1 //===-- SystemZISelLowering.cpp - SystemZ DAG lowering implementation -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SystemZTargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "systemz-lower"
16 #include "SystemZISelLowering.h"
17 #include "SystemZCallingConv.h"
18 #include "SystemZConstantPoolValue.h"
19 #include "SystemZMachineFunctionInfo.h"
20 #include "SystemZTargetMachine.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 // Represents a sequence for extracting a 0/1 value from an IPM result:
31 // (((X ^ XORValue) + AddValue) >> Bit)
32 struct IPMConversion {
33 IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
34 : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
41 // Represents information about a comparison.
43 Comparison(SDValue Op0In, SDValue Op1In)
44 : Op0(Op0In), Op1(Op1In), Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
46 // The operands to the comparison.
49 // The opcode that should be used to compare Op0 and Op1.
52 // A SystemZICMP value. Only used for integer comparisons.
55 // The mask of CC values that Opcode can produce.
58 // The mask of CC values for which the original condition is true.
61 } // end anonymous namespace
63 // Classify VT as either 32 or 64 bit.
64 static bool is32Bit(EVT VT) {
65 switch (VT.getSimpleVT().SimpleTy) {
71 llvm_unreachable("Unsupported type");
75 // Return a version of MachineOperand that can be safely used before the
77 static MachineOperand earlyUseOperand(MachineOperand Op) {
83 SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm)
84 : TargetLowering(tm, new TargetLoweringObjectFileELF()),
85 Subtarget(*tm.getSubtargetImpl()), TM(tm) {
86 MVT PtrVT = getPointerTy();
88 // Set up the register classes.
89 if (Subtarget.hasHighWord())
90 addRegisterClass(MVT::i32, &SystemZ::GRX32BitRegClass);
92 addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
93 addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
94 addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
95 addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
96 addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
98 // Compute derived properties from the register classes
99 computeRegisterProperties();
101 // Set up special registers.
102 setExceptionPointerRegister(SystemZ::R6D);
103 setExceptionSelectorRegister(SystemZ::R7D);
104 setStackPointerRegisterToSaveRestore(SystemZ::R15D);
106 // TODO: It may be better to default to latency-oriented scheduling, however
107 // LLVM's current latency-oriented scheduler can't handle physreg definitions
108 // such as SystemZ has with CC, so set this to the register-pressure
109 // scheduler, because it can.
110 setSchedulingPreference(Sched::RegPressure);
112 setBooleanContents(ZeroOrOneBooleanContent);
113 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
115 // Instructions are strings of 2-byte aligned 2-byte values.
116 setMinFunctionAlignment(2);
118 // Handle operations that are handled in a similar way for all types.
119 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
120 I <= MVT::LAST_FP_VALUETYPE;
122 MVT VT = MVT::SimpleValueType(I);
123 if (isTypeLegal(VT)) {
124 // Lower SET_CC into an IPM-based sequence.
125 setOperationAction(ISD::SETCC, VT, Custom);
127 // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
128 setOperationAction(ISD::SELECT, VT, Expand);
130 // Lower SELECT_CC and BR_CC into separate comparisons and branches.
131 setOperationAction(ISD::SELECT_CC, VT, Custom);
132 setOperationAction(ISD::BR_CC, VT, Custom);
136 // Expand jump table branches as address arithmetic followed by an
138 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
140 // Expand BRCOND into a BR_CC (see above).
141 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
143 // Handle integer types.
144 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
145 I <= MVT::LAST_INTEGER_VALUETYPE;
147 MVT VT = MVT::SimpleValueType(I);
148 if (isTypeLegal(VT)) {
149 // Expand individual DIV and REMs into DIVREMs.
150 setOperationAction(ISD::SDIV, VT, Expand);
151 setOperationAction(ISD::UDIV, VT, Expand);
152 setOperationAction(ISD::SREM, VT, Expand);
153 setOperationAction(ISD::UREM, VT, Expand);
154 setOperationAction(ISD::SDIVREM, VT, Custom);
155 setOperationAction(ISD::UDIVREM, VT, Custom);
157 // Lower ATOMIC_LOAD and ATOMIC_STORE into normal volatile loads and
158 // stores, putting a serialization instruction after the stores.
159 setOperationAction(ISD::ATOMIC_LOAD, VT, Custom);
160 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
162 // Lower ATOMIC_LOAD_SUB into ATOMIC_LOAD_ADD if LAA and LAAG are
163 // available, or if the operand is constant.
164 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
166 // No special instructions for these.
167 setOperationAction(ISD::CTPOP, VT, Expand);
168 setOperationAction(ISD::CTTZ, VT, Expand);
169 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
170 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
171 setOperationAction(ISD::ROTR, VT, Expand);
173 // Use *MUL_LOHI where possible instead of MULH*.
174 setOperationAction(ISD::MULHS, VT, Expand);
175 setOperationAction(ISD::MULHU, VT, Expand);
176 setOperationAction(ISD::SMUL_LOHI, VT, Custom);
177 setOperationAction(ISD::UMUL_LOHI, VT, Custom);
179 // We have instructions for signed but not unsigned FP conversion.
180 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
184 // Type legalization will convert 8- and 16-bit atomic operations into
185 // forms that operate on i32s (but still keeping the original memory VT).
186 // Lower them into full i32 operations.
187 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
188 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
189 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
190 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
191 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
192 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
193 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
194 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
195 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
196 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
197 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
198 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
200 // We have instructions for signed but not unsigned FP conversion.
201 // Handle unsigned 32-bit types as signed 64-bit types.
202 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
203 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
205 // We have native support for a 64-bit CTLZ, via FLOGR.
206 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
207 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
209 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
210 setOperationAction(ISD::OR, MVT::i64, Custom);
212 // Give LowerOperation the chance to optimize SIGN_EXTEND sequences.
213 setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
215 // FIXME: Can we support these natively?
216 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
217 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
218 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
220 // We have native instructions for i8, i16 and i32 extensions, but not i1.
221 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
222 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
223 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
224 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
226 // Handle the various types of symbolic address.
227 setOperationAction(ISD::ConstantPool, PtrVT, Custom);
228 setOperationAction(ISD::GlobalAddress, PtrVT, Custom);
229 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
230 setOperationAction(ISD::BlockAddress, PtrVT, Custom);
231 setOperationAction(ISD::JumpTable, PtrVT, Custom);
233 // We need to handle dynamic allocations specially because of the
234 // 160-byte area at the bottom of the stack.
235 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
237 // Use custom expanders so that we can force the function to use
239 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
240 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
242 // Handle prefetches with PFD or PFDRL.
243 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
245 // Handle floating-point types.
246 for (unsigned I = MVT::FIRST_FP_VALUETYPE;
247 I <= MVT::LAST_FP_VALUETYPE;
249 MVT VT = MVT::SimpleValueType(I);
250 if (isTypeLegal(VT)) {
251 // We can use FI for FRINT.
252 setOperationAction(ISD::FRINT, VT, Legal);
254 // We can use the extended form of FI for other rounding operations.
255 if (Subtarget.hasFPExtension()) {
256 setOperationAction(ISD::FNEARBYINT, VT, Legal);
257 setOperationAction(ISD::FFLOOR, VT, Legal);
258 setOperationAction(ISD::FCEIL, VT, Legal);
259 setOperationAction(ISD::FTRUNC, VT, Legal);
260 setOperationAction(ISD::FROUND, VT, Legal);
263 // No special instructions for these.
264 setOperationAction(ISD::FSIN, VT, Expand);
265 setOperationAction(ISD::FCOS, VT, Expand);
266 setOperationAction(ISD::FREM, VT, Expand);
270 // We have fused multiply-addition for f32 and f64 but not f128.
271 setOperationAction(ISD::FMA, MVT::f32, Legal);
272 setOperationAction(ISD::FMA, MVT::f64, Legal);
273 setOperationAction(ISD::FMA, MVT::f128, Expand);
275 // Needed so that we don't try to implement f128 constant loads using
276 // a load-and-extend of a f80 constant (in cases where the constant
277 // would fit in an f80).
278 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
280 // Floating-point truncation and stores need to be done separately.
281 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
282 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
283 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
285 // We have 64-bit FPR<->GPR moves, but need special handling for
287 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
288 setOperationAction(ISD::BITCAST, MVT::f32, Custom);
290 // VASTART and VACOPY need to deal with the SystemZ-specific varargs
291 // structure, but VAEND is a no-op.
292 setOperationAction(ISD::VASTART, MVT::Other, Custom);
293 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
294 setOperationAction(ISD::VAEND, MVT::Other, Expand);
296 // We want to use MVC in preference to even a single load/store pair.
297 MaxStoresPerMemcpy = 0;
298 MaxStoresPerMemcpyOptSize = 0;
300 // The main memset sequence is a byte store followed by an MVC.
301 // Two STC or MV..I stores win over that, but the kind of fused stores
302 // generated by target-independent code don't when the byte value is
303 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
304 // than "STC;MVC". Handle the choice in target-specific code instead.
305 MaxStoresPerMemset = 0;
306 MaxStoresPerMemsetOptSize = 0;
309 EVT SystemZTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
312 return VT.changeVectorElementTypeToInteger();
315 bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
316 VT = VT.getScalarType();
321 switch (VT.getSimpleVT().SimpleTy) {
334 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
335 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
336 return Imm.isZero() || Imm.isNegZero();
339 bool SystemZTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
342 // Unaligned accesses should never be slower than the expanded version.
343 // We check specifically for aligned accesses in the few cases where
344 // they are required.
350 bool SystemZTargetLowering::isLegalAddressingMode(const AddrMode &AM,
352 // Punt on globals for now, although they can be used in limited
353 // RELATIVE LONG cases.
357 // Require a 20-bit signed offset.
358 if (!isInt<20>(AM.BaseOffs))
361 // Indexing is OK but no scale factor can be applied.
362 return AM.Scale == 0 || AM.Scale == 1;
365 bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
366 if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
368 unsigned FromBits = FromType->getPrimitiveSizeInBits();
369 unsigned ToBits = ToType->getPrimitiveSizeInBits();
370 return FromBits > ToBits;
373 bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
374 if (!FromVT.isInteger() || !ToVT.isInteger())
376 unsigned FromBits = FromVT.getSizeInBits();
377 unsigned ToBits = ToVT.getSizeInBits();
378 return FromBits > ToBits;
381 //===----------------------------------------------------------------------===//
382 // Inline asm support
383 //===----------------------------------------------------------------------===//
385 TargetLowering::ConstraintType
386 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
387 if (Constraint.size() == 1) {
388 switch (Constraint[0]) {
389 case 'a': // Address register
390 case 'd': // Data register (equivalent to 'r')
391 case 'f': // Floating-point register
392 case 'h': // High-part register
393 case 'r': // General-purpose register
394 return C_RegisterClass;
396 case 'Q': // Memory with base and unsigned 12-bit displacement
397 case 'R': // Likewise, plus an index
398 case 'S': // Memory with base and signed 20-bit displacement
399 case 'T': // Likewise, plus an index
400 case 'm': // Equivalent to 'T'.
403 case 'I': // Unsigned 8-bit constant
404 case 'J': // Unsigned 12-bit constant
405 case 'K': // Signed 16-bit constant
406 case 'L': // Signed 20-bit displacement (on all targets we support)
407 case 'M': // 0x7fffffff
414 return TargetLowering::getConstraintType(Constraint);
417 TargetLowering::ConstraintWeight SystemZTargetLowering::
418 getSingleConstraintMatchWeight(AsmOperandInfo &info,
419 const char *constraint) const {
420 ConstraintWeight weight = CW_Invalid;
421 Value *CallOperandVal = info.CallOperandVal;
422 // If we don't have a value, we can't do a match,
423 // but allow it at the lowest weight.
424 if (CallOperandVal == NULL)
426 Type *type = CallOperandVal->getType();
427 // Look at the constraint type.
428 switch (*constraint) {
430 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
433 case 'a': // Address register
434 case 'd': // Data register (equivalent to 'r')
435 case 'h': // High-part register
436 case 'r': // General-purpose register
437 if (CallOperandVal->getType()->isIntegerTy())
438 weight = CW_Register;
441 case 'f': // Floating-point register
442 if (type->isFloatingPointTy())
443 weight = CW_Register;
446 case 'I': // Unsigned 8-bit constant
447 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
448 if (isUInt<8>(C->getZExtValue()))
449 weight = CW_Constant;
452 case 'J': // Unsigned 12-bit constant
453 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
454 if (isUInt<12>(C->getZExtValue()))
455 weight = CW_Constant;
458 case 'K': // Signed 16-bit constant
459 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
460 if (isInt<16>(C->getSExtValue()))
461 weight = CW_Constant;
464 case 'L': // Signed 20-bit displacement (on all targets we support)
465 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
466 if (isInt<20>(C->getSExtValue()))
467 weight = CW_Constant;
470 case 'M': // 0x7fffffff
471 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
472 if (C->getZExtValue() == 0x7fffffff)
473 weight = CW_Constant;
479 // Parse a "{tNNN}" register constraint for which the register type "t"
480 // has already been verified. MC is the class associated with "t" and
481 // Map maps 0-based register numbers to LLVM register numbers.
482 static std::pair<unsigned, const TargetRegisterClass *>
483 parseRegisterNumber(const std::string &Constraint,
484 const TargetRegisterClass *RC, const unsigned *Map) {
485 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
486 if (isdigit(Constraint[2])) {
487 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2);
488 unsigned Index = atoi(Suffix.c_str());
489 if (Index < 16 && Map[Index])
490 return std::make_pair(Map[Index], RC);
492 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
495 std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering::
496 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const {
497 if (Constraint.size() == 1) {
498 // GCC Constraint Letters
499 switch (Constraint[0]) {
501 case 'd': // Data register (equivalent to 'r')
502 case 'r': // General-purpose register
504 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
505 else if (VT == MVT::i128)
506 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
507 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
509 case 'a': // Address register
511 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
512 else if (VT == MVT::i128)
513 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
514 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
516 case 'h': // High-part register (an LLVM extension)
517 return std::make_pair(0U, &SystemZ::GRH32BitRegClass);
519 case 'f': // Floating-point register
521 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
522 else if (VT == MVT::f128)
523 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
524 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
527 if (Constraint[0] == '{') {
528 // We need to override the default register parsing for GPRs and FPRs
529 // because the interpretation depends on VT. The internal names of
530 // the registers are also different from the external names
531 // (F0D and F0S instead of F0, etc.).
532 if (Constraint[1] == 'r') {
534 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
535 SystemZMC::GR32Regs);
537 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
538 SystemZMC::GR128Regs);
539 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
540 SystemZMC::GR64Regs);
542 if (Constraint[1] == 'f') {
544 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
545 SystemZMC::FP32Regs);
547 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
548 SystemZMC::FP128Regs);
549 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
550 SystemZMC::FP64Regs);
553 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
556 void SystemZTargetLowering::
557 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
558 std::vector<SDValue> &Ops,
559 SelectionDAG &DAG) const {
560 // Only support length 1 constraints for now.
561 if (Constraint.length() == 1) {
562 switch (Constraint[0]) {
563 case 'I': // Unsigned 8-bit constant
564 if (auto *C = dyn_cast<ConstantSDNode>(Op))
565 if (isUInt<8>(C->getZExtValue()))
566 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
570 case 'J': // Unsigned 12-bit constant
571 if (auto *C = dyn_cast<ConstantSDNode>(Op))
572 if (isUInt<12>(C->getZExtValue()))
573 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
577 case 'K': // Signed 16-bit constant
578 if (auto *C = dyn_cast<ConstantSDNode>(Op))
579 if (isInt<16>(C->getSExtValue()))
580 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
584 case 'L': // Signed 20-bit displacement (on all targets we support)
585 if (auto *C = dyn_cast<ConstantSDNode>(Op))
586 if (isInt<20>(C->getSExtValue()))
587 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
591 case 'M': // 0x7fffffff
592 if (auto *C = dyn_cast<ConstantSDNode>(Op))
593 if (C->getZExtValue() == 0x7fffffff)
594 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
599 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
602 //===----------------------------------------------------------------------===//
603 // Calling conventions
604 //===----------------------------------------------------------------------===//
606 #include "SystemZGenCallingConv.inc"
608 bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
609 Type *ToType) const {
610 return isTruncateFree(FromType, ToType);
613 bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
614 if (!CI->isTailCall())
619 // Value is a value that has been passed to us in the location described by VA
620 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
621 // any loads onto Chain.
622 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
623 CCValAssign &VA, SDValue Chain,
625 // If the argument has been promoted from a smaller type, insert an
626 // assertion to capture this.
627 if (VA.getLocInfo() == CCValAssign::SExt)
628 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
629 DAG.getValueType(VA.getValVT()));
630 else if (VA.getLocInfo() == CCValAssign::ZExt)
631 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
632 DAG.getValueType(VA.getValVT()));
635 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
636 else if (VA.getLocInfo() == CCValAssign::Indirect)
637 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
638 MachinePointerInfo(), false, false, false, 0);
640 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
644 // Value is a value of type VA.getValVT() that we need to copy into
645 // the location described by VA. Return a copy of Value converted to
646 // VA.getValVT(). The caller is responsible for handling indirect values.
647 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
648 CCValAssign &VA, SDValue Value) {
649 switch (VA.getLocInfo()) {
650 case CCValAssign::SExt:
651 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
652 case CCValAssign::ZExt:
653 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
654 case CCValAssign::AExt:
655 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
656 case CCValAssign::Full:
659 llvm_unreachable("Unhandled getLocInfo()");
663 SDValue SystemZTargetLowering::
664 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
665 const SmallVectorImpl<ISD::InputArg> &Ins,
666 SDLoc DL, SelectionDAG &DAG,
667 SmallVectorImpl<SDValue> &InVals) const {
668 MachineFunction &MF = DAG.getMachineFunction();
669 MachineFrameInfo *MFI = MF.getFrameInfo();
670 MachineRegisterInfo &MRI = MF.getRegInfo();
671 SystemZMachineFunctionInfo *FuncInfo =
672 MF.getInfo<SystemZMachineFunctionInfo>();
673 auto *TFL = static_cast<const SystemZFrameLowering *>(TM.getFrameLowering());
675 // Assign locations to all of the incoming arguments.
676 SmallVector<CCValAssign, 16> ArgLocs;
677 CCState CCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
678 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
680 unsigned NumFixedGPRs = 0;
681 unsigned NumFixedFPRs = 0;
682 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
684 CCValAssign &VA = ArgLocs[I];
685 EVT LocVT = VA.getLocVT();
687 // Arguments passed in registers
688 const TargetRegisterClass *RC;
689 switch (LocVT.getSimpleVT().SimpleTy) {
691 // Integers smaller than i64 should be promoted to i64.
692 llvm_unreachable("Unexpected argument type");
695 RC = &SystemZ::GR32BitRegClass;
699 RC = &SystemZ::GR64BitRegClass;
703 RC = &SystemZ::FP32BitRegClass;
707 RC = &SystemZ::FP64BitRegClass;
711 unsigned VReg = MRI.createVirtualRegister(RC);
712 MRI.addLiveIn(VA.getLocReg(), VReg);
713 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
715 assert(VA.isMemLoc() && "Argument not register or memory");
717 // Create the frame index object for this incoming parameter.
718 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
719 VA.getLocMemOffset(), true);
721 // Create the SelectionDAG nodes corresponding to a load
722 // from this parameter. Unpromoted ints and floats are
723 // passed as right-justified 8-byte values.
724 EVT PtrVT = getPointerTy();
725 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
726 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
727 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4));
728 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
729 MachinePointerInfo::getFixedStack(FI),
730 false, false, false, 0);
733 // Convert the value of the argument register into the value that's
735 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
739 // Save the number of non-varargs registers for later use by va_start, etc.
740 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
741 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
743 // Likewise the address (in the form of a frame index) of where the
744 // first stack vararg would be. The 1-byte size here is arbitrary.
745 int64_t StackSize = CCInfo.getNextStackOffset();
746 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
748 // ...and a similar frame index for the caller-allocated save area
749 // that will be used to store the incoming registers.
750 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
751 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
752 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
754 // Store the FPR varargs in the reserved frame slots. (We store the
755 // GPRs as part of the prologue.)
756 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
757 SDValue MemOps[SystemZ::NumArgFPRs];
758 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
759 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
760 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
761 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
762 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
763 &SystemZ::FP64BitRegClass);
764 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
765 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
766 MachinePointerInfo::getFixedStack(FI),
770 // Join the stores, which are independent of one another.
771 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
772 &MemOps[NumFixedFPRs],
773 SystemZ::NumArgFPRs - NumFixedFPRs);
780 static bool canUseSiblingCall(CCState ArgCCInfo,
781 SmallVectorImpl<CCValAssign> &ArgLocs) {
782 // Punt if there are any indirect or stack arguments, or if the call
783 // needs the call-saved argument register R6.
784 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
785 CCValAssign &VA = ArgLocs[I];
786 if (VA.getLocInfo() == CCValAssign::Indirect)
790 unsigned Reg = VA.getLocReg();
791 if (Reg == SystemZ::R6H || Reg == SystemZ::R6L || Reg == SystemZ::R6D)
798 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
799 SmallVectorImpl<SDValue> &InVals) const {
800 SelectionDAG &DAG = CLI.DAG;
802 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
803 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
804 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
805 SDValue Chain = CLI.Chain;
806 SDValue Callee = CLI.Callee;
807 bool &IsTailCall = CLI.IsTailCall;
808 CallingConv::ID CallConv = CLI.CallConv;
809 bool IsVarArg = CLI.IsVarArg;
810 MachineFunction &MF = DAG.getMachineFunction();
811 EVT PtrVT = getPointerTy();
813 // Analyze the operands of the call, assigning locations to each operand.
814 SmallVector<CCValAssign, 16> ArgLocs;
815 CCState ArgCCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
816 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
818 // We don't support GuaranteedTailCallOpt, only automatically-detected
820 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs))
823 // Get a count of how many bytes are to be pushed on the stack.
824 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
826 // Mark the start of the call.
828 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true),
831 // Copy argument values to their designated locations.
832 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
833 SmallVector<SDValue, 8> MemOpChains;
835 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
836 CCValAssign &VA = ArgLocs[I];
837 SDValue ArgValue = OutVals[I];
839 if (VA.getLocInfo() == CCValAssign::Indirect) {
840 // Store the argument in a stack slot and pass its address.
841 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
842 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
843 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
844 MachinePointerInfo::getFixedStack(FI),
846 ArgValue = SpillSlot;
848 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
851 // Queue up the argument copies and emit them at the end.
852 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
854 assert(VA.isMemLoc() && "Argument not register or memory");
856 // Work out the address of the stack slot. Unpromoted ints and
857 // floats are passed as right-justified 8-byte values.
858 if (!StackPtr.getNode())
859 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
860 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
861 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
863 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
864 DAG.getIntPtrConstant(Offset));
867 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
868 MachinePointerInfo(),
873 // Join the stores, which are independent of one another.
874 if (!MemOpChains.empty())
875 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
876 &MemOpChains[0], MemOpChains.size());
878 // Accept direct calls by converting symbolic call addresses to the
879 // associated Target* opcodes. Force %r1 to be used for indirect
882 if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
883 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
884 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
885 } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
886 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
887 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
888 } else if (IsTailCall) {
889 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
890 Glue = Chain.getValue(1);
891 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
894 // Build a sequence of copy-to-reg nodes, chained and glued together.
895 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
896 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
897 RegsToPass[I].second, Glue);
898 Glue = Chain.getValue(1);
901 // The first call operand is the chain and the second is the target address.
902 SmallVector<SDValue, 8> Ops;
903 Ops.push_back(Chain);
904 Ops.push_back(Callee);
906 // Add argument registers to the end of the list so that they are
907 // known live into the call.
908 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
909 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
910 RegsToPass[I].second.getValueType()));
912 // Glue the call to the argument copies, if any.
917 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
919 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, &Ops[0], Ops.size());
920 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
921 Glue = Chain.getValue(1);
923 // Mark the end of the call, which is glued to the call itself.
924 Chain = DAG.getCALLSEQ_END(Chain,
925 DAG.getConstant(NumBytes, PtrVT, true),
926 DAG.getConstant(0, PtrVT, true),
928 Glue = Chain.getValue(1);
930 // Assign locations to each value returned by this call.
931 SmallVector<CCValAssign, 16> RetLocs;
932 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
933 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
935 // Copy all of the result registers out of their specified physreg.
936 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
937 CCValAssign &VA = RetLocs[I];
939 // Copy the value out, gluing the copy to the end of the call sequence.
940 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
941 VA.getLocVT(), Glue);
942 Chain = RetValue.getValue(1);
943 Glue = RetValue.getValue(2);
945 // Convert the value of the return register into the value that's
947 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
954 SystemZTargetLowering::LowerReturn(SDValue Chain,
955 CallingConv::ID CallConv, bool IsVarArg,
956 const SmallVectorImpl<ISD::OutputArg> &Outs,
957 const SmallVectorImpl<SDValue> &OutVals,
958 SDLoc DL, SelectionDAG &DAG) const {
959 MachineFunction &MF = DAG.getMachineFunction();
961 // Assign locations to each returned value.
962 SmallVector<CCValAssign, 16> RetLocs;
963 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
964 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
966 // Quick exit for void returns
968 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
970 // Copy the result values into the output registers.
972 SmallVector<SDValue, 4> RetOps;
973 RetOps.push_back(Chain);
974 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
975 CCValAssign &VA = RetLocs[I];
976 SDValue RetValue = OutVals[I];
978 // Make the return register live on exit.
979 assert(VA.isRegLoc() && "Can only return in registers!");
981 // Promote the value as required.
982 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
984 // Chain and glue the copies together.
985 unsigned Reg = VA.getLocReg();
986 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
987 Glue = Chain.getValue(1);
988 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
991 // Update chain and glue.
994 RetOps.push_back(Glue);
996 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other,
997 RetOps.data(), RetOps.size());
1000 SDValue SystemZTargetLowering::
1001 prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, SelectionDAG &DAG) const {
1002 return DAG.getNode(SystemZISD::SERIALIZE, DL, MVT::Other, Chain);
1005 // CC is a comparison that will be implemented using an integer or
1006 // floating-point comparison. Return the condition code mask for
1007 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
1008 // unsigned comparisons and clear for signed ones. In the floating-point
1009 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
1010 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
1012 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
1013 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
1014 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
1018 llvm_unreachable("Invalid integer condition!");
1027 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
1028 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
1033 // Return a sequence for getting a 1 from an IPM result when CC has a
1034 // value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
1035 // The handling of CC values outside CCValid doesn't matter.
1036 static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
1037 // Deal with cases where the result can be taken directly from a bit
1038 // of the IPM result.
1039 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
1040 return IPMConversion(0, 0, SystemZ::IPM_CC);
1041 if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
1042 return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
1044 // Deal with cases where we can add a value to force the sign bit
1045 // to contain the right value. Putting the bit in 31 means we can
1046 // use SRL rather than RISBG(L), and also makes it easier to get a
1047 // 0/-1 value, so it has priority over the other tests below.
1049 // These sequences rely on the fact that the upper two bits of the
1050 // IPM result are zero.
1051 uint64_t TopBit = uint64_t(1) << 31;
1052 if (CCMask == (CCValid & SystemZ::CCMASK_0))
1053 return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
1054 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
1055 return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
1056 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1058 | SystemZ::CCMASK_2)))
1059 return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
1060 if (CCMask == (CCValid & SystemZ::CCMASK_3))
1061 return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
1062 if (CCMask == (CCValid & (SystemZ::CCMASK_1
1064 | SystemZ::CCMASK_3)))
1065 return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
1067 // Next try inverting the value and testing a bit. 0/1 could be
1068 // handled this way too, but we dealt with that case above.
1069 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
1070 return IPMConversion(-1, 0, SystemZ::IPM_CC);
1072 // Handle cases where adding a value forces a non-sign bit to contain
1074 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
1075 return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
1076 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
1077 return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
1079 // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
1080 // can be done by inverting the low CC bit and applying one of the
1081 // sign-based extractions above.
1082 if (CCMask == (CCValid & SystemZ::CCMASK_1))
1083 return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
1084 if (CCMask == (CCValid & SystemZ::CCMASK_2))
1085 return IPMConversion(1 << SystemZ::IPM_CC,
1086 TopBit - (3 << SystemZ::IPM_CC), 31);
1087 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1089 | SystemZ::CCMASK_3)))
1090 return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
1091 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1093 | SystemZ::CCMASK_3)))
1094 return IPMConversion(1 << SystemZ::IPM_CC,
1095 TopBit - (1 << SystemZ::IPM_CC), 31);
1097 llvm_unreachable("Unexpected CC combination");
1100 // If C can be converted to a comparison against zero, adjust the operands
1102 static void adjustZeroCmp(SelectionDAG &DAG, Comparison &C) {
1103 if (C.ICmpType == SystemZICMP::UnsignedOnly)
1106 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
1110 int64_t Value = ConstOp1->getSExtValue();
1111 if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
1112 (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
1113 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
1114 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
1115 C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1116 C.Op1 = DAG.getConstant(0, C.Op1.getValueType());
1120 // If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
1121 // adjust the operands as necessary.
1122 static void adjustSubwordCmp(SelectionDAG &DAG, Comparison &C) {
1123 // For us to make any changes, it must a comparison between a single-use
1124 // load and a constant.
1125 if (!C.Op0.hasOneUse() ||
1126 C.Op0.getOpcode() != ISD::LOAD ||
1127 C.Op1.getOpcode() != ISD::Constant)
1130 // We must have an 8- or 16-bit load.
1131 auto *Load = cast<LoadSDNode>(C.Op0);
1132 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
1133 if (NumBits != 8 && NumBits != 16)
1136 // The load must be an extending one and the constant must be within the
1137 // range of the unextended value.
1138 auto *ConstOp1 = cast<ConstantSDNode>(C.Op1);
1139 uint64_t Value = ConstOp1->getZExtValue();
1140 uint64_t Mask = (1 << NumBits) - 1;
1141 if (Load->getExtensionType() == ISD::SEXTLOAD) {
1142 // Make sure that ConstOp1 is in range of C.Op0.
1143 int64_t SignedValue = ConstOp1->getSExtValue();
1144 if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
1146 if (C.ICmpType != SystemZICMP::SignedOnly) {
1147 // Unsigned comparison between two sign-extended values is equivalent
1148 // to unsigned comparison between two zero-extended values.
1150 } else if (NumBits == 8) {
1151 // Try to treat the comparison as unsigned, so that we can use CLI.
1152 // Adjust CCMask and Value as necessary.
1153 if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
1154 // Test whether the high bit of the byte is set.
1155 Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
1156 else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
1157 // Test whether the high bit of the byte is clear.
1158 Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
1160 // No instruction exists for this combination.
1162 C.ICmpType = SystemZICMP::UnsignedOnly;
1164 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
1167 assert(C.ICmpType == SystemZICMP::Any &&
1168 "Signedness shouldn't matter here.");
1172 // Make sure that the first operand is an i32 of the right extension type.
1173 ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
1176 if (C.Op0.getValueType() != MVT::i32 ||
1177 Load->getExtensionType() != ExtType)
1178 C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
1179 Load->getChain(), Load->getBasePtr(),
1180 Load->getPointerInfo(), Load->getMemoryVT(),
1181 Load->isVolatile(), Load->isNonTemporal(),
1182 Load->getAlignment());
1184 // Make sure that the second operand is an i32 with the right value.
1185 if (C.Op1.getValueType() != MVT::i32 ||
1186 Value != ConstOp1->getZExtValue())
1187 C.Op1 = DAG.getConstant(Value, MVT::i32);
1190 // Return true if Op is either an unextended load, or a load suitable
1191 // for integer register-memory comparisons of type ICmpType.
1192 static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
1193 auto *Load = dyn_cast<LoadSDNode>(Op.getNode());
1195 // There are no instructions to compare a register with a memory byte.
1196 if (Load->getMemoryVT() == MVT::i8)
1198 // Otherwise decide on extension type.
1199 switch (Load->getExtensionType()) {
1200 case ISD::NON_EXTLOAD:
1203 return ICmpType != SystemZICMP::UnsignedOnly;
1205 return ICmpType != SystemZICMP::SignedOnly;
1213 // Return true if it is better to swap the operands of C.
1214 static bool shouldSwapCmpOperands(const Comparison &C) {
1215 // Leave f128 comparisons alone, since they have no memory forms.
1216 if (C.Op0.getValueType() == MVT::f128)
1219 // Always keep a floating-point constant second, since comparisons with
1220 // zero can use LOAD TEST and comparisons with other constants make a
1221 // natural memory operand.
1222 if (isa<ConstantFPSDNode>(C.Op1))
1225 // Never swap comparisons with zero since there are many ways to optimize
1227 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1228 if (ConstOp1 && ConstOp1->getZExtValue() == 0)
1231 // Also keep natural memory operands second if the loaded value is
1232 // only used here. Several comparisons have memory forms.
1233 if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
1236 // Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
1237 // In that case we generally prefer the memory to be second.
1238 if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
1239 // The only exceptions are when the second operand is a constant and
1240 // we can use things like CHHSI.
1243 // The unsigned memory-immediate instructions can handle 16-bit
1244 // unsigned integers.
1245 if (C.ICmpType != SystemZICMP::SignedOnly &&
1246 isUInt<16>(ConstOp1->getZExtValue()))
1248 // The signed memory-immediate instructions can handle 16-bit
1250 if (C.ICmpType != SystemZICMP::UnsignedOnly &&
1251 isInt<16>(ConstOp1->getSExtValue()))
1256 // Try to promote the use of CGFR and CLGFR.
1257 unsigned Opcode0 = C.Op0.getOpcode();
1258 if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
1260 if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
1262 if (C.ICmpType != SystemZICMP::SignedOnly &&
1263 Opcode0 == ISD::AND &&
1264 C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
1265 cast<ConstantSDNode>(C.Op0.getOperand(1))->getZExtValue() == 0xffffffff)
1271 // Return a version of comparison CC mask CCMask in which the LT and GT
1272 // actions are swapped.
1273 static unsigned reverseCCMask(unsigned CCMask) {
1274 return ((CCMask & SystemZ::CCMASK_CMP_EQ) |
1275 (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
1276 (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
1277 (CCMask & SystemZ::CCMASK_CMP_UO));
1280 // Check whether C tests for equality between X and Y and whether X - Y
1281 // or Y - X is also computed. In that case it's better to compare the
1282 // result of the subtraction against zero.
1283 static void adjustForSubtraction(SelectionDAG &DAG, Comparison &C) {
1284 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1285 C.CCMask == SystemZ::CCMASK_CMP_NE) {
1286 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1288 if (N->getOpcode() == ISD::SUB &&
1289 ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
1290 (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
1291 C.Op0 = SDValue(N, 0);
1292 C.Op1 = DAG.getConstant(0, N->getValueType(0));
1299 // Check whether C compares a floating-point value with zero and if that
1300 // floating-point value is also negated. In this case we can use the
1301 // negation to set CC, so avoiding separate LOAD AND TEST and
1302 // LOAD (NEGATIVE/COMPLEMENT) instructions.
1303 static void adjustForFNeg(Comparison &C) {
1304 auto *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
1305 if (C1 && C1->isZero()) {
1306 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1308 if (N->getOpcode() == ISD::FNEG) {
1309 C.Op0 = SDValue(N, 0);
1310 C.CCMask = reverseCCMask(C.CCMask);
1317 // Check whether C compares (shl X, 32) with 0 and whether X is
1318 // also sign-extended. In that case it is better to test the result
1319 // of the sign extension using LTGFR.
1321 // This case is important because InstCombine transforms a comparison
1322 // with (sext (trunc X)) into a comparison with (shl X, 32).
1323 static void adjustForLTGFR(Comparison &C) {
1324 // Check for a comparison between (shl X, 32) and 0.
1325 if (C.Op0.getOpcode() == ISD::SHL &&
1326 C.Op0.getValueType() == MVT::i64 &&
1327 C.Op1.getOpcode() == ISD::Constant &&
1328 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1329 auto *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
1330 if (C1 && C1->getZExtValue() == 32) {
1331 SDValue ShlOp0 = C.Op0.getOperand(0);
1332 // See whether X has any SIGN_EXTEND_INREG uses.
1333 for (auto I = ShlOp0->use_begin(), E = ShlOp0->use_end(); I != E; ++I) {
1335 if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
1336 cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
1337 C.Op0 = SDValue(N, 0);
1345 // If C compares the truncation of an extending load, try to compare
1346 // the untruncated value instead. This exposes more opportunities to
1348 static void adjustICmpTruncate(SelectionDAG &DAG, Comparison &C) {
1349 if (C.Op0.getOpcode() == ISD::TRUNCATE &&
1350 C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
1351 C.Op1.getOpcode() == ISD::Constant &&
1352 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1353 auto *L = cast<LoadSDNode>(C.Op0.getOperand(0));
1354 if (L->getMemoryVT().getStoreSizeInBits()
1355 <= C.Op0.getValueType().getSizeInBits()) {
1356 unsigned Type = L->getExtensionType();
1357 if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
1358 (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
1359 C.Op0 = C.Op0.getOperand(0);
1360 C.Op1 = DAG.getConstant(0, C.Op0.getValueType());
1366 // Return true if shift operation N has an in-range constant shift value.
1367 // Store it in ShiftVal if so.
1368 static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
1369 auto *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
1373 uint64_t Amount = Shift->getZExtValue();
1374 if (Amount >= N.getValueType().getSizeInBits())
1381 // Check whether an AND with Mask is suitable for a TEST UNDER MASK
1382 // instruction and whether the CC value is descriptive enough to handle
1383 // a comparison of type Opcode between the AND result and CmpVal.
1384 // CCMask says which comparison result is being tested and BitSize is
1385 // the number of bits in the operands. If TEST UNDER MASK can be used,
1386 // return the corresponding CC mask, otherwise return 0.
1387 static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask,
1388 uint64_t Mask, uint64_t CmpVal,
1389 unsigned ICmpType) {
1390 assert(Mask != 0 && "ANDs with zero should have been removed by now");
1392 // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.
1393 if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&
1394 !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))
1397 // Work out the masks for the lowest and highest bits.
1398 unsigned HighShift = 63 - countLeadingZeros(Mask);
1399 uint64_t High = uint64_t(1) << HighShift;
1400 uint64_t Low = uint64_t(1) << countTrailingZeros(Mask);
1402 // Signed ordered comparisons are effectively unsigned if the sign
1404 bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly);
1406 // Check for equality comparisons with 0, or the equivalent.
1408 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1409 return SystemZ::CCMASK_TM_ALL_0;
1410 if (CCMask == SystemZ::CCMASK_CMP_NE)
1411 return SystemZ::CCMASK_TM_SOME_1;
1413 if (EffectivelyUnsigned && CmpVal <= Low) {
1414 if (CCMask == SystemZ::CCMASK_CMP_LT)
1415 return SystemZ::CCMASK_TM_ALL_0;
1416 if (CCMask == SystemZ::CCMASK_CMP_GE)
1417 return SystemZ::CCMASK_TM_SOME_1;
1419 if (EffectivelyUnsigned && CmpVal < Low) {
1420 if (CCMask == SystemZ::CCMASK_CMP_LE)
1421 return SystemZ::CCMASK_TM_ALL_0;
1422 if (CCMask == SystemZ::CCMASK_CMP_GT)
1423 return SystemZ::CCMASK_TM_SOME_1;
1426 // Check for equality comparisons with the mask, or the equivalent.
1427 if (CmpVal == Mask) {
1428 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1429 return SystemZ::CCMASK_TM_ALL_1;
1430 if (CCMask == SystemZ::CCMASK_CMP_NE)
1431 return SystemZ::CCMASK_TM_SOME_0;
1433 if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) {
1434 if (CCMask == SystemZ::CCMASK_CMP_GT)
1435 return SystemZ::CCMASK_TM_ALL_1;
1436 if (CCMask == SystemZ::CCMASK_CMP_LE)
1437 return SystemZ::CCMASK_TM_SOME_0;
1439 if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) {
1440 if (CCMask == SystemZ::CCMASK_CMP_GE)
1441 return SystemZ::CCMASK_TM_ALL_1;
1442 if (CCMask == SystemZ::CCMASK_CMP_LT)
1443 return SystemZ::CCMASK_TM_SOME_0;
1446 // Check for ordered comparisons with the top bit.
1447 if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) {
1448 if (CCMask == SystemZ::CCMASK_CMP_LE)
1449 return SystemZ::CCMASK_TM_MSB_0;
1450 if (CCMask == SystemZ::CCMASK_CMP_GT)
1451 return SystemZ::CCMASK_TM_MSB_1;
1453 if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) {
1454 if (CCMask == SystemZ::CCMASK_CMP_LT)
1455 return SystemZ::CCMASK_TM_MSB_0;
1456 if (CCMask == SystemZ::CCMASK_CMP_GE)
1457 return SystemZ::CCMASK_TM_MSB_1;
1460 // If there are just two bits, we can do equality checks for Low and High
1462 if (Mask == Low + High) {
1463 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low)
1464 return SystemZ::CCMASK_TM_MIXED_MSB_0;
1465 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low)
1466 return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY;
1467 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High)
1468 return SystemZ::CCMASK_TM_MIXED_MSB_1;
1469 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High)
1470 return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY;
1473 // Looks like we've exhausted our options.
1477 // See whether C can be implemented as a TEST UNDER MASK instruction.
1478 // Update the arguments with the TM version if so.
1479 static void adjustForTestUnderMask(SelectionDAG &DAG, Comparison &C) {
1480 // Check that we have a comparison with a constant.
1481 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1484 uint64_t CmpVal = ConstOp1->getZExtValue();
1486 // Check whether the nonconstant input is an AND with a constant mask.
1489 ConstantSDNode *Mask = 0;
1490 if (C.Op0.getOpcode() == ISD::AND) {
1491 NewC.Op0 = C.Op0.getOperand(0);
1492 NewC.Op1 = C.Op0.getOperand(1);
1493 Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
1496 MaskVal = Mask->getZExtValue();
1498 // There is no instruction to compare with a 64-bit immediate
1499 // so use TMHH instead if possible. We need an unsigned ordered
1500 // comparison with an i64 immediate.
1501 if (NewC.Op0.getValueType() != MVT::i64 ||
1502 NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
1503 NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
1504 NewC.ICmpType == SystemZICMP::SignedOnly)
1506 // Convert LE and GT comparisons into LT and GE.
1507 if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
1508 NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
1509 if (CmpVal == uint64_t(-1))
1512 NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1514 // If the low N bits of Op1 are zero than the low N bits of Op0 can
1515 // be masked off without changing the result.
1516 MaskVal = -(CmpVal & -CmpVal);
1517 NewC.ICmpType = SystemZICMP::UnsignedOnly;
1520 // Check whether the combination of mask, comparison value and comparison
1521 // type are suitable.
1522 unsigned BitSize = NewC.Op0.getValueType().getSizeInBits();
1523 unsigned NewCCMask, ShiftVal;
1524 if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1525 NewC.Op0.getOpcode() == ISD::SHL &&
1526 isSimpleShift(NewC.Op0, ShiftVal) &&
1527 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1528 MaskVal >> ShiftVal,
1530 SystemZICMP::Any))) {
1531 NewC.Op0 = NewC.Op0.getOperand(0);
1532 MaskVal >>= ShiftVal;
1533 } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1534 NewC.Op0.getOpcode() == ISD::SRL &&
1535 isSimpleShift(NewC.Op0, ShiftVal) &&
1536 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1537 MaskVal << ShiftVal,
1539 SystemZICMP::UnsignedOnly))) {
1540 NewC.Op0 = NewC.Op0.getOperand(0);
1541 MaskVal <<= ShiftVal;
1543 NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
1549 // Go ahead and make the change.
1550 C.Opcode = SystemZISD::TM;
1552 if (Mask && Mask->getZExtValue() == MaskVal)
1553 C.Op1 = SDValue(Mask, 0);
1555 C.Op1 = DAG.getConstant(MaskVal, C.Op0.getValueType());
1556 C.CCValid = SystemZ::CCMASK_TM;
1557 C.CCMask = NewCCMask;
1560 // Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
1561 static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
1562 ISD::CondCode Cond) {
1563 Comparison C(CmpOp0, CmpOp1);
1564 C.CCMask = CCMaskForCondCode(Cond);
1565 if (C.Op0.getValueType().isFloatingPoint()) {
1566 C.CCValid = SystemZ::CCMASK_FCMP;
1567 C.Opcode = SystemZISD::FCMP;
1570 C.CCValid = SystemZ::CCMASK_ICMP;
1571 C.Opcode = SystemZISD::ICMP;
1572 // Choose the type of comparison. Equality and inequality tests can
1573 // use either signed or unsigned comparisons. The choice also doesn't
1574 // matter if both sign bits are known to be clear. In those cases we
1575 // want to give the main isel code the freedom to choose whichever
1577 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1578 C.CCMask == SystemZ::CCMASK_CMP_NE ||
1579 (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
1580 C.ICmpType = SystemZICMP::Any;
1581 else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
1582 C.ICmpType = SystemZICMP::UnsignedOnly;
1584 C.ICmpType = SystemZICMP::SignedOnly;
1585 C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
1586 adjustZeroCmp(DAG, C);
1587 adjustSubwordCmp(DAG, C);
1588 adjustForSubtraction(DAG, C);
1590 adjustICmpTruncate(DAG, C);
1593 if (shouldSwapCmpOperands(C)) {
1594 std::swap(C.Op0, C.Op1);
1595 C.CCMask = reverseCCMask(C.CCMask);
1598 adjustForTestUnderMask(DAG, C);
1602 // Emit the comparison instruction described by C.
1603 static SDValue emitCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1604 if (C.Opcode == SystemZISD::ICMP)
1605 return DAG.getNode(SystemZISD::ICMP, DL, MVT::Glue, C.Op0, C.Op1,
1606 DAG.getConstant(C.ICmpType, MVT::i32));
1607 if (C.Opcode == SystemZISD::TM) {
1608 bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
1609 bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
1610 return DAG.getNode(SystemZISD::TM, DL, MVT::Glue, C.Op0, C.Op1,
1611 DAG.getConstant(RegisterOnly, MVT::i32));
1613 return DAG.getNode(C.Opcode, DL, MVT::Glue, C.Op0, C.Op1);
1616 // Implement a 32-bit *MUL_LOHI operation by extending both operands to
1617 // 64 bits. Extend is the extension type to use. Store the high part
1618 // in Hi and the low part in Lo.
1619 static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL,
1620 unsigned Extend, SDValue Op0, SDValue Op1,
1621 SDValue &Hi, SDValue &Lo) {
1622 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
1623 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
1624 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
1625 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, DAG.getConstant(32, MVT::i64));
1626 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
1627 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
1630 // Lower a binary operation that produces two VT results, one in each
1631 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
1632 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
1633 // on the extended Op0 and (unextended) Op1. Store the even register result
1634 // in Even and the odd register result in Odd.
1635 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
1636 unsigned Extend, unsigned Opcode,
1637 SDValue Op0, SDValue Op1,
1638 SDValue &Even, SDValue &Odd) {
1639 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
1640 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
1641 SDValue(In128, 0), Op1);
1642 bool Is32Bit = is32Bit(VT);
1643 Even = DAG.getTargetExtractSubreg(SystemZ::even128(Is32Bit), DL, VT, Result);
1644 Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
1647 // Return an i32 value that is 1 if the CC value produced by Glue is
1648 // in the mask CCMask and 0 otherwise. CC is known to have a value
1649 // in CCValid, so other values can be ignored.
1650 static SDValue emitSETCC(SelectionDAG &DAG, SDLoc DL, SDValue Glue,
1651 unsigned CCValid, unsigned CCMask) {
1652 IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
1653 SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
1655 if (Conversion.XORValue)
1656 Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
1657 DAG.getConstant(Conversion.XORValue, MVT::i32));
1659 if (Conversion.AddValue)
1660 Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
1661 DAG.getConstant(Conversion.AddValue, MVT::i32));
1663 // The SHR/AND sequence should get optimized to an RISBG.
1664 Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
1665 DAG.getConstant(Conversion.Bit, MVT::i32));
1666 if (Conversion.Bit != 31)
1667 Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
1668 DAG.getConstant(1, MVT::i32));
1672 SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
1673 SelectionDAG &DAG) const {
1674 SDValue CmpOp0 = Op.getOperand(0);
1675 SDValue CmpOp1 = Op.getOperand(1);
1676 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1679 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1680 SDValue Glue = emitCmp(DAG, DL, C);
1681 return emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1684 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1685 SDValue Chain = Op.getOperand(0);
1686 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1687 SDValue CmpOp0 = Op.getOperand(2);
1688 SDValue CmpOp1 = Op.getOperand(3);
1689 SDValue Dest = Op.getOperand(4);
1692 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1693 SDValue Glue = emitCmp(DAG, DL, C);
1694 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
1695 Chain, DAG.getConstant(C.CCValid, MVT::i32),
1696 DAG.getConstant(C.CCMask, MVT::i32), Dest, Glue);
1699 // Return true if Pos is CmpOp and Neg is the negative of CmpOp,
1700 // allowing Pos and Neg to be wider than CmpOp.
1701 static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
1702 return (Neg.getOpcode() == ISD::SUB &&
1703 Neg.getOperand(0).getOpcode() == ISD::Constant &&
1704 cast<ConstantSDNode>(Neg.getOperand(0))->getZExtValue() == 0 &&
1705 Neg.getOperand(1) == Pos &&
1707 (Pos.getOpcode() == ISD::SIGN_EXTEND &&
1708 Pos.getOperand(0) == CmpOp)));
1711 // Return the absolute or negative absolute of Op; IsNegative decides which.
1712 static SDValue getAbsolute(SelectionDAG &DAG, SDLoc DL, SDValue Op,
1714 Op = DAG.getNode(SystemZISD::IABS, DL, Op.getValueType(), Op);
1716 Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
1717 DAG.getConstant(0, Op.getValueType()), Op);
1721 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
1722 SelectionDAG &DAG) const {
1723 SDValue CmpOp0 = Op.getOperand(0);
1724 SDValue CmpOp1 = Op.getOperand(1);
1725 SDValue TrueOp = Op.getOperand(2);
1726 SDValue FalseOp = Op.getOperand(3);
1727 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1730 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1732 // Check for absolute and negative-absolute selections, including those
1733 // where the comparison value is sign-extended (for LPGFR and LNGFR).
1734 // This check supplements the one in DAGCombiner.
1735 if (C.Opcode == SystemZISD::ICMP &&
1736 C.CCMask != SystemZ::CCMASK_CMP_EQ &&
1737 C.CCMask != SystemZ::CCMASK_CMP_NE &&
1738 C.Op1.getOpcode() == ISD::Constant &&
1739 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1740 if (isAbsolute(C.Op0, TrueOp, FalseOp))
1741 return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
1742 if (isAbsolute(C.Op0, FalseOp, TrueOp))
1743 return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
1746 SDValue Glue = emitCmp(DAG, DL, C);
1748 // Special case for handling -1/0 results. The shifts we use here
1749 // should get optimized with the IPM conversion sequence.
1750 auto *TrueC = dyn_cast<ConstantSDNode>(TrueOp);
1751 auto *FalseC = dyn_cast<ConstantSDNode>(FalseOp);
1752 if (TrueC && FalseC) {
1753 int64_t TrueVal = TrueC->getSExtValue();
1754 int64_t FalseVal = FalseC->getSExtValue();
1755 if ((TrueVal == -1 && FalseVal == 0) || (TrueVal == 0 && FalseVal == -1)) {
1756 // Invert the condition if we want -1 on false.
1758 C.CCMask ^= C.CCValid;
1759 SDValue Result = emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1760 EVT VT = Op.getValueType();
1761 // Extend the result to VT. Upper bits are ignored.
1763 Result = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Result);
1764 // Sign-extend from the low bit.
1765 SDValue ShAmt = DAG.getConstant(VT.getSizeInBits() - 1, MVT::i32);
1766 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Result, ShAmt);
1767 return DAG.getNode(ISD::SRA, DL, VT, Shl, ShAmt);
1771 SmallVector<SDValue, 5> Ops;
1772 Ops.push_back(TrueOp);
1773 Ops.push_back(FalseOp);
1774 Ops.push_back(DAG.getConstant(C.CCValid, MVT::i32));
1775 Ops.push_back(DAG.getConstant(C.CCMask, MVT::i32));
1776 Ops.push_back(Glue);
1778 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
1779 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size());
1782 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
1783 SelectionDAG &DAG) const {
1785 const GlobalValue *GV = Node->getGlobal();
1786 int64_t Offset = Node->getOffset();
1787 EVT PtrVT = getPointerTy();
1788 Reloc::Model RM = TM.getRelocationModel();
1789 CodeModel::Model CM = TM.getCodeModel();
1792 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
1793 // Assign anchors at 1<<12 byte boundaries.
1794 uint64_t Anchor = Offset & ~uint64_t(0xfff);
1795 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor);
1796 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1798 // The offset can be folded into the address if it is aligned to a halfword.
1800 if (Offset != 0 && (Offset & 1) == 0) {
1801 SDValue Full = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor + Offset);
1802 Result = DAG.getNode(SystemZISD::PCREL_OFFSET, DL, PtrVT, Full, Result);
1806 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
1807 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1808 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
1809 MachinePointerInfo::getGOT(), false, false, false, 0);
1812 // If there was a non-zero offset that we didn't fold, create an explicit
1815 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
1816 DAG.getConstant(Offset, PtrVT));
1821 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
1822 SelectionDAG &DAG) const {
1824 const GlobalValue *GV = Node->getGlobal();
1825 EVT PtrVT = getPointerTy();
1826 TLSModel::Model model = TM.getTLSModel(GV);
1828 if (model != TLSModel::LocalExec)
1829 llvm_unreachable("only local-exec TLS mode supported");
1831 // The high part of the thread pointer is in access register 0.
1832 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1833 DAG.getConstant(0, MVT::i32));
1834 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
1836 // The low part of the thread pointer is in access register 1.
1837 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1838 DAG.getConstant(1, MVT::i32));
1839 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
1841 // Merge them into a single 64-bit address.
1842 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
1843 DAG.getConstant(32, PtrVT));
1844 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
1846 // Get the offset of GA from the thread pointer.
1847 SystemZConstantPoolValue *CPV =
1848 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
1850 // Force the offset into the constant pool and load it from there.
1851 SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8);
1852 SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1853 CPAddr, MachinePointerInfo::getConstantPool(),
1854 false, false, false, 0);
1856 // Add the base and offset together.
1857 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
1860 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
1861 SelectionDAG &DAG) const {
1863 const BlockAddress *BA = Node->getBlockAddress();
1864 int64_t Offset = Node->getOffset();
1865 EVT PtrVT = getPointerTy();
1867 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
1868 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1872 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
1873 SelectionDAG &DAG) const {
1875 EVT PtrVT = getPointerTy();
1876 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1878 // Use LARL to load the address of the table.
1879 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1882 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
1883 SelectionDAG &DAG) const {
1885 EVT PtrVT = getPointerTy();
1888 if (CP->isMachineConstantPoolEntry())
1889 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1890 CP->getAlignment());
1892 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1893 CP->getAlignment(), CP->getOffset());
1895 // Use LARL to load the address of the constant pool entry.
1896 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1899 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
1900 SelectionDAG &DAG) const {
1902 SDValue In = Op.getOperand(0);
1903 EVT InVT = In.getValueType();
1904 EVT ResVT = Op.getValueType();
1906 if (InVT == MVT::i32 && ResVT == MVT::f32) {
1908 if (Subtarget.hasHighWord()) {
1909 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL,
1911 In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
1912 MVT::i64, SDValue(U64, 0), In);
1914 In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
1915 In64 = DAG.getNode(ISD::SHL, DL, MVT::i64, In64,
1916 DAG.getConstant(32, MVT::i64));
1918 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, In64);
1919 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32,
1920 DL, MVT::f32, Out64);
1922 if (InVT == MVT::f32 && ResVT == MVT::i32) {
1923 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
1924 SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
1925 MVT::f64, SDValue(U64, 0), In);
1926 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
1927 if (Subtarget.hasHighWord())
1928 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32, DL,
1930 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64,
1931 DAG.getConstant(32, MVT::i64));
1932 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
1934 llvm_unreachable("Unexpected bitcast combination");
1937 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
1938 SelectionDAG &DAG) const {
1939 MachineFunction &MF = DAG.getMachineFunction();
1940 SystemZMachineFunctionInfo *FuncInfo =
1941 MF.getInfo<SystemZMachineFunctionInfo>();
1942 EVT PtrVT = getPointerTy();
1944 SDValue Chain = Op.getOperand(0);
1945 SDValue Addr = Op.getOperand(1);
1946 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1949 // The initial values of each field.
1950 const unsigned NumFields = 4;
1951 SDValue Fields[NumFields] = {
1952 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT),
1953 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT),
1954 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
1955 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
1958 // Store each field into its respective slot.
1959 SDValue MemOps[NumFields];
1960 unsigned Offset = 0;
1961 for (unsigned I = 0; I < NumFields; ++I) {
1962 SDValue FieldAddr = Addr;
1964 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
1965 DAG.getIntPtrConstant(Offset));
1966 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
1967 MachinePointerInfo(SV, Offset),
1971 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps, NumFields);
1974 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
1975 SelectionDAG &DAG) const {
1976 SDValue Chain = Op.getOperand(0);
1977 SDValue DstPtr = Op.getOperand(1);
1978 SDValue SrcPtr = Op.getOperand(2);
1979 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
1980 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
1983 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32),
1984 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
1985 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
1988 SDValue SystemZTargetLowering::
1989 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
1990 SDValue Chain = Op.getOperand(0);
1991 SDValue Size = Op.getOperand(1);
1994 unsigned SPReg = getStackPointerRegisterToSaveRestore();
1996 // Get a reference to the stack pointer.
1997 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
1999 // Get the new stack pointer value.
2000 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
2002 // Copy the new stack pointer back.
2003 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
2005 // The allocated data lives above the 160 bytes allocated for the standard
2006 // frame, plus any outgoing stack arguments. We don't know how much that
2007 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
2008 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
2009 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
2011 SDValue Ops[2] = { Result, Chain };
2012 return DAG.getMergeValues(Ops, 2, DL);
2015 SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
2016 SelectionDAG &DAG) const {
2017 EVT VT = Op.getValueType();
2021 // Just do a normal 64-bit multiplication and extract the results.
2022 // We define this so that it can be used for constant division.
2023 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
2024 Op.getOperand(1), Ops[1], Ops[0]);
2026 // Do a full 128-bit multiplication based on UMUL_LOHI64:
2028 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
2030 // but using the fact that the upper halves are either all zeros
2033 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
2035 // and grouping the right terms together since they are quicker than the
2038 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
2039 SDValue C63 = DAG.getConstant(63, MVT::i64);
2040 SDValue LL = Op.getOperand(0);
2041 SDValue RL = Op.getOperand(1);
2042 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
2043 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
2044 // UMUL_LOHI64 returns the low result in the odd register and the high
2045 // result in the even register. SMUL_LOHI is defined to return the
2046 // low half first, so the results are in reverse order.
2047 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2048 LL, RL, Ops[1], Ops[0]);
2049 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
2050 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
2051 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
2052 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
2054 return DAG.getMergeValues(Ops, 2, DL);
2057 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
2058 SelectionDAG &DAG) const {
2059 EVT VT = Op.getValueType();
2063 // Just do a normal 64-bit multiplication and extract the results.
2064 // We define this so that it can be used for constant division.
2065 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
2066 Op.getOperand(1), Ops[1], Ops[0]);
2068 // UMUL_LOHI64 returns the low result in the odd register and the high
2069 // result in the even register. UMUL_LOHI is defined to return the
2070 // low half first, so the results are in reverse order.
2071 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2072 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2073 return DAG.getMergeValues(Ops, 2, DL);
2076 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
2077 SelectionDAG &DAG) const {
2078 SDValue Op0 = Op.getOperand(0);
2079 SDValue Op1 = Op.getOperand(1);
2080 EVT VT = Op.getValueType();
2084 // We use DSGF for 32-bit division.
2086 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
2087 Opcode = SystemZISD::SDIVREM32;
2088 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
2089 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
2090 Opcode = SystemZISD::SDIVREM32;
2092 Opcode = SystemZISD::SDIVREM64;
2094 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
2095 // input is "don't care". The instruction returns the remainder in
2096 // the even register and the quotient in the odd register.
2098 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
2099 Op0, Op1, Ops[1], Ops[0]);
2100 return DAG.getMergeValues(Ops, 2, DL);
2103 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
2104 SelectionDAG &DAG) const {
2105 EVT VT = Op.getValueType();
2108 // DL(G) uses a double-width dividend, so we need to clear the even
2109 // register in the GR128 input. The instruction returns the remainder
2110 // in the even register and the quotient in the odd register.
2113 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
2114 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2116 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
2117 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2118 return DAG.getMergeValues(Ops, 2, DL);
2121 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
2122 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
2124 // Get the known-zero masks for each operand.
2125 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
2126 APInt KnownZero[2], KnownOne[2];
2127 DAG.ComputeMaskedBits(Ops[0], KnownZero[0], KnownOne[0]);
2128 DAG.ComputeMaskedBits(Ops[1], KnownZero[1], KnownOne[1]);
2130 // See if the upper 32 bits of one operand and the lower 32 bits of the
2131 // other are known zero. They are the low and high operands respectively.
2132 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
2133 KnownZero[1].getZExtValue() };
2135 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
2137 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
2142 SDValue LowOp = Ops[Low];
2143 SDValue HighOp = Ops[High];
2145 // If the high part is a constant, we're better off using IILH.
2146 if (HighOp.getOpcode() == ISD::Constant)
2149 // If the low part is a constant that is outside the range of LHI,
2150 // then we're better off using IILF.
2151 if (LowOp.getOpcode() == ISD::Constant) {
2152 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
2153 if (!isInt<16>(Value))
2157 // Check whether the high part is an AND that doesn't change the
2158 // high 32 bits and just masks out low bits. We can skip it if so.
2159 if (HighOp.getOpcode() == ISD::AND &&
2160 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
2161 SDValue HighOp0 = HighOp.getOperand(0);
2162 uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
2163 if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
2167 // Take advantage of the fact that all GR32 operations only change the
2168 // low 32 bits by truncating Low to an i32 and inserting it directly
2169 // using a subreg. The interesting cases are those where the truncation
2172 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
2173 return DAG.getTargetInsertSubreg(SystemZ::subreg_l32, DL,
2174 MVT::i64, HighOp, Low32);
2177 SDValue SystemZTargetLowering::lowerSIGN_EXTEND(SDValue Op,
2178 SelectionDAG &DAG) const {
2179 // Convert (sext (ashr (shl X, C1), C2)) to
2180 // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
2181 // cheap as narrower ones.
2182 SDValue N0 = Op.getOperand(0);
2183 EVT VT = Op.getValueType();
2184 if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
2185 auto *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2186 SDValue Inner = N0.getOperand(0);
2187 if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
2188 auto *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1));
2190 unsigned Extra = (VT.getSizeInBits() -
2191 N0.getValueType().getSizeInBits());
2192 unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
2193 unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
2194 EVT ShiftVT = N0.getOperand(1).getValueType();
2195 SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
2196 Inner.getOperand(0));
2197 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
2198 DAG.getConstant(NewShlAmt, ShiftVT));
2199 return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
2200 DAG.getConstant(NewSraAmt, ShiftVT));
2207 // Op is an atomic load. Lower it into a normal volatile load.
2208 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
2209 SelectionDAG &DAG) const {
2210 auto *Node = cast<AtomicSDNode>(Op.getNode());
2211 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), Op.getValueType(),
2212 Node->getChain(), Node->getBasePtr(),
2213 Node->getMemoryVT(), Node->getMemOperand());
2216 // Op is an atomic store. Lower it into a normal volatile store followed
2217 // by a serialization.
2218 SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
2219 SelectionDAG &DAG) const {
2220 auto *Node = cast<AtomicSDNode>(Op.getNode());
2221 SDValue Chain = DAG.getTruncStore(Node->getChain(), SDLoc(Op), Node->getVal(),
2222 Node->getBasePtr(), Node->getMemoryVT(),
2223 Node->getMemOperand());
2224 return SDValue(DAG.getMachineNode(SystemZ::Serialize, SDLoc(Op), MVT::Other,
2228 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
2229 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
2230 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
2232 unsigned Opcode) const {
2233 auto *Node = cast<AtomicSDNode>(Op.getNode());
2235 // 32-bit operations need no code outside the main loop.
2236 EVT NarrowVT = Node->getMemoryVT();
2237 EVT WideVT = MVT::i32;
2238 if (NarrowVT == WideVT)
2241 int64_t BitSize = NarrowVT.getSizeInBits();
2242 SDValue ChainIn = Node->getChain();
2243 SDValue Addr = Node->getBasePtr();
2244 SDValue Src2 = Node->getVal();
2245 MachineMemOperand *MMO = Node->getMemOperand();
2247 EVT PtrVT = Addr.getValueType();
2249 // Convert atomic subtracts of constants into additions.
2250 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
2251 if (auto *Const = dyn_cast<ConstantSDNode>(Src2)) {
2252 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
2253 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
2256 // Get the address of the containing word.
2257 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2258 DAG.getConstant(-4, PtrVT));
2260 // Get the number of bits that the word must be rotated left in order
2261 // to bring the field to the top bits of a GR32.
2262 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2263 DAG.getConstant(3, PtrVT));
2264 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2266 // Get the complementing shift amount, for rotating a field in the top
2267 // bits back to its proper position.
2268 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2269 DAG.getConstant(0, WideVT), BitShift);
2271 // Extend the source operand to 32 bits and prepare it for the inner loop.
2272 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
2273 // operations require the source to be shifted in advance. (This shift
2274 // can be folded if the source is constant.) For AND and NAND, the lower
2275 // bits must be set, while for other opcodes they should be left clear.
2276 if (Opcode != SystemZISD::ATOMIC_SWAPW)
2277 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
2278 DAG.getConstant(32 - BitSize, WideVT));
2279 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
2280 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
2281 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
2282 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT));
2284 // Construct the ATOMIC_LOADW_* node.
2285 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2286 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
2287 DAG.getConstant(BitSize, WideVT) };
2288 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
2289 array_lengthof(Ops),
2292 // Rotate the result of the final CS so that the field is in the lower
2293 // bits of a GR32, then truncate it.
2294 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
2295 DAG.getConstant(BitSize, WideVT));
2296 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
2298 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
2299 return DAG.getMergeValues(RetOps, 2, DL);
2302 // Op is an ATOMIC_LOAD_SUB operation. Lower 8- and 16-bit operations
2303 // into ATOMIC_LOADW_SUBs and decide whether to convert 32- and 64-bit
2304 // operations into additions.
2305 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
2306 SelectionDAG &DAG) const {
2307 auto *Node = cast<AtomicSDNode>(Op.getNode());
2308 EVT MemVT = Node->getMemoryVT();
2309 if (MemVT == MVT::i32 || MemVT == MVT::i64) {
2310 // A full-width operation.
2311 assert(Op.getValueType() == MemVT && "Mismatched VTs");
2312 SDValue Src2 = Node->getVal();
2316 if (auto *Op2 = dyn_cast<ConstantSDNode>(Src2)) {
2317 // Use an addition if the operand is constant and either LAA(G) is
2318 // available or the negative value is in the range of A(G)FHI.
2319 int64_t Value = (-Op2->getAPIntValue()).getSExtValue();
2320 if (isInt<32>(Value) || TM.getSubtargetImpl()->hasInterlockedAccess1())
2321 NegSrc2 = DAG.getConstant(Value, MemVT);
2322 } else if (TM.getSubtargetImpl()->hasInterlockedAccess1())
2323 // Use LAA(G) if available.
2324 NegSrc2 = DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, MemVT),
2327 if (NegSrc2.getNode())
2328 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
2329 Node->getChain(), Node->getBasePtr(), NegSrc2,
2330 Node->getMemOperand(), Node->getOrdering(),
2331 Node->getSynchScope());
2333 // Use the node as-is.
2337 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
2340 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
2341 // into a fullword ATOMIC_CMP_SWAPW operation.
2342 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
2343 SelectionDAG &DAG) const {
2344 auto *Node = cast<AtomicSDNode>(Op.getNode());
2346 // We have native support for 32-bit compare and swap.
2347 EVT NarrowVT = Node->getMemoryVT();
2348 EVT WideVT = MVT::i32;
2349 if (NarrowVT == WideVT)
2352 int64_t BitSize = NarrowVT.getSizeInBits();
2353 SDValue ChainIn = Node->getOperand(0);
2354 SDValue Addr = Node->getOperand(1);
2355 SDValue CmpVal = Node->getOperand(2);
2356 SDValue SwapVal = Node->getOperand(3);
2357 MachineMemOperand *MMO = Node->getMemOperand();
2359 EVT PtrVT = Addr.getValueType();
2361 // Get the address of the containing word.
2362 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2363 DAG.getConstant(-4, PtrVT));
2365 // Get the number of bits that the word must be rotated left in order
2366 // to bring the field to the top bits of a GR32.
2367 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2368 DAG.getConstant(3, PtrVT));
2369 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2371 // Get the complementing shift amount, for rotating a field in the top
2372 // bits back to its proper position.
2373 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2374 DAG.getConstant(0, WideVT), BitShift);
2376 // Construct the ATOMIC_CMP_SWAPW node.
2377 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2378 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
2379 NegBitShift, DAG.getConstant(BitSize, WideVT) };
2380 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
2381 VTList, Ops, array_lengthof(Ops),
2386 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
2387 SelectionDAG &DAG) const {
2388 MachineFunction &MF = DAG.getMachineFunction();
2389 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2390 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
2391 SystemZ::R15D, Op.getValueType());
2394 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
2395 SelectionDAG &DAG) const {
2396 MachineFunction &MF = DAG.getMachineFunction();
2397 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2398 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
2399 SystemZ::R15D, Op.getOperand(1));
2402 SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op,
2403 SelectionDAG &DAG) const {
2404 bool IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2406 // Just preserve the chain.
2407 return Op.getOperand(0);
2409 bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2410 unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
2411 auto *Node = cast<MemIntrinsicSDNode>(Op.getNode());
2414 DAG.getConstant(Code, MVT::i32),
2417 return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, SDLoc(Op),
2418 Node->getVTList(), Ops, array_lengthof(Ops),
2419 Node->getMemoryVT(), Node->getMemOperand());
2422 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
2423 SelectionDAG &DAG) const {
2424 switch (Op.getOpcode()) {
2426 return lowerBR_CC(Op, DAG);
2427 case ISD::SELECT_CC:
2428 return lowerSELECT_CC(Op, DAG);
2430 return lowerSETCC(Op, DAG);
2431 case ISD::GlobalAddress:
2432 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
2433 case ISD::GlobalTLSAddress:
2434 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
2435 case ISD::BlockAddress:
2436 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
2437 case ISD::JumpTable:
2438 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
2439 case ISD::ConstantPool:
2440 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
2442 return lowerBITCAST(Op, DAG);
2444 return lowerVASTART(Op, DAG);
2446 return lowerVACOPY(Op, DAG);
2447 case ISD::DYNAMIC_STACKALLOC:
2448 return lowerDYNAMIC_STACKALLOC(Op, DAG);
2449 case ISD::SMUL_LOHI:
2450 return lowerSMUL_LOHI(Op, DAG);
2451 case ISD::UMUL_LOHI:
2452 return lowerUMUL_LOHI(Op, DAG);
2454 return lowerSDIVREM(Op, DAG);
2456 return lowerUDIVREM(Op, DAG);
2458 return lowerOR(Op, DAG);
2459 case ISD::SIGN_EXTEND:
2460 return lowerSIGN_EXTEND(Op, DAG);
2461 case ISD::ATOMIC_SWAP:
2462 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
2463 case ISD::ATOMIC_STORE:
2464 return lowerATOMIC_STORE(Op, DAG);
2465 case ISD::ATOMIC_LOAD:
2466 return lowerATOMIC_LOAD(Op, DAG);
2467 case ISD::ATOMIC_LOAD_ADD:
2468 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
2469 case ISD::ATOMIC_LOAD_SUB:
2470 return lowerATOMIC_LOAD_SUB(Op, DAG);
2471 case ISD::ATOMIC_LOAD_AND:
2472 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
2473 case ISD::ATOMIC_LOAD_OR:
2474 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
2475 case ISD::ATOMIC_LOAD_XOR:
2476 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
2477 case ISD::ATOMIC_LOAD_NAND:
2478 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
2479 case ISD::ATOMIC_LOAD_MIN:
2480 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
2481 case ISD::ATOMIC_LOAD_MAX:
2482 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
2483 case ISD::ATOMIC_LOAD_UMIN:
2484 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
2485 case ISD::ATOMIC_LOAD_UMAX:
2486 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
2487 case ISD::ATOMIC_CMP_SWAP:
2488 return lowerATOMIC_CMP_SWAP(Op, DAG);
2489 case ISD::STACKSAVE:
2490 return lowerSTACKSAVE(Op, DAG);
2491 case ISD::STACKRESTORE:
2492 return lowerSTACKRESTORE(Op, DAG);
2494 return lowerPREFETCH(Op, DAG);
2496 llvm_unreachable("Unexpected node to lower");
2500 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
2501 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
2506 OPCODE(PCREL_WRAPPER);
2507 OPCODE(PCREL_OFFSET);
2513 OPCODE(SELECT_CCMASK);
2514 OPCODE(ADJDYNALLOC);
2515 OPCODE(EXTRACT_ACCESS);
2516 OPCODE(UMUL_LOHI64);
2532 OPCODE(SEARCH_STRING);
2535 OPCODE(ATOMIC_SWAPW);
2536 OPCODE(ATOMIC_LOADW_ADD);
2537 OPCODE(ATOMIC_LOADW_SUB);
2538 OPCODE(ATOMIC_LOADW_AND);
2539 OPCODE(ATOMIC_LOADW_OR);
2540 OPCODE(ATOMIC_LOADW_XOR);
2541 OPCODE(ATOMIC_LOADW_NAND);
2542 OPCODE(ATOMIC_LOADW_MIN);
2543 OPCODE(ATOMIC_LOADW_MAX);
2544 OPCODE(ATOMIC_LOADW_UMIN);
2545 OPCODE(ATOMIC_LOADW_UMAX);
2546 OPCODE(ATOMIC_CMP_SWAPW);
2553 //===----------------------------------------------------------------------===//
2555 //===----------------------------------------------------------------------===//
2557 // Create a new basic block after MBB.
2558 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
2559 MachineFunction &MF = *MBB->getParent();
2560 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
2561 MF.insert(std::next(MachineFunction::iterator(MBB)), NewMBB);
2565 // Split MBB after MI and return the new block (the one that contains
2566 // instructions after MI).
2567 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
2568 MachineBasicBlock *MBB) {
2569 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2570 NewMBB->splice(NewMBB->begin(), MBB,
2571 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
2572 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2576 // Split MBB before MI and return the new block (the one that contains MI).
2577 static MachineBasicBlock *splitBlockBefore(MachineInstr *MI,
2578 MachineBasicBlock *MBB) {
2579 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2580 NewMBB->splice(NewMBB->begin(), MBB, MI, MBB->end());
2581 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2585 // Force base value Base into a register before MI. Return the register.
2586 static unsigned forceReg(MachineInstr *MI, MachineOperand &Base,
2587 const SystemZInstrInfo *TII) {
2589 return Base.getReg();
2591 MachineBasicBlock *MBB = MI->getParent();
2592 MachineFunction &MF = *MBB->getParent();
2593 MachineRegisterInfo &MRI = MF.getRegInfo();
2595 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
2596 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LA), Reg)
2597 .addOperand(Base).addImm(0).addReg(0);
2601 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
2603 SystemZTargetLowering::emitSelect(MachineInstr *MI,
2604 MachineBasicBlock *MBB) const {
2605 const SystemZInstrInfo *TII = TM.getInstrInfo();
2607 unsigned DestReg = MI->getOperand(0).getReg();
2608 unsigned TrueReg = MI->getOperand(1).getReg();
2609 unsigned FalseReg = MI->getOperand(2).getReg();
2610 unsigned CCValid = MI->getOperand(3).getImm();
2611 unsigned CCMask = MI->getOperand(4).getImm();
2612 DebugLoc DL = MI->getDebugLoc();
2614 MachineBasicBlock *StartMBB = MBB;
2615 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2616 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2619 // BRC CCMask, JoinMBB
2620 // # fallthrough to FalseMBB
2622 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2623 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2624 MBB->addSuccessor(JoinMBB);
2625 MBB->addSuccessor(FalseMBB);
2628 // # fallthrough to JoinMBB
2630 MBB->addSuccessor(JoinMBB);
2633 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
2636 BuildMI(*MBB, MI, DL, TII->get(SystemZ::PHI), DestReg)
2637 .addReg(TrueReg).addMBB(StartMBB)
2638 .addReg(FalseReg).addMBB(FalseMBB);
2640 MI->eraseFromParent();
2644 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
2645 // StoreOpcode is the store to use and Invert says whether the store should
2646 // happen when the condition is false rather than true. If a STORE ON
2647 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
2649 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
2650 MachineBasicBlock *MBB,
2651 unsigned StoreOpcode, unsigned STOCOpcode,
2652 bool Invert) const {
2653 const SystemZInstrInfo *TII = TM.getInstrInfo();
2655 unsigned SrcReg = MI->getOperand(0).getReg();
2656 MachineOperand Base = MI->getOperand(1);
2657 int64_t Disp = MI->getOperand(2).getImm();
2658 unsigned IndexReg = MI->getOperand(3).getReg();
2659 unsigned CCValid = MI->getOperand(4).getImm();
2660 unsigned CCMask = MI->getOperand(5).getImm();
2661 DebugLoc DL = MI->getDebugLoc();
2663 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
2665 // Use STOCOpcode if possible. We could use different store patterns in
2666 // order to avoid matching the index register, but the performance trade-offs
2667 // might be more complicated in that case.
2668 if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) {
2671 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
2672 .addReg(SrcReg).addOperand(Base).addImm(Disp)
2673 .addImm(CCValid).addImm(CCMask);
2674 MI->eraseFromParent();
2678 // Get the condition needed to branch around the store.
2682 MachineBasicBlock *StartMBB = MBB;
2683 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2684 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2687 // BRC CCMask, JoinMBB
2688 // # fallthrough to FalseMBB
2690 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2691 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2692 MBB->addSuccessor(JoinMBB);
2693 MBB->addSuccessor(FalseMBB);
2696 // store %SrcReg, %Disp(%Index,%Base)
2697 // # fallthrough to JoinMBB
2699 BuildMI(MBB, DL, TII->get(StoreOpcode))
2700 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
2701 MBB->addSuccessor(JoinMBB);
2703 MI->eraseFromParent();
2707 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
2708 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
2709 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
2710 // BitSize is the width of the field in bits, or 0 if this is a partword
2711 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
2712 // is one of the operands. Invert says whether the field should be
2713 // inverted after performing BinOpcode (e.g. for NAND).
2715 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
2716 MachineBasicBlock *MBB,
2719 bool Invert) const {
2720 const SystemZInstrInfo *TII = TM.getInstrInfo();
2721 MachineFunction &MF = *MBB->getParent();
2722 MachineRegisterInfo &MRI = MF.getRegInfo();
2723 bool IsSubWord = (BitSize < 32);
2725 // Extract the operands. Base can be a register or a frame index.
2726 // Src2 can be a register or immediate.
2727 unsigned Dest = MI->getOperand(0).getReg();
2728 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2729 int64_t Disp = MI->getOperand(2).getImm();
2730 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
2731 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2732 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2733 DebugLoc DL = MI->getDebugLoc();
2735 BitSize = MI->getOperand(6).getImm();
2737 // Subword operations use 32-bit registers.
2738 const TargetRegisterClass *RC = (BitSize <= 32 ?
2739 &SystemZ::GR32BitRegClass :
2740 &SystemZ::GR64BitRegClass);
2741 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2742 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2744 // Get the right opcodes for the displacement.
2745 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2746 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2747 assert(LOpcode && CSOpcode && "Displacement out of range");
2749 // Create virtual registers for temporary results.
2750 unsigned OrigVal = MRI.createVirtualRegister(RC);
2751 unsigned OldVal = MRI.createVirtualRegister(RC);
2752 unsigned NewVal = (BinOpcode || IsSubWord ?
2753 MRI.createVirtualRegister(RC) : Src2.getReg());
2754 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2755 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2757 // Insert a basic block for the main loop.
2758 MachineBasicBlock *StartMBB = MBB;
2759 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2760 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2764 // %OrigVal = L Disp(%Base)
2765 // # fall through to LoopMMB
2767 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2768 .addOperand(Base).addImm(Disp).addReg(0);
2769 MBB->addSuccessor(LoopMBB);
2772 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
2773 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2774 // %RotatedNewVal = OP %RotatedOldVal, %Src2
2775 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2776 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2778 // # fall through to DoneMMB
2780 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2781 .addReg(OrigVal).addMBB(StartMBB)
2782 .addReg(Dest).addMBB(LoopMBB);
2784 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2785 .addReg(OldVal).addReg(BitShift).addImm(0);
2787 // Perform the operation normally and then invert every bit of the field.
2788 unsigned Tmp = MRI.createVirtualRegister(RC);
2789 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
2790 .addReg(RotatedOldVal).addOperand(Src2);
2792 // XILF with the upper BitSize bits set.
2793 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2794 .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize)));
2795 else if (BitSize == 32)
2796 // XILF with every bit set.
2797 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2798 .addReg(Tmp).addImm(~uint32_t(0));
2800 // Use LCGR and add -1 to the result, which is more compact than
2801 // an XILF, XILH pair.
2802 unsigned Tmp2 = MRI.createVirtualRegister(RC);
2803 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
2804 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
2805 .addReg(Tmp2).addImm(-1);
2807 } else if (BinOpcode)
2808 // A simply binary operation.
2809 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
2810 .addReg(RotatedOldVal).addOperand(Src2);
2812 // Use RISBG to rotate Src2 into position and use it to replace the
2813 // field in RotatedOldVal.
2814 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
2815 .addReg(RotatedOldVal).addReg(Src2.getReg())
2816 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
2818 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2819 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2820 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2821 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2822 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2823 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2824 MBB->addSuccessor(LoopMBB);
2825 MBB->addSuccessor(DoneMBB);
2827 MI->eraseFromParent();
2831 // Implement EmitInstrWithCustomInserter for pseudo
2832 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
2833 // instruction that should be used to compare the current field with the
2834 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
2835 // for when the current field should be kept. BitSize is the width of
2836 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
2838 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
2839 MachineBasicBlock *MBB,
2840 unsigned CompareOpcode,
2841 unsigned KeepOldMask,
2842 unsigned BitSize) const {
2843 const SystemZInstrInfo *TII = TM.getInstrInfo();
2844 MachineFunction &MF = *MBB->getParent();
2845 MachineRegisterInfo &MRI = MF.getRegInfo();
2846 bool IsSubWord = (BitSize < 32);
2848 // Extract the operands. Base can be a register or a frame index.
2849 unsigned Dest = MI->getOperand(0).getReg();
2850 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2851 int64_t Disp = MI->getOperand(2).getImm();
2852 unsigned Src2 = MI->getOperand(3).getReg();
2853 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2854 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2855 DebugLoc DL = MI->getDebugLoc();
2857 BitSize = MI->getOperand(6).getImm();
2859 // Subword operations use 32-bit registers.
2860 const TargetRegisterClass *RC = (BitSize <= 32 ?
2861 &SystemZ::GR32BitRegClass :
2862 &SystemZ::GR64BitRegClass);
2863 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2864 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2866 // Get the right opcodes for the displacement.
2867 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2868 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2869 assert(LOpcode && CSOpcode && "Displacement out of range");
2871 // Create virtual registers for temporary results.
2872 unsigned OrigVal = MRI.createVirtualRegister(RC);
2873 unsigned OldVal = MRI.createVirtualRegister(RC);
2874 unsigned NewVal = MRI.createVirtualRegister(RC);
2875 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2876 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
2877 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2879 // Insert 3 basic blocks for the loop.
2880 MachineBasicBlock *StartMBB = MBB;
2881 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2882 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2883 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
2884 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
2888 // %OrigVal = L Disp(%Base)
2889 // # fall through to LoopMMB
2891 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2892 .addOperand(Base).addImm(Disp).addReg(0);
2893 MBB->addSuccessor(LoopMBB);
2896 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
2897 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2898 // CompareOpcode %RotatedOldVal, %Src2
2899 // BRC KeepOldMask, UpdateMBB
2901 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2902 .addReg(OrigVal).addMBB(StartMBB)
2903 .addReg(Dest).addMBB(UpdateMBB);
2905 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2906 .addReg(OldVal).addReg(BitShift).addImm(0);
2907 BuildMI(MBB, DL, TII->get(CompareOpcode))
2908 .addReg(RotatedOldVal).addReg(Src2);
2909 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2910 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
2911 MBB->addSuccessor(UpdateMBB);
2912 MBB->addSuccessor(UseAltMBB);
2915 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
2916 // # fall through to UpdateMMB
2919 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
2920 .addReg(RotatedOldVal).addReg(Src2)
2921 .addImm(32).addImm(31 + BitSize).addImm(0);
2922 MBB->addSuccessor(UpdateMBB);
2925 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
2926 // [ %RotatedAltVal, UseAltMBB ]
2927 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2928 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2930 // # fall through to DoneMMB
2932 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
2933 .addReg(RotatedOldVal).addMBB(LoopMBB)
2934 .addReg(RotatedAltVal).addMBB(UseAltMBB);
2936 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2937 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2938 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2939 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2940 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2941 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2942 MBB->addSuccessor(LoopMBB);
2943 MBB->addSuccessor(DoneMBB);
2945 MI->eraseFromParent();
2949 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
2952 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
2953 MachineBasicBlock *MBB) const {
2954 const SystemZInstrInfo *TII = TM.getInstrInfo();
2955 MachineFunction &MF = *MBB->getParent();
2956 MachineRegisterInfo &MRI = MF.getRegInfo();
2958 // Extract the operands. Base can be a register or a frame index.
2959 unsigned Dest = MI->getOperand(0).getReg();
2960 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2961 int64_t Disp = MI->getOperand(2).getImm();
2962 unsigned OrigCmpVal = MI->getOperand(3).getReg();
2963 unsigned OrigSwapVal = MI->getOperand(4).getReg();
2964 unsigned BitShift = MI->getOperand(5).getReg();
2965 unsigned NegBitShift = MI->getOperand(6).getReg();
2966 int64_t BitSize = MI->getOperand(7).getImm();
2967 DebugLoc DL = MI->getDebugLoc();
2969 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
2971 // Get the right opcodes for the displacement.
2972 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
2973 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
2974 assert(LOpcode && CSOpcode && "Displacement out of range");
2976 // Create virtual registers for temporary results.
2977 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
2978 unsigned OldVal = MRI.createVirtualRegister(RC);
2979 unsigned CmpVal = MRI.createVirtualRegister(RC);
2980 unsigned SwapVal = MRI.createVirtualRegister(RC);
2981 unsigned StoreVal = MRI.createVirtualRegister(RC);
2982 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
2983 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
2984 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
2986 // Insert 2 basic blocks for the loop.
2987 MachineBasicBlock *StartMBB = MBB;
2988 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2989 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2990 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
2994 // %OrigOldVal = L Disp(%Base)
2995 // # fall through to LoopMMB
2997 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
2998 .addOperand(Base).addImm(Disp).addReg(0);
2999 MBB->addSuccessor(LoopMBB);
3002 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
3003 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
3004 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
3005 // %Dest = RLL %OldVal, BitSize(%BitShift)
3006 // ^^ The low BitSize bits contain the field
3008 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
3009 // ^^ Replace the upper 32-BitSize bits of the
3010 // comparison value with those that we loaded,
3011 // so that we can use a full word comparison.
3012 // CR %Dest, %RetryCmpVal
3014 // # Fall through to SetMBB
3016 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
3017 .addReg(OrigOldVal).addMBB(StartMBB)
3018 .addReg(RetryOldVal).addMBB(SetMBB);
3019 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
3020 .addReg(OrigCmpVal).addMBB(StartMBB)
3021 .addReg(RetryCmpVal).addMBB(SetMBB);
3022 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
3023 .addReg(OrigSwapVal).addMBB(StartMBB)
3024 .addReg(RetrySwapVal).addMBB(SetMBB);
3025 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
3026 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
3027 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
3028 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3029 BuildMI(MBB, DL, TII->get(SystemZ::CR))
3030 .addReg(Dest).addReg(RetryCmpVal);
3031 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3032 .addImm(SystemZ::CCMASK_ICMP)
3033 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
3034 MBB->addSuccessor(DoneMBB);
3035 MBB->addSuccessor(SetMBB);
3038 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
3039 // ^^ Replace the upper 32-BitSize bits of the new
3040 // value with those that we loaded.
3041 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
3042 // ^^ Rotate the new field to its proper position.
3043 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
3045 // # fall through to ExitMMB
3047 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
3048 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3049 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
3050 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
3051 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
3052 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
3053 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3054 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
3055 MBB->addSuccessor(LoopMBB);
3056 MBB->addSuccessor(DoneMBB);
3058 MI->eraseFromParent();
3062 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
3063 // if the high register of the GR128 value must be cleared or false if
3064 // it's "don't care". SubReg is subreg_l32 when extending a GR32
3065 // and subreg_l64 when extending a GR64.
3067 SystemZTargetLowering::emitExt128(MachineInstr *MI,
3068 MachineBasicBlock *MBB,
3069 bool ClearEven, unsigned SubReg) const {
3070 const SystemZInstrInfo *TII = TM.getInstrInfo();
3071 MachineFunction &MF = *MBB->getParent();
3072 MachineRegisterInfo &MRI = MF.getRegInfo();
3073 DebugLoc DL = MI->getDebugLoc();
3075 unsigned Dest = MI->getOperand(0).getReg();
3076 unsigned Src = MI->getOperand(1).getReg();
3077 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3079 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
3081 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3082 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
3084 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
3086 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
3087 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_h64);
3090 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
3091 .addReg(In128).addReg(Src).addImm(SubReg);
3093 MI->eraseFromParent();
3098 SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
3099 MachineBasicBlock *MBB,
3100 unsigned Opcode) const {
3101 const SystemZInstrInfo *TII = TM.getInstrInfo();
3102 MachineFunction &MF = *MBB->getParent();
3103 MachineRegisterInfo &MRI = MF.getRegInfo();
3104 DebugLoc DL = MI->getDebugLoc();
3106 MachineOperand DestBase = earlyUseOperand(MI->getOperand(0));
3107 uint64_t DestDisp = MI->getOperand(1).getImm();
3108 MachineOperand SrcBase = earlyUseOperand(MI->getOperand(2));
3109 uint64_t SrcDisp = MI->getOperand(3).getImm();
3110 uint64_t Length = MI->getOperand(4).getImm();
3112 // When generating more than one CLC, all but the last will need to
3113 // branch to the end when a difference is found.
3114 MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ?
3115 splitBlockAfter(MI, MBB) : 0);
3117 // Check for the loop form, in which operand 5 is the trip count.
3118 if (MI->getNumExplicitOperands() > 5) {
3119 bool HaveSingleBase = DestBase.isIdenticalTo(SrcBase);
3121 uint64_t StartCountReg = MI->getOperand(5).getReg();
3122 uint64_t StartSrcReg = forceReg(MI, SrcBase, TII);
3123 uint64_t StartDestReg = (HaveSingleBase ? StartSrcReg :
3124 forceReg(MI, DestBase, TII));
3126 const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass;
3127 uint64_t ThisSrcReg = MRI.createVirtualRegister(RC);
3128 uint64_t ThisDestReg = (HaveSingleBase ? ThisSrcReg :
3129 MRI.createVirtualRegister(RC));
3130 uint64_t NextSrcReg = MRI.createVirtualRegister(RC);
3131 uint64_t NextDestReg = (HaveSingleBase ? NextSrcReg :
3132 MRI.createVirtualRegister(RC));
3134 RC = &SystemZ::GR64BitRegClass;
3135 uint64_t ThisCountReg = MRI.createVirtualRegister(RC);
3136 uint64_t NextCountReg = MRI.createVirtualRegister(RC);
3138 MachineBasicBlock *StartMBB = MBB;
3139 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3140 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3141 MachineBasicBlock *NextMBB = (EndMBB ? emitBlockAfter(LoopMBB) : LoopMBB);
3144 // # fall through to LoopMMB
3145 MBB->addSuccessor(LoopMBB);
3148 // %ThisDestReg = phi [ %StartDestReg, StartMBB ],
3149 // [ %NextDestReg, NextMBB ]
3150 // %ThisSrcReg = phi [ %StartSrcReg, StartMBB ],
3151 // [ %NextSrcReg, NextMBB ]
3152 // %ThisCountReg = phi [ %StartCountReg, StartMBB ],
3153 // [ %NextCountReg, NextMBB ]
3154 // ( PFD 2, 768+DestDisp(%ThisDestReg) )
3155 // Opcode DestDisp(256,%ThisDestReg), SrcDisp(%ThisSrcReg)
3158 // The prefetch is used only for MVC. The JLH is used only for CLC.
3161 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisDestReg)
3162 .addReg(StartDestReg).addMBB(StartMBB)
3163 .addReg(NextDestReg).addMBB(NextMBB);
3164 if (!HaveSingleBase)
3165 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisSrcReg)
3166 .addReg(StartSrcReg).addMBB(StartMBB)
3167 .addReg(NextSrcReg).addMBB(NextMBB);
3168 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisCountReg)
3169 .addReg(StartCountReg).addMBB(StartMBB)
3170 .addReg(NextCountReg).addMBB(NextMBB);
3171 if (Opcode == SystemZ::MVC)
3172 BuildMI(MBB, DL, TII->get(SystemZ::PFD))
3173 .addImm(SystemZ::PFD_WRITE)
3174 .addReg(ThisDestReg).addImm(DestDisp + 768).addReg(0);
3175 BuildMI(MBB, DL, TII->get(Opcode))
3176 .addReg(ThisDestReg).addImm(DestDisp).addImm(256)
3177 .addReg(ThisSrcReg).addImm(SrcDisp);
3179 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3180 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3182 MBB->addSuccessor(EndMBB);
3183 MBB->addSuccessor(NextMBB);
3187 // %NextDestReg = LA 256(%ThisDestReg)
3188 // %NextSrcReg = LA 256(%ThisSrcReg)
3189 // %NextCountReg = AGHI %ThisCountReg, -1
3190 // CGHI %NextCountReg, 0
3192 // # fall through to DoneMMB
3194 // The AGHI, CGHI and JLH should be converted to BRCTG by later passes.
3197 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextDestReg)
3198 .addReg(ThisDestReg).addImm(256).addReg(0);
3199 if (!HaveSingleBase)
3200 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextSrcReg)
3201 .addReg(ThisSrcReg).addImm(256).addReg(0);
3202 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), NextCountReg)
3203 .addReg(ThisCountReg).addImm(-1);
3204 BuildMI(MBB, DL, TII->get(SystemZ::CGHI))
3205 .addReg(NextCountReg).addImm(0);
3206 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3207 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3209 MBB->addSuccessor(LoopMBB);
3210 MBB->addSuccessor(DoneMBB);
3212 DestBase = MachineOperand::CreateReg(NextDestReg, false);
3213 SrcBase = MachineOperand::CreateReg(NextSrcReg, false);
3217 // Handle any remaining bytes with straight-line code.
3218 while (Length > 0) {
3219 uint64_t ThisLength = std::min(Length, uint64_t(256));
3220 // The previous iteration might have created out-of-range displacements.
3221 // Apply them using LAY if so.
3222 if (!isUInt<12>(DestDisp)) {
3223 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3224 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3225 .addOperand(DestBase).addImm(DestDisp).addReg(0);
3226 DestBase = MachineOperand::CreateReg(Reg, false);
3229 if (!isUInt<12>(SrcDisp)) {
3230 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3231 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3232 .addOperand(SrcBase).addImm(SrcDisp).addReg(0);
3233 SrcBase = MachineOperand::CreateReg(Reg, false);
3236 BuildMI(*MBB, MI, DL, TII->get(Opcode))
3237 .addOperand(DestBase).addImm(DestDisp).addImm(ThisLength)
3238 .addOperand(SrcBase).addImm(SrcDisp);
3239 DestDisp += ThisLength;
3240 SrcDisp += ThisLength;
3241 Length -= ThisLength;
3242 // If there's another CLC to go, branch to the end if a difference
3244 if (EndMBB && Length > 0) {
3245 MachineBasicBlock *NextMBB = splitBlockBefore(MI, MBB);
3246 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3247 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3249 MBB->addSuccessor(EndMBB);
3250 MBB->addSuccessor(NextMBB);
3255 MBB->addSuccessor(EndMBB);
3257 MBB->addLiveIn(SystemZ::CC);
3260 MI->eraseFromParent();
3264 // Decompose string pseudo-instruction MI into a loop that continually performs
3265 // Opcode until CC != 3.
3267 SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
3268 MachineBasicBlock *MBB,
3269 unsigned Opcode) const {
3270 const SystemZInstrInfo *TII = TM.getInstrInfo();
3271 MachineFunction &MF = *MBB->getParent();
3272 MachineRegisterInfo &MRI = MF.getRegInfo();
3273 DebugLoc DL = MI->getDebugLoc();
3275 uint64_t End1Reg = MI->getOperand(0).getReg();
3276 uint64_t Start1Reg = MI->getOperand(1).getReg();
3277 uint64_t Start2Reg = MI->getOperand(2).getReg();
3278 uint64_t CharReg = MI->getOperand(3).getReg();
3280 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass;
3281 uint64_t This1Reg = MRI.createVirtualRegister(RC);
3282 uint64_t This2Reg = MRI.createVirtualRegister(RC);
3283 uint64_t End2Reg = MRI.createVirtualRegister(RC);
3285 MachineBasicBlock *StartMBB = MBB;
3286 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3287 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3290 // # fall through to LoopMMB
3291 MBB->addSuccessor(LoopMBB);
3294 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ]
3295 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ]
3297 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0L
3299 // # fall through to DoneMMB
3301 // The load of R0L can be hoisted by post-RA LICM.
3304 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg)
3305 .addReg(Start1Reg).addMBB(StartMBB)
3306 .addReg(End1Reg).addMBB(LoopMBB);
3307 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg)
3308 .addReg(Start2Reg).addMBB(StartMBB)
3309 .addReg(End2Reg).addMBB(LoopMBB);
3310 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0L).addReg(CharReg);
3311 BuildMI(MBB, DL, TII->get(Opcode))
3312 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define)
3313 .addReg(This1Reg).addReg(This2Reg);
3314 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3315 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB);
3316 MBB->addSuccessor(LoopMBB);
3317 MBB->addSuccessor(DoneMBB);
3319 DoneMBB->addLiveIn(SystemZ::CC);
3321 MI->eraseFromParent();
3325 MachineBasicBlock *SystemZTargetLowering::
3326 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
3327 switch (MI->getOpcode()) {
3328 case SystemZ::Select32Mux:
3329 case SystemZ::Select32:
3330 case SystemZ::SelectF32:
3331 case SystemZ::Select64:
3332 case SystemZ::SelectF64:
3333 case SystemZ::SelectF128:
3334 return emitSelect(MI, MBB);
3336 case SystemZ::CondStore8Mux:
3337 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, false);
3338 case SystemZ::CondStore8MuxInv:
3339 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, true);
3340 case SystemZ::CondStore16Mux:
3341 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, false);
3342 case SystemZ::CondStore16MuxInv:
3343 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, true);
3344 case SystemZ::CondStore8:
3345 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
3346 case SystemZ::CondStore8Inv:
3347 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
3348 case SystemZ::CondStore16:
3349 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
3350 case SystemZ::CondStore16Inv:
3351 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
3352 case SystemZ::CondStore32:
3353 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
3354 case SystemZ::CondStore32Inv:
3355 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
3356 case SystemZ::CondStore64:
3357 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
3358 case SystemZ::CondStore64Inv:
3359 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
3360 case SystemZ::CondStoreF32:
3361 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
3362 case SystemZ::CondStoreF32Inv:
3363 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
3364 case SystemZ::CondStoreF64:
3365 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
3366 case SystemZ::CondStoreF64Inv:
3367 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
3369 case SystemZ::AEXT128_64:
3370 return emitExt128(MI, MBB, false, SystemZ::subreg_l64);
3371 case SystemZ::ZEXT128_32:
3372 return emitExt128(MI, MBB, true, SystemZ::subreg_l32);
3373 case SystemZ::ZEXT128_64:
3374 return emitExt128(MI, MBB, true, SystemZ::subreg_l64);
3376 case SystemZ::ATOMIC_SWAPW:
3377 return emitAtomicLoadBinary(MI, MBB, 0, 0);
3378 case SystemZ::ATOMIC_SWAP_32:
3379 return emitAtomicLoadBinary(MI, MBB, 0, 32);
3380 case SystemZ::ATOMIC_SWAP_64:
3381 return emitAtomicLoadBinary(MI, MBB, 0, 64);
3383 case SystemZ::ATOMIC_LOADW_AR:
3384 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
3385 case SystemZ::ATOMIC_LOADW_AFI:
3386 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
3387 case SystemZ::ATOMIC_LOAD_AR:
3388 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
3389 case SystemZ::ATOMIC_LOAD_AHI:
3390 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
3391 case SystemZ::ATOMIC_LOAD_AFI:
3392 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
3393 case SystemZ::ATOMIC_LOAD_AGR:
3394 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
3395 case SystemZ::ATOMIC_LOAD_AGHI:
3396 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
3397 case SystemZ::ATOMIC_LOAD_AGFI:
3398 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
3400 case SystemZ::ATOMIC_LOADW_SR:
3401 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
3402 case SystemZ::ATOMIC_LOAD_SR:
3403 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
3404 case SystemZ::ATOMIC_LOAD_SGR:
3405 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
3407 case SystemZ::ATOMIC_LOADW_NR:
3408 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
3409 case SystemZ::ATOMIC_LOADW_NILH:
3410 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0);
3411 case SystemZ::ATOMIC_LOAD_NR:
3412 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
3413 case SystemZ::ATOMIC_LOAD_NILL:
3414 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32);
3415 case SystemZ::ATOMIC_LOAD_NILH:
3416 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32);
3417 case SystemZ::ATOMIC_LOAD_NILF:
3418 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32);
3419 case SystemZ::ATOMIC_LOAD_NGR:
3420 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
3421 case SystemZ::ATOMIC_LOAD_NILL64:
3422 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64);
3423 case SystemZ::ATOMIC_LOAD_NILH64:
3424 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64);
3425 case SystemZ::ATOMIC_LOAD_NIHL64:
3426 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64);
3427 case SystemZ::ATOMIC_LOAD_NIHH64:
3428 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64);
3429 case SystemZ::ATOMIC_LOAD_NILF64:
3430 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64);
3431 case SystemZ::ATOMIC_LOAD_NIHF64:
3432 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64);
3434 case SystemZ::ATOMIC_LOADW_OR:
3435 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
3436 case SystemZ::ATOMIC_LOADW_OILH:
3437 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 0);
3438 case SystemZ::ATOMIC_LOAD_OR:
3439 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
3440 case SystemZ::ATOMIC_LOAD_OILL:
3441 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 32);
3442 case SystemZ::ATOMIC_LOAD_OILH:
3443 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 32);
3444 case SystemZ::ATOMIC_LOAD_OILF:
3445 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 32);
3446 case SystemZ::ATOMIC_LOAD_OGR:
3447 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
3448 case SystemZ::ATOMIC_LOAD_OILL64:
3449 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL64, 64);
3450 case SystemZ::ATOMIC_LOAD_OILH64:
3451 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH64, 64);
3452 case SystemZ::ATOMIC_LOAD_OIHL64:
3453 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL64, 64);
3454 case SystemZ::ATOMIC_LOAD_OIHH64:
3455 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH64, 64);
3456 case SystemZ::ATOMIC_LOAD_OILF64:
3457 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF64, 64);
3458 case SystemZ::ATOMIC_LOAD_OIHF64:
3459 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF64, 64);
3461 case SystemZ::ATOMIC_LOADW_XR:
3462 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
3463 case SystemZ::ATOMIC_LOADW_XILF:
3464 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 0);
3465 case SystemZ::ATOMIC_LOAD_XR:
3466 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
3467 case SystemZ::ATOMIC_LOAD_XILF:
3468 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 32);
3469 case SystemZ::ATOMIC_LOAD_XGR:
3470 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
3471 case SystemZ::ATOMIC_LOAD_XILF64:
3472 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF64, 64);
3473 case SystemZ::ATOMIC_LOAD_XIHF64:
3474 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF64, 64);
3476 case SystemZ::ATOMIC_LOADW_NRi:
3477 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
3478 case SystemZ::ATOMIC_LOADW_NILHi:
3479 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0, true);
3480 case SystemZ::ATOMIC_LOAD_NRi:
3481 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
3482 case SystemZ::ATOMIC_LOAD_NILLi:
3483 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32, true);
3484 case SystemZ::ATOMIC_LOAD_NILHi:
3485 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32, true);
3486 case SystemZ::ATOMIC_LOAD_NILFi:
3487 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32, true);
3488 case SystemZ::ATOMIC_LOAD_NGRi:
3489 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
3490 case SystemZ::ATOMIC_LOAD_NILL64i:
3491 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64, true);
3492 case SystemZ::ATOMIC_LOAD_NILH64i:
3493 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64, true);
3494 case SystemZ::ATOMIC_LOAD_NIHL64i:
3495 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64, true);
3496 case SystemZ::ATOMIC_LOAD_NIHH64i:
3497 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64, true);
3498 case SystemZ::ATOMIC_LOAD_NILF64i:
3499 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64, true);
3500 case SystemZ::ATOMIC_LOAD_NIHF64i:
3501 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64, true);
3503 case SystemZ::ATOMIC_LOADW_MIN:
3504 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3505 SystemZ::CCMASK_CMP_LE, 0);
3506 case SystemZ::ATOMIC_LOAD_MIN_32:
3507 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3508 SystemZ::CCMASK_CMP_LE, 32);
3509 case SystemZ::ATOMIC_LOAD_MIN_64:
3510 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3511 SystemZ::CCMASK_CMP_LE, 64);
3513 case SystemZ::ATOMIC_LOADW_MAX:
3514 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3515 SystemZ::CCMASK_CMP_GE, 0);
3516 case SystemZ::ATOMIC_LOAD_MAX_32:
3517 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3518 SystemZ::CCMASK_CMP_GE, 32);
3519 case SystemZ::ATOMIC_LOAD_MAX_64:
3520 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3521 SystemZ::CCMASK_CMP_GE, 64);
3523 case SystemZ::ATOMIC_LOADW_UMIN:
3524 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3525 SystemZ::CCMASK_CMP_LE, 0);
3526 case SystemZ::ATOMIC_LOAD_UMIN_32:
3527 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3528 SystemZ::CCMASK_CMP_LE, 32);
3529 case SystemZ::ATOMIC_LOAD_UMIN_64:
3530 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3531 SystemZ::CCMASK_CMP_LE, 64);
3533 case SystemZ::ATOMIC_LOADW_UMAX:
3534 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3535 SystemZ::CCMASK_CMP_GE, 0);
3536 case SystemZ::ATOMIC_LOAD_UMAX_32:
3537 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3538 SystemZ::CCMASK_CMP_GE, 32);
3539 case SystemZ::ATOMIC_LOAD_UMAX_64:
3540 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3541 SystemZ::CCMASK_CMP_GE, 64);
3543 case SystemZ::ATOMIC_CMP_SWAPW:
3544 return emitAtomicCmpSwapW(MI, MBB);
3545 case SystemZ::MVCSequence:
3546 case SystemZ::MVCLoop:
3547 return emitMemMemWrapper(MI, MBB, SystemZ::MVC);
3548 case SystemZ::NCSequence:
3549 case SystemZ::NCLoop:
3550 return emitMemMemWrapper(MI, MBB, SystemZ::NC);
3551 case SystemZ::OCSequence:
3552 case SystemZ::OCLoop:
3553 return emitMemMemWrapper(MI, MBB, SystemZ::OC);
3554 case SystemZ::XCSequence:
3555 case SystemZ::XCLoop:
3556 return emitMemMemWrapper(MI, MBB, SystemZ::XC);
3557 case SystemZ::CLCSequence:
3558 case SystemZ::CLCLoop:
3559 return emitMemMemWrapper(MI, MBB, SystemZ::CLC);
3560 case SystemZ::CLSTLoop:
3561 return emitStringWrapper(MI, MBB, SystemZ::CLST);
3562 case SystemZ::MVSTLoop:
3563 return emitStringWrapper(MI, MBB, SystemZ::MVST);
3564 case SystemZ::SRSTLoop:
3565 return emitStringWrapper(MI, MBB, SystemZ::SRST);
3567 llvm_unreachable("Unexpected instr type to insert");