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"
28 // Classify VT as either 32 or 64 bit.
29 static bool is32Bit(EVT VT) {
30 switch (VT.getSimpleVT().SimpleTy) {
36 llvm_unreachable("Unsupported type");
40 // Return a version of MachineOperand that can be safely used before the
42 static MachineOperand earlyUseOperand(MachineOperand Op) {
48 SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm)
49 : TargetLowering(tm, new TargetLoweringObjectFileELF()),
50 Subtarget(*tm.getSubtargetImpl()), TM(tm) {
51 MVT PtrVT = getPointerTy();
53 // Set up the register classes.
54 addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
55 addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
56 addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
57 addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
58 addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
60 // Compute derived properties from the register classes
61 computeRegisterProperties();
63 // Set up special registers.
64 setExceptionPointerRegister(SystemZ::R6D);
65 setExceptionSelectorRegister(SystemZ::R7D);
66 setStackPointerRegisterToSaveRestore(SystemZ::R15D);
68 // TODO: It may be better to default to latency-oriented scheduling, however
69 // LLVM's current latency-oriented scheduler can't handle physreg definitions
70 // such as SystemZ has with CC, so set this to the register-pressure
71 // scheduler, because it can.
72 setSchedulingPreference(Sched::RegPressure);
74 setBooleanContents(ZeroOrOneBooleanContent);
75 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
77 // Instructions are strings of 2-byte aligned 2-byte values.
78 setMinFunctionAlignment(2);
80 // Handle operations that are handled in a similar way for all types.
81 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
82 I <= MVT::LAST_FP_VALUETYPE;
84 MVT VT = MVT::SimpleValueType(I);
85 if (isTypeLegal(VT)) {
86 // Expand SETCC(X, Y, COND) into SELECT_CC(X, Y, 1, 0, COND).
87 setOperationAction(ISD::SETCC, VT, Expand);
89 // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
90 setOperationAction(ISD::SELECT, VT, Expand);
92 // Lower SELECT_CC and BR_CC into separate comparisons and branches.
93 setOperationAction(ISD::SELECT_CC, VT, Custom);
94 setOperationAction(ISD::BR_CC, VT, Custom);
98 // Expand jump table branches as address arithmetic followed by an
100 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
102 // Expand BRCOND into a BR_CC (see above).
103 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
105 // Handle integer types.
106 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
107 I <= MVT::LAST_INTEGER_VALUETYPE;
109 MVT VT = MVT::SimpleValueType(I);
110 if (isTypeLegal(VT)) {
111 // Expand individual DIV and REMs into DIVREMs.
112 setOperationAction(ISD::SDIV, VT, Expand);
113 setOperationAction(ISD::UDIV, VT, Expand);
114 setOperationAction(ISD::SREM, VT, Expand);
115 setOperationAction(ISD::UREM, VT, Expand);
116 setOperationAction(ISD::SDIVREM, VT, Custom);
117 setOperationAction(ISD::UDIVREM, VT, Custom);
119 // Expand ATOMIC_LOAD and ATOMIC_STORE using ATOMIC_CMP_SWAP.
120 // FIXME: probably much too conservative.
121 setOperationAction(ISD::ATOMIC_LOAD, VT, Expand);
122 setOperationAction(ISD::ATOMIC_STORE, VT, Expand);
124 // No special instructions for these.
125 setOperationAction(ISD::CTPOP, VT, Expand);
126 setOperationAction(ISD::CTTZ, VT, Expand);
127 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
128 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
129 setOperationAction(ISD::ROTR, VT, Expand);
131 // Use *MUL_LOHI where possible instead of MULH*.
132 setOperationAction(ISD::MULHS, VT, Expand);
133 setOperationAction(ISD::MULHU, VT, Expand);
134 setOperationAction(ISD::SMUL_LOHI, VT, Custom);
135 setOperationAction(ISD::UMUL_LOHI, VT, Custom);
137 // We have instructions for signed but not unsigned FP conversion.
138 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
142 // Type legalization will convert 8- and 16-bit atomic operations into
143 // forms that operate on i32s (but still keeping the original memory VT).
144 // Lower them into full i32 operations.
145 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
146 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
147 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
148 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
149 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
150 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
151 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
152 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
153 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
154 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
155 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
156 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
158 // We have instructions for signed but not unsigned FP conversion.
159 // Handle unsigned 32-bit types as signed 64-bit types.
160 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
161 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
163 // We have native support for a 64-bit CTLZ, via FLOGR.
164 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
165 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
167 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
168 setOperationAction(ISD::OR, MVT::i64, Custom);
170 // FIXME: Can we support these natively?
171 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
172 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
173 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
175 // We have native instructions for i8, i16 and i32 extensions, but not i1.
176 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
177 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
178 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
179 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
181 // Handle the various types of symbolic address.
182 setOperationAction(ISD::ConstantPool, PtrVT, Custom);
183 setOperationAction(ISD::GlobalAddress, PtrVT, Custom);
184 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
185 setOperationAction(ISD::BlockAddress, PtrVT, Custom);
186 setOperationAction(ISD::JumpTable, PtrVT, Custom);
188 // We need to handle dynamic allocations specially because of the
189 // 160-byte area at the bottom of the stack.
190 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
192 // Use custom expanders so that we can force the function to use
194 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
195 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
197 // Handle floating-point types.
198 for (unsigned I = MVT::FIRST_FP_VALUETYPE;
199 I <= MVT::LAST_FP_VALUETYPE;
201 MVT VT = MVT::SimpleValueType(I);
202 if (isTypeLegal(VT)) {
203 // We can use FI for FRINT.
204 setOperationAction(ISD::FRINT, VT, Legal);
206 // We can use the extended form of FI for other rounding operations.
207 if (Subtarget.hasFPExtension()) {
208 setOperationAction(ISD::FNEARBYINT, VT, Legal);
209 setOperationAction(ISD::FFLOOR, VT, Legal);
210 setOperationAction(ISD::FCEIL, VT, Legal);
211 setOperationAction(ISD::FTRUNC, VT, Legal);
212 setOperationAction(ISD::FROUND, VT, Legal);
215 // No special instructions for these.
216 setOperationAction(ISD::FSIN, VT, Expand);
217 setOperationAction(ISD::FCOS, VT, Expand);
218 setOperationAction(ISD::FREM, VT, Expand);
222 // We have fused multiply-addition for f32 and f64 but not f128.
223 setOperationAction(ISD::FMA, MVT::f32, Legal);
224 setOperationAction(ISD::FMA, MVT::f64, Legal);
225 setOperationAction(ISD::FMA, MVT::f128, Expand);
227 // Needed so that we don't try to implement f128 constant loads using
228 // a load-and-extend of a f80 constant (in cases where the constant
229 // would fit in an f80).
230 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
232 // Floating-point truncation and stores need to be done separately.
233 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
234 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
235 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
237 // We have 64-bit FPR<->GPR moves, but need special handling for
239 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
240 setOperationAction(ISD::BITCAST, MVT::f32, Custom);
242 // VASTART and VACOPY need to deal with the SystemZ-specific varargs
243 // structure, but VAEND is a no-op.
244 setOperationAction(ISD::VASTART, MVT::Other, Custom);
245 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
246 setOperationAction(ISD::VAEND, MVT::Other, Expand);
248 // We want to use MVC in preference to even a single load/store pair.
249 MaxStoresPerMemcpy = 0;
250 MaxStoresPerMemcpyOptSize = 0;
252 // The main memset sequence is a byte store followed by an MVC.
253 // Two STC or MV..I stores win over that, but the kind of fused stores
254 // generated by target-independent code don't when the byte value is
255 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
256 // than "STC;MVC". Handle the choice in target-specific code instead.
257 MaxStoresPerMemset = 0;
258 MaxStoresPerMemsetOptSize = 0;
262 SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
263 VT = VT.getScalarType();
268 switch (VT.getSimpleVT().SimpleTy) {
281 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
282 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
283 return Imm.isZero() || Imm.isNegZero();
286 bool SystemZTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
288 // Unaligned accesses should never be slower than the expanded version.
289 // We check specifically for aligned accesses in the few cases where
290 // they are required.
296 bool SystemZTargetLowering::isLegalAddressingMode(const AddrMode &AM,
298 // Punt on globals for now, although they can be used in limited
299 // RELATIVE LONG cases.
303 // Require a 20-bit signed offset.
304 if (!isInt<20>(AM.BaseOffs))
307 // Indexing is OK but no scale factor can be applied.
308 return AM.Scale == 0 || AM.Scale == 1;
311 bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
312 if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
314 unsigned FromBits = FromType->getPrimitiveSizeInBits();
315 unsigned ToBits = ToType->getPrimitiveSizeInBits();
316 return FromBits > ToBits;
319 bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
320 if (!FromVT.isInteger() || !ToVT.isInteger())
322 unsigned FromBits = FromVT.getSizeInBits();
323 unsigned ToBits = ToVT.getSizeInBits();
324 return FromBits > ToBits;
327 //===----------------------------------------------------------------------===//
328 // Inline asm support
329 //===----------------------------------------------------------------------===//
331 TargetLowering::ConstraintType
332 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
333 if (Constraint.size() == 1) {
334 switch (Constraint[0]) {
335 case 'a': // Address register
336 case 'd': // Data register (equivalent to 'r')
337 case 'f': // Floating-point register
338 case 'r': // General-purpose register
339 return C_RegisterClass;
341 case 'Q': // Memory with base and unsigned 12-bit displacement
342 case 'R': // Likewise, plus an index
343 case 'S': // Memory with base and signed 20-bit displacement
344 case 'T': // Likewise, plus an index
345 case 'm': // Equivalent to 'T'.
348 case 'I': // Unsigned 8-bit constant
349 case 'J': // Unsigned 12-bit constant
350 case 'K': // Signed 16-bit constant
351 case 'L': // Signed 20-bit displacement (on all targets we support)
352 case 'M': // 0x7fffffff
359 return TargetLowering::getConstraintType(Constraint);
362 TargetLowering::ConstraintWeight SystemZTargetLowering::
363 getSingleConstraintMatchWeight(AsmOperandInfo &info,
364 const char *constraint) const {
365 ConstraintWeight weight = CW_Invalid;
366 Value *CallOperandVal = info.CallOperandVal;
367 // If we don't have a value, we can't do a match,
368 // but allow it at the lowest weight.
369 if (CallOperandVal == NULL)
371 Type *type = CallOperandVal->getType();
372 // Look at the constraint type.
373 switch (*constraint) {
375 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
378 case 'a': // Address register
379 case 'd': // Data register (equivalent to 'r')
380 case 'r': // General-purpose register
381 if (CallOperandVal->getType()->isIntegerTy())
382 weight = CW_Register;
385 case 'f': // Floating-point register
386 if (type->isFloatingPointTy())
387 weight = CW_Register;
390 case 'I': // Unsigned 8-bit constant
391 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
392 if (isUInt<8>(C->getZExtValue()))
393 weight = CW_Constant;
396 case 'J': // Unsigned 12-bit constant
397 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
398 if (isUInt<12>(C->getZExtValue()))
399 weight = CW_Constant;
402 case 'K': // Signed 16-bit constant
403 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
404 if (isInt<16>(C->getSExtValue()))
405 weight = CW_Constant;
408 case 'L': // Signed 20-bit displacement (on all targets we support)
409 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
410 if (isInt<20>(C->getSExtValue()))
411 weight = CW_Constant;
414 case 'M': // 0x7fffffff
415 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
416 if (C->getZExtValue() == 0x7fffffff)
417 weight = CW_Constant;
423 // Parse a "{tNNN}" register constraint for which the register type "t"
424 // has already been verified. MC is the class associated with "t" and
425 // Map maps 0-based register numbers to LLVM register numbers.
426 static std::pair<unsigned, const TargetRegisterClass *>
427 parseRegisterNumber(const std::string &Constraint,
428 const TargetRegisterClass *RC, const unsigned *Map) {
429 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
430 if (isdigit(Constraint[2])) {
431 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2);
432 unsigned Index = atoi(Suffix.c_str());
433 if (Index < 16 && Map[Index])
434 return std::make_pair(Map[Index], RC);
436 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
439 std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering::
440 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const {
441 if (Constraint.size() == 1) {
442 // GCC Constraint Letters
443 switch (Constraint[0]) {
445 case 'd': // Data register (equivalent to 'r')
446 case 'r': // General-purpose register
448 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
449 else if (VT == MVT::i128)
450 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
451 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
453 case 'a': // Address register
455 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
456 else if (VT == MVT::i128)
457 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
458 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
460 case 'f': // Floating-point register
462 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
463 else if (VT == MVT::f128)
464 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
465 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
468 if (Constraint[0] == '{') {
469 // We need to override the default register parsing for GPRs and FPRs
470 // because the interpretation depends on VT. The internal names of
471 // the registers are also different from the external names
472 // (F0D and F0S instead of F0, etc.).
473 if (Constraint[1] == 'r') {
475 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
476 SystemZMC::GR32Regs);
478 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
479 SystemZMC::GR128Regs);
480 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
481 SystemZMC::GR64Regs);
483 if (Constraint[1] == 'f') {
485 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
486 SystemZMC::FP32Regs);
488 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
489 SystemZMC::FP128Regs);
490 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
491 SystemZMC::FP64Regs);
494 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
497 void SystemZTargetLowering::
498 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
499 std::vector<SDValue> &Ops,
500 SelectionDAG &DAG) const {
501 // Only support length 1 constraints for now.
502 if (Constraint.length() == 1) {
503 switch (Constraint[0]) {
504 case 'I': // Unsigned 8-bit constant
505 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
506 if (isUInt<8>(C->getZExtValue()))
507 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
511 case 'J': // Unsigned 12-bit constant
512 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
513 if (isUInt<12>(C->getZExtValue()))
514 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
518 case 'K': // Signed 16-bit constant
519 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
520 if (isInt<16>(C->getSExtValue()))
521 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
525 case 'L': // Signed 20-bit displacement (on all targets we support)
526 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
527 if (isInt<20>(C->getSExtValue()))
528 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
532 case 'M': // 0x7fffffff
533 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
534 if (C->getZExtValue() == 0x7fffffff)
535 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
540 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
543 //===----------------------------------------------------------------------===//
544 // Calling conventions
545 //===----------------------------------------------------------------------===//
547 #include "SystemZGenCallingConv.inc"
549 bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
550 Type *ToType) const {
551 return isTruncateFree(FromType, ToType);
554 bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
555 if (!CI->isTailCall())
560 // Value is a value that has been passed to us in the location described by VA
561 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
562 // any loads onto Chain.
563 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
564 CCValAssign &VA, SDValue Chain,
566 // If the argument has been promoted from a smaller type, insert an
567 // assertion to capture this.
568 if (VA.getLocInfo() == CCValAssign::SExt)
569 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
570 DAG.getValueType(VA.getValVT()));
571 else if (VA.getLocInfo() == CCValAssign::ZExt)
572 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
573 DAG.getValueType(VA.getValVT()));
576 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
577 else if (VA.getLocInfo() == CCValAssign::Indirect)
578 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
579 MachinePointerInfo(), false, false, false, 0);
581 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
585 // Value is a value of type VA.getValVT() that we need to copy into
586 // the location described by VA. Return a copy of Value converted to
587 // VA.getValVT(). The caller is responsible for handling indirect values.
588 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
589 CCValAssign &VA, SDValue Value) {
590 switch (VA.getLocInfo()) {
591 case CCValAssign::SExt:
592 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
593 case CCValAssign::ZExt:
594 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
595 case CCValAssign::AExt:
596 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
597 case CCValAssign::Full:
600 llvm_unreachable("Unhandled getLocInfo()");
604 SDValue SystemZTargetLowering::
605 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
606 const SmallVectorImpl<ISD::InputArg> &Ins,
607 SDLoc DL, SelectionDAG &DAG,
608 SmallVectorImpl<SDValue> &InVals) const {
609 MachineFunction &MF = DAG.getMachineFunction();
610 MachineFrameInfo *MFI = MF.getFrameInfo();
611 MachineRegisterInfo &MRI = MF.getRegInfo();
612 SystemZMachineFunctionInfo *FuncInfo =
613 MF.getInfo<SystemZMachineFunctionInfo>();
614 const SystemZFrameLowering *TFL =
615 static_cast<const SystemZFrameLowering *>(TM.getFrameLowering());
617 // Assign locations to all of the incoming arguments.
618 SmallVector<CCValAssign, 16> ArgLocs;
619 CCState CCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
620 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
622 unsigned NumFixedGPRs = 0;
623 unsigned NumFixedFPRs = 0;
624 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
626 CCValAssign &VA = ArgLocs[I];
627 EVT LocVT = VA.getLocVT();
629 // Arguments passed in registers
630 const TargetRegisterClass *RC;
631 switch (LocVT.getSimpleVT().SimpleTy) {
633 // Integers smaller than i64 should be promoted to i64.
634 llvm_unreachable("Unexpected argument type");
637 RC = &SystemZ::GR32BitRegClass;
641 RC = &SystemZ::GR64BitRegClass;
645 RC = &SystemZ::FP32BitRegClass;
649 RC = &SystemZ::FP64BitRegClass;
653 unsigned VReg = MRI.createVirtualRegister(RC);
654 MRI.addLiveIn(VA.getLocReg(), VReg);
655 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
657 assert(VA.isMemLoc() && "Argument not register or memory");
659 // Create the frame index object for this incoming parameter.
660 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
661 VA.getLocMemOffset(), true);
663 // Create the SelectionDAG nodes corresponding to a load
664 // from this parameter. Unpromoted ints and floats are
665 // passed as right-justified 8-byte values.
666 EVT PtrVT = getPointerTy();
667 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
668 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
669 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4));
670 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
671 MachinePointerInfo::getFixedStack(FI),
672 false, false, false, 0);
675 // Convert the value of the argument register into the value that's
677 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
681 // Save the number of non-varargs registers for later use by va_start, etc.
682 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
683 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
685 // Likewise the address (in the form of a frame index) of where the
686 // first stack vararg would be. The 1-byte size here is arbitrary.
687 int64_t StackSize = CCInfo.getNextStackOffset();
688 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
690 // ...and a similar frame index for the caller-allocated save area
691 // that will be used to store the incoming registers.
692 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
693 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
694 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
696 // Store the FPR varargs in the reserved frame slots. (We store the
697 // GPRs as part of the prologue.)
698 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
699 SDValue MemOps[SystemZ::NumArgFPRs];
700 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
701 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
702 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
703 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
704 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
705 &SystemZ::FP64BitRegClass);
706 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
707 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
708 MachinePointerInfo::getFixedStack(FI),
712 // Join the stores, which are independent of one another.
713 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
714 &MemOps[NumFixedFPRs],
715 SystemZ::NumArgFPRs - NumFixedFPRs);
722 static bool canUseSiblingCall(CCState ArgCCInfo,
723 SmallVectorImpl<CCValAssign> &ArgLocs) {
724 // Punt if there are any indirect or stack arguments, or if the call
725 // needs the call-saved argument register R6.
726 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
727 CCValAssign &VA = ArgLocs[I];
728 if (VA.getLocInfo() == CCValAssign::Indirect)
732 unsigned Reg = VA.getLocReg();
733 if (Reg == SystemZ::R6W || Reg == SystemZ::R6D)
740 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
741 SmallVectorImpl<SDValue> &InVals) const {
742 SelectionDAG &DAG = CLI.DAG;
744 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
745 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
746 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
747 SDValue Chain = CLI.Chain;
748 SDValue Callee = CLI.Callee;
749 bool &IsTailCall = CLI.IsTailCall;
750 CallingConv::ID CallConv = CLI.CallConv;
751 bool IsVarArg = CLI.IsVarArg;
752 MachineFunction &MF = DAG.getMachineFunction();
753 EVT PtrVT = getPointerTy();
755 // Analyze the operands of the call, assigning locations to each operand.
756 SmallVector<CCValAssign, 16> ArgLocs;
757 CCState ArgCCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
758 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
760 // We don't support GuaranteedTailCallOpt, only automatically-detected
762 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs))
765 // Get a count of how many bytes are to be pushed on the stack.
766 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
768 // Mark the start of the call.
770 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true),
773 // Copy argument values to their designated locations.
774 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
775 SmallVector<SDValue, 8> MemOpChains;
777 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
778 CCValAssign &VA = ArgLocs[I];
779 SDValue ArgValue = OutVals[I];
781 if (VA.getLocInfo() == CCValAssign::Indirect) {
782 // Store the argument in a stack slot and pass its address.
783 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
784 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
785 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
786 MachinePointerInfo::getFixedStack(FI),
788 ArgValue = SpillSlot;
790 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
793 // Queue up the argument copies and emit them at the end.
794 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
796 assert(VA.isMemLoc() && "Argument not register or memory");
798 // Work out the address of the stack slot. Unpromoted ints and
799 // floats are passed as right-justified 8-byte values.
800 if (!StackPtr.getNode())
801 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
802 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
803 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
805 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
806 DAG.getIntPtrConstant(Offset));
809 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
810 MachinePointerInfo(),
815 // Join the stores, which are independent of one another.
816 if (!MemOpChains.empty())
817 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
818 &MemOpChains[0], MemOpChains.size());
820 // Accept direct calls by converting symbolic call addresses to the
821 // associated Target* opcodes. Force %r1 to be used for indirect
824 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
825 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
826 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
827 } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
828 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
829 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
830 } else if (IsTailCall) {
831 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
832 Glue = Chain.getValue(1);
833 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
836 // Build a sequence of copy-to-reg nodes, chained and glued together.
837 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
838 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
839 RegsToPass[I].second, Glue);
840 Glue = Chain.getValue(1);
843 // The first call operand is the chain and the second is the target address.
844 SmallVector<SDValue, 8> Ops;
845 Ops.push_back(Chain);
846 Ops.push_back(Callee);
848 // Add argument registers to the end of the list so that they are
849 // known live into the call.
850 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
851 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
852 RegsToPass[I].second.getValueType()));
854 // Glue the call to the argument copies, if any.
859 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
861 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, &Ops[0], Ops.size());
862 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
863 Glue = Chain.getValue(1);
865 // Mark the end of the call, which is glued to the call itself.
866 Chain = DAG.getCALLSEQ_END(Chain,
867 DAG.getConstant(NumBytes, PtrVT, true),
868 DAG.getConstant(0, PtrVT, true),
870 Glue = Chain.getValue(1);
872 // Assign locations to each value returned by this call.
873 SmallVector<CCValAssign, 16> RetLocs;
874 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
875 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
877 // Copy all of the result registers out of their specified physreg.
878 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
879 CCValAssign &VA = RetLocs[I];
881 // Copy the value out, gluing the copy to the end of the call sequence.
882 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
883 VA.getLocVT(), Glue);
884 Chain = RetValue.getValue(1);
885 Glue = RetValue.getValue(2);
887 // Convert the value of the return register into the value that's
889 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
896 SystemZTargetLowering::LowerReturn(SDValue Chain,
897 CallingConv::ID CallConv, bool IsVarArg,
898 const SmallVectorImpl<ISD::OutputArg> &Outs,
899 const SmallVectorImpl<SDValue> &OutVals,
900 SDLoc DL, SelectionDAG &DAG) const {
901 MachineFunction &MF = DAG.getMachineFunction();
903 // Assign locations to each returned value.
904 SmallVector<CCValAssign, 16> RetLocs;
905 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
906 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
908 // Quick exit for void returns
910 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
912 // Copy the result values into the output registers.
914 SmallVector<SDValue, 4> RetOps;
915 RetOps.push_back(Chain);
916 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
917 CCValAssign &VA = RetLocs[I];
918 SDValue RetValue = OutVals[I];
920 // Make the return register live on exit.
921 assert(VA.isRegLoc() && "Can only return in registers!");
923 // Promote the value as required.
924 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
926 // Chain and glue the copies together.
927 unsigned Reg = VA.getLocReg();
928 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
929 Glue = Chain.getValue(1);
930 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
933 // Update chain and glue.
936 RetOps.push_back(Glue);
938 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other,
939 RetOps.data(), RetOps.size());
942 // CC is a comparison that will be implemented using an integer or
943 // floating-point comparison. Return the condition code mask for
944 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
945 // unsigned comparisons and clear for signed ones. In the floating-point
946 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
947 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
949 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
950 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
951 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
955 llvm_unreachable("Invalid integer condition!");
964 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
965 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
970 // If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1
971 // can be converted to a comparison against zero, adjust the operands
973 static void adjustZeroCmp(SelectionDAG &DAG, bool &IsUnsigned,
974 SDValue &CmpOp0, SDValue &CmpOp1,
979 ConstantSDNode *ConstOp1 = dyn_cast<ConstantSDNode>(CmpOp1.getNode());
983 int64_t Value = ConstOp1->getSExtValue();
984 if ((Value == -1 && CCMask == SystemZ::CCMASK_CMP_GT) ||
985 (Value == -1 && CCMask == SystemZ::CCMASK_CMP_LE) ||
986 (Value == 1 && CCMask == SystemZ::CCMASK_CMP_LT) ||
987 (Value == 1 && CCMask == SystemZ::CCMASK_CMP_GE)) {
988 CCMask ^= SystemZ::CCMASK_CMP_EQ;
989 CmpOp1 = DAG.getConstant(0, CmpOp1.getValueType());
993 // If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1
994 // is suitable for CLI(Y), CHHSI or CLHHSI, adjust the operands as necessary.
995 static void adjustSubwordCmp(SelectionDAG &DAG, bool &IsUnsigned,
996 SDValue &CmpOp0, SDValue &CmpOp1,
998 // For us to make any changes, it must a comparison between a single-use
999 // load and a constant.
1000 if (!CmpOp0.hasOneUse() ||
1001 CmpOp0.getOpcode() != ISD::LOAD ||
1002 CmpOp1.getOpcode() != ISD::Constant)
1005 // We must have an 8- or 16-bit load.
1006 LoadSDNode *Load = cast<LoadSDNode>(CmpOp0);
1007 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
1008 if (NumBits != 8 && NumBits != 16)
1011 // The load must be an extending one and the constant must be within the
1012 // range of the unextended value.
1013 ConstantSDNode *Constant = cast<ConstantSDNode>(CmpOp1);
1014 uint64_t Value = Constant->getZExtValue();
1015 uint64_t Mask = (1 << NumBits) - 1;
1016 if (Load->getExtensionType() == ISD::SEXTLOAD) {
1017 int64_t SignedValue = Constant->getSExtValue();
1018 if (uint64_t(SignedValue) + (1ULL << (NumBits - 1)) > Mask)
1020 // Unsigned comparison between two sign-extended values is equivalent
1021 // to unsigned comparison between two zero-extended values.
1024 else if (CCMask == SystemZ::CCMASK_CMP_EQ ||
1025 CCMask == SystemZ::CCMASK_CMP_NE)
1026 // Any choice of IsUnsigned is OK for equality comparisons.
1027 // We could use either CHHSI or CLHHSI for 16-bit comparisons,
1028 // but since we use CLHHSI for zero extensions, it seems better
1029 // to be consistent and do the same here.
1030 Value &= Mask, IsUnsigned = true;
1031 else if (NumBits == 8) {
1032 // Try to treat the comparison as unsigned, so that we can use CLI.
1033 // Adjust CCMask and Value as necessary.
1034 if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_LT)
1035 // Test whether the high bit of the byte is set.
1036 Value = 127, CCMask = SystemZ::CCMASK_CMP_GT, IsUnsigned = true;
1037 else if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_GE)
1038 // Test whether the high bit of the byte is clear.
1039 Value = 128, CCMask = SystemZ::CCMASK_CMP_LT, IsUnsigned = true;
1041 // No instruction exists for this combination.
1044 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
1047 // Signed comparison between two zero-extended values is equivalent
1048 // to unsigned comparison.
1053 // Make sure that the first operand is an i32 of the right extension type.
1054 ISD::LoadExtType ExtType = IsUnsigned ? ISD::ZEXTLOAD : ISD::SEXTLOAD;
1055 if (CmpOp0.getValueType() != MVT::i32 ||
1056 Load->getExtensionType() != ExtType)
1057 CmpOp0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
1058 Load->getChain(), Load->getBasePtr(),
1059 Load->getPointerInfo(), Load->getMemoryVT(),
1060 Load->isVolatile(), Load->isNonTemporal(),
1061 Load->getAlignment());
1063 // Make sure that the second operand is an i32 with the right value.
1064 if (CmpOp1.getValueType() != MVT::i32 ||
1065 Value != Constant->getZExtValue())
1066 CmpOp1 = DAG.getConstant(Value, MVT::i32);
1069 // Return true if a comparison described by CCMask, CmpOp0 and CmpOp1
1070 // is an equality comparison that is better implemented using unsigned
1071 // rather than signed comparison instructions.
1072 static bool preferUnsignedComparison(SelectionDAG &DAG, SDValue CmpOp0,
1073 SDValue CmpOp1, unsigned CCMask) {
1074 // The test must be for equality or inequality.
1075 if (CCMask != SystemZ::CCMASK_CMP_EQ && CCMask != SystemZ::CCMASK_CMP_NE)
1078 if (CmpOp1.getOpcode() == ISD::Constant) {
1079 uint64_t Value = cast<ConstantSDNode>(CmpOp1)->getSExtValue();
1081 // If we're comparing with memory, prefer unsigned comparisons for
1082 // values that are in the unsigned 16-bit range but not the signed
1083 // 16-bit range. We want to use CLFHSI and CLGHSI.
1084 if (CmpOp0.hasOneUse() &&
1085 ISD::isNormalLoad(CmpOp0.getNode()) &&
1086 (Value >= 32768 && Value < 65536))
1089 // Use unsigned comparisons for values that are in the CLGFI range
1090 // but not in the CGFI range.
1091 if (CmpOp0.getValueType() == MVT::i64 && (Value >> 31) == 1)
1097 // Prefer CL for zero-extended loads.
1098 if (CmpOp1.getOpcode() == ISD::ZERO_EXTEND ||
1099 ISD::isZEXTLoad(CmpOp1.getNode()))
1102 // ...and for "in-register" zero extensions.
1103 if (CmpOp1.getOpcode() == ISD::AND && CmpOp1.getValueType() == MVT::i64) {
1104 SDValue Mask = CmpOp1.getOperand(1);
1105 if (Mask.getOpcode() == ISD::Constant &&
1106 cast<ConstantSDNode>(Mask)->getZExtValue() == 0xffffffff)
1113 // Return a target node that compares CmpOp0 with CmpOp1 and stores a
1114 // 2-bit result in CC. Set CCValid to the CCMASK_* of all possible
1115 // 2-bit results and CCMask to the subset of those results that are
1116 // associated with Cond.
1117 static SDValue emitCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
1118 ISD::CondCode Cond, unsigned &CCValid,
1120 bool IsUnsigned = false;
1121 CCMask = CCMaskForCondCode(Cond);
1122 if (CmpOp0.getValueType().isFloatingPoint())
1123 CCValid = SystemZ::CCMASK_FCMP;
1125 IsUnsigned = CCMask & SystemZ::CCMASK_CMP_UO;
1126 CCValid = SystemZ::CCMASK_ICMP;
1128 adjustZeroCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask);
1129 adjustSubwordCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask);
1130 if (preferUnsignedComparison(DAG, CmpOp0, CmpOp1, CCMask))
1135 return DAG.getNode((IsUnsigned ? SystemZISD::UCMP : SystemZISD::CMP),
1136 DL, MVT::Glue, CmpOp0, CmpOp1);
1139 // Implement a 32-bit *MUL_LOHI operation by extending both operands to
1140 // 64 bits. Extend is the extension type to use. Store the high part
1141 // in Hi and the low part in Lo.
1142 static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL,
1143 unsigned Extend, SDValue Op0, SDValue Op1,
1144 SDValue &Hi, SDValue &Lo) {
1145 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
1146 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
1147 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
1148 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, DAG.getConstant(32, MVT::i64));
1149 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
1150 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
1153 // Lower a binary operation that produces two VT results, one in each
1154 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
1155 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
1156 // on the extended Op0 and (unextended) Op1. Store the even register result
1157 // in Even and the odd register result in Odd.
1158 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
1159 unsigned Extend, unsigned Opcode,
1160 SDValue Op0, SDValue Op1,
1161 SDValue &Even, SDValue &Odd) {
1162 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
1163 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
1164 SDValue(In128, 0), Op1);
1165 bool Is32Bit = is32Bit(VT);
1166 SDValue SubReg0 = DAG.getTargetConstant(SystemZ::even128(Is32Bit), VT);
1167 SDValue SubReg1 = DAG.getTargetConstant(SystemZ::odd128(Is32Bit), VT);
1168 SDNode *Reg0 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
1169 VT, Result, SubReg0);
1170 SDNode *Reg1 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
1171 VT, Result, SubReg1);
1172 Even = SDValue(Reg0, 0);
1173 Odd = SDValue(Reg1, 0);
1176 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1177 SDValue Chain = Op.getOperand(0);
1178 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1179 SDValue CmpOp0 = Op.getOperand(2);
1180 SDValue CmpOp1 = Op.getOperand(3);
1181 SDValue Dest = Op.getOperand(4);
1184 unsigned CCValid, CCMask;
1185 SDValue Flags = emitCmp(DAG, CmpOp0, CmpOp1, CC, CCValid, CCMask);
1186 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
1187 Chain, DAG.getConstant(CCValid, MVT::i32),
1188 DAG.getConstant(CCMask, MVT::i32), Dest, Flags);
1191 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
1192 SelectionDAG &DAG) const {
1193 SDValue CmpOp0 = Op.getOperand(0);
1194 SDValue CmpOp1 = Op.getOperand(1);
1195 SDValue TrueOp = Op.getOperand(2);
1196 SDValue FalseOp = Op.getOperand(3);
1197 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1200 unsigned CCValid, CCMask;
1201 SDValue Flags = emitCmp(DAG, CmpOp0, CmpOp1, CC, CCValid, CCMask);
1203 SmallVector<SDValue, 5> Ops;
1204 Ops.push_back(TrueOp);
1205 Ops.push_back(FalseOp);
1206 Ops.push_back(DAG.getConstant(CCValid, MVT::i32));
1207 Ops.push_back(DAG.getConstant(CCMask, MVT::i32));
1208 Ops.push_back(Flags);
1210 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
1211 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size());
1214 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
1215 SelectionDAG &DAG) const {
1217 const GlobalValue *GV = Node->getGlobal();
1218 int64_t Offset = Node->getOffset();
1219 EVT PtrVT = getPointerTy();
1220 Reloc::Model RM = TM.getRelocationModel();
1221 CodeModel::Model CM = TM.getCodeModel();
1224 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
1225 // Make sure that the offset is aligned to a halfword. If it isn't,
1226 // create an "anchor" at the previous 12-bit boundary.
1227 // FIXME check whether there is a better way of handling this.
1229 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT,
1230 Offset & ~uint64_t(0xfff));
1233 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Offset);
1236 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1238 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
1239 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1240 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
1241 MachinePointerInfo::getGOT(), false, false, false, 0);
1244 // If there was a non-zero offset that we didn't fold, create an explicit
1247 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
1248 DAG.getConstant(Offset, PtrVT));
1253 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
1254 SelectionDAG &DAG) const {
1256 const GlobalValue *GV = Node->getGlobal();
1257 EVT PtrVT = getPointerTy();
1258 TLSModel::Model model = TM.getTLSModel(GV);
1260 if (model != TLSModel::LocalExec)
1261 llvm_unreachable("only local-exec TLS mode supported");
1263 // The high part of the thread pointer is in access register 0.
1264 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1265 DAG.getConstant(0, MVT::i32));
1266 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
1268 // The low part of the thread pointer is in access register 1.
1269 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1270 DAG.getConstant(1, MVT::i32));
1271 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
1273 // Merge them into a single 64-bit address.
1274 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
1275 DAG.getConstant(32, PtrVT));
1276 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
1278 // Get the offset of GA from the thread pointer.
1279 SystemZConstantPoolValue *CPV =
1280 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
1282 // Force the offset into the constant pool and load it from there.
1283 SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8);
1284 SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1285 CPAddr, MachinePointerInfo::getConstantPool(),
1286 false, false, false, 0);
1288 // Add the base and offset together.
1289 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
1292 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
1293 SelectionDAG &DAG) const {
1295 const BlockAddress *BA = Node->getBlockAddress();
1296 int64_t Offset = Node->getOffset();
1297 EVT PtrVT = getPointerTy();
1299 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
1300 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1304 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
1305 SelectionDAG &DAG) const {
1307 EVT PtrVT = getPointerTy();
1308 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1310 // Use LARL to load the address of the table.
1311 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1314 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
1315 SelectionDAG &DAG) const {
1317 EVT PtrVT = getPointerTy();
1320 if (CP->isMachineConstantPoolEntry())
1321 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1322 CP->getAlignment());
1324 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1325 CP->getAlignment(), CP->getOffset());
1327 // Use LARL to load the address of the constant pool entry.
1328 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1331 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
1332 SelectionDAG &DAG) const {
1334 SDValue In = Op.getOperand(0);
1335 EVT InVT = In.getValueType();
1336 EVT ResVT = Op.getValueType();
1338 SDValue SubReg32 = DAG.getTargetConstant(SystemZ::subreg_32bit, MVT::i64);
1339 SDValue Shift32 = DAG.getConstant(32, MVT::i64);
1340 if (InVT == MVT::i32 && ResVT == MVT::f32) {
1341 SDValue In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
1342 SDValue Shift = DAG.getNode(ISD::SHL, DL, MVT::i64, In64, Shift32);
1343 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, Shift);
1344 SDNode *Out = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
1345 MVT::f32, Out64, SubReg32);
1346 return SDValue(Out, 0);
1348 if (InVT == MVT::f32 && ResVT == MVT::i32) {
1349 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
1350 SDNode *In64 = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
1351 MVT::f64, SDValue(U64, 0), In, SubReg32);
1352 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, SDValue(In64, 0));
1353 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64, Shift32);
1354 SDValue Out = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
1357 llvm_unreachable("Unexpected bitcast combination");
1360 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
1361 SelectionDAG &DAG) const {
1362 MachineFunction &MF = DAG.getMachineFunction();
1363 SystemZMachineFunctionInfo *FuncInfo =
1364 MF.getInfo<SystemZMachineFunctionInfo>();
1365 EVT PtrVT = getPointerTy();
1367 SDValue Chain = Op.getOperand(0);
1368 SDValue Addr = Op.getOperand(1);
1369 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1372 // The initial values of each field.
1373 const unsigned NumFields = 4;
1374 SDValue Fields[NumFields] = {
1375 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT),
1376 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT),
1377 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
1378 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
1381 // Store each field into its respective slot.
1382 SDValue MemOps[NumFields];
1383 unsigned Offset = 0;
1384 for (unsigned I = 0; I < NumFields; ++I) {
1385 SDValue FieldAddr = Addr;
1387 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
1388 DAG.getIntPtrConstant(Offset));
1389 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
1390 MachinePointerInfo(SV, Offset),
1394 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps, NumFields);
1397 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
1398 SelectionDAG &DAG) const {
1399 SDValue Chain = Op.getOperand(0);
1400 SDValue DstPtr = Op.getOperand(1);
1401 SDValue SrcPtr = Op.getOperand(2);
1402 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
1403 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
1406 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32),
1407 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
1408 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
1411 SDValue SystemZTargetLowering::
1412 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
1413 SDValue Chain = Op.getOperand(0);
1414 SDValue Size = Op.getOperand(1);
1417 unsigned SPReg = getStackPointerRegisterToSaveRestore();
1419 // Get a reference to the stack pointer.
1420 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
1422 // Get the new stack pointer value.
1423 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
1425 // Copy the new stack pointer back.
1426 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
1428 // The allocated data lives above the 160 bytes allocated for the standard
1429 // frame, plus any outgoing stack arguments. We don't know how much that
1430 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
1431 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
1432 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
1434 SDValue Ops[2] = { Result, Chain };
1435 return DAG.getMergeValues(Ops, 2, DL);
1438 SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
1439 SelectionDAG &DAG) const {
1440 EVT VT = Op.getValueType();
1444 // Just do a normal 64-bit multiplication and extract the results.
1445 // We define this so that it can be used for constant division.
1446 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
1447 Op.getOperand(1), Ops[1], Ops[0]);
1449 // Do a full 128-bit multiplication based on UMUL_LOHI64:
1451 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
1453 // but using the fact that the upper halves are either all zeros
1456 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
1458 // and grouping the right terms together since they are quicker than the
1461 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
1462 SDValue C63 = DAG.getConstant(63, MVT::i64);
1463 SDValue LL = Op.getOperand(0);
1464 SDValue RL = Op.getOperand(1);
1465 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
1466 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
1467 // UMUL_LOHI64 returns the low result in the odd register and the high
1468 // result in the even register. SMUL_LOHI is defined to return the
1469 // low half first, so the results are in reverse order.
1470 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
1471 LL, RL, Ops[1], Ops[0]);
1472 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
1473 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
1474 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
1475 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
1477 return DAG.getMergeValues(Ops, 2, DL);
1480 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
1481 SelectionDAG &DAG) const {
1482 EVT VT = Op.getValueType();
1486 // Just do a normal 64-bit multiplication and extract the results.
1487 // We define this so that it can be used for constant division.
1488 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
1489 Op.getOperand(1), Ops[1], Ops[0]);
1491 // UMUL_LOHI64 returns the low result in the odd register and the high
1492 // result in the even register. UMUL_LOHI is defined to return the
1493 // low half first, so the results are in reverse order.
1494 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
1495 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
1496 return DAG.getMergeValues(Ops, 2, DL);
1499 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
1500 SelectionDAG &DAG) const {
1501 SDValue Op0 = Op.getOperand(0);
1502 SDValue Op1 = Op.getOperand(1);
1503 EVT VT = Op.getValueType();
1507 // We use DSGF for 32-bit division.
1509 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
1510 Opcode = SystemZISD::SDIVREM32;
1511 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
1512 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
1513 Opcode = SystemZISD::SDIVREM32;
1515 Opcode = SystemZISD::SDIVREM64;
1517 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
1518 // input is "don't care". The instruction returns the remainder in
1519 // the even register and the quotient in the odd register.
1521 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
1522 Op0, Op1, Ops[1], Ops[0]);
1523 return DAG.getMergeValues(Ops, 2, DL);
1526 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
1527 SelectionDAG &DAG) const {
1528 EVT VT = Op.getValueType();
1531 // DL(G) uses a double-width dividend, so we need to clear the even
1532 // register in the GR128 input. The instruction returns the remainder
1533 // in the even register and the quotient in the odd register.
1536 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
1537 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
1539 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
1540 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
1541 return DAG.getMergeValues(Ops, 2, DL);
1544 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
1545 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
1547 // Get the known-zero masks for each operand.
1548 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
1549 APInt KnownZero[2], KnownOne[2];
1550 DAG.ComputeMaskedBits(Ops[0], KnownZero[0], KnownOne[0]);
1551 DAG.ComputeMaskedBits(Ops[1], KnownZero[1], KnownOne[1]);
1553 // See if the upper 32 bits of one operand and the lower 32 bits of the
1554 // other are known zero. They are the low and high operands respectively.
1555 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
1556 KnownZero[1].getZExtValue() };
1558 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
1560 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
1565 SDValue LowOp = Ops[Low];
1566 SDValue HighOp = Ops[High];
1568 // If the high part is a constant, we're better off using IILH.
1569 if (HighOp.getOpcode() == ISD::Constant)
1572 // If the low part is a constant that is outside the range of LHI,
1573 // then we're better off using IILF.
1574 if (LowOp.getOpcode() == ISD::Constant) {
1575 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
1576 if (!isInt<16>(Value))
1580 // Check whether the high part is an AND that doesn't change the
1581 // high 32 bits and just masks out low bits. We can skip it if so.
1582 if (HighOp.getOpcode() == ISD::AND &&
1583 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
1584 ConstantSDNode *MaskNode = cast<ConstantSDNode>(HighOp.getOperand(1));
1585 uint64_t Mask = MaskNode->getZExtValue() | Masks[High];
1586 if ((Mask >> 32) == 0xffffffff)
1587 HighOp = HighOp.getOperand(0);
1590 // Take advantage of the fact that all GR32 operations only change the
1591 // low 32 bits by truncating Low to an i32 and inserting it directly
1592 // using a subreg. The interesting cases are those where the truncation
1595 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
1596 SDValue SubReg32 = DAG.getTargetConstant(SystemZ::subreg_32bit, MVT::i64);
1597 SDNode *Result = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
1598 MVT::i64, HighOp, Low32, SubReg32);
1599 return SDValue(Result, 0);
1602 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
1603 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
1604 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
1606 unsigned Opcode) const {
1607 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
1609 // 32-bit operations need no code outside the main loop.
1610 EVT NarrowVT = Node->getMemoryVT();
1611 EVT WideVT = MVT::i32;
1612 if (NarrowVT == WideVT)
1615 int64_t BitSize = NarrowVT.getSizeInBits();
1616 SDValue ChainIn = Node->getChain();
1617 SDValue Addr = Node->getBasePtr();
1618 SDValue Src2 = Node->getVal();
1619 MachineMemOperand *MMO = Node->getMemOperand();
1621 EVT PtrVT = Addr.getValueType();
1623 // Convert atomic subtracts of constants into additions.
1624 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
1625 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Src2)) {
1626 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
1627 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
1630 // Get the address of the containing word.
1631 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
1632 DAG.getConstant(-4, PtrVT));
1634 // Get the number of bits that the word must be rotated left in order
1635 // to bring the field to the top bits of a GR32.
1636 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
1637 DAG.getConstant(3, PtrVT));
1638 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
1640 // Get the complementing shift amount, for rotating a field in the top
1641 // bits back to its proper position.
1642 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
1643 DAG.getConstant(0, WideVT), BitShift);
1645 // Extend the source operand to 32 bits and prepare it for the inner loop.
1646 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
1647 // operations require the source to be shifted in advance. (This shift
1648 // can be folded if the source is constant.) For AND and NAND, the lower
1649 // bits must be set, while for other opcodes they should be left clear.
1650 if (Opcode != SystemZISD::ATOMIC_SWAPW)
1651 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
1652 DAG.getConstant(32 - BitSize, WideVT));
1653 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
1654 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
1655 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
1656 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT));
1658 // Construct the ATOMIC_LOADW_* node.
1659 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
1660 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
1661 DAG.getConstant(BitSize, WideVT) };
1662 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
1663 array_lengthof(Ops),
1666 // Rotate the result of the final CS so that the field is in the lower
1667 // bits of a GR32, then truncate it.
1668 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
1669 DAG.getConstant(BitSize, WideVT));
1670 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
1672 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
1673 return DAG.getMergeValues(RetOps, 2, DL);
1676 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
1677 // into a fullword ATOMIC_CMP_SWAPW operation.
1678 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
1679 SelectionDAG &DAG) const {
1680 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
1682 // We have native support for 32-bit compare and swap.
1683 EVT NarrowVT = Node->getMemoryVT();
1684 EVT WideVT = MVT::i32;
1685 if (NarrowVT == WideVT)
1688 int64_t BitSize = NarrowVT.getSizeInBits();
1689 SDValue ChainIn = Node->getOperand(0);
1690 SDValue Addr = Node->getOperand(1);
1691 SDValue CmpVal = Node->getOperand(2);
1692 SDValue SwapVal = Node->getOperand(3);
1693 MachineMemOperand *MMO = Node->getMemOperand();
1695 EVT PtrVT = Addr.getValueType();
1697 // Get the address of the containing word.
1698 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
1699 DAG.getConstant(-4, PtrVT));
1701 // Get the number of bits that the word must be rotated left in order
1702 // to bring the field to the top bits of a GR32.
1703 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
1704 DAG.getConstant(3, PtrVT));
1705 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
1707 // Get the complementing shift amount, for rotating a field in the top
1708 // bits back to its proper position.
1709 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
1710 DAG.getConstant(0, WideVT), BitShift);
1712 // Construct the ATOMIC_CMP_SWAPW node.
1713 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
1714 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
1715 NegBitShift, DAG.getConstant(BitSize, WideVT) };
1716 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
1717 VTList, Ops, array_lengthof(Ops),
1722 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
1723 SelectionDAG &DAG) const {
1724 MachineFunction &MF = DAG.getMachineFunction();
1725 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
1726 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
1727 SystemZ::R15D, Op.getValueType());
1730 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
1731 SelectionDAG &DAG) const {
1732 MachineFunction &MF = DAG.getMachineFunction();
1733 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
1734 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
1735 SystemZ::R15D, Op.getOperand(1));
1738 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
1739 SelectionDAG &DAG) const {
1740 switch (Op.getOpcode()) {
1742 return lowerBR_CC(Op, DAG);
1743 case ISD::SELECT_CC:
1744 return lowerSELECT_CC(Op, DAG);
1745 case ISD::GlobalAddress:
1746 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
1747 case ISD::GlobalTLSAddress:
1748 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
1749 case ISD::BlockAddress:
1750 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
1751 case ISD::JumpTable:
1752 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
1753 case ISD::ConstantPool:
1754 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
1756 return lowerBITCAST(Op, DAG);
1758 return lowerVASTART(Op, DAG);
1760 return lowerVACOPY(Op, DAG);
1761 case ISD::DYNAMIC_STACKALLOC:
1762 return lowerDYNAMIC_STACKALLOC(Op, DAG);
1763 case ISD::SMUL_LOHI:
1764 return lowerSMUL_LOHI(Op, DAG);
1765 case ISD::UMUL_LOHI:
1766 return lowerUMUL_LOHI(Op, DAG);
1768 return lowerSDIVREM(Op, DAG);
1770 return lowerUDIVREM(Op, DAG);
1772 return lowerOR(Op, DAG);
1773 case ISD::ATOMIC_SWAP:
1774 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_SWAPW);
1775 case ISD::ATOMIC_LOAD_ADD:
1776 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
1777 case ISD::ATOMIC_LOAD_SUB:
1778 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
1779 case ISD::ATOMIC_LOAD_AND:
1780 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
1781 case ISD::ATOMIC_LOAD_OR:
1782 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
1783 case ISD::ATOMIC_LOAD_XOR:
1784 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
1785 case ISD::ATOMIC_LOAD_NAND:
1786 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
1787 case ISD::ATOMIC_LOAD_MIN:
1788 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
1789 case ISD::ATOMIC_LOAD_MAX:
1790 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
1791 case ISD::ATOMIC_LOAD_UMIN:
1792 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
1793 case ISD::ATOMIC_LOAD_UMAX:
1794 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
1795 case ISD::ATOMIC_CMP_SWAP:
1796 return lowerATOMIC_CMP_SWAP(Op, DAG);
1797 case ISD::STACKSAVE:
1798 return lowerSTACKSAVE(Op, DAG);
1799 case ISD::STACKRESTORE:
1800 return lowerSTACKRESTORE(Op, DAG);
1802 llvm_unreachable("Unexpected node to lower");
1806 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
1807 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
1812 OPCODE(PCREL_WRAPPER);
1816 OPCODE(SELECT_CCMASK);
1817 OPCODE(ADJDYNALLOC);
1818 OPCODE(EXTRACT_ACCESS);
1819 OPCODE(UMUL_LOHI64);
1827 OPCODE(SEARCH_STRING);
1829 OPCODE(ATOMIC_SWAPW);
1830 OPCODE(ATOMIC_LOADW_ADD);
1831 OPCODE(ATOMIC_LOADW_SUB);
1832 OPCODE(ATOMIC_LOADW_AND);
1833 OPCODE(ATOMIC_LOADW_OR);
1834 OPCODE(ATOMIC_LOADW_XOR);
1835 OPCODE(ATOMIC_LOADW_NAND);
1836 OPCODE(ATOMIC_LOADW_MIN);
1837 OPCODE(ATOMIC_LOADW_MAX);
1838 OPCODE(ATOMIC_LOADW_UMIN);
1839 OPCODE(ATOMIC_LOADW_UMAX);
1840 OPCODE(ATOMIC_CMP_SWAPW);
1846 //===----------------------------------------------------------------------===//
1848 //===----------------------------------------------------------------------===//
1850 // Create a new basic block after MBB.
1851 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
1852 MachineFunction &MF = *MBB->getParent();
1853 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
1854 MF.insert(llvm::next(MachineFunction::iterator(MBB)), NewMBB);
1858 // Split MBB after MI and return the new block (the one that contains
1859 // instructions after MI).
1860 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
1861 MachineBasicBlock *MBB) {
1862 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
1863 NewMBB->splice(NewMBB->begin(), MBB,
1864 llvm::next(MachineBasicBlock::iterator(MI)),
1866 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
1870 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
1872 SystemZTargetLowering::emitSelect(MachineInstr *MI,
1873 MachineBasicBlock *MBB) const {
1874 const SystemZInstrInfo *TII = TM.getInstrInfo();
1876 unsigned DestReg = MI->getOperand(0).getReg();
1877 unsigned TrueReg = MI->getOperand(1).getReg();
1878 unsigned FalseReg = MI->getOperand(2).getReg();
1879 unsigned CCValid = MI->getOperand(3).getImm();
1880 unsigned CCMask = MI->getOperand(4).getImm();
1881 DebugLoc DL = MI->getDebugLoc();
1883 MachineBasicBlock *StartMBB = MBB;
1884 MachineBasicBlock *JoinMBB = splitBlockAfter(MI, MBB);
1885 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
1888 // BRC CCMask, JoinMBB
1889 // # fallthrough to FalseMBB
1891 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
1892 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
1893 MBB->addSuccessor(JoinMBB);
1894 MBB->addSuccessor(FalseMBB);
1897 // # fallthrough to JoinMBB
1899 MBB->addSuccessor(JoinMBB);
1902 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
1905 BuildMI(*MBB, MBB->begin(), DL, TII->get(SystemZ::PHI), DestReg)
1906 .addReg(TrueReg).addMBB(StartMBB)
1907 .addReg(FalseReg).addMBB(FalseMBB);
1909 MI->eraseFromParent();
1913 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
1914 // StoreOpcode is the store to use and Invert says whether the store should
1915 // happen when the condition is false rather than true. If a STORE ON
1916 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
1918 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
1919 MachineBasicBlock *MBB,
1920 unsigned StoreOpcode, unsigned STOCOpcode,
1921 bool Invert) const {
1922 const SystemZInstrInfo *TII = TM.getInstrInfo();
1924 unsigned SrcReg = MI->getOperand(0).getReg();
1925 MachineOperand Base = MI->getOperand(1);
1926 int64_t Disp = MI->getOperand(2).getImm();
1927 unsigned IndexReg = MI->getOperand(3).getReg();
1928 unsigned CCValid = MI->getOperand(4).getImm();
1929 unsigned CCMask = MI->getOperand(5).getImm();
1930 DebugLoc DL = MI->getDebugLoc();
1932 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
1934 // Use STOCOpcode if possible. We could use different store patterns in
1935 // order to avoid matching the index register, but the performance trade-offs
1936 // might be more complicated in that case.
1937 if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) {
1940 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
1941 .addReg(SrcReg).addOperand(Base).addImm(Disp)
1942 .addImm(CCValid).addImm(CCMask);
1943 MI->eraseFromParent();
1947 // Get the condition needed to branch around the store.
1951 MachineBasicBlock *StartMBB = MBB;
1952 MachineBasicBlock *JoinMBB = splitBlockAfter(MI, MBB);
1953 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
1956 // BRC CCMask, JoinMBB
1957 // # fallthrough to FalseMBB
1959 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
1960 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
1961 MBB->addSuccessor(JoinMBB);
1962 MBB->addSuccessor(FalseMBB);
1965 // store %SrcReg, %Disp(%Index,%Base)
1966 // # fallthrough to JoinMBB
1968 BuildMI(MBB, DL, TII->get(StoreOpcode))
1969 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
1970 MBB->addSuccessor(JoinMBB);
1972 MI->eraseFromParent();
1976 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
1977 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
1978 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
1979 // BitSize is the width of the field in bits, or 0 if this is a partword
1980 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
1981 // is one of the operands. Invert says whether the field should be
1982 // inverted after performing BinOpcode (e.g. for NAND).
1984 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
1985 MachineBasicBlock *MBB,
1988 bool Invert) const {
1989 const SystemZInstrInfo *TII = TM.getInstrInfo();
1990 MachineFunction &MF = *MBB->getParent();
1991 MachineRegisterInfo &MRI = MF.getRegInfo();
1992 bool IsSubWord = (BitSize < 32);
1994 // Extract the operands. Base can be a register or a frame index.
1995 // Src2 can be a register or immediate.
1996 unsigned Dest = MI->getOperand(0).getReg();
1997 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
1998 int64_t Disp = MI->getOperand(2).getImm();
1999 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
2000 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2001 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2002 DebugLoc DL = MI->getDebugLoc();
2004 BitSize = MI->getOperand(6).getImm();
2006 // Subword operations use 32-bit registers.
2007 const TargetRegisterClass *RC = (BitSize <= 32 ?
2008 &SystemZ::GR32BitRegClass :
2009 &SystemZ::GR64BitRegClass);
2010 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2011 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2013 // Get the right opcodes for the displacement.
2014 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2015 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2016 assert(LOpcode && CSOpcode && "Displacement out of range");
2018 // Create virtual registers for temporary results.
2019 unsigned OrigVal = MRI.createVirtualRegister(RC);
2020 unsigned OldVal = MRI.createVirtualRegister(RC);
2021 unsigned NewVal = (BinOpcode || IsSubWord ?
2022 MRI.createVirtualRegister(RC) : Src2.getReg());
2023 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2024 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2026 // Insert a basic block for the main loop.
2027 MachineBasicBlock *StartMBB = MBB;
2028 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
2029 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2033 // %OrigVal = L Disp(%Base)
2034 // # fall through to LoopMMB
2036 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2037 .addOperand(Base).addImm(Disp).addReg(0);
2038 MBB->addSuccessor(LoopMBB);
2041 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
2042 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2043 // %RotatedNewVal = OP %RotatedOldVal, %Src2
2044 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2045 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2047 // # fall through to DoneMMB
2049 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2050 .addReg(OrigVal).addMBB(StartMBB)
2051 .addReg(Dest).addMBB(LoopMBB);
2053 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2054 .addReg(OldVal).addReg(BitShift).addImm(0);
2056 // Perform the operation normally and then invert every bit of the field.
2057 unsigned Tmp = MRI.createVirtualRegister(RC);
2058 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
2059 .addReg(RotatedOldVal).addOperand(Src2);
2061 // XILF with the upper BitSize bits set.
2062 BuildMI(MBB, DL, TII->get(SystemZ::XILF32), RotatedNewVal)
2063 .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize)));
2064 else if (BitSize == 32)
2065 // XILF with every bit set.
2066 BuildMI(MBB, DL, TII->get(SystemZ::XILF32), RotatedNewVal)
2067 .addReg(Tmp).addImm(~uint32_t(0));
2069 // Use LCGR and add -1 to the result, which is more compact than
2070 // an XILF, XILH pair.
2071 unsigned Tmp2 = MRI.createVirtualRegister(RC);
2072 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
2073 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
2074 .addReg(Tmp2).addImm(-1);
2076 } else if (BinOpcode)
2077 // A simply binary operation.
2078 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
2079 .addReg(RotatedOldVal).addOperand(Src2);
2081 // Use RISBG to rotate Src2 into position and use it to replace the
2082 // field in RotatedOldVal.
2083 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
2084 .addReg(RotatedOldVal).addReg(Src2.getReg())
2085 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
2087 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2088 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2089 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2090 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2091 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2092 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2093 MBB->addSuccessor(LoopMBB);
2094 MBB->addSuccessor(DoneMBB);
2096 MI->eraseFromParent();
2100 // Implement EmitInstrWithCustomInserter for pseudo
2101 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
2102 // instruction that should be used to compare the current field with the
2103 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
2104 // for when the current field should be kept. BitSize is the width of
2105 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
2107 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
2108 MachineBasicBlock *MBB,
2109 unsigned CompareOpcode,
2110 unsigned KeepOldMask,
2111 unsigned BitSize) const {
2112 const SystemZInstrInfo *TII = TM.getInstrInfo();
2113 MachineFunction &MF = *MBB->getParent();
2114 MachineRegisterInfo &MRI = MF.getRegInfo();
2115 bool IsSubWord = (BitSize < 32);
2117 // Extract the operands. Base can be a register or a frame index.
2118 unsigned Dest = MI->getOperand(0).getReg();
2119 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2120 int64_t Disp = MI->getOperand(2).getImm();
2121 unsigned Src2 = MI->getOperand(3).getReg();
2122 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2123 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2124 DebugLoc DL = MI->getDebugLoc();
2126 BitSize = MI->getOperand(6).getImm();
2128 // Subword operations use 32-bit registers.
2129 const TargetRegisterClass *RC = (BitSize <= 32 ?
2130 &SystemZ::GR32BitRegClass :
2131 &SystemZ::GR64BitRegClass);
2132 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2133 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2135 // Get the right opcodes for the displacement.
2136 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2137 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2138 assert(LOpcode && CSOpcode && "Displacement out of range");
2140 // Create virtual registers for temporary results.
2141 unsigned OrigVal = MRI.createVirtualRegister(RC);
2142 unsigned OldVal = MRI.createVirtualRegister(RC);
2143 unsigned NewVal = MRI.createVirtualRegister(RC);
2144 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2145 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
2146 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2148 // Insert 3 basic blocks for the loop.
2149 MachineBasicBlock *StartMBB = MBB;
2150 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
2151 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2152 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
2153 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
2157 // %OrigVal = L Disp(%Base)
2158 // # fall through to LoopMMB
2160 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2161 .addOperand(Base).addImm(Disp).addReg(0);
2162 MBB->addSuccessor(LoopMBB);
2165 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
2166 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2167 // CompareOpcode %RotatedOldVal, %Src2
2168 // BRC KeepOldMask, UpdateMBB
2170 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2171 .addReg(OrigVal).addMBB(StartMBB)
2172 .addReg(Dest).addMBB(UpdateMBB);
2174 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2175 .addReg(OldVal).addReg(BitShift).addImm(0);
2176 BuildMI(MBB, DL, TII->get(CompareOpcode))
2177 .addReg(RotatedOldVal).addReg(Src2);
2178 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2179 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
2180 MBB->addSuccessor(UpdateMBB);
2181 MBB->addSuccessor(UseAltMBB);
2184 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
2185 // # fall through to UpdateMMB
2188 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
2189 .addReg(RotatedOldVal).addReg(Src2)
2190 .addImm(32).addImm(31 + BitSize).addImm(0);
2191 MBB->addSuccessor(UpdateMBB);
2194 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
2195 // [ %RotatedAltVal, UseAltMBB ]
2196 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2197 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2199 // # fall through to DoneMMB
2201 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
2202 .addReg(RotatedOldVal).addMBB(LoopMBB)
2203 .addReg(RotatedAltVal).addMBB(UseAltMBB);
2205 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2206 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2207 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2208 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2209 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2210 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2211 MBB->addSuccessor(LoopMBB);
2212 MBB->addSuccessor(DoneMBB);
2214 MI->eraseFromParent();
2218 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
2221 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
2222 MachineBasicBlock *MBB) const {
2223 const SystemZInstrInfo *TII = TM.getInstrInfo();
2224 MachineFunction &MF = *MBB->getParent();
2225 MachineRegisterInfo &MRI = MF.getRegInfo();
2227 // Extract the operands. Base can be a register or a frame index.
2228 unsigned Dest = MI->getOperand(0).getReg();
2229 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2230 int64_t Disp = MI->getOperand(2).getImm();
2231 unsigned OrigCmpVal = MI->getOperand(3).getReg();
2232 unsigned OrigSwapVal = MI->getOperand(4).getReg();
2233 unsigned BitShift = MI->getOperand(5).getReg();
2234 unsigned NegBitShift = MI->getOperand(6).getReg();
2235 int64_t BitSize = MI->getOperand(7).getImm();
2236 DebugLoc DL = MI->getDebugLoc();
2238 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
2240 // Get the right opcodes for the displacement.
2241 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
2242 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
2243 assert(LOpcode && CSOpcode && "Displacement out of range");
2245 // Create virtual registers for temporary results.
2246 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
2247 unsigned OldVal = MRI.createVirtualRegister(RC);
2248 unsigned CmpVal = MRI.createVirtualRegister(RC);
2249 unsigned SwapVal = MRI.createVirtualRegister(RC);
2250 unsigned StoreVal = MRI.createVirtualRegister(RC);
2251 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
2252 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
2253 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
2255 // Insert 2 basic blocks for the loop.
2256 MachineBasicBlock *StartMBB = MBB;
2257 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
2258 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2259 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
2263 // %OrigOldVal = L Disp(%Base)
2264 // # fall through to LoopMMB
2266 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
2267 .addOperand(Base).addImm(Disp).addReg(0);
2268 MBB->addSuccessor(LoopMBB);
2271 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
2272 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
2273 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
2274 // %Dest = RLL %OldVal, BitSize(%BitShift)
2275 // ^^ The low BitSize bits contain the field
2277 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
2278 // ^^ Replace the upper 32-BitSize bits of the
2279 // comparison value with those that we loaded,
2280 // so that we can use a full word comparison.
2281 // CR %Dest, %RetryCmpVal
2283 // # Fall through to SetMBB
2285 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2286 .addReg(OrigOldVal).addMBB(StartMBB)
2287 .addReg(RetryOldVal).addMBB(SetMBB);
2288 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
2289 .addReg(OrigCmpVal).addMBB(StartMBB)
2290 .addReg(RetryCmpVal).addMBB(SetMBB);
2291 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
2292 .addReg(OrigSwapVal).addMBB(StartMBB)
2293 .addReg(RetrySwapVal).addMBB(SetMBB);
2294 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
2295 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
2296 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
2297 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
2298 BuildMI(MBB, DL, TII->get(SystemZ::CR))
2299 .addReg(Dest).addReg(RetryCmpVal);
2300 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2301 .addImm(SystemZ::CCMASK_ICMP)
2302 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
2303 MBB->addSuccessor(DoneMBB);
2304 MBB->addSuccessor(SetMBB);
2307 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
2308 // ^^ Replace the upper 32-BitSize bits of the new
2309 // value with those that we loaded.
2310 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
2311 // ^^ Rotate the new field to its proper position.
2312 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
2314 // # fall through to ExitMMB
2316 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
2317 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
2318 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
2319 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
2320 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
2321 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
2322 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2323 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2324 MBB->addSuccessor(LoopMBB);
2325 MBB->addSuccessor(DoneMBB);
2327 MI->eraseFromParent();
2331 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
2332 // if the high register of the GR128 value must be cleared or false if
2333 // it's "don't care". SubReg is subreg_odd32 when extending a GR32
2334 // and subreg_odd when extending a GR64.
2336 SystemZTargetLowering::emitExt128(MachineInstr *MI,
2337 MachineBasicBlock *MBB,
2338 bool ClearEven, unsigned SubReg) const {
2339 const SystemZInstrInfo *TII = TM.getInstrInfo();
2340 MachineFunction &MF = *MBB->getParent();
2341 MachineRegisterInfo &MRI = MF.getRegInfo();
2342 DebugLoc DL = MI->getDebugLoc();
2344 unsigned Dest = MI->getOperand(0).getReg();
2345 unsigned Src = MI->getOperand(1).getReg();
2346 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
2348 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
2350 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
2351 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
2353 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
2355 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
2356 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_high);
2359 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
2360 .addReg(In128).addReg(Src).addImm(SubReg);
2362 MI->eraseFromParent();
2367 SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
2368 MachineBasicBlock *MBB,
2369 unsigned Opcode) const {
2370 const SystemZInstrInfo *TII = TM.getInstrInfo();
2371 DebugLoc DL = MI->getDebugLoc();
2373 MachineOperand DestBase = MI->getOperand(0);
2374 uint64_t DestDisp = MI->getOperand(1).getImm();
2375 MachineOperand SrcBase = MI->getOperand(2);
2376 uint64_t SrcDisp = MI->getOperand(3).getImm();
2377 uint64_t Length = MI->getOperand(4).getImm();
2379 BuildMI(*MBB, MI, DL, TII->get(Opcode))
2380 .addOperand(DestBase).addImm(DestDisp).addImm(Length)
2381 .addOperand(SrcBase).addImm(SrcDisp);
2383 MI->eraseFromParent();
2387 // Decompose string pseudo-instruction MI into a loop that continually performs
2388 // Opcode until CC != 3.
2390 SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
2391 MachineBasicBlock *MBB,
2392 unsigned Opcode) const {
2393 const SystemZInstrInfo *TII = TM.getInstrInfo();
2394 MachineFunction &MF = *MBB->getParent();
2395 MachineRegisterInfo &MRI = MF.getRegInfo();
2396 DebugLoc DL = MI->getDebugLoc();
2398 uint64_t End1Reg = MI->getOperand(0).getReg();
2399 uint64_t Start1Reg = MI->getOperand(1).getReg();
2400 uint64_t Start2Reg = MI->getOperand(2).getReg();
2401 uint64_t CharReg = MI->getOperand(3).getReg();
2403 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass;
2404 uint64_t This1Reg = MRI.createVirtualRegister(RC);
2405 uint64_t This2Reg = MRI.createVirtualRegister(RC);
2406 uint64_t End2Reg = MRI.createVirtualRegister(RC);
2408 MachineBasicBlock *StartMBB = MBB;
2409 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
2410 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2413 // # fall through to LoopMMB
2414 MBB->addSuccessor(LoopMBB);
2417 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ]
2418 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ]
2420 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0W
2422 // # fall through to DoneMMB
2424 // The load of R0W can be hoisted by post-RA LICM.
2427 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg)
2428 .addReg(Start1Reg).addMBB(StartMBB)
2429 .addReg(End1Reg).addMBB(LoopMBB);
2430 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg)
2431 .addReg(Start2Reg).addMBB(StartMBB)
2432 .addReg(End2Reg).addMBB(LoopMBB);
2433 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0W).addReg(CharReg);
2434 BuildMI(MBB, DL, TII->get(Opcode))
2435 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define)
2436 .addReg(This1Reg).addReg(This2Reg);
2437 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2438 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB);
2439 MBB->addSuccessor(LoopMBB);
2440 MBB->addSuccessor(DoneMBB);
2442 DoneMBB->addLiveIn(SystemZ::CC);
2444 MI->eraseFromParent();
2448 MachineBasicBlock *SystemZTargetLowering::
2449 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
2450 switch (MI->getOpcode()) {
2451 case SystemZ::Select32:
2452 case SystemZ::SelectF32:
2453 case SystemZ::Select64:
2454 case SystemZ::SelectF64:
2455 case SystemZ::SelectF128:
2456 return emitSelect(MI, MBB);
2458 case SystemZ::CondStore8_32:
2459 return emitCondStore(MI, MBB, SystemZ::STC32, 0, false);
2460 case SystemZ::CondStore8_32Inv:
2461 return emitCondStore(MI, MBB, SystemZ::STC32, 0, true);
2462 case SystemZ::CondStore16_32:
2463 return emitCondStore(MI, MBB, SystemZ::STH32, 0, false);
2464 case SystemZ::CondStore16_32Inv:
2465 return emitCondStore(MI, MBB, SystemZ::STH32, 0, true);
2466 case SystemZ::CondStore32_32:
2467 return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, false);
2468 case SystemZ::CondStore32_32Inv:
2469 return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, true);
2470 case SystemZ::CondStore8:
2471 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
2472 case SystemZ::CondStore8Inv:
2473 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
2474 case SystemZ::CondStore16:
2475 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
2476 case SystemZ::CondStore16Inv:
2477 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
2478 case SystemZ::CondStore32:
2479 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
2480 case SystemZ::CondStore32Inv:
2481 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
2482 case SystemZ::CondStore64:
2483 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
2484 case SystemZ::CondStore64Inv:
2485 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
2486 case SystemZ::CondStoreF32:
2487 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
2488 case SystemZ::CondStoreF32Inv:
2489 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
2490 case SystemZ::CondStoreF64:
2491 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
2492 case SystemZ::CondStoreF64Inv:
2493 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
2495 case SystemZ::AEXT128_64:
2496 return emitExt128(MI, MBB, false, SystemZ::subreg_low);
2497 case SystemZ::ZEXT128_32:
2498 return emitExt128(MI, MBB, true, SystemZ::subreg_low32);
2499 case SystemZ::ZEXT128_64:
2500 return emitExt128(MI, MBB, true, SystemZ::subreg_low);
2502 case SystemZ::ATOMIC_SWAPW:
2503 return emitAtomicLoadBinary(MI, MBB, 0, 0);
2504 case SystemZ::ATOMIC_SWAP_32:
2505 return emitAtomicLoadBinary(MI, MBB, 0, 32);
2506 case SystemZ::ATOMIC_SWAP_64:
2507 return emitAtomicLoadBinary(MI, MBB, 0, 64);
2509 case SystemZ::ATOMIC_LOADW_AR:
2510 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
2511 case SystemZ::ATOMIC_LOADW_AFI:
2512 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
2513 case SystemZ::ATOMIC_LOAD_AR:
2514 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
2515 case SystemZ::ATOMIC_LOAD_AHI:
2516 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
2517 case SystemZ::ATOMIC_LOAD_AFI:
2518 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
2519 case SystemZ::ATOMIC_LOAD_AGR:
2520 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
2521 case SystemZ::ATOMIC_LOAD_AGHI:
2522 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
2523 case SystemZ::ATOMIC_LOAD_AGFI:
2524 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
2526 case SystemZ::ATOMIC_LOADW_SR:
2527 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
2528 case SystemZ::ATOMIC_LOAD_SR:
2529 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
2530 case SystemZ::ATOMIC_LOAD_SGR:
2531 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
2533 case SystemZ::ATOMIC_LOADW_NR:
2534 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
2535 case SystemZ::ATOMIC_LOADW_NILH:
2536 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 0);
2537 case SystemZ::ATOMIC_LOAD_NR:
2538 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
2539 case SystemZ::ATOMIC_LOAD_NILL32:
2540 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL32, 32);
2541 case SystemZ::ATOMIC_LOAD_NILH32:
2542 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 32);
2543 case SystemZ::ATOMIC_LOAD_NILF32:
2544 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF32, 32);
2545 case SystemZ::ATOMIC_LOAD_NGR:
2546 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
2547 case SystemZ::ATOMIC_LOAD_NILL:
2548 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 64);
2549 case SystemZ::ATOMIC_LOAD_NILH:
2550 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 64);
2551 case SystemZ::ATOMIC_LOAD_NIHL:
2552 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL, 64);
2553 case SystemZ::ATOMIC_LOAD_NIHH:
2554 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH, 64);
2555 case SystemZ::ATOMIC_LOAD_NILF:
2556 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 64);
2557 case SystemZ::ATOMIC_LOAD_NIHF:
2558 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF, 64);
2560 case SystemZ::ATOMIC_LOADW_OR:
2561 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
2562 case SystemZ::ATOMIC_LOADW_OILH:
2563 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH32, 0);
2564 case SystemZ::ATOMIC_LOAD_OR:
2565 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
2566 case SystemZ::ATOMIC_LOAD_OILL32:
2567 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL32, 32);
2568 case SystemZ::ATOMIC_LOAD_OILH32:
2569 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH32, 32);
2570 case SystemZ::ATOMIC_LOAD_OILF32:
2571 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF32, 32);
2572 case SystemZ::ATOMIC_LOAD_OGR:
2573 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
2574 case SystemZ::ATOMIC_LOAD_OILL:
2575 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 64);
2576 case SystemZ::ATOMIC_LOAD_OILH:
2577 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 64);
2578 case SystemZ::ATOMIC_LOAD_OIHL:
2579 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL, 64);
2580 case SystemZ::ATOMIC_LOAD_OIHH:
2581 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH, 64);
2582 case SystemZ::ATOMIC_LOAD_OILF:
2583 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 64);
2584 case SystemZ::ATOMIC_LOAD_OIHF:
2585 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF, 64);
2587 case SystemZ::ATOMIC_LOADW_XR:
2588 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
2589 case SystemZ::ATOMIC_LOADW_XILF:
2590 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF32, 0);
2591 case SystemZ::ATOMIC_LOAD_XR:
2592 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
2593 case SystemZ::ATOMIC_LOAD_XILF32:
2594 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF32, 32);
2595 case SystemZ::ATOMIC_LOAD_XGR:
2596 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
2597 case SystemZ::ATOMIC_LOAD_XILF:
2598 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 64);
2599 case SystemZ::ATOMIC_LOAD_XIHF:
2600 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF, 64);
2602 case SystemZ::ATOMIC_LOADW_NRi:
2603 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
2604 case SystemZ::ATOMIC_LOADW_NILHi:
2605 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 0, true);
2606 case SystemZ::ATOMIC_LOAD_NRi:
2607 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
2608 case SystemZ::ATOMIC_LOAD_NILL32i:
2609 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL32, 32, true);
2610 case SystemZ::ATOMIC_LOAD_NILH32i:
2611 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 32, true);
2612 case SystemZ::ATOMIC_LOAD_NILF32i:
2613 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF32, 32, true);
2614 case SystemZ::ATOMIC_LOAD_NGRi:
2615 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
2616 case SystemZ::ATOMIC_LOAD_NILLi:
2617 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 64, true);
2618 case SystemZ::ATOMIC_LOAD_NILHi:
2619 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 64, true);
2620 case SystemZ::ATOMIC_LOAD_NIHLi:
2621 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL, 64, true);
2622 case SystemZ::ATOMIC_LOAD_NIHHi:
2623 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH, 64, true);
2624 case SystemZ::ATOMIC_LOAD_NILFi:
2625 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 64, true);
2626 case SystemZ::ATOMIC_LOAD_NIHFi:
2627 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF, 64, true);
2629 case SystemZ::ATOMIC_LOADW_MIN:
2630 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2631 SystemZ::CCMASK_CMP_LE, 0);
2632 case SystemZ::ATOMIC_LOAD_MIN_32:
2633 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2634 SystemZ::CCMASK_CMP_LE, 32);
2635 case SystemZ::ATOMIC_LOAD_MIN_64:
2636 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
2637 SystemZ::CCMASK_CMP_LE, 64);
2639 case SystemZ::ATOMIC_LOADW_MAX:
2640 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2641 SystemZ::CCMASK_CMP_GE, 0);
2642 case SystemZ::ATOMIC_LOAD_MAX_32:
2643 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2644 SystemZ::CCMASK_CMP_GE, 32);
2645 case SystemZ::ATOMIC_LOAD_MAX_64:
2646 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
2647 SystemZ::CCMASK_CMP_GE, 64);
2649 case SystemZ::ATOMIC_LOADW_UMIN:
2650 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2651 SystemZ::CCMASK_CMP_LE, 0);
2652 case SystemZ::ATOMIC_LOAD_UMIN_32:
2653 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2654 SystemZ::CCMASK_CMP_LE, 32);
2655 case SystemZ::ATOMIC_LOAD_UMIN_64:
2656 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
2657 SystemZ::CCMASK_CMP_LE, 64);
2659 case SystemZ::ATOMIC_LOADW_UMAX:
2660 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2661 SystemZ::CCMASK_CMP_GE, 0);
2662 case SystemZ::ATOMIC_LOAD_UMAX_32:
2663 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2664 SystemZ::CCMASK_CMP_GE, 32);
2665 case SystemZ::ATOMIC_LOAD_UMAX_64:
2666 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
2667 SystemZ::CCMASK_CMP_GE, 64);
2669 case SystemZ::ATOMIC_CMP_SWAPW:
2670 return emitAtomicCmpSwapW(MI, MBB);
2671 case SystemZ::MVCWrapper:
2672 return emitMemMemWrapper(MI, MBB, SystemZ::MVC);
2673 case SystemZ::CLCWrapper:
2674 return emitMemMemWrapper(MI, MBB, SystemZ::CLC);
2675 case SystemZ::CLSTLoop:
2676 return emitStringWrapper(MI, MBB, SystemZ::CLST);
2677 case SystemZ::MVSTLoop:
2678 return emitStringWrapper(MI, MBB, SystemZ::MVST);
2679 case SystemZ::SRSTLoop:
2680 return emitStringWrapper(MI, MBB, SystemZ::SRST);
2682 llvm_unreachable("Unexpected instr type to insert");