1 //! Elaboration phase: lowers EGraph back to sequences of operations 2 //! in CFG nodes. 3 4 use super::Stats; 5 use super::cost::Cost; 6 use crate::ctxhash::NullCtx; 7 use crate::dominator_tree::DominatorTree; 8 use crate::hash_map::Entry as HashEntry; 9 use crate::inst_predicates::is_pure_for_egraph; 10 use crate::ir::{Block, Function, Inst, Value, ValueDef}; 11 use crate::loop_analysis::{Loop, LoopAnalysis}; 12 use crate::scoped_hash_map::ScopedHashMap; 13 use crate::trace; 14 use alloc::vec::Vec; 15 use cranelift_control::ControlPlane; 16 use cranelift_entity::{SecondaryMap, packed_option::ReservedValue}; 17 use rustc_hash::{FxHashMap, FxHashSet}; 18 use smallvec::{SmallVec, smallvec}; 19 20 pub(crate) struct Elaborator<'a> { 21 func: &'a mut Function, 22 domtree: &'a DominatorTree, 23 loop_analysis: &'a LoopAnalysis, 24 /// Map from Value that is produced by a pure Inst (and was thus 25 /// not in the side-effecting skeleton) to the value produced by 26 /// an elaborated inst (placed in the layout) to whose results we 27 /// refer in the final code. 28 /// 29 /// The first time we use some result of an instruction during 30 /// elaboration, we can place it and insert an identity map (inst 31 /// results to that same inst's results) in this scoped 32 /// map. Within that block and its dom-tree children, that mapping 33 /// is visible and we can continue to use it. This allows us to 34 /// avoid cloning the instruction. However, if we pop that scope 35 /// and use it somewhere else as well, we will need to 36 /// duplicate. We detect this case by checking, when a value that 37 /// we want is not present in this map, whether the producing inst 38 /// is already placed in the Layout. If so, we duplicate, and 39 /// insert non-identity mappings from the original inst's results 40 /// to the cloned inst's results. 41 /// 42 /// Note that as values may refer to unions that represent a subset 43 /// of a larger eclass, it's not valid to walk towards the root of a 44 /// union tree: doing so would potentially equate values that fall 45 /// on different branches of the dominator tree. 46 value_to_elaborated_value: ScopedHashMap<Value, ElaboratedValue>, 47 /// Map from Value to the best (lowest-cost) Value in its eclass 48 /// (tree of union value-nodes). 49 value_to_best_value: SecondaryMap<Value, BestEntry>, 50 /// Stack of blocks and loops in current elaboration path. 51 loop_stack: SmallVec<[LoopStackEntry; 8]>, 52 /// The current block into which we are elaborating. 53 cur_block: Block, 54 /// Values that opt rules have indicated should be rematerialized 55 /// in every block they are used (e.g., immediates or other 56 /// "cheap-to-compute" ops). 57 remat_values: &'a FxHashSet<Value>, 58 /// Explicitly-unrolled value elaboration stack. 59 elab_stack: Vec<ElabStackEntry>, 60 /// Results from the elab stack. 61 elab_result_stack: Vec<ElaboratedValue>, 62 /// Explicitly-unrolled block elaboration stack. 63 block_stack: Vec<BlockStackEntry>, 64 /// Copies of values that have been rematerialized. 65 remat_copies: FxHashMap<(Block, Value), Value>, 66 /// Stats for various events during egraph processing, to help 67 /// with optimization of this infrastructure. 68 stats: &'a mut Stats, 69 /// Chaos-mode control-plane so we can test that we still get 70 /// correct results when our heuristics make bad decisions. 71 ctrl_plane: &'a mut ControlPlane, 72 } 73 74 #[derive(Clone, Copy, Debug, PartialEq, Eq)] 75 struct BestEntry(Cost, Value); 76 77 impl PartialOrd for BestEntry { 78 fn partial_cmp(&self, other: &Self) -> Option<core::cmp::Ordering> { 79 Some(self.cmp(other)) 80 } 81 } 82 83 impl Ord for BestEntry { 84 #[inline] 85 fn cmp(&self, other: &Self) -> std::cmp::Ordering { 86 self.0.cmp(&other.0).then_with(|| { 87 // Note that this comparison is reversed. When costs are equal, 88 // prefer the value with the bigger index. This is a heuristic that 89 // prefers results of rewrites to the original value, since we 90 // expect that our rewrites are generally improvements. 91 self.1.cmp(&other.1).reverse() 92 }) 93 } 94 } 95 96 #[derive(Clone, Copy, Debug)] 97 struct ElaboratedValue { 98 in_block: Block, 99 value: Value, 100 } 101 102 #[derive(Clone, Debug)] 103 struct LoopStackEntry { 104 /// The loop identifier. 105 lp: Loop, 106 /// The hoist point: a block that immediately dominates this 107 /// loop. May not be an immediate predecessor, but will be a valid 108 /// point to place all loop-invariant ops: they must depend only 109 /// on inputs that dominate the loop, so are available at (the end 110 /// of) this block. 111 hoist_block: Block, 112 /// The depth in the scope map. 113 scope_depth: u32, 114 } 115 116 #[derive(Clone, Debug)] 117 enum ElabStackEntry { 118 /// Next action is to resolve this value into an elaborated inst 119 /// (placed into the layout) that produces the value, and 120 /// recursively elaborate the insts that produce its args. 121 /// 122 /// Any inserted ops should be inserted before `before`, which is 123 /// the instruction demanding this value. 124 Start { value: Value, before: Inst }, 125 /// Args have been pushed; waiting for results. 126 PendingInst { 127 inst: Inst, 128 result_idx: usize, 129 num_args: usize, 130 before: Inst, 131 }, 132 } 133 134 #[derive(Clone, Debug)] 135 enum BlockStackEntry { 136 Elaborate { block: Block, idom: Option<Block> }, 137 Pop, 138 } 139 140 impl<'a> Elaborator<'a> { 141 pub(crate) fn new( 142 func: &'a mut Function, 143 domtree: &'a DominatorTree, 144 loop_analysis: &'a LoopAnalysis, 145 remat_values: &'a FxHashSet<Value>, 146 stats: &'a mut Stats, 147 ctrl_plane: &'a mut ControlPlane, 148 ) -> Self { 149 let num_values = func.dfg.num_values(); 150 let mut value_to_best_value = 151 SecondaryMap::with_default(BestEntry(Cost::infinity(), Value::reserved_value())); 152 value_to_best_value.resize(num_values); 153 Self { 154 func, 155 domtree, 156 loop_analysis, 157 value_to_elaborated_value: ScopedHashMap::with_capacity(num_values), 158 value_to_best_value, 159 loop_stack: smallvec![], 160 cur_block: Block::reserved_value(), 161 remat_values, 162 elab_stack: vec![], 163 elab_result_stack: vec![], 164 block_stack: vec![], 165 remat_copies: FxHashMap::default(), 166 stats, 167 ctrl_plane, 168 } 169 } 170 171 fn start_block(&mut self, idom: Option<Block>, block: Block) { 172 trace!( 173 "start_block: block {:?} with idom {:?} at loop depth {:?} scope depth {}", 174 block, 175 idom, 176 self.loop_stack.len(), 177 self.value_to_elaborated_value.depth() 178 ); 179 180 // Pop any loop levels we're no longer in. 181 while let Some(inner_loop) = self.loop_stack.last() { 182 if self.loop_analysis.is_in_loop(block, inner_loop.lp) { 183 break; 184 } 185 self.loop_stack.pop(); 186 } 187 188 // Note that if the *entry* block is a loop header, we will 189 // not make note of the loop here because it will not have an 190 // immediate dominator. We must disallow this case because we 191 // will skip adding the `LoopStackEntry` here but our 192 // `LoopAnalysis` will otherwise still make note of this loop 193 // and loop depths will not match. 194 if let Some(idom) = idom { 195 if let Some(lp) = self.loop_analysis.is_loop_header(block) { 196 self.loop_stack.push(LoopStackEntry { 197 lp, 198 // Any code hoisted out of this loop will have code 199 // placed in `idom`, and will have def mappings 200 // inserted in to the scoped hashmap at that block's 201 // level. 202 hoist_block: idom, 203 scope_depth: (self.value_to_elaborated_value.depth() - 1) as u32, 204 }); 205 trace!( 206 " -> loop header, pushing; depth now {}", 207 self.loop_stack.len() 208 ); 209 } 210 } else { 211 debug_assert!( 212 self.loop_analysis.is_loop_header(block).is_none(), 213 "Entry block (domtree root) cannot be a loop header!" 214 ); 215 } 216 217 trace!("block {}: loop stack is {:?}", block, self.loop_stack); 218 219 self.cur_block = block; 220 } 221 222 fn compute_best_values(&mut self) { 223 let best = &mut self.value_to_best_value; 224 225 // We can't make random decisions inside the fixpoint loop below because 226 // that could cause values to change on every iteration of the loop, 227 // which would make the loop never terminate. So in chaos testing 228 // mode we need a form of making suboptimal decisions that is fully 229 // deterministic. We choose to simply make the worst decision we know 230 // how to do instead of the best. 231 let use_worst = self.ctrl_plane.get_decision(); 232 233 // Do a fixpoint loop to compute the best value for each eclass. 234 // 235 // The maximum number of iterations is the length of the longest chain 236 // of `vNN -> vMM` edges in the dataflow graph where `NN < MM`, so this 237 // is *technically* quadratic, but `cranelift-frontend` won't construct 238 // any such edges. NaN canonicalization will introduce some of these 239 // edges, but they are chains of only two or three edges. So in 240 // practice, we *never* do more than a handful of iterations here unless 241 // (a) we parsed the CLIF from text and the text was funkily numbered, 242 // which we don't really care about, or (b) the CLIF producer did 243 // something weird, in which case it is their responsibility to stop 244 // doing that. 245 trace!( 246 "Entering fixpoint loop to compute the {} values for each eclass", 247 if use_worst { 248 "worst (chaos mode)" 249 } else { 250 "best" 251 } 252 ); 253 let mut keep_going = true; 254 while keep_going { 255 keep_going = false; 256 trace!( 257 "fixpoint iteration {}", 258 self.stats.elaborate_best_cost_fixpoint_iters 259 ); 260 self.stats.elaborate_best_cost_fixpoint_iters += 1; 261 262 for (value, def) in self.func.dfg.values_and_defs() { 263 trace!("computing best for value {:?} def {:?}", value, def); 264 let orig_best_value = best[value]; 265 266 match def { 267 ValueDef::Union(x, y) => { 268 // Pick the best of the two options based on 269 // min-cost. This works because each element of `best` 270 // is a `(cost, value)` tuple; `cost` comes first so 271 // the natural comparison works based on cost, and 272 // breaks ties based on value number. 273 best[value] = if use_worst { 274 if best[x].1.is_reserved_value() { 275 best[y] 276 } else if best[y].1.is_reserved_value() { 277 best[x] 278 } else { 279 std::cmp::max(best[x], best[y]) 280 } 281 } else { 282 std::cmp::min(best[x], best[y]) 283 }; 284 trace!( 285 " -> best of union({:?}, {:?}) = {:?}", 286 best[x], best[y], best[value] 287 ); 288 } 289 ValueDef::Param(_, _) => { 290 best[value] = BestEntry(Cost::zero(), value); 291 } 292 // If the Inst is inserted into the layout (which is, 293 // at this point, only the side-effecting skeleton), 294 // then it must be computed and thus we give it zero 295 // cost. 296 ValueDef::Result(inst, _) => { 297 if let Some(_) = self.func.layout.inst_block(inst) { 298 best[value] = BestEntry(Cost::zero(), value); 299 } else { 300 let inst_data = &self.func.dfg.insts[inst]; 301 // N.B.: at this point we know that the opcode is 302 // pure, so `pure_op_cost`'s precondition is 303 // satisfied. 304 let cost = Cost::of_pure_op( 305 inst_data.opcode(), 306 self.func.dfg.inst_values(inst).map(|value| best[value].0), 307 ); 308 best[value] = BestEntry(cost, value); 309 trace!(" -> cost of value {} = {:?}", value, cost); 310 } 311 } 312 }; 313 314 // Keep on iterating the fixpoint loop while we are finding new 315 // best values. 316 keep_going |= orig_best_value != best[value]; 317 } 318 } 319 320 if cfg!(any(feature = "trace-log", debug_assertions)) { 321 trace!("finished fixpoint loop to compute best value for each eclass"); 322 for value in self.func.dfg.values() { 323 trace!("-> best for eclass {:?}: {:?}", value, best[value]); 324 debug_assert_ne!(best[value].1, Value::reserved_value()); 325 // You might additionally be expecting an assert that the best 326 // cost is not infinity, however infinite cost *can* happen in 327 // practice. First, note that our cost function doesn't know 328 // about any shared structure in the dataflow graph, it only 329 // sums operand costs. (And trying to avoid that by deduping a 330 // single operation's operands is a losing game because you can 331 // always just add one indirection and go from `add(x, x)` to 332 // `add(foo(x), bar(x))` to hide the shared structure.) Given 333 // that blindness to sharing, we can make cost grow 334 // exponentially with a linear sequence of operations: 335 // 336 // v0 = iconst.i32 1 ;; cost = 1 337 // v1 = iadd v0, v0 ;; cost = 3 + 1 + 1 338 // v2 = iadd v1, v1 ;; cost = 3 + 5 + 5 339 // v3 = iadd v2, v2 ;; cost = 3 + 13 + 13 340 // v4 = iadd v3, v3 ;; cost = 3 + 29 + 29 341 // v5 = iadd v4, v4 ;; cost = 3 + 61 + 61 342 // v6 = iadd v5, v5 ;; cost = 3 + 125 + 125 343 // ;; etc... 344 // 345 // Such a chain can cause cost to saturate to infinity. How do 346 // we choose which e-node is best when there are multiple that 347 // have saturated to infinity? It doesn't matter. As long as 348 // invariant (2) for optimization rules is upheld by our rule 349 // set (see `cranelift/codegen/src/opts/README.md`) it is safe 350 // to choose *any* e-node in the e-class. At worst we will 351 // produce suboptimal code, but never an incorrectness. 352 } 353 } 354 } 355 356 /// Elaborate use of an eclass, inserting any needed new 357 /// instructions before the given inst `before`. Should only be 358 /// given values corresponding to results of instructions or 359 /// blockparams. 360 fn elaborate_eclass_use(&mut self, value: Value, before: Inst) -> ElaboratedValue { 361 debug_assert_ne!(value, Value::reserved_value()); 362 363 // Kick off the process by requesting this result 364 // value. 365 self.elab_stack 366 .push(ElabStackEntry::Start { value, before }); 367 368 // Now run the explicit-stack recursion until we reach 369 // the root. 370 self.process_elab_stack(); 371 debug_assert_eq!(self.elab_result_stack.len(), 1); 372 self.elab_result_stack.pop().unwrap() 373 } 374 375 /// Possibly rematerialize the instruction producing the value in 376 /// `arg` and rewrite `arg` to refer to it, if needed. Returns 377 /// `true` if a rewrite occurred. 378 fn maybe_remat_arg( 379 remat_values: &FxHashSet<Value>, 380 func: &mut Function, 381 remat_copies: &mut FxHashMap<(Block, Value), Value>, 382 insert_block: Block, 383 before: Inst, 384 arg: &mut ElaboratedValue, 385 stats: &mut Stats, 386 ) -> bool { 387 // TODO (#7313): we may want to consider recursive 388 // rematerialization as well. We could process the arguments of 389 // the rematerialized instruction up to a certain depth. This 390 // would affect, e.g., adds-with-one-constant-arg, which are 391 // currently rematerialized. Right now we don't do this, to 392 // avoid the need for another fixpoint loop here. 393 if arg.in_block != insert_block && remat_values.contains(&arg.value) { 394 let new_value = match remat_copies.entry((insert_block, arg.value)) { 395 HashEntry::Occupied(o) => *o.get(), 396 HashEntry::Vacant(v) => { 397 let inst = func.dfg.value_def(arg.value).inst().unwrap(); 398 debug_assert_eq!(func.dfg.inst_results(inst).len(), 1); 399 let new_inst = func.dfg.clone_inst(inst); 400 func.layout.insert_inst(new_inst, before); 401 let new_result = func.dfg.inst_results(new_inst)[0]; 402 *v.insert(new_result) 403 } 404 }; 405 trace!("rematerialized {} as {}", arg.value, new_value); 406 arg.value = new_value; 407 stats.elaborate_remat += 1; 408 true 409 } else { 410 false 411 } 412 } 413 414 fn process_elab_stack(&mut self) { 415 while let Some(entry) = self.elab_stack.pop() { 416 match entry { 417 ElabStackEntry::Start { value, before } => { 418 debug_assert!(self.func.dfg.value_is_real(value)); 419 420 self.stats.elaborate_visit_node += 1; 421 422 // Get the best option; we use `value` (latest 423 // value) here so we have a full view of the 424 // eclass. 425 trace!("looking up best value for {}", value); 426 let BestEntry(_, best_value) = self.value_to_best_value[value]; 427 trace!("elaborate: value {} -> best {}", value, best_value); 428 debug_assert_ne!(best_value, Value::reserved_value()); 429 430 if let Some(elab_val) = 431 self.value_to_elaborated_value.get(&NullCtx, &best_value) 432 { 433 // Value is available; use it. 434 trace!("elaborate: value {} -> {:?}", value, elab_val); 435 self.stats.elaborate_memoize_hit += 1; 436 self.elab_result_stack.push(*elab_val); 437 continue; 438 } 439 440 self.stats.elaborate_memoize_miss += 1; 441 442 // Now resolve the value to its definition to see 443 // how we can compute it. 444 let (inst, result_idx) = match self.func.dfg.value_def(best_value) { 445 ValueDef::Result(inst, result_idx) => { 446 trace!( 447 " -> value {} is result {} of {}", 448 best_value, result_idx, inst 449 ); 450 (inst, result_idx) 451 } 452 ValueDef::Param(in_block, _) => { 453 // We don't need to do anything to compute 454 // this value; just push its result on the 455 // result stack (blockparams are already 456 // available). 457 trace!(" -> value {} is a blockparam", best_value); 458 self.elab_result_stack.push(ElaboratedValue { 459 in_block, 460 value: best_value, 461 }); 462 continue; 463 } 464 ValueDef::Union(_, _) => { 465 panic!("Should never have a Union value as the best value"); 466 } 467 }; 468 469 trace!( 470 " -> result {} of inst {:?}", 471 result_idx, self.func.dfg.insts[inst] 472 ); 473 474 // We're going to need to use this instruction 475 // result, placing the instruction into the 476 // layout. First, enqueue all args to be 477 // elaborated. Push state to receive the results 478 // and later elab this inst. 479 let num_args = self.func.dfg.inst_values(inst).count(); 480 self.elab_stack.push(ElabStackEntry::PendingInst { 481 inst, 482 result_idx, 483 num_args, 484 before, 485 }); 486 487 // Push args in reverse order so we process the 488 // first arg first. 489 for arg in self.func.dfg.inst_values(inst).rev() { 490 debug_assert_ne!(arg, Value::reserved_value()); 491 self.elab_stack 492 .push(ElabStackEntry::Start { value: arg, before }); 493 } 494 } 495 496 ElabStackEntry::PendingInst { 497 inst, 498 result_idx, 499 num_args, 500 before, 501 } => { 502 trace!( 503 "PendingInst: {} result {} args {} before {}", 504 inst, result_idx, num_args, before 505 ); 506 507 // We should have all args resolved at this 508 // point. Grab them and drain them out, removing 509 // them. 510 let arg_idx = self.elab_result_stack.len() - num_args; 511 let arg_values = &mut self.elab_result_stack[arg_idx..]; 512 513 // Compute max loop depth. 514 // 515 // Note that if there are no arguments then this instruction 516 // is allowed to get hoisted up one loop. This is not 517 // usually used since no-argument values are things like 518 // constants which are typically rematerialized, but for the 519 // `vconst` instruction 128-bit constants aren't as easily 520 // rematerialized. They're hoisted out of inner loops but 521 // not to the function entry which may run the risk of 522 // placing too much register pressure on the entire 523 // function. This is modeled with the `.saturating_sub(1)` 524 // as the default if there's otherwise no maximum. 525 let loop_hoist_level = arg_values 526 .iter() 527 .map(|&value| { 528 // Find the outermost loop level at which 529 // the value's defining block *is not* a 530 // member. This is the loop-nest level 531 // whose hoist-block we hoist to. 532 let hoist_level = self 533 .loop_stack 534 .iter() 535 .position(|loop_entry| { 536 !self.loop_analysis.is_in_loop(value.in_block, loop_entry.lp) 537 }) 538 .unwrap_or(self.loop_stack.len()); 539 trace!( 540 " -> arg: elab_value {:?} hoist level {:?}", 541 value, hoist_level 542 ); 543 hoist_level 544 }) 545 .max() 546 .unwrap_or(self.loop_stack.len().saturating_sub(1)); 547 trace!( 548 " -> loop hoist level: {:?}; cur loop depth: {:?}, loop_stack: {:?}", 549 loop_hoist_level, 550 self.loop_stack.len(), 551 self.loop_stack, 552 ); 553 554 // We know that this is a pure inst, because 555 // non-pure roots have already been placed in the 556 // value-to-elab'd-value map, so they will not 557 // reach this stage of processing. 558 // 559 // We now must determine the location at which we 560 // place the instruction. This is the current 561 // block *unless* we hoist above a loop when all 562 // args are loop-invariant (and this op is pure). 563 let (scope_depth, before, insert_block) = if loop_hoist_level 564 == self.loop_stack.len() 565 { 566 // Depends on some value at the current 567 // loop depth, or remat forces it here: 568 // place it at the current location. 569 ( 570 self.value_to_elaborated_value.depth(), 571 before, 572 self.func.layout.inst_block(before).unwrap(), 573 ) 574 } else { 575 // Does not depend on any args at current 576 // loop depth: hoist out of loop. 577 self.stats.elaborate_licm_hoist += 1; 578 let data = &self.loop_stack[loop_hoist_level]; 579 // `data.hoist_block` should dominate `before`'s block. 580 let before_block = self.func.layout.inst_block(before).unwrap(); 581 debug_assert!(self.domtree.block_dominates(data.hoist_block, before_block)); 582 // Determine the instruction at which we 583 // insert in `data.hoist_block`. 584 let before = self.func.layout.last_inst(data.hoist_block).unwrap(); 585 (data.scope_depth as usize, before, data.hoist_block) 586 }; 587 588 trace!( 589 " -> decided to place: before {} insert_block {}", 590 before, insert_block 591 ); 592 593 // Now that we have the location for the 594 // instruction, check if any of its args are remat 595 // values. If so, and if we don't have a copy of 596 // the rematerializing instruction for this block 597 // yet, create one. 598 let mut remat_arg = false; 599 for arg_value in arg_values.iter_mut() { 600 if Self::maybe_remat_arg( 601 &self.remat_values, 602 &mut self.func, 603 &mut self.remat_copies, 604 insert_block, 605 before, 606 arg_value, 607 &mut self.stats, 608 ) { 609 remat_arg = true; 610 } 611 } 612 613 // Now we need to place `inst` at the computed 614 // location (just before `before`). Note that 615 // `inst` may already have been placed somewhere 616 // else, because a pure node may be elaborated at 617 // more than one place. In this case, we need to 618 // duplicate the instruction (and return the 619 // `Value`s for that duplicated instance instead). 620 // 621 // Also clone if we rematerialized, because we 622 // don't want to rewrite the args in the original 623 // copy. 624 trace!("need inst {} before {}", inst, before); 625 let inst = if self.func.layout.inst_block(inst).is_some() || remat_arg { 626 // Clone the inst! 627 let new_inst = self.func.dfg.clone_inst(inst); 628 trace!( 629 " -> inst {} already has a location; cloned to {}", 630 inst, new_inst 631 ); 632 // Create mappings in the 633 // value-to-elab'd-value map from original 634 // results to cloned results. 635 for (&result, &new_result) in self 636 .func 637 .dfg 638 .inst_results(inst) 639 .iter() 640 .zip(self.func.dfg.inst_results(new_inst).iter()) 641 { 642 let elab_value = ElaboratedValue { 643 value: new_result, 644 in_block: insert_block, 645 }; 646 let best_result = self.value_to_best_value[result]; 647 self.value_to_elaborated_value.insert_if_absent_with_depth( 648 &NullCtx, 649 best_result.1, 650 elab_value, 651 scope_depth, 652 ); 653 654 self.value_to_best_value[new_result] = best_result; 655 656 trace!( 657 " -> cloned inst has new result {} for orig {}", 658 new_result, result 659 ); 660 } 661 new_inst 662 } else { 663 trace!(" -> no location; using original inst"); 664 // Create identity mappings from result values 665 // to themselves in this scope, since we're 666 // using the original inst. 667 for &result in self.func.dfg.inst_results(inst) { 668 let elab_value = ElaboratedValue { 669 value: result, 670 in_block: insert_block, 671 }; 672 let best_result = self.value_to_best_value[result]; 673 self.value_to_elaborated_value.insert_if_absent_with_depth( 674 &NullCtx, 675 best_result.1, 676 elab_value, 677 scope_depth, 678 ); 679 trace!(" -> inserting identity mapping for {}", result); 680 } 681 inst 682 }; 683 684 // Place the inst just before `before`. 685 assert!( 686 is_pure_for_egraph(self.func, inst), 687 "something has gone very wrong if we are elaborating effectful \ 688 instructions, they should have remained in the skeleton" 689 ); 690 self.func.layout.insert_inst(inst, before); 691 692 // Update the inst's arguments. 693 self.func 694 .dfg 695 .overwrite_inst_values(inst, arg_values.into_iter().map(|ev| ev.value)); 696 697 // Now that we've consumed the arg values, pop 698 // them off the stack. 699 self.elab_result_stack.truncate(arg_idx); 700 701 // Push the requested result index of the 702 // instruction onto the elab-results stack. 703 self.elab_result_stack.push(ElaboratedValue { 704 in_block: insert_block, 705 value: self.func.dfg.inst_results(inst)[result_idx], 706 }); 707 } 708 } 709 } 710 } 711 712 fn elaborate_block(&mut self, elab_values: &mut Vec<Value>, idom: Option<Block>, block: Block) { 713 trace!("elaborate_block: block {}", block); 714 self.start_block(idom, block); 715 716 // Iterate over the side-effecting skeleton using the linked 717 // list in Layout. We will insert instructions that are 718 // elaborated *before* `inst`, so we can always use its 719 // next-link to continue the iteration. 720 let mut next_inst = self.func.layout.first_inst(block); 721 let mut first_branch = None; 722 while let Some(inst) = next_inst { 723 trace!( 724 "elaborating inst {} with results {:?}", 725 inst, 726 self.func.dfg.inst_results(inst) 727 ); 728 // Record the first branch we see in the block; all 729 // elaboration for args of *any* branch must be inserted 730 // before the *first* branch, because the branch group 731 // must remain contiguous at the end of the block. 732 if self.func.dfg.insts[inst].opcode().is_branch() && first_branch == None { 733 first_branch = Some(inst); 734 } 735 736 // Determine where elaboration inserts insts. 737 let before = first_branch.unwrap_or(inst); 738 trace!(" -> inserting before {}", before); 739 740 elab_values.extend(self.func.dfg.inst_values(inst)); 741 for arg in elab_values.iter_mut() { 742 trace!(" -> arg {}", *arg); 743 // Elaborate the arg, placing any newly-inserted insts 744 // before `before`. Get the updated value, which may 745 // be different than the original. 746 let mut new_arg = self.elaborate_eclass_use(*arg, before); 747 Self::maybe_remat_arg( 748 &self.remat_values, 749 &mut self.func, 750 &mut self.remat_copies, 751 block, 752 inst, 753 &mut new_arg, 754 &mut self.stats, 755 ); 756 trace!(" -> rewrote arg to {:?}", new_arg); 757 *arg = new_arg.value; 758 } 759 self.func 760 .dfg 761 .overwrite_inst_values(inst, elab_values.drain(..)); 762 763 // We need to put the results of this instruction in the 764 // map now. 765 for &result in self.func.dfg.inst_results(inst) { 766 trace!(" -> result {}", result); 767 let best_result = self.value_to_best_value[result]; 768 self.value_to_elaborated_value.insert_if_absent( 769 &NullCtx, 770 best_result.1, 771 ElaboratedValue { 772 in_block: block, 773 value: result, 774 }, 775 ); 776 } 777 778 next_inst = self.func.layout.next_inst(inst); 779 } 780 } 781 782 fn elaborate_domtree(&mut self, domtree: &DominatorTree) { 783 self.block_stack.push(BlockStackEntry::Elaborate { 784 block: self.func.layout.entry_block().unwrap(), 785 idom: None, 786 }); 787 788 // A temporary workspace for elaborate_block, allocated here to maximize the use of the 789 // allocation. 790 let mut elab_values = Vec::new(); 791 792 while let Some(top) = self.block_stack.pop() { 793 match top { 794 BlockStackEntry::Elaborate { block, idom } => { 795 self.block_stack.push(BlockStackEntry::Pop); 796 self.value_to_elaborated_value.increment_depth(); 797 798 self.elaborate_block(&mut elab_values, idom, block); 799 800 // Push children. We are doing a preorder 801 // traversal so we do this after processing this 802 // block above. 803 let block_stack_end = self.block_stack.len(); 804 for child in self.ctrl_plane.shuffled(domtree.children(block)) { 805 self.block_stack.push(BlockStackEntry::Elaborate { 806 block: child, 807 idom: Some(block), 808 }); 809 } 810 // Reverse what we just pushed so we elaborate in 811 // original block order. (The domtree iter is a 812 // single-ended iter over a singly-linked list so 813 // we can't `.rev()` above.) 814 self.block_stack[block_stack_end..].reverse(); 815 } 816 BlockStackEntry::Pop => { 817 self.value_to_elaborated_value.decrement_depth(); 818 } 819 } 820 } 821 } 822 823 pub(crate) fn elaborate(&mut self) { 824 self.stats.elaborate_func += 1; 825 self.stats.elaborate_func_pre_insts += self.func.dfg.num_insts() as u64; 826 self.compute_best_values(); 827 self.elaborate_domtree(&self.domtree); 828 self.stats.elaborate_func_post_insts += self.func.dfg.num_insts() as u64; 829 } 830 } 831