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 #include "SystemZISelLowering.h"
15 #include "SystemZCallingConv.h"
16 #include "SystemZConstantPoolValue.h"
17 #include "SystemZMachineFunctionInfo.h"
18 #include "SystemZTargetMachine.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
27 #define DEBUG_TYPE "systemz-lower"
30 // Represents a sequence for extracting a 0/1 value from an IPM result:
31 // (((X ^ XORValue) + AddValue) >> Bit)
32 struct IPMConversion {
33 IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
34 : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
41 // Represents information about a comparison.
43 Comparison(SDValue Op0In, SDValue Op1In)
44 : Op0(Op0In), Op1(Op1In), Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
46 // The operands to the comparison.
49 // The opcode that should be used to compare Op0 and Op1.
52 // A SystemZICMP value. Only used for integer comparisons.
55 // The mask of CC values that Opcode can produce.
58 // The mask of CC values for which the original condition is true.
61 } // end anonymous namespace
63 // Classify VT as either 32 or 64 bit.
64 static bool is32Bit(EVT VT) {
65 switch (VT.getSimpleVT().SimpleTy) {
71 llvm_unreachable("Unsupported type");
75 // Return a version of MachineOperand that can be safely used before the
77 static MachineOperand earlyUseOperand(MachineOperand Op) {
83 SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &tm,
84 const SystemZSubtarget &STI)
85 : TargetLowering(tm), Subtarget(STI) {
86 MVT PtrVT = getPointerTy();
88 // Set up the register classes.
89 if (Subtarget.hasHighWord())
90 addRegisterClass(MVT::i32, &SystemZ::GRX32BitRegClass);
92 addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
93 addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
94 addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
95 addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
96 addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
98 // Compute derived properties from the register classes
99 computeRegisterProperties(Subtarget.getRegisterInfo());
101 // Set up special registers.
102 setExceptionPointerRegister(SystemZ::R6D);
103 setExceptionSelectorRegister(SystemZ::R7D);
104 setStackPointerRegisterToSaveRestore(SystemZ::R15D);
106 // TODO: It may be better to default to latency-oriented scheduling, however
107 // LLVM's current latency-oriented scheduler can't handle physreg definitions
108 // such as SystemZ has with CC, so set this to the register-pressure
109 // scheduler, because it can.
110 setSchedulingPreference(Sched::RegPressure);
112 setBooleanContents(ZeroOrOneBooleanContent);
113 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
115 // Instructions are strings of 2-byte aligned 2-byte values.
116 setMinFunctionAlignment(2);
118 // Handle operations that are handled in a similar way for all types.
119 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
120 I <= MVT::LAST_FP_VALUETYPE;
122 MVT VT = MVT::SimpleValueType(I);
123 if (isTypeLegal(VT)) {
124 // Lower SET_CC into an IPM-based sequence.
125 setOperationAction(ISD::SETCC, VT, Custom);
127 // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
128 setOperationAction(ISD::SELECT, VT, Expand);
130 // Lower SELECT_CC and BR_CC into separate comparisons and branches.
131 setOperationAction(ISD::SELECT_CC, VT, Custom);
132 setOperationAction(ISD::BR_CC, VT, Custom);
136 // Expand jump table branches as address arithmetic followed by an
138 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
140 // Expand BRCOND into a BR_CC (see above).
141 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
143 // Handle integer types.
144 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
145 I <= MVT::LAST_INTEGER_VALUETYPE;
147 MVT VT = MVT::SimpleValueType(I);
148 if (isTypeLegal(VT)) {
149 // Expand individual DIV and REMs into DIVREMs.
150 setOperationAction(ISD::SDIV, VT, Expand);
151 setOperationAction(ISD::UDIV, VT, Expand);
152 setOperationAction(ISD::SREM, VT, Expand);
153 setOperationAction(ISD::UREM, VT, Expand);
154 setOperationAction(ISD::SDIVREM, VT, Custom);
155 setOperationAction(ISD::UDIVREM, VT, Custom);
157 // Lower ATOMIC_LOAD and ATOMIC_STORE into normal volatile loads and
158 // stores, putting a serialization instruction after the stores.
159 setOperationAction(ISD::ATOMIC_LOAD, VT, Custom);
160 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
162 // Lower ATOMIC_LOAD_SUB into ATOMIC_LOAD_ADD if LAA and LAAG are
163 // available, or if the operand is constant.
164 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
166 // No special instructions for these.
167 setOperationAction(ISD::CTPOP, VT, Expand);
168 setOperationAction(ISD::CTTZ, VT, Expand);
169 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
170 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
171 setOperationAction(ISD::ROTR, VT, Expand);
173 // Use *MUL_LOHI where possible instead of MULH*.
174 setOperationAction(ISD::MULHS, VT, Expand);
175 setOperationAction(ISD::MULHU, VT, Expand);
176 setOperationAction(ISD::SMUL_LOHI, VT, Custom);
177 setOperationAction(ISD::UMUL_LOHI, VT, Custom);
179 // Only z196 and above have native support for conversions to unsigned.
180 if (!Subtarget.hasFPExtension())
181 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
185 // Type legalization will convert 8- and 16-bit atomic operations into
186 // forms that operate on i32s (but still keeping the original memory VT).
187 // Lower them into full i32 operations.
188 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
189 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
190 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
191 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
192 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
193 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
194 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
195 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
196 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
197 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
198 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
199 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
201 // z10 has instructions for signed but not unsigned FP conversion.
202 // Handle unsigned 32-bit types as signed 64-bit types.
203 if (!Subtarget.hasFPExtension()) {
204 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
205 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
208 // We have native support for a 64-bit CTLZ, via FLOGR.
209 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
210 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
212 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
213 setOperationAction(ISD::OR, MVT::i64, Custom);
215 // FIXME: Can we support these natively?
216 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
217 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
218 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
220 // We have native instructions for i8, i16 and i32 extensions, but not i1.
221 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
222 for (MVT VT : MVT::integer_valuetypes()) {
223 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
224 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
225 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
228 // Handle the various types of symbolic address.
229 setOperationAction(ISD::ConstantPool, PtrVT, Custom);
230 setOperationAction(ISD::GlobalAddress, PtrVT, Custom);
231 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
232 setOperationAction(ISD::BlockAddress, PtrVT, Custom);
233 setOperationAction(ISD::JumpTable, PtrVT, Custom);
235 // We need to handle dynamic allocations specially because of the
236 // 160-byte area at the bottom of the stack.
237 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
239 // Use custom expanders so that we can force the function to use
241 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
242 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
244 // Handle prefetches with PFD or PFDRL.
245 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
247 // Handle floating-point types.
248 for (unsigned I = MVT::FIRST_FP_VALUETYPE;
249 I <= MVT::LAST_FP_VALUETYPE;
251 MVT VT = MVT::SimpleValueType(I);
252 if (isTypeLegal(VT)) {
253 // We can use FI for FRINT.
254 setOperationAction(ISD::FRINT, VT, Legal);
256 // We can use the extended form of FI for other rounding operations.
257 if (Subtarget.hasFPExtension()) {
258 setOperationAction(ISD::FNEARBYINT, VT, Legal);
259 setOperationAction(ISD::FFLOOR, VT, Legal);
260 setOperationAction(ISD::FCEIL, VT, Legal);
261 setOperationAction(ISD::FTRUNC, VT, Legal);
262 setOperationAction(ISD::FROUND, VT, Legal);
265 // No special instructions for these.
266 setOperationAction(ISD::FSIN, VT, Expand);
267 setOperationAction(ISD::FCOS, VT, Expand);
268 setOperationAction(ISD::FREM, VT, Expand);
272 // We have fused multiply-addition for f32 and f64 but not f128.
273 setOperationAction(ISD::FMA, MVT::f32, Legal);
274 setOperationAction(ISD::FMA, MVT::f64, Legal);
275 setOperationAction(ISD::FMA, MVT::f128, Expand);
277 // Needed so that we don't try to implement f128 constant loads using
278 // a load-and-extend of a f80 constant (in cases where the constant
279 // would fit in an f80).
280 for (MVT VT : MVT::fp_valuetypes())
281 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
283 // Floating-point truncation and stores need to be done separately.
284 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
285 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
286 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
288 // We have 64-bit FPR<->GPR moves, but need special handling for
290 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
291 setOperationAction(ISD::BITCAST, MVT::f32, Custom);
293 // VASTART and VACOPY need to deal with the SystemZ-specific varargs
294 // structure, but VAEND is a no-op.
295 setOperationAction(ISD::VASTART, MVT::Other, Custom);
296 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
297 setOperationAction(ISD::VAEND, MVT::Other, Expand);
299 // Codes for which we want to perform some z-specific combinations.
300 setTargetDAGCombine(ISD::SIGN_EXTEND);
302 // We want to use MVC in preference to even a single load/store pair.
303 MaxStoresPerMemcpy = 0;
304 MaxStoresPerMemcpyOptSize = 0;
306 // The main memset sequence is a byte store followed by an MVC.
307 // Two STC or MV..I stores win over that, but the kind of fused stores
308 // generated by target-independent code don't when the byte value is
309 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
310 // than "STC;MVC". Handle the choice in target-specific code instead.
311 MaxStoresPerMemset = 0;
312 MaxStoresPerMemsetOptSize = 0;
315 EVT SystemZTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
318 return VT.changeVectorElementTypeToInteger();
321 bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
322 VT = VT.getScalarType();
327 switch (VT.getSimpleVT().SimpleTy) {
340 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
341 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
342 return Imm.isZero() || Imm.isNegZero();
345 bool SystemZTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
349 // Unaligned accesses should never be slower than the expanded version.
350 // We check specifically for aligned accesses in the few cases where
351 // they are required.
357 bool SystemZTargetLowering::isLegalAddressingMode(const AddrMode &AM,
359 // Punt on globals for now, although they can be used in limited
360 // RELATIVE LONG cases.
364 // Require a 20-bit signed offset.
365 if (!isInt<20>(AM.BaseOffs))
368 // Indexing is OK but no scale factor can be applied.
369 return AM.Scale == 0 || AM.Scale == 1;
372 bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
373 if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
375 unsigned FromBits = FromType->getPrimitiveSizeInBits();
376 unsigned ToBits = ToType->getPrimitiveSizeInBits();
377 return FromBits > ToBits;
380 bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
381 if (!FromVT.isInteger() || !ToVT.isInteger())
383 unsigned FromBits = FromVT.getSizeInBits();
384 unsigned ToBits = ToVT.getSizeInBits();
385 return FromBits > ToBits;
388 //===----------------------------------------------------------------------===//
389 // Inline asm support
390 //===----------------------------------------------------------------------===//
392 TargetLowering::ConstraintType
393 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
394 if (Constraint.size() == 1) {
395 switch (Constraint[0]) {
396 case 'a': // Address register
397 case 'd': // Data register (equivalent to 'r')
398 case 'f': // Floating-point register
399 case 'h': // High-part register
400 case 'r': // General-purpose register
401 return C_RegisterClass;
403 case 'Q': // Memory with base and unsigned 12-bit displacement
404 case 'R': // Likewise, plus an index
405 case 'S': // Memory with base and signed 20-bit displacement
406 case 'T': // Likewise, plus an index
407 case 'm': // Equivalent to 'T'.
410 case 'I': // Unsigned 8-bit constant
411 case 'J': // Unsigned 12-bit constant
412 case 'K': // Signed 16-bit constant
413 case 'L': // Signed 20-bit displacement (on all targets we support)
414 case 'M': // 0x7fffffff
421 return TargetLowering::getConstraintType(Constraint);
424 TargetLowering::ConstraintWeight SystemZTargetLowering::
425 getSingleConstraintMatchWeight(AsmOperandInfo &info,
426 const char *constraint) const {
427 ConstraintWeight weight = CW_Invalid;
428 Value *CallOperandVal = info.CallOperandVal;
429 // If we don't have a value, we can't do a match,
430 // but allow it at the lowest weight.
433 Type *type = CallOperandVal->getType();
434 // Look at the constraint type.
435 switch (*constraint) {
437 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
440 case 'a': // Address register
441 case 'd': // Data register (equivalent to 'r')
442 case 'h': // High-part register
443 case 'r': // General-purpose register
444 if (CallOperandVal->getType()->isIntegerTy())
445 weight = CW_Register;
448 case 'f': // Floating-point register
449 if (type->isFloatingPointTy())
450 weight = CW_Register;
453 case 'I': // Unsigned 8-bit constant
454 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
455 if (isUInt<8>(C->getZExtValue()))
456 weight = CW_Constant;
459 case 'J': // Unsigned 12-bit constant
460 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
461 if (isUInt<12>(C->getZExtValue()))
462 weight = CW_Constant;
465 case 'K': // Signed 16-bit constant
466 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
467 if (isInt<16>(C->getSExtValue()))
468 weight = CW_Constant;
471 case 'L': // Signed 20-bit displacement (on all targets we support)
472 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
473 if (isInt<20>(C->getSExtValue()))
474 weight = CW_Constant;
477 case 'M': // 0x7fffffff
478 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
479 if (C->getZExtValue() == 0x7fffffff)
480 weight = CW_Constant;
486 // Parse a "{tNNN}" register constraint for which the register type "t"
487 // has already been verified. MC is the class associated with "t" and
488 // Map maps 0-based register numbers to LLVM register numbers.
489 static std::pair<unsigned, const TargetRegisterClass *>
490 parseRegisterNumber(const std::string &Constraint,
491 const TargetRegisterClass *RC, const unsigned *Map) {
492 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
493 if (isdigit(Constraint[2])) {
494 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2);
495 unsigned Index = atoi(Suffix.c_str());
496 if (Index < 16 && Map[Index])
497 return std::make_pair(Map[Index], RC);
499 return std::make_pair(0U, nullptr);
502 std::pair<unsigned, const TargetRegisterClass *>
503 SystemZTargetLowering::getRegForInlineAsmConstraint(
504 const TargetRegisterInfo *TRI, const std::string &Constraint,
506 if (Constraint.size() == 1) {
507 // GCC Constraint Letters
508 switch (Constraint[0]) {
510 case 'd': // Data register (equivalent to 'r')
511 case 'r': // General-purpose register
513 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
514 else if (VT == MVT::i128)
515 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
516 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
518 case 'a': // Address register
520 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
521 else if (VT == MVT::i128)
522 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
523 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
525 case 'h': // High-part register (an LLVM extension)
526 return std::make_pair(0U, &SystemZ::GRH32BitRegClass);
528 case 'f': // Floating-point register
530 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
531 else if (VT == MVT::f128)
532 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
533 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
536 if (Constraint[0] == '{') {
537 // We need to override the default register parsing for GPRs and FPRs
538 // because the interpretation depends on VT. The internal names of
539 // the registers are also different from the external names
540 // (F0D and F0S instead of F0, etc.).
541 if (Constraint[1] == 'r') {
543 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
544 SystemZMC::GR32Regs);
546 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
547 SystemZMC::GR128Regs);
548 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
549 SystemZMC::GR64Regs);
551 if (Constraint[1] == 'f') {
553 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
554 SystemZMC::FP32Regs);
556 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
557 SystemZMC::FP128Regs);
558 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
559 SystemZMC::FP64Regs);
562 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
565 void SystemZTargetLowering::
566 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
567 std::vector<SDValue> &Ops,
568 SelectionDAG &DAG) const {
569 // Only support length 1 constraints for now.
570 if (Constraint.length() == 1) {
571 switch (Constraint[0]) {
572 case 'I': // Unsigned 8-bit constant
573 if (auto *C = dyn_cast<ConstantSDNode>(Op))
574 if (isUInt<8>(C->getZExtValue()))
575 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
579 case 'J': // Unsigned 12-bit constant
580 if (auto *C = dyn_cast<ConstantSDNode>(Op))
581 if (isUInt<12>(C->getZExtValue()))
582 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
586 case 'K': // Signed 16-bit constant
587 if (auto *C = dyn_cast<ConstantSDNode>(Op))
588 if (isInt<16>(C->getSExtValue()))
589 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
593 case 'L': // Signed 20-bit displacement (on all targets we support)
594 if (auto *C = dyn_cast<ConstantSDNode>(Op))
595 if (isInt<20>(C->getSExtValue()))
596 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
600 case 'M': // 0x7fffffff
601 if (auto *C = dyn_cast<ConstantSDNode>(Op))
602 if (C->getZExtValue() == 0x7fffffff)
603 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
608 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
611 //===----------------------------------------------------------------------===//
612 // Calling conventions
613 //===----------------------------------------------------------------------===//
615 #include "SystemZGenCallingConv.inc"
617 bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
618 Type *ToType) const {
619 return isTruncateFree(FromType, ToType);
622 bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
623 if (!CI->isTailCall())
628 // Value is a value that has been passed to us in the location described by VA
629 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
630 // any loads onto Chain.
631 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
632 CCValAssign &VA, SDValue Chain,
634 // If the argument has been promoted from a smaller type, insert an
635 // assertion to capture this.
636 if (VA.getLocInfo() == CCValAssign::SExt)
637 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
638 DAG.getValueType(VA.getValVT()));
639 else if (VA.getLocInfo() == CCValAssign::ZExt)
640 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
641 DAG.getValueType(VA.getValVT()));
644 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
645 else if (VA.getLocInfo() == CCValAssign::Indirect)
646 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
647 MachinePointerInfo(), false, false, false, 0);
649 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
653 // Value is a value of type VA.getValVT() that we need to copy into
654 // the location described by VA. Return a copy of Value converted to
655 // VA.getValVT(). The caller is responsible for handling indirect values.
656 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
657 CCValAssign &VA, SDValue Value) {
658 switch (VA.getLocInfo()) {
659 case CCValAssign::SExt:
660 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
661 case CCValAssign::ZExt:
662 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
663 case CCValAssign::AExt:
664 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
665 case CCValAssign::Full:
668 llvm_unreachable("Unhandled getLocInfo()");
672 SDValue SystemZTargetLowering::
673 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
674 const SmallVectorImpl<ISD::InputArg> &Ins,
675 SDLoc DL, SelectionDAG &DAG,
676 SmallVectorImpl<SDValue> &InVals) const {
677 MachineFunction &MF = DAG.getMachineFunction();
678 MachineFrameInfo *MFI = MF.getFrameInfo();
679 MachineRegisterInfo &MRI = MF.getRegInfo();
680 SystemZMachineFunctionInfo *FuncInfo =
681 MF.getInfo<SystemZMachineFunctionInfo>();
683 static_cast<const SystemZFrameLowering *>(Subtarget.getFrameLowering());
685 // Assign locations to all of the incoming arguments.
686 SmallVector<CCValAssign, 16> ArgLocs;
687 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
688 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
690 unsigned NumFixedGPRs = 0;
691 unsigned NumFixedFPRs = 0;
692 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
694 CCValAssign &VA = ArgLocs[I];
695 EVT LocVT = VA.getLocVT();
697 // Arguments passed in registers
698 const TargetRegisterClass *RC;
699 switch (LocVT.getSimpleVT().SimpleTy) {
701 // Integers smaller than i64 should be promoted to i64.
702 llvm_unreachable("Unexpected argument type");
705 RC = &SystemZ::GR32BitRegClass;
709 RC = &SystemZ::GR64BitRegClass;
713 RC = &SystemZ::FP32BitRegClass;
717 RC = &SystemZ::FP64BitRegClass;
721 unsigned VReg = MRI.createVirtualRegister(RC);
722 MRI.addLiveIn(VA.getLocReg(), VReg);
723 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
725 assert(VA.isMemLoc() && "Argument not register or memory");
727 // Create the frame index object for this incoming parameter.
728 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
729 VA.getLocMemOffset(), true);
731 // Create the SelectionDAG nodes corresponding to a load
732 // from this parameter. Unpromoted ints and floats are
733 // passed as right-justified 8-byte values.
734 EVT PtrVT = getPointerTy();
735 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
736 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
737 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4));
738 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
739 MachinePointerInfo::getFixedStack(FI),
740 false, false, false, 0);
743 // Convert the value of the argument register into the value that's
745 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
749 // Save the number of non-varargs registers for later use by va_start, etc.
750 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
751 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
753 // Likewise the address (in the form of a frame index) of where the
754 // first stack vararg would be. The 1-byte size here is arbitrary.
755 int64_t StackSize = CCInfo.getNextStackOffset();
756 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
758 // ...and a similar frame index for the caller-allocated save area
759 // that will be used to store the incoming registers.
760 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
761 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
762 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
764 // Store the FPR varargs in the reserved frame slots. (We store the
765 // GPRs as part of the prologue.)
766 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
767 SDValue MemOps[SystemZ::NumArgFPRs];
768 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
769 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
770 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
771 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
772 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
773 &SystemZ::FP64BitRegClass);
774 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
775 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
776 MachinePointerInfo::getFixedStack(FI),
780 // Join the stores, which are independent of one another.
781 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
782 makeArrayRef(&MemOps[NumFixedFPRs],
783 SystemZ::NumArgFPRs-NumFixedFPRs));
790 static bool canUseSiblingCall(const CCState &ArgCCInfo,
791 SmallVectorImpl<CCValAssign> &ArgLocs) {
792 // Punt if there are any indirect or stack arguments, or if the call
793 // needs the call-saved argument register R6.
794 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
795 CCValAssign &VA = ArgLocs[I];
796 if (VA.getLocInfo() == CCValAssign::Indirect)
800 unsigned Reg = VA.getLocReg();
801 if (Reg == SystemZ::R6H || Reg == SystemZ::R6L || Reg == SystemZ::R6D)
808 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
809 SmallVectorImpl<SDValue> &InVals) const {
810 SelectionDAG &DAG = CLI.DAG;
812 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
813 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
814 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
815 SDValue Chain = CLI.Chain;
816 SDValue Callee = CLI.Callee;
817 bool &IsTailCall = CLI.IsTailCall;
818 CallingConv::ID CallConv = CLI.CallConv;
819 bool IsVarArg = CLI.IsVarArg;
820 MachineFunction &MF = DAG.getMachineFunction();
821 EVT PtrVT = getPointerTy();
823 // Analyze the operands of the call, assigning locations to each operand.
824 SmallVector<CCValAssign, 16> ArgLocs;
825 CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
826 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
828 // We don't support GuaranteedTailCallOpt, only automatically-detected
830 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs))
833 // Get a count of how many bytes are to be pushed on the stack.
834 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
836 // Mark the start of the call.
838 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true),
841 // Copy argument values to their designated locations.
842 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
843 SmallVector<SDValue, 8> MemOpChains;
845 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
846 CCValAssign &VA = ArgLocs[I];
847 SDValue ArgValue = OutVals[I];
849 if (VA.getLocInfo() == CCValAssign::Indirect) {
850 // Store the argument in a stack slot and pass its address.
851 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
852 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
853 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
854 MachinePointerInfo::getFixedStack(FI),
856 ArgValue = SpillSlot;
858 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
861 // Queue up the argument copies and emit them at the end.
862 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
864 assert(VA.isMemLoc() && "Argument not register or memory");
866 // Work out the address of the stack slot. Unpromoted ints and
867 // floats are passed as right-justified 8-byte values.
868 if (!StackPtr.getNode())
869 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
870 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
871 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
873 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
874 DAG.getIntPtrConstant(Offset));
877 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
878 MachinePointerInfo(),
883 // Join the stores, which are independent of one another.
884 if (!MemOpChains.empty())
885 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
887 // Accept direct calls by converting symbolic call addresses to the
888 // associated Target* opcodes. Force %r1 to be used for indirect
891 if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
892 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
893 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
894 } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
895 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
896 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
897 } else if (IsTailCall) {
898 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
899 Glue = Chain.getValue(1);
900 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
903 // Build a sequence of copy-to-reg nodes, chained and glued together.
904 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
905 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
906 RegsToPass[I].second, Glue);
907 Glue = Chain.getValue(1);
910 // The first call operand is the chain and the second is the target address.
911 SmallVector<SDValue, 8> Ops;
912 Ops.push_back(Chain);
913 Ops.push_back(Callee);
915 // Add argument registers to the end of the list so that they are
916 // known live into the call.
917 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
918 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
919 RegsToPass[I].second.getValueType()));
921 // Add a register mask operand representing the call-preserved registers.
922 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
923 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
924 assert(Mask && "Missing call preserved mask for calling convention");
925 Ops.push_back(DAG.getRegisterMask(Mask));
927 // Glue the call to the argument copies, if any.
932 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
934 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, Ops);
935 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, Ops);
936 Glue = Chain.getValue(1);
938 // Mark the end of the call, which is glued to the call itself.
939 Chain = DAG.getCALLSEQ_END(Chain,
940 DAG.getConstant(NumBytes, PtrVT, true),
941 DAG.getConstant(0, PtrVT, true),
943 Glue = Chain.getValue(1);
945 // Assign locations to each value returned by this call.
946 SmallVector<CCValAssign, 16> RetLocs;
947 CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
948 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
950 // Copy all of the result registers out of their specified physreg.
951 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
952 CCValAssign &VA = RetLocs[I];
954 // Copy the value out, gluing the copy to the end of the call sequence.
955 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
956 VA.getLocVT(), Glue);
957 Chain = RetValue.getValue(1);
958 Glue = RetValue.getValue(2);
960 // Convert the value of the return register into the value that's
962 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
969 SystemZTargetLowering::LowerReturn(SDValue Chain,
970 CallingConv::ID CallConv, bool IsVarArg,
971 const SmallVectorImpl<ISD::OutputArg> &Outs,
972 const SmallVectorImpl<SDValue> &OutVals,
973 SDLoc DL, SelectionDAG &DAG) const {
974 MachineFunction &MF = DAG.getMachineFunction();
976 // Assign locations to each returned value.
977 SmallVector<CCValAssign, 16> RetLocs;
978 CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
979 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
981 // Quick exit for void returns
983 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
985 // Copy the result values into the output registers.
987 SmallVector<SDValue, 4> RetOps;
988 RetOps.push_back(Chain);
989 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
990 CCValAssign &VA = RetLocs[I];
991 SDValue RetValue = OutVals[I];
993 // Make the return register live on exit.
994 assert(VA.isRegLoc() && "Can only return in registers!");
996 // Promote the value as required.
997 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
999 // Chain and glue the copies together.
1000 unsigned Reg = VA.getLocReg();
1001 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
1002 Glue = Chain.getValue(1);
1003 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
1006 // Update chain and glue.
1009 RetOps.push_back(Glue);
1011 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, RetOps);
1014 SDValue SystemZTargetLowering::
1015 prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, SelectionDAG &DAG) const {
1016 return DAG.getNode(SystemZISD::SERIALIZE, DL, MVT::Other, Chain);
1019 // CC is a comparison that will be implemented using an integer or
1020 // floating-point comparison. Return the condition code mask for
1021 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
1022 // unsigned comparisons and clear for signed ones. In the floating-point
1023 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
1024 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
1026 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
1027 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
1028 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
1032 llvm_unreachable("Invalid integer condition!");
1041 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
1042 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
1047 // Return a sequence for getting a 1 from an IPM result when CC has a
1048 // value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
1049 // The handling of CC values outside CCValid doesn't matter.
1050 static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
1051 // Deal with cases where the result can be taken directly from a bit
1052 // of the IPM result.
1053 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
1054 return IPMConversion(0, 0, SystemZ::IPM_CC);
1055 if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
1056 return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
1058 // Deal with cases where we can add a value to force the sign bit
1059 // to contain the right value. Putting the bit in 31 means we can
1060 // use SRL rather than RISBG(L), and also makes it easier to get a
1061 // 0/-1 value, so it has priority over the other tests below.
1063 // These sequences rely on the fact that the upper two bits of the
1064 // IPM result are zero.
1065 uint64_t TopBit = uint64_t(1) << 31;
1066 if (CCMask == (CCValid & SystemZ::CCMASK_0))
1067 return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
1068 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
1069 return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
1070 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1072 | SystemZ::CCMASK_2)))
1073 return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
1074 if (CCMask == (CCValid & SystemZ::CCMASK_3))
1075 return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
1076 if (CCMask == (CCValid & (SystemZ::CCMASK_1
1078 | SystemZ::CCMASK_3)))
1079 return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
1081 // Next try inverting the value and testing a bit. 0/1 could be
1082 // handled this way too, but we dealt with that case above.
1083 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
1084 return IPMConversion(-1, 0, SystemZ::IPM_CC);
1086 // Handle cases where adding a value forces a non-sign bit to contain
1088 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
1089 return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
1090 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
1091 return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
1093 // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
1094 // can be done by inverting the low CC bit and applying one of the
1095 // sign-based extractions above.
1096 if (CCMask == (CCValid & SystemZ::CCMASK_1))
1097 return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
1098 if (CCMask == (CCValid & SystemZ::CCMASK_2))
1099 return IPMConversion(1 << SystemZ::IPM_CC,
1100 TopBit - (3 << SystemZ::IPM_CC), 31);
1101 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1103 | SystemZ::CCMASK_3)))
1104 return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
1105 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1107 | SystemZ::CCMASK_3)))
1108 return IPMConversion(1 << SystemZ::IPM_CC,
1109 TopBit - (1 << SystemZ::IPM_CC), 31);
1111 llvm_unreachable("Unexpected CC combination");
1114 // If C can be converted to a comparison against zero, adjust the operands
1116 static void adjustZeroCmp(SelectionDAG &DAG, Comparison &C) {
1117 if (C.ICmpType == SystemZICMP::UnsignedOnly)
1120 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
1124 int64_t Value = ConstOp1->getSExtValue();
1125 if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
1126 (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
1127 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
1128 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
1129 C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1130 C.Op1 = DAG.getConstant(0, C.Op1.getValueType());
1134 // If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
1135 // adjust the operands as necessary.
1136 static void adjustSubwordCmp(SelectionDAG &DAG, Comparison &C) {
1137 // For us to make any changes, it must a comparison between a single-use
1138 // load and a constant.
1139 if (!C.Op0.hasOneUse() ||
1140 C.Op0.getOpcode() != ISD::LOAD ||
1141 C.Op1.getOpcode() != ISD::Constant)
1144 // We must have an 8- or 16-bit load.
1145 auto *Load = cast<LoadSDNode>(C.Op0);
1146 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
1147 if (NumBits != 8 && NumBits != 16)
1150 // The load must be an extending one and the constant must be within the
1151 // range of the unextended value.
1152 auto *ConstOp1 = cast<ConstantSDNode>(C.Op1);
1153 uint64_t Value = ConstOp1->getZExtValue();
1154 uint64_t Mask = (1 << NumBits) - 1;
1155 if (Load->getExtensionType() == ISD::SEXTLOAD) {
1156 // Make sure that ConstOp1 is in range of C.Op0.
1157 int64_t SignedValue = ConstOp1->getSExtValue();
1158 if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
1160 if (C.ICmpType != SystemZICMP::SignedOnly) {
1161 // Unsigned comparison between two sign-extended values is equivalent
1162 // to unsigned comparison between two zero-extended values.
1164 } else if (NumBits == 8) {
1165 // Try to treat the comparison as unsigned, so that we can use CLI.
1166 // Adjust CCMask and Value as necessary.
1167 if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
1168 // Test whether the high bit of the byte is set.
1169 Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
1170 else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
1171 // Test whether the high bit of the byte is clear.
1172 Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
1174 // No instruction exists for this combination.
1176 C.ICmpType = SystemZICMP::UnsignedOnly;
1178 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
1181 assert(C.ICmpType == SystemZICMP::Any &&
1182 "Signedness shouldn't matter here.");
1186 // Make sure that the first operand is an i32 of the right extension type.
1187 ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
1190 if (C.Op0.getValueType() != MVT::i32 ||
1191 Load->getExtensionType() != ExtType)
1192 C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
1193 Load->getChain(), Load->getBasePtr(),
1194 Load->getPointerInfo(), Load->getMemoryVT(),
1195 Load->isVolatile(), Load->isNonTemporal(),
1196 Load->isInvariant(), Load->getAlignment());
1198 // Make sure that the second operand is an i32 with the right value.
1199 if (C.Op1.getValueType() != MVT::i32 ||
1200 Value != ConstOp1->getZExtValue())
1201 C.Op1 = DAG.getConstant(Value, MVT::i32);
1204 // Return true if Op is either an unextended load, or a load suitable
1205 // for integer register-memory comparisons of type ICmpType.
1206 static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
1207 auto *Load = dyn_cast<LoadSDNode>(Op.getNode());
1209 // There are no instructions to compare a register with a memory byte.
1210 if (Load->getMemoryVT() == MVT::i8)
1212 // Otherwise decide on extension type.
1213 switch (Load->getExtensionType()) {
1214 case ISD::NON_EXTLOAD:
1217 return ICmpType != SystemZICMP::UnsignedOnly;
1219 return ICmpType != SystemZICMP::SignedOnly;
1227 // Return true if it is better to swap the operands of C.
1228 static bool shouldSwapCmpOperands(const Comparison &C) {
1229 // Leave f128 comparisons alone, since they have no memory forms.
1230 if (C.Op0.getValueType() == MVT::f128)
1233 // Always keep a floating-point constant second, since comparisons with
1234 // zero can use LOAD TEST and comparisons with other constants make a
1235 // natural memory operand.
1236 if (isa<ConstantFPSDNode>(C.Op1))
1239 // Never swap comparisons with zero since there are many ways to optimize
1241 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1242 if (ConstOp1 && ConstOp1->getZExtValue() == 0)
1245 // Also keep natural memory operands second if the loaded value is
1246 // only used here. Several comparisons have memory forms.
1247 if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
1250 // Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
1251 // In that case we generally prefer the memory to be second.
1252 if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
1253 // The only exceptions are when the second operand is a constant and
1254 // we can use things like CHHSI.
1257 // The unsigned memory-immediate instructions can handle 16-bit
1258 // unsigned integers.
1259 if (C.ICmpType != SystemZICMP::SignedOnly &&
1260 isUInt<16>(ConstOp1->getZExtValue()))
1262 // The signed memory-immediate instructions can handle 16-bit
1264 if (C.ICmpType != SystemZICMP::UnsignedOnly &&
1265 isInt<16>(ConstOp1->getSExtValue()))
1270 // Try to promote the use of CGFR and CLGFR.
1271 unsigned Opcode0 = C.Op0.getOpcode();
1272 if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
1274 if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
1276 if (C.ICmpType != SystemZICMP::SignedOnly &&
1277 Opcode0 == ISD::AND &&
1278 C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
1279 cast<ConstantSDNode>(C.Op0.getOperand(1))->getZExtValue() == 0xffffffff)
1285 // Return a version of comparison CC mask CCMask in which the LT and GT
1286 // actions are swapped.
1287 static unsigned reverseCCMask(unsigned CCMask) {
1288 return ((CCMask & SystemZ::CCMASK_CMP_EQ) |
1289 (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
1290 (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
1291 (CCMask & SystemZ::CCMASK_CMP_UO));
1294 // Check whether C tests for equality between X and Y and whether X - Y
1295 // or Y - X is also computed. In that case it's better to compare the
1296 // result of the subtraction against zero.
1297 static void adjustForSubtraction(SelectionDAG &DAG, Comparison &C) {
1298 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1299 C.CCMask == SystemZ::CCMASK_CMP_NE) {
1300 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1302 if (N->getOpcode() == ISD::SUB &&
1303 ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
1304 (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
1305 C.Op0 = SDValue(N, 0);
1306 C.Op1 = DAG.getConstant(0, N->getValueType(0));
1313 // Check whether C compares a floating-point value with zero and if that
1314 // floating-point value is also negated. In this case we can use the
1315 // negation to set CC, so avoiding separate LOAD AND TEST and
1316 // LOAD (NEGATIVE/COMPLEMENT) instructions.
1317 static void adjustForFNeg(Comparison &C) {
1318 auto *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
1319 if (C1 && C1->isZero()) {
1320 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1322 if (N->getOpcode() == ISD::FNEG) {
1323 C.Op0 = SDValue(N, 0);
1324 C.CCMask = reverseCCMask(C.CCMask);
1331 // Check whether C compares (shl X, 32) with 0 and whether X is
1332 // also sign-extended. In that case it is better to test the result
1333 // of the sign extension using LTGFR.
1335 // This case is important because InstCombine transforms a comparison
1336 // with (sext (trunc X)) into a comparison with (shl X, 32).
1337 static void adjustForLTGFR(Comparison &C) {
1338 // Check for a comparison between (shl X, 32) and 0.
1339 if (C.Op0.getOpcode() == ISD::SHL &&
1340 C.Op0.getValueType() == MVT::i64 &&
1341 C.Op1.getOpcode() == ISD::Constant &&
1342 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1343 auto *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
1344 if (C1 && C1->getZExtValue() == 32) {
1345 SDValue ShlOp0 = C.Op0.getOperand(0);
1346 // See whether X has any SIGN_EXTEND_INREG uses.
1347 for (auto I = ShlOp0->use_begin(), E = ShlOp0->use_end(); I != E; ++I) {
1349 if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
1350 cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
1351 C.Op0 = SDValue(N, 0);
1359 // If C compares the truncation of an extending load, try to compare
1360 // the untruncated value instead. This exposes more opportunities to
1362 static void adjustICmpTruncate(SelectionDAG &DAG, Comparison &C) {
1363 if (C.Op0.getOpcode() == ISD::TRUNCATE &&
1364 C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
1365 C.Op1.getOpcode() == ISD::Constant &&
1366 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1367 auto *L = cast<LoadSDNode>(C.Op0.getOperand(0));
1368 if (L->getMemoryVT().getStoreSizeInBits()
1369 <= C.Op0.getValueType().getSizeInBits()) {
1370 unsigned Type = L->getExtensionType();
1371 if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
1372 (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
1373 C.Op0 = C.Op0.getOperand(0);
1374 C.Op1 = DAG.getConstant(0, C.Op0.getValueType());
1380 // Return true if shift operation N has an in-range constant shift value.
1381 // Store it in ShiftVal if so.
1382 static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
1383 auto *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
1387 uint64_t Amount = Shift->getZExtValue();
1388 if (Amount >= N.getValueType().getSizeInBits())
1395 // Check whether an AND with Mask is suitable for a TEST UNDER MASK
1396 // instruction and whether the CC value is descriptive enough to handle
1397 // a comparison of type Opcode between the AND result and CmpVal.
1398 // CCMask says which comparison result is being tested and BitSize is
1399 // the number of bits in the operands. If TEST UNDER MASK can be used,
1400 // return the corresponding CC mask, otherwise return 0.
1401 static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask,
1402 uint64_t Mask, uint64_t CmpVal,
1403 unsigned ICmpType) {
1404 assert(Mask != 0 && "ANDs with zero should have been removed by now");
1406 // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.
1407 if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&
1408 !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))
1411 // Work out the masks for the lowest and highest bits.
1412 unsigned HighShift = 63 - countLeadingZeros(Mask);
1413 uint64_t High = uint64_t(1) << HighShift;
1414 uint64_t Low = uint64_t(1) << countTrailingZeros(Mask);
1416 // Signed ordered comparisons are effectively unsigned if the sign
1418 bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly);
1420 // Check for equality comparisons with 0, or the equivalent.
1422 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1423 return SystemZ::CCMASK_TM_ALL_0;
1424 if (CCMask == SystemZ::CCMASK_CMP_NE)
1425 return SystemZ::CCMASK_TM_SOME_1;
1427 if (EffectivelyUnsigned && CmpVal <= Low) {
1428 if (CCMask == SystemZ::CCMASK_CMP_LT)
1429 return SystemZ::CCMASK_TM_ALL_0;
1430 if (CCMask == SystemZ::CCMASK_CMP_GE)
1431 return SystemZ::CCMASK_TM_SOME_1;
1433 if (EffectivelyUnsigned && CmpVal < Low) {
1434 if (CCMask == SystemZ::CCMASK_CMP_LE)
1435 return SystemZ::CCMASK_TM_ALL_0;
1436 if (CCMask == SystemZ::CCMASK_CMP_GT)
1437 return SystemZ::CCMASK_TM_SOME_1;
1440 // Check for equality comparisons with the mask, or the equivalent.
1441 if (CmpVal == Mask) {
1442 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1443 return SystemZ::CCMASK_TM_ALL_1;
1444 if (CCMask == SystemZ::CCMASK_CMP_NE)
1445 return SystemZ::CCMASK_TM_SOME_0;
1447 if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) {
1448 if (CCMask == SystemZ::CCMASK_CMP_GT)
1449 return SystemZ::CCMASK_TM_ALL_1;
1450 if (CCMask == SystemZ::CCMASK_CMP_LE)
1451 return SystemZ::CCMASK_TM_SOME_0;
1453 if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) {
1454 if (CCMask == SystemZ::CCMASK_CMP_GE)
1455 return SystemZ::CCMASK_TM_ALL_1;
1456 if (CCMask == SystemZ::CCMASK_CMP_LT)
1457 return SystemZ::CCMASK_TM_SOME_0;
1460 // Check for ordered comparisons with the top bit.
1461 if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) {
1462 if (CCMask == SystemZ::CCMASK_CMP_LE)
1463 return SystemZ::CCMASK_TM_MSB_0;
1464 if (CCMask == SystemZ::CCMASK_CMP_GT)
1465 return SystemZ::CCMASK_TM_MSB_1;
1467 if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) {
1468 if (CCMask == SystemZ::CCMASK_CMP_LT)
1469 return SystemZ::CCMASK_TM_MSB_0;
1470 if (CCMask == SystemZ::CCMASK_CMP_GE)
1471 return SystemZ::CCMASK_TM_MSB_1;
1474 // If there are just two bits, we can do equality checks for Low and High
1476 if (Mask == Low + High) {
1477 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low)
1478 return SystemZ::CCMASK_TM_MIXED_MSB_0;
1479 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low)
1480 return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY;
1481 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High)
1482 return SystemZ::CCMASK_TM_MIXED_MSB_1;
1483 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High)
1484 return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY;
1487 // Looks like we've exhausted our options.
1491 // See whether C can be implemented as a TEST UNDER MASK instruction.
1492 // Update the arguments with the TM version if so.
1493 static void adjustForTestUnderMask(SelectionDAG &DAG, Comparison &C) {
1494 // Check that we have a comparison with a constant.
1495 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1498 uint64_t CmpVal = ConstOp1->getZExtValue();
1500 // Check whether the nonconstant input is an AND with a constant mask.
1503 ConstantSDNode *Mask = nullptr;
1504 if (C.Op0.getOpcode() == ISD::AND) {
1505 NewC.Op0 = C.Op0.getOperand(0);
1506 NewC.Op1 = C.Op0.getOperand(1);
1507 Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
1510 MaskVal = Mask->getZExtValue();
1512 // There is no instruction to compare with a 64-bit immediate
1513 // so use TMHH instead if possible. We need an unsigned ordered
1514 // comparison with an i64 immediate.
1515 if (NewC.Op0.getValueType() != MVT::i64 ||
1516 NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
1517 NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
1518 NewC.ICmpType == SystemZICMP::SignedOnly)
1520 // Convert LE and GT comparisons into LT and GE.
1521 if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
1522 NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
1523 if (CmpVal == uint64_t(-1))
1526 NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1528 // If the low N bits of Op1 are zero than the low N bits of Op0 can
1529 // be masked off without changing the result.
1530 MaskVal = -(CmpVal & -CmpVal);
1531 NewC.ICmpType = SystemZICMP::UnsignedOnly;
1534 // Check whether the combination of mask, comparison value and comparison
1535 // type are suitable.
1536 unsigned BitSize = NewC.Op0.getValueType().getSizeInBits();
1537 unsigned NewCCMask, ShiftVal;
1538 if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1539 NewC.Op0.getOpcode() == ISD::SHL &&
1540 isSimpleShift(NewC.Op0, ShiftVal) &&
1541 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1542 MaskVal >> ShiftVal,
1544 SystemZICMP::Any))) {
1545 NewC.Op0 = NewC.Op0.getOperand(0);
1546 MaskVal >>= ShiftVal;
1547 } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1548 NewC.Op0.getOpcode() == ISD::SRL &&
1549 isSimpleShift(NewC.Op0, ShiftVal) &&
1550 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1551 MaskVal << ShiftVal,
1553 SystemZICMP::UnsignedOnly))) {
1554 NewC.Op0 = NewC.Op0.getOperand(0);
1555 MaskVal <<= ShiftVal;
1557 NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
1563 // Go ahead and make the change.
1564 C.Opcode = SystemZISD::TM;
1566 if (Mask && Mask->getZExtValue() == MaskVal)
1567 C.Op1 = SDValue(Mask, 0);
1569 C.Op1 = DAG.getConstant(MaskVal, C.Op0.getValueType());
1570 C.CCValid = SystemZ::CCMASK_TM;
1571 C.CCMask = NewCCMask;
1574 // Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
1575 static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
1576 ISD::CondCode Cond) {
1577 Comparison C(CmpOp0, CmpOp1);
1578 C.CCMask = CCMaskForCondCode(Cond);
1579 if (C.Op0.getValueType().isFloatingPoint()) {
1580 C.CCValid = SystemZ::CCMASK_FCMP;
1581 C.Opcode = SystemZISD::FCMP;
1584 C.CCValid = SystemZ::CCMASK_ICMP;
1585 C.Opcode = SystemZISD::ICMP;
1586 // Choose the type of comparison. Equality and inequality tests can
1587 // use either signed or unsigned comparisons. The choice also doesn't
1588 // matter if both sign bits are known to be clear. In those cases we
1589 // want to give the main isel code the freedom to choose whichever
1591 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1592 C.CCMask == SystemZ::CCMASK_CMP_NE ||
1593 (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
1594 C.ICmpType = SystemZICMP::Any;
1595 else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
1596 C.ICmpType = SystemZICMP::UnsignedOnly;
1598 C.ICmpType = SystemZICMP::SignedOnly;
1599 C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
1600 adjustZeroCmp(DAG, C);
1601 adjustSubwordCmp(DAG, C);
1602 adjustForSubtraction(DAG, C);
1604 adjustICmpTruncate(DAG, C);
1607 if (shouldSwapCmpOperands(C)) {
1608 std::swap(C.Op0, C.Op1);
1609 C.CCMask = reverseCCMask(C.CCMask);
1612 adjustForTestUnderMask(DAG, C);
1616 // Emit the comparison instruction described by C.
1617 static SDValue emitCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1618 if (C.Opcode == SystemZISD::ICMP)
1619 return DAG.getNode(SystemZISD::ICMP, DL, MVT::Glue, C.Op0, C.Op1,
1620 DAG.getConstant(C.ICmpType, MVT::i32));
1621 if (C.Opcode == SystemZISD::TM) {
1622 bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
1623 bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
1624 return DAG.getNode(SystemZISD::TM, DL, MVT::Glue, C.Op0, C.Op1,
1625 DAG.getConstant(RegisterOnly, MVT::i32));
1627 return DAG.getNode(C.Opcode, DL, MVT::Glue, C.Op0, C.Op1);
1630 // Implement a 32-bit *MUL_LOHI operation by extending both operands to
1631 // 64 bits. Extend is the extension type to use. Store the high part
1632 // in Hi and the low part in Lo.
1633 static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL,
1634 unsigned Extend, SDValue Op0, SDValue Op1,
1635 SDValue &Hi, SDValue &Lo) {
1636 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
1637 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
1638 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
1639 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, DAG.getConstant(32, MVT::i64));
1640 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
1641 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
1644 // Lower a binary operation that produces two VT results, one in each
1645 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
1646 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
1647 // on the extended Op0 and (unextended) Op1. Store the even register result
1648 // in Even and the odd register result in Odd.
1649 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
1650 unsigned Extend, unsigned Opcode,
1651 SDValue Op0, SDValue Op1,
1652 SDValue &Even, SDValue &Odd) {
1653 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
1654 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
1655 SDValue(In128, 0), Op1);
1656 bool Is32Bit = is32Bit(VT);
1657 Even = DAG.getTargetExtractSubreg(SystemZ::even128(Is32Bit), DL, VT, Result);
1658 Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
1661 // Return an i32 value that is 1 if the CC value produced by Glue is
1662 // in the mask CCMask and 0 otherwise. CC is known to have a value
1663 // in CCValid, so other values can be ignored.
1664 static SDValue emitSETCC(SelectionDAG &DAG, SDLoc DL, SDValue Glue,
1665 unsigned CCValid, unsigned CCMask) {
1666 IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
1667 SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
1669 if (Conversion.XORValue)
1670 Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
1671 DAG.getConstant(Conversion.XORValue, MVT::i32));
1673 if (Conversion.AddValue)
1674 Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
1675 DAG.getConstant(Conversion.AddValue, MVT::i32));
1677 // The SHR/AND sequence should get optimized to an RISBG.
1678 Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
1679 DAG.getConstant(Conversion.Bit, MVT::i32));
1680 if (Conversion.Bit != 31)
1681 Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
1682 DAG.getConstant(1, MVT::i32));
1686 SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
1687 SelectionDAG &DAG) const {
1688 SDValue CmpOp0 = Op.getOperand(0);
1689 SDValue CmpOp1 = Op.getOperand(1);
1690 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1693 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1694 SDValue Glue = emitCmp(DAG, DL, C);
1695 return emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1698 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1699 SDValue Chain = Op.getOperand(0);
1700 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1701 SDValue CmpOp0 = Op.getOperand(2);
1702 SDValue CmpOp1 = Op.getOperand(3);
1703 SDValue Dest = Op.getOperand(4);
1706 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1707 SDValue Glue = emitCmp(DAG, DL, C);
1708 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
1709 Chain, DAG.getConstant(C.CCValid, MVT::i32),
1710 DAG.getConstant(C.CCMask, MVT::i32), Dest, Glue);
1713 // Return true if Pos is CmpOp and Neg is the negative of CmpOp,
1714 // allowing Pos and Neg to be wider than CmpOp.
1715 static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
1716 return (Neg.getOpcode() == ISD::SUB &&
1717 Neg.getOperand(0).getOpcode() == ISD::Constant &&
1718 cast<ConstantSDNode>(Neg.getOperand(0))->getZExtValue() == 0 &&
1719 Neg.getOperand(1) == Pos &&
1721 (Pos.getOpcode() == ISD::SIGN_EXTEND &&
1722 Pos.getOperand(0) == CmpOp)));
1725 // Return the absolute or negative absolute of Op; IsNegative decides which.
1726 static SDValue getAbsolute(SelectionDAG &DAG, SDLoc DL, SDValue Op,
1728 Op = DAG.getNode(SystemZISD::IABS, DL, Op.getValueType(), Op);
1730 Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
1731 DAG.getConstant(0, Op.getValueType()), Op);
1735 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
1736 SelectionDAG &DAG) const {
1737 SDValue CmpOp0 = Op.getOperand(0);
1738 SDValue CmpOp1 = Op.getOperand(1);
1739 SDValue TrueOp = Op.getOperand(2);
1740 SDValue FalseOp = Op.getOperand(3);
1741 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1744 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1746 // Check for absolute and negative-absolute selections, including those
1747 // where the comparison value is sign-extended (for LPGFR and LNGFR).
1748 // This check supplements the one in DAGCombiner.
1749 if (C.Opcode == SystemZISD::ICMP &&
1750 C.CCMask != SystemZ::CCMASK_CMP_EQ &&
1751 C.CCMask != SystemZ::CCMASK_CMP_NE &&
1752 C.Op1.getOpcode() == ISD::Constant &&
1753 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1754 if (isAbsolute(C.Op0, TrueOp, FalseOp))
1755 return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
1756 if (isAbsolute(C.Op0, FalseOp, TrueOp))
1757 return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
1760 SDValue Glue = emitCmp(DAG, DL, C);
1762 // Special case for handling -1/0 results. The shifts we use here
1763 // should get optimized with the IPM conversion sequence.
1764 auto *TrueC = dyn_cast<ConstantSDNode>(TrueOp);
1765 auto *FalseC = dyn_cast<ConstantSDNode>(FalseOp);
1766 if (TrueC && FalseC) {
1767 int64_t TrueVal = TrueC->getSExtValue();
1768 int64_t FalseVal = FalseC->getSExtValue();
1769 if ((TrueVal == -1 && FalseVal == 0) || (TrueVal == 0 && FalseVal == -1)) {
1770 // Invert the condition if we want -1 on false.
1772 C.CCMask ^= C.CCValid;
1773 SDValue Result = emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1774 EVT VT = Op.getValueType();
1775 // Extend the result to VT. Upper bits are ignored.
1777 Result = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Result);
1778 // Sign-extend from the low bit.
1779 SDValue ShAmt = DAG.getConstant(VT.getSizeInBits() - 1, MVT::i32);
1780 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Result, ShAmt);
1781 return DAG.getNode(ISD::SRA, DL, VT, Shl, ShAmt);
1785 SDValue Ops[] = {TrueOp, FalseOp, DAG.getConstant(C.CCValid, MVT::i32),
1786 DAG.getConstant(C.CCMask, MVT::i32), Glue};
1788 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
1789 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, Ops);
1792 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
1793 SelectionDAG &DAG) const {
1795 const GlobalValue *GV = Node->getGlobal();
1796 int64_t Offset = Node->getOffset();
1797 EVT PtrVT = getPointerTy();
1798 Reloc::Model RM = DAG.getTarget().getRelocationModel();
1799 CodeModel::Model CM = DAG.getTarget().getCodeModel();
1802 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
1803 // Assign anchors at 1<<12 byte boundaries.
1804 uint64_t Anchor = Offset & ~uint64_t(0xfff);
1805 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor);
1806 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1808 // The offset can be folded into the address if it is aligned to a halfword.
1810 if (Offset != 0 && (Offset & 1) == 0) {
1811 SDValue Full = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor + Offset);
1812 Result = DAG.getNode(SystemZISD::PCREL_OFFSET, DL, PtrVT, Full, Result);
1816 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
1817 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1818 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
1819 MachinePointerInfo::getGOT(), false, false, false, 0);
1822 // If there was a non-zero offset that we didn't fold, create an explicit
1825 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
1826 DAG.getConstant(Offset, PtrVT));
1831 SDValue SystemZTargetLowering::lowerTLSGetOffset(GlobalAddressSDNode *Node,
1834 SDValue GOTOffset) const {
1836 EVT PtrVT = getPointerTy();
1837 SDValue Chain = DAG.getEntryNode();
1840 // __tls_get_offset takes the GOT offset in %r2 and the GOT in %r12.
1841 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1842 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R12D, GOT, Glue);
1843 Glue = Chain.getValue(1);
1844 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R2D, GOTOffset, Glue);
1845 Glue = Chain.getValue(1);
1847 // The first call operand is the chain and the second is the TLS symbol.
1848 SmallVector<SDValue, 8> Ops;
1849 Ops.push_back(Chain);
1850 Ops.push_back(DAG.getTargetGlobalAddress(Node->getGlobal(), DL,
1851 Node->getValueType(0),
1854 // Add argument registers to the end of the list so that they are
1855 // known live into the call.
1856 Ops.push_back(DAG.getRegister(SystemZ::R2D, PtrVT));
1857 Ops.push_back(DAG.getRegister(SystemZ::R12D, PtrVT));
1859 // Add a register mask operand representing the call-preserved registers.
1860 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
1861 const uint32_t *Mask = TRI->getCallPreservedMask(CallingConv::C);
1862 assert(Mask && "Missing call preserved mask for calling convention");
1863 Ops.push_back(DAG.getRegisterMask(Mask));
1865 // Glue the call to the argument copies.
1866 Ops.push_back(Glue);
1869 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1870 Chain = DAG.getNode(Opcode, DL, NodeTys, Ops);
1871 Glue = Chain.getValue(1);
1873 // Copy the return value from %r2.
1874 return DAG.getCopyFromReg(Chain, DL, SystemZ::R2D, PtrVT, Glue);
1877 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
1878 SelectionDAG &DAG) const {
1880 const GlobalValue *GV = Node->getGlobal();
1881 EVT PtrVT = getPointerTy();
1882 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
1884 // The high part of the thread pointer is in access register 0.
1885 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1886 DAG.getConstant(0, MVT::i32));
1887 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
1889 // The low part of the thread pointer is in access register 1.
1890 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1891 DAG.getConstant(1, MVT::i32));
1892 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
1894 // Merge them into a single 64-bit address.
1895 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
1896 DAG.getConstant(32, PtrVT));
1897 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
1899 // Get the offset of GA from the thread pointer, based on the TLS model.
1902 case TLSModel::GeneralDynamic: {
1903 // Load the GOT offset of the tls_index (module ID / per-symbol offset).
1904 SystemZConstantPoolValue *CPV =
1905 SystemZConstantPoolValue::Create(GV, SystemZCP::TLSGD);
1907 Offset = DAG.getConstantPool(CPV, PtrVT, 8);
1908 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1909 Offset, MachinePointerInfo::getConstantPool(),
1910 false, false, false, 0);
1912 // Call __tls_get_offset to retrieve the offset.
1913 Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_GDCALL, Offset);
1917 case TLSModel::LocalDynamic: {
1918 // Load the GOT offset of the module ID.
1919 SystemZConstantPoolValue *CPV =
1920 SystemZConstantPoolValue::Create(GV, SystemZCP::TLSLDM);
1922 Offset = DAG.getConstantPool(CPV, PtrVT, 8);
1923 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1924 Offset, MachinePointerInfo::getConstantPool(),
1925 false, false, false, 0);
1927 // Call __tls_get_offset to retrieve the module base offset.
1928 Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_LDCALL, Offset);
1930 // Note: The SystemZLDCleanupPass will remove redundant computations
1931 // of the module base offset. Count total number of local-dynamic
1932 // accesses to trigger execution of that pass.
1933 SystemZMachineFunctionInfo* MFI =
1934 DAG.getMachineFunction().getInfo<SystemZMachineFunctionInfo>();
1935 MFI->incNumLocalDynamicTLSAccesses();
1937 // Add the per-symbol offset.
1938 CPV = SystemZConstantPoolValue::Create(GV, SystemZCP::DTPOFF);
1940 SDValue DTPOffset = DAG.getConstantPool(CPV, PtrVT, 8);
1941 DTPOffset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1942 DTPOffset, MachinePointerInfo::getConstantPool(),
1943 false, false, false, 0);
1945 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Offset, DTPOffset);
1949 case TLSModel::InitialExec: {
1950 // Load the offset from the GOT.
1951 Offset = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1952 SystemZII::MO_INDNTPOFF);
1953 Offset = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Offset);
1954 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1955 Offset, MachinePointerInfo::getGOT(),
1956 false, false, false, 0);
1960 case TLSModel::LocalExec: {
1961 // Force the offset into the constant pool and load it from there.
1962 SystemZConstantPoolValue *CPV =
1963 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
1965 Offset = DAG.getConstantPool(CPV, PtrVT, 8);
1966 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1967 Offset, MachinePointerInfo::getConstantPool(),
1968 false, false, false, 0);
1973 // Add the base and offset together.
1974 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
1977 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
1978 SelectionDAG &DAG) const {
1980 const BlockAddress *BA = Node->getBlockAddress();
1981 int64_t Offset = Node->getOffset();
1982 EVT PtrVT = getPointerTy();
1984 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
1985 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1989 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
1990 SelectionDAG &DAG) const {
1992 EVT PtrVT = getPointerTy();
1993 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1995 // Use LARL to load the address of the table.
1996 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1999 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
2000 SelectionDAG &DAG) const {
2002 EVT PtrVT = getPointerTy();
2005 if (CP->isMachineConstantPoolEntry())
2006 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2007 CP->getAlignment());
2009 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2010 CP->getAlignment(), CP->getOffset());
2012 // Use LARL to load the address of the constant pool entry.
2013 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
2016 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
2017 SelectionDAG &DAG) const {
2019 SDValue In = Op.getOperand(0);
2020 EVT InVT = In.getValueType();
2021 EVT ResVT = Op.getValueType();
2023 if (InVT == MVT::i32 && ResVT == MVT::f32) {
2025 if (Subtarget.hasHighWord()) {
2026 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL,
2028 In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
2029 MVT::i64, SDValue(U64, 0), In);
2031 In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
2032 In64 = DAG.getNode(ISD::SHL, DL, MVT::i64, In64,
2033 DAG.getConstant(32, MVT::i64));
2035 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, In64);
2036 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32,
2037 DL, MVT::f32, Out64);
2039 if (InVT == MVT::f32 && ResVT == MVT::i32) {
2040 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
2041 SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
2042 MVT::f64, SDValue(U64, 0), In);
2043 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
2044 if (Subtarget.hasHighWord())
2045 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32, DL,
2047 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64,
2048 DAG.getConstant(32, MVT::i64));
2049 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
2051 llvm_unreachable("Unexpected bitcast combination");
2054 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
2055 SelectionDAG &DAG) const {
2056 MachineFunction &MF = DAG.getMachineFunction();
2057 SystemZMachineFunctionInfo *FuncInfo =
2058 MF.getInfo<SystemZMachineFunctionInfo>();
2059 EVT PtrVT = getPointerTy();
2061 SDValue Chain = Op.getOperand(0);
2062 SDValue Addr = Op.getOperand(1);
2063 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2066 // The initial values of each field.
2067 const unsigned NumFields = 4;
2068 SDValue Fields[NumFields] = {
2069 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT),
2070 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT),
2071 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
2072 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
2075 // Store each field into its respective slot.
2076 SDValue MemOps[NumFields];
2077 unsigned Offset = 0;
2078 for (unsigned I = 0; I < NumFields; ++I) {
2079 SDValue FieldAddr = Addr;
2081 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
2082 DAG.getIntPtrConstant(Offset));
2083 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
2084 MachinePointerInfo(SV, Offset),
2088 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2091 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
2092 SelectionDAG &DAG) const {
2093 SDValue Chain = Op.getOperand(0);
2094 SDValue DstPtr = Op.getOperand(1);
2095 SDValue SrcPtr = Op.getOperand(2);
2096 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
2097 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
2100 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32),
2101 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
2102 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
2105 SDValue SystemZTargetLowering::
2106 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
2107 SDValue Chain = Op.getOperand(0);
2108 SDValue Size = Op.getOperand(1);
2111 unsigned SPReg = getStackPointerRegisterToSaveRestore();
2113 // Get a reference to the stack pointer.
2114 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
2116 // Get the new stack pointer value.
2117 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
2119 // Copy the new stack pointer back.
2120 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
2122 // The allocated data lives above the 160 bytes allocated for the standard
2123 // frame, plus any outgoing stack arguments. We don't know how much that
2124 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
2125 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
2126 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
2128 SDValue Ops[2] = { Result, Chain };
2129 return DAG.getMergeValues(Ops, DL);
2132 SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
2133 SelectionDAG &DAG) const {
2134 EVT VT = Op.getValueType();
2138 // Just do a normal 64-bit multiplication and extract the results.
2139 // We define this so that it can be used for constant division.
2140 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
2141 Op.getOperand(1), Ops[1], Ops[0]);
2143 // Do a full 128-bit multiplication based on UMUL_LOHI64:
2145 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
2147 // but using the fact that the upper halves are either all zeros
2150 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
2152 // and grouping the right terms together since they are quicker than the
2155 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
2156 SDValue C63 = DAG.getConstant(63, MVT::i64);
2157 SDValue LL = Op.getOperand(0);
2158 SDValue RL = Op.getOperand(1);
2159 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
2160 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
2161 // UMUL_LOHI64 returns the low result in the odd register and the high
2162 // result in the even register. SMUL_LOHI is defined to return the
2163 // low half first, so the results are in reverse order.
2164 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2165 LL, RL, Ops[1], Ops[0]);
2166 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
2167 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
2168 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
2169 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
2171 return DAG.getMergeValues(Ops, DL);
2174 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
2175 SelectionDAG &DAG) const {
2176 EVT VT = Op.getValueType();
2180 // Just do a normal 64-bit multiplication and extract the results.
2181 // We define this so that it can be used for constant division.
2182 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
2183 Op.getOperand(1), Ops[1], Ops[0]);
2185 // UMUL_LOHI64 returns the low result in the odd register and the high
2186 // result in the even register. UMUL_LOHI is defined to return the
2187 // low half first, so the results are in reverse order.
2188 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2189 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2190 return DAG.getMergeValues(Ops, DL);
2193 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
2194 SelectionDAG &DAG) const {
2195 SDValue Op0 = Op.getOperand(0);
2196 SDValue Op1 = Op.getOperand(1);
2197 EVT VT = Op.getValueType();
2201 // We use DSGF for 32-bit division.
2203 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
2204 Opcode = SystemZISD::SDIVREM32;
2205 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
2206 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
2207 Opcode = SystemZISD::SDIVREM32;
2209 Opcode = SystemZISD::SDIVREM64;
2211 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
2212 // input is "don't care". The instruction returns the remainder in
2213 // the even register and the quotient in the odd register.
2215 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
2216 Op0, Op1, Ops[1], Ops[0]);
2217 return DAG.getMergeValues(Ops, DL);
2220 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
2221 SelectionDAG &DAG) const {
2222 EVT VT = Op.getValueType();
2225 // DL(G) uses a double-width dividend, so we need to clear the even
2226 // register in the GR128 input. The instruction returns the remainder
2227 // in the even register and the quotient in the odd register.
2230 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
2231 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2233 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
2234 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2235 return DAG.getMergeValues(Ops, DL);
2238 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
2239 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
2241 // Get the known-zero masks for each operand.
2242 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
2243 APInt KnownZero[2], KnownOne[2];
2244 DAG.computeKnownBits(Ops[0], KnownZero[0], KnownOne[0]);
2245 DAG.computeKnownBits(Ops[1], KnownZero[1], KnownOne[1]);
2247 // See if the upper 32 bits of one operand and the lower 32 bits of the
2248 // other are known zero. They are the low and high operands respectively.
2249 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
2250 KnownZero[1].getZExtValue() };
2252 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
2254 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
2259 SDValue LowOp = Ops[Low];
2260 SDValue HighOp = Ops[High];
2262 // If the high part is a constant, we're better off using IILH.
2263 if (HighOp.getOpcode() == ISD::Constant)
2266 // If the low part is a constant that is outside the range of LHI,
2267 // then we're better off using IILF.
2268 if (LowOp.getOpcode() == ISD::Constant) {
2269 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
2270 if (!isInt<16>(Value))
2274 // Check whether the high part is an AND that doesn't change the
2275 // high 32 bits and just masks out low bits. We can skip it if so.
2276 if (HighOp.getOpcode() == ISD::AND &&
2277 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
2278 SDValue HighOp0 = HighOp.getOperand(0);
2279 uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
2280 if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
2284 // Take advantage of the fact that all GR32 operations only change the
2285 // low 32 bits by truncating Low to an i32 and inserting it directly
2286 // using a subreg. The interesting cases are those where the truncation
2289 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
2290 return DAG.getTargetInsertSubreg(SystemZ::subreg_l32, DL,
2291 MVT::i64, HighOp, Low32);
2294 // Op is an atomic load. Lower it into a normal volatile load.
2295 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
2296 SelectionDAG &DAG) const {
2297 auto *Node = cast<AtomicSDNode>(Op.getNode());
2298 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), Op.getValueType(),
2299 Node->getChain(), Node->getBasePtr(),
2300 Node->getMemoryVT(), Node->getMemOperand());
2303 // Op is an atomic store. Lower it into a normal volatile store followed
2304 // by a serialization.
2305 SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
2306 SelectionDAG &DAG) const {
2307 auto *Node = cast<AtomicSDNode>(Op.getNode());
2308 SDValue Chain = DAG.getTruncStore(Node->getChain(), SDLoc(Op), Node->getVal(),
2309 Node->getBasePtr(), Node->getMemoryVT(),
2310 Node->getMemOperand());
2311 return SDValue(DAG.getMachineNode(SystemZ::Serialize, SDLoc(Op), MVT::Other,
2315 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
2316 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
2317 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
2319 unsigned Opcode) const {
2320 auto *Node = cast<AtomicSDNode>(Op.getNode());
2322 // 32-bit operations need no code outside the main loop.
2323 EVT NarrowVT = Node->getMemoryVT();
2324 EVT WideVT = MVT::i32;
2325 if (NarrowVT == WideVT)
2328 int64_t BitSize = NarrowVT.getSizeInBits();
2329 SDValue ChainIn = Node->getChain();
2330 SDValue Addr = Node->getBasePtr();
2331 SDValue Src2 = Node->getVal();
2332 MachineMemOperand *MMO = Node->getMemOperand();
2334 EVT PtrVT = Addr.getValueType();
2336 // Convert atomic subtracts of constants into additions.
2337 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
2338 if (auto *Const = dyn_cast<ConstantSDNode>(Src2)) {
2339 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
2340 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
2343 // Get the address of the containing word.
2344 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2345 DAG.getConstant(-4, PtrVT));
2347 // Get the number of bits that the word must be rotated left in order
2348 // to bring the field to the top bits of a GR32.
2349 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2350 DAG.getConstant(3, PtrVT));
2351 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2353 // Get the complementing shift amount, for rotating a field in the top
2354 // bits back to its proper position.
2355 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2356 DAG.getConstant(0, WideVT), BitShift);
2358 // Extend the source operand to 32 bits and prepare it for the inner loop.
2359 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
2360 // operations require the source to be shifted in advance. (This shift
2361 // can be folded if the source is constant.) For AND and NAND, the lower
2362 // bits must be set, while for other opcodes they should be left clear.
2363 if (Opcode != SystemZISD::ATOMIC_SWAPW)
2364 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
2365 DAG.getConstant(32 - BitSize, WideVT));
2366 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
2367 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
2368 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
2369 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT));
2371 // Construct the ATOMIC_LOADW_* node.
2372 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2373 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
2374 DAG.getConstant(BitSize, WideVT) };
2375 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
2378 // Rotate the result of the final CS so that the field is in the lower
2379 // bits of a GR32, then truncate it.
2380 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
2381 DAG.getConstant(BitSize, WideVT));
2382 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
2384 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
2385 return DAG.getMergeValues(RetOps, DL);
2388 // Op is an ATOMIC_LOAD_SUB operation. Lower 8- and 16-bit operations
2389 // into ATOMIC_LOADW_SUBs and decide whether to convert 32- and 64-bit
2390 // operations into additions.
2391 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
2392 SelectionDAG &DAG) const {
2393 auto *Node = cast<AtomicSDNode>(Op.getNode());
2394 EVT MemVT = Node->getMemoryVT();
2395 if (MemVT == MVT::i32 || MemVT == MVT::i64) {
2396 // A full-width operation.
2397 assert(Op.getValueType() == MemVT && "Mismatched VTs");
2398 SDValue Src2 = Node->getVal();
2402 if (auto *Op2 = dyn_cast<ConstantSDNode>(Src2)) {
2403 // Use an addition if the operand is constant and either LAA(G) is
2404 // available or the negative value is in the range of A(G)FHI.
2405 int64_t Value = (-Op2->getAPIntValue()).getSExtValue();
2406 if (isInt<32>(Value) || Subtarget.hasInterlockedAccess1())
2407 NegSrc2 = DAG.getConstant(Value, MemVT);
2408 } else if (Subtarget.hasInterlockedAccess1())
2409 // Use LAA(G) if available.
2410 NegSrc2 = DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, MemVT),
2413 if (NegSrc2.getNode())
2414 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
2415 Node->getChain(), Node->getBasePtr(), NegSrc2,
2416 Node->getMemOperand(), Node->getOrdering(),
2417 Node->getSynchScope());
2419 // Use the node as-is.
2423 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
2426 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
2427 // into a fullword ATOMIC_CMP_SWAPW operation.
2428 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
2429 SelectionDAG &DAG) const {
2430 auto *Node = cast<AtomicSDNode>(Op.getNode());
2432 // We have native support for 32-bit compare and swap.
2433 EVT NarrowVT = Node->getMemoryVT();
2434 EVT WideVT = MVT::i32;
2435 if (NarrowVT == WideVT)
2438 int64_t BitSize = NarrowVT.getSizeInBits();
2439 SDValue ChainIn = Node->getOperand(0);
2440 SDValue Addr = Node->getOperand(1);
2441 SDValue CmpVal = Node->getOperand(2);
2442 SDValue SwapVal = Node->getOperand(3);
2443 MachineMemOperand *MMO = Node->getMemOperand();
2445 EVT PtrVT = Addr.getValueType();
2447 // Get the address of the containing word.
2448 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2449 DAG.getConstant(-4, PtrVT));
2451 // Get the number of bits that the word must be rotated left in order
2452 // to bring the field to the top bits of a GR32.
2453 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2454 DAG.getConstant(3, PtrVT));
2455 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2457 // Get the complementing shift amount, for rotating a field in the top
2458 // bits back to its proper position.
2459 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2460 DAG.getConstant(0, WideVT), BitShift);
2462 // Construct the ATOMIC_CMP_SWAPW node.
2463 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2464 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
2465 NegBitShift, DAG.getConstant(BitSize, WideVT) };
2466 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
2467 VTList, Ops, NarrowVT, MMO);
2471 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
2472 SelectionDAG &DAG) const {
2473 MachineFunction &MF = DAG.getMachineFunction();
2474 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2475 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
2476 SystemZ::R15D, Op.getValueType());
2479 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
2480 SelectionDAG &DAG) const {
2481 MachineFunction &MF = DAG.getMachineFunction();
2482 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2483 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
2484 SystemZ::R15D, Op.getOperand(1));
2487 SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op,
2488 SelectionDAG &DAG) const {
2489 bool IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2491 // Just preserve the chain.
2492 return Op.getOperand(0);
2494 bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2495 unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
2496 auto *Node = cast<MemIntrinsicSDNode>(Op.getNode());
2499 DAG.getConstant(Code, MVT::i32),
2502 return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, SDLoc(Op),
2503 Node->getVTList(), Ops,
2504 Node->getMemoryVT(), Node->getMemOperand());
2507 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
2508 SelectionDAG &DAG) const {
2509 switch (Op.getOpcode()) {
2511 return lowerBR_CC(Op, DAG);
2512 case ISD::SELECT_CC:
2513 return lowerSELECT_CC(Op, DAG);
2515 return lowerSETCC(Op, DAG);
2516 case ISD::GlobalAddress:
2517 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
2518 case ISD::GlobalTLSAddress:
2519 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
2520 case ISD::BlockAddress:
2521 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
2522 case ISD::JumpTable:
2523 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
2524 case ISD::ConstantPool:
2525 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
2527 return lowerBITCAST(Op, DAG);
2529 return lowerVASTART(Op, DAG);
2531 return lowerVACOPY(Op, DAG);
2532 case ISD::DYNAMIC_STACKALLOC:
2533 return lowerDYNAMIC_STACKALLOC(Op, DAG);
2534 case ISD::SMUL_LOHI:
2535 return lowerSMUL_LOHI(Op, DAG);
2536 case ISD::UMUL_LOHI:
2537 return lowerUMUL_LOHI(Op, DAG);
2539 return lowerSDIVREM(Op, DAG);
2541 return lowerUDIVREM(Op, DAG);
2543 return lowerOR(Op, DAG);
2544 case ISD::ATOMIC_SWAP:
2545 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
2546 case ISD::ATOMIC_STORE:
2547 return lowerATOMIC_STORE(Op, DAG);
2548 case ISD::ATOMIC_LOAD:
2549 return lowerATOMIC_LOAD(Op, DAG);
2550 case ISD::ATOMIC_LOAD_ADD:
2551 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
2552 case ISD::ATOMIC_LOAD_SUB:
2553 return lowerATOMIC_LOAD_SUB(Op, DAG);
2554 case ISD::ATOMIC_LOAD_AND:
2555 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
2556 case ISD::ATOMIC_LOAD_OR:
2557 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
2558 case ISD::ATOMIC_LOAD_XOR:
2559 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
2560 case ISD::ATOMIC_LOAD_NAND:
2561 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
2562 case ISD::ATOMIC_LOAD_MIN:
2563 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
2564 case ISD::ATOMIC_LOAD_MAX:
2565 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
2566 case ISD::ATOMIC_LOAD_UMIN:
2567 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
2568 case ISD::ATOMIC_LOAD_UMAX:
2569 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
2570 case ISD::ATOMIC_CMP_SWAP:
2571 return lowerATOMIC_CMP_SWAP(Op, DAG);
2572 case ISD::STACKSAVE:
2573 return lowerSTACKSAVE(Op, DAG);
2574 case ISD::STACKRESTORE:
2575 return lowerSTACKRESTORE(Op, DAG);
2577 return lowerPREFETCH(Op, DAG);
2579 llvm_unreachable("Unexpected node to lower");
2583 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
2584 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
2589 OPCODE(PCREL_WRAPPER);
2590 OPCODE(PCREL_OFFSET);
2596 OPCODE(SELECT_CCMASK);
2597 OPCODE(ADJDYNALLOC);
2598 OPCODE(EXTRACT_ACCESS);
2599 OPCODE(UMUL_LOHI64);
2615 OPCODE(SEARCH_STRING);
2618 OPCODE(ATOMIC_SWAPW);
2619 OPCODE(ATOMIC_LOADW_ADD);
2620 OPCODE(ATOMIC_LOADW_SUB);
2621 OPCODE(ATOMIC_LOADW_AND);
2622 OPCODE(ATOMIC_LOADW_OR);
2623 OPCODE(ATOMIC_LOADW_XOR);
2624 OPCODE(ATOMIC_LOADW_NAND);
2625 OPCODE(ATOMIC_LOADW_MIN);
2626 OPCODE(ATOMIC_LOADW_MAX);
2627 OPCODE(ATOMIC_LOADW_UMIN);
2628 OPCODE(ATOMIC_LOADW_UMAX);
2629 OPCODE(ATOMIC_CMP_SWAPW);
2636 SDValue SystemZTargetLowering::PerformDAGCombine(SDNode *N,
2637 DAGCombinerInfo &DCI) const {
2638 SelectionDAG &DAG = DCI.DAG;
2639 unsigned Opcode = N->getOpcode();
2640 if (Opcode == ISD::SIGN_EXTEND) {
2641 // Convert (sext (ashr (shl X, C1), C2)) to
2642 // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
2643 // cheap as narrower ones.
2644 SDValue N0 = N->getOperand(0);
2645 EVT VT = N->getValueType(0);
2646 if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
2647 auto *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2648 SDValue Inner = N0.getOperand(0);
2649 if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
2650 if (auto *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1))) {
2651 unsigned Extra = (VT.getSizeInBits() -
2652 N0.getValueType().getSizeInBits());
2653 unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
2654 unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
2655 EVT ShiftVT = N0.getOperand(1).getValueType();
2656 SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
2657 Inner.getOperand(0));
2658 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
2659 DAG.getConstant(NewShlAmt, ShiftVT));
2660 return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
2661 DAG.getConstant(NewSraAmt, ShiftVT));
2669 //===----------------------------------------------------------------------===//
2671 //===----------------------------------------------------------------------===//
2673 // Create a new basic block after MBB.
2674 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
2675 MachineFunction &MF = *MBB->getParent();
2676 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
2677 MF.insert(std::next(MachineFunction::iterator(MBB)), NewMBB);
2681 // Split MBB after MI and return the new block (the one that contains
2682 // instructions after MI).
2683 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
2684 MachineBasicBlock *MBB) {
2685 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2686 NewMBB->splice(NewMBB->begin(), MBB,
2687 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
2688 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2692 // Split MBB before MI and return the new block (the one that contains MI).
2693 static MachineBasicBlock *splitBlockBefore(MachineInstr *MI,
2694 MachineBasicBlock *MBB) {
2695 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2696 NewMBB->splice(NewMBB->begin(), MBB, MI, MBB->end());
2697 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2701 // Force base value Base into a register before MI. Return the register.
2702 static unsigned forceReg(MachineInstr *MI, MachineOperand &Base,
2703 const SystemZInstrInfo *TII) {
2705 return Base.getReg();
2707 MachineBasicBlock *MBB = MI->getParent();
2708 MachineFunction &MF = *MBB->getParent();
2709 MachineRegisterInfo &MRI = MF.getRegInfo();
2711 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
2712 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LA), Reg)
2713 .addOperand(Base).addImm(0).addReg(0);
2717 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
2719 SystemZTargetLowering::emitSelect(MachineInstr *MI,
2720 MachineBasicBlock *MBB) const {
2721 const SystemZInstrInfo *TII =
2722 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
2724 unsigned DestReg = MI->getOperand(0).getReg();
2725 unsigned TrueReg = MI->getOperand(1).getReg();
2726 unsigned FalseReg = MI->getOperand(2).getReg();
2727 unsigned CCValid = MI->getOperand(3).getImm();
2728 unsigned CCMask = MI->getOperand(4).getImm();
2729 DebugLoc DL = MI->getDebugLoc();
2731 MachineBasicBlock *StartMBB = MBB;
2732 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2733 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2736 // BRC CCMask, JoinMBB
2737 // # fallthrough to FalseMBB
2739 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2740 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2741 MBB->addSuccessor(JoinMBB);
2742 MBB->addSuccessor(FalseMBB);
2745 // # fallthrough to JoinMBB
2747 MBB->addSuccessor(JoinMBB);
2750 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
2753 BuildMI(*MBB, MI, DL, TII->get(SystemZ::PHI), DestReg)
2754 .addReg(TrueReg).addMBB(StartMBB)
2755 .addReg(FalseReg).addMBB(FalseMBB);
2757 MI->eraseFromParent();
2761 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
2762 // StoreOpcode is the store to use and Invert says whether the store should
2763 // happen when the condition is false rather than true. If a STORE ON
2764 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
2766 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
2767 MachineBasicBlock *MBB,
2768 unsigned StoreOpcode, unsigned STOCOpcode,
2769 bool Invert) const {
2770 const SystemZInstrInfo *TII =
2771 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
2773 unsigned SrcReg = MI->getOperand(0).getReg();
2774 MachineOperand Base = MI->getOperand(1);
2775 int64_t Disp = MI->getOperand(2).getImm();
2776 unsigned IndexReg = MI->getOperand(3).getReg();
2777 unsigned CCValid = MI->getOperand(4).getImm();
2778 unsigned CCMask = MI->getOperand(5).getImm();
2779 DebugLoc DL = MI->getDebugLoc();
2781 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
2783 // Use STOCOpcode if possible. We could use different store patterns in
2784 // order to avoid matching the index register, but the performance trade-offs
2785 // might be more complicated in that case.
2786 if (STOCOpcode && !IndexReg && Subtarget.hasLoadStoreOnCond()) {
2789 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
2790 .addReg(SrcReg).addOperand(Base).addImm(Disp)
2791 .addImm(CCValid).addImm(CCMask);
2792 MI->eraseFromParent();
2796 // Get the condition needed to branch around the store.
2800 MachineBasicBlock *StartMBB = MBB;
2801 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2802 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2805 // BRC CCMask, JoinMBB
2806 // # fallthrough to FalseMBB
2808 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2809 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2810 MBB->addSuccessor(JoinMBB);
2811 MBB->addSuccessor(FalseMBB);
2814 // store %SrcReg, %Disp(%Index,%Base)
2815 // # fallthrough to JoinMBB
2817 BuildMI(MBB, DL, TII->get(StoreOpcode))
2818 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
2819 MBB->addSuccessor(JoinMBB);
2821 MI->eraseFromParent();
2825 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
2826 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
2827 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
2828 // BitSize is the width of the field in bits, or 0 if this is a partword
2829 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
2830 // is one of the operands. Invert says whether the field should be
2831 // inverted after performing BinOpcode (e.g. for NAND).
2833 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
2834 MachineBasicBlock *MBB,
2837 bool Invert) const {
2838 MachineFunction &MF = *MBB->getParent();
2839 const SystemZInstrInfo *TII =
2840 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
2841 MachineRegisterInfo &MRI = MF.getRegInfo();
2842 bool IsSubWord = (BitSize < 32);
2844 // Extract the operands. Base can be a register or a frame index.
2845 // Src2 can be a register or immediate.
2846 unsigned Dest = MI->getOperand(0).getReg();
2847 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2848 int64_t Disp = MI->getOperand(2).getImm();
2849 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
2850 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2851 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2852 DebugLoc DL = MI->getDebugLoc();
2854 BitSize = MI->getOperand(6).getImm();
2856 // Subword operations use 32-bit registers.
2857 const TargetRegisterClass *RC = (BitSize <= 32 ?
2858 &SystemZ::GR32BitRegClass :
2859 &SystemZ::GR64BitRegClass);
2860 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2861 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2863 // Get the right opcodes for the displacement.
2864 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2865 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2866 assert(LOpcode && CSOpcode && "Displacement out of range");
2868 // Create virtual registers for temporary results.
2869 unsigned OrigVal = MRI.createVirtualRegister(RC);
2870 unsigned OldVal = MRI.createVirtualRegister(RC);
2871 unsigned NewVal = (BinOpcode || IsSubWord ?
2872 MRI.createVirtualRegister(RC) : Src2.getReg());
2873 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2874 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2876 // Insert a basic block for the main loop.
2877 MachineBasicBlock *StartMBB = MBB;
2878 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2879 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2883 // %OrigVal = L Disp(%Base)
2884 // # fall through to LoopMMB
2886 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2887 .addOperand(Base).addImm(Disp).addReg(0);
2888 MBB->addSuccessor(LoopMBB);
2891 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
2892 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2893 // %RotatedNewVal = OP %RotatedOldVal, %Src2
2894 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2895 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2897 // # fall through to DoneMMB
2899 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2900 .addReg(OrigVal).addMBB(StartMBB)
2901 .addReg(Dest).addMBB(LoopMBB);
2903 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2904 .addReg(OldVal).addReg(BitShift).addImm(0);
2906 // Perform the operation normally and then invert every bit of the field.
2907 unsigned Tmp = MRI.createVirtualRegister(RC);
2908 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
2909 .addReg(RotatedOldVal).addOperand(Src2);
2911 // XILF with the upper BitSize bits set.
2912 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2913 .addReg(Tmp).addImm(-1U << (32 - BitSize));
2915 // Use LCGR and add -1 to the result, which is more compact than
2916 // an XILF, XILH pair.
2917 unsigned Tmp2 = MRI.createVirtualRegister(RC);
2918 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
2919 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
2920 .addReg(Tmp2).addImm(-1);
2922 } else if (BinOpcode)
2923 // A simply binary operation.
2924 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
2925 .addReg(RotatedOldVal).addOperand(Src2);
2927 // Use RISBG to rotate Src2 into position and use it to replace the
2928 // field in RotatedOldVal.
2929 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
2930 .addReg(RotatedOldVal).addReg(Src2.getReg())
2931 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
2933 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2934 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2935 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2936 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2937 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2938 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2939 MBB->addSuccessor(LoopMBB);
2940 MBB->addSuccessor(DoneMBB);
2942 MI->eraseFromParent();
2946 // Implement EmitInstrWithCustomInserter for pseudo
2947 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
2948 // instruction that should be used to compare the current field with the
2949 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
2950 // for when the current field should be kept. BitSize is the width of
2951 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
2953 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
2954 MachineBasicBlock *MBB,
2955 unsigned CompareOpcode,
2956 unsigned KeepOldMask,
2957 unsigned BitSize) const {
2958 MachineFunction &MF = *MBB->getParent();
2959 const SystemZInstrInfo *TII =
2960 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
2961 MachineRegisterInfo &MRI = MF.getRegInfo();
2962 bool IsSubWord = (BitSize < 32);
2964 // Extract the operands. Base can be a register or a frame index.
2965 unsigned Dest = MI->getOperand(0).getReg();
2966 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2967 int64_t Disp = MI->getOperand(2).getImm();
2968 unsigned Src2 = MI->getOperand(3).getReg();
2969 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2970 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2971 DebugLoc DL = MI->getDebugLoc();
2973 BitSize = MI->getOperand(6).getImm();
2975 // Subword operations use 32-bit registers.
2976 const TargetRegisterClass *RC = (BitSize <= 32 ?
2977 &SystemZ::GR32BitRegClass :
2978 &SystemZ::GR64BitRegClass);
2979 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2980 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2982 // Get the right opcodes for the displacement.
2983 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2984 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2985 assert(LOpcode && CSOpcode && "Displacement out of range");
2987 // Create virtual registers for temporary results.
2988 unsigned OrigVal = MRI.createVirtualRegister(RC);
2989 unsigned OldVal = MRI.createVirtualRegister(RC);
2990 unsigned NewVal = MRI.createVirtualRegister(RC);
2991 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2992 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
2993 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2995 // Insert 3 basic blocks for the loop.
2996 MachineBasicBlock *StartMBB = MBB;
2997 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2998 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2999 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
3000 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
3004 // %OrigVal = L Disp(%Base)
3005 // # fall through to LoopMMB
3007 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
3008 .addOperand(Base).addImm(Disp).addReg(0);
3009 MBB->addSuccessor(LoopMBB);
3012 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
3013 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
3014 // CompareOpcode %RotatedOldVal, %Src2
3015 // BRC KeepOldMask, UpdateMBB
3017 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
3018 .addReg(OrigVal).addMBB(StartMBB)
3019 .addReg(Dest).addMBB(UpdateMBB);
3021 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
3022 .addReg(OldVal).addReg(BitShift).addImm(0);
3023 BuildMI(MBB, DL, TII->get(CompareOpcode))
3024 .addReg(RotatedOldVal).addReg(Src2);
3025 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3026 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
3027 MBB->addSuccessor(UpdateMBB);
3028 MBB->addSuccessor(UseAltMBB);
3031 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
3032 // # fall through to UpdateMMB
3035 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
3036 .addReg(RotatedOldVal).addReg(Src2)
3037 .addImm(32).addImm(31 + BitSize).addImm(0);
3038 MBB->addSuccessor(UpdateMBB);
3041 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
3042 // [ %RotatedAltVal, UseAltMBB ]
3043 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
3044 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
3046 // # fall through to DoneMMB
3048 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
3049 .addReg(RotatedOldVal).addMBB(LoopMBB)
3050 .addReg(RotatedAltVal).addMBB(UseAltMBB);
3052 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
3053 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
3054 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
3055 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
3056 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3057 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
3058 MBB->addSuccessor(LoopMBB);
3059 MBB->addSuccessor(DoneMBB);
3061 MI->eraseFromParent();
3065 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
3068 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
3069 MachineBasicBlock *MBB) const {
3070 MachineFunction &MF = *MBB->getParent();
3071 const SystemZInstrInfo *TII =
3072 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
3073 MachineRegisterInfo &MRI = MF.getRegInfo();
3075 // Extract the operands. Base can be a register or a frame index.
3076 unsigned Dest = MI->getOperand(0).getReg();
3077 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
3078 int64_t Disp = MI->getOperand(2).getImm();
3079 unsigned OrigCmpVal = MI->getOperand(3).getReg();
3080 unsigned OrigSwapVal = MI->getOperand(4).getReg();
3081 unsigned BitShift = MI->getOperand(5).getReg();
3082 unsigned NegBitShift = MI->getOperand(6).getReg();
3083 int64_t BitSize = MI->getOperand(7).getImm();
3084 DebugLoc DL = MI->getDebugLoc();
3086 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
3088 // Get the right opcodes for the displacement.
3089 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
3090 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
3091 assert(LOpcode && CSOpcode && "Displacement out of range");
3093 // Create virtual registers for temporary results.
3094 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
3095 unsigned OldVal = MRI.createVirtualRegister(RC);
3096 unsigned CmpVal = MRI.createVirtualRegister(RC);
3097 unsigned SwapVal = MRI.createVirtualRegister(RC);
3098 unsigned StoreVal = MRI.createVirtualRegister(RC);
3099 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
3100 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
3101 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
3103 // Insert 2 basic blocks for the loop.
3104 MachineBasicBlock *StartMBB = MBB;
3105 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3106 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3107 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
3111 // %OrigOldVal = L Disp(%Base)
3112 // # fall through to LoopMMB
3114 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
3115 .addOperand(Base).addImm(Disp).addReg(0);
3116 MBB->addSuccessor(LoopMBB);
3119 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
3120 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
3121 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
3122 // %Dest = RLL %OldVal, BitSize(%BitShift)
3123 // ^^ The low BitSize bits contain the field
3125 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
3126 // ^^ Replace the upper 32-BitSize bits of the
3127 // comparison value with those that we loaded,
3128 // so that we can use a full word comparison.
3129 // CR %Dest, %RetryCmpVal
3131 // # Fall through to SetMBB
3133 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
3134 .addReg(OrigOldVal).addMBB(StartMBB)
3135 .addReg(RetryOldVal).addMBB(SetMBB);
3136 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
3137 .addReg(OrigCmpVal).addMBB(StartMBB)
3138 .addReg(RetryCmpVal).addMBB(SetMBB);
3139 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
3140 .addReg(OrigSwapVal).addMBB(StartMBB)
3141 .addReg(RetrySwapVal).addMBB(SetMBB);
3142 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
3143 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
3144 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
3145 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3146 BuildMI(MBB, DL, TII->get(SystemZ::CR))
3147 .addReg(Dest).addReg(RetryCmpVal);
3148 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3149 .addImm(SystemZ::CCMASK_ICMP)
3150 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
3151 MBB->addSuccessor(DoneMBB);
3152 MBB->addSuccessor(SetMBB);
3155 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
3156 // ^^ Replace the upper 32-BitSize bits of the new
3157 // value with those that we loaded.
3158 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
3159 // ^^ Rotate the new field to its proper position.
3160 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
3162 // # fall through to ExitMMB
3164 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
3165 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3166 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
3167 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
3168 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
3169 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
3170 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3171 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
3172 MBB->addSuccessor(LoopMBB);
3173 MBB->addSuccessor(DoneMBB);
3175 MI->eraseFromParent();
3179 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
3180 // if the high register of the GR128 value must be cleared or false if
3181 // it's "don't care". SubReg is subreg_l32 when extending a GR32
3182 // and subreg_l64 when extending a GR64.
3184 SystemZTargetLowering::emitExt128(MachineInstr *MI,
3185 MachineBasicBlock *MBB,
3186 bool ClearEven, unsigned SubReg) const {
3187 MachineFunction &MF = *MBB->getParent();
3188 const SystemZInstrInfo *TII =
3189 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
3190 MachineRegisterInfo &MRI = MF.getRegInfo();
3191 DebugLoc DL = MI->getDebugLoc();
3193 unsigned Dest = MI->getOperand(0).getReg();
3194 unsigned Src = MI->getOperand(1).getReg();
3195 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3197 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
3199 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3200 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
3202 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
3204 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
3205 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_h64);
3208 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
3209 .addReg(In128).addReg(Src).addImm(SubReg);
3211 MI->eraseFromParent();
3216 SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
3217 MachineBasicBlock *MBB,
3218 unsigned Opcode) const {
3219 MachineFunction &MF = *MBB->getParent();
3220 const SystemZInstrInfo *TII =
3221 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
3222 MachineRegisterInfo &MRI = MF.getRegInfo();
3223 DebugLoc DL = MI->getDebugLoc();
3225 MachineOperand DestBase = earlyUseOperand(MI->getOperand(0));
3226 uint64_t DestDisp = MI->getOperand(1).getImm();
3227 MachineOperand SrcBase = earlyUseOperand(MI->getOperand(2));
3228 uint64_t SrcDisp = MI->getOperand(3).getImm();
3229 uint64_t Length = MI->getOperand(4).getImm();
3231 // When generating more than one CLC, all but the last will need to
3232 // branch to the end when a difference is found.
3233 MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ?
3234 splitBlockAfter(MI, MBB) : nullptr);
3236 // Check for the loop form, in which operand 5 is the trip count.
3237 if (MI->getNumExplicitOperands() > 5) {
3238 bool HaveSingleBase = DestBase.isIdenticalTo(SrcBase);
3240 uint64_t StartCountReg = MI->getOperand(5).getReg();
3241 uint64_t StartSrcReg = forceReg(MI, SrcBase, TII);
3242 uint64_t StartDestReg = (HaveSingleBase ? StartSrcReg :
3243 forceReg(MI, DestBase, TII));
3245 const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass;
3246 uint64_t ThisSrcReg = MRI.createVirtualRegister(RC);
3247 uint64_t ThisDestReg = (HaveSingleBase ? ThisSrcReg :
3248 MRI.createVirtualRegister(RC));
3249 uint64_t NextSrcReg = MRI.createVirtualRegister(RC);
3250 uint64_t NextDestReg = (HaveSingleBase ? NextSrcReg :
3251 MRI.createVirtualRegister(RC));
3253 RC = &SystemZ::GR64BitRegClass;
3254 uint64_t ThisCountReg = MRI.createVirtualRegister(RC);
3255 uint64_t NextCountReg = MRI.createVirtualRegister(RC);
3257 MachineBasicBlock *StartMBB = MBB;
3258 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3259 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3260 MachineBasicBlock *NextMBB = (EndMBB ? emitBlockAfter(LoopMBB) : LoopMBB);
3263 // # fall through to LoopMMB
3264 MBB->addSuccessor(LoopMBB);
3267 // %ThisDestReg = phi [ %StartDestReg, StartMBB ],
3268 // [ %NextDestReg, NextMBB ]
3269 // %ThisSrcReg = phi [ %StartSrcReg, StartMBB ],
3270 // [ %NextSrcReg, NextMBB ]
3271 // %ThisCountReg = phi [ %StartCountReg, StartMBB ],
3272 // [ %NextCountReg, NextMBB ]
3273 // ( PFD 2, 768+DestDisp(%ThisDestReg) )
3274 // Opcode DestDisp(256,%ThisDestReg), SrcDisp(%ThisSrcReg)
3277 // The prefetch is used only for MVC. The JLH is used only for CLC.
3280 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisDestReg)
3281 .addReg(StartDestReg).addMBB(StartMBB)
3282 .addReg(NextDestReg).addMBB(NextMBB);
3283 if (!HaveSingleBase)
3284 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisSrcReg)
3285 .addReg(StartSrcReg).addMBB(StartMBB)
3286 .addReg(NextSrcReg).addMBB(NextMBB);
3287 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisCountReg)
3288 .addReg(StartCountReg).addMBB(StartMBB)
3289 .addReg(NextCountReg).addMBB(NextMBB);
3290 if (Opcode == SystemZ::MVC)
3291 BuildMI(MBB, DL, TII->get(SystemZ::PFD))
3292 .addImm(SystemZ::PFD_WRITE)
3293 .addReg(ThisDestReg).addImm(DestDisp + 768).addReg(0);
3294 BuildMI(MBB, DL, TII->get(Opcode))
3295 .addReg(ThisDestReg).addImm(DestDisp).addImm(256)
3296 .addReg(ThisSrcReg).addImm(SrcDisp);
3298 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3299 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3301 MBB->addSuccessor(EndMBB);
3302 MBB->addSuccessor(NextMBB);
3306 // %NextDestReg = LA 256(%ThisDestReg)
3307 // %NextSrcReg = LA 256(%ThisSrcReg)
3308 // %NextCountReg = AGHI %ThisCountReg, -1
3309 // CGHI %NextCountReg, 0
3311 // # fall through to DoneMMB
3313 // The AGHI, CGHI and JLH should be converted to BRCTG by later passes.
3316 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextDestReg)
3317 .addReg(ThisDestReg).addImm(256).addReg(0);
3318 if (!HaveSingleBase)
3319 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextSrcReg)
3320 .addReg(ThisSrcReg).addImm(256).addReg(0);
3321 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), NextCountReg)
3322 .addReg(ThisCountReg).addImm(-1);
3323 BuildMI(MBB, DL, TII->get(SystemZ::CGHI))
3324 .addReg(NextCountReg).addImm(0);
3325 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3326 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3328 MBB->addSuccessor(LoopMBB);
3329 MBB->addSuccessor(DoneMBB);
3331 DestBase = MachineOperand::CreateReg(NextDestReg, false);
3332 SrcBase = MachineOperand::CreateReg(NextSrcReg, false);
3336 // Handle any remaining bytes with straight-line code.
3337 while (Length > 0) {
3338 uint64_t ThisLength = std::min(Length, uint64_t(256));
3339 // The previous iteration might have created out-of-range displacements.
3340 // Apply them using LAY if so.
3341 if (!isUInt<12>(DestDisp)) {
3342 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3343 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3344 .addOperand(DestBase).addImm(DestDisp).addReg(0);
3345 DestBase = MachineOperand::CreateReg(Reg, false);
3348 if (!isUInt<12>(SrcDisp)) {
3349 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3350 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3351 .addOperand(SrcBase).addImm(SrcDisp).addReg(0);
3352 SrcBase = MachineOperand::CreateReg(Reg, false);
3355 BuildMI(*MBB, MI, DL, TII->get(Opcode))
3356 .addOperand(DestBase).addImm(DestDisp).addImm(ThisLength)
3357 .addOperand(SrcBase).addImm(SrcDisp);
3358 DestDisp += ThisLength;
3359 SrcDisp += ThisLength;
3360 Length -= ThisLength;
3361 // If there's another CLC to go, branch to the end if a difference
3363 if (EndMBB && Length > 0) {
3364 MachineBasicBlock *NextMBB = splitBlockBefore(MI, MBB);
3365 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3366 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3368 MBB->addSuccessor(EndMBB);
3369 MBB->addSuccessor(NextMBB);
3374 MBB->addSuccessor(EndMBB);
3376 MBB->addLiveIn(SystemZ::CC);
3379 MI->eraseFromParent();
3383 // Decompose string pseudo-instruction MI into a loop that continually performs
3384 // Opcode until CC != 3.
3386 SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
3387 MachineBasicBlock *MBB,
3388 unsigned Opcode) const {
3389 MachineFunction &MF = *MBB->getParent();
3390 const SystemZInstrInfo *TII =
3391 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
3392 MachineRegisterInfo &MRI = MF.getRegInfo();
3393 DebugLoc DL = MI->getDebugLoc();
3395 uint64_t End1Reg = MI->getOperand(0).getReg();
3396 uint64_t Start1Reg = MI->getOperand(1).getReg();
3397 uint64_t Start2Reg = MI->getOperand(2).getReg();
3398 uint64_t CharReg = MI->getOperand(3).getReg();
3400 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass;
3401 uint64_t This1Reg = MRI.createVirtualRegister(RC);
3402 uint64_t This2Reg = MRI.createVirtualRegister(RC);
3403 uint64_t End2Reg = MRI.createVirtualRegister(RC);
3405 MachineBasicBlock *StartMBB = MBB;
3406 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3407 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3410 // # fall through to LoopMMB
3411 MBB->addSuccessor(LoopMBB);
3414 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ]
3415 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ]
3417 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0L
3419 // # fall through to DoneMMB
3421 // The load of R0L can be hoisted by post-RA LICM.
3424 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg)
3425 .addReg(Start1Reg).addMBB(StartMBB)
3426 .addReg(End1Reg).addMBB(LoopMBB);
3427 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg)
3428 .addReg(Start2Reg).addMBB(StartMBB)
3429 .addReg(End2Reg).addMBB(LoopMBB);
3430 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0L).addReg(CharReg);
3431 BuildMI(MBB, DL, TII->get(Opcode))
3432 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define)
3433 .addReg(This1Reg).addReg(This2Reg);
3434 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3435 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB);
3436 MBB->addSuccessor(LoopMBB);
3437 MBB->addSuccessor(DoneMBB);
3439 DoneMBB->addLiveIn(SystemZ::CC);
3441 MI->eraseFromParent();
3445 MachineBasicBlock *SystemZTargetLowering::
3446 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
3447 switch (MI->getOpcode()) {
3448 case SystemZ::Select32Mux:
3449 case SystemZ::Select32:
3450 case SystemZ::SelectF32:
3451 case SystemZ::Select64:
3452 case SystemZ::SelectF64:
3453 case SystemZ::SelectF128:
3454 return emitSelect(MI, MBB);
3456 case SystemZ::CondStore8Mux:
3457 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, false);
3458 case SystemZ::CondStore8MuxInv:
3459 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, true);
3460 case SystemZ::CondStore16Mux:
3461 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, false);
3462 case SystemZ::CondStore16MuxInv:
3463 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, true);
3464 case SystemZ::CondStore8:
3465 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
3466 case SystemZ::CondStore8Inv:
3467 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
3468 case SystemZ::CondStore16:
3469 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
3470 case SystemZ::CondStore16Inv:
3471 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
3472 case SystemZ::CondStore32:
3473 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
3474 case SystemZ::CondStore32Inv:
3475 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
3476 case SystemZ::CondStore64:
3477 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
3478 case SystemZ::CondStore64Inv:
3479 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
3480 case SystemZ::CondStoreF32:
3481 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
3482 case SystemZ::CondStoreF32Inv:
3483 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
3484 case SystemZ::CondStoreF64:
3485 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
3486 case SystemZ::CondStoreF64Inv:
3487 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
3489 case SystemZ::AEXT128_64:
3490 return emitExt128(MI, MBB, false, SystemZ::subreg_l64);
3491 case SystemZ::ZEXT128_32:
3492 return emitExt128(MI, MBB, true, SystemZ::subreg_l32);
3493 case SystemZ::ZEXT128_64:
3494 return emitExt128(MI, MBB, true, SystemZ::subreg_l64);
3496 case SystemZ::ATOMIC_SWAPW:
3497 return emitAtomicLoadBinary(MI, MBB, 0, 0);
3498 case SystemZ::ATOMIC_SWAP_32:
3499 return emitAtomicLoadBinary(MI, MBB, 0, 32);
3500 case SystemZ::ATOMIC_SWAP_64:
3501 return emitAtomicLoadBinary(MI, MBB, 0, 64);
3503 case SystemZ::ATOMIC_LOADW_AR:
3504 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
3505 case SystemZ::ATOMIC_LOADW_AFI:
3506 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
3507 case SystemZ::ATOMIC_LOAD_AR:
3508 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
3509 case SystemZ::ATOMIC_LOAD_AHI:
3510 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
3511 case SystemZ::ATOMIC_LOAD_AFI:
3512 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
3513 case SystemZ::ATOMIC_LOAD_AGR:
3514 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
3515 case SystemZ::ATOMIC_LOAD_AGHI:
3516 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
3517 case SystemZ::ATOMIC_LOAD_AGFI:
3518 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
3520 case SystemZ::ATOMIC_LOADW_SR:
3521 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
3522 case SystemZ::ATOMIC_LOAD_SR:
3523 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
3524 case SystemZ::ATOMIC_LOAD_SGR:
3525 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
3527 case SystemZ::ATOMIC_LOADW_NR:
3528 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
3529 case SystemZ::ATOMIC_LOADW_NILH:
3530 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0);
3531 case SystemZ::ATOMIC_LOAD_NR:
3532 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
3533 case SystemZ::ATOMIC_LOAD_NILL:
3534 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32);
3535 case SystemZ::ATOMIC_LOAD_NILH:
3536 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32);
3537 case SystemZ::ATOMIC_LOAD_NILF:
3538 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32);
3539 case SystemZ::ATOMIC_LOAD_NGR:
3540 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
3541 case SystemZ::ATOMIC_LOAD_NILL64:
3542 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64);
3543 case SystemZ::ATOMIC_LOAD_NILH64:
3544 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64);
3545 case SystemZ::ATOMIC_LOAD_NIHL64:
3546 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64);
3547 case SystemZ::ATOMIC_LOAD_NIHH64:
3548 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64);
3549 case SystemZ::ATOMIC_LOAD_NILF64:
3550 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64);
3551 case SystemZ::ATOMIC_LOAD_NIHF64:
3552 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64);
3554 case SystemZ::ATOMIC_LOADW_OR:
3555 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
3556 case SystemZ::ATOMIC_LOADW_OILH:
3557 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 0);
3558 case SystemZ::ATOMIC_LOAD_OR:
3559 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
3560 case SystemZ::ATOMIC_LOAD_OILL:
3561 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 32);
3562 case SystemZ::ATOMIC_LOAD_OILH:
3563 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 32);
3564 case SystemZ::ATOMIC_LOAD_OILF:
3565 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 32);
3566 case SystemZ::ATOMIC_LOAD_OGR:
3567 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
3568 case SystemZ::ATOMIC_LOAD_OILL64:
3569 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL64, 64);
3570 case SystemZ::ATOMIC_LOAD_OILH64:
3571 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH64, 64);
3572 case SystemZ::ATOMIC_LOAD_OIHL64:
3573 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL64, 64);
3574 case SystemZ::ATOMIC_LOAD_OIHH64:
3575 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH64, 64);
3576 case SystemZ::ATOMIC_LOAD_OILF64:
3577 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF64, 64);
3578 case SystemZ::ATOMIC_LOAD_OIHF64:
3579 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF64, 64);
3581 case SystemZ::ATOMIC_LOADW_XR:
3582 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
3583 case SystemZ::ATOMIC_LOADW_XILF:
3584 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 0);
3585 case SystemZ::ATOMIC_LOAD_XR:
3586 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
3587 case SystemZ::ATOMIC_LOAD_XILF:
3588 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 32);
3589 case SystemZ::ATOMIC_LOAD_XGR:
3590 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
3591 case SystemZ::ATOMIC_LOAD_XILF64:
3592 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF64, 64);
3593 case SystemZ::ATOMIC_LOAD_XIHF64:
3594 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF64, 64);
3596 case SystemZ::ATOMIC_LOADW_NRi:
3597 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
3598 case SystemZ::ATOMIC_LOADW_NILHi:
3599 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0, true);
3600 case SystemZ::ATOMIC_LOAD_NRi:
3601 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
3602 case SystemZ::ATOMIC_LOAD_NILLi:
3603 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32, true);
3604 case SystemZ::ATOMIC_LOAD_NILHi:
3605 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32, true);
3606 case SystemZ::ATOMIC_LOAD_NILFi:
3607 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32, true);
3608 case SystemZ::ATOMIC_LOAD_NGRi:
3609 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
3610 case SystemZ::ATOMIC_LOAD_NILL64i:
3611 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64, true);
3612 case SystemZ::ATOMIC_LOAD_NILH64i:
3613 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64, true);
3614 case SystemZ::ATOMIC_LOAD_NIHL64i:
3615 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64, true);
3616 case SystemZ::ATOMIC_LOAD_NIHH64i:
3617 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64, true);
3618 case SystemZ::ATOMIC_LOAD_NILF64i:
3619 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64, true);
3620 case SystemZ::ATOMIC_LOAD_NIHF64i:
3621 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64, true);
3623 case SystemZ::ATOMIC_LOADW_MIN:
3624 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3625 SystemZ::CCMASK_CMP_LE, 0);
3626 case SystemZ::ATOMIC_LOAD_MIN_32:
3627 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3628 SystemZ::CCMASK_CMP_LE, 32);
3629 case SystemZ::ATOMIC_LOAD_MIN_64:
3630 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3631 SystemZ::CCMASK_CMP_LE, 64);
3633 case SystemZ::ATOMIC_LOADW_MAX:
3634 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3635 SystemZ::CCMASK_CMP_GE, 0);
3636 case SystemZ::ATOMIC_LOAD_MAX_32:
3637 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3638 SystemZ::CCMASK_CMP_GE, 32);
3639 case SystemZ::ATOMIC_LOAD_MAX_64:
3640 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3641 SystemZ::CCMASK_CMP_GE, 64);
3643 case SystemZ::ATOMIC_LOADW_UMIN:
3644 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3645 SystemZ::CCMASK_CMP_LE, 0);
3646 case SystemZ::ATOMIC_LOAD_UMIN_32:
3647 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3648 SystemZ::CCMASK_CMP_LE, 32);
3649 case SystemZ::ATOMIC_LOAD_UMIN_64:
3650 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3651 SystemZ::CCMASK_CMP_LE, 64);
3653 case SystemZ::ATOMIC_LOADW_UMAX:
3654 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3655 SystemZ::CCMASK_CMP_GE, 0);
3656 case SystemZ::ATOMIC_LOAD_UMAX_32:
3657 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3658 SystemZ::CCMASK_CMP_GE, 32);
3659 case SystemZ::ATOMIC_LOAD_UMAX_64:
3660 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3661 SystemZ::CCMASK_CMP_GE, 64);
3663 case SystemZ::ATOMIC_CMP_SWAPW:
3664 return emitAtomicCmpSwapW(MI, MBB);
3665 case SystemZ::MVCSequence:
3666 case SystemZ::MVCLoop:
3667 return emitMemMemWrapper(MI, MBB, SystemZ::MVC);
3668 case SystemZ::NCSequence:
3669 case SystemZ::NCLoop:
3670 return emitMemMemWrapper(MI, MBB, SystemZ::NC);
3671 case SystemZ::OCSequence:
3672 case SystemZ::OCLoop:
3673 return emitMemMemWrapper(MI, MBB, SystemZ::OC);
3674 case SystemZ::XCSequence:
3675 case SystemZ::XCLoop:
3676 return emitMemMemWrapper(MI, MBB, SystemZ::XC);
3677 case SystemZ::CLCSequence:
3678 case SystemZ::CLCLoop:
3679 return emitMemMemWrapper(MI, MBB, SystemZ::CLC);
3680 case SystemZ::CLSTLoop:
3681 return emitStringWrapper(MI, MBB, SystemZ::CLST);
3682 case SystemZ::MVSTLoop:
3683 return emitStringWrapper(MI, MBB, SystemZ::MVST);
3684 case SystemZ::SRSTLoop:
3685 return emitStringWrapper(MI, MBB, SystemZ::SRST);
3687 llvm_unreachable("Unexpected instr type to insert");