1 //===-- WebAssemblyCFGStackify.cpp - CFG Stackification -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief This file implements a CFG stacking pass.
13 /// This pass reorders the blocks in a function to put them into a reverse
14 /// post-order [0], with special care to keep the order as similar as possible
15 /// to the original order, and to keep loops contiguous even in the case of
18 /// Then, it inserts BLOCK and LOOP markers to mark the start of scopes, since
19 /// scope boundaries serve as the labels for WebAssembly's control transfers.
21 /// This is sufficient to convert arbitrary CFGs into a form that works on
22 /// WebAssembly, provided that all loops are single-entry.
24 /// [0] https://en.wikipedia.org/wiki/Depth-first_search#Vertex_orderings
26 //===----------------------------------------------------------------------===//
28 #include "WebAssembly.h"
29 #include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
30 #include "WebAssemblySubtarget.h"
31 #include "llvm/ADT/SCCIterator.h"
32 #include "llvm/ADT/SetVector.h"
33 #include "llvm/CodeGen/MachineDominators.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/MachineLoopInfo.h"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
42 #define DEBUG_TYPE "wasm-cfg-stackify"
45 class WebAssemblyCFGStackify final : public MachineFunctionPass {
46 const char *getPassName() const override {
47 return "WebAssembly CFG Stackify";
50 void getAnalysisUsage(AnalysisUsage &AU) const override {
52 AU.addRequired<MachineDominatorTree>();
53 AU.addPreserved<MachineDominatorTree>();
54 AU.addRequired<MachineLoopInfo>();
55 AU.addPreserved<MachineLoopInfo>();
56 MachineFunctionPass::getAnalysisUsage(AU);
59 bool runOnMachineFunction(MachineFunction &MF) override;
62 static char ID; // Pass identification, replacement for typeid
63 WebAssemblyCFGStackify() : MachineFunctionPass(ID) {}
65 } // end anonymous namespace
67 char WebAssemblyCFGStackify::ID = 0;
68 FunctionPass *llvm::createWebAssemblyCFGStackify() {
69 return new WebAssemblyCFGStackify();
72 static void EliminateMultipleEntryLoops(MachineFunction &MF,
73 const MachineLoopInfo &MLI) {
74 SmallPtrSet<MachineBasicBlock *, 8> InSet;
75 for (scc_iterator<MachineFunction *> I = scc_begin(&MF), E = scc_end(&MF);
77 const std::vector<MachineBasicBlock *> &CurrentSCC = *I;
80 if (CurrentSCC.size() == 1)
83 InSet.insert(CurrentSCC.begin(), CurrentSCC.end());
84 MachineBasicBlock *Header = nullptr;
85 for (MachineBasicBlock *MBB : CurrentSCC) {
86 for (MachineBasicBlock *Pred : MBB->predecessors()) {
87 if (InSet.count(Pred))
93 // TODO: Implement multiple-entry loops.
94 report_fatal_error("multiple-entry loops are not supported yet");
97 assert(MLI.isLoopHeader(Header));
104 /// Post-order traversal stack entry.
105 struct POStackEntry {
106 MachineBasicBlock *MBB;
107 SmallVector<MachineBasicBlock *, 0> Succs;
109 POStackEntry(MachineBasicBlock *MBB, MachineFunction &MF,
110 const MachineLoopInfo &MLI);
112 } // end anonymous namespace
114 static bool LoopContains(const MachineLoop *Loop,
115 const MachineBasicBlock *MBB) {
116 return Loop ? Loop->contains(MBB) : true;
119 POStackEntry::POStackEntry(MachineBasicBlock *MBB, MachineFunction &MF,
120 const MachineLoopInfo &MLI)
121 : MBB(MBB), Succs(MBB->successors()) {
122 // RPO is not a unique form, since at every basic block with multiple
123 // successors, the DFS has to pick which order to visit the successors in.
124 // Sort them strategically (see below).
125 MachineLoop *Loop = MLI.getLoopFor(MBB);
126 MachineFunction::iterator Next = next(MachineFunction::iterator(MBB));
127 MachineBasicBlock *LayoutSucc = Next == MF.end() ? nullptr : &*Next;
129 Succs.begin(), Succs.end(),
130 [=, &MLI](const MachineBasicBlock *A, const MachineBasicBlock *B) {
134 // Keep loops contiguous by preferring the block that's in the same
136 bool LoopContainsA = LoopContains(Loop, A);
137 bool LoopContainsB = LoopContains(Loop, B);
138 if (LoopContainsA && !LoopContainsB)
140 if (!LoopContainsA && LoopContainsB)
143 // Minimize perturbation by preferring the block which is the immediate
150 // TODO: More sophisticated orderings may be profitable here.
156 /// Return the "bottom" block of a loop. This differs from
157 /// MachineLoop::getBottomBlock in that it works even if the loop is
159 static MachineBasicBlock *LoopBottom(const MachineLoop *Loop) {
160 MachineBasicBlock *Bottom = Loop->getHeader();
161 for (MachineBasicBlock *MBB : Loop->blocks())
162 if (MBB->getNumber() > Bottom->getNumber())
167 /// Sort the blocks in RPO, taking special care to make sure that loops are
168 /// contiguous even in the case of split backedges.
170 /// TODO: Determine whether RPO is actually worthwhile, or whether we should
171 /// move to just a stable-topological-sort-based approach that would preserve
172 /// more of the original order.
173 static void SortBlocks(MachineFunction &MF, const MachineLoopInfo &MLI) {
174 // Note that we do our own RPO rather than using
175 // "llvm/ADT/PostOrderIterator.h" because we want control over the order that
176 // successors are visited in (see above). Also, we can sort the blocks in the
177 // MachineFunction as we go.
178 SmallPtrSet<MachineBasicBlock *, 16> Visited;
179 SmallVector<POStackEntry, 16> Stack;
181 MachineBasicBlock *EntryBlock = &*MF.begin();
182 Visited.insert(EntryBlock);
183 Stack.push_back(POStackEntry(EntryBlock, MF, MLI));
186 POStackEntry &Entry = Stack.back();
187 SmallVectorImpl<MachineBasicBlock *> &Succs = Entry.Succs;
188 if (!Succs.empty()) {
189 MachineBasicBlock *Succ = Succs.pop_back_val();
190 if (Visited.insert(Succ).second)
191 Stack.push_back(POStackEntry(Succ, MF, MLI));
195 // Put the block in its position in the MachineFunction.
196 MachineBasicBlock &MBB = *Entry.MBB;
197 MBB.moveBefore(&*MF.begin());
199 // Branch instructions may utilize a fallthrough, so update them if a
200 // fallthrough has been added or removed.
201 if (!MBB.empty() && MBB.back().isTerminator() && !MBB.back().isBranch() &&
202 !MBB.back().isBarrier())
204 "Non-branch terminator with fallthrough cannot yet be rewritten");
205 if (MBB.empty() || !MBB.back().isTerminator() || MBB.back().isBranch())
206 MBB.updateTerminator();
213 // Now that we've sorted the blocks in RPO, renumber them.
217 SmallSetVector<MachineLoop *, 8> OnStack;
219 // Insert a sentinel representing the degenerate loop that starts at the
220 // function entry block and includes the entire function as a "loop" that
222 OnStack.insert(nullptr);
224 for (auto &MBB : MF) {
225 assert(MBB.getNumber() >= 0 && "Renumbered blocks should be non-negative.");
227 MachineLoop *Loop = MLI.getLoopFor(&MBB);
228 if (Loop && &MBB == Loop->getHeader()) {
229 // Loop header. The loop predecessor should be sorted above, and the other
230 // predecessors should be backedges below.
231 for (auto Pred : MBB.predecessors())
233 (Pred->getNumber() < MBB.getNumber() || Loop->contains(Pred)) &&
234 "Loop header predecessors must be loop predecessors or backedges");
235 assert(OnStack.insert(Loop) && "Loops should be declared at most once.");
237 // Not a loop header. All predecessors should be sorted above.
238 for (auto Pred : MBB.predecessors())
239 assert(Pred->getNumber() < MBB.getNumber() &&
240 "Non-loop-header predecessors should be topologically sorted");
241 assert(OnStack.count(MLI.getLoopFor(&MBB)) &&
242 "Blocks must be nested in their loops");
244 while (OnStack.size() > 1 && &MBB == LoopBottom(OnStack.back()))
247 assert(OnStack.pop_back_val() == nullptr &&
248 "The function entry block shouldn't actually be a loop header");
249 assert(OnStack.empty() &&
250 "Control flow stack pushes and pops should be balanced.");
254 /// Test whether Pred has any terminators explicitly branching to MBB, as
255 /// opposed to falling through. Note that it's possible (eg. in unoptimized
256 /// code) for a branch instruction to both branch to a block and fallthrough
257 /// to it, so we check the actual branch operands to see if there are any
258 /// explicit mentions.
259 static bool ExplicitlyBranchesTo(MachineBasicBlock *Pred,
260 MachineBasicBlock *MBB) {
261 for (MachineInstr &MI : Pred->terminators())
262 for (MachineOperand &MO : MI.explicit_operands())
263 if (MO.isMBB() && MO.getMBB() == MBB)
268 /// Insert a BLOCK marker for branches to MBB (if needed).
269 static void PlaceBlockMarker(MachineBasicBlock &MBB, MachineFunction &MF,
270 SmallVectorImpl<MachineBasicBlock *> &ScopeTops,
271 const WebAssemblyInstrInfo &TII,
272 const MachineLoopInfo &MLI,
273 MachineDominatorTree &MDT) {
274 // First compute the nearest common dominator of all forward non-fallthrough
275 // predecessors so that we minimize the time that the BLOCK is on the stack,
276 // which reduces overall stack height.
277 MachineBasicBlock *Header = nullptr;
278 bool IsBranchedTo = false;
279 int MBBNumber = MBB.getNumber();
280 for (MachineBasicBlock *Pred : MBB.predecessors())
281 if (Pred->getNumber() < MBBNumber) {
282 Header = Header ? MDT.findNearestCommonDominator(Header, Pred) : Pred;
283 if (ExplicitlyBranchesTo(Pred, &MBB))
291 assert(&MBB != &MF.front() && "Header blocks shouldn't have predecessors");
292 MachineBasicBlock *LayoutPred = &*prev(MachineFunction::iterator(&MBB));
294 // If the nearest common dominator is inside a more deeply nested context,
295 // walk out to the nearest scope which isn't more deeply nested.
296 for (MachineFunction::iterator I(LayoutPred), E(Header); I != E; --I) {
297 if (MachineBasicBlock *ScopeTop = ScopeTops[I->getNumber()]) {
298 if (ScopeTop->getNumber() > Header->getNumber()) {
299 // Skip over an intervening scope.
300 I = next(MachineFunction::iterator(ScopeTop));
302 // We found a scope level at an appropriate depth.
309 // If there's a loop which ends just before MBB which contains Header, we can
310 // reuse its label instead of inserting a new BLOCK.
311 for (MachineLoop *Loop = MLI.getLoopFor(LayoutPred);
312 Loop && Loop->contains(LayoutPred); Loop = Loop->getParentLoop())
313 if (Loop && LoopBottom(Loop) == LayoutPred && Loop->contains(Header))
316 // Decide where in Header to put the BLOCK.
317 MachineBasicBlock::iterator InsertPos;
318 MachineLoop *HeaderLoop = MLI.getLoopFor(Header);
319 if (HeaderLoop && MBB.getNumber() > LoopBottom(HeaderLoop)->getNumber()) {
320 // Header is the header of a loop that does not lexically contain MBB, so
321 // the BLOCK needs to be above the LOOP.
322 InsertPos = Header->begin();
324 // Otherwise, insert the BLOCK as late in Header as we can, but before the
325 // beginning of the local expression tree and any nested BLOCKs.
326 InsertPos = Header->getFirstTerminator();
327 while (InsertPos != Header->begin() &&
328 prev(InsertPos)->definesRegister(WebAssembly::EXPR_STACK) &&
329 prev(InsertPos)->getOpcode() != WebAssembly::LOOP &&
330 prev(InsertPos)->getOpcode() != WebAssembly::END_BLOCK &&
331 prev(InsertPos)->getOpcode() != WebAssembly::END_LOOP)
336 BuildMI(*Header, InsertPos, DebugLoc(), TII.get(WebAssembly::BLOCK));
338 // Mark the end of the block.
339 InsertPos = MBB.begin();
340 while (InsertPos != MBB.end() &&
341 InsertPos->getOpcode() == WebAssembly::END_LOOP)
343 BuildMI(MBB, InsertPos, DebugLoc(), TII.get(WebAssembly::END_BLOCK));
345 // Track the farthest-spanning scope that ends at this point.
346 int Number = MBB.getNumber();
347 if (!ScopeTops[Number] ||
348 ScopeTops[Number]->getNumber() > Header->getNumber())
349 ScopeTops[Number] = Header;
352 /// Insert a LOOP marker for a loop starting at MBB (if it's a loop header).
353 static void PlaceLoopMarker(
354 MachineBasicBlock &MBB, MachineFunction &MF,
355 SmallVectorImpl<MachineBasicBlock *> &ScopeTops,
356 DenseMap<const MachineInstr *, const MachineBasicBlock *> &LoopTops,
357 const WebAssemblyInstrInfo &TII, const MachineLoopInfo &MLI) {
358 MachineLoop *Loop = MLI.getLoopFor(&MBB);
359 if (!Loop || Loop->getHeader() != &MBB)
362 // The operand of a LOOP is the first block after the loop. If the loop is the
363 // bottom of the function, insert a dummy block at the end.
364 MachineBasicBlock *Bottom = LoopBottom(Loop);
365 auto Iter = next(MachineFunction::iterator(Bottom));
366 if (Iter == MF.end()) {
367 MachineBasicBlock *Label = MF.CreateMachineBasicBlock();
368 // Give it a fake predecessor so that AsmPrinter prints its label.
369 Label->addSuccessor(Label);
371 Iter = next(MachineFunction::iterator(Bottom));
373 MachineBasicBlock *AfterLoop = &*Iter;
375 // Mark the beginning of the loop (after the end of any existing loop that
377 auto InsertPos = MBB.begin();
378 while (InsertPos != MBB.end() &&
379 InsertPos->getOpcode() == WebAssembly::END_LOOP)
381 BuildMI(MBB, InsertPos, DebugLoc(), TII.get(WebAssembly::LOOP));
383 // Mark the end of the loop.
384 MachineInstr *End = BuildMI(*AfterLoop, AfterLoop->begin(), DebugLoc(),
385 TII.get(WebAssembly::END_LOOP));
386 LoopTops[End] = &MBB;
388 assert((!ScopeTops[AfterLoop->getNumber()] ||
389 ScopeTops[AfterLoop->getNumber()]->getNumber() < MBB.getNumber()) &&
390 "With RPO we should visit the outer-most loop for a block first.");
391 if (!ScopeTops[AfterLoop->getNumber()])
392 ScopeTops[AfterLoop->getNumber()] = &MBB;
396 GetDepth(const SmallVectorImpl<const MachineBasicBlock *> &Stack,
397 const MachineBasicBlock *MBB) {
399 for (auto X : reverse(Stack)) {
404 assert(Depth < Stack.size() && "Branch destination should be in scope");
408 /// Insert LOOP and BLOCK markers at appropriate places.
409 static void PlaceMarkers(MachineFunction &MF, const MachineLoopInfo &MLI,
410 const WebAssemblyInstrInfo &TII,
411 MachineDominatorTree &MDT) {
412 // For each block whose label represents the end of a scope, record the block
413 // which holds the beginning of the scope. This will allow us to quickly skip
414 // over scoped regions when walking blocks. We allocate one more than the
415 // number of blocks in the function to accommodate for the possible fake block
416 // we may insert at the end.
417 SmallVector<MachineBasicBlock *, 8> ScopeTops(MF.getNumBlockIDs() + 1);
419 // For eacn LOOP_END, the corresponding LOOP.
420 DenseMap<const MachineInstr *, const MachineBasicBlock *> LoopTops;
422 for (auto &MBB : MF) {
423 // Place the LOOP for MBB if MBB is the header of a loop.
424 PlaceLoopMarker(MBB, MF, ScopeTops, LoopTops, TII, MLI);
426 // Place the BLOCK for MBB if MBB is branched to from above.
427 PlaceBlockMarker(MBB, MF, ScopeTops, TII, MLI, MDT);
430 // Now rewrite references to basic blocks to be depth immediates.
431 SmallVector<const MachineBasicBlock *, 8> Stack;
432 for (auto &MBB : reverse(MF)) {
433 for (auto &MI : reverse(MBB)) {
434 switch (MI.getOpcode()) {
435 case WebAssembly::BLOCK:
436 assert(ScopeTops[Stack.back()->getNumber()] == &MBB &&
437 "Block should be balanced");
440 case WebAssembly::LOOP:
441 assert(Stack.back() == &MBB && "Loop top should be balanced");
445 case WebAssembly::END_BLOCK:
446 Stack.push_back(&MBB);
448 case WebAssembly::END_LOOP:
449 Stack.push_back(&MBB);
450 Stack.push_back(LoopTops[&MI]);
453 if (MI.isTerminator()) {
454 // Rewrite MBB operands to be depth immediates.
455 SmallVector<MachineOperand, 4> Ops(MI.operands());
456 while (MI.getNumOperands() > 0)
457 MI.RemoveOperand(MI.getNumOperands() - 1);
458 for (auto MO : Ops) {
460 MO = MachineOperand::CreateImm(GetDepth(Stack, MO.getMBB()));
461 MI.addOperand(MF, MO);
468 assert(Stack.empty() && "Control flow should be balanced");
471 bool WebAssemblyCFGStackify::runOnMachineFunction(MachineFunction &MF) {
472 DEBUG(dbgs() << "********** CFG Stackifying **********\n"
473 "********** Function: "
474 << MF.getName() << '\n');
476 const auto &MLI = getAnalysis<MachineLoopInfo>();
477 auto &MDT = getAnalysis<MachineDominatorTree>();
478 const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
480 // RPO sorting needs all loops to be single-entry.
481 EliminateMultipleEntryLoops(MF, MLI);
483 // Sort the blocks in RPO, with contiguous loops.
486 // Place the BLOCK and LOOP markers to indicate the beginnings of scopes.
487 PlaceMarkers(MF, MLI, TII, MDT);