xref: /freebsd/contrib/llvm-project/llvm/lib/Target/SystemZ/SystemZISelLowering.h (revision f81cdf24ba5436367377f7c8e8f51f6df2a75ca7)
1 //===-- SystemZISelLowering.h - SystemZ DAG lowering interface --*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that SystemZ uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_SYSTEMZ_SYSTEMZISELLOWERING_H
15 #define LLVM_LIB_TARGET_SYSTEMZ_SYSTEMZISELLOWERING_H
16 
17 #include "SystemZ.h"
18 #include "SystemZInstrInfo.h"
19 #include "llvm/CodeGen/MachineBasicBlock.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/CodeGen/TargetLowering.h"
22 #include <optional>
23 
24 namespace llvm {
25 namespace SystemZISD {
26 enum NodeType : unsigned {
27   FIRST_NUMBER = ISD::BUILTIN_OP_END,
28 
29   // Return with a glue operand.  Operand 0 is the chain operand.
30   RET_GLUE,
31 
32   // Calls a function.  Operand 0 is the chain operand and operand 1
33   // is the target address.  The arguments start at operand 2.
34   // There is an optional glue operand at the end.
35   CALL,
36   SIBCALL,
37 
38   // TLS calls.  Like regular calls, except operand 1 is the TLS symbol.
39   // (The call target is implicitly __tls_get_offset.)
40   TLS_GDCALL,
41   TLS_LDCALL,
42 
43   // Wraps a TargetGlobalAddress that should be loaded using PC-relative
44   // accesses (LARL).  Operand 0 is the address.
45   PCREL_WRAPPER,
46 
47   // Used in cases where an offset is applied to a TargetGlobalAddress.
48   // Operand 0 is the full TargetGlobalAddress and operand 1 is a
49   // PCREL_WRAPPER for an anchor point.  This is used so that we can
50   // cheaply refer to either the full address or the anchor point
51   // as a register base.
52   PCREL_OFFSET,
53 
54   // Integer comparisons.  There are three operands: the two values
55   // to compare, and an integer of type SystemZICMP.
56   ICMP,
57 
58   // Floating-point comparisons.  The two operands are the values to compare.
59   FCMP,
60 
61   // Test under mask.  The first operand is ANDed with the second operand
62   // and the condition codes are set on the result.  The third operand is
63   // a boolean that is true if the condition codes need to distinguish
64   // between CCMASK_TM_MIXED_MSB_0 and CCMASK_TM_MIXED_MSB_1 (which the
65   // register forms do but the memory forms don't).
66   TM,
67 
68   // Branches if a condition is true.  Operand 0 is the chain operand;
69   // operand 1 is the 4-bit condition-code mask, with bit N in
70   // big-endian order meaning "branch if CC=N"; operand 2 is the
71   // target block and operand 3 is the flag operand.
72   BR_CCMASK,
73 
74   // Selects between operand 0 and operand 1.  Operand 2 is the
75   // mask of condition-code values for which operand 0 should be
76   // chosen over operand 1; it has the same form as BR_CCMASK.
77   // Operand 3 is the flag operand.
78   SELECT_CCMASK,
79 
80   // Evaluates to the gap between the stack pointer and the
81   // base of the dynamically-allocatable area.
82   ADJDYNALLOC,
83 
84   // For allocating stack space when using stack clash protector.
85   // Allocation is performed by block, and each block is probed.
86   PROBED_ALLOCA,
87 
88   // Count number of bits set in operand 0 per byte.
89   POPCNT,
90 
91   // Wrappers around the ISD opcodes of the same name.  The output is GR128.
92   // Input operands may be GR64 or GR32, depending on the instruction.
93   SMUL_LOHI,
94   UMUL_LOHI,
95   SDIVREM,
96   UDIVREM,
97 
98   // Add/subtract with overflow/carry.  These have the same operands as
99   // the corresponding standard operations, except with the carry flag
100   // replaced by a condition code value.
101   SADDO, SSUBO, UADDO, USUBO, ADDCARRY, SUBCARRY,
102 
103   // Set the condition code from a boolean value in operand 0.
104   // Operand 1 is a mask of all condition-code values that may result of this
105   // operation, operand 2 is a mask of condition-code values that may result
106   // if the boolean is true.
107   // Note that this operation is always optimized away, we will never
108   // generate any code for it.
109   GET_CCMASK,
110 
111   // Use a series of MVCs to copy bytes from one memory location to another.
112   // The operands are:
113   // - the target address
114   // - the source address
115   // - the constant length
116   //
117   // This isn't a memory opcode because we'd need to attach two
118   // MachineMemOperands rather than one.
119   MVC,
120 
121   // Similar to MVC, but for logic operations (AND, OR, XOR).
122   NC,
123   OC,
124   XC,
125 
126   // Use CLC to compare two blocks of memory, with the same comments
127   // as for MVC.
128   CLC,
129 
130   // Use MVC to set a block of memory after storing the first byte.
131   MEMSET_MVC,
132 
133   // Use an MVST-based sequence to implement stpcpy().
134   STPCPY,
135 
136   // Use a CLST-based sequence to implement strcmp().  The two input operands
137   // are the addresses of the strings to compare.
138   STRCMP,
139 
140   // Use an SRST-based sequence to search a block of memory.  The first
141   // operand is the end address, the second is the start, and the third
142   // is the character to search for.  CC is set to 1 on success and 2
143   // on failure.
144   SEARCH_STRING,
145 
146   // Store the CC value in bits 29 and 28 of an integer.
147   IPM,
148 
149   // Transaction begin.  The first operand is the chain, the second
150   // the TDB pointer, and the third the immediate control field.
151   // Returns CC value and chain.
152   TBEGIN,
153   TBEGIN_NOFLOAT,
154 
155   // Transaction end.  Just the chain operand.  Returns CC value and chain.
156   TEND,
157 
158   // Create a vector constant by filling byte N of the result with bit
159   // 15-N of the single operand.
160   BYTE_MASK,
161 
162   // Create a vector constant by replicating an element-sized RISBG-style mask.
163   // The first operand specifies the starting set bit and the second operand
164   // specifies the ending set bit.  Both operands count from the MSB of the
165   // element.
166   ROTATE_MASK,
167 
168   // Replicate a GPR scalar value into all elements of a vector.
169   REPLICATE,
170 
171   // Create a vector from two i64 GPRs.
172   JOIN_DWORDS,
173 
174   // Replicate one element of a vector into all elements.  The first operand
175   // is the vector and the second is the index of the element to replicate.
176   SPLAT,
177 
178   // Interleave elements from the high half of operand 0 and the high half
179   // of operand 1.
180   MERGE_HIGH,
181 
182   // Likewise for the low halves.
183   MERGE_LOW,
184 
185   // Concatenate the vectors in the first two operands, shift them left
186   // by the third operand, and take the first half of the result.
187   SHL_DOUBLE,
188 
189   // Take one element of the first v2i64 operand and the one element of
190   // the second v2i64 operand and concatenate them to form a v2i64 result.
191   // The third operand is a 4-bit value of the form 0A0B, where A and B
192   // are the element selectors for the first operand and second operands
193   // respectively.
194   PERMUTE_DWORDS,
195 
196   // Perform a general vector permute on vector operands 0 and 1.
197   // Each byte of operand 2 controls the corresponding byte of the result,
198   // in the same way as a byte-level VECTOR_SHUFFLE mask.
199   PERMUTE,
200 
201   // Pack vector operands 0 and 1 into a single vector with half-sized elements.
202   PACK,
203 
204   // Likewise, but saturate the result and set CC.  PACKS_CC does signed
205   // saturation and PACKLS_CC does unsigned saturation.
206   PACKS_CC,
207   PACKLS_CC,
208 
209   // Unpack the first half of vector operand 0 into double-sized elements.
210   // UNPACK_HIGH sign-extends and UNPACKL_HIGH zero-extends.
211   UNPACK_HIGH,
212   UNPACKL_HIGH,
213 
214   // Likewise for the second half.
215   UNPACK_LOW,
216   UNPACKL_LOW,
217 
218   // Shift each element of vector operand 0 by the number of bits specified
219   // by scalar operand 1.
220   VSHL_BY_SCALAR,
221   VSRL_BY_SCALAR,
222   VSRA_BY_SCALAR,
223 
224   // For each element of the output type, sum across all sub-elements of
225   // operand 0 belonging to the corresponding element, and add in the
226   // rightmost sub-element of the corresponding element of operand 1.
227   VSUM,
228 
229   // Compare integer vector operands 0 and 1 to produce the usual 0/-1
230   // vector result.  VICMPE is for equality, VICMPH for "signed greater than"
231   // and VICMPHL for "unsigned greater than".
232   VICMPE,
233   VICMPH,
234   VICMPHL,
235 
236   // Likewise, but also set the condition codes on the result.
237   VICMPES,
238   VICMPHS,
239   VICMPHLS,
240 
241   // Compare floating-point vector operands 0 and 1 to produce the usual 0/-1
242   // vector result.  VFCMPE is for "ordered and equal", VFCMPH for "ordered and
243   // greater than" and VFCMPHE for "ordered and greater than or equal to".
244   VFCMPE,
245   VFCMPH,
246   VFCMPHE,
247 
248   // Likewise, but also set the condition codes on the result.
249   VFCMPES,
250   VFCMPHS,
251   VFCMPHES,
252 
253   // Test floating-point data class for vectors.
254   VFTCI,
255 
256   // Extend the even f32 elements of vector operand 0 to produce a vector
257   // of f64 elements.
258   VEXTEND,
259 
260   // Round the f64 elements of vector operand 0 to f32s and store them in the
261   // even elements of the result.
262   VROUND,
263 
264   // AND the two vector operands together and set CC based on the result.
265   VTM,
266 
267   // String operations that set CC as a side-effect.
268   VFAE_CC,
269   VFAEZ_CC,
270   VFEE_CC,
271   VFEEZ_CC,
272   VFENE_CC,
273   VFENEZ_CC,
274   VISTR_CC,
275   VSTRC_CC,
276   VSTRCZ_CC,
277   VSTRS_CC,
278   VSTRSZ_CC,
279 
280   // Test Data Class.
281   //
282   // Operand 0: the value to test
283   // Operand 1: the bit mask
284   TDC,
285 
286   // z/OS XPLINK ADA Entry
287   // Wraps a TargetGlobalAddress that should be loaded from a function's
288   // AssociatedData Area (ADA). Tha ADA is passed to the function by the
289   // caller in the XPLink ABI defined register R5.
290   // Operand 0: the GlobalValue/External Symbol
291   // Operand 1: the ADA register
292   // Operand 2: the offset (0 for the first and 8 for the second element in the
293   // function descriptor)
294   ADA_ENTRY,
295 
296   // Strict variants of scalar floating-point comparisons.
297   // Quiet and signaling versions.
298   STRICT_FCMP = ISD::FIRST_TARGET_STRICTFP_OPCODE,
299   STRICT_FCMPS,
300 
301   // Strict variants of vector floating-point comparisons.
302   // Quiet and signaling versions.
303   STRICT_VFCMPE,
304   STRICT_VFCMPH,
305   STRICT_VFCMPHE,
306   STRICT_VFCMPES,
307   STRICT_VFCMPHS,
308   STRICT_VFCMPHES,
309 
310   // Strict variants of VEXTEND and VROUND.
311   STRICT_VEXTEND,
312   STRICT_VROUND,
313 
314   // Wrappers around the inner loop of an 8- or 16-bit ATOMIC_SWAP or
315   // ATOMIC_LOAD_<op>.
316   //
317   // Operand 0: the address of the containing 32-bit-aligned field
318   // Operand 1: the second operand of <op>, in the high bits of an i32
319   //            for everything except ATOMIC_SWAPW
320   // Operand 2: how many bits to rotate the i32 left to bring the first
321   //            operand into the high bits
322   // Operand 3: the negative of operand 2, for rotating the other way
323   // Operand 4: the width of the field in bits (8 or 16)
324   ATOMIC_SWAPW = ISD::FIRST_TARGET_MEMORY_OPCODE,
325   ATOMIC_LOADW_ADD,
326   ATOMIC_LOADW_SUB,
327   ATOMIC_LOADW_AND,
328   ATOMIC_LOADW_OR,
329   ATOMIC_LOADW_XOR,
330   ATOMIC_LOADW_NAND,
331   ATOMIC_LOADW_MIN,
332   ATOMIC_LOADW_MAX,
333   ATOMIC_LOADW_UMIN,
334   ATOMIC_LOADW_UMAX,
335 
336   // A wrapper around the inner loop of an ATOMIC_CMP_SWAP.
337   //
338   // Operand 0: the address of the containing 32-bit-aligned field
339   // Operand 1: the compare value, in the low bits of an i32
340   // Operand 2: the swap value, in the low bits of an i32
341   // Operand 3: how many bits to rotate the i32 left to bring the first
342   //            operand into the high bits
343   // Operand 4: the negative of operand 2, for rotating the other way
344   // Operand 5: the width of the field in bits (8 or 16)
345   ATOMIC_CMP_SWAPW,
346 
347   // Atomic compare-and-swap returning CC value.
348   // Val, CC, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmp, swap)
349   ATOMIC_CMP_SWAP,
350 
351   // 128-bit atomic load.
352   // Val, OUTCHAIN = ATOMIC_LOAD_128(INCHAIN, ptr)
353   ATOMIC_LOAD_128,
354 
355   // 128-bit atomic store.
356   // OUTCHAIN = ATOMIC_STORE_128(INCHAIN, val, ptr)
357   ATOMIC_STORE_128,
358 
359   // 128-bit atomic compare-and-swap.
360   // Val, CC, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmp, swap)
361   ATOMIC_CMP_SWAP_128,
362 
363   // Byte swapping load/store.  Same operands as regular load/store.
364   LRV, STRV,
365 
366   // Element swapping load/store.  Same operands as regular load/store.
367   VLER, VSTER,
368 
369   // Prefetch from the second operand using the 4-bit control code in
370   // the first operand.  The code is 1 for a load prefetch and 2 for
371   // a store prefetch.
372   PREFETCH
373 };
374 
375 // Return true if OPCODE is some kind of PC-relative address.
376 inline bool isPCREL(unsigned Opcode) {
377   return Opcode == PCREL_WRAPPER || Opcode == PCREL_OFFSET;
378 }
379 } // end namespace SystemZISD
380 
381 namespace SystemZICMP {
382 // Describes whether an integer comparison needs to be signed or unsigned,
383 // or whether either type is OK.
384 enum {
385   Any,
386   UnsignedOnly,
387   SignedOnly
388 };
389 } // end namespace SystemZICMP
390 
391 class SystemZSubtarget;
392 
393 class SystemZTargetLowering : public TargetLowering {
394 public:
395   explicit SystemZTargetLowering(const TargetMachine &TM,
396                                  const SystemZSubtarget &STI);
397 
398   bool useSoftFloat() const override;
399 
400   // Override TargetLowering.
401   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
402     return MVT::i32;
403   }
404   MVT getVectorIdxTy(const DataLayout &DL) const override {
405     // Only the lower 12 bits of an element index are used, so we don't
406     // want to clobber the upper 32 bits of a GPR unnecessarily.
407     return MVT::i32;
408   }
409   TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT)
410     const override {
411     // Widen subvectors to the full width rather than promoting integer
412     // elements.  This is better because:
413     //
414     // (a) it means that we can handle the ABI for passing and returning
415     //     sub-128 vectors without having to handle them as legal types.
416     //
417     // (b) we don't have instructions to extend on load and truncate on store,
418     //     so promoting the integers is less efficient.
419     //
420     // (c) there are no multiplication instructions for the widest integer
421     //     type (v2i64).
422     if (VT.getScalarSizeInBits() % 8 == 0)
423       return TypeWidenVector;
424     return TargetLoweringBase::getPreferredVectorAction(VT);
425   }
426   unsigned
427   getNumRegisters(LLVMContext &Context, EVT VT,
428                   std::optional<MVT> RegisterVT) const override {
429     // i128 inline assembly operand.
430     if (VT == MVT::i128 && RegisterVT && *RegisterVT == MVT::Untyped)
431       return 1;
432     return TargetLowering::getNumRegisters(Context, VT);
433   }
434   bool isCheapToSpeculateCtlz(Type *) const override { return true; }
435   bool preferZeroCompareBranch() const override { return true; }
436   bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override {
437     ConstantInt* Mask = dyn_cast<ConstantInt>(AndI.getOperand(1));
438     return Mask && Mask->getValue().isIntN(16);
439   }
440   bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
441     return VT.isScalarInteger();
442   }
443   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &,
444                          EVT) const override;
445   bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
446                                   EVT VT) const override;
447   bool isFPImmLegal(const APFloat &Imm, EVT VT,
448                     bool ForCodeSize) const override;
449   bool ShouldShrinkFPConstant(EVT VT) const override {
450     // Do not shrink 64-bit FP constpool entries since LDEB is slower than
451     // LD, and having the full constant in memory enables reg/mem opcodes.
452     return VT != MVT::f64;
453   }
454   bool hasInlineStackProbe(const MachineFunction &MF) const override;
455   bool isLegalICmpImmediate(int64_t Imm) const override;
456   bool isLegalAddImmediate(int64_t Imm) const override;
457   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
458                              unsigned AS,
459                              Instruction *I = nullptr) const override;
460   bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, Align Alignment,
461                                       MachineMemOperand::Flags Flags,
462                                       unsigned *Fast) const override;
463   bool
464   findOptimalMemOpLowering(std::vector<EVT> &MemOps, unsigned Limit,
465                            const MemOp &Op, unsigned DstAS, unsigned SrcAS,
466                            const AttributeList &FuncAttributes) const override;
467   EVT getOptimalMemOpType(const MemOp &Op,
468                           const AttributeList &FuncAttributes) const override;
469   bool isTruncateFree(Type *, Type *) const override;
470   bool isTruncateFree(EVT, EVT) const override;
471 
472   bool shouldFormOverflowOp(unsigned Opcode, EVT VT,
473                             bool MathUsed) const override {
474     // Form add and sub with overflow intrinsics regardless of any extra
475     // users of the math result.
476     return VT == MVT::i32 || VT == MVT::i64;
477   }
478 
479   bool shouldConsiderGEPOffsetSplit() const override { return true; }
480 
481   const char *getTargetNodeName(unsigned Opcode) const override;
482   std::pair<unsigned, const TargetRegisterClass *>
483   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
484                                StringRef Constraint, MVT VT) const override;
485   TargetLowering::ConstraintType
486   getConstraintType(StringRef Constraint) const override;
487   TargetLowering::ConstraintWeight
488     getSingleConstraintMatchWeight(AsmOperandInfo &info,
489                                    const char *constraint) const override;
490   void LowerAsmOperandForConstraint(SDValue Op,
491                                     std::string &Constraint,
492                                     std::vector<SDValue> &Ops,
493                                     SelectionDAG &DAG) const override;
494 
495   unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
496     if (ConstraintCode.size() == 1) {
497       switch(ConstraintCode[0]) {
498       default:
499         break;
500       case 'o':
501         return InlineAsm::Constraint_o;
502       case 'Q':
503         return InlineAsm::Constraint_Q;
504       case 'R':
505         return InlineAsm::Constraint_R;
506       case 'S':
507         return InlineAsm::Constraint_S;
508       case 'T':
509         return InlineAsm::Constraint_T;
510       }
511     } else if (ConstraintCode.size() == 2 && ConstraintCode[0] == 'Z') {
512       switch (ConstraintCode[1]) {
513       default:
514         break;
515       case 'Q':
516         return InlineAsm::Constraint_ZQ;
517       case 'R':
518         return InlineAsm::Constraint_ZR;
519       case 'S':
520         return InlineAsm::Constraint_ZS;
521       case 'T':
522         return InlineAsm::Constraint_ZT;
523       }
524     }
525     return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
526   }
527 
528   Register getRegisterByName(const char *RegName, LLT VT,
529                              const MachineFunction &MF) const override;
530 
531   /// If a physical register, this returns the register that receives the
532   /// exception address on entry to an EH pad.
533   Register
534   getExceptionPointerRegister(const Constant *PersonalityFn) const override {
535     return SystemZ::R6D;
536   }
537 
538   /// If a physical register, this returns the register that receives the
539   /// exception typeid on entry to a landing pad.
540   Register
541   getExceptionSelectorRegister(const Constant *PersonalityFn) const override {
542     return SystemZ::R7D;
543   }
544 
545   /// Override to support customized stack guard loading.
546   bool useLoadStackGuardNode() const override {
547     return true;
548   }
549   void insertSSPDeclarations(Module &M) const override {
550   }
551 
552   MachineBasicBlock *
553   EmitInstrWithCustomInserter(MachineInstr &MI,
554                               MachineBasicBlock *BB) const override;
555   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
556   void LowerOperationWrapper(SDNode *N, SmallVectorImpl<SDValue> &Results,
557                              SelectionDAG &DAG) const override;
558   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
559                           SelectionDAG &DAG) const override;
560   const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
561   bool allowTruncateForTailCall(Type *, Type *) const override;
562   bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
563   bool splitValueIntoRegisterParts(
564       SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
565       unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC)
566       const override;
567   SDValue joinRegisterPartsIntoValue(
568       SelectionDAG & DAG, const SDLoc &DL, const SDValue *Parts,
569       unsigned NumParts, MVT PartVT, EVT ValueVT,
570       std::optional<CallingConv::ID> CC) const override;
571   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
572                                bool isVarArg,
573                                const SmallVectorImpl<ISD::InputArg> &Ins,
574                                const SDLoc &DL, SelectionDAG &DAG,
575                                SmallVectorImpl<SDValue> &InVals) const override;
576   SDValue LowerCall(CallLoweringInfo &CLI,
577                     SmallVectorImpl<SDValue> &InVals) const override;
578 
579   std::pair<SDValue, SDValue>
580   makeExternalCall(SDValue Chain, SelectionDAG &DAG, const char *CalleeName,
581                    EVT RetVT, ArrayRef<SDValue> Ops, CallingConv::ID CallConv,
582                    bool IsSigned, SDLoc DL, bool DoesNotReturn,
583                    bool IsReturnValueUsed) const;
584 
585   bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
586                       bool isVarArg,
587                       const SmallVectorImpl<ISD::OutputArg> &Outs,
588                       LLVMContext &Context) const override;
589   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
590                       const SmallVectorImpl<ISD::OutputArg> &Outs,
591                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
592                       SelectionDAG &DAG) const override;
593   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
594 
595   /// Determine which of the bits specified in Mask are known to be either
596   /// zero or one and return them in the KnownZero/KnownOne bitsets.
597   void computeKnownBitsForTargetNode(const SDValue Op,
598                                      KnownBits &Known,
599                                      const APInt &DemandedElts,
600                                      const SelectionDAG &DAG,
601                                      unsigned Depth = 0) const override;
602 
603   /// Determine the number of bits in the operation that are sign bits.
604   unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
605                                            const APInt &DemandedElts,
606                                            const SelectionDAG &DAG,
607                                            unsigned Depth) const override;
608 
609   bool isGuaranteedNotToBeUndefOrPoisonForTargetNode(
610       SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
611       bool PoisonOnly, unsigned Depth) const override;
612 
613   ISD::NodeType getExtendForAtomicOps() const override {
614     return ISD::ANY_EXTEND;
615   }
616   ISD::NodeType getExtendForAtomicCmpSwapArg() const override {
617     return ISD::ZERO_EXTEND;
618   }
619 
620   bool supportSwiftError() const override {
621     return true;
622   }
623 
624   unsigned getStackProbeSize(const MachineFunction &MF) const;
625 
626 private:
627   const SystemZSubtarget &Subtarget;
628 
629   // Implement LowerOperation for individual opcodes.
630   SDValue getVectorCmp(SelectionDAG &DAG, unsigned Opcode,
631                        const SDLoc &DL, EVT VT,
632                        SDValue CmpOp0, SDValue CmpOp1, SDValue Chain) const;
633   SDValue lowerVectorSETCC(SelectionDAG &DAG, const SDLoc &DL,
634                            EVT VT, ISD::CondCode CC,
635                            SDValue CmpOp0, SDValue CmpOp1,
636                            SDValue Chain = SDValue(),
637                            bool IsSignaling = false) const;
638   SDValue lowerSETCC(SDValue Op, SelectionDAG &DAG) const;
639   SDValue lowerSTRICT_FSETCC(SDValue Op, SelectionDAG &DAG,
640                              bool IsSignaling) const;
641   SDValue lowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
642   SDValue lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
643   SDValue lowerGlobalAddress(GlobalAddressSDNode *Node,
644                              SelectionDAG &DAG) const;
645   SDValue lowerTLSGetOffset(GlobalAddressSDNode *Node,
646                             SelectionDAG &DAG, unsigned Opcode,
647                             SDValue GOTOffset) const;
648   SDValue lowerThreadPointer(const SDLoc &DL, SelectionDAG &DAG) const;
649   SDValue lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
650                                 SelectionDAG &DAG) const;
651   SDValue lowerBlockAddress(BlockAddressSDNode *Node,
652                             SelectionDAG &DAG) const;
653   SDValue lowerJumpTable(JumpTableSDNode *JT, SelectionDAG &DAG) const;
654   SDValue lowerConstantPool(ConstantPoolSDNode *CP, SelectionDAG &DAG) const;
655   SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
656   SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
657   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
658   SDValue lowerVASTART_ELF(SDValue Op, SelectionDAG &DAG) const;
659   SDValue lowerVASTART_XPLINK(SDValue Op, SelectionDAG &DAG) const;
660   SDValue lowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
661   SDValue lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
662   SDValue lowerDYNAMIC_STACKALLOC_ELF(SDValue Op, SelectionDAG &DAG) const;
663   SDValue lowerDYNAMIC_STACKALLOC_XPLINK(SDValue Op, SelectionDAG &DAG) const;
664   SDValue lowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
665   SDValue lowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const;
666   SDValue lowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const;
667   SDValue lowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
668   SDValue lowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
669   SDValue lowerXALUO(SDValue Op, SelectionDAG &DAG) const;
670   SDValue lowerUADDSUBO_CARRY(SDValue Op, SelectionDAG &DAG) const;
671   SDValue lowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
672   SDValue lowerOR(SDValue Op, SelectionDAG &DAG) const;
673   SDValue lowerCTPOP(SDValue Op, SelectionDAG &DAG) const;
674   SDValue lowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const;
675   SDValue lowerATOMIC_LOAD(SDValue Op, SelectionDAG &DAG) const;
676   SDValue lowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) const;
677   SDValue lowerATOMIC_LOAD_OP(SDValue Op, SelectionDAG &DAG,
678                               unsigned Opcode) const;
679   SDValue lowerATOMIC_LOAD_SUB(SDValue Op, SelectionDAG &DAG) const;
680   SDValue lowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
681   SDValue lowerSTACKSAVE(SDValue Op, SelectionDAG &DAG) const;
682   SDValue lowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
683   SDValue lowerPREFETCH(SDValue Op, SelectionDAG &DAG) const;
684   SDValue lowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
685   SDValue lowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
686   bool isVectorElementLoad(SDValue Op) const;
687   SDValue buildVector(SelectionDAG &DAG, const SDLoc &DL, EVT VT,
688                       SmallVectorImpl<SDValue> &Elems) const;
689   SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
690   SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
691   SDValue lowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
692   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
693   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
694   SDValue lowerSIGN_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const;
695   SDValue lowerZERO_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const;
696   SDValue lowerShift(SDValue Op, SelectionDAG &DAG, unsigned ByScalar) const;
697   SDValue lowerIS_FPCLASS(SDValue Op, SelectionDAG &DAG) const;
698   SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
699 
700   bool canTreatAsByteVector(EVT VT) const;
701   SDValue combineExtract(const SDLoc &DL, EVT ElemVT, EVT VecVT, SDValue OrigOp,
702                          unsigned Index, DAGCombinerInfo &DCI,
703                          bool Force) const;
704   SDValue combineTruncateExtract(const SDLoc &DL, EVT TruncVT, SDValue Op,
705                                  DAGCombinerInfo &DCI) const;
706   SDValue combineZERO_EXTEND(SDNode *N, DAGCombinerInfo &DCI) const;
707   SDValue combineSIGN_EXTEND(SDNode *N, DAGCombinerInfo &DCI) const;
708   SDValue combineSIGN_EXTEND_INREG(SDNode *N, DAGCombinerInfo &DCI) const;
709   SDValue combineMERGE(SDNode *N, DAGCombinerInfo &DCI) const;
710   bool canLoadStoreByteSwapped(EVT VT) const;
711   SDValue combineLOAD(SDNode *N, DAGCombinerInfo &DCI) const;
712   SDValue combineSTORE(SDNode *N, DAGCombinerInfo &DCI) const;
713   SDValue combineVECTOR_SHUFFLE(SDNode *N, DAGCombinerInfo &DCI) const;
714   SDValue combineEXTRACT_VECTOR_ELT(SDNode *N, DAGCombinerInfo &DCI) const;
715   SDValue combineJOIN_DWORDS(SDNode *N, DAGCombinerInfo &DCI) const;
716   SDValue combineFP_ROUND(SDNode *N, DAGCombinerInfo &DCI) const;
717   SDValue combineFP_EXTEND(SDNode *N, DAGCombinerInfo &DCI) const;
718   SDValue combineINT_TO_FP(SDNode *N, DAGCombinerInfo &DCI) const;
719   SDValue combineBSWAP(SDNode *N, DAGCombinerInfo &DCI) const;
720   SDValue combineBR_CCMASK(SDNode *N, DAGCombinerInfo &DCI) const;
721   SDValue combineSELECT_CCMASK(SDNode *N, DAGCombinerInfo &DCI) const;
722   SDValue combineGET_CCMASK(SDNode *N, DAGCombinerInfo &DCI) const;
723   SDValue combineIntDIVREM(SDNode *N, DAGCombinerInfo &DCI) const;
724   SDValue combineINTRINSIC(SDNode *N, DAGCombinerInfo &DCI) const;
725 
726   SDValue unwrapAddress(SDValue N) const override;
727 
728   // If the last instruction before MBBI in MBB was some form of COMPARE,
729   // try to replace it with a COMPARE AND BRANCH just before MBBI.
730   // CCMask and Target are the BRC-like operands for the branch.
731   // Return true if the change was made.
732   bool convertPrevCompareToBranch(MachineBasicBlock *MBB,
733                                   MachineBasicBlock::iterator MBBI,
734                                   unsigned CCMask,
735                                   MachineBasicBlock *Target) const;
736 
737   // Implement EmitInstrWithCustomInserter for individual operation types.
738   MachineBasicBlock *emitSelect(MachineInstr &MI, MachineBasicBlock *BB) const;
739   MachineBasicBlock *emitCondStore(MachineInstr &MI, MachineBasicBlock *BB,
740                                    unsigned StoreOpcode, unsigned STOCOpcode,
741                                    bool Invert) const;
742   MachineBasicBlock *emitPair128(MachineInstr &MI,
743                                  MachineBasicBlock *MBB) const;
744   MachineBasicBlock *emitExt128(MachineInstr &MI, MachineBasicBlock *MBB,
745                                 bool ClearEven) const;
746   MachineBasicBlock *emitAtomicLoadBinary(MachineInstr &MI,
747                                           MachineBasicBlock *BB,
748                                           unsigned BinOpcode, unsigned BitSize,
749                                           bool Invert = false) const;
750   MachineBasicBlock *emitAtomicLoadMinMax(MachineInstr &MI,
751                                           MachineBasicBlock *MBB,
752                                           unsigned CompareOpcode,
753                                           unsigned KeepOldMask,
754                                           unsigned BitSize) const;
755   MachineBasicBlock *emitAtomicCmpSwapW(MachineInstr &MI,
756                                         MachineBasicBlock *BB) const;
757   MachineBasicBlock *emitMemMemWrapper(MachineInstr &MI, MachineBasicBlock *BB,
758                                        unsigned Opcode,
759                                        bool IsMemset = false) const;
760   MachineBasicBlock *emitStringWrapper(MachineInstr &MI, MachineBasicBlock *BB,
761                                        unsigned Opcode) const;
762   MachineBasicBlock *emitTransactionBegin(MachineInstr &MI,
763                                           MachineBasicBlock *MBB,
764                                           unsigned Opcode, bool NoFloat) const;
765   MachineBasicBlock *emitLoadAndTestCmp0(MachineInstr &MI,
766                                          MachineBasicBlock *MBB,
767                                          unsigned Opcode) const;
768   MachineBasicBlock *emitProbedAlloca(MachineInstr &MI,
769                                       MachineBasicBlock *MBB) const;
770 
771   SDValue getBackchainAddress(SDValue SP, SelectionDAG &DAG) const;
772 
773   MachineMemOperand::Flags
774   getTargetMMOFlags(const Instruction &I) const override;
775   const TargetRegisterClass *getRepRegClassFor(MVT VT) const override;
776 };
777 
778 struct SystemZVectorConstantInfo {
779 private:
780   APInt IntBits;             // The 128 bits as an integer.
781   APInt SplatBits;           // Smallest splat value.
782   APInt SplatUndef;          // Bits correspoding to undef operands of the BVN.
783   unsigned SplatBitSize = 0;
784   bool isFP128 = false;
785 public:
786   unsigned Opcode = 0;
787   SmallVector<unsigned, 2> OpVals;
788   MVT VecVT;
789   SystemZVectorConstantInfo(APInt IntImm);
790   SystemZVectorConstantInfo(APFloat FPImm)
791       : SystemZVectorConstantInfo(FPImm.bitcastToAPInt()) {
792     isFP128 = (&FPImm.getSemantics() == &APFloat::IEEEquad());
793   }
794   SystemZVectorConstantInfo(BuildVectorSDNode *BVN);
795   bool isVectorConstantLegal(const SystemZSubtarget &Subtarget);
796 };
797 
798 } // end namespace llvm
799 
800 #endif
801