xref: /freebsd/contrib/llvm-project/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp (revision d56accc7c3dcc897489b6a07834763a03b9f3d68)
1 //===- RISCVInsertVSETVLI.cpp - Insert VSETVLI instructions ---------------===//
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 file implements a function pass that inserts VSETVLI instructions where
10 // needed.
11 //
12 // This pass consists of 3 phases:
13 //
14 // Phase 1 collects how each basic block affects VL/VTYPE.
15 //
16 // Phase 2 uses the information from phase 1 to do a data flow analysis to
17 // propagate the VL/VTYPE changes through the function. This gives us the
18 // VL/VTYPE at the start of each basic block.
19 //
20 // Phase 3 inserts VSETVLI instructions in each basic block. Information from
21 // phase 2 is used to prevent inserting a VSETVLI before the first vector
22 // instruction in the block if possible.
23 //
24 //===----------------------------------------------------------------------===//
25 
26 #include "RISCV.h"
27 #include "RISCVSubtarget.h"
28 #include "llvm/CodeGen/LiveIntervals.h"
29 #include "llvm/CodeGen/MachineFunctionPass.h"
30 #include <queue>
31 using namespace llvm;
32 
33 #define DEBUG_TYPE "riscv-insert-vsetvli"
34 #define RISCV_INSERT_VSETVLI_NAME "RISCV Insert VSETVLI pass"
35 
36 static cl::opt<bool> DisableInsertVSETVLPHIOpt(
37     "riscv-disable-insert-vsetvl-phi-opt", cl::init(false), cl::Hidden,
38     cl::desc("Disable looking through phis when inserting vsetvlis."));
39 
40 namespace {
41 
42 class VSETVLIInfo {
43   union {
44     Register AVLReg;
45     unsigned AVLImm;
46   };
47 
48   enum : uint8_t {
49     Uninitialized,
50     AVLIsReg,
51     AVLIsImm,
52     Unknown,
53   } State = Uninitialized;
54 
55   // Fields from VTYPE.
56   RISCVII::VLMUL VLMul = RISCVII::LMUL_1;
57   uint8_t SEW = 0;
58   uint8_t TailAgnostic : 1;
59   uint8_t MaskAgnostic : 1;
60   uint8_t MaskRegOp : 1;
61   uint8_t StoreOp : 1;
62   uint8_t ScalarMovOp : 1;
63   uint8_t SEWLMULRatioOnly : 1;
64 
65 public:
66   VSETVLIInfo()
67       : AVLImm(0), TailAgnostic(false), MaskAgnostic(false), MaskRegOp(false),
68         StoreOp(false), ScalarMovOp(false), SEWLMULRatioOnly(false) {}
69 
70   static VSETVLIInfo getUnknown() {
71     VSETVLIInfo Info;
72     Info.setUnknown();
73     return Info;
74   }
75 
76   bool isValid() const { return State != Uninitialized; }
77   void setUnknown() { State = Unknown; }
78   bool isUnknown() const { return State == Unknown; }
79 
80   void setAVLReg(Register Reg) {
81     AVLReg = Reg;
82     State = AVLIsReg;
83   }
84 
85   void setAVLImm(unsigned Imm) {
86     AVLImm = Imm;
87     State = AVLIsImm;
88   }
89 
90   bool hasAVLImm() const { return State == AVLIsImm; }
91   bool hasAVLReg() const { return State == AVLIsReg; }
92   Register getAVLReg() const {
93     assert(hasAVLReg());
94     return AVLReg;
95   }
96   unsigned getAVLImm() const {
97     assert(hasAVLImm());
98     return AVLImm;
99   }
100   bool hasZeroAVL() const {
101     if (hasAVLImm())
102       return getAVLImm() == 0;
103     return false;
104   }
105   bool hasNonZeroAVL() const {
106     if (hasAVLImm())
107       return getAVLImm() > 0;
108     if (hasAVLReg())
109       return getAVLReg() == RISCV::X0;
110     return false;
111   }
112 
113   bool hasSameAVL(const VSETVLIInfo &Other) const {
114     assert(isValid() && Other.isValid() &&
115            "Can't compare invalid VSETVLIInfos");
116     assert(!isUnknown() && !Other.isUnknown() &&
117            "Can't compare AVL in unknown state");
118     if (hasAVLReg() && Other.hasAVLReg())
119       return getAVLReg() == Other.getAVLReg();
120 
121     if (hasAVLImm() && Other.hasAVLImm())
122       return getAVLImm() == Other.getAVLImm();
123 
124     return false;
125   }
126 
127   void setVTYPE(unsigned VType) {
128     assert(isValid() && !isUnknown() &&
129            "Can't set VTYPE for uninitialized or unknown");
130     VLMul = RISCVVType::getVLMUL(VType);
131     SEW = RISCVVType::getSEW(VType);
132     TailAgnostic = RISCVVType::isTailAgnostic(VType);
133     MaskAgnostic = RISCVVType::isMaskAgnostic(VType);
134   }
135   void setVTYPE(RISCVII::VLMUL L, unsigned S, bool TA, bool MA, bool MRO,
136                 bool IsStore, bool IsScalarMovOp) {
137     assert(isValid() && !isUnknown() &&
138            "Can't set VTYPE for uninitialized or unknown");
139     VLMul = L;
140     SEW = S;
141     TailAgnostic = TA;
142     MaskAgnostic = MA;
143     MaskRegOp = MRO;
144     StoreOp = IsStore;
145     ScalarMovOp = IsScalarMovOp;
146   }
147 
148   unsigned encodeVTYPE() const {
149     assert(isValid() && !isUnknown() && !SEWLMULRatioOnly &&
150            "Can't encode VTYPE for uninitialized or unknown");
151     return RISCVVType::encodeVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
152   }
153 
154   bool hasSEWLMULRatioOnly() const { return SEWLMULRatioOnly; }
155 
156   bool hasSameSEW(const VSETVLIInfo &Other) const {
157     assert(isValid() && Other.isValid() &&
158            "Can't compare invalid VSETVLIInfos");
159     assert(!isUnknown() && !Other.isUnknown() &&
160            "Can't compare VTYPE in unknown state");
161     assert(!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly &&
162            "Can't compare when only LMUL/SEW ratio is valid.");
163     return SEW == Other.SEW;
164   }
165 
166   bool hasSameVTYPE(const VSETVLIInfo &Other) const {
167     assert(isValid() && Other.isValid() &&
168            "Can't compare invalid VSETVLIInfos");
169     assert(!isUnknown() && !Other.isUnknown() &&
170            "Can't compare VTYPE in unknown state");
171     assert(!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly &&
172            "Can't compare when only LMUL/SEW ratio is valid.");
173     return std::tie(VLMul, SEW, TailAgnostic, MaskAgnostic) ==
174            std::tie(Other.VLMul, Other.SEW, Other.TailAgnostic,
175                     Other.MaskAgnostic);
176   }
177 
178   static unsigned getSEWLMULRatio(unsigned SEW, RISCVII::VLMUL VLMul) {
179     unsigned LMul;
180     bool Fractional;
181     std::tie(LMul, Fractional) = RISCVVType::decodeVLMUL(VLMul);
182 
183     // Convert LMul to a fixed point value with 3 fractional bits.
184     LMul = Fractional ? (8 / LMul) : (LMul * 8);
185 
186     assert(SEW >= 8 && "Unexpected SEW value");
187     return (SEW * 8) / LMul;
188   }
189 
190   unsigned getSEWLMULRatio() const {
191     assert(isValid() && !isUnknown() &&
192            "Can't use VTYPE for uninitialized or unknown");
193     return getSEWLMULRatio(SEW, VLMul);
194   }
195 
196   // Check if the VTYPE for these two VSETVLIInfos produce the same VLMAX.
197   bool hasSameVLMAX(const VSETVLIInfo &Other) const {
198     assert(isValid() && Other.isValid() &&
199            "Can't compare invalid VSETVLIInfos");
200     assert(!isUnknown() && !Other.isUnknown() &&
201            "Can't compare VTYPE in unknown state");
202     return getSEWLMULRatio() == Other.getSEWLMULRatio();
203   }
204 
205   bool hasSamePolicy(const VSETVLIInfo &Other) const {
206     assert(isValid() && Other.isValid() &&
207            "Can't compare invalid VSETVLIInfos");
208     assert(!isUnknown() && !Other.isUnknown() &&
209            "Can't compare VTYPE in unknown state");
210     return TailAgnostic == Other.TailAgnostic &&
211            MaskAgnostic == Other.MaskAgnostic;
212   }
213 
214   bool hasCompatibleVTYPE(const VSETVLIInfo &InstrInfo, bool Strict) const {
215     // Simple case, see if full VTYPE matches.
216     if (hasSameVTYPE(InstrInfo))
217       return true;
218 
219     if (Strict)
220       return false;
221 
222     // If this is a mask reg operation, it only cares about VLMAX.
223     // FIXME: Mask reg operations are probably ok if "this" VLMAX is larger
224     // than "InstrInfo".
225     // FIXME: The policy bits can probably be ignored for mask reg operations.
226     if (InstrInfo.MaskRegOp && hasSameVLMAX(InstrInfo) &&
227         TailAgnostic == InstrInfo.TailAgnostic &&
228         MaskAgnostic == InstrInfo.MaskAgnostic)
229       return true;
230 
231     return false;
232   }
233 
234   // Determine whether the vector instructions requirements represented by
235   // InstrInfo are compatible with the previous vsetvli instruction represented
236   // by this.
237   bool isCompatible(const VSETVLIInfo &InstrInfo, bool Strict) const {
238     assert(isValid() && InstrInfo.isValid() &&
239            "Can't compare invalid VSETVLIInfos");
240     assert(!InstrInfo.SEWLMULRatioOnly &&
241            "Expected a valid VTYPE for instruction!");
242     // Nothing is compatible with Unknown.
243     if (isUnknown() || InstrInfo.isUnknown())
244       return false;
245 
246     // If only our VLMAX ratio is valid, then this isn't compatible.
247     if (SEWLMULRatioOnly)
248       return false;
249 
250     // If the instruction doesn't need an AVLReg and the SEW matches, consider
251     // it compatible.
252     if (!Strict && InstrInfo.hasAVLReg() &&
253         InstrInfo.AVLReg == RISCV::NoRegister) {
254       if (SEW == InstrInfo.SEW)
255         return true;
256     }
257 
258     // For vmv.s.x and vfmv.s.f, there is only two behaviors, VL = 0 and VL > 0.
259     // So it's compatible when we could make sure that both VL be the same
260     // situation.
261     if (!Strict && InstrInfo.ScalarMovOp && InstrInfo.hasAVLImm() &&
262         ((hasNonZeroAVL() && InstrInfo.hasNonZeroAVL()) ||
263          (hasZeroAVL() && InstrInfo.hasZeroAVL())) &&
264         hasSameSEW(InstrInfo) && hasSamePolicy(InstrInfo))
265       return true;
266 
267     // The AVL must match.
268     if (!hasSameAVL(InstrInfo))
269       return false;
270 
271     if (hasCompatibleVTYPE(InstrInfo, Strict))
272       return true;
273 
274     // Strict matches must ensure a full VTYPE match.
275     if (Strict)
276       return false;
277 
278     // Store instructions don't use the policy fields.
279     // TODO: Move into hasCompatibleVTYPE?
280     if (InstrInfo.StoreOp && VLMul == InstrInfo.VLMul && SEW == InstrInfo.SEW)
281       return true;
282 
283     // Anything else is not compatible.
284     return false;
285   }
286 
287   bool isCompatibleWithLoadStoreEEW(unsigned EEW,
288                                     const VSETVLIInfo &InstrInfo) const {
289     assert(isValid() && InstrInfo.isValid() &&
290            "Can't compare invalid VSETVLIInfos");
291     assert(!InstrInfo.SEWLMULRatioOnly &&
292            "Expected a valid VTYPE for instruction!");
293     assert(EEW == InstrInfo.SEW && "Mismatched EEW/SEW for store");
294 
295     if (isUnknown() || hasSEWLMULRatioOnly())
296       return false;
297 
298     if (!hasSameAVL(InstrInfo))
299       return false;
300 
301     // Stores can ignore the tail and mask policies.
302     if (!InstrInfo.StoreOp && (TailAgnostic != InstrInfo.TailAgnostic ||
303                                MaskAgnostic != InstrInfo.MaskAgnostic))
304       return false;
305 
306     return getSEWLMULRatio() == getSEWLMULRatio(EEW, InstrInfo.VLMul);
307   }
308 
309   bool operator==(const VSETVLIInfo &Other) const {
310     // Uninitialized is only equal to another Uninitialized.
311     if (!isValid())
312       return !Other.isValid();
313     if (!Other.isValid())
314       return !isValid();
315 
316     // Unknown is only equal to another Unknown.
317     if (isUnknown())
318       return Other.isUnknown();
319     if (Other.isUnknown())
320       return isUnknown();
321 
322     if (!hasSameAVL(Other))
323       return false;
324 
325     // If only the VLMAX is valid, check that it is the same.
326     if (SEWLMULRatioOnly && Other.SEWLMULRatioOnly)
327       return hasSameVLMAX(Other);
328 
329     // If the full VTYPE is valid, check that it is the same.
330     if (!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly)
331       return hasSameVTYPE(Other);
332 
333     // If the SEWLMULRatioOnly bits are different, then they aren't equal.
334     return false;
335   }
336 
337   bool operator!=(const VSETVLIInfo &Other) const {
338     return !(*this == Other);
339   }
340 
341   // Calculate the VSETVLIInfo visible to a block assuming this and Other are
342   // both predecessors.
343   VSETVLIInfo intersect(const VSETVLIInfo &Other) const {
344     // If the new value isn't valid, ignore it.
345     if (!Other.isValid())
346       return *this;
347 
348     // If this value isn't valid, this must be the first predecessor, use it.
349     if (!isValid())
350       return Other;
351 
352     // If either is unknown, the result is unknown.
353     if (isUnknown() || Other.isUnknown())
354       return VSETVLIInfo::getUnknown();
355 
356     // If we have an exact, match return this.
357     if (*this == Other)
358       return *this;
359 
360     // Not an exact match, but maybe the AVL and VLMAX are the same. If so,
361     // return an SEW/LMUL ratio only value.
362     if (hasSameAVL(Other) && hasSameVLMAX(Other)) {
363       VSETVLIInfo MergeInfo = *this;
364       MergeInfo.SEWLMULRatioOnly = true;
365       return MergeInfo;
366     }
367 
368     // Otherwise the result is unknown.
369     return VSETVLIInfo::getUnknown();
370   }
371 
372   // Calculate the VSETVLIInfo visible at the end of the block assuming this
373   // is the predecessor value, and Other is change for this block.
374   VSETVLIInfo merge(const VSETVLIInfo &Other) const {
375     assert(isValid() && "Can only merge with a valid VSETVLInfo");
376 
377     // Nothing changed from the predecessor, keep it.
378     if (!Other.isValid())
379       return *this;
380 
381     // If the change is compatible with the input, we won't create a VSETVLI
382     // and should keep the predecessor.
383     if (isCompatible(Other, /*Strict*/ true))
384       return *this;
385 
386     // Otherwise just use whatever is in this block.
387     return Other;
388   }
389 };
390 
391 struct BlockData {
392   // The VSETVLIInfo that represents the net changes to the VL/VTYPE registers
393   // made by this block. Calculated in Phase 1.
394   VSETVLIInfo Change;
395 
396   // The VSETVLIInfo that represents the VL/VTYPE settings on exit from this
397   // block. Calculated in Phase 2.
398   VSETVLIInfo Exit;
399 
400   // The VSETVLIInfo that represents the VL/VTYPE settings from all predecessor
401   // blocks. Calculated in Phase 2, and used by Phase 3.
402   VSETVLIInfo Pred;
403 
404   // Keeps track of whether the block is already in the queue.
405   bool InQueue = false;
406 
407   BlockData() {}
408 };
409 
410 class RISCVInsertVSETVLI : public MachineFunctionPass {
411   const TargetInstrInfo *TII;
412   MachineRegisterInfo *MRI;
413 
414   std::vector<BlockData> BlockInfo;
415   std::queue<const MachineBasicBlock *> WorkList;
416 
417 public:
418   static char ID;
419 
420   RISCVInsertVSETVLI() : MachineFunctionPass(ID) {
421     initializeRISCVInsertVSETVLIPass(*PassRegistry::getPassRegistry());
422   }
423   bool runOnMachineFunction(MachineFunction &MF) override;
424 
425   void getAnalysisUsage(AnalysisUsage &AU) const override {
426     AU.setPreservesCFG();
427     MachineFunctionPass::getAnalysisUsage(AU);
428   }
429 
430   StringRef getPassName() const override { return RISCV_INSERT_VSETVLI_NAME; }
431 
432 private:
433   bool needVSETVLI(const VSETVLIInfo &Require, const VSETVLIInfo &CurInfo);
434   bool needVSETVLIPHI(const VSETVLIInfo &Require, const MachineBasicBlock &MBB);
435   void insertVSETVLI(MachineBasicBlock &MBB, MachineInstr &MI,
436                      const VSETVLIInfo &Info, const VSETVLIInfo &PrevInfo);
437 
438   bool computeVLVTYPEChanges(const MachineBasicBlock &MBB);
439   void computeIncomingVLVTYPE(const MachineBasicBlock &MBB);
440   void emitVSETVLIs(MachineBasicBlock &MBB);
441 };
442 
443 } // end anonymous namespace
444 
445 char RISCVInsertVSETVLI::ID = 0;
446 
447 INITIALIZE_PASS(RISCVInsertVSETVLI, DEBUG_TYPE, RISCV_INSERT_VSETVLI_NAME,
448                 false, false)
449 
450 static MachineInstr *elideCopies(MachineInstr *MI,
451                                  const MachineRegisterInfo *MRI) {
452   while (true) {
453     if (!MI->isFullCopy())
454       return MI;
455     if (!Register::isVirtualRegister(MI->getOperand(1).getReg()))
456       return nullptr;
457     MI = MRI->getVRegDef(MI->getOperand(1).getReg());
458     if (!MI)
459       return nullptr;
460   }
461 }
462 
463 static bool isScalarMoveInstr(const MachineInstr &MI) {
464   switch (MI.getOpcode()) {
465   default:
466     return false;
467   case RISCV::PseudoVMV_S_X_M1:
468   case RISCV::PseudoVMV_S_X_M2:
469   case RISCV::PseudoVMV_S_X_M4:
470   case RISCV::PseudoVMV_S_X_M8:
471   case RISCV::PseudoVMV_S_X_MF2:
472   case RISCV::PseudoVMV_S_X_MF4:
473   case RISCV::PseudoVMV_S_X_MF8:
474   case RISCV::PseudoVFMV_S_F16_M1:
475   case RISCV::PseudoVFMV_S_F16_M2:
476   case RISCV::PseudoVFMV_S_F16_M4:
477   case RISCV::PseudoVFMV_S_F16_M8:
478   case RISCV::PseudoVFMV_S_F16_MF2:
479   case RISCV::PseudoVFMV_S_F16_MF4:
480   case RISCV::PseudoVFMV_S_F32_M1:
481   case RISCV::PseudoVFMV_S_F32_M2:
482   case RISCV::PseudoVFMV_S_F32_M4:
483   case RISCV::PseudoVFMV_S_F32_M8:
484   case RISCV::PseudoVFMV_S_F32_MF2:
485   case RISCV::PseudoVFMV_S_F64_M1:
486   case RISCV::PseudoVFMV_S_F64_M2:
487   case RISCV::PseudoVFMV_S_F64_M4:
488   case RISCV::PseudoVFMV_S_F64_M8:
489     return true;
490   }
491 }
492 
493 static VSETVLIInfo computeInfoForInstr(const MachineInstr &MI, uint64_t TSFlags,
494                                        const MachineRegisterInfo *MRI) {
495   VSETVLIInfo InstrInfo;
496   unsigned NumOperands = MI.getNumExplicitOperands();
497   bool HasPolicy = RISCVII::hasVecPolicyOp(TSFlags);
498 
499   // Default to tail agnostic unless the destination is tied to a source.
500   // Unless the source is undef. In that case the user would have some control
501   // over the tail values. Some pseudo instructions force a tail agnostic policy
502   // despite having a tied def.
503   bool ForceTailAgnostic = RISCVII::doesForceTailAgnostic(TSFlags);
504   bool TailAgnostic = true;
505   // If the instruction has policy argument, use the argument.
506   if (HasPolicy) {
507     const MachineOperand &Op = MI.getOperand(MI.getNumExplicitOperands() - 1);
508     TailAgnostic = Op.getImm() & 0x1;
509   }
510 
511   unsigned UseOpIdx;
512   if (!(ForceTailAgnostic || (HasPolicy && TailAgnostic)) &&
513       MI.isRegTiedToUseOperand(0, &UseOpIdx)) {
514     TailAgnostic = false;
515     // If the tied operand is an IMPLICIT_DEF we can keep TailAgnostic.
516     const MachineOperand &UseMO = MI.getOperand(UseOpIdx);
517     MachineInstr *UseMI = MRI->getVRegDef(UseMO.getReg());
518     if (UseMI) {
519       UseMI = elideCopies(UseMI, MRI);
520       if (UseMI && UseMI->isImplicitDef())
521         TailAgnostic = true;
522     }
523   }
524 
525   // Remove the tail policy so we can find the SEW and VL.
526   if (HasPolicy)
527     --NumOperands;
528 
529   RISCVII::VLMUL VLMul = RISCVII::getLMul(TSFlags);
530 
531   unsigned Log2SEW = MI.getOperand(NumOperands - 1).getImm();
532   // A Log2SEW of 0 is an operation on mask registers only.
533   bool MaskRegOp = Log2SEW == 0;
534   unsigned SEW = Log2SEW ? 1 << Log2SEW : 8;
535   assert(RISCVVType::isValidSEW(SEW) && "Unexpected SEW");
536 
537   // If there are no explicit defs, this is a store instruction which can
538   // ignore the tail and mask policies.
539   bool StoreOp = MI.getNumExplicitDefs() == 0;
540   bool ScalarMovOp = isScalarMoveInstr(MI);
541 
542   if (RISCVII::hasVLOp(TSFlags)) {
543     const MachineOperand &VLOp = MI.getOperand(NumOperands - 2);
544     if (VLOp.isImm()) {
545       int64_t Imm = VLOp.getImm();
546       // Conver the VLMax sentintel to X0 register.
547       if (Imm == RISCV::VLMaxSentinel)
548         InstrInfo.setAVLReg(RISCV::X0);
549       else
550         InstrInfo.setAVLImm(Imm);
551     } else {
552       InstrInfo.setAVLReg(VLOp.getReg());
553     }
554   } else
555     InstrInfo.setAVLReg(RISCV::NoRegister);
556   InstrInfo.setVTYPE(VLMul, SEW, /*TailAgnostic*/ TailAgnostic,
557                      /*MaskAgnostic*/ false, MaskRegOp, StoreOp, ScalarMovOp);
558 
559   return InstrInfo;
560 }
561 
562 void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB, MachineInstr &MI,
563                                        const VSETVLIInfo &Info,
564                                        const VSETVLIInfo &PrevInfo) {
565   DebugLoc DL = MI.getDebugLoc();
566 
567   // Use X0, X0 form if the AVL is the same and the SEW+LMUL gives the same
568   // VLMAX.
569   if (PrevInfo.isValid() && !PrevInfo.isUnknown() &&
570       Info.hasSameAVL(PrevInfo) && Info.hasSameVLMAX(PrevInfo)) {
571     BuildMI(MBB, MI, DL, TII->get(RISCV::PseudoVSETVLIX0))
572         .addReg(RISCV::X0, RegState::Define | RegState::Dead)
573         .addReg(RISCV::X0, RegState::Kill)
574         .addImm(Info.encodeVTYPE())
575         .addReg(RISCV::VL, RegState::Implicit);
576     return;
577   }
578 
579   if (Info.hasAVLImm()) {
580     BuildMI(MBB, MI, DL, TII->get(RISCV::PseudoVSETIVLI))
581         .addReg(RISCV::X0, RegState::Define | RegState::Dead)
582         .addImm(Info.getAVLImm())
583         .addImm(Info.encodeVTYPE());
584     return;
585   }
586 
587   Register AVLReg = Info.getAVLReg();
588   if (AVLReg == RISCV::NoRegister) {
589     // We can only use x0, x0 if there's no chance of the vtype change causing
590     // the previous vl to become invalid.
591     if (PrevInfo.isValid() && !PrevInfo.isUnknown() &&
592         Info.hasSameVLMAX(PrevInfo)) {
593       BuildMI(MBB, MI, DL, TII->get(RISCV::PseudoVSETVLIX0))
594           .addReg(RISCV::X0, RegState::Define | RegState::Dead)
595           .addReg(RISCV::X0, RegState::Kill)
596           .addImm(Info.encodeVTYPE())
597           .addReg(RISCV::VL, RegState::Implicit);
598       return;
599     }
600     // Otherwise use an AVL of 0 to avoid depending on previous vl.
601     BuildMI(MBB, MI, DL, TII->get(RISCV::PseudoVSETIVLI))
602         .addReg(RISCV::X0, RegState::Define | RegState::Dead)
603         .addImm(0)
604         .addImm(Info.encodeVTYPE());
605     return;
606   }
607 
608   if (AVLReg.isVirtual())
609     MRI->constrainRegClass(AVLReg, &RISCV::GPRNoX0RegClass);
610 
611   // Use X0 as the DestReg unless AVLReg is X0. We also need to change the
612   // opcode if the AVLReg is X0 as they have different register classes for
613   // the AVL operand.
614   Register DestReg = RISCV::X0;
615   unsigned Opcode = RISCV::PseudoVSETVLI;
616   if (AVLReg == RISCV::X0) {
617     DestReg = MRI->createVirtualRegister(&RISCV::GPRRegClass);
618     Opcode = RISCV::PseudoVSETVLIX0;
619   }
620   BuildMI(MBB, MI, DL, TII->get(Opcode))
621       .addReg(DestReg, RegState::Define | RegState::Dead)
622       .addReg(AVLReg)
623       .addImm(Info.encodeVTYPE());
624 }
625 
626 // Return a VSETVLIInfo representing the changes made by this VSETVLI or
627 // VSETIVLI instruction.
628 static VSETVLIInfo getInfoForVSETVLI(const MachineInstr &MI) {
629   VSETVLIInfo NewInfo;
630   if (MI.getOpcode() == RISCV::PseudoVSETIVLI) {
631     NewInfo.setAVLImm(MI.getOperand(1).getImm());
632   } else {
633     assert(MI.getOpcode() == RISCV::PseudoVSETVLI ||
634            MI.getOpcode() == RISCV::PseudoVSETVLIX0);
635     Register AVLReg = MI.getOperand(1).getReg();
636     assert((AVLReg != RISCV::X0 || MI.getOperand(0).getReg() != RISCV::X0) &&
637            "Can't handle X0, X0 vsetvli yet");
638     NewInfo.setAVLReg(AVLReg);
639   }
640   NewInfo.setVTYPE(MI.getOperand(2).getImm());
641 
642   return NewInfo;
643 }
644 
645 bool RISCVInsertVSETVLI::needVSETVLI(const VSETVLIInfo &Require,
646                                      const VSETVLIInfo &CurInfo) {
647   if (CurInfo.isCompatible(Require, /*Strict*/ false))
648     return false;
649 
650   // We didn't find a compatible value. If our AVL is a virtual register,
651   // it might be defined by a VSET(I)VLI. If it has the same VTYPE we need
652   // and the last VL/VTYPE we observed is the same, we don't need a
653   // VSETVLI here.
654   if (!CurInfo.isUnknown() && Require.hasAVLReg() &&
655       Require.getAVLReg().isVirtual() && !CurInfo.hasSEWLMULRatioOnly() &&
656       CurInfo.hasCompatibleVTYPE(Require, /*Strict*/ false)) {
657     if (MachineInstr *DefMI = MRI->getVRegDef(Require.getAVLReg())) {
658       if (DefMI->getOpcode() == RISCV::PseudoVSETVLI ||
659           DefMI->getOpcode() == RISCV::PseudoVSETVLIX0 ||
660           DefMI->getOpcode() == RISCV::PseudoVSETIVLI) {
661         VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
662         if (DefInfo.hasSameAVL(CurInfo) && DefInfo.hasSameVTYPE(CurInfo))
663           return false;
664       }
665     }
666   }
667 
668   return true;
669 }
670 
671 bool canSkipVSETVLIForLoadStore(const MachineInstr &MI,
672                                 const VSETVLIInfo &Require,
673                                 const VSETVLIInfo &CurInfo) {
674   unsigned EEW;
675   switch (MI.getOpcode()) {
676   default:
677     return false;
678   case RISCV::PseudoVLE8_V_M1:
679   case RISCV::PseudoVLE8_V_M1_MASK:
680   case RISCV::PseudoVLE8_V_M2:
681   case RISCV::PseudoVLE8_V_M2_MASK:
682   case RISCV::PseudoVLE8_V_M4:
683   case RISCV::PseudoVLE8_V_M4_MASK:
684   case RISCV::PseudoVLE8_V_M8:
685   case RISCV::PseudoVLE8_V_M8_MASK:
686   case RISCV::PseudoVLE8_V_MF2:
687   case RISCV::PseudoVLE8_V_MF2_MASK:
688   case RISCV::PseudoVLE8_V_MF4:
689   case RISCV::PseudoVLE8_V_MF4_MASK:
690   case RISCV::PseudoVLE8_V_MF8:
691   case RISCV::PseudoVLE8_V_MF8_MASK:
692   case RISCV::PseudoVLSE8_V_M1:
693   case RISCV::PseudoVLSE8_V_M1_MASK:
694   case RISCV::PseudoVLSE8_V_M2:
695   case RISCV::PseudoVLSE8_V_M2_MASK:
696   case RISCV::PseudoVLSE8_V_M4:
697   case RISCV::PseudoVLSE8_V_M4_MASK:
698   case RISCV::PseudoVLSE8_V_M8:
699   case RISCV::PseudoVLSE8_V_M8_MASK:
700   case RISCV::PseudoVLSE8_V_MF2:
701   case RISCV::PseudoVLSE8_V_MF2_MASK:
702   case RISCV::PseudoVLSE8_V_MF4:
703   case RISCV::PseudoVLSE8_V_MF4_MASK:
704   case RISCV::PseudoVLSE8_V_MF8:
705   case RISCV::PseudoVLSE8_V_MF8_MASK:
706   case RISCV::PseudoVSE8_V_M1:
707   case RISCV::PseudoVSE8_V_M1_MASK:
708   case RISCV::PseudoVSE8_V_M2:
709   case RISCV::PseudoVSE8_V_M2_MASK:
710   case RISCV::PseudoVSE8_V_M4:
711   case RISCV::PseudoVSE8_V_M4_MASK:
712   case RISCV::PseudoVSE8_V_M8:
713   case RISCV::PseudoVSE8_V_M8_MASK:
714   case RISCV::PseudoVSE8_V_MF2:
715   case RISCV::PseudoVSE8_V_MF2_MASK:
716   case RISCV::PseudoVSE8_V_MF4:
717   case RISCV::PseudoVSE8_V_MF4_MASK:
718   case RISCV::PseudoVSE8_V_MF8:
719   case RISCV::PseudoVSE8_V_MF8_MASK:
720   case RISCV::PseudoVSSE8_V_M1:
721   case RISCV::PseudoVSSE8_V_M1_MASK:
722   case RISCV::PseudoVSSE8_V_M2:
723   case RISCV::PseudoVSSE8_V_M2_MASK:
724   case RISCV::PseudoVSSE8_V_M4:
725   case RISCV::PseudoVSSE8_V_M4_MASK:
726   case RISCV::PseudoVSSE8_V_M8:
727   case RISCV::PseudoVSSE8_V_M8_MASK:
728   case RISCV::PseudoVSSE8_V_MF2:
729   case RISCV::PseudoVSSE8_V_MF2_MASK:
730   case RISCV::PseudoVSSE8_V_MF4:
731   case RISCV::PseudoVSSE8_V_MF4_MASK:
732   case RISCV::PseudoVSSE8_V_MF8:
733   case RISCV::PseudoVSSE8_V_MF8_MASK:
734     EEW = 8;
735     break;
736   case RISCV::PseudoVLE16_V_M1:
737   case RISCV::PseudoVLE16_V_M1_MASK:
738   case RISCV::PseudoVLE16_V_M2:
739   case RISCV::PseudoVLE16_V_M2_MASK:
740   case RISCV::PseudoVLE16_V_M4:
741   case RISCV::PseudoVLE16_V_M4_MASK:
742   case RISCV::PseudoVLE16_V_M8:
743   case RISCV::PseudoVLE16_V_M8_MASK:
744   case RISCV::PseudoVLE16_V_MF2:
745   case RISCV::PseudoVLE16_V_MF2_MASK:
746   case RISCV::PseudoVLE16_V_MF4:
747   case RISCV::PseudoVLE16_V_MF4_MASK:
748   case RISCV::PseudoVLSE16_V_M1:
749   case RISCV::PseudoVLSE16_V_M1_MASK:
750   case RISCV::PseudoVLSE16_V_M2:
751   case RISCV::PseudoVLSE16_V_M2_MASK:
752   case RISCV::PseudoVLSE16_V_M4:
753   case RISCV::PseudoVLSE16_V_M4_MASK:
754   case RISCV::PseudoVLSE16_V_M8:
755   case RISCV::PseudoVLSE16_V_M8_MASK:
756   case RISCV::PseudoVLSE16_V_MF2:
757   case RISCV::PseudoVLSE16_V_MF2_MASK:
758   case RISCV::PseudoVLSE16_V_MF4:
759   case RISCV::PseudoVLSE16_V_MF4_MASK:
760   case RISCV::PseudoVSE16_V_M1:
761   case RISCV::PseudoVSE16_V_M1_MASK:
762   case RISCV::PseudoVSE16_V_M2:
763   case RISCV::PseudoVSE16_V_M2_MASK:
764   case RISCV::PseudoVSE16_V_M4:
765   case RISCV::PseudoVSE16_V_M4_MASK:
766   case RISCV::PseudoVSE16_V_M8:
767   case RISCV::PseudoVSE16_V_M8_MASK:
768   case RISCV::PseudoVSE16_V_MF2:
769   case RISCV::PseudoVSE16_V_MF2_MASK:
770   case RISCV::PseudoVSE16_V_MF4:
771   case RISCV::PseudoVSE16_V_MF4_MASK:
772   case RISCV::PseudoVSSE16_V_M1:
773   case RISCV::PseudoVSSE16_V_M1_MASK:
774   case RISCV::PseudoVSSE16_V_M2:
775   case RISCV::PseudoVSSE16_V_M2_MASK:
776   case RISCV::PseudoVSSE16_V_M4:
777   case RISCV::PseudoVSSE16_V_M4_MASK:
778   case RISCV::PseudoVSSE16_V_M8:
779   case RISCV::PseudoVSSE16_V_M8_MASK:
780   case RISCV::PseudoVSSE16_V_MF2:
781   case RISCV::PseudoVSSE16_V_MF2_MASK:
782   case RISCV::PseudoVSSE16_V_MF4:
783   case RISCV::PseudoVSSE16_V_MF4_MASK:
784     EEW = 16;
785     break;
786   case RISCV::PseudoVLE32_V_M1:
787   case RISCV::PseudoVLE32_V_M1_MASK:
788   case RISCV::PseudoVLE32_V_M2:
789   case RISCV::PseudoVLE32_V_M2_MASK:
790   case RISCV::PseudoVLE32_V_M4:
791   case RISCV::PseudoVLE32_V_M4_MASK:
792   case RISCV::PseudoVLE32_V_M8:
793   case RISCV::PseudoVLE32_V_M8_MASK:
794   case RISCV::PseudoVLE32_V_MF2:
795   case RISCV::PseudoVLE32_V_MF2_MASK:
796   case RISCV::PseudoVLSE32_V_M1:
797   case RISCV::PseudoVLSE32_V_M1_MASK:
798   case RISCV::PseudoVLSE32_V_M2:
799   case RISCV::PseudoVLSE32_V_M2_MASK:
800   case RISCV::PseudoVLSE32_V_M4:
801   case RISCV::PseudoVLSE32_V_M4_MASK:
802   case RISCV::PseudoVLSE32_V_M8:
803   case RISCV::PseudoVLSE32_V_M8_MASK:
804   case RISCV::PseudoVLSE32_V_MF2:
805   case RISCV::PseudoVLSE32_V_MF2_MASK:
806   case RISCV::PseudoVSE32_V_M1:
807   case RISCV::PseudoVSE32_V_M1_MASK:
808   case RISCV::PseudoVSE32_V_M2:
809   case RISCV::PseudoVSE32_V_M2_MASK:
810   case RISCV::PseudoVSE32_V_M4:
811   case RISCV::PseudoVSE32_V_M4_MASK:
812   case RISCV::PseudoVSE32_V_M8:
813   case RISCV::PseudoVSE32_V_M8_MASK:
814   case RISCV::PseudoVSE32_V_MF2:
815   case RISCV::PseudoVSE32_V_MF2_MASK:
816   case RISCV::PseudoVSSE32_V_M1:
817   case RISCV::PseudoVSSE32_V_M1_MASK:
818   case RISCV::PseudoVSSE32_V_M2:
819   case RISCV::PseudoVSSE32_V_M2_MASK:
820   case RISCV::PseudoVSSE32_V_M4:
821   case RISCV::PseudoVSSE32_V_M4_MASK:
822   case RISCV::PseudoVSSE32_V_M8:
823   case RISCV::PseudoVSSE32_V_M8_MASK:
824   case RISCV::PseudoVSSE32_V_MF2:
825   case RISCV::PseudoVSSE32_V_MF2_MASK:
826     EEW = 32;
827     break;
828   case RISCV::PseudoVLE64_V_M1:
829   case RISCV::PseudoVLE64_V_M1_MASK:
830   case RISCV::PseudoVLE64_V_M2:
831   case RISCV::PseudoVLE64_V_M2_MASK:
832   case RISCV::PseudoVLE64_V_M4:
833   case RISCV::PseudoVLE64_V_M4_MASK:
834   case RISCV::PseudoVLE64_V_M8:
835   case RISCV::PseudoVLE64_V_M8_MASK:
836   case RISCV::PseudoVLSE64_V_M1:
837   case RISCV::PseudoVLSE64_V_M1_MASK:
838   case RISCV::PseudoVLSE64_V_M2:
839   case RISCV::PseudoVLSE64_V_M2_MASK:
840   case RISCV::PseudoVLSE64_V_M4:
841   case RISCV::PseudoVLSE64_V_M4_MASK:
842   case RISCV::PseudoVLSE64_V_M8:
843   case RISCV::PseudoVLSE64_V_M8_MASK:
844   case RISCV::PseudoVSE64_V_M1:
845   case RISCV::PseudoVSE64_V_M1_MASK:
846   case RISCV::PseudoVSE64_V_M2:
847   case RISCV::PseudoVSE64_V_M2_MASK:
848   case RISCV::PseudoVSE64_V_M4:
849   case RISCV::PseudoVSE64_V_M4_MASK:
850   case RISCV::PseudoVSE64_V_M8:
851   case RISCV::PseudoVSE64_V_M8_MASK:
852   case RISCV::PseudoVSSE64_V_M1:
853   case RISCV::PseudoVSSE64_V_M1_MASK:
854   case RISCV::PseudoVSSE64_V_M2:
855   case RISCV::PseudoVSSE64_V_M2_MASK:
856   case RISCV::PseudoVSSE64_V_M4:
857   case RISCV::PseudoVSSE64_V_M4_MASK:
858   case RISCV::PseudoVSSE64_V_M8:
859   case RISCV::PseudoVSSE64_V_M8_MASK:
860     EEW = 64;
861     break;
862   }
863 
864   return CurInfo.isCompatibleWithLoadStoreEEW(EEW, Require);
865 }
866 
867 bool RISCVInsertVSETVLI::computeVLVTYPEChanges(const MachineBasicBlock &MBB) {
868   bool HadVectorOp = false;
869 
870   BlockData &BBInfo = BlockInfo[MBB.getNumber()];
871   for (const MachineInstr &MI : MBB) {
872     // If this is an explicit VSETVLI or VSETIVLI, update our state.
873     if (MI.getOpcode() == RISCV::PseudoVSETVLI ||
874         MI.getOpcode() == RISCV::PseudoVSETVLIX0 ||
875         MI.getOpcode() == RISCV::PseudoVSETIVLI) {
876       HadVectorOp = true;
877       BBInfo.Change = getInfoForVSETVLI(MI);
878       continue;
879     }
880 
881     uint64_t TSFlags = MI.getDesc().TSFlags;
882     if (RISCVII::hasSEWOp(TSFlags)) {
883       HadVectorOp = true;
884 
885       VSETVLIInfo NewInfo = computeInfoForInstr(MI, TSFlags, MRI);
886 
887       if (!BBInfo.Change.isValid()) {
888         BBInfo.Change = NewInfo;
889       } else {
890         // If this instruction isn't compatible with the previous VL/VTYPE
891         // we need to insert a VSETVLI.
892         // If this is a unit-stride or strided load/store, we may be able to use
893         // the EMUL=(EEW/SEW)*LMUL relationship to avoid changing vtype.
894         // NOTE: We only do this if the vtype we're comparing against was
895         // created in this block. We need the first and third phase to treat
896         // the store the same way.
897         if (!canSkipVSETVLIForLoadStore(MI, NewInfo, BBInfo.Change) &&
898             needVSETVLI(NewInfo, BBInfo.Change))
899           BBInfo.Change = NewInfo;
900       }
901     }
902 
903     // If this is something that updates VL/VTYPE that we don't know about, set
904     // the state to unknown.
905     if (MI.isCall() || MI.isInlineAsm() || MI.modifiesRegister(RISCV::VL) ||
906         MI.modifiesRegister(RISCV::VTYPE)) {
907       BBInfo.Change = VSETVLIInfo::getUnknown();
908     }
909   }
910 
911   // Initial exit state is whatever change we found in the block.
912   BBInfo.Exit = BBInfo.Change;
913 
914   return HadVectorOp;
915 }
916 
917 void RISCVInsertVSETVLI::computeIncomingVLVTYPE(const MachineBasicBlock &MBB) {
918   BlockData &BBInfo = BlockInfo[MBB.getNumber()];
919 
920   BBInfo.InQueue = false;
921 
922   VSETVLIInfo InInfo;
923   if (MBB.pred_empty()) {
924     // There are no predecessors, so use the default starting status.
925     InInfo.setUnknown();
926   } else {
927     for (MachineBasicBlock *P : MBB.predecessors())
928       InInfo = InInfo.intersect(BlockInfo[P->getNumber()].Exit);
929   }
930 
931   // If we don't have any valid predecessor value, wait until we do.
932   if (!InInfo.isValid())
933     return;
934 
935   BBInfo.Pred = InInfo;
936 
937   VSETVLIInfo TmpStatus = BBInfo.Pred.merge(BBInfo.Change);
938 
939   // If the new exit value matches the old exit value, we don't need to revisit
940   // any blocks.
941   if (BBInfo.Exit == TmpStatus)
942     return;
943 
944   BBInfo.Exit = TmpStatus;
945 
946   // Add the successors to the work list so we can propagate the changed exit
947   // status.
948   for (MachineBasicBlock *S : MBB.successors())
949     if (!BlockInfo[S->getNumber()].InQueue)
950       WorkList.push(S);
951 }
952 
953 // If we weren't able to prove a vsetvli was directly unneeded, it might still
954 // be/ unneeded if the AVL is a phi node where all incoming values are VL
955 // outputs from the last VSETVLI in their respective basic blocks.
956 bool RISCVInsertVSETVLI::needVSETVLIPHI(const VSETVLIInfo &Require,
957                                         const MachineBasicBlock &MBB) {
958   if (DisableInsertVSETVLPHIOpt)
959     return true;
960 
961   if (!Require.hasAVLReg())
962     return true;
963 
964   Register AVLReg = Require.getAVLReg();
965   if (!AVLReg.isVirtual())
966     return true;
967 
968   // We need the AVL to be produce by a PHI node in this basic block.
969   MachineInstr *PHI = MRI->getVRegDef(AVLReg);
970   if (!PHI || PHI->getOpcode() != RISCV::PHI || PHI->getParent() != &MBB)
971     return true;
972 
973   for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps;
974        PHIOp += 2) {
975     Register InReg = PHI->getOperand(PHIOp).getReg();
976     MachineBasicBlock *PBB = PHI->getOperand(PHIOp + 1).getMBB();
977     const BlockData &PBBInfo = BlockInfo[PBB->getNumber()];
978     // If the exit from the predecessor has the VTYPE we are looking for
979     // we might be able to avoid a VSETVLI.
980     if (PBBInfo.Exit.isUnknown() ||
981         !PBBInfo.Exit.hasCompatibleVTYPE(Require, /*Strict*/ false))
982       return true;
983 
984     // We need the PHI input to the be the output of a VSET(I)VLI.
985     MachineInstr *DefMI = MRI->getVRegDef(InReg);
986     if (!DefMI || (DefMI->getOpcode() != RISCV::PseudoVSETVLI &&
987                    DefMI->getOpcode() != RISCV::PseudoVSETVLIX0 &&
988                    DefMI->getOpcode() != RISCV::PseudoVSETIVLI))
989       return true;
990 
991     // We found a VSET(I)VLI make sure it matches the output of the
992     // predecessor block.
993     VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
994     if (!DefInfo.hasSameAVL(PBBInfo.Exit) ||
995         !DefInfo.hasSameVTYPE(PBBInfo.Exit))
996       return true;
997   }
998 
999   // If all the incoming values to the PHI checked out, we don't need
1000   // to insert a VSETVLI.
1001   return false;
1002 }
1003 
1004 void RISCVInsertVSETVLI::emitVSETVLIs(MachineBasicBlock &MBB) {
1005   VSETVLIInfo CurInfo;
1006   // Only be set if current VSETVLIInfo is from an explicit VSET(I)VLI.
1007   MachineInstr *PrevVSETVLIMI = nullptr;
1008 
1009   for (MachineInstr &MI : MBB) {
1010     // If this is an explicit VSETVLI or VSETIVLI, update our state.
1011     if (MI.getOpcode() == RISCV::PseudoVSETVLI ||
1012         MI.getOpcode() == RISCV::PseudoVSETVLIX0 ||
1013         MI.getOpcode() == RISCV::PseudoVSETIVLI) {
1014       // Conservatively, mark the VL and VTYPE as live.
1015       assert(MI.getOperand(3).getReg() == RISCV::VL &&
1016              MI.getOperand(4).getReg() == RISCV::VTYPE &&
1017              "Unexpected operands where VL and VTYPE should be");
1018       MI.getOperand(3).setIsDead(false);
1019       MI.getOperand(4).setIsDead(false);
1020       CurInfo = getInfoForVSETVLI(MI);
1021       PrevVSETVLIMI = &MI;
1022       continue;
1023     }
1024 
1025     uint64_t TSFlags = MI.getDesc().TSFlags;
1026     if (RISCVII::hasSEWOp(TSFlags)) {
1027       VSETVLIInfo NewInfo = computeInfoForInstr(MI, TSFlags, MRI);
1028       if (RISCVII::hasVLOp(TSFlags)) {
1029         unsigned Offset = 2;
1030         if (RISCVII::hasVecPolicyOp(TSFlags))
1031           Offset = 3;
1032         MachineOperand &VLOp =
1033             MI.getOperand(MI.getNumExplicitOperands() - Offset);
1034         if (VLOp.isReg()) {
1035           // Erase the AVL operand from the instruction.
1036           VLOp.setReg(RISCV::NoRegister);
1037           VLOp.setIsKill(false);
1038         }
1039         MI.addOperand(MachineOperand::CreateReg(RISCV::VL, /*isDef*/ false,
1040                                                 /*isImp*/ true));
1041       }
1042       MI.addOperand(MachineOperand::CreateReg(RISCV::VTYPE, /*isDef*/ false,
1043                                               /*isImp*/ true));
1044 
1045       if (!CurInfo.isValid()) {
1046         // We haven't found any vector instructions or VL/VTYPE changes yet,
1047         // use the predecessor information.
1048         assert(BlockInfo[MBB.getNumber()].Pred.isValid() &&
1049                "Expected a valid predecessor state.");
1050         if (needVSETVLI(NewInfo, BlockInfo[MBB.getNumber()].Pred) &&
1051             needVSETVLIPHI(NewInfo, MBB)) {
1052           insertVSETVLI(MBB, MI, NewInfo, BlockInfo[MBB.getNumber()].Pred);
1053           CurInfo = NewInfo;
1054         }
1055       } else {
1056         // If this instruction isn't compatible with the previous VL/VTYPE
1057         // we need to insert a VSETVLI.
1058         // If this is a unit-stride or strided load/store, we may be able to use
1059         // the EMUL=(EEW/SEW)*LMUL relationship to avoid changing vtype.
1060         // NOTE: We can't use predecessor information for the store. We must
1061         // treat it the same as the first phase so that we produce the correct
1062         // vl/vtype for succesor blocks.
1063         if (!canSkipVSETVLIForLoadStore(MI, NewInfo, CurInfo) &&
1064             needVSETVLI(NewInfo, CurInfo)) {
1065           // If the previous VL/VTYPE is set by VSETVLI and do not use, Merge it
1066           // with current VL/VTYPE.
1067           bool NeedInsertVSETVLI = true;
1068           if (PrevVSETVLIMI) {
1069             bool HasSameAVL =
1070                 CurInfo.hasSameAVL(NewInfo) ||
1071                 (NewInfo.hasAVLReg() && NewInfo.getAVLReg().isVirtual() &&
1072                  NewInfo.getAVLReg() == PrevVSETVLIMI->getOperand(0).getReg());
1073             // If these two VSETVLI have the same AVL and the same VLMAX,
1074             // we could merge these two VSETVLI.
1075             if (HasSameAVL &&
1076                 CurInfo.getSEWLMULRatio() == NewInfo.getSEWLMULRatio()) {
1077               PrevVSETVLIMI->getOperand(2).setImm(NewInfo.encodeVTYPE());
1078               NeedInsertVSETVLI = false;
1079             }
1080             if (isScalarMoveInstr(MI) &&
1081                 ((CurInfo.hasNonZeroAVL() && NewInfo.hasNonZeroAVL()) ||
1082                  (CurInfo.hasZeroAVL() && NewInfo.hasZeroAVL())) &&
1083                 NewInfo.hasSameVLMAX(CurInfo)) {
1084               PrevVSETVLIMI->getOperand(2).setImm(NewInfo.encodeVTYPE());
1085               NeedInsertVSETVLI = false;
1086             }
1087           }
1088           if (NeedInsertVSETVLI)
1089             insertVSETVLI(MBB, MI, NewInfo, CurInfo);
1090           CurInfo = NewInfo;
1091         }
1092       }
1093       PrevVSETVLIMI = nullptr;
1094     }
1095 
1096     // If this is something updates VL/VTYPE that we don't know about, set
1097     // the state to unknown.
1098     if (MI.isCall() || MI.isInlineAsm() || MI.modifiesRegister(RISCV::VL) ||
1099         MI.modifiesRegister(RISCV::VTYPE)) {
1100       CurInfo = VSETVLIInfo::getUnknown();
1101       PrevVSETVLIMI = nullptr;
1102     }
1103 
1104     // If we reach the end of the block and our current info doesn't match the
1105     // expected info, insert a vsetvli to correct.
1106     if (MI.isTerminator()) {
1107       const VSETVLIInfo &ExitInfo = BlockInfo[MBB.getNumber()].Exit;
1108       if (CurInfo.isValid() && ExitInfo.isValid() && !ExitInfo.isUnknown() &&
1109           CurInfo != ExitInfo) {
1110         insertVSETVLI(MBB, MI, ExitInfo, CurInfo);
1111         CurInfo = ExitInfo;
1112       }
1113     }
1114   }
1115 }
1116 
1117 bool RISCVInsertVSETVLI::runOnMachineFunction(MachineFunction &MF) {
1118   // Skip if the vector extension is not enabled.
1119   const RISCVSubtarget &ST = MF.getSubtarget<RISCVSubtarget>();
1120   if (!ST.hasVInstructions())
1121     return false;
1122 
1123   TII = ST.getInstrInfo();
1124   MRI = &MF.getRegInfo();
1125 
1126   assert(BlockInfo.empty() && "Expect empty block infos");
1127   BlockInfo.resize(MF.getNumBlockIDs());
1128 
1129   bool HaveVectorOp = false;
1130 
1131   // Phase 1 - determine how VL/VTYPE are affected by the each block.
1132   for (const MachineBasicBlock &MBB : MF)
1133     HaveVectorOp |= computeVLVTYPEChanges(MBB);
1134 
1135   // If we didn't find any instructions that need VSETVLI, we're done.
1136   if (HaveVectorOp) {
1137     // Phase 2 - determine the exit VL/VTYPE from each block. We add all
1138     // blocks to the list here, but will also add any that need to be revisited
1139     // during Phase 2 processing.
1140     for (const MachineBasicBlock &MBB : MF) {
1141       WorkList.push(&MBB);
1142       BlockInfo[MBB.getNumber()].InQueue = true;
1143     }
1144     while (!WorkList.empty()) {
1145       const MachineBasicBlock &MBB = *WorkList.front();
1146       WorkList.pop();
1147       computeIncomingVLVTYPE(MBB);
1148     }
1149 
1150     // Phase 3 - add any vsetvli instructions needed in the block. Use the
1151     // Phase 2 information to avoid adding vsetvlis before the first vector
1152     // instruction in the block if the VL/VTYPE is satisfied by its
1153     // predecessors.
1154     for (MachineBasicBlock &MBB : MF)
1155       emitVSETVLIs(MBB);
1156   }
1157 
1158   BlockInfo.clear();
1159 
1160   return HaveVectorOp;
1161 }
1162 
1163 /// Returns an instance of the Insert VSETVLI pass.
1164 FunctionPass *llvm::createRISCVInsertVSETVLIPass() {
1165   return new RISCVInsertVSETVLI();
1166 }
1167