[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / Target / RISCV / RISCVInsertVSETVLI.cpp
bloba14f9a283547370ed26fdfc8334a50f9c8912592
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 and expands the vl outputs of VLEFF/VLSEGFF to PseudoReadVL
11 // instructions.
13 // This pass consists of 3 phases:
15 // Phase 1 collects how each basic block affects VL/VTYPE.
17 // Phase 2 uses the information from phase 1 to do a data flow analysis to
18 // propagate the VL/VTYPE changes through the function. This gives us the
19 // VL/VTYPE at the start of each basic block.
21 // Phase 3 inserts VSETVLI instructions in each basic block. Information from
22 // phase 2 is used to prevent inserting a VSETVLI before the first vector
23 // instruction in the block if possible.
25 //===----------------------------------------------------------------------===//
27 #include "RISCV.h"
28 #include "RISCVSubtarget.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/CodeGen/LiveIntervals.h"
31 #include "llvm/CodeGen/MachineFunctionPass.h"
32 #include <queue>
33 using namespace llvm;
35 #define DEBUG_TYPE "riscv-insert-vsetvli"
36 #define RISCV_INSERT_VSETVLI_NAME "RISC-V Insert VSETVLI pass"
38 STATISTIC(NumInsertedVSETVL, "Number of VSETVL inst inserted");
39 STATISTIC(NumRemovedVSETVL, "Number of VSETVL inst removed");
41 static cl::opt<bool> DisableInsertVSETVLPHIOpt(
42 "riscv-disable-insert-vsetvl-phi-opt", cl::init(false), cl::Hidden,
43 cl::desc("Disable looking through phis when inserting vsetvlis."));
45 static cl::opt<bool> UseStrictAsserts(
46 "riscv-insert-vsetvl-strict-asserts", cl::init(true), cl::Hidden,
47 cl::desc("Enable strict assertion checking for the dataflow algorithm"));
49 namespace {
51 static unsigned getVLOpNum(const MachineInstr &MI) {
52 return RISCVII::getVLOpNum(MI.getDesc());
55 static unsigned getSEWOpNum(const MachineInstr &MI) {
56 return RISCVII::getSEWOpNum(MI.getDesc());
59 static bool isVectorConfigInstr(const MachineInstr &MI) {
60 return MI.getOpcode() == RISCV::PseudoVSETVLI ||
61 MI.getOpcode() == RISCV::PseudoVSETVLIX0 ||
62 MI.getOpcode() == RISCV::PseudoVSETIVLI;
65 /// Return true if this is 'vsetvli x0, x0, vtype' which preserves
66 /// VL and only sets VTYPE.
67 static bool isVLPreservingConfig(const MachineInstr &MI) {
68 if (MI.getOpcode() != RISCV::PseudoVSETVLIX0)
69 return false;
70 assert(RISCV::X0 == MI.getOperand(1).getReg());
71 return RISCV::X0 == MI.getOperand(0).getReg();
74 static bool isFloatScalarMoveOrScalarSplatInstr(const MachineInstr &MI) {
75 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
76 default:
77 return false;
78 case RISCV::VFMV_S_F:
79 case RISCV::VFMV_V_F:
80 return true;
84 static bool isScalarExtractInstr(const MachineInstr &MI) {
85 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
86 default:
87 return false;
88 case RISCV::VMV_X_S:
89 case RISCV::VFMV_F_S:
90 return true;
94 static bool isScalarInsertInstr(const MachineInstr &MI) {
95 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
96 default:
97 return false;
98 case RISCV::VMV_S_X:
99 case RISCV::VFMV_S_F:
100 return true;
104 static bool isScalarSplatInstr(const MachineInstr &MI) {
105 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
106 default:
107 return false;
108 case RISCV::VMV_V_I:
109 case RISCV::VMV_V_X:
110 case RISCV::VFMV_V_F:
111 return true;
115 static bool isVSlideInstr(const MachineInstr &MI) {
116 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
117 default:
118 return false;
119 case RISCV::VSLIDEDOWN_VX:
120 case RISCV::VSLIDEDOWN_VI:
121 case RISCV::VSLIDEUP_VX:
122 case RISCV::VSLIDEUP_VI:
123 return true;
127 /// Get the EEW for a load or store instruction. Return std::nullopt if MI is
128 /// not a load or store which ignores SEW.
129 static std::optional<unsigned> getEEWForLoadStore(const MachineInstr &MI) {
130 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
131 default:
132 return std::nullopt;
133 case RISCV::VLE8_V:
134 case RISCV::VLSE8_V:
135 case RISCV::VSE8_V:
136 case RISCV::VSSE8_V:
137 return 8;
138 case RISCV::VLE16_V:
139 case RISCV::VLSE16_V:
140 case RISCV::VSE16_V:
141 case RISCV::VSSE16_V:
142 return 16;
143 case RISCV::VLE32_V:
144 case RISCV::VLSE32_V:
145 case RISCV::VSE32_V:
146 case RISCV::VSSE32_V:
147 return 32;
148 case RISCV::VLE64_V:
149 case RISCV::VLSE64_V:
150 case RISCV::VSE64_V:
151 case RISCV::VSSE64_V:
152 return 64;
156 static bool isNonZeroLoadImmediate(MachineInstr &MI) {
157 return MI.getOpcode() == RISCV::ADDI &&
158 MI.getOperand(1).isReg() && MI.getOperand(2).isImm() &&
159 MI.getOperand(1).getReg() == RISCV::X0 &&
160 MI.getOperand(2).getImm() != 0;
163 /// Return true if this is an operation on mask registers. Note that
164 /// this includes both arithmetic/logical ops and load/store (vlm/vsm).
165 static bool isMaskRegOp(const MachineInstr &MI) {
166 if (!RISCVII::hasSEWOp(MI.getDesc().TSFlags))
167 return false;
168 const unsigned Log2SEW = MI.getOperand(getSEWOpNum(MI)).getImm();
169 // A Log2SEW of 0 is an operation on mask registers only.
170 return Log2SEW == 0;
173 /// Return true if the inactive elements in the result are entirely undefined.
174 /// Note that this is different from "agnostic" as defined by the vector
175 /// specification. Agnostic requires each lane to either be undisturbed, or
176 /// take the value -1; no other value is allowed.
177 static bool hasUndefinedMergeOp(const MachineInstr &MI,
178 const MachineRegisterInfo &MRI) {
180 unsigned UseOpIdx;
181 if (!MI.isRegTiedToUseOperand(0, &UseOpIdx))
182 // If there is no passthrough operand, then the pass through
183 // lanes are undefined.
184 return true;
186 // If the tied operand is NoReg, an IMPLICIT_DEF, or a REG_SEQEUENCE whose
187 // operands are solely IMPLICIT_DEFS, then the pass through lanes are
188 // undefined.
189 const MachineOperand &UseMO = MI.getOperand(UseOpIdx);
190 if (UseMO.getReg() == RISCV::NoRegister)
191 return true;
193 if (MachineInstr *UseMI = MRI.getVRegDef(UseMO.getReg())) {
194 if (UseMI->isImplicitDef())
195 return true;
197 if (UseMI->isRegSequence()) {
198 for (unsigned i = 1, e = UseMI->getNumOperands(); i < e; i += 2) {
199 MachineInstr *SourceMI = MRI.getVRegDef(UseMI->getOperand(i).getReg());
200 if (!SourceMI || !SourceMI->isImplicitDef())
201 return false;
203 return true;
206 return false;
209 /// Which subfields of VL or VTYPE have values we need to preserve?
210 struct DemandedFields {
211 // Some unknown property of VL is used. If demanded, must preserve entire
212 // value.
213 bool VLAny = false;
214 // Only zero vs non-zero is used. If demanded, can change non-zero values.
215 bool VLZeroness = false;
216 // What properties of SEW we need to preserve.
217 enum : uint8_t {
218 SEWEqual = 3, // The exact value of SEW needs to be preserved.
219 SEWGreaterThanOrEqual = 2, // SEW can be changed as long as it's greater
220 // than or equal to the original value.
221 SEWGreaterThanOrEqualAndLessThan64 =
222 1, // SEW can be changed as long as it's greater
223 // than or equal to the original value, but must be less
224 // than 64.
225 SEWNone = 0 // We don't need to preserve SEW at all.
226 } SEW = SEWNone;
227 bool LMUL = false;
228 bool SEWLMULRatio = false;
229 bool TailPolicy = false;
230 bool MaskPolicy = false;
232 // Return true if any part of VTYPE was used
233 bool usedVTYPE() const {
234 return SEW || LMUL || SEWLMULRatio || TailPolicy || MaskPolicy;
237 // Return true if any property of VL was used
238 bool usedVL() {
239 return VLAny || VLZeroness;
242 // Mark all VTYPE subfields and properties as demanded
243 void demandVTYPE() {
244 SEW = SEWEqual;
245 LMUL = true;
246 SEWLMULRatio = true;
247 TailPolicy = true;
248 MaskPolicy = true;
251 // Mark all VL properties as demanded
252 void demandVL() {
253 VLAny = true;
254 VLZeroness = true;
257 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
258 /// Support for debugging, callable in GDB: V->dump()
259 LLVM_DUMP_METHOD void dump() const {
260 print(dbgs());
261 dbgs() << "\n";
264 /// Implement operator<<.
265 void print(raw_ostream &OS) const {
266 OS << "{";
267 OS << "VLAny=" << VLAny << ", ";
268 OS << "VLZeroness=" << VLZeroness << ", ";
269 OS << "SEW=";
270 switch (SEW) {
271 case SEWEqual:
272 OS << "SEWEqual";
273 break;
274 case SEWGreaterThanOrEqual:
275 OS << "SEWGreaterThanOrEqual";
276 break;
277 case SEWGreaterThanOrEqualAndLessThan64:
278 OS << "SEWGreaterThanOrEqualAndLessThan64";
279 break;
280 case SEWNone:
281 OS << "SEWNone";
282 break;
284 OS << ", ";
285 OS << "LMUL=" << LMUL << ", ";
286 OS << "SEWLMULRatio=" << SEWLMULRatio << ", ";
287 OS << "TailPolicy=" << TailPolicy << ", ";
288 OS << "MaskPolicy=" << MaskPolicy;
289 OS << "}";
291 #endif
294 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
295 LLVM_ATTRIBUTE_USED
296 inline raw_ostream &operator<<(raw_ostream &OS, const DemandedFields &DF) {
297 DF.print(OS);
298 return OS;
300 #endif
302 /// Return true if moving from CurVType to NewVType is
303 /// indistinguishable from the perspective of an instruction (or set
304 /// of instructions) which use only the Used subfields and properties.
305 static bool areCompatibleVTYPEs(uint64_t CurVType, uint64_t NewVType,
306 const DemandedFields &Used) {
307 switch (Used.SEW) {
308 case DemandedFields::SEWNone:
309 break;
310 case DemandedFields::SEWEqual:
311 if (RISCVVType::getSEW(CurVType) != RISCVVType::getSEW(NewVType))
312 return false;
313 break;
314 case DemandedFields::SEWGreaterThanOrEqual:
315 if (RISCVVType::getSEW(NewVType) < RISCVVType::getSEW(CurVType))
316 return false;
317 break;
318 case DemandedFields::SEWGreaterThanOrEqualAndLessThan64:
319 if (RISCVVType::getSEW(NewVType) < RISCVVType::getSEW(CurVType) ||
320 RISCVVType::getSEW(NewVType) >= 64)
321 return false;
322 break;
325 if (Used.LMUL &&
326 RISCVVType::getVLMUL(CurVType) != RISCVVType::getVLMUL(NewVType))
327 return false;
329 if (Used.SEWLMULRatio) {
330 auto Ratio1 = RISCVVType::getSEWLMULRatio(RISCVVType::getSEW(CurVType),
331 RISCVVType::getVLMUL(CurVType));
332 auto Ratio2 = RISCVVType::getSEWLMULRatio(RISCVVType::getSEW(NewVType),
333 RISCVVType::getVLMUL(NewVType));
334 if (Ratio1 != Ratio2)
335 return false;
338 if (Used.TailPolicy && RISCVVType::isTailAgnostic(CurVType) !=
339 RISCVVType::isTailAgnostic(NewVType))
340 return false;
341 if (Used.MaskPolicy && RISCVVType::isMaskAgnostic(CurVType) !=
342 RISCVVType::isMaskAgnostic(NewVType))
343 return false;
344 return true;
347 /// Return the fields and properties demanded by the provided instruction.
348 DemandedFields getDemanded(const MachineInstr &MI,
349 const MachineRegisterInfo *MRI,
350 const RISCVSubtarget *ST) {
351 // Warning: This function has to work on both the lowered (i.e. post
352 // emitVSETVLIs) and pre-lowering forms. The main implication of this is
353 // that it can't use the value of a SEW, VL, or Policy operand as they might
354 // be stale after lowering.
356 // Most instructions don't use any of these subfeilds.
357 DemandedFields Res;
358 // Start conservative if registers are used
359 if (MI.isCall() || MI.isInlineAsm() || MI.readsRegister(RISCV::VL))
360 Res.demandVL();
361 if (MI.isCall() || MI.isInlineAsm() || MI.readsRegister(RISCV::VTYPE))
362 Res.demandVTYPE();
363 // Start conservative on the unlowered form too
364 uint64_t TSFlags = MI.getDesc().TSFlags;
365 if (RISCVII::hasSEWOp(TSFlags)) {
366 Res.demandVTYPE();
367 if (RISCVII::hasVLOp(TSFlags))
368 Res.demandVL();
370 // Behavior is independent of mask policy.
371 if (!RISCVII::usesMaskPolicy(TSFlags))
372 Res.MaskPolicy = false;
375 // Loads and stores with implicit EEW do not demand SEW or LMUL directly.
376 // They instead demand the ratio of the two which is used in computing
377 // EMUL, but which allows us the flexibility to change SEW and LMUL
378 // provided we don't change the ratio.
379 // Note: We assume that the instructions initial SEW is the EEW encoded
380 // in the opcode. This is asserted when constructing the VSETVLIInfo.
381 if (getEEWForLoadStore(MI)) {
382 Res.SEW = DemandedFields::SEWNone;
383 Res.LMUL = false;
386 // Store instructions don't use the policy fields.
387 if (RISCVII::hasSEWOp(TSFlags) && MI.getNumExplicitDefs() == 0) {
388 Res.TailPolicy = false;
389 Res.MaskPolicy = false;
392 // If this is a mask reg operation, it only cares about VLMAX.
393 // TODO: Possible extensions to this logic
394 // * Probably ok if available VLMax is larger than demanded
395 // * The policy bits can probably be ignored..
396 if (isMaskRegOp(MI)) {
397 Res.SEW = DemandedFields::SEWNone;
398 Res.LMUL = false;
401 // For vmv.s.x and vfmv.s.f, there are only two behaviors, VL = 0 and VL > 0.
402 if (isScalarInsertInstr(MI)) {
403 Res.LMUL = false;
404 Res.SEWLMULRatio = false;
405 Res.VLAny = false;
406 // For vmv.s.x and vfmv.s.f, if the merge operand is *undefined*, we don't
407 // need to preserve any other bits and are thus compatible with any larger,
408 // etype and can disregard policy bits. Warning: It's tempting to try doing
409 // this for any tail agnostic operation, but we can't as TA requires
410 // tail lanes to either be the original value or -1. We are writing
411 // unknown bits to the lanes here.
412 if (hasUndefinedMergeOp(MI, *MRI)) {
413 if (isFloatScalarMoveOrScalarSplatInstr(MI) && !ST->hasVInstructionsF64())
414 Res.SEW = DemandedFields::SEWGreaterThanOrEqualAndLessThan64;
415 else
416 Res.SEW = DemandedFields::SEWGreaterThanOrEqual;
417 Res.TailPolicy = false;
421 // vmv.x.s, and vmv.f.s are unconditional and ignore everything except SEW.
422 if (isScalarExtractInstr(MI)) {
423 assert(!RISCVII::hasVLOp(TSFlags));
424 Res.LMUL = false;
425 Res.SEWLMULRatio = false;
426 Res.TailPolicy = false;
427 Res.MaskPolicy = false;
430 return Res;
433 /// Defines the abstract state with which the forward dataflow models the
434 /// values of the VL and VTYPE registers after insertion.
435 class VSETVLIInfo {
436 union {
437 Register AVLReg;
438 unsigned AVLImm;
441 enum : uint8_t {
442 Uninitialized,
443 AVLIsReg,
444 AVLIsImm,
445 Unknown,
446 } State = Uninitialized;
448 // Fields from VTYPE.
449 RISCVII::VLMUL VLMul = RISCVII::LMUL_1;
450 uint8_t SEW = 0;
451 uint8_t TailAgnostic : 1;
452 uint8_t MaskAgnostic : 1;
453 uint8_t SEWLMULRatioOnly : 1;
455 public:
456 VSETVLIInfo()
457 : AVLImm(0), TailAgnostic(false), MaskAgnostic(false),
458 SEWLMULRatioOnly(false) {}
460 static VSETVLIInfo getUnknown() {
461 VSETVLIInfo Info;
462 Info.setUnknown();
463 return Info;
466 bool isValid() const { return State != Uninitialized; }
467 void setUnknown() { State = Unknown; }
468 bool isUnknown() const { return State == Unknown; }
470 void setAVLReg(Register Reg) {
471 AVLReg = Reg;
472 State = AVLIsReg;
475 void setAVLImm(unsigned Imm) {
476 AVLImm = Imm;
477 State = AVLIsImm;
480 bool hasAVLImm() const { return State == AVLIsImm; }
481 bool hasAVLReg() const { return State == AVLIsReg; }
482 Register getAVLReg() const {
483 assert(hasAVLReg());
484 return AVLReg;
486 unsigned getAVLImm() const {
487 assert(hasAVLImm());
488 return AVLImm;
491 void setAVL(VSETVLIInfo Info) {
492 assert(Info.isValid());
493 if (Info.isUnknown())
494 setUnknown();
495 else if (Info.hasAVLReg())
496 setAVLReg(Info.getAVLReg());
497 else {
498 assert(Info.hasAVLImm());
499 setAVLImm(Info.getAVLImm());
503 unsigned getSEW() const { return SEW; }
504 RISCVII::VLMUL getVLMUL() const { return VLMul; }
505 bool getTailAgnostic() const { return TailAgnostic; }
506 bool getMaskAgnostic() const { return MaskAgnostic; }
508 bool hasNonZeroAVL(const MachineRegisterInfo &MRI) const {
509 if (hasAVLImm())
510 return getAVLImm() > 0;
511 if (hasAVLReg()) {
512 if (getAVLReg() == RISCV::X0)
513 return true;
514 if (MachineInstr *MI = MRI.getVRegDef(getAVLReg());
515 MI && isNonZeroLoadImmediate(*MI))
516 return true;
517 return false;
519 return false;
522 bool hasEquallyZeroAVL(const VSETVLIInfo &Other,
523 const MachineRegisterInfo &MRI) const {
524 if (hasSameAVL(Other))
525 return true;
526 return (hasNonZeroAVL(MRI) && Other.hasNonZeroAVL(MRI));
529 bool hasSameAVL(const VSETVLIInfo &Other) const {
530 if (hasAVLReg() && Other.hasAVLReg())
531 return getAVLReg() == Other.getAVLReg();
533 if (hasAVLImm() && Other.hasAVLImm())
534 return getAVLImm() == Other.getAVLImm();
536 return false;
539 void setVTYPE(unsigned VType) {
540 assert(isValid() && !isUnknown() &&
541 "Can't set VTYPE for uninitialized or unknown");
542 VLMul = RISCVVType::getVLMUL(VType);
543 SEW = RISCVVType::getSEW(VType);
544 TailAgnostic = RISCVVType::isTailAgnostic(VType);
545 MaskAgnostic = RISCVVType::isMaskAgnostic(VType);
547 void setVTYPE(RISCVII::VLMUL L, unsigned S, bool TA, bool MA) {
548 assert(isValid() && !isUnknown() &&
549 "Can't set VTYPE for uninitialized or unknown");
550 VLMul = L;
551 SEW = S;
552 TailAgnostic = TA;
553 MaskAgnostic = MA;
556 void setVLMul(RISCVII::VLMUL VLMul) { this->VLMul = VLMul; }
558 unsigned encodeVTYPE() const {
559 assert(isValid() && !isUnknown() && !SEWLMULRatioOnly &&
560 "Can't encode VTYPE for uninitialized or unknown");
561 return RISCVVType::encodeVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
564 bool hasSEWLMULRatioOnly() const { return SEWLMULRatioOnly; }
566 bool hasSameVTYPE(const VSETVLIInfo &Other) const {
567 assert(isValid() && Other.isValid() &&
568 "Can't compare invalid VSETVLIInfos");
569 assert(!isUnknown() && !Other.isUnknown() &&
570 "Can't compare VTYPE in unknown state");
571 assert(!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly &&
572 "Can't compare when only LMUL/SEW ratio is valid.");
573 return std::tie(VLMul, SEW, TailAgnostic, MaskAgnostic) ==
574 std::tie(Other.VLMul, Other.SEW, Other.TailAgnostic,
575 Other.MaskAgnostic);
578 unsigned getSEWLMULRatio() const {
579 assert(isValid() && !isUnknown() &&
580 "Can't use VTYPE for uninitialized or unknown");
581 return RISCVVType::getSEWLMULRatio(SEW, VLMul);
584 // Check if the VTYPE for these two VSETVLIInfos produce the same VLMAX.
585 // Note that having the same VLMAX ensures that both share the same
586 // function from AVL to VL; that is, they must produce the same VL value
587 // for any given AVL value.
588 bool hasSameVLMAX(const VSETVLIInfo &Other) const {
589 assert(isValid() && Other.isValid() &&
590 "Can't compare invalid VSETVLIInfos");
591 assert(!isUnknown() && !Other.isUnknown() &&
592 "Can't compare VTYPE in unknown state");
593 return getSEWLMULRatio() == Other.getSEWLMULRatio();
596 bool hasCompatibleVTYPE(const DemandedFields &Used,
597 const VSETVLIInfo &Require) const {
598 return areCompatibleVTYPEs(Require.encodeVTYPE(), encodeVTYPE(), Used);
601 // Determine whether the vector instructions requirements represented by
602 // Require are compatible with the previous vsetvli instruction represented
603 // by this. MI is the instruction whose requirements we're considering.
604 bool isCompatible(const DemandedFields &Used, const VSETVLIInfo &Require,
605 const MachineRegisterInfo &MRI) const {
606 assert(isValid() && Require.isValid() &&
607 "Can't compare invalid VSETVLIInfos");
608 assert(!Require.SEWLMULRatioOnly &&
609 "Expected a valid VTYPE for instruction!");
610 // Nothing is compatible with Unknown.
611 if (isUnknown() || Require.isUnknown())
612 return false;
614 // If only our VLMAX ratio is valid, then this isn't compatible.
615 if (SEWLMULRatioOnly)
616 return false;
618 if (Used.VLAny && !hasSameAVL(Require))
619 return false;
621 if (Used.VLZeroness && !hasEquallyZeroAVL(Require, MRI))
622 return false;
624 return hasCompatibleVTYPE(Used, Require);
627 bool operator==(const VSETVLIInfo &Other) const {
628 // Uninitialized is only equal to another Uninitialized.
629 if (!isValid())
630 return !Other.isValid();
631 if (!Other.isValid())
632 return !isValid();
634 // Unknown is only equal to another Unknown.
635 if (isUnknown())
636 return Other.isUnknown();
637 if (Other.isUnknown())
638 return isUnknown();
640 if (!hasSameAVL(Other))
641 return false;
643 // If the SEWLMULRatioOnly bits are different, then they aren't equal.
644 if (SEWLMULRatioOnly != Other.SEWLMULRatioOnly)
645 return false;
647 // If only the VLMAX is valid, check that it is the same.
648 if (SEWLMULRatioOnly)
649 return hasSameVLMAX(Other);
651 // If the full VTYPE is valid, check that it is the same.
652 return hasSameVTYPE(Other);
655 bool operator!=(const VSETVLIInfo &Other) const {
656 return !(*this == Other);
659 // Calculate the VSETVLIInfo visible to a block assuming this and Other are
660 // both predecessors.
661 VSETVLIInfo intersect(const VSETVLIInfo &Other) const {
662 // If the new value isn't valid, ignore it.
663 if (!Other.isValid())
664 return *this;
666 // If this value isn't valid, this must be the first predecessor, use it.
667 if (!isValid())
668 return Other;
670 // If either is unknown, the result is unknown.
671 if (isUnknown() || Other.isUnknown())
672 return VSETVLIInfo::getUnknown();
674 // If we have an exact, match return this.
675 if (*this == Other)
676 return *this;
678 // Not an exact match, but maybe the AVL and VLMAX are the same. If so,
679 // return an SEW/LMUL ratio only value.
680 if (hasSameAVL(Other) && hasSameVLMAX(Other)) {
681 VSETVLIInfo MergeInfo = *this;
682 MergeInfo.SEWLMULRatioOnly = true;
683 return MergeInfo;
686 // Otherwise the result is unknown.
687 return VSETVLIInfo::getUnknown();
690 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
691 /// Support for debugging, callable in GDB: V->dump()
692 LLVM_DUMP_METHOD void dump() const {
693 print(dbgs());
694 dbgs() << "\n";
697 /// Implement operator<<.
698 /// @{
699 void print(raw_ostream &OS) const {
700 OS << "{";
701 if (!isValid())
702 OS << "Uninitialized";
703 if (isUnknown())
704 OS << "unknown";
705 if (hasAVLReg())
706 OS << "AVLReg=" << (unsigned)AVLReg;
707 if (hasAVLImm())
708 OS << "AVLImm=" << (unsigned)AVLImm;
709 OS << ", "
710 << "VLMul=" << (unsigned)VLMul << ", "
711 << "SEW=" << (unsigned)SEW << ", "
712 << "TailAgnostic=" << (bool)TailAgnostic << ", "
713 << "MaskAgnostic=" << (bool)MaskAgnostic << ", "
714 << "SEWLMULRatioOnly=" << (bool)SEWLMULRatioOnly << "}";
716 #endif
719 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
720 LLVM_ATTRIBUTE_USED
721 inline raw_ostream &operator<<(raw_ostream &OS, const VSETVLIInfo &V) {
722 V.print(OS);
723 return OS;
725 #endif
727 struct BlockData {
728 // The VSETVLIInfo that represents the VL/VTYPE settings on exit from this
729 // block. Calculated in Phase 2.
730 VSETVLIInfo Exit;
732 // The VSETVLIInfo that represents the VL/VTYPE settings from all predecessor
733 // blocks. Calculated in Phase 2, and used by Phase 3.
734 VSETVLIInfo Pred;
736 // Keeps track of whether the block is already in the queue.
737 bool InQueue = false;
739 BlockData() = default;
742 class RISCVInsertVSETVLI : public MachineFunctionPass {
743 const RISCVSubtarget *ST;
744 const TargetInstrInfo *TII;
745 MachineRegisterInfo *MRI;
747 std::vector<BlockData> BlockInfo;
748 std::queue<const MachineBasicBlock *> WorkList;
750 public:
751 static char ID;
753 RISCVInsertVSETVLI() : MachineFunctionPass(ID) {}
754 bool runOnMachineFunction(MachineFunction &MF) override;
756 void getAnalysisUsage(AnalysisUsage &AU) const override {
757 AU.setPreservesCFG();
758 MachineFunctionPass::getAnalysisUsage(AU);
761 StringRef getPassName() const override { return RISCV_INSERT_VSETVLI_NAME; }
763 private:
764 bool needVSETVLI(const MachineInstr &MI, const VSETVLIInfo &Require,
765 const VSETVLIInfo &CurInfo) const;
766 bool needVSETVLIPHI(const VSETVLIInfo &Require,
767 const MachineBasicBlock &MBB) const;
768 void insertVSETVLI(MachineBasicBlock &MBB, MachineInstr &MI,
769 const VSETVLIInfo &Info, const VSETVLIInfo &PrevInfo);
770 void insertVSETVLI(MachineBasicBlock &MBB,
771 MachineBasicBlock::iterator InsertPt, DebugLoc DL,
772 const VSETVLIInfo &Info, const VSETVLIInfo &PrevInfo);
774 void transferBefore(VSETVLIInfo &Info, const MachineInstr &MI) const;
775 void transferAfter(VSETVLIInfo &Info, const MachineInstr &MI) const;
776 bool computeVLVTYPEChanges(const MachineBasicBlock &MBB,
777 VSETVLIInfo &Info) const;
778 void computeIncomingVLVTYPE(const MachineBasicBlock &MBB);
779 void emitVSETVLIs(MachineBasicBlock &MBB);
780 void doLocalPostpass(MachineBasicBlock &MBB);
781 void doPRE(MachineBasicBlock &MBB);
782 void insertReadVL(MachineBasicBlock &MBB);
785 } // end anonymous namespace
787 char RISCVInsertVSETVLI::ID = 0;
789 INITIALIZE_PASS(RISCVInsertVSETVLI, DEBUG_TYPE, RISCV_INSERT_VSETVLI_NAME,
790 false, false)
792 // Return a VSETVLIInfo representing the changes made by this VSETVLI or
793 // VSETIVLI instruction.
794 static VSETVLIInfo getInfoForVSETVLI(const MachineInstr &MI) {
795 VSETVLIInfo NewInfo;
796 if (MI.getOpcode() == RISCV::PseudoVSETIVLI) {
797 NewInfo.setAVLImm(MI.getOperand(1).getImm());
798 } else {
799 assert(MI.getOpcode() == RISCV::PseudoVSETVLI ||
800 MI.getOpcode() == RISCV::PseudoVSETVLIX0);
801 Register AVLReg = MI.getOperand(1).getReg();
802 assert((AVLReg != RISCV::X0 || MI.getOperand(0).getReg() != RISCV::X0) &&
803 "Can't handle X0, X0 vsetvli yet");
804 NewInfo.setAVLReg(AVLReg);
806 NewInfo.setVTYPE(MI.getOperand(2).getImm());
808 return NewInfo;
811 static unsigned computeVLMAX(unsigned VLEN, unsigned SEW,
812 RISCVII::VLMUL VLMul) {
813 auto [LMul, Fractional] = RISCVVType::decodeVLMUL(VLMul);
814 if (Fractional)
815 VLEN = VLEN / LMul;
816 else
817 VLEN = VLEN * LMul;
818 return VLEN/SEW;
821 static VSETVLIInfo computeInfoForInstr(const MachineInstr &MI, uint64_t TSFlags,
822 const RISCVSubtarget &ST,
823 const MachineRegisterInfo *MRI) {
824 VSETVLIInfo InstrInfo;
826 bool TailAgnostic = true;
827 bool MaskAgnostic = true;
828 if (!hasUndefinedMergeOp(MI, *MRI)) {
829 // Start with undisturbed.
830 TailAgnostic = false;
831 MaskAgnostic = false;
833 // If there is a policy operand, use it.
834 if (RISCVII::hasVecPolicyOp(TSFlags)) {
835 const MachineOperand &Op = MI.getOperand(MI.getNumExplicitOperands() - 1);
836 uint64_t Policy = Op.getImm();
837 assert(Policy <= (RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC) &&
838 "Invalid Policy Value");
839 TailAgnostic = Policy & RISCVII::TAIL_AGNOSTIC;
840 MaskAgnostic = Policy & RISCVII::MASK_AGNOSTIC;
843 // Some pseudo instructions force a tail agnostic policy despite having a
844 // tied def.
845 if (RISCVII::doesForceTailAgnostic(TSFlags))
846 TailAgnostic = true;
848 if (!RISCVII::usesMaskPolicy(TSFlags))
849 MaskAgnostic = true;
852 RISCVII::VLMUL VLMul = RISCVII::getLMul(TSFlags);
854 unsigned Log2SEW = MI.getOperand(getSEWOpNum(MI)).getImm();
855 // A Log2SEW of 0 is an operation on mask registers only.
856 unsigned SEW = Log2SEW ? 1 << Log2SEW : 8;
857 assert(RISCVVType::isValidSEW(SEW) && "Unexpected SEW");
859 if (RISCVII::hasVLOp(TSFlags)) {
860 const MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
861 if (VLOp.isImm()) {
862 int64_t Imm = VLOp.getImm();
863 // Conver the VLMax sentintel to X0 register.
864 if (Imm == RISCV::VLMaxSentinel) {
865 // If we know the exact VLEN, see if we can use the constant encoding
866 // for the VLMAX instead. This reduces register pressure slightly.
867 const unsigned VLMAX = computeVLMAX(ST.getRealMaxVLen(), SEW, VLMul);
868 if (ST.getRealMinVLen() == ST.getRealMaxVLen() && VLMAX <= 31)
869 InstrInfo.setAVLImm(VLMAX);
870 else
871 InstrInfo.setAVLReg(RISCV::X0);
873 else
874 InstrInfo.setAVLImm(Imm);
875 } else {
876 InstrInfo.setAVLReg(VLOp.getReg());
878 } else {
879 assert(isScalarExtractInstr(MI));
880 InstrInfo.setAVLReg(RISCV::NoRegister);
882 #ifndef NDEBUG
883 if (std::optional<unsigned> EEW = getEEWForLoadStore(MI)) {
884 assert(SEW == EEW && "Initial SEW doesn't match expected EEW");
886 #endif
887 InstrInfo.setVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
889 // If AVL is defined by a vsetvli with the same VLMAX, we can replace the
890 // AVL operand with the AVL of the defining vsetvli. We avoid general
891 // register AVLs to avoid extending live ranges without being sure we can
892 // kill the original source reg entirely.
893 if (InstrInfo.hasAVLReg() && InstrInfo.getAVLReg().isVirtual()) {
894 MachineInstr *DefMI = MRI->getVRegDef(InstrInfo.getAVLReg());
895 if (DefMI && isVectorConfigInstr(*DefMI)) {
896 VSETVLIInfo DefInstrInfo = getInfoForVSETVLI(*DefMI);
897 if (DefInstrInfo.hasSameVLMAX(InstrInfo) &&
898 (DefInstrInfo.hasAVLImm() || DefInstrInfo.getAVLReg() == RISCV::X0)) {
899 InstrInfo.setAVL(DefInstrInfo);
904 return InstrInfo;
907 void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB, MachineInstr &MI,
908 const VSETVLIInfo &Info,
909 const VSETVLIInfo &PrevInfo) {
910 DebugLoc DL = MI.getDebugLoc();
911 insertVSETVLI(MBB, MachineBasicBlock::iterator(&MI), DL, Info, PrevInfo);
914 void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
915 MachineBasicBlock::iterator InsertPt, DebugLoc DL,
916 const VSETVLIInfo &Info, const VSETVLIInfo &PrevInfo) {
918 ++NumInsertedVSETVL;
919 if (PrevInfo.isValid() && !PrevInfo.isUnknown()) {
920 // Use X0, X0 form if the AVL is the same and the SEW+LMUL gives the same
921 // VLMAX.
922 if (Info.hasSameAVL(PrevInfo) && Info.hasSameVLMAX(PrevInfo)) {
923 BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLIX0))
924 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
925 .addReg(RISCV::X0, RegState::Kill)
926 .addImm(Info.encodeVTYPE())
927 .addReg(RISCV::VL, RegState::Implicit);
928 return;
931 // If our AVL is a virtual register, it might be defined by a VSET(I)VLI. If
932 // it has the same VLMAX we want and the last VL/VTYPE we observed is the
933 // same, we can use the X0, X0 form.
934 if (Info.hasSameVLMAX(PrevInfo) && Info.hasAVLReg() &&
935 Info.getAVLReg().isVirtual()) {
936 if (MachineInstr *DefMI = MRI->getVRegDef(Info.getAVLReg())) {
937 if (isVectorConfigInstr(*DefMI)) {
938 VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
939 if (DefInfo.hasSameAVL(PrevInfo) && DefInfo.hasSameVLMAX(PrevInfo)) {
940 BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLIX0))
941 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
942 .addReg(RISCV::X0, RegState::Kill)
943 .addImm(Info.encodeVTYPE())
944 .addReg(RISCV::VL, RegState::Implicit);
945 return;
952 if (Info.hasAVLImm()) {
953 BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETIVLI))
954 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
955 .addImm(Info.getAVLImm())
956 .addImm(Info.encodeVTYPE());
957 return;
960 Register AVLReg = Info.getAVLReg();
961 if (AVLReg == RISCV::NoRegister) {
962 // We can only use x0, x0 if there's no chance of the vtype change causing
963 // the previous vl to become invalid.
964 if (PrevInfo.isValid() && !PrevInfo.isUnknown() &&
965 Info.hasSameVLMAX(PrevInfo)) {
966 BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLIX0))
967 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
968 .addReg(RISCV::X0, RegState::Kill)
969 .addImm(Info.encodeVTYPE())
970 .addReg(RISCV::VL, RegState::Implicit);
971 return;
973 // Otherwise use an AVL of 1 to avoid depending on previous vl.
974 BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETIVLI))
975 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
976 .addImm(1)
977 .addImm(Info.encodeVTYPE());
978 return;
981 if (AVLReg.isVirtual())
982 MRI->constrainRegClass(AVLReg, &RISCV::GPRNoX0RegClass);
984 // Use X0 as the DestReg unless AVLReg is X0. We also need to change the
985 // opcode if the AVLReg is X0 as they have different register classes for
986 // the AVL operand.
987 Register DestReg = RISCV::X0;
988 unsigned Opcode = RISCV::PseudoVSETVLI;
989 if (AVLReg == RISCV::X0) {
990 DestReg = MRI->createVirtualRegister(&RISCV::GPRRegClass);
991 Opcode = RISCV::PseudoVSETVLIX0;
993 BuildMI(MBB, InsertPt, DL, TII->get(Opcode))
994 .addReg(DestReg, RegState::Define | RegState::Dead)
995 .addReg(AVLReg)
996 .addImm(Info.encodeVTYPE());
999 static bool isLMUL1OrSmaller(RISCVII::VLMUL LMUL) {
1000 auto [LMul, Fractional] = RISCVVType::decodeVLMUL(LMUL);
1001 return Fractional || LMul == 1;
1004 /// Return true if a VSETVLI is required to transition from CurInfo to Require
1005 /// before MI.
1006 bool RISCVInsertVSETVLI::needVSETVLI(const MachineInstr &MI,
1007 const VSETVLIInfo &Require,
1008 const VSETVLIInfo &CurInfo) const {
1009 assert(Require == computeInfoForInstr(MI, MI.getDesc().TSFlags, *ST, MRI));
1011 if (!CurInfo.isValid() || CurInfo.isUnknown() || CurInfo.hasSEWLMULRatioOnly())
1012 return true;
1014 DemandedFields Used = getDemanded(MI, MRI, ST);
1016 // A slidedown/slideup with an *undefined* merge op can freely clobber
1017 // elements not copied from the source vector (e.g. masked off, tail, or
1018 // slideup's prefix). Notes:
1019 // * We can't modify SEW here since the slide amount is in units of SEW.
1020 // * VL=1 is special only because we have existing support for zero vs
1021 // non-zero VL. We could generalize this if we had a VL > C predicate.
1022 // * The LMUL1 restriction is for machines whose latency may depend on VL.
1023 // * As above, this is only legal for tail "undefined" not "agnostic".
1024 if (isVSlideInstr(MI) && Require.hasAVLImm() && Require.getAVLImm() == 1 &&
1025 isLMUL1OrSmaller(CurInfo.getVLMUL()) && hasUndefinedMergeOp(MI, *MRI)) {
1026 Used.VLAny = false;
1027 Used.VLZeroness = true;
1028 Used.LMUL = false;
1029 Used.TailPolicy = false;
1032 // A tail undefined vmv.v.i/x or vfmv.v.f with VL=1 can be treated in the same
1033 // semantically as vmv.s.x. This is particularly useful since we don't have an
1034 // immediate form of vmv.s.x, and thus frequently use vmv.v.i in it's place.
1035 // Since a splat is non-constant time in LMUL, we do need to be careful to not
1036 // increase the number of active vector registers (unlike for vmv.s.x.)
1037 if (isScalarSplatInstr(MI) && Require.hasAVLImm() && Require.getAVLImm() == 1 &&
1038 isLMUL1OrSmaller(CurInfo.getVLMUL()) && hasUndefinedMergeOp(MI, *MRI)) {
1039 Used.LMUL = false;
1040 Used.SEWLMULRatio = false;
1041 Used.VLAny = false;
1042 if (isFloatScalarMoveOrScalarSplatInstr(MI) && !ST->hasVInstructionsF64())
1043 Used.SEW = DemandedFields::SEWGreaterThanOrEqualAndLessThan64;
1044 else
1045 Used.SEW = DemandedFields::SEWGreaterThanOrEqual;
1046 Used.TailPolicy = false;
1049 if (CurInfo.isCompatible(Used, Require, *MRI))
1050 return false;
1052 // We didn't find a compatible value. If our AVL is a virtual register,
1053 // it might be defined by a VSET(I)VLI. If it has the same VLMAX we need
1054 // and the last VL/VTYPE we observed is the same, we don't need a
1055 // VSETVLI here.
1056 if (Require.hasAVLReg() && Require.getAVLReg().isVirtual() &&
1057 CurInfo.hasCompatibleVTYPE(Used, Require)) {
1058 if (MachineInstr *DefMI = MRI->getVRegDef(Require.getAVLReg())) {
1059 if (isVectorConfigInstr(*DefMI)) {
1060 VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
1061 if (DefInfo.hasSameAVL(CurInfo) && DefInfo.hasSameVLMAX(CurInfo))
1062 return false;
1067 return true;
1070 // If we don't use LMUL or the SEW/LMUL ratio, then adjust LMUL so that we
1071 // maintain the SEW/LMUL ratio. This allows us to eliminate VL toggles in more
1072 // places.
1073 static VSETVLIInfo adjustIncoming(VSETVLIInfo PrevInfo, VSETVLIInfo NewInfo,
1074 DemandedFields &Demanded) {
1075 VSETVLIInfo Info = NewInfo;
1077 if (!Demanded.LMUL && !Demanded.SEWLMULRatio && PrevInfo.isValid() &&
1078 !PrevInfo.isUnknown()) {
1079 if (auto NewVLMul = RISCVVType::getSameRatioLMUL(
1080 PrevInfo.getSEW(), PrevInfo.getVLMUL(), Info.getSEW()))
1081 Info.setVLMul(*NewVLMul);
1082 Demanded.LMUL = true;
1085 return Info;
1088 // Given an incoming state reaching MI, minimally modifies that state so that it
1089 // is compatible with MI. The resulting state is guaranteed to be semantically
1090 // legal for MI, but may not be the state requested by MI.
1091 void RISCVInsertVSETVLI::transferBefore(VSETVLIInfo &Info,
1092 const MachineInstr &MI) const {
1093 uint64_t TSFlags = MI.getDesc().TSFlags;
1094 if (!RISCVII::hasSEWOp(TSFlags))
1095 return;
1097 const VSETVLIInfo NewInfo = computeInfoForInstr(MI, TSFlags, *ST, MRI);
1098 assert(NewInfo.isValid() && !NewInfo.isUnknown());
1099 if (Info.isValid() && !needVSETVLI(MI, NewInfo, Info))
1100 return;
1102 const VSETVLIInfo PrevInfo = Info;
1103 if (!Info.isValid() || Info.isUnknown())
1104 Info = NewInfo;
1106 DemandedFields Demanded = getDemanded(MI, MRI, ST);
1107 const VSETVLIInfo IncomingInfo = adjustIncoming(PrevInfo, NewInfo, Demanded);
1109 // If MI only demands that VL has the same zeroness, we only need to set the
1110 // AVL if the zeroness differs. This removes a vsetvli entirely if the types
1111 // match or allows use of cheaper avl preserving variant if VLMAX doesn't
1112 // change. If VLMAX might change, we couldn't use the 'vsetvli x0, x0, vtype"
1113 // variant, so we avoid the transform to prevent extending live range of an
1114 // avl register operand.
1115 // TODO: We can probably relax this for immediates.
1116 bool EquallyZero = IncomingInfo.hasEquallyZeroAVL(PrevInfo, *MRI) &&
1117 IncomingInfo.hasSameVLMAX(PrevInfo);
1118 if (Demanded.VLAny || (Demanded.VLZeroness && !EquallyZero))
1119 Info.setAVL(IncomingInfo);
1121 Info.setVTYPE(
1122 ((Demanded.LMUL || Demanded.SEWLMULRatio) ? IncomingInfo : Info)
1123 .getVLMUL(),
1124 ((Demanded.SEW || Demanded.SEWLMULRatio) ? IncomingInfo : Info).getSEW(),
1125 // Prefer tail/mask agnostic since it can be relaxed to undisturbed later
1126 // if needed.
1127 (Demanded.TailPolicy ? IncomingInfo : Info).getTailAgnostic() ||
1128 IncomingInfo.getTailAgnostic(),
1129 (Demanded.MaskPolicy ? IncomingInfo : Info).getMaskAgnostic() ||
1130 IncomingInfo.getMaskAgnostic());
1132 // If we only knew the sew/lmul ratio previously, replace the VTYPE but keep
1133 // the AVL.
1134 if (Info.hasSEWLMULRatioOnly()) {
1135 VSETVLIInfo RatiolessInfo = IncomingInfo;
1136 RatiolessInfo.setAVL(Info);
1137 Info = RatiolessInfo;
1141 // Given a state with which we evaluated MI (see transferBefore above for why
1142 // this might be different that the state MI requested), modify the state to
1143 // reflect the changes MI might make.
1144 void RISCVInsertVSETVLI::transferAfter(VSETVLIInfo &Info,
1145 const MachineInstr &MI) const {
1146 if (isVectorConfigInstr(MI)) {
1147 Info = getInfoForVSETVLI(MI);
1148 return;
1151 if (RISCV::isFaultFirstLoad(MI)) {
1152 // Update AVL to vl-output of the fault first load.
1153 Info.setAVLReg(MI.getOperand(1).getReg());
1154 return;
1157 // If this is something that updates VL/VTYPE that we don't know about, set
1158 // the state to unknown.
1159 if (MI.isCall() || MI.isInlineAsm() || MI.modifiesRegister(RISCV::VL) ||
1160 MI.modifiesRegister(RISCV::VTYPE))
1161 Info = VSETVLIInfo::getUnknown();
1164 bool RISCVInsertVSETVLI::computeVLVTYPEChanges(const MachineBasicBlock &MBB,
1165 VSETVLIInfo &Info) const {
1166 bool HadVectorOp = false;
1168 Info = BlockInfo[MBB.getNumber()].Pred;
1169 for (const MachineInstr &MI : MBB) {
1170 transferBefore(Info, MI);
1172 if (isVectorConfigInstr(MI) || RISCVII::hasSEWOp(MI.getDesc().TSFlags))
1173 HadVectorOp = true;
1175 transferAfter(Info, MI);
1178 return HadVectorOp;
1181 void RISCVInsertVSETVLI::computeIncomingVLVTYPE(const MachineBasicBlock &MBB) {
1183 BlockData &BBInfo = BlockInfo[MBB.getNumber()];
1185 BBInfo.InQueue = false;
1187 // Start with the previous entry so that we keep the most conservative state
1188 // we have ever found.
1189 VSETVLIInfo InInfo = BBInfo.Pred;
1190 if (MBB.pred_empty()) {
1191 // There are no predecessors, so use the default starting status.
1192 InInfo.setUnknown();
1193 } else {
1194 for (MachineBasicBlock *P : MBB.predecessors())
1195 InInfo = InInfo.intersect(BlockInfo[P->getNumber()].Exit);
1198 // If we don't have any valid predecessor value, wait until we do.
1199 if (!InInfo.isValid())
1200 return;
1202 // If no change, no need to rerun block
1203 if (InInfo == BBInfo.Pred)
1204 return;
1206 BBInfo.Pred = InInfo;
1207 LLVM_DEBUG(dbgs() << "Entry state of " << printMBBReference(MBB)
1208 << " changed to " << BBInfo.Pred << "\n");
1210 // Note: It's tempting to cache the state changes here, but due to the
1211 // compatibility checks performed a blocks output state can change based on
1212 // the input state. To cache, we'd have to add logic for finding
1213 // never-compatible state changes.
1214 VSETVLIInfo TmpStatus;
1215 computeVLVTYPEChanges(MBB, TmpStatus);
1217 // If the new exit value matches the old exit value, we don't need to revisit
1218 // any blocks.
1219 if (BBInfo.Exit == TmpStatus)
1220 return;
1222 BBInfo.Exit = TmpStatus;
1223 LLVM_DEBUG(dbgs() << "Exit state of " << printMBBReference(MBB)
1224 << " changed to " << BBInfo.Exit << "\n");
1226 // Add the successors to the work list so we can propagate the changed exit
1227 // status.
1228 for (MachineBasicBlock *S : MBB.successors())
1229 if (!BlockInfo[S->getNumber()].InQueue) {
1230 BlockInfo[S->getNumber()].InQueue = true;
1231 WorkList.push(S);
1235 // If we weren't able to prove a vsetvli was directly unneeded, it might still
1236 // be unneeded if the AVL is a phi node where all incoming values are VL
1237 // outputs from the last VSETVLI in their respective basic blocks.
1238 bool RISCVInsertVSETVLI::needVSETVLIPHI(const VSETVLIInfo &Require,
1239 const MachineBasicBlock &MBB) const {
1240 if (DisableInsertVSETVLPHIOpt)
1241 return true;
1243 if (!Require.hasAVLReg())
1244 return true;
1246 Register AVLReg = Require.getAVLReg();
1247 if (!AVLReg.isVirtual())
1248 return true;
1250 // We need the AVL to be produce by a PHI node in this basic block.
1251 MachineInstr *PHI = MRI->getVRegDef(AVLReg);
1252 if (!PHI || PHI->getOpcode() != RISCV::PHI || PHI->getParent() != &MBB)
1253 return true;
1255 for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps;
1256 PHIOp += 2) {
1257 Register InReg = PHI->getOperand(PHIOp).getReg();
1258 MachineBasicBlock *PBB = PHI->getOperand(PHIOp + 1).getMBB();
1259 const BlockData &PBBInfo = BlockInfo[PBB->getNumber()];
1260 // If the exit from the predecessor has the VTYPE we are looking for
1261 // we might be able to avoid a VSETVLI.
1262 if (PBBInfo.Exit.isUnknown() || !PBBInfo.Exit.hasSameVTYPE(Require))
1263 return true;
1265 // We need the PHI input to the be the output of a VSET(I)VLI.
1266 MachineInstr *DefMI = MRI->getVRegDef(InReg);
1267 if (!DefMI || !isVectorConfigInstr(*DefMI))
1268 return true;
1270 // We found a VSET(I)VLI make sure it matches the output of the
1271 // predecessor block.
1272 VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
1273 if (!DefInfo.hasSameAVL(PBBInfo.Exit) ||
1274 !DefInfo.hasSameVTYPE(PBBInfo.Exit))
1275 return true;
1278 // If all the incoming values to the PHI checked out, we don't need
1279 // to insert a VSETVLI.
1280 return false;
1283 void RISCVInsertVSETVLI::emitVSETVLIs(MachineBasicBlock &MBB) {
1284 VSETVLIInfo CurInfo = BlockInfo[MBB.getNumber()].Pred;
1285 // Track whether the prefix of the block we've scanned is transparent
1286 // (meaning has not yet changed the abstract state).
1287 bool PrefixTransparent = true;
1288 for (MachineInstr &MI : MBB) {
1289 const VSETVLIInfo PrevInfo = CurInfo;
1290 transferBefore(CurInfo, MI);
1292 // If this is an explicit VSETVLI or VSETIVLI, update our state.
1293 if (isVectorConfigInstr(MI)) {
1294 // Conservatively, mark the VL and VTYPE as live.
1295 assert(MI.getOperand(3).getReg() == RISCV::VL &&
1296 MI.getOperand(4).getReg() == RISCV::VTYPE &&
1297 "Unexpected operands where VL and VTYPE should be");
1298 MI.getOperand(3).setIsDead(false);
1299 MI.getOperand(4).setIsDead(false);
1300 PrefixTransparent = false;
1303 uint64_t TSFlags = MI.getDesc().TSFlags;
1304 if (RISCVII::hasSEWOp(TSFlags)) {
1305 if (PrevInfo != CurInfo) {
1306 // If this is the first implicit state change, and the state change
1307 // requested can be proven to produce the same register contents, we
1308 // can skip emitting the actual state change and continue as if we
1309 // had since we know the GPR result of the implicit state change
1310 // wouldn't be used and VL/VTYPE registers are correct. Note that
1311 // we *do* need to model the state as if it changed as while the
1312 // register contents are unchanged, the abstract model can change.
1313 if (!PrefixTransparent || needVSETVLIPHI(CurInfo, MBB))
1314 insertVSETVLI(MBB, MI, CurInfo, PrevInfo);
1315 PrefixTransparent = false;
1318 if (RISCVII::hasVLOp(TSFlags)) {
1319 MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
1320 if (VLOp.isReg()) {
1321 Register Reg = VLOp.getReg();
1322 MachineInstr *VLOpDef = MRI->getVRegDef(Reg);
1324 // Erase the AVL operand from the instruction.
1325 VLOp.setReg(RISCV::NoRegister);
1326 VLOp.setIsKill(false);
1328 // If the AVL was an immediate > 31, then it would have been emitted
1329 // as an ADDI. However, the ADDI might not have been used in the
1330 // vsetvli, or a vsetvli might not have been emitted, so it may be
1331 // dead now.
1332 if (VLOpDef && TII->isAddImmediate(*VLOpDef, Reg) &&
1333 MRI->use_nodbg_empty(Reg))
1334 VLOpDef->eraseFromParent();
1336 MI.addOperand(MachineOperand::CreateReg(RISCV::VL, /*isDef*/ false,
1337 /*isImp*/ true));
1339 MI.addOperand(MachineOperand::CreateReg(RISCV::VTYPE, /*isDef*/ false,
1340 /*isImp*/ true));
1343 if (MI.isCall() || MI.isInlineAsm() || MI.modifiesRegister(RISCV::VL) ||
1344 MI.modifiesRegister(RISCV::VTYPE))
1345 PrefixTransparent = false;
1347 transferAfter(CurInfo, MI);
1350 // If we reach the end of the block and our current info doesn't match the
1351 // expected info, insert a vsetvli to correct.
1352 if (!UseStrictAsserts) {
1353 const VSETVLIInfo &ExitInfo = BlockInfo[MBB.getNumber()].Exit;
1354 if (CurInfo.isValid() && ExitInfo.isValid() && !ExitInfo.isUnknown() &&
1355 CurInfo != ExitInfo) {
1356 // Note there's an implicit assumption here that terminators never use
1357 // or modify VL or VTYPE. Also, fallthrough will return end().
1358 auto InsertPt = MBB.getFirstInstrTerminator();
1359 insertVSETVLI(MBB, InsertPt, MBB.findDebugLoc(InsertPt), ExitInfo,
1360 CurInfo);
1361 CurInfo = ExitInfo;
1365 if (UseStrictAsserts && CurInfo.isValid()) {
1366 const auto &Info = BlockInfo[MBB.getNumber()];
1367 if (CurInfo != Info.Exit) {
1368 LLVM_DEBUG(dbgs() << "in block " << printMBBReference(MBB) << "\n");
1369 LLVM_DEBUG(dbgs() << " begin state: " << Info.Pred << "\n");
1370 LLVM_DEBUG(dbgs() << " expected end state: " << Info.Exit << "\n");
1371 LLVM_DEBUG(dbgs() << " actual end state: " << CurInfo << "\n");
1373 assert(CurInfo == Info.Exit &&
1374 "InsertVSETVLI dataflow invariant violated");
1378 /// Perform simple partial redundancy elimination of the VSETVLI instructions
1379 /// we're about to insert by looking for cases where we can PRE from the
1380 /// beginning of one block to the end of one of its predecessors. Specifically,
1381 /// this is geared to catch the common case of a fixed length vsetvl in a single
1382 /// block loop when it could execute once in the preheader instead.
1383 void RISCVInsertVSETVLI::doPRE(MachineBasicBlock &MBB) {
1384 if (!BlockInfo[MBB.getNumber()].Pred.isUnknown())
1385 return;
1387 MachineBasicBlock *UnavailablePred = nullptr;
1388 VSETVLIInfo AvailableInfo;
1389 for (MachineBasicBlock *P : MBB.predecessors()) {
1390 const VSETVLIInfo &PredInfo = BlockInfo[P->getNumber()].Exit;
1391 if (PredInfo.isUnknown()) {
1392 if (UnavailablePred)
1393 return;
1394 UnavailablePred = P;
1395 } else if (!AvailableInfo.isValid()) {
1396 AvailableInfo = PredInfo;
1397 } else if (AvailableInfo != PredInfo) {
1398 return;
1402 // Unreachable, single pred, or full redundancy. Note that FRE is handled by
1403 // phase 3.
1404 if (!UnavailablePred || !AvailableInfo.isValid())
1405 return;
1407 // If we don't know the exact VTYPE, we can't copy the vsetvli to the exit of
1408 // the unavailable pred.
1409 if (AvailableInfo.hasSEWLMULRatioOnly())
1410 return;
1412 // Critical edge - TODO: consider splitting?
1413 if (UnavailablePred->succ_size() != 1)
1414 return;
1416 // If the AVL value is a register (other than our VLMAX sentinel),
1417 // we need to prove the value is available at the point we're going
1418 // to insert the vsetvli at.
1419 if (AvailableInfo.hasAVLReg() && RISCV::X0 != AvailableInfo.getAVLReg()) {
1420 MachineInstr *AVLDefMI = MRI->getVRegDef(AvailableInfo.getAVLReg());
1421 if (!AVLDefMI)
1422 return;
1423 // This is an inline dominance check which covers the case of
1424 // UnavailablePred being the preheader of a loop.
1425 if (AVLDefMI->getParent() != UnavailablePred)
1426 return;
1427 for (auto &TermMI : UnavailablePred->terminators())
1428 if (&TermMI == AVLDefMI)
1429 return;
1432 // Model the effect of changing the input state of the block MBB to
1433 // AvailableInfo. We're looking for two issues here; one legality,
1434 // one profitability.
1435 // 1) If the block doesn't use some of the fields from VL or VTYPE, we
1436 // may hit the end of the block with a different end state. We can
1437 // not make this change without reflowing later blocks as well.
1438 // 2) If we don't actually remove a transition, inserting a vsetvli
1439 // into the predecessor block would be correct, but unprofitable.
1440 VSETVLIInfo OldInfo = BlockInfo[MBB.getNumber()].Pred;
1441 VSETVLIInfo CurInfo = AvailableInfo;
1442 int TransitionsRemoved = 0;
1443 for (const MachineInstr &MI : MBB) {
1444 const VSETVLIInfo LastInfo = CurInfo;
1445 const VSETVLIInfo LastOldInfo = OldInfo;
1446 transferBefore(CurInfo, MI);
1447 transferBefore(OldInfo, MI);
1448 if (CurInfo == LastInfo)
1449 TransitionsRemoved++;
1450 if (LastOldInfo == OldInfo)
1451 TransitionsRemoved--;
1452 transferAfter(CurInfo, MI);
1453 transferAfter(OldInfo, MI);
1454 if (CurInfo == OldInfo)
1455 // Convergence. All transitions after this must match by construction.
1456 break;
1458 if (CurInfo != OldInfo || TransitionsRemoved <= 0)
1459 // Issues 1 and 2 above
1460 return;
1462 // Finally, update both data flow state and insert the actual vsetvli.
1463 // Doing both keeps the code in sync with the dataflow results, which
1464 // is critical for correctness of phase 3.
1465 auto OldExit = BlockInfo[UnavailablePred->getNumber()].Exit;
1466 LLVM_DEBUG(dbgs() << "PRE VSETVLI from " << MBB.getName() << " to "
1467 << UnavailablePred->getName() << " with state "
1468 << AvailableInfo << "\n");
1469 BlockInfo[UnavailablePred->getNumber()].Exit = AvailableInfo;
1470 BlockInfo[MBB.getNumber()].Pred = AvailableInfo;
1472 // Note there's an implicit assumption here that terminators never use
1473 // or modify VL or VTYPE. Also, fallthrough will return end().
1474 auto InsertPt = UnavailablePred->getFirstInstrTerminator();
1475 insertVSETVLI(*UnavailablePred, InsertPt,
1476 UnavailablePred->findDebugLoc(InsertPt),
1477 AvailableInfo, OldExit);
1480 static void doUnion(DemandedFields &A, DemandedFields B) {
1481 A.VLAny |= B.VLAny;
1482 A.VLZeroness |= B.VLZeroness;
1483 A.SEW = std::max(A.SEW, B.SEW);
1484 A.LMUL |= B.LMUL;
1485 A.SEWLMULRatio |= B.SEWLMULRatio;
1486 A.TailPolicy |= B.TailPolicy;
1487 A.MaskPolicy |= B.MaskPolicy;
1490 // Return true if we can mutate PrevMI to match MI without changing any the
1491 // fields which would be observed.
1492 static bool canMutatePriorConfig(const MachineInstr &PrevMI,
1493 const MachineInstr &MI,
1494 const DemandedFields &Used,
1495 const MachineRegisterInfo &MRI) {
1496 // If the VL values aren't equal, return false if either a) the former is
1497 // demanded, or b) we can't rewrite the former to be the later for
1498 // implementation reasons.
1499 if (!isVLPreservingConfig(MI)) {
1500 if (Used.VLAny)
1501 return false;
1503 if (Used.VLZeroness) {
1504 if (isVLPreservingConfig(PrevMI))
1505 return false;
1506 if (!getInfoForVSETVLI(PrevMI).hasEquallyZeroAVL(getInfoForVSETVLI(MI),
1507 MRI))
1508 return false;
1511 auto &AVL = MI.getOperand(1);
1512 auto &PrevAVL = PrevMI.getOperand(1);
1513 assert(MRI.isSSA());
1515 // If the AVL is a register, we need to make sure MI's AVL dominates PrevMI.
1516 // For now just check that PrevMI uses the same virtual register.
1517 if (AVL.isReg() && AVL.getReg() != RISCV::X0) {
1518 if (AVL.getReg().isPhysical())
1519 return false;
1520 if (!PrevAVL.isReg() || PrevAVL.getReg() != AVL.getReg())
1521 return false;
1525 if (!PrevMI.getOperand(2).isImm() || !MI.getOperand(2).isImm())
1526 return false;
1528 auto PriorVType = PrevMI.getOperand(2).getImm();
1529 auto VType = MI.getOperand(2).getImm();
1530 return areCompatibleVTYPEs(PriorVType, VType, Used);
1533 void RISCVInsertVSETVLI::doLocalPostpass(MachineBasicBlock &MBB) {
1534 MachineInstr *NextMI = nullptr;
1535 // We can have arbitrary code in successors, so VL and VTYPE
1536 // must be considered demanded.
1537 DemandedFields Used;
1538 Used.demandVL();
1539 Used.demandVTYPE();
1540 SmallVector<MachineInstr*> ToDelete;
1541 for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) {
1543 if (!isVectorConfigInstr(MI)) {
1544 doUnion(Used, getDemanded(MI, MRI, ST));
1545 continue;
1548 Register VRegDef = MI.getOperand(0).getReg();
1549 if (VRegDef != RISCV::X0 &&
1550 !(VRegDef.isVirtual() && MRI->use_nodbg_empty(VRegDef)))
1551 Used.demandVL();
1553 if (NextMI) {
1554 if (!Used.usedVL() && !Used.usedVTYPE()) {
1555 ToDelete.push_back(&MI);
1556 // Leave NextMI unchanged
1557 continue;
1558 } else if (canMutatePriorConfig(MI, *NextMI, Used, *MRI)) {
1559 if (!isVLPreservingConfig(*NextMI)) {
1560 MI.getOperand(0).setReg(NextMI->getOperand(0).getReg());
1561 MI.getOperand(0).setIsDead(false);
1562 Register OldVLReg;
1563 if (MI.getOperand(1).isReg())
1564 OldVLReg = MI.getOperand(1).getReg();
1565 if (NextMI->getOperand(1).isImm())
1566 MI.getOperand(1).ChangeToImmediate(NextMI->getOperand(1).getImm());
1567 else
1568 MI.getOperand(1).ChangeToRegister(NextMI->getOperand(1).getReg(), false);
1569 if (OldVLReg) {
1570 MachineInstr *VLOpDef = MRI->getUniqueVRegDef(OldVLReg);
1571 if (VLOpDef && TII->isAddImmediate(*VLOpDef, OldVLReg) &&
1572 MRI->use_nodbg_empty(OldVLReg))
1573 VLOpDef->eraseFromParent();
1575 MI.setDesc(NextMI->getDesc());
1577 MI.getOperand(2).setImm(NextMI->getOperand(2).getImm());
1578 ToDelete.push_back(NextMI);
1579 // fallthrough
1582 NextMI = &MI;
1583 Used = getDemanded(MI, MRI, ST);
1586 NumRemovedVSETVL += ToDelete.size();
1587 for (auto *MI : ToDelete)
1588 MI->eraseFromParent();
1591 void RISCVInsertVSETVLI::insertReadVL(MachineBasicBlock &MBB) {
1592 for (auto I = MBB.begin(), E = MBB.end(); I != E;) {
1593 MachineInstr &MI = *I++;
1594 if (RISCV::isFaultFirstLoad(MI)) {
1595 Register VLOutput = MI.getOperand(1).getReg();
1596 if (!MRI->use_nodbg_empty(VLOutput))
1597 BuildMI(MBB, I, MI.getDebugLoc(), TII->get(RISCV::PseudoReadVL),
1598 VLOutput);
1599 // We don't use the vl output of the VLEFF/VLSEGFF anymore.
1600 MI.getOperand(1).setReg(RISCV::X0);
1605 bool RISCVInsertVSETVLI::runOnMachineFunction(MachineFunction &MF) {
1606 // Skip if the vector extension is not enabled.
1607 ST = &MF.getSubtarget<RISCVSubtarget>();
1608 if (!ST->hasVInstructions())
1609 return false;
1611 LLVM_DEBUG(dbgs() << "Entering InsertVSETVLI for " << MF.getName() << "\n");
1613 TII = ST->getInstrInfo();
1614 MRI = &MF.getRegInfo();
1616 assert(BlockInfo.empty() && "Expect empty block infos");
1617 BlockInfo.resize(MF.getNumBlockIDs());
1619 bool HaveVectorOp = false;
1621 // Phase 1 - determine how VL/VTYPE are affected by the each block.
1622 for (const MachineBasicBlock &MBB : MF) {
1623 VSETVLIInfo TmpStatus;
1624 HaveVectorOp |= computeVLVTYPEChanges(MBB, TmpStatus);
1625 // Initial exit state is whatever change we found in the block.
1626 BlockData &BBInfo = BlockInfo[MBB.getNumber()];
1627 BBInfo.Exit = TmpStatus;
1628 LLVM_DEBUG(dbgs() << "Initial exit state of " << printMBBReference(MBB)
1629 << " is " << BBInfo.Exit << "\n");
1633 // If we didn't find any instructions that need VSETVLI, we're done.
1634 if (!HaveVectorOp) {
1635 BlockInfo.clear();
1636 return false;
1639 // Phase 2 - determine the exit VL/VTYPE from each block. We add all
1640 // blocks to the list here, but will also add any that need to be revisited
1641 // during Phase 2 processing.
1642 for (const MachineBasicBlock &MBB : MF) {
1643 WorkList.push(&MBB);
1644 BlockInfo[MBB.getNumber()].InQueue = true;
1646 while (!WorkList.empty()) {
1647 const MachineBasicBlock &MBB = *WorkList.front();
1648 WorkList.pop();
1649 computeIncomingVLVTYPE(MBB);
1652 // Perform partial redundancy elimination of vsetvli transitions.
1653 for (MachineBasicBlock &MBB : MF)
1654 doPRE(MBB);
1656 // Phase 3 - add any vsetvli instructions needed in the block. Use the
1657 // Phase 2 information to avoid adding vsetvlis before the first vector
1658 // instruction in the block if the VL/VTYPE is satisfied by its
1659 // predecessors.
1660 for (MachineBasicBlock &MBB : MF)
1661 emitVSETVLIs(MBB);
1663 // Now that all vsetvlis are explicit, go through and do block local
1664 // DSE and peephole based demanded fields based transforms. Note that
1665 // this *must* be done outside the main dataflow so long as we allow
1666 // any cross block analysis within the dataflow. We can't have both
1667 // demanded fields based mutation and non-local analysis in the
1668 // dataflow at the same time without introducing inconsistencies.
1669 for (MachineBasicBlock &MBB : MF)
1670 doLocalPostpass(MBB);
1672 // Insert PseudoReadVL after VLEFF/VLSEGFF and replace it with the vl output
1673 // of VLEFF/VLSEGFF.
1674 for (MachineBasicBlock &MBB : MF)
1675 insertReadVL(MBB);
1677 BlockInfo.clear();
1678 return HaveVectorOp;
1681 /// Returns an instance of the Insert VSETVLI pass.
1682 FunctionPass *llvm::createRISCVInsertVSETVLIPass() {
1683 return new RISCVInsertVSETVLI();