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.
12 //
13 // This pass consists of 3 phases:
14 //
15 // Phase 1 collects how each basic block affects VL/VTYPE.
16 //
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.
20 //
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.
24 //
25 //===----------------------------------------------------------------------===//
26
27 #include "RISCV.h"
28 #include "RISCVSubtarget.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/CodeGen/LiveDebugVariables.h"
31 #include "llvm/CodeGen/LiveIntervals.h"
32 #include "llvm/CodeGen/LiveStacks.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include <queue>
35 using namespace llvm;
36
37 #define DEBUG_TYPE "riscv-insert-vsetvli"
38 #define RISCV_INSERT_VSETVLI_NAME "RISC-V Insert VSETVLI pass"
39 #define RISCV_COALESCE_VSETVLI_NAME "RISC-V Coalesce VSETVLI pass"
40
41 STATISTIC(NumInsertedVSETVL, "Number of VSETVL inst inserted");
42 STATISTIC(NumCoalescedVSETVL, "Number of VSETVL inst coalesced");
43
44 namespace {
45
46 /// Given a virtual register \p Reg, return the corresponding VNInfo for it.
47 /// This will return nullptr if the virtual register is an implicit_def or
48 /// if LiveIntervals is not available.
getVNInfoFromReg(Register Reg,const MachineInstr & MI,const LiveIntervals * LIS)49 static VNInfo *getVNInfoFromReg(Register Reg, const MachineInstr &MI,
50 const LiveIntervals *LIS) {
51 assert(Reg.isVirtual());
52 if (!LIS)
53 return nullptr;
54 auto &LI = LIS->getInterval(Reg);
55 SlotIndex SI = LIS->getSlotIndexes()->getInstructionIndex(MI);
56 return LI.getVNInfoBefore(SI);
57 }
58
getVLOpNum(const MachineInstr & MI)59 static unsigned getVLOpNum(const MachineInstr &MI) {
60 return RISCVII::getVLOpNum(MI.getDesc());
61 }
62
getSEWOpNum(const MachineInstr & MI)63 static unsigned getSEWOpNum(const MachineInstr &MI) {
64 return RISCVII::getSEWOpNum(MI.getDesc());
65 }
66
isVectorConfigInstr(const MachineInstr & MI)67 static bool isVectorConfigInstr(const MachineInstr &MI) {
68 return MI.getOpcode() == RISCV::PseudoVSETVLI ||
69 MI.getOpcode() == RISCV::PseudoVSETVLIX0 ||
70 MI.getOpcode() == RISCV::PseudoVSETIVLI;
71 }
72
73 /// Return true if this is 'vsetvli x0, x0, vtype' which preserves
74 /// VL and only sets VTYPE.
isVLPreservingConfig(const MachineInstr & MI)75 static bool isVLPreservingConfig(const MachineInstr &MI) {
76 if (MI.getOpcode() != RISCV::PseudoVSETVLIX0)
77 return false;
78 assert(RISCV::X0 == MI.getOperand(1).getReg());
79 return RISCV::X0 == MI.getOperand(0).getReg();
80 }
81
isFloatScalarMoveOrScalarSplatInstr(const MachineInstr & MI)82 static bool isFloatScalarMoveOrScalarSplatInstr(const MachineInstr &MI) {
83 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
84 default:
85 return false;
86 case RISCV::VFMV_S_F:
87 case RISCV::VFMV_V_F:
88 return true;
89 }
90 }
91
isScalarExtractInstr(const MachineInstr & MI)92 static bool isScalarExtractInstr(const MachineInstr &MI) {
93 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
94 default:
95 return false;
96 case RISCV::VMV_X_S:
97 case RISCV::VFMV_F_S:
98 return true;
99 }
100 }
101
isScalarInsertInstr(const MachineInstr & MI)102 static bool isScalarInsertInstr(const MachineInstr &MI) {
103 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
104 default:
105 return false;
106 case RISCV::VMV_S_X:
107 case RISCV::VFMV_S_F:
108 return true;
109 }
110 }
111
isScalarSplatInstr(const MachineInstr & MI)112 static bool isScalarSplatInstr(const MachineInstr &MI) {
113 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
114 default:
115 return false;
116 case RISCV::VMV_V_I:
117 case RISCV::VMV_V_X:
118 case RISCV::VFMV_V_F:
119 return true;
120 }
121 }
122
isVSlideInstr(const MachineInstr & MI)123 static bool isVSlideInstr(const MachineInstr &MI) {
124 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
125 default:
126 return false;
127 case RISCV::VSLIDEDOWN_VX:
128 case RISCV::VSLIDEDOWN_VI:
129 case RISCV::VSLIDEUP_VX:
130 case RISCV::VSLIDEUP_VI:
131 return true;
132 }
133 }
134
135 /// Get the EEW for a load or store instruction. Return std::nullopt if MI is
136 /// not a load or store which ignores SEW.
getEEWForLoadStore(const MachineInstr & MI)137 static std::optional<unsigned> getEEWForLoadStore(const MachineInstr &MI) {
138 switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
139 default:
140 return std::nullopt;
141 case RISCV::VLE8_V:
142 case RISCV::VLSE8_V:
143 case RISCV::VSE8_V:
144 case RISCV::VSSE8_V:
145 return 8;
146 case RISCV::VLE16_V:
147 case RISCV::VLSE16_V:
148 case RISCV::VSE16_V:
149 case RISCV::VSSE16_V:
150 return 16;
151 case RISCV::VLE32_V:
152 case RISCV::VLSE32_V:
153 case RISCV::VSE32_V:
154 case RISCV::VSSE32_V:
155 return 32;
156 case RISCV::VLE64_V:
157 case RISCV::VLSE64_V:
158 case RISCV::VSE64_V:
159 case RISCV::VSSE64_V:
160 return 64;
161 }
162 }
163
isNonZeroLoadImmediate(const MachineInstr & MI)164 static bool isNonZeroLoadImmediate(const MachineInstr &MI) {
165 return MI.getOpcode() == RISCV::ADDI &&
166 MI.getOperand(1).isReg() && MI.getOperand(2).isImm() &&
167 MI.getOperand(1).getReg() == RISCV::X0 &&
168 MI.getOperand(2).getImm() != 0;
169 }
170
171 /// Return true if this is an operation on mask registers. Note that
172 /// this includes both arithmetic/logical ops and load/store (vlm/vsm).
isMaskRegOp(const MachineInstr & MI)173 static bool isMaskRegOp(const MachineInstr &MI) {
174 if (!RISCVII::hasSEWOp(MI.getDesc().TSFlags))
175 return false;
176 const unsigned Log2SEW = MI.getOperand(getSEWOpNum(MI)).getImm();
177 // A Log2SEW of 0 is an operation on mask registers only.
178 return Log2SEW == 0;
179 }
180
181 /// Return true if the inactive elements in the result are entirely undefined.
182 /// Note that this is different from "agnostic" as defined by the vector
183 /// specification. Agnostic requires each lane to either be undisturbed, or
184 /// take the value -1; no other value is allowed.
hasUndefinedMergeOp(const MachineInstr & MI)185 static bool hasUndefinedMergeOp(const MachineInstr &MI) {
186
187 unsigned UseOpIdx;
188 if (!MI.isRegTiedToUseOperand(0, &UseOpIdx))
189 // If there is no passthrough operand, then the pass through
190 // lanes are undefined.
191 return true;
192
193 // All undefined passthrus should be $noreg: see
194 // RISCVDAGToDAGISel::doPeepholeNoRegPassThru
195 const MachineOperand &UseMO = MI.getOperand(UseOpIdx);
196 return UseMO.getReg() == RISCV::NoRegister || UseMO.isUndef();
197 }
198
199 /// Which subfields of VL or VTYPE have values we need to preserve?
200 struct DemandedFields {
201 // Some unknown property of VL is used. If demanded, must preserve entire
202 // value.
203 bool VLAny = false;
204 // Only zero vs non-zero is used. If demanded, can change non-zero values.
205 bool VLZeroness = false;
206 // What properties of SEW we need to preserve.
207 enum : uint8_t {
208 SEWEqual = 3, // The exact value of SEW needs to be preserved.
209 SEWGreaterThanOrEqual = 2, // SEW can be changed as long as it's greater
210 // than or equal to the original value.
211 SEWGreaterThanOrEqualAndLessThan64 =
212 1, // SEW can be changed as long as it's greater
213 // than or equal to the original value, but must be less
214 // than 64.
215 SEWNone = 0 // We don't need to preserve SEW at all.
216 } SEW = SEWNone;
217 enum : uint8_t {
218 LMULEqual = 2, // The exact value of LMUL needs to be preserved.
219 LMULLessThanOrEqualToM1 = 1, // We can use any LMUL <= M1.
220 LMULNone = 0 // We don't need to preserve LMUL at all.
221 } LMUL = LMULNone;
222 bool SEWLMULRatio = false;
223 bool TailPolicy = false;
224 bool MaskPolicy = false;
225
226 // Return true if any part of VTYPE was used
usedVTYPE__anoncddf45c50111::DemandedFields227 bool usedVTYPE() const {
228 return SEW || LMUL || SEWLMULRatio || TailPolicy || MaskPolicy;
229 }
230
231 // Return true if any property of VL was used
usedVL__anoncddf45c50111::DemandedFields232 bool usedVL() {
233 return VLAny || VLZeroness;
234 }
235
236 // Mark all VTYPE subfields and properties as demanded
demandVTYPE__anoncddf45c50111::DemandedFields237 void demandVTYPE() {
238 SEW = SEWEqual;
239 LMUL = LMULEqual;
240 SEWLMULRatio = true;
241 TailPolicy = true;
242 MaskPolicy = true;
243 }
244
245 // Mark all VL properties as demanded
demandVL__anoncddf45c50111::DemandedFields246 void demandVL() {
247 VLAny = true;
248 VLZeroness = true;
249 }
250
all__anoncddf45c50111::DemandedFields251 static DemandedFields all() {
252 DemandedFields DF;
253 DF.demandVTYPE();
254 DF.demandVL();
255 return DF;
256 }
257
258 // Make this the result of demanding both the fields in this and B.
doUnion__anoncddf45c50111::DemandedFields259 void doUnion(const DemandedFields &B) {
260 VLAny |= B.VLAny;
261 VLZeroness |= B.VLZeroness;
262 SEW = std::max(SEW, B.SEW);
263 LMUL = std::max(LMUL, B.LMUL);
264 SEWLMULRatio |= B.SEWLMULRatio;
265 TailPolicy |= B.TailPolicy;
266 MaskPolicy |= B.MaskPolicy;
267 }
268
269 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
270 /// Support for debugging, callable in GDB: V->dump()
dump__anoncddf45c50111::DemandedFields271 LLVM_DUMP_METHOD void dump() const {
272 print(dbgs());
273 dbgs() << "\n";
274 }
275
276 /// Implement operator<<.
print__anoncddf45c50111::DemandedFields277 void print(raw_ostream &OS) const {
278 OS << "{";
279 OS << "VLAny=" << VLAny << ", ";
280 OS << "VLZeroness=" << VLZeroness << ", ";
281 OS << "SEW=";
282 switch (SEW) {
283 case SEWEqual:
284 OS << "SEWEqual";
285 break;
286 case SEWGreaterThanOrEqual:
287 OS << "SEWGreaterThanOrEqual";
288 break;
289 case SEWGreaterThanOrEqualAndLessThan64:
290 OS << "SEWGreaterThanOrEqualAndLessThan64";
291 break;
292 case SEWNone:
293 OS << "SEWNone";
294 break;
295 };
296 OS << ", ";
297 OS << "LMUL=";
298 switch (LMUL) {
299 case LMULEqual:
300 OS << "LMULEqual";
301 break;
302 case LMULLessThanOrEqualToM1:
303 OS << "LMULLessThanOrEqualToM1";
304 break;
305 case LMULNone:
306 OS << "LMULNone";
307 break;
308 };
309 OS << ", ";
310 OS << "SEWLMULRatio=" << SEWLMULRatio << ", ";
311 OS << "TailPolicy=" << TailPolicy << ", ";
312 OS << "MaskPolicy=" << MaskPolicy;
313 OS << "}";
314 }
315 #endif
316 };
317
318 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
319 LLVM_ATTRIBUTE_USED
operator <<(raw_ostream & OS,const DemandedFields & DF)320 inline raw_ostream &operator<<(raw_ostream &OS, const DemandedFields &DF) {
321 DF.print(OS);
322 return OS;
323 }
324 #endif
325
isLMUL1OrSmaller(RISCVII::VLMUL LMUL)326 static bool isLMUL1OrSmaller(RISCVII::VLMUL LMUL) {
327 auto [LMul, Fractional] = RISCVVType::decodeVLMUL(LMUL);
328 return Fractional || LMul == 1;
329 }
330
331 /// Return true if moving from CurVType to NewVType is
332 /// indistinguishable from the perspective of an instruction (or set
333 /// of instructions) which use only the Used subfields and properties.
areCompatibleVTYPEs(uint64_t CurVType,uint64_t NewVType,const DemandedFields & Used)334 static bool areCompatibleVTYPEs(uint64_t CurVType, uint64_t NewVType,
335 const DemandedFields &Used) {
336 switch (Used.SEW) {
337 case DemandedFields::SEWNone:
338 break;
339 case DemandedFields::SEWEqual:
340 if (RISCVVType::getSEW(CurVType) != RISCVVType::getSEW(NewVType))
341 return false;
342 break;
343 case DemandedFields::SEWGreaterThanOrEqual:
344 if (RISCVVType::getSEW(NewVType) < RISCVVType::getSEW(CurVType))
345 return false;
346 break;
347 case DemandedFields::SEWGreaterThanOrEqualAndLessThan64:
348 if (RISCVVType::getSEW(NewVType) < RISCVVType::getSEW(CurVType) ||
349 RISCVVType::getSEW(NewVType) >= 64)
350 return false;
351 break;
352 }
353
354 switch (Used.LMUL) {
355 case DemandedFields::LMULNone:
356 break;
357 case DemandedFields::LMULEqual:
358 if (RISCVVType::getVLMUL(CurVType) != RISCVVType::getVLMUL(NewVType))
359 return false;
360 break;
361 case DemandedFields::LMULLessThanOrEqualToM1:
362 if (!isLMUL1OrSmaller(RISCVVType::getVLMUL(NewVType)))
363 return false;
364 break;
365 }
366
367 if (Used.SEWLMULRatio) {
368 auto Ratio1 = RISCVVType::getSEWLMULRatio(RISCVVType::getSEW(CurVType),
369 RISCVVType::getVLMUL(CurVType));
370 auto Ratio2 = RISCVVType::getSEWLMULRatio(RISCVVType::getSEW(NewVType),
371 RISCVVType::getVLMUL(NewVType));
372 if (Ratio1 != Ratio2)
373 return false;
374 }
375
376 if (Used.TailPolicy && RISCVVType::isTailAgnostic(CurVType) !=
377 RISCVVType::isTailAgnostic(NewVType))
378 return false;
379 if (Used.MaskPolicy && RISCVVType::isMaskAgnostic(CurVType) !=
380 RISCVVType::isMaskAgnostic(NewVType))
381 return false;
382 return true;
383 }
384
385 /// Return the fields and properties demanded by the provided instruction.
getDemanded(const MachineInstr & MI,const RISCVSubtarget * ST)386 DemandedFields getDemanded(const MachineInstr &MI, const RISCVSubtarget *ST) {
387 // This function works in coalesceVSETVLI too. We can still use the value of a
388 // SEW, VL, or Policy operand even though it might not be the exact value in
389 // the VL or VTYPE, since we only care about what the instruction originally
390 // demanded.
391
392 // Most instructions don't use any of these subfeilds.
393 DemandedFields Res;
394 // Start conservative if registers are used
395 if (MI.isCall() || MI.isInlineAsm() ||
396 MI.readsRegister(RISCV::VL, /*TRI=*/nullptr))
397 Res.demandVL();
398 if (MI.isCall() || MI.isInlineAsm() ||
399 MI.readsRegister(RISCV::VTYPE, /*TRI=*/nullptr))
400 Res.demandVTYPE();
401 // Start conservative on the unlowered form too
402 uint64_t TSFlags = MI.getDesc().TSFlags;
403 if (RISCVII::hasSEWOp(TSFlags)) {
404 Res.demandVTYPE();
405 if (RISCVII::hasVLOp(TSFlags))
406 if (const MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
407 !VLOp.isReg() || !VLOp.isUndef())
408 Res.demandVL();
409
410 // Behavior is independent of mask policy.
411 if (!RISCVII::usesMaskPolicy(TSFlags))
412 Res.MaskPolicy = false;
413 }
414
415 // Loads and stores with implicit EEW do not demand SEW or LMUL directly.
416 // They instead demand the ratio of the two which is used in computing
417 // EMUL, but which allows us the flexibility to change SEW and LMUL
418 // provided we don't change the ratio.
419 // Note: We assume that the instructions initial SEW is the EEW encoded
420 // in the opcode. This is asserted when constructing the VSETVLIInfo.
421 if (getEEWForLoadStore(MI)) {
422 Res.SEW = DemandedFields::SEWNone;
423 Res.LMUL = DemandedFields::LMULNone;
424 }
425
426 // Store instructions don't use the policy fields.
427 if (RISCVII::hasSEWOp(TSFlags) && MI.getNumExplicitDefs() == 0) {
428 Res.TailPolicy = false;
429 Res.MaskPolicy = false;
430 }
431
432 // If this is a mask reg operation, it only cares about VLMAX.
433 // TODO: Possible extensions to this logic
434 // * Probably ok if available VLMax is larger than demanded
435 // * The policy bits can probably be ignored..
436 if (isMaskRegOp(MI)) {
437 Res.SEW = DemandedFields::SEWNone;
438 Res.LMUL = DemandedFields::LMULNone;
439 }
440
441 // For vmv.s.x and vfmv.s.f, there are only two behaviors, VL = 0 and VL > 0.
442 if (isScalarInsertInstr(MI)) {
443 Res.LMUL = DemandedFields::LMULNone;
444 Res.SEWLMULRatio = false;
445 Res.VLAny = false;
446 // For vmv.s.x and vfmv.s.f, if the merge operand is *undefined*, we don't
447 // need to preserve any other bits and are thus compatible with any larger,
448 // etype and can disregard policy bits. Warning: It's tempting to try doing
449 // this for any tail agnostic operation, but we can't as TA requires
450 // tail lanes to either be the original value or -1. We are writing
451 // unknown bits to the lanes here.
452 if (hasUndefinedMergeOp(MI)) {
453 if (isFloatScalarMoveOrScalarSplatInstr(MI) && !ST->hasVInstructionsF64())
454 Res.SEW = DemandedFields::SEWGreaterThanOrEqualAndLessThan64;
455 else
456 Res.SEW = DemandedFields::SEWGreaterThanOrEqual;
457 Res.TailPolicy = false;
458 }
459 }
460
461 // vmv.x.s, and vmv.f.s are unconditional and ignore everything except SEW.
462 if (isScalarExtractInstr(MI)) {
463 assert(!RISCVII::hasVLOp(TSFlags));
464 Res.LMUL = DemandedFields::LMULNone;
465 Res.SEWLMULRatio = false;
466 Res.TailPolicy = false;
467 Res.MaskPolicy = false;
468 }
469
470 if (RISCVII::hasVLOp(MI.getDesc().TSFlags)) {
471 const MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
472 // A slidedown/slideup with an *undefined* merge op can freely clobber
473 // elements not copied from the source vector (e.g. masked off, tail, or
474 // slideup's prefix). Notes:
475 // * We can't modify SEW here since the slide amount is in units of SEW.
476 // * VL=1 is special only because we have existing support for zero vs
477 // non-zero VL. We could generalize this if we had a VL > C predicate.
478 // * The LMUL1 restriction is for machines whose latency may depend on VL.
479 // * As above, this is only legal for tail "undefined" not "agnostic".
480 if (isVSlideInstr(MI) && VLOp.isImm() && VLOp.getImm() == 1 &&
481 hasUndefinedMergeOp(MI)) {
482 Res.VLAny = false;
483 Res.VLZeroness = true;
484 Res.LMUL = DemandedFields::LMULLessThanOrEqualToM1;
485 Res.TailPolicy = false;
486 }
487
488 // A tail undefined vmv.v.i/x or vfmv.v.f with VL=1 can be treated in the
489 // same semantically as vmv.s.x. This is particularly useful since we don't
490 // have an immediate form of vmv.s.x, and thus frequently use vmv.v.i in
491 // it's place. Since a splat is non-constant time in LMUL, we do need to be
492 // careful to not increase the number of active vector registers (unlike for
493 // vmv.s.x.)
494 if (isScalarSplatInstr(MI) && VLOp.isImm() && VLOp.getImm() == 1 &&
495 hasUndefinedMergeOp(MI)) {
496 Res.LMUL = DemandedFields::LMULLessThanOrEqualToM1;
497 Res.SEWLMULRatio = false;
498 Res.VLAny = false;
499 if (isFloatScalarMoveOrScalarSplatInstr(MI) && !ST->hasVInstructionsF64())
500 Res.SEW = DemandedFields::SEWGreaterThanOrEqualAndLessThan64;
501 else
502 Res.SEW = DemandedFields::SEWGreaterThanOrEqual;
503 Res.TailPolicy = false;
504 }
505 }
506
507 return Res;
508 }
509
510 /// Defines the abstract state with which the forward dataflow models the
511 /// values of the VL and VTYPE registers after insertion.
512 class VSETVLIInfo {
513 struct AVLDef {
514 // Every AVLDef should have a VNInfo, unless we're running without
515 // LiveIntervals in which case this will be nullptr.
516 const VNInfo *ValNo;
517 Register DefReg;
518 };
519 union {
520 AVLDef AVLRegDef;
521 unsigned AVLImm;
522 };
523
524 enum : uint8_t {
525 Uninitialized,
526 AVLIsReg,
527 AVLIsImm,
528 AVLIsVLMAX,
529 Unknown, // AVL and VTYPE are fully unknown
530 } State = Uninitialized;
531
532 // Fields from VTYPE.
533 RISCVII::VLMUL VLMul = RISCVII::LMUL_1;
534 uint8_t SEW = 0;
535 uint8_t TailAgnostic : 1;
536 uint8_t MaskAgnostic : 1;
537 uint8_t SEWLMULRatioOnly : 1;
538
539 public:
VSETVLIInfo()540 VSETVLIInfo()
541 : AVLImm(0), TailAgnostic(false), MaskAgnostic(false),
542 SEWLMULRatioOnly(false) {}
543
getUnknown()544 static VSETVLIInfo getUnknown() {
545 VSETVLIInfo Info;
546 Info.setUnknown();
547 return Info;
548 }
549
isValid() const550 bool isValid() const { return State != Uninitialized; }
setUnknown()551 void setUnknown() { State = Unknown; }
isUnknown() const552 bool isUnknown() const { return State == Unknown; }
553
setAVLRegDef(const VNInfo * VNInfo,Register AVLReg)554 void setAVLRegDef(const VNInfo *VNInfo, Register AVLReg) {
555 assert(AVLReg.isVirtual());
556 AVLRegDef.ValNo = VNInfo;
557 AVLRegDef.DefReg = AVLReg;
558 State = AVLIsReg;
559 }
560
setAVLImm(unsigned Imm)561 void setAVLImm(unsigned Imm) {
562 AVLImm = Imm;
563 State = AVLIsImm;
564 }
565
setAVLVLMAX()566 void setAVLVLMAX() { State = AVLIsVLMAX; }
567
hasAVLImm() const568 bool hasAVLImm() const { return State == AVLIsImm; }
hasAVLReg() const569 bool hasAVLReg() const { return State == AVLIsReg; }
hasAVLVLMAX() const570 bool hasAVLVLMAX() const { return State == AVLIsVLMAX; }
getAVLReg() const571 Register getAVLReg() const {
572 assert(hasAVLReg() && AVLRegDef.DefReg.isVirtual());
573 return AVLRegDef.DefReg;
574 }
getAVLImm() const575 unsigned getAVLImm() const {
576 assert(hasAVLImm());
577 return AVLImm;
578 }
getAVLVNInfo() const579 const VNInfo *getAVLVNInfo() const {
580 assert(hasAVLReg());
581 return AVLRegDef.ValNo;
582 }
583 // Most AVLIsReg infos will have a single defining MachineInstr, unless it was
584 // a PHI node. In that case getAVLVNInfo()->def will point to the block
585 // boundary slot and this will return nullptr. If LiveIntervals isn't
586 // available, nullptr is also returned.
getAVLDefMI(const LiveIntervals * LIS) const587 const MachineInstr *getAVLDefMI(const LiveIntervals *LIS) const {
588 assert(hasAVLReg());
589 if (!LIS || getAVLVNInfo()->isPHIDef())
590 return nullptr;
591 auto *MI = LIS->getInstructionFromIndex(getAVLVNInfo()->def);
592 assert(MI);
593 return MI;
594 }
595
setAVL(VSETVLIInfo Info)596 void setAVL(VSETVLIInfo Info) {
597 assert(Info.isValid());
598 if (Info.isUnknown())
599 setUnknown();
600 else if (Info.hasAVLReg())
601 setAVLRegDef(Info.getAVLVNInfo(), Info.getAVLReg());
602 else if (Info.hasAVLVLMAX())
603 setAVLVLMAX();
604 else {
605 assert(Info.hasAVLImm());
606 setAVLImm(Info.getAVLImm());
607 }
608 }
609
getSEW() const610 unsigned getSEW() const { return SEW; }
getVLMUL() const611 RISCVII::VLMUL getVLMUL() const { return VLMul; }
getTailAgnostic() const612 bool getTailAgnostic() const { return TailAgnostic; }
getMaskAgnostic() const613 bool getMaskAgnostic() const { return MaskAgnostic; }
614
hasNonZeroAVL(const LiveIntervals * LIS) const615 bool hasNonZeroAVL(const LiveIntervals *LIS) const {
616 if (hasAVLImm())
617 return getAVLImm() > 0;
618 if (hasAVLReg()) {
619 if (auto *DefMI = getAVLDefMI(LIS))
620 return isNonZeroLoadImmediate(*DefMI);
621 }
622 if (hasAVLVLMAX())
623 return true;
624 return false;
625 }
626
hasEquallyZeroAVL(const VSETVLIInfo & Other,const LiveIntervals * LIS) const627 bool hasEquallyZeroAVL(const VSETVLIInfo &Other,
628 const LiveIntervals *LIS) const {
629 if (hasSameAVL(Other))
630 return true;
631 return (hasNonZeroAVL(LIS) && Other.hasNonZeroAVL(LIS));
632 }
633
hasSameAVLLatticeValue(const VSETVLIInfo & Other) const634 bool hasSameAVLLatticeValue(const VSETVLIInfo &Other) const {
635 if (hasAVLReg() && Other.hasAVLReg()) {
636 assert(!getAVLVNInfo() == !Other.getAVLVNInfo() &&
637 "we either have intervals or we don't");
638 if (!getAVLVNInfo())
639 return getAVLReg() == Other.getAVLReg();
640 return getAVLVNInfo()->id == Other.getAVLVNInfo()->id &&
641 getAVLReg() == Other.getAVLReg();
642 }
643
644 if (hasAVLImm() && Other.hasAVLImm())
645 return getAVLImm() == Other.getAVLImm();
646
647 if (hasAVLVLMAX())
648 return Other.hasAVLVLMAX() && hasSameVLMAX(Other);
649
650 return false;
651 }
652
653 // Return true if the two lattice values are guaranteed to have
654 // the same AVL value at runtime.
hasSameAVL(const VSETVLIInfo & Other) const655 bool hasSameAVL(const VSETVLIInfo &Other) const {
656 // Without LiveIntervals, we don't know which instruction defines a
657 // register. Since a register may be redefined, this means all AVLIsReg
658 // states must be treated as possibly distinct.
659 if (hasAVLReg() && Other.hasAVLReg()) {
660 assert(!getAVLVNInfo() == !Other.getAVLVNInfo() &&
661 "we either have intervals or we don't");
662 if (!getAVLVNInfo())
663 return false;
664 }
665 return hasSameAVLLatticeValue(Other);
666 }
667
setVTYPE(unsigned VType)668 void setVTYPE(unsigned VType) {
669 assert(isValid() && !isUnknown() &&
670 "Can't set VTYPE for uninitialized or unknown");
671 VLMul = RISCVVType::getVLMUL(VType);
672 SEW = RISCVVType::getSEW(VType);
673 TailAgnostic = RISCVVType::isTailAgnostic(VType);
674 MaskAgnostic = RISCVVType::isMaskAgnostic(VType);
675 }
setVTYPE(RISCVII::VLMUL L,unsigned S,bool TA,bool MA)676 void setVTYPE(RISCVII::VLMUL L, unsigned S, bool TA, bool MA) {
677 assert(isValid() && !isUnknown() &&
678 "Can't set VTYPE for uninitialized or unknown");
679 VLMul = L;
680 SEW = S;
681 TailAgnostic = TA;
682 MaskAgnostic = MA;
683 }
684
setVLMul(RISCVII::VLMUL VLMul)685 void setVLMul(RISCVII::VLMUL VLMul) { this->VLMul = VLMul; }
686
encodeVTYPE() const687 unsigned encodeVTYPE() const {
688 assert(isValid() && !isUnknown() && !SEWLMULRatioOnly &&
689 "Can't encode VTYPE for uninitialized or unknown");
690 return RISCVVType::encodeVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
691 }
692
hasSEWLMULRatioOnly() const693 bool hasSEWLMULRatioOnly() const { return SEWLMULRatioOnly; }
694
hasSameVTYPE(const VSETVLIInfo & Other) const695 bool hasSameVTYPE(const VSETVLIInfo &Other) const {
696 assert(isValid() && Other.isValid() &&
697 "Can't compare invalid VSETVLIInfos");
698 assert(!isUnknown() && !Other.isUnknown() &&
699 "Can't compare VTYPE in unknown state");
700 assert(!SEWLMULRatioOnly && !Other.SEWLMULRatioOnly &&
701 "Can't compare when only LMUL/SEW ratio is valid.");
702 return std::tie(VLMul, SEW, TailAgnostic, MaskAgnostic) ==
703 std::tie(Other.VLMul, Other.SEW, Other.TailAgnostic,
704 Other.MaskAgnostic);
705 }
706
getSEWLMULRatio() const707 unsigned getSEWLMULRatio() const {
708 assert(isValid() && !isUnknown() &&
709 "Can't use VTYPE for uninitialized or unknown");
710 return RISCVVType::getSEWLMULRatio(SEW, VLMul);
711 }
712
713 // Check if the VTYPE for these two VSETVLIInfos produce the same VLMAX.
714 // Note that having the same VLMAX ensures that both share the same
715 // function from AVL to VL; that is, they must produce the same VL value
716 // for any given AVL value.
hasSameVLMAX(const VSETVLIInfo & Other) const717 bool hasSameVLMAX(const VSETVLIInfo &Other) const {
718 assert(isValid() && Other.isValid() &&
719 "Can't compare invalid VSETVLIInfos");
720 assert(!isUnknown() && !Other.isUnknown() &&
721 "Can't compare VTYPE in unknown state");
722 return getSEWLMULRatio() == Other.getSEWLMULRatio();
723 }
724
hasCompatibleVTYPE(const DemandedFields & Used,const VSETVLIInfo & Require) const725 bool hasCompatibleVTYPE(const DemandedFields &Used,
726 const VSETVLIInfo &Require) const {
727 return areCompatibleVTYPEs(Require.encodeVTYPE(), encodeVTYPE(), Used);
728 }
729
730 // Determine whether the vector instructions requirements represented by
731 // Require are compatible with the previous vsetvli instruction represented
732 // by this. MI is the instruction whose requirements we're considering.
isCompatible(const DemandedFields & Used,const VSETVLIInfo & Require,const LiveIntervals * LIS) const733 bool isCompatible(const DemandedFields &Used, const VSETVLIInfo &Require,
734 const LiveIntervals *LIS) const {
735 assert(isValid() && Require.isValid() &&
736 "Can't compare invalid VSETVLIInfos");
737 // Nothing is compatible with Unknown.
738 if (isUnknown() || Require.isUnknown())
739 return false;
740
741 // If only our VLMAX ratio is valid, then this isn't compatible.
742 if (SEWLMULRatioOnly || Require.SEWLMULRatioOnly)
743 return false;
744
745 if (Used.VLAny && !(hasSameAVL(Require) && hasSameVLMAX(Require)))
746 return false;
747
748 if (Used.VLZeroness && !hasEquallyZeroAVL(Require, LIS))
749 return false;
750
751 return hasCompatibleVTYPE(Used, Require);
752 }
753
operator ==(const VSETVLIInfo & Other) const754 bool operator==(const VSETVLIInfo &Other) const {
755 // Uninitialized is only equal to another Uninitialized.
756 if (!isValid())
757 return !Other.isValid();
758 if (!Other.isValid())
759 return !isValid();
760
761 // Unknown is only equal to another Unknown.
762 if (isUnknown())
763 return Other.isUnknown();
764 if (Other.isUnknown())
765 return isUnknown();
766
767 if (!hasSameAVLLatticeValue(Other))
768 return false;
769
770 // If the SEWLMULRatioOnly bits are different, then they aren't equal.
771 if (SEWLMULRatioOnly != Other.SEWLMULRatioOnly)
772 return false;
773
774 // If only the VLMAX is valid, check that it is the same.
775 if (SEWLMULRatioOnly)
776 return hasSameVLMAX(Other);
777
778 // If the full VTYPE is valid, check that it is the same.
779 return hasSameVTYPE(Other);
780 }
781
operator !=(const VSETVLIInfo & Other) const782 bool operator!=(const VSETVLIInfo &Other) const {
783 return !(*this == Other);
784 }
785
786 // Calculate the VSETVLIInfo visible to a block assuming this and Other are
787 // both predecessors.
intersect(const VSETVLIInfo & Other) const788 VSETVLIInfo intersect(const VSETVLIInfo &Other) const {
789 // If the new value isn't valid, ignore it.
790 if (!Other.isValid())
791 return *this;
792
793 // If this value isn't valid, this must be the first predecessor, use it.
794 if (!isValid())
795 return Other;
796
797 // If either is unknown, the result is unknown.
798 if (isUnknown() || Other.isUnknown())
799 return VSETVLIInfo::getUnknown();
800
801 // If we have an exact, match return this.
802 if (*this == Other)
803 return *this;
804
805 // Not an exact match, but maybe the AVL and VLMAX are the same. If so,
806 // return an SEW/LMUL ratio only value.
807 if (hasSameAVL(Other) && hasSameVLMAX(Other)) {
808 VSETVLIInfo MergeInfo = *this;
809 MergeInfo.SEWLMULRatioOnly = true;
810 return MergeInfo;
811 }
812
813 // Otherwise the result is unknown.
814 return VSETVLIInfo::getUnknown();
815 }
816
817 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
818 /// Support for debugging, callable in GDB: V->dump()
dump() const819 LLVM_DUMP_METHOD void dump() const {
820 print(dbgs());
821 dbgs() << "\n";
822 }
823
824 /// Implement operator<<.
825 /// @{
print(raw_ostream & OS) const826 void print(raw_ostream &OS) const {
827 OS << "{";
828 if (!isValid())
829 OS << "Uninitialized";
830 if (isUnknown())
831 OS << "unknown";
832 if (hasAVLReg())
833 OS << "AVLReg=" << llvm::printReg(getAVLReg());
834 if (hasAVLImm())
835 OS << "AVLImm=" << (unsigned)AVLImm;
836 if (hasAVLVLMAX())
837 OS << "AVLVLMAX";
838 OS << ", "
839 << "VLMul=" << (unsigned)VLMul << ", "
840 << "SEW=" << (unsigned)SEW << ", "
841 << "TailAgnostic=" << (bool)TailAgnostic << ", "
842 << "MaskAgnostic=" << (bool)MaskAgnostic << ", "
843 << "SEWLMULRatioOnly=" << (bool)SEWLMULRatioOnly << "}";
844 }
845 #endif
846 };
847
848 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
849 LLVM_ATTRIBUTE_USED
operator <<(raw_ostream & OS,const VSETVLIInfo & V)850 inline raw_ostream &operator<<(raw_ostream &OS, const VSETVLIInfo &V) {
851 V.print(OS);
852 return OS;
853 }
854 #endif
855
856 struct BlockData {
857 // The VSETVLIInfo that represents the VL/VTYPE settings on exit from this
858 // block. Calculated in Phase 2.
859 VSETVLIInfo Exit;
860
861 // The VSETVLIInfo that represents the VL/VTYPE settings from all predecessor
862 // blocks. Calculated in Phase 2, and used by Phase 3.
863 VSETVLIInfo Pred;
864
865 // Keeps track of whether the block is already in the queue.
866 bool InQueue = false;
867
868 BlockData() = default;
869 };
870
871 class RISCVInsertVSETVLI : public MachineFunctionPass {
872 const RISCVSubtarget *ST;
873 const TargetInstrInfo *TII;
874 MachineRegisterInfo *MRI;
875 // Possibly null!
876 LiveIntervals *LIS;
877
878 std::vector<BlockData> BlockInfo;
879 std::queue<const MachineBasicBlock *> WorkList;
880
881 public:
882 static char ID;
883
RISCVInsertVSETVLI()884 RISCVInsertVSETVLI() : MachineFunctionPass(ID) {}
885 bool runOnMachineFunction(MachineFunction &MF) override;
886
getAnalysisUsage(AnalysisUsage & AU) const887 void getAnalysisUsage(AnalysisUsage &AU) const override {
888 AU.setPreservesCFG();
889
890 AU.addUsedIfAvailable<LiveIntervalsWrapperPass>();
891 AU.addPreserved<LiveIntervalsWrapperPass>();
892 AU.addPreserved<SlotIndexesWrapperPass>();
893 AU.addPreserved<LiveDebugVariables>();
894 AU.addPreserved<LiveStacks>();
895
896 MachineFunctionPass::getAnalysisUsage(AU);
897 }
898
getPassName() const899 StringRef getPassName() const override { return RISCV_INSERT_VSETVLI_NAME; }
900
901 private:
902 bool needVSETVLI(const DemandedFields &Used, const VSETVLIInfo &Require,
903 const VSETVLIInfo &CurInfo) const;
904 bool needVSETVLIPHI(const VSETVLIInfo &Require,
905 const MachineBasicBlock &MBB) const;
906 void insertVSETVLI(MachineBasicBlock &MBB,
907 MachineBasicBlock::iterator InsertPt, DebugLoc DL,
908 const VSETVLIInfo &Info, const VSETVLIInfo &PrevInfo);
909
910 void transferBefore(VSETVLIInfo &Info, const MachineInstr &MI) const;
911 void transferAfter(VSETVLIInfo &Info, const MachineInstr &MI) const;
912 bool computeVLVTYPEChanges(const MachineBasicBlock &MBB,
913 VSETVLIInfo &Info) const;
914 void computeIncomingVLVTYPE(const MachineBasicBlock &MBB);
915 void emitVSETVLIs(MachineBasicBlock &MBB);
916 void doPRE(MachineBasicBlock &MBB);
917 void insertReadVL(MachineBasicBlock &MBB);
918
919 bool canMutatePriorConfig(const MachineInstr &PrevMI, const MachineInstr &MI,
920 const DemandedFields &Used) const;
921 void coalesceVSETVLIs(MachineBasicBlock &MBB) const;
922
923 VSETVLIInfo getInfoForVSETVLI(const MachineInstr &MI) const;
924 VSETVLIInfo computeInfoForInstr(const MachineInstr &MI) const;
925 void forwardVSETVLIAVL(VSETVLIInfo &Info) const;
926 };
927
928 } // end anonymous namespace
929
930 char RISCVInsertVSETVLI::ID = 0;
931 char &llvm::RISCVInsertVSETVLIID = RISCVInsertVSETVLI::ID;
932
INITIALIZE_PASS(RISCVInsertVSETVLI,DEBUG_TYPE,RISCV_INSERT_VSETVLI_NAME,false,false)933 INITIALIZE_PASS(RISCVInsertVSETVLI, DEBUG_TYPE, RISCV_INSERT_VSETVLI_NAME,
934 false, false)
935
936 // If the AVL is defined by a vsetvli's output vl with the same VLMAX, we can
937 // replace the AVL operand with the AVL of the defining vsetvli. E.g.
938 //
939 // %vl = PseudoVSETVLI %avl:gpr, SEW=32, LMUL=M1
940 // $x0 = PseudoVSETVLI %vl:gpr, SEW=32, LMUL=M1
941 // ->
942 // %vl = PseudoVSETVLI %avl:gpr, SEW=32, LMUL=M1
943 // $x0 = PseudoVSETVLI %avl:gpr, SEW=32, LMUL=M1
944 void RISCVInsertVSETVLI::forwardVSETVLIAVL(VSETVLIInfo &Info) const {
945 if (!Info.hasAVLReg())
946 return;
947 const MachineInstr *DefMI = Info.getAVLDefMI(LIS);
948 if (!DefMI || !isVectorConfigInstr(*DefMI))
949 return;
950 VSETVLIInfo DefInstrInfo = getInfoForVSETVLI(*DefMI);
951 if (!DefInstrInfo.hasSameVLMAX(Info))
952 return;
953 Info.setAVL(DefInstrInfo);
954 }
955
956 // Return a VSETVLIInfo representing the changes made by this VSETVLI or
957 // VSETIVLI instruction.
958 VSETVLIInfo
getInfoForVSETVLI(const MachineInstr & MI) const959 RISCVInsertVSETVLI::getInfoForVSETVLI(const MachineInstr &MI) const {
960 VSETVLIInfo NewInfo;
961 if (MI.getOpcode() == RISCV::PseudoVSETIVLI) {
962 NewInfo.setAVLImm(MI.getOperand(1).getImm());
963 } else {
964 assert(MI.getOpcode() == RISCV::PseudoVSETVLI ||
965 MI.getOpcode() == RISCV::PseudoVSETVLIX0);
966 Register AVLReg = MI.getOperand(1).getReg();
967 assert((AVLReg != RISCV::X0 || MI.getOperand(0).getReg() != RISCV::X0) &&
968 "Can't handle X0, X0 vsetvli yet");
969 if (AVLReg == RISCV::X0)
970 NewInfo.setAVLVLMAX();
971 else if (MI.getOperand(1).isUndef())
972 // Otherwise use an AVL of 1 to avoid depending on previous vl.
973 NewInfo.setAVLImm(1);
974 else {
975 VNInfo *VNI = getVNInfoFromReg(AVLReg, MI, LIS);
976 NewInfo.setAVLRegDef(VNI, AVLReg);
977 }
978 }
979 NewInfo.setVTYPE(MI.getOperand(2).getImm());
980
981 forwardVSETVLIAVL(NewInfo);
982
983 return NewInfo;
984 }
985
computeVLMAX(unsigned VLEN,unsigned SEW,RISCVII::VLMUL VLMul)986 static unsigned computeVLMAX(unsigned VLEN, unsigned SEW,
987 RISCVII::VLMUL VLMul) {
988 auto [LMul, Fractional] = RISCVVType::decodeVLMUL(VLMul);
989 if (Fractional)
990 VLEN = VLEN / LMul;
991 else
992 VLEN = VLEN * LMul;
993 return VLEN/SEW;
994 }
995
996 VSETVLIInfo
computeInfoForInstr(const MachineInstr & MI) const997 RISCVInsertVSETVLI::computeInfoForInstr(const MachineInstr &MI) const {
998 VSETVLIInfo InstrInfo;
999 const uint64_t TSFlags = MI.getDesc().TSFlags;
1000
1001 bool TailAgnostic = true;
1002 bool MaskAgnostic = true;
1003 if (!hasUndefinedMergeOp(MI)) {
1004 // Start with undisturbed.
1005 TailAgnostic = false;
1006 MaskAgnostic = false;
1007
1008 // If there is a policy operand, use it.
1009 if (RISCVII::hasVecPolicyOp(TSFlags)) {
1010 const MachineOperand &Op = MI.getOperand(MI.getNumExplicitOperands() - 1);
1011 uint64_t Policy = Op.getImm();
1012 assert(Policy <= (RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC) &&
1013 "Invalid Policy Value");
1014 TailAgnostic = Policy & RISCVII::TAIL_AGNOSTIC;
1015 MaskAgnostic = Policy & RISCVII::MASK_AGNOSTIC;
1016 }
1017
1018 // Some pseudo instructions force a tail agnostic policy despite having a
1019 // tied def.
1020 if (RISCVII::doesForceTailAgnostic(TSFlags))
1021 TailAgnostic = true;
1022
1023 if (!RISCVII::usesMaskPolicy(TSFlags))
1024 MaskAgnostic = true;
1025 }
1026
1027 RISCVII::VLMUL VLMul = RISCVII::getLMul(TSFlags);
1028
1029 unsigned Log2SEW = MI.getOperand(getSEWOpNum(MI)).getImm();
1030 // A Log2SEW of 0 is an operation on mask registers only.
1031 unsigned SEW = Log2SEW ? 1 << Log2SEW : 8;
1032 assert(RISCVVType::isValidSEW(SEW) && "Unexpected SEW");
1033
1034 if (RISCVII::hasVLOp(TSFlags)) {
1035 const MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
1036 if (VLOp.isImm()) {
1037 int64_t Imm = VLOp.getImm();
1038 // Conver the VLMax sentintel to X0 register.
1039 if (Imm == RISCV::VLMaxSentinel) {
1040 // If we know the exact VLEN, see if we can use the constant encoding
1041 // for the VLMAX instead. This reduces register pressure slightly.
1042 const unsigned VLMAX = computeVLMAX(ST->getRealMaxVLen(), SEW, VLMul);
1043 if (ST->getRealMinVLen() == ST->getRealMaxVLen() && VLMAX <= 31)
1044 InstrInfo.setAVLImm(VLMAX);
1045 else
1046 InstrInfo.setAVLVLMAX();
1047 }
1048 else
1049 InstrInfo.setAVLImm(Imm);
1050 } else if (VLOp.isUndef()) {
1051 // Otherwise use an AVL of 1 to avoid depending on previous vl.
1052 InstrInfo.setAVLImm(1);
1053 } else {
1054 VNInfo *VNI = getVNInfoFromReg(VLOp.getReg(), MI, LIS);
1055 InstrInfo.setAVLRegDef(VNI, VLOp.getReg());
1056 }
1057 } else {
1058 assert(isScalarExtractInstr(MI));
1059 // Pick a random value for state tracking purposes, will be ignored via
1060 // the demanded fields mechanism
1061 InstrInfo.setAVLImm(1);
1062 }
1063 #ifndef NDEBUG
1064 if (std::optional<unsigned> EEW = getEEWForLoadStore(MI)) {
1065 assert(SEW == EEW && "Initial SEW doesn't match expected EEW");
1066 }
1067 #endif
1068 InstrInfo.setVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
1069
1070 forwardVSETVLIAVL(InstrInfo);
1071
1072 return InstrInfo;
1073 }
1074
insertVSETVLI(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertPt,DebugLoc DL,const VSETVLIInfo & Info,const VSETVLIInfo & PrevInfo)1075 void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
1076 MachineBasicBlock::iterator InsertPt, DebugLoc DL,
1077 const VSETVLIInfo &Info, const VSETVLIInfo &PrevInfo) {
1078
1079 ++NumInsertedVSETVL;
1080 if (PrevInfo.isValid() && !PrevInfo.isUnknown()) {
1081 // Use X0, X0 form if the AVL is the same and the SEW+LMUL gives the same
1082 // VLMAX.
1083 if (Info.hasSameAVL(PrevInfo) && Info.hasSameVLMAX(PrevInfo)) {
1084 auto MI = BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLIX0))
1085 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
1086 .addReg(RISCV::X0, RegState::Kill)
1087 .addImm(Info.encodeVTYPE())
1088 .addReg(RISCV::VL, RegState::Implicit);
1089 if (LIS)
1090 LIS->InsertMachineInstrInMaps(*MI);
1091 return;
1092 }
1093
1094 // If our AVL is a virtual register, it might be defined by a VSET(I)VLI. If
1095 // it has the same VLMAX we want and the last VL/VTYPE we observed is the
1096 // same, we can use the X0, X0 form.
1097 if (Info.hasSameVLMAX(PrevInfo) && Info.hasAVLReg()) {
1098 if (const MachineInstr *DefMI = Info.getAVLDefMI(LIS);
1099 DefMI && isVectorConfigInstr(*DefMI)) {
1100 VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
1101 if (DefInfo.hasSameAVL(PrevInfo) && DefInfo.hasSameVLMAX(PrevInfo)) {
1102 auto MI = BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLIX0))
1103 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
1104 .addReg(RISCV::X0, RegState::Kill)
1105 .addImm(Info.encodeVTYPE())
1106 .addReg(RISCV::VL, RegState::Implicit);
1107 if (LIS)
1108 LIS->InsertMachineInstrInMaps(*MI);
1109 return;
1110 }
1111 }
1112 }
1113 }
1114
1115 if (Info.hasAVLImm()) {
1116 auto MI = BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETIVLI))
1117 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
1118 .addImm(Info.getAVLImm())
1119 .addImm(Info.encodeVTYPE());
1120 if (LIS)
1121 LIS->InsertMachineInstrInMaps(*MI);
1122 return;
1123 }
1124
1125 if (Info.hasAVLVLMAX()) {
1126 Register DestReg = MRI->createVirtualRegister(&RISCV::GPRRegClass);
1127 auto MI = BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLIX0))
1128 .addReg(DestReg, RegState::Define | RegState::Dead)
1129 .addReg(RISCV::X0, RegState::Kill)
1130 .addImm(Info.encodeVTYPE());
1131 if (LIS) {
1132 LIS->InsertMachineInstrInMaps(*MI);
1133 LIS->createAndComputeVirtRegInterval(DestReg);
1134 }
1135 return;
1136 }
1137
1138 Register AVLReg = Info.getAVLReg();
1139 MRI->constrainRegClass(AVLReg, &RISCV::GPRNoX0RegClass);
1140 auto MI = BuildMI(MBB, InsertPt, DL, TII->get(RISCV::PseudoVSETVLI))
1141 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
1142 .addReg(AVLReg)
1143 .addImm(Info.encodeVTYPE());
1144 if (LIS) {
1145 LIS->InsertMachineInstrInMaps(*MI);
1146 LiveInterval &LI = LIS->getInterval(AVLReg);
1147 SlotIndex SI = LIS->getInstructionIndex(*MI).getRegSlot();
1148 // If the AVL value isn't live at MI, do a quick check to see if it's easily
1149 // extendable. Otherwise, we need to copy it.
1150 if (LI.getVNInfoBefore(SI) != Info.getAVLVNInfo()) {
1151 if (!LI.liveAt(SI) && LI.containsOneValue())
1152 LIS->extendToIndices(LI, SI);
1153 else {
1154 Register AVLCopyReg =
1155 MRI->createVirtualRegister(&RISCV::GPRNoX0RegClass);
1156 MachineBasicBlock::iterator II;
1157 if (Info.getAVLVNInfo()->isPHIDef())
1158 II = LIS->getMBBFromIndex(Info.getAVLVNInfo()->def)->getFirstNonPHI();
1159 else {
1160 II = LIS->getInstructionFromIndex(Info.getAVLVNInfo()->def);
1161 II = std::next(II);
1162 }
1163 assert(II.isValid());
1164 auto AVLCopy =
1165 BuildMI(*II->getParent(), II, DL, TII->get(RISCV::COPY), AVLCopyReg)
1166 .addReg(AVLReg);
1167 LIS->InsertMachineInstrInMaps(*AVLCopy);
1168 MI->getOperand(1).setReg(AVLCopyReg);
1169 LIS->createAndComputeVirtRegInterval(AVLCopyReg);
1170 }
1171 }
1172 }
1173 }
1174
1175 /// Return true if a VSETVLI is required to transition from CurInfo to Require
1176 /// given a set of DemandedFields \p Used.
needVSETVLI(const DemandedFields & Used,const VSETVLIInfo & Require,const VSETVLIInfo & CurInfo) const1177 bool RISCVInsertVSETVLI::needVSETVLI(const DemandedFields &Used,
1178 const VSETVLIInfo &Require,
1179 const VSETVLIInfo &CurInfo) const {
1180 if (!CurInfo.isValid() || CurInfo.isUnknown() || CurInfo.hasSEWLMULRatioOnly())
1181 return true;
1182
1183 if (CurInfo.isCompatible(Used, Require, LIS))
1184 return false;
1185
1186 return true;
1187 }
1188
1189 // If we don't use LMUL or the SEW/LMUL ratio, then adjust LMUL so that we
1190 // maintain the SEW/LMUL ratio. This allows us to eliminate VL toggles in more
1191 // places.
adjustIncoming(VSETVLIInfo PrevInfo,VSETVLIInfo NewInfo,DemandedFields & Demanded)1192 static VSETVLIInfo adjustIncoming(VSETVLIInfo PrevInfo, VSETVLIInfo NewInfo,
1193 DemandedFields &Demanded) {
1194 VSETVLIInfo Info = NewInfo;
1195
1196 if (!Demanded.LMUL && !Demanded.SEWLMULRatio && PrevInfo.isValid() &&
1197 !PrevInfo.isUnknown()) {
1198 if (auto NewVLMul = RISCVVType::getSameRatioLMUL(
1199 PrevInfo.getSEW(), PrevInfo.getVLMUL(), Info.getSEW()))
1200 Info.setVLMul(*NewVLMul);
1201 Demanded.LMUL = DemandedFields::LMULEqual;
1202 }
1203
1204 return Info;
1205 }
1206
1207 // Given an incoming state reaching MI, minimally modifies that state so that it
1208 // is compatible with MI. The resulting state is guaranteed to be semantically
1209 // legal for MI, but may not be the state requested by MI.
transferBefore(VSETVLIInfo & Info,const MachineInstr & MI) const1210 void RISCVInsertVSETVLI::transferBefore(VSETVLIInfo &Info,
1211 const MachineInstr &MI) const {
1212 if (!RISCVII::hasSEWOp(MI.getDesc().TSFlags))
1213 return;
1214
1215 DemandedFields Demanded = getDemanded(MI, ST);
1216
1217 const VSETVLIInfo NewInfo = computeInfoForInstr(MI);
1218 assert(NewInfo.isValid() && !NewInfo.isUnknown());
1219 if (Info.isValid() && !needVSETVLI(Demanded, NewInfo, Info))
1220 return;
1221
1222 const VSETVLIInfo PrevInfo = Info;
1223 if (!Info.isValid() || Info.isUnknown())
1224 Info = NewInfo;
1225
1226 const VSETVLIInfo IncomingInfo = adjustIncoming(PrevInfo, NewInfo, Demanded);
1227
1228 // If MI only demands that VL has the same zeroness, we only need to set the
1229 // AVL if the zeroness differs. This removes a vsetvli entirely if the types
1230 // match or allows use of cheaper avl preserving variant if VLMAX doesn't
1231 // change. If VLMAX might change, we couldn't use the 'vsetvli x0, x0, vtype"
1232 // variant, so we avoid the transform to prevent extending live range of an
1233 // avl register operand.
1234 // TODO: We can probably relax this for immediates.
1235 bool EquallyZero = IncomingInfo.hasEquallyZeroAVL(PrevInfo, LIS) &&
1236 IncomingInfo.hasSameVLMAX(PrevInfo);
1237 if (Demanded.VLAny || (Demanded.VLZeroness && !EquallyZero))
1238 Info.setAVL(IncomingInfo);
1239
1240 Info.setVTYPE(
1241 ((Demanded.LMUL || Demanded.SEWLMULRatio) ? IncomingInfo : Info)
1242 .getVLMUL(),
1243 ((Demanded.SEW || Demanded.SEWLMULRatio) ? IncomingInfo : Info).getSEW(),
1244 // Prefer tail/mask agnostic since it can be relaxed to undisturbed later
1245 // if needed.
1246 (Demanded.TailPolicy ? IncomingInfo : Info).getTailAgnostic() ||
1247 IncomingInfo.getTailAgnostic(),
1248 (Demanded.MaskPolicy ? IncomingInfo : Info).getMaskAgnostic() ||
1249 IncomingInfo.getMaskAgnostic());
1250
1251 // If we only knew the sew/lmul ratio previously, replace the VTYPE but keep
1252 // the AVL.
1253 if (Info.hasSEWLMULRatioOnly()) {
1254 VSETVLIInfo RatiolessInfo = IncomingInfo;
1255 RatiolessInfo.setAVL(Info);
1256 Info = RatiolessInfo;
1257 }
1258 }
1259
1260 // Given a state with which we evaluated MI (see transferBefore above for why
1261 // this might be different that the state MI requested), modify the state to
1262 // reflect the changes MI might make.
transferAfter(VSETVLIInfo & Info,const MachineInstr & MI) const1263 void RISCVInsertVSETVLI::transferAfter(VSETVLIInfo &Info,
1264 const MachineInstr &MI) const {
1265 if (isVectorConfigInstr(MI)) {
1266 Info = getInfoForVSETVLI(MI);
1267 return;
1268 }
1269
1270 if (RISCV::isFaultFirstLoad(MI)) {
1271 // Update AVL to vl-output of the fault first load.
1272 assert(MI.getOperand(1).getReg().isVirtual());
1273 if (LIS) {
1274 auto &LI = LIS->getInterval(MI.getOperand(1).getReg());
1275 SlotIndex SI =
1276 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot();
1277 VNInfo *VNI = LI.getVNInfoAt(SI);
1278 Info.setAVLRegDef(VNI, MI.getOperand(1).getReg());
1279 } else
1280 Info.setAVLRegDef(nullptr, MI.getOperand(1).getReg());
1281 return;
1282 }
1283
1284 // If this is something that updates VL/VTYPE that we don't know about, set
1285 // the state to unknown.
1286 if (MI.isCall() || MI.isInlineAsm() ||
1287 MI.modifiesRegister(RISCV::VL, /*TRI=*/nullptr) ||
1288 MI.modifiesRegister(RISCV::VTYPE, /*TRI=*/nullptr))
1289 Info = VSETVLIInfo::getUnknown();
1290 }
1291
computeVLVTYPEChanges(const MachineBasicBlock & MBB,VSETVLIInfo & Info) const1292 bool RISCVInsertVSETVLI::computeVLVTYPEChanges(const MachineBasicBlock &MBB,
1293 VSETVLIInfo &Info) const {
1294 bool HadVectorOp = false;
1295
1296 Info = BlockInfo[MBB.getNumber()].Pred;
1297 for (const MachineInstr &MI : MBB) {
1298 transferBefore(Info, MI);
1299
1300 if (isVectorConfigInstr(MI) || RISCVII::hasSEWOp(MI.getDesc().TSFlags))
1301 HadVectorOp = true;
1302
1303 transferAfter(Info, MI);
1304 }
1305
1306 return HadVectorOp;
1307 }
1308
computeIncomingVLVTYPE(const MachineBasicBlock & MBB)1309 void RISCVInsertVSETVLI::computeIncomingVLVTYPE(const MachineBasicBlock &MBB) {
1310
1311 BlockData &BBInfo = BlockInfo[MBB.getNumber()];
1312
1313 BBInfo.InQueue = false;
1314
1315 // Start with the previous entry so that we keep the most conservative state
1316 // we have ever found.
1317 VSETVLIInfo InInfo = BBInfo.Pred;
1318 if (MBB.pred_empty()) {
1319 // There are no predecessors, so use the default starting status.
1320 InInfo.setUnknown();
1321 } else {
1322 for (MachineBasicBlock *P : MBB.predecessors())
1323 InInfo = InInfo.intersect(BlockInfo[P->getNumber()].Exit);
1324 }
1325
1326 // If we don't have any valid predecessor value, wait until we do.
1327 if (!InInfo.isValid())
1328 return;
1329
1330 // If no change, no need to rerun block
1331 if (InInfo == BBInfo.Pred)
1332 return;
1333
1334 BBInfo.Pred = InInfo;
1335 LLVM_DEBUG(dbgs() << "Entry state of " << printMBBReference(MBB)
1336 << " changed to " << BBInfo.Pred << "\n");
1337
1338 // Note: It's tempting to cache the state changes here, but due to the
1339 // compatibility checks performed a blocks output state can change based on
1340 // the input state. To cache, we'd have to add logic for finding
1341 // never-compatible state changes.
1342 VSETVLIInfo TmpStatus;
1343 computeVLVTYPEChanges(MBB, TmpStatus);
1344
1345 // If the new exit value matches the old exit value, we don't need to revisit
1346 // any blocks.
1347 if (BBInfo.Exit == TmpStatus)
1348 return;
1349
1350 BBInfo.Exit = TmpStatus;
1351 LLVM_DEBUG(dbgs() << "Exit state of " << printMBBReference(MBB)
1352 << " changed to " << BBInfo.Exit << "\n");
1353
1354 // Add the successors to the work list so we can propagate the changed exit
1355 // status.
1356 for (MachineBasicBlock *S : MBB.successors())
1357 if (!BlockInfo[S->getNumber()].InQueue) {
1358 BlockInfo[S->getNumber()].InQueue = true;
1359 WorkList.push(S);
1360 }
1361 }
1362
1363 // If we weren't able to prove a vsetvli was directly unneeded, it might still
1364 // be unneeded if the AVL was a phi node where all incoming values are VL
1365 // outputs from the last VSETVLI in their respective basic blocks.
needVSETVLIPHI(const VSETVLIInfo & Require,const MachineBasicBlock & MBB) const1366 bool RISCVInsertVSETVLI::needVSETVLIPHI(const VSETVLIInfo &Require,
1367 const MachineBasicBlock &MBB) const {
1368 if (!Require.hasAVLReg())
1369 return true;
1370
1371 if (!LIS)
1372 return true;
1373
1374 // We need the AVL to have been produced by a PHI node in this basic block.
1375 const VNInfo *Valno = Require.getAVLVNInfo();
1376 if (!Valno->isPHIDef() || LIS->getMBBFromIndex(Valno->def) != &MBB)
1377 return true;
1378
1379 const LiveRange &LR = LIS->getInterval(Require.getAVLReg());
1380
1381 for (auto *PBB : MBB.predecessors()) {
1382 const VSETVLIInfo &PBBExit = BlockInfo[PBB->getNumber()].Exit;
1383
1384 // We need the PHI input to the be the output of a VSET(I)VLI.
1385 const VNInfo *Value = LR.getVNInfoBefore(LIS->getMBBEndIdx(PBB));
1386 if (!Value)
1387 return true;
1388 MachineInstr *DefMI = LIS->getInstructionFromIndex(Value->def);
1389 if (!DefMI || !isVectorConfigInstr(*DefMI))
1390 return true;
1391
1392 // We found a VSET(I)VLI make sure it matches the output of the
1393 // predecessor block.
1394 VSETVLIInfo DefInfo = getInfoForVSETVLI(*DefMI);
1395 if (DefInfo != PBBExit)
1396 return true;
1397
1398 // Require has the same VL as PBBExit, so if the exit from the
1399 // predecessor has the VTYPE we are looking for we might be able
1400 // to avoid a VSETVLI.
1401 if (PBBExit.isUnknown() || !PBBExit.hasSameVTYPE(Require))
1402 return true;
1403 }
1404
1405 // If all the incoming values to the PHI checked out, we don't need
1406 // to insert a VSETVLI.
1407 return false;
1408 }
1409
emitVSETVLIs(MachineBasicBlock & MBB)1410 void RISCVInsertVSETVLI::emitVSETVLIs(MachineBasicBlock &MBB) {
1411 VSETVLIInfo CurInfo = BlockInfo[MBB.getNumber()].Pred;
1412 // Track whether the prefix of the block we've scanned is transparent
1413 // (meaning has not yet changed the abstract state).
1414 bool PrefixTransparent = true;
1415 for (MachineInstr &MI : MBB) {
1416 const VSETVLIInfo PrevInfo = CurInfo;
1417 transferBefore(CurInfo, MI);
1418
1419 // If this is an explicit VSETVLI or VSETIVLI, update our state.
1420 if (isVectorConfigInstr(MI)) {
1421 // Conservatively, mark the VL and VTYPE as live.
1422 assert(MI.getOperand(3).getReg() == RISCV::VL &&
1423 MI.getOperand(4).getReg() == RISCV::VTYPE &&
1424 "Unexpected operands where VL and VTYPE should be");
1425 MI.getOperand(3).setIsDead(false);
1426 MI.getOperand(4).setIsDead(false);
1427 PrefixTransparent = false;
1428 }
1429
1430 uint64_t TSFlags = MI.getDesc().TSFlags;
1431 if (RISCVII::hasSEWOp(TSFlags)) {
1432 if (!PrevInfo.isCompatible(DemandedFields::all(), CurInfo, LIS)) {
1433 // If this is the first implicit state change, and the state change
1434 // requested can be proven to produce the same register contents, we
1435 // can skip emitting the actual state change and continue as if we
1436 // had since we know the GPR result of the implicit state change
1437 // wouldn't be used and VL/VTYPE registers are correct. Note that
1438 // we *do* need to model the state as if it changed as while the
1439 // register contents are unchanged, the abstract model can change.
1440 if (!PrefixTransparent || needVSETVLIPHI(CurInfo, MBB))
1441 insertVSETVLI(MBB, MI, MI.getDebugLoc(), CurInfo, PrevInfo);
1442 PrefixTransparent = false;
1443 }
1444
1445 if (RISCVII::hasVLOp(TSFlags)) {
1446 MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
1447 if (VLOp.isReg()) {
1448 Register Reg = VLOp.getReg();
1449
1450 // Erase the AVL operand from the instruction.
1451 VLOp.setReg(RISCV::NoRegister);
1452 VLOp.setIsKill(false);
1453 if (LIS) {
1454 LiveInterval &LI = LIS->getInterval(Reg);
1455 SmallVector<MachineInstr *> DeadMIs;
1456 LIS->shrinkToUses(&LI, &DeadMIs);
1457 // We might have separate components that need split due to
1458 // needVSETVLIPHI causing us to skip inserting a new VL def.
1459 SmallVector<LiveInterval *> SplitLIs;
1460 LIS->splitSeparateComponents(LI, SplitLIs);
1461
1462 // If the AVL was an immediate > 31, then it would have been emitted
1463 // as an ADDI. However, the ADDI might not have been used in the
1464 // vsetvli, or a vsetvli might not have been emitted, so it may be
1465 // dead now.
1466 for (MachineInstr *DeadMI : DeadMIs) {
1467 if (!TII->isAddImmediate(*DeadMI, Reg))
1468 continue;
1469 LIS->RemoveMachineInstrFromMaps(*DeadMI);
1470 DeadMI->eraseFromParent();
1471 }
1472 }
1473 }
1474 MI.addOperand(MachineOperand::CreateReg(RISCV::VL, /*isDef*/ false,
1475 /*isImp*/ true));
1476 }
1477 MI.addOperand(MachineOperand::CreateReg(RISCV::VTYPE, /*isDef*/ false,
1478 /*isImp*/ true));
1479 }
1480
1481 if (MI.isCall() || MI.isInlineAsm() ||
1482 MI.modifiesRegister(RISCV::VL, /*TRI=*/nullptr) ||
1483 MI.modifiesRegister(RISCV::VTYPE, /*TRI=*/nullptr))
1484 PrefixTransparent = false;
1485
1486 transferAfter(CurInfo, MI);
1487 }
1488
1489 const auto &Info = BlockInfo[MBB.getNumber()];
1490 if (CurInfo != Info.Exit) {
1491 LLVM_DEBUG(dbgs() << "in block " << printMBBReference(MBB) << "\n");
1492 LLVM_DEBUG(dbgs() << " begin state: " << Info.Pred << "\n");
1493 LLVM_DEBUG(dbgs() << " expected end state: " << Info.Exit << "\n");
1494 LLVM_DEBUG(dbgs() << " actual end state: " << CurInfo << "\n");
1495 }
1496 assert(CurInfo == Info.Exit && "InsertVSETVLI dataflow invariant violated");
1497 }
1498
1499 /// Perform simple partial redundancy elimination of the VSETVLI instructions
1500 /// we're about to insert by looking for cases where we can PRE from the
1501 /// beginning of one block to the end of one of its predecessors. Specifically,
1502 /// this is geared to catch the common case of a fixed length vsetvl in a single
1503 /// block loop when it could execute once in the preheader instead.
doPRE(MachineBasicBlock & MBB)1504 void RISCVInsertVSETVLI::doPRE(MachineBasicBlock &MBB) {
1505 if (!BlockInfo[MBB.getNumber()].Pred.isUnknown())
1506 return;
1507
1508 MachineBasicBlock *UnavailablePred = nullptr;
1509 VSETVLIInfo AvailableInfo;
1510 for (MachineBasicBlock *P : MBB.predecessors()) {
1511 const VSETVLIInfo &PredInfo = BlockInfo[P->getNumber()].Exit;
1512 if (PredInfo.isUnknown()) {
1513 if (UnavailablePred)
1514 return;
1515 UnavailablePred = P;
1516 } else if (!AvailableInfo.isValid()) {
1517 AvailableInfo = PredInfo;
1518 } else if (AvailableInfo != PredInfo) {
1519 return;
1520 }
1521 }
1522
1523 // Unreachable, single pred, or full redundancy. Note that FRE is handled by
1524 // phase 3.
1525 if (!UnavailablePred || !AvailableInfo.isValid())
1526 return;
1527
1528 if (!LIS)
1529 return;
1530
1531 // If we don't know the exact VTYPE, we can't copy the vsetvli to the exit of
1532 // the unavailable pred.
1533 if (AvailableInfo.hasSEWLMULRatioOnly())
1534 return;
1535
1536 // Critical edge - TODO: consider splitting?
1537 if (UnavailablePred->succ_size() != 1)
1538 return;
1539
1540 // If the AVL value is a register (other than our VLMAX sentinel),
1541 // we need to prove the value is available at the point we're going
1542 // to insert the vsetvli at.
1543 if (AvailableInfo.hasAVLReg()) {
1544 SlotIndex SI = AvailableInfo.getAVLVNInfo()->def;
1545 // This is an inline dominance check which covers the case of
1546 // UnavailablePred being the preheader of a loop.
1547 if (LIS->getMBBFromIndex(SI) != UnavailablePred)
1548 return;
1549 if (!UnavailablePred->terminators().empty() &&
1550 SI >= LIS->getInstructionIndex(*UnavailablePred->getFirstTerminator()))
1551 return;
1552 }
1553
1554 // Model the effect of changing the input state of the block MBB to
1555 // AvailableInfo. We're looking for two issues here; one legality,
1556 // one profitability.
1557 // 1) If the block doesn't use some of the fields from VL or VTYPE, we
1558 // may hit the end of the block with a different end state. We can
1559 // not make this change without reflowing later blocks as well.
1560 // 2) If we don't actually remove a transition, inserting a vsetvli
1561 // into the predecessor block would be correct, but unprofitable.
1562 VSETVLIInfo OldInfo = BlockInfo[MBB.getNumber()].Pred;
1563 VSETVLIInfo CurInfo = AvailableInfo;
1564 int TransitionsRemoved = 0;
1565 for (const MachineInstr &MI : MBB) {
1566 const VSETVLIInfo LastInfo = CurInfo;
1567 const VSETVLIInfo LastOldInfo = OldInfo;
1568 transferBefore(CurInfo, MI);
1569 transferBefore(OldInfo, MI);
1570 if (CurInfo == LastInfo)
1571 TransitionsRemoved++;
1572 if (LastOldInfo == OldInfo)
1573 TransitionsRemoved--;
1574 transferAfter(CurInfo, MI);
1575 transferAfter(OldInfo, MI);
1576 if (CurInfo == OldInfo)
1577 // Convergence. All transitions after this must match by construction.
1578 break;
1579 }
1580 if (CurInfo != OldInfo || TransitionsRemoved <= 0)
1581 // Issues 1 and 2 above
1582 return;
1583
1584 // Finally, update both data flow state and insert the actual vsetvli.
1585 // Doing both keeps the code in sync with the dataflow results, which
1586 // is critical for correctness of phase 3.
1587 auto OldExit = BlockInfo[UnavailablePred->getNumber()].Exit;
1588 LLVM_DEBUG(dbgs() << "PRE VSETVLI from " << MBB.getName() << " to "
1589 << UnavailablePred->getName() << " with state "
1590 << AvailableInfo << "\n");
1591 BlockInfo[UnavailablePred->getNumber()].Exit = AvailableInfo;
1592 BlockInfo[MBB.getNumber()].Pred = AvailableInfo;
1593
1594 // Note there's an implicit assumption here that terminators never use
1595 // or modify VL or VTYPE. Also, fallthrough will return end().
1596 auto InsertPt = UnavailablePred->getFirstInstrTerminator();
1597 insertVSETVLI(*UnavailablePred, InsertPt,
1598 UnavailablePred->findDebugLoc(InsertPt),
1599 AvailableInfo, OldExit);
1600 }
1601
1602 // Return true if we can mutate PrevMI to match MI without changing any the
1603 // fields which would be observed.
canMutatePriorConfig(const MachineInstr & PrevMI,const MachineInstr & MI,const DemandedFields & Used) const1604 bool RISCVInsertVSETVLI::canMutatePriorConfig(
1605 const MachineInstr &PrevMI, const MachineInstr &MI,
1606 const DemandedFields &Used) const {
1607 // If the VL values aren't equal, return false if either a) the former is
1608 // demanded, or b) we can't rewrite the former to be the later for
1609 // implementation reasons.
1610 if (!isVLPreservingConfig(MI)) {
1611 if (Used.VLAny)
1612 return false;
1613
1614 if (Used.VLZeroness) {
1615 if (isVLPreservingConfig(PrevMI))
1616 return false;
1617 if (!getInfoForVSETVLI(PrevMI).hasEquallyZeroAVL(getInfoForVSETVLI(MI),
1618 LIS))
1619 return false;
1620 }
1621
1622 auto &AVL = MI.getOperand(1);
1623 auto &PrevAVL = PrevMI.getOperand(1);
1624
1625 // If the AVL is a register, we need to make sure MI's AVL dominates PrevMI.
1626 // For now just check that PrevMI uses the same virtual register.
1627 if (AVL.isReg() && AVL.getReg() != RISCV::X0 &&
1628 (!MRI->hasOneDef(AVL.getReg()) || !PrevAVL.isReg() ||
1629 PrevAVL.getReg() != AVL.getReg()))
1630 return false;
1631 }
1632
1633 assert(PrevMI.getOperand(2).isImm() && MI.getOperand(2).isImm());
1634 auto PriorVType = PrevMI.getOperand(2).getImm();
1635 auto VType = MI.getOperand(2).getImm();
1636 return areCompatibleVTYPEs(PriorVType, VType, Used);
1637 }
1638
coalesceVSETVLIs(MachineBasicBlock & MBB) const1639 void RISCVInsertVSETVLI::coalesceVSETVLIs(MachineBasicBlock &MBB) const {
1640 MachineInstr *NextMI = nullptr;
1641 // We can have arbitrary code in successors, so VL and VTYPE
1642 // must be considered demanded.
1643 DemandedFields Used;
1644 Used.demandVL();
1645 Used.demandVTYPE();
1646 SmallVector<MachineInstr*> ToDelete;
1647
1648 // Update LIS and cleanup dead AVLs given a value which has
1649 // has had one use (as an AVL) removed.
1650 auto afterDroppedAVLUse = [&](Register OldVLReg) {
1651 if (LIS)
1652 LIS->shrinkToUses(&LIS->getInterval(OldVLReg));
1653
1654 MachineInstr *VLOpDef = MRI->getUniqueVRegDef(OldVLReg);
1655 if (VLOpDef && TII->isAddImmediate(*VLOpDef, OldVLReg) &&
1656 MRI->use_nodbg_empty(OldVLReg)) {
1657 if (LIS) {
1658 LIS->removeInterval(OldVLReg);
1659 LIS->RemoveMachineInstrFromMaps(*VLOpDef);
1660 }
1661 VLOpDef->eraseFromParent();
1662 }
1663 };
1664
1665 for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) {
1666
1667 if (!isVectorConfigInstr(MI)) {
1668 Used.doUnion(getDemanded(MI, ST));
1669 if (MI.isCall() || MI.isInlineAsm() ||
1670 MI.modifiesRegister(RISCV::VL, /*TRI=*/nullptr) ||
1671 MI.modifiesRegister(RISCV::VTYPE, /*TRI=*/nullptr))
1672 NextMI = nullptr;
1673 continue;
1674 }
1675
1676 if (!MI.getOperand(0).isDead())
1677 Used.demandVL();
1678
1679 if (NextMI) {
1680 if (!Used.usedVL() && !Used.usedVTYPE()) {
1681 ToDelete.push_back(&MI);
1682 // Leave NextMI unchanged
1683 continue;
1684 }
1685
1686 if (canMutatePriorConfig(MI, *NextMI, Used)) {
1687 if (!isVLPreservingConfig(*NextMI)) {
1688 Register DefReg = NextMI->getOperand(0).getReg();
1689
1690 MI.getOperand(0).setReg(DefReg);
1691 MI.getOperand(0).setIsDead(false);
1692
1693 // The def of DefReg moved to MI, so extend the LiveInterval up to
1694 // it.
1695 if (DefReg.isVirtual() && LIS) {
1696 LiveInterval &DefLI = LIS->getInterval(DefReg);
1697 SlotIndex MISlot = LIS->getInstructionIndex(MI).getRegSlot();
1698 VNInfo *DefVNI = DefLI.getVNInfoAt(DefLI.beginIndex());
1699 LiveInterval::Segment S(MISlot, DefLI.beginIndex(), DefVNI);
1700 DefLI.addSegment(S);
1701 DefVNI->def = MISlot;
1702 // Mark DefLI as spillable if it was previously unspillable
1703 DefLI.setWeight(0);
1704
1705 // DefReg may have had no uses, in which case we need to shrink
1706 // the LiveInterval up to MI.
1707 LIS->shrinkToUses(&DefLI);
1708 }
1709
1710 Register OldVLReg;
1711 if (MI.getOperand(1).isReg())
1712 OldVLReg = MI.getOperand(1).getReg();
1713 if (NextMI->getOperand(1).isImm())
1714 MI.getOperand(1).ChangeToImmediate(NextMI->getOperand(1).getImm());
1715 else
1716 MI.getOperand(1).ChangeToRegister(NextMI->getOperand(1).getReg(), false);
1717 if (OldVLReg && OldVLReg.isVirtual())
1718 afterDroppedAVLUse(OldVLReg);
1719
1720 MI.setDesc(NextMI->getDesc());
1721 }
1722 MI.getOperand(2).setImm(NextMI->getOperand(2).getImm());
1723 ToDelete.push_back(NextMI);
1724 // fallthrough
1725 }
1726 }
1727 NextMI = &MI;
1728 Used = getDemanded(MI, ST);
1729 }
1730
1731 NumCoalescedVSETVL += ToDelete.size();
1732 for (auto *MI : ToDelete) {
1733 if (LIS)
1734 LIS->RemoveMachineInstrFromMaps(*MI);
1735 Register OldAVLReg;
1736 if (MI->getOperand(1).isReg())
1737 OldAVLReg = MI->getOperand(1).getReg();
1738 MI->eraseFromParent();
1739 if (OldAVLReg && OldAVLReg.isVirtual())
1740 afterDroppedAVLUse(OldAVLReg);
1741 }
1742 }
1743
insertReadVL(MachineBasicBlock & MBB)1744 void RISCVInsertVSETVLI::insertReadVL(MachineBasicBlock &MBB) {
1745 for (auto I = MBB.begin(), E = MBB.end(); I != E;) {
1746 MachineInstr &MI = *I++;
1747 if (RISCV::isFaultFirstLoad(MI)) {
1748 Register VLOutput = MI.getOperand(1).getReg();
1749 assert(VLOutput.isVirtual());
1750 if (!MI.getOperand(1).isDead()) {
1751 auto ReadVLMI = BuildMI(MBB, I, MI.getDebugLoc(),
1752 TII->get(RISCV::PseudoReadVL), VLOutput);
1753 // Move the LiveInterval's definition down to PseudoReadVL.
1754 if (LIS) {
1755 SlotIndex NewDefSI =
1756 LIS->InsertMachineInstrInMaps(*ReadVLMI).getRegSlot();
1757 LiveInterval &DefLI = LIS->getInterval(VLOutput);
1758 VNInfo *DefVNI = DefLI.getVNInfoAt(DefLI.beginIndex());
1759 DefLI.removeSegment(DefLI.beginIndex(), NewDefSI);
1760 DefVNI->def = NewDefSI;
1761 }
1762 }
1763 // We don't use the vl output of the VLEFF/VLSEGFF anymore.
1764 MI.getOperand(1).setReg(RISCV::X0);
1765 }
1766 }
1767 }
1768
runOnMachineFunction(MachineFunction & MF)1769 bool RISCVInsertVSETVLI::runOnMachineFunction(MachineFunction &MF) {
1770 // Skip if the vector extension is not enabled.
1771 ST = &MF.getSubtarget<RISCVSubtarget>();
1772 if (!ST->hasVInstructions())
1773 return false;
1774
1775 LLVM_DEBUG(dbgs() << "Entering InsertVSETVLI for " << MF.getName() << "\n");
1776
1777 TII = ST->getInstrInfo();
1778 MRI = &MF.getRegInfo();
1779 auto *LISWrapper = getAnalysisIfAvailable<LiveIntervalsWrapperPass>();
1780 LIS = LISWrapper ? &LISWrapper->getLIS() : nullptr;
1781
1782 assert(BlockInfo.empty() && "Expect empty block infos");
1783 BlockInfo.resize(MF.getNumBlockIDs());
1784
1785 bool HaveVectorOp = false;
1786
1787 // Phase 1 - determine how VL/VTYPE are affected by the each block.
1788 for (const MachineBasicBlock &MBB : MF) {
1789 VSETVLIInfo TmpStatus;
1790 HaveVectorOp |= computeVLVTYPEChanges(MBB, TmpStatus);
1791 // Initial exit state is whatever change we found in the block.
1792 BlockData &BBInfo = BlockInfo[MBB.getNumber()];
1793 BBInfo.Exit = TmpStatus;
1794 LLVM_DEBUG(dbgs() << "Initial exit state of " << printMBBReference(MBB)
1795 << " is " << BBInfo.Exit << "\n");
1796
1797 }
1798
1799 // If we didn't find any instructions that need VSETVLI, we're done.
1800 if (!HaveVectorOp) {
1801 BlockInfo.clear();
1802 return false;
1803 }
1804
1805 // Phase 2 - determine the exit VL/VTYPE from each block. We add all
1806 // blocks to the list here, but will also add any that need to be revisited
1807 // during Phase 2 processing.
1808 for (const MachineBasicBlock &MBB : MF) {
1809 WorkList.push(&MBB);
1810 BlockInfo[MBB.getNumber()].InQueue = true;
1811 }
1812 while (!WorkList.empty()) {
1813 const MachineBasicBlock &MBB = *WorkList.front();
1814 WorkList.pop();
1815 computeIncomingVLVTYPE(MBB);
1816 }
1817
1818 // Perform partial redundancy elimination of vsetvli transitions.
1819 for (MachineBasicBlock &MBB : MF)
1820 doPRE(MBB);
1821
1822 // Phase 3 - add any vsetvli instructions needed in the block. Use the
1823 // Phase 2 information to avoid adding vsetvlis before the first vector
1824 // instruction in the block if the VL/VTYPE is satisfied by its
1825 // predecessors.
1826 for (MachineBasicBlock &MBB : MF)
1827 emitVSETVLIs(MBB);
1828
1829 // Now that all vsetvlis are explicit, go through and do block local
1830 // DSE and peephole based demanded fields based transforms. Note that
1831 // this *must* be done outside the main dataflow so long as we allow
1832 // any cross block analysis within the dataflow. We can't have both
1833 // demanded fields based mutation and non-local analysis in the
1834 // dataflow at the same time without introducing inconsistencies.
1835 for (MachineBasicBlock &MBB : MF)
1836 coalesceVSETVLIs(MBB);
1837
1838 // Insert PseudoReadVL after VLEFF/VLSEGFF and replace it with the vl output
1839 // of VLEFF/VLSEGFF.
1840 for (MachineBasicBlock &MBB : MF)
1841 insertReadVL(MBB);
1842
1843 BlockInfo.clear();
1844 return HaveVectorOp;
1845 }
1846
1847 /// Returns an instance of the Insert VSETVLI pass.
createRISCVInsertVSETVLIPass()1848 FunctionPass *llvm::createRISCVInsertVSETVLIPass() {
1849 return new RISCVInsertVSETVLI();
1850 }
1851