1 //===-- SystemZISelLowering.cpp - SystemZ DAG lowering implementation -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SystemZTargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "systemz-lower"
16 #include "SystemZISelLowering.h"
17 #include "SystemZCallingConv.h"
18 #include "SystemZConstantPoolValue.h"
19 #include "SystemZMachineFunctionInfo.h"
20 #include "SystemZTargetMachine.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 // Represents a sequence for extracting a 0/1 value from an IPM result:
31 // (((X ^ XORValue) + AddValue) >> Bit)
32 struct IPMConversion {
33 IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
34 : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
41 // Represents information about a comparison.
43 Comparison(SDValue Op0In, SDValue Op1In)
44 : Op0(Op0In), Op1(Op1In), Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
46 // The operands to the comparison.
49 // The opcode that should be used to compare Op0 and Op1.
52 // A SystemZICMP value. Only used for integer comparisons.
55 // The mask of CC values that Opcode can produce.
58 // The mask of CC values for which the original condition is true.
61 } // end anonymous namespace
63 // Classify VT as either 32 or 64 bit.
64 static bool is32Bit(EVT VT) {
65 switch (VT.getSimpleVT().SimpleTy) {
71 llvm_unreachable("Unsupported type");
75 // Return a version of MachineOperand that can be safely used before the
77 static MachineOperand earlyUseOperand(MachineOperand Op) {
83 SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm)
84 : TargetLowering(tm, new TargetLoweringObjectFileELF()),
85 Subtarget(*tm.getSubtargetImpl()), TM(tm) {
86 MVT PtrVT = getPointerTy();
88 // Set up the register classes.
89 if (Subtarget.hasHighWord())
90 addRegisterClass(MVT::i32, &SystemZ::GRX32BitRegClass);
92 addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
93 addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
94 addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
95 addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
96 addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
98 // Compute derived properties from the register classes
99 computeRegisterProperties();
101 // Set up special registers.
102 setExceptionPointerRegister(SystemZ::R6D);
103 setExceptionSelectorRegister(SystemZ::R7D);
104 setStackPointerRegisterToSaveRestore(SystemZ::R15D);
106 // TODO: It may be better to default to latency-oriented scheduling, however
107 // LLVM's current latency-oriented scheduler can't handle physreg definitions
108 // such as SystemZ has with CC, so set this to the register-pressure
109 // scheduler, because it can.
110 setSchedulingPreference(Sched::RegPressure);
112 setBooleanContents(ZeroOrOneBooleanContent);
113 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
115 // Instructions are strings of 2-byte aligned 2-byte values.
116 setMinFunctionAlignment(2);
118 // Handle operations that are handled in a similar way for all types.
119 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
120 I <= MVT::LAST_FP_VALUETYPE;
122 MVT VT = MVT::SimpleValueType(I);
123 if (isTypeLegal(VT)) {
124 // Lower SET_CC into an IPM-based sequence.
125 setOperationAction(ISD::SETCC, VT, Custom);
127 // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
128 setOperationAction(ISD::SELECT, VT, Expand);
130 // Lower SELECT_CC and BR_CC into separate comparisons and branches.
131 setOperationAction(ISD::SELECT_CC, VT, Custom);
132 setOperationAction(ISD::BR_CC, VT, Custom);
136 // Expand jump table branches as address arithmetic followed by an
138 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
140 // Expand BRCOND into a BR_CC (see above).
141 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
143 // Handle integer types.
144 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
145 I <= MVT::LAST_INTEGER_VALUETYPE;
147 MVT VT = MVT::SimpleValueType(I);
148 if (isTypeLegal(VT)) {
149 // Expand individual DIV and REMs into DIVREMs.
150 setOperationAction(ISD::SDIV, VT, Expand);
151 setOperationAction(ISD::UDIV, VT, Expand);
152 setOperationAction(ISD::SREM, VT, Expand);
153 setOperationAction(ISD::UREM, VT, Expand);
154 setOperationAction(ISD::SDIVREM, VT, Custom);
155 setOperationAction(ISD::UDIVREM, VT, Custom);
157 // Lower ATOMIC_LOAD and ATOMIC_STORE into normal volatile loads and
158 // stores, putting a serialization instruction after the stores.
159 setOperationAction(ISD::ATOMIC_LOAD, VT, Custom);
160 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
162 // Lower ATOMIC_LOAD_SUB into ATOMIC_LOAD_ADD if LAA and LAAG are
163 // available, or if the operand is constant.
164 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
166 // No special instructions for these.
167 setOperationAction(ISD::CTPOP, VT, Expand);
168 setOperationAction(ISD::CTTZ, VT, Expand);
169 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
170 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
171 setOperationAction(ISD::ROTR, VT, Expand);
173 // Use *MUL_LOHI where possible instead of MULH*.
174 setOperationAction(ISD::MULHS, VT, Expand);
175 setOperationAction(ISD::MULHU, VT, Expand);
176 setOperationAction(ISD::SMUL_LOHI, VT, Custom);
177 setOperationAction(ISD::UMUL_LOHI, VT, Custom);
179 // We have instructions for signed but not unsigned FP conversion.
180 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
184 // Type legalization will convert 8- and 16-bit atomic operations into
185 // forms that operate on i32s (but still keeping the original memory VT).
186 // Lower them into full i32 operations.
187 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
188 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
189 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
190 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
191 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
192 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
193 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
194 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
195 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
196 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
197 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
198 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
200 // We have instructions for signed but not unsigned FP conversion.
201 // Handle unsigned 32-bit types as signed 64-bit types.
202 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
203 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
205 // We have native support for a 64-bit CTLZ, via FLOGR.
206 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
207 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
209 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
210 setOperationAction(ISD::OR, MVT::i64, Custom);
212 // Give LowerOperation the chance to optimize SIGN_EXTEND sequences.
213 setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
215 // FIXME: Can we support these natively?
216 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
217 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
218 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
220 // We have native instructions for i8, i16 and i32 extensions, but not i1.
221 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
222 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
223 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
224 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
226 // Handle the various types of symbolic address.
227 setOperationAction(ISD::ConstantPool, PtrVT, Custom);
228 setOperationAction(ISD::GlobalAddress, PtrVT, Custom);
229 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
230 setOperationAction(ISD::BlockAddress, PtrVT, Custom);
231 setOperationAction(ISD::JumpTable, PtrVT, Custom);
233 // We need to handle dynamic allocations specially because of the
234 // 160-byte area at the bottom of the stack.
235 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
237 // Use custom expanders so that we can force the function to use
239 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
240 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
242 // Handle prefetches with PFD or PFDRL.
243 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
245 // Handle floating-point types.
246 for (unsigned I = MVT::FIRST_FP_VALUETYPE;
247 I <= MVT::LAST_FP_VALUETYPE;
249 MVT VT = MVT::SimpleValueType(I);
250 if (isTypeLegal(VT)) {
251 // We can use FI for FRINT.
252 setOperationAction(ISD::FRINT, VT, Legal);
254 // We can use the extended form of FI for other rounding operations.
255 if (Subtarget.hasFPExtension()) {
256 setOperationAction(ISD::FNEARBYINT, VT, Legal);
257 setOperationAction(ISD::FFLOOR, VT, Legal);
258 setOperationAction(ISD::FCEIL, VT, Legal);
259 setOperationAction(ISD::FTRUNC, VT, Legal);
260 setOperationAction(ISD::FROUND, VT, Legal);
263 // No special instructions for these.
264 setOperationAction(ISD::FSIN, VT, Expand);
265 setOperationAction(ISD::FCOS, VT, Expand);
266 setOperationAction(ISD::FREM, VT, Expand);
270 // We have fused multiply-addition for f32 and f64 but not f128.
271 setOperationAction(ISD::FMA, MVT::f32, Legal);
272 setOperationAction(ISD::FMA, MVT::f64, Legal);
273 setOperationAction(ISD::FMA, MVT::f128, Expand);
275 // Needed so that we don't try to implement f128 constant loads using
276 // a load-and-extend of a f80 constant (in cases where the constant
277 // would fit in an f80).
278 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
280 // Floating-point truncation and stores need to be done separately.
281 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
282 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
283 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
285 // We have 64-bit FPR<->GPR moves, but need special handling for
287 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
288 setOperationAction(ISD::BITCAST, MVT::f32, Custom);
290 // VASTART and VACOPY need to deal with the SystemZ-specific varargs
291 // structure, but VAEND is a no-op.
292 setOperationAction(ISD::VASTART, MVT::Other, Custom);
293 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
294 setOperationAction(ISD::VAEND, MVT::Other, Expand);
296 // We want to use MVC in preference to even a single load/store pair.
297 MaxStoresPerMemcpy = 0;
298 MaxStoresPerMemcpyOptSize = 0;
300 // The main memset sequence is a byte store followed by an MVC.
301 // Two STC or MV..I stores win over that, but the kind of fused stores
302 // generated by target-independent code don't when the byte value is
303 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
304 // than "STC;MVC". Handle the choice in target-specific code instead.
305 MaxStoresPerMemset = 0;
306 MaxStoresPerMemsetOptSize = 0;
309 EVT SystemZTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
312 return VT.changeVectorElementTypeToInteger();
315 bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
316 VT = VT.getScalarType();
321 switch (VT.getSimpleVT().SimpleTy) {
334 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
335 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
336 return Imm.isZero() || Imm.isNegZero();
339 bool SystemZTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
342 // Unaligned accesses should never be slower than the expanded version.
343 // We check specifically for aligned accesses in the few cases where
344 // they are required.
350 bool SystemZTargetLowering::isLegalAddressingMode(const AddrMode &AM,
352 // Punt on globals for now, although they can be used in limited
353 // RELATIVE LONG cases.
357 // Require a 20-bit signed offset.
358 if (!isInt<20>(AM.BaseOffs))
361 // Indexing is OK but no scale factor can be applied.
362 return AM.Scale == 0 || AM.Scale == 1;
365 bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
366 if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
368 unsigned FromBits = FromType->getPrimitiveSizeInBits();
369 unsigned ToBits = ToType->getPrimitiveSizeInBits();
370 return FromBits > ToBits;
373 bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
374 if (!FromVT.isInteger() || !ToVT.isInteger())
376 unsigned FromBits = FromVT.getSizeInBits();
377 unsigned ToBits = ToVT.getSizeInBits();
378 return FromBits > ToBits;
381 //===----------------------------------------------------------------------===//
382 // Inline asm support
383 //===----------------------------------------------------------------------===//
385 TargetLowering::ConstraintType
386 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
387 if (Constraint.size() == 1) {
388 switch (Constraint[0]) {
389 case 'a': // Address register
390 case 'd': // Data register (equivalent to 'r')
391 case 'f': // Floating-point register
392 case 'h': // High-part register
393 case 'r': // General-purpose register
394 return C_RegisterClass;
396 case 'Q': // Memory with base and unsigned 12-bit displacement
397 case 'R': // Likewise, plus an index
398 case 'S': // Memory with base and signed 20-bit displacement
399 case 'T': // Likewise, plus an index
400 case 'm': // Equivalent to 'T'.
403 case 'I': // Unsigned 8-bit constant
404 case 'J': // Unsigned 12-bit constant
405 case 'K': // Signed 16-bit constant
406 case 'L': // Signed 20-bit displacement (on all targets we support)
407 case 'M': // 0x7fffffff
414 return TargetLowering::getConstraintType(Constraint);
417 TargetLowering::ConstraintWeight SystemZTargetLowering::
418 getSingleConstraintMatchWeight(AsmOperandInfo &info,
419 const char *constraint) const {
420 ConstraintWeight weight = CW_Invalid;
421 Value *CallOperandVal = info.CallOperandVal;
422 // If we don't have a value, we can't do a match,
423 // but allow it at the lowest weight.
424 if (CallOperandVal == NULL)
426 Type *type = CallOperandVal->getType();
427 // Look at the constraint type.
428 switch (*constraint) {
430 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
433 case 'a': // Address register
434 case 'd': // Data register (equivalent to 'r')
435 case 'h': // High-part register
436 case 'r': // General-purpose register
437 if (CallOperandVal->getType()->isIntegerTy())
438 weight = CW_Register;
441 case 'f': // Floating-point register
442 if (type->isFloatingPointTy())
443 weight = CW_Register;
446 case 'I': // Unsigned 8-bit constant
447 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
448 if (isUInt<8>(C->getZExtValue()))
449 weight = CW_Constant;
452 case 'J': // Unsigned 12-bit constant
453 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
454 if (isUInt<12>(C->getZExtValue()))
455 weight = CW_Constant;
458 case 'K': // Signed 16-bit constant
459 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
460 if (isInt<16>(C->getSExtValue()))
461 weight = CW_Constant;
464 case 'L': // Signed 20-bit displacement (on all targets we support)
465 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
466 if (isInt<20>(C->getSExtValue()))
467 weight = CW_Constant;
470 case 'M': // 0x7fffffff
471 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
472 if (C->getZExtValue() == 0x7fffffff)
473 weight = CW_Constant;
479 // Parse a "{tNNN}" register constraint for which the register type "t"
480 // has already been verified. MC is the class associated with "t" and
481 // Map maps 0-based register numbers to LLVM register numbers.
482 static std::pair<unsigned, const TargetRegisterClass *>
483 parseRegisterNumber(const std::string &Constraint,
484 const TargetRegisterClass *RC, const unsigned *Map) {
485 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
486 if (isdigit(Constraint[2])) {
487 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2);
488 unsigned Index = atoi(Suffix.c_str());
489 if (Index < 16 && Map[Index])
490 return std::make_pair(Map[Index], RC);
492 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
495 std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering::
496 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const {
497 if (Constraint.size() == 1) {
498 // GCC Constraint Letters
499 switch (Constraint[0]) {
501 case 'd': // Data register (equivalent to 'r')
502 case 'r': // General-purpose register
504 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
505 else if (VT == MVT::i128)
506 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
507 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
509 case 'a': // Address register
511 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
512 else if (VT == MVT::i128)
513 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
514 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
516 case 'h': // High-part register (an LLVM extension)
517 return std::make_pair(0U, &SystemZ::GRH32BitRegClass);
519 case 'f': // Floating-point register
521 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
522 else if (VT == MVT::f128)
523 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
524 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
527 if (Constraint[0] == '{') {
528 // We need to override the default register parsing for GPRs and FPRs
529 // because the interpretation depends on VT. The internal names of
530 // the registers are also different from the external names
531 // (F0D and F0S instead of F0, etc.).
532 if (Constraint[1] == 'r') {
534 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
535 SystemZMC::GR32Regs);
537 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
538 SystemZMC::GR128Regs);
539 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
540 SystemZMC::GR64Regs);
542 if (Constraint[1] == 'f') {
544 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
545 SystemZMC::FP32Regs);
547 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
548 SystemZMC::FP128Regs);
549 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
550 SystemZMC::FP64Regs);
553 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
556 void SystemZTargetLowering::
557 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
558 std::vector<SDValue> &Ops,
559 SelectionDAG &DAG) const {
560 // Only support length 1 constraints for now.
561 if (Constraint.length() == 1) {
562 switch (Constraint[0]) {
563 case 'I': // Unsigned 8-bit constant
564 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
565 if (isUInt<8>(C->getZExtValue()))
566 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
570 case 'J': // Unsigned 12-bit constant
571 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
572 if (isUInt<12>(C->getZExtValue()))
573 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
577 case 'K': // Signed 16-bit constant
578 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
579 if (isInt<16>(C->getSExtValue()))
580 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
584 case 'L': // Signed 20-bit displacement (on all targets we support)
585 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
586 if (isInt<20>(C->getSExtValue()))
587 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
591 case 'M': // 0x7fffffff
592 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
593 if (C->getZExtValue() == 0x7fffffff)
594 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
599 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
602 //===----------------------------------------------------------------------===//
603 // Calling conventions
604 //===----------------------------------------------------------------------===//
606 #include "SystemZGenCallingConv.inc"
608 bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
609 Type *ToType) const {
610 return isTruncateFree(FromType, ToType);
613 bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
614 if (!CI->isTailCall())
619 // Value is a value that has been passed to us in the location described by VA
620 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
621 // any loads onto Chain.
622 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
623 CCValAssign &VA, SDValue Chain,
625 // If the argument has been promoted from a smaller type, insert an
626 // assertion to capture this.
627 if (VA.getLocInfo() == CCValAssign::SExt)
628 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
629 DAG.getValueType(VA.getValVT()));
630 else if (VA.getLocInfo() == CCValAssign::ZExt)
631 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
632 DAG.getValueType(VA.getValVT()));
635 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
636 else if (VA.getLocInfo() == CCValAssign::Indirect)
637 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
638 MachinePointerInfo(), false, false, false, 0);
640 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
644 // Value is a value of type VA.getValVT() that we need to copy into
645 // the location described by VA. Return a copy of Value converted to
646 // VA.getValVT(). The caller is responsible for handling indirect values.
647 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
648 CCValAssign &VA, SDValue Value) {
649 switch (VA.getLocInfo()) {
650 case CCValAssign::SExt:
651 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
652 case CCValAssign::ZExt:
653 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
654 case CCValAssign::AExt:
655 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
656 case CCValAssign::Full:
659 llvm_unreachable("Unhandled getLocInfo()");
663 SDValue SystemZTargetLowering::
664 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
665 const SmallVectorImpl<ISD::InputArg> &Ins,
666 SDLoc DL, SelectionDAG &DAG,
667 SmallVectorImpl<SDValue> &InVals) const {
668 MachineFunction &MF = DAG.getMachineFunction();
669 MachineFrameInfo *MFI = MF.getFrameInfo();
670 MachineRegisterInfo &MRI = MF.getRegInfo();
671 SystemZMachineFunctionInfo *FuncInfo =
672 MF.getInfo<SystemZMachineFunctionInfo>();
673 const SystemZFrameLowering *TFL =
674 static_cast<const SystemZFrameLowering *>(TM.getFrameLowering());
676 // Assign locations to all of the incoming arguments.
677 SmallVector<CCValAssign, 16> ArgLocs;
678 CCState CCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
679 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
681 unsigned NumFixedGPRs = 0;
682 unsigned NumFixedFPRs = 0;
683 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
685 CCValAssign &VA = ArgLocs[I];
686 EVT LocVT = VA.getLocVT();
688 // Arguments passed in registers
689 const TargetRegisterClass *RC;
690 switch (LocVT.getSimpleVT().SimpleTy) {
692 // Integers smaller than i64 should be promoted to i64.
693 llvm_unreachable("Unexpected argument type");
696 RC = &SystemZ::GR32BitRegClass;
700 RC = &SystemZ::GR64BitRegClass;
704 RC = &SystemZ::FP32BitRegClass;
708 RC = &SystemZ::FP64BitRegClass;
712 unsigned VReg = MRI.createVirtualRegister(RC);
713 MRI.addLiveIn(VA.getLocReg(), VReg);
714 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
716 assert(VA.isMemLoc() && "Argument not register or memory");
718 // Create the frame index object for this incoming parameter.
719 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
720 VA.getLocMemOffset(), true);
722 // Create the SelectionDAG nodes corresponding to a load
723 // from this parameter. Unpromoted ints and floats are
724 // passed as right-justified 8-byte values.
725 EVT PtrVT = getPointerTy();
726 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
727 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
728 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4));
729 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
730 MachinePointerInfo::getFixedStack(FI),
731 false, false, false, 0);
734 // Convert the value of the argument register into the value that's
736 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
740 // Save the number of non-varargs registers for later use by va_start, etc.
741 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
742 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
744 // Likewise the address (in the form of a frame index) of where the
745 // first stack vararg would be. The 1-byte size here is arbitrary.
746 int64_t StackSize = CCInfo.getNextStackOffset();
747 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
749 // ...and a similar frame index for the caller-allocated save area
750 // that will be used to store the incoming registers.
751 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
752 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
753 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
755 // Store the FPR varargs in the reserved frame slots. (We store the
756 // GPRs as part of the prologue.)
757 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
758 SDValue MemOps[SystemZ::NumArgFPRs];
759 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
760 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
761 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
762 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
763 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
764 &SystemZ::FP64BitRegClass);
765 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
766 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
767 MachinePointerInfo::getFixedStack(FI),
771 // Join the stores, which are independent of one another.
772 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
773 &MemOps[NumFixedFPRs],
774 SystemZ::NumArgFPRs - NumFixedFPRs);
781 static bool canUseSiblingCall(CCState ArgCCInfo,
782 SmallVectorImpl<CCValAssign> &ArgLocs) {
783 // Punt if there are any indirect or stack arguments, or if the call
784 // needs the call-saved argument register R6.
785 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
786 CCValAssign &VA = ArgLocs[I];
787 if (VA.getLocInfo() == CCValAssign::Indirect)
791 unsigned Reg = VA.getLocReg();
792 if (Reg == SystemZ::R6H || Reg == SystemZ::R6L || Reg == SystemZ::R6D)
799 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
800 SmallVectorImpl<SDValue> &InVals) const {
801 SelectionDAG &DAG = CLI.DAG;
803 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
804 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
805 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
806 SDValue Chain = CLI.Chain;
807 SDValue Callee = CLI.Callee;
808 bool &IsTailCall = CLI.IsTailCall;
809 CallingConv::ID CallConv = CLI.CallConv;
810 bool IsVarArg = CLI.IsVarArg;
811 MachineFunction &MF = DAG.getMachineFunction();
812 EVT PtrVT = getPointerTy();
814 // Analyze the operands of the call, assigning locations to each operand.
815 SmallVector<CCValAssign, 16> ArgLocs;
816 CCState ArgCCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
817 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
819 // We don't support GuaranteedTailCallOpt, only automatically-detected
821 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs))
824 // Get a count of how many bytes are to be pushed on the stack.
825 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
827 // Mark the start of the call.
829 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true),
832 // Copy argument values to their designated locations.
833 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
834 SmallVector<SDValue, 8> MemOpChains;
836 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
837 CCValAssign &VA = ArgLocs[I];
838 SDValue ArgValue = OutVals[I];
840 if (VA.getLocInfo() == CCValAssign::Indirect) {
841 // Store the argument in a stack slot and pass its address.
842 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
843 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
844 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
845 MachinePointerInfo::getFixedStack(FI),
847 ArgValue = SpillSlot;
849 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
852 // Queue up the argument copies and emit them at the end.
853 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
855 assert(VA.isMemLoc() && "Argument not register or memory");
857 // Work out the address of the stack slot. Unpromoted ints and
858 // floats are passed as right-justified 8-byte values.
859 if (!StackPtr.getNode())
860 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
861 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
862 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
864 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
865 DAG.getIntPtrConstant(Offset));
868 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
869 MachinePointerInfo(),
874 // Join the stores, which are independent of one another.
875 if (!MemOpChains.empty())
876 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
877 &MemOpChains[0], MemOpChains.size());
879 // Accept direct calls by converting symbolic call addresses to the
880 // associated Target* opcodes. Force %r1 to be used for indirect
883 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
884 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
885 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
886 } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
887 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
888 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
889 } else if (IsTailCall) {
890 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
891 Glue = Chain.getValue(1);
892 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
895 // Build a sequence of copy-to-reg nodes, chained and glued together.
896 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
897 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
898 RegsToPass[I].second, Glue);
899 Glue = Chain.getValue(1);
902 // The first call operand is the chain and the second is the target address.
903 SmallVector<SDValue, 8> Ops;
904 Ops.push_back(Chain);
905 Ops.push_back(Callee);
907 // Add argument registers to the end of the list so that they are
908 // known live into the call.
909 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
910 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
911 RegsToPass[I].second.getValueType()));
913 // Glue the call to the argument copies, if any.
918 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
920 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, &Ops[0], Ops.size());
921 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
922 Glue = Chain.getValue(1);
924 // Mark the end of the call, which is glued to the call itself.
925 Chain = DAG.getCALLSEQ_END(Chain,
926 DAG.getConstant(NumBytes, PtrVT, true),
927 DAG.getConstant(0, PtrVT, true),
929 Glue = Chain.getValue(1);
931 // Assign locations to each value returned by this call.
932 SmallVector<CCValAssign, 16> RetLocs;
933 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
934 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
936 // Copy all of the result registers out of their specified physreg.
937 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
938 CCValAssign &VA = RetLocs[I];
940 // Copy the value out, gluing the copy to the end of the call sequence.
941 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
942 VA.getLocVT(), Glue);
943 Chain = RetValue.getValue(1);
944 Glue = RetValue.getValue(2);
946 // Convert the value of the return register into the value that's
948 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
955 SystemZTargetLowering::LowerReturn(SDValue Chain,
956 CallingConv::ID CallConv, bool IsVarArg,
957 const SmallVectorImpl<ISD::OutputArg> &Outs,
958 const SmallVectorImpl<SDValue> &OutVals,
959 SDLoc DL, SelectionDAG &DAG) const {
960 MachineFunction &MF = DAG.getMachineFunction();
962 // Assign locations to each returned value.
963 SmallVector<CCValAssign, 16> RetLocs;
964 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
965 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
967 // Quick exit for void returns
969 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
971 // Copy the result values into the output registers.
973 SmallVector<SDValue, 4> RetOps;
974 RetOps.push_back(Chain);
975 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
976 CCValAssign &VA = RetLocs[I];
977 SDValue RetValue = OutVals[I];
979 // Make the return register live on exit.
980 assert(VA.isRegLoc() && "Can only return in registers!");
982 // Promote the value as required.
983 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
985 // Chain and glue the copies together.
986 unsigned Reg = VA.getLocReg();
987 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
988 Glue = Chain.getValue(1);
989 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
992 // Update chain and glue.
995 RetOps.push_back(Glue);
997 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other,
998 RetOps.data(), RetOps.size());
1001 SDValue SystemZTargetLowering::
1002 prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, SelectionDAG &DAG) const {
1003 return DAG.getNode(SystemZISD::SERIALIZE, DL, MVT::Other, Chain);
1006 // CC is a comparison that will be implemented using an integer or
1007 // floating-point comparison. Return the condition code mask for
1008 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
1009 // unsigned comparisons and clear for signed ones. In the floating-point
1010 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
1011 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
1013 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
1014 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
1015 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
1019 llvm_unreachable("Invalid integer condition!");
1028 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
1029 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
1034 // Return a sequence for getting a 1 from an IPM result when CC has a
1035 // value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
1036 // The handling of CC values outside CCValid doesn't matter.
1037 static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
1038 // Deal with cases where the result can be taken directly from a bit
1039 // of the IPM result.
1040 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
1041 return IPMConversion(0, 0, SystemZ::IPM_CC);
1042 if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
1043 return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
1045 // Deal with cases where we can add a value to force the sign bit
1046 // to contain the right value. Putting the bit in 31 means we can
1047 // use SRL rather than RISBG(L), and also makes it easier to get a
1048 // 0/-1 value, so it has priority over the other tests below.
1050 // These sequences rely on the fact that the upper two bits of the
1051 // IPM result are zero.
1052 uint64_t TopBit = uint64_t(1) << 31;
1053 if (CCMask == (CCValid & SystemZ::CCMASK_0))
1054 return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
1055 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
1056 return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
1057 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1059 | SystemZ::CCMASK_2)))
1060 return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
1061 if (CCMask == (CCValid & SystemZ::CCMASK_3))
1062 return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
1063 if (CCMask == (CCValid & (SystemZ::CCMASK_1
1065 | SystemZ::CCMASK_3)))
1066 return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
1068 // Next try inverting the value and testing a bit. 0/1 could be
1069 // handled this way too, but we dealt with that case above.
1070 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
1071 return IPMConversion(-1, 0, SystemZ::IPM_CC);
1073 // Handle cases where adding a value forces a non-sign bit to contain
1075 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
1076 return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
1077 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
1078 return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
1080 // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
1081 // can be done by inverting the low CC bit and applying one of the
1082 // sign-based extractions above.
1083 if (CCMask == (CCValid & SystemZ::CCMASK_1))
1084 return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
1085 if (CCMask == (CCValid & SystemZ::CCMASK_2))
1086 return IPMConversion(1 << SystemZ::IPM_CC,
1087 TopBit - (3 << SystemZ::IPM_CC), 31);
1088 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1090 | SystemZ::CCMASK_3)))
1091 return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
1092 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1094 | SystemZ::CCMASK_3)))
1095 return IPMConversion(1 << SystemZ::IPM_CC,
1096 TopBit - (1 << SystemZ::IPM_CC), 31);
1098 llvm_unreachable("Unexpected CC combination");
1101 // If C can be converted to a comparison against zero, adjust the operands
1103 static void adjustZeroCmp(SelectionDAG &DAG, Comparison &C) {
1104 if (C.ICmpType == SystemZICMP::UnsignedOnly)
1107 ConstantSDNode *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
1111 int64_t Value = ConstOp1->getSExtValue();
1112 if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
1113 (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
1114 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
1115 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
1116 C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1117 C.Op1 = DAG.getConstant(0, C.Op1.getValueType());
1121 // If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
1122 // adjust the operands as necessary.
1123 static void adjustSubwordCmp(SelectionDAG &DAG, Comparison &C) {
1124 // For us to make any changes, it must a comparison between a single-use
1125 // load and a constant.
1126 if (!C.Op0.hasOneUse() ||
1127 C.Op0.getOpcode() != ISD::LOAD ||
1128 C.Op1.getOpcode() != ISD::Constant)
1131 // We must have an 8- or 16-bit load.
1132 LoadSDNode *Load = cast<LoadSDNode>(C.Op0);
1133 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
1134 if (NumBits != 8 && NumBits != 16)
1137 // The load must be an extending one and the constant must be within the
1138 // range of the unextended value.
1139 ConstantSDNode *ConstOp1 = cast<ConstantSDNode>(C.Op1);
1140 uint64_t Value = ConstOp1->getZExtValue();
1141 uint64_t Mask = (1 << NumBits) - 1;
1142 if (Load->getExtensionType() == ISD::SEXTLOAD) {
1143 // Make sure that ConstOp1 is in range of C.Op0.
1144 int64_t SignedValue = ConstOp1->getSExtValue();
1145 if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
1147 if (C.ICmpType != SystemZICMP::SignedOnly) {
1148 // Unsigned comparison between two sign-extended values is equivalent
1149 // to unsigned comparison between two zero-extended values.
1151 } else if (NumBits == 8) {
1152 // Try to treat the comparison as unsigned, so that we can use CLI.
1153 // Adjust CCMask and Value as necessary.
1154 if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
1155 // Test whether the high bit of the byte is set.
1156 Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
1157 else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
1158 // Test whether the high bit of the byte is clear.
1159 Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
1161 // No instruction exists for this combination.
1163 C.ICmpType = SystemZICMP::UnsignedOnly;
1165 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
1168 assert(C.ICmpType == SystemZICMP::Any &&
1169 "Signedness shouldn't matter here.");
1173 // Make sure that the first operand is an i32 of the right extension type.
1174 ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
1177 if (C.Op0.getValueType() != MVT::i32 ||
1178 Load->getExtensionType() != ExtType)
1179 C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
1180 Load->getChain(), Load->getBasePtr(),
1181 Load->getPointerInfo(), Load->getMemoryVT(),
1182 Load->isVolatile(), Load->isNonTemporal(),
1183 Load->getAlignment());
1185 // Make sure that the second operand is an i32 with the right value.
1186 if (C.Op1.getValueType() != MVT::i32 ||
1187 Value != ConstOp1->getZExtValue())
1188 C.Op1 = DAG.getConstant(Value, MVT::i32);
1191 // Return true if Op is either an unextended load, or a load suitable
1192 // for integer register-memory comparisons of type ICmpType.
1193 static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
1194 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op.getNode());
1196 // There are no instructions to compare a register with a memory byte.
1197 if (Load->getMemoryVT() == MVT::i8)
1199 // Otherwise decide on extension type.
1200 switch (Load->getExtensionType()) {
1201 case ISD::NON_EXTLOAD:
1204 return ICmpType != SystemZICMP::UnsignedOnly;
1206 return ICmpType != SystemZICMP::SignedOnly;
1214 // Return true if it is better to swap the operands of C.
1215 static bool shouldSwapCmpOperands(const Comparison &C) {
1216 // Leave f128 comparisons alone, since they have no memory forms.
1217 if (C.Op0.getValueType() == MVT::f128)
1220 // Always keep a floating-point constant second, since comparisons with
1221 // zero can use LOAD TEST and comparisons with other constants make a
1222 // natural memory operand.
1223 if (isa<ConstantFPSDNode>(C.Op1))
1226 // Never swap comparisons with zero since there are many ways to optimize
1228 ConstantSDNode *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1229 if (ConstOp1 && ConstOp1->getZExtValue() == 0)
1232 // Also keep natural memory operands second if the loaded value is
1233 // only used here. Several comparisons have memory forms.
1234 if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
1237 // Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
1238 // In that case we generally prefer the memory to be second.
1239 if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
1240 // The only exceptions are when the second operand is a constant and
1241 // we can use things like CHHSI.
1244 // The unsigned memory-immediate instructions can handle 16-bit
1245 // unsigned integers.
1246 if (C.ICmpType != SystemZICMP::SignedOnly &&
1247 isUInt<16>(ConstOp1->getZExtValue()))
1249 // The signed memory-immediate instructions can handle 16-bit
1251 if (C.ICmpType != SystemZICMP::UnsignedOnly &&
1252 isInt<16>(ConstOp1->getSExtValue()))
1257 // Try to promote the use of CGFR and CLGFR.
1258 unsigned Opcode0 = C.Op0.getOpcode();
1259 if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
1261 if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
1263 if (C.ICmpType != SystemZICMP::SignedOnly &&
1264 Opcode0 == ISD::AND &&
1265 C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
1266 cast<ConstantSDNode>(C.Op0.getOperand(1))->getZExtValue() == 0xffffffff)
1272 // Return a version of comparison CC mask CCMask in which the LT and GT
1273 // actions are swapped.
1274 static unsigned reverseCCMask(unsigned CCMask) {
1275 return ((CCMask & SystemZ::CCMASK_CMP_EQ) |
1276 (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
1277 (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
1278 (CCMask & SystemZ::CCMASK_CMP_UO));
1281 // Check whether C tests for equality between X and Y and whether X - Y
1282 // or Y - X is also computed. In that case it's better to compare the
1283 // result of the subtraction against zero.
1284 static void adjustForSubtraction(SelectionDAG &DAG, Comparison &C) {
1285 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1286 C.CCMask == SystemZ::CCMASK_CMP_NE) {
1287 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1289 if (N->getOpcode() == ISD::SUB &&
1290 ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
1291 (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
1292 C.Op0 = SDValue(N, 0);
1293 C.Op1 = DAG.getConstant(0, N->getValueType(0));
1300 // Check whether C compares a floating-point value with zero and if that
1301 // floating-point value is also negated. In this case we can use the
1302 // negation to set CC, so avoiding separate LOAD AND TEST and
1303 // LOAD (NEGATIVE/COMPLEMENT) instructions.
1304 static void adjustForFNeg(Comparison &C) {
1305 ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
1306 if (C1 && C1->isZero()) {
1307 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1309 if (N->getOpcode() == ISD::FNEG) {
1310 C.Op0 = SDValue(N, 0);
1311 C.CCMask = reverseCCMask(C.CCMask);
1318 // Check whether C compares (shl X, 32) with 0 and whether X is
1319 // also sign-extended. In that case it is better to test the result
1320 // of the sign extension using LTGFR.
1322 // This case is important because InstCombine transforms a comparison
1323 // with (sext (trunc X)) into a comparison with (shl X, 32).
1324 static void adjustForLTGFR(Comparison &C) {
1325 // Check for a comparison between (shl X, 32) and 0.
1326 if (C.Op0.getOpcode() == ISD::SHL &&
1327 C.Op0.getValueType() == MVT::i64 &&
1328 C.Op1.getOpcode() == ISD::Constant &&
1329 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1330 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
1331 if (C1 && C1->getZExtValue() == 32) {
1332 SDValue ShlOp0 = C.Op0.getOperand(0);
1333 // See whether X has any SIGN_EXTEND_INREG uses.
1334 for (auto I = ShlOp0->use_begin(), E = ShlOp0->use_end(); I != E; ++I) {
1336 if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
1337 cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
1338 C.Op0 = SDValue(N, 0);
1346 // If C compares the truncation of an extending load, try to compare
1347 // the untruncated value instead. This exposes more opportunities to
1349 static void adjustICmpTruncate(SelectionDAG &DAG, Comparison &C) {
1350 if (C.Op0.getOpcode() == ISD::TRUNCATE &&
1351 C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
1352 C.Op1.getOpcode() == ISD::Constant &&
1353 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1354 LoadSDNode *L = cast<LoadSDNode>(C.Op0.getOperand(0));
1355 if (L->getMemoryVT().getStoreSizeInBits()
1356 <= C.Op0.getValueType().getSizeInBits()) {
1357 unsigned Type = L->getExtensionType();
1358 if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
1359 (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
1360 C.Op0 = C.Op0.getOperand(0);
1361 C.Op1 = DAG.getConstant(0, C.Op0.getValueType());
1367 // Return true if shift operation N has an in-range constant shift value.
1368 // Store it in ShiftVal if so.
1369 static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
1370 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
1374 uint64_t Amount = Shift->getZExtValue();
1375 if (Amount >= N.getValueType().getSizeInBits())
1382 // Check whether an AND with Mask is suitable for a TEST UNDER MASK
1383 // instruction and whether the CC value is descriptive enough to handle
1384 // a comparison of type Opcode between the AND result and CmpVal.
1385 // CCMask says which comparison result is being tested and BitSize is
1386 // the number of bits in the operands. If TEST UNDER MASK can be used,
1387 // return the corresponding CC mask, otherwise return 0.
1388 static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask,
1389 uint64_t Mask, uint64_t CmpVal,
1390 unsigned ICmpType) {
1391 assert(Mask != 0 && "ANDs with zero should have been removed by now");
1393 // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.
1394 if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&
1395 !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))
1398 // Work out the masks for the lowest and highest bits.
1399 unsigned HighShift = 63 - countLeadingZeros(Mask);
1400 uint64_t High = uint64_t(1) << HighShift;
1401 uint64_t Low = uint64_t(1) << countTrailingZeros(Mask);
1403 // Signed ordered comparisons are effectively unsigned if the sign
1405 bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly);
1407 // Check for equality comparisons with 0, or the equivalent.
1409 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1410 return SystemZ::CCMASK_TM_ALL_0;
1411 if (CCMask == SystemZ::CCMASK_CMP_NE)
1412 return SystemZ::CCMASK_TM_SOME_1;
1414 if (EffectivelyUnsigned && CmpVal <= Low) {
1415 if (CCMask == SystemZ::CCMASK_CMP_LT)
1416 return SystemZ::CCMASK_TM_ALL_0;
1417 if (CCMask == SystemZ::CCMASK_CMP_GE)
1418 return SystemZ::CCMASK_TM_SOME_1;
1420 if (EffectivelyUnsigned && CmpVal < Low) {
1421 if (CCMask == SystemZ::CCMASK_CMP_LE)
1422 return SystemZ::CCMASK_TM_ALL_0;
1423 if (CCMask == SystemZ::CCMASK_CMP_GT)
1424 return SystemZ::CCMASK_TM_SOME_1;
1427 // Check for equality comparisons with the mask, or the equivalent.
1428 if (CmpVal == Mask) {
1429 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1430 return SystemZ::CCMASK_TM_ALL_1;
1431 if (CCMask == SystemZ::CCMASK_CMP_NE)
1432 return SystemZ::CCMASK_TM_SOME_0;
1434 if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) {
1435 if (CCMask == SystemZ::CCMASK_CMP_GT)
1436 return SystemZ::CCMASK_TM_ALL_1;
1437 if (CCMask == SystemZ::CCMASK_CMP_LE)
1438 return SystemZ::CCMASK_TM_SOME_0;
1440 if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) {
1441 if (CCMask == SystemZ::CCMASK_CMP_GE)
1442 return SystemZ::CCMASK_TM_ALL_1;
1443 if (CCMask == SystemZ::CCMASK_CMP_LT)
1444 return SystemZ::CCMASK_TM_SOME_0;
1447 // Check for ordered comparisons with the top bit.
1448 if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) {
1449 if (CCMask == SystemZ::CCMASK_CMP_LE)
1450 return SystemZ::CCMASK_TM_MSB_0;
1451 if (CCMask == SystemZ::CCMASK_CMP_GT)
1452 return SystemZ::CCMASK_TM_MSB_1;
1454 if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) {
1455 if (CCMask == SystemZ::CCMASK_CMP_LT)
1456 return SystemZ::CCMASK_TM_MSB_0;
1457 if (CCMask == SystemZ::CCMASK_CMP_GE)
1458 return SystemZ::CCMASK_TM_MSB_1;
1461 // If there are just two bits, we can do equality checks for Low and High
1463 if (Mask == Low + High) {
1464 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low)
1465 return SystemZ::CCMASK_TM_MIXED_MSB_0;
1466 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low)
1467 return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY;
1468 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High)
1469 return SystemZ::CCMASK_TM_MIXED_MSB_1;
1470 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High)
1471 return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY;
1474 // Looks like we've exhausted our options.
1478 // See whether C can be implemented as a TEST UNDER MASK instruction.
1479 // Update the arguments with the TM version if so.
1480 static void adjustForTestUnderMask(SelectionDAG &DAG, Comparison &C) {
1481 // Check that we have a comparison with a constant.
1482 ConstantSDNode *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1485 uint64_t CmpVal = ConstOp1->getZExtValue();
1487 // Check whether the nonconstant input is an AND with a constant mask.
1490 ConstantSDNode *Mask = 0;
1491 if (C.Op0.getOpcode() == ISD::AND) {
1492 NewC.Op0 = C.Op0.getOperand(0);
1493 NewC.Op1 = C.Op0.getOperand(1);
1494 Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
1497 MaskVal = Mask->getZExtValue();
1499 // There is no instruction to compare with a 64-bit immediate
1500 // so use TMHH instead if possible. We need an unsigned ordered
1501 // comparison with an i64 immediate.
1502 if (NewC.Op0.getValueType() != MVT::i64 ||
1503 NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
1504 NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
1505 NewC.ICmpType == SystemZICMP::SignedOnly)
1507 // Convert LE and GT comparisons into LT and GE.
1508 if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
1509 NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
1510 if (CmpVal == uint64_t(-1))
1513 NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1515 // If the low N bits of Op1 are zero than the low N bits of Op0 can
1516 // be masked off without changing the result.
1517 MaskVal = -(CmpVal & -CmpVal);
1518 NewC.ICmpType = SystemZICMP::UnsignedOnly;
1521 // Check whether the combination of mask, comparison value and comparison
1522 // type are suitable.
1523 unsigned BitSize = NewC.Op0.getValueType().getSizeInBits();
1524 unsigned NewCCMask, ShiftVal;
1525 if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1526 NewC.Op0.getOpcode() == ISD::SHL &&
1527 isSimpleShift(NewC.Op0, ShiftVal) &&
1528 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1529 MaskVal >> ShiftVal,
1531 SystemZICMP::Any))) {
1532 NewC.Op0 = NewC.Op0.getOperand(0);
1533 MaskVal >>= ShiftVal;
1534 } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1535 NewC.Op0.getOpcode() == ISD::SRL &&
1536 isSimpleShift(NewC.Op0, ShiftVal) &&
1537 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1538 MaskVal << ShiftVal,
1540 SystemZICMP::UnsignedOnly))) {
1541 NewC.Op0 = NewC.Op0.getOperand(0);
1542 MaskVal <<= ShiftVal;
1544 NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
1550 // Go ahead and make the change.
1551 C.Opcode = SystemZISD::TM;
1553 if (Mask && Mask->getZExtValue() == MaskVal)
1554 C.Op1 = SDValue(Mask, 0);
1556 C.Op1 = DAG.getConstant(MaskVal, C.Op0.getValueType());
1557 C.CCValid = SystemZ::CCMASK_TM;
1558 C.CCMask = NewCCMask;
1561 // Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
1562 static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
1563 ISD::CondCode Cond) {
1564 Comparison C(CmpOp0, CmpOp1);
1565 C.CCMask = CCMaskForCondCode(Cond);
1566 if (C.Op0.getValueType().isFloatingPoint()) {
1567 C.CCValid = SystemZ::CCMASK_FCMP;
1568 C.Opcode = SystemZISD::FCMP;
1571 C.CCValid = SystemZ::CCMASK_ICMP;
1572 C.Opcode = SystemZISD::ICMP;
1573 // Choose the type of comparison. Equality and inequality tests can
1574 // use either signed or unsigned comparisons. The choice also doesn't
1575 // matter if both sign bits are known to be clear. In those cases we
1576 // want to give the main isel code the freedom to choose whichever
1578 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1579 C.CCMask == SystemZ::CCMASK_CMP_NE ||
1580 (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
1581 C.ICmpType = SystemZICMP::Any;
1582 else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
1583 C.ICmpType = SystemZICMP::UnsignedOnly;
1585 C.ICmpType = SystemZICMP::SignedOnly;
1586 C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
1587 adjustZeroCmp(DAG, C);
1588 adjustSubwordCmp(DAG, C);
1589 adjustForSubtraction(DAG, C);
1591 adjustICmpTruncate(DAG, C);
1594 if (shouldSwapCmpOperands(C)) {
1595 std::swap(C.Op0, C.Op1);
1596 C.CCMask = reverseCCMask(C.CCMask);
1599 adjustForTestUnderMask(DAG, C);
1603 // Emit the comparison instruction described by C.
1604 static SDValue emitCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1605 if (C.Opcode == SystemZISD::ICMP)
1606 return DAG.getNode(SystemZISD::ICMP, DL, MVT::Glue, C.Op0, C.Op1,
1607 DAG.getConstant(C.ICmpType, MVT::i32));
1608 if (C.Opcode == SystemZISD::TM) {
1609 bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
1610 bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
1611 return DAG.getNode(SystemZISD::TM, DL, MVT::Glue, C.Op0, C.Op1,
1612 DAG.getConstant(RegisterOnly, MVT::i32));
1614 return DAG.getNode(C.Opcode, DL, MVT::Glue, C.Op0, C.Op1);
1617 // Implement a 32-bit *MUL_LOHI operation by extending both operands to
1618 // 64 bits. Extend is the extension type to use. Store the high part
1619 // in Hi and the low part in Lo.
1620 static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL,
1621 unsigned Extend, SDValue Op0, SDValue Op1,
1622 SDValue &Hi, SDValue &Lo) {
1623 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
1624 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
1625 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
1626 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, DAG.getConstant(32, MVT::i64));
1627 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
1628 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
1631 // Lower a binary operation that produces two VT results, one in each
1632 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
1633 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
1634 // on the extended Op0 and (unextended) Op1. Store the even register result
1635 // in Even and the odd register result in Odd.
1636 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
1637 unsigned Extend, unsigned Opcode,
1638 SDValue Op0, SDValue Op1,
1639 SDValue &Even, SDValue &Odd) {
1640 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
1641 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
1642 SDValue(In128, 0), Op1);
1643 bool Is32Bit = is32Bit(VT);
1644 Even = DAG.getTargetExtractSubreg(SystemZ::even128(Is32Bit), DL, VT, Result);
1645 Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
1648 // Return an i32 value that is 1 if the CC value produced by Glue is
1649 // in the mask CCMask and 0 otherwise. CC is known to have a value
1650 // in CCValid, so other values can be ignored.
1651 static SDValue emitSETCC(SelectionDAG &DAG, SDLoc DL, SDValue Glue,
1652 unsigned CCValid, unsigned CCMask) {
1653 IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
1654 SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
1656 if (Conversion.XORValue)
1657 Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
1658 DAG.getConstant(Conversion.XORValue, MVT::i32));
1660 if (Conversion.AddValue)
1661 Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
1662 DAG.getConstant(Conversion.AddValue, MVT::i32));
1664 // The SHR/AND sequence should get optimized to an RISBG.
1665 Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
1666 DAG.getConstant(Conversion.Bit, MVT::i32));
1667 if (Conversion.Bit != 31)
1668 Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
1669 DAG.getConstant(1, MVT::i32));
1673 SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
1674 SelectionDAG &DAG) const {
1675 SDValue CmpOp0 = Op.getOperand(0);
1676 SDValue CmpOp1 = Op.getOperand(1);
1677 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1680 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1681 SDValue Glue = emitCmp(DAG, DL, C);
1682 return emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1685 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1686 SDValue Chain = Op.getOperand(0);
1687 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1688 SDValue CmpOp0 = Op.getOperand(2);
1689 SDValue CmpOp1 = Op.getOperand(3);
1690 SDValue Dest = Op.getOperand(4);
1693 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1694 SDValue Glue = emitCmp(DAG, DL, C);
1695 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
1696 Chain, DAG.getConstant(C.CCValid, MVT::i32),
1697 DAG.getConstant(C.CCMask, MVT::i32), Dest, Glue);
1700 // Return true if Pos is CmpOp and Neg is the negative of CmpOp,
1701 // allowing Pos and Neg to be wider than CmpOp.
1702 static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
1703 return (Neg.getOpcode() == ISD::SUB &&
1704 Neg.getOperand(0).getOpcode() == ISD::Constant &&
1705 cast<ConstantSDNode>(Neg.getOperand(0))->getZExtValue() == 0 &&
1706 Neg.getOperand(1) == Pos &&
1708 (Pos.getOpcode() == ISD::SIGN_EXTEND &&
1709 Pos.getOperand(0) == CmpOp)));
1712 // Return the absolute or negative absolute of Op; IsNegative decides which.
1713 static SDValue getAbsolute(SelectionDAG &DAG, SDLoc DL, SDValue Op,
1715 Op = DAG.getNode(SystemZISD::IABS, DL, Op.getValueType(), Op);
1717 Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
1718 DAG.getConstant(0, Op.getValueType()), Op);
1722 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
1723 SelectionDAG &DAG) const {
1724 SDValue CmpOp0 = Op.getOperand(0);
1725 SDValue CmpOp1 = Op.getOperand(1);
1726 SDValue TrueOp = Op.getOperand(2);
1727 SDValue FalseOp = Op.getOperand(3);
1728 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1731 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1733 // Check for absolute and negative-absolute selections, including those
1734 // where the comparison value is sign-extended (for LPGFR and LNGFR).
1735 // This check supplements the one in DAGCombiner.
1736 if (C.Opcode == SystemZISD::ICMP &&
1737 C.CCMask != SystemZ::CCMASK_CMP_EQ &&
1738 C.CCMask != SystemZ::CCMASK_CMP_NE &&
1739 C.Op1.getOpcode() == ISD::Constant &&
1740 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1741 if (isAbsolute(C.Op0, TrueOp, FalseOp))
1742 return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
1743 if (isAbsolute(C.Op0, FalseOp, TrueOp))
1744 return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
1747 SDValue Glue = emitCmp(DAG, DL, C);
1749 // Special case for handling -1/0 results. The shifts we use here
1750 // should get optimized with the IPM conversion sequence.
1751 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp);
1752 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp);
1753 if (TrueC && FalseC) {
1754 int64_t TrueVal = TrueC->getSExtValue();
1755 int64_t FalseVal = FalseC->getSExtValue();
1756 if ((TrueVal == -1 && FalseVal == 0) || (TrueVal == 0 && FalseVal == -1)) {
1757 // Invert the condition if we want -1 on false.
1759 C.CCMask ^= C.CCValid;
1760 SDValue Result = emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1761 EVT VT = Op.getValueType();
1762 // Extend the result to VT. Upper bits are ignored.
1764 Result = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Result);
1765 // Sign-extend from the low bit.
1766 SDValue ShAmt = DAG.getConstant(VT.getSizeInBits() - 1, MVT::i32);
1767 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Result, ShAmt);
1768 return DAG.getNode(ISD::SRA, DL, VT, Shl, ShAmt);
1772 SmallVector<SDValue, 5> Ops;
1773 Ops.push_back(TrueOp);
1774 Ops.push_back(FalseOp);
1775 Ops.push_back(DAG.getConstant(C.CCValid, MVT::i32));
1776 Ops.push_back(DAG.getConstant(C.CCMask, MVT::i32));
1777 Ops.push_back(Glue);
1779 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
1780 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size());
1783 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
1784 SelectionDAG &DAG) const {
1786 const GlobalValue *GV = Node->getGlobal();
1787 int64_t Offset = Node->getOffset();
1788 EVT PtrVT = getPointerTy();
1789 Reloc::Model RM = TM.getRelocationModel();
1790 CodeModel::Model CM = TM.getCodeModel();
1793 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
1794 // Assign anchors at 1<<12 byte boundaries.
1795 uint64_t Anchor = Offset & ~uint64_t(0xfff);
1796 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor);
1797 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1799 // The offset can be folded into the address if it is aligned to a halfword.
1801 if (Offset != 0 && (Offset & 1) == 0) {
1802 SDValue Full = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor + Offset);
1803 Result = DAG.getNode(SystemZISD::PCREL_OFFSET, DL, PtrVT, Full, Result);
1807 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
1808 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1809 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
1810 MachinePointerInfo::getGOT(), false, false, false, 0);
1813 // If there was a non-zero offset that we didn't fold, create an explicit
1816 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
1817 DAG.getConstant(Offset, PtrVT));
1822 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
1823 SelectionDAG &DAG) const {
1825 const GlobalValue *GV = Node->getGlobal();
1826 EVT PtrVT = getPointerTy();
1827 TLSModel::Model model = TM.getTLSModel(GV);
1829 if (model != TLSModel::LocalExec)
1830 llvm_unreachable("only local-exec TLS mode supported");
1832 // The high part of the thread pointer is in access register 0.
1833 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1834 DAG.getConstant(0, MVT::i32));
1835 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
1837 // The low part of the thread pointer is in access register 1.
1838 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1839 DAG.getConstant(1, MVT::i32));
1840 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
1842 // Merge them into a single 64-bit address.
1843 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
1844 DAG.getConstant(32, PtrVT));
1845 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
1847 // Get the offset of GA from the thread pointer.
1848 SystemZConstantPoolValue *CPV =
1849 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
1851 // Force the offset into the constant pool and load it from there.
1852 SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8);
1853 SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1854 CPAddr, MachinePointerInfo::getConstantPool(),
1855 false, false, false, 0);
1857 // Add the base and offset together.
1858 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
1861 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
1862 SelectionDAG &DAG) const {
1864 const BlockAddress *BA = Node->getBlockAddress();
1865 int64_t Offset = Node->getOffset();
1866 EVT PtrVT = getPointerTy();
1868 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
1869 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1873 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
1874 SelectionDAG &DAG) const {
1876 EVT PtrVT = getPointerTy();
1877 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1879 // Use LARL to load the address of the table.
1880 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1883 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
1884 SelectionDAG &DAG) const {
1886 EVT PtrVT = getPointerTy();
1889 if (CP->isMachineConstantPoolEntry())
1890 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1891 CP->getAlignment());
1893 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1894 CP->getAlignment(), CP->getOffset());
1896 // Use LARL to load the address of the constant pool entry.
1897 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1900 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
1901 SelectionDAG &DAG) const {
1903 SDValue In = Op.getOperand(0);
1904 EVT InVT = In.getValueType();
1905 EVT ResVT = Op.getValueType();
1907 if (InVT == MVT::i32 && ResVT == MVT::f32) {
1909 if (Subtarget.hasHighWord()) {
1910 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL,
1912 In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
1913 MVT::i64, SDValue(U64, 0), In);
1915 In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
1916 In64 = DAG.getNode(ISD::SHL, DL, MVT::i64, In64,
1917 DAG.getConstant(32, MVT::i64));
1919 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, In64);
1920 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32,
1921 DL, MVT::f32, Out64);
1923 if (InVT == MVT::f32 && ResVT == MVT::i32) {
1924 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
1925 SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
1926 MVT::f64, SDValue(U64, 0), In);
1927 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
1928 if (Subtarget.hasHighWord())
1929 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32, DL,
1931 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64,
1932 DAG.getConstant(32, MVT::i64));
1933 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
1935 llvm_unreachable("Unexpected bitcast combination");
1938 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
1939 SelectionDAG &DAG) const {
1940 MachineFunction &MF = DAG.getMachineFunction();
1941 SystemZMachineFunctionInfo *FuncInfo =
1942 MF.getInfo<SystemZMachineFunctionInfo>();
1943 EVT PtrVT = getPointerTy();
1945 SDValue Chain = Op.getOperand(0);
1946 SDValue Addr = Op.getOperand(1);
1947 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1950 // The initial values of each field.
1951 const unsigned NumFields = 4;
1952 SDValue Fields[NumFields] = {
1953 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT),
1954 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT),
1955 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
1956 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
1959 // Store each field into its respective slot.
1960 SDValue MemOps[NumFields];
1961 unsigned Offset = 0;
1962 for (unsigned I = 0; I < NumFields; ++I) {
1963 SDValue FieldAddr = Addr;
1965 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
1966 DAG.getIntPtrConstant(Offset));
1967 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
1968 MachinePointerInfo(SV, Offset),
1972 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps, NumFields);
1975 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
1976 SelectionDAG &DAG) const {
1977 SDValue Chain = Op.getOperand(0);
1978 SDValue DstPtr = Op.getOperand(1);
1979 SDValue SrcPtr = Op.getOperand(2);
1980 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
1981 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
1984 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32),
1985 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
1986 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
1989 SDValue SystemZTargetLowering::
1990 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
1991 SDValue Chain = Op.getOperand(0);
1992 SDValue Size = Op.getOperand(1);
1995 unsigned SPReg = getStackPointerRegisterToSaveRestore();
1997 // Get a reference to the stack pointer.
1998 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
2000 // Get the new stack pointer value.
2001 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
2003 // Copy the new stack pointer back.
2004 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
2006 // The allocated data lives above the 160 bytes allocated for the standard
2007 // frame, plus any outgoing stack arguments. We don't know how much that
2008 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
2009 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
2010 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
2012 SDValue Ops[2] = { Result, Chain };
2013 return DAG.getMergeValues(Ops, 2, DL);
2016 SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
2017 SelectionDAG &DAG) const {
2018 EVT VT = Op.getValueType();
2022 // Just do a normal 64-bit multiplication and extract the results.
2023 // We define this so that it can be used for constant division.
2024 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
2025 Op.getOperand(1), Ops[1], Ops[0]);
2027 // Do a full 128-bit multiplication based on UMUL_LOHI64:
2029 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
2031 // but using the fact that the upper halves are either all zeros
2034 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
2036 // and grouping the right terms together since they are quicker than the
2039 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
2040 SDValue C63 = DAG.getConstant(63, MVT::i64);
2041 SDValue LL = Op.getOperand(0);
2042 SDValue RL = Op.getOperand(1);
2043 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
2044 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
2045 // UMUL_LOHI64 returns the low result in the odd register and the high
2046 // result in the even register. SMUL_LOHI is defined to return the
2047 // low half first, so the results are in reverse order.
2048 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2049 LL, RL, Ops[1], Ops[0]);
2050 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
2051 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
2052 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
2053 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
2055 return DAG.getMergeValues(Ops, 2, DL);
2058 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
2059 SelectionDAG &DAG) const {
2060 EVT VT = Op.getValueType();
2064 // Just do a normal 64-bit multiplication and extract the results.
2065 // We define this so that it can be used for constant division.
2066 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
2067 Op.getOperand(1), Ops[1], Ops[0]);
2069 // UMUL_LOHI64 returns the low result in the odd register and the high
2070 // result in the even register. UMUL_LOHI is defined to return the
2071 // low half first, so the results are in reverse order.
2072 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2073 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2074 return DAG.getMergeValues(Ops, 2, DL);
2077 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
2078 SelectionDAG &DAG) const {
2079 SDValue Op0 = Op.getOperand(0);
2080 SDValue Op1 = Op.getOperand(1);
2081 EVT VT = Op.getValueType();
2085 // We use DSGF for 32-bit division.
2087 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
2088 Opcode = SystemZISD::SDIVREM32;
2089 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
2090 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
2091 Opcode = SystemZISD::SDIVREM32;
2093 Opcode = SystemZISD::SDIVREM64;
2095 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
2096 // input is "don't care". The instruction returns the remainder in
2097 // the even register and the quotient in the odd register.
2099 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
2100 Op0, Op1, Ops[1], Ops[0]);
2101 return DAG.getMergeValues(Ops, 2, DL);
2104 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
2105 SelectionDAG &DAG) const {
2106 EVT VT = Op.getValueType();
2109 // DL(G) uses a double-width dividend, so we need to clear the even
2110 // register in the GR128 input. The instruction returns the remainder
2111 // in the even register and the quotient in the odd register.
2114 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
2115 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2117 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
2118 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2119 return DAG.getMergeValues(Ops, 2, DL);
2122 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
2123 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
2125 // Get the known-zero masks for each operand.
2126 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
2127 APInt KnownZero[2], KnownOne[2];
2128 DAG.ComputeMaskedBits(Ops[0], KnownZero[0], KnownOne[0]);
2129 DAG.ComputeMaskedBits(Ops[1], KnownZero[1], KnownOne[1]);
2131 // See if the upper 32 bits of one operand and the lower 32 bits of the
2132 // other are known zero. They are the low and high operands respectively.
2133 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
2134 KnownZero[1].getZExtValue() };
2136 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
2138 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
2143 SDValue LowOp = Ops[Low];
2144 SDValue HighOp = Ops[High];
2146 // If the high part is a constant, we're better off using IILH.
2147 if (HighOp.getOpcode() == ISD::Constant)
2150 // If the low part is a constant that is outside the range of LHI,
2151 // then we're better off using IILF.
2152 if (LowOp.getOpcode() == ISD::Constant) {
2153 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
2154 if (!isInt<16>(Value))
2158 // Check whether the high part is an AND that doesn't change the
2159 // high 32 bits and just masks out low bits. We can skip it if so.
2160 if (HighOp.getOpcode() == ISD::AND &&
2161 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
2162 SDValue HighOp0 = HighOp.getOperand(0);
2163 uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
2164 if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
2168 // Take advantage of the fact that all GR32 operations only change the
2169 // low 32 bits by truncating Low to an i32 and inserting it directly
2170 // using a subreg. The interesting cases are those where the truncation
2173 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
2174 return DAG.getTargetInsertSubreg(SystemZ::subreg_l32, DL,
2175 MVT::i64, HighOp, Low32);
2178 SDValue SystemZTargetLowering::lowerSIGN_EXTEND(SDValue Op,
2179 SelectionDAG &DAG) const {
2180 // Convert (sext (ashr (shl X, C1), C2)) to
2181 // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
2182 // cheap as narrower ones.
2183 SDValue N0 = Op.getOperand(0);
2184 EVT VT = Op.getValueType();
2185 if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
2186 ConstantSDNode *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2187 SDValue Inner = N0.getOperand(0);
2188 if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
2189 ConstantSDNode *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1));
2191 unsigned Extra = (VT.getSizeInBits() -
2192 N0.getValueType().getSizeInBits());
2193 unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
2194 unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
2195 EVT ShiftVT = N0.getOperand(1).getValueType();
2196 SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
2197 Inner.getOperand(0));
2198 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
2199 DAG.getConstant(NewShlAmt, ShiftVT));
2200 return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
2201 DAG.getConstant(NewSraAmt, ShiftVT));
2208 // Op is an atomic load. Lower it into a normal volatile load.
2209 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
2210 SelectionDAG &DAG) const {
2211 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
2212 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), Op.getValueType(),
2213 Node->getChain(), Node->getBasePtr(),
2214 Node->getMemoryVT(), Node->getMemOperand());
2217 // Op is an atomic store. Lower it into a normal volatile store followed
2218 // by a serialization.
2219 SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
2220 SelectionDAG &DAG) const {
2221 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
2222 SDValue Chain = DAG.getTruncStore(Node->getChain(), SDLoc(Op), Node->getVal(),
2223 Node->getBasePtr(), Node->getMemoryVT(),
2224 Node->getMemOperand());
2225 return SDValue(DAG.getMachineNode(SystemZ::Serialize, SDLoc(Op), MVT::Other,
2229 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
2230 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
2231 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
2233 unsigned Opcode) const {
2234 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
2236 // 32-bit operations need no code outside the main loop.
2237 EVT NarrowVT = Node->getMemoryVT();
2238 EVT WideVT = MVT::i32;
2239 if (NarrowVT == WideVT)
2242 int64_t BitSize = NarrowVT.getSizeInBits();
2243 SDValue ChainIn = Node->getChain();
2244 SDValue Addr = Node->getBasePtr();
2245 SDValue Src2 = Node->getVal();
2246 MachineMemOperand *MMO = Node->getMemOperand();
2248 EVT PtrVT = Addr.getValueType();
2250 // Convert atomic subtracts of constants into additions.
2251 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
2252 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Src2)) {
2253 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
2254 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
2257 // Get the address of the containing word.
2258 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2259 DAG.getConstant(-4, PtrVT));
2261 // Get the number of bits that the word must be rotated left in order
2262 // to bring the field to the top bits of a GR32.
2263 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2264 DAG.getConstant(3, PtrVT));
2265 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2267 // Get the complementing shift amount, for rotating a field in the top
2268 // bits back to its proper position.
2269 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2270 DAG.getConstant(0, WideVT), BitShift);
2272 // Extend the source operand to 32 bits and prepare it for the inner loop.
2273 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
2274 // operations require the source to be shifted in advance. (This shift
2275 // can be folded if the source is constant.) For AND and NAND, the lower
2276 // bits must be set, while for other opcodes they should be left clear.
2277 if (Opcode != SystemZISD::ATOMIC_SWAPW)
2278 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
2279 DAG.getConstant(32 - BitSize, WideVT));
2280 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
2281 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
2282 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
2283 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT));
2285 // Construct the ATOMIC_LOADW_* node.
2286 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2287 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
2288 DAG.getConstant(BitSize, WideVT) };
2289 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
2290 array_lengthof(Ops),
2293 // Rotate the result of the final CS so that the field is in the lower
2294 // bits of a GR32, then truncate it.
2295 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
2296 DAG.getConstant(BitSize, WideVT));
2297 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
2299 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
2300 return DAG.getMergeValues(RetOps, 2, DL);
2303 // Op is an ATOMIC_LOAD_SUB operation. Lower 8- and 16-bit operations
2304 // into ATOMIC_LOADW_SUBs and decide whether to convert 32- and 64-bit
2305 // operations into additions.
2306 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
2307 SelectionDAG &DAG) const {
2308 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
2309 EVT MemVT = Node->getMemoryVT();
2310 if (MemVT == MVT::i32 || MemVT == MVT::i64) {
2311 // A full-width operation.
2312 assert(Op.getValueType() == MemVT && "Mismatched VTs");
2313 SDValue Src2 = Node->getVal();
2317 if (ConstantSDNode *Op2 = dyn_cast<ConstantSDNode>(Src2)) {
2318 // Use an addition if the operand is constant and either LAA(G) is
2319 // available or the negative value is in the range of A(G)FHI.
2320 int64_t Value = (-Op2->getAPIntValue()).getSExtValue();
2321 if (isInt<32>(Value) || TM.getSubtargetImpl()->hasInterlockedAccess1())
2322 NegSrc2 = DAG.getConstant(Value, MemVT);
2323 } else if (TM.getSubtargetImpl()->hasInterlockedAccess1())
2324 // Use LAA(G) if available.
2325 NegSrc2 = DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, MemVT),
2328 if (NegSrc2.getNode())
2329 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
2330 Node->getChain(), Node->getBasePtr(), NegSrc2,
2331 Node->getMemOperand(), Node->getOrdering(),
2332 Node->getSynchScope());
2334 // Use the node as-is.
2338 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
2341 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
2342 // into a fullword ATOMIC_CMP_SWAPW operation.
2343 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
2344 SelectionDAG &DAG) const {
2345 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
2347 // We have native support for 32-bit compare and swap.
2348 EVT NarrowVT = Node->getMemoryVT();
2349 EVT WideVT = MVT::i32;
2350 if (NarrowVT == WideVT)
2353 int64_t BitSize = NarrowVT.getSizeInBits();
2354 SDValue ChainIn = Node->getOperand(0);
2355 SDValue Addr = Node->getOperand(1);
2356 SDValue CmpVal = Node->getOperand(2);
2357 SDValue SwapVal = Node->getOperand(3);
2358 MachineMemOperand *MMO = Node->getMemOperand();
2360 EVT PtrVT = Addr.getValueType();
2362 // Get the address of the containing word.
2363 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2364 DAG.getConstant(-4, PtrVT));
2366 // Get the number of bits that the word must be rotated left in order
2367 // to bring the field to the top bits of a GR32.
2368 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2369 DAG.getConstant(3, PtrVT));
2370 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2372 // Get the complementing shift amount, for rotating a field in the top
2373 // bits back to its proper position.
2374 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2375 DAG.getConstant(0, WideVT), BitShift);
2377 // Construct the ATOMIC_CMP_SWAPW node.
2378 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2379 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
2380 NegBitShift, DAG.getConstant(BitSize, WideVT) };
2381 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
2382 VTList, Ops, array_lengthof(Ops),
2387 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
2388 SelectionDAG &DAG) const {
2389 MachineFunction &MF = DAG.getMachineFunction();
2390 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2391 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
2392 SystemZ::R15D, Op.getValueType());
2395 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
2396 SelectionDAG &DAG) const {
2397 MachineFunction &MF = DAG.getMachineFunction();
2398 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2399 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
2400 SystemZ::R15D, Op.getOperand(1));
2403 SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op,
2404 SelectionDAG &DAG) const {
2405 bool IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2407 // Just preserve the chain.
2408 return Op.getOperand(0);
2410 bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2411 unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
2412 MemIntrinsicSDNode *Node = cast<MemIntrinsicSDNode>(Op.getNode());
2415 DAG.getConstant(Code, MVT::i32),
2418 return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, SDLoc(Op),
2419 Node->getVTList(), Ops, array_lengthof(Ops),
2420 Node->getMemoryVT(), Node->getMemOperand());
2423 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
2424 SelectionDAG &DAG) const {
2425 switch (Op.getOpcode()) {
2427 return lowerBR_CC(Op, DAG);
2428 case ISD::SELECT_CC:
2429 return lowerSELECT_CC(Op, DAG);
2431 return lowerSETCC(Op, DAG);
2432 case ISD::GlobalAddress:
2433 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
2434 case ISD::GlobalTLSAddress:
2435 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
2436 case ISD::BlockAddress:
2437 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
2438 case ISD::JumpTable:
2439 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
2440 case ISD::ConstantPool:
2441 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
2443 return lowerBITCAST(Op, DAG);
2445 return lowerVASTART(Op, DAG);
2447 return lowerVACOPY(Op, DAG);
2448 case ISD::DYNAMIC_STACKALLOC:
2449 return lowerDYNAMIC_STACKALLOC(Op, DAG);
2450 case ISD::SMUL_LOHI:
2451 return lowerSMUL_LOHI(Op, DAG);
2452 case ISD::UMUL_LOHI:
2453 return lowerUMUL_LOHI(Op, DAG);
2455 return lowerSDIVREM(Op, DAG);
2457 return lowerUDIVREM(Op, DAG);
2459 return lowerOR(Op, DAG);
2460 case ISD::SIGN_EXTEND:
2461 return lowerSIGN_EXTEND(Op, DAG);
2462 case ISD::ATOMIC_SWAP:
2463 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
2464 case ISD::ATOMIC_STORE:
2465 return lowerATOMIC_STORE(Op, DAG);
2466 case ISD::ATOMIC_LOAD:
2467 return lowerATOMIC_LOAD(Op, DAG);
2468 case ISD::ATOMIC_LOAD_ADD:
2469 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
2470 case ISD::ATOMIC_LOAD_SUB:
2471 return lowerATOMIC_LOAD_SUB(Op, DAG);
2472 case ISD::ATOMIC_LOAD_AND:
2473 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
2474 case ISD::ATOMIC_LOAD_OR:
2475 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
2476 case ISD::ATOMIC_LOAD_XOR:
2477 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
2478 case ISD::ATOMIC_LOAD_NAND:
2479 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
2480 case ISD::ATOMIC_LOAD_MIN:
2481 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
2482 case ISD::ATOMIC_LOAD_MAX:
2483 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
2484 case ISD::ATOMIC_LOAD_UMIN:
2485 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
2486 case ISD::ATOMIC_LOAD_UMAX:
2487 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
2488 case ISD::ATOMIC_CMP_SWAP:
2489 return lowerATOMIC_CMP_SWAP(Op, DAG);
2490 case ISD::STACKSAVE:
2491 return lowerSTACKSAVE(Op, DAG);
2492 case ISD::STACKRESTORE:
2493 return lowerSTACKRESTORE(Op, DAG);
2495 return lowerPREFETCH(Op, DAG);
2497 llvm_unreachable("Unexpected node to lower");
2501 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
2502 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
2507 OPCODE(PCREL_WRAPPER);
2508 OPCODE(PCREL_OFFSET);
2514 OPCODE(SELECT_CCMASK);
2515 OPCODE(ADJDYNALLOC);
2516 OPCODE(EXTRACT_ACCESS);
2517 OPCODE(UMUL_LOHI64);
2533 OPCODE(SEARCH_STRING);
2536 OPCODE(ATOMIC_SWAPW);
2537 OPCODE(ATOMIC_LOADW_ADD);
2538 OPCODE(ATOMIC_LOADW_SUB);
2539 OPCODE(ATOMIC_LOADW_AND);
2540 OPCODE(ATOMIC_LOADW_OR);
2541 OPCODE(ATOMIC_LOADW_XOR);
2542 OPCODE(ATOMIC_LOADW_NAND);
2543 OPCODE(ATOMIC_LOADW_MIN);
2544 OPCODE(ATOMIC_LOADW_MAX);
2545 OPCODE(ATOMIC_LOADW_UMIN);
2546 OPCODE(ATOMIC_LOADW_UMAX);
2547 OPCODE(ATOMIC_CMP_SWAPW);
2554 //===----------------------------------------------------------------------===//
2556 //===----------------------------------------------------------------------===//
2558 // Create a new basic block after MBB.
2559 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
2560 MachineFunction &MF = *MBB->getParent();
2561 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
2562 MF.insert(std::next(MachineFunction::iterator(MBB)), NewMBB);
2566 // Split MBB after MI and return the new block (the one that contains
2567 // instructions after MI).
2568 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
2569 MachineBasicBlock *MBB) {
2570 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2571 NewMBB->splice(NewMBB->begin(), MBB,
2572 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
2573 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2577 // Split MBB before MI and return the new block (the one that contains MI).
2578 static MachineBasicBlock *splitBlockBefore(MachineInstr *MI,
2579 MachineBasicBlock *MBB) {
2580 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2581 NewMBB->splice(NewMBB->begin(), MBB, MI, MBB->end());
2582 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2586 // Force base value Base into a register before MI. Return the register.
2587 static unsigned forceReg(MachineInstr *MI, MachineOperand &Base,
2588 const SystemZInstrInfo *TII) {
2590 return Base.getReg();
2592 MachineBasicBlock *MBB = MI->getParent();
2593 MachineFunction &MF = *MBB->getParent();
2594 MachineRegisterInfo &MRI = MF.getRegInfo();
2596 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
2597 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LA), Reg)
2598 .addOperand(Base).addImm(0).addReg(0);
2602 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
2604 SystemZTargetLowering::emitSelect(MachineInstr *MI,
2605 MachineBasicBlock *MBB) const {
2606 const SystemZInstrInfo *TII = TM.getInstrInfo();
2608 unsigned DestReg = MI->getOperand(0).getReg();
2609 unsigned TrueReg = MI->getOperand(1).getReg();
2610 unsigned FalseReg = MI->getOperand(2).getReg();
2611 unsigned CCValid = MI->getOperand(3).getImm();
2612 unsigned CCMask = MI->getOperand(4).getImm();
2613 DebugLoc DL = MI->getDebugLoc();
2615 MachineBasicBlock *StartMBB = MBB;
2616 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2617 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2620 // BRC CCMask, JoinMBB
2621 // # fallthrough to FalseMBB
2623 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2624 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2625 MBB->addSuccessor(JoinMBB);
2626 MBB->addSuccessor(FalseMBB);
2629 // # fallthrough to JoinMBB
2631 MBB->addSuccessor(JoinMBB);
2634 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
2637 BuildMI(*MBB, MI, DL, TII->get(SystemZ::PHI), DestReg)
2638 .addReg(TrueReg).addMBB(StartMBB)
2639 .addReg(FalseReg).addMBB(FalseMBB);
2641 MI->eraseFromParent();
2645 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
2646 // StoreOpcode is the store to use and Invert says whether the store should
2647 // happen when the condition is false rather than true. If a STORE ON
2648 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
2650 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
2651 MachineBasicBlock *MBB,
2652 unsigned StoreOpcode, unsigned STOCOpcode,
2653 bool Invert) const {
2654 const SystemZInstrInfo *TII = TM.getInstrInfo();
2656 unsigned SrcReg = MI->getOperand(0).getReg();
2657 MachineOperand Base = MI->getOperand(1);
2658 int64_t Disp = MI->getOperand(2).getImm();
2659 unsigned IndexReg = MI->getOperand(3).getReg();
2660 unsigned CCValid = MI->getOperand(4).getImm();
2661 unsigned CCMask = MI->getOperand(5).getImm();
2662 DebugLoc DL = MI->getDebugLoc();
2664 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
2666 // Use STOCOpcode if possible. We could use different store patterns in
2667 // order to avoid matching the index register, but the performance trade-offs
2668 // might be more complicated in that case.
2669 if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) {
2672 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
2673 .addReg(SrcReg).addOperand(Base).addImm(Disp)
2674 .addImm(CCValid).addImm(CCMask);
2675 MI->eraseFromParent();
2679 // Get the condition needed to branch around the store.
2683 MachineBasicBlock *StartMBB = MBB;
2684 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2685 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2688 // BRC CCMask, JoinMBB
2689 // # fallthrough to FalseMBB
2691 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2692 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2693 MBB->addSuccessor(JoinMBB);
2694 MBB->addSuccessor(FalseMBB);
2697 // store %SrcReg, %Disp(%Index,%Base)
2698 // # fallthrough to JoinMBB
2700 BuildMI(MBB, DL, TII->get(StoreOpcode))
2701 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
2702 MBB->addSuccessor(JoinMBB);
2704 MI->eraseFromParent();
2708 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
2709 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
2710 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
2711 // BitSize is the width of the field in bits, or 0 if this is a partword
2712 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
2713 // is one of the operands. Invert says whether the field should be
2714 // inverted after performing BinOpcode (e.g. for NAND).
2716 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
2717 MachineBasicBlock *MBB,
2720 bool Invert) const {
2721 const SystemZInstrInfo *TII = TM.getInstrInfo();
2722 MachineFunction &MF = *MBB->getParent();
2723 MachineRegisterInfo &MRI = MF.getRegInfo();
2724 bool IsSubWord = (BitSize < 32);
2726 // Extract the operands. Base can be a register or a frame index.
2727 // Src2 can be a register or immediate.
2728 unsigned Dest = MI->getOperand(0).getReg();
2729 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2730 int64_t Disp = MI->getOperand(2).getImm();
2731 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
2732 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2733 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2734 DebugLoc DL = MI->getDebugLoc();
2736 BitSize = MI->getOperand(6).getImm();
2738 // Subword operations use 32-bit registers.
2739 const TargetRegisterClass *RC = (BitSize <= 32 ?
2740 &SystemZ::GR32BitRegClass :
2741 &SystemZ::GR64BitRegClass);
2742 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2743 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2745 // Get the right opcodes for the displacement.
2746 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2747 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2748 assert(LOpcode && CSOpcode && "Displacement out of range");
2750 // Create virtual registers for temporary results.
2751 unsigned OrigVal = MRI.createVirtualRegister(RC);
2752 unsigned OldVal = MRI.createVirtualRegister(RC);
2753 unsigned NewVal = (BinOpcode || IsSubWord ?
2754 MRI.createVirtualRegister(RC) : Src2.getReg());
2755 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2756 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2758 // Insert a basic block for the main loop.
2759 MachineBasicBlock *StartMBB = MBB;
2760 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2761 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2765 // %OrigVal = L Disp(%Base)
2766 // # fall through to LoopMMB
2768 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2769 .addOperand(Base).addImm(Disp).addReg(0);
2770 MBB->addSuccessor(LoopMBB);
2773 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
2774 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2775 // %RotatedNewVal = OP %RotatedOldVal, %Src2
2776 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2777 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2779 // # fall through to DoneMMB
2781 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2782 .addReg(OrigVal).addMBB(StartMBB)
2783 .addReg(Dest).addMBB(LoopMBB);
2785 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2786 .addReg(OldVal).addReg(BitShift).addImm(0);
2788 // Perform the operation normally and then invert every bit of the field.
2789 unsigned Tmp = MRI.createVirtualRegister(RC);
2790 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
2791 .addReg(RotatedOldVal).addOperand(Src2);
2793 // XILF with the upper BitSize bits set.
2794 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2795 .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize)));
2796 else if (BitSize == 32)
2797 // XILF with every bit set.
2798 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2799 .addReg(Tmp).addImm(~uint32_t(0));
2801 // Use LCGR and add -1 to the result, which is more compact than
2802 // an XILF, XILH pair.
2803 unsigned Tmp2 = MRI.createVirtualRegister(RC);
2804 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
2805 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
2806 .addReg(Tmp2).addImm(-1);
2808 } else if (BinOpcode)
2809 // A simply binary operation.
2810 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
2811 .addReg(RotatedOldVal).addOperand(Src2);
2813 // Use RISBG to rotate Src2 into position and use it to replace the
2814 // field in RotatedOldVal.
2815 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
2816 .addReg(RotatedOldVal).addReg(Src2.getReg())
2817 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
2819 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2820 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2821 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2822 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2823 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2824 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2825 MBB->addSuccessor(LoopMBB);
2826 MBB->addSuccessor(DoneMBB);
2828 MI->eraseFromParent();
2832 // Implement EmitInstrWithCustomInserter for pseudo
2833 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
2834 // instruction that should be used to compare the current field with the
2835 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
2836 // for when the current field should be kept. BitSize is the width of
2837 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
2839 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
2840 MachineBasicBlock *MBB,
2841 unsigned CompareOpcode,
2842 unsigned KeepOldMask,
2843 unsigned BitSize) const {
2844 const SystemZInstrInfo *TII = TM.getInstrInfo();
2845 MachineFunction &MF = *MBB->getParent();
2846 MachineRegisterInfo &MRI = MF.getRegInfo();
2847 bool IsSubWord = (BitSize < 32);
2849 // Extract the operands. Base can be a register or a frame index.
2850 unsigned Dest = MI->getOperand(0).getReg();
2851 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2852 int64_t Disp = MI->getOperand(2).getImm();
2853 unsigned Src2 = MI->getOperand(3).getReg();
2854 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2855 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2856 DebugLoc DL = MI->getDebugLoc();
2858 BitSize = MI->getOperand(6).getImm();
2860 // Subword operations use 32-bit registers.
2861 const TargetRegisterClass *RC = (BitSize <= 32 ?
2862 &SystemZ::GR32BitRegClass :
2863 &SystemZ::GR64BitRegClass);
2864 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2865 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2867 // Get the right opcodes for the displacement.
2868 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2869 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2870 assert(LOpcode && CSOpcode && "Displacement out of range");
2872 // Create virtual registers for temporary results.
2873 unsigned OrigVal = MRI.createVirtualRegister(RC);
2874 unsigned OldVal = MRI.createVirtualRegister(RC);
2875 unsigned NewVal = MRI.createVirtualRegister(RC);
2876 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2877 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
2878 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2880 // Insert 3 basic blocks for the loop.
2881 MachineBasicBlock *StartMBB = MBB;
2882 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2883 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2884 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
2885 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
2889 // %OrigVal = L Disp(%Base)
2890 // # fall through to LoopMMB
2892 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2893 .addOperand(Base).addImm(Disp).addReg(0);
2894 MBB->addSuccessor(LoopMBB);
2897 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
2898 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2899 // CompareOpcode %RotatedOldVal, %Src2
2900 // BRC KeepOldMask, UpdateMBB
2902 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2903 .addReg(OrigVal).addMBB(StartMBB)
2904 .addReg(Dest).addMBB(UpdateMBB);
2906 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2907 .addReg(OldVal).addReg(BitShift).addImm(0);
2908 BuildMI(MBB, DL, TII->get(CompareOpcode))
2909 .addReg(RotatedOldVal).addReg(Src2);
2910 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2911 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
2912 MBB->addSuccessor(UpdateMBB);
2913 MBB->addSuccessor(UseAltMBB);
2916 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
2917 // # fall through to UpdateMMB
2920 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
2921 .addReg(RotatedOldVal).addReg(Src2)
2922 .addImm(32).addImm(31 + BitSize).addImm(0);
2923 MBB->addSuccessor(UpdateMBB);
2926 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
2927 // [ %RotatedAltVal, UseAltMBB ]
2928 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2929 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2931 // # fall through to DoneMMB
2933 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
2934 .addReg(RotatedOldVal).addMBB(LoopMBB)
2935 .addReg(RotatedAltVal).addMBB(UseAltMBB);
2937 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2938 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2939 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2940 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2941 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2942 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2943 MBB->addSuccessor(LoopMBB);
2944 MBB->addSuccessor(DoneMBB);
2946 MI->eraseFromParent();
2950 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
2953 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
2954 MachineBasicBlock *MBB) const {
2955 const SystemZInstrInfo *TII = TM.getInstrInfo();
2956 MachineFunction &MF = *MBB->getParent();
2957 MachineRegisterInfo &MRI = MF.getRegInfo();
2959 // Extract the operands. Base can be a register or a frame index.
2960 unsigned Dest = MI->getOperand(0).getReg();
2961 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2962 int64_t Disp = MI->getOperand(2).getImm();
2963 unsigned OrigCmpVal = MI->getOperand(3).getReg();
2964 unsigned OrigSwapVal = MI->getOperand(4).getReg();
2965 unsigned BitShift = MI->getOperand(5).getReg();
2966 unsigned NegBitShift = MI->getOperand(6).getReg();
2967 int64_t BitSize = MI->getOperand(7).getImm();
2968 DebugLoc DL = MI->getDebugLoc();
2970 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
2972 // Get the right opcodes for the displacement.
2973 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
2974 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
2975 assert(LOpcode && CSOpcode && "Displacement out of range");
2977 // Create virtual registers for temporary results.
2978 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
2979 unsigned OldVal = MRI.createVirtualRegister(RC);
2980 unsigned CmpVal = MRI.createVirtualRegister(RC);
2981 unsigned SwapVal = MRI.createVirtualRegister(RC);
2982 unsigned StoreVal = MRI.createVirtualRegister(RC);
2983 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
2984 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
2985 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
2987 // Insert 2 basic blocks for the loop.
2988 MachineBasicBlock *StartMBB = MBB;
2989 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2990 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2991 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
2995 // %OrigOldVal = L Disp(%Base)
2996 // # fall through to LoopMMB
2998 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
2999 .addOperand(Base).addImm(Disp).addReg(0);
3000 MBB->addSuccessor(LoopMBB);
3003 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
3004 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
3005 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
3006 // %Dest = RLL %OldVal, BitSize(%BitShift)
3007 // ^^ The low BitSize bits contain the field
3009 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
3010 // ^^ Replace the upper 32-BitSize bits of the
3011 // comparison value with those that we loaded,
3012 // so that we can use a full word comparison.
3013 // CR %Dest, %RetryCmpVal
3015 // # Fall through to SetMBB
3017 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
3018 .addReg(OrigOldVal).addMBB(StartMBB)
3019 .addReg(RetryOldVal).addMBB(SetMBB);
3020 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
3021 .addReg(OrigCmpVal).addMBB(StartMBB)
3022 .addReg(RetryCmpVal).addMBB(SetMBB);
3023 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
3024 .addReg(OrigSwapVal).addMBB(StartMBB)
3025 .addReg(RetrySwapVal).addMBB(SetMBB);
3026 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
3027 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
3028 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
3029 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3030 BuildMI(MBB, DL, TII->get(SystemZ::CR))
3031 .addReg(Dest).addReg(RetryCmpVal);
3032 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3033 .addImm(SystemZ::CCMASK_ICMP)
3034 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
3035 MBB->addSuccessor(DoneMBB);
3036 MBB->addSuccessor(SetMBB);
3039 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
3040 // ^^ Replace the upper 32-BitSize bits of the new
3041 // value with those that we loaded.
3042 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
3043 // ^^ Rotate the new field to its proper position.
3044 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
3046 // # fall through to ExitMMB
3048 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
3049 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3050 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
3051 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
3052 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
3053 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
3054 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3055 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
3056 MBB->addSuccessor(LoopMBB);
3057 MBB->addSuccessor(DoneMBB);
3059 MI->eraseFromParent();
3063 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
3064 // if the high register of the GR128 value must be cleared or false if
3065 // it's "don't care". SubReg is subreg_l32 when extending a GR32
3066 // and subreg_l64 when extending a GR64.
3068 SystemZTargetLowering::emitExt128(MachineInstr *MI,
3069 MachineBasicBlock *MBB,
3070 bool ClearEven, unsigned SubReg) const {
3071 const SystemZInstrInfo *TII = TM.getInstrInfo();
3072 MachineFunction &MF = *MBB->getParent();
3073 MachineRegisterInfo &MRI = MF.getRegInfo();
3074 DebugLoc DL = MI->getDebugLoc();
3076 unsigned Dest = MI->getOperand(0).getReg();
3077 unsigned Src = MI->getOperand(1).getReg();
3078 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3080 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
3082 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3083 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
3085 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
3087 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
3088 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_h64);
3091 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
3092 .addReg(In128).addReg(Src).addImm(SubReg);
3094 MI->eraseFromParent();
3099 SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
3100 MachineBasicBlock *MBB,
3101 unsigned Opcode) const {
3102 const SystemZInstrInfo *TII = TM.getInstrInfo();
3103 MachineFunction &MF = *MBB->getParent();
3104 MachineRegisterInfo &MRI = MF.getRegInfo();
3105 DebugLoc DL = MI->getDebugLoc();
3107 MachineOperand DestBase = earlyUseOperand(MI->getOperand(0));
3108 uint64_t DestDisp = MI->getOperand(1).getImm();
3109 MachineOperand SrcBase = earlyUseOperand(MI->getOperand(2));
3110 uint64_t SrcDisp = MI->getOperand(3).getImm();
3111 uint64_t Length = MI->getOperand(4).getImm();
3113 // When generating more than one CLC, all but the last will need to
3114 // branch to the end when a difference is found.
3115 MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ?
3116 splitBlockAfter(MI, MBB) : 0);
3118 // Check for the loop form, in which operand 5 is the trip count.
3119 if (MI->getNumExplicitOperands() > 5) {
3120 bool HaveSingleBase = DestBase.isIdenticalTo(SrcBase);
3122 uint64_t StartCountReg = MI->getOperand(5).getReg();
3123 uint64_t StartSrcReg = forceReg(MI, SrcBase, TII);
3124 uint64_t StartDestReg = (HaveSingleBase ? StartSrcReg :
3125 forceReg(MI, DestBase, TII));
3127 const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass;
3128 uint64_t ThisSrcReg = MRI.createVirtualRegister(RC);
3129 uint64_t ThisDestReg = (HaveSingleBase ? ThisSrcReg :
3130 MRI.createVirtualRegister(RC));
3131 uint64_t NextSrcReg = MRI.createVirtualRegister(RC);
3132 uint64_t NextDestReg = (HaveSingleBase ? NextSrcReg :
3133 MRI.createVirtualRegister(RC));
3135 RC = &SystemZ::GR64BitRegClass;
3136 uint64_t ThisCountReg = MRI.createVirtualRegister(RC);
3137 uint64_t NextCountReg = MRI.createVirtualRegister(RC);
3139 MachineBasicBlock *StartMBB = MBB;
3140 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3141 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3142 MachineBasicBlock *NextMBB = (EndMBB ? emitBlockAfter(LoopMBB) : LoopMBB);
3145 // # fall through to LoopMMB
3146 MBB->addSuccessor(LoopMBB);
3149 // %ThisDestReg = phi [ %StartDestReg, StartMBB ],
3150 // [ %NextDestReg, NextMBB ]
3151 // %ThisSrcReg = phi [ %StartSrcReg, StartMBB ],
3152 // [ %NextSrcReg, NextMBB ]
3153 // %ThisCountReg = phi [ %StartCountReg, StartMBB ],
3154 // [ %NextCountReg, NextMBB ]
3155 // ( PFD 2, 768+DestDisp(%ThisDestReg) )
3156 // Opcode DestDisp(256,%ThisDestReg), SrcDisp(%ThisSrcReg)
3159 // The prefetch is used only for MVC. The JLH is used only for CLC.
3162 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisDestReg)
3163 .addReg(StartDestReg).addMBB(StartMBB)
3164 .addReg(NextDestReg).addMBB(NextMBB);
3165 if (!HaveSingleBase)
3166 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisSrcReg)
3167 .addReg(StartSrcReg).addMBB(StartMBB)
3168 .addReg(NextSrcReg).addMBB(NextMBB);
3169 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisCountReg)
3170 .addReg(StartCountReg).addMBB(StartMBB)
3171 .addReg(NextCountReg).addMBB(NextMBB);
3172 if (Opcode == SystemZ::MVC)
3173 BuildMI(MBB, DL, TII->get(SystemZ::PFD))
3174 .addImm(SystemZ::PFD_WRITE)
3175 .addReg(ThisDestReg).addImm(DestDisp + 768).addReg(0);
3176 BuildMI(MBB, DL, TII->get(Opcode))
3177 .addReg(ThisDestReg).addImm(DestDisp).addImm(256)
3178 .addReg(ThisSrcReg).addImm(SrcDisp);
3180 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3181 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3183 MBB->addSuccessor(EndMBB);
3184 MBB->addSuccessor(NextMBB);
3188 // %NextDestReg = LA 256(%ThisDestReg)
3189 // %NextSrcReg = LA 256(%ThisSrcReg)
3190 // %NextCountReg = AGHI %ThisCountReg, -1
3191 // CGHI %NextCountReg, 0
3193 // # fall through to DoneMMB
3195 // The AGHI, CGHI and JLH should be converted to BRCTG by later passes.
3198 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextDestReg)
3199 .addReg(ThisDestReg).addImm(256).addReg(0);
3200 if (!HaveSingleBase)
3201 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextSrcReg)
3202 .addReg(ThisSrcReg).addImm(256).addReg(0);
3203 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), NextCountReg)
3204 .addReg(ThisCountReg).addImm(-1);
3205 BuildMI(MBB, DL, TII->get(SystemZ::CGHI))
3206 .addReg(NextCountReg).addImm(0);
3207 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3208 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3210 MBB->addSuccessor(LoopMBB);
3211 MBB->addSuccessor(DoneMBB);
3213 DestBase = MachineOperand::CreateReg(NextDestReg, false);
3214 SrcBase = MachineOperand::CreateReg(NextSrcReg, false);
3218 // Handle any remaining bytes with straight-line code.
3219 while (Length > 0) {
3220 uint64_t ThisLength = std::min(Length, uint64_t(256));
3221 // The previous iteration might have created out-of-range displacements.
3222 // Apply them using LAY if so.
3223 if (!isUInt<12>(DestDisp)) {
3224 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3225 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3226 .addOperand(DestBase).addImm(DestDisp).addReg(0);
3227 DestBase = MachineOperand::CreateReg(Reg, false);
3230 if (!isUInt<12>(SrcDisp)) {
3231 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3232 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3233 .addOperand(SrcBase).addImm(SrcDisp).addReg(0);
3234 SrcBase = MachineOperand::CreateReg(Reg, false);
3237 BuildMI(*MBB, MI, DL, TII->get(Opcode))
3238 .addOperand(DestBase).addImm(DestDisp).addImm(ThisLength)
3239 .addOperand(SrcBase).addImm(SrcDisp);
3240 DestDisp += ThisLength;
3241 SrcDisp += ThisLength;
3242 Length -= ThisLength;
3243 // If there's another CLC to go, branch to the end if a difference
3245 if (EndMBB && Length > 0) {
3246 MachineBasicBlock *NextMBB = splitBlockBefore(MI, MBB);
3247 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3248 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3250 MBB->addSuccessor(EndMBB);
3251 MBB->addSuccessor(NextMBB);
3256 MBB->addSuccessor(EndMBB);
3258 MBB->addLiveIn(SystemZ::CC);
3261 MI->eraseFromParent();
3265 // Decompose string pseudo-instruction MI into a loop that continually performs
3266 // Opcode until CC != 3.
3268 SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
3269 MachineBasicBlock *MBB,
3270 unsigned Opcode) const {
3271 const SystemZInstrInfo *TII = TM.getInstrInfo();
3272 MachineFunction &MF = *MBB->getParent();
3273 MachineRegisterInfo &MRI = MF.getRegInfo();
3274 DebugLoc DL = MI->getDebugLoc();
3276 uint64_t End1Reg = MI->getOperand(0).getReg();
3277 uint64_t Start1Reg = MI->getOperand(1).getReg();
3278 uint64_t Start2Reg = MI->getOperand(2).getReg();
3279 uint64_t CharReg = MI->getOperand(3).getReg();
3281 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass;
3282 uint64_t This1Reg = MRI.createVirtualRegister(RC);
3283 uint64_t This2Reg = MRI.createVirtualRegister(RC);
3284 uint64_t End2Reg = MRI.createVirtualRegister(RC);
3286 MachineBasicBlock *StartMBB = MBB;
3287 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3288 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3291 // # fall through to LoopMMB
3292 MBB->addSuccessor(LoopMBB);
3295 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ]
3296 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ]
3298 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0L
3300 // # fall through to DoneMMB
3302 // The load of R0L can be hoisted by post-RA LICM.
3305 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg)
3306 .addReg(Start1Reg).addMBB(StartMBB)
3307 .addReg(End1Reg).addMBB(LoopMBB);
3308 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg)
3309 .addReg(Start2Reg).addMBB(StartMBB)
3310 .addReg(End2Reg).addMBB(LoopMBB);
3311 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0L).addReg(CharReg);
3312 BuildMI(MBB, DL, TII->get(Opcode))
3313 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define)
3314 .addReg(This1Reg).addReg(This2Reg);
3315 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3316 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB);
3317 MBB->addSuccessor(LoopMBB);
3318 MBB->addSuccessor(DoneMBB);
3320 DoneMBB->addLiveIn(SystemZ::CC);
3322 MI->eraseFromParent();
3326 MachineBasicBlock *SystemZTargetLowering::
3327 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
3328 switch (MI->getOpcode()) {
3329 case SystemZ::Select32Mux:
3330 case SystemZ::Select32:
3331 case SystemZ::SelectF32:
3332 case SystemZ::Select64:
3333 case SystemZ::SelectF64:
3334 case SystemZ::SelectF128:
3335 return emitSelect(MI, MBB);
3337 case SystemZ::CondStore8Mux:
3338 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, false);
3339 case SystemZ::CondStore8MuxInv:
3340 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, true);
3341 case SystemZ::CondStore16Mux:
3342 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, false);
3343 case SystemZ::CondStore16MuxInv:
3344 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, true);
3345 case SystemZ::CondStore8:
3346 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
3347 case SystemZ::CondStore8Inv:
3348 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
3349 case SystemZ::CondStore16:
3350 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
3351 case SystemZ::CondStore16Inv:
3352 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
3353 case SystemZ::CondStore32:
3354 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
3355 case SystemZ::CondStore32Inv:
3356 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
3357 case SystemZ::CondStore64:
3358 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
3359 case SystemZ::CondStore64Inv:
3360 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
3361 case SystemZ::CondStoreF32:
3362 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
3363 case SystemZ::CondStoreF32Inv:
3364 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
3365 case SystemZ::CondStoreF64:
3366 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
3367 case SystemZ::CondStoreF64Inv:
3368 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
3370 case SystemZ::AEXT128_64:
3371 return emitExt128(MI, MBB, false, SystemZ::subreg_l64);
3372 case SystemZ::ZEXT128_32:
3373 return emitExt128(MI, MBB, true, SystemZ::subreg_l32);
3374 case SystemZ::ZEXT128_64:
3375 return emitExt128(MI, MBB, true, SystemZ::subreg_l64);
3377 case SystemZ::ATOMIC_SWAPW:
3378 return emitAtomicLoadBinary(MI, MBB, 0, 0);
3379 case SystemZ::ATOMIC_SWAP_32:
3380 return emitAtomicLoadBinary(MI, MBB, 0, 32);
3381 case SystemZ::ATOMIC_SWAP_64:
3382 return emitAtomicLoadBinary(MI, MBB, 0, 64);
3384 case SystemZ::ATOMIC_LOADW_AR:
3385 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
3386 case SystemZ::ATOMIC_LOADW_AFI:
3387 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
3388 case SystemZ::ATOMIC_LOAD_AR:
3389 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
3390 case SystemZ::ATOMIC_LOAD_AHI:
3391 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
3392 case SystemZ::ATOMIC_LOAD_AFI:
3393 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
3394 case SystemZ::ATOMIC_LOAD_AGR:
3395 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
3396 case SystemZ::ATOMIC_LOAD_AGHI:
3397 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
3398 case SystemZ::ATOMIC_LOAD_AGFI:
3399 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
3401 case SystemZ::ATOMIC_LOADW_SR:
3402 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
3403 case SystemZ::ATOMIC_LOAD_SR:
3404 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
3405 case SystemZ::ATOMIC_LOAD_SGR:
3406 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
3408 case SystemZ::ATOMIC_LOADW_NR:
3409 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
3410 case SystemZ::ATOMIC_LOADW_NILH:
3411 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0);
3412 case SystemZ::ATOMIC_LOAD_NR:
3413 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
3414 case SystemZ::ATOMIC_LOAD_NILL:
3415 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32);
3416 case SystemZ::ATOMIC_LOAD_NILH:
3417 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32);
3418 case SystemZ::ATOMIC_LOAD_NILF:
3419 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32);
3420 case SystemZ::ATOMIC_LOAD_NGR:
3421 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
3422 case SystemZ::ATOMIC_LOAD_NILL64:
3423 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64);
3424 case SystemZ::ATOMIC_LOAD_NILH64:
3425 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64);
3426 case SystemZ::ATOMIC_LOAD_NIHL64:
3427 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64);
3428 case SystemZ::ATOMIC_LOAD_NIHH64:
3429 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64);
3430 case SystemZ::ATOMIC_LOAD_NILF64:
3431 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64);
3432 case SystemZ::ATOMIC_LOAD_NIHF64:
3433 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64);
3435 case SystemZ::ATOMIC_LOADW_OR:
3436 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
3437 case SystemZ::ATOMIC_LOADW_OILH:
3438 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 0);
3439 case SystemZ::ATOMIC_LOAD_OR:
3440 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
3441 case SystemZ::ATOMIC_LOAD_OILL:
3442 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 32);
3443 case SystemZ::ATOMIC_LOAD_OILH:
3444 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 32);
3445 case SystemZ::ATOMIC_LOAD_OILF:
3446 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 32);
3447 case SystemZ::ATOMIC_LOAD_OGR:
3448 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
3449 case SystemZ::ATOMIC_LOAD_OILL64:
3450 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL64, 64);
3451 case SystemZ::ATOMIC_LOAD_OILH64:
3452 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH64, 64);
3453 case SystemZ::ATOMIC_LOAD_OIHL64:
3454 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL64, 64);
3455 case SystemZ::ATOMIC_LOAD_OIHH64:
3456 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH64, 64);
3457 case SystemZ::ATOMIC_LOAD_OILF64:
3458 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF64, 64);
3459 case SystemZ::ATOMIC_LOAD_OIHF64:
3460 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF64, 64);
3462 case SystemZ::ATOMIC_LOADW_XR:
3463 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
3464 case SystemZ::ATOMIC_LOADW_XILF:
3465 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 0);
3466 case SystemZ::ATOMIC_LOAD_XR:
3467 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
3468 case SystemZ::ATOMIC_LOAD_XILF:
3469 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 32);
3470 case SystemZ::ATOMIC_LOAD_XGR:
3471 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
3472 case SystemZ::ATOMIC_LOAD_XILF64:
3473 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF64, 64);
3474 case SystemZ::ATOMIC_LOAD_XIHF64:
3475 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF64, 64);
3477 case SystemZ::ATOMIC_LOADW_NRi:
3478 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
3479 case SystemZ::ATOMIC_LOADW_NILHi:
3480 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0, true);
3481 case SystemZ::ATOMIC_LOAD_NRi:
3482 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
3483 case SystemZ::ATOMIC_LOAD_NILLi:
3484 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32, true);
3485 case SystemZ::ATOMIC_LOAD_NILHi:
3486 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32, true);
3487 case SystemZ::ATOMIC_LOAD_NILFi:
3488 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32, true);
3489 case SystemZ::ATOMIC_LOAD_NGRi:
3490 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
3491 case SystemZ::ATOMIC_LOAD_NILL64i:
3492 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64, true);
3493 case SystemZ::ATOMIC_LOAD_NILH64i:
3494 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64, true);
3495 case SystemZ::ATOMIC_LOAD_NIHL64i:
3496 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64, true);
3497 case SystemZ::ATOMIC_LOAD_NIHH64i:
3498 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64, true);
3499 case SystemZ::ATOMIC_LOAD_NILF64i:
3500 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64, true);
3501 case SystemZ::ATOMIC_LOAD_NIHF64i:
3502 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64, true);
3504 case SystemZ::ATOMIC_LOADW_MIN:
3505 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3506 SystemZ::CCMASK_CMP_LE, 0);
3507 case SystemZ::ATOMIC_LOAD_MIN_32:
3508 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3509 SystemZ::CCMASK_CMP_LE, 32);
3510 case SystemZ::ATOMIC_LOAD_MIN_64:
3511 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3512 SystemZ::CCMASK_CMP_LE, 64);
3514 case SystemZ::ATOMIC_LOADW_MAX:
3515 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3516 SystemZ::CCMASK_CMP_GE, 0);
3517 case SystemZ::ATOMIC_LOAD_MAX_32:
3518 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3519 SystemZ::CCMASK_CMP_GE, 32);
3520 case SystemZ::ATOMIC_LOAD_MAX_64:
3521 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3522 SystemZ::CCMASK_CMP_GE, 64);
3524 case SystemZ::ATOMIC_LOADW_UMIN:
3525 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3526 SystemZ::CCMASK_CMP_LE, 0);
3527 case SystemZ::ATOMIC_LOAD_UMIN_32:
3528 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3529 SystemZ::CCMASK_CMP_LE, 32);
3530 case SystemZ::ATOMIC_LOAD_UMIN_64:
3531 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3532 SystemZ::CCMASK_CMP_LE, 64);
3534 case SystemZ::ATOMIC_LOADW_UMAX:
3535 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3536 SystemZ::CCMASK_CMP_GE, 0);
3537 case SystemZ::ATOMIC_LOAD_UMAX_32:
3538 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3539 SystemZ::CCMASK_CMP_GE, 32);
3540 case SystemZ::ATOMIC_LOAD_UMAX_64:
3541 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3542 SystemZ::CCMASK_CMP_GE, 64);
3544 case SystemZ::ATOMIC_CMP_SWAPW:
3545 return emitAtomicCmpSwapW(MI, MBB);
3546 case SystemZ::MVCSequence:
3547 case SystemZ::MVCLoop:
3548 return emitMemMemWrapper(MI, MBB, SystemZ::MVC);
3549 case SystemZ::NCSequence:
3550 case SystemZ::NCLoop:
3551 return emitMemMemWrapper(MI, MBB, SystemZ::NC);
3552 case SystemZ::OCSequence:
3553 case SystemZ::OCLoop:
3554 return emitMemMemWrapper(MI, MBB, SystemZ::OC);
3555 case SystemZ::XCSequence:
3556 case SystemZ::XCLoop:
3557 return emitMemMemWrapper(MI, MBB, SystemZ::XC);
3558 case SystemZ::CLCSequence:
3559 case SystemZ::CLCLoop:
3560 return emitMemMemWrapper(MI, MBB, SystemZ::CLC);
3561 case SystemZ::CLSTLoop:
3562 return emitStringWrapper(MI, MBB, SystemZ::CLST);
3563 case SystemZ::MVSTLoop:
3564 return emitStringWrapper(MI, MBB, SystemZ::MVST);
3565 case SystemZ::SRSTLoop:
3566 return emitStringWrapper(MI, MBB, SystemZ::SRST);
3568 llvm_unreachable("Unexpected instr type to insert");