xref: /llvm-project/llvm/lib/Target/WebAssembly/WebAssemblyCFGStackify.cpp (revision e76fa9eccaec169913725487c1760341ff4853d4)
1 //===-- WebAssemblyCFGStackify.cpp - CFG Stackification -------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 ///
10 /// \file
11 /// This file implements a CFG stacking pass.
12 ///
13 /// This pass inserts BLOCK, LOOP, and TRY markers to mark the start of scopes,
14 /// since scope boundaries serve as the labels for WebAssembly's control
15 /// transfers.
16 ///
17 /// This is sufficient to convert arbitrary CFGs into a form that works on
18 /// WebAssembly, provided that all loops are single-entry.
19 ///
20 /// In case we use exceptions, this pass also fixes mismatches in unwind
21 /// destinations created during transforming CFG into wasm structured format.
22 ///
23 //===----------------------------------------------------------------------===//
24 
25 #include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
26 #include "WebAssembly.h"
27 #include "WebAssemblyExceptionInfo.h"
28 #include "WebAssemblyMachineFunctionInfo.h"
29 #include "WebAssemblySubtarget.h"
30 #include "WebAssemblyUtilities.h"
31 #include "llvm/CodeGen/MachineDominators.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/CodeGen/WasmEHFuncInfo.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
41 using namespace llvm;
42 
43 #define DEBUG_TYPE "wasm-cfg-stackify"
44 
45 namespace {
46 class WebAssemblyCFGStackify final : public MachineFunctionPass {
47   StringRef getPassName() const override { return "WebAssembly CFG Stackify"; }
48 
49   void getAnalysisUsage(AnalysisUsage &AU) const override {
50     AU.addRequired<MachineDominatorTree>();
51     AU.addRequired<MachineLoopInfo>();
52     AU.addRequired<WebAssemblyExceptionInfo>();
53     MachineFunctionPass::getAnalysisUsage(AU);
54   }
55 
56   bool runOnMachineFunction(MachineFunction &MF) override;
57 
58   // For each block whose label represents the end of a scope, record the block
59   // which holds the beginning of the scope. This will allow us to quickly skip
60   // over scoped regions when walking blocks.
61   SmallVector<MachineBasicBlock *, 8> ScopeTops;
62 
63   void placeMarkers(MachineFunction &MF);
64   void placeBlockMarker(MachineBasicBlock &MBB);
65   void placeLoopMarker(MachineBasicBlock &MBB);
66   void placeTryMarker(MachineBasicBlock &MBB);
67   void rewriteDepthImmediates(MachineFunction &MF);
68   void fixEndsAtEndOfFunction(MachineFunction &MF);
69 
70   // For each BLOCK|LOOP|TRY, the corresponding END_(BLOCK|LOOP|TRY).
71   DenseMap<const MachineInstr *, MachineInstr *> BeginToEnd;
72   // For each END_(BLOCK|LOOP|TRY), the corresponding BLOCK|LOOP|TRY.
73   DenseMap<const MachineInstr *, MachineInstr *> EndToBegin;
74   // <TRY marker, EH pad> map
75   DenseMap<const MachineInstr *, MachineBasicBlock *> TryToEHPad;
76   // <EH pad, TRY marker> map
77   DenseMap<const MachineBasicBlock *, MachineInstr *> EHPadToTry;
78   // <LOOP|TRY marker, Loop/exception bottom BB> map
79   DenseMap<const MachineInstr *, MachineBasicBlock *> BeginToBottom;
80 
81   // Helper functions to register / unregister scope information created by
82   // marker instructions.
83   void registerScope(MachineInstr *Begin, MachineInstr *End);
84   void registerTryScope(MachineInstr *Begin, MachineInstr *End,
85                         MachineBasicBlock *EHPad);
86   void unregisterScope(MachineInstr *Begin);
87 
88   MachineBasicBlock *getBottom(const MachineInstr *Begin);
89 
90 public:
91   static char ID; // Pass identification, replacement for typeid
92   WebAssemblyCFGStackify() : MachineFunctionPass(ID) {}
93   ~WebAssemblyCFGStackify() override { releaseMemory(); }
94   void releaseMemory() override;
95 };
96 } // end anonymous namespace
97 
98 char WebAssemblyCFGStackify::ID = 0;
99 INITIALIZE_PASS(WebAssemblyCFGStackify, DEBUG_TYPE,
100                 "Insert BLOCK and LOOP markers for WebAssembly scopes",
101                 false, false)
102 
103 FunctionPass *llvm::createWebAssemblyCFGStackify() {
104   return new WebAssemblyCFGStackify();
105 }
106 
107 /// Test whether Pred has any terminators explicitly branching to MBB, as
108 /// opposed to falling through. Note that it's possible (eg. in unoptimized
109 /// code) for a branch instruction to both branch to a block and fallthrough
110 /// to it, so we check the actual branch operands to see if there are any
111 /// explicit mentions.
112 static bool ExplicitlyBranchesTo(MachineBasicBlock *Pred,
113                                  MachineBasicBlock *MBB) {
114   for (MachineInstr &MI : Pred->terminators())
115     // Even if a rethrow takes a BB argument, it is not a branch
116     if (!WebAssembly::isRethrow(MI))
117       for (MachineOperand &MO : MI.explicit_operands())
118         if (MO.isMBB() && MO.getMBB() == MBB)
119           return true;
120   return false;
121 }
122 
123 // Returns an iterator to the earliest position possible within the MBB,
124 // satisfying the restrictions given by BeforeSet and AfterSet. BeforeSet
125 // contains instructions that should go before the marker, and AfterSet contains
126 // ones that should go after the marker. In this function, AfterSet is only
127 // used for sanity checking.
128 static MachineBasicBlock::iterator
129 GetEarliestInsertPos(MachineBasicBlock *MBB,
130                      const SmallPtrSet<const MachineInstr *, 4> &BeforeSet,
131                      const SmallPtrSet<const MachineInstr *, 4> &AfterSet) {
132   auto InsertPos = MBB->end();
133   while (InsertPos != MBB->begin()) {
134     if (BeforeSet.count(&*std::prev(InsertPos))) {
135 #ifndef NDEBUG
136       // Sanity check
137       for (auto Pos = InsertPos, E = MBB->begin(); Pos != E; --Pos)
138         assert(!AfterSet.count(&*std::prev(Pos)));
139 #endif
140       break;
141     }
142     --InsertPos;
143   }
144   return InsertPos;
145 }
146 
147 // Returns an iterator to the latest position possible within the MBB,
148 // satisfying the restrictions given by BeforeSet and AfterSet. BeforeSet
149 // contains instructions that should go before the marker, and AfterSet contains
150 // ones that should go after the marker. In this function, BeforeSet is only
151 // used for sanity checking.
152 static MachineBasicBlock::iterator
153 GetLatestInsertPos(MachineBasicBlock *MBB,
154                    const SmallPtrSet<const MachineInstr *, 4> &BeforeSet,
155                    const SmallPtrSet<const MachineInstr *, 4> &AfterSet) {
156   auto InsertPos = MBB->begin();
157   while (InsertPos != MBB->end()) {
158     if (AfterSet.count(&*InsertPos)) {
159 #ifndef NDEBUG
160       // Sanity check
161       for (auto Pos = InsertPos, E = MBB->end(); Pos != E; ++Pos)
162         assert(!BeforeSet.count(&*Pos));
163 #endif
164       break;
165     }
166     ++InsertPos;
167   }
168   return InsertPos;
169 }
170 
171 void WebAssemblyCFGStackify::registerScope(MachineInstr *Begin,
172                                            MachineInstr *End) {
173   BeginToEnd[Begin] = End;
174   EndToBegin[End] = Begin;
175 }
176 
177 void WebAssemblyCFGStackify::registerTryScope(MachineInstr *Begin,
178                                               MachineInstr *End,
179                                               MachineBasicBlock *EHPad) {
180   registerScope(Begin, End);
181   TryToEHPad[Begin] = EHPad;
182   EHPadToTry[EHPad] = Begin;
183 }
184 
185 void WebAssemblyCFGStackify::unregisterScope(MachineInstr *Begin) {
186   assert(BeginToEnd.count(Begin));
187   MachineInstr *End = BeginToEnd[Begin];
188   assert(EndToBegin.count(End));
189   BeginToEnd.erase(Begin);
190   EndToBegin.erase(End);
191   MachineBasicBlock *EHPad = TryToEHPad.lookup(Begin);
192   if (EHPad) {
193     assert(EHPadToTry.count(EHPad));
194     TryToEHPad.erase(Begin);
195     EHPadToTry.erase(EHPad);
196   }
197   MachineBasicBlock *Bottom = BeginToBottom.lookup(Begin);
198   if (Bottom)
199     BeginToBottom.erase(Begin);
200 }
201 
202 // Given a LOOP/TRY marker, returns its bottom BB. Use cached information if any
203 // to prevent recomputation.
204 MachineBasicBlock *
205 WebAssemblyCFGStackify::getBottom(const MachineInstr *Begin) {
206   const auto &MLI = getAnalysis<MachineLoopInfo>();
207   const auto &WEI = getAnalysis<WebAssemblyExceptionInfo>();
208   if (BeginToBottom.count(Begin))
209     return BeginToBottom[Begin];
210   if (Begin->getOpcode() == WebAssembly::LOOP) {
211     MachineLoop *L = MLI.getLoopFor(Begin->getParent());
212     assert(L);
213     BeginToBottom[Begin] = WebAssembly::getBottom(L);
214   } else if (Begin->getOpcode() == WebAssembly::TRY) {
215     WebAssemblyException *WE = WEI.getExceptionFor(TryToEHPad[Begin]);
216     assert(WE);
217     BeginToBottom[Begin] = WebAssembly::getBottom(WE);
218   } else
219     assert(false);
220   return BeginToBottom[Begin];
221 }
222 
223 /// Insert a BLOCK marker for branches to MBB (if needed).
224 void WebAssemblyCFGStackify::placeBlockMarker(MachineBasicBlock &MBB) {
225   // This should have been handled in placeTryMarker.
226   if (MBB.isEHPad())
227     return;
228 
229   MachineFunction &MF = *MBB.getParent();
230   auto &MDT = getAnalysis<MachineDominatorTree>();
231   const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
232   const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
233 
234   // First compute the nearest common dominator of all forward non-fallthrough
235   // predecessors so that we minimize the time that the BLOCK is on the stack,
236   // which reduces overall stack height.
237   MachineBasicBlock *Header = nullptr;
238   bool IsBranchedTo = false;
239   int MBBNumber = MBB.getNumber();
240   for (MachineBasicBlock *Pred : MBB.predecessors()) {
241     if (Pred->getNumber() < MBBNumber) {
242       Header = Header ? MDT.findNearestCommonDominator(Header, Pred) : Pred;
243       if (ExplicitlyBranchesTo(Pred, &MBB))
244         IsBranchedTo = true;
245     }
246   }
247   if (!Header)
248     return;
249   if (!IsBranchedTo)
250     return;
251 
252   assert(&MBB != &MF.front() && "Header blocks shouldn't have predecessors");
253   MachineBasicBlock *LayoutPred = &*std::prev(MachineFunction::iterator(&MBB));
254 
255   // If the nearest common dominator is inside a more deeply nested context,
256   // walk out to the nearest scope which isn't more deeply nested.
257   for (MachineFunction::iterator I(LayoutPred), E(Header); I != E; --I) {
258     if (MachineBasicBlock *ScopeTop = ScopeTops[I->getNumber()]) {
259       if (ScopeTop->getNumber() > Header->getNumber()) {
260         // Skip over an intervening scope.
261         I = std::next(MachineFunction::iterator(ScopeTop));
262       } else {
263         // We found a scope level at an appropriate depth.
264         Header = ScopeTop;
265         break;
266       }
267     }
268   }
269 
270   // Decide where in Header to put the BLOCK.
271 
272   // Instructions that should go before the BLOCK.
273   SmallPtrSet<const MachineInstr *, 4> BeforeSet;
274   // Instructions that should go after the BLOCK.
275   SmallPtrSet<const MachineInstr *, 4> AfterSet;
276   for (const auto &MI : *Header) {
277     // If there is a previously placed LOOP/TRY marker and the bottom block of
278     // the loop/exception is above MBB, it should be after the BLOCK, because
279     // the loop/exception is nested in this block. Otherwise it should be before
280     // the BLOCK.
281     if (MI.getOpcode() == WebAssembly::LOOP ||
282         MI.getOpcode() == WebAssembly::TRY) {
283       if (MBB.getNumber() > getBottom(&MI)->getNumber())
284         AfterSet.insert(&MI);
285 #ifndef NDEBUG
286       else
287         BeforeSet.insert(&MI);
288 #endif
289     }
290 
291     // All previously inserted BLOCK markers should be after the BLOCK because
292     // they are all nested blocks.
293     if (MI.getOpcode() == WebAssembly::BLOCK)
294       AfterSet.insert(&MI);
295 
296 #ifndef NDEBUG
297     // All END_(BLOCK|LOOP|TRY) markers should be before the BLOCK.
298     if (MI.getOpcode() == WebAssembly::END_BLOCK ||
299         MI.getOpcode() == WebAssembly::END_LOOP ||
300         MI.getOpcode() == WebAssembly::END_TRY)
301       BeforeSet.insert(&MI);
302 #endif
303 
304     // Terminators should go after the BLOCK.
305     if (MI.isTerminator())
306       AfterSet.insert(&MI);
307   }
308 
309   // Local expression tree should go after the BLOCK.
310   for (auto I = Header->getFirstTerminator(), E = Header->begin(); I != E;
311        --I) {
312     if (WebAssembly::isChild(*std::prev(I), MFI))
313       AfterSet.insert(&*std::prev(I));
314     else
315       break;
316   }
317 
318   // Add the BLOCK.
319   auto InsertPos = GetLatestInsertPos(Header, BeforeSet, AfterSet);
320   MachineInstr *Begin =
321       BuildMI(*Header, InsertPos, Header->findDebugLoc(InsertPos),
322               TII.get(WebAssembly::BLOCK))
323           .addImm(int64_t(WebAssembly::ExprType::Void));
324 
325   // Decide where in Header to put the END_BLOCK.
326   BeforeSet.clear();
327   AfterSet.clear();
328   for (auto &MI : MBB) {
329 #ifndef NDEBUG
330     // END_BLOCK should precede existing LOOP and TRY markers.
331     if (MI.getOpcode() == WebAssembly::LOOP ||
332         MI.getOpcode() == WebAssembly::TRY)
333       AfterSet.insert(&MI);
334 #endif
335 
336     // If there is a previously placed END_LOOP marker and the header of the
337     // loop is above this block's header, the END_LOOP should be placed after
338     // the BLOCK, because the loop contains this block. Otherwise the END_LOOP
339     // should be placed before the BLOCK. The same for END_TRY.
340     if (MI.getOpcode() == WebAssembly::END_LOOP ||
341         MI.getOpcode() == WebAssembly::END_TRY) {
342       if (EndToBegin[&MI]->getParent()->getNumber() >= Header->getNumber())
343         BeforeSet.insert(&MI);
344 #ifndef NDEBUG
345       else
346         AfterSet.insert(&MI);
347 #endif
348     }
349   }
350 
351   // Mark the end of the block.
352   InsertPos = GetEarliestInsertPos(&MBB, BeforeSet, AfterSet);
353   MachineInstr *End = BuildMI(MBB, InsertPos, MBB.findPrevDebugLoc(InsertPos),
354                               TII.get(WebAssembly::END_BLOCK));
355   registerScope(Begin, End);
356 
357   // Track the farthest-spanning scope that ends at this point.
358   int Number = MBB.getNumber();
359   if (!ScopeTops[Number] ||
360       ScopeTops[Number]->getNumber() > Header->getNumber())
361     ScopeTops[Number] = Header;
362 }
363 
364 /// Insert a LOOP marker for a loop starting at MBB (if it's a loop header).
365 void WebAssemblyCFGStackify::placeLoopMarker(MachineBasicBlock &MBB) {
366   MachineFunction &MF = *MBB.getParent();
367   const auto &MLI = getAnalysis<MachineLoopInfo>();
368   const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
369 
370   MachineLoop *Loop = MLI.getLoopFor(&MBB);
371   if (!Loop || Loop->getHeader() != &MBB)
372     return;
373 
374   // The operand of a LOOP is the first block after the loop. If the loop is the
375   // bottom of the function, insert a dummy block at the end.
376   MachineBasicBlock *Bottom = WebAssembly::getBottom(Loop);
377   auto Iter = std::next(MachineFunction::iterator(Bottom));
378   if (Iter == MF.end()) {
379     MachineBasicBlock *Label = MF.CreateMachineBasicBlock();
380     // Give it a fake predecessor so that AsmPrinter prints its label.
381     Label->addSuccessor(Label);
382     MF.push_back(Label);
383     Iter = std::next(MachineFunction::iterator(Bottom));
384   }
385   MachineBasicBlock *AfterLoop = &*Iter;
386 
387   // Decide where in Header to put the LOOP.
388   SmallPtrSet<const MachineInstr *, 4> BeforeSet;
389   SmallPtrSet<const MachineInstr *, 4> AfterSet;
390   for (const auto &MI : MBB) {
391     // LOOP marker should be after any existing loop that ends here. Otherwise
392     // we assume the instruction belongs to the loop.
393     if (MI.getOpcode() == WebAssembly::END_LOOP)
394       BeforeSet.insert(&MI);
395 #ifndef NDEBUG
396     else
397       AfterSet.insert(&MI);
398 #endif
399   }
400 
401   // Mark the beginning of the loop.
402   auto InsertPos = GetEarliestInsertPos(&MBB, BeforeSet, AfterSet);
403   MachineInstr *Begin = BuildMI(MBB, InsertPos, MBB.findDebugLoc(InsertPos),
404                                 TII.get(WebAssembly::LOOP))
405                             .addImm(int64_t(WebAssembly::ExprType::Void));
406 
407   // Decide where in Header to put the END_LOOP.
408   BeforeSet.clear();
409   AfterSet.clear();
410 #ifndef NDEBUG
411   for (const auto &MI : MBB)
412     // Existing END_LOOP markers belong to parent loops of this loop
413     if (MI.getOpcode() == WebAssembly::END_LOOP)
414       AfterSet.insert(&MI);
415 #endif
416 
417   // Mark the end of the loop (using arbitrary debug location that branched to
418   // the loop end as its location).
419   InsertPos = GetEarliestInsertPos(AfterLoop, BeforeSet, AfterSet);
420   DebugLoc EndDL = (*AfterLoop->pred_rbegin())->findBranchDebugLoc();
421   MachineInstr *End =
422       BuildMI(*AfterLoop, InsertPos, EndDL, TII.get(WebAssembly::END_LOOP));
423   registerScope(Begin, End);
424 
425   assert((!ScopeTops[AfterLoop->getNumber()] ||
426           ScopeTops[AfterLoop->getNumber()]->getNumber() < MBB.getNumber()) &&
427          "With block sorting the outermost loop for a block should be first.");
428   if (!ScopeTops[AfterLoop->getNumber()])
429     ScopeTops[AfterLoop->getNumber()] = &MBB;
430 }
431 
432 void WebAssemblyCFGStackify::placeTryMarker(MachineBasicBlock &MBB) {
433   if (!MBB.isEHPad())
434     return;
435 
436   // catch_all terminate pad is grouped together with catch terminate pad and
437   // does not need a separate TRY and END_TRY marker.
438   if (WebAssembly::isCatchAllTerminatePad(MBB))
439     return;
440 
441   MachineFunction &MF = *MBB.getParent();
442   auto &MDT = getAnalysis<MachineDominatorTree>();
443   const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
444   const auto &WEI = getAnalysis<WebAssemblyExceptionInfo>();
445   const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
446 
447   // Compute the nearest common dominator of all unwind predecessors
448   MachineBasicBlock *Header = nullptr;
449   int MBBNumber = MBB.getNumber();
450   for (auto *Pred : MBB.predecessors()) {
451     if (Pred->getNumber() < MBBNumber) {
452       Header = Header ? MDT.findNearestCommonDominator(Header, Pred) : Pred;
453       assert(!ExplicitlyBranchesTo(Pred, &MBB) &&
454              "Explicit branch to an EH pad!");
455     }
456   }
457   if (!Header)
458     return;
459 
460   // If this try is at the bottom of the function, insert a dummy block at the
461   // end.
462   WebAssemblyException *WE = WEI.getExceptionFor(&MBB);
463   assert(WE);
464   MachineBasicBlock *Bottom = WebAssembly::getBottom(WE);
465 
466   auto Iter = std::next(MachineFunction::iterator(Bottom));
467   if (Iter == MF.end()) {
468     MachineBasicBlock *Label = MF.CreateMachineBasicBlock();
469     // Give it a fake predecessor so that AsmPrinter prints its label.
470     Label->addSuccessor(Label);
471     MF.push_back(Label);
472     Iter = std::next(MachineFunction::iterator(Bottom));
473   }
474   MachineBasicBlock *AfterTry = &*Iter;
475 
476   assert(AfterTry != &MF.front());
477   MachineBasicBlock *LayoutPred =
478       &*std::prev(MachineFunction::iterator(AfterTry));
479 
480   // If the nearest common dominator is inside a more deeply nested context,
481   // walk out to the nearest scope which isn't more deeply nested.
482   for (MachineFunction::iterator I(LayoutPred), E(Header); I != E; --I) {
483     if (MachineBasicBlock *ScopeTop = ScopeTops[I->getNumber()]) {
484       if (ScopeTop->getNumber() > Header->getNumber()) {
485         // Skip over an intervening scope.
486         I = std::next(MachineFunction::iterator(ScopeTop));
487       } else {
488         // We found a scope level at an appropriate depth.
489         Header = ScopeTop;
490         break;
491       }
492     }
493   }
494 
495   // Decide where in Header to put the TRY.
496 
497   // Instructions that should go before the BLOCK.
498   SmallPtrSet<const MachineInstr *, 4> BeforeSet;
499   // Instructions that should go after the BLOCK.
500   SmallPtrSet<const MachineInstr *, 4> AfterSet;
501   for (const auto &MI : *Header) {
502     // If there is a previously placed LOOP marker and the bottom block of
503     // the loop is above MBB, the LOOP should be after the TRY, because the
504     // loop is nested in this try. Otherwise it should be before the TRY.
505     if (MI.getOpcode() == WebAssembly::LOOP) {
506       if (MBB.getNumber() > Bottom->getNumber())
507         AfterSet.insert(&MI);
508 #ifndef NDEBUG
509       else
510         BeforeSet.insert(&MI);
511 #endif
512     }
513 
514     // All previously inserted TRY markers should be after the TRY because they
515     // are all nested trys.
516     if (MI.getOpcode() == WebAssembly::TRY)
517       AfterSet.insert(&MI);
518 
519 #ifndef NDEBUG
520     // All END_(LOOP/TRY) markers should be before the TRY.
521     if (MI.getOpcode() == WebAssembly::END_LOOP ||
522         MI.getOpcode() == WebAssembly::END_TRY)
523       BeforeSet.insert(&MI);
524 #endif
525 
526     // Terminators should go after the TRY.
527     if (MI.isTerminator())
528       AfterSet.insert(&MI);
529   }
530 
531   // Local expression tree should go after the TRY.
532   for (auto I = Header->getFirstTerminator(), E = Header->begin(); I != E;
533        --I) {
534     if (WebAssembly::isChild(*std::prev(I), MFI))
535       AfterSet.insert(&*std::prev(I));
536     else
537       break;
538   }
539 
540   // If Header unwinds to MBB (= Header contains 'invoke'), the try block should
541   // contain the call within it. So the call should go after the TRY. The
542   // exception is when the header's terminator is a rethrow instruction, in
543   // which case that instruction, not a call instruction before it, is gonna
544   // throw.
545   if (MBB.isPredecessor(Header)) {
546     auto TermPos = Header->getFirstTerminator();
547     if (TermPos == Header->end() || !WebAssembly::isRethrow(*TermPos)) {
548       for (const auto &MI : reverse(*Header)) {
549         if (MI.isCall()) {
550           AfterSet.insert(&MI);
551           break;
552         }
553       }
554     }
555   }
556 
557   // Add the TRY.
558   auto InsertPos = GetLatestInsertPos(Header, BeforeSet, AfterSet);
559   MachineInstr *Begin =
560       BuildMI(*Header, InsertPos, Header->findDebugLoc(InsertPos),
561               TII.get(WebAssembly::TRY))
562           .addImm(int64_t(WebAssembly::ExprType::Void));
563 
564   // Decide where in Header to put the END_TRY.
565   BeforeSet.clear();
566   AfterSet.clear();
567   for (const auto &MI : *AfterTry) {
568 #ifndef NDEBUG
569     // END_TRY should precede existing LOOP markers.
570     if (MI.getOpcode() == WebAssembly::LOOP)
571       AfterSet.insert(&MI);
572 
573     // All END_TRY markers placed earlier belong to exceptions that contains
574     // this one.
575     if (MI.getOpcode() == WebAssembly::END_TRY)
576       AfterSet.insert(&MI);
577 #endif
578 
579     // If there is a previously placed END_LOOP marker and its header is after
580     // where TRY marker is, this loop is contained within the 'catch' part, so
581     // the END_TRY marker should go after that. Otherwise, the whole try-catch
582     // is contained within this loop, so the END_TRY should go before that.
583     if (MI.getOpcode() == WebAssembly::END_LOOP) {
584       if (EndToBegin[&MI]->getParent()->getNumber() >= Header->getNumber())
585         BeforeSet.insert(&MI);
586 #ifndef NDEBUG
587       else
588         AfterSet.insert(&MI);
589 #endif
590     }
591   }
592 
593   // Mark the end of the TRY.
594   InsertPos = GetEarliestInsertPos(AfterTry, BeforeSet, AfterSet);
595   MachineInstr *End =
596       BuildMI(*AfterTry, InsertPos, Bottom->findBranchDebugLoc(),
597               TII.get(WebAssembly::END_TRY));
598   registerTryScope(Begin, End, &MBB);
599 
600   // Track the farthest-spanning scope that ends at this point.
601   int Number = AfterTry->getNumber();
602   if (!ScopeTops[Number] ||
603       ScopeTops[Number]->getNumber() > Header->getNumber())
604     ScopeTops[Number] = Header;
605 }
606 
607 static unsigned
608 GetDepth(const SmallVectorImpl<const MachineBasicBlock *> &Stack,
609          const MachineBasicBlock *MBB) {
610   unsigned Depth = 0;
611   for (auto X : reverse(Stack)) {
612     if (X == MBB)
613       break;
614     ++Depth;
615   }
616   assert(Depth < Stack.size() && "Branch destination should be in scope");
617   return Depth;
618 }
619 
620 /// In normal assembly languages, when the end of a function is unreachable,
621 /// because the function ends in an infinite loop or a noreturn call or similar,
622 /// it isn't necessary to worry about the function return type at the end of
623 /// the function, because it's never reached. However, in WebAssembly, blocks
624 /// that end at the function end need to have a return type signature that
625 /// matches the function signature, even though it's unreachable. This function
626 /// checks for such cases and fixes up the signatures.
627 void WebAssemblyCFGStackify::fixEndsAtEndOfFunction(MachineFunction &MF) {
628   const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
629   assert(MFI.getResults().size() <= 1);
630 
631   if (MFI.getResults().empty())
632     return;
633 
634   WebAssembly::ExprType retType;
635   switch (MFI.getResults().front().SimpleTy) {
636   case MVT::i32: retType = WebAssembly::ExprType::I32; break;
637   case MVT::i64: retType = WebAssembly::ExprType::I64; break;
638   case MVT::f32: retType = WebAssembly::ExprType::F32; break;
639   case MVT::f64: retType = WebAssembly::ExprType::F64; break;
640   case MVT::v16i8:
641   case MVT::v8i16:
642   case MVT::v4i32:
643   case MVT::v2i64:
644   case MVT::v4f32:
645   case MVT::v2f64:
646     retType = WebAssembly::ExprType::V128;
647     break;
648   case MVT::ExceptRef: retType = WebAssembly::ExprType::ExceptRef; break;
649   default: llvm_unreachable("unexpected return type");
650   }
651 
652   for (MachineBasicBlock &MBB : reverse(MF)) {
653     for (MachineInstr &MI : reverse(MBB)) {
654       if (MI.isPosition() || MI.isDebugInstr())
655         continue;
656       if (MI.getOpcode() == WebAssembly::END_BLOCK) {
657         EndToBegin[&MI]->getOperand(0).setImm(int32_t(retType));
658         continue;
659       }
660       if (MI.getOpcode() == WebAssembly::END_LOOP) {
661         EndToBegin[&MI]->getOperand(0).setImm(int32_t(retType));
662         continue;
663       }
664       // Something other than an `end`. We're done.
665       return;
666     }
667   }
668 }
669 
670 // WebAssembly functions end with an end instruction, as if the function body
671 // were a block.
672 static void AppendEndToFunction(
673     MachineFunction &MF,
674     const WebAssemblyInstrInfo &TII) {
675   BuildMI(MF.back(), MF.back().end(),
676           MF.back().findPrevDebugLoc(MF.back().end()),
677           TII.get(WebAssembly::END_FUNCTION));
678 }
679 
680 /// Insert LOOP/TRY/BLOCK markers at appropriate places.
681 void WebAssemblyCFGStackify::placeMarkers(MachineFunction &MF) {
682   const MCAsmInfo *MCAI = MF.getTarget().getMCAsmInfo();
683   // We allocate one more than the number of blocks in the function to
684   // accommodate for the possible fake block we may insert at the end.
685   ScopeTops.resize(MF.getNumBlockIDs() + 1);
686   // Place the LOOP for MBB if MBB is the header of a loop.
687   for (auto &MBB : MF)
688     placeLoopMarker(MBB);
689   // Place the TRY for MBB if MBB is the EH pad of an exception.
690   if (MCAI->getExceptionHandlingType() == ExceptionHandling::Wasm &&
691       MF.getFunction().hasPersonalityFn())
692     for (auto &MBB : MF)
693       placeTryMarker(MBB);
694   // Place the BLOCK for MBB if MBB is branched to from above.
695   for (auto &MBB : MF)
696     placeBlockMarker(MBB);
697 }
698 
699 void WebAssemblyCFGStackify::rewriteDepthImmediates(MachineFunction &MF) {
700   const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
701   // Now rewrite references to basic blocks to be depth immediates.
702   // We need two stacks: one for normal scopes and the other for EH pad scopes.
703   // EH pad stack is used to rewrite depths in rethrow instructions.
704   SmallVector<const MachineBasicBlock *, 8> Stack;
705   SmallVector<const MachineBasicBlock *, 8> EHPadStack;
706   for (auto &MBB : reverse(MF)) {
707     for (auto I = MBB.rbegin(), E = MBB.rend(); I != E; ++I) {
708       MachineInstr &MI = *I;
709       switch (MI.getOpcode()) {
710       case WebAssembly::BLOCK:
711         assert(ScopeTops[Stack.back()->getNumber()]->getNumber() <=
712                    MBB.getNumber() &&
713                "Block/try should be balanced");
714         Stack.pop_back();
715         break;
716 
717       case WebAssembly::TRY:
718         assert(ScopeTops[Stack.back()->getNumber()]->getNumber() <=
719                    MBB.getNumber() &&
720                "Block/try marker should be balanced");
721         Stack.pop_back();
722         EHPadStack.pop_back();
723         break;
724 
725       case WebAssembly::CATCH_I32:
726       case WebAssembly::CATCH_I64:
727       case WebAssembly::CATCH_ALL:
728         EHPadStack.push_back(&MBB);
729         break;
730 
731       case WebAssembly::LOOP:
732         assert(Stack.back() == &MBB && "Loop top should be balanced");
733         Stack.pop_back();
734         break;
735 
736       case WebAssembly::END_BLOCK:
737       case WebAssembly::END_TRY:
738         Stack.push_back(&MBB);
739         break;
740 
741       case WebAssembly::END_LOOP:
742         Stack.push_back(EndToBegin[&MI]->getParent());
743         break;
744 
745       case WebAssembly::RETHROW: {
746         // Rewrite MBB operands to be depth immediates.
747         unsigned EHPadDepth = GetDepth(EHPadStack, MI.getOperand(0).getMBB());
748         MI.RemoveOperand(0);
749         MI.addOperand(MF, MachineOperand::CreateImm(EHPadDepth));
750         break;
751       }
752 
753       case WebAssembly::RETHROW_TO_CALLER: {
754         MachineInstr *Rethrow =
755             BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(WebAssembly::RETHROW))
756                 .addImm(Stack.size());
757         MI.eraseFromParent();
758         I = MachineBasicBlock::reverse_iterator(Rethrow);
759         break;
760       }
761 
762       default:
763         if (MI.isTerminator()) {
764           // Rewrite MBB operands to be depth immediates.
765           SmallVector<MachineOperand, 4> Ops(MI.operands());
766           while (MI.getNumOperands() > 0)
767             MI.RemoveOperand(MI.getNumOperands() - 1);
768           for (auto MO : Ops) {
769             if (MO.isMBB())
770               MO = MachineOperand::CreateImm(GetDepth(Stack, MO.getMBB()));
771             MI.addOperand(MF, MO);
772           }
773         }
774         break;
775       }
776     }
777   }
778   assert(Stack.empty() && "Control flow should be balanced");
779 }
780 
781 void WebAssemblyCFGStackify::releaseMemory() {
782   ScopeTops.clear();
783   BeginToEnd.clear();
784   EndToBegin.clear();
785   TryToEHPad.clear();
786   EHPadToTry.clear();
787   BeginToBottom.clear();
788 }
789 
790 bool WebAssemblyCFGStackify::runOnMachineFunction(MachineFunction &MF) {
791   LLVM_DEBUG(dbgs() << "********** CFG Stackifying **********\n"
792                        "********** Function: "
793                     << MF.getName() << '\n');
794 
795   releaseMemory();
796 
797   // Liveness is not tracked for VALUE_STACK physreg.
798   MF.getRegInfo().invalidateLiveness();
799 
800   // Place the BLOCK/LOOP/TRY markers to indicate the beginnings of scopes.
801   placeMarkers(MF);
802 
803   // Convert MBB operands in terminators to relative depth immediates.
804   rewriteDepthImmediates(MF);
805 
806   // Fix up block/loop/try signatures at the end of the function to conform to
807   // WebAssembly's rules.
808   fixEndsAtEndOfFunction(MF);
809 
810   // Add an end instruction at the end of the function body.
811   const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
812   if (!MF.getSubtarget<WebAssemblySubtarget>()
813            .getTargetTriple()
814            .isOSBinFormatELF())
815     AppendEndToFunction(MF, TII);
816 
817   return true;
818 }
819