doing more 242 /// detailed "current minimum" / sort-by-deadline trickery. 243 island_deadline: CodeOffset, 244 /// How many bytes are needed in the worst case for an island, given all 245 /// pending constants and fixups. 246 island_worst_case_size: CodeOffset, 247 /// Latest branches, to facilitate in-place editing for better fallthrough 248 /// behavior and empty-block removal. 249 latest_branches: SmallVec<[MachBranch; 4]>, 250 /// All labels at the current offset (emission tail). This is lazily 251 /// cleared: it is actually accurate as long as the current offset is 252 /// `labels_at_tail_off`, but if `cur_offset()` has grown larger, it should 253 /// be considered as empty. 254 /// 255 /// For correctness, this *must* be complete (i.e., the vector must contain 256 /// all labels whose offsets are resolved to the current tail), because we 257 /// rely on it to update labels when we truncate branches. 258 labels_at_tail: SmallVec<[MachLabel; 4]>, 259 /// The last offset at which `labels_at_tail` is valid. It is conceptually 260 /// always describing the tail of the buffer, but we do not clear 261 /// `labels_at_tail` eagerly when the tail grows, rather we lazily clear it 262 /// when the offset has grown past this (`labels_at_tail_off`) point. 263 /// Always <= `cur_offset()`. 264 labels_at_tail_off: CodeOffset, 265 /// Map used constants to their [MachLabel]. 266 constant_labels: SecondaryMap<VCodeConstant, MachLabel>, 267 } 268 269 impl MachBufferFinalized<Stencil> { 270 pub(crate) fn apply_params(self, params: &FunctionParameters) -> MachBufferFinalized<Final> { 271 MachBufferFinalized { 272 data: self.data, 273 relocs: self.relocs, 274 traps: self.traps, 275 call_sites: self.call_sites, 276 srclocs: self 277 .srclocs 278 .into_iter() 279 .map(|srcloc| srcloc.apply_params(params)) 280 .collect(), 281 stack_maps: self.stack_maps, 282 unwind_info: self.unwind_info, 283 } 284 } 285 } 286 287 /// A `MachBuffer` once emission is completed: holds generated code and records, 288 /// without fixups. This allows the type to be independent of the backend. 289 #[derive(PartialEq, Debug, Clone)] 290 #[cfg_attr(feature = "enable-serde", derive(serde::Serialize, serde::Deserialize))] 291 pub struct MachBufferFinalized<T: CompilePhase> { 292 /// The buffer contents, as raw bytes. 293 pub(crate) data: SmallVec<[u8; 1024]>, 294 /// Any relocations referring to this code. Note that only *external* 295 /// relocations are tracked here; references to labels within the buffer are 296 /// resolved before emission. 297 pub(crate) relocs: SmallVec<[MachReloc; 16]>, 298 /// Any trap records referring to this code. 299 pub(crate) traps: SmallVec<[MachTrap; 16]>, 300 /// Any call site records referring to this code. 301 pub(crate) call_sites: SmallVec<[MachCallSite; 16]>, 302 /// Any source location mappings referring to this code. 303 pub(crate) srclocs: SmallVec<[T::MachSrcLocType; 64]>, 304 /// Any stack maps referring to this code. 305 pub(crate) stack_maps: SmallVec<[MachStackMap; 8]>, 306 /// Any unwind info at a given location. 307 pub unwind_info: SmallVec<[(CodeOffset, UnwindInst); 8]>, 308 } 309 310 const UNKNOWN_LABEL_OFFSET: CodeOffset = 0xffff_ffff; 311 const UNKNOWN_LABEL: MachLabel = MachLabel(0xffff_ffff); 312 313 /// Threshold on max length of `labels_at_this_branch` list to avoid 314 /// unbounded quadratic behavior (see comment below at use-site). 315 const LABEL_LIST_THRESHOLD: usize = 100; 316 317 /// A label refers to some offset in a `MachBuffer`. It may not be resolved at 318 /// the point at which it is used by emitted code; the buffer records "fixups" 319 /// for references to the label, and will come back and patch the code 320 /// appropriately when the label's location is eventually known. 321 #[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)] 322 pub struct MachLabel(u32); 323 entity_impl!(MachLabel); 324 325 impl MachLabel { 326 /// Get a label for a block. (The first N MachLabels are always reseved for 327 /// the N blocks in the vcode.) 328 pub fn from_block(bindex: BlockIndex) -> MachLabel { 329 MachLabel(bindex.index() as u32) 330 } 331 332 /// Get the numeric label index. 333 pub fn get(self) -> u32 { 334 self.0 335 } 336 337 /// Creates a string representing this label, for convenience. 338 pub fn to_string(&self) -> String { 339 format!("label{}", self.0) 340 } 341 } 342 343 impl Default for MachLabel { 344 fn default() -> Self { 345 UNKNOWN_LABEL 346 } 347 } 348 349 /// A stack map extent, when creating a stack map. 350 pub enum StackMapExtent { 351 /// The stack map starts at this instruction, and ends after the number of upcoming bytes 352 /// (note: this is a code offset diff). 353 UpcomingBytes(CodeOffset), 354 355 /// The stack map started at the given offset and ends at the current one. This helps 356 /// architectures where the instruction size has not a fixed length. 357 StartedAtOffset(CodeOffset), 358 } 359 360 impl<I: VCodeInst> MachBuffer<I> { 361 /// Create a new section, known to start at `start_offset` and with a size limited to 362 /// `length_limit`. 363 pub fn new() -> MachBuffer<I> { 364 MachBuffer { 365 data: SmallVec::new(), 366 relocs: SmallVec::new(), 367 traps: SmallVec::new(), 368 call_sites: SmallVec::new(), 369 srclocs: SmallVec::new(), 370 stack_maps: SmallVec::new(), 371 unwind_info: SmallVec::new(), 372 cur_srcloc: None, 373 label_offsets: SmallVec::new(), 374 label_aliases: SmallVec::new(), 375 pending_constants: SmallVec::new(), 376 fixup_records: SmallVec::new(), 377 island_deadline: UNKNOWN_LABEL_OFFSET, 378 island_worst_case_size: 0, 379 latest_branches: SmallVec::new(), 380 labels_at_tail: SmallVec::new(), 381 labels_at_tail_off: 0, 382 constant_labels: SecondaryMap::new(), 383 } 384 } 385 386 /// Current offset from start of buffer. 387 pub fn cur_offset(&self) -> CodeOffset { 388 self.data.len() as CodeOffset 389 } 390 391 /// Add a byte. 392 pub fn put1(&mut self, value: u8) { 393 trace!("MachBuffer: put byte @ {}: {:x}", self.cur_offset(), value); 394 self.data.push(value); 395 396 // Post-invariant: conceptual-labels_at_tail contains a complete and 397 // precise list of labels bound at `cur_offset()`. We have advanced 398 // `cur_offset()`, hence if it had been equal to `labels_at_tail_off` 399 // before, it is not anymore (and it cannot become equal, because 400 // `labels_at_tail_off` is always <= `cur_offset()`). Thus the list is 401 // conceptually empty (even though it is only lazily cleared). No labels 402 // can be bound at this new offset (by invariant on `label_offsets`). 403 // Hence the invariant holds. 404 } 405 406 /// Add 2 bytes. 407 pub fn put2(&mut self, value: u16) { 408 trace!( 409 "MachBuffer: put 16-bit word @ {}: {:x}", 410 self.cur_offset(), 411 value 412 ); 413 let bytes = value.to_le_bytes(); 414 self.data.extend_from_slice(&bytes[..]); 415 416 // Post-invariant: as for `put1()`. 417 } 418 419 /// Add 4 bytes. 420 pub fn put4(&mut self, value: u32) { 421 trace!( 422 "MachBuffer: put 32-bit word @ {}: {:x}", 423 self.cur_offset(), 424 value 425 ); 426 let bytes = value.to_le_bytes(); 427 self.data.extend_from_slice(&bytes[..]); 428 429 // Post-invariant: as for `put1()`. 430 } 431 432 /// Add 8 bytes. 433 pub fn put8(&mut self, value: u64) { 434 trace!( 435 "MachBuffer: put 64-bit word @ {}: {:x}", 436 self.cur_offset(), 437 value 438 ); 439 let bytes = value.to_le_bytes(); 440 self.data.extend_from_slice(&bytes[..]); 441 442 // Post-invariant: as for `put1()`. 443 } 444 445 /// Add a slice of bytes. 446 pub fn put_data(&mut self, data: &[u8]) { 447 trace!( 448 "MachBuffer: put data @ {}: len {}", 449 self.cur_offset(), 450 data.len() 451 ); 452 self.data.extend_from_slice(data); 453 454 // Post-invariant: as for `put1()`. 455 } 456 457 /// Reserve appended space and return a mutable slice referring to it. 458 pub fn get_appended_space(&mut self, len: usize) -> &mut [u8] { 459 trace!("MachBuffer: put data @ {}: len {}", self.cur_offset(), len); 460 let off = self.data.len(); 461 let new_len = self.data.len() + len; 462 self.data.resize(new_len, 0); 463 &mut self.data[off..] 464 465 // Post-invariant: as for `put1()`. 466 } 467 468 /// Align up to the given alignment. 469 pub fn align_to(&mut self, align_to: CodeOffset) { 470 trace!("MachBuffer: align to {}", align_to); 471 assert!(align_to.is_power_of_two()); 472 while self.cur_offset() & (align_to - 1) != 0 { 473 self.put1(0); 474 } 475 476 // Post-invariant: as for `put1()`. 477 } 478 479 /// Allocate a `Label` to refer to some offset. May not be bound to a fixed 480 /// offset yet. 481 pub fn get_label(&mut self) -> MachLabel { 482 let l = self.label_offsets.len() as u32; 483 self.label_offsets.push(UNKNOWN_LABEL_OFFSET); 484 self.label_aliases.push(UNKNOWN_LABEL); 485 trace!("MachBuffer: new label -> {:?}", MachLabel(l)); 486 MachLabel(l) 487 488 // Post-invariant: the only mutation is to add a new label; it has no 489 // bound offset yet, so it trivially satisfies all invariants. 490 } 491 492 /// Reserve the first N MachLabels for blocks. 493 pub fn reserve_labels_for_blocks(&mut self, blocks: usize) { 494 trace!("MachBuffer: first {} labels are for blocks", blocks); 495 debug_assert!(self.label_offsets.is_empty()); 496 self.label_offsets.resize(blocks, UNKNOWN_LABEL_OFFSET); 497 self.label_aliases.resize(blocks, UNKNOWN_LABEL); 498 499 // Post-invariant: as for `get_label()`. 500 } 501 502 /// Reserve the next N MachLabels for constants. 503 pub fn reserve_labels_for_constants(&mut self, constants: &VCodeConstants) { 504 trace!( 505 "MachBuffer: next {} labels are for constants", 506 constants.len() 507 ); 508 for c in constants.keys() { 509 self.constant_labels[c] = self.get_label(); 510 } 511 512 // Post-invariant: as for `get_label()`. 513 } 514 515 /// Retrieve the reserved label for a constant. 516 pub fn get_label_for_constant(&self, constant: VCodeConstant) -> MachLabel { 517 self.constant_labels[constant] 518 } 519 520 /// Bind a label to the current offset. A label can only be bound once. 521 pub fn bind_label(&mut self, label: MachLabel) { 522 trace!( 523 "MachBuffer: bind label {:?} at offset {}", 524 label, 525 self.cur_offset() 526 ); 527 debug_assert_eq!(self.label_offsets[label.0 as usize], UNKNOWN_LABEL_OFFSET); 528 debug_assert_eq!(self.label_aliases[label.0 as usize], UNKNOWN_LABEL); 529 let offset = self.cur_offset(); 530 self.label_offsets[label.0 as usize] = offset; 531 self.lazily_clear_labels_at_tail(); 532 self.labels_at_tail.push(label); 533 534 // Invariants hold: bound offset of label is <= cur_offset (in fact it 535 // is equal). If the `labels_at_tail` list was complete and precise 536 // before, it is still, because we have bound this label to the current 537 // offset and added it to the list (which contains all labels at the 538 // current offset). 539 540 self.optimize_branches(); 541 542 // Post-invariant: by `optimize_branches()` (see argument there). 543 } 544 545 /// Lazily clear `labels_at_tail` if the tail offset has moved beyond the 546 /// offset that it applies to. 547 fn lazily_clear_labels_at_tail(&mut self) { 548 let offset = self.cur_offset(); 549 if offset > self.labels_at_tail_off { 550 self.labels_at_tail_off = offset; 551 self.labels_at_tail.clear(); 552 } 553 554 // Post-invariant: either labels_at_tail_off was at cur_offset, and 555 // state is untouched, or was less than cur_offset, in which case the 556 // labels_at_tail list was conceptually empty, and is now actually 557 // empty. 558 } 559 560 /// Resolve a label to an offset, if known. May return `UNKNOWN_LABEL_OFFSET`. 561 pub(crate) fn resolve_label_offset(&self, mut label: MachLabel) -> CodeOffset { 562 let mut iters = 0; 563 while self.label_aliases[label.0 as usize] != UNKNOWN_LABEL { 564 label = self.label_aliases[label.0 as usize]; 565 // To protect against an infinite loop (despite our assurances to 566 // ourselves that the invariants make this impossible), assert out 567 // after 1M iterations. The number of basic blocks is limited 568 // in most contexts anyway so this should be impossible to hit with 569 // a legitimate input. 570 iters += 1; 571 assert!(iters < 1_000_000, "Unexpected cycle in label aliases"); 572 } 573 self.label_offsets[label.0 as usize] 574 575 // Post-invariant: no mutations. 576 } 577 578 /// Emit a reference to the given label with the given reference type (i.e., 579 /// branch-instruction format) at the current offset. This is like a 580 /// relocation, but handled internally. 581 /// 582 /// This can be called before the branch is actually emitted; fixups will 583 /// not happen until an island is emitted or the buffer is finished. 584 pub fn use_label_at_offset(&mut self, offset: CodeOffset, label: MachLabel, kind: I::LabelUse) { 585 trace!( 586 "MachBuffer: use_label_at_offset: offset {} label {:?} kind {:?}", 587 offset, 588 label, 589 kind 590 ); 591 592 // Add the fixup, and update the worst-case island size based on a 593 // veneer for this label use. 594 self.fixup_records.push(MachLabelFixup { 595 label, 596 offset, 597 kind, 598 }); 599 if kind.supports_veneer() { 600 self.island_worst_case_size += kind.veneer_size(); 601 self.island_worst_case_size &= !(I::LabelUse::ALIGN - 1); 602 } 603 let deadline = offset.saturating_add(kind.max_pos_range()); 604 if deadline < self.island_deadline { 605 self.island_deadline = deadline; 606 } 607 608 // Post-invariant: no mutations to branches/labels data structures. 609 } 610 611 /// Inform the buffer of an unconditional branch at the given offset, 612 /// targetting the given label. May be used to optimize branches. 613 /// The last added label-use must correspond to this branch. 614 /// This must be called when the current offset is equal to `start`; i.e., 615 /// before actually emitting the branch. This implies that for a branch that 616 /// uses a label and is eligible for optimizations by the MachBuffer, the 617 /// proper sequence is: 618 /// 619 /// - Call `use_label_at_offset()` to emit the fixup record. 620 /// - Call `add_uncond_branch()` to make note of the branch. 621 /// - Emit the bytes for the branch's machine code. 622 /// 623 /// Additional requirement: no labels may be bound between `start` and `end` 624 /// (exclusive on both ends). 625 pub fn add_uncond_branch(&mut self, start: CodeOffset, end: CodeOffset, target: MachLabel) { 626 assert!(self.cur_offset() == start); 627 debug_assert!(end > start); 628 assert!(!self.fixup_records.is_empty()); 629 let fixup = self.fixup_records.len() - 1; 630 self.lazily_clear_labels_at_tail(); 631 self.latest_branches.push(MachBranch { 632 start, 633 end, 634 target, 635 fixup, 636 inverted: None, 637 labels_at_this_branch: self.labels_at_tail.clone(), 638 }); 639 640 // Post-invariant: we asserted branch start is current tail; the list of 641 // labels at branch is cloned from list of labels at current tail. 642 } 643 644 /// Inform the buffer of a conditional branch at the given offset, 645 /// targetting the given label. May be used to optimize branches. 646 /// The last added label-use must correspond to this branch. 647 /// 648 /// Additional requirement: no labels may be bound between `start` and `end` 649 /// (exclusive on both ends). 650 pub fn add_cond_branch( 651 &mut self, 652 start: CodeOffset, 653 end: CodeOffset, 654 target: MachLabel, 655 inverted: &[u8], 656 ) { 657 assert!(self.cur_offset() == start); 658 debug_assert!(end > start); 659 assert!(!self.fixup_records.is_empty()); 660 debug_assert!(inverted.len() == (end - start) as usize); 661 let fixup = self.fixup_records.len() - 1; 662 let inverted = Some(SmallVec::from(inverted)); 663 self.lazily_clear_labels_at_tail(); 664 self.latest_branches.push(MachBranch { 665 start, 666 end, 667 target, 668 fixup, 669 inverted, 670 labels_at_this_branch: self.labels_at_tail.clone(), 671 }); 672 673 // Post-invariant: we asserted branch start is current tail; labels at 674 // branch list is cloned from list of labels at current tail. 675 } 676 677 fn truncate_last_branch(&mut self) { 678 self.lazily_clear_labels_at_tail(); 679 // Invariants hold at this point. 680 681 let b = self.latest_branches.pop().unwrap(); 682 assert!(b.end == self.cur_offset()); 683 684 // State: 685 // [PRE CODE] 686 // Offset b.start, b.labels_at_this_branch: 687 // [BRANCH CODE] 688 // cur_off, self.labels_at_tail --> 689 // (end of buffer) 690 self.data.truncate(b.start as usize); 691 self.fixup_records.truncate(b.fixup); 692 while let Some(mut last_srcloc) = self.srclocs.last_mut() { 693 if last_srcloc.end <= b.start { 694 break; 695 } 696 if last_srcloc.start < b.start { 697 last_srcloc.end = b.start; 698 break; 699 } 700 self.srclocs.pop(); 701 } 702 // State: 703 // [PRE CODE] 704 // cur_off, Offset b.start, b.labels_at_this_branch: 705 // (end of buffer) 706 // 707 // self.labels_at_tail --> (past end of buffer) 708 let cur_off = self.cur_offset(); 709 self.labels_at_tail_off = cur_off; 710 // State: 711 // [PRE CODE] 712 // cur_off, Offset b.start, b.labels_at_this_branch, 713 // self.labels_at_tail: 714 // (end of buffer) 715 // 716 // resolve_label_offset(l) for l in labels_at_tail: 717 // (past end of buffer) 718 719 trace!( 720 "truncate_last_branch: truncated {:?}; off now {}", 721 b, 722 cur_off 723 ); 724 725 // Fix up resolved label offsets for labels at tail. 726 for &l in &self.labels_at_tail { 727 self.label_offsets[l.0 as usize] = cur_off; 728 } 729 // Old labels_at_this_branch are now at cur_off. 730 self.labels_at_tail 731 .extend(b.labels_at_this_branch.into_iter()); 732 733 // Post-invariant: this operation is defined to truncate the buffer, 734 // which moves cur_off backward, and to move labels at the end of the 735 // buffer back to the start-of-branch offset. 736 // 737 // latest_branches satisfies all invariants: 738 // - it has no branches past the end of the buffer (branches are in 739 // order, we removed the last one, and we truncated the buffer to just 740 // before the start of that branch) 741 // - no labels were moved to lower offsets than the (new) cur_off, so 742 // the labels_at_this_branch list for any other branch need not change. 743 // 744 // labels_at_tail satisfies all invariants: 745 // - all labels that were at the tail after the truncated branch are 746 // moved backward to just before the branch, which becomes the new tail; 747 // thus every element in the list should remain (ensured by `.extend()` 748 // above). 749 // - all labels that refer to the new tail, which is the start-offset of 750 // the truncated branch, must be present. The `labels_at_this_branch` 751 // list in the truncated branch's record is a complete and precise list 752 // of exactly these labels; we append these to labels_at_tail. 753 // - labels_at_tail_off is at cur_off after truncation occurs, so the 754 // list is valid (not to be lazily cleared). 755 // 756 // The stated operation was performed: 757 // - For each label at the end of the buffer prior to this method, it 758 // now resolves to the new (truncated) end of the buffer: it must have 759 // been in `labels_at_tail` (this list is precise and complete, and 760 // the tail was at the end of the truncated branch on entry), and we 761 // iterate over this list and set `label_offsets` to the new tail. 762 // None of these labels could have been an alias (by invariant), so 763 // `label_offsets` is authoritative for each. 764 // - No other labels will be past the end of the buffer, because of the 765 // requirement that no labels be bound to the middle of branch ranges 766 // (see comments to `add_{cond,uncond}_branch()`). 767 // - The buffer is truncated to just before the last branch, and the 768 // fixup record referring to that last branch is removed. 769 } 770 771 fn optimize_branches(&mut self) { 772 self.lazily_clear_labels_at_tail(); 773 // Invariants valid at this point. 774 775 trace!( 776 "enter optimize_branches:\n b = {:?}\n l = {:?}\n f = {:?}", 777 self.latest_branches, 778 self.labels_at_tail, 779 self.fixup_records 780 ); 781 782 // We continue to munch on branches at the tail of the buffer until no 783 // more rules apply. Note that the loop only continues if a branch is 784 // actually truncated (or if labels are redirected away from a branch), 785 // so this always makes progress. 786 while let Some(b) = self.latest_branches.last() { 787 let cur_off = self.cur_offset(); 788 trace!("optimize_branches: last branch {:?} at off {}", b, cur_off); 789 // If there has been any code emission since the end of the last branch or 790 // label definition, then there's nothing we can edit (because we 791 // don't move code once placed, only back up and overwrite), so 792 // clear the records and finish. 793 if b.end < cur_off { 794 break; 795 } 796 797 // If the "labels at this branch" list on this branch is 798 // longer than a threshold, don't do any simplification, 799 // and let the branch remain to separate those labels from 800 // the current tail. This avoids quadratic behavior (see 801 // #3468): otherwise, if a long string of "goto next; 802 // next:" patterns are emitted, all of the labels will 803 // coalesce into a long list of aliases for the current 804 // buffer tail. We must track all aliases of the current 805 // tail for correctness, but we are also allowed to skip 806 // optimization (removal) of any branch, so we take the 807 // escape hatch here and let it stand. In effect this 808 // "spreads" the many thousands of labels in the 809 // pathological case among an actual (harmless but 810 // suboptimal) instruction once per N labels. 811 if b.labels_at_this_branch.len() > LABEL_LIST_THRESHOLD { 812 break; 813 } 814 815 // Invariant: we are looking at a branch that ends at the tail of 816 // the buffer. 817 818 // For any branch, conditional or unconditional: 819 // - If the target is a label at the current offset, then remove 820 // the conditional branch, and reset all labels that targetted 821 // the current offset (end of branch) to the truncated 822 // end-of-code. 823 // 824 // Preserves execution semantics: a branch to its own fallthrough 825 // address is equivalent to a no-op; in both cases, nextPC is the 826 // fallthrough. 827 if self.resolve_label_offset(b.target) == cur_off { 828 trace!("branch with target == cur off; truncating"); 829 self.truncate_last_branch(); 830 continue; 831 } 832 833 // If latest is an unconditional branch: 834 // 835 // - If the branch's target is not its own start address, then for 836 // each label at the start of branch, make the label an alias of the 837 // branch target, and remove the label from the "labels at this 838 // branch" list. 839 // 840 // - Preserves execution semantics: an unconditional branch's 841 // only effect is to set PC to a new PC; this change simply 842 // collapses one step in the step-semantics. 843 // 844 // - Post-invariant: the labels that were bound to the start of 845 // this branch become aliases, so they must not be present in any 846 // labels-at-this-branch list or the labels-at-tail list. The 847 // labels are removed form the latest-branch record's 848 // labels-at-this-branch list, and are never placed in the 849 // labels-at-tail list. Furthermore, it is correct that they are 850 // not in either list, because they are now aliases, and labels 851 // that are aliases remain aliases forever. 852 // 853 // - If there is a prior unconditional branch that ends just before 854 // this one begins, and this branch has no labels bound to its 855 // start, then we can truncate this branch, because it is entirely 856 // unreachable (we have redirected all labels that make it 857 // reachable otherwise). Do so and continue around the loop. 858 // 859 // - Preserves execution semantics: the branch is unreachable, 860 // because execution can only flow into an instruction from the 861 // prior instruction's fallthrough or from a branch bound to that 862 // instruction's start offset. Unconditional branches have no 863 // fallthrough, so if the prior instruction is an unconditional 864 // branch, no fallthrough entry can happen. The 865 // labels-at-this-branch list is complete (by invariant), so if it 866 // is empty, then the instruction is entirely unreachable. Thus, 867 // it can be removed. 868 // 869 // - Post-invariant: ensured by truncate_last_branch(). 870 // 871 // - If there is a prior conditional branch whose target label 872 // resolves to the current offset (branches around the 873 // unconditional branch), then remove the unconditional branch, 874 // and make the target of the unconditional the target of the 875 // conditional instead. 876 // 877 // - Preserves execution semantics: previously we had: 878 // 879 // L1: 880 // cond_br L2 881 // br L3 882 // L2: 883 // (end of buffer) 884 // 885 // by removing the last branch, we have: 886 // 887 // L1: 888 // cond_br L2 889 // L2: 890 // (end of buffer) 891 // 892 // we then fix up the records for the conditional branch to 893 // have: 894 // 895 // L1: 896 // cond_br.inverted L3 897 // L2: 898 // 899 // In the original code, control flow reaches L2 when the 900 // conditional branch's predicate is true, and L3 otherwise. In 901 // the optimized code, the same is true. 902 // 903 // - Post-invariant: all edits to latest_branches and 904 // labels_at_tail are performed by `truncate_last_branch()`, 905 // which maintains the invariants at each step. 906 907 if b.is_uncond() { 908 // Set any label equal to current branch's start as an alias of 909 // the branch's target, if the target is not the branch itself 910 // (i.e., an infinite loop). 911 // 912 // We cannot perform this aliasing if the target of this branch 913 // ultimately aliases back here; if so, we need to keep this 914 // branch, so break out of this loop entirely (and clear the 915 // latest-branches list below). 916 // 917 // Note that this check is what prevents cycles from forming in 918 // `self.label_aliases`. To see why, consider an arbitrary start 919 // state: 920 // 921 // label_aliases[L1] = L2, label_aliases[L2] = L3, ..., up to 922 // Ln, which is not aliased. 923 // 924 // We would create a cycle if we assigned label_aliases[Ln] 925 // = L1. Note that the below assignment is the only write 926 // to label_aliases. 927 // 928 // By our other invariants, we have that Ln (`l` below) 929 // resolves to the offset `b.start`, because it is in the 930 // set `b.labels_at_this_branch`. 931 // 932 // If L1 were already aliased, through some arbitrarily deep 933 // chain, to Ln, then it must also resolve to this offset 934 // `b.start`. 935 // 936 // By checking the resolution of `L1` against this offset, 937 // and aborting this branch-simplification if they are 938 // equal, we prevent the below assignment from ever creating 939 // a cycle. 940 if self.resolve_label_offset(b.target) != b.start { 941 let redirected = b.labels_at_this_branch.len(); 942 for &l in &b.labels_at_this_branch { 943 trace!( 944 " -> label at start of branch {:?} redirected to target {:?}", 945 l, 946 b.target 947 ); 948 self.label_aliases[l.0 as usize] = b.target; 949 // NOTE: we continue to ensure the invariant that labels 950 // pointing to tail of buffer are in `labels_at_tail` 951 // because we already ensured above that the last branch 952 // cannot have a target of `cur_off`; so we never have 953 // to put the label into `labels_at_tail` when moving it 954 // here. 955 } 956 // Maintain invariant: all branches have been redirected 957 // and are no longer pointing at the start of this branch. 958 let mut_b = self.latest_branches.last_mut().unwrap(); 959 mut_b.labels_at_this_branch.clear(); 960 961 if redirected > 0 { 962 trace!(" -> after label redirects, restarting loop"); 963 continue; 964 } 965 } else { 966 break; 967 } 968 969 let b = self.latest_branches.last().unwrap(); 970 971 // Examine any immediately preceding branch. 972 if self.latest_branches.len() > 1 { 973 let prev_b = &self.latest_branches[self.latest_branches.len() - 2]; 974 trace!(" -> more than one branch; prev_b = {:?}", prev_b); 975 // This uncond is immediately after another uncond; we 976 // should have already redirected labels to this uncond away 977 // (but check to be sure); so we can truncate this uncond. 978 if prev_b.is_uncond() 979 && prev_b.end == b.start 980 && b.labels_at_this_branch.is_empty() 981 { 982 trace!(" -> uncond follows another uncond; truncating"); 983 self.truncate_last_branch(); 984 continue; 985 } 986 987 // This uncond is immediately after a conditional, and the 988 // conditional's target is the end of this uncond, and we've 989 // already redirected labels to this uncond away; so we can 990 // truncate this uncond, flip the sense of the conditional, and 991 // set the conditional's target (in `latest_branches` and in 992 // `fixup_records`) to the uncond's target. 993 if prev_b.is_cond() 994 && prev_b.end == b.start 995 && self.resolve_label_offset(prev_b.target) == cur_off 996 { 997 trace!(" -> uncond follows a conditional, and conditional's target resolves to current offset"); 998 // Save the target of the uncond (this becomes the 999 // target of the cond), and truncate the uncond. 1000 let target = b.target; 1001 let data = prev_b.inverted.clone().unwrap(); 1002 self.truncate_last_branch(); 1003 1004 // Mutate the code and cond branch. 1005 let off_before_edit = self.cur_offset(); 1006 let prev_b = self.latest_branches.last_mut().unwrap(); 1007 let not_inverted = SmallVec::from( 1008 &self.data[(prev_b.start as usize)..(prev_b.end as usize)], 1009 ); 1010 1011 // Low-level edit: replaces bytes of branch with 1012 // inverted form. cur_off remains the same afterward, so 1013 // we do not need to modify label data structures. 1014 self.data.truncate(prev_b.start as usize); 1015 self.data.extend_from_slice(&data[..]); 1016 1017 // Save the original code as the inversion of the 1018 // inverted branch, in case we later edit this branch 1019 // again. 1020 prev_b.inverted = Some(not_inverted); 1021 self.fixup_records[prev_b.fixup].label = target; 1022 trace!(" -> reassigning target of condbr to {:?}", target); 1023 prev_b.target = target; 1024 debug_assert_eq!(off_before_edit, self.cur_offset()); 1025 continue; 1026 } 1027 } 1028 } 1029 1030 // If we couldn't do anything with the last branch, then break. 1031 break; 1032 } 1033 1034 self.purge_latest_branches(); 1035 1036 trace!( 1037 "leave optimize_branches:\n b = {:?}\n l = {:?}\n f = {:?}", 1038 self.latest_branches, 1039 self.labels_at_tail, 1040 self.fixup_records 1041 ); 1042 } 1043 1044 fn purge_latest_branches(&mut self) { 1045 // All of our branch simplification rules work only if a branch ends at 1046 // the tail of the buffer, with no following code; and branches are in 1047 // order in latest_branches; so if the last entry ends prior to 1048 // cur_offset, then clear all entries. 1049 let cur_off = self.cur_offset(); 1050 if let Some(l) = self.latest_branches.last() { 1051 if l.end < cur_off { 1052 trace!("purge_latest_branches: removing branch {:?}", l); 1053 self.latest_branches.clear(); 1054 } 1055 } 1056 1057 // Post-invariant: no invariant requires any branch to appear in 1058 // `latest_branches`; it is always optional. The list-clear above thus 1059 // preserves all semantics. 1060 } 1061 1062 /// Emit a constant at some point in the future, binding the given label to 1063 /// its offset. The constant will be placed at most `max_distance` from the 1064 /// current offset. 1065 pub fn defer_constant( 1066 &mut self, 1067 label: MachLabel, 1068 align: CodeOffset, 1069 data: &[u8], 1070 max_distance: CodeOffset, 1071 ) { 1072 trace!( 1073 "defer_constant: eventually emit {} bytes aligned to {} at label {:?}", 1074 data.len(), 1075 align, 1076 label 1077 ); 1078 let deadline = self.cur_offset().saturating_add(max_distance); 1079 self.island_worst_case_size += data.len() as CodeOffset; 1080 self.island_worst_case_size = 1081 (self.island_worst_case_size + I::LabelUse::ALIGN - 1) & !(I::LabelUse::ALIGN - 1); 1082 self.pending_constants.push(MachLabelConstant { 1083 label, 1084 align, 1085 data: SmallVec::from(data), 1086 }); 1087 if deadline < self.island_deadline { 1088 self.island_deadline = deadline; 1089 } 1090 } 1091 1092 /// Is an island needed within the next N bytes? 1093 pub fn island_needed(&self, distance: CodeOffset) -> bool { 1094 self.worst_case_end_of_island(distance) > self.island_deadline 1095 } 1096 1097 /// Returns the maximal offset that islands can reach if `distance` more 1098 /// bytes are appended. 1099 /// 1100 /// This is used to determine if veneers need insertions since jumps that 1101 /// can't reach past this point must get a veneer of some form. 1102 fn worst_case_end_of_island(&self, distance: CodeOffset) -> CodeOffset { 1103 self.cur_offset() 1104 .saturating_add(distance) 1105 .saturating_add(self.island_worst_case_size) 1106 } 1107 1108 /// Emit all pending constants and required pending veneers. 1109 /// 1110 /// Should only be called if `island_needed()` returns true, i.e., if we 1111 /// actually reach a deadline. It's not necessarily a problem to do so 1112 /// otherwise but it may result in unnecessary work during emission. 1113 pub fn emit_island(&mut self, distance: CodeOffset) { 1114 self.emit_island_maybe_forced(false, distance); 1115 } 1116 1117 /// Same as `emit_island`, but an internal API with a `force_veneers` 1118 /// argument to force all veneers to always get emitted for debugging. 1119 fn emit_island_maybe_forced(&mut self, force_veneers: bool, distance: CodeOffset) { 1120 // We're going to purge fixups, so no latest-branch editing can happen 1121 // anymore. 1122 self.latest_branches.clear(); 1123 1124 // Reset internal calculations about islands since we're going to 1125 // change the calculus as we apply fixups. The `forced_threshold` is 1126 // used here to determine whether jumps to unknown labels will require 1127 // a veneer or not. 1128 let forced_threshold = self.worst_case_end_of_island(distance); 1129 self.island_deadline = UNKNOWN_LABEL_OFFSET; 1130 self.island_worst_case_size = 0; 1131 1132 // First flush out all constants so we have more labels in case fixups 1133 // are applied against these labels. 1134 for MachLabelConstant { label, align, data } in mem::take(&mut self.pending_constants) { 1135 self.align_to(align); 1136 self.bind_label(label); 1137 self.put_data(&data[..]); 1138 } 1139 1140 for fixup in mem::take(&mut self.fixup_records) { 1141 trace!("emit_island: fixup {:?}", fixup); 1142 let MachLabelFixup { 1143 label, 1144 offset, 1145 kind, 1146 } = fixup; 1147 let label_offset = self.resolve_label_offset(label); 1148 let start = offset as usize; 1149 let end = (offset + kind.patch_size()) as usize; 1150 1151 if label_offset != UNKNOWN_LABEL_OFFSET { 1152 // If the offset of the label for this fixup is known then 1153 // we're going to do something here-and-now. We're either going 1154 // to patch the original offset because it's an in-bounds jump, 1155 // or we're going to generate a veneer, patch the fixup to jump 1156 // to the veneer, and then keep going. 1157 // 1158 // If the label comes after the original fixup, then we should 1159 // be guaranteed that the jump is in-bounds. Otherwise there's 1160 // a bug somewhere because this method wasn't called soon 1161 // enough. All forward-jumps are tracked and should get veneers 1162 // before their deadline comes and they're unable to jump 1163 // further. 1164 // 1165 // Otherwise if the label is before the fixup, then that's a 1166 // backwards jump. If it's past the maximum negative range 1167 // then we'll emit a veneer that to jump forward to which can 1168 // then jump backwards. 1169 let veneer_required = if label_offset >= offset { 1170 assert!((label_offset - offset) <= kind.max_pos_range()); 1171 false 1172 } else { 1173 (offset - label_offset) > kind.max_neg_range() 1174 }; 1175 trace!( 1176 " -> label_offset = {}, known, required = {} (pos {} neg {})", 1177 label_offset, 1178 veneer_required, 1179 kind.max_pos_range(), 1180 kind.max_neg_range() 1181 ); 1182 1183 if (force_veneers && kind.supports_veneer()) || veneer_required { 1184 self.emit_veneer(label, offset, kind); 1185 } else { 1186 let slice = &mut self.data[start..end]; 1187 trace!("patching in-range!"); 1188 kind.patch(slice, offset, label_offset); 1189 } 1190 } else { 1191 // If the offset of this label is not known at this time then 1192 // there's one of two possibilities: 1193 // 1194 // * First we may be about to exceed the maximum jump range of 1195 // this fixup. In that case a veneer is inserted to buy some 1196 // more budget for the forward-jump. It's guaranteed that the 1197 // label will eventually come after where we're at, so we know 1198 // that the forward jump is necessary. 1199 // 1200 // * Otherwise we're still within range of the forward jump but 1201 // the precise target isn't known yet. In that case we 1202 // enqueue the fixup to get processed later. 1203 if forced_threshold - offset > kind.max_pos_range() { 1204 self.emit_veneer(label, offset, kind); 1205 } else { 1206 self.use_label_at_offset(offset, label, kind); 1207 } 1208 } 1209 } 1210 } 1211 1212 /// Emits a "veneer" the `kind` code at `offset` to jump to `label`. 1213 /// 1214 /// This will generate extra machine code, using `kind`, to get a 1215 /// larger-jump-kind than `kind` allows. The code at `offset` is then 1216 /// patched to jump to our new code, and then the new code is enqueued for 1217 /// a fixup to get processed at some later time. 1218 fn emit_veneer(&mut self, label: MachLabel, offset: CodeOffset, kind: I::LabelUse) { 1219 // If this `kind` doesn't support a veneer then that's a bug in the 1220 // backend because we need to implement support for such a veneer. 1221 assert!( 1222 kind.supports_veneer(), 1223 "jump beyond the range of {:?} but a veneer isn't supported", 1224 kind, 1225 ); 1226 1227 // Allocate space for a veneer in the island. 1228 self.align_to(I::LabelUse::ALIGN); 1229 let veneer_offset = self.cur_offset(); 1230 trace!("making a veneer at {}", veneer_offset); 1231 let start = offset as usize; 1232 let end = (offset + kind.patch_size()) as usize; 1233 let slice = &mut self.data[start..end]; 1234 // Patch the original label use to refer to the veneer. 1235 trace!( 1236 "patching original at offset {} to veneer offset {}", 1237 offset, 1238 veneer_offset 1239 ); 1240 kind.patch(slice, offset, veneer_offset); 1241 // Generate the veneer. 1242 let veneer_slice = self.get_appended_space(kind.veneer_size() as usize); 1243 let (veneer_fixup_off, veneer_label_use) = 1244 kind.generate_veneer(veneer_slice, veneer_offset); 1245 trace!( 1246 "generated veneer; fixup offset {}, label_use {:?}", 1247 veneer_fixup_off, 1248 veneer_label_use 1249 ); 1250 // Register a new use of `label` with our new veneer fixup and offset. 1251 // This'll recalculate deadlines accordingly and enqueue this fixup to 1252 // get processed at some later time. 1253 self.use_label_at_offset(veneer_fixup_off, label, veneer_label_use); 1254 } 1255 1256 fn finish_emission_maybe_forcing_veneers(&mut self, force_veneers: bool) { 1257 while !self.pending_constants.is_empty() || !self.fixup_records.is_empty() { 1258 // `emit_island()` will emit any pending veneers and constants, and 1259 // as a side-effect, will also take care of any fixups with resolved 1260 // labels eagerly. 1261 self.emit_island_maybe_forced(force_veneers, u32::MAX); 1262 } 1263 1264 // Ensure that all labels have been fixed up after the last island is emitted. This is a 1265 // full (release-mode) assert because an unresolved label means the emitted code is 1266 // incorrect. 1267 assert!(self.fixup_records.is_empty()); 1268 } 1269 1270 /// Finish any deferred emissions and/or fixups. 1271 pub fn finish(mut self) -> MachBufferFinalized<Stencil> { 1272 let _tt = timing::vcode_emit_finish(); 1273 1274 // Do any optimizations on branches at tail of buffer, as if we 1275 // had bound one last label. 1276 self.optimize_branches(); 1277 1278 self.finish_emission_maybe_forcing_veneers(false); 1279 1280 let mut srclocs = self.srclocs; 1281 srclocs.sort_by_key(|entry| entry.start); 1282 1283 MachBufferFinalized { 1284 data: self.data, 1285 relocs: self.relocs, 1286 traps: self.traps, 1287 call_sites: self.call_sites, 1288 srclocs, 1289 stack_maps: self.stack_maps, 1290 unwind_info: self.unwind_info, 1291 } 1292 } 1293 1294 /// Add an external relocation at the current offset. 1295 pub fn add_reloc(&mut self, kind: Reloc, name: &ExternalName, addend: Addend) { 1296 let name = name.clone(); 1297 // FIXME(#3277): This should use `I::LabelUse::from_reloc` to optionally 1298 // generate a label-use statement to track whether an island is possibly 1299 // needed to escape this function to actually get to the external name. 1300 // This is most likely to come up on AArch64 where calls between 1301 // functions use a 26-bit signed offset which gives +/- 64MB. This means 1302 // that if a function is 128MB in size and there's a call in the middle 1303 // it's impossible to reach the actual target. Also, while it's 1304 // technically possible to jump to the start of a function and then jump 1305 // further, island insertion below always inserts islands after 1306 // previously appended code so for Cranelift's own implementation this 1307 // is also a problem for 64MB functions on AArch64 which start with a 1308 // call instruction, those won't be able to escape. 1309 // 1310 // Ideally what needs to happen here is that a `LabelUse` is 1311 // transparently generated (or call-sites of this function are audited 1312 // to generate a `LabelUse` instead) and tracked internally. The actual 1313 // relocation would then change over time if and when a veneer is 1314 // inserted, where the relocation here would be patched by this 1315 // `MachBuffer` to jump to the veneer. The problem, though, is that all 1316 // this still needs to end up, in the case of a singular function, 1317 // generating a final relocation pointing either to this particular 1318 // relocation or to the veneer inserted. Additionally 1319 // `MachBuffer` needs the concept of a label which will never be 1320 // resolved, so `emit_island` doesn't trip over not actually ever 1321 // knowning what some labels are. Currently the loop in 1322 // `finish_emission_maybe_forcing_veneers` would otherwise infinitely 1323 // loop. 1324 // 1325 // For now this means that because relocs aren't tracked at all that 1326 // AArch64 functions have a rough size limits of 64MB. For now that's 1327 // somewhat reasonable and the failure mode is a panic in `MachBuffer` 1328 // when a relocation can't otherwise be resolved later, so it shouldn't 1329 // actually result in any memory unsafety or anything like that. 1330 self.relocs.push(MachReloc { 1331 offset: self.data.len() as CodeOffset, 1332 kind, 1333 name, 1334 addend, 1335 }); 1336 } 1337 1338 /// Add a trap record at the current offset. 1339 pub fn add_trap(&mut self, code: TrapCode) { 1340 self.traps.push(MachTrap { 1341 offset: self.data.len() as CodeOffset, 1342 code, 1343 }); 1344 } 1345 1346 /// Add a call-site record at the current offset. 1347 pub fn add_call_site(&mut self, opcode: Opcode) { 1348 debug_assert!( 1349 opcode.is_call(), 1350 "adding call site info for a non-call instruction." 1351 ); 1352 self.call_sites.push(MachCallSite { 1353 ret_addr: self.data.len() as CodeOffset, 1354 opcode, 1355 }); 1356 } 1357 1358 /// Add an unwind record at the current offset. 1359 pub fn add_unwind(&mut self, unwind: UnwindInst) { 1360 self.unwind_info.push((self.cur_offset(), unwind)); 1361 } 1362 1363 /// Set the `SourceLoc` for code from this offset until the offset at the 1364 /// next call to `end_srcloc()`. 1365 pub fn start_srcloc(&mut self, loc: RelSourceLoc) { 1366 self.cur_srcloc = Some((self.cur_offset(), loc)); 1367 } 1368 1369 /// Mark the end of the `SourceLoc` segment started at the last 1370 /// `start_srcloc()` call. 1371 pub fn end_srcloc(&mut self) { 1372 let (start, loc) = self 1373 .cur_srcloc 1374 .take() 1375 .expect("end_srcloc() called without start_srcloc()"); 1376 let end = self.cur_offset(); 1377 // Skip zero-length extends. 1378 debug_assert!(end >= start); 1379 if end > start { 1380 self.srclocs.push(MachSrcLoc { start, end, loc }); 1381 } 1382 } 1383 1384 /// Add stack map metadata for this program point: a set of stack offsets 1385 /// (from SP upward) that contain live references. 1386 /// 1387 /// The `offset_to_fp` value is the offset from the nominal SP (at which the `stack_offsets` 1388 /// are based) and the FP value. By subtracting `offset_to_fp` from each `stack_offsets` 1389 /// element, one can obtain live-reference offsets from FP instead. 1390 pub fn add_stack_map(&mut self, extent: StackMapExtent, stack_map: StackMap) { 1391 let (start, end) = match extent { 1392 StackMapExtent::UpcomingBytes(insn_len) => { 1393 let start_offset = self.cur_offset(); 1394 (start_offset, start_offset + insn_len) 1395 } 1396 StackMapExtent::StartedAtOffset(start_offset) => { 1397 let end_offset = self.cur_offset(); 1398 debug_assert!(end_offset >= start_offset); 1399 (start_offset, end_offset) 1400 } 1401 }; 1402 trace!("Adding stack map for offsets {start:#x}..{end:#x}"); 1403 self.stack_maps.push(MachStackMap { 1404 offset: start, 1405 offset_end: end, 1406 stack_map, 1407 }); 1408 } 1409 } 1410 1411 impl<T: CompilePhase> MachBufferFinalized<T> { 1412 /// Get a list of source location mapping tuples in sorted-by-start-offset order. 1413 pub fn get_srclocs_sorted(&self) -> &[T::MachSrcLocType] { 1414 &self.srclocs[..] 1415 } 1416 1417 /// Get the total required size for the code. 1418 pub fn total_size(&self) -> CodeOffset { 1419 self.data.len() as CodeOffset 1420 } 1421 1422 /// Return the code in this mach buffer as a hex string for testing purposes. 1423 pub fn stringify_code_bytes(&self) -> String { 1424 // This is pretty lame, but whatever .. 1425 use std::fmt::Write; 1426 let mut s = String::with_capacity(self.data.len() * 2); 1427 for b in &self.data { 1428 write!(&mut s, "{:02X}", b).unwrap(); 1429 } 1430 s 1431 } 1432 1433 /// Get the code bytes. 1434 pub fn data(&self) -> &[u8] { 1435 // N.B.: we emit every section into the .text section as far as 1436 // the `CodeSink` is concerned; we do not bother to segregate 1437 // the contents into the actual program text, the jumptable and the 1438 // rodata (constant pool). This allows us to generate code assuming 1439 // that these will not be relocated relative to each other, and avoids 1440 // having to designate each section as belonging in one of the three 1441 // fixed categories defined by `CodeSink`. If this becomes a problem 1442 // later (e.g. because of memory permissions or similar), we can 1443 // add this designation and segregate the output; take care, however, 1444 // to add the appropriate relocations in this case. 1445 1446 &self.data[..] 1447 } 1448 1449 /// Get the list of external relocations for this code. 1450 pub fn relocs(&self) -> &[MachReloc] { 1451 &self.relocs[..] 1452 } 1453 1454 /// Get the list of trap records for this code. 1455 pub fn traps(&self) -> &[MachTrap] { 1456 &self.traps[..] 1457 } 1458 1459 /// Get the stack map metadata for this code. 1460 pub fn stack_maps(&self) -> &[MachStackMap] { 1461 &self.stack_maps[..] 1462 } 1463 1464 /// Get the list of call sites for this code. 1465 pub fn call_sites(&self) -> &[MachCallSite] { 1466 &self.call_sites[..] 1467 } 1468 } 1469 1470 /// A constant that is deferred to the next constant-pool opportunity. 1471 struct MachLabelConstant { 1472 /// This label will refer to the constant's offset. 1473 label: MachLabel, 1474 /// Required alignment. 1475 align: CodeOffset, 1476 /// This data will be emitted when able. 1477 data: SmallVec<[u8; 16]>, 1478 } 1479 1480 /// A fixup to perform on the buffer once code is emitted. Fixups always refer 1481 /// to labels and patch the code based on label offsets. Hence, they are like 1482 /// relocations, but internal to one buffer. 1483 #[derive(Debug)] 1484 struct MachLabelFixup<I: VCodeInst> { 1485 /// The label whose offset controls this fixup. 1486 label: MachLabel, 1487 /// The offset to fix up / patch to refer to this label. 1488 offset: CodeOffset, 1489 /// The kind of fixup. This is architecture-specific; each architecture may have, 1490 /// e.g., several types of branch instructions, each with differently-sized 1491 /// offset fields and different places within the instruction to place the 1492 /// bits. 1493 kind: I::LabelUse, 1494 } 1495 1496 /// A relocation resulting from a compilation. 1497 #[derive(Clone, Debug, PartialEq)] 1498 #[cfg_attr(feature = "enable-serde", derive(serde::Serialize, serde::Deserialize))] 1499 pub struct MachReloc { 1500 /// The offset at which the relocation applies, *relative to the 1501 /// containing section*. 1502 pub offset: CodeOffset, 1503 /// The kind of relocation. 1504 pub kind: Reloc, 1505 /// The external symbol / name to which this relocation refers. 1506 pub name: ExternalName, 1507 /// The addend to add to the symbol value. 1508 pub addend: i64, 1509 } 1510 1511 /// A trap record resulting from a compilation. 1512 #[derive(Clone, Debug, PartialEq)] 1513 #[cfg_attr(feature = "enable-serde", derive(serde::Serialize, serde::Deserialize))] 1514 pub struct MachTrap { 1515 /// The offset at which the trap instruction occurs, *relative to the 1516 /// containing section*. 1517 pub offset: CodeOffset, 1518 /// The trap code. 1519 pub code: TrapCode, 1520 } 1521 1522 /// A call site record resulting from a compilation. 1523 #[derive(Clone, Debug, PartialEq)] 1524 #[cfg_attr(feature = "enable-serde", derive(serde::Serialize, serde::Deserialize))] 1525 pub struct MachCallSite { 1526 /// The offset of the call's return address, *relative to the containing section*. 1527 pub ret_addr: CodeOffset, 1528 /// The call's opcode. 1529 pub opcode: Opcode, 1530 } 1531 1532 /// A source-location mapping resulting from a compilation. 1533 #[derive(PartialEq, Debug, Clone)] 1534 #[cfg_attr(feature = "enable-serde", derive(serde::Serialize, serde::Deserialize))] 1535 pub struct MachSrcLoc<T: CompilePhase> { 1536 /// The start of the region of code corresponding to a source location. 1537 /// This is relative to the start of the function, not to the start of the 1538 /// section. 1539 pub start: CodeOffset, 1540 /// The end of the region of code corresponding to a source location. 1541 /// This is relative to the start of the section, not to the start of the 1542 /// section. 1543 pub end: CodeOffset, 1544 /// The source location. 1545 pub loc: T::SourceLocType, 1546 } 1547 1548 impl MachSrcLoc<Stencil> { 1549 fn apply_params(self, params: &FunctionParameters) -> MachSrcLoc<Final> { 1550 MachSrcLoc { 1551 start: self.start, 1552 end: self.end, 1553 loc: self.loc.expand(params.base_srcloc()), 1554 } 1555 } 1556 } 1557 1558 /// Record of stack map metadata: stack offsets containing references. 1559 #[derive(Clone, Debug, PartialEq)] 1560 #[cfg_attr(feature = "enable-serde", derive(serde::Serialize, serde::Deserialize))] 1561 pub struct MachStackMap { 1562 /// The code offset at which this stack map applies. 1563 pub offset: CodeOffset, 1564 /// The code offset just past the "end" of the instruction: that is, the 1565 /// offset of the first byte of the following instruction, or equivalently, 1566 /// the start offset plus the instruction length. 1567 pub offset_end: CodeOffset, 1568 /// The stack map itself. 1569 pub stack_map: StackMap, 1570 } 1571 1572 /// Record of branch instruction in the buffer, to facilitate editing. 1573 #[derive(Clone, Debug)] 1574 struct MachBranch { 1575 start: CodeOffset, 1576 end: CodeOffset, 1577 target: MachLabel, 1578 fixup: usize, 1579 inverted: Option<SmallVec<[u8; 8]>>, 1580 /// All labels pointing to the start of this branch. For correctness, this 1581 /// *must* be complete (i.e., must contain all labels whose resolved offsets 1582 /// are at the start of this branch): we rely on being able to redirect all 1583 /// labels that could jump to this branch before removing it, if it is 1584 /// otherwise unreachable. 1585 labels_at_this_branch: SmallVec<[MachLabel; 4]>, 1586 } 1587 1588 impl MachBranch { 1589 fn is_cond(&self) -> bool { 1590 self.inverted.is_some() 1591 } 1592 fn is_uncond(&self) -> bool { 1593 self.inverted.is_none() 1594 } 1595 } 1596 1597 /// Implementation of the `TextSectionBuilder` trait backed by `MachBuffer`. 1598 /// 1599 /// Note that `MachBuffer` was primarily written for intra-function references 1600 /// of jumps between basic blocks, but it's also quite usable for entire text 1601 /// sections and resolving references between functions themselves. This 1602 /// builder interprets "blocks" as labeled functions for the purposes of 1603 /// resolving labels internally in the buffer. 1604 pub struct MachTextSectionBuilder<I: VCodeInst> { 1605 buf: MachBuffer<I>, 1606 next_func: usize, 1607 force_veneers: bool, 1608 } 1609 1610 impl<I: VCodeInst> MachTextSectionBuilder<I> { 1611 pub fn new(num_funcs: u32) -> MachTextSectionBuilder<I> { 1612 let mut buf = MachBuffer::new(); 1613 buf.reserve_labels_for_blocks(num_funcs as usize); 1614 MachTextSectionBuilder { 1615 buf, 1616 next_func: 0, 1617 force_veneers: false, 1618 } 1619 } 1620 } 1621 1622 impl<I: VCodeInst> TextSectionBuilder for MachTextSectionBuilder<I> { 1623 fn append(&mut self, named: bool, func: &[u8], align: Option<u32>) -> u64 { 1624 // Conditionally emit an island if it's necessary to resolve jumps 1625 // between functions which are too far away. 1626 let size = func.len() as u32; 1627 if self.force_veneers || self.buf.island_needed(size) { 1628 self.buf.emit_island_maybe_forced(self.force_veneers, size); 1629 } 1630 1631 self.buf.align_to(align.unwrap_or(I::LabelUse::ALIGN)); 1632 let pos = self.buf.cur_offset(); 1633 if named { 1634 self.buf 1635 .bind_label(MachLabel::from_block(BlockIndex::new(self.next_func))); 1636 self.next_func += 1; 1637 } 1638 self.buf.put_data(func); 1639 u64::from(pos) 1640 } 1641 1642 fn resolve_reloc(&mut self, offset: u64, reloc: Reloc, addend: Addend, target: u32) -> bool { 1643 let label = MachLabel::from_block(BlockIndex::new(target as usize)); 1644 let offset = u32::try_from(offset).unwrap(); 1645 match I::LabelUse::from_reloc(reloc, addend) { 1646 Some(label_use) => { 1647 self.buf.use_label_at_offset(offset, label, label_use); 1648 true 1649 } 1650 None => false, 1651 } 1652 } 1653 1654 fn force_veneers(&mut self) { 1655 self.force_veneers = true; 1656 } 1657 1658 fn finish(&mut self) -> Vec<u8> { 1659 // Double-check all functions were pushed. 1660 assert_eq!(self.next_func, self.buf.label_offsets.len()); 1661 1662 // Finish up any veneers, if necessary. 1663 self.buf 1664 .finish_emission_maybe_forcing_veneers(self.force_veneers); 1665 1666 // We don't need the data any more, so return it to the caller. 1667 mem::take(&mut self.buf.data).into_vec() 1668 } 1669 } 1670 1671 // We use an actual instruction definition to do tests, so we depend on the `arm64` feature here. 1672 #[cfg(all(test, feature = "arm64"))] 1673 mod test { 1674 use cranelift_entity::EntityRef as _; 1675 1676 use super::*; 1677 use crate::ir::UserExternalNameRef; 1678 use crate::isa::aarch64::inst::xreg; 1679 use crate::isa::aarch64::inst::{BranchTarget, CondBrKind, EmitInfo, Inst}; 1680 use crate::machinst::MachInstEmit; 1681 use crate::settings; 1682 use std::default::Default; 1683 use std::vec::Vec; 1684 1685 fn label(n: u32) -> MachLabel { 1686 MachLabel::from_block(BlockIndex::new(n as usize)) 1687 } 1688 fn target(n: u32) -> BranchTarget { 1689 BranchTarget::Label(label(n)) 1690 } 1691 1692 #[test] 1693 fn test_elide_jump_to_next() { 1694 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 1695 let mut buf = MachBuffer::new(); 1696 let mut state = Default::default(); 1697 1698 buf.reserve_labels_for_blocks(2); 1699 buf.bind_label(label(0)); 1700 let inst = Inst::Jump { dest: target(1) }; 1701 inst.emit(&[], &mut buf, &info, &mut state); 1702 buf.bind_label(label(1)); 1703 let buf = buf.finish(); 1704 assert_eq!(0, buf.total_size()); 1705 } 1706 1707 #[test] 1708 fn test_elide_trivial_jump_blocks() { 1709 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 1710 let mut buf = MachBuffer::new(); 1711 let mut state = Default::default(); 1712 1713 buf.reserve_labels_for_blocks(4); 1714 1715 buf.bind_label(label(0)); 1716 let inst = Inst::CondBr { 1717 kind: CondBrKind::NotZero(xreg(0)), 1718 taken: target(1), 1719 not_taken: target(2), 1720 }; 1721 inst.emit(&[], &mut buf, &info, &mut state); 1722 1723 buf.bind_label(label(1)); 1724 let inst = Inst::Jump { dest: target(3) }; 1725 inst.emit(&[], &mut buf, &info, &mut state); 1726 1727 buf.bind_label(label(2)); 1728 let inst = Inst::Jump { dest: target(3) }; 1729 inst.emit(&[], &mut buf, &info, &mut state); 1730 1731 buf.bind_label(label(3)); 1732 1733 let buf = buf.finish(); 1734 assert_eq!(0, buf.total_size()); 1735 } 1736 1737 #[test] 1738 fn test_flip_cond() { 1739 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 1740 let mut buf = MachBuffer::new(); 1741 let mut state = Default::default(); 1742 1743 buf.reserve_labels_for_blocks(4); 1744 1745 buf.bind_label(label(0)); 1746 let inst = Inst::CondBr { 1747 kind: CondBrKind::NotZero(xreg(0)), 1748 taken: target(1), 1749 not_taken: target(2), 1750 }; 1751 inst.emit(&[], &mut buf, &info, &mut state); 1752 1753 buf.bind_label(label(1)); 1754 let inst = Inst::Udf { 1755 trap_code: TrapCode::Interrupt, 1756 }; 1757 inst.emit(&[], &mut buf, &info, &mut state); 1758 1759 buf.bind_label(label(2)); 1760 let inst = Inst::Nop4; 1761 inst.emit(&[], &mut buf, &info, &mut state); 1762 1763 buf.bind_label(label(3)); 1764 1765 let buf = buf.finish(); 1766 1767 let mut buf2 = MachBuffer::new(); 1768 let mut state = Default::default(); 1769 let inst = Inst::TrapIf { 1770 kind: CondBrKind::NotZero(xreg(0)), 1771 trap_code: TrapCode::Interrupt, 1772 }; 1773 inst.emit(&[], &mut buf2, &info, &mut state); 1774 let inst = Inst::Nop4; 1775 inst.emit(&[], &mut buf2, &info, &mut state); 1776 1777 let buf2 = buf2.finish(); 1778 1779 assert_eq!(buf.data, buf2.data); 1780 } 1781 1782 #[test] 1783 fn test_island() { 1784 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 1785 let mut buf = MachBuffer::new(); 1786 let mut state = Default::default(); 1787 1788 buf.reserve_labels_for_blocks(4); 1789 1790 buf.bind_label(label(0)); 1791 let inst = Inst::CondBr { 1792 kind: CondBrKind::NotZero(xreg(0)), 1793 taken: target(2), 1794 not_taken: target(3), 1795 }; 1796 inst.emit(&[], &mut buf, &info, &mut state); 1797 1798 buf.bind_label(label(1)); 1799 while buf.cur_offset() < 2000000 { 1800 if buf.island_needed(0) { 1801 buf.emit_island(0); 1802 } 1803 let inst = Inst::Nop4; 1804 inst.emit(&[], &mut buf, &info, &mut state); 1805 } 1806 1807 buf.bind_label(label(2)); 1808 let inst = Inst::Nop4; 1809 inst.emit(&[], &mut buf, &info, &mut state); 1810 1811 buf.bind_label(label(3)); 1812 let inst = Inst::Nop4; 1813 inst.emit(&[], &mut buf, &info, &mut state); 1814 1815 let buf = buf.finish(); 1816 1817 assert_eq!(2000000 + 8, buf.total_size()); 1818 1819 let mut buf2 = MachBuffer::new(); 1820 let mut state = Default::default(); 1821 let inst = Inst::CondBr { 1822 kind: CondBrKind::NotZero(xreg(0)), 1823 1824 // This conditionally taken branch has a 19-bit constant, shifted 1825 // to the left by two, giving us a 21-bit range in total. Half of 1826 // this range positive so the we should be around 1 << 20 bytes 1827 // away for our jump target. 1828 // 1829 // There are two pending fixups by the time we reach this point, 1830 // one for this 19-bit jump and one for the unconditional 26-bit 1831 // jump below. A 19-bit veneer is 4 bytes large and the 26-bit 1832 // veneer is 20 bytes large, which means that pessimistically 1833 // assuming we'll need two veneers we need 24 bytes of extra 1834 // space, meaning that the actual island should come 24-bytes 1835 // before the deadline. 1836 taken: BranchTarget::ResolvedOffset((1 << 20) - 4 - 20), 1837 1838 // This branch is in-range so no veneers should be needed, it should 1839 // go directly to the target. 1840 not_taken: BranchTarget::ResolvedOffset(2000000 + 4 - 4), 1841 }; 1842 inst.emit(&[], &mut buf2, &info, &mut state); 1843 1844 let buf2 = buf2.finish(); 1845 1846 assert_eq!(&buf.data[0..8], &buf2.data[..]); 1847 } 1848 1849 #[test] 1850 fn test_island_backward() { 1851 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 1852 let mut buf = MachBuffer::new(); 1853 let mut state = Default::default(); 1854 1855 buf.reserve_labels_for_blocks(4); 1856 1857 buf.bind_label(label(0)); 1858 let inst = Inst::Nop4; 1859 inst.emit(&[], &mut buf, &info, &mut state); 1860 1861 buf.bind_label(label(1)); 1862 let inst = Inst::Nop4; 1863 inst.emit(&[], &mut buf, &info, &mut state); 1864 1865 buf.bind_label(label(2)); 1866 while buf.cur_offset() < 2000000 { 1867 let inst = Inst::Nop4; 1868 inst.emit(&[], &mut buf, &info, &mut state); 1869 } 1870 1871 buf.bind_label(label(3)); 1872 let inst = Inst::CondBr { 1873 kind: CondBrKind::NotZero(xreg(0)), 1874 taken: target(0), 1875 not_taken: target(1), 1876 }; 1877 inst.emit(&[], &mut buf, &info, &mut state); 1878 1879 let buf = buf.finish(); 1880 1881 assert_eq!(2000000 + 12, buf.total_size()); 1882 1883 let mut buf2 = MachBuffer::new(); 1884 let mut state = Default::default(); 1885 let inst = Inst::CondBr { 1886 kind: CondBrKind::NotZero(xreg(0)), 1887 taken: BranchTarget::ResolvedOffset(8), 1888 not_taken: BranchTarget::ResolvedOffset(4 - (2000000 + 4)), 1889 }; 1890 inst.emit(&[], &mut buf2, &info, &mut state); 1891 let inst = Inst::Jump { 1892 dest: BranchTarget::ResolvedOffset(-(2000000 + 8)), 1893 }; 1894 inst.emit(&[], &mut buf2, &info, &mut state); 1895 1896 let buf2 = buf2.finish(); 1897 1898 assert_eq!(&buf.data[2000000..], &buf2.data[..]); 1899 } 1900 1901 #[test] 1902 fn test_multiple_redirect() { 1903 // label0: 1904 // cbz x0, label1 1905 // b label2 1906 // label1: 1907 // b label3 1908 // label2: 1909 // nop 1910 // nop 1911 // b label0 1912 // label3: 1913 // b label4 1914 // label4: 1915 // b label5 1916 // label5: 1917 // b label7 1918 // label6: 1919 // nop 1920 // label7: 1921 // ret 1922 // 1923 // -- should become: 1924 // 1925 // label0: 1926 // cbz x0, label7 1927 // label2: 1928 // nop 1929 // nop 1930 // b label0 1931 // label6: 1932 // nop 1933 // label7: 1934 // ret 1935 1936 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 1937 let mut buf = MachBuffer::new(); 1938 let mut state = Default::default(); 1939 1940 buf.reserve_labels_for_blocks(8); 1941 1942 buf.bind_label(label(0)); 1943 let inst = Inst::CondBr { 1944 kind: CondBrKind::Zero(xreg(0)), 1945 taken: target(1), 1946 not_taken: target(2), 1947 }; 1948 inst.emit(&[], &mut buf, &info, &mut state); 1949 1950 buf.bind_label(label(1)); 1951 let inst = Inst::Jump { dest: target(3) }; 1952 inst.emit(&[], &mut buf, &info, &mut state); 1953 1954 buf.bind_label(label(2)); 1955 let inst = Inst::Nop4; 1956 inst.emit(&[], &mut buf, &info, &mut state); 1957 inst.emit(&[], &mut buf, &info, &mut state); 1958 let inst = Inst::Jump { dest: target(0) }; 1959 inst.emit(&[], &mut buf, &info, &mut state); 1960 1961 buf.bind_label(label(3)); 1962 let inst = Inst::Jump { dest: target(4) }; 1963 inst.emit(&[], &mut buf, &info, &mut state); 1964 1965 buf.bind_label(label(4)); 1966 let inst = Inst::Jump { dest: target(5) }; 1967 inst.emit(&[], &mut buf, &info, &mut state); 1968 1969 buf.bind_label(label(5)); 1970 let inst = Inst::Jump { dest: target(7) }; 1971 inst.emit(&[], &mut buf, &info, &mut state); 1972 1973 buf.bind_label(label(6)); 1974 let inst = Inst::Nop4; 1975 inst.emit(&[], &mut buf, &info, &mut state); 1976 1977 buf.bind_label(label(7)); 1978 let inst = Inst::Ret { rets: vec![] }; 1979 inst.emit(&[], &mut buf, &info, &mut state); 1980 1981 let buf = buf.finish(); 1982 1983 let golden_data = vec![ 1984 0xa0, 0x00, 0x00, 0xb4, // cbz x0, 0x14 1985 0x1f, 0x20, 0x03, 0xd5, // nop 1986 0x1f, 0x20, 0x03, 0xd5, // nop 1987 0xfd, 0xff, 0xff, 0x17, // b 0 1988 0x1f, 0x20, 0x03, 0xd5, // nop 1989 0xc0, 0x03, 0x5f, 0xd6, // ret 1990 ]; 1991 1992 assert_eq!(&golden_data[..], &buf.data[..]); 1993 } 1994 1995 #[test] 1996 fn test_handle_branch_cycle() { 1997 // label0: 1998 // b label1 1999 // label1: 2000 // b label2 2001 // label2: 2002 // b label3 2003 // label3: 2004 // b label4 2005 // label4: 2006 // b label1 // note: not label0 (to make it interesting). 2007 // 2008 // -- should become: 2009 // 2010 // label0, label1, ..., label4: 2011 // b label0 2012 let info = EmitInfo::new(settings::Flags::new(settings::builder())); 2013 let mut buf = MachBuffer::new(); 2014 let mut state = Default::default(); 2015 2016 buf.reserve_labels_for_blocks(5); 2017 2018 buf.bind_label(label(0)); 2019 let inst = Inst::Jump { dest: target(1) }; 2020 inst.emit(&[], &mut buf, &info, &mut state); 2021 2022 buf.bind_label(label(1)); 2023 let inst = Inst::Jump { dest: target(2) }; 2024 inst.emit(&[], &mut buf, &info, &mut state); 2025 2026 buf.bind_label(label(2)); 2027 let inst = Inst::Jump { dest: target(3) }; 2028 inst.emit(&[], &mut buf, &info, &mut state); 2029 2030 buf.bind_label(label(3)); 2031 let inst = Inst::Jump { dest: target(4) }; 2032 inst.emit(&[], &mut buf, &info, &mut state); 2033 2034 buf.bind_label(label(4)); 2035 let inst = Inst::Jump { dest: target(1) }; 2036 inst.emit(&[], &mut buf, &info, &mut state); 2037 2038 let buf = buf.finish(); 2039 2040 let golden_data = vec![ 2041 0x00, 0x00, 0x00, 0x14, // b 0 2042 ]; 2043 2044 assert_eq!(&golden_data[..], &buf.data[..]); 2045 } 2046 2047 #[test] 2048 fn metadata_records() { 2049 let mut buf = MachBuffer::<Inst>::new(); 2050 2051 buf.reserve_labels_for_blocks(1); 2052 2053 buf.bind_label(label(0)); 2054 buf.put1(1); 2055 buf.add_trap(TrapCode::HeapOutOfBounds); 2056 buf.put1(2); 2057 buf.add_trap(TrapCode::IntegerOverflow); 2058 buf.add_trap(TrapCode::IntegerDivisionByZero); 2059 buf.add_call_site(Opcode::Call); 2060 buf.add_reloc( 2061 Reloc::Abs4, 2062 &ExternalName::User(UserExternalNameRef::new(0)), 2063 0, 2064 ); 2065 buf.put1(3); 2066 buf.add_reloc( 2067 Reloc::Abs8, 2068 &ExternalName::User(UserExternalNameRef::new(1)), 2069 1, 2070 ); 2071 buf.put1(4); 2072 2073 let buf = buf.finish(); 2074 2075 assert_eq!(buf.data(), &[1, 2, 3, 4]); 2076 assert_eq!( 2077 buf.traps() 2078 .iter() 2079 .map(|trap| (trap.offset, trap.code)) 2080 .collect::<Vec<_>>(), 2081 vec![ 2082 (1, TrapCode::HeapOutOfBounds), 2083 (2, TrapCode::IntegerOverflow), 2084 (2, TrapCode::IntegerDivisionByZero) 2085 ] 2086 ); 2087 assert_eq!( 2088 buf.call_sites() 2089 .iter() 2090 .map(|call_site| (call_site.ret_addr, call_site.opcode)) 2091 .collect::<Vec<_>>(), 2092 vec![(2, Opcode::Call)] 2093 ); 2094 assert_eq!( 2095 buf.relocs() 2096 .iter() 2097 .map(|reloc| (reloc.offset, reloc.kind)) 2098 .collect::<Vec<_>>(), 2099 vec![(2, Reloc::Abs4), (3, Reloc::Abs8)] 2100 ); 2101 } 2102 } 2103