xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
1 //=== AArch64PostLegalizerCombiner.cpp --------------------------*- C++ -*-===//
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 /// \file
10 /// Post-legalization combines on generic MachineInstrs.
11 ///
12 /// The combines here must preserve instruction legality.
13 ///
14 /// Lowering combines (e.g. pseudo matching) should be handled by
15 /// AArch64PostLegalizerLowering.
16 ///
17 /// Combines which don't rely on instruction legality should go in the
18 /// AArch64PreLegalizerCombiner.
19 ///
20 //===----------------------------------------------------------------------===//
21 
22 #include "AArch64TargetMachine.h"
23 #include "llvm/CodeGen/GlobalISel/CSEInfo.h"
24 #include "llvm/CodeGen/GlobalISel/Combiner.h"
25 #include "llvm/CodeGen/GlobalISel/CombinerHelper.h"
26 #include "llvm/CodeGen/GlobalISel/CombinerInfo.h"
27 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h"
28 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
29 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
30 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
31 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
32 #include "llvm/CodeGen/GlobalISel/Utils.h"
33 #include "llvm/CodeGen/MachineDominators.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/TargetOpcodes.h"
37 #include "llvm/CodeGen/TargetPassConfig.h"
38 #include "llvm/Support/Debug.h"
39 
40 #define DEBUG_TYPE "aarch64-postlegalizer-combiner"
41 
42 using namespace llvm;
43 using namespace MIPatternMatch;
44 
45 /// This combine tries do what performExtractVectorEltCombine does in SDAG.
46 /// Rewrite for pairwise fadd pattern
47 ///   (s32 (g_extract_vector_elt
48 ///           (g_fadd (vXs32 Other)
49 ///                  (g_vector_shuffle (vXs32 Other) undef <1,X,...> )) 0))
50 /// ->
51 ///   (s32 (g_fadd (g_extract_vector_elt (vXs32 Other) 0)
52 ///              (g_extract_vector_elt (vXs32 Other) 1))
53 bool matchExtractVecEltPairwiseAdd(
54     MachineInstr &MI, MachineRegisterInfo &MRI,
55     std::tuple<unsigned, LLT, Register> &MatchInfo) {
56   Register Src1 = MI.getOperand(1).getReg();
57   Register Src2 = MI.getOperand(2).getReg();
58   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
59 
60   auto Cst = getIConstantVRegValWithLookThrough(Src2, MRI);
61   if (!Cst || Cst->Value != 0)
62     return false;
63   // SDAG also checks for FullFP16, but this looks to be beneficial anyway.
64 
65   // Now check for an fadd operation. TODO: expand this for integer add?
66   auto *FAddMI = getOpcodeDef(TargetOpcode::G_FADD, Src1, MRI);
67   if (!FAddMI)
68     return false;
69 
70   // If we add support for integer add, must restrict these types to just s64.
71   unsigned DstSize = DstTy.getSizeInBits();
72   if (DstSize != 16 && DstSize != 32 && DstSize != 64)
73     return false;
74 
75   Register Src1Op1 = FAddMI->getOperand(1).getReg();
76   Register Src1Op2 = FAddMI->getOperand(2).getReg();
77   MachineInstr *Shuffle =
78       getOpcodeDef(TargetOpcode::G_SHUFFLE_VECTOR, Src1Op2, MRI);
79   MachineInstr *Other = MRI.getVRegDef(Src1Op1);
80   if (!Shuffle) {
81     Shuffle = getOpcodeDef(TargetOpcode::G_SHUFFLE_VECTOR, Src1Op1, MRI);
82     Other = MRI.getVRegDef(Src1Op2);
83   }
84 
85   // We're looking for a shuffle that moves the second element to index 0.
86   if (Shuffle && Shuffle->getOperand(3).getShuffleMask()[0] == 1 &&
87       Other == MRI.getVRegDef(Shuffle->getOperand(1).getReg())) {
88     std::get<0>(MatchInfo) = TargetOpcode::G_FADD;
89     std::get<1>(MatchInfo) = DstTy;
90     std::get<2>(MatchInfo) = Other->getOperand(0).getReg();
91     return true;
92   }
93   return false;
94 }
95 
96 bool applyExtractVecEltPairwiseAdd(
97     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
98     std::tuple<unsigned, LLT, Register> &MatchInfo) {
99   unsigned Opc = std::get<0>(MatchInfo);
100   assert(Opc == TargetOpcode::G_FADD && "Unexpected opcode!");
101   // We want to generate two extracts of elements 0 and 1, and add them.
102   LLT Ty = std::get<1>(MatchInfo);
103   Register Src = std::get<2>(MatchInfo);
104   LLT s64 = LLT::scalar(64);
105   B.setInstrAndDebugLoc(MI);
106   auto Elt0 = B.buildExtractVectorElement(Ty, Src, B.buildConstant(s64, 0));
107   auto Elt1 = B.buildExtractVectorElement(Ty, Src, B.buildConstant(s64, 1));
108   B.buildInstr(Opc, {MI.getOperand(0).getReg()}, {Elt0, Elt1});
109   MI.eraseFromParent();
110   return true;
111 }
112 
113 static bool isSignExtended(Register R, MachineRegisterInfo &MRI) {
114   // TODO: check if extended build vector as well.
115   unsigned Opc = MRI.getVRegDef(R)->getOpcode();
116   return Opc == TargetOpcode::G_SEXT || Opc == TargetOpcode::G_SEXT_INREG;
117 }
118 
119 static bool isZeroExtended(Register R, MachineRegisterInfo &MRI) {
120   // TODO: check if extended build vector as well.
121   return MRI.getVRegDef(R)->getOpcode() == TargetOpcode::G_ZEXT;
122 }
123 
124 bool matchAArch64MulConstCombine(
125     MachineInstr &MI, MachineRegisterInfo &MRI,
126     std::function<void(MachineIRBuilder &B, Register DstReg)> &ApplyFn) {
127   assert(MI.getOpcode() == TargetOpcode::G_MUL);
128   Register LHS = MI.getOperand(1).getReg();
129   Register RHS = MI.getOperand(2).getReg();
130   Register Dst = MI.getOperand(0).getReg();
131   const LLT Ty = MRI.getType(LHS);
132 
133   // The below optimizations require a constant RHS.
134   auto Const = getIConstantVRegValWithLookThrough(RHS, MRI);
135   if (!Const)
136     return false;
137 
138   APInt ConstValue = Const->Value.sext(Ty.getSizeInBits());
139   // The following code is ported from AArch64ISelLowering.
140   // Multiplication of a power of two plus/minus one can be done more
141   // cheaply as as shift+add/sub. For now, this is true unilaterally. If
142   // future CPUs have a cheaper MADD instruction, this may need to be
143   // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
144   // 64-bit is 5 cycles, so this is always a win.
145   // More aggressively, some multiplications N0 * C can be lowered to
146   // shift+add+shift if the constant C = A * B where A = 2^N + 1 and B = 2^M,
147   // e.g. 6=3*2=(2+1)*2.
148   // TODO: consider lowering more cases, e.g. C = 14, -6, -14 or even 45
149   // which equals to (1+2)*16-(1+2).
150   // TrailingZeroes is used to test if the mul can be lowered to
151   // shift+add+shift.
152   unsigned TrailingZeroes = ConstValue.countTrailingZeros();
153   if (TrailingZeroes) {
154     // Conservatively do not lower to shift+add+shift if the mul might be
155     // folded into smul or umul.
156     if (MRI.hasOneNonDBGUse(LHS) &&
157         (isSignExtended(LHS, MRI) || isZeroExtended(LHS, MRI)))
158       return false;
159     // Conservatively do not lower to shift+add+shift if the mul might be
160     // folded into madd or msub.
161     if (MRI.hasOneNonDBGUse(Dst)) {
162       MachineInstr &UseMI = *MRI.use_instr_begin(Dst);
163       unsigned UseOpc = UseMI.getOpcode();
164       if (UseOpc == TargetOpcode::G_ADD || UseOpc == TargetOpcode::G_PTR_ADD ||
165           UseOpc == TargetOpcode::G_SUB)
166         return false;
167     }
168   }
169   // Use ShiftedConstValue instead of ConstValue to support both shift+add/sub
170   // and shift+add+shift.
171   APInt ShiftedConstValue = ConstValue.ashr(TrailingZeroes);
172 
173   unsigned ShiftAmt, AddSubOpc;
174   // Is the shifted value the LHS operand of the add/sub?
175   bool ShiftValUseIsLHS = true;
176   // Do we need to negate the result?
177   bool NegateResult = false;
178 
179   if (ConstValue.isNonNegative()) {
180     // (mul x, 2^N + 1) => (add (shl x, N), x)
181     // (mul x, 2^N - 1) => (sub (shl x, N), x)
182     // (mul x, (2^N + 1) * 2^M) => (shl (add (shl x, N), x), M)
183     APInt SCVMinus1 = ShiftedConstValue - 1;
184     APInt CVPlus1 = ConstValue + 1;
185     if (SCVMinus1.isPowerOf2()) {
186       ShiftAmt = SCVMinus1.logBase2();
187       AddSubOpc = TargetOpcode::G_ADD;
188     } else if (CVPlus1.isPowerOf2()) {
189       ShiftAmt = CVPlus1.logBase2();
190       AddSubOpc = TargetOpcode::G_SUB;
191     } else
192       return false;
193   } else {
194     // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
195     // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
196     APInt CVNegPlus1 = -ConstValue + 1;
197     APInt CVNegMinus1 = -ConstValue - 1;
198     if (CVNegPlus1.isPowerOf2()) {
199       ShiftAmt = CVNegPlus1.logBase2();
200       AddSubOpc = TargetOpcode::G_SUB;
201       ShiftValUseIsLHS = false;
202     } else if (CVNegMinus1.isPowerOf2()) {
203       ShiftAmt = CVNegMinus1.logBase2();
204       AddSubOpc = TargetOpcode::G_ADD;
205       NegateResult = true;
206     } else
207       return false;
208   }
209 
210   if (NegateResult && TrailingZeroes)
211     return false;
212 
213   ApplyFn = [=](MachineIRBuilder &B, Register DstReg) {
214     auto Shift = B.buildConstant(LLT::scalar(64), ShiftAmt);
215     auto ShiftedVal = B.buildShl(Ty, LHS, Shift);
216 
217     Register AddSubLHS = ShiftValUseIsLHS ? ShiftedVal.getReg(0) : LHS;
218     Register AddSubRHS = ShiftValUseIsLHS ? LHS : ShiftedVal.getReg(0);
219     auto Res = B.buildInstr(AddSubOpc, {Ty}, {AddSubLHS, AddSubRHS});
220     assert(!(NegateResult && TrailingZeroes) &&
221            "NegateResult and TrailingZeroes cannot both be true for now.");
222     // Negate the result.
223     if (NegateResult) {
224       B.buildSub(DstReg, B.buildConstant(Ty, 0), Res);
225       return;
226     }
227     // Shift the result.
228     if (TrailingZeroes) {
229       B.buildShl(DstReg, Res, B.buildConstant(LLT::scalar(64), TrailingZeroes));
230       return;
231     }
232     B.buildCopy(DstReg, Res.getReg(0));
233   };
234   return true;
235 }
236 
237 bool applyAArch64MulConstCombine(
238     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
239     std::function<void(MachineIRBuilder &B, Register DstReg)> &ApplyFn) {
240   B.setInstrAndDebugLoc(MI);
241   ApplyFn(B, MI.getOperand(0).getReg());
242   MI.eraseFromParent();
243   return true;
244 }
245 
246 /// Try to fold a G_MERGE_VALUES of 2 s32 sources, where the second source
247 /// is a zero, into a G_ZEXT of the first.
248 bool matchFoldMergeToZext(MachineInstr &MI, MachineRegisterInfo &MRI) {
249   auto &Merge = cast<GMerge>(MI);
250   LLT SrcTy = MRI.getType(Merge.getSourceReg(0));
251   if (SrcTy != LLT::scalar(32) || Merge.getNumSources() != 2)
252     return false;
253   return mi_match(Merge.getSourceReg(1), MRI, m_SpecificICst(0));
254 }
255 
256 void applyFoldMergeToZext(MachineInstr &MI, MachineRegisterInfo &MRI,
257                           MachineIRBuilder &B, GISelChangeObserver &Observer) {
258   // Mutate %d(s64) = G_MERGE_VALUES %a(s32), 0(s32)
259   //  ->
260   // %d(s64) = G_ZEXT %a(s32)
261   Observer.changingInstr(MI);
262   MI.setDesc(B.getTII().get(TargetOpcode::G_ZEXT));
263   MI.removeOperand(2);
264   Observer.changedInstr(MI);
265 }
266 
267 /// \returns True if a G_ANYEXT instruction \p MI should be mutated to a G_ZEXT
268 /// instruction.
269 static bool matchMutateAnyExtToZExt(MachineInstr &MI, MachineRegisterInfo &MRI) {
270   // If this is coming from a scalar compare then we can use a G_ZEXT instead of
271   // a G_ANYEXT:
272   //
273   // %cmp:_(s32) = G_[I|F]CMP ... <-- produces 0/1.
274   // %ext:_(s64) = G_ANYEXT %cmp(s32)
275   //
276   // By doing this, we can leverage more KnownBits combines.
277   assert(MI.getOpcode() == TargetOpcode::G_ANYEXT);
278   Register Dst = MI.getOperand(0).getReg();
279   Register Src = MI.getOperand(1).getReg();
280   return MRI.getType(Dst).isScalar() &&
281          mi_match(Src, MRI,
282                   m_any_of(m_GICmp(m_Pred(), m_Reg(), m_Reg()),
283                            m_GFCmp(m_Pred(), m_Reg(), m_Reg())));
284 }
285 
286 static void applyMutateAnyExtToZExt(MachineInstr &MI, MachineRegisterInfo &MRI,
287                               MachineIRBuilder &B,
288                               GISelChangeObserver &Observer) {
289   Observer.changingInstr(MI);
290   MI.setDesc(B.getTII().get(TargetOpcode::G_ZEXT));
291   Observer.changedInstr(MI);
292 }
293 
294 /// Match a 128b store of zero and split it into two 64 bit stores, for
295 /// size/performance reasons.
296 static bool matchSplitStoreZero128(MachineInstr &MI, MachineRegisterInfo &MRI) {
297   GStore &Store = cast<GStore>(MI);
298   if (!Store.isSimple())
299     return false;
300   LLT ValTy = MRI.getType(Store.getValueReg());
301   if (!ValTy.isVector() || ValTy.getSizeInBits() != 128)
302     return false;
303   if (ValTy.getSizeInBits() != Store.getMemSizeInBits())
304     return false; // Don't split truncating stores.
305   if (!MRI.hasOneNonDBGUse(Store.getValueReg()))
306     return false;
307   auto MaybeCst = isConstantOrConstantSplatVector(
308       *MRI.getVRegDef(Store.getValueReg()), MRI);
309   return MaybeCst && MaybeCst->isZero();
310 }
311 
312 static void applySplitStoreZero128(MachineInstr &MI, MachineRegisterInfo &MRI,
313                                    MachineIRBuilder &B,
314                                    GISelChangeObserver &Observer) {
315   B.setInstrAndDebugLoc(MI);
316   GStore &Store = cast<GStore>(MI);
317   assert(MRI.getType(Store.getValueReg()).isVector() &&
318          "Expected a vector store value");
319   LLT NewTy = LLT::scalar(64);
320   Register PtrReg = Store.getPointerReg();
321   auto Zero = B.buildConstant(NewTy, 0);
322   auto HighPtr = B.buildPtrAdd(MRI.getType(PtrReg), PtrReg,
323                                B.buildConstant(LLT::scalar(64), 8));
324   auto &MF = *MI.getMF();
325   auto *LowMMO = MF.getMachineMemOperand(&Store.getMMO(), 0, NewTy);
326   auto *HighMMO = MF.getMachineMemOperand(&Store.getMMO(), 8, NewTy);
327   B.buildStore(Zero, PtrReg, *LowMMO);
328   B.buildStore(Zero, HighPtr, *HighMMO);
329   Store.eraseFromParent();
330 }
331 
332 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_DEPS
333 #include "AArch64GenPostLegalizeGICombiner.inc"
334 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_DEPS
335 
336 namespace {
337 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_H
338 #include "AArch64GenPostLegalizeGICombiner.inc"
339 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_H
340 
341 class AArch64PostLegalizerCombinerInfo : public CombinerInfo {
342   GISelKnownBits *KB;
343   MachineDominatorTree *MDT;
344 
345 public:
346   AArch64GenPostLegalizerCombinerHelperRuleConfig GeneratedRuleCfg;
347 
348   AArch64PostLegalizerCombinerInfo(bool EnableOpt, bool OptSize, bool MinSize,
349                                    GISelKnownBits *KB,
350                                    MachineDominatorTree *MDT)
351       : CombinerInfo(/*AllowIllegalOps*/ true, /*ShouldLegalizeIllegal*/ false,
352                      /*LegalizerInfo*/ nullptr, EnableOpt, OptSize, MinSize),
353         KB(KB), MDT(MDT) {
354     if (!GeneratedRuleCfg.parseCommandLineOption())
355       report_fatal_error("Invalid rule identifier");
356   }
357 
358   bool combine(GISelChangeObserver &Observer, MachineInstr &MI,
359                MachineIRBuilder &B) const override;
360 };
361 
362 bool AArch64PostLegalizerCombinerInfo::combine(GISelChangeObserver &Observer,
363                                                MachineInstr &MI,
364                                                MachineIRBuilder &B) const {
365   const auto *LI =
366       MI.getParent()->getParent()->getSubtarget().getLegalizerInfo();
367   CombinerHelper Helper(Observer, B, KB, MDT, LI);
368   AArch64GenPostLegalizerCombinerHelper Generated(GeneratedRuleCfg);
369   return Generated.tryCombineAll(Observer, MI, B, Helper);
370 }
371 
372 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_CPP
373 #include "AArch64GenPostLegalizeGICombiner.inc"
374 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_CPP
375 
376 class AArch64PostLegalizerCombiner : public MachineFunctionPass {
377 public:
378   static char ID;
379 
380   AArch64PostLegalizerCombiner(bool IsOptNone = false);
381 
382   StringRef getPassName() const override {
383     return "AArch64PostLegalizerCombiner";
384   }
385 
386   bool runOnMachineFunction(MachineFunction &MF) override;
387   void getAnalysisUsage(AnalysisUsage &AU) const override;
388 
389 private:
390   bool IsOptNone;
391 };
392 } // end anonymous namespace
393 
394 void AArch64PostLegalizerCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
395   AU.addRequired<TargetPassConfig>();
396   AU.setPreservesCFG();
397   getSelectionDAGFallbackAnalysisUsage(AU);
398   AU.addRequired<GISelKnownBitsAnalysis>();
399   AU.addPreserved<GISelKnownBitsAnalysis>();
400   if (!IsOptNone) {
401     AU.addRequired<MachineDominatorTree>();
402     AU.addPreserved<MachineDominatorTree>();
403     AU.addRequired<GISelCSEAnalysisWrapperPass>();
404     AU.addPreserved<GISelCSEAnalysisWrapperPass>();
405   }
406   MachineFunctionPass::getAnalysisUsage(AU);
407 }
408 
409 AArch64PostLegalizerCombiner::AArch64PostLegalizerCombiner(bool IsOptNone)
410     : MachineFunctionPass(ID), IsOptNone(IsOptNone) {
411   initializeAArch64PostLegalizerCombinerPass(*PassRegistry::getPassRegistry());
412 }
413 
414 bool AArch64PostLegalizerCombiner::runOnMachineFunction(MachineFunction &MF) {
415   if (MF.getProperties().hasProperty(
416           MachineFunctionProperties::Property::FailedISel))
417     return false;
418   assert(MF.getProperties().hasProperty(
419              MachineFunctionProperties::Property::Legalized) &&
420          "Expected a legalized function?");
421   auto *TPC = &getAnalysis<TargetPassConfig>();
422   const Function &F = MF.getFunction();
423   bool EnableOpt =
424       MF.getTarget().getOptLevel() != CodeGenOpt::None && !skipFunction(F);
425   GISelKnownBits *KB = &getAnalysis<GISelKnownBitsAnalysis>().get(MF);
426   MachineDominatorTree *MDT =
427       IsOptNone ? nullptr : &getAnalysis<MachineDominatorTree>();
428   AArch64PostLegalizerCombinerInfo PCInfo(EnableOpt, F.hasOptSize(),
429                                           F.hasMinSize(), KB, MDT);
430   GISelCSEAnalysisWrapper &Wrapper =
431       getAnalysis<GISelCSEAnalysisWrapperPass>().getCSEWrapper();
432   auto *CSEInfo = &Wrapper.get(TPC->getCSEConfig());
433   Combiner C(PCInfo, TPC);
434   return C.combineMachineInstrs(MF, CSEInfo);
435 }
436 
437 char AArch64PostLegalizerCombiner::ID = 0;
438 INITIALIZE_PASS_BEGIN(AArch64PostLegalizerCombiner, DEBUG_TYPE,
439                       "Combine AArch64 MachineInstrs after legalization", false,
440                       false)
441 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
442 INITIALIZE_PASS_DEPENDENCY(GISelKnownBitsAnalysis)
443 INITIALIZE_PASS_END(AArch64PostLegalizerCombiner, DEBUG_TYPE,
444                     "Combine AArch64 MachineInstrs after legalization", false,
445                     false)
446 
447 namespace llvm {
448 FunctionPass *createAArch64PostLegalizerCombiner(bool IsOptNone) {
449   return new AArch64PostLegalizerCombiner(IsOptNone);
450 }
451 } // end namespace llvm
452