/Users/buildslave/jenkins/sharedspace/clang-stage2-coverage-R@2/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===// |
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 | | #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" |
11 | | #include "llvm/ADT/DenseMap.h" |
12 | | #include "llvm/ADT/STLExtras.h" |
13 | | #include "llvm/ADT/Sequence.h" |
14 | | #include "llvm/ADT/SetVector.h" |
15 | | #include "llvm/ADT/SmallPtrSet.h" |
16 | | #include "llvm/ADT/SmallVector.h" |
17 | | #include "llvm/ADT/Statistic.h" |
18 | | #include "llvm/ADT/Twine.h" |
19 | | #include "llvm/Analysis/AssumptionCache.h" |
20 | | #include "llvm/Analysis/LoopAnalysisManager.h" |
21 | | #include "llvm/Analysis/LoopInfo.h" |
22 | | #include "llvm/Analysis/LoopPass.h" |
23 | | #include "llvm/IR/BasicBlock.h" |
24 | | #include "llvm/IR/Constant.h" |
25 | | #include "llvm/IR/Constants.h" |
26 | | #include "llvm/IR/Dominators.h" |
27 | | #include "llvm/IR/Function.h" |
28 | | #include "llvm/IR/InstrTypes.h" |
29 | | #include "llvm/IR/Instruction.h" |
30 | | #include "llvm/IR/Instructions.h" |
31 | | #include "llvm/IR/Use.h" |
32 | | #include "llvm/IR/Value.h" |
33 | | #include "llvm/Pass.h" |
34 | | #include "llvm/Support/Casting.h" |
35 | | #include "llvm/Support/Debug.h" |
36 | | #include "llvm/Support/ErrorHandling.h" |
37 | | #include "llvm/Support/GenericDomTree.h" |
38 | | #include "llvm/Support/raw_ostream.h" |
39 | | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
40 | | #include "llvm/Transforms/Utils/LoopUtils.h" |
41 | | #include <algorithm> |
42 | | #include <cassert> |
43 | | #include <iterator> |
44 | | #include <utility> |
45 | | |
46 | | #define DEBUG_TYPE "simple-loop-unswitch" |
47 | | |
48 | | using namespace llvm; |
49 | | |
50 | | STATISTIC(NumBranches, "Number of branches unswitched"); |
51 | | STATISTIC(NumSwitches, "Number of switches unswitched"); |
52 | | STATISTIC(NumTrivial, "Number of unswitches that are trivial"); |
53 | | |
54 | | static void replaceLoopUsesWithConstant(Loop &L, Value &LIC, |
55 | 13 | Constant &Replacement) { |
56 | 13 | assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); |
57 | 13 | |
58 | 13 | // Replace uses of LIC in the loop with the given constant. |
59 | 26 | for (auto UI = LIC.use_begin(), UE = LIC.use_end(); UI != UE26 ;) { |
60 | 13 | // Grab the use and walk past it so we can clobber it in the use list. |
61 | 13 | Use *U = &*UI++; |
62 | 13 | Instruction *UserI = dyn_cast<Instruction>(U->getUser()); |
63 | 13 | if (!UserI || 13 !L.contains(UserI)13 ) |
64 | 13 | continue; |
65 | 0 |
|
66 | 0 | // Replace this use within the loop body. |
67 | 0 | *U = &Replacement; |
68 | 0 | } |
69 | 13 | } |
70 | | |
71 | | /// Update the dominator tree after removing one exiting predecessor of a loop |
72 | | /// exit block. |
73 | | static void updateLoopExitIDom(BasicBlock *LoopExitBB, Loop &L, |
74 | 5 | DominatorTree &DT) { |
75 | 5 | assert(pred_begin(LoopExitBB) != pred_end(LoopExitBB) && |
76 | 5 | "Cannot have empty predecessors of the loop exit block if we split " |
77 | 5 | "off a block to unswitch!"); |
78 | 5 | |
79 | 5 | BasicBlock *IDom = *pred_begin(LoopExitBB); |
80 | 5 | // Walk all of the other predecessors finding the nearest common dominator |
81 | 5 | // until all predecessors are covered or we reach the loop header. The loop |
82 | 5 | // header necessarily dominates all loop exit blocks in loop simplified form |
83 | 5 | // so we can early-exit the moment we hit that block. |
84 | 5 | for (auto PI = std::next(pred_begin(LoopExitBB)), PE = pred_end(LoopExitBB); |
85 | 5 | PI != PE && 5 IDom != L.getHeader()0 ; ++PI0 ) |
86 | 0 | IDom = DT.findNearestCommonDominator(IDom, *PI); |
87 | 5 | |
88 | 5 | DT.changeImmediateDominator(LoopExitBB, IDom); |
89 | 5 | } |
90 | | |
91 | | /// Update the dominator tree after unswitching a particular former exit block. |
92 | | /// |
93 | | /// This handles the full update of the dominator tree after hoisting a block |
94 | | /// that previously was an exit block (or split off of an exit block) up to be |
95 | | /// reached from the new immediate dominator of the preheader. |
96 | | /// |
97 | | /// The common case is simple -- we just move the unswitched block to have an |
98 | | /// immediate dominator of the old preheader. But in complex cases, there may |
99 | | /// be other blocks reachable from the unswitched block that are immediately |
100 | | /// dominated by some node between the unswitched one and the old preheader. |
101 | | /// All of these also need to be hoisted in the dominator tree. We also want to |
102 | | /// minimize queries to the dominator tree because each step of this |
103 | | /// invalidates any DFS numbers that would make queries fast. |
104 | | static void updateDTAfterUnswitch(BasicBlock *UnswitchedBB, BasicBlock *OldPH, |
105 | 69 | DominatorTree &DT) { |
106 | 69 | DomTreeNode *OldPHNode = DT[OldPH]; |
107 | 69 | DomTreeNode *UnswitchedNode = DT[UnswitchedBB]; |
108 | 69 | // If the dominator tree has already been updated for this unswitched node, |
109 | 69 | // we're done. This makes it easier to use this routine if there are multiple |
110 | 69 | // paths to the same unswitched destination. |
111 | 69 | if (UnswitchedNode->getIDom() == OldPHNode) |
112 | 33 | return; |
113 | 36 | |
114 | 36 | // First collect the domtree nodes that we are hoisting over. These are the |
115 | 36 | // set of nodes which may have children that need to be hoisted as well. |
116 | 36 | SmallPtrSet<DomTreeNode *, 4> DomChain; |
117 | 117 | for (auto *IDom = UnswitchedNode->getIDom(); IDom != OldPHNode; |
118 | 81 | IDom = IDom->getIDom()) |
119 | 81 | DomChain.insert(IDom); |
120 | 36 | |
121 | 36 | // The unswitched block ends up immediately dominated by the old preheader -- |
122 | 36 | // regardless of whether it is the loop exit block or split off of the loop |
123 | 36 | // exit block. |
124 | 36 | DT.changeImmediateDominator(UnswitchedNode, OldPHNode); |
125 | 36 | |
126 | 36 | // For everything that moves up the dominator tree, we need to examine the |
127 | 36 | // dominator frontier to see if it additionally should move up the dominator |
128 | 36 | // tree. This lambda appends the dominator frontier for a node on the |
129 | 36 | // worklist. |
130 | 36 | // |
131 | 36 | // Note that we don't currently use the IDFCalculator here for two reasons: |
132 | 36 | // 1) It computes dominator tree levels for the entire function on each run |
133 | 36 | // of 'compute'. While this isn't terrible, given that we expect to update |
134 | 36 | // relatively small subtrees of the domtree, it isn't necessarily the right |
135 | 36 | // tradeoff. |
136 | 36 | // 2) The interface doesn't fit this usage well. It doesn't operate in |
137 | 36 | // append-only, and builds several sets that we don't need. |
138 | 36 | // |
139 | 36 | // FIXME: Neither of these issues are a big deal and could be addressed with |
140 | 36 | // some amount of refactoring of IDFCalculator. That would allow us to share |
141 | 36 | // the core logic here (which is solving the same core problem). |
142 | 36 | SmallSetVector<BasicBlock *, 4> Worklist; |
143 | 36 | SmallVector<DomTreeNode *, 4> DomNodes; |
144 | 36 | SmallPtrSet<BasicBlock *, 4> DomSet; |
145 | 43 | auto AppendDomFrontier = [&](DomTreeNode *Node) { |
146 | 43 | assert(DomNodes.empty() && "Must start with no dominator nodes."); |
147 | 43 | assert(DomSet.empty() && "Must start with an empty dominator set."); |
148 | 43 | |
149 | 43 | // First flatten this subtree into sequence of nodes by doing a pre-order |
150 | 43 | // walk. |
151 | 43 | DomNodes.push_back(Node); |
152 | 43 | // We intentionally re-evaluate the size as each node can add new children. |
153 | 43 | // Because this is a tree walk, this cannot add any duplicates. |
154 | 90 | for (int i = 0; i < (int)DomNodes.size()90 ; ++i47 ) |
155 | 47 | DomNodes.insert(DomNodes.end(), DomNodes[i]->begin(), DomNodes[i]->end()); |
156 | 43 | |
157 | 43 | // Now create a set of the basic blocks so we can quickly test for |
158 | 43 | // dominated successors. We could in theory use the DFS numbers of the |
159 | 43 | // dominator tree for this, but we want this to remain predictably fast |
160 | 43 | // even while we mutate the dominator tree in ways that would invalidate |
161 | 43 | // the DFS numbering. |
162 | 43 | for (DomTreeNode *InnerN : DomNodes) |
163 | 47 | DomSet.insert(InnerN->getBlock()); |
164 | 43 | |
165 | 43 | // Now re-walk the nodes, appending every successor of every node that isn't |
166 | 43 | // in the set. Note that we don't append the node itself, even though if it |
167 | 43 | // is a successor it does not strictly dominate itself and thus it would be |
168 | 43 | // part of the dominance frontier. The reason we don't append it is that |
169 | 43 | // the node passed in came *from* the worklist and so it has already been |
170 | 43 | // processed. |
171 | 43 | for (DomTreeNode *InnerN : DomNodes) |
172 | 47 | for (BasicBlock *SuccBB : successors(InnerN->getBlock())) |
173 | 15 | if (15 !DomSet.count(SuccBB)15 ) |
174 | 11 | Worklist.insert(SuccBB); |
175 | 43 | |
176 | 43 | DomNodes.clear(); |
177 | 43 | DomSet.clear(); |
178 | 43 | }; |
179 | 36 | |
180 | 36 | // Append the initial dom frontier nodes. |
181 | 36 | AppendDomFrontier(UnswitchedNode); |
182 | 36 | |
183 | 36 | // Walk the worklist. We grow the list in the loop and so must recompute size. |
184 | 47 | for (int i = 0; i < (int)Worklist.size()47 ; ++i11 ) { |
185 | 11 | auto *BB = Worklist[i]; |
186 | 11 | |
187 | 11 | DomTreeNode *Node = DT[BB]; |
188 | 11 | assert(!DomChain.count(Node) && |
189 | 11 | "Cannot be dominated by a block you can reach!"); |
190 | 11 | |
191 | 11 | // If this block had an immediate dominator somewhere in the chain |
192 | 11 | // we hoisted over, then its position in the domtree needs to move as it is |
193 | 11 | // reachable from a node hoisted over this chain. |
194 | 11 | if (!DomChain.count(Node->getIDom())) |
195 | 4 | continue; |
196 | 7 | |
197 | 7 | DT.changeImmediateDominator(Node, OldPHNode); |
198 | 7 | |
199 | 7 | // Now add this node's dominator frontier to the worklist as well. |
200 | 7 | AppendDomFrontier(Node); |
201 | 7 | } |
202 | 69 | } |
203 | | |
204 | | /// Check that all the LCSSA PHI nodes in the loop exit block have trivial |
205 | | /// incoming values along this edge. |
206 | | static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB, |
207 | 72 | BasicBlock &ExitBB) { |
208 | 88 | for (Instruction &I : ExitBB) { |
209 | 88 | auto *PN = dyn_cast<PHINode>(&I); |
210 | 88 | if (!PN) |
211 | 88 | // No more PHIs to check. |
212 | 69 | return true; |
213 | 19 | |
214 | 19 | // If the incoming value for this edge isn't loop invariant the unswitch |
215 | 19 | // won't be trivial. |
216 | 19 | if (19 !L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))19 ) |
217 | 3 | return false; |
218 | 0 | } |
219 | 0 | llvm_unreachable0 ("Basic blocks should never be empty!"); |
220 | 0 | } |
221 | | |
222 | | /// Rewrite the PHI nodes in an unswitched loop exit basic block. |
223 | | /// |
224 | | /// Requires that the loop exit and unswitched basic block are the same, and |
225 | | /// that the exiting block was a unique predecessor of that block. Rewrites the |
226 | | /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial |
227 | | /// PHI nodes from the old preheader that now contains the unswitched |
228 | | /// terminator. |
229 | | static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, |
230 | | BasicBlock &OldExitingBB, |
231 | 31 | BasicBlock &OldPH) { |
232 | 38 | for (Instruction &I : UnswitchedBB) { |
233 | 38 | auto *PN = dyn_cast<PHINode>(&I); |
234 | 38 | if (!PN) |
235 | 38 | // No more PHIs to check. |
236 | 31 | break; |
237 | 7 | |
238 | 7 | // When the loop exit is directly unswitched we just need to update the |
239 | 7 | // incoming basic block. We loop to handle weird cases with repeated |
240 | 7 | // incoming blocks, but expect to typically only have one operand here. |
241 | 7 | for (auto i : seq<int>(0, PN->getNumOperands())) 7 { |
242 | 9 | assert(PN->getIncomingBlock(i) == &OldExitingBB && |
243 | 9 | "Found incoming block different from unique predecessor!"); |
244 | 9 | PN->setIncomingBlock(i, &OldPH); |
245 | 9 | } |
246 | 38 | } |
247 | 31 | } |
248 | | |
249 | | /// Rewrite the PHI nodes in the loop exit basic block and the split off |
250 | | /// unswitched block. |
251 | | /// |
252 | | /// Because the exit block remains an exit from the loop, this rewrites the |
253 | | /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI |
254 | | /// nodes into the unswitched basic block to select between the value in the |
255 | | /// old preheader and the loop exit. |
256 | | static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, |
257 | | BasicBlock &UnswitchedBB, |
258 | | BasicBlock &OldExitingBB, |
259 | 5 | BasicBlock &OldPH) { |
260 | 5 | assert(&ExitBB != &UnswitchedBB && |
261 | 5 | "Must have different loop exit and unswitched blocks!"); |
262 | 5 | Instruction *InsertPt = &*UnswitchedBB.begin(); |
263 | 10 | for (Instruction &I : ExitBB) { |
264 | 10 | auto *PN = dyn_cast<PHINode>(&I); |
265 | 10 | if (!PN) |
266 | 10 | // No more PHIs to check. |
267 | 5 | break; |
268 | 5 | |
269 | 5 | auto *NewPN = PHINode::Create(PN->getType(), /*NumReservedValues*/ 2, |
270 | 5 | PN->getName() + ".split", InsertPt); |
271 | 5 | |
272 | 5 | // Walk backwards over the old PHI node's inputs to minimize the cost of |
273 | 5 | // removing each one. We have to do this weird loop manually so that we |
274 | 5 | // create the same number of new incoming edges in the new PHI as we expect |
275 | 5 | // each case-based edge to be included in the unswitched switch in some |
276 | 5 | // cases. |
277 | 5 | // FIXME: This is really, really gross. It would be much cleaner if LLVM |
278 | 5 | // allowed us to create a single entry for a predecessor block without |
279 | 5 | // having separate entries for each "edge" even though these edges are |
280 | 5 | // required to produce identical results. |
281 | 17 | for (int i = PN->getNumIncomingValues() - 1; i >= 017 ; --i12 ) { |
282 | 12 | if (PN->getIncomingBlock(i) != &OldExitingBB) |
283 | 5 | continue; |
284 | 7 | |
285 | 7 | Value *Incoming = PN->removeIncomingValue(i); |
286 | 7 | NewPN->addIncoming(Incoming, &OldPH); |
287 | 7 | } |
288 | 10 | |
289 | 10 | // Now replace the old PHI with the new one and wire the old one in as an |
290 | 10 | // input to the new one. |
291 | 10 | PN->replaceAllUsesWith(NewPN); |
292 | 10 | NewPN->addIncoming(PN, &ExitBB); |
293 | 10 | } |
294 | 5 | } |
295 | | |
296 | | /// Unswitch a trivial branch if the condition is loop invariant. |
297 | | /// |
298 | | /// This routine should only be called when loop code leading to the branch has |
299 | | /// been validated as trivial (no side effects). This routine checks if the |
300 | | /// condition is invariant and one of the successors is a loop exit. This |
301 | | /// allows us to unswitch without duplicating the loop, making it trivial. |
302 | | /// |
303 | | /// If this routine fails to unswitch the branch it returns false. |
304 | | /// |
305 | | /// If the branch can be unswitched, this routine splits the preheader and |
306 | | /// hoists the branch above that split. Preserves loop simplified form |
307 | | /// (splitting the exit block as necessary). It simplifies the branch within |
308 | | /// the loop to an unconditional branch but doesn't remove it entirely. Further |
309 | | /// cleanup can be done with some simplify-cfg like pass. |
310 | | static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, |
311 | 30 | LoopInfo &LI) { |
312 | 30 | assert(BI.isConditional() && "Can only unswitch a conditional branch!"); |
313 | 30 | DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n"); |
314 | 30 | |
315 | 30 | Value *LoopCond = BI.getCondition(); |
316 | 30 | |
317 | 30 | // Need a trivial loop condition to unswitch. |
318 | 30 | if (!L.isLoopInvariant(LoopCond)) |
319 | 11 | return false; |
320 | 19 | |
321 | 19 | // FIXME: We should compute this once at the start and update it! |
322 | 19 | SmallVector<BasicBlock *, 16> ExitBlocks; |
323 | 19 | L.getExitBlocks(ExitBlocks); |
324 | 19 | SmallPtrSet<BasicBlock *, 16> ExitBlockSet(ExitBlocks.begin(), |
325 | 19 | ExitBlocks.end()); |
326 | 19 | |
327 | 19 | // Check to see if a successor of the branch is guaranteed to |
328 | 19 | // exit through a unique exit block without having any |
329 | 19 | // side-effects. If so, determine the value of Cond that causes |
330 | 19 | // it to do this. |
331 | 19 | ConstantInt *CondVal = ConstantInt::getTrue(BI.getContext()); |
332 | 19 | ConstantInt *Replacement = ConstantInt::getFalse(BI.getContext()); |
333 | 19 | int LoopExitSuccIdx = 0; |
334 | 19 | auto *LoopExitBB = BI.getSuccessor(0); |
335 | 19 | if (!ExitBlockSet.count(LoopExitBB)19 ) { |
336 | 15 | std::swap(CondVal, Replacement); |
337 | 15 | LoopExitSuccIdx = 1; |
338 | 15 | LoopExitBB = BI.getSuccessor(1); |
339 | 15 | if (!ExitBlockSet.count(LoopExitBB)) |
340 | 4 | return false; |
341 | 15 | } |
342 | 15 | auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); |
343 | 15 | assert(L.contains(ContinueBB) && |
344 | 15 | "Cannot have both successors exit and still be in the loop!"); |
345 | 15 | |
346 | 15 | auto *ParentBB = BI.getParent(); |
347 | 15 | if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) |
348 | 2 | return false; |
349 | 13 | |
350 | 13 | DEBUG13 (dbgs() << " unswitching trivial branch when: " << CondVal |
351 | 13 | << " == " << LoopCond << "\n"); |
352 | 13 | |
353 | 13 | // Split the preheader, so that we know that there is a safe place to insert |
354 | 13 | // the conditional branch. We will change the preheader to have a conditional |
355 | 13 | // branch on LoopCond. |
356 | 13 | BasicBlock *OldPH = L.getLoopPreheader(); |
357 | 13 | BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI); |
358 | 13 | |
359 | 13 | // Now that we have a place to insert the conditional branch, create a place |
360 | 13 | // to branch to: this is the exit block out of the loop that we are |
361 | 13 | // unswitching. We need to split this if there are other loop predecessors. |
362 | 13 | // Because the loop is in simplified form, *any* other predecessor is enough. |
363 | 13 | BasicBlock *UnswitchedBB; |
364 | 13 | if (BasicBlock *PredBB13 = LoopExitBB->getUniquePredecessor()) { |
365 | 10 | (void)PredBB; |
366 | 10 | assert(PredBB == BI.getParent() && |
367 | 10 | "A branch's parent isn't a predecessor!"); |
368 | 10 | UnswitchedBB = LoopExitBB; |
369 | 13 | } else { |
370 | 3 | UnswitchedBB = SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI); |
371 | 3 | } |
372 | 13 | |
373 | 13 | // Now splice the branch to gate reaching the new preheader and re-point its |
374 | 13 | // successors. |
375 | 13 | OldPH->getInstList().splice(std::prev(OldPH->end()), |
376 | 13 | BI.getParent()->getInstList(), BI); |
377 | 13 | OldPH->getTerminator()->eraseFromParent(); |
378 | 13 | BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); |
379 | 13 | BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); |
380 | 13 | |
381 | 13 | // Create a new unconditional branch that will continue the loop as a new |
382 | 13 | // terminator. |
383 | 13 | BranchInst::Create(ContinueBB, ParentBB); |
384 | 13 | |
385 | 13 | // Rewrite the relevant PHI nodes. |
386 | 13 | if (UnswitchedBB == LoopExitBB) |
387 | 10 | rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); |
388 | 13 | else |
389 | 3 | rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, |
390 | 3 | *ParentBB, *OldPH); |
391 | 13 | |
392 | 13 | // Now we need to update the dominator tree. |
393 | 13 | updateDTAfterUnswitch(UnswitchedBB, OldPH, DT); |
394 | 13 | // But if we split something off of the loop exit block then we also removed |
395 | 13 | // one of the predecessors for the loop exit block and may need to update its |
396 | 13 | // idom. |
397 | 13 | if (UnswitchedBB != LoopExitBB) |
398 | 3 | updateLoopExitIDom(LoopExitBB, L, DT); |
399 | 30 | |
400 | 30 | // Since this is an i1 condition we can also trivially replace uses of it |
401 | 30 | // within the loop with a constant. |
402 | 30 | replaceLoopUsesWithConstant(L, *LoopCond, *Replacement); |
403 | 30 | |
404 | 30 | ++NumTrivial; |
405 | 30 | ++NumBranches; |
406 | 30 | return true; |
407 | 30 | } |
408 | | |
409 | | /// Unswitch a trivial switch if the condition is loop invariant. |
410 | | /// |
411 | | /// This routine should only be called when loop code leading to the switch has |
412 | | /// been validated as trivial (no side effects). This routine checks if the |
413 | | /// condition is invariant and that at least one of the successors is a loop |
414 | | /// exit. This allows us to unswitch without duplicating the loop, making it |
415 | | /// trivial. |
416 | | /// |
417 | | /// If this routine fails to unswitch the switch it returns false. |
418 | | /// |
419 | | /// If the switch can be unswitched, this routine splits the preheader and |
420 | | /// copies the switch above that split. If the default case is one of the |
421 | | /// exiting cases, it copies the non-exiting cases and points them at the new |
422 | | /// preheader. If the default case is not exiting, it copies the exiting cases |
423 | | /// and points the default at the preheader. It preserves loop simplified form |
424 | | /// (splitting the exit blocks as necessary). It simplifies the switch within |
425 | | /// the loop by removing now-dead cases. If the default case is one of those |
426 | | /// unswitched, it replaces its destination with a new basic block containing |
427 | | /// only unreachable. Such basic blocks, while technically loop exits, are not |
428 | | /// considered for unswitching so this is a stable transform and the same |
429 | | /// switch will not be revisited. If after unswitching there is only a single |
430 | | /// in-loop successor, the switch is further simplified to an unconditional |
431 | | /// branch. Still more cleanup can be done with some simplify-cfg like pass. |
432 | | static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, |
433 | 10 | LoopInfo &LI) { |
434 | 10 | DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n"); |
435 | 10 | Value *LoopCond = SI.getCondition(); |
436 | 10 | |
437 | 10 | // If this isn't switching on an invariant condition, we can't unswitch it. |
438 | 10 | if (!L.isLoopInvariant(LoopCond)) |
439 | 1 | return false; |
440 | 9 | |
441 | 9 | auto *ParentBB = SI.getParent(); |
442 | 9 | |
443 | 9 | // FIXME: We should compute this once at the start and update it! |
444 | 9 | SmallVector<BasicBlock *, 16> ExitBlocks; |
445 | 9 | L.getExitBlocks(ExitBlocks); |
446 | 9 | SmallPtrSet<BasicBlock *, 16> ExitBlockSet(ExitBlocks.begin(), |
447 | 9 | ExitBlocks.end()); |
448 | 9 | |
449 | 9 | SmallVector<int, 4> ExitCaseIndices; |
450 | 61 | for (auto Case : SI.cases()) { |
451 | 61 | auto *SuccBB = Case.getCaseSuccessor(); |
452 | 61 | if (ExitBlockSet.count(SuccBB) && |
453 | 53 | areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB)) |
454 | 52 | ExitCaseIndices.push_back(Case.getCaseIndex()); |
455 | 61 | } |
456 | 9 | BasicBlock *DefaultExitBB = nullptr; |
457 | 9 | if (ExitBlockSet.count(SI.getDefaultDest()) && |
458 | 4 | areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) && |
459 | 4 | !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator())) |
460 | 4 | DefaultExitBB = SI.getDefaultDest(); |
461 | 5 | else if (5 ExitCaseIndices.empty()5 ) |
462 | 1 | return false; |
463 | 8 | |
464 | 8 | DEBUG8 (dbgs() << " unswitching trivial cases...\n"); |
465 | 8 | |
466 | 8 | SmallVector<std::pair<ConstantInt *, BasicBlock *>, 4> ExitCases; |
467 | 8 | ExitCases.reserve(ExitCaseIndices.size()); |
468 | 8 | // We walk the case indices backwards so that we remove the last case first |
469 | 8 | // and don't disrupt the earlier indices. |
470 | 52 | for (unsigned Index : reverse(ExitCaseIndices)) { |
471 | 52 | auto CaseI = SI.case_begin() + Index; |
472 | 52 | // Save the value of this case. |
473 | 52 | ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()}); |
474 | 52 | // Delete the unswitched cases. |
475 | 52 | SI.removeCase(CaseI); |
476 | 52 | } |
477 | 8 | |
478 | 8 | // Check if after this all of the remaining cases point at the same |
479 | 8 | // successor. |
480 | 8 | BasicBlock *CommonSuccBB = nullptr; |
481 | 8 | if (SI.getNumCases() > 0 && |
482 | 4 | std::all_of(std::next(SI.case_begin()), SI.case_end(), |
483 | 3 | [&SI](const SwitchInst::CaseHandle &Case) { |
484 | 3 | return Case.getCaseSuccessor() == |
485 | 3 | SI.case_begin()->getCaseSuccessor(); |
486 | 3 | })) |
487 | 2 | CommonSuccBB = SI.case_begin()->getCaseSuccessor(); |
488 | 8 | |
489 | 8 | if (DefaultExitBB8 ) { |
490 | 4 | // We can't remove the default edge so replace it with an edge to either |
491 | 4 | // the single common remaining successor (if we have one) or an unreachable |
492 | 4 | // block. |
493 | 4 | if (CommonSuccBB4 ) { |
494 | 2 | SI.setDefaultDest(CommonSuccBB); |
495 | 4 | } else { |
496 | 2 | BasicBlock *UnreachableBB = BasicBlock::Create( |
497 | 2 | ParentBB->getContext(), |
498 | 2 | Twine(ParentBB->getName()) + ".unreachable_default", |
499 | 2 | ParentBB->getParent()); |
500 | 2 | new UnreachableInst(ParentBB->getContext(), UnreachableBB); |
501 | 2 | SI.setDefaultDest(UnreachableBB); |
502 | 2 | DT.addNewBlock(UnreachableBB, ParentBB); |
503 | 2 | } |
504 | 8 | } else { |
505 | 4 | // If we're not unswitching the default, we need it to match any cases to |
506 | 4 | // have a common successor or if we have no cases it is the common |
507 | 4 | // successor. |
508 | 4 | if (SI.getNumCases() == 0) |
509 | 4 | CommonSuccBB = SI.getDefaultDest(); |
510 | 0 | else if (0 SI.getDefaultDest() != CommonSuccBB0 ) |
511 | 0 | CommonSuccBB = nullptr; |
512 | 4 | } |
513 | 8 | |
514 | 8 | // Split the preheader, so that we know that there is a safe place to insert |
515 | 8 | // the switch. |
516 | 8 | BasicBlock *OldPH = L.getLoopPreheader(); |
517 | 8 | BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI); |
518 | 8 | OldPH->getTerminator()->eraseFromParent(); |
519 | 8 | |
520 | 8 | // Now add the unswitched switch. |
521 | 8 | auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); |
522 | 8 | |
523 | 8 | // Rewrite the IR for the unswitched basic blocks. This requires two steps. |
524 | 8 | // First, we split any exit blocks with remaining in-loop predecessors. Then |
525 | 8 | // we update the PHIs in one of two ways depending on if there was a split. |
526 | 8 | // We walk in reverse so that we split in the same order as the cases |
527 | 8 | // appeared. This is purely for convenience of reading the resulting IR, but |
528 | 8 | // it doesn't cost anything really. |
529 | 8 | SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; |
530 | 8 | SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; |
531 | 8 | // Handle the default exit if necessary. |
532 | 8 | // FIXME: It'd be great if we could merge this with the loop below but LLVM's |
533 | 8 | // ranges aren't quite powerful enough yet. |
534 | 8 | if (DefaultExitBB8 ) { |
535 | 4 | if (pred_empty(DefaultExitBB)4 ) { |
536 | 4 | UnswitchedExitBBs.insert(DefaultExitBB); |
537 | 4 | rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); |
538 | 4 | } else { |
539 | 0 | auto *SplitBB = |
540 | 0 | SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI); |
541 | 0 | rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB, |
542 | 0 | *ParentBB, *OldPH); |
543 | 0 | updateLoopExitIDom(DefaultExitBB, L, DT); |
544 | 0 | DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; |
545 | 0 | } |
546 | 4 | } |
547 | 8 | // Note that we must use a reference in the for loop so that we update the |
548 | 8 | // container. |
549 | 52 | for (auto &CasePair : reverse(ExitCases)) { |
550 | 52 | // Grab a reference to the exit block in the pair so that we can update it. |
551 | 52 | BasicBlock *ExitBB = CasePair.second; |
552 | 52 | |
553 | 52 | // If this case is the last edge into the exit block, we can simply reuse it |
554 | 52 | // as it will no longer be a loop exit. No mapping necessary. |
555 | 52 | if (pred_empty(ExitBB)52 ) { |
556 | 49 | // Only rewrite once. |
557 | 49 | if (UnswitchedExitBBs.insert(ExitBB).second) |
558 | 17 | rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); |
559 | 49 | continue; |
560 | 49 | } |
561 | 3 | |
562 | 3 | // Otherwise we need to split the exit block so that we retain an exit |
563 | 3 | // block from the loop and a target for the unswitched condition. |
564 | 3 | BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; |
565 | 3 | if (!SplitExitBB3 ) { |
566 | 2 | // If this is the first time we see this, do the split and remember it. |
567 | 2 | SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI); |
568 | 2 | rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB, |
569 | 2 | *ParentBB, *OldPH); |
570 | 2 | updateLoopExitIDom(ExitBB, L, DT); |
571 | 2 | } |
572 | 52 | // Update the case pair to point to the split block. |
573 | 52 | CasePair.second = SplitExitBB; |
574 | 52 | } |
575 | 8 | |
576 | 8 | // Now add the unswitched cases. We do this in reverse order as we built them |
577 | 8 | // in reverse order. |
578 | 52 | for (auto CasePair : reverse(ExitCases)) { |
579 | 52 | ConstantInt *CaseVal = CasePair.first; |
580 | 52 | BasicBlock *UnswitchedBB = CasePair.second; |
581 | 52 | |
582 | 52 | NewSI->addCase(CaseVal, UnswitchedBB); |
583 | 52 | updateDTAfterUnswitch(UnswitchedBB, OldPH, DT); |
584 | 52 | } |
585 | 8 | |
586 | 8 | // If the default was unswitched, re-point it and add explicit cases for |
587 | 8 | // entering the loop. |
588 | 8 | if (DefaultExitBB8 ) { |
589 | 4 | NewSI->setDefaultDest(DefaultExitBB); |
590 | 4 | updateDTAfterUnswitch(DefaultExitBB, OldPH, DT); |
591 | 4 | |
592 | 4 | // We removed all the exit cases, so we just copy the cases to the |
593 | 4 | // unswitched switch. |
594 | 4 | for (auto Case : SI.cases()) |
595 | 8 | NewSI->addCase(Case.getCaseValue(), NewPH); |
596 | 4 | } |
597 | 8 | |
598 | 8 | // If we ended up with a common successor for every path through the switch |
599 | 8 | // after unswitching, rewrite it to an unconditional branch to make it easy |
600 | 8 | // to recognize. Otherwise we potentially have to recognize the default case |
601 | 8 | // pointing at unreachable and other complexity. |
602 | 8 | if (CommonSuccBB8 ) { |
603 | 6 | BasicBlock *BB = SI.getParent(); |
604 | 6 | SI.eraseFromParent(); |
605 | 6 | BranchInst::Create(CommonSuccBB, BB); |
606 | 6 | } |
607 | 10 | |
608 | 10 | DT.verifyDomTree(); |
609 | 10 | ++NumTrivial; |
610 | 10 | ++NumSwitches; |
611 | 10 | return true; |
612 | 10 | } |
613 | | |
614 | | /// This routine scans the loop to find a branch or switch which occurs before |
615 | | /// any side effects occur. These can potentially be unswitched without |
616 | | /// duplicating the loop. If a branch or switch is successfully unswitched the |
617 | | /// scanning continues to see if subsequent branches or switches have become |
618 | | /// trivial. Once all trivial candidates have been unswitched, this routine |
619 | | /// returns. |
620 | | /// |
621 | | /// The return value indicates whether anything was unswitched (and therefore |
622 | | /// changed). |
623 | | static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, |
624 | 90 | LoopInfo &LI) { |
625 | 90 | bool Changed = false; |
626 | 90 | |
627 | 90 | // If loop header has only one reachable successor we should keep looking for |
628 | 90 | // trivial condition candidates in the successor as well. An alternative is |
629 | 90 | // to constant fold conditions and merge successors into loop header (then we |
630 | 90 | // only need to check header's terminator). The reason for not doing this in |
631 | 90 | // LoopUnswitch pass is that it could potentially break LoopPassManager's |
632 | 90 | // invariants. Folding dead branches could either eliminate the current loop |
633 | 90 | // or make other loops unreachable. LCSSA form might also not be preserved |
634 | 90 | // after deleting branches. The following code keeps traversing loop header's |
635 | 90 | // successors until it finds the trivial condition candidate (condition that |
636 | 90 | // is not a constant). Since unswitching generates branches with constant |
637 | 90 | // conditions, this scenario could be very common in practice. |
638 | 90 | BasicBlock *CurrentBB = L.getHeader(); |
639 | 90 | SmallPtrSet<BasicBlock *, 8> Visited; |
640 | 90 | Visited.insert(CurrentBB); |
641 | 109 | do { |
642 | 109 | // Check if there are any side-effecting instructions (e.g. stores, calls, |
643 | 109 | // volatile loads) in the part of the loop that the code *would* execute |
644 | 109 | // without unswitching. |
645 | 109 | if (llvm::any_of(*CurrentBB, |
646 | 267 | [](Instruction &I) { return I.mayHaveSideEffects(); })) |
647 | 43 | return Changed; |
648 | 66 | |
649 | 66 | TerminatorInst *CurrentTerm = CurrentBB->getTerminator(); |
650 | 66 | |
651 | 66 | if (auto *SI66 = dyn_cast<SwitchInst>(CurrentTerm)) { |
652 | 11 | // Don't bother trying to unswitch past a switch with a constant |
653 | 11 | // condition. This should be removed prior to running this pass by |
654 | 11 | // simplify-cfg. |
655 | 11 | if (isa<Constant>(SI->getCondition())) |
656 | 1 | return Changed; |
657 | 10 | |
658 | 10 | if (10 !unswitchTrivialSwitch(L, *SI, DT, LI)10 ) |
659 | 10 | // Coludn't unswitch this one so we're done. |
660 | 2 | return Changed; |
661 | 8 | |
662 | 8 | // Mark that we managed to unswitch something. |
663 | 8 | Changed = true; |
664 | 8 | |
665 | 8 | // If unswitching turned the terminator into an unconditional branch then |
666 | 8 | // we can continue. The unswitching logic specifically works to fold any |
667 | 8 | // cases it can into an unconditional branch to make it easier to |
668 | 8 | // recognize here. |
669 | 8 | auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); |
670 | 8 | if (!BI || 8 BI->isConditional()6 ) |
671 | 2 | return Changed; |
672 | 6 | |
673 | 6 | CurrentBB = BI->getSuccessor(0); |
674 | 6 | continue; |
675 | 6 | } |
676 | 55 | |
677 | 55 | auto *BI = dyn_cast<BranchInst>(CurrentTerm); |
678 | 55 | if (!BI) |
679 | 55 | // We do not understand other terminator instructions. |
680 | 1 | return Changed; |
681 | 54 | |
682 | 54 | // Don't bother trying to unswitch past an unconditional branch or a branch |
683 | 54 | // with a constant value. These should be removed by simplify-cfg prior to |
684 | 54 | // running this pass. |
685 | 54 | if (54 !BI->isConditional() || 54 isa<Constant>(BI->getCondition())43 ) |
686 | 24 | return Changed; |
687 | 30 | |
688 | 30 | // Found a trivial condition candidate: non-foldable conditional branch. If |
689 | 30 | // we fail to unswitch this, we can't do anything else that is trivial. |
690 | 30 | if (30 !unswitchTrivialBranch(L, *BI, DT, LI)30 ) |
691 | 17 | return Changed; |
692 | 13 | |
693 | 13 | // Mark that we managed to unswitch something. |
694 | 13 | Changed = true; |
695 | 13 | |
696 | 13 | // We unswitched the branch. This should always leave us with an |
697 | 13 | // unconditional branch that we can follow now. |
698 | 13 | BI = cast<BranchInst>(CurrentBB->getTerminator()); |
699 | 13 | assert(!BI->isConditional() && |
700 | 13 | "Cannot form a conditional branch by unswitching1"); |
701 | 13 | CurrentBB = BI->getSuccessor(0); |
702 | 13 | |
703 | 13 | // When continuing, if we exit the loop or reach a previous visited block, |
704 | 13 | // then we can not reach any trivial condition candidates (unfoldable |
705 | 13 | // branch instructions or switch instructions) and no unswitch can happen. |
706 | 90 | } while (L.contains(CurrentBB) && 19 Visited.insert(CurrentBB).second19 ); |
707 | 90 | |
708 | 0 | return Changed; |
709 | 90 | } |
710 | | |
711 | | /// Unswitch control flow predicated on loop invariant conditions. |
712 | | /// |
713 | | /// This first hoists all branches or switches which are trivial (IE, do not |
714 | | /// require duplicating any part of the loop) out of the loop body. It then |
715 | | /// looks at other loop invariant control flows and tries to unswitch those as |
716 | | /// well by cloning the loop if the result is small enough. |
717 | | static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, |
718 | 90 | AssumptionCache &AC) { |
719 | 90 | assert(L.isLCSSAForm(DT) && |
720 | 90 | "Loops must be in LCSSA form before unswitching."); |
721 | 90 | bool Changed = false; |
722 | 90 | |
723 | 90 | // Must be in loop simplified form: we need a preheader and dedicated exits. |
724 | 90 | if (!L.isLoopSimplifyForm()) |
725 | 0 | return false; |
726 | 90 | |
727 | 90 | // Try trivial unswitch first before loop over other basic blocks in the loop. |
728 | 90 | Changed |= unswitchAllTrivialConditions(L, DT, LI); |
729 | 90 | |
730 | 90 | // FIXME: Add support for non-trivial unswitching by cloning the loop. |
731 | 90 | |
732 | 90 | return Changed; |
733 | 90 | } |
734 | | |
735 | | PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, |
736 | | LoopStandardAnalysisResults &AR, |
737 | 50 | LPMUpdater &U) { |
738 | 50 | Function &F = *L.getHeader()->getParent(); |
739 | 50 | (void)F; |
740 | 50 | |
741 | 50 | DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L << "\n"); |
742 | 50 | |
743 | 50 | if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC)) |
744 | 40 | return PreservedAnalyses::all(); |
745 | 10 | |
746 | | #ifndef NDEBUG |
747 | | // Historically this pass has had issues with the dominator tree so verify it |
748 | | // in asserts builds. |
749 | | AR.DT.verifyDomTree(); |
750 | | #endif |
751 | 0 | return getLoopPassPreservedAnalyses(); |
752 | 10 | } |
753 | | |
754 | | namespace { |
755 | | |
756 | | class SimpleLoopUnswitchLegacyPass : public LoopPass { |
757 | | public: |
758 | | static char ID; // Pass ID, replacement for typeid |
759 | | |
760 | 24 | explicit SimpleLoopUnswitchLegacyPass() : LoopPass(ID) { |
761 | 24 | initializeSimpleLoopUnswitchLegacyPassPass( |
762 | 24 | *PassRegistry::getPassRegistry()); |
763 | 24 | } |
764 | | |
765 | | bool runOnLoop(Loop *L, LPPassManager &LPM) override; |
766 | | |
767 | 24 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
768 | 24 | AU.addRequired<AssumptionCacheTracker>(); |
769 | 24 | getLoopAnalysisUsage(AU); |
770 | 24 | } |
771 | | }; |
772 | | |
773 | | } // end anonymous namespace |
774 | | |
775 | 40 | bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { |
776 | 40 | if (skipLoop(L)) |
777 | 0 | return false; |
778 | 40 | |
779 | 40 | Function &F = *L->getHeader()->getParent(); |
780 | 40 | |
781 | 40 | DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L << "\n"); |
782 | 40 | |
783 | 40 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
784 | 40 | auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
785 | 40 | auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
786 | 40 | |
787 | 40 | bool Changed = unswitchLoop(*L, DT, LI, AC); |
788 | 40 | |
789 | | #ifndef NDEBUG |
790 | | // Historically this pass has had issues with the dominator tree so verify it |
791 | | // in asserts builds. |
792 | | DT.verifyDomTree(); |
793 | | #endif |
794 | | return Changed; |
795 | 40 | } |
796 | | |
797 | | char SimpleLoopUnswitchLegacyPass::ID = 0; |
798 | 24.6k | INITIALIZE_PASS_BEGIN24.6k (SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
|
799 | 24.6k | "Simple unswitch loops", false, false) |
800 | 24.6k | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
801 | 24.6k | INITIALIZE_PASS_DEPENDENCY(LoopPass) |
802 | 24.6k | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
803 | 24.6k | INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", |
804 | | "Simple unswitch loops", false, false) |
805 | | |
806 | 0 | Pass *llvm::createSimpleLoopUnswitchLegacyPass() { |
807 | 0 | return new SimpleLoopUnswitchLegacyPass(); |
808 | 0 | } |