Bug Summary

File:lib/Transforms/Scalar/SimpleLoopUnswitch.cpp
Warning:line 2472, column 24
The left operand of '>=' is a garbage value

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name SimpleLoopUnswitch.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn345461/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/lib/Transforms/Scalar -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-10-27-211344-32123-1 -x c++ /build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp -faddrsig
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/CFG.h"
21#include "llvm/Analysis/CodeMetrics.h"
22#include "llvm/Analysis/GuardUtils.h"
23#include "llvm/Analysis/InstructionSimplify.h"
24#include "llvm/Analysis/LoopAnalysisManager.h"
25#include "llvm/Analysis/LoopInfo.h"
26#include "llvm/Analysis/LoopIterator.h"
27#include "llvm/Analysis/LoopPass.h"
28#include "llvm/Analysis/Utils/Local.h"
29#include "llvm/IR/BasicBlock.h"
30#include "llvm/IR/Constant.h"
31#include "llvm/IR/Constants.h"
32#include "llvm/IR/Dominators.h"
33#include "llvm/IR/Function.h"
34#include "llvm/IR/InstrTypes.h"
35#include "llvm/IR/Instruction.h"
36#include "llvm/IR/Instructions.h"
37#include "llvm/IR/IntrinsicInst.h"
38#include "llvm/IR/Use.h"
39#include "llvm/IR/Value.h"
40#include "llvm/Pass.h"
41#include "llvm/Support/Casting.h"
42#include "llvm/Support/Debug.h"
43#include "llvm/Support/ErrorHandling.h"
44#include "llvm/Support/GenericDomTree.h"
45#include "llvm/Support/raw_ostream.h"
46#include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
47#include "llvm/Transforms/Utils/BasicBlockUtils.h"
48#include "llvm/Transforms/Utils/Cloning.h"
49#include "llvm/Transforms/Utils/LoopUtils.h"
50#include "llvm/Transforms/Utils/ValueMapper.h"
51#include <algorithm>
52#include <cassert>
53#include <iterator>
54#include <numeric>
55#include <utility>
56
57#define DEBUG_TYPE"simple-loop-unswitch" "simple-loop-unswitch"
58
59using namespace llvm;
60
61STATISTIC(NumBranches, "Number of branches unswitched")static llvm::Statistic NumBranches = {"simple-loop-unswitch",
"NumBranches", "Number of branches unswitched", {0}, {false}
}
;
62STATISTIC(NumSwitches, "Number of switches unswitched")static llvm::Statistic NumSwitches = {"simple-loop-unswitch",
"NumSwitches", "Number of switches unswitched", {0}, {false}
}
;
63STATISTIC(NumGuards, "Number of guards turned into branches for unswitching")static llvm::Statistic NumGuards = {"simple-loop-unswitch", "NumGuards"
, "Number of guards turned into branches for unswitching", {0
}, {false}}
;
64STATISTIC(NumTrivial, "Number of unswitches that are trivial")static llvm::Statistic NumTrivial = {"simple-loop-unswitch", "NumTrivial"
, "Number of unswitches that are trivial", {0}, {false}}
;
65
66static cl::opt<bool> EnableNonTrivialUnswitch(
67 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
68 cl::desc("Forcibly enables non-trivial loop unswitching rather than "
69 "following the configuration passed into the pass."));
70
71static cl::opt<int>
72 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
73 cl::desc("The cost threshold for unswitching a loop."));
74
75static cl::opt<bool> UnswitchGuards(
76 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
77 cl::desc("If enabled, simple loop unswitching will also consider "
78 "llvm.experimental.guard intrinsics as unswitch candidates."));
79
80/// Collect all of the loop invariant input values transitively used by the
81/// homogeneous instruction graph from a given root.
82///
83/// This essentially walks from a root recursively through loop variant operands
84/// which have the exact same opcode and finds all inputs which are loop
85/// invariant. For some operations these can be re-associated and unswitched out
86/// of the loop entirely.
87static TinyPtrVector<Value *>
88collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root,
89 LoopInfo &LI) {
90 assert(!L.isLoopInvariant(&Root) &&((!L.isLoopInvariant(&Root) && "Only need to walk the graph if root itself is not invariant."
) ? static_cast<void> (0) : __assert_fail ("!L.isLoopInvariant(&Root) && \"Only need to walk the graph if root itself is not invariant.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 91, __PRETTY_FUNCTION__))
91 "Only need to walk the graph if root itself is not invariant.")((!L.isLoopInvariant(&Root) && "Only need to walk the graph if root itself is not invariant."
) ? static_cast<void> (0) : __assert_fail ("!L.isLoopInvariant(&Root) && \"Only need to walk the graph if root itself is not invariant.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 91, __PRETTY_FUNCTION__))
;
92 TinyPtrVector<Value *> Invariants;
93
94 // Build a worklist and recurse through operators collecting invariants.
95 SmallVector<Instruction *, 4> Worklist;
96 SmallPtrSet<Instruction *, 8> Visited;
97 Worklist.push_back(&Root);
98 Visited.insert(&Root);
99 do {
100 Instruction &I = *Worklist.pop_back_val();
101 for (Value *OpV : I.operand_values()) {
102 // Skip constants as unswitching isn't interesting for them.
103 if (isa<Constant>(OpV))
104 continue;
105
106 // Add it to our result if loop invariant.
107 if (L.isLoopInvariant(OpV)) {
108 Invariants.push_back(OpV);
109 continue;
110 }
111
112 // If not an instruction with the same opcode, nothing we can do.
113 Instruction *OpI = dyn_cast<Instruction>(OpV);
114 if (!OpI || OpI->getOpcode() != Root.getOpcode())
115 continue;
116
117 // Visit this operand.
118 if (Visited.insert(OpI).second)
119 Worklist.push_back(OpI);
120 }
121 } while (!Worklist.empty());
122
123 return Invariants;
124}
125
126static void replaceLoopInvariantUses(Loop &L, Value *Invariant,
127 Constant &Replacement) {
128 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?")((!isa<Constant>(Invariant) && "Why are we unswitching on a constant?"
) ? static_cast<void> (0) : __assert_fail ("!isa<Constant>(Invariant) && \"Why are we unswitching on a constant?\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 128, __PRETTY_FUNCTION__))
;
129
130 // Replace uses of LIC in the loop with the given constant.
131 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); UI != UE;) {
132 // Grab the use and walk past it so we can clobber it in the use list.
133 Use *U = &*UI++;
134 Instruction *UserI = dyn_cast<Instruction>(U->getUser());
135
136 // Replace this use within the loop body.
137 if (UserI && L.contains(UserI))
138 U->set(&Replacement);
139 }
140}
141
142/// Check that all the LCSSA PHI nodes in the loop exit block have trivial
143/// incoming values along this edge.
144static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB,
145 BasicBlock &ExitBB) {
146 for (Instruction &I : ExitBB) {
147 auto *PN = dyn_cast<PHINode>(&I);
148 if (!PN)
149 // No more PHIs to check.
150 return true;
151
152 // If the incoming value for this edge isn't loop invariant the unswitch
153 // won't be trivial.
154 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
155 return false;
156 }
157 llvm_unreachable("Basic blocks should never be empty!")::llvm::llvm_unreachable_internal("Basic blocks should never be empty!"
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 157)
;
158}
159
160/// Insert code to test a set of loop invariant values, and conditionally branch
161/// on them.
162static void buildPartialUnswitchConditionalBranch(BasicBlock &BB,
163 ArrayRef<Value *> Invariants,
164 bool Direction,
165 BasicBlock &UnswitchedSucc,
166 BasicBlock &NormalSucc) {
167 IRBuilder<> IRB(&BB);
168 Value *Cond = Invariants.front();
169 for (Value *Invariant :
170 make_range(std::next(Invariants.begin()), Invariants.end()))
171 if (Direction)
172 Cond = IRB.CreateOr(Cond, Invariant);
173 else
174 Cond = IRB.CreateAnd(Cond, Invariant);
175
176 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
177 Direction ? &NormalSucc : &UnswitchedSucc);
178}
179
180/// Rewrite the PHI nodes in an unswitched loop exit basic block.
181///
182/// Requires that the loop exit and unswitched basic block are the same, and
183/// that the exiting block was a unique predecessor of that block. Rewrites the
184/// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
185/// PHI nodes from the old preheader that now contains the unswitched
186/// terminator.
187static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
188 BasicBlock &OldExitingBB,
189 BasicBlock &OldPH) {
190 for (PHINode &PN : UnswitchedBB.phis()) {
191 // When the loop exit is directly unswitched we just need to update the
192 // incoming basic block. We loop to handle weird cases with repeated
193 // incoming blocks, but expect to typically only have one operand here.
194 for (auto i : seq<int>(0, PN.getNumOperands())) {
195 assert(PN.getIncomingBlock(i) == &OldExitingBB &&((PN.getIncomingBlock(i) == &OldExitingBB && "Found incoming block different from unique predecessor!"
) ? static_cast<void> (0) : __assert_fail ("PN.getIncomingBlock(i) == &OldExitingBB && \"Found incoming block different from unique predecessor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 196, __PRETTY_FUNCTION__))
196 "Found incoming block different from unique predecessor!")((PN.getIncomingBlock(i) == &OldExitingBB && "Found incoming block different from unique predecessor!"
) ? static_cast<void> (0) : __assert_fail ("PN.getIncomingBlock(i) == &OldExitingBB && \"Found incoming block different from unique predecessor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 196, __PRETTY_FUNCTION__))
;
197 PN.setIncomingBlock(i, &OldPH);
198 }
199 }
200}
201
202/// Rewrite the PHI nodes in the loop exit basic block and the split off
203/// unswitched block.
204///
205/// Because the exit block remains an exit from the loop, this rewrites the
206/// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
207/// nodes into the unswitched basic block to select between the value in the
208/// old preheader and the loop exit.
209static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
210 BasicBlock &UnswitchedBB,
211 BasicBlock &OldExitingBB,
212 BasicBlock &OldPH,
213 bool FullUnswitch) {
214 assert(&ExitBB != &UnswitchedBB &&((&ExitBB != &UnswitchedBB && "Must have different loop exit and unswitched blocks!"
) ? static_cast<void> (0) : __assert_fail ("&ExitBB != &UnswitchedBB && \"Must have different loop exit and unswitched blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 215, __PRETTY_FUNCTION__))
215 "Must have different loop exit and unswitched blocks!")((&ExitBB != &UnswitchedBB && "Must have different loop exit and unswitched blocks!"
) ? static_cast<void> (0) : __assert_fail ("&ExitBB != &UnswitchedBB && \"Must have different loop exit and unswitched blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 215, __PRETTY_FUNCTION__))
;
216 Instruction *InsertPt = &*UnswitchedBB.begin();
217 for (PHINode &PN : ExitBB.phis()) {
218 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
219 PN.getName() + ".split", InsertPt);
220
221 // Walk backwards over the old PHI node's inputs to minimize the cost of
222 // removing each one. We have to do this weird loop manually so that we
223 // create the same number of new incoming edges in the new PHI as we expect
224 // each case-based edge to be included in the unswitched switch in some
225 // cases.
226 // FIXME: This is really, really gross. It would be much cleaner if LLVM
227 // allowed us to create a single entry for a predecessor block without
228 // having separate entries for each "edge" even though these edges are
229 // required to produce identical results.
230 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
231 if (PN.getIncomingBlock(i) != &OldExitingBB)
232 continue;
233
234 Value *Incoming = PN.getIncomingValue(i);
235 if (FullUnswitch)
236 // No more edge from the old exiting block to the exit block.
237 PN.removeIncomingValue(i);
238
239 NewPN->addIncoming(Incoming, &OldPH);
240 }
241
242 // Now replace the old PHI with the new one and wire the old one in as an
243 // input to the new one.
244 PN.replaceAllUsesWith(NewPN);
245 NewPN->addIncoming(&PN, &ExitBB);
246 }
247}
248
249/// Hoist the current loop up to the innermost loop containing a remaining exit.
250///
251/// Because we've removed an exit from the loop, we may have changed the set of
252/// loops reachable and need to move the current loop up the loop nest or even
253/// to an entirely separate nest.
254static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
255 DominatorTree &DT, LoopInfo &LI) {
256 // If the loop is already at the top level, we can't hoist it anywhere.
257 Loop *OldParentL = L.getParentLoop();
258 if (!OldParentL)
259 return;
260
261 SmallVector<BasicBlock *, 4> Exits;
262 L.getExitBlocks(Exits);
263 Loop *NewParentL = nullptr;
264 for (auto *ExitBB : Exits)
265 if (Loop *ExitL = LI.getLoopFor(ExitBB))
266 if (!NewParentL || NewParentL->contains(ExitL))
267 NewParentL = ExitL;
268
269 if (NewParentL == OldParentL)
270 return;
271
272 // The new parent loop (if different) should always contain the old one.
273 if (NewParentL)
274 assert(NewParentL->contains(OldParentL) &&((NewParentL->contains(OldParentL) && "Can only hoist this loop up the nest!"
) ? static_cast<void> (0) : __assert_fail ("NewParentL->contains(OldParentL) && \"Can only hoist this loop up the nest!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 275, __PRETTY_FUNCTION__))
275 "Can only hoist this loop up the nest!")((NewParentL->contains(OldParentL) && "Can only hoist this loop up the nest!"
) ? static_cast<void> (0) : __assert_fail ("NewParentL->contains(OldParentL) && \"Can only hoist this loop up the nest!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 275, __PRETTY_FUNCTION__))
;
276
277 // The preheader will need to move with the body of this loop. However,
278 // because it isn't in this loop we also need to update the primary loop map.
279 assert(OldParentL == LI.getLoopFor(&Preheader) &&((OldParentL == LI.getLoopFor(&Preheader) && "Parent loop of this loop should contain this loop's preheader!"
) ? static_cast<void> (0) : __assert_fail ("OldParentL == LI.getLoopFor(&Preheader) && \"Parent loop of this loop should contain this loop's preheader!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 280, __PRETTY_FUNCTION__))
280 "Parent loop of this loop should contain this loop's preheader!")((OldParentL == LI.getLoopFor(&Preheader) && "Parent loop of this loop should contain this loop's preheader!"
) ? static_cast<void> (0) : __assert_fail ("OldParentL == LI.getLoopFor(&Preheader) && \"Parent loop of this loop should contain this loop's preheader!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 280, __PRETTY_FUNCTION__))
;
281 LI.changeLoopFor(&Preheader, NewParentL);
282
283 // Remove this loop from its old parent.
284 OldParentL->removeChildLoop(&L);
285
286 // Add the loop either to the new parent or as a top-level loop.
287 if (NewParentL)
288 NewParentL->addChildLoop(&L);
289 else
290 LI.addTopLevelLoop(&L);
291
292 // Remove this loops blocks from the old parent and every other loop up the
293 // nest until reaching the new parent. Also update all of these
294 // no-longer-containing loops to reflect the nesting change.
295 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
296 OldContainingL = OldContainingL->getParentLoop()) {
297 llvm::erase_if(OldContainingL->getBlocksVector(),
298 [&](const BasicBlock *BB) {
299 return BB == &Preheader || L.contains(BB);
300 });
301
302 OldContainingL->getBlocksSet().erase(&Preheader);
303 for (BasicBlock *BB : L.blocks())
304 OldContainingL->getBlocksSet().erase(BB);
305
306 // Because we just hoisted a loop out of this one, we have essentially
307 // created new exit paths from it. That means we need to form LCSSA PHI
308 // nodes for values used in the no-longer-nested loop.
309 formLCSSA(*OldContainingL, DT, &LI, nullptr);
310
311 // We shouldn't need to form dedicated exits because the exit introduced
312 // here is the (just split by unswitching) preheader. However, after trivial
313 // unswitching it is possible to get new non-dedicated exits out of parent
314 // loop so let's conservatively form dedicated exit blocks and figure out
315 // if we can optimize later.
316 formDedicatedExitBlocks(OldContainingL, &DT, &LI, /*PreserveLCSSA*/ true);
317 }
318}
319
320/// Unswitch a trivial branch if the condition is loop invariant.
321///
322/// This routine should only be called when loop code leading to the branch has
323/// been validated as trivial (no side effects). This routine checks if the
324/// condition is invariant and one of the successors is a loop exit. This
325/// allows us to unswitch without duplicating the loop, making it trivial.
326///
327/// If this routine fails to unswitch the branch it returns false.
328///
329/// If the branch can be unswitched, this routine splits the preheader and
330/// hoists the branch above that split. Preserves loop simplified form
331/// (splitting the exit block as necessary). It simplifies the branch within
332/// the loop to an unconditional branch but doesn't remove it entirely. Further
333/// cleanup can be done with some simplify-cfg like pass.
334///
335/// If `SE` is not null, it will be updated based on the potential loop SCEVs
336/// invalidated by this.
337static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
338 LoopInfo &LI, ScalarEvolution *SE) {
339 assert(BI.isConditional() && "Can only unswitch a conditional branch!")((BI.isConditional() && "Can only unswitch a conditional branch!"
) ? static_cast<void> (0) : __assert_fail ("BI.isConditional() && \"Can only unswitch a conditional branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 339, __PRETTY_FUNCTION__))
;
340 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Trying to unswitch branch: "
<< BI << "\n"; } } while (false)
;
341
342 // The loop invariant values that we want to unswitch.
343 TinyPtrVector<Value *> Invariants;
344
345 // When true, we're fully unswitching the branch rather than just unswitching
346 // some input conditions to the branch.
347 bool FullUnswitch = false;
348
349 if (L.isLoopInvariant(BI.getCondition())) {
350 Invariants.push_back(BI.getCondition());
351 FullUnswitch = true;
352 } else {
353 if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition()))
354 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
355 if (Invariants.empty())
356 // Couldn't find invariant inputs!
357 return false;
358 }
359
360 // Check that one of the branch's successors exits, and which one.
361 bool ExitDirection = true;
362 int LoopExitSuccIdx = 0;
363 auto *LoopExitBB = BI.getSuccessor(0);
364 if (L.contains(LoopExitBB)) {
365 ExitDirection = false;
366 LoopExitSuccIdx = 1;
367 LoopExitBB = BI.getSuccessor(1);
368 if (L.contains(LoopExitBB))
369 return false;
370 }
371 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
372 auto *ParentBB = BI.getParent();
373 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB))
374 return false;
375
376 // When unswitching only part of the branch's condition, we need the exit
377 // block to be reached directly from the partially unswitched input. This can
378 // be done when the exit block is along the true edge and the branch condition
379 // is a graph of `or` operations, or the exit block is along the false edge
380 // and the condition is a graph of `and` operations.
381 if (!FullUnswitch) {
382 if (ExitDirection) {
383 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::Or)
384 return false;
385 } else {
386 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::And)
387 return false;
388 }
389 }
390
391 LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
392 dbgs() << " unswitching trivial invariant conditions for: " << BIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
393 << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
394 for (Value *Invariant : Invariants) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
395 dbgs() << " " << *Invariant << " == true";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
396 if (Invariant != Invariants.back())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
397 dbgs() << " ||";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
398 dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
399 }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
400 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { { dbgs() << " unswitching trivial invariant conditions for: "
<< BI << "\n"; for (Value *Invariant : Invariants
) { dbgs() << " " << *Invariant << " == true"
; if (Invariant != Invariants.back()) dbgs() << " ||"; dbgs
() << "\n"; } }; } } while (false)
;
401
402 // If we have scalar evolutions, we need to invalidate them including this
403 // loop and the loop containing the exit block.
404 if (SE) {
405 if (Loop *ExitL = LI.getLoopFor(LoopExitBB))
406 SE->forgetLoop(ExitL);
407 else
408 // Forget the entire nest as this exits the entire nest.
409 SE->forgetTopmostLoop(&L);
410 }
411
412 // Split the preheader, so that we know that there is a safe place to insert
413 // the conditional branch. We will change the preheader to have a conditional
414 // branch on LoopCond.
415 BasicBlock *OldPH = L.getLoopPreheader();
416 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI);
417
418 // Now that we have a place to insert the conditional branch, create a place
419 // to branch to: this is the exit block out of the loop that we are
420 // unswitching. We need to split this if there are other loop predecessors.
421 // Because the loop is in simplified form, *any* other predecessor is enough.
422 BasicBlock *UnswitchedBB;
423 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
424 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&((LoopExitBB->getUniquePredecessor() == BI.getParent() &&
"A branch's parent isn't a predecessor!") ? static_cast<void
> (0) : __assert_fail ("LoopExitBB->getUniquePredecessor() == BI.getParent() && \"A branch's parent isn't a predecessor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 425, __PRETTY_FUNCTION__))
425 "A branch's parent isn't a predecessor!")((LoopExitBB->getUniquePredecessor() == BI.getParent() &&
"A branch's parent isn't a predecessor!") ? static_cast<void
> (0) : __assert_fail ("LoopExitBB->getUniquePredecessor() == BI.getParent() && \"A branch's parent isn't a predecessor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 425, __PRETTY_FUNCTION__))
;
426 UnswitchedBB = LoopExitBB;
427 } else {
428 UnswitchedBB = SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI);
429 }
430
431 // Actually move the invariant uses into the unswitched position. If possible,
432 // we do this by moving the instructions, but when doing partial unswitching
433 // we do it by building a new merge of the values in the unswitched position.
434 OldPH->getTerminator()->eraseFromParent();
435 if (FullUnswitch) {
436 // If fully unswitching, we can use the existing branch instruction.
437 // Splice it into the old PH to gate reaching the new preheader and re-point
438 // its successors.
439 OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(),
440 BI);
441 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
442 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
443
444 // Create a new unconditional branch that will continue the loop as a new
445 // terminator.
446 BranchInst::Create(ContinueBB, ParentBB);
447 } else {
448 // Only unswitching a subset of inputs to the condition, so we will need to
449 // build a new branch that merges the invariant inputs.
450 if (ExitDirection)
451 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Must have an `or` of `i1`s for the condition!"
) ? static_cast<void> (0) : __assert_fail ("cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Must have an `or` of `i1`s for the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 453, __PRETTY_FUNCTION__))
452 Instruction::Or &&((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Must have an `or` of `i1`s for the condition!"
) ? static_cast<void> (0) : __assert_fail ("cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Must have an `or` of `i1`s for the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 453, __PRETTY_FUNCTION__))
453 "Must have an `or` of `i1`s for the condition!")((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Must have an `or` of `i1`s for the condition!"
) ? static_cast<void> (0) : __assert_fail ("cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Must have an `or` of `i1`s for the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 453, __PRETTY_FUNCTION__))
;
454 else
455 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::And && "Must have an `and` of `i1`s for the condition!"
) ? static_cast<void> (0) : __assert_fail ("cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::And && \"Must have an `and` of `i1`s for the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 457, __PRETTY_FUNCTION__))
456 Instruction::And &&((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::And && "Must have an `and` of `i1`s for the condition!"
) ? static_cast<void> (0) : __assert_fail ("cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::And && \"Must have an `and` of `i1`s for the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 457, __PRETTY_FUNCTION__))
457 "Must have an `and` of `i1`s for the condition!")((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::And && "Must have an `and` of `i1`s for the condition!"
) ? static_cast<void> (0) : __assert_fail ("cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::And && \"Must have an `and` of `i1`s for the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 457, __PRETTY_FUNCTION__))
;
458 buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection,
459 *UnswitchedBB, *NewPH);
460 }
461
462 // Rewrite the relevant PHI nodes.
463 if (UnswitchedBB == LoopExitBB)
464 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
465 else
466 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
467 *ParentBB, *OldPH, FullUnswitch);
468
469 // Now we need to update the dominator tree.
470 SmallVector<DominatorTree::UpdateType, 2> DTUpdates;
471 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedBB});
472 if (FullUnswitch)
473 DTUpdates.push_back({DT.Delete, ParentBB, LoopExitBB});
474 DT.applyUpdates(DTUpdates);
475
476 // The constant we can replace all of our invariants with inside the loop
477 // body. If any of the invariants have a value other than this the loop won't
478 // be entered.
479 ConstantInt *Replacement = ExitDirection
480 ? ConstantInt::getFalse(BI.getContext())
481 : ConstantInt::getTrue(BI.getContext());
482
483 // Since this is an i1 condition we can also trivially replace uses of it
484 // within the loop with a constant.
485 for (Value *Invariant : Invariants)
486 replaceLoopInvariantUses(L, Invariant, *Replacement);
487
488 // If this was full unswitching, we may have changed the nesting relationship
489 // for this loop so hoist it to its correct parent if needed.
490 if (FullUnswitch)
491 hoistLoopToNewParent(L, *NewPH, DT, LI);
492
493 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " done: unswitching trivial branch...\n"
; } } while (false)
;
494 ++NumTrivial;
495 ++NumBranches;
496 return true;
497}
498
499/// Unswitch a trivial switch if the condition is loop invariant.
500///
501/// This routine should only be called when loop code leading to the switch has
502/// been validated as trivial (no side effects). This routine checks if the
503/// condition is invariant and that at least one of the successors is a loop
504/// exit. This allows us to unswitch without duplicating the loop, making it
505/// trivial.
506///
507/// If this routine fails to unswitch the switch it returns false.
508///
509/// If the switch can be unswitched, this routine splits the preheader and
510/// copies the switch above that split. If the default case is one of the
511/// exiting cases, it copies the non-exiting cases and points them at the new
512/// preheader. If the default case is not exiting, it copies the exiting cases
513/// and points the default at the preheader. It preserves loop simplified form
514/// (splitting the exit blocks as necessary). It simplifies the switch within
515/// the loop by removing now-dead cases. If the default case is one of those
516/// unswitched, it replaces its destination with a new basic block containing
517/// only unreachable. Such basic blocks, while technically loop exits, are not
518/// considered for unswitching so this is a stable transform and the same
519/// switch will not be revisited. If after unswitching there is only a single
520/// in-loop successor, the switch is further simplified to an unconditional
521/// branch. Still more cleanup can be done with some simplify-cfg like pass.
522///
523/// If `SE` is not null, it will be updated based on the potential loop SCEVs
524/// invalidated by this.
525static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
526 LoopInfo &LI, ScalarEvolution *SE) {
527 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Trying to unswitch switch: "
<< SI << "\n"; } } while (false)
;
528 Value *LoopCond = SI.getCondition();
529
530 // If this isn't switching on an invariant condition, we can't unswitch it.
531 if (!L.isLoopInvariant(LoopCond))
532 return false;
533
534 auto *ParentBB = SI.getParent();
535
536 SmallVector<int, 4> ExitCaseIndices;
537 for (auto Case : SI.cases()) {
538 auto *SuccBB = Case.getCaseSuccessor();
539 if (!L.contains(SuccBB) &&
540 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB))
541 ExitCaseIndices.push_back(Case.getCaseIndex());
542 }
543 BasicBlock *DefaultExitBB = nullptr;
544 if (!L.contains(SI.getDefaultDest()) &&
545 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) &&
546 !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator()))
547 DefaultExitBB = SI.getDefaultDest();
548 else if (ExitCaseIndices.empty())
549 return false;
550
551 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " unswitching trivial switch...\n"
; } } while (false)
;
552
553 // We may need to invalidate SCEVs for the outermost loop reached by any of
554 // the exits.
555 Loop *OuterL = &L;
556
557 if (DefaultExitBB) {
558 // Clear out the default destination temporarily to allow accurate
559 // predecessor lists to be examined below.
560 SI.setDefaultDest(nullptr);
561 // Check the loop containing this exit.
562 Loop *ExitL = LI.getLoopFor(DefaultExitBB);
563 if (!ExitL || ExitL->contains(OuterL))
564 OuterL = ExitL;
565 }
566
567 // Store the exit cases into a separate data structure and remove them from
568 // the switch.
569 SmallVector<std::pair<ConstantInt *, BasicBlock *>, 4> ExitCases;
570 ExitCases.reserve(ExitCaseIndices.size());
571 // We walk the case indices backwards so that we remove the last case first
572 // and don't disrupt the earlier indices.
573 for (unsigned Index : reverse(ExitCaseIndices)) {
574 auto CaseI = SI.case_begin() + Index;
575 // Compute the outer loop from this exit.
576 Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor());
577 if (!ExitL || ExitL->contains(OuterL))
578 OuterL = ExitL;
579 // Save the value of this case.
580 ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()});
581 // Delete the unswitched cases.
582 SI.removeCase(CaseI);
583 }
584
585 if (SE) {
586 if (OuterL)
587 SE->forgetLoop(OuterL);
588 else
589 SE->forgetTopmostLoop(&L);
590 }
591
592 // Check if after this all of the remaining cases point at the same
593 // successor.
594 BasicBlock *CommonSuccBB = nullptr;
595 if (SI.getNumCases() > 0 &&
596 std::all_of(std::next(SI.case_begin()), SI.case_end(),
597 [&SI](const SwitchInst::CaseHandle &Case) {
598 return Case.getCaseSuccessor() ==
599 SI.case_begin()->getCaseSuccessor();
600 }))
601 CommonSuccBB = SI.case_begin()->getCaseSuccessor();
602 if (!DefaultExitBB) {
603 // If we're not unswitching the default, we need it to match any cases to
604 // have a common successor or if we have no cases it is the common
605 // successor.
606 if (SI.getNumCases() == 0)
607 CommonSuccBB = SI.getDefaultDest();
608 else if (SI.getDefaultDest() != CommonSuccBB)
609 CommonSuccBB = nullptr;
610 }
611
612 // Split the preheader, so that we know that there is a safe place to insert
613 // the switch.
614 BasicBlock *OldPH = L.getLoopPreheader();
615 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI);
616 OldPH->getTerminator()->eraseFromParent();
617
618 // Now add the unswitched switch.
619 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
620
621 // Rewrite the IR for the unswitched basic blocks. This requires two steps.
622 // First, we split any exit blocks with remaining in-loop predecessors. Then
623 // we update the PHIs in one of two ways depending on if there was a split.
624 // We walk in reverse so that we split in the same order as the cases
625 // appeared. This is purely for convenience of reading the resulting IR, but
626 // it doesn't cost anything really.
627 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
628 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
629 // Handle the default exit if necessary.
630 // FIXME: It'd be great if we could merge this with the loop below but LLVM's
631 // ranges aren't quite powerful enough yet.
632 if (DefaultExitBB) {
633 if (pred_empty(DefaultExitBB)) {
634 UnswitchedExitBBs.insert(DefaultExitBB);
635 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
636 } else {
637 auto *SplitBB =
638 SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI);
639 rewritePHINodesForExitAndUnswitchedBlocks(
640 *DefaultExitBB, *SplitBB, *ParentBB, *OldPH, /*FullUnswitch*/ true);
641 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
642 }
643 }
644 // Note that we must use a reference in the for loop so that we update the
645 // container.
646 for (auto &CasePair : reverse(ExitCases)) {
647 // Grab a reference to the exit block in the pair so that we can update it.
648 BasicBlock *ExitBB = CasePair.second;
649
650 // If this case is the last edge into the exit block, we can simply reuse it
651 // as it will no longer be a loop exit. No mapping necessary.
652 if (pred_empty(ExitBB)) {
653 // Only rewrite once.
654 if (UnswitchedExitBBs.insert(ExitBB).second)
655 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
656 continue;
657 }
658
659 // Otherwise we need to split the exit block so that we retain an exit
660 // block from the loop and a target for the unswitched condition.
661 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
662 if (!SplitExitBB) {
663 // If this is the first time we see this, do the split and remember it.
664 SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI);
665 rewritePHINodesForExitAndUnswitchedBlocks(
666 *ExitBB, *SplitExitBB, *ParentBB, *OldPH, /*FullUnswitch*/ true);
667 }
668 // Update the case pair to point to the split block.
669 CasePair.second = SplitExitBB;
670 }
671
672 // Now add the unswitched cases. We do this in reverse order as we built them
673 // in reverse order.
674 for (auto CasePair : reverse(ExitCases)) {
675 ConstantInt *CaseVal = CasePair.first;
676 BasicBlock *UnswitchedBB = CasePair.second;
677
678 NewSI->addCase(CaseVal, UnswitchedBB);
679 }
680
681 // If the default was unswitched, re-point it and add explicit cases for
682 // entering the loop.
683 if (DefaultExitBB) {
684 NewSI->setDefaultDest(DefaultExitBB);
685
686 // We removed all the exit cases, so we just copy the cases to the
687 // unswitched switch.
688 for (auto Case : SI.cases())
689 NewSI->addCase(Case.getCaseValue(), NewPH);
690 }
691
692 // If we ended up with a common successor for every path through the switch
693 // after unswitching, rewrite it to an unconditional branch to make it easy
694 // to recognize. Otherwise we potentially have to recognize the default case
695 // pointing at unreachable and other complexity.
696 if (CommonSuccBB) {
697 BasicBlock *BB = SI.getParent();
698 // We may have had multiple edges to this common successor block, so remove
699 // them as predecessors. We skip the first one, either the default or the
700 // actual first case.
701 bool SkippedFirst = DefaultExitBB == nullptr;
702 for (auto Case : SI.cases()) {
703 assert(Case.getCaseSuccessor() == CommonSuccBB &&((Case.getCaseSuccessor() == CommonSuccBB && "Non-common successor!"
) ? static_cast<void> (0) : __assert_fail ("Case.getCaseSuccessor() == CommonSuccBB && \"Non-common successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 704, __PRETTY_FUNCTION__))
704 "Non-common successor!")((Case.getCaseSuccessor() == CommonSuccBB && "Non-common successor!"
) ? static_cast<void> (0) : __assert_fail ("Case.getCaseSuccessor() == CommonSuccBB && \"Non-common successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 704, __PRETTY_FUNCTION__))
;
705 (void)Case;
706 if (!SkippedFirst) {
707 SkippedFirst = true;
708 continue;
709 }
710 CommonSuccBB->removePredecessor(BB,
711 /*DontDeleteUselessPHIs*/ true);
712 }
713 // Now nuke the switch and replace it with a direct branch.
714 SI.eraseFromParent();
715 BranchInst::Create(CommonSuccBB, BB);
716 } else if (DefaultExitBB) {
717 assert(SI.getNumCases() > 0 &&((SI.getNumCases() > 0 && "If we had no cases we'd have a common successor!"
) ? static_cast<void> (0) : __assert_fail ("SI.getNumCases() > 0 && \"If we had no cases we'd have a common successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 718, __PRETTY_FUNCTION__))
718 "If we had no cases we'd have a common successor!")((SI.getNumCases() > 0 && "If we had no cases we'd have a common successor!"
) ? static_cast<void> (0) : __assert_fail ("SI.getNumCases() > 0 && \"If we had no cases we'd have a common successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 718, __PRETTY_FUNCTION__))
;
719 // Move the last case to the default successor. This is valid as if the
720 // default got unswitched it cannot be reached. This has the advantage of
721 // being simple and keeping the number of edges from this switch to
722 // successors the same, and avoiding any PHI update complexity.
723 auto LastCaseI = std::prev(SI.case_end());
724 SI.setDefaultDest(LastCaseI->getCaseSuccessor());
725 SI.removeCase(LastCaseI);
726 }
727
728 // Walk the unswitched exit blocks and the unswitched split blocks and update
729 // the dominator tree based on the CFG edits. While we are walking unordered
730 // containers here, the API for applyUpdates takes an unordered list of
731 // updates and requires them to not contain duplicates.
732 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
733 for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
734 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
735 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
736 }
737 for (auto SplitUnswitchedPair : SplitExitBBMap) {
738 auto *UnswitchedBB = SplitUnswitchedPair.second;
739 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedBB});
740 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedBB});
741 }
742 DT.applyUpdates(DTUpdates);
743 assert(DT.verify(DominatorTree::VerificationLevel::Fast))((DT.verify(DominatorTree::VerificationLevel::Fast)) ? static_cast
<void> (0) : __assert_fail ("DT.verify(DominatorTree::VerificationLevel::Fast)"
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 743, __PRETTY_FUNCTION__))
;
744
745 // We may have changed the nesting relationship for this loop so hoist it to
746 // its correct parent if needed.
747 hoistLoopToNewParent(L, *NewPH, DT, LI);
748
749 ++NumTrivial;
750 ++NumSwitches;
751 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " done: unswitching trivial switch...\n"
; } } while (false)
;
752 return true;
753}
754
755/// This routine scans the loop to find a branch or switch which occurs before
756/// any side effects occur. These can potentially be unswitched without
757/// duplicating the loop. If a branch or switch is successfully unswitched the
758/// scanning continues to see if subsequent branches or switches have become
759/// trivial. Once all trivial candidates have been unswitched, this routine
760/// returns.
761///
762/// The return value indicates whether anything was unswitched (and therefore
763/// changed).
764///
765/// If `SE` is not null, it will be updated based on the potential loop SCEVs
766/// invalidated by this.
767static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
768 LoopInfo &LI, ScalarEvolution *SE) {
769 bool Changed = false;
770
771 // If loop header has only one reachable successor we should keep looking for
772 // trivial condition candidates in the successor as well. An alternative is
773 // to constant fold conditions and merge successors into loop header (then we
774 // only need to check header's terminator). The reason for not doing this in
775 // LoopUnswitch pass is that it could potentially break LoopPassManager's
776 // invariants. Folding dead branches could either eliminate the current loop
777 // or make other loops unreachable. LCSSA form might also not be preserved
778 // after deleting branches. The following code keeps traversing loop header's
779 // successors until it finds the trivial condition candidate (condition that
780 // is not a constant). Since unswitching generates branches with constant
781 // conditions, this scenario could be very common in practice.
782 BasicBlock *CurrentBB = L.getHeader();
783 SmallPtrSet<BasicBlock *, 8> Visited;
784 Visited.insert(CurrentBB);
785 do {
786 // Check if there are any side-effecting instructions (e.g. stores, calls,
787 // volatile loads) in the part of the loop that the code *would* execute
788 // without unswitching.
789 if (llvm::any_of(*CurrentBB,
790 [](Instruction &I) { return I.mayHaveSideEffects(); }))
791 return Changed;
792
793 Instruction *CurrentTerm = CurrentBB->getTerminator();
794
795 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
796 // Don't bother trying to unswitch past a switch with a constant
797 // condition. This should be removed prior to running this pass by
798 // simplify-cfg.
799 if (isa<Constant>(SI->getCondition()))
800 return Changed;
801
802 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE))
803 // Couldn't unswitch this one so we're done.
804 return Changed;
805
806 // Mark that we managed to unswitch something.
807 Changed = true;
808
809 // If unswitching turned the terminator into an unconditional branch then
810 // we can continue. The unswitching logic specifically works to fold any
811 // cases it can into an unconditional branch to make it easier to
812 // recognize here.
813 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
814 if (!BI || BI->isConditional())
815 return Changed;
816
817 CurrentBB = BI->getSuccessor(0);
818 continue;
819 }
820
821 auto *BI = dyn_cast<BranchInst>(CurrentTerm);
822 if (!BI)
823 // We do not understand other terminator instructions.
824 return Changed;
825
826 // Don't bother trying to unswitch past an unconditional branch or a branch
827 // with a constant value. These should be removed by simplify-cfg prior to
828 // running this pass.
829 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
830 return Changed;
831
832 // Found a trivial condition candidate: non-foldable conditional branch. If
833 // we fail to unswitch this, we can't do anything else that is trivial.
834 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE))
835 return Changed;
836
837 // Mark that we managed to unswitch something.
838 Changed = true;
839
840 // If we only unswitched some of the conditions feeding the branch, we won't
841 // have collapsed it to a single successor.
842 BI = cast<BranchInst>(CurrentBB->getTerminator());
843 if (BI->isConditional())
844 return Changed;
845
846 // Follow the newly unconditional branch into its successor.
847 CurrentBB = BI->getSuccessor(0);
848
849 // When continuing, if we exit the loop or reach a previous visited block,
850 // then we can not reach any trivial condition candidates (unfoldable
851 // branch instructions or switch instructions) and no unswitch can happen.
852 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
853
854 return Changed;
855}
856
857/// Build the cloned blocks for an unswitched copy of the given loop.
858///
859/// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
860/// after the split block (`SplitBB`) that will be used to select between the
861/// cloned and original loop.
862///
863/// This routine handles cloning all of the necessary loop blocks and exit
864/// blocks including rewriting their instructions and the relevant PHI nodes.
865/// Any loop blocks or exit blocks which are dominated by a different successor
866/// than the one for this clone of the loop blocks can be trivially skipped. We
867/// use the `DominatingSucc` map to determine whether a block satisfies that
868/// property with a simple map lookup.
869///
870/// It also correctly creates the unconditional branch in the cloned
871/// unswitched parent block to only point at the unswitched successor.
872///
873/// This does not handle most of the necessary updates to `LoopInfo`. Only exit
874/// block splitting is correctly reflected in `LoopInfo`, essentially all of
875/// the cloned blocks (and their loops) are left without full `LoopInfo`
876/// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
877/// blocks to them but doesn't create the cloned `DominatorTree` structure and
878/// instead the caller must recompute an accurate DT. It *does* correctly
879/// update the `AssumptionCache` provided in `AC`.
880static BasicBlock *buildClonedLoopBlocks(
881 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
882 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
883 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
884 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
885 ValueToValueMapTy &VMap,
886 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
887 DominatorTree &DT, LoopInfo &LI) {
888 SmallVector<BasicBlock *, 4> NewBlocks;
889 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
890
891 // We will need to clone a bunch of blocks, wrap up the clone operation in
892 // a helper.
893 auto CloneBlock = [&](BasicBlock *OldBB) {
894 // Clone the basic block and insert it before the new preheader.
895 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
896 NewBB->moveBefore(LoopPH);
897
898 // Record this block and the mapping.
899 NewBlocks.push_back(NewBB);
900 VMap[OldBB] = NewBB;
901
902 return NewBB;
903 };
904
905 // We skip cloning blocks when they have a dominating succ that is not the
906 // succ we are cloning for.
907 auto SkipBlock = [&](BasicBlock *BB) {
908 auto It = DominatingSucc.find(BB);
909 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
910 };
911
912 // First, clone the preheader.
913 auto *ClonedPH = CloneBlock(LoopPH);
914
915 // Then clone all the loop blocks, skipping the ones that aren't necessary.
916 for (auto *LoopBB : L.blocks())
917 if (!SkipBlock(LoopBB))
918 CloneBlock(LoopBB);
919
920 // Split all the loop exit edges so that when we clone the exit blocks, if
921 // any of the exit blocks are *also* a preheader for some other loop, we
922 // don't create multiple predecessors entering the loop header.
923 for (auto *ExitBB : ExitBlocks) {
924 if (SkipBlock(ExitBB))
925 continue;
926
927 // When we are going to clone an exit, we don't need to clone all the
928 // instructions in the exit block and we want to ensure we have an easy
929 // place to merge the CFG, so split the exit first. This is always safe to
930 // do because there cannot be any non-loop predecessors of a loop exit in
931 // loop simplified form.
932 auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI);
933
934 // Rearrange the names to make it easier to write test cases by having the
935 // exit block carry the suffix rather than the merge block carrying the
936 // suffix.
937 MergeBB->takeName(ExitBB);
938 ExitBB->setName(Twine(MergeBB->getName()) + ".split");
939
940 // Now clone the original exit block.
941 auto *ClonedExitBB = CloneBlock(ExitBB);
942 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&((ClonedExitBB->getTerminator()->getNumSuccessors() == 1
&& "Exit block should have been split to have one successor!"
) ? static_cast<void> (0) : __assert_fail ("ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && \"Exit block should have been split to have one successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 943, __PRETTY_FUNCTION__))
943 "Exit block should have been split to have one successor!")((ClonedExitBB->getTerminator()->getNumSuccessors() == 1
&& "Exit block should have been split to have one successor!"
) ? static_cast<void> (0) : __assert_fail ("ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && \"Exit block should have been split to have one successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 943, __PRETTY_FUNCTION__))
;
944 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&((ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB
&& "Cloned exit block has the wrong successor!") ? static_cast
<void> (0) : __assert_fail ("ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && \"Cloned exit block has the wrong successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 945, __PRETTY_FUNCTION__))
945 "Cloned exit block has the wrong successor!")((ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB
&& "Cloned exit block has the wrong successor!") ? static_cast
<void> (0) : __assert_fail ("ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && \"Cloned exit block has the wrong successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 945, __PRETTY_FUNCTION__))
;
946
947 // Remap any cloned instructions and create a merge phi node for them.
948 for (auto ZippedInsts : llvm::zip_first(
949 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
950 llvm::make_range(ClonedExitBB->begin(),
951 std::prev(ClonedExitBB->end())))) {
952 Instruction &I = std::get<0>(ZippedInsts);
953 Instruction &ClonedI = std::get<1>(ZippedInsts);
954
955 // The only instructions in the exit block should be PHI nodes and
956 // potentially a landing pad.
957 assert((((isa<PHINode>(I) || isa<LandingPadInst>(I) || isa
<CatchPadInst>(I)) && "Bad instruction in exit block!"
) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && \"Bad instruction in exit block!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 959, __PRETTY_FUNCTION__))
958 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&(((isa<PHINode>(I) || isa<LandingPadInst>(I) || isa
<CatchPadInst>(I)) && "Bad instruction in exit block!"
) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && \"Bad instruction in exit block!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 959, __PRETTY_FUNCTION__))
959 "Bad instruction in exit block!")(((isa<PHINode>(I) || isa<LandingPadInst>(I) || isa
<CatchPadInst>(I)) && "Bad instruction in exit block!"
) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && \"Bad instruction in exit block!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 959, __PRETTY_FUNCTION__))
;
960 // We should have a value map between the instruction and its clone.
961 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!")((VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!"
) ? static_cast<void> (0) : __assert_fail ("VMap.lookup(&I) == &ClonedI && \"Mismatch in the value map!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 961, __PRETTY_FUNCTION__))
;
962
963 auto *MergePN =
964 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi",
965 &*MergeBB->getFirstInsertionPt());
966 I.replaceAllUsesWith(MergePN);
967 MergePN->addIncoming(&I, ExitBB);
968 MergePN->addIncoming(&ClonedI, ClonedExitBB);
969 }
970 }
971
972 // Rewrite the instructions in the cloned blocks to refer to the instructions
973 // in the cloned blocks. We have to do this as a second pass so that we have
974 // everything available. Also, we have inserted new instructions which may
975 // include assume intrinsics, so we update the assumption cache while
976 // processing this.
977 for (auto *ClonedBB : NewBlocks)
978 for (Instruction &I : *ClonedBB) {
979 RemapInstruction(&I, VMap,
980 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
981 if (auto *II = dyn_cast<IntrinsicInst>(&I))
982 if (II->getIntrinsicID() == Intrinsic::assume)
983 AC.registerAssumption(II);
984 }
985
986 // Update any PHI nodes in the cloned successors of the skipped blocks to not
987 // have spurious incoming values.
988 for (auto *LoopBB : L.blocks())
989 if (SkipBlock(LoopBB))
990 for (auto *SuccBB : successors(LoopBB))
991 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
992 for (PHINode &PN : ClonedSuccBB->phis())
993 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
994
995 // Remove the cloned parent as a predecessor of any successor we ended up
996 // cloning other than the unswitched one.
997 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
998 for (auto *SuccBB : successors(ParentBB)) {
999 if (SuccBB == UnswitchedSuccBB)
1000 continue;
1001
1002 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1003 if (!ClonedSuccBB)
1004 continue;
1005
1006 ClonedSuccBB->removePredecessor(ClonedParentBB,
1007 /*DontDeleteUselessPHIs*/ true);
1008 }
1009
1010 // Replace the cloned branch with an unconditional branch to the cloned
1011 // unswitched successor.
1012 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1013 ClonedParentBB->getTerminator()->eraseFromParent();
1014 BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1015
1016 // If there are duplicate entries in the PHI nodes because of multiple edges
1017 // to the unswitched successor, we need to nuke all but one as we replaced it
1018 // with a direct branch.
1019 for (PHINode &PN : ClonedSuccBB->phis()) {
1020 bool Found = false;
1021 // Loop over the incoming operands backwards so we can easily delete as we
1022 // go without invalidating the index.
1023 for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1024 if (PN.getIncomingBlock(i) != ClonedParentBB)
1025 continue;
1026 if (!Found) {
1027 Found = true;
1028 continue;
1029 }
1030 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1031 }
1032 }
1033
1034 // Record the domtree updates for the new blocks.
1035 SmallPtrSet<BasicBlock *, 4> SuccSet;
1036 for (auto *ClonedBB : NewBlocks) {
1037 for (auto *SuccBB : successors(ClonedBB))
1038 if (SuccSet.insert(SuccBB).second)
1039 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1040 SuccSet.clear();
1041 }
1042
1043 return ClonedPH;
1044}
1045
1046/// Recursively clone the specified loop and all of its children.
1047///
1048/// The target parent loop for the clone should be provided, or can be null if
1049/// the clone is a top-level loop. While cloning, all the blocks are mapped
1050/// with the provided value map. The entire original loop must be present in
1051/// the value map. The cloned loop is returned.
1052static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1053 const ValueToValueMapTy &VMap, LoopInfo &LI) {
1054 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1055 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!")((ClonedL.getBlocks().empty() && "Must start with an empty loop!"
) ? static_cast<void> (0) : __assert_fail ("ClonedL.getBlocks().empty() && \"Must start with an empty loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1055, __PRETTY_FUNCTION__))
;
1056 ClonedL.reserveBlocks(OrigL.getNumBlocks());
1057 for (auto *BB : OrigL.blocks()) {
1058 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1059 ClonedL.addBlockEntry(ClonedBB);
1060 if (LI.getLoopFor(BB) == &OrigL)
1061 LI.changeLoopFor(ClonedBB, &ClonedL);
1062 }
1063 };
1064
1065 // We specially handle the first loop because it may get cloned into
1066 // a different parent and because we most commonly are cloning leaf loops.
1067 Loop *ClonedRootL = LI.AllocateLoop();
1068 if (RootParentL)
1069 RootParentL->addChildLoop(ClonedRootL);
1070 else
1071 LI.addTopLevelLoop(ClonedRootL);
1072 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1073
1074 if (OrigRootL.empty())
1075 return ClonedRootL;
1076
1077 // If we have a nest, we can quickly clone the entire loop nest using an
1078 // iterative approach because it is a tree. We keep the cloned parent in the
1079 // data structure to avoid repeatedly querying through a map to find it.
1080 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1081 // Build up the loops to clone in reverse order as we'll clone them from the
1082 // back.
1083 for (Loop *ChildL : llvm::reverse(OrigRootL))
1084 LoopsToClone.push_back({ClonedRootL, ChildL});
1085 do {
1086 Loop *ClonedParentL, *L;
1087 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1088 Loop *ClonedL = LI.AllocateLoop();
1089 ClonedParentL->addChildLoop(ClonedL);
1090 AddClonedBlocksToLoop(*L, *ClonedL);
1091 for (Loop *ChildL : llvm::reverse(*L))
1092 LoopsToClone.push_back({ClonedL, ChildL});
1093 } while (!LoopsToClone.empty());
1094
1095 return ClonedRootL;
1096}
1097
1098/// Build the cloned loops of an original loop from unswitching.
1099///
1100/// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1101/// operation. We need to re-verify that there even is a loop (as the backedge
1102/// may not have been cloned), and even if there are remaining backedges the
1103/// backedge set may be different. However, we know that each child loop is
1104/// undisturbed, we only need to find where to place each child loop within
1105/// either any parent loop or within a cloned version of the original loop.
1106///
1107/// Because child loops may end up cloned outside of any cloned version of the
1108/// original loop, multiple cloned sibling loops may be created. All of them
1109/// are returned so that the newly introduced loop nest roots can be
1110/// identified.
1111static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1112 const ValueToValueMapTy &VMap, LoopInfo &LI,
1113 SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1114 Loop *ClonedL = nullptr;
1115
1116 auto *OrigPH = OrigL.getLoopPreheader();
1117 auto *OrigHeader = OrigL.getHeader();
1118
1119 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1120 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1121
1122 // We need to know the loops of the cloned exit blocks to even compute the
1123 // accurate parent loop. If we only clone exits to some parent of the
1124 // original parent, we want to clone into that outer loop. We also keep track
1125 // of the loops that our cloned exit blocks participate in.
1126 Loop *ParentL = nullptr;
1127 SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1128 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1129 ClonedExitsInLoops.reserve(ExitBlocks.size());
1130 for (auto *ExitBB : ExitBlocks)
1131 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1132 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1133 ExitLoopMap[ClonedExitBB] = ExitL;
1134 ClonedExitsInLoops.push_back(ClonedExitBB);
1135 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1136 ParentL = ExitL;
1137 }
1138 assert((!ParentL || ParentL == OrigL.getParentLoop() ||(((!ParentL || ParentL == OrigL.getParentLoop() || ParentL->
contains(OrigL.getParentLoop())) && "The computed parent loop should always contain (or be) the parent of "
"the original loop.") ? static_cast<void> (0) : __assert_fail
("(!ParentL || ParentL == OrigL.getParentLoop() || ParentL->contains(OrigL.getParentLoop())) && \"The computed parent loop should always contain (or be) the parent of \" \"the original loop.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1141, __PRETTY_FUNCTION__))
1139 ParentL->contains(OrigL.getParentLoop())) &&(((!ParentL || ParentL == OrigL.getParentLoop() || ParentL->
contains(OrigL.getParentLoop())) && "The computed parent loop should always contain (or be) the parent of "
"the original loop.") ? static_cast<void> (0) : __assert_fail
("(!ParentL || ParentL == OrigL.getParentLoop() || ParentL->contains(OrigL.getParentLoop())) && \"The computed parent loop should always contain (or be) the parent of \" \"the original loop.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1141, __PRETTY_FUNCTION__))
1140 "The computed parent loop should always contain (or be) the parent of "(((!ParentL || ParentL == OrigL.getParentLoop() || ParentL->
contains(OrigL.getParentLoop())) && "The computed parent loop should always contain (or be) the parent of "
"the original loop.") ? static_cast<void> (0) : __assert_fail
("(!ParentL || ParentL == OrigL.getParentLoop() || ParentL->contains(OrigL.getParentLoop())) && \"The computed parent loop should always contain (or be) the parent of \" \"the original loop.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1141, __PRETTY_FUNCTION__))
1141 "the original loop.")(((!ParentL || ParentL == OrigL.getParentLoop() || ParentL->
contains(OrigL.getParentLoop())) && "The computed parent loop should always contain (or be) the parent of "
"the original loop.") ? static_cast<void> (0) : __assert_fail
("(!ParentL || ParentL == OrigL.getParentLoop() || ParentL->contains(OrigL.getParentLoop())) && \"The computed parent loop should always contain (or be) the parent of \" \"the original loop.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1141, __PRETTY_FUNCTION__))
;
1142
1143 // We build the set of blocks dominated by the cloned header from the set of
1144 // cloned blocks out of the original loop. While not all of these will
1145 // necessarily be in the cloned loop, it is enough to establish that they
1146 // aren't in unreachable cycles, etc.
1147 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1148 for (auto *BB : OrigL.blocks())
1149 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1150 ClonedLoopBlocks.insert(ClonedBB);
1151
1152 // Rebuild the set of blocks that will end up in the cloned loop. We may have
1153 // skipped cloning some region of this loop which can in turn skip some of
1154 // the backedges so we have to rebuild the blocks in the loop based on the
1155 // backedges that remain after cloning.
1156 SmallVector<BasicBlock *, 16> Worklist;
1157 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1158 for (auto *Pred : predecessors(ClonedHeader)) {
1159 // The only possible non-loop header predecessor is the preheader because
1160 // we know we cloned the loop in simplified form.
1161 if (Pred == ClonedPH)
1162 continue;
1163
1164 // Because the loop was in simplified form, the only non-loop predecessor
1165 // should be the preheader.
1166 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "((ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
"header other than the preheader " "that is not part of the loop!"
) ? static_cast<void> (0) : __assert_fail ("ClonedLoopBlocks.count(Pred) && \"Found a predecessor of the loop \" \"header other than the preheader \" \"that is not part of the loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1168, __PRETTY_FUNCTION__))
1167 "header other than the preheader "((ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
"header other than the preheader " "that is not part of the loop!"
) ? static_cast<void> (0) : __assert_fail ("ClonedLoopBlocks.count(Pred) && \"Found a predecessor of the loop \" \"header other than the preheader \" \"that is not part of the loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1168, __PRETTY_FUNCTION__))
1168 "that is not part of the loop!")((ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
"header other than the preheader " "that is not part of the loop!"
) ? static_cast<void> (0) : __assert_fail ("ClonedLoopBlocks.count(Pred) && \"Found a predecessor of the loop \" \"header other than the preheader \" \"that is not part of the loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1168, __PRETTY_FUNCTION__))
;
1169
1170 // Insert this block into the loop set and on the first visit (and if it
1171 // isn't the header we're currently walking) put it into the worklist to
1172 // recurse through.
1173 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1174 Worklist.push_back(Pred);
1175 }
1176
1177 // If we had any backedges then there *is* a cloned loop. Put the header into
1178 // the loop set and then walk the worklist backwards to find all the blocks
1179 // that remain within the loop after cloning.
1180 if (!BlocksInClonedLoop.empty()) {
1181 BlocksInClonedLoop.insert(ClonedHeader);
1182
1183 while (!Worklist.empty()) {
1184 BasicBlock *BB = Worklist.pop_back_val();
1185 assert(BlocksInClonedLoop.count(BB) &&((BlocksInClonedLoop.count(BB) && "Didn't put block into the loop set!"
) ? static_cast<void> (0) : __assert_fail ("BlocksInClonedLoop.count(BB) && \"Didn't put block into the loop set!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1186, __PRETTY_FUNCTION__))
1186 "Didn't put block into the loop set!")((BlocksInClonedLoop.count(BB) && "Didn't put block into the loop set!"
) ? static_cast<void> (0) : __assert_fail ("BlocksInClonedLoop.count(BB) && \"Didn't put block into the loop set!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1186, __PRETTY_FUNCTION__))
;
1187
1188 // Insert any predecessors that are in the possible set into the cloned
1189 // set, and if the insert is successful, add them to the worklist. Note
1190 // that we filter on the blocks that are definitely reachable via the
1191 // backedge to the loop header so we may prune out dead code within the
1192 // cloned loop.
1193 for (auto *Pred : predecessors(BB))
1194 if (ClonedLoopBlocks.count(Pred) &&
1195 BlocksInClonedLoop.insert(Pred).second)
1196 Worklist.push_back(Pred);
1197 }
1198
1199 ClonedL = LI.AllocateLoop();
1200 if (ParentL) {
1201 ParentL->addBasicBlockToLoop(ClonedPH, LI);
1202 ParentL->addChildLoop(ClonedL);
1203 } else {
1204 LI.addTopLevelLoop(ClonedL);
1205 }
1206 NonChildClonedLoops.push_back(ClonedL);
1207
1208 ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1209 // We don't want to just add the cloned loop blocks based on how we
1210 // discovered them. The original order of blocks was carefully built in
1211 // a way that doesn't rely on predecessor ordering. Rather than re-invent
1212 // that logic, we just re-walk the original blocks (and those of the child
1213 // loops) and filter them as we add them into the cloned loop.
1214 for (auto *BB : OrigL.blocks()) {
1215 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1216 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1217 continue;
1218
1219 // Directly add the blocks that are only in this loop.
1220 if (LI.getLoopFor(BB) == &OrigL) {
1221 ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1222 continue;
1223 }
1224
1225 // We want to manually add it to this loop and parents.
1226 // Registering it with LoopInfo will happen when we clone the top
1227 // loop for this block.
1228 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1229 PL->addBlockEntry(ClonedBB);
1230 }
1231
1232 // Now add each child loop whose header remains within the cloned loop. All
1233 // of the blocks within the loop must satisfy the same constraints as the
1234 // header so once we pass the header checks we can just clone the entire
1235 // child loop nest.
1236 for (Loop *ChildL : OrigL) {
1237 auto *ClonedChildHeader =
1238 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1239 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1240 continue;
1241
1242#ifndef NDEBUG
1243 // We should never have a cloned child loop header but fail to have
1244 // all of the blocks for that child loop.
1245 for (auto *ChildLoopBB : ChildL->blocks())
1246 assert(BlocksInClonedLoop.count(((BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup
(ChildLoopBB))) && "Child cloned loop has a header within the cloned outer "
"loop but not all of its blocks!") ? static_cast<void>
(0) : __assert_fail ("BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && \"Child cloned loop has a header within the cloned outer \" \"loop but not all of its blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1249, __PRETTY_FUNCTION__))
1247 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&((BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup
(ChildLoopBB))) && "Child cloned loop has a header within the cloned outer "
"loop but not all of its blocks!") ? static_cast<void>
(0) : __assert_fail ("BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && \"Child cloned loop has a header within the cloned outer \" \"loop but not all of its blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1249, __PRETTY_FUNCTION__))
1248 "Child cloned loop has a header within the cloned outer "((BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup
(ChildLoopBB))) && "Child cloned loop has a header within the cloned outer "
"loop but not all of its blocks!") ? static_cast<void>
(0) : __assert_fail ("BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && \"Child cloned loop has a header within the cloned outer \" \"loop but not all of its blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1249, __PRETTY_FUNCTION__))
1249 "loop but not all of its blocks!")((BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup
(ChildLoopBB))) && "Child cloned loop has a header within the cloned outer "
"loop but not all of its blocks!") ? static_cast<void>
(0) : __assert_fail ("BlocksInClonedLoop.count( cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && \"Child cloned loop has a header within the cloned outer \" \"loop but not all of its blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1249, __PRETTY_FUNCTION__))
;
1250#endif
1251
1252 cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1253 }
1254 }
1255
1256 // Now that we've handled all the components of the original loop that were
1257 // cloned into a new loop, we still need to handle anything from the original
1258 // loop that wasn't in a cloned loop.
1259
1260 // Figure out what blocks are left to place within any loop nest containing
1261 // the unswitched loop. If we never formed a loop, the cloned PH is one of
1262 // them.
1263 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1264 if (BlocksInClonedLoop.empty())
1265 UnloopedBlockSet.insert(ClonedPH);
1266 for (auto *ClonedBB : ClonedLoopBlocks)
1267 if (!BlocksInClonedLoop.count(ClonedBB))
1268 UnloopedBlockSet.insert(ClonedBB);
1269
1270 // Copy the cloned exits and sort them in ascending loop depth, we'll work
1271 // backwards across these to process them inside out. The order shouldn't
1272 // matter as we're just trying to build up the map from inside-out; we use
1273 // the map in a more stably ordered way below.
1274 auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1275 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1276 return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1277 ExitLoopMap.lookup(RHS)->getLoopDepth();
1278 });
1279
1280 // Populate the existing ExitLoopMap with everything reachable from each
1281 // exit, starting from the inner most exit.
1282 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1283 assert(Worklist.empty() && "Didn't clear worklist!")((Worklist.empty() && "Didn't clear worklist!") ? static_cast
<void> (0) : __assert_fail ("Worklist.empty() && \"Didn't clear worklist!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1283, __PRETTY_FUNCTION__))
;
1284
1285 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1286 Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1287
1288 // Walk the CFG back until we hit the cloned PH adding everything reachable
1289 // and in the unlooped set to this exit block's loop.
1290 Worklist.push_back(ExitBB);
1291 do {
1292 BasicBlock *BB = Worklist.pop_back_val();
1293 // We can stop recursing at the cloned preheader (if we get there).
1294 if (BB == ClonedPH)
1295 continue;
1296
1297 for (BasicBlock *PredBB : predecessors(BB)) {
1298 // If this pred has already been moved to our set or is part of some
1299 // (inner) loop, no update needed.
1300 if (!UnloopedBlockSet.erase(PredBB)) {
1301 assert((((BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB
)) && "Predecessor not mapped to a loop!") ? static_cast
<void> (0) : __assert_fail ("(BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && \"Predecessor not mapped to a loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1303, __PRETTY_FUNCTION__))
1302 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&(((BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB
)) && "Predecessor not mapped to a loop!") ? static_cast
<void> (0) : __assert_fail ("(BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && \"Predecessor not mapped to a loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1303, __PRETTY_FUNCTION__))
1303 "Predecessor not mapped to a loop!")(((BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB
)) && "Predecessor not mapped to a loop!") ? static_cast
<void> (0) : __assert_fail ("(BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && \"Predecessor not mapped to a loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1303, __PRETTY_FUNCTION__))
;
1304 continue;
1305 }
1306
1307 // We just insert into the loop set here. We'll add these blocks to the
1308 // exit loop after we build up the set in an order that doesn't rely on
1309 // predecessor order (which in turn relies on use list order).
1310 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1311 (void)Inserted;
1312 assert(Inserted && "Should only visit an unlooped block once!")((Inserted && "Should only visit an unlooped block once!"
) ? static_cast<void> (0) : __assert_fail ("Inserted && \"Should only visit an unlooped block once!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1312, __PRETTY_FUNCTION__))
;
1313
1314 // And recurse through to its predecessors.
1315 Worklist.push_back(PredBB);
1316 }
1317 } while (!Worklist.empty());
1318 }
1319
1320 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned
1321 // blocks to their outer loops, walk the cloned blocks and the cloned exits
1322 // in their original order adding them to the correct loop.
1323
1324 // We need a stable insertion order. We use the order of the original loop
1325 // order and map into the correct parent loop.
1326 for (auto *BB : llvm::concat<BasicBlock *const>(
1327 makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1328 if (Loop *OuterL = ExitLoopMap.lookup(BB))
1329 OuterL->addBasicBlockToLoop(BB, LI);
1330
1331#ifndef NDEBUG
1332 for (auto &BBAndL : ExitLoopMap) {
1333 auto *BB = BBAndL.first;
1334 auto *OuterL = BBAndL.second;
1335 assert(LI.getLoopFor(BB) == OuterL &&((LI.getLoopFor(BB) == OuterL && "Failed to put all blocks into outer loops!"
) ? static_cast<void> (0) : __assert_fail ("LI.getLoopFor(BB) == OuterL && \"Failed to put all blocks into outer loops!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1336, __PRETTY_FUNCTION__))
1336 "Failed to put all blocks into outer loops!")((LI.getLoopFor(BB) == OuterL && "Failed to put all blocks into outer loops!"
) ? static_cast<void> (0) : __assert_fail ("LI.getLoopFor(BB) == OuterL && \"Failed to put all blocks into outer loops!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1336, __PRETTY_FUNCTION__))
;
1337 }
1338#endif
1339
1340 // Now that all the blocks are placed into the correct containing loop in the
1341 // absence of child loops, find all the potentially cloned child loops and
1342 // clone them into whatever outer loop we placed their header into.
1343 for (Loop *ChildL : OrigL) {
1344 auto *ClonedChildHeader =
1345 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1346 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1347 continue;
1348
1349#ifndef NDEBUG
1350 for (auto *ChildLoopBB : ChildL->blocks())
1351 assert(VMap.count(ChildLoopBB) &&((VMap.count(ChildLoopBB) && "Cloned a child loop header but not all of that loops blocks!"
) ? static_cast<void> (0) : __assert_fail ("VMap.count(ChildLoopBB) && \"Cloned a child loop header but not all of that loops blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1352, __PRETTY_FUNCTION__))
1352 "Cloned a child loop header but not all of that loops blocks!")((VMap.count(ChildLoopBB) && "Cloned a child loop header but not all of that loops blocks!"
) ? static_cast<void> (0) : __assert_fail ("VMap.count(ChildLoopBB) && \"Cloned a child loop header but not all of that loops blocks!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1352, __PRETTY_FUNCTION__))
;
1353#endif
1354
1355 NonChildClonedLoops.push_back(cloneLoopNest(
1356 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1357 }
1358}
1359
1360static void
1361deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1362 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1363 DominatorTree &DT) {
1364 // Find all the dead clones, and remove them from their successors.
1365 SmallVector<BasicBlock *, 16> DeadBlocks;
1366 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1367 for (auto &VMap : VMaps)
1368 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1369 if (!DT.isReachableFromEntry(ClonedBB)) {
1370 for (BasicBlock *SuccBB : successors(ClonedBB))
1371 SuccBB->removePredecessor(ClonedBB);
1372 DeadBlocks.push_back(ClonedBB);
1373 }
1374
1375 // Drop any remaining references to break cycles.
1376 for (BasicBlock *BB : DeadBlocks)
1377 BB->dropAllReferences();
1378 // Erase them from the IR.
1379 for (BasicBlock *BB : DeadBlocks)
1380 BB->eraseFromParent();
1381}
1382
1383static void
1384deleteDeadBlocksFromLoop(Loop &L,
1385 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1386 DominatorTree &DT, LoopInfo &LI) {
1387 // Find all the dead blocks tied to this loop, and remove them from their
1388 // successors.
1389 SmallPtrSet<BasicBlock *, 16> DeadBlockSet;
1390
1391 // Start with loop/exit blocks and get a transitive closure of reachable dead
1392 // blocks.
1393 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1394 ExitBlocks.end());
1395 DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1396 while (!DeathCandidates.empty()) {
1397 auto *BB = DeathCandidates.pop_back_val();
1398 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1399 for (BasicBlock *SuccBB : successors(BB)) {
1400 SuccBB->removePredecessor(BB);
1401 DeathCandidates.push_back(SuccBB);
1402 }
1403 DeadBlockSet.insert(BB);
1404 }
1405 }
1406
1407 // Filter out the dead blocks from the exit blocks list so that it can be
1408 // used in the caller.
1409 llvm::erase_if(ExitBlocks,
1410 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1411
1412 // Walk from this loop up through its parents removing all of the dead blocks.
1413 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1414 for (auto *BB : DeadBlockSet)
1415 ParentL->getBlocksSet().erase(BB);
1416 llvm::erase_if(ParentL->getBlocksVector(),
1417 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1418 }
1419
1420 // Now delete the dead child loops. This raw delete will clear them
1421 // recursively.
1422 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1423 if (!DeadBlockSet.count(ChildL->getHeader()))
1424 return false;
1425
1426 assert(llvm::all_of(ChildL->blocks(),((llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB
) { return DeadBlockSet.count(ChildBB); }) && "If the child loop header is dead all blocks in the child loop must "
"be dead as well!") ? static_cast<void> (0) : __assert_fail
("llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB) { return DeadBlockSet.count(ChildBB); }) && \"If the child loop header is dead all blocks in the child loop must \" \"be dead as well!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1431, __PRETTY_FUNCTION__))
1427 [&](BasicBlock *ChildBB) {((llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB
) { return DeadBlockSet.count(ChildBB); }) && "If the child loop header is dead all blocks in the child loop must "
"be dead as well!") ? static_cast<void> (0) : __assert_fail
("llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB) { return DeadBlockSet.count(ChildBB); }) && \"If the child loop header is dead all blocks in the child loop must \" \"be dead as well!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1431, __PRETTY_FUNCTION__))
1428 return DeadBlockSet.count(ChildBB);((llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB
) { return DeadBlockSet.count(ChildBB); }) && "If the child loop header is dead all blocks in the child loop must "
"be dead as well!") ? static_cast<void> (0) : __assert_fail
("llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB) { return DeadBlockSet.count(ChildBB); }) && \"If the child loop header is dead all blocks in the child loop must \" \"be dead as well!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1431, __PRETTY_FUNCTION__))
1429 }) &&((llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB
) { return DeadBlockSet.count(ChildBB); }) && "If the child loop header is dead all blocks in the child loop must "
"be dead as well!") ? static_cast<void> (0) : __assert_fail
("llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB) { return DeadBlockSet.count(ChildBB); }) && \"If the child loop header is dead all blocks in the child loop must \" \"be dead as well!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1431, __PRETTY_FUNCTION__))
1430 "If the child loop header is dead all blocks in the child loop must "((llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB
) { return DeadBlockSet.count(ChildBB); }) && "If the child loop header is dead all blocks in the child loop must "
"be dead as well!") ? static_cast<void> (0) : __assert_fail
("llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB) { return DeadBlockSet.count(ChildBB); }) && \"If the child loop header is dead all blocks in the child loop must \" \"be dead as well!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1431, __PRETTY_FUNCTION__))
1431 "be dead as well!")((llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB
) { return DeadBlockSet.count(ChildBB); }) && "If the child loop header is dead all blocks in the child loop must "
"be dead as well!") ? static_cast<void> (0) : __assert_fail
("llvm::all_of(ChildL->blocks(), [&](BasicBlock *ChildBB) { return DeadBlockSet.count(ChildBB); }) && \"If the child loop header is dead all blocks in the child loop must \" \"be dead as well!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1431, __PRETTY_FUNCTION__))
;
1432 LI.destroy(ChildL);
1433 return true;
1434 });
1435
1436 // Remove the loop mappings for the dead blocks and drop all the references
1437 // from these blocks to others to handle cyclic references as we start
1438 // deleting the blocks themselves.
1439 for (auto *BB : DeadBlockSet) {
1440 // Check that the dominator tree has already been updated.
1441 assert(!DT.getNode(BB) && "Should already have cleared domtree!")((!DT.getNode(BB) && "Should already have cleared domtree!"
) ? static_cast<void> (0) : __assert_fail ("!DT.getNode(BB) && \"Should already have cleared domtree!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1441, __PRETTY_FUNCTION__))
;
1442 LI.changeLoopFor(BB, nullptr);
1443 BB->dropAllReferences();
1444 }
1445
1446 // Actually delete the blocks now that they've been fully unhooked from the
1447 // IR.
1448 for (auto *BB : DeadBlockSet)
1449 BB->eraseFromParent();
1450}
1451
1452/// Recompute the set of blocks in a loop after unswitching.
1453///
1454/// This walks from the original headers predecessors to rebuild the loop. We
1455/// take advantage of the fact that new blocks can't have been added, and so we
1456/// filter by the original loop's blocks. This also handles potentially
1457/// unreachable code that we don't want to explore but might be found examining
1458/// the predecessors of the header.
1459///
1460/// If the original loop is no longer a loop, this will return an empty set. If
1461/// it remains a loop, all the blocks within it will be added to the set
1462/// (including those blocks in inner loops).
1463static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1464 LoopInfo &LI) {
1465 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1466
1467 auto *PH = L.getLoopPreheader();
1468 auto *Header = L.getHeader();
1469
1470 // A worklist to use while walking backwards from the header.
1471 SmallVector<BasicBlock *, 16> Worklist;
1472
1473 // First walk the predecessors of the header to find the backedges. This will
1474 // form the basis of our walk.
1475 for (auto *Pred : predecessors(Header)) {
1476 // Skip the preheader.
1477 if (Pred == PH)
1478 continue;
1479
1480 // Because the loop was in simplified form, the only non-loop predecessor
1481 // is the preheader.
1482 assert(L.contains(Pred) && "Found a predecessor of the loop header other "((L.contains(Pred) && "Found a predecessor of the loop header other "
"than the preheader that is not part of the " "loop!") ? static_cast
<void> (0) : __assert_fail ("L.contains(Pred) && \"Found a predecessor of the loop header other \" \"than the preheader that is not part of the \" \"loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1484, __PRETTY_FUNCTION__))
1483 "than the preheader that is not part of the "((L.contains(Pred) && "Found a predecessor of the loop header other "
"than the preheader that is not part of the " "loop!") ? static_cast
<void> (0) : __assert_fail ("L.contains(Pred) && \"Found a predecessor of the loop header other \" \"than the preheader that is not part of the \" \"loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1484, __PRETTY_FUNCTION__))
1484 "loop!")((L.contains(Pred) && "Found a predecessor of the loop header other "
"than the preheader that is not part of the " "loop!") ? static_cast
<void> (0) : __assert_fail ("L.contains(Pred) && \"Found a predecessor of the loop header other \" \"than the preheader that is not part of the \" \"loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1484, __PRETTY_FUNCTION__))
;
1485
1486 // Insert this block into the loop set and on the first visit and, if it
1487 // isn't the header we're currently walking, put it into the worklist to
1488 // recurse through.
1489 if (LoopBlockSet.insert(Pred).second && Pred != Header)
1490 Worklist.push_back(Pred);
1491 }
1492
1493 // If no backedges were found, we're done.
1494 if (LoopBlockSet.empty())
1495 return LoopBlockSet;
1496
1497 // We found backedges, recurse through them to identify the loop blocks.
1498 while (!Worklist.empty()) {
1499 BasicBlock *BB = Worklist.pop_back_val();
1500 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!")((LoopBlockSet.count(BB) && "Didn't put block into the loop set!"
) ? static_cast<void> (0) : __assert_fail ("LoopBlockSet.count(BB) && \"Didn't put block into the loop set!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1500, __PRETTY_FUNCTION__))
;
1501
1502 // No need to walk past the header.
1503 if (BB == Header)
1504 continue;
1505
1506 // Because we know the inner loop structure remains valid we can use the
1507 // loop structure to jump immediately across the entire nested loop.
1508 // Further, because it is in loop simplified form, we can directly jump
1509 // to its preheader afterward.
1510 if (Loop *InnerL = LI.getLoopFor(BB))
1511 if (InnerL != &L) {
1512 assert(L.contains(InnerL) &&((L.contains(InnerL) && "Should not reach a loop *outside* this loop!"
) ? static_cast<void> (0) : __assert_fail ("L.contains(InnerL) && \"Should not reach a loop *outside* this loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1513, __PRETTY_FUNCTION__))
1513 "Should not reach a loop *outside* this loop!")((L.contains(InnerL) && "Should not reach a loop *outside* this loop!"
) ? static_cast<void> (0) : __assert_fail ("L.contains(InnerL) && \"Should not reach a loop *outside* this loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1513, __PRETTY_FUNCTION__))
;
1514 // The preheader is the only possible predecessor of the loop so
1515 // insert it into the set and check whether it was already handled.
1516 auto *InnerPH = InnerL->getLoopPreheader();
1517 assert(L.contains(InnerPH) && "Cannot contain an inner loop block "((L.contains(InnerPH) && "Cannot contain an inner loop block "
"but not contain the inner loop " "preheader!") ? static_cast
<void> (0) : __assert_fail ("L.contains(InnerPH) && \"Cannot contain an inner loop block \" \"but not contain the inner loop \" \"preheader!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1519, __PRETTY_FUNCTION__))
1518 "but not contain the inner loop "((L.contains(InnerPH) && "Cannot contain an inner loop block "
"but not contain the inner loop " "preheader!") ? static_cast
<void> (0) : __assert_fail ("L.contains(InnerPH) && \"Cannot contain an inner loop block \" \"but not contain the inner loop \" \"preheader!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1519, __PRETTY_FUNCTION__))
1519 "preheader!")((L.contains(InnerPH) && "Cannot contain an inner loop block "
"but not contain the inner loop " "preheader!") ? static_cast
<void> (0) : __assert_fail ("L.contains(InnerPH) && \"Cannot contain an inner loop block \" \"but not contain the inner loop \" \"preheader!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1519, __PRETTY_FUNCTION__))
;
1520 if (!LoopBlockSet.insert(InnerPH).second)
1521 // The only way to reach the preheader is through the loop body
1522 // itself so if it has been visited the loop is already handled.
1523 continue;
1524
1525 // Insert all of the blocks (other than those already present) into
1526 // the loop set. We expect at least the block that led us to find the
1527 // inner loop to be in the block set, but we may also have other loop
1528 // blocks if they were already enqueued as predecessors of some other
1529 // outer loop block.
1530 for (auto *InnerBB : InnerL->blocks()) {
1531 if (InnerBB == BB) {
1532 assert(LoopBlockSet.count(InnerBB) &&((LoopBlockSet.count(InnerBB) && "Block should already be in the set!"
) ? static_cast<void> (0) : __assert_fail ("LoopBlockSet.count(InnerBB) && \"Block should already be in the set!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1533, __PRETTY_FUNCTION__))
1533 "Block should already be in the set!")((LoopBlockSet.count(InnerBB) && "Block should already be in the set!"
) ? static_cast<void> (0) : __assert_fail ("LoopBlockSet.count(InnerBB) && \"Block should already be in the set!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1533, __PRETTY_FUNCTION__))
;
1534 continue;
1535 }
1536
1537 LoopBlockSet.insert(InnerBB);
1538 }
1539
1540 // Add the preheader to the worklist so we will continue past the
1541 // loop body.
1542 Worklist.push_back(InnerPH);
1543 continue;
1544 }
1545
1546 // Insert any predecessors that were in the original loop into the new
1547 // set, and if the insert is successful, add them to the worklist.
1548 for (auto *Pred : predecessors(BB))
1549 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1550 Worklist.push_back(Pred);
1551 }
1552
1553 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!")((LoopBlockSet.count(Header) && "Cannot fail to add the header!"
) ? static_cast<void> (0) : __assert_fail ("LoopBlockSet.count(Header) && \"Cannot fail to add the header!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1553, __PRETTY_FUNCTION__))
;
1554
1555 // We've found all the blocks participating in the loop, return our completed
1556 // set.
1557 return LoopBlockSet;
1558}
1559
1560/// Rebuild a loop after unswitching removes some subset of blocks and edges.
1561///
1562/// The removal may have removed some child loops entirely but cannot have
1563/// disturbed any remaining child loops. However, they may need to be hoisted
1564/// to the parent loop (or to be top-level loops). The original loop may be
1565/// completely removed.
1566///
1567/// The sibling loops resulting from this update are returned. If the original
1568/// loop remains a valid loop, it will be the first entry in this list with all
1569/// of the newly sibling loops following it.
1570///
1571/// Returns true if the loop remains a loop after unswitching, and false if it
1572/// is no longer a loop after unswitching (and should not continue to be
1573/// referenced).
1574static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1575 LoopInfo &LI,
1576 SmallVectorImpl<Loop *> &HoistedLoops) {
1577 auto *PH = L.getLoopPreheader();
1578
1579 // Compute the actual parent loop from the exit blocks. Because we may have
1580 // pruned some exits the loop may be different from the original parent.
1581 Loop *ParentL = nullptr;
1582 SmallVector<Loop *, 4> ExitLoops;
1583 SmallVector<BasicBlock *, 4> ExitsInLoops;
1584 ExitsInLoops.reserve(ExitBlocks.size());
1585 for (auto *ExitBB : ExitBlocks)
1586 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1587 ExitLoops.push_back(ExitL);
1588 ExitsInLoops.push_back(ExitBB);
1589 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1590 ParentL = ExitL;
1591 }
1592
1593 // Recompute the blocks participating in this loop. This may be empty if it
1594 // is no longer a loop.
1595 auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1596
1597 // If we still have a loop, we need to re-set the loop's parent as the exit
1598 // block set changing may have moved it within the loop nest. Note that this
1599 // can only happen when this loop has a parent as it can only hoist the loop
1600 // *up* the nest.
1601 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1602 // Remove this loop's (original) blocks from all of the intervening loops.
1603 for (Loop *IL = L.getParentLoop(); IL != ParentL;
1604 IL = IL->getParentLoop()) {
1605 IL->getBlocksSet().erase(PH);
1606 for (auto *BB : L.blocks())
1607 IL->getBlocksSet().erase(BB);
1608 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1609 return BB == PH || L.contains(BB);
1610 });
1611 }
1612
1613 LI.changeLoopFor(PH, ParentL);
1614 L.getParentLoop()->removeChildLoop(&L);
1615 if (ParentL)
1616 ParentL->addChildLoop(&L);
1617 else
1618 LI.addTopLevelLoop(&L);
1619 }
1620
1621 // Now we update all the blocks which are no longer within the loop.
1622 auto &Blocks = L.getBlocksVector();
1623 auto BlocksSplitI =
1624 LoopBlockSet.empty()
1625 ? Blocks.begin()
1626 : std::stable_partition(
1627 Blocks.begin(), Blocks.end(),
1628 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1629
1630 // Before we erase the list of unlooped blocks, build a set of them.
1631 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1632 if (LoopBlockSet.empty())
1633 UnloopedBlocks.insert(PH);
1634
1635 // Now erase these blocks from the loop.
1636 for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1637 L.getBlocksSet().erase(BB);
1638 Blocks.erase(BlocksSplitI, Blocks.end());
1639
1640 // Sort the exits in ascending loop depth, we'll work backwards across these
1641 // to process them inside out.
1642 std::stable_sort(ExitsInLoops.begin(), ExitsInLoops.end(),
1643 [&](BasicBlock *LHS, BasicBlock *RHS) {
1644 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1645 });
1646
1647 // We'll build up a set for each exit loop.
1648 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1649 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1650
1651 auto RemoveUnloopedBlocksFromLoop =
1652 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1653 for (auto *BB : UnloopedBlocks)
1654 L.getBlocksSet().erase(BB);
1655 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1656 return UnloopedBlocks.count(BB);
1657 });
1658 };
1659
1660 SmallVector<BasicBlock *, 16> Worklist;
1661 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1662 assert(Worklist.empty() && "Didn't clear worklist!")((Worklist.empty() && "Didn't clear worklist!") ? static_cast
<void> (0) : __assert_fail ("Worklist.empty() && \"Didn't clear worklist!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1662, __PRETTY_FUNCTION__))
;
1663 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!")((NewExitLoopBlocks.empty() && "Didn't clear loop set!"
) ? static_cast<void> (0) : __assert_fail ("NewExitLoopBlocks.empty() && \"Didn't clear loop set!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1663, __PRETTY_FUNCTION__))
;
1664
1665 // Grab the next exit block, in decreasing loop depth order.
1666 BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1667 Loop &ExitL = *LI.getLoopFor(ExitBB);
1668 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!")((ExitL.contains(&L) && "Exit loop must contain the inner loop!"
) ? static_cast<void> (0) : __assert_fail ("ExitL.contains(&L) && \"Exit loop must contain the inner loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1668, __PRETTY_FUNCTION__))
;
1669
1670 // Erase all of the unlooped blocks from the loops between the previous
1671 // exit loop and this exit loop. This works because the ExitInLoops list is
1672 // sorted in increasing order of loop depth and thus we visit loops in
1673 // decreasing order of loop depth.
1674 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1675 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1676
1677 // Walk the CFG back until we hit the cloned PH adding everything reachable
1678 // and in the unlooped set to this exit block's loop.
1679 Worklist.push_back(ExitBB);
1680 do {
1681 BasicBlock *BB = Worklist.pop_back_val();
1682 // We can stop recursing at the cloned preheader (if we get there).
1683 if (BB == PH)
1684 continue;
1685
1686 for (BasicBlock *PredBB : predecessors(BB)) {
1687 // If this pred has already been moved to our set or is part of some
1688 // (inner) loop, no update needed.
1689 if (!UnloopedBlocks.erase(PredBB)) {
1690 assert((NewExitLoopBlocks.count(PredBB) ||(((NewExitLoopBlocks.count(PredBB) || ExitL.contains(LI.getLoopFor
(PredBB))) && "Predecessor not in a nested loop (or already visited)!"
) ? static_cast<void> (0) : __assert_fail ("(NewExitLoopBlocks.count(PredBB) || ExitL.contains(LI.getLoopFor(PredBB))) && \"Predecessor not in a nested loop (or already visited)!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1692, __PRETTY_FUNCTION__))
1691 ExitL.contains(LI.getLoopFor(PredBB))) &&(((NewExitLoopBlocks.count(PredBB) || ExitL.contains(LI.getLoopFor
(PredBB))) && "Predecessor not in a nested loop (or already visited)!"
) ? static_cast<void> (0) : __assert_fail ("(NewExitLoopBlocks.count(PredBB) || ExitL.contains(LI.getLoopFor(PredBB))) && \"Predecessor not in a nested loop (or already visited)!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1692, __PRETTY_FUNCTION__))
1692 "Predecessor not in a nested loop (or already visited)!")(((NewExitLoopBlocks.count(PredBB) || ExitL.contains(LI.getLoopFor
(PredBB))) && "Predecessor not in a nested loop (or already visited)!"
) ? static_cast<void> (0) : __assert_fail ("(NewExitLoopBlocks.count(PredBB) || ExitL.contains(LI.getLoopFor(PredBB))) && \"Predecessor not in a nested loop (or already visited)!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1692, __PRETTY_FUNCTION__))
;
1693 continue;
1694 }
1695
1696 // We just insert into the loop set here. We'll add these blocks to the
1697 // exit loop after we build up the set in a deterministic order rather
1698 // than the predecessor-influenced visit order.
1699 bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
1700 (void)Inserted;
1701 assert(Inserted && "Should only visit an unlooped block once!")((Inserted && "Should only visit an unlooped block once!"
) ? static_cast<void> (0) : __assert_fail ("Inserted && \"Should only visit an unlooped block once!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1701, __PRETTY_FUNCTION__))
;
1702
1703 // And recurse through to its predecessors.
1704 Worklist.push_back(PredBB);
1705 }
1706 } while (!Worklist.empty());
1707
1708 // If blocks in this exit loop were directly part of the original loop (as
1709 // opposed to a child loop) update the map to point to this exit loop. This
1710 // just updates a map and so the fact that the order is unstable is fine.
1711 for (auto *BB : NewExitLoopBlocks)
1712 if (Loop *BBL = LI.getLoopFor(BB))
1713 if (BBL == &L || !L.contains(BBL))
1714 LI.changeLoopFor(BB, &ExitL);
1715
1716 // We will remove the remaining unlooped blocks from this loop in the next
1717 // iteration or below.
1718 NewExitLoopBlocks.clear();
1719 }
1720
1721 // Any remaining unlooped blocks are no longer part of any loop unless they
1722 // are part of some child loop.
1723 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
1724 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1725 for (auto *BB : UnloopedBlocks)
1726 if (Loop *BBL = LI.getLoopFor(BB))
1727 if (BBL == &L || !L.contains(BBL))
1728 LI.changeLoopFor(BB, nullptr);
1729
1730 // Sink all the child loops whose headers are no longer in the loop set to
1731 // the parent (or to be top level loops). We reach into the loop and directly
1732 // update its subloop vector to make this batch update efficient.
1733 auto &SubLoops = L.getSubLoopsVector();
1734 auto SubLoopsSplitI =
1735 LoopBlockSet.empty()
1736 ? SubLoops.begin()
1737 : std::stable_partition(
1738 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
1739 return LoopBlockSet.count(SubL->getHeader());
1740 });
1741 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
1742 HoistedLoops.push_back(HoistedL);
1743 HoistedL->setParentLoop(nullptr);
1744
1745 // To compute the new parent of this hoisted loop we look at where we
1746 // placed the preheader above. We can't lookup the header itself because we
1747 // retained the mapping from the header to the hoisted loop. But the
1748 // preheader and header should have the exact same new parent computed
1749 // based on the set of exit blocks from the original loop as the preheader
1750 // is a predecessor of the header and so reached in the reverse walk. And
1751 // because the loops were all in simplified form the preheader of the
1752 // hoisted loop can't be part of some *other* loop.
1753 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
1754 NewParentL->addChildLoop(HoistedL);
1755 else
1756 LI.addTopLevelLoop(HoistedL);
1757 }
1758 SubLoops.erase(SubLoopsSplitI, SubLoops.end());
1759
1760 // Actually delete the loop if nothing remained within it.
1761 if (Blocks.empty()) {
1762 assert(SubLoops.empty() &&((SubLoops.empty() && "Failed to remove all subloops from the original loop!"
) ? static_cast<void> (0) : __assert_fail ("SubLoops.empty() && \"Failed to remove all subloops from the original loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1763, __PRETTY_FUNCTION__))
1763 "Failed to remove all subloops from the original loop!")((SubLoops.empty() && "Failed to remove all subloops from the original loop!"
) ? static_cast<void> (0) : __assert_fail ("SubLoops.empty() && \"Failed to remove all subloops from the original loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1763, __PRETTY_FUNCTION__))
;
1764 if (Loop *ParentL = L.getParentLoop())
1765 ParentL->removeChildLoop(llvm::find(*ParentL, &L));
1766 else
1767 LI.removeLoop(llvm::find(LI, &L));
1768 LI.destroy(&L);
1769 return false;
1770 }
1771
1772 return true;
1773}
1774
1775/// Helper to visit a dominator subtree, invoking a callable on each node.
1776///
1777/// Returning false at any point will stop walking past that node of the tree.
1778template <typename CallableT>
1779void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
1780 SmallVector<DomTreeNode *, 4> DomWorklist;
1781 DomWorklist.push_back(DT[BB]);
1782#ifndef NDEBUG
1783 SmallPtrSet<DomTreeNode *, 4> Visited;
1784 Visited.insert(DT[BB]);
1785#endif
1786 do {
1787 DomTreeNode *N = DomWorklist.pop_back_val();
1788
1789 // Visit this node.
1790 if (!Callable(N->getBlock()))
1791 continue;
1792
1793 // Accumulate the child nodes.
1794 for (DomTreeNode *ChildN : *N) {
1795 assert(Visited.insert(ChildN).second &&((Visited.insert(ChildN).second && "Cannot visit a node twice when walking a tree!"
) ? static_cast<void> (0) : __assert_fail ("Visited.insert(ChildN).second && \"Cannot visit a node twice when walking a tree!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1796, __PRETTY_FUNCTION__))
1796 "Cannot visit a node twice when walking a tree!")((Visited.insert(ChildN).second && "Cannot visit a node twice when walking a tree!"
) ? static_cast<void> (0) : __assert_fail ("Visited.insert(ChildN).second && \"Cannot visit a node twice when walking a tree!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1796, __PRETTY_FUNCTION__))
;
1797 DomWorklist.push_back(ChildN);
1798 }
1799 } while (!DomWorklist.empty());
1800}
1801
1802static void unswitchNontrivialInvariants(
1803 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
1804 SmallVectorImpl<BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI,
1805 AssumptionCache &AC, function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
1806 ScalarEvolution *SE) {
1807 auto *ParentBB = TI.getParent();
1808 BranchInst *BI = dyn_cast<BranchInst>(&TI);
1809 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
1810
1811 // We can only unswitch switches, conditional branches with an invariant
1812 // condition, or combining invariant conditions with an instruction.
1813 assert((SI || BI->isConditional()) &&(((SI || BI->isConditional()) && "Can only unswitch switches and conditional branch!"
) ? static_cast<void> (0) : __assert_fail ("(SI || BI->isConditional()) && \"Can only unswitch switches and conditional branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1814, __PRETTY_FUNCTION__))
1814 "Can only unswitch switches and conditional branch!")(((SI || BI->isConditional()) && "Can only unswitch switches and conditional branch!"
) ? static_cast<void> (0) : __assert_fail ("(SI || BI->isConditional()) && \"Can only unswitch switches and conditional branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1814, __PRETTY_FUNCTION__))
;
1815 bool FullUnswitch = SI || BI->getCondition() == Invariants[0];
1816 if (FullUnswitch)
1817 assert(Invariants.size() == 1 &&((Invariants.size() == 1 && "Cannot have other invariants with full unswitching!"
) ? static_cast<void> (0) : __assert_fail ("Invariants.size() == 1 && \"Cannot have other invariants with full unswitching!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1818, __PRETTY_FUNCTION__))
1818 "Cannot have other invariants with full unswitching!")((Invariants.size() == 1 && "Cannot have other invariants with full unswitching!"
) ? static_cast<void> (0) : __assert_fail ("Invariants.size() == 1 && \"Cannot have other invariants with full unswitching!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1818, __PRETTY_FUNCTION__))
;
1819 else
1820 assert(isa<Instruction>(BI->getCondition()) &&((isa<Instruction>(BI->getCondition()) && "Partial unswitching requires an instruction as the condition!"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(BI->getCondition()) && \"Partial unswitching requires an instruction as the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1821, __PRETTY_FUNCTION__))
1821 "Partial unswitching requires an instruction as the condition!")((isa<Instruction>(BI->getCondition()) && "Partial unswitching requires an instruction as the condition!"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(BI->getCondition()) && \"Partial unswitching requires an instruction as the condition!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1821, __PRETTY_FUNCTION__))
;
1822
1823 // Constant and BBs tracking the cloned and continuing successor. When we are
1824 // unswitching the entire condition, this can just be trivially chosen to
1825 // unswitch towards `true`. However, when we are unswitching a set of
1826 // invariants combined with `and` or `or`, the combining operation determines
1827 // the best direction to unswitch: we want to unswitch the direction that will
1828 // collapse the branch.
1829 bool Direction = true;
1830 int ClonedSucc = 0;
1831 if (!FullUnswitch) {
1832 if (cast<Instruction>(BI->getCondition())->getOpcode() != Instruction::Or) {
1833 assert(cast<Instruction>(BI->getCondition())->getOpcode() ==((cast<Instruction>(BI->getCondition())->getOpcode
() == Instruction::And && "Only `or` and `and` instructions can combine invariants being "
"unswitched.") ? static_cast<void> (0) : __assert_fail
("cast<Instruction>(BI->getCondition())->getOpcode() == Instruction::And && \"Only `or` and `and` instructions can combine invariants being \" \"unswitched.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1836, __PRETTY_FUNCTION__))
1834 Instruction::And &&((cast<Instruction>(BI->getCondition())->getOpcode
() == Instruction::And && "Only `or` and `and` instructions can combine invariants being "
"unswitched.") ? static_cast<void> (0) : __assert_fail
("cast<Instruction>(BI->getCondition())->getOpcode() == Instruction::And && \"Only `or` and `and` instructions can combine invariants being \" \"unswitched.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1836, __PRETTY_FUNCTION__))
1835 "Only `or` and `and` instructions can combine invariants being "((cast<Instruction>(BI->getCondition())->getOpcode
() == Instruction::And && "Only `or` and `and` instructions can combine invariants being "
"unswitched.") ? static_cast<void> (0) : __assert_fail
("cast<Instruction>(BI->getCondition())->getOpcode() == Instruction::And && \"Only `or` and `and` instructions can combine invariants being \" \"unswitched.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1836, __PRETTY_FUNCTION__))
1836 "unswitched.")((cast<Instruction>(BI->getCondition())->getOpcode
() == Instruction::And && "Only `or` and `and` instructions can combine invariants being "
"unswitched.") ? static_cast<void> (0) : __assert_fail
("cast<Instruction>(BI->getCondition())->getOpcode() == Instruction::And && \"Only `or` and `and` instructions can combine invariants being \" \"unswitched.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1836, __PRETTY_FUNCTION__))
;
1837 Direction = false;
1838 ClonedSucc = 1;
1839 }
1840 }
1841
1842 BasicBlock *RetainedSuccBB =
1843 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
1844 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
1845 if (BI)
1846 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
1847 else
1848 for (auto Case : SI->cases())
1849 if (Case.getCaseSuccessor() != RetainedSuccBB)
1850 UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
1851
1852 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&((!UnswitchedSuccBBs.count(RetainedSuccBB) && "Should not unswitch the same successor we are retaining!"
) ? static_cast<void> (0) : __assert_fail ("!UnswitchedSuccBBs.count(RetainedSuccBB) && \"Should not unswitch the same successor we are retaining!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1853, __PRETTY_FUNCTION__))
1853 "Should not unswitch the same successor we are retaining!")((!UnswitchedSuccBBs.count(RetainedSuccBB) && "Should not unswitch the same successor we are retaining!"
) ? static_cast<void> (0) : __assert_fail ("!UnswitchedSuccBBs.count(RetainedSuccBB) && \"Should not unswitch the same successor we are retaining!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1853, __PRETTY_FUNCTION__))
;
1854
1855 // The branch should be in this exact loop. Any inner loop's invariant branch
1856 // should be handled by unswitching that inner loop. The caller of this
1857 // routine should filter out any candidates that remain (but were skipped for
1858 // whatever reason).
1859 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!")((LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!"
) ? static_cast<void> (0) : __assert_fail ("LI.getLoopFor(ParentBB) == &L && \"Branch in an inner loop!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1859, __PRETTY_FUNCTION__))
;
1860
1861 // Compute the parent loop now before we start hacking on things.
1862 Loop *ParentL = L.getParentLoop();
1863
1864 // Compute the outer-most loop containing one of our exit blocks. This is the
1865 // furthest up our loopnest which can be mutated, which we will use below to
1866 // update things.
1867 Loop *OuterExitL = &L;
1868 for (auto *ExitBB : ExitBlocks) {
1869 Loop *NewOuterExitL = LI.getLoopFor(ExitBB);
1870 if (!NewOuterExitL) {
1871 // We exited the entire nest with this block, so we're done.
1872 OuterExitL = nullptr;
1873 break;
1874 }
1875 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
1876 OuterExitL = NewOuterExitL;
1877 }
1878
1879 // At this point, we're definitely going to unswitch something so invalidate
1880 // any cached information in ScalarEvolution for the outer most loop
1881 // containing an exit block and all nested loops.
1882 if (SE) {
1883 if (OuterExitL)
1884 SE->forgetLoop(OuterExitL);
1885 else
1886 SE->forgetTopmostLoop(&L);
1887 }
1888
1889 // If the edge from this terminator to a successor dominates that successor,
1890 // store a map from each block in its dominator subtree to it. This lets us
1891 // tell when cloning for a particular successor if a block is dominated by
1892 // some *other* successor with a single data structure. We use this to
1893 // significantly reduce cloning.
1894 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
1895 for (auto *SuccBB : llvm::concat<BasicBlock *const>(
1896 makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs))
1897 if (SuccBB->getUniquePredecessor() ||
1898 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
1899 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
1900 }))
1901 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
1902 DominatingSucc[BB] = SuccBB;
1903 return true;
1904 });
1905
1906 // Split the preheader, so that we know that there is a safe place to insert
1907 // the conditional branch. We will change the preheader to have a conditional
1908 // branch on LoopCond. The original preheader will become the split point
1909 // between the unswitched versions, and we will have a new preheader for the
1910 // original loop.
1911 BasicBlock *SplitBB = L.getLoopPreheader();
1912 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI);
1913
1914 // Keep track of the dominator tree updates needed.
1915 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1916
1917 // Clone the loop for each unswitched successor.
1918 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
1919 VMaps.reserve(UnswitchedSuccBBs.size());
1920 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
1921 for (auto *SuccBB : UnswitchedSuccBBs) {
1922 VMaps.emplace_back(new ValueToValueMapTy());
1923 ClonedPHs[SuccBB] = buildClonedLoopBlocks(
1924 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
1925 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI);
1926 }
1927
1928 // The stitching of the branched code back together depends on whether we're
1929 // doing full unswitching or not with the exception that we always want to
1930 // nuke the initial terminator placed in the split block.
1931 SplitBB->getTerminator()->eraseFromParent();
1932 if (FullUnswitch) {
1933 // First we need to unhook the successor relationship as we'll be replacing
1934 // the terminator with a direct branch. This is much simpler for branches
1935 // than switches so we handle those first.
1936 if (BI) {
1937 // Remove the parent as a predecessor of the unswitched successor.
1938 assert(UnswitchedSuccBBs.size() == 1 &&((UnswitchedSuccBBs.size() == 1 && "Only one possible unswitched block for a branch!"
) ? static_cast<void> (0) : __assert_fail ("UnswitchedSuccBBs.size() == 1 && \"Only one possible unswitched block for a branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1939, __PRETTY_FUNCTION__))
1939 "Only one possible unswitched block for a branch!")((UnswitchedSuccBBs.size() == 1 && "Only one possible unswitched block for a branch!"
) ? static_cast<void> (0) : __assert_fail ("UnswitchedSuccBBs.size() == 1 && \"Only one possible unswitched block for a branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1939, __PRETTY_FUNCTION__))
;
1940 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
1941 UnswitchedSuccBB->removePredecessor(ParentBB,
1942 /*DontDeleteUselessPHIs*/ true);
1943 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
1944 } else {
1945 // Note that we actually want to remove the parent block as a predecessor
1946 // of *every* case successor. The case successor is either unswitched,
1947 // completely eliminating an edge from the parent to that successor, or it
1948 // is a duplicate edge to the retained successor as the retained successor
1949 // is always the default successor and as we'll replace this with a direct
1950 // branch we no longer need the duplicate entries in the PHI nodes.
1951 assert(SI->getDefaultDest() == RetainedSuccBB &&((SI->getDefaultDest() == RetainedSuccBB && "Not retaining default successor!"
) ? static_cast<void> (0) : __assert_fail ("SI->getDefaultDest() == RetainedSuccBB && \"Not retaining default successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1952, __PRETTY_FUNCTION__))
1952 "Not retaining default successor!")((SI->getDefaultDest() == RetainedSuccBB && "Not retaining default successor!"
) ? static_cast<void> (0) : __assert_fail ("SI->getDefaultDest() == RetainedSuccBB && \"Not retaining default successor!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1952, __PRETTY_FUNCTION__))
;
1953 for (auto &Case : SI->cases())
1954 Case.getCaseSuccessor()->removePredecessor(
1955 ParentBB,
1956 /*DontDeleteUselessPHIs*/ true);
1957
1958 // We need to use the set to populate domtree updates as even when there
1959 // are multiple cases pointing at the same successor we only want to
1960 // remove and insert one edge in the domtree.
1961 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
1962 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
1963 }
1964
1965 // Now that we've unhooked the successor relationship, splice the terminator
1966 // from the original loop to the split.
1967 SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI);
1968
1969 // Now wire up the terminator to the preheaders.
1970 if (BI) {
1971 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
1972 BI->setSuccessor(ClonedSucc, ClonedPH);
1973 BI->setSuccessor(1 - ClonedSucc, LoopPH);
1974 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
1975 } else {
1976 assert(SI && "Must either be a branch or switch!")((SI && "Must either be a branch or switch!") ? static_cast
<void> (0) : __assert_fail ("SI && \"Must either be a branch or switch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1976, __PRETTY_FUNCTION__))
;
1977
1978 // Walk the cases and directly update their successors.
1979 SI->setDefaultDest(LoopPH);
1980 for (auto &Case : SI->cases())
1981 if (Case.getCaseSuccessor() == RetainedSuccBB)
1982 Case.setSuccessor(LoopPH);
1983 else
1984 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
1985
1986 // We need to use the set to populate domtree updates as even when there
1987 // are multiple cases pointing at the same successor we only want to
1988 // remove and insert one edge in the domtree.
1989 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
1990 DTUpdates.push_back(
1991 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
1992 }
1993
1994 // Create a new unconditional branch to the continuing block (as opposed to
1995 // the one cloned).
1996 BranchInst::Create(RetainedSuccBB, ParentBB);
1997 } else {
1998 assert(BI && "Only branches have partial unswitching.")((BI && "Only branches have partial unswitching.") ? static_cast
<void> (0) : __assert_fail ("BI && \"Only branches have partial unswitching.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 1998, __PRETTY_FUNCTION__))
;
1999 assert(UnswitchedSuccBBs.size() == 1 &&((UnswitchedSuccBBs.size() == 1 && "Only one possible unswitched block for a branch!"
) ? static_cast<void> (0) : __assert_fail ("UnswitchedSuccBBs.size() == 1 && \"Only one possible unswitched block for a branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2000, __PRETTY_FUNCTION__))
2000 "Only one possible unswitched block for a branch!")((UnswitchedSuccBBs.size() == 1 && "Only one possible unswitched block for a branch!"
) ? static_cast<void> (0) : __assert_fail ("UnswitchedSuccBBs.size() == 1 && \"Only one possible unswitched block for a branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2000, __PRETTY_FUNCTION__))
;
2001 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2002 // When doing a partial unswitch, we have to do a bit more work to build up
2003 // the branch in the split block.
2004 buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction,
2005 *ClonedPH, *LoopPH);
2006 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2007 }
2008
2009 // Apply the updates accumulated above to get an up-to-date dominator tree.
2010 DT.applyUpdates(DTUpdates);
2011
2012 // Now that we have an accurate dominator tree, first delete the dead cloned
2013 // blocks so that we can accurately build any cloned loops. It is important to
2014 // not delete the blocks from the original loop yet because we still want to
2015 // reference the original loop to understand the cloned loop's structure.
2016 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT);
2017
2018 // Build the cloned loop structure itself. This may be substantially
2019 // different from the original structure due to the simplified CFG. This also
2020 // handles inserting all the cloned blocks into the correct loops.
2021 SmallVector<Loop *, 4> NonChildClonedLoops;
2022 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2023 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2024
2025 // Now that our cloned loops have been built, we can update the original loop.
2026 // First we delete the dead blocks from it and then we rebuild the loop
2027 // structure taking these deletions into account.
2028 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI);
2029 SmallVector<Loop *, 4> HoistedLoops;
2030 bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops);
2031
2032 // This transformation has a high risk of corrupting the dominator tree, and
2033 // the below steps to rebuild loop structures will result in hard to debug
2034 // errors in that case so verify that the dominator tree is sane first.
2035 // FIXME: Remove this when the bugs stop showing up and rely on existing
2036 // verification steps.
2037 assert(DT.verify(DominatorTree::VerificationLevel::Fast))((DT.verify(DominatorTree::VerificationLevel::Fast)) ? static_cast
<void> (0) : __assert_fail ("DT.verify(DominatorTree::VerificationLevel::Fast)"
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2037, __PRETTY_FUNCTION__))
;
2038
2039 if (BI) {
2040 // If we unswitched a branch which collapses the condition to a known
2041 // constant we want to replace all the uses of the invariants within both
2042 // the original and cloned blocks. We do this here so that we can use the
2043 // now updated dominator tree to identify which side the users are on.
2044 assert(UnswitchedSuccBBs.size() == 1 &&((UnswitchedSuccBBs.size() == 1 && "Only one possible unswitched block for a branch!"
) ? static_cast<void> (0) : __assert_fail ("UnswitchedSuccBBs.size() == 1 && \"Only one possible unswitched block for a branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2045, __PRETTY_FUNCTION__))
2045 "Only one possible unswitched block for a branch!")((UnswitchedSuccBBs.size() == 1 && "Only one possible unswitched block for a branch!"
) ? static_cast<void> (0) : __assert_fail ("UnswitchedSuccBBs.size() == 1 && \"Only one possible unswitched block for a branch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2045, __PRETTY_FUNCTION__))
;
2046 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2047 ConstantInt *UnswitchedReplacement =
2048 Direction ? ConstantInt::getTrue(BI->getContext())
2049 : ConstantInt::getFalse(BI->getContext());
2050 ConstantInt *ContinueReplacement =
2051 Direction ? ConstantInt::getFalse(BI->getContext())
2052 : ConstantInt::getTrue(BI->getContext());
2053 for (Value *Invariant : Invariants)
2054 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end();
2055 UI != UE;) {
2056 // Grab the use and walk past it so we can clobber it in the use list.
2057 Use *U = &*UI++;
2058 Instruction *UserI = dyn_cast<Instruction>(U->getUser());
2059 if (!UserI)
2060 continue;
2061
2062 // Replace it with the 'continue' side if in the main loop body, and the
2063 // unswitched if in the cloned blocks.
2064 if (DT.dominates(LoopPH, UserI->getParent()))
2065 U->set(ContinueReplacement);
2066 else if (DT.dominates(ClonedPH, UserI->getParent()))
2067 U->set(UnswitchedReplacement);
2068 }
2069 }
2070
2071 // We can change which blocks are exit blocks of all the cloned sibling
2072 // loops, the current loop, and any parent loops which shared exit blocks
2073 // with the current loop. As a consequence, we need to re-form LCSSA for
2074 // them. But we shouldn't need to re-form LCSSA for any child loops.
2075 // FIXME: This could be made more efficient by tracking which exit blocks are
2076 // new, and focusing on them, but that isn't likely to be necessary.
2077 //
2078 // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2079 // loop nest and update every loop that could have had its exits changed. We
2080 // also need to cover any intervening loops. We add all of these loops to
2081 // a list and sort them by loop depth to achieve this without updating
2082 // unnecessary loops.
2083 auto UpdateLoop = [&](Loop &UpdateL) {
2084#ifndef NDEBUG
2085 UpdateL.verifyLoop();
2086 for (Loop *ChildL : UpdateL) {
2087 ChildL->verifyLoop();
2088 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&((ChildL->isRecursivelyLCSSAForm(DT, LI) && "Perturbed a child loop's LCSSA form!"
) ? static_cast<void> (0) : __assert_fail ("ChildL->isRecursivelyLCSSAForm(DT, LI) && \"Perturbed a child loop's LCSSA form!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2089, __PRETTY_FUNCTION__))
2089 "Perturbed a child loop's LCSSA form!")((ChildL->isRecursivelyLCSSAForm(DT, LI) && "Perturbed a child loop's LCSSA form!"
) ? static_cast<void> (0) : __assert_fail ("ChildL->isRecursivelyLCSSAForm(DT, LI) && \"Perturbed a child loop's LCSSA form!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2089, __PRETTY_FUNCTION__))
;
2090 }
2091#endif
2092 // First build LCSSA for this loop so that we can preserve it when
2093 // forming dedicated exits. We don't want to perturb some other loop's
2094 // LCSSA while doing that CFG edit.
2095 formLCSSA(UpdateL, DT, &LI, nullptr);
2096
2097 // For loops reached by this loop's original exit blocks we may
2098 // introduced new, non-dedicated exits. At least try to re-form dedicated
2099 // exits for these loops. This may fail if they couldn't have dedicated
2100 // exits to start with.
2101 formDedicatedExitBlocks(&UpdateL, &DT, &LI, /*PreserveLCSSA*/ true);
2102 };
2103
2104 // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2105 // and we can do it in any order as they don't nest relative to each other.
2106 //
2107 // Also check if any of the loops we have updated have become top-level loops
2108 // as that will necessitate widening the outer loop scope.
2109 for (Loop *UpdatedL :
2110 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2111 UpdateLoop(*UpdatedL);
2112 if (!UpdatedL->getParentLoop())
2113 OuterExitL = nullptr;
2114 }
2115 if (IsStillLoop) {
2116 UpdateLoop(L);
2117 if (!L.getParentLoop())
2118 OuterExitL = nullptr;
2119 }
2120
2121 // If the original loop had exit blocks, walk up through the outer most loop
2122 // of those exit blocks to update LCSSA and form updated dedicated exits.
2123 if (OuterExitL != &L)
2124 for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2125 OuterL = OuterL->getParentLoop())
2126 UpdateLoop(*OuterL);
2127
2128#ifndef NDEBUG
2129 // Verify the entire loop structure to catch any incorrect updates before we
2130 // progress in the pass pipeline.
2131 LI.verify(DT);
2132#endif
2133
2134 // Now that we've unswitched something, make callbacks to report the changes.
2135 // For that we need to merge together the updated loops and the cloned loops
2136 // and check whether the original loop survived.
2137 SmallVector<Loop *, 4> SibLoops;
2138 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2139 if (UpdatedL->getParentLoop() == ParentL)
2140 SibLoops.push_back(UpdatedL);
2141 UnswitchCB(IsStillLoop, SibLoops);
2142
2143 ++NumBranches;
2144}
2145
2146/// Recursively compute the cost of a dominator subtree based on the per-block
2147/// cost map provided.
2148///
2149/// The recursive computation is memozied into the provided DT-indexed cost map
2150/// to allow querying it for most nodes in the domtree without it becoming
2151/// quadratic.
2152static int
2153computeDomSubtreeCost(DomTreeNode &N,
2154 const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap,
2155 SmallDenseMap<DomTreeNode *, int, 4> &DTCostMap) {
2156 // Don't accumulate cost (or recurse through) blocks not in our block cost
2157 // map and thus not part of the duplication cost being considered.
2158 auto BBCostIt = BBCostMap.find(N.getBlock());
2159 if (BBCostIt == BBCostMap.end())
2160 return 0;
2161
2162 // Lookup this node to see if we already computed its cost.
2163 auto DTCostIt = DTCostMap.find(&N);
2164 if (DTCostIt != DTCostMap.end())
2165 return DTCostIt->second;
2166
2167 // If not, we have to compute it. We can't use insert above and update
2168 // because computing the cost may insert more things into the map.
2169 int Cost = std::accumulate(
2170 N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) {
2171 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2172 });
2173 bool Inserted = DTCostMap.insert({&N, Cost}).second;
2174 (void)Inserted;
2175 assert(Inserted && "Should not insert a node while visiting children!")((Inserted && "Should not insert a node while visiting children!"
) ? static_cast<void> (0) : __assert_fail ("Inserted && \"Should not insert a node while visiting children!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2175, __PRETTY_FUNCTION__))
;
2176 return Cost;
2177}
2178
2179/// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2180/// making the following replacement:
2181///
2182/// --code before guard--
2183/// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2184/// --code after guard--
2185///
2186/// into
2187///
2188/// --code before guard--
2189/// br i1 %cond, label %guarded, label %deopt
2190///
2191/// guarded:
2192/// --code after guard--
2193///
2194/// deopt:
2195/// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2196/// unreachable
2197///
2198/// It also makes all relevant DT and LI updates, so that all structures are in
2199/// valid state after this transform.
2200static BranchInst *
2201turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2202 SmallVectorImpl<BasicBlock *> &ExitBlocks,
2203 DominatorTree &DT, LoopInfo &LI) {
2204 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2205 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Turning " <<
*GI << " into a branch.\n"; } } while (false)
;
2206 BasicBlock *CheckBB = GI->getParent();
2207
2208 // Remove all CheckBB's successors from DomTree. A block can be seen among
2209 // successors more than once, but for DomTree it should be added only once.
2210 SmallPtrSet<BasicBlock *, 4> Successors;
2211 for (auto *Succ : successors(CheckBB))
2212 if (Successors.insert(Succ).second)
2213 DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ});
2214
2215 Instruction *DeoptBlockTerm =
2216 SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true);
2217 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2218 // SplitBlockAndInsertIfThen inserts control flow that branches to
2219 // DeoptBlockTerm if the condition is true. We want the opposite.
2220 CheckBI->swapSuccessors();
2221
2222 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2223 GuardedBlock->setName("guarded");
2224 CheckBI->getSuccessor(1)->setName("deopt");
2225
2226 // We now have a new exit block.
2227 ExitBlocks.push_back(CheckBI->getSuccessor(1));
2228
2229 GI->moveBefore(DeoptBlockTerm);
2230 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
2231
2232 // Add new successors of CheckBB into DomTree.
2233 for (auto *Succ : successors(CheckBB))
2234 DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ});
2235
2236 // Now the blocks that used to be CheckBB's successors are GuardedBlock's
2237 // successors.
2238 for (auto *Succ : Successors)
2239 DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ});
2240
2241 // Make proper changes to DT.
2242 DT.applyUpdates(DTUpdates);
2243 // Inform LI of a new loop block.
2244 L.addBasicBlockToLoop(GuardedBlock, LI);
2245
2246 ++NumGuards;
2247 return CheckBI;
2248}
2249
2250static bool
2251unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
2252 AssumptionCache &AC, TargetTransformInfo &TTI,
2253 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2254 ScalarEvolution *SE) {
2255 // Collect all invariant conditions within this loop (as opposed to an inner
2256 // loop which would be handled when visiting that inner loop).
2257 SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4>
2258 UnswitchCandidates;
2259
2260 // Whether or not we should also collect guards in the loop.
2261 bool CollectGuards = false;
2262 if (UnswitchGuards) {
1
Assuming the condition is false
2
Taking false branch
2263 auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2264 Intrinsic::getName(Intrinsic::experimental_guard));
2265 if (GuardDecl && !GuardDecl->use_empty())
2266 CollectGuards = true;
2267 }
2268
2269 for (auto *BB : L.blocks()) {
3
Assuming '__begin1' is equal to '__end1'
2270 if (LI.getLoopFor(BB) != &L)
2271 continue;
2272
2273 if (CollectGuards)
2274 for (auto &I : *BB)
2275 if (isGuard(&I)) {
2276 auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0);
2277 // TODO: Support AND, OR conditions and partial unswitching.
2278 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2279 UnswitchCandidates.push_back({&I, {Cond}});
2280 }
2281
2282 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2283 // We can only consider fully loop-invariant switch conditions as we need
2284 // to completely eliminate the switch after unswitching.
2285 if (!isa<Constant>(SI->getCondition()) &&
2286 L.isLoopInvariant(SI->getCondition()))
2287 UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2288 continue;
2289 }
2290
2291 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2292 if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) ||
2293 BI->getSuccessor(0) == BI->getSuccessor(1))
2294 continue;
2295
2296 if (L.isLoopInvariant(BI->getCondition())) {
2297 UnswitchCandidates.push_back({BI, {BI->getCondition()}});
2298 continue;
2299 }
2300
2301 Instruction &CondI = *cast<Instruction>(BI->getCondition());
2302 if (CondI.getOpcode() != Instruction::And &&
2303 CondI.getOpcode() != Instruction::Or)
2304 continue;
2305
2306 TinyPtrVector<Value *> Invariants =
2307 collectHomogenousInstGraphLoopInvariants(L, CondI, LI);
2308 if (Invariants.empty())
2309 continue;
2310
2311 UnswitchCandidates.push_back({BI, std::move(Invariants)});
2312 }
2313
2314 // If we didn't find any candidates, we're done.
2315 if (UnswitchCandidates.empty())
4
Taking false branch
2316 return false;
2317
2318 // Check if there are irreducible CFG cycles in this loop. If so, we cannot
2319 // easily unswitch non-trivial edges out of the loop. Doing so might turn the
2320 // irreducible control flow into reducible control flow and introduce new
2321 // loops "out of thin air". If we ever discover important use cases for doing
2322 // this, we can add support to loop unswitch, but it is a lot of complexity
2323 // for what seems little or no real world benefit.
2324 LoopBlocksRPO RPOT(&L);
2325 RPOT.perform(&LI);
2326 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
5
Taking false branch
2327 return false;
2328
2329 SmallVector<BasicBlock *, 4> ExitBlocks;
2330 L.getUniqueExitBlocks(ExitBlocks);
2331
2332 // We cannot unswitch if exit blocks contain a cleanuppad instruction as we
2333 // don't know how to split those exit blocks.
2334 // FIXME: We should teach SplitBlock to handle this and remove this
2335 // restriction.
2336 for (auto *ExitBB : ExitBlocks)
6
Assuming '__begin1' is equal to '__end1'
2337 if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) {
2338 dbgs() << "Cannot unswitch because of cleanuppad in exit block\n";
2339 return false;
2340 }
2341
2342 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Considering " <<
UnswitchCandidates.size() << " non-trivial loop invariant conditions for unswitching.\n"
; } } while (false)
2343 dbgs() << "Considering " << UnswitchCandidates.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Considering " <<
UnswitchCandidates.size() << " non-trivial loop invariant conditions for unswitching.\n"
; } } while (false)
2344 << " non-trivial loop invariant conditions for unswitching.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Considering " <<
UnswitchCandidates.size() << " non-trivial loop invariant conditions for unswitching.\n"
; } } while (false)
;
2345
2346 // Given that unswitching these terminators will require duplicating parts of
2347 // the loop, so we need to be able to model that cost. Compute the ephemeral
2348 // values and set up a data structure to hold per-BB costs. We cache each
2349 // block's cost so that we don't recompute this when considering different
2350 // subsets of the loop for duplication during unswitching.
2351 SmallPtrSet<const Value *, 4> EphValues;
2352 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
2353 SmallDenseMap<BasicBlock *, int, 4> BBCostMap;
2354
2355 // Compute the cost of each block, as well as the total loop cost. Also, bail
2356 // out if we see instructions which are incompatible with loop unswitching
2357 // (convergent, noduplicate, or cross-basic-block tokens).
2358 // FIXME: We might be able to safely handle some of these in non-duplicated
2359 // regions.
2360 int LoopCost = 0;
2361 for (auto *BB : L.blocks()) {
7
Assuming '__begin1' is equal to '__end1'
2362 int Cost = 0;
2363 for (auto &I : *BB) {
2364 if (EphValues.count(&I))
2365 continue;
2366
2367 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
2368 return false;
2369 if (auto CS = CallSite(&I))
2370 if (CS.isConvergent() || CS.cannotDuplicate())
2371 return false;
2372
2373 Cost += TTI.getUserCost(&I);
2374 }
2375 assert(Cost >= 0 && "Must not have negative costs!")((Cost >= 0 && "Must not have negative costs!") ? static_cast
<void> (0) : __assert_fail ("Cost >= 0 && \"Must not have negative costs!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2375, __PRETTY_FUNCTION__))
;
2376 LoopCost += Cost;
2377 assert(LoopCost >= 0 && "Must not have negative loop costs!")((LoopCost >= 0 && "Must not have negative loop costs!"
) ? static_cast<void> (0) : __assert_fail ("LoopCost >= 0 && \"Must not have negative loop costs!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2377, __PRETTY_FUNCTION__))
;
2378 BBCostMap[BB] = Cost;
2379 }
2380 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Total loop cost: "
<< LoopCost << "\n"; } } while (false)
;
2381
2382 // Now we find the best candidate by searching for the one with the following
2383 // properties in order:
2384 //
2385 // 1) An unswitching cost below the threshold
2386 // 2) The smallest number of duplicated unswitch candidates (to avoid
2387 // creating redundant subsequent unswitching)
2388 // 3) The smallest cost after unswitching.
2389 //
2390 // We prioritize reducing fanout of unswitch candidates provided the cost
2391 // remains below the threshold because this has a multiplicative effect.
2392 //
2393 // This requires memoizing each dominator subtree to avoid redundant work.
2394 //
2395 // FIXME: Need to actually do the number of candidates part above.
2396 SmallDenseMap<DomTreeNode *, int, 4> DTCostMap;
2397 // Given a terminator which might be unswitched, computes the non-duplicated
2398 // cost for that terminator.
2399 auto ComputeUnswitchedCost = [&](Instruction &TI, bool FullUnswitch) {
2400 BasicBlock &BB = *TI.getParent();
2401 SmallPtrSet<BasicBlock *, 4> Visited;
2402
2403 int Cost = LoopCost;
2404 for (BasicBlock *SuccBB : successors(&BB)) {
2405 // Don't count successors more than once.
2406 if (!Visited.insert(SuccBB).second)
2407 continue;
2408
2409 // If this is a partial unswitch candidate, then it must be a conditional
2410 // branch with a condition of either `or` or `and`. In that case, one of
2411 // the successors is necessarily duplicated, so don't even try to remove
2412 // its cost.
2413 if (!FullUnswitch) {
2414 auto &BI = cast<BranchInst>(TI);
2415 if (cast<Instruction>(BI.getCondition())->getOpcode() ==
2416 Instruction::And) {
2417 if (SuccBB == BI.getSuccessor(1))
2418 continue;
2419 } else {
2420 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Only `and` and `or` conditions can result in a partial "
"unswitch!") ? static_cast<void> (0) : __assert_fail (
"cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Only `and` and `or` conditions can result in a partial \" \"unswitch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2423, __PRETTY_FUNCTION__))
2421 Instruction::Or &&((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Only `and` and `or` conditions can result in a partial "
"unswitch!") ? static_cast<void> (0) : __assert_fail (
"cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Only `and` and `or` conditions can result in a partial \" \"unswitch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2423, __PRETTY_FUNCTION__))
2422 "Only `and` and `or` conditions can result in a partial "((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Only `and` and `or` conditions can result in a partial "
"unswitch!") ? static_cast<void> (0) : __assert_fail (
"cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Only `and` and `or` conditions can result in a partial \" \"unswitch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2423, __PRETTY_FUNCTION__))
2423 "unswitch!")((cast<Instruction>(BI.getCondition())->getOpcode() ==
Instruction::Or && "Only `and` and `or` conditions can result in a partial "
"unswitch!") ? static_cast<void> (0) : __assert_fail (
"cast<Instruction>(BI.getCondition())->getOpcode() == Instruction::Or && \"Only `and` and `or` conditions can result in a partial \" \"unswitch!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2423, __PRETTY_FUNCTION__))
;
2424 if (SuccBB == BI.getSuccessor(0))
2425 continue;
2426 }
2427 }
2428
2429 // This successor's domtree will not need to be duplicated after
2430 // unswitching if the edge to the successor dominates it (and thus the
2431 // entire tree). This essentially means there is no other path into this
2432 // subtree and so it will end up live in only one clone of the loop.
2433 if (SuccBB->getUniquePredecessor() ||
2434 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2435 return PredBB == &BB || DT.dominates(SuccBB, PredBB);
2436 })) {
2437 Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
2438 assert(Cost >= 0 &&((Cost >= 0 && "Non-duplicated cost should never exceed total loop cost!"
) ? static_cast<void> (0) : __assert_fail ("Cost >= 0 && \"Non-duplicated cost should never exceed total loop cost!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2439, __PRETTY_FUNCTION__))
2439 "Non-duplicated cost should never exceed total loop cost!")((Cost >= 0 && "Non-duplicated cost should never exceed total loop cost!"
) ? static_cast<void> (0) : __assert_fail ("Cost >= 0 && \"Non-duplicated cost should never exceed total loop cost!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2439, __PRETTY_FUNCTION__))
;
2440 }
2441 }
2442
2443 // Now scale the cost by the number of unique successors minus one. We
2444 // subtract one because there is already at least one copy of the entire
2445 // loop. This is computing the new cost of unswitching a condition.
2446 // Note that guards always have 2 unique successors that are implicit and
2447 // will be materialized if we decide to unswitch it.
2448 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
2449 assert(SuccessorsCount > 1 &&((SuccessorsCount > 1 && "Cannot unswitch a condition without multiple distinct successors!"
) ? static_cast<void> (0) : __assert_fail ("SuccessorsCount > 1 && \"Cannot unswitch a condition without multiple distinct successors!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2450, __PRETTY_FUNCTION__))
2450 "Cannot unswitch a condition without multiple distinct successors!")((SuccessorsCount > 1 && "Cannot unswitch a condition without multiple distinct successors!"
) ? static_cast<void> (0) : __assert_fail ("SuccessorsCount > 1 && \"Cannot unswitch a condition without multiple distinct successors!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2450, __PRETTY_FUNCTION__))
;
2451 return Cost * (SuccessorsCount - 1);
2452 };
2453 Instruction *BestUnswitchTI = nullptr;
2454 int BestUnswitchCost;
8
'BestUnswitchCost' declared without an initial value
2455 ArrayRef<Value *> BestUnswitchInvariants;
2456 for (auto &TerminatorAndInvariants : UnswitchCandidates) {
9
Assuming '__begin1' is equal to '__end1'
2457 Instruction &TI = *TerminatorAndInvariants.first;
2458 ArrayRef<Value *> Invariants = TerminatorAndInvariants.second;
2459 BranchInst *BI = dyn_cast<BranchInst>(&TI);
2460 int CandidateCost = ComputeUnswitchedCost(
2461 TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 &&
2462 Invariants[0] == BI->getCondition()));
2463 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Computed cost of "
<< CandidateCost << " for unswitch candidate: " <<
TI << "\n"; } } while (false)
2464 << " for unswitch candidate: " << TI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Computed cost of "
<< CandidateCost << " for unswitch candidate: " <<
TI << "\n"; } } while (false)
;
2465 if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) {
2466 BestUnswitchTI = &TI;
2467 BestUnswitchCost = CandidateCost;
2468 BestUnswitchInvariants = Invariants;
2469 }
2470 }
2471
2472 if (BestUnswitchCost >= UnswitchThreshold) {
10
The left operand of '>=' is a garbage value
2473 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Cannot unswitch, lowest cost found: "
<< BestUnswitchCost << "\n"; } } while (false)
2474 << BestUnswitchCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Cannot unswitch, lowest cost found: "
<< BestUnswitchCost << "\n"; } } while (false)
;
2475 return false;
2476 }
2477
2478 // If the best candidate is a guard, turn it into a branch.
2479 if (isGuard(BestUnswitchTI))
2480 BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L,
2481 ExitBlocks, DT, LI);
2482
2483 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Unswitching non-trivial (cost = "
<< BestUnswitchCost << ") terminator: " <<
*BestUnswitchTI << "\n"; } } while (false)
2484 << BestUnswitchCost << ") terminator: " << *BestUnswitchTIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Unswitching non-trivial (cost = "
<< BestUnswitchCost << ") terminator: " <<
*BestUnswitchTI << "\n"; } } while (false)
2485 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << " Unswitching non-trivial (cost = "
<< BestUnswitchCost << ") terminator: " <<
*BestUnswitchTI << "\n"; } } while (false)
;
2486 unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants,
2487 ExitBlocks, DT, LI, AC, UnswitchCB, SE);
2488 return true;
2489}
2490
2491/// Unswitch control flow predicated on loop invariant conditions.
2492///
2493/// This first hoists all branches or switches which are trivial (IE, do not
2494/// require duplicating any part of the loop) out of the loop body. It then
2495/// looks at other loop invariant control flows and tries to unswitch those as
2496/// well by cloning the loop if the result is small enough.
2497///
2498/// The `DT`, `LI`, `AC`, `TTI` parameters are required analyses that are also
2499/// updated based on the unswitch.
2500///
2501/// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
2502/// true, we will attempt to do non-trivial unswitching as well as trivial
2503/// unswitching.
2504///
2505/// The `UnswitchCB` callback provided will be run after unswitching is
2506/// complete, with the first parameter set to `true` if the provided loop
2507/// remains a loop, and a list of new sibling loops created.
2508///
2509/// If `SE` is non-null, we will update that analysis based on the unswitching
2510/// done.
2511static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
2512 AssumptionCache &AC, TargetTransformInfo &TTI,
2513 bool NonTrivial,
2514 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2515 ScalarEvolution *SE) {
2516 assert(L.isRecursivelyLCSSAForm(DT, LI) &&((L.isRecursivelyLCSSAForm(DT, LI) && "Loops must be in LCSSA form before unswitching."
) ? static_cast<void> (0) : __assert_fail ("L.isRecursivelyLCSSAForm(DT, LI) && \"Loops must be in LCSSA form before unswitching.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2517, __PRETTY_FUNCTION__))
2517 "Loops must be in LCSSA form before unswitching.")((L.isRecursivelyLCSSAForm(DT, LI) && "Loops must be in LCSSA form before unswitching."
) ? static_cast<void> (0) : __assert_fail ("L.isRecursivelyLCSSAForm(DT, LI) && \"Loops must be in LCSSA form before unswitching.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2517, __PRETTY_FUNCTION__))
;
2518 bool Changed = false;
2519
2520 // Must be in loop simplified form: we need a preheader and dedicated exits.
2521 if (!L.isLoopSimplifyForm())
2522 return false;
2523
2524 // Try trivial unswitch first before loop over other basic blocks in the loop.
2525 if (unswitchAllTrivialConditions(L, DT, LI, SE)) {
2526 // If we unswitched successfully we will want to clean up the loop before
2527 // processing it further so just mark it as unswitched and return.
2528 UnswitchCB(/*CurrentLoopValid*/ true, {});
2529 return true;
2530 }
2531
2532 // If we're not doing non-trivial unswitching, we're done. We both accept
2533 // a parameter but also check a local flag that can be used for testing
2534 // a debugging.
2535 if (!NonTrivial && !EnableNonTrivialUnswitch)
2536 return false;
2537
2538 // For non-trivial unswitching, because it often creates new loops, we rely on
2539 // the pass manager to iterate on the loops rather than trying to immediately
2540 // reach a fixed point. There is no substantial advantage to iterating
2541 // internally, and if any of the new loops are simplified enough to contain
2542 // trivial unswitching we want to prefer those.
2543
2544 // Try to unswitch the best invariant condition. We prefer this full unswitch to
2545 // a partial unswitch when possible below the threshold.
2546 if (unswitchBestCondition(L, DT, LI, AC, TTI, UnswitchCB, SE))
2547 return true;
2548
2549 // No other opportunities to unswitch.
2550 return Changed;
2551}
2552
2553PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
2554 LoopStandardAnalysisResults &AR,
2555 LPMUpdater &U) {
2556 Function &F = *L.getHeader()->getParent();
2557 (void)F;
2558
2559 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << Ldo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Unswitching loop in "
<< F.getName() << ": " << L << "\n";
} } while (false)
2560 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Unswitching loop in "
<< F.getName() << ": " << L << "\n";
} } while (false)
;
2561
2562 // Save the current loop name in a variable so that we can report it even
2563 // after it has been deleted.
2564 std::string LoopName = L.getName();
2565
2566 auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid,
2567 ArrayRef<Loop *> NewLoops) {
2568 // If we did a non-trivial unswitch, we have added new (cloned) loops.
2569 if (!NewLoops.empty())
2570 U.addSiblingLoops(NewLoops);
2571
2572 // If the current loop remains valid, we should revisit it to catch any
2573 // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2574 if (CurrentLoopValid)
2575 U.revisitCurrentLoop();
2576 else
2577 U.markLoopAsDeleted(L, LoopName);
2578 };
2579
2580 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, UnswitchCB,
2581 &AR.SE))
2582 return PreservedAnalyses::all();
2583
2584 // Historically this pass has had issues with the dominator tree so verify it
2585 // in asserts builds.
2586 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast))((AR.DT.verify(DominatorTree::VerificationLevel::Fast)) ? static_cast
<void> (0) : __assert_fail ("AR.DT.verify(DominatorTree::VerificationLevel::Fast)"
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2586, __PRETTY_FUNCTION__))
;
2587 return getLoopPassPreservedAnalyses();
2588}
2589
2590namespace {
2591
2592class SimpleLoopUnswitchLegacyPass : public LoopPass {
2593 bool NonTrivial;
2594
2595public:
2596 static char ID; // Pass ID, replacement for typeid
2597
2598 explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false)
2599 : LoopPass(ID), NonTrivial(NonTrivial) {
2600 initializeSimpleLoopUnswitchLegacyPassPass(
2601 *PassRegistry::getPassRegistry());
2602 }
2603
2604 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
2605
2606 void getAnalysisUsage(AnalysisUsage &AU) const override {
2607 AU.addRequired<AssumptionCacheTracker>();
2608 AU.addRequired<TargetTransformInfoWrapperPass>();
2609 getLoopAnalysisUsage(AU);
2610 }
2611};
2612
2613} // end anonymous namespace
2614
2615bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
2616 if (skipLoop(L))
2617 return false;
2618
2619 Function &F = *L->getHeader()->getParent();
2620
2621 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *Ldo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Unswitching loop in "
<< F.getName() << ": " << *L << "\n"
; } } while (false)
2622 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simple-loop-unswitch")) { dbgs() << "Unswitching loop in "
<< F.getName() << ": " << *L << "\n"
; } } while (false)
;
2623
2624 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2625 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2626 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2627 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2628
2629 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
2630 auto *SE = SEWP ? &SEWP->getSE() : nullptr;
2631
2632 auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid,
2633 ArrayRef<Loop *> NewLoops) {
2634 // If we did a non-trivial unswitch, we have added new (cloned) loops.
2635 for (auto *NewL : NewLoops)
2636 LPM.addLoop(*NewL);
2637
2638 // If the current loop remains valid, re-add it to the queue. This is
2639 // a little wasteful as we'll finish processing the current loop as well,
2640 // but it is the best we can do in the old PM.
2641 if (CurrentLoopValid)
2642 LPM.addLoop(*L);
2643 else
2644 LPM.markLoopAsDeleted(*L);
2645 };
2646
2647 bool Changed = unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, UnswitchCB, SE);
2648
2649 // If anything was unswitched, also clear any cached information about this
2650 // loop.
2651 LPM.deleteSimpleAnalysisLoop(L);
2652
2653 // Historically this pass has had issues with the dominator tree so verify it
2654 // in asserts builds.
2655 assert(DT.verify(DominatorTree::VerificationLevel::Fast))((DT.verify(DominatorTree::VerificationLevel::Fast)) ? static_cast
<void> (0) : __assert_fail ("DT.verify(DominatorTree::VerificationLevel::Fast)"
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp"
, 2655, __PRETTY_FUNCTION__))
;
2656
2657 return Changed;
2658}
2659
2660char SimpleLoopUnswitchLegacyPass::ID = 0;
2661INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",static void *initializeSimpleLoopUnswitchLegacyPassPassOnce(PassRegistry
&Registry) {
2662 "Simple unswitch loops", false, false)static void *initializeSimpleLoopUnswitchLegacyPassPassOnce(PassRegistry
&Registry) {
2663INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
2664INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
2665INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
2666INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry);
2667INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
2668INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",PassInfo *PI = new PassInfo( "Simple unswitch loops", "simple-loop-unswitch"
, &SimpleLoopUnswitchLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<SimpleLoopUnswitchLegacyPass>), false,
false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeSimpleLoopUnswitchLegacyPassPassFlag
; void llvm::initializeSimpleLoopUnswitchLegacyPassPass(PassRegistry
&Registry) { llvm::call_once(InitializeSimpleLoopUnswitchLegacyPassPassFlag
, initializeSimpleLoopUnswitchLegacyPassPassOnce, std::ref(Registry
)); }
2669 "Simple unswitch loops", false, false)PassInfo *PI = new PassInfo( "Simple unswitch loops", "simple-loop-unswitch"
, &SimpleLoopUnswitchLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<SimpleLoopUnswitchLegacyPass>), false,
false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeSimpleLoopUnswitchLegacyPassPassFlag
; void llvm::initializeSimpleLoopUnswitchLegacyPassPass(PassRegistry
&Registry) { llvm::call_once(InitializeSimpleLoopUnswitchLegacyPassPassFlag
, initializeSimpleLoopUnswitchLegacyPassPassOnce, std::ref(Registry
)); }
2670
2671Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) {
2672 return new SimpleLoopUnswitchLegacyPass(NonTrivial);
2673}