1//=- AArch64ConditionOptimizer.cpp - Remove useless comparisons for AArch64 -=//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass tries to make consecutive compares of values use same operands to
10// allow CSE pass to remove duplicated instructions. For this it analyzes
11// branches and adjusts comparisons with immediate values by converting:
12// * GE -> GT
13// * GT -> GE
14// * LT -> LE
15// * LE -> LT
16// and adjusting immediate values appropriately. It basically corrects two
17// immediate values towards each other to make them equal.
18//
19// Consider the following example in C:
20//
21// if ((a < 5 && ...) || (a > 5 && ...)) {
22// ~~~~~ ~~~~~
23// ^ ^
24// x y
25//
26// Here both "x" and "y" expressions compare "a" with "5". When "x" evaluates
27// to "false", "y" can just check flags set by the first comparison. As a
28// result of the canonicalization employed by
29// SelectionDAGBuilder::visitSwitchCase, DAGCombine, and other target-specific
30// code, assembly ends up in the form that is not CSE friendly:
31//
32// ...
33// cmp w8, #4
34// b.gt .LBB0_3
35// ...
36// .LBB0_3:
37// cmp w8, #6
38// b.lt .LBB0_6
39// ...
40//
41// Same assembly after the pass:
42//
43// ...
44// cmp w8, #5
45// b.ge .LBB0_3
46// ...
47// .LBB0_3:
48// cmp w8, #5 // <-- CSE pass removes this instruction
49// b.le .LBB0_6
50// ...
51//
52// Currently only SUBS and ADDS followed by b.?? are supported.
53//
54// TODO: maybe handle TBNZ/TBZ the same way as CMP when used instead for "a < 0"
55// TODO: handle other conditional instructions (e.g. CSET)
56// TODO: allow second branching to be anything if it doesn't require adjusting
57//
58//===----------------------------------------------------------------------===//
59
60#include "AArch64.h"
61#include "MCTargetDesc/AArch64AddressingModes.h"
62#include "Utils/AArch64BaseInfo.h"
63#include "llvm/ADT/ArrayRef.h"
64#include "llvm/ADT/DepthFirstIterator.h"
65#include "llvm/ADT/SmallVector.h"
66#include "llvm/ADT/Statistic.h"
67#include "llvm/CodeGen/MachineBasicBlock.h"
68#include "llvm/CodeGen/MachineDominators.h"
69#include "llvm/CodeGen/MachineFunction.h"
70#include "llvm/CodeGen/MachineFunctionPass.h"
71#include "llvm/CodeGen/MachineInstr.h"
72#include "llvm/CodeGen/MachineInstrBuilder.h"
73#include "llvm/CodeGen/MachineOperand.h"
74#include "llvm/CodeGen/MachineRegisterInfo.h"
75#include "llvm/CodeGen/TargetInstrInfo.h"
76#include "llvm/CodeGen/TargetSubtargetInfo.h"
77#include "llvm/InitializePasses.h"
78#include "llvm/Pass.h"
79#include "llvm/Support/Debug.h"
80#include "llvm/Support/ErrorHandling.h"
81#include "llvm/Support/raw_ostream.h"
82#include <cassert>
83#include <cstdlib>
84#include <tuple>
85
86using namespace llvm;
87
88#define DEBUG_TYPE "aarch64-condopt"
89
90STATISTIC(NumConditionsAdjusted, "Number of conditions adjusted");
91
92namespace {
93
94class AArch64ConditionOptimizer : public MachineFunctionPass {
95 const TargetInstrInfo *TII;
96 MachineDominatorTree *DomTree;
97 const MachineRegisterInfo *MRI;
98
99public:
100 // Stores immediate, compare instruction opcode and branch condition (in this
101 // order) of adjusted comparison.
102 using CmpInfo = std::tuple<int, unsigned, AArch64CC::CondCode>;
103
104 static char ID;
105
106 AArch64ConditionOptimizer() : MachineFunctionPass(ID) {
107 initializeAArch64ConditionOptimizerPass(*PassRegistry::getPassRegistry());
108 }
109
110 void getAnalysisUsage(AnalysisUsage &AU) const override;
111 MachineInstr *findSuitableCompare(MachineBasicBlock *MBB);
112 CmpInfo adjustCmp(MachineInstr *CmpMI, AArch64CC::CondCode Cmp);
113 void modifyCmp(MachineInstr *CmpMI, const CmpInfo &Info);
114 bool adjustTo(MachineInstr *CmpMI, AArch64CC::CondCode Cmp, MachineInstr *To,
115 int ToImm);
116 bool runOnMachineFunction(MachineFunction &MF) override;
117
118 StringRef getPassName() const override {
119 return "AArch64 Condition Optimizer";
120 }
121};
122
123} // end anonymous namespace
124
125char AArch64ConditionOptimizer::ID = 0;
126
127INITIALIZE_PASS_BEGIN(AArch64ConditionOptimizer, "aarch64-condopt",
128 "AArch64 CondOpt Pass", false, false)
129INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
130INITIALIZE_PASS_END(AArch64ConditionOptimizer, "aarch64-condopt",
131 "AArch64 CondOpt Pass", false, false)
132
133FunctionPass *llvm::createAArch64ConditionOptimizerPass() {
134 return new AArch64ConditionOptimizer();
135}
136
137void AArch64ConditionOptimizer::getAnalysisUsage(AnalysisUsage &AU) const {
138 AU.addRequired<MachineDominatorTree>();
139 AU.addPreserved<MachineDominatorTree>();
140 MachineFunctionPass::getAnalysisUsage(AU);
141}
142
143// Finds compare instruction that corresponds to supported types of branching.
144// Returns the instruction or nullptr on failures or detecting unsupported
145// instructions.
146MachineInstr *AArch64ConditionOptimizer::findSuitableCompare(
147 MachineBasicBlock *MBB) {
148 MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
149 if (Term == MBB->end())
150 return nullptr;
151
152 if (Term->getOpcode() != AArch64::Bcc)
153 return nullptr;
154
155 // Since we may modify cmp of this MBB, make sure NZCV does not live out.
156 for (auto *SuccBB : MBB->successors())
157 if (SuccBB->isLiveIn(AArch64::Reg: NZCV))
158 return nullptr;
159
160 // Now find the instruction controlling the terminator.
161 for (MachineBasicBlock::iterator B = MBB->begin(), It = Term; It != B;) {
162 It = prev_nodbg(It, Begin: B);
163 MachineInstr &I = *It;
164 assert(!I.isTerminator() && "Spurious terminator");
165 // Check if there is any use of NZCV between CMP and Bcc.
166 if (I.readsRegister(AArch64::Reg: NZCV, /*TRI=*/nullptr))
167 return nullptr;
168 switch (I.getOpcode()) {
169 // cmp is an alias for subs with a dead destination register.
170 case AArch64::SUBSWri:
171 case AArch64::SUBSXri:
172 // cmn is an alias for adds with a dead destination register.
173 case AArch64::ADDSWri:
174 case AArch64::ADDSXri: {
175 unsigned ShiftAmt = AArch64_AM::getShiftValue(Imm: I.getOperand(i: 3).getImm());
176 if (!I.getOperand(i: 2).isImm()) {
177 LLVM_DEBUG(dbgs() << "Immediate of cmp is symbolic, " << I << '\n');
178 return nullptr;
179 } else if (I.getOperand(i: 2).getImm() << ShiftAmt >= 0xfff) {
180 LLVM_DEBUG(dbgs() << "Immediate of cmp may be out of range, " << I
181 << '\n');
182 return nullptr;
183 } else if (!MRI->use_nodbg_empty(RegNo: I.getOperand(i: 0).getReg())) {
184 LLVM_DEBUG(dbgs() << "Destination of cmp is not dead, " << I << '\n');
185 return nullptr;
186 }
187 return &I;
188 }
189 // Prevent false positive case like:
190 // cmp w19, #0
191 // cinc w0, w19, gt
192 // ...
193 // fcmp d8, #0.0
194 // b.gt .LBB0_5
195 case AArch64::FCMPDri:
196 case AArch64::FCMPSri:
197 case AArch64::FCMPESri:
198 case AArch64::FCMPEDri:
199
200 case AArch64::SUBSWrr:
201 case AArch64::SUBSXrr:
202 case AArch64::ADDSWrr:
203 case AArch64::ADDSXrr:
204 case AArch64::FCMPSrr:
205 case AArch64::FCMPDrr:
206 case AArch64::FCMPESrr:
207 case AArch64::FCMPEDrr:
208 // Skip comparison instructions without immediate operands.
209 return nullptr;
210 }
211 }
212 LLVM_DEBUG(dbgs() << "Flags not defined in " << printMBBReference(*MBB)
213 << '\n');
214 return nullptr;
215}
216
217// Changes opcode adds <-> subs considering register operand width.
218static int getComplementOpc(int Opc) {
219 switch (Opc) {
220 case AArch64::ADDSWri: return AArch64::SUBSWri;
221 case AArch64::ADDSXri: return AArch64::SUBSXri;
222 case AArch64::SUBSWri: return AArch64::ADDSWri;
223 case AArch64::SUBSXri: return AArch64::ADDSXri;
224 default:
225 llvm_unreachable("Unexpected opcode");
226 }
227}
228
229// Changes form of comparison inclusive <-> exclusive.
230static AArch64CC::CondCode getAdjustedCmp(AArch64CC::CondCode Cmp) {
231 switch (Cmp) {
232 case AArch64CC::GT: return AArch64CC::GE;
233 case AArch64CC::GE: return AArch64CC::GT;
234 case AArch64CC::LT: return AArch64CC::LE;
235 case AArch64CC::LE: return AArch64CC::LT;
236 default:
237 llvm_unreachable("Unexpected condition code");
238 }
239}
240
241// Transforms GT -> GE, GE -> GT, LT -> LE, LE -> LT by updating comparison
242// operator and condition code.
243AArch64ConditionOptimizer::CmpInfo AArch64ConditionOptimizer::adjustCmp(
244 MachineInstr *CmpMI, AArch64CC::CondCode Cmp) {
245 unsigned Opc = CmpMI->getOpcode();
246
247 // CMN (compare with negative immediate) is an alias to ADDS (as
248 // "operand - negative" == "operand + positive")
249 bool Negative = (Opc == AArch64::ADDSWri || Opc == AArch64::ADDSXri);
250
251 int Correction = (Cmp == AArch64CC::GT) ? 1 : -1;
252 // Negate Correction value for comparison with negative immediate (CMN).
253 if (Negative) {
254 Correction = -Correction;
255 }
256
257 const int OldImm = (int)CmpMI->getOperand(i: 2).getImm();
258 const int NewImm = std::abs(x: OldImm + Correction);
259
260 // Handle +0 -> -1 and -0 -> +1 (CMN with 0 immediate) transitions by
261 // adjusting compare instruction opcode.
262 if (OldImm == 0 && ((Negative && Correction == 1) ||
263 (!Negative && Correction == -1))) {
264 Opc = getComplementOpc(Opc);
265 }
266
267 return CmpInfo(NewImm, Opc, getAdjustedCmp(Cmp));
268}
269
270// Applies changes to comparison instruction suggested by adjustCmp().
271void AArch64ConditionOptimizer::modifyCmp(MachineInstr *CmpMI,
272 const CmpInfo &Info) {
273 int Imm;
274 unsigned Opc;
275 AArch64CC::CondCode Cmp;
276 std::tie(args&: Imm, args&: Opc, args&: Cmp) = Info;
277
278 MachineBasicBlock *const MBB = CmpMI->getParent();
279
280 // Change immediate in comparison instruction (ADDS or SUBS).
281 BuildMI(BB&: *MBB, I: CmpMI, MIMD: CmpMI->getDebugLoc(), MCID: TII->get(Opcode: Opc))
282 .add(MO: CmpMI->getOperand(i: 0))
283 .add(MO: CmpMI->getOperand(i: 1))
284 .addImm(Val: Imm)
285 .add(MO: CmpMI->getOperand(i: 3));
286 CmpMI->eraseFromParent();
287
288 // The fact that this comparison was picked ensures that it's related to the
289 // first terminator instruction.
290 MachineInstr &BrMI = *MBB->getFirstTerminator();
291
292 // Change condition in branch instruction.
293 BuildMI(*MBB, BrMI, BrMI.getDebugLoc(), TII->get(AArch64::Opcode: Bcc))
294 .addImm(Cmp)
295 .add(BrMI.getOperand(i: 1));
296 BrMI.eraseFromParent();
297
298 ++NumConditionsAdjusted;
299}
300
301// Parse a condition code returned by analyzeBranch, and compute the CondCode
302// corresponding to TBB.
303// Returns true if parsing was successful, otherwise false is returned.
304static bool parseCond(ArrayRef<MachineOperand> Cond, AArch64CC::CondCode &CC) {
305 // A normal br.cond simply has the condition code.
306 if (Cond[0].getImm() != -1) {
307 assert(Cond.size() == 1 && "Unknown Cond array format");
308 CC = (AArch64CC::CondCode)(int)Cond[0].getImm();
309 return true;
310 }
311 return false;
312}
313
314// Adjusts one cmp instruction to another one if result of adjustment will allow
315// CSE. Returns true if compare instruction was changed, otherwise false is
316// returned.
317bool AArch64ConditionOptimizer::adjustTo(MachineInstr *CmpMI,
318 AArch64CC::CondCode Cmp, MachineInstr *To, int ToImm)
319{
320 CmpInfo Info = adjustCmp(CmpMI, Cmp);
321 if (std::get<0>(t&: Info) == ToImm && std::get<1>(t&: Info) == To->getOpcode()) {
322 modifyCmp(CmpMI, Info);
323 return true;
324 }
325 return false;
326}
327
328bool AArch64ConditionOptimizer::runOnMachineFunction(MachineFunction &MF) {
329 LLVM_DEBUG(dbgs() << "********** AArch64 Conditional Compares **********\n"
330 << "********** Function: " << MF.getName() << '\n');
331 if (skipFunction(F: MF.getFunction()))
332 return false;
333
334 TII = MF.getSubtarget().getInstrInfo();
335 DomTree = &getAnalysis<MachineDominatorTree>();
336 MRI = &MF.getRegInfo();
337
338 bool Changed = false;
339
340 // Visit blocks in dominator tree pre-order. The pre-order enables multiple
341 // cmp-conversions from the same head block.
342 // Note that updateDomTree() modifies the children of the DomTree node
343 // currently being visited. The df_iterator supports that; it doesn't look at
344 // child_begin() / child_end() until after a node has been visited.
345 for (MachineDomTreeNode *I : depth_first(G: DomTree)) {
346 MachineBasicBlock *HBB = I->getBlock();
347
348 SmallVector<MachineOperand, 4> HeadCond;
349 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
350 if (TII->analyzeBranch(MBB&: *HBB, TBB, FBB, Cond&: HeadCond)) {
351 continue;
352 }
353
354 // Equivalence check is to skip loops.
355 if (!TBB || TBB == HBB) {
356 continue;
357 }
358
359 SmallVector<MachineOperand, 4> TrueCond;
360 MachineBasicBlock *TBB_TBB = nullptr, *TBB_FBB = nullptr;
361 if (TII->analyzeBranch(MBB&: *TBB, TBB&: TBB_TBB, FBB&: TBB_FBB, Cond&: TrueCond)) {
362 continue;
363 }
364
365 MachineInstr *HeadCmpMI = findSuitableCompare(MBB: HBB);
366 if (!HeadCmpMI) {
367 continue;
368 }
369
370 MachineInstr *TrueCmpMI = findSuitableCompare(MBB: TBB);
371 if (!TrueCmpMI) {
372 continue;
373 }
374
375 AArch64CC::CondCode HeadCmp;
376 if (HeadCond.empty() || !parseCond(Cond: HeadCond, CC&: HeadCmp)) {
377 continue;
378 }
379
380 AArch64CC::CondCode TrueCmp;
381 if (TrueCond.empty() || !parseCond(Cond: TrueCond, CC&: TrueCmp)) {
382 continue;
383 }
384
385 const int HeadImm = (int)HeadCmpMI->getOperand(i: 2).getImm();
386 const int TrueImm = (int)TrueCmpMI->getOperand(i: 2).getImm();
387
388 LLVM_DEBUG(dbgs() << "Head branch:\n");
389 LLVM_DEBUG(dbgs() << "\tcondition: " << AArch64CC::getCondCodeName(HeadCmp)
390 << '\n');
391 LLVM_DEBUG(dbgs() << "\timmediate: " << HeadImm << '\n');
392
393 LLVM_DEBUG(dbgs() << "True branch:\n");
394 LLVM_DEBUG(dbgs() << "\tcondition: " << AArch64CC::getCondCodeName(TrueCmp)
395 << '\n');
396 LLVM_DEBUG(dbgs() << "\timmediate: " << TrueImm << '\n');
397
398 if (((HeadCmp == AArch64CC::GT && TrueCmp == AArch64CC::LT) ||
399 (HeadCmp == AArch64CC::LT && TrueCmp == AArch64CC::GT)) &&
400 std::abs(x: TrueImm - HeadImm) == 2) {
401 // This branch transforms machine instructions that correspond to
402 //
403 // 1) (a > {TrueImm} && ...) || (a < {HeadImm} && ...)
404 // 2) (a < {TrueImm} && ...) || (a > {HeadImm} && ...)
405 //
406 // into
407 //
408 // 1) (a >= {NewImm} && ...) || (a <= {NewImm} && ...)
409 // 2) (a <= {NewImm} && ...) || (a >= {NewImm} && ...)
410
411 CmpInfo HeadCmpInfo = adjustCmp(CmpMI: HeadCmpMI, Cmp: HeadCmp);
412 CmpInfo TrueCmpInfo = adjustCmp(CmpMI: TrueCmpMI, Cmp: TrueCmp);
413 if (std::get<0>(t&: HeadCmpInfo) == std::get<0>(t&: TrueCmpInfo) &&
414 std::get<1>(t&: HeadCmpInfo) == std::get<1>(t&: TrueCmpInfo)) {
415 modifyCmp(CmpMI: HeadCmpMI, Info: HeadCmpInfo);
416 modifyCmp(CmpMI: TrueCmpMI, Info: TrueCmpInfo);
417 Changed = true;
418 }
419 } else if (((HeadCmp == AArch64CC::GT && TrueCmp == AArch64CC::GT) ||
420 (HeadCmp == AArch64CC::LT && TrueCmp == AArch64CC::LT)) &&
421 std::abs(x: TrueImm - HeadImm) == 1) {
422 // This branch transforms machine instructions that correspond to
423 //
424 // 1) (a > {TrueImm} && ...) || (a > {HeadImm} && ...)
425 // 2) (a < {TrueImm} && ...) || (a < {HeadImm} && ...)
426 //
427 // into
428 //
429 // 1) (a <= {NewImm} && ...) || (a > {NewImm} && ...)
430 // 2) (a < {NewImm} && ...) || (a >= {NewImm} && ...)
431
432 // GT -> GE transformation increases immediate value, so picking the
433 // smaller one; LT -> LE decreases immediate value so invert the choice.
434 bool adjustHeadCond = (HeadImm < TrueImm);
435 if (HeadCmp == AArch64CC::LT) {
436 adjustHeadCond = !adjustHeadCond;
437 }
438
439 if (adjustHeadCond) {
440 Changed |= adjustTo(CmpMI: HeadCmpMI, Cmp: HeadCmp, To: TrueCmpMI, ToImm: TrueImm);
441 } else {
442 Changed |= adjustTo(CmpMI: TrueCmpMI, Cmp: TrueCmp, To: HeadCmpMI, ToImm: HeadImm);
443 }
444 }
445 // Other transformation cases almost never occur due to generation of < or >
446 // comparisons instead of <= and >=.
447 }
448
449 return Changed;
450}
451

source code of llvm/lib/Target/AArch64/AArch64ConditionOptimizer.cpp