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 and a wider multiplication otherwise.
132 setOperationAction(ISD::MULHS, VT, Expand);
133 setOperationAction(ISD::MULHU, VT, Expand);
135 // We have instructions for signed but not unsigned FP conversion.
136 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
140 // Type legalization will convert 8- and 16-bit atomic operations into
141 // forms that operate on i32s (but still keeping the original memory VT).
142 // Lower them into full i32 operations.
143 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
144 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
145 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
146 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
147 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
148 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
149 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
150 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
151 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
152 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
153 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
154 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
156 // We have instructions for signed but not unsigned FP conversion.
157 // Handle unsigned 32-bit types as signed 64-bit types.
158 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
159 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
161 // We have native support for a 64-bit CTLZ, via FLOGR.
162 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
163 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
165 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
166 setOperationAction(ISD::OR, MVT::i64, Custom);
168 // The architecture has 32-bit SMUL_LOHI and UMUL_LOHI (MR and MLR),
169 // but they aren't really worth using. There is no 64-bit SMUL_LOHI,
170 // but there is a 64-bit UMUL_LOHI: MLGR.
171 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
172 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
173 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
174 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Custom);
176 // FIXME: Can we support these natively?
177 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
178 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
179 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
181 // We have native instructions for i8, i16 and i32 extensions, but not i1.
182 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
183 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
184 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
185 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
187 // Handle the various types of symbolic address.
188 setOperationAction(ISD::ConstantPool, PtrVT, Custom);
189 setOperationAction(ISD::GlobalAddress, PtrVT, Custom);
190 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
191 setOperationAction(ISD::BlockAddress, PtrVT, Custom);
192 setOperationAction(ISD::JumpTable, PtrVT, Custom);
194 // We need to handle dynamic allocations specially because of the
195 // 160-byte area at the bottom of the stack.
196 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
198 // Use custom expanders so that we can force the function to use
200 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
201 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
203 // Handle floating-point types.
204 for (unsigned I = MVT::FIRST_FP_VALUETYPE;
205 I <= MVT::LAST_FP_VALUETYPE;
207 MVT VT = MVT::SimpleValueType(I);
208 if (isTypeLegal(VT)) {
209 // We can use FI for FRINT.
210 setOperationAction(ISD::FRINT, VT, Legal);
212 // No special instructions for these.
213 setOperationAction(ISD::FSIN, VT, Expand);
214 setOperationAction(ISD::FCOS, VT, Expand);
215 setOperationAction(ISD::FREM, VT, Expand);
219 // We have fused multiply-addition for f32 and f64 but not f128.
220 setOperationAction(ISD::FMA, MVT::f32, Legal);
221 setOperationAction(ISD::FMA, MVT::f64, Legal);
222 setOperationAction(ISD::FMA, MVT::f128, Expand);
224 // Needed so that we don't try to implement f128 constant loads using
225 // a load-and-extend of a f80 constant (in cases where the constant
226 // would fit in an f80).
227 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
229 // Floating-point truncation and stores need to be done separately.
230 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
231 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
232 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
234 // We have 64-bit FPR<->GPR moves, but need special handling for
236 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
237 setOperationAction(ISD::BITCAST, MVT::f32, Custom);
239 // VASTART and VACOPY need to deal with the SystemZ-specific varargs
240 // structure, but VAEND is a no-op.
241 setOperationAction(ISD::VASTART, MVT::Other, Custom);
242 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
243 setOperationAction(ISD::VAEND, MVT::Other, Expand);
245 // We want to use MVC in preference to even a single load/store pair.
246 MaxStoresPerMemcpy = 0;
247 MaxStoresPerMemcpyOptSize = 0;
249 // The main memset sequence is a byte store followed by an MVC.
250 // Two STC or MV..I stores win over that, but the kind of fused stores
251 // generated by target-independent code don't when the byte value is
252 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
253 // than "STC;MVC". Handle the choice in target-specific code instead.
254 MaxStoresPerMemset = 0;
255 MaxStoresPerMemsetOptSize = 0;
259 SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
260 VT = VT.getScalarType();
265 switch (VT.getSimpleVT().SimpleTy) {
278 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
279 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
280 return Imm.isZero() || Imm.isNegZero();
283 bool SystemZTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
285 // Unaligned accesses should never be slower than the expanded version.
286 // We check specifically for aligned accesses in the few cases where
287 // they are required.
293 //===----------------------------------------------------------------------===//
294 // Inline asm support
295 //===----------------------------------------------------------------------===//
297 TargetLowering::ConstraintType
298 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
299 if (Constraint.size() == 1) {
300 switch (Constraint[0]) {
301 case 'a': // Address register
302 case 'd': // Data register (equivalent to 'r')
303 case 'f': // Floating-point register
304 case 'r': // General-purpose register
305 return C_RegisterClass;
307 case 'Q': // Memory with base and unsigned 12-bit displacement
308 case 'R': // Likewise, plus an index
309 case 'S': // Memory with base and signed 20-bit displacement
310 case 'T': // Likewise, plus an index
311 case 'm': // Equivalent to 'T'.
314 case 'I': // Unsigned 8-bit constant
315 case 'J': // Unsigned 12-bit constant
316 case 'K': // Signed 16-bit constant
317 case 'L': // Signed 20-bit displacement (on all targets we support)
318 case 'M': // 0x7fffffff
325 return TargetLowering::getConstraintType(Constraint);
328 TargetLowering::ConstraintWeight SystemZTargetLowering::
329 getSingleConstraintMatchWeight(AsmOperandInfo &info,
330 const char *constraint) const {
331 ConstraintWeight weight = CW_Invalid;
332 Value *CallOperandVal = info.CallOperandVal;
333 // If we don't have a value, we can't do a match,
334 // but allow it at the lowest weight.
335 if (CallOperandVal == NULL)
337 Type *type = CallOperandVal->getType();
338 // Look at the constraint type.
339 switch (*constraint) {
341 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
344 case 'a': // Address register
345 case 'd': // Data register (equivalent to 'r')
346 case 'r': // General-purpose register
347 if (CallOperandVal->getType()->isIntegerTy())
348 weight = CW_Register;
351 case 'f': // Floating-point register
352 if (type->isFloatingPointTy())
353 weight = CW_Register;
356 case 'I': // Unsigned 8-bit constant
357 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
358 if (isUInt<8>(C->getZExtValue()))
359 weight = CW_Constant;
362 case 'J': // Unsigned 12-bit constant
363 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
364 if (isUInt<12>(C->getZExtValue()))
365 weight = CW_Constant;
368 case 'K': // Signed 16-bit constant
369 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
370 if (isInt<16>(C->getSExtValue()))
371 weight = CW_Constant;
374 case 'L': // Signed 20-bit displacement (on all targets we support)
375 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
376 if (isInt<20>(C->getSExtValue()))
377 weight = CW_Constant;
380 case 'M': // 0x7fffffff
381 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
382 if (C->getZExtValue() == 0x7fffffff)
383 weight = CW_Constant;
389 // Parse a "{tNNN}" register constraint for which the register type "t"
390 // has already been verified. MC is the class associated with "t" and
391 // Map maps 0-based register numbers to LLVM register numbers.
392 static std::pair<unsigned, const TargetRegisterClass *>
393 parseRegisterNumber(const std::string &Constraint,
394 const TargetRegisterClass *RC, const unsigned *Map) {
395 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
396 if (isdigit(Constraint[2])) {
397 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2);
398 unsigned Index = atoi(Suffix.c_str());
399 if (Index < 16 && Map[Index])
400 return std::make_pair(Map[Index], RC);
402 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
405 std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering::
406 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const {
407 if (Constraint.size() == 1) {
408 // GCC Constraint Letters
409 switch (Constraint[0]) {
411 case 'd': // Data register (equivalent to 'r')
412 case 'r': // General-purpose register
414 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
415 else if (VT == MVT::i128)
416 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
417 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
419 case 'a': // Address register
421 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
422 else if (VT == MVT::i128)
423 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
424 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
426 case 'f': // Floating-point register
428 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
429 else if (VT == MVT::f128)
430 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
431 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
434 if (Constraint[0] == '{') {
435 // We need to override the default register parsing for GPRs and FPRs
436 // because the interpretation depends on VT. The internal names of
437 // the registers are also different from the external names
438 // (F0D and F0S instead of F0, etc.).
439 if (Constraint[1] == 'r') {
441 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
442 SystemZMC::GR32Regs);
444 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
445 SystemZMC::GR128Regs);
446 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
447 SystemZMC::GR64Regs);
449 if (Constraint[1] == 'f') {
451 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
452 SystemZMC::FP32Regs);
454 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
455 SystemZMC::FP128Regs);
456 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
457 SystemZMC::FP64Regs);
460 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
463 void SystemZTargetLowering::
464 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
465 std::vector<SDValue> &Ops,
466 SelectionDAG &DAG) const {
467 // Only support length 1 constraints for now.
468 if (Constraint.length() == 1) {
469 switch (Constraint[0]) {
470 case 'I': // Unsigned 8-bit constant
471 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
472 if (isUInt<8>(C->getZExtValue()))
473 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
477 case 'J': // Unsigned 12-bit constant
478 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
479 if (isUInt<12>(C->getZExtValue()))
480 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
484 case 'K': // Signed 16-bit constant
485 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
486 if (isInt<16>(C->getSExtValue()))
487 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
491 case 'L': // Signed 20-bit displacement (on all targets we support)
492 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
493 if (isInt<20>(C->getSExtValue()))
494 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
498 case 'M': // 0x7fffffff
499 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
500 if (C->getZExtValue() == 0x7fffffff)
501 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
506 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
509 //===----------------------------------------------------------------------===//
510 // Calling conventions
511 //===----------------------------------------------------------------------===//
513 #include "SystemZGenCallingConv.inc"
515 // Value is a value that has been passed to us in the location described by VA
516 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
517 // any loads onto Chain.
518 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
519 CCValAssign &VA, SDValue Chain,
521 // If the argument has been promoted from a smaller type, insert an
522 // assertion to capture this.
523 if (VA.getLocInfo() == CCValAssign::SExt)
524 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
525 DAG.getValueType(VA.getValVT()));
526 else if (VA.getLocInfo() == CCValAssign::ZExt)
527 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
528 DAG.getValueType(VA.getValVT()));
531 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
532 else if (VA.getLocInfo() == CCValAssign::Indirect)
533 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
534 MachinePointerInfo(), false, false, false, 0);
536 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
540 // Value is a value of type VA.getValVT() that we need to copy into
541 // the location described by VA. Return a copy of Value converted to
542 // VA.getValVT(). The caller is responsible for handling indirect values.
543 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
544 CCValAssign &VA, SDValue Value) {
545 switch (VA.getLocInfo()) {
546 case CCValAssign::SExt:
547 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
548 case CCValAssign::ZExt:
549 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
550 case CCValAssign::AExt:
551 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
552 case CCValAssign::Full:
555 llvm_unreachable("Unhandled getLocInfo()");
559 SDValue SystemZTargetLowering::
560 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
561 const SmallVectorImpl<ISD::InputArg> &Ins,
562 SDLoc DL, SelectionDAG &DAG,
563 SmallVectorImpl<SDValue> &InVals) const {
564 MachineFunction &MF = DAG.getMachineFunction();
565 MachineFrameInfo *MFI = MF.getFrameInfo();
566 MachineRegisterInfo &MRI = MF.getRegInfo();
567 SystemZMachineFunctionInfo *FuncInfo =
568 MF.getInfo<SystemZMachineFunctionInfo>();
569 const SystemZFrameLowering *TFL =
570 static_cast<const SystemZFrameLowering *>(TM.getFrameLowering());
572 // Assign locations to all of the incoming arguments.
573 SmallVector<CCValAssign, 16> ArgLocs;
574 CCState CCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
575 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
577 unsigned NumFixedGPRs = 0;
578 unsigned NumFixedFPRs = 0;
579 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
581 CCValAssign &VA = ArgLocs[I];
582 EVT LocVT = VA.getLocVT();
584 // Arguments passed in registers
585 const TargetRegisterClass *RC;
586 switch (LocVT.getSimpleVT().SimpleTy) {
588 // Integers smaller than i64 should be promoted to i64.
589 llvm_unreachable("Unexpected argument type");
592 RC = &SystemZ::GR32BitRegClass;
596 RC = &SystemZ::GR64BitRegClass;
600 RC = &SystemZ::FP32BitRegClass;
604 RC = &SystemZ::FP64BitRegClass;
608 unsigned VReg = MRI.createVirtualRegister(RC);
609 MRI.addLiveIn(VA.getLocReg(), VReg);
610 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
612 assert(VA.isMemLoc() && "Argument not register or memory");
614 // Create the frame index object for this incoming parameter.
615 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
616 VA.getLocMemOffset(), true);
618 // Create the SelectionDAG nodes corresponding to a load
619 // from this parameter. Unpromoted ints and floats are
620 // passed as right-justified 8-byte values.
621 EVT PtrVT = getPointerTy();
622 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
623 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
624 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4));
625 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
626 MachinePointerInfo::getFixedStack(FI),
627 false, false, false, 0);
630 // Convert the value of the argument register into the value that's
632 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
636 // Save the number of non-varargs registers for later use by va_start, etc.
637 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
638 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
640 // Likewise the address (in the form of a frame index) of where the
641 // first stack vararg would be. The 1-byte size here is arbitrary.
642 int64_t StackSize = CCInfo.getNextStackOffset();
643 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
645 // ...and a similar frame index for the caller-allocated save area
646 // that will be used to store the incoming registers.
647 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
648 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
649 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
651 // Store the FPR varargs in the reserved frame slots. (We store the
652 // GPRs as part of the prologue.)
653 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
654 SDValue MemOps[SystemZ::NumArgFPRs];
655 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
656 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
657 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
658 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
659 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
660 &SystemZ::FP64BitRegClass);
661 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
662 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
663 MachinePointerInfo::getFixedStack(FI),
667 // Join the stores, which are independent of one another.
668 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
669 &MemOps[NumFixedFPRs],
670 SystemZ::NumArgFPRs - NumFixedFPRs);
678 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
679 SmallVectorImpl<SDValue> &InVals) const {
680 SelectionDAG &DAG = CLI.DAG;
682 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
683 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
684 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
685 SDValue Chain = CLI.Chain;
686 SDValue Callee = CLI.Callee;
687 bool &isTailCall = CLI.IsTailCall;
688 CallingConv::ID CallConv = CLI.CallConv;
689 bool IsVarArg = CLI.IsVarArg;
690 MachineFunction &MF = DAG.getMachineFunction();
691 EVT PtrVT = getPointerTy();
693 // SystemZ target does not yet support tail call optimization.
696 // Analyze the operands of the call, assigning locations to each operand.
697 SmallVector<CCValAssign, 16> ArgLocs;
698 CCState ArgCCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
699 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
701 // Get a count of how many bytes are to be pushed on the stack.
702 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
704 // Mark the start of the call.
705 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true),
708 // Copy argument values to their designated locations.
709 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
710 SmallVector<SDValue, 8> MemOpChains;
712 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
713 CCValAssign &VA = ArgLocs[I];
714 SDValue ArgValue = OutVals[I];
716 if (VA.getLocInfo() == CCValAssign::Indirect) {
717 // Store the argument in a stack slot and pass its address.
718 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
719 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
720 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
721 MachinePointerInfo::getFixedStack(FI),
723 ArgValue = SpillSlot;
725 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
728 // Queue up the argument copies and emit them at the end.
729 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
731 assert(VA.isMemLoc() && "Argument not register or memory");
733 // Work out the address of the stack slot. Unpromoted ints and
734 // floats are passed as right-justified 8-byte values.
735 if (!StackPtr.getNode())
736 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
737 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
738 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
740 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
741 DAG.getIntPtrConstant(Offset));
744 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
745 MachinePointerInfo(),
750 // Join the stores, which are independent of one another.
751 if (!MemOpChains.empty())
752 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
753 &MemOpChains[0], MemOpChains.size());
755 // Build a sequence of copy-to-reg nodes, chained and glued together.
757 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
758 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
759 RegsToPass[I].second, Glue);
760 Glue = Chain.getValue(1);
763 // Accept direct calls by converting symbolic call addresses to the
764 // associated Target* opcodes.
765 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
766 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
767 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
768 } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
769 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
770 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
773 // The first call operand is the chain and the second is the target address.
774 SmallVector<SDValue, 8> Ops;
775 Ops.push_back(Chain);
776 Ops.push_back(Callee);
778 // Add argument registers to the end of the list so that they are
779 // known live into the call.
780 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
781 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
782 RegsToPass[I].second.getValueType()));
784 // Glue the call to the argument copies, if any.
789 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
790 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
791 Glue = Chain.getValue(1);
793 // Mark the end of the call, which is glued to the call itself.
794 Chain = DAG.getCALLSEQ_END(Chain,
795 DAG.getConstant(NumBytes, PtrVT, true),
796 DAG.getConstant(0, PtrVT, true),
798 Glue = Chain.getValue(1);
800 // Assign locations to each value returned by this call.
801 SmallVector<CCValAssign, 16> RetLocs;
802 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
803 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
805 // Copy all of the result registers out of their specified physreg.
806 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
807 CCValAssign &VA = RetLocs[I];
809 // Copy the value out, gluing the copy to the end of the call sequence.
810 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
811 VA.getLocVT(), Glue);
812 Chain = RetValue.getValue(1);
813 Glue = RetValue.getValue(2);
815 // Convert the value of the return register into the value that's
817 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
824 SystemZTargetLowering::LowerReturn(SDValue Chain,
825 CallingConv::ID CallConv, bool IsVarArg,
826 const SmallVectorImpl<ISD::OutputArg> &Outs,
827 const SmallVectorImpl<SDValue> &OutVals,
828 SDLoc DL, SelectionDAG &DAG) const {
829 MachineFunction &MF = DAG.getMachineFunction();
831 // Assign locations to each returned value.
832 SmallVector<CCValAssign, 16> RetLocs;
833 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
834 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
836 // Quick exit for void returns
838 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
840 // Copy the result values into the output registers.
842 SmallVector<SDValue, 4> RetOps;
843 RetOps.push_back(Chain);
844 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
845 CCValAssign &VA = RetLocs[I];
846 SDValue RetValue = OutVals[I];
848 // Make the return register live on exit.
849 assert(VA.isRegLoc() && "Can only return in registers!");
851 // Promote the value as required.
852 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
854 // Chain and glue the copies together.
855 unsigned Reg = VA.getLocReg();
856 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
857 Glue = Chain.getValue(1);
858 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
861 // Update chain and glue.
864 RetOps.push_back(Glue);
866 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other,
867 RetOps.data(), RetOps.size());
870 // CC is a comparison that will be implemented using an integer or
871 // floating-point comparison. Return the condition code mask for
872 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
873 // unsigned comparisons and clear for signed ones. In the floating-point
874 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
875 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
877 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
878 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
879 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
883 llvm_unreachable("Invalid integer condition!");
892 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
893 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
898 // If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1
899 // is suitable for CLI(Y), CHHSI or CLHHSI, adjust the operands as necessary.
900 static void adjustSubwordCmp(SelectionDAG &DAG, bool &IsUnsigned,
901 SDValue &CmpOp0, SDValue &CmpOp1,
903 // For us to make any changes, it must a comparison between a single-use
904 // load and a constant.
905 if (!CmpOp0.hasOneUse() ||
906 CmpOp0.getOpcode() != ISD::LOAD ||
907 CmpOp1.getOpcode() != ISD::Constant)
910 // We must have an 8- or 16-bit load.
911 LoadSDNode *Load = cast<LoadSDNode>(CmpOp0);
912 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
913 if (NumBits != 8 && NumBits != 16)
916 // The load must be an extending one and the constant must be within the
917 // range of the unextended value.
918 ConstantSDNode *Constant = cast<ConstantSDNode>(CmpOp1);
919 uint64_t Value = Constant->getZExtValue();
920 uint64_t Mask = (1 << NumBits) - 1;
921 if (Load->getExtensionType() == ISD::SEXTLOAD) {
922 int64_t SignedValue = Constant->getSExtValue();
923 if (uint64_t(SignedValue) + (1ULL << (NumBits - 1)) > Mask)
925 // Unsigned comparison between two sign-extended values is equivalent
926 // to unsigned comparison between two zero-extended values.
929 else if (CCMask == SystemZ::CCMASK_CMP_EQ ||
930 CCMask == SystemZ::CCMASK_CMP_NE)
931 // Any choice of IsUnsigned is OK for equality comparisons.
932 // We could use either CHHSI or CLHHSI for 16-bit comparisons,
933 // but since we use CLHHSI for zero extensions, it seems better
934 // to be consistent and do the same here.
935 Value &= Mask, IsUnsigned = true;
936 else if (NumBits == 8) {
937 // Try to treat the comparison as unsigned, so that we can use CLI.
938 // Adjust CCMask and Value as necessary.
939 if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_LT)
940 // Test whether the high bit of the byte is set.
941 Value = 127, CCMask = SystemZ::CCMASK_CMP_GT, IsUnsigned = true;
942 else if (SignedValue == -1 && CCMask == SystemZ::CCMASK_CMP_GT)
943 // Test whether the high bit of the byte is clear.
944 Value = 128, CCMask = SystemZ::CCMASK_CMP_LT, IsUnsigned = true;
946 // No instruction exists for this combination.
949 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
952 // Signed comparison between two zero-extended values is equivalent
953 // to unsigned comparison.
958 // Make sure that the first operand is an i32 of the right extension type.
959 ISD::LoadExtType ExtType = IsUnsigned ? ISD::ZEXTLOAD : ISD::SEXTLOAD;
960 if (CmpOp0.getValueType() != MVT::i32 ||
961 Load->getExtensionType() != ExtType)
962 CmpOp0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
963 Load->getChain(), Load->getBasePtr(),
964 Load->getPointerInfo(), Load->getMemoryVT(),
965 Load->isVolatile(), Load->isNonTemporal(),
966 Load->getAlignment());
968 // Make sure that the second operand is an i32 with the right value.
969 if (CmpOp1.getValueType() != MVT::i32 ||
970 Value != Constant->getZExtValue())
971 CmpOp1 = DAG.getConstant(Value, MVT::i32);
974 // Return true if a comparison described by CCMask, CmpOp0 and CmpOp1
975 // is an equality comparison that is better implemented using unsigned
976 // rather than signed comparison instructions.
977 static bool preferUnsignedComparison(SelectionDAG &DAG, SDValue CmpOp0,
978 SDValue CmpOp1, unsigned CCMask) {
979 // The test must be for equality or inequality.
980 if (CCMask != SystemZ::CCMASK_CMP_EQ && CCMask != SystemZ::CCMASK_CMP_NE)
983 if (CmpOp1.getOpcode() == ISD::Constant) {
984 uint64_t Value = cast<ConstantSDNode>(CmpOp1)->getSExtValue();
986 // If we're comparing with memory, prefer unsigned comparisons for
987 // values that are in the unsigned 16-bit range but not the signed
988 // 16-bit range. We want to use CLFHSI and CLGHSI.
989 if (CmpOp0.hasOneUse() &&
990 ISD::isNormalLoad(CmpOp0.getNode()) &&
991 (Value >= 32768 && Value < 65536))
994 // Use unsigned comparisons for values that are in the CLGFI range
995 // but not in the CGFI range.
996 if (CmpOp0.getValueType() == MVT::i64 && (Value >> 31) == 1)
1002 // Prefer CL for zero-extended loads.
1003 if (CmpOp1.getOpcode() == ISD::ZERO_EXTEND ||
1004 ISD::isZEXTLoad(CmpOp1.getNode()))
1007 // ...and for "in-register" zero extensions.
1008 if (CmpOp1.getOpcode() == ISD::AND && CmpOp1.getValueType() == MVT::i64) {
1009 SDValue Mask = CmpOp1.getOperand(1);
1010 if (Mask.getOpcode() == ISD::Constant &&
1011 cast<ConstantSDNode>(Mask)->getZExtValue() == 0xffffffff)
1018 // Return a target node that compares CmpOp0 with CmpOp1 and stores a
1019 // 2-bit result in CC. Set CCValid to the CCMASK_* of all possible
1020 // 2-bit results and CCMask to the subset of those results that are
1021 // associated with Cond.
1022 static SDValue emitCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
1023 ISD::CondCode Cond, unsigned &CCValid,
1025 bool IsUnsigned = false;
1026 CCMask = CCMaskForCondCode(Cond);
1027 if (CmpOp0.getValueType().isFloatingPoint())
1028 CCValid = SystemZ::CCMASK_FCMP;
1030 IsUnsigned = CCMask & SystemZ::CCMASK_CMP_UO;
1031 CCValid = SystemZ::CCMASK_ICMP;
1033 adjustSubwordCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask);
1034 if (preferUnsignedComparison(DAG, CmpOp0, CmpOp1, CCMask))
1039 return DAG.getNode((IsUnsigned ? SystemZISD::UCMP : SystemZISD::CMP),
1040 DL, MVT::Glue, CmpOp0, CmpOp1);
1043 // Lower a binary operation that produces two VT results, one in each
1044 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
1045 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
1046 // on the extended Op0 and (unextended) Op1. Store the even register result
1047 // in Even and the odd register result in Odd.
1048 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
1049 unsigned Extend, unsigned Opcode,
1050 SDValue Op0, SDValue Op1,
1051 SDValue &Even, SDValue &Odd) {
1052 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
1053 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
1054 SDValue(In128, 0), Op1);
1055 bool Is32Bit = is32Bit(VT);
1056 SDValue SubReg0 = DAG.getTargetConstant(SystemZ::even128(Is32Bit), VT);
1057 SDValue SubReg1 = DAG.getTargetConstant(SystemZ::odd128(Is32Bit), VT);
1058 SDNode *Reg0 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
1059 VT, Result, SubReg0);
1060 SDNode *Reg1 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
1061 VT, Result, SubReg1);
1062 Even = SDValue(Reg0, 0);
1063 Odd = SDValue(Reg1, 0);
1066 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1067 SDValue Chain = Op.getOperand(0);
1068 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1069 SDValue CmpOp0 = Op.getOperand(2);
1070 SDValue CmpOp1 = Op.getOperand(3);
1071 SDValue Dest = Op.getOperand(4);
1074 unsigned CCValid, CCMask;
1075 SDValue Flags = emitCmp(DAG, CmpOp0, CmpOp1, CC, CCValid, CCMask);
1076 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
1077 Chain, DAG.getConstant(CCValid, MVT::i32),
1078 DAG.getConstant(CCMask, MVT::i32), Dest, Flags);
1081 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
1082 SelectionDAG &DAG) const {
1083 SDValue CmpOp0 = Op.getOperand(0);
1084 SDValue CmpOp1 = Op.getOperand(1);
1085 SDValue TrueOp = Op.getOperand(2);
1086 SDValue FalseOp = Op.getOperand(3);
1087 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1090 unsigned CCValid, CCMask;
1091 SDValue Flags = emitCmp(DAG, CmpOp0, CmpOp1, CC, CCValid, CCMask);
1093 SmallVector<SDValue, 5> Ops;
1094 Ops.push_back(TrueOp);
1095 Ops.push_back(FalseOp);
1096 Ops.push_back(DAG.getConstant(CCValid, MVT::i32));
1097 Ops.push_back(DAG.getConstant(CCMask, MVT::i32));
1098 Ops.push_back(Flags);
1100 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
1101 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size());
1104 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
1105 SelectionDAG &DAG) const {
1107 const GlobalValue *GV = Node->getGlobal();
1108 int64_t Offset = Node->getOffset();
1109 EVT PtrVT = getPointerTy();
1110 Reloc::Model RM = TM.getRelocationModel();
1111 CodeModel::Model CM = TM.getCodeModel();
1114 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
1115 // Make sure that the offset is aligned to a halfword. If it isn't,
1116 // create an "anchor" at the previous 12-bit boundary.
1117 // FIXME check whether there is a better way of handling this.
1119 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT,
1120 Offset & ~uint64_t(0xfff));
1123 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Offset);
1126 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1128 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
1129 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1130 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
1131 MachinePointerInfo::getGOT(), false, false, false, 0);
1134 // If there was a non-zero offset that we didn't fold, create an explicit
1137 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
1138 DAG.getConstant(Offset, PtrVT));
1143 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
1144 SelectionDAG &DAG) const {
1146 const GlobalValue *GV = Node->getGlobal();
1147 EVT PtrVT = getPointerTy();
1148 TLSModel::Model model = TM.getTLSModel(GV);
1150 if (model != TLSModel::LocalExec)
1151 llvm_unreachable("only local-exec TLS mode supported");
1153 // The high part of the thread pointer is in access register 0.
1154 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1155 DAG.getConstant(0, MVT::i32));
1156 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
1158 // The low part of the thread pointer is in access register 1.
1159 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1160 DAG.getConstant(1, MVT::i32));
1161 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
1163 // Merge them into a single 64-bit address.
1164 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
1165 DAG.getConstant(32, PtrVT));
1166 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
1168 // Get the offset of GA from the thread pointer.
1169 SystemZConstantPoolValue *CPV =
1170 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
1172 // Force the offset into the constant pool and load it from there.
1173 SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8);
1174 SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1175 CPAddr, MachinePointerInfo::getConstantPool(),
1176 false, false, false, 0);
1178 // Add the base and offset together.
1179 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
1182 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
1183 SelectionDAG &DAG) const {
1185 const BlockAddress *BA = Node->getBlockAddress();
1186 int64_t Offset = Node->getOffset();
1187 EVT PtrVT = getPointerTy();
1189 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
1190 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1194 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
1195 SelectionDAG &DAG) const {
1197 EVT PtrVT = getPointerTy();
1198 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1200 // Use LARL to load the address of the table.
1201 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1204 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
1205 SelectionDAG &DAG) const {
1207 EVT PtrVT = getPointerTy();
1210 if (CP->isMachineConstantPoolEntry())
1211 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1212 CP->getAlignment());
1214 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1215 CP->getAlignment(), CP->getOffset());
1217 // Use LARL to load the address of the constant pool entry.
1218 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1221 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
1222 SelectionDAG &DAG) const {
1224 SDValue In = Op.getOperand(0);
1225 EVT InVT = In.getValueType();
1226 EVT ResVT = Op.getValueType();
1228 SDValue SubReg32 = DAG.getTargetConstant(SystemZ::subreg_32bit, MVT::i64);
1229 SDValue Shift32 = DAG.getConstant(32, MVT::i64);
1230 if (InVT == MVT::i32 && ResVT == MVT::f32) {
1231 SDValue In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
1232 SDValue Shift = DAG.getNode(ISD::SHL, DL, MVT::i64, In64, Shift32);
1233 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, Shift);
1234 SDNode *Out = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
1235 MVT::f32, Out64, SubReg32);
1236 return SDValue(Out, 0);
1238 if (InVT == MVT::f32 && ResVT == MVT::i32) {
1239 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
1240 SDNode *In64 = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
1241 MVT::f64, SDValue(U64, 0), In, SubReg32);
1242 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, SDValue(In64, 0));
1243 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64, Shift32);
1244 SDValue Out = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
1247 llvm_unreachable("Unexpected bitcast combination");
1250 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
1251 SelectionDAG &DAG) const {
1252 MachineFunction &MF = DAG.getMachineFunction();
1253 SystemZMachineFunctionInfo *FuncInfo =
1254 MF.getInfo<SystemZMachineFunctionInfo>();
1255 EVT PtrVT = getPointerTy();
1257 SDValue Chain = Op.getOperand(0);
1258 SDValue Addr = Op.getOperand(1);
1259 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1262 // The initial values of each field.
1263 const unsigned NumFields = 4;
1264 SDValue Fields[NumFields] = {
1265 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT),
1266 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT),
1267 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
1268 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
1271 // Store each field into its respective slot.
1272 SDValue MemOps[NumFields];
1273 unsigned Offset = 0;
1274 for (unsigned I = 0; I < NumFields; ++I) {
1275 SDValue FieldAddr = Addr;
1277 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
1278 DAG.getIntPtrConstant(Offset));
1279 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
1280 MachinePointerInfo(SV, Offset),
1284 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps, NumFields);
1287 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
1288 SelectionDAG &DAG) const {
1289 SDValue Chain = Op.getOperand(0);
1290 SDValue DstPtr = Op.getOperand(1);
1291 SDValue SrcPtr = Op.getOperand(2);
1292 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
1293 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
1296 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32),
1297 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
1298 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
1301 SDValue SystemZTargetLowering::
1302 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
1303 SDValue Chain = Op.getOperand(0);
1304 SDValue Size = Op.getOperand(1);
1307 unsigned SPReg = getStackPointerRegisterToSaveRestore();
1309 // Get a reference to the stack pointer.
1310 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
1312 // Get the new stack pointer value.
1313 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
1315 // Copy the new stack pointer back.
1316 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
1318 // The allocated data lives above the 160 bytes allocated for the standard
1319 // frame, plus any outgoing stack arguments. We don't know how much that
1320 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
1321 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
1322 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
1324 SDValue Ops[2] = { Result, Chain };
1325 return DAG.getMergeValues(Ops, 2, DL);
1328 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
1329 SelectionDAG &DAG) const {
1330 EVT VT = Op.getValueType();
1332 assert(!is32Bit(VT) && "Only support 64-bit UMUL_LOHI");
1334 // UMUL_LOHI64 returns the low result in the odd register and the high
1335 // result in the even register. UMUL_LOHI is defined to return the
1336 // low half first, so the results are in reverse order.
1338 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
1339 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
1340 return DAG.getMergeValues(Ops, 2, DL);
1343 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
1344 SelectionDAG &DAG) const {
1345 SDValue Op0 = Op.getOperand(0);
1346 SDValue Op1 = Op.getOperand(1);
1347 EVT VT = Op.getValueType();
1351 // We use DSGF for 32-bit division.
1353 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
1354 Opcode = SystemZISD::SDIVREM32;
1355 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
1356 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
1357 Opcode = SystemZISD::SDIVREM32;
1359 Opcode = SystemZISD::SDIVREM64;
1361 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
1362 // input is "don't care". The instruction returns the remainder in
1363 // the even register and the quotient in the odd register.
1365 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
1366 Op0, Op1, Ops[1], Ops[0]);
1367 return DAG.getMergeValues(Ops, 2, DL);
1370 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
1371 SelectionDAG &DAG) const {
1372 EVT VT = Op.getValueType();
1375 // DL(G) uses a double-width dividend, so we need to clear the even
1376 // register in the GR128 input. The instruction returns the remainder
1377 // in the even register and the quotient in the odd register.
1380 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
1381 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
1383 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
1384 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
1385 return DAG.getMergeValues(Ops, 2, DL);
1388 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
1389 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
1391 // Get the known-zero masks for each operand.
1392 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
1393 APInt KnownZero[2], KnownOne[2];
1394 DAG.ComputeMaskedBits(Ops[0], KnownZero[0], KnownOne[0]);
1395 DAG.ComputeMaskedBits(Ops[1], KnownZero[1], KnownOne[1]);
1397 // See if the upper 32 bits of one operand and the lower 32 bits of the
1398 // other are known zero. They are the low and high operands respectively.
1399 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
1400 KnownZero[1].getZExtValue() };
1402 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
1404 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
1409 SDValue LowOp = Ops[Low];
1410 SDValue HighOp = Ops[High];
1412 // If the high part is a constant, we're better off using IILH.
1413 if (HighOp.getOpcode() == ISD::Constant)
1416 // If the low part is a constant that is outside the range of LHI,
1417 // then we're better off using IILF.
1418 if (LowOp.getOpcode() == ISD::Constant) {
1419 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
1420 if (!isInt<16>(Value))
1424 // Check whether the high part is an AND that doesn't change the
1425 // high 32 bits and just masks out low bits. We can skip it if so.
1426 if (HighOp.getOpcode() == ISD::AND &&
1427 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
1428 ConstantSDNode *MaskNode = cast<ConstantSDNode>(HighOp.getOperand(1));
1429 uint64_t Mask = MaskNode->getZExtValue() | Masks[High];
1430 if ((Mask >> 32) == 0xffffffff)
1431 HighOp = HighOp.getOperand(0);
1434 // Take advantage of the fact that all GR32 operations only change the
1435 // low 32 bits by truncating Low to an i32 and inserting it directly
1436 // using a subreg. The interesting cases are those where the truncation
1439 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
1440 SDValue SubReg32 = DAG.getTargetConstant(SystemZ::subreg_32bit, MVT::i64);
1441 SDNode *Result = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
1442 MVT::i64, HighOp, Low32, SubReg32);
1443 return SDValue(Result, 0);
1446 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
1447 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
1448 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
1450 unsigned Opcode) const {
1451 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
1453 // 32-bit operations need no code outside the main loop.
1454 EVT NarrowVT = Node->getMemoryVT();
1455 EVT WideVT = MVT::i32;
1456 if (NarrowVT == WideVT)
1459 int64_t BitSize = NarrowVT.getSizeInBits();
1460 SDValue ChainIn = Node->getChain();
1461 SDValue Addr = Node->getBasePtr();
1462 SDValue Src2 = Node->getVal();
1463 MachineMemOperand *MMO = Node->getMemOperand();
1465 EVT PtrVT = Addr.getValueType();
1467 // Convert atomic subtracts of constants into additions.
1468 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
1469 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Src2)) {
1470 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
1471 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
1474 // Get the address of the containing word.
1475 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
1476 DAG.getConstant(-4, PtrVT));
1478 // Get the number of bits that the word must be rotated left in order
1479 // to bring the field to the top bits of a GR32.
1480 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
1481 DAG.getConstant(3, PtrVT));
1482 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
1484 // Get the complementing shift amount, for rotating a field in the top
1485 // bits back to its proper position.
1486 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
1487 DAG.getConstant(0, WideVT), BitShift);
1489 // Extend the source operand to 32 bits and prepare it for the inner loop.
1490 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
1491 // operations require the source to be shifted in advance. (This shift
1492 // can be folded if the source is constant.) For AND and NAND, the lower
1493 // bits must be set, while for other opcodes they should be left clear.
1494 if (Opcode != SystemZISD::ATOMIC_SWAPW)
1495 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
1496 DAG.getConstant(32 - BitSize, WideVT));
1497 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
1498 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
1499 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
1500 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT));
1502 // Construct the ATOMIC_LOADW_* node.
1503 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
1504 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
1505 DAG.getConstant(BitSize, WideVT) };
1506 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
1507 array_lengthof(Ops),
1510 // Rotate the result of the final CS so that the field is in the lower
1511 // bits of a GR32, then truncate it.
1512 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
1513 DAG.getConstant(BitSize, WideVT));
1514 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
1516 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
1517 return DAG.getMergeValues(RetOps, 2, DL);
1520 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
1521 // into a fullword ATOMIC_CMP_SWAPW operation.
1522 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
1523 SelectionDAG &DAG) const {
1524 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
1526 // We have native support for 32-bit compare and swap.
1527 EVT NarrowVT = Node->getMemoryVT();
1528 EVT WideVT = MVT::i32;
1529 if (NarrowVT == WideVT)
1532 int64_t BitSize = NarrowVT.getSizeInBits();
1533 SDValue ChainIn = Node->getOperand(0);
1534 SDValue Addr = Node->getOperand(1);
1535 SDValue CmpVal = Node->getOperand(2);
1536 SDValue SwapVal = Node->getOperand(3);
1537 MachineMemOperand *MMO = Node->getMemOperand();
1539 EVT PtrVT = Addr.getValueType();
1541 // Get the address of the containing word.
1542 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
1543 DAG.getConstant(-4, PtrVT));
1545 // Get the number of bits that the word must be rotated left in order
1546 // to bring the field to the top bits of a GR32.
1547 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
1548 DAG.getConstant(3, PtrVT));
1549 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
1551 // Get the complementing shift amount, for rotating a field in the top
1552 // bits back to its proper position.
1553 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
1554 DAG.getConstant(0, WideVT), BitShift);
1556 // Construct the ATOMIC_CMP_SWAPW node.
1557 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
1558 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
1559 NegBitShift, DAG.getConstant(BitSize, WideVT) };
1560 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
1561 VTList, Ops, array_lengthof(Ops),
1566 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
1567 SelectionDAG &DAG) const {
1568 MachineFunction &MF = DAG.getMachineFunction();
1569 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
1570 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
1571 SystemZ::R15D, Op.getValueType());
1574 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
1575 SelectionDAG &DAG) const {
1576 MachineFunction &MF = DAG.getMachineFunction();
1577 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
1578 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
1579 SystemZ::R15D, Op.getOperand(1));
1582 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
1583 SelectionDAG &DAG) const {
1584 switch (Op.getOpcode()) {
1586 return lowerBR_CC(Op, DAG);
1587 case ISD::SELECT_CC:
1588 return lowerSELECT_CC(Op, DAG);
1589 case ISD::GlobalAddress:
1590 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
1591 case ISD::GlobalTLSAddress:
1592 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
1593 case ISD::BlockAddress:
1594 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
1595 case ISD::JumpTable:
1596 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
1597 case ISD::ConstantPool:
1598 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
1600 return lowerBITCAST(Op, DAG);
1602 return lowerVASTART(Op, DAG);
1604 return lowerVACOPY(Op, DAG);
1605 case ISD::DYNAMIC_STACKALLOC:
1606 return lowerDYNAMIC_STACKALLOC(Op, DAG);
1607 case ISD::UMUL_LOHI:
1608 return lowerUMUL_LOHI(Op, DAG);
1610 return lowerSDIVREM(Op, DAG);
1612 return lowerUDIVREM(Op, DAG);
1614 return lowerOR(Op, DAG);
1615 case ISD::ATOMIC_SWAP:
1616 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_SWAPW);
1617 case ISD::ATOMIC_LOAD_ADD:
1618 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
1619 case ISD::ATOMIC_LOAD_SUB:
1620 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
1621 case ISD::ATOMIC_LOAD_AND:
1622 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
1623 case ISD::ATOMIC_LOAD_OR:
1624 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
1625 case ISD::ATOMIC_LOAD_XOR:
1626 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
1627 case ISD::ATOMIC_LOAD_NAND:
1628 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
1629 case ISD::ATOMIC_LOAD_MIN:
1630 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
1631 case ISD::ATOMIC_LOAD_MAX:
1632 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
1633 case ISD::ATOMIC_LOAD_UMIN:
1634 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
1635 case ISD::ATOMIC_LOAD_UMAX:
1636 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
1637 case ISD::ATOMIC_CMP_SWAP:
1638 return lowerATOMIC_CMP_SWAP(Op, DAG);
1639 case ISD::STACKSAVE:
1640 return lowerSTACKSAVE(Op, DAG);
1641 case ISD::STACKRESTORE:
1642 return lowerSTACKRESTORE(Op, DAG);
1644 llvm_unreachable("Unexpected node to lower");
1648 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
1649 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
1653 OPCODE(PCREL_WRAPPER);
1657 OPCODE(SELECT_CCMASK);
1658 OPCODE(ADJDYNALLOC);
1659 OPCODE(EXTRACT_ACCESS);
1660 OPCODE(UMUL_LOHI64);
1665 OPCODE(ATOMIC_SWAPW);
1666 OPCODE(ATOMIC_LOADW_ADD);
1667 OPCODE(ATOMIC_LOADW_SUB);
1668 OPCODE(ATOMIC_LOADW_AND);
1669 OPCODE(ATOMIC_LOADW_OR);
1670 OPCODE(ATOMIC_LOADW_XOR);
1671 OPCODE(ATOMIC_LOADW_NAND);
1672 OPCODE(ATOMIC_LOADW_MIN);
1673 OPCODE(ATOMIC_LOADW_MAX);
1674 OPCODE(ATOMIC_LOADW_UMIN);
1675 OPCODE(ATOMIC_LOADW_UMAX);
1676 OPCODE(ATOMIC_CMP_SWAPW);
1682 //===----------------------------------------------------------------------===//
1684 //===----------------------------------------------------------------------===//
1686 // Create a new basic block after MBB.
1687 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
1688 MachineFunction &MF = *MBB->getParent();
1689 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
1690 MF.insert(llvm::next(MachineFunction::iterator(MBB)), NewMBB);
1694 // Split MBB after MI and return the new block (the one that contains
1695 // instructions after MI).
1696 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
1697 MachineBasicBlock *MBB) {
1698 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
1699 NewMBB->splice(NewMBB->begin(), MBB,
1700 llvm::next(MachineBasicBlock::iterator(MI)),
1702 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
1706 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
1708 SystemZTargetLowering::emitSelect(MachineInstr *MI,
1709 MachineBasicBlock *MBB) const {
1710 const SystemZInstrInfo *TII = TM.getInstrInfo();
1712 unsigned DestReg = MI->getOperand(0).getReg();
1713 unsigned TrueReg = MI->getOperand(1).getReg();
1714 unsigned FalseReg = MI->getOperand(2).getReg();
1715 unsigned CCValid = MI->getOperand(3).getImm();
1716 unsigned CCMask = MI->getOperand(4).getImm();
1717 DebugLoc DL = MI->getDebugLoc();
1719 MachineBasicBlock *StartMBB = MBB;
1720 MachineBasicBlock *JoinMBB = splitBlockAfter(MI, MBB);
1721 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
1724 // BRC CCMask, JoinMBB
1725 // # fallthrough to FalseMBB
1727 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
1728 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
1729 MBB->addSuccessor(JoinMBB);
1730 MBB->addSuccessor(FalseMBB);
1733 // # fallthrough to JoinMBB
1735 MBB->addSuccessor(JoinMBB);
1738 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
1741 BuildMI(*MBB, MBB->begin(), DL, TII->get(SystemZ::PHI), DestReg)
1742 .addReg(TrueReg).addMBB(StartMBB)
1743 .addReg(FalseReg).addMBB(FalseMBB);
1745 MI->eraseFromParent();
1749 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
1750 // StoreOpcode is the store to use and Invert says whether the store should
1751 // happen when the condition is false rather than true. If a STORE ON
1752 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
1754 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
1755 MachineBasicBlock *MBB,
1756 unsigned StoreOpcode, unsigned STOCOpcode,
1757 bool Invert) const {
1758 const SystemZInstrInfo *TII = TM.getInstrInfo();
1760 unsigned SrcReg = MI->getOperand(0).getReg();
1761 MachineOperand Base = MI->getOperand(1);
1762 int64_t Disp = MI->getOperand(2).getImm();
1763 unsigned IndexReg = MI->getOperand(3).getReg();
1764 unsigned CCValid = MI->getOperand(4).getImm();
1765 unsigned CCMask = MI->getOperand(5).getImm();
1766 DebugLoc DL = MI->getDebugLoc();
1768 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
1770 // Use STOCOpcode if possible. We could use different store patterns in
1771 // order to avoid matching the index register, but the performance trade-offs
1772 // might be more complicated in that case.
1773 if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) {
1776 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
1777 .addReg(SrcReg).addOperand(Base).addImm(Disp).addImm(CCMask);
1778 MI->eraseFromParent();
1782 // Get the condition needed to branch around the store.
1786 MachineBasicBlock *StartMBB = MBB;
1787 MachineBasicBlock *JoinMBB = splitBlockAfter(MI, MBB);
1788 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
1791 // BRC CCMask, JoinMBB
1792 // # fallthrough to FalseMBB
1794 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
1795 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
1796 MBB->addSuccessor(JoinMBB);
1797 MBB->addSuccessor(FalseMBB);
1800 // store %SrcReg, %Disp(%Index,%Base)
1801 // # fallthrough to JoinMBB
1803 BuildMI(MBB, DL, TII->get(StoreOpcode))
1804 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
1805 MBB->addSuccessor(JoinMBB);
1807 MI->eraseFromParent();
1811 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
1812 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
1813 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
1814 // BitSize is the width of the field in bits, or 0 if this is a partword
1815 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
1816 // is one of the operands. Invert says whether the field should be
1817 // inverted after performing BinOpcode (e.g. for NAND).
1819 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
1820 MachineBasicBlock *MBB,
1823 bool Invert) const {
1824 const SystemZInstrInfo *TII = TM.getInstrInfo();
1825 MachineFunction &MF = *MBB->getParent();
1826 MachineRegisterInfo &MRI = MF.getRegInfo();
1827 bool IsSubWord = (BitSize < 32);
1829 // Extract the operands. Base can be a register or a frame index.
1830 // Src2 can be a register or immediate.
1831 unsigned Dest = MI->getOperand(0).getReg();
1832 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
1833 int64_t Disp = MI->getOperand(2).getImm();
1834 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
1835 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
1836 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
1837 DebugLoc DL = MI->getDebugLoc();
1839 BitSize = MI->getOperand(6).getImm();
1841 // Subword operations use 32-bit registers.
1842 const TargetRegisterClass *RC = (BitSize <= 32 ?
1843 &SystemZ::GR32BitRegClass :
1844 &SystemZ::GR64BitRegClass);
1845 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
1846 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
1848 // Get the right opcodes for the displacement.
1849 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
1850 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
1851 assert(LOpcode && CSOpcode && "Displacement out of range");
1853 // Create virtual registers for temporary results.
1854 unsigned OrigVal = MRI.createVirtualRegister(RC);
1855 unsigned OldVal = MRI.createVirtualRegister(RC);
1856 unsigned NewVal = (BinOpcode || IsSubWord ?
1857 MRI.createVirtualRegister(RC) : Src2.getReg());
1858 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
1859 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
1861 // Insert a basic block for the main loop.
1862 MachineBasicBlock *StartMBB = MBB;
1863 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
1864 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
1868 // %OrigVal = L Disp(%Base)
1869 // # fall through to LoopMMB
1871 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
1872 .addOperand(Base).addImm(Disp).addReg(0);
1873 MBB->addSuccessor(LoopMBB);
1876 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
1877 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
1878 // %RotatedNewVal = OP %RotatedOldVal, %Src2
1879 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
1880 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
1882 // # fall through to DoneMMB
1884 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
1885 .addReg(OrigVal).addMBB(StartMBB)
1886 .addReg(Dest).addMBB(LoopMBB);
1888 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
1889 .addReg(OldVal).addReg(BitShift).addImm(0);
1891 // Perform the operation normally and then invert every bit of the field.
1892 unsigned Tmp = MRI.createVirtualRegister(RC);
1893 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
1894 .addReg(RotatedOldVal).addOperand(Src2);
1896 // XILF with the upper BitSize bits set.
1897 BuildMI(MBB, DL, TII->get(SystemZ::XILF32), RotatedNewVal)
1898 .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize)));
1899 else if (BitSize == 32)
1900 // XILF with every bit set.
1901 BuildMI(MBB, DL, TII->get(SystemZ::XILF32), RotatedNewVal)
1902 .addReg(Tmp).addImm(~uint32_t(0));
1904 // Use LCGR and add -1 to the result, which is more compact than
1905 // an XILF, XILH pair.
1906 unsigned Tmp2 = MRI.createVirtualRegister(RC);
1907 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
1908 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
1909 .addReg(Tmp2).addImm(-1);
1911 } else if (BinOpcode)
1912 // A simply binary operation.
1913 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
1914 .addReg(RotatedOldVal).addOperand(Src2);
1916 // Use RISBG to rotate Src2 into position and use it to replace the
1917 // field in RotatedOldVal.
1918 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
1919 .addReg(RotatedOldVal).addReg(Src2.getReg())
1920 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
1922 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
1923 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
1924 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
1925 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
1926 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
1927 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
1928 MBB->addSuccessor(LoopMBB);
1929 MBB->addSuccessor(DoneMBB);
1931 MI->eraseFromParent();
1935 // Implement EmitInstrWithCustomInserter for pseudo
1936 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
1937 // instruction that should be used to compare the current field with the
1938 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
1939 // for when the current field should be kept. BitSize is the width of
1940 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
1942 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
1943 MachineBasicBlock *MBB,
1944 unsigned CompareOpcode,
1945 unsigned KeepOldMask,
1946 unsigned BitSize) const {
1947 const SystemZInstrInfo *TII = TM.getInstrInfo();
1948 MachineFunction &MF = *MBB->getParent();
1949 MachineRegisterInfo &MRI = MF.getRegInfo();
1950 bool IsSubWord = (BitSize < 32);
1952 // Extract the operands. Base can be a register or a frame index.
1953 unsigned Dest = MI->getOperand(0).getReg();
1954 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
1955 int64_t Disp = MI->getOperand(2).getImm();
1956 unsigned Src2 = MI->getOperand(3).getReg();
1957 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
1958 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
1959 DebugLoc DL = MI->getDebugLoc();
1961 BitSize = MI->getOperand(6).getImm();
1963 // Subword operations use 32-bit registers.
1964 const TargetRegisterClass *RC = (BitSize <= 32 ?
1965 &SystemZ::GR32BitRegClass :
1966 &SystemZ::GR64BitRegClass);
1967 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
1968 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
1970 // Get the right opcodes for the displacement.
1971 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
1972 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
1973 assert(LOpcode && CSOpcode && "Displacement out of range");
1975 // Create virtual registers for temporary results.
1976 unsigned OrigVal = MRI.createVirtualRegister(RC);
1977 unsigned OldVal = MRI.createVirtualRegister(RC);
1978 unsigned NewVal = MRI.createVirtualRegister(RC);
1979 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
1980 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
1981 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
1983 // Insert 3 basic blocks for the loop.
1984 MachineBasicBlock *StartMBB = MBB;
1985 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
1986 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
1987 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
1988 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
1992 // %OrigVal = L Disp(%Base)
1993 // # fall through to LoopMMB
1995 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
1996 .addOperand(Base).addImm(Disp).addReg(0);
1997 MBB->addSuccessor(LoopMBB);
2000 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
2001 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2002 // CompareOpcode %RotatedOldVal, %Src2
2003 // BRC KeepOldMask, UpdateMBB
2005 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2006 .addReg(OrigVal).addMBB(StartMBB)
2007 .addReg(Dest).addMBB(UpdateMBB);
2009 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2010 .addReg(OldVal).addReg(BitShift).addImm(0);
2011 BuildMI(MBB, DL, TII->get(CompareOpcode))
2012 .addReg(RotatedOldVal).addReg(Src2);
2013 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2014 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
2015 MBB->addSuccessor(UpdateMBB);
2016 MBB->addSuccessor(UseAltMBB);
2019 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
2020 // # fall through to UpdateMMB
2023 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
2024 .addReg(RotatedOldVal).addReg(Src2)
2025 .addImm(32).addImm(31 + BitSize).addImm(0);
2026 MBB->addSuccessor(UpdateMBB);
2029 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
2030 // [ %RotatedAltVal, UseAltMBB ]
2031 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2032 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2034 // # fall through to DoneMMB
2036 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
2037 .addReg(RotatedOldVal).addMBB(LoopMBB)
2038 .addReg(RotatedAltVal).addMBB(UseAltMBB);
2040 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2041 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2042 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2043 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2044 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2045 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2046 MBB->addSuccessor(LoopMBB);
2047 MBB->addSuccessor(DoneMBB);
2049 MI->eraseFromParent();
2053 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
2056 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
2057 MachineBasicBlock *MBB) const {
2058 const SystemZInstrInfo *TII = TM.getInstrInfo();
2059 MachineFunction &MF = *MBB->getParent();
2060 MachineRegisterInfo &MRI = MF.getRegInfo();
2062 // Extract the operands. Base can be a register or a frame index.
2063 unsigned Dest = MI->getOperand(0).getReg();
2064 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2065 int64_t Disp = MI->getOperand(2).getImm();
2066 unsigned OrigCmpVal = MI->getOperand(3).getReg();
2067 unsigned OrigSwapVal = MI->getOperand(4).getReg();
2068 unsigned BitShift = MI->getOperand(5).getReg();
2069 unsigned NegBitShift = MI->getOperand(6).getReg();
2070 int64_t BitSize = MI->getOperand(7).getImm();
2071 DebugLoc DL = MI->getDebugLoc();
2073 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
2075 // Get the right opcodes for the displacement.
2076 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
2077 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
2078 assert(LOpcode && CSOpcode && "Displacement out of range");
2080 // Create virtual registers for temporary results.
2081 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
2082 unsigned OldVal = MRI.createVirtualRegister(RC);
2083 unsigned CmpVal = MRI.createVirtualRegister(RC);
2084 unsigned SwapVal = MRI.createVirtualRegister(RC);
2085 unsigned StoreVal = MRI.createVirtualRegister(RC);
2086 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
2087 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
2088 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
2090 // Insert 2 basic blocks for the loop.
2091 MachineBasicBlock *StartMBB = MBB;
2092 MachineBasicBlock *DoneMBB = splitBlockAfter(MI, MBB);
2093 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2094 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
2098 // %OrigOldVal = L Disp(%Base)
2099 // # fall through to LoopMMB
2101 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
2102 .addOperand(Base).addImm(Disp).addReg(0);
2103 MBB->addSuccessor(LoopMBB);
2106 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
2107 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
2108 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
2109 // %Dest = RLL %OldVal, BitSize(%BitShift)
2110 // ^^ The low BitSize bits contain the field
2112 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
2113 // ^^ Replace the upper 32-BitSize bits of the
2114 // comparison value with those that we loaded,
2115 // so that we can use a full word comparison.
2116 // CR %Dest, %RetryCmpVal
2118 // # Fall through to SetMBB
2120 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2121 .addReg(OrigOldVal).addMBB(StartMBB)
2122 .addReg(RetryOldVal).addMBB(SetMBB);
2123 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
2124 .addReg(OrigCmpVal).addMBB(StartMBB)
2125 .addReg(RetryCmpVal).addMBB(SetMBB);
2126 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
2127 .addReg(OrigSwapVal).addMBB(StartMBB)
2128 .addReg(RetrySwapVal).addMBB(SetMBB);
2129 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
2130 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
2131 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
2132 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
2133 BuildMI(MBB, DL, TII->get(SystemZ::CR))
2134 .addReg(Dest).addReg(RetryCmpVal);
2135 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2136 .addImm(SystemZ::CCMASK_ICMP)
2137 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
2138 MBB->addSuccessor(DoneMBB);
2139 MBB->addSuccessor(SetMBB);
2142 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
2143 // ^^ Replace the upper 32-BitSize bits of the new
2144 // value with those that we loaded.
2145 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
2146 // ^^ Rotate the new field to its proper position.
2147 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
2149 // # fall through to ExitMMB
2151 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
2152 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
2153 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
2154 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
2155 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
2156 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
2157 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2158 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2159 MBB->addSuccessor(LoopMBB);
2160 MBB->addSuccessor(DoneMBB);
2162 MI->eraseFromParent();
2166 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
2167 // if the high register of the GR128 value must be cleared or false if
2168 // it's "don't care". SubReg is subreg_odd32 when extending a GR32
2169 // and subreg_odd when extending a GR64.
2171 SystemZTargetLowering::emitExt128(MachineInstr *MI,
2172 MachineBasicBlock *MBB,
2173 bool ClearEven, unsigned SubReg) const {
2174 const SystemZInstrInfo *TII = TM.getInstrInfo();
2175 MachineFunction &MF = *MBB->getParent();
2176 MachineRegisterInfo &MRI = MF.getRegInfo();
2177 DebugLoc DL = MI->getDebugLoc();
2179 unsigned Dest = MI->getOperand(0).getReg();
2180 unsigned Src = MI->getOperand(1).getReg();
2181 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
2183 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
2185 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
2186 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
2188 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
2190 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
2191 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_high);
2194 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
2195 .addReg(In128).addReg(Src).addImm(SubReg);
2197 MI->eraseFromParent();
2202 SystemZTargetLowering::emitMVCWrapper(MachineInstr *MI,
2203 MachineBasicBlock *MBB) const {
2204 const SystemZInstrInfo *TII = TM.getInstrInfo();
2205 DebugLoc DL = MI->getDebugLoc();
2207 MachineOperand DestBase = MI->getOperand(0);
2208 uint64_t DestDisp = MI->getOperand(1).getImm();
2209 MachineOperand SrcBase = MI->getOperand(2);
2210 uint64_t SrcDisp = MI->getOperand(3).getImm();
2211 uint64_t Length = MI->getOperand(4).getImm();
2213 BuildMI(*MBB, MI, DL, TII->get(SystemZ::MVC))
2214 .addOperand(DestBase).addImm(DestDisp).addImm(Length)
2215 .addOperand(SrcBase).addImm(SrcDisp);
2217 MI->eraseFromParent();
2221 MachineBasicBlock *SystemZTargetLowering::
2222 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
2223 switch (MI->getOpcode()) {
2224 case SystemZ::Select32:
2225 case SystemZ::SelectF32:
2226 case SystemZ::Select64:
2227 case SystemZ::SelectF64:
2228 case SystemZ::SelectF128:
2229 return emitSelect(MI, MBB);
2231 case SystemZ::CondStore8_32:
2232 return emitCondStore(MI, MBB, SystemZ::STC32, 0, false);
2233 case SystemZ::CondStore8_32Inv:
2234 return emitCondStore(MI, MBB, SystemZ::STC32, 0, true);
2235 case SystemZ::CondStore16_32:
2236 return emitCondStore(MI, MBB, SystemZ::STH32, 0, false);
2237 case SystemZ::CondStore16_32Inv:
2238 return emitCondStore(MI, MBB, SystemZ::STH32, 0, true);
2239 case SystemZ::CondStore32_32:
2240 return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, false);
2241 case SystemZ::CondStore32_32Inv:
2242 return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, true);
2243 case SystemZ::CondStore8:
2244 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
2245 case SystemZ::CondStore8Inv:
2246 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
2247 case SystemZ::CondStore16:
2248 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
2249 case SystemZ::CondStore16Inv:
2250 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
2251 case SystemZ::CondStore32:
2252 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
2253 case SystemZ::CondStore32Inv:
2254 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
2255 case SystemZ::CondStore64:
2256 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
2257 case SystemZ::CondStore64Inv:
2258 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
2259 case SystemZ::CondStoreF32:
2260 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
2261 case SystemZ::CondStoreF32Inv:
2262 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
2263 case SystemZ::CondStoreF64:
2264 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
2265 case SystemZ::CondStoreF64Inv:
2266 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
2268 case SystemZ::AEXT128_64:
2269 return emitExt128(MI, MBB, false, SystemZ::subreg_low);
2270 case SystemZ::ZEXT128_32:
2271 return emitExt128(MI, MBB, true, SystemZ::subreg_low32);
2272 case SystemZ::ZEXT128_64:
2273 return emitExt128(MI, MBB, true, SystemZ::subreg_low);
2275 case SystemZ::ATOMIC_SWAPW:
2276 return emitAtomicLoadBinary(MI, MBB, 0, 0);
2277 case SystemZ::ATOMIC_SWAP_32:
2278 return emitAtomicLoadBinary(MI, MBB, 0, 32);
2279 case SystemZ::ATOMIC_SWAP_64:
2280 return emitAtomicLoadBinary(MI, MBB, 0, 64);
2282 case SystemZ::ATOMIC_LOADW_AR:
2283 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
2284 case SystemZ::ATOMIC_LOADW_AFI:
2285 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
2286 case SystemZ::ATOMIC_LOAD_AR:
2287 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
2288 case SystemZ::ATOMIC_LOAD_AHI:
2289 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
2290 case SystemZ::ATOMIC_LOAD_AFI:
2291 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
2292 case SystemZ::ATOMIC_LOAD_AGR:
2293 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
2294 case SystemZ::ATOMIC_LOAD_AGHI:
2295 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
2296 case SystemZ::ATOMIC_LOAD_AGFI:
2297 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
2299 case SystemZ::ATOMIC_LOADW_SR:
2300 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
2301 case SystemZ::ATOMIC_LOAD_SR:
2302 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
2303 case SystemZ::ATOMIC_LOAD_SGR:
2304 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
2306 case SystemZ::ATOMIC_LOADW_NR:
2307 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
2308 case SystemZ::ATOMIC_LOADW_NILH:
2309 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 0);
2310 case SystemZ::ATOMIC_LOAD_NR:
2311 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
2312 case SystemZ::ATOMIC_LOAD_NILL32:
2313 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL32, 32);
2314 case SystemZ::ATOMIC_LOAD_NILH32:
2315 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 32);
2316 case SystemZ::ATOMIC_LOAD_NILF32:
2317 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF32, 32);
2318 case SystemZ::ATOMIC_LOAD_NGR:
2319 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
2320 case SystemZ::ATOMIC_LOAD_NILL:
2321 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 64);
2322 case SystemZ::ATOMIC_LOAD_NILH:
2323 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 64);
2324 case SystemZ::ATOMIC_LOAD_NIHL:
2325 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL, 64);
2326 case SystemZ::ATOMIC_LOAD_NIHH:
2327 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH, 64);
2328 case SystemZ::ATOMIC_LOAD_NILF:
2329 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 64);
2330 case SystemZ::ATOMIC_LOAD_NIHF:
2331 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF, 64);
2333 case SystemZ::ATOMIC_LOADW_OR:
2334 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
2335 case SystemZ::ATOMIC_LOADW_OILH:
2336 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH32, 0);
2337 case SystemZ::ATOMIC_LOAD_OR:
2338 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
2339 case SystemZ::ATOMIC_LOAD_OILL32:
2340 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL32, 32);
2341 case SystemZ::ATOMIC_LOAD_OILH32:
2342 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH32, 32);
2343 case SystemZ::ATOMIC_LOAD_OILF32:
2344 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF32, 32);
2345 case SystemZ::ATOMIC_LOAD_OGR:
2346 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
2347 case SystemZ::ATOMIC_LOAD_OILL:
2348 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 64);
2349 case SystemZ::ATOMIC_LOAD_OILH:
2350 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 64);
2351 case SystemZ::ATOMIC_LOAD_OIHL:
2352 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL, 64);
2353 case SystemZ::ATOMIC_LOAD_OIHH:
2354 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH, 64);
2355 case SystemZ::ATOMIC_LOAD_OILF:
2356 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 64);
2357 case SystemZ::ATOMIC_LOAD_OIHF:
2358 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF, 64);
2360 case SystemZ::ATOMIC_LOADW_XR:
2361 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
2362 case SystemZ::ATOMIC_LOADW_XILF:
2363 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF32, 0);
2364 case SystemZ::ATOMIC_LOAD_XR:
2365 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
2366 case SystemZ::ATOMIC_LOAD_XILF32:
2367 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF32, 32);
2368 case SystemZ::ATOMIC_LOAD_XGR:
2369 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
2370 case SystemZ::ATOMIC_LOAD_XILF:
2371 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 64);
2372 case SystemZ::ATOMIC_LOAD_XIHF:
2373 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF, 64);
2375 case SystemZ::ATOMIC_LOADW_NRi:
2376 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
2377 case SystemZ::ATOMIC_LOADW_NILHi:
2378 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 0, true);
2379 case SystemZ::ATOMIC_LOAD_NRi:
2380 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
2381 case SystemZ::ATOMIC_LOAD_NILL32i:
2382 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL32, 32, true);
2383 case SystemZ::ATOMIC_LOAD_NILH32i:
2384 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 32, true);
2385 case SystemZ::ATOMIC_LOAD_NILF32i:
2386 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF32, 32, true);
2387 case SystemZ::ATOMIC_LOAD_NGRi:
2388 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
2389 case SystemZ::ATOMIC_LOAD_NILLi:
2390 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 64, true);
2391 case SystemZ::ATOMIC_LOAD_NILHi:
2392 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 64, true);
2393 case SystemZ::ATOMIC_LOAD_NIHLi:
2394 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL, 64, true);
2395 case SystemZ::ATOMIC_LOAD_NIHHi:
2396 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH, 64, true);
2397 case SystemZ::ATOMIC_LOAD_NILFi:
2398 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 64, true);
2399 case SystemZ::ATOMIC_LOAD_NIHFi:
2400 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF, 64, true);
2402 case SystemZ::ATOMIC_LOADW_MIN:
2403 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2404 SystemZ::CCMASK_CMP_LE, 0);
2405 case SystemZ::ATOMIC_LOAD_MIN_32:
2406 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2407 SystemZ::CCMASK_CMP_LE, 32);
2408 case SystemZ::ATOMIC_LOAD_MIN_64:
2409 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
2410 SystemZ::CCMASK_CMP_LE, 64);
2412 case SystemZ::ATOMIC_LOADW_MAX:
2413 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2414 SystemZ::CCMASK_CMP_GE, 0);
2415 case SystemZ::ATOMIC_LOAD_MAX_32:
2416 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
2417 SystemZ::CCMASK_CMP_GE, 32);
2418 case SystemZ::ATOMIC_LOAD_MAX_64:
2419 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
2420 SystemZ::CCMASK_CMP_GE, 64);
2422 case SystemZ::ATOMIC_LOADW_UMIN:
2423 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2424 SystemZ::CCMASK_CMP_LE, 0);
2425 case SystemZ::ATOMIC_LOAD_UMIN_32:
2426 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2427 SystemZ::CCMASK_CMP_LE, 32);
2428 case SystemZ::ATOMIC_LOAD_UMIN_64:
2429 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
2430 SystemZ::CCMASK_CMP_LE, 64);
2432 case SystemZ::ATOMIC_LOADW_UMAX:
2433 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2434 SystemZ::CCMASK_CMP_GE, 0);
2435 case SystemZ::ATOMIC_LOAD_UMAX_32:
2436 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
2437 SystemZ::CCMASK_CMP_GE, 32);
2438 case SystemZ::ATOMIC_LOAD_UMAX_64:
2439 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
2440 SystemZ::CCMASK_CMP_GE, 64);
2442 case SystemZ::ATOMIC_CMP_SWAPW:
2443 return emitAtomicCmpSwapW(MI, MBB);
2444 case SystemZ::MVCWrapper:
2445 return emitMVCWrapper(MI, MBB);
2447 llvm_unreachable("Unexpected instr type to insert");