1 //===-- SystemZISelDAGToDAG.cpp - A dag to dag inst selector for SystemZ --===//
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
7 //===----------------------------------------------------------------------===//
9 // This file defines an instruction selector for the SystemZ target.
11 //===----------------------------------------------------------------------===//
13 #include "SystemZTargetMachine.h"
14 #include "SystemZISelLowering.h"
15 #include "llvm/Analysis/AliasAnalysis.h"
16 #include "llvm/CodeGen/SelectionDAGISel.h"
17 #include "llvm/Support/Debug.h"
18 #include "llvm/Support/KnownBits.h"
19 #include "llvm/Support/raw_ostream.h"
23 #define DEBUG_TYPE "systemz-isel"
26 // Used to build addressing modes.
27 struct SystemZAddressingMode {
28 // The shape of the address.
33 // base+displacement+index for load and store operands
36 // base+displacement+index for load address operands
39 // base+displacement+index+ADJDYNALLOC
44 // The type of displacement. The enum names here correspond directly
45 // to the definitions in SystemZOperand.td. We could split them into
46 // flags -- single/pair, 128-bit, etc. -- but it hardly seems worth it.
56 // The parts of the address. The address is equivalent to:
58 // Base + Disp + Index + (IncludesDynAlloc ? ADJDYNALLOC : 0)
62 bool IncludesDynAlloc;
64 SystemZAddressingMode(AddrForm form, DispRange dr)
65 : Form(form), DR(dr), Base(), Disp(0), Index(),
66 IncludesDynAlloc(false) {}
68 // True if the address can have an index register.
69 bool hasIndexField() { return Form != FormBD; }
71 // True if the address can (and must) include ADJDYNALLOC.
72 bool isDynAlloc() { return Form == FormBDXDynAlloc; }
74 void dump(const llvm::SelectionDAG *DAG) {
75 errs() << "SystemZAddressingMode " << this << '\n';
79 Base.getNode()->dump(DAG);
83 if (hasIndexField()) {
86 Index.getNode()->dump(DAG);
91 errs() << " Disp " << Disp;
93 errs() << " + ADJDYNALLOC";
98 // Return a mask with Count low bits set.
99 static uint64_t allOnes(unsigned int Count) {
103 return (uint64_t(1) << Count) - 1;
106 // Represents operands 2 to 5 of the ROTATE AND ... SELECTED BITS operation
107 // given by Opcode. The operands are: Input (R2), Start (I3), End (I4) and
108 // Rotate (I5). The combined operand value is effectively:
110 // (or (rotl Input, Rotate), ~Mask)
114 // (and (rotl Input, Rotate), Mask)
116 // otherwise. The output value has BitSize bits, although Input may be
117 // narrower (in which case the upper bits are don't care), or wider (in which
118 // case the result will be truncated as part of the operation).
119 struct RxSBGOperands {
120 RxSBGOperands(unsigned Op, SDValue N)
121 : Opcode(Op), BitSize(N.getValueSizeInBits()),
122 Mask(allOnes(BitSize)), Input(N), Start(64 - BitSize), End(63),
134 class SystemZDAGToDAGISel : public SelectionDAGISel {
135 const SystemZSubtarget *Subtarget;
137 // Used by SystemZOperands.td to create integer constants.
138 inline SDValue getImm(const SDNode *Node, uint64_t Imm) const {
139 return CurDAG->getTargetConstant(Imm, SDLoc(Node), Node->getValueType(0));
142 const SystemZTargetMachine &getTargetMachine() const {
143 return static_cast<const SystemZTargetMachine &>(TM);
146 const SystemZInstrInfo *getInstrInfo() const {
147 return Subtarget->getInstrInfo();
150 // Try to fold more of the base or index of AM into AM, where IsBase
151 // selects between the base and index.
152 bool expandAddress(SystemZAddressingMode &AM, bool IsBase) const;
154 // Try to describe N in AM, returning true on success.
155 bool selectAddress(SDValue N, SystemZAddressingMode &AM) const;
157 // Extract individual target operands from matched address AM.
158 void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
159 SDValue &Base, SDValue &Disp) const;
160 void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
161 SDValue &Base, SDValue &Disp, SDValue &Index) const;
163 // Try to match Addr as a FormBD address with displacement type DR.
164 // Return true on success, storing the base and displacement in
165 // Base and Disp respectively.
166 bool selectBDAddr(SystemZAddressingMode::DispRange DR, SDValue Addr,
167 SDValue &Base, SDValue &Disp) const;
169 // Try to match Addr as a FormBDX address with displacement type DR.
170 // Return true on success and if the result had no index. Store the
171 // base and displacement in Base and Disp respectively.
172 bool selectMVIAddr(SystemZAddressingMode::DispRange DR, SDValue Addr,
173 SDValue &Base, SDValue &Disp) const;
175 // Try to match Addr as a FormBDX* address of form Form with
176 // displacement type DR. Return true on success, storing the base,
177 // displacement and index in Base, Disp and Index respectively.
178 bool selectBDXAddr(SystemZAddressingMode::AddrForm Form,
179 SystemZAddressingMode::DispRange DR, SDValue Addr,
180 SDValue &Base, SDValue &Disp, SDValue &Index) const;
182 // PC-relative address matching routines used by SystemZOperands.td.
183 bool selectPCRelAddress(SDValue Addr, SDValue &Target) const {
184 if (SystemZISD::isPCREL(Addr.getOpcode())) {
185 Target = Addr.getOperand(0);
191 // BD matching routines used by SystemZOperands.td.
192 bool selectBDAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp) const {
193 return selectBDAddr(SystemZAddressingMode::Disp12Only, Addr, Base, Disp);
195 bool selectBDAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
196 return selectBDAddr(SystemZAddressingMode::Disp12Pair, Addr, Base, Disp);
198 bool selectBDAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp) const {
199 return selectBDAddr(SystemZAddressingMode::Disp20Only, Addr, Base, Disp);
201 bool selectBDAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
202 return selectBDAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp);
205 // MVI matching routines used by SystemZOperands.td.
206 bool selectMVIAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
207 return selectMVIAddr(SystemZAddressingMode::Disp12Pair, Addr, Base, Disp);
209 bool selectMVIAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
210 return selectMVIAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp);
213 // BDX matching routines used by SystemZOperands.td.
214 bool selectBDXAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
215 SDValue &Index) const {
216 return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
217 SystemZAddressingMode::Disp12Only,
218 Addr, Base, Disp, Index);
220 bool selectBDXAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
221 SDValue &Index) const {
222 return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
223 SystemZAddressingMode::Disp12Pair,
224 Addr, Base, Disp, Index);
226 bool selectDynAlloc12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
227 SDValue &Index) const {
228 return selectBDXAddr(SystemZAddressingMode::FormBDXDynAlloc,
229 SystemZAddressingMode::Disp12Only,
230 Addr, Base, Disp, Index);
232 bool selectBDXAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp,
233 SDValue &Index) const {
234 return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
235 SystemZAddressingMode::Disp20Only,
236 Addr, Base, Disp, Index);
238 bool selectBDXAddr20Only128(SDValue Addr, SDValue &Base, SDValue &Disp,
239 SDValue &Index) const {
240 return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
241 SystemZAddressingMode::Disp20Only128,
242 Addr, Base, Disp, Index);
244 bool selectBDXAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
245 SDValue &Index) const {
246 return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
247 SystemZAddressingMode::Disp20Pair,
248 Addr, Base, Disp, Index);
250 bool selectLAAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
251 SDValue &Index) const {
252 return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
253 SystemZAddressingMode::Disp12Pair,
254 Addr, Base, Disp, Index);
256 bool selectLAAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
257 SDValue &Index) const {
258 return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
259 SystemZAddressingMode::Disp20Pair,
260 Addr, Base, Disp, Index);
263 // Try to match Addr as an address with a base, 12-bit displacement
264 // and index, where the index is element Elem of a vector.
265 // Return true on success, storing the base, displacement and vector
266 // in Base, Disp and Index respectively.
267 bool selectBDVAddr12Only(SDValue Addr, SDValue Elem, SDValue &Base,
268 SDValue &Disp, SDValue &Index) const;
270 // Check whether (or Op (and X InsertMask)) is effectively an insertion
271 // of X into bits InsertMask of some Y != Op. Return true if so and
273 bool detectOrAndInsertion(SDValue &Op, uint64_t InsertMask) const;
275 // Try to update RxSBG so that only the bits of RxSBG.Input in Mask are used.
276 // Return true on success.
277 bool refineRxSBGMask(RxSBGOperands &RxSBG, uint64_t Mask) const;
279 // Try to fold some of RxSBG.Input into other fields of RxSBG.
280 // Return true on success.
281 bool expandRxSBG(RxSBGOperands &RxSBG) const;
283 // Return an undefined value of type VT.
284 SDValue getUNDEF(const SDLoc &DL, EVT VT) const;
286 // Convert N to VT, if it isn't already.
287 SDValue convertTo(const SDLoc &DL, EVT VT, SDValue N) const;
289 // Try to implement AND or shift node N using RISBG with the zero flag set.
290 // Return the selected node on success, otherwise return null.
291 bool tryRISBGZero(SDNode *N);
293 // Try to use RISBG or Opcode to implement OR or XOR node N.
294 // Return the selected node on success, otherwise return null.
295 bool tryRxSBG(SDNode *N, unsigned Opcode);
297 // If Op0 is null, then Node is a constant that can be loaded using:
299 // (Opcode UpperVal LowerVal)
301 // If Op0 is nonnull, then Node can be implemented using:
303 // (Opcode (Opcode Op0 UpperVal) LowerVal)
304 void splitLargeImmediate(unsigned Opcode, SDNode *Node, SDValue Op0,
305 uint64_t UpperVal, uint64_t LowerVal);
307 void loadVectorConstant(const SystemZVectorConstantInfo &VCI,
310 // Try to use gather instruction Opcode to implement vector insertion N.
311 bool tryGather(SDNode *N, unsigned Opcode);
313 // Try to use scatter instruction Opcode to implement store Store.
314 bool tryScatter(StoreSDNode *Store, unsigned Opcode);
316 // Change a chain of {load; op; store} of the same value into a simple op
317 // through memory of that value, if the uses of the modified value and its
318 // address are suitable.
319 bool tryFoldLoadStoreIntoMemOperand(SDNode *Node);
321 // Return true if Load and Store are loads and stores of the same size
322 // and are guaranteed not to overlap. Such operations can be implemented
323 // using block (SS-format) instructions.
325 // Partial overlap would lead to incorrect code, since the block operations
326 // are logically bytewise, even though they have a fast path for the
327 // non-overlapping case. We also need to avoid full overlap (i.e. two
328 // addresses that might be equal at run time) because although that case
329 // would be handled correctly, it might be implemented by millicode.
330 bool canUseBlockOperation(StoreSDNode *Store, LoadSDNode *Load) const;
332 // N is a (store (load Y), X) pattern. Return true if it can use an MVC
334 bool storeLoadCanUseMVC(SDNode *N) const;
336 // N is a (store (op (load A[0]), (load A[1])), X) pattern. Return true
337 // if A[1 - I] == X and if N can use a block operation like NC from A[I]
339 bool storeLoadCanUseBlockBinary(SDNode *N, unsigned I) const;
341 // Try to expand a boolean SELECT_CCMASK using an IPM sequence.
342 SDValue expandSelectBoolean(SDNode *Node);
345 SystemZDAGToDAGISel(SystemZTargetMachine &TM, CodeGenOpt::Level OptLevel)
346 : SelectionDAGISel(TM, OptLevel) {}
348 bool runOnMachineFunction(MachineFunction &MF) override {
349 Subtarget = &MF.getSubtarget<SystemZSubtarget>();
350 return SelectionDAGISel::runOnMachineFunction(MF);
353 // Override MachineFunctionPass.
354 StringRef getPassName() const override {
355 return "SystemZ DAG->DAG Pattern Instruction Selection";
358 // Override SelectionDAGISel.
359 void Select(SDNode *Node) override;
360 bool SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
361 std::vector<SDValue> &OutOps) override;
362 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
363 void PreprocessISelDAG() override;
365 // Include the pieces autogenerated from the target description.
366 #include "SystemZGenDAGISel.inc"
368 } // end anonymous namespace
370 FunctionPass *llvm::createSystemZISelDag(SystemZTargetMachine &TM,
371 CodeGenOpt::Level OptLevel) {
372 return new SystemZDAGToDAGISel(TM, OptLevel);
375 // Return true if Val should be selected as a displacement for an address
376 // with range DR. Here we're interested in the range of both the instruction
377 // described by DR and of any pairing instruction.
378 static bool selectDisp(SystemZAddressingMode::DispRange DR, int64_t Val) {
380 case SystemZAddressingMode::Disp12Only:
381 return isUInt<12>(Val);
383 case SystemZAddressingMode::Disp12Pair:
384 case SystemZAddressingMode::Disp20Only:
385 case SystemZAddressingMode::Disp20Pair:
386 return isInt<20>(Val);
388 case SystemZAddressingMode::Disp20Only128:
389 return isInt<20>(Val) && isInt<20>(Val + 8);
391 llvm_unreachable("Unhandled displacement range");
394 // Change the base or index in AM to Value, where IsBase selects
395 // between the base and index.
396 static void changeComponent(SystemZAddressingMode &AM, bool IsBase,
404 // The base or index of AM is equivalent to Value + ADJDYNALLOC,
405 // where IsBase selects between the base and index. Try to fold the
406 // ADJDYNALLOC into AM.
407 static bool expandAdjDynAlloc(SystemZAddressingMode &AM, bool IsBase,
409 if (AM.isDynAlloc() && !AM.IncludesDynAlloc) {
410 changeComponent(AM, IsBase, Value);
411 AM.IncludesDynAlloc = true;
417 // The base of AM is equivalent to Base + Index. Try to use Index as
418 // the index register.
419 static bool expandIndex(SystemZAddressingMode &AM, SDValue Base,
421 if (AM.hasIndexField() && !AM.Index.getNode()) {
429 // The base or index of AM is equivalent to Op0 + Op1, where IsBase selects
430 // between the base and index. Try to fold Op1 into AM's displacement.
431 static bool expandDisp(SystemZAddressingMode &AM, bool IsBase,
432 SDValue Op0, uint64_t Op1) {
433 // First try adjusting the displacement.
434 int64_t TestDisp = AM.Disp + Op1;
435 if (selectDisp(AM.DR, TestDisp)) {
436 changeComponent(AM, IsBase, Op0);
441 // We could consider forcing the displacement into a register and
442 // using it as an index, but it would need to be carefully tuned.
446 bool SystemZDAGToDAGISel::expandAddress(SystemZAddressingMode &AM,
448 SDValue N = IsBase ? AM.Base : AM.Index;
449 unsigned Opcode = N.getOpcode();
450 if (Opcode == ISD::TRUNCATE) {
452 Opcode = N.getOpcode();
454 if (Opcode == ISD::ADD || CurDAG->isBaseWithConstantOffset(N)) {
455 SDValue Op0 = N.getOperand(0);
456 SDValue Op1 = N.getOperand(1);
458 unsigned Op0Code = Op0->getOpcode();
459 unsigned Op1Code = Op1->getOpcode();
461 if (Op0Code == SystemZISD::ADJDYNALLOC)
462 return expandAdjDynAlloc(AM, IsBase, Op1);
463 if (Op1Code == SystemZISD::ADJDYNALLOC)
464 return expandAdjDynAlloc(AM, IsBase, Op0);
466 if (Op0Code == ISD::Constant)
467 return expandDisp(AM, IsBase, Op1,
468 cast<ConstantSDNode>(Op0)->getSExtValue());
469 if (Op1Code == ISD::Constant)
470 return expandDisp(AM, IsBase, Op0,
471 cast<ConstantSDNode>(Op1)->getSExtValue());
473 if (IsBase && expandIndex(AM, Op0, Op1))
476 if (Opcode == SystemZISD::PCREL_OFFSET) {
477 SDValue Full = N.getOperand(0);
478 SDValue Base = N.getOperand(1);
479 SDValue Anchor = Base.getOperand(0);
480 uint64_t Offset = (cast<GlobalAddressSDNode>(Full)->getOffset() -
481 cast<GlobalAddressSDNode>(Anchor)->getOffset());
482 return expandDisp(AM, IsBase, Base, Offset);
487 // Return true if an instruction with displacement range DR should be
488 // used for displacement value Val. selectDisp(DR, Val) must already hold.
489 static bool isValidDisp(SystemZAddressingMode::DispRange DR, int64_t Val) {
490 assert(selectDisp(DR, Val) && "Invalid displacement");
492 case SystemZAddressingMode::Disp12Only:
493 case SystemZAddressingMode::Disp20Only:
494 case SystemZAddressingMode::Disp20Only128:
497 case SystemZAddressingMode::Disp12Pair:
498 // Use the other instruction if the displacement is too large.
499 return isUInt<12>(Val);
501 case SystemZAddressingMode::Disp20Pair:
502 // Use the other instruction if the displacement is small enough.
503 return !isUInt<12>(Val);
505 llvm_unreachable("Unhandled displacement range");
508 // Return true if Base + Disp + Index should be performed by LA(Y).
509 static bool shouldUseLA(SDNode *Base, int64_t Disp, SDNode *Index) {
510 // Don't use LA(Y) for constants.
514 // Always use LA(Y) for frame addresses, since we know that the destination
515 // register is almost always (perhaps always) going to be different from
516 // the frame register.
517 if (Base->getOpcode() == ISD::FrameIndex)
521 // Always use LA(Y) if there is a base, displacement and index.
525 // Always use LA if the displacement is small enough. It should always
526 // be no worse than AGHI (and better if it avoids a move).
527 if (isUInt<12>(Disp))
530 // For similar reasons, always use LAY if the constant is too big for AGHI.
531 // LAY should be no worse than AGFI.
532 if (!isInt<16>(Disp))
535 // Don't use LA for plain registers.
539 // Don't use LA for plain addition if the index operand is only used
540 // once. It should be a natural two-operand addition in that case.
541 if (Index->hasOneUse())
544 // Prefer addition if the second operation is sign-extended, in the
545 // hope of using AGF.
546 unsigned IndexOpcode = Index->getOpcode();
547 if (IndexOpcode == ISD::SIGN_EXTEND ||
548 IndexOpcode == ISD::SIGN_EXTEND_INREG)
552 // Don't use LA for two-operand addition if either operand is only
553 // used once. The addition instructions are better in that case.
554 if (Base->hasOneUse())
560 // Return true if Addr is suitable for AM, updating AM if so.
561 bool SystemZDAGToDAGISel::selectAddress(SDValue Addr,
562 SystemZAddressingMode &AM) const {
563 // Start out assuming that the address will need to be loaded separately,
564 // then try to extend it as much as we can.
567 // First try treating the address as a constant.
568 if (Addr.getOpcode() == ISD::Constant &&
569 expandDisp(AM, true, SDValue(),
570 cast<ConstantSDNode>(Addr)->getSExtValue()))
572 // Also see if it's a bare ADJDYNALLOC.
573 else if (Addr.getOpcode() == SystemZISD::ADJDYNALLOC &&
574 expandAdjDynAlloc(AM, true, SDValue()))
577 // Otherwise try expanding each component.
578 while (expandAddress(AM, true) ||
579 (AM.Index.getNode() && expandAddress(AM, false)))
582 // Reject cases where it isn't profitable to use LA(Y).
583 if (AM.Form == SystemZAddressingMode::FormBDXLA &&
584 !shouldUseLA(AM.Base.getNode(), AM.Disp, AM.Index.getNode()))
587 // Reject cases where the other instruction in a pair should be used.
588 if (!isValidDisp(AM.DR, AM.Disp))
591 // Make sure that ADJDYNALLOC is included where necessary.
592 if (AM.isDynAlloc() && !AM.IncludesDynAlloc)
595 LLVM_DEBUG(AM.dump(CurDAG));
599 // Insert a node into the DAG at least before Pos. This will reposition
600 // the node as needed, and will assign it a node ID that is <= Pos's ID.
601 // Note that this does *not* preserve the uniqueness of node IDs!
602 // The selection DAG must no longer depend on their uniqueness when this
604 static void insertDAGNode(SelectionDAG *DAG, SDNode *Pos, SDValue N) {
605 if (N->getNodeId() == -1 ||
606 (SelectionDAGISel::getUninvalidatedNodeId(N.getNode()) >
607 SelectionDAGISel::getUninvalidatedNodeId(Pos))) {
608 DAG->RepositionNode(Pos->getIterator(), N.getNode());
609 // Mark Node as invalid for pruning as after this it may be a successor to a
610 // selected node but otherwise be in the same position of Pos.
611 // Conservatively mark it with the same -abs(Id) to assure node id
612 // invariant is preserved.
613 N->setNodeId(Pos->getNodeId());
614 SelectionDAGISel::InvalidateNodeId(N.getNode());
618 void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
619 EVT VT, SDValue &Base,
620 SDValue &Disp) const {
623 // Register 0 means "no base". This is mostly useful for shifts.
624 Base = CurDAG->getRegister(0, VT);
625 else if (Base.getOpcode() == ISD::FrameIndex) {
626 // Lower a FrameIndex to a TargetFrameIndex.
627 int64_t FrameIndex = cast<FrameIndexSDNode>(Base)->getIndex();
628 Base = CurDAG->getTargetFrameIndex(FrameIndex, VT);
629 } else if (Base.getValueType() != VT) {
630 // Truncate values from i64 to i32, for shifts.
631 assert(VT == MVT::i32 && Base.getValueType() == MVT::i64 &&
632 "Unexpected truncation");
634 SDValue Trunc = CurDAG->getNode(ISD::TRUNCATE, DL, VT, Base);
635 insertDAGNode(CurDAG, Base.getNode(), Trunc);
639 // Lower the displacement to a TargetConstant.
640 Disp = CurDAG->getTargetConstant(AM.Disp, SDLoc(Base), VT);
643 void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
644 EVT VT, SDValue &Base,
646 SDValue &Index) const {
647 getAddressOperands(AM, VT, Base, Disp);
650 if (!Index.getNode())
651 // Register 0 means "no index".
652 Index = CurDAG->getRegister(0, VT);
655 bool SystemZDAGToDAGISel::selectBDAddr(SystemZAddressingMode::DispRange DR,
656 SDValue Addr, SDValue &Base,
657 SDValue &Disp) const {
658 SystemZAddressingMode AM(SystemZAddressingMode::FormBD, DR);
659 if (!selectAddress(Addr, AM))
662 getAddressOperands(AM, Addr.getValueType(), Base, Disp);
666 bool SystemZDAGToDAGISel::selectMVIAddr(SystemZAddressingMode::DispRange DR,
667 SDValue Addr, SDValue &Base,
668 SDValue &Disp) const {
669 SystemZAddressingMode AM(SystemZAddressingMode::FormBDXNormal, DR);
670 if (!selectAddress(Addr, AM) || AM.Index.getNode())
673 getAddressOperands(AM, Addr.getValueType(), Base, Disp);
677 bool SystemZDAGToDAGISel::selectBDXAddr(SystemZAddressingMode::AddrForm Form,
678 SystemZAddressingMode::DispRange DR,
679 SDValue Addr, SDValue &Base,
680 SDValue &Disp, SDValue &Index) const {
681 SystemZAddressingMode AM(Form, DR);
682 if (!selectAddress(Addr, AM))
685 getAddressOperands(AM, Addr.getValueType(), Base, Disp, Index);
689 bool SystemZDAGToDAGISel::selectBDVAddr12Only(SDValue Addr, SDValue Elem,
692 SDValue &Index) const {
694 if (selectBDXAddr12Only(Addr, Regs[0], Disp, Regs[1]) &&
695 Regs[0].getNode() && Regs[1].getNode()) {
696 for (unsigned int I = 0; I < 2; ++I) {
699 // We can't tell here whether the index vector has the right type
700 // for the access; the caller needs to do that instead.
701 if (Index.getOpcode() == ISD::ZERO_EXTEND)
702 Index = Index.getOperand(0);
703 if (Index.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
704 Index.getOperand(1) == Elem) {
705 Index = Index.getOperand(0);
713 bool SystemZDAGToDAGISel::detectOrAndInsertion(SDValue &Op,
714 uint64_t InsertMask) const {
715 // We're only interested in cases where the insertion is into some operand
716 // of Op, rather than into Op itself. The only useful case is an AND.
717 if (Op.getOpcode() != ISD::AND)
720 // We need a constant mask.
721 auto *MaskNode = dyn_cast<ConstantSDNode>(Op.getOperand(1).getNode());
725 // It's not an insertion of Op.getOperand(0) if the two masks overlap.
726 uint64_t AndMask = MaskNode->getZExtValue();
727 if (InsertMask & AndMask)
730 // It's only an insertion if all bits are covered or are known to be zero.
731 // The inner check covers all cases but is more expensive.
732 uint64_t Used = allOnes(Op.getValueSizeInBits());
733 if (Used != (AndMask | InsertMask)) {
734 KnownBits Known = CurDAG->computeKnownBits(Op.getOperand(0));
735 if (Used != (AndMask | InsertMask | Known.Zero.getZExtValue()))
739 Op = Op.getOperand(0);
743 bool SystemZDAGToDAGISel::refineRxSBGMask(RxSBGOperands &RxSBG,
744 uint64_t Mask) const {
745 const SystemZInstrInfo *TII = getInstrInfo();
746 if (RxSBG.Rotate != 0)
747 Mask = (Mask << RxSBG.Rotate) | (Mask >> (64 - RxSBG.Rotate));
749 if (TII->isRxSBGMask(Mask, RxSBG.BitSize, RxSBG.Start, RxSBG.End)) {
756 // Return true if any bits of (RxSBG.Input & Mask) are significant.
757 static bool maskMatters(RxSBGOperands &RxSBG, uint64_t Mask) {
758 // Rotate the mask in the same way as RxSBG.Input is rotated.
759 if (RxSBG.Rotate != 0)
760 Mask = ((Mask << RxSBG.Rotate) | (Mask >> (64 - RxSBG.Rotate)));
761 return (Mask & RxSBG.Mask) != 0;
764 bool SystemZDAGToDAGISel::expandRxSBG(RxSBGOperands &RxSBG) const {
765 SDValue N = RxSBG.Input;
766 unsigned Opcode = N.getOpcode();
768 case ISD::TRUNCATE: {
769 if (RxSBG.Opcode == SystemZ::RNSBG)
771 uint64_t BitSize = N.getValueSizeInBits();
772 uint64_t Mask = allOnes(BitSize);
773 if (!refineRxSBGMask(RxSBG, Mask))
775 RxSBG.Input = N.getOperand(0);
779 if (RxSBG.Opcode == SystemZ::RNSBG)
782 auto *MaskNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
786 SDValue Input = N.getOperand(0);
787 uint64_t Mask = MaskNode->getZExtValue();
788 if (!refineRxSBGMask(RxSBG, Mask)) {
789 // If some bits of Input are already known zeros, those bits will have
790 // been removed from the mask. See if adding them back in makes the
792 KnownBits Known = CurDAG->computeKnownBits(Input);
793 Mask |= Known.Zero.getZExtValue();
794 if (!refineRxSBGMask(RxSBG, Mask))
802 if (RxSBG.Opcode != SystemZ::RNSBG)
805 auto *MaskNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
809 SDValue Input = N.getOperand(0);
810 uint64_t Mask = ~MaskNode->getZExtValue();
811 if (!refineRxSBGMask(RxSBG, Mask)) {
812 // If some bits of Input are already known ones, those bits will have
813 // been removed from the mask. See if adding them back in makes the
815 KnownBits Known = CurDAG->computeKnownBits(Input);
816 Mask &= ~Known.One.getZExtValue();
817 if (!refineRxSBGMask(RxSBG, Mask))
825 // Any 64-bit rotate left can be merged into the RxSBG.
826 if (RxSBG.BitSize != 64 || N.getValueType() != MVT::i64)
828 auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
832 RxSBG.Rotate = (RxSBG.Rotate + CountNode->getZExtValue()) & 63;
833 RxSBG.Input = N.getOperand(0);
837 case ISD::ANY_EXTEND:
838 // Bits above the extended operand are don't-care.
839 RxSBG.Input = N.getOperand(0);
842 case ISD::ZERO_EXTEND:
843 if (RxSBG.Opcode != SystemZ::RNSBG) {
844 // Restrict the mask to the extended operand.
845 unsigned InnerBitSize = N.getOperand(0).getValueSizeInBits();
846 if (!refineRxSBGMask(RxSBG, allOnes(InnerBitSize)))
849 RxSBG.Input = N.getOperand(0);
854 case ISD::SIGN_EXTEND: {
855 // Check that the extension bits are don't-care (i.e. are masked out
856 // by the final mask).
857 unsigned BitSize = N.getValueSizeInBits();
858 unsigned InnerBitSize = N.getOperand(0).getValueSizeInBits();
859 if (maskMatters(RxSBG, allOnes(BitSize) - allOnes(InnerBitSize))) {
860 // In the case where only the sign bit is active, increase Rotate with
861 // the extension width.
862 if (RxSBG.Mask == 1 && RxSBG.Rotate == 1)
863 RxSBG.Rotate += (BitSize - InnerBitSize);
868 RxSBG.Input = N.getOperand(0);
873 auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
877 uint64_t Count = CountNode->getZExtValue();
878 unsigned BitSize = N.getValueSizeInBits();
879 if (Count < 1 || Count >= BitSize)
882 if (RxSBG.Opcode == SystemZ::RNSBG) {
883 // Treat (shl X, count) as (rotl X, size-count) as long as the bottom
884 // count bits from RxSBG.Input are ignored.
885 if (maskMatters(RxSBG, allOnes(Count)))
888 // Treat (shl X, count) as (and (rotl X, count), ~0<<count).
889 if (!refineRxSBGMask(RxSBG, allOnes(BitSize - Count) << Count))
893 RxSBG.Rotate = (RxSBG.Rotate + Count) & 63;
894 RxSBG.Input = N.getOperand(0);
900 auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
904 uint64_t Count = CountNode->getZExtValue();
905 unsigned BitSize = N.getValueSizeInBits();
906 if (Count < 1 || Count >= BitSize)
909 if (RxSBG.Opcode == SystemZ::RNSBG || Opcode == ISD::SRA) {
910 // Treat (srl|sra X, count) as (rotl X, size-count) as long as the top
911 // count bits from RxSBG.Input are ignored.
912 if (maskMatters(RxSBG, allOnes(Count) << (BitSize - Count)))
915 // Treat (srl X, count), mask) as (and (rotl X, size-count), ~0>>count),
916 // which is similar to SLL above.
917 if (!refineRxSBGMask(RxSBG, allOnes(BitSize - Count)))
921 RxSBG.Rotate = (RxSBG.Rotate - Count) & 63;
922 RxSBG.Input = N.getOperand(0);
930 SDValue SystemZDAGToDAGISel::getUNDEF(const SDLoc &DL, EVT VT) const {
931 SDNode *N = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, VT);
932 return SDValue(N, 0);
935 SDValue SystemZDAGToDAGISel::convertTo(const SDLoc &DL, EVT VT,
937 if (N.getValueType() == MVT::i32 && VT == MVT::i64)
938 return CurDAG->getTargetInsertSubreg(SystemZ::subreg_l32,
939 DL, VT, getUNDEF(DL, MVT::i64), N);
940 if (N.getValueType() == MVT::i64 && VT == MVT::i32)
941 return CurDAG->getTargetExtractSubreg(SystemZ::subreg_l32, DL, VT, N);
942 assert(N.getValueType() == VT && "Unexpected value types");
946 bool SystemZDAGToDAGISel::tryRISBGZero(SDNode *N) {
948 EVT VT = N->getValueType(0);
949 if (!VT.isInteger() || VT.getSizeInBits() > 64)
951 RxSBGOperands RISBG(SystemZ::RISBG, SDValue(N, 0));
953 while (expandRxSBG(RISBG))
954 // The widening or narrowing is expected to be free.
955 // Counting widening or narrowing as a saved operation will result in
956 // preferring an R*SBG over a simple shift/logical instruction.
957 if (RISBG.Input.getOpcode() != ISD::ANY_EXTEND &&
958 RISBG.Input.getOpcode() != ISD::TRUNCATE)
963 // Prefer to use normal shift instructions over RISBG, since they can handle
964 // all cases and are sometimes shorter.
965 if (Count == 1 && N->getOpcode() != ISD::AND)
968 // Prefer register extensions like LLC over RISBG. Also prefer to start
969 // out with normal ANDs if one instruction would be enough. We can convert
970 // these ANDs into an RISBG later if a three-address instruction is useful.
971 if (RISBG.Rotate == 0) {
972 bool PreferAnd = false;
973 // Prefer AND for any 32-bit and-immediate operation.
976 // As well as for any 64-bit operation that can be implemented via LLC(R),
977 // LLH(R), LLGT(R), or one of the and-immediate instructions.
978 else if (RISBG.Mask == 0xff ||
979 RISBG.Mask == 0xffff ||
980 RISBG.Mask == 0x7fffffff ||
981 SystemZ::isImmLF(~RISBG.Mask) ||
982 SystemZ::isImmHF(~RISBG.Mask))
984 // And likewise for the LLZRGF instruction, which doesn't have a register
985 // to register version.
986 else if (auto *Load = dyn_cast<LoadSDNode>(RISBG.Input)) {
987 if (Load->getMemoryVT() == MVT::i32 &&
988 (Load->getExtensionType() == ISD::EXTLOAD ||
989 Load->getExtensionType() == ISD::ZEXTLOAD) &&
990 RISBG.Mask == 0xffffff00 &&
991 Subtarget->hasLoadAndZeroRightmostByte())
995 // Replace the current node with an AND. Note that the current node
996 // might already be that same AND, in which case it is already CSE'd
997 // with it, and we must not call ReplaceNode.
998 SDValue In = convertTo(DL, VT, RISBG.Input);
999 SDValue Mask = CurDAG->getConstant(RISBG.Mask, DL, VT);
1000 SDValue New = CurDAG->getNode(ISD::AND, DL, VT, In, Mask);
1001 if (N != New.getNode()) {
1002 insertDAGNode(CurDAG, N, Mask);
1003 insertDAGNode(CurDAG, N, New);
1004 ReplaceNode(N, New.getNode());
1007 // Now, select the machine opcode to implement this operation.
1008 if (!N->isMachineOpcode())
1014 unsigned Opcode = SystemZ::RISBG;
1015 // Prefer RISBGN if available, since it does not clobber CC.
1016 if (Subtarget->hasMiscellaneousExtensions())
1017 Opcode = SystemZ::RISBGN;
1018 EVT OpcodeVT = MVT::i64;
1019 if (VT == MVT::i32 && Subtarget->hasHighWord() &&
1020 // We can only use the 32-bit instructions if all source bits are
1021 // in the low 32 bits without wrapping, both after rotation (because
1022 // of the smaller range for Start and End) and before rotation
1023 // (because the input value is truncated).
1024 RISBG.Start >= 32 && RISBG.End >= RISBG.Start &&
1025 ((RISBG.Start + RISBG.Rotate) & 63) >= 32 &&
1026 ((RISBG.End + RISBG.Rotate) & 63) >=
1027 ((RISBG.Start + RISBG.Rotate) & 63)) {
1028 Opcode = SystemZ::RISBMux;
1029 OpcodeVT = MVT::i32;
1034 getUNDEF(DL, OpcodeVT),
1035 convertTo(DL, OpcodeVT, RISBG.Input),
1036 CurDAG->getTargetConstant(RISBG.Start, DL, MVT::i32),
1037 CurDAG->getTargetConstant(RISBG.End | 128, DL, MVT::i32),
1038 CurDAG->getTargetConstant(RISBG.Rotate, DL, MVT::i32)
1040 SDValue New = convertTo(
1041 DL, VT, SDValue(CurDAG->getMachineNode(Opcode, DL, OpcodeVT, Ops), 0));
1042 ReplaceNode(N, New.getNode());
1046 bool SystemZDAGToDAGISel::tryRxSBG(SDNode *N, unsigned Opcode) {
1048 EVT VT = N->getValueType(0);
1049 if (!VT.isInteger() || VT.getSizeInBits() > 64)
1051 // Try treating each operand of N as the second operand of the RxSBG
1052 // and see which goes deepest.
1053 RxSBGOperands RxSBG[] = {
1054 RxSBGOperands(Opcode, N->getOperand(0)),
1055 RxSBGOperands(Opcode, N->getOperand(1))
1057 unsigned Count[] = { 0, 0 };
1058 for (unsigned I = 0; I < 2; ++I)
1059 while (expandRxSBG(RxSBG[I]))
1060 // The widening or narrowing is expected to be free.
1061 // Counting widening or narrowing as a saved operation will result in
1062 // preferring an R*SBG over a simple shift/logical instruction.
1063 if (RxSBG[I].Input.getOpcode() != ISD::ANY_EXTEND &&
1064 RxSBG[I].Input.getOpcode() != ISD::TRUNCATE)
1067 // Do nothing if neither operand is suitable.
1068 if (Count[0] == 0 && Count[1] == 0)
1071 // Pick the deepest second operand.
1072 unsigned I = Count[0] > Count[1] ? 0 : 1;
1073 SDValue Op0 = N->getOperand(I ^ 1);
1075 // Prefer IC for character insertions from memory.
1076 if (Opcode == SystemZ::ROSBG && (RxSBG[I].Mask & 0xff) == 0)
1077 if (auto *Load = dyn_cast<LoadSDNode>(Op0.getNode()))
1078 if (Load->getMemoryVT() == MVT::i8)
1081 // See whether we can avoid an AND in the first operand by converting
1083 if (Opcode == SystemZ::ROSBG && detectOrAndInsertion(Op0, RxSBG[I].Mask)) {
1084 Opcode = SystemZ::RISBG;
1085 // Prefer RISBGN if available, since it does not clobber CC.
1086 if (Subtarget->hasMiscellaneousExtensions())
1087 Opcode = SystemZ::RISBGN;
1091 convertTo(DL, MVT::i64, Op0),
1092 convertTo(DL, MVT::i64, RxSBG[I].Input),
1093 CurDAG->getTargetConstant(RxSBG[I].Start, DL, MVT::i32),
1094 CurDAG->getTargetConstant(RxSBG[I].End, DL, MVT::i32),
1095 CurDAG->getTargetConstant(RxSBG[I].Rotate, DL, MVT::i32)
1097 SDValue New = convertTo(
1098 DL, VT, SDValue(CurDAG->getMachineNode(Opcode, DL, MVT::i64, Ops), 0));
1099 ReplaceNode(N, New.getNode());
1103 void SystemZDAGToDAGISel::splitLargeImmediate(unsigned Opcode, SDNode *Node,
1104 SDValue Op0, uint64_t UpperVal,
1105 uint64_t LowerVal) {
1106 EVT VT = Node->getValueType(0);
1108 SDValue Upper = CurDAG->getConstant(UpperVal, DL, VT);
1110 Upper = CurDAG->getNode(Opcode, DL, VT, Op0, Upper);
1113 // When we haven't passed in Op0, Upper will be a constant. In order to
1114 // prevent folding back to the large immediate in `Or = getNode(...)` we run
1115 // SelectCode first and end up with an opaque machine node. This means that
1116 // we need to use a handle to keep track of Upper in case it gets CSE'd by
1119 // Note that in the case where Op0 is passed in we could just call
1120 // SelectCode(Upper) later, along with the SelectCode(Or), and avoid needing
1121 // the handle at all, but it's fine to do it here.
1123 // TODO: This is a pretty hacky way to do this. Can we do something that
1124 // doesn't require a two paragraph explanation?
1125 HandleSDNode Handle(Upper);
1126 SelectCode(Upper.getNode());
1127 Upper = Handle.getValue();
1130 SDValue Lower = CurDAG->getConstant(LowerVal, DL, VT);
1131 SDValue Or = CurDAG->getNode(Opcode, DL, VT, Upper, Lower);
1133 ReplaceNode(Node, Or.getNode());
1135 SelectCode(Or.getNode());
1138 void SystemZDAGToDAGISel::loadVectorConstant(
1139 const SystemZVectorConstantInfo &VCI, SDNode *Node) {
1140 assert((VCI.Opcode == SystemZISD::BYTE_MASK ||
1141 VCI.Opcode == SystemZISD::REPLICATE ||
1142 VCI.Opcode == SystemZISD::ROTATE_MASK) &&
1144 assert(VCI.VecVT.getSizeInBits() == 128 && "Expected a vector type");
1145 EVT VT = Node->getValueType(0);
1147 SmallVector<SDValue, 2> Ops;
1148 for (unsigned OpVal : VCI.OpVals)
1149 Ops.push_back(CurDAG->getConstant(OpVal, DL, MVT::i32));
1150 SDValue Op = CurDAG->getNode(VCI.Opcode, DL, VCI.VecVT, Ops);
1152 if (VCI.VecVT == VT.getSimpleVT())
1153 ReplaceNode(Node, Op.getNode());
1154 else if (VT.getSizeInBits() == 128) {
1155 SDValue BitCast = CurDAG->getNode(ISD::BITCAST, DL, VT, Op);
1156 ReplaceNode(Node, BitCast.getNode());
1157 SelectCode(BitCast.getNode());
1158 } else { // float or double
1159 unsigned SubRegIdx =
1160 (VT.getSizeInBits() == 32 ? SystemZ::subreg_h32 : SystemZ::subreg_h64);
1162 Node, CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, Op).getNode());
1164 SelectCode(Op.getNode());
1167 bool SystemZDAGToDAGISel::tryGather(SDNode *N, unsigned Opcode) {
1168 SDValue ElemV = N->getOperand(2);
1169 auto *ElemN = dyn_cast<ConstantSDNode>(ElemV);
1173 unsigned Elem = ElemN->getZExtValue();
1174 EVT VT = N->getValueType(0);
1175 if (Elem >= VT.getVectorNumElements())
1178 auto *Load = dyn_cast<LoadSDNode>(N->getOperand(1));
1179 if (!Load || !Load->hasNUsesOfValue(1, 0))
1181 if (Load->getMemoryVT().getSizeInBits() !=
1182 Load->getValueType(0).getSizeInBits())
1185 SDValue Base, Disp, Index;
1186 if (!selectBDVAddr12Only(Load->getBasePtr(), ElemV, Base, Disp, Index) ||
1187 Index.getValueType() != VT.changeVectorElementTypeToInteger())
1192 N->getOperand(0), Base, Disp, Index,
1193 CurDAG->getTargetConstant(Elem, DL, MVT::i32), Load->getChain()
1195 SDNode *Res = CurDAG->getMachineNode(Opcode, DL, VT, MVT::Other, Ops);
1196 ReplaceUses(SDValue(Load, 1), SDValue(Res, 1));
1197 ReplaceNode(N, Res);
1201 bool SystemZDAGToDAGISel::tryScatter(StoreSDNode *Store, unsigned Opcode) {
1202 SDValue Value = Store->getValue();
1203 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
1205 if (Store->getMemoryVT().getSizeInBits() != Value.getValueSizeInBits())
1208 SDValue ElemV = Value.getOperand(1);
1209 auto *ElemN = dyn_cast<ConstantSDNode>(ElemV);
1213 SDValue Vec = Value.getOperand(0);
1214 EVT VT = Vec.getValueType();
1215 unsigned Elem = ElemN->getZExtValue();
1216 if (Elem >= VT.getVectorNumElements())
1219 SDValue Base, Disp, Index;
1220 if (!selectBDVAddr12Only(Store->getBasePtr(), ElemV, Base, Disp, Index) ||
1221 Index.getValueType() != VT.changeVectorElementTypeToInteger())
1226 Vec, Base, Disp, Index, CurDAG->getTargetConstant(Elem, DL, MVT::i32),
1229 ReplaceNode(Store, CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops));
1233 // Check whether or not the chain ending in StoreNode is suitable for doing
1234 // the {load; op; store} to modify transformation.
1235 static bool isFusableLoadOpStorePattern(StoreSDNode *StoreNode,
1236 SDValue StoredVal, SelectionDAG *CurDAG,
1237 LoadSDNode *&LoadNode,
1238 SDValue &InputChain) {
1239 // Is the stored value result 0 of the operation?
1240 if (StoredVal.getResNo() != 0)
1243 // Are there other uses of the loaded value than the operation?
1244 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0))
1247 // Is the store non-extending and non-indexed?
1248 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
1251 SDValue Load = StoredVal->getOperand(0);
1252 // Is the stored value a non-extending and non-indexed load?
1253 if (!ISD::isNormalLoad(Load.getNode()))
1256 // Return LoadNode by reference.
1257 LoadNode = cast<LoadSDNode>(Load);
1259 // Is store the only read of the loaded value?
1260 if (!Load.hasOneUse())
1263 // Is the address of the store the same as the load?
1264 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
1265 LoadNode->getOffset() != StoreNode->getOffset())
1268 // Check if the chain is produced by the load or is a TokenFactor with
1269 // the load output chain as an operand. Return InputChain by reference.
1270 SDValue Chain = StoreNode->getChain();
1272 bool ChainCheck = false;
1273 if (Chain == Load.getValue(1)) {
1275 InputChain = LoadNode->getChain();
1276 } else if (Chain.getOpcode() == ISD::TokenFactor) {
1277 SmallVector<SDValue, 4> ChainOps;
1278 SmallVector<const SDNode *, 4> LoopWorklist;
1279 SmallPtrSet<const SDNode *, 16> Visited;
1280 const unsigned int Max = 1024;
1281 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
1282 SDValue Op = Chain.getOperand(i);
1283 if (Op == Load.getValue(1)) {
1285 // Drop Load, but keep its chain. No cycle check necessary.
1286 ChainOps.push_back(Load.getOperand(0));
1289 LoopWorklist.push_back(Op.getNode());
1290 ChainOps.push_back(Op);
1294 // Add the other operand of StoredVal to worklist.
1295 for (SDValue Op : StoredVal->ops())
1296 if (Op.getNode() != LoadNode)
1297 LoopWorklist.push_back(Op.getNode());
1299 // Check if Load is reachable from any of the nodes in the worklist.
1300 if (SDNode::hasPredecessorHelper(Load.getNode(), Visited, LoopWorklist, Max,
1304 // Make a new TokenFactor with all the other input chains except
1306 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
1307 MVT::Other, ChainOps);
1316 // Change a chain of {load; op; store} of the same value into a simple op
1317 // through memory of that value, if the uses of the modified value and its
1318 // address are suitable.
1320 // The tablegen pattern memory operand pattern is currently not able to match
1321 // the case where the CC on the original operation are used.
1323 // See the equivalent routine in X86ISelDAGToDAG for further comments.
1324 bool SystemZDAGToDAGISel::tryFoldLoadStoreIntoMemOperand(SDNode *Node) {
1325 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
1326 SDValue StoredVal = StoreNode->getOperand(1);
1327 unsigned Opc = StoredVal->getOpcode();
1328 SDLoc DL(StoreNode);
1330 // Before we try to select anything, make sure this is memory operand size
1331 // and opcode we can handle. Note that this must match the code below that
1332 // actually lowers the opcodes.
1333 EVT MemVT = StoreNode->getMemoryVT();
1334 unsigned NewOpc = 0;
1335 bool NegateOperand = false;
1339 case SystemZISD::SSUBO:
1340 NegateOperand = true;
1342 case SystemZISD::SADDO:
1343 if (MemVT == MVT::i32)
1344 NewOpc = SystemZ::ASI;
1345 else if (MemVT == MVT::i64)
1346 NewOpc = SystemZ::AGSI;
1350 case SystemZISD::USUBO:
1351 NegateOperand = true;
1353 case SystemZISD::UADDO:
1354 if (MemVT == MVT::i32)
1355 NewOpc = SystemZ::ALSI;
1356 else if (MemVT == MVT::i64)
1357 NewOpc = SystemZ::ALGSI;
1363 LoadSDNode *LoadNode = nullptr;
1365 if (!isFusableLoadOpStorePattern(StoreNode, StoredVal, CurDAG, LoadNode,
1369 SDValue Operand = StoredVal.getOperand(1);
1370 auto *OperandC = dyn_cast<ConstantSDNode>(Operand);
1373 auto OperandV = OperandC->getAPIntValue();
1375 OperandV = -OperandV;
1376 if (OperandV.getMinSignedBits() > 8)
1378 Operand = CurDAG->getTargetConstant(OperandV, DL, MemVT);
1381 if (!selectBDAddr20Only(StoreNode->getBasePtr(), Base, Disp))
1384 SDValue Ops[] = { Base, Disp, Operand, InputChain };
1385 MachineSDNode *Result =
1386 CurDAG->getMachineNode(NewOpc, DL, MVT::i32, MVT::Other, Ops);
1387 CurDAG->setNodeMemRefs(
1388 Result, {StoreNode->getMemOperand(), LoadNode->getMemOperand()});
1390 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
1391 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
1392 CurDAG->RemoveDeadNode(Node);
1396 bool SystemZDAGToDAGISel::canUseBlockOperation(StoreSDNode *Store,
1397 LoadSDNode *Load) const {
1398 // Check that the two memory operands have the same size.
1399 if (Load->getMemoryVT() != Store->getMemoryVT())
1402 // Volatility stops an access from being decomposed.
1403 if (Load->isVolatile() || Store->isVolatile())
1406 // There's no chance of overlap if the load is invariant.
1407 if (Load->isInvariant() && Load->isDereferenceable())
1410 // Otherwise we need to check whether there's an alias.
1411 const Value *V1 = Load->getMemOperand()->getValue();
1412 const Value *V2 = Store->getMemOperand()->getValue();
1417 uint64_t Size = Load->getMemoryVT().getStoreSize();
1418 int64_t End1 = Load->getSrcValueOffset() + Size;
1419 int64_t End2 = Store->getSrcValueOffset() + Size;
1420 if (V1 == V2 && End1 == End2)
1423 return !AA->alias(MemoryLocation(V1, End1, Load->getAAInfo()),
1424 MemoryLocation(V2, End2, Store->getAAInfo()));
1427 bool SystemZDAGToDAGISel::storeLoadCanUseMVC(SDNode *N) const {
1428 auto *Store = cast<StoreSDNode>(N);
1429 auto *Load = cast<LoadSDNode>(Store->getValue());
1431 // Prefer not to use MVC if either address can use ... RELATIVE LONG
1433 uint64_t Size = Load->getMemoryVT().getStoreSize();
1434 if (Size > 1 && Size <= 8) {
1435 // Prefer LHRL, LRL and LGRL.
1436 if (SystemZISD::isPCREL(Load->getBasePtr().getOpcode()))
1438 // Prefer STHRL, STRL and STGRL.
1439 if (SystemZISD::isPCREL(Store->getBasePtr().getOpcode()))
1443 return canUseBlockOperation(Store, Load);
1446 bool SystemZDAGToDAGISel::storeLoadCanUseBlockBinary(SDNode *N,
1448 auto *StoreA = cast<StoreSDNode>(N);
1449 auto *LoadA = cast<LoadSDNode>(StoreA->getValue().getOperand(1 - I));
1450 auto *LoadB = cast<LoadSDNode>(StoreA->getValue().getOperand(I));
1451 return !LoadA->isVolatile() && canUseBlockOperation(StoreA, LoadB);
1454 void SystemZDAGToDAGISel::Select(SDNode *Node) {
1455 // If we have a custom node, we already have selected!
1456 if (Node->isMachineOpcode()) {
1457 LLVM_DEBUG(errs() << "== "; Node->dump(CurDAG); errs() << "\n");
1458 Node->setNodeId(-1);
1462 unsigned Opcode = Node->getOpcode();
1465 if (Node->getOperand(1).getOpcode() != ISD::Constant)
1466 if (tryRxSBG(Node, SystemZ::ROSBG))
1471 if (Node->getOperand(1).getOpcode() != ISD::Constant)
1472 if (tryRxSBG(Node, SystemZ::RXSBG))
1476 // If this is a 64-bit operation in which both 32-bit halves are nonzero,
1477 // split the operation into two. If both operands here happen to be
1478 // constant, leave this to common code to optimize.
1479 if (Node->getValueType(0) == MVT::i64 &&
1480 Node->getOperand(0).getOpcode() != ISD::Constant)
1481 if (auto *Op1 = dyn_cast<ConstantSDNode>(Node->getOperand(1))) {
1482 uint64_t Val = Op1->getZExtValue();
1483 // Don't split the operation if we can match one of the combined
1484 // logical operations provided by miscellaneous-extensions-3.
1485 if (Subtarget->hasMiscellaneousExtensions3()) {
1486 unsigned ChildOpcode = Node->getOperand(0).getOpcode();
1487 // Check whether this expression matches NAND/NOR/NXOR.
1488 if (Val == (uint64_t)-1 && Opcode == ISD::XOR)
1489 if (ChildOpcode == ISD::AND || ChildOpcode == ISD::OR ||
1490 ChildOpcode == ISD::XOR)
1492 // Check whether this expression matches OR-with-complement.
1493 if (Opcode == ISD::OR && ChildOpcode == ISD::XOR) {
1494 auto Op0 = Node->getOperand(0);
1495 if (auto *Op0Op1 = dyn_cast<ConstantSDNode>(Op0->getOperand(1)))
1496 if (Op0Op1->getZExtValue() == (uint64_t)-1)
1500 if (!SystemZ::isImmLF(Val) && !SystemZ::isImmHF(Val)) {
1501 splitLargeImmediate(Opcode, Node, Node->getOperand(0),
1502 Val - uint32_t(Val), uint32_t(Val));
1509 if (Node->getOperand(1).getOpcode() != ISD::Constant)
1510 if (tryRxSBG(Node, SystemZ::RNSBG))
1516 case ISD::ZERO_EXTEND:
1517 if (tryRISBGZero(Node))
1522 // If this is a 64-bit constant that is out of the range of LLILF,
1523 // LLIHF and LGFI, split it into two 32-bit pieces.
1524 if (Node->getValueType(0) == MVT::i64) {
1525 uint64_t Val = cast<ConstantSDNode>(Node)->getZExtValue();
1526 if (!SystemZ::isImmLF(Val) && !SystemZ::isImmHF(Val) && !isInt<32>(Val)) {
1527 splitLargeImmediate(ISD::OR, Node, SDValue(), Val - uint32_t(Val),
1534 case SystemZISD::SELECT_CCMASK: {
1535 SDValue Op0 = Node->getOperand(0);
1536 SDValue Op1 = Node->getOperand(1);
1537 // Prefer to put any load first, so that it can be matched as a
1538 // conditional load. Likewise for constants in range for LOCHI.
1539 if ((Op1.getOpcode() == ISD::LOAD && Op0.getOpcode() != ISD::LOAD) ||
1540 (Subtarget->hasLoadStoreOnCond2() &&
1541 Node->getValueType(0).isInteger() &&
1542 Op1.getOpcode() == ISD::Constant &&
1543 isInt<16>(cast<ConstantSDNode>(Op1)->getSExtValue()) &&
1544 !(Op0.getOpcode() == ISD::Constant &&
1545 isInt<16>(cast<ConstantSDNode>(Op0)->getSExtValue())))) {
1546 SDValue CCValid = Node->getOperand(2);
1547 SDValue CCMask = Node->getOperand(3);
1548 uint64_t ConstCCValid =
1549 cast<ConstantSDNode>(CCValid.getNode())->getZExtValue();
1550 uint64_t ConstCCMask =
1551 cast<ConstantSDNode>(CCMask.getNode())->getZExtValue();
1552 // Invert the condition.
1553 CCMask = CurDAG->getConstant(ConstCCValid ^ ConstCCMask, SDLoc(Node),
1554 CCMask.getValueType());
1555 SDValue Op4 = Node->getOperand(4);
1556 SDNode *UpdatedNode =
1557 CurDAG->UpdateNodeOperands(Node, Op1, Op0, CCValid, CCMask, Op4);
1558 if (UpdatedNode != Node) {
1559 // In case this node already exists then replace Node with it.
1560 ReplaceNode(Node, UpdatedNode);
1567 case ISD::INSERT_VECTOR_ELT: {
1568 EVT VT = Node->getValueType(0);
1569 unsigned ElemBitSize = VT.getScalarSizeInBits();
1570 if (ElemBitSize == 32) {
1571 if (tryGather(Node, SystemZ::VGEF))
1573 } else if (ElemBitSize == 64) {
1574 if (tryGather(Node, SystemZ::VGEG))
1580 case ISD::BUILD_VECTOR: {
1581 auto *BVN = cast<BuildVectorSDNode>(Node);
1582 SystemZVectorConstantInfo VCI(BVN);
1583 if (VCI.isVectorConstantLegal(*Subtarget)) {
1584 loadVectorConstant(VCI, Node);
1590 case ISD::ConstantFP: {
1591 APFloat Imm = cast<ConstantFPSDNode>(Node)->getValueAPF();
1592 if (Imm.isZero() || Imm.isNegZero())
1594 SystemZVectorConstantInfo VCI(Imm);
1595 bool Success = VCI.isVectorConstantLegal(*Subtarget); (void)Success;
1596 assert(Success && "Expected legal FP immediate");
1597 loadVectorConstant(VCI, Node);
1602 if (tryFoldLoadStoreIntoMemOperand(Node))
1604 auto *Store = cast<StoreSDNode>(Node);
1605 unsigned ElemBitSize = Store->getValue().getValueSizeInBits();
1606 if (ElemBitSize == 32) {
1607 if (tryScatter(Store, SystemZ::VSCEF))
1609 } else if (ElemBitSize == 64) {
1610 if (tryScatter(Store, SystemZ::VSCEG))
1620 bool SystemZDAGToDAGISel::
1621 SelectInlineAsmMemoryOperand(const SDValue &Op,
1622 unsigned ConstraintID,
1623 std::vector<SDValue> &OutOps) {
1624 SystemZAddressingMode::AddrForm Form;
1625 SystemZAddressingMode::DispRange DispRange;
1626 SDValue Base, Disp, Index;
1628 switch(ConstraintID) {
1630 llvm_unreachable("Unexpected asm memory constraint");
1631 case InlineAsm::Constraint_i:
1632 case InlineAsm::Constraint_Q:
1633 // Accept an address with a short displacement, but no index.
1634 Form = SystemZAddressingMode::FormBD;
1635 DispRange = SystemZAddressingMode::Disp12Only;
1637 case InlineAsm::Constraint_R:
1638 // Accept an address with a short displacement and an index.
1639 Form = SystemZAddressingMode::FormBDXNormal;
1640 DispRange = SystemZAddressingMode::Disp12Only;
1642 case InlineAsm::Constraint_S:
1643 // Accept an address with a long displacement, but no index.
1644 Form = SystemZAddressingMode::FormBD;
1645 DispRange = SystemZAddressingMode::Disp20Only;
1647 case InlineAsm::Constraint_T:
1648 case InlineAsm::Constraint_m:
1649 case InlineAsm::Constraint_o:
1650 // Accept an address with a long displacement and an index.
1651 // m works the same as T, as this is the most general case.
1652 // We don't really have any special handling of "offsettable"
1653 // memory addresses, so just treat o the same as m.
1654 Form = SystemZAddressingMode::FormBDXNormal;
1655 DispRange = SystemZAddressingMode::Disp20Only;
1659 if (selectBDXAddr(Form, DispRange, Op, Base, Disp, Index)) {
1660 const TargetRegisterClass *TRC =
1661 Subtarget->getRegisterInfo()->getPointerRegClass(*MF);
1663 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), DL, MVT::i32);
1665 // Make sure that the base address doesn't go into %r0.
1666 // If it's a TargetFrameIndex or a fixed register, we shouldn't do anything.
1667 if (Base.getOpcode() != ISD::TargetFrameIndex &&
1668 Base.getOpcode() != ISD::Register) {
1670 SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1671 DL, Base.getValueType(),
1675 // Make sure that the index register isn't assigned to %r0 either.
1676 if (Index.getOpcode() != ISD::Register) {
1678 SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1679 DL, Index.getValueType(),
1683 OutOps.push_back(Base);
1684 OutOps.push_back(Disp);
1685 OutOps.push_back(Index);
1692 // IsProfitableToFold - Returns true if is profitable to fold the specific
1693 // operand node N of U during instruction selection that starts at Root.
1695 SystemZDAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
1696 SDNode *Root) const {
1697 // We want to avoid folding a LOAD into an ICMP node if as a result
1698 // we would be forced to spill the condition code into a GPR.
1699 if (N.getOpcode() == ISD::LOAD && U->getOpcode() == SystemZISD::ICMP) {
1700 if (!N.hasOneUse() || !U->hasOneUse())
1703 // The user of the CC value will usually be a CopyToReg into the
1704 // physical CC register, which in turn is glued and chained to the
1705 // actual instruction that uses the CC value. Bail out if we have
1706 // anything else than that.
1707 SDNode *CCUser = *U->use_begin();
1708 SDNode *CCRegUser = nullptr;
1709 if (CCUser->getOpcode() == ISD::CopyToReg ||
1710 cast<RegisterSDNode>(CCUser->getOperand(1))->getReg() == SystemZ::CC) {
1711 for (auto *U : CCUser->uses()) {
1712 if (CCRegUser == nullptr)
1714 else if (CCRegUser != U)
1718 if (CCRegUser == nullptr)
1721 // If the actual instruction is a branch, the only thing that remains to be
1722 // checked is whether the CCUser chain is a predecessor of the load.
1723 if (CCRegUser->isMachineOpcode() &&
1724 CCRegUser->getMachineOpcode() == SystemZ::BRC)
1725 return !N->isPredecessorOf(CCUser->getOperand(0).getNode());
1727 // Otherwise, the instruction may have multiple operands, and we need to
1728 // verify that none of them are a predecessor of the load. This is exactly
1729 // the same check that would be done by common code if the CC setter were
1730 // glued to the CC user, so simply invoke that check here.
1731 if (!IsLegalToFold(N, U, CCRegUser, OptLevel, false))
1739 // Represents a sequence for extracting a 0/1 value from an IPM result:
1740 // (((X ^ XORValue) + AddValue) >> Bit)
1741 struct IPMConversion {
1742 IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
1743 : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
1749 } // end anonymous namespace
1751 // Return a sequence for getting a 1 from an IPM result when CC has a
1752 // value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
1753 // The handling of CC values outside CCValid doesn't matter.
1754 static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
1755 // Deal with cases where the result can be taken directly from a bit
1756 // of the IPM result.
1757 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
1758 return IPMConversion(0, 0, SystemZ::IPM_CC);
1759 if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
1760 return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
1762 // Deal with cases where we can add a value to force the sign bit
1763 // to contain the right value. Putting the bit in 31 means we can
1764 // use SRL rather than RISBG(L), and also makes it easier to get a
1765 // 0/-1 value, so it has priority over the other tests below.
1767 // These sequences rely on the fact that the upper two bits of the
1768 // IPM result are zero.
1769 uint64_t TopBit = uint64_t(1) << 31;
1770 if (CCMask == (CCValid & SystemZ::CCMASK_0))
1771 return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
1772 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
1773 return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
1774 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1776 | SystemZ::CCMASK_2)))
1777 return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
1778 if (CCMask == (CCValid & SystemZ::CCMASK_3))
1779 return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
1780 if (CCMask == (CCValid & (SystemZ::CCMASK_1
1782 | SystemZ::CCMASK_3)))
1783 return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
1785 // Next try inverting the value and testing a bit. 0/1 could be
1786 // handled this way too, but we dealt with that case above.
1787 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
1788 return IPMConversion(-1, 0, SystemZ::IPM_CC);
1790 // Handle cases where adding a value forces a non-sign bit to contain
1792 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
1793 return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
1794 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
1795 return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
1797 // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
1798 // can be done by inverting the low CC bit and applying one of the
1799 // sign-based extractions above.
1800 if (CCMask == (CCValid & SystemZ::CCMASK_1))
1801 return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
1802 if (CCMask == (CCValid & SystemZ::CCMASK_2))
1803 return IPMConversion(1 << SystemZ::IPM_CC,
1804 TopBit - (3 << SystemZ::IPM_CC), 31);
1805 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1807 | SystemZ::CCMASK_3)))
1808 return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
1809 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1811 | SystemZ::CCMASK_3)))
1812 return IPMConversion(1 << SystemZ::IPM_CC,
1813 TopBit - (1 << SystemZ::IPM_CC), 31);
1815 llvm_unreachable("Unexpected CC combination");
1818 SDValue SystemZDAGToDAGISel::expandSelectBoolean(SDNode *Node) {
1819 auto *TrueOp = dyn_cast<ConstantSDNode>(Node->getOperand(0));
1820 auto *FalseOp = dyn_cast<ConstantSDNode>(Node->getOperand(1));
1821 if (!TrueOp || !FalseOp)
1823 if (FalseOp->getZExtValue() != 0)
1825 if (TrueOp->getSExtValue() != 1 && TrueOp->getSExtValue() != -1)
1828 auto *CCValidOp = dyn_cast<ConstantSDNode>(Node->getOperand(2));
1829 auto *CCMaskOp = dyn_cast<ConstantSDNode>(Node->getOperand(3));
1830 if (!CCValidOp || !CCMaskOp)
1832 int CCValid = CCValidOp->getZExtValue();
1833 int CCMask = CCMaskOp->getZExtValue();
1836 SDValue CCReg = Node->getOperand(4);
1837 IPMConversion IPM = getIPMConversion(CCValid, CCMask);
1838 SDValue Result = CurDAG->getNode(SystemZISD::IPM, DL, MVT::i32, CCReg);
1841 Result = CurDAG->getNode(ISD::XOR, DL, MVT::i32, Result,
1842 CurDAG->getConstant(IPM.XORValue, DL, MVT::i32));
1845 Result = CurDAG->getNode(ISD::ADD, DL, MVT::i32, Result,
1846 CurDAG->getConstant(IPM.AddValue, DL, MVT::i32));
1848 EVT VT = Node->getValueType(0);
1849 if (VT == MVT::i32 && IPM.Bit == 31) {
1850 unsigned ShiftOp = TrueOp->getSExtValue() == 1 ? ISD::SRL : ISD::SRA;
1851 Result = CurDAG->getNode(ShiftOp, DL, MVT::i32, Result,
1852 CurDAG->getConstant(IPM.Bit, DL, MVT::i32));
1855 Result = CurDAG->getNode(ISD::ANY_EXTEND, DL, VT, Result);
1857 if (TrueOp->getSExtValue() == 1) {
1858 // The SHR/AND sequence should get optimized to an RISBG.
1859 Result = CurDAG->getNode(ISD::SRL, DL, VT, Result,
1860 CurDAG->getConstant(IPM.Bit, DL, MVT::i32));
1861 Result = CurDAG->getNode(ISD::AND, DL, VT, Result,
1862 CurDAG->getConstant(1, DL, VT));
1864 // Sign-extend from IPM.Bit using a pair of shifts.
1865 int ShlAmt = VT.getSizeInBits() - 1 - IPM.Bit;
1866 int SraAmt = VT.getSizeInBits() - 1;
1867 Result = CurDAG->getNode(ISD::SHL, DL, VT, Result,
1868 CurDAG->getConstant(ShlAmt, DL, MVT::i32));
1869 Result = CurDAG->getNode(ISD::SRA, DL, VT, Result,
1870 CurDAG->getConstant(SraAmt, DL, MVT::i32));
1877 void SystemZDAGToDAGISel::PreprocessISelDAG() {
1878 // If we have conditional immediate loads, we always prefer
1879 // using those over an IPM sequence.
1880 if (Subtarget->hasLoadStoreOnCond2())
1883 bool MadeChange = false;
1885 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1886 E = CurDAG->allnodes_end();
1893 switch (N->getOpcode()) {
1895 case SystemZISD::SELECT_CCMASK:
1896 Res = expandSelectBoolean(N);
1901 LLVM_DEBUG(dbgs() << "SystemZ DAG preprocessing replacing:\nOld: ");
1902 LLVM_DEBUG(N->dump(CurDAG));
1903 LLVM_DEBUG(dbgs() << "\nNew: ");
1904 LLVM_DEBUG(Res.getNode()->dump(CurDAG));
1905 LLVM_DEBUG(dbgs() << "\n");
1907 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res);
1913 CurDAG->RemoveDeadNodes();