1 //===-- SystemZISelLowering.cpp - SystemZ DAG lowering implementation -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SystemZTargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "SystemZISelLowering.h"
15 #include "SystemZCallingConv.h"
16 #include "SystemZConstantPoolValue.h"
17 #include "SystemZMachineFunctionInfo.h"
18 #include "SystemZTargetMachine.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
27 #define DEBUG_TYPE "systemz-lower"
30 // Represents a sequence for extracting a 0/1 value from an IPM result:
31 // (((X ^ XORValue) + AddValue) >> Bit)
32 struct IPMConversion {
33 IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
34 : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
41 // Represents information about a comparison.
43 Comparison(SDValue Op0In, SDValue Op1In)
44 : Op0(Op0In), Op1(Op1In), Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
46 // The operands to the comparison.
49 // The opcode that should be used to compare Op0 and Op1.
52 // A SystemZICMP value. Only used for integer comparisons.
55 // The mask of CC values that Opcode can produce.
58 // The mask of CC values for which the original condition is true.
61 } // end anonymous namespace
63 // Classify VT as either 32 or 64 bit.
64 static bool is32Bit(EVT VT) {
65 switch (VT.getSimpleVT().SimpleTy) {
71 llvm_unreachable("Unsupported type");
75 // Return a version of MachineOperand that can be safely used before the
77 static MachineOperand earlyUseOperand(MachineOperand Op) {
83 SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &tm)
84 : TargetLowering(tm, new TargetLoweringObjectFileELF()),
85 Subtarget(tm.getSubtarget<SystemZSubtarget>()) {
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 // Only z196 and above have native support for conversions to unsigned.
180 if (!Subtarget.hasFPExtension())
181 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
185 // Type legalization will convert 8- and 16-bit atomic operations into
186 // forms that operate on i32s (but still keeping the original memory VT).
187 // Lower them into full i32 operations.
188 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
189 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
190 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
191 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
192 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
193 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
194 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
195 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
196 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
197 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
198 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
199 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
201 // z10 has instructions for signed but not unsigned FP conversion.
202 // Handle unsigned 32-bit types as signed 64-bit types.
203 if (!Subtarget.hasFPExtension()) {
204 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
205 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
208 // We have native support for a 64-bit CTLZ, via FLOGR.
209 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
210 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
212 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
213 setOperationAction(ISD::OR, MVT::i64, Custom);
215 // FIXME: Can we support these natively?
216 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
217 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
218 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
220 // We have native instructions for i8, i16 and i32 extensions, but not i1.
221 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 // Codes for which we want to perform some z-specific combinations.
297 setTargetDAGCombine(ISD::SIGN_EXTEND);
299 // We want to use MVC in preference to even a single load/store pair.
300 MaxStoresPerMemcpy = 0;
301 MaxStoresPerMemcpyOptSize = 0;
303 // The main memset sequence is a byte store followed by an MVC.
304 // Two STC or MV..I stores win over that, but the kind of fused stores
305 // generated by target-independent code don't when the byte value is
306 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
307 // than "STC;MVC". Handle the choice in target-specific code instead.
308 MaxStoresPerMemset = 0;
309 MaxStoresPerMemsetOptSize = 0;
312 EVT SystemZTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
315 return VT.changeVectorElementTypeToInteger();
318 bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
319 VT = VT.getScalarType();
324 switch (VT.getSimpleVT().SimpleTy) {
337 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
338 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
339 return Imm.isZero() || Imm.isNegZero();
342 bool SystemZTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
346 // Unaligned accesses should never be slower than the expanded version.
347 // We check specifically for aligned accesses in the few cases where
348 // they are required.
354 bool SystemZTargetLowering::isLegalAddressingMode(const AddrMode &AM,
356 // Punt on globals for now, although they can be used in limited
357 // RELATIVE LONG cases.
361 // Require a 20-bit signed offset.
362 if (!isInt<20>(AM.BaseOffs))
365 // Indexing is OK but no scale factor can be applied.
366 return AM.Scale == 0 || AM.Scale == 1;
369 bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
370 if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
372 unsigned FromBits = FromType->getPrimitiveSizeInBits();
373 unsigned ToBits = ToType->getPrimitiveSizeInBits();
374 return FromBits > ToBits;
377 bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
378 if (!FromVT.isInteger() || !ToVT.isInteger())
380 unsigned FromBits = FromVT.getSizeInBits();
381 unsigned ToBits = ToVT.getSizeInBits();
382 return FromBits > ToBits;
385 //===----------------------------------------------------------------------===//
386 // Inline asm support
387 //===----------------------------------------------------------------------===//
389 TargetLowering::ConstraintType
390 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
391 if (Constraint.size() == 1) {
392 switch (Constraint[0]) {
393 case 'a': // Address register
394 case 'd': // Data register (equivalent to 'r')
395 case 'f': // Floating-point register
396 case 'h': // High-part register
397 case 'r': // General-purpose register
398 return C_RegisterClass;
400 case 'Q': // Memory with base and unsigned 12-bit displacement
401 case 'R': // Likewise, plus an index
402 case 'S': // Memory with base and signed 20-bit displacement
403 case 'T': // Likewise, plus an index
404 case 'm': // Equivalent to 'T'.
407 case 'I': // Unsigned 8-bit constant
408 case 'J': // Unsigned 12-bit constant
409 case 'K': // Signed 16-bit constant
410 case 'L': // Signed 20-bit displacement (on all targets we support)
411 case 'M': // 0x7fffffff
418 return TargetLowering::getConstraintType(Constraint);
421 TargetLowering::ConstraintWeight SystemZTargetLowering::
422 getSingleConstraintMatchWeight(AsmOperandInfo &info,
423 const char *constraint) const {
424 ConstraintWeight weight = CW_Invalid;
425 Value *CallOperandVal = info.CallOperandVal;
426 // If we don't have a value, we can't do a match,
427 // but allow it at the lowest weight.
430 Type *type = CallOperandVal->getType();
431 // Look at the constraint type.
432 switch (*constraint) {
434 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
437 case 'a': // Address register
438 case 'd': // Data register (equivalent to 'r')
439 case 'h': // High-part register
440 case 'r': // General-purpose register
441 if (CallOperandVal->getType()->isIntegerTy())
442 weight = CW_Register;
445 case 'f': // Floating-point register
446 if (type->isFloatingPointTy())
447 weight = CW_Register;
450 case 'I': // Unsigned 8-bit constant
451 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
452 if (isUInt<8>(C->getZExtValue()))
453 weight = CW_Constant;
456 case 'J': // Unsigned 12-bit constant
457 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
458 if (isUInt<12>(C->getZExtValue()))
459 weight = CW_Constant;
462 case 'K': // Signed 16-bit constant
463 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
464 if (isInt<16>(C->getSExtValue()))
465 weight = CW_Constant;
468 case 'L': // Signed 20-bit displacement (on all targets we support)
469 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
470 if (isInt<20>(C->getSExtValue()))
471 weight = CW_Constant;
474 case 'M': // 0x7fffffff
475 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
476 if (C->getZExtValue() == 0x7fffffff)
477 weight = CW_Constant;
483 // Parse a "{tNNN}" register constraint for which the register type "t"
484 // has already been verified. MC is the class associated with "t" and
485 // Map maps 0-based register numbers to LLVM register numbers.
486 static std::pair<unsigned, const TargetRegisterClass *>
487 parseRegisterNumber(const std::string &Constraint,
488 const TargetRegisterClass *RC, const unsigned *Map) {
489 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
490 if (isdigit(Constraint[2])) {
491 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2);
492 unsigned Index = atoi(Suffix.c_str());
493 if (Index < 16 && Map[Index])
494 return std::make_pair(Map[Index], RC);
496 return std::make_pair(0U, nullptr);
499 std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering::
500 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const {
501 if (Constraint.size() == 1) {
502 // GCC Constraint Letters
503 switch (Constraint[0]) {
505 case 'd': // Data register (equivalent to 'r')
506 case 'r': // General-purpose register
508 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
509 else if (VT == MVT::i128)
510 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
511 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
513 case 'a': // Address register
515 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
516 else if (VT == MVT::i128)
517 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
518 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
520 case 'h': // High-part register (an LLVM extension)
521 return std::make_pair(0U, &SystemZ::GRH32BitRegClass);
523 case 'f': // Floating-point register
525 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
526 else if (VT == MVT::f128)
527 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
528 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
531 if (Constraint[0] == '{') {
532 // We need to override the default register parsing for GPRs and FPRs
533 // because the interpretation depends on VT. The internal names of
534 // the registers are also different from the external names
535 // (F0D and F0S instead of F0, etc.).
536 if (Constraint[1] == 'r') {
538 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
539 SystemZMC::GR32Regs);
541 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
542 SystemZMC::GR128Regs);
543 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
544 SystemZMC::GR64Regs);
546 if (Constraint[1] == 'f') {
548 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
549 SystemZMC::FP32Regs);
551 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
552 SystemZMC::FP128Regs);
553 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
554 SystemZMC::FP64Regs);
557 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
560 void SystemZTargetLowering::
561 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
562 std::vector<SDValue> &Ops,
563 SelectionDAG &DAG) const {
564 // Only support length 1 constraints for now.
565 if (Constraint.length() == 1) {
566 switch (Constraint[0]) {
567 case 'I': // Unsigned 8-bit constant
568 if (auto *C = dyn_cast<ConstantSDNode>(Op))
569 if (isUInt<8>(C->getZExtValue()))
570 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
574 case 'J': // Unsigned 12-bit constant
575 if (auto *C = dyn_cast<ConstantSDNode>(Op))
576 if (isUInt<12>(C->getZExtValue()))
577 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
581 case 'K': // Signed 16-bit constant
582 if (auto *C = dyn_cast<ConstantSDNode>(Op))
583 if (isInt<16>(C->getSExtValue()))
584 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
588 case 'L': // Signed 20-bit displacement (on all targets we support)
589 if (auto *C = dyn_cast<ConstantSDNode>(Op))
590 if (isInt<20>(C->getSExtValue()))
591 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
595 case 'M': // 0x7fffffff
596 if (auto *C = dyn_cast<ConstantSDNode>(Op))
597 if (C->getZExtValue() == 0x7fffffff)
598 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
603 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
606 //===----------------------------------------------------------------------===//
607 // Calling conventions
608 //===----------------------------------------------------------------------===//
610 #include "SystemZGenCallingConv.inc"
612 bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
613 Type *ToType) const {
614 return isTruncateFree(FromType, ToType);
617 bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
618 if (!CI->isTailCall())
623 // Value is a value that has been passed to us in the location described by VA
624 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
625 // any loads onto Chain.
626 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
627 CCValAssign &VA, SDValue Chain,
629 // If the argument has been promoted from a smaller type, insert an
630 // assertion to capture this.
631 if (VA.getLocInfo() == CCValAssign::SExt)
632 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
633 DAG.getValueType(VA.getValVT()));
634 else if (VA.getLocInfo() == CCValAssign::ZExt)
635 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
636 DAG.getValueType(VA.getValVT()));
639 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
640 else if (VA.getLocInfo() == CCValAssign::Indirect)
641 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
642 MachinePointerInfo(), false, false, false, 0);
644 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
648 // Value is a value of type VA.getValVT() that we need to copy into
649 // the location described by VA. Return a copy of Value converted to
650 // VA.getValVT(). The caller is responsible for handling indirect values.
651 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
652 CCValAssign &VA, SDValue Value) {
653 switch (VA.getLocInfo()) {
654 case CCValAssign::SExt:
655 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
656 case CCValAssign::ZExt:
657 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
658 case CCValAssign::AExt:
659 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
660 case CCValAssign::Full:
663 llvm_unreachable("Unhandled getLocInfo()");
667 SDValue SystemZTargetLowering::
668 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
669 const SmallVectorImpl<ISD::InputArg> &Ins,
670 SDLoc DL, SelectionDAG &DAG,
671 SmallVectorImpl<SDValue> &InVals) const {
672 MachineFunction &MF = DAG.getMachineFunction();
673 MachineFrameInfo *MFI = MF.getFrameInfo();
674 MachineRegisterInfo &MRI = MF.getRegInfo();
675 SystemZMachineFunctionInfo *FuncInfo =
676 MF.getInfo<SystemZMachineFunctionInfo>();
677 auto *TFL = static_cast<const SystemZFrameLowering *>(
678 DAG.getTarget().getSubtargetImpl()->getFrameLowering());
680 // Assign locations to all of the incoming arguments.
681 SmallVector<CCValAssign, 16> ArgLocs;
682 CCState CCInfo(CallConv, IsVarArg, MF, DAG.getTarget(), ArgLocs,
684 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
686 unsigned NumFixedGPRs = 0;
687 unsigned NumFixedFPRs = 0;
688 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
690 CCValAssign &VA = ArgLocs[I];
691 EVT LocVT = VA.getLocVT();
693 // Arguments passed in registers
694 const TargetRegisterClass *RC;
695 switch (LocVT.getSimpleVT().SimpleTy) {
697 // Integers smaller than i64 should be promoted to i64.
698 llvm_unreachable("Unexpected argument type");
701 RC = &SystemZ::GR32BitRegClass;
705 RC = &SystemZ::GR64BitRegClass;
709 RC = &SystemZ::FP32BitRegClass;
713 RC = &SystemZ::FP64BitRegClass;
717 unsigned VReg = MRI.createVirtualRegister(RC);
718 MRI.addLiveIn(VA.getLocReg(), VReg);
719 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
721 assert(VA.isMemLoc() && "Argument not register or memory");
723 // Create the frame index object for this incoming parameter.
724 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
725 VA.getLocMemOffset(), true);
727 // Create the SelectionDAG nodes corresponding to a load
728 // from this parameter. Unpromoted ints and floats are
729 // passed as right-justified 8-byte values.
730 EVT PtrVT = getPointerTy();
731 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
732 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
733 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4));
734 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
735 MachinePointerInfo::getFixedStack(FI),
736 false, false, false, 0);
739 // Convert the value of the argument register into the value that's
741 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
745 // Save the number of non-varargs registers for later use by va_start, etc.
746 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
747 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
749 // Likewise the address (in the form of a frame index) of where the
750 // first stack vararg would be. The 1-byte size here is arbitrary.
751 int64_t StackSize = CCInfo.getNextStackOffset();
752 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
754 // ...and a similar frame index for the caller-allocated save area
755 // that will be used to store the incoming registers.
756 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
757 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
758 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
760 // Store the FPR varargs in the reserved frame slots. (We store the
761 // GPRs as part of the prologue.)
762 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
763 SDValue MemOps[SystemZ::NumArgFPRs];
764 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
765 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
766 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
767 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
768 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
769 &SystemZ::FP64BitRegClass);
770 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
771 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
772 MachinePointerInfo::getFixedStack(FI),
776 // Join the stores, which are independent of one another.
777 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
778 makeArrayRef(&MemOps[NumFixedFPRs],
779 SystemZ::NumArgFPRs-NumFixedFPRs));
786 static bool canUseSiblingCall(CCState ArgCCInfo,
787 SmallVectorImpl<CCValAssign> &ArgLocs) {
788 // Punt if there are any indirect or stack arguments, or if the call
789 // needs the call-saved argument register R6.
790 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
791 CCValAssign &VA = ArgLocs[I];
792 if (VA.getLocInfo() == CCValAssign::Indirect)
796 unsigned Reg = VA.getLocReg();
797 if (Reg == SystemZ::R6H || Reg == SystemZ::R6L || Reg == SystemZ::R6D)
804 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
805 SmallVectorImpl<SDValue> &InVals) const {
806 SelectionDAG &DAG = CLI.DAG;
808 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
809 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
810 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
811 SDValue Chain = CLI.Chain;
812 SDValue Callee = CLI.Callee;
813 bool &IsTailCall = CLI.IsTailCall;
814 CallingConv::ID CallConv = CLI.CallConv;
815 bool IsVarArg = CLI.IsVarArg;
816 MachineFunction &MF = DAG.getMachineFunction();
817 EVT PtrVT = getPointerTy();
819 // Analyze the operands of the call, assigning locations to each operand.
820 SmallVector<CCValAssign, 16> ArgLocs;
821 CCState ArgCCInfo(CallConv, IsVarArg, MF, DAG.getTarget(), ArgLocs,
823 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
825 // We don't support GuaranteedTailCallOpt, only automatically-detected
827 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs))
830 // Get a count of how many bytes are to be pushed on the stack.
831 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
833 // Mark the start of the call.
835 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true),
838 // Copy argument values to their designated locations.
839 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
840 SmallVector<SDValue, 8> MemOpChains;
842 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
843 CCValAssign &VA = ArgLocs[I];
844 SDValue ArgValue = OutVals[I];
846 if (VA.getLocInfo() == CCValAssign::Indirect) {
847 // Store the argument in a stack slot and pass its address.
848 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
849 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
850 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
851 MachinePointerInfo::getFixedStack(FI),
853 ArgValue = SpillSlot;
855 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
858 // Queue up the argument copies and emit them at the end.
859 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
861 assert(VA.isMemLoc() && "Argument not register or memory");
863 // Work out the address of the stack slot. Unpromoted ints and
864 // floats are passed as right-justified 8-byte values.
865 if (!StackPtr.getNode())
866 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
867 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
868 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
870 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
871 DAG.getIntPtrConstant(Offset));
874 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
875 MachinePointerInfo(),
880 // Join the stores, which are independent of one another.
881 if (!MemOpChains.empty())
882 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
884 // Accept direct calls by converting symbolic call addresses to the
885 // associated Target* opcodes. Force %r1 to be used for indirect
888 if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
889 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
890 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
891 } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
892 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
893 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
894 } else if (IsTailCall) {
895 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
896 Glue = Chain.getValue(1);
897 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
900 // Build a sequence of copy-to-reg nodes, chained and glued together.
901 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
902 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
903 RegsToPass[I].second, Glue);
904 Glue = Chain.getValue(1);
907 // The first call operand is the chain and the second is the target address.
908 SmallVector<SDValue, 8> Ops;
909 Ops.push_back(Chain);
910 Ops.push_back(Callee);
912 // Add argument registers to the end of the list so that they are
913 // known live into the call.
914 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
915 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
916 RegsToPass[I].second.getValueType()));
918 // Add a register mask operand representing the call-preserved registers.
919 const TargetRegisterInfo *TRI =
920 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
921 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
922 assert(Mask && "Missing call preserved mask for calling convention");
923 Ops.push_back(DAG.getRegisterMask(Mask));
925 // Glue the call to the argument copies, if any.
930 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
932 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, Ops);
933 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, Ops);
934 Glue = Chain.getValue(1);
936 // Mark the end of the call, which is glued to the call itself.
937 Chain = DAG.getCALLSEQ_END(Chain,
938 DAG.getConstant(NumBytes, PtrVT, true),
939 DAG.getConstant(0, PtrVT, true),
941 Glue = Chain.getValue(1);
943 // Assign locations to each value returned by this call.
944 SmallVector<CCValAssign, 16> RetLocs;
945 CCState RetCCInfo(CallConv, IsVarArg, MF, DAG.getTarget(), RetLocs,
947 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
949 // Copy all of the result registers out of their specified physreg.
950 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
951 CCValAssign &VA = RetLocs[I];
953 // Copy the value out, gluing the copy to the end of the call sequence.
954 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
955 VA.getLocVT(), Glue);
956 Chain = RetValue.getValue(1);
957 Glue = RetValue.getValue(2);
959 // Convert the value of the return register into the value that's
961 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
968 SystemZTargetLowering::LowerReturn(SDValue Chain,
969 CallingConv::ID CallConv, bool IsVarArg,
970 const SmallVectorImpl<ISD::OutputArg> &Outs,
971 const SmallVectorImpl<SDValue> &OutVals,
972 SDLoc DL, SelectionDAG &DAG) const {
973 MachineFunction &MF = DAG.getMachineFunction();
975 // Assign locations to each returned value.
976 SmallVector<CCValAssign, 16> RetLocs;
977 CCState RetCCInfo(CallConv, IsVarArg, MF, DAG.getTarget(), RetLocs,
979 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
981 // Quick exit for void returns
983 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
985 // Copy the result values into the output registers.
987 SmallVector<SDValue, 4> RetOps;
988 RetOps.push_back(Chain);
989 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
990 CCValAssign &VA = RetLocs[I];
991 SDValue RetValue = OutVals[I];
993 // Make the return register live on exit.
994 assert(VA.isRegLoc() && "Can only return in registers!");
996 // Promote the value as required.
997 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
999 // Chain and glue the copies together.
1000 unsigned Reg = VA.getLocReg();
1001 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
1002 Glue = Chain.getValue(1);
1003 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
1006 // Update chain and glue.
1009 RetOps.push_back(Glue);
1011 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, RetOps);
1014 SDValue SystemZTargetLowering::
1015 prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, SelectionDAG &DAG) const {
1016 return DAG.getNode(SystemZISD::SERIALIZE, DL, MVT::Other, Chain);
1019 // CC is a comparison that will be implemented using an integer or
1020 // floating-point comparison. Return the condition code mask for
1021 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
1022 // unsigned comparisons and clear for signed ones. In the floating-point
1023 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
1024 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
1026 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
1027 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
1028 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
1032 llvm_unreachable("Invalid integer condition!");
1041 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
1042 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
1047 // Return a sequence for getting a 1 from an IPM result when CC has a
1048 // value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
1049 // The handling of CC values outside CCValid doesn't matter.
1050 static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
1051 // Deal with cases where the result can be taken directly from a bit
1052 // of the IPM result.
1053 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
1054 return IPMConversion(0, 0, SystemZ::IPM_CC);
1055 if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
1056 return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
1058 // Deal with cases where we can add a value to force the sign bit
1059 // to contain the right value. Putting the bit in 31 means we can
1060 // use SRL rather than RISBG(L), and also makes it easier to get a
1061 // 0/-1 value, so it has priority over the other tests below.
1063 // These sequences rely on the fact that the upper two bits of the
1064 // IPM result are zero.
1065 uint64_t TopBit = uint64_t(1) << 31;
1066 if (CCMask == (CCValid & SystemZ::CCMASK_0))
1067 return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
1068 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
1069 return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
1070 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1072 | SystemZ::CCMASK_2)))
1073 return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
1074 if (CCMask == (CCValid & SystemZ::CCMASK_3))
1075 return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
1076 if (CCMask == (CCValid & (SystemZ::CCMASK_1
1078 | SystemZ::CCMASK_3)))
1079 return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
1081 // Next try inverting the value and testing a bit. 0/1 could be
1082 // handled this way too, but we dealt with that case above.
1083 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
1084 return IPMConversion(-1, 0, SystemZ::IPM_CC);
1086 // Handle cases where adding a value forces a non-sign bit to contain
1088 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
1089 return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
1090 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
1091 return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
1093 // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
1094 // can be done by inverting the low CC bit and applying one of the
1095 // sign-based extractions above.
1096 if (CCMask == (CCValid & SystemZ::CCMASK_1))
1097 return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
1098 if (CCMask == (CCValid & SystemZ::CCMASK_2))
1099 return IPMConversion(1 << SystemZ::IPM_CC,
1100 TopBit - (3 << SystemZ::IPM_CC), 31);
1101 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1103 | SystemZ::CCMASK_3)))
1104 return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
1105 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1107 | SystemZ::CCMASK_3)))
1108 return IPMConversion(1 << SystemZ::IPM_CC,
1109 TopBit - (1 << SystemZ::IPM_CC), 31);
1111 llvm_unreachable("Unexpected CC combination");
1114 // If C can be converted to a comparison against zero, adjust the operands
1116 static void adjustZeroCmp(SelectionDAG &DAG, Comparison &C) {
1117 if (C.ICmpType == SystemZICMP::UnsignedOnly)
1120 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
1124 int64_t Value = ConstOp1->getSExtValue();
1125 if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
1126 (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
1127 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
1128 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
1129 C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1130 C.Op1 = DAG.getConstant(0, C.Op1.getValueType());
1134 // If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
1135 // adjust the operands as necessary.
1136 static void adjustSubwordCmp(SelectionDAG &DAG, Comparison &C) {
1137 // For us to make any changes, it must a comparison between a single-use
1138 // load and a constant.
1139 if (!C.Op0.hasOneUse() ||
1140 C.Op0.getOpcode() != ISD::LOAD ||
1141 C.Op1.getOpcode() != ISD::Constant)
1144 // We must have an 8- or 16-bit load.
1145 auto *Load = cast<LoadSDNode>(C.Op0);
1146 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
1147 if (NumBits != 8 && NumBits != 16)
1150 // The load must be an extending one and the constant must be within the
1151 // range of the unextended value.
1152 auto *ConstOp1 = cast<ConstantSDNode>(C.Op1);
1153 uint64_t Value = ConstOp1->getZExtValue();
1154 uint64_t Mask = (1 << NumBits) - 1;
1155 if (Load->getExtensionType() == ISD::SEXTLOAD) {
1156 // Make sure that ConstOp1 is in range of C.Op0.
1157 int64_t SignedValue = ConstOp1->getSExtValue();
1158 if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
1160 if (C.ICmpType != SystemZICMP::SignedOnly) {
1161 // Unsigned comparison between two sign-extended values is equivalent
1162 // to unsigned comparison between two zero-extended values.
1164 } else if (NumBits == 8) {
1165 // Try to treat the comparison as unsigned, so that we can use CLI.
1166 // Adjust CCMask and Value as necessary.
1167 if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
1168 // Test whether the high bit of the byte is set.
1169 Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
1170 else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
1171 // Test whether the high bit of the byte is clear.
1172 Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
1174 // No instruction exists for this combination.
1176 C.ICmpType = SystemZICMP::UnsignedOnly;
1178 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
1181 assert(C.ICmpType == SystemZICMP::Any &&
1182 "Signedness shouldn't matter here.");
1186 // Make sure that the first operand is an i32 of the right extension type.
1187 ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
1190 if (C.Op0.getValueType() != MVT::i32 ||
1191 Load->getExtensionType() != ExtType)
1192 C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
1193 Load->getChain(), Load->getBasePtr(),
1194 Load->getPointerInfo(), Load->getMemoryVT(),
1195 Load->isVolatile(), Load->isNonTemporal(),
1196 Load->isInvariant(), Load->getAlignment());
1198 // Make sure that the second operand is an i32 with the right value.
1199 if (C.Op1.getValueType() != MVT::i32 ||
1200 Value != ConstOp1->getZExtValue())
1201 C.Op1 = DAG.getConstant(Value, MVT::i32);
1204 // Return true if Op is either an unextended load, or a load suitable
1205 // for integer register-memory comparisons of type ICmpType.
1206 static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
1207 auto *Load = dyn_cast<LoadSDNode>(Op.getNode());
1209 // There are no instructions to compare a register with a memory byte.
1210 if (Load->getMemoryVT() == MVT::i8)
1212 // Otherwise decide on extension type.
1213 switch (Load->getExtensionType()) {
1214 case ISD::NON_EXTLOAD:
1217 return ICmpType != SystemZICMP::UnsignedOnly;
1219 return ICmpType != SystemZICMP::SignedOnly;
1227 // Return true if it is better to swap the operands of C.
1228 static bool shouldSwapCmpOperands(const Comparison &C) {
1229 // Leave f128 comparisons alone, since they have no memory forms.
1230 if (C.Op0.getValueType() == MVT::f128)
1233 // Always keep a floating-point constant second, since comparisons with
1234 // zero can use LOAD TEST and comparisons with other constants make a
1235 // natural memory operand.
1236 if (isa<ConstantFPSDNode>(C.Op1))
1239 // Never swap comparisons with zero since there are many ways to optimize
1241 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1242 if (ConstOp1 && ConstOp1->getZExtValue() == 0)
1245 // Also keep natural memory operands second if the loaded value is
1246 // only used here. Several comparisons have memory forms.
1247 if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
1250 // Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
1251 // In that case we generally prefer the memory to be second.
1252 if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
1253 // The only exceptions are when the second operand is a constant and
1254 // we can use things like CHHSI.
1257 // The unsigned memory-immediate instructions can handle 16-bit
1258 // unsigned integers.
1259 if (C.ICmpType != SystemZICMP::SignedOnly &&
1260 isUInt<16>(ConstOp1->getZExtValue()))
1262 // The signed memory-immediate instructions can handle 16-bit
1264 if (C.ICmpType != SystemZICMP::UnsignedOnly &&
1265 isInt<16>(ConstOp1->getSExtValue()))
1270 // Try to promote the use of CGFR and CLGFR.
1271 unsigned Opcode0 = C.Op0.getOpcode();
1272 if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
1274 if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
1276 if (C.ICmpType != SystemZICMP::SignedOnly &&
1277 Opcode0 == ISD::AND &&
1278 C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
1279 cast<ConstantSDNode>(C.Op0.getOperand(1))->getZExtValue() == 0xffffffff)
1285 // Return a version of comparison CC mask CCMask in which the LT and GT
1286 // actions are swapped.
1287 static unsigned reverseCCMask(unsigned CCMask) {
1288 return ((CCMask & SystemZ::CCMASK_CMP_EQ) |
1289 (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
1290 (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
1291 (CCMask & SystemZ::CCMASK_CMP_UO));
1294 // Check whether C tests for equality between X and Y and whether X - Y
1295 // or Y - X is also computed. In that case it's better to compare the
1296 // result of the subtraction against zero.
1297 static void adjustForSubtraction(SelectionDAG &DAG, Comparison &C) {
1298 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1299 C.CCMask == SystemZ::CCMASK_CMP_NE) {
1300 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1302 if (N->getOpcode() == ISD::SUB &&
1303 ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
1304 (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
1305 C.Op0 = SDValue(N, 0);
1306 C.Op1 = DAG.getConstant(0, N->getValueType(0));
1313 // Check whether C compares a floating-point value with zero and if that
1314 // floating-point value is also negated. In this case we can use the
1315 // negation to set CC, so avoiding separate LOAD AND TEST and
1316 // LOAD (NEGATIVE/COMPLEMENT) instructions.
1317 static void adjustForFNeg(Comparison &C) {
1318 auto *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
1319 if (C1 && C1->isZero()) {
1320 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1322 if (N->getOpcode() == ISD::FNEG) {
1323 C.Op0 = SDValue(N, 0);
1324 C.CCMask = reverseCCMask(C.CCMask);
1331 // Check whether C compares (shl X, 32) with 0 and whether X is
1332 // also sign-extended. In that case it is better to test the result
1333 // of the sign extension using LTGFR.
1335 // This case is important because InstCombine transforms a comparison
1336 // with (sext (trunc X)) into a comparison with (shl X, 32).
1337 static void adjustForLTGFR(Comparison &C) {
1338 // Check for a comparison between (shl X, 32) and 0.
1339 if (C.Op0.getOpcode() == ISD::SHL &&
1340 C.Op0.getValueType() == MVT::i64 &&
1341 C.Op1.getOpcode() == ISD::Constant &&
1342 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1343 auto *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
1344 if (C1 && C1->getZExtValue() == 32) {
1345 SDValue ShlOp0 = C.Op0.getOperand(0);
1346 // See whether X has any SIGN_EXTEND_INREG uses.
1347 for (auto I = ShlOp0->use_begin(), E = ShlOp0->use_end(); I != E; ++I) {
1349 if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
1350 cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
1351 C.Op0 = SDValue(N, 0);
1359 // If C compares the truncation of an extending load, try to compare
1360 // the untruncated value instead. This exposes more opportunities to
1362 static void adjustICmpTruncate(SelectionDAG &DAG, Comparison &C) {
1363 if (C.Op0.getOpcode() == ISD::TRUNCATE &&
1364 C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
1365 C.Op1.getOpcode() == ISD::Constant &&
1366 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1367 auto *L = cast<LoadSDNode>(C.Op0.getOperand(0));
1368 if (L->getMemoryVT().getStoreSizeInBits()
1369 <= C.Op0.getValueType().getSizeInBits()) {
1370 unsigned Type = L->getExtensionType();
1371 if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
1372 (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
1373 C.Op0 = C.Op0.getOperand(0);
1374 C.Op1 = DAG.getConstant(0, C.Op0.getValueType());
1380 // Return true if shift operation N has an in-range constant shift value.
1381 // Store it in ShiftVal if so.
1382 static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
1383 auto *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
1387 uint64_t Amount = Shift->getZExtValue();
1388 if (Amount >= N.getValueType().getSizeInBits())
1395 // Check whether an AND with Mask is suitable for a TEST UNDER MASK
1396 // instruction and whether the CC value is descriptive enough to handle
1397 // a comparison of type Opcode between the AND result and CmpVal.
1398 // CCMask says which comparison result is being tested and BitSize is
1399 // the number of bits in the operands. If TEST UNDER MASK can be used,
1400 // return the corresponding CC mask, otherwise return 0.
1401 static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask,
1402 uint64_t Mask, uint64_t CmpVal,
1403 unsigned ICmpType) {
1404 assert(Mask != 0 && "ANDs with zero should have been removed by now");
1406 // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.
1407 if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&
1408 !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))
1411 // Work out the masks for the lowest and highest bits.
1412 unsigned HighShift = 63 - countLeadingZeros(Mask);
1413 uint64_t High = uint64_t(1) << HighShift;
1414 uint64_t Low = uint64_t(1) << countTrailingZeros(Mask);
1416 // Signed ordered comparisons are effectively unsigned if the sign
1418 bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly);
1420 // Check for equality comparisons with 0, or the equivalent.
1422 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1423 return SystemZ::CCMASK_TM_ALL_0;
1424 if (CCMask == SystemZ::CCMASK_CMP_NE)
1425 return SystemZ::CCMASK_TM_SOME_1;
1427 if (EffectivelyUnsigned && CmpVal <= Low) {
1428 if (CCMask == SystemZ::CCMASK_CMP_LT)
1429 return SystemZ::CCMASK_TM_ALL_0;
1430 if (CCMask == SystemZ::CCMASK_CMP_GE)
1431 return SystemZ::CCMASK_TM_SOME_1;
1433 if (EffectivelyUnsigned && CmpVal < Low) {
1434 if (CCMask == SystemZ::CCMASK_CMP_LE)
1435 return SystemZ::CCMASK_TM_ALL_0;
1436 if (CCMask == SystemZ::CCMASK_CMP_GT)
1437 return SystemZ::CCMASK_TM_SOME_1;
1440 // Check for equality comparisons with the mask, or the equivalent.
1441 if (CmpVal == Mask) {
1442 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1443 return SystemZ::CCMASK_TM_ALL_1;
1444 if (CCMask == SystemZ::CCMASK_CMP_NE)
1445 return SystemZ::CCMASK_TM_SOME_0;
1447 if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) {
1448 if (CCMask == SystemZ::CCMASK_CMP_GT)
1449 return SystemZ::CCMASK_TM_ALL_1;
1450 if (CCMask == SystemZ::CCMASK_CMP_LE)
1451 return SystemZ::CCMASK_TM_SOME_0;
1453 if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) {
1454 if (CCMask == SystemZ::CCMASK_CMP_GE)
1455 return SystemZ::CCMASK_TM_ALL_1;
1456 if (CCMask == SystemZ::CCMASK_CMP_LT)
1457 return SystemZ::CCMASK_TM_SOME_0;
1460 // Check for ordered comparisons with the top bit.
1461 if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) {
1462 if (CCMask == SystemZ::CCMASK_CMP_LE)
1463 return SystemZ::CCMASK_TM_MSB_0;
1464 if (CCMask == SystemZ::CCMASK_CMP_GT)
1465 return SystemZ::CCMASK_TM_MSB_1;
1467 if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) {
1468 if (CCMask == SystemZ::CCMASK_CMP_LT)
1469 return SystemZ::CCMASK_TM_MSB_0;
1470 if (CCMask == SystemZ::CCMASK_CMP_GE)
1471 return SystemZ::CCMASK_TM_MSB_1;
1474 // If there are just two bits, we can do equality checks for Low and High
1476 if (Mask == Low + High) {
1477 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low)
1478 return SystemZ::CCMASK_TM_MIXED_MSB_0;
1479 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low)
1480 return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY;
1481 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High)
1482 return SystemZ::CCMASK_TM_MIXED_MSB_1;
1483 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High)
1484 return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY;
1487 // Looks like we've exhausted our options.
1491 // See whether C can be implemented as a TEST UNDER MASK instruction.
1492 // Update the arguments with the TM version if so.
1493 static void adjustForTestUnderMask(SelectionDAG &DAG, Comparison &C) {
1494 // Check that we have a comparison with a constant.
1495 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1498 uint64_t CmpVal = ConstOp1->getZExtValue();
1500 // Check whether the nonconstant input is an AND with a constant mask.
1503 ConstantSDNode *Mask = nullptr;
1504 if (C.Op0.getOpcode() == ISD::AND) {
1505 NewC.Op0 = C.Op0.getOperand(0);
1506 NewC.Op1 = C.Op0.getOperand(1);
1507 Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
1510 MaskVal = Mask->getZExtValue();
1512 // There is no instruction to compare with a 64-bit immediate
1513 // so use TMHH instead if possible. We need an unsigned ordered
1514 // comparison with an i64 immediate.
1515 if (NewC.Op0.getValueType() != MVT::i64 ||
1516 NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
1517 NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
1518 NewC.ICmpType == SystemZICMP::SignedOnly)
1520 // Convert LE and GT comparisons into LT and GE.
1521 if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
1522 NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
1523 if (CmpVal == uint64_t(-1))
1526 NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1528 // If the low N bits of Op1 are zero than the low N bits of Op0 can
1529 // be masked off without changing the result.
1530 MaskVal = -(CmpVal & -CmpVal);
1531 NewC.ICmpType = SystemZICMP::UnsignedOnly;
1534 // Check whether the combination of mask, comparison value and comparison
1535 // type are suitable.
1536 unsigned BitSize = NewC.Op0.getValueType().getSizeInBits();
1537 unsigned NewCCMask, ShiftVal;
1538 if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1539 NewC.Op0.getOpcode() == ISD::SHL &&
1540 isSimpleShift(NewC.Op0, ShiftVal) &&
1541 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1542 MaskVal >> ShiftVal,
1544 SystemZICMP::Any))) {
1545 NewC.Op0 = NewC.Op0.getOperand(0);
1546 MaskVal >>= ShiftVal;
1547 } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1548 NewC.Op0.getOpcode() == ISD::SRL &&
1549 isSimpleShift(NewC.Op0, ShiftVal) &&
1550 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1551 MaskVal << ShiftVal,
1553 SystemZICMP::UnsignedOnly))) {
1554 NewC.Op0 = NewC.Op0.getOperand(0);
1555 MaskVal <<= ShiftVal;
1557 NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
1563 // Go ahead and make the change.
1564 C.Opcode = SystemZISD::TM;
1566 if (Mask && Mask->getZExtValue() == MaskVal)
1567 C.Op1 = SDValue(Mask, 0);
1569 C.Op1 = DAG.getConstant(MaskVal, C.Op0.getValueType());
1570 C.CCValid = SystemZ::CCMASK_TM;
1571 C.CCMask = NewCCMask;
1574 // Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
1575 static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
1576 ISD::CondCode Cond) {
1577 Comparison C(CmpOp0, CmpOp1);
1578 C.CCMask = CCMaskForCondCode(Cond);
1579 if (C.Op0.getValueType().isFloatingPoint()) {
1580 C.CCValid = SystemZ::CCMASK_FCMP;
1581 C.Opcode = SystemZISD::FCMP;
1584 C.CCValid = SystemZ::CCMASK_ICMP;
1585 C.Opcode = SystemZISD::ICMP;
1586 // Choose the type of comparison. Equality and inequality tests can
1587 // use either signed or unsigned comparisons. The choice also doesn't
1588 // matter if both sign bits are known to be clear. In those cases we
1589 // want to give the main isel code the freedom to choose whichever
1591 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1592 C.CCMask == SystemZ::CCMASK_CMP_NE ||
1593 (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
1594 C.ICmpType = SystemZICMP::Any;
1595 else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
1596 C.ICmpType = SystemZICMP::UnsignedOnly;
1598 C.ICmpType = SystemZICMP::SignedOnly;
1599 C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
1600 adjustZeroCmp(DAG, C);
1601 adjustSubwordCmp(DAG, C);
1602 adjustForSubtraction(DAG, C);
1604 adjustICmpTruncate(DAG, C);
1607 if (shouldSwapCmpOperands(C)) {
1608 std::swap(C.Op0, C.Op1);
1609 C.CCMask = reverseCCMask(C.CCMask);
1612 adjustForTestUnderMask(DAG, C);
1616 // Emit the comparison instruction described by C.
1617 static SDValue emitCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1618 if (C.Opcode == SystemZISD::ICMP)
1619 return DAG.getNode(SystemZISD::ICMP, DL, MVT::Glue, C.Op0, C.Op1,
1620 DAG.getConstant(C.ICmpType, MVT::i32));
1621 if (C.Opcode == SystemZISD::TM) {
1622 bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
1623 bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
1624 return DAG.getNode(SystemZISD::TM, DL, MVT::Glue, C.Op0, C.Op1,
1625 DAG.getConstant(RegisterOnly, MVT::i32));
1627 return DAG.getNode(C.Opcode, DL, MVT::Glue, C.Op0, C.Op1);
1630 // Implement a 32-bit *MUL_LOHI operation by extending both operands to
1631 // 64 bits. Extend is the extension type to use. Store the high part
1632 // in Hi and the low part in Lo.
1633 static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL,
1634 unsigned Extend, SDValue Op0, SDValue Op1,
1635 SDValue &Hi, SDValue &Lo) {
1636 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
1637 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
1638 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
1639 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, DAG.getConstant(32, MVT::i64));
1640 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
1641 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
1644 // Lower a binary operation that produces two VT results, one in each
1645 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
1646 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
1647 // on the extended Op0 and (unextended) Op1. Store the even register result
1648 // in Even and the odd register result in Odd.
1649 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
1650 unsigned Extend, unsigned Opcode,
1651 SDValue Op0, SDValue Op1,
1652 SDValue &Even, SDValue &Odd) {
1653 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
1654 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
1655 SDValue(In128, 0), Op1);
1656 bool Is32Bit = is32Bit(VT);
1657 Even = DAG.getTargetExtractSubreg(SystemZ::even128(Is32Bit), DL, VT, Result);
1658 Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
1661 // Return an i32 value that is 1 if the CC value produced by Glue is
1662 // in the mask CCMask and 0 otherwise. CC is known to have a value
1663 // in CCValid, so other values can be ignored.
1664 static SDValue emitSETCC(SelectionDAG &DAG, SDLoc DL, SDValue Glue,
1665 unsigned CCValid, unsigned CCMask) {
1666 IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
1667 SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
1669 if (Conversion.XORValue)
1670 Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
1671 DAG.getConstant(Conversion.XORValue, MVT::i32));
1673 if (Conversion.AddValue)
1674 Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
1675 DAG.getConstant(Conversion.AddValue, MVT::i32));
1677 // The SHR/AND sequence should get optimized to an RISBG.
1678 Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
1679 DAG.getConstant(Conversion.Bit, MVT::i32));
1680 if (Conversion.Bit != 31)
1681 Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
1682 DAG.getConstant(1, MVT::i32));
1686 SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
1687 SelectionDAG &DAG) const {
1688 SDValue CmpOp0 = Op.getOperand(0);
1689 SDValue CmpOp1 = Op.getOperand(1);
1690 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1693 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1694 SDValue Glue = emitCmp(DAG, DL, C);
1695 return emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1698 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1699 SDValue Chain = Op.getOperand(0);
1700 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1701 SDValue CmpOp0 = Op.getOperand(2);
1702 SDValue CmpOp1 = Op.getOperand(3);
1703 SDValue Dest = Op.getOperand(4);
1706 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1707 SDValue Glue = emitCmp(DAG, DL, C);
1708 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
1709 Chain, DAG.getConstant(C.CCValid, MVT::i32),
1710 DAG.getConstant(C.CCMask, MVT::i32), Dest, Glue);
1713 // Return true if Pos is CmpOp and Neg is the negative of CmpOp,
1714 // allowing Pos and Neg to be wider than CmpOp.
1715 static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
1716 return (Neg.getOpcode() == ISD::SUB &&
1717 Neg.getOperand(0).getOpcode() == ISD::Constant &&
1718 cast<ConstantSDNode>(Neg.getOperand(0))->getZExtValue() == 0 &&
1719 Neg.getOperand(1) == Pos &&
1721 (Pos.getOpcode() == ISD::SIGN_EXTEND &&
1722 Pos.getOperand(0) == CmpOp)));
1725 // Return the absolute or negative absolute of Op; IsNegative decides which.
1726 static SDValue getAbsolute(SelectionDAG &DAG, SDLoc DL, SDValue Op,
1728 Op = DAG.getNode(SystemZISD::IABS, DL, Op.getValueType(), Op);
1730 Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
1731 DAG.getConstant(0, Op.getValueType()), Op);
1735 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
1736 SelectionDAG &DAG) const {
1737 SDValue CmpOp0 = Op.getOperand(0);
1738 SDValue CmpOp1 = Op.getOperand(1);
1739 SDValue TrueOp = Op.getOperand(2);
1740 SDValue FalseOp = Op.getOperand(3);
1741 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1744 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
1746 // Check for absolute and negative-absolute selections, including those
1747 // where the comparison value is sign-extended (for LPGFR and LNGFR).
1748 // This check supplements the one in DAGCombiner.
1749 if (C.Opcode == SystemZISD::ICMP &&
1750 C.CCMask != SystemZ::CCMASK_CMP_EQ &&
1751 C.CCMask != SystemZ::CCMASK_CMP_NE &&
1752 C.Op1.getOpcode() == ISD::Constant &&
1753 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1754 if (isAbsolute(C.Op0, TrueOp, FalseOp))
1755 return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
1756 if (isAbsolute(C.Op0, FalseOp, TrueOp))
1757 return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
1760 SDValue Glue = emitCmp(DAG, DL, C);
1762 // Special case for handling -1/0 results. The shifts we use here
1763 // should get optimized with the IPM conversion sequence.
1764 auto *TrueC = dyn_cast<ConstantSDNode>(TrueOp);
1765 auto *FalseC = dyn_cast<ConstantSDNode>(FalseOp);
1766 if (TrueC && FalseC) {
1767 int64_t TrueVal = TrueC->getSExtValue();
1768 int64_t FalseVal = FalseC->getSExtValue();
1769 if ((TrueVal == -1 && FalseVal == 0) || (TrueVal == 0 && FalseVal == -1)) {
1770 // Invert the condition if we want -1 on false.
1772 C.CCMask ^= C.CCValid;
1773 SDValue Result = emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
1774 EVT VT = Op.getValueType();
1775 // Extend the result to VT. Upper bits are ignored.
1777 Result = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Result);
1778 // Sign-extend from the low bit.
1779 SDValue ShAmt = DAG.getConstant(VT.getSizeInBits() - 1, MVT::i32);
1780 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Result, ShAmt);
1781 return DAG.getNode(ISD::SRA, DL, VT, Shl, ShAmt);
1785 SmallVector<SDValue, 5> Ops;
1786 Ops.push_back(TrueOp);
1787 Ops.push_back(FalseOp);
1788 Ops.push_back(DAG.getConstant(C.CCValid, MVT::i32));
1789 Ops.push_back(DAG.getConstant(C.CCMask, MVT::i32));
1790 Ops.push_back(Glue);
1792 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
1793 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, Ops);
1796 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
1797 SelectionDAG &DAG) const {
1799 const GlobalValue *GV = Node->getGlobal();
1800 int64_t Offset = Node->getOffset();
1801 EVT PtrVT = getPointerTy();
1802 Reloc::Model RM = DAG.getTarget().getRelocationModel();
1803 CodeModel::Model CM = DAG.getTarget().getCodeModel();
1806 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
1807 // Assign anchors at 1<<12 byte boundaries.
1808 uint64_t Anchor = Offset & ~uint64_t(0xfff);
1809 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor);
1810 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1812 // The offset can be folded into the address if it is aligned to a halfword.
1814 if (Offset != 0 && (Offset & 1) == 0) {
1815 SDValue Full = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor + Offset);
1816 Result = DAG.getNode(SystemZISD::PCREL_OFFSET, DL, PtrVT, Full, Result);
1820 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
1821 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1822 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
1823 MachinePointerInfo::getGOT(), false, false, false, 0);
1826 // If there was a non-zero offset that we didn't fold, create an explicit
1829 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
1830 DAG.getConstant(Offset, PtrVT));
1835 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
1836 SelectionDAG &DAG) const {
1838 const GlobalValue *GV = Node->getGlobal();
1839 EVT PtrVT = getPointerTy();
1840 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
1842 if (model != TLSModel::LocalExec)
1843 llvm_unreachable("only local-exec TLS mode supported");
1845 // The high part of the thread pointer is in access register 0.
1846 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1847 DAG.getConstant(0, MVT::i32));
1848 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
1850 // The low part of the thread pointer is in access register 1.
1851 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
1852 DAG.getConstant(1, MVT::i32));
1853 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
1855 // Merge them into a single 64-bit address.
1856 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
1857 DAG.getConstant(32, PtrVT));
1858 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
1860 // Get the offset of GA from the thread pointer.
1861 SystemZConstantPoolValue *CPV =
1862 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
1864 // Force the offset into the constant pool and load it from there.
1865 SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8);
1866 SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
1867 CPAddr, MachinePointerInfo::getConstantPool(),
1868 false, false, false, 0);
1870 // Add the base and offset together.
1871 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
1874 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
1875 SelectionDAG &DAG) const {
1877 const BlockAddress *BA = Node->getBlockAddress();
1878 int64_t Offset = Node->getOffset();
1879 EVT PtrVT = getPointerTy();
1881 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
1882 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1886 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
1887 SelectionDAG &DAG) const {
1889 EVT PtrVT = getPointerTy();
1890 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1892 // Use LARL to load the address of the table.
1893 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1896 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
1897 SelectionDAG &DAG) const {
1899 EVT PtrVT = getPointerTy();
1902 if (CP->isMachineConstantPoolEntry())
1903 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1904 CP->getAlignment());
1906 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1907 CP->getAlignment(), CP->getOffset());
1909 // Use LARL to load the address of the constant pool entry.
1910 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
1913 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
1914 SelectionDAG &DAG) const {
1916 SDValue In = Op.getOperand(0);
1917 EVT InVT = In.getValueType();
1918 EVT ResVT = Op.getValueType();
1920 if (InVT == MVT::i32 && ResVT == MVT::f32) {
1922 if (Subtarget.hasHighWord()) {
1923 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL,
1925 In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
1926 MVT::i64, SDValue(U64, 0), In);
1928 In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
1929 In64 = DAG.getNode(ISD::SHL, DL, MVT::i64, In64,
1930 DAG.getConstant(32, MVT::i64));
1932 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, In64);
1933 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32,
1934 DL, MVT::f32, Out64);
1936 if (InVT == MVT::f32 && ResVT == MVT::i32) {
1937 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
1938 SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
1939 MVT::f64, SDValue(U64, 0), In);
1940 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
1941 if (Subtarget.hasHighWord())
1942 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32, DL,
1944 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64,
1945 DAG.getConstant(32, MVT::i64));
1946 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
1948 llvm_unreachable("Unexpected bitcast combination");
1951 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
1952 SelectionDAG &DAG) const {
1953 MachineFunction &MF = DAG.getMachineFunction();
1954 SystemZMachineFunctionInfo *FuncInfo =
1955 MF.getInfo<SystemZMachineFunctionInfo>();
1956 EVT PtrVT = getPointerTy();
1958 SDValue Chain = Op.getOperand(0);
1959 SDValue Addr = Op.getOperand(1);
1960 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1963 // The initial values of each field.
1964 const unsigned NumFields = 4;
1965 SDValue Fields[NumFields] = {
1966 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT),
1967 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT),
1968 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
1969 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
1972 // Store each field into its respective slot.
1973 SDValue MemOps[NumFields];
1974 unsigned Offset = 0;
1975 for (unsigned I = 0; I < NumFields; ++I) {
1976 SDValue FieldAddr = Addr;
1978 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
1979 DAG.getIntPtrConstant(Offset));
1980 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
1981 MachinePointerInfo(SV, Offset),
1985 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
1988 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
1989 SelectionDAG &DAG) const {
1990 SDValue Chain = Op.getOperand(0);
1991 SDValue DstPtr = Op.getOperand(1);
1992 SDValue SrcPtr = Op.getOperand(2);
1993 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
1994 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
1997 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32),
1998 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
1999 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
2002 SDValue SystemZTargetLowering::
2003 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
2004 SDValue Chain = Op.getOperand(0);
2005 SDValue Size = Op.getOperand(1);
2008 unsigned SPReg = getStackPointerRegisterToSaveRestore();
2010 // Get a reference to the stack pointer.
2011 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
2013 // Get the new stack pointer value.
2014 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
2016 // Copy the new stack pointer back.
2017 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
2019 // The allocated data lives above the 160 bytes allocated for the standard
2020 // frame, plus any outgoing stack arguments. We don't know how much that
2021 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
2022 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
2023 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
2025 SDValue Ops[2] = { Result, Chain };
2026 return DAG.getMergeValues(Ops, DL);
2029 SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
2030 SelectionDAG &DAG) const {
2031 EVT VT = Op.getValueType();
2035 // Just do a normal 64-bit multiplication and extract the results.
2036 // We define this so that it can be used for constant division.
2037 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
2038 Op.getOperand(1), Ops[1], Ops[0]);
2040 // Do a full 128-bit multiplication based on UMUL_LOHI64:
2042 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
2044 // but using the fact that the upper halves are either all zeros
2047 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
2049 // and grouping the right terms together since they are quicker than the
2052 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
2053 SDValue C63 = DAG.getConstant(63, MVT::i64);
2054 SDValue LL = Op.getOperand(0);
2055 SDValue RL = Op.getOperand(1);
2056 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
2057 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
2058 // UMUL_LOHI64 returns the low result in the odd register and the high
2059 // result in the even register. SMUL_LOHI is defined to return the
2060 // low half first, so the results are in reverse order.
2061 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2062 LL, RL, Ops[1], Ops[0]);
2063 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
2064 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
2065 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
2066 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
2068 return DAG.getMergeValues(Ops, DL);
2071 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
2072 SelectionDAG &DAG) const {
2073 EVT VT = Op.getValueType();
2077 // Just do a normal 64-bit multiplication and extract the results.
2078 // We define this so that it can be used for constant division.
2079 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
2080 Op.getOperand(1), Ops[1], Ops[0]);
2082 // UMUL_LOHI64 returns the low result in the odd register and the high
2083 // result in the even register. UMUL_LOHI is defined to return the
2084 // low half first, so the results are in reverse order.
2085 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2086 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2087 return DAG.getMergeValues(Ops, DL);
2090 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
2091 SelectionDAG &DAG) const {
2092 SDValue Op0 = Op.getOperand(0);
2093 SDValue Op1 = Op.getOperand(1);
2094 EVT VT = Op.getValueType();
2098 // We use DSGF for 32-bit division.
2100 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
2101 Opcode = SystemZISD::SDIVREM32;
2102 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
2103 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
2104 Opcode = SystemZISD::SDIVREM32;
2106 Opcode = SystemZISD::SDIVREM64;
2108 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
2109 // input is "don't care". The instruction returns the remainder in
2110 // the even register and the quotient in the odd register.
2112 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
2113 Op0, Op1, Ops[1], Ops[0]);
2114 return DAG.getMergeValues(Ops, DL);
2117 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
2118 SelectionDAG &DAG) const {
2119 EVT VT = Op.getValueType();
2122 // DL(G) uses a double-width dividend, so we need to clear the even
2123 // register in the GR128 input. The instruction returns the remainder
2124 // in the even register and the quotient in the odd register.
2127 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
2128 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2130 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
2131 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2132 return DAG.getMergeValues(Ops, DL);
2135 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
2136 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
2138 // Get the known-zero masks for each operand.
2139 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
2140 APInt KnownZero[2], KnownOne[2];
2141 DAG.computeKnownBits(Ops[0], KnownZero[0], KnownOne[0]);
2142 DAG.computeKnownBits(Ops[1], KnownZero[1], KnownOne[1]);
2144 // See if the upper 32 bits of one operand and the lower 32 bits of the
2145 // other are known zero. They are the low and high operands respectively.
2146 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
2147 KnownZero[1].getZExtValue() };
2149 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
2151 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
2156 SDValue LowOp = Ops[Low];
2157 SDValue HighOp = Ops[High];
2159 // If the high part is a constant, we're better off using IILH.
2160 if (HighOp.getOpcode() == ISD::Constant)
2163 // If the low part is a constant that is outside the range of LHI,
2164 // then we're better off using IILF.
2165 if (LowOp.getOpcode() == ISD::Constant) {
2166 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
2167 if (!isInt<16>(Value))
2171 // Check whether the high part is an AND that doesn't change the
2172 // high 32 bits and just masks out low bits. We can skip it if so.
2173 if (HighOp.getOpcode() == ISD::AND &&
2174 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
2175 SDValue HighOp0 = HighOp.getOperand(0);
2176 uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
2177 if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
2181 // Take advantage of the fact that all GR32 operations only change the
2182 // low 32 bits by truncating Low to an i32 and inserting it directly
2183 // using a subreg. The interesting cases are those where the truncation
2186 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
2187 return DAG.getTargetInsertSubreg(SystemZ::subreg_l32, DL,
2188 MVT::i64, HighOp, Low32);
2191 // Op is an atomic load. Lower it into a normal volatile load.
2192 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
2193 SelectionDAG &DAG) const {
2194 auto *Node = cast<AtomicSDNode>(Op.getNode());
2195 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), Op.getValueType(),
2196 Node->getChain(), Node->getBasePtr(),
2197 Node->getMemoryVT(), Node->getMemOperand());
2200 // Op is an atomic store. Lower it into a normal volatile store followed
2201 // by a serialization.
2202 SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
2203 SelectionDAG &DAG) const {
2204 auto *Node = cast<AtomicSDNode>(Op.getNode());
2205 SDValue Chain = DAG.getTruncStore(Node->getChain(), SDLoc(Op), Node->getVal(),
2206 Node->getBasePtr(), Node->getMemoryVT(),
2207 Node->getMemOperand());
2208 return SDValue(DAG.getMachineNode(SystemZ::Serialize, SDLoc(Op), MVT::Other,
2212 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
2213 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
2214 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
2216 unsigned Opcode) const {
2217 auto *Node = cast<AtomicSDNode>(Op.getNode());
2219 // 32-bit operations need no code outside the main loop.
2220 EVT NarrowVT = Node->getMemoryVT();
2221 EVT WideVT = MVT::i32;
2222 if (NarrowVT == WideVT)
2225 int64_t BitSize = NarrowVT.getSizeInBits();
2226 SDValue ChainIn = Node->getChain();
2227 SDValue Addr = Node->getBasePtr();
2228 SDValue Src2 = Node->getVal();
2229 MachineMemOperand *MMO = Node->getMemOperand();
2231 EVT PtrVT = Addr.getValueType();
2233 // Convert atomic subtracts of constants into additions.
2234 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
2235 if (auto *Const = dyn_cast<ConstantSDNode>(Src2)) {
2236 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
2237 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
2240 // Get the address of the containing word.
2241 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2242 DAG.getConstant(-4, PtrVT));
2244 // Get the number of bits that the word must be rotated left in order
2245 // to bring the field to the top bits of a GR32.
2246 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2247 DAG.getConstant(3, PtrVT));
2248 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2250 // Get the complementing shift amount, for rotating a field in the top
2251 // bits back to its proper position.
2252 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2253 DAG.getConstant(0, WideVT), BitShift);
2255 // Extend the source operand to 32 bits and prepare it for the inner loop.
2256 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
2257 // operations require the source to be shifted in advance. (This shift
2258 // can be folded if the source is constant.) For AND and NAND, the lower
2259 // bits must be set, while for other opcodes they should be left clear.
2260 if (Opcode != SystemZISD::ATOMIC_SWAPW)
2261 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
2262 DAG.getConstant(32 - BitSize, WideVT));
2263 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
2264 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
2265 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
2266 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT));
2268 // Construct the ATOMIC_LOADW_* node.
2269 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2270 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
2271 DAG.getConstant(BitSize, WideVT) };
2272 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
2275 // Rotate the result of the final CS so that the field is in the lower
2276 // bits of a GR32, then truncate it.
2277 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
2278 DAG.getConstant(BitSize, WideVT));
2279 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
2281 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
2282 return DAG.getMergeValues(RetOps, DL);
2285 // Op is an ATOMIC_LOAD_SUB operation. Lower 8- and 16-bit operations
2286 // into ATOMIC_LOADW_SUBs and decide whether to convert 32- and 64-bit
2287 // operations into additions.
2288 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
2289 SelectionDAG &DAG) const {
2290 auto *Node = cast<AtomicSDNode>(Op.getNode());
2291 EVT MemVT = Node->getMemoryVT();
2292 if (MemVT == MVT::i32 || MemVT == MVT::i64) {
2293 // A full-width operation.
2294 assert(Op.getValueType() == MemVT && "Mismatched VTs");
2295 SDValue Src2 = Node->getVal();
2299 if (auto *Op2 = dyn_cast<ConstantSDNode>(Src2)) {
2300 // Use an addition if the operand is constant and either LAA(G) is
2301 // available or the negative value is in the range of A(G)FHI.
2302 int64_t Value = (-Op2->getAPIntValue()).getSExtValue();
2303 if (isInt<32>(Value) || Subtarget.hasInterlockedAccess1())
2304 NegSrc2 = DAG.getConstant(Value, MemVT);
2305 } else if (Subtarget.hasInterlockedAccess1())
2306 // Use LAA(G) if available.
2307 NegSrc2 = DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, MemVT),
2310 if (NegSrc2.getNode())
2311 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
2312 Node->getChain(), Node->getBasePtr(), NegSrc2,
2313 Node->getMemOperand(), Node->getOrdering(),
2314 Node->getSynchScope());
2316 // Use the node as-is.
2320 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
2323 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
2324 // into a fullword ATOMIC_CMP_SWAPW operation.
2325 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
2326 SelectionDAG &DAG) const {
2327 auto *Node = cast<AtomicSDNode>(Op.getNode());
2329 // We have native support for 32-bit compare and swap.
2330 EVT NarrowVT = Node->getMemoryVT();
2331 EVT WideVT = MVT::i32;
2332 if (NarrowVT == WideVT)
2335 int64_t BitSize = NarrowVT.getSizeInBits();
2336 SDValue ChainIn = Node->getOperand(0);
2337 SDValue Addr = Node->getOperand(1);
2338 SDValue CmpVal = Node->getOperand(2);
2339 SDValue SwapVal = Node->getOperand(3);
2340 MachineMemOperand *MMO = Node->getMemOperand();
2342 EVT PtrVT = Addr.getValueType();
2344 // Get the address of the containing word.
2345 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
2346 DAG.getConstant(-4, PtrVT));
2348 // Get the number of bits that the word must be rotated left in order
2349 // to bring the field to the top bits of a GR32.
2350 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
2351 DAG.getConstant(3, PtrVT));
2352 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
2354 // Get the complementing shift amount, for rotating a field in the top
2355 // bits back to its proper position.
2356 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
2357 DAG.getConstant(0, WideVT), BitShift);
2359 // Construct the ATOMIC_CMP_SWAPW node.
2360 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
2361 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
2362 NegBitShift, DAG.getConstant(BitSize, WideVT) };
2363 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
2364 VTList, Ops, NarrowVT, MMO);
2368 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
2369 SelectionDAG &DAG) const {
2370 MachineFunction &MF = DAG.getMachineFunction();
2371 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2372 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
2373 SystemZ::R15D, Op.getValueType());
2376 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
2377 SelectionDAG &DAG) const {
2378 MachineFunction &MF = DAG.getMachineFunction();
2379 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
2380 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
2381 SystemZ::R15D, Op.getOperand(1));
2384 SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op,
2385 SelectionDAG &DAG) const {
2386 bool IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2388 // Just preserve the chain.
2389 return Op.getOperand(0);
2391 bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2392 unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
2393 auto *Node = cast<MemIntrinsicSDNode>(Op.getNode());
2396 DAG.getConstant(Code, MVT::i32),
2399 return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, SDLoc(Op),
2400 Node->getVTList(), Ops,
2401 Node->getMemoryVT(), Node->getMemOperand());
2404 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
2405 SelectionDAG &DAG) const {
2406 switch (Op.getOpcode()) {
2408 return lowerBR_CC(Op, DAG);
2409 case ISD::SELECT_CC:
2410 return lowerSELECT_CC(Op, DAG);
2412 return lowerSETCC(Op, DAG);
2413 case ISD::GlobalAddress:
2414 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
2415 case ISD::GlobalTLSAddress:
2416 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
2417 case ISD::BlockAddress:
2418 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
2419 case ISD::JumpTable:
2420 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
2421 case ISD::ConstantPool:
2422 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
2424 return lowerBITCAST(Op, DAG);
2426 return lowerVASTART(Op, DAG);
2428 return lowerVACOPY(Op, DAG);
2429 case ISD::DYNAMIC_STACKALLOC:
2430 return lowerDYNAMIC_STACKALLOC(Op, DAG);
2431 case ISD::SMUL_LOHI:
2432 return lowerSMUL_LOHI(Op, DAG);
2433 case ISD::UMUL_LOHI:
2434 return lowerUMUL_LOHI(Op, DAG);
2436 return lowerSDIVREM(Op, DAG);
2438 return lowerUDIVREM(Op, DAG);
2440 return lowerOR(Op, DAG);
2441 case ISD::ATOMIC_SWAP:
2442 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
2443 case ISD::ATOMIC_STORE:
2444 return lowerATOMIC_STORE(Op, DAG);
2445 case ISD::ATOMIC_LOAD:
2446 return lowerATOMIC_LOAD(Op, DAG);
2447 case ISD::ATOMIC_LOAD_ADD:
2448 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
2449 case ISD::ATOMIC_LOAD_SUB:
2450 return lowerATOMIC_LOAD_SUB(Op, DAG);
2451 case ISD::ATOMIC_LOAD_AND:
2452 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
2453 case ISD::ATOMIC_LOAD_OR:
2454 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
2455 case ISD::ATOMIC_LOAD_XOR:
2456 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
2457 case ISD::ATOMIC_LOAD_NAND:
2458 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
2459 case ISD::ATOMIC_LOAD_MIN:
2460 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
2461 case ISD::ATOMIC_LOAD_MAX:
2462 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
2463 case ISD::ATOMIC_LOAD_UMIN:
2464 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
2465 case ISD::ATOMIC_LOAD_UMAX:
2466 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
2467 case ISD::ATOMIC_CMP_SWAP:
2468 return lowerATOMIC_CMP_SWAP(Op, DAG);
2469 case ISD::STACKSAVE:
2470 return lowerSTACKSAVE(Op, DAG);
2471 case ISD::STACKRESTORE:
2472 return lowerSTACKRESTORE(Op, DAG);
2474 return lowerPREFETCH(Op, DAG);
2476 llvm_unreachable("Unexpected node to lower");
2480 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
2481 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
2486 OPCODE(PCREL_WRAPPER);
2487 OPCODE(PCREL_OFFSET);
2493 OPCODE(SELECT_CCMASK);
2494 OPCODE(ADJDYNALLOC);
2495 OPCODE(EXTRACT_ACCESS);
2496 OPCODE(UMUL_LOHI64);
2512 OPCODE(SEARCH_STRING);
2515 OPCODE(ATOMIC_SWAPW);
2516 OPCODE(ATOMIC_LOADW_ADD);
2517 OPCODE(ATOMIC_LOADW_SUB);
2518 OPCODE(ATOMIC_LOADW_AND);
2519 OPCODE(ATOMIC_LOADW_OR);
2520 OPCODE(ATOMIC_LOADW_XOR);
2521 OPCODE(ATOMIC_LOADW_NAND);
2522 OPCODE(ATOMIC_LOADW_MIN);
2523 OPCODE(ATOMIC_LOADW_MAX);
2524 OPCODE(ATOMIC_LOADW_UMIN);
2525 OPCODE(ATOMIC_LOADW_UMAX);
2526 OPCODE(ATOMIC_CMP_SWAPW);
2533 SDValue SystemZTargetLowering::PerformDAGCombine(SDNode *N,
2534 DAGCombinerInfo &DCI) const {
2535 SelectionDAG &DAG = DCI.DAG;
2536 unsigned Opcode = N->getOpcode();
2537 if (Opcode == ISD::SIGN_EXTEND) {
2538 // Convert (sext (ashr (shl X, C1), C2)) to
2539 // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
2540 // cheap as narrower ones.
2541 SDValue N0 = N->getOperand(0);
2542 EVT VT = N->getValueType(0);
2543 if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
2544 auto *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2545 SDValue Inner = N0.getOperand(0);
2546 if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
2547 if (auto *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1))) {
2548 unsigned Extra = (VT.getSizeInBits() -
2549 N0.getValueType().getSizeInBits());
2550 unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
2551 unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
2552 EVT ShiftVT = N0.getOperand(1).getValueType();
2553 SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
2554 Inner.getOperand(0));
2555 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
2556 DAG.getConstant(NewShlAmt, ShiftVT));
2557 return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
2558 DAG.getConstant(NewSraAmt, ShiftVT));
2566 //===----------------------------------------------------------------------===//
2568 //===----------------------------------------------------------------------===//
2570 // Create a new basic block after MBB.
2571 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
2572 MachineFunction &MF = *MBB->getParent();
2573 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
2574 MF.insert(std::next(MachineFunction::iterator(MBB)), NewMBB);
2578 // Split MBB after MI and return the new block (the one that contains
2579 // instructions after MI).
2580 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
2581 MachineBasicBlock *MBB) {
2582 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2583 NewMBB->splice(NewMBB->begin(), MBB,
2584 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
2585 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2589 // Split MBB before MI and return the new block (the one that contains MI).
2590 static MachineBasicBlock *splitBlockBefore(MachineInstr *MI,
2591 MachineBasicBlock *MBB) {
2592 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
2593 NewMBB->splice(NewMBB->begin(), MBB, MI, MBB->end());
2594 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
2598 // Force base value Base into a register before MI. Return the register.
2599 static unsigned forceReg(MachineInstr *MI, MachineOperand &Base,
2600 const SystemZInstrInfo *TII) {
2602 return Base.getReg();
2604 MachineBasicBlock *MBB = MI->getParent();
2605 MachineFunction &MF = *MBB->getParent();
2606 MachineRegisterInfo &MRI = MF.getRegInfo();
2608 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
2609 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LA), Reg)
2610 .addOperand(Base).addImm(0).addReg(0);
2614 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
2616 SystemZTargetLowering::emitSelect(MachineInstr *MI,
2617 MachineBasicBlock *MBB) const {
2618 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
2619 MBB->getParent()->getTarget().getSubtargetImpl()->getInstrInfo());
2621 unsigned DestReg = MI->getOperand(0).getReg();
2622 unsigned TrueReg = MI->getOperand(1).getReg();
2623 unsigned FalseReg = MI->getOperand(2).getReg();
2624 unsigned CCValid = MI->getOperand(3).getImm();
2625 unsigned CCMask = MI->getOperand(4).getImm();
2626 DebugLoc DL = MI->getDebugLoc();
2628 MachineBasicBlock *StartMBB = MBB;
2629 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2630 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2633 // BRC CCMask, JoinMBB
2634 // # fallthrough to FalseMBB
2636 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2637 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2638 MBB->addSuccessor(JoinMBB);
2639 MBB->addSuccessor(FalseMBB);
2642 // # fallthrough to JoinMBB
2644 MBB->addSuccessor(JoinMBB);
2647 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
2650 BuildMI(*MBB, MI, DL, TII->get(SystemZ::PHI), DestReg)
2651 .addReg(TrueReg).addMBB(StartMBB)
2652 .addReg(FalseReg).addMBB(FalseMBB);
2654 MI->eraseFromParent();
2658 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
2659 // StoreOpcode is the store to use and Invert says whether the store should
2660 // happen when the condition is false rather than true. If a STORE ON
2661 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
2663 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
2664 MachineBasicBlock *MBB,
2665 unsigned StoreOpcode, unsigned STOCOpcode,
2666 bool Invert) const {
2667 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
2668 MBB->getParent()->getTarget().getSubtargetImpl()->getInstrInfo());
2670 unsigned SrcReg = MI->getOperand(0).getReg();
2671 MachineOperand Base = MI->getOperand(1);
2672 int64_t Disp = MI->getOperand(2).getImm();
2673 unsigned IndexReg = MI->getOperand(3).getReg();
2674 unsigned CCValid = MI->getOperand(4).getImm();
2675 unsigned CCMask = MI->getOperand(5).getImm();
2676 DebugLoc DL = MI->getDebugLoc();
2678 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
2680 // Use STOCOpcode if possible. We could use different store patterns in
2681 // order to avoid matching the index register, but the performance trade-offs
2682 // might be more complicated in that case.
2683 if (STOCOpcode && !IndexReg && Subtarget.hasLoadStoreOnCond()) {
2686 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
2687 .addReg(SrcReg).addOperand(Base).addImm(Disp)
2688 .addImm(CCValid).addImm(CCMask);
2689 MI->eraseFromParent();
2693 // Get the condition needed to branch around the store.
2697 MachineBasicBlock *StartMBB = MBB;
2698 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
2699 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
2702 // BRC CCMask, JoinMBB
2703 // # fallthrough to FalseMBB
2705 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2706 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
2707 MBB->addSuccessor(JoinMBB);
2708 MBB->addSuccessor(FalseMBB);
2711 // store %SrcReg, %Disp(%Index,%Base)
2712 // # fallthrough to JoinMBB
2714 BuildMI(MBB, DL, TII->get(StoreOpcode))
2715 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
2716 MBB->addSuccessor(JoinMBB);
2718 MI->eraseFromParent();
2722 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
2723 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
2724 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
2725 // BitSize is the width of the field in bits, or 0 if this is a partword
2726 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
2727 // is one of the operands. Invert says whether the field should be
2728 // inverted after performing BinOpcode (e.g. for NAND).
2730 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
2731 MachineBasicBlock *MBB,
2734 bool Invert) const {
2735 MachineFunction &MF = *MBB->getParent();
2736 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
2737 MF.getTarget().getSubtargetImpl()->getInstrInfo());
2738 MachineRegisterInfo &MRI = MF.getRegInfo();
2739 bool IsSubWord = (BitSize < 32);
2741 // Extract the operands. Base can be a register or a frame index.
2742 // Src2 can be a register or immediate.
2743 unsigned Dest = MI->getOperand(0).getReg();
2744 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2745 int64_t Disp = MI->getOperand(2).getImm();
2746 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
2747 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2748 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2749 DebugLoc DL = MI->getDebugLoc();
2751 BitSize = MI->getOperand(6).getImm();
2753 // Subword operations use 32-bit registers.
2754 const TargetRegisterClass *RC = (BitSize <= 32 ?
2755 &SystemZ::GR32BitRegClass :
2756 &SystemZ::GR64BitRegClass);
2757 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2758 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2760 // Get the right opcodes for the displacement.
2761 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2762 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2763 assert(LOpcode && CSOpcode && "Displacement out of range");
2765 // Create virtual registers for temporary results.
2766 unsigned OrigVal = MRI.createVirtualRegister(RC);
2767 unsigned OldVal = MRI.createVirtualRegister(RC);
2768 unsigned NewVal = (BinOpcode || IsSubWord ?
2769 MRI.createVirtualRegister(RC) : Src2.getReg());
2770 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2771 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2773 // Insert a basic block for the main loop.
2774 MachineBasicBlock *StartMBB = MBB;
2775 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2776 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2780 // %OrigVal = L Disp(%Base)
2781 // # fall through to LoopMMB
2783 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2784 .addOperand(Base).addImm(Disp).addReg(0);
2785 MBB->addSuccessor(LoopMBB);
2788 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
2789 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2790 // %RotatedNewVal = OP %RotatedOldVal, %Src2
2791 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2792 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2794 // # fall through to DoneMMB
2796 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2797 .addReg(OrigVal).addMBB(StartMBB)
2798 .addReg(Dest).addMBB(LoopMBB);
2800 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2801 .addReg(OldVal).addReg(BitShift).addImm(0);
2803 // Perform the operation normally and then invert every bit of the field.
2804 unsigned Tmp = MRI.createVirtualRegister(RC);
2805 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
2806 .addReg(RotatedOldVal).addOperand(Src2);
2808 // XILF with the upper BitSize bits set.
2809 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2810 .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize)));
2811 else if (BitSize == 32)
2812 // XILF with every bit set.
2813 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
2814 .addReg(Tmp).addImm(~uint32_t(0));
2816 // Use LCGR and add -1 to the result, which is more compact than
2817 // an XILF, XILH pair.
2818 unsigned Tmp2 = MRI.createVirtualRegister(RC);
2819 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
2820 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
2821 .addReg(Tmp2).addImm(-1);
2823 } else if (BinOpcode)
2824 // A simply binary operation.
2825 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
2826 .addReg(RotatedOldVal).addOperand(Src2);
2828 // Use RISBG to rotate Src2 into position and use it to replace the
2829 // field in RotatedOldVal.
2830 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
2831 .addReg(RotatedOldVal).addReg(Src2.getReg())
2832 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
2834 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2835 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2836 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2837 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2838 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2839 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2840 MBB->addSuccessor(LoopMBB);
2841 MBB->addSuccessor(DoneMBB);
2843 MI->eraseFromParent();
2847 // Implement EmitInstrWithCustomInserter for pseudo
2848 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
2849 // instruction that should be used to compare the current field with the
2850 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
2851 // for when the current field should be kept. BitSize is the width of
2852 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
2854 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
2855 MachineBasicBlock *MBB,
2856 unsigned CompareOpcode,
2857 unsigned KeepOldMask,
2858 unsigned BitSize) const {
2859 MachineFunction &MF = *MBB->getParent();
2860 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
2861 MF.getTarget().getSubtargetImpl()->getInstrInfo());
2862 MachineRegisterInfo &MRI = MF.getRegInfo();
2863 bool IsSubWord = (BitSize < 32);
2865 // Extract the operands. Base can be a register or a frame index.
2866 unsigned Dest = MI->getOperand(0).getReg();
2867 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2868 int64_t Disp = MI->getOperand(2).getImm();
2869 unsigned Src2 = MI->getOperand(3).getReg();
2870 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
2871 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
2872 DebugLoc DL = MI->getDebugLoc();
2874 BitSize = MI->getOperand(6).getImm();
2876 // Subword operations use 32-bit registers.
2877 const TargetRegisterClass *RC = (BitSize <= 32 ?
2878 &SystemZ::GR32BitRegClass :
2879 &SystemZ::GR64BitRegClass);
2880 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
2881 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
2883 // Get the right opcodes for the displacement.
2884 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
2885 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
2886 assert(LOpcode && CSOpcode && "Displacement out of range");
2888 // Create virtual registers for temporary results.
2889 unsigned OrigVal = MRI.createVirtualRegister(RC);
2890 unsigned OldVal = MRI.createVirtualRegister(RC);
2891 unsigned NewVal = MRI.createVirtualRegister(RC);
2892 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
2893 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
2894 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
2896 // Insert 3 basic blocks for the loop.
2897 MachineBasicBlock *StartMBB = MBB;
2898 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
2899 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
2900 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
2901 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
2905 // %OrigVal = L Disp(%Base)
2906 // # fall through to LoopMMB
2908 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
2909 .addOperand(Base).addImm(Disp).addReg(0);
2910 MBB->addSuccessor(LoopMBB);
2913 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
2914 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
2915 // CompareOpcode %RotatedOldVal, %Src2
2916 // BRC KeepOldMask, UpdateMBB
2918 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
2919 .addReg(OrigVal).addMBB(StartMBB)
2920 .addReg(Dest).addMBB(UpdateMBB);
2922 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
2923 .addReg(OldVal).addReg(BitShift).addImm(0);
2924 BuildMI(MBB, DL, TII->get(CompareOpcode))
2925 .addReg(RotatedOldVal).addReg(Src2);
2926 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2927 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
2928 MBB->addSuccessor(UpdateMBB);
2929 MBB->addSuccessor(UseAltMBB);
2932 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
2933 // # fall through to UpdateMMB
2936 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
2937 .addReg(RotatedOldVal).addReg(Src2)
2938 .addImm(32).addImm(31 + BitSize).addImm(0);
2939 MBB->addSuccessor(UpdateMBB);
2942 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
2943 // [ %RotatedAltVal, UseAltMBB ]
2944 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
2945 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
2947 // # fall through to DoneMMB
2949 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
2950 .addReg(RotatedOldVal).addMBB(LoopMBB)
2951 .addReg(RotatedAltVal).addMBB(UseAltMBB);
2953 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
2954 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
2955 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
2956 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
2957 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
2958 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
2959 MBB->addSuccessor(LoopMBB);
2960 MBB->addSuccessor(DoneMBB);
2962 MI->eraseFromParent();
2966 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
2969 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
2970 MachineBasicBlock *MBB) const {
2971 MachineFunction &MF = *MBB->getParent();
2972 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
2973 MF.getTarget().getSubtargetImpl()->getInstrInfo());
2974 MachineRegisterInfo &MRI = MF.getRegInfo();
2976 // Extract the operands. Base can be a register or a frame index.
2977 unsigned Dest = MI->getOperand(0).getReg();
2978 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
2979 int64_t Disp = MI->getOperand(2).getImm();
2980 unsigned OrigCmpVal = MI->getOperand(3).getReg();
2981 unsigned OrigSwapVal = MI->getOperand(4).getReg();
2982 unsigned BitShift = MI->getOperand(5).getReg();
2983 unsigned NegBitShift = MI->getOperand(6).getReg();
2984 int64_t BitSize = MI->getOperand(7).getImm();
2985 DebugLoc DL = MI->getDebugLoc();
2987 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
2989 // Get the right opcodes for the displacement.
2990 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
2991 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
2992 assert(LOpcode && CSOpcode && "Displacement out of range");
2994 // Create virtual registers for temporary results.
2995 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
2996 unsigned OldVal = MRI.createVirtualRegister(RC);
2997 unsigned CmpVal = MRI.createVirtualRegister(RC);
2998 unsigned SwapVal = MRI.createVirtualRegister(RC);
2999 unsigned StoreVal = MRI.createVirtualRegister(RC);
3000 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
3001 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
3002 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
3004 // Insert 2 basic blocks for the loop.
3005 MachineBasicBlock *StartMBB = MBB;
3006 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3007 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3008 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
3012 // %OrigOldVal = L Disp(%Base)
3013 // # fall through to LoopMMB
3015 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
3016 .addOperand(Base).addImm(Disp).addReg(0);
3017 MBB->addSuccessor(LoopMBB);
3020 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
3021 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
3022 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
3023 // %Dest = RLL %OldVal, BitSize(%BitShift)
3024 // ^^ The low BitSize bits contain the field
3026 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
3027 // ^^ Replace the upper 32-BitSize bits of the
3028 // comparison value with those that we loaded,
3029 // so that we can use a full word comparison.
3030 // CR %Dest, %RetryCmpVal
3032 // # Fall through to SetMBB
3034 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
3035 .addReg(OrigOldVal).addMBB(StartMBB)
3036 .addReg(RetryOldVal).addMBB(SetMBB);
3037 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
3038 .addReg(OrigCmpVal).addMBB(StartMBB)
3039 .addReg(RetryCmpVal).addMBB(SetMBB);
3040 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
3041 .addReg(OrigSwapVal).addMBB(StartMBB)
3042 .addReg(RetrySwapVal).addMBB(SetMBB);
3043 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
3044 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
3045 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
3046 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3047 BuildMI(MBB, DL, TII->get(SystemZ::CR))
3048 .addReg(Dest).addReg(RetryCmpVal);
3049 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3050 .addImm(SystemZ::CCMASK_ICMP)
3051 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
3052 MBB->addSuccessor(DoneMBB);
3053 MBB->addSuccessor(SetMBB);
3056 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
3057 // ^^ Replace the upper 32-BitSize bits of the new
3058 // value with those that we loaded.
3059 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
3060 // ^^ Rotate the new field to its proper position.
3061 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
3063 // # fall through to ExitMMB
3065 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
3066 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
3067 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
3068 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
3069 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
3070 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
3071 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3072 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
3073 MBB->addSuccessor(LoopMBB);
3074 MBB->addSuccessor(DoneMBB);
3076 MI->eraseFromParent();
3080 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
3081 // if the high register of the GR128 value must be cleared or false if
3082 // it's "don't care". SubReg is subreg_l32 when extending a GR32
3083 // and subreg_l64 when extending a GR64.
3085 SystemZTargetLowering::emitExt128(MachineInstr *MI,
3086 MachineBasicBlock *MBB,
3087 bool ClearEven, unsigned SubReg) const {
3088 MachineFunction &MF = *MBB->getParent();
3089 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
3090 MF.getTarget().getSubtargetImpl()->getInstrInfo());
3091 MachineRegisterInfo &MRI = MF.getRegInfo();
3092 DebugLoc DL = MI->getDebugLoc();
3094 unsigned Dest = MI->getOperand(0).getReg();
3095 unsigned Src = MI->getOperand(1).getReg();
3096 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3098 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
3100 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
3101 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
3103 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
3105 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
3106 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_h64);
3109 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
3110 .addReg(In128).addReg(Src).addImm(SubReg);
3112 MI->eraseFromParent();
3117 SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
3118 MachineBasicBlock *MBB,
3119 unsigned Opcode) const {
3120 MachineFunction &MF = *MBB->getParent();
3121 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
3122 MF.getTarget().getSubtargetImpl()->getInstrInfo());
3123 MachineRegisterInfo &MRI = MF.getRegInfo();
3124 DebugLoc DL = MI->getDebugLoc();
3126 MachineOperand DestBase = earlyUseOperand(MI->getOperand(0));
3127 uint64_t DestDisp = MI->getOperand(1).getImm();
3128 MachineOperand SrcBase = earlyUseOperand(MI->getOperand(2));
3129 uint64_t SrcDisp = MI->getOperand(3).getImm();
3130 uint64_t Length = MI->getOperand(4).getImm();
3132 // When generating more than one CLC, all but the last will need to
3133 // branch to the end when a difference is found.
3134 MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ?
3135 splitBlockAfter(MI, MBB) : nullptr);
3137 // Check for the loop form, in which operand 5 is the trip count.
3138 if (MI->getNumExplicitOperands() > 5) {
3139 bool HaveSingleBase = DestBase.isIdenticalTo(SrcBase);
3141 uint64_t StartCountReg = MI->getOperand(5).getReg();
3142 uint64_t StartSrcReg = forceReg(MI, SrcBase, TII);
3143 uint64_t StartDestReg = (HaveSingleBase ? StartSrcReg :
3144 forceReg(MI, DestBase, TII));
3146 const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass;
3147 uint64_t ThisSrcReg = MRI.createVirtualRegister(RC);
3148 uint64_t ThisDestReg = (HaveSingleBase ? ThisSrcReg :
3149 MRI.createVirtualRegister(RC));
3150 uint64_t NextSrcReg = MRI.createVirtualRegister(RC);
3151 uint64_t NextDestReg = (HaveSingleBase ? NextSrcReg :
3152 MRI.createVirtualRegister(RC));
3154 RC = &SystemZ::GR64BitRegClass;
3155 uint64_t ThisCountReg = MRI.createVirtualRegister(RC);
3156 uint64_t NextCountReg = MRI.createVirtualRegister(RC);
3158 MachineBasicBlock *StartMBB = MBB;
3159 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3160 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3161 MachineBasicBlock *NextMBB = (EndMBB ? emitBlockAfter(LoopMBB) : LoopMBB);
3164 // # fall through to LoopMMB
3165 MBB->addSuccessor(LoopMBB);
3168 // %ThisDestReg = phi [ %StartDestReg, StartMBB ],
3169 // [ %NextDestReg, NextMBB ]
3170 // %ThisSrcReg = phi [ %StartSrcReg, StartMBB ],
3171 // [ %NextSrcReg, NextMBB ]
3172 // %ThisCountReg = phi [ %StartCountReg, StartMBB ],
3173 // [ %NextCountReg, NextMBB ]
3174 // ( PFD 2, 768+DestDisp(%ThisDestReg) )
3175 // Opcode DestDisp(256,%ThisDestReg), SrcDisp(%ThisSrcReg)
3178 // The prefetch is used only for MVC. The JLH is used only for CLC.
3181 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisDestReg)
3182 .addReg(StartDestReg).addMBB(StartMBB)
3183 .addReg(NextDestReg).addMBB(NextMBB);
3184 if (!HaveSingleBase)
3185 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisSrcReg)
3186 .addReg(StartSrcReg).addMBB(StartMBB)
3187 .addReg(NextSrcReg).addMBB(NextMBB);
3188 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisCountReg)
3189 .addReg(StartCountReg).addMBB(StartMBB)
3190 .addReg(NextCountReg).addMBB(NextMBB);
3191 if (Opcode == SystemZ::MVC)
3192 BuildMI(MBB, DL, TII->get(SystemZ::PFD))
3193 .addImm(SystemZ::PFD_WRITE)
3194 .addReg(ThisDestReg).addImm(DestDisp + 768).addReg(0);
3195 BuildMI(MBB, DL, TII->get(Opcode))
3196 .addReg(ThisDestReg).addImm(DestDisp).addImm(256)
3197 .addReg(ThisSrcReg).addImm(SrcDisp);
3199 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3200 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3202 MBB->addSuccessor(EndMBB);
3203 MBB->addSuccessor(NextMBB);
3207 // %NextDestReg = LA 256(%ThisDestReg)
3208 // %NextSrcReg = LA 256(%ThisSrcReg)
3209 // %NextCountReg = AGHI %ThisCountReg, -1
3210 // CGHI %NextCountReg, 0
3212 // # fall through to DoneMMB
3214 // The AGHI, CGHI and JLH should be converted to BRCTG by later passes.
3217 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextDestReg)
3218 .addReg(ThisDestReg).addImm(256).addReg(0);
3219 if (!HaveSingleBase)
3220 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextSrcReg)
3221 .addReg(ThisSrcReg).addImm(256).addReg(0);
3222 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), NextCountReg)
3223 .addReg(ThisCountReg).addImm(-1);
3224 BuildMI(MBB, DL, TII->get(SystemZ::CGHI))
3225 .addReg(NextCountReg).addImm(0);
3226 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3227 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3229 MBB->addSuccessor(LoopMBB);
3230 MBB->addSuccessor(DoneMBB);
3232 DestBase = MachineOperand::CreateReg(NextDestReg, false);
3233 SrcBase = MachineOperand::CreateReg(NextSrcReg, false);
3237 // Handle any remaining bytes with straight-line code.
3238 while (Length > 0) {
3239 uint64_t ThisLength = std::min(Length, uint64_t(256));
3240 // The previous iteration might have created out-of-range displacements.
3241 // Apply them using LAY if so.
3242 if (!isUInt<12>(DestDisp)) {
3243 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3244 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3245 .addOperand(DestBase).addImm(DestDisp).addReg(0);
3246 DestBase = MachineOperand::CreateReg(Reg, false);
3249 if (!isUInt<12>(SrcDisp)) {
3250 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
3251 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
3252 .addOperand(SrcBase).addImm(SrcDisp).addReg(0);
3253 SrcBase = MachineOperand::CreateReg(Reg, false);
3256 BuildMI(*MBB, MI, DL, TII->get(Opcode))
3257 .addOperand(DestBase).addImm(DestDisp).addImm(ThisLength)
3258 .addOperand(SrcBase).addImm(SrcDisp);
3259 DestDisp += ThisLength;
3260 SrcDisp += ThisLength;
3261 Length -= ThisLength;
3262 // If there's another CLC to go, branch to the end if a difference
3264 if (EndMBB && Length > 0) {
3265 MachineBasicBlock *NextMBB = splitBlockBefore(MI, MBB);
3266 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3267 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
3269 MBB->addSuccessor(EndMBB);
3270 MBB->addSuccessor(NextMBB);
3275 MBB->addSuccessor(EndMBB);
3277 MBB->addLiveIn(SystemZ::CC);
3280 MI->eraseFromParent();
3284 // Decompose string pseudo-instruction MI into a loop that continually performs
3285 // Opcode until CC != 3.
3287 SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
3288 MachineBasicBlock *MBB,
3289 unsigned Opcode) const {
3290 MachineFunction &MF = *MBB->getParent();
3291 const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(
3292 MF.getTarget().getSubtargetImpl()->getInstrInfo());
3293 MachineRegisterInfo &MRI = MF.getRegInfo();
3294 DebugLoc DL = MI->getDebugLoc();
3296 uint64_t End1Reg = MI->getOperand(0).getReg();
3297 uint64_t Start1Reg = MI->getOperand(1).getReg();
3298 uint64_t Start2Reg = MI->getOperand(2).getReg();
3299 uint64_t CharReg = MI->getOperand(3).getReg();
3301 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass;
3302 uint64_t This1Reg = MRI.createVirtualRegister(RC);
3303 uint64_t This2Reg = MRI.createVirtualRegister(RC);
3304 uint64_t End2Reg = MRI.createVirtualRegister(RC);
3306 MachineBasicBlock *StartMBB = MBB;
3307 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
3308 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
3311 // # fall through to LoopMMB
3312 MBB->addSuccessor(LoopMBB);
3315 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ]
3316 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ]
3318 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0L
3320 // # fall through to DoneMMB
3322 // The load of R0L can be hoisted by post-RA LICM.
3325 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg)
3326 .addReg(Start1Reg).addMBB(StartMBB)
3327 .addReg(End1Reg).addMBB(LoopMBB);
3328 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg)
3329 .addReg(Start2Reg).addMBB(StartMBB)
3330 .addReg(End2Reg).addMBB(LoopMBB);
3331 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0L).addReg(CharReg);
3332 BuildMI(MBB, DL, TII->get(Opcode))
3333 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define)
3334 .addReg(This1Reg).addReg(This2Reg);
3335 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
3336 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB);
3337 MBB->addSuccessor(LoopMBB);
3338 MBB->addSuccessor(DoneMBB);
3340 DoneMBB->addLiveIn(SystemZ::CC);
3342 MI->eraseFromParent();
3346 MachineBasicBlock *SystemZTargetLowering::
3347 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
3348 switch (MI->getOpcode()) {
3349 case SystemZ::Select32Mux:
3350 case SystemZ::Select32:
3351 case SystemZ::SelectF32:
3352 case SystemZ::Select64:
3353 case SystemZ::SelectF64:
3354 case SystemZ::SelectF128:
3355 return emitSelect(MI, MBB);
3357 case SystemZ::CondStore8Mux:
3358 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, false);
3359 case SystemZ::CondStore8MuxInv:
3360 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, true);
3361 case SystemZ::CondStore16Mux:
3362 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, false);
3363 case SystemZ::CondStore16MuxInv:
3364 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, true);
3365 case SystemZ::CondStore8:
3366 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
3367 case SystemZ::CondStore8Inv:
3368 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
3369 case SystemZ::CondStore16:
3370 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
3371 case SystemZ::CondStore16Inv:
3372 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
3373 case SystemZ::CondStore32:
3374 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
3375 case SystemZ::CondStore32Inv:
3376 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
3377 case SystemZ::CondStore64:
3378 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
3379 case SystemZ::CondStore64Inv:
3380 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
3381 case SystemZ::CondStoreF32:
3382 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
3383 case SystemZ::CondStoreF32Inv:
3384 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
3385 case SystemZ::CondStoreF64:
3386 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
3387 case SystemZ::CondStoreF64Inv:
3388 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
3390 case SystemZ::AEXT128_64:
3391 return emitExt128(MI, MBB, false, SystemZ::subreg_l64);
3392 case SystemZ::ZEXT128_32:
3393 return emitExt128(MI, MBB, true, SystemZ::subreg_l32);
3394 case SystemZ::ZEXT128_64:
3395 return emitExt128(MI, MBB, true, SystemZ::subreg_l64);
3397 case SystemZ::ATOMIC_SWAPW:
3398 return emitAtomicLoadBinary(MI, MBB, 0, 0);
3399 case SystemZ::ATOMIC_SWAP_32:
3400 return emitAtomicLoadBinary(MI, MBB, 0, 32);
3401 case SystemZ::ATOMIC_SWAP_64:
3402 return emitAtomicLoadBinary(MI, MBB, 0, 64);
3404 case SystemZ::ATOMIC_LOADW_AR:
3405 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
3406 case SystemZ::ATOMIC_LOADW_AFI:
3407 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
3408 case SystemZ::ATOMIC_LOAD_AR:
3409 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
3410 case SystemZ::ATOMIC_LOAD_AHI:
3411 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
3412 case SystemZ::ATOMIC_LOAD_AFI:
3413 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
3414 case SystemZ::ATOMIC_LOAD_AGR:
3415 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
3416 case SystemZ::ATOMIC_LOAD_AGHI:
3417 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
3418 case SystemZ::ATOMIC_LOAD_AGFI:
3419 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
3421 case SystemZ::ATOMIC_LOADW_SR:
3422 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
3423 case SystemZ::ATOMIC_LOAD_SR:
3424 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
3425 case SystemZ::ATOMIC_LOAD_SGR:
3426 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
3428 case SystemZ::ATOMIC_LOADW_NR:
3429 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
3430 case SystemZ::ATOMIC_LOADW_NILH:
3431 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0);
3432 case SystemZ::ATOMIC_LOAD_NR:
3433 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
3434 case SystemZ::ATOMIC_LOAD_NILL:
3435 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32);
3436 case SystemZ::ATOMIC_LOAD_NILH:
3437 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32);
3438 case SystemZ::ATOMIC_LOAD_NILF:
3439 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32);
3440 case SystemZ::ATOMIC_LOAD_NGR:
3441 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
3442 case SystemZ::ATOMIC_LOAD_NILL64:
3443 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64);
3444 case SystemZ::ATOMIC_LOAD_NILH64:
3445 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64);
3446 case SystemZ::ATOMIC_LOAD_NIHL64:
3447 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64);
3448 case SystemZ::ATOMIC_LOAD_NIHH64:
3449 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64);
3450 case SystemZ::ATOMIC_LOAD_NILF64:
3451 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64);
3452 case SystemZ::ATOMIC_LOAD_NIHF64:
3453 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64);
3455 case SystemZ::ATOMIC_LOADW_OR:
3456 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
3457 case SystemZ::ATOMIC_LOADW_OILH:
3458 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 0);
3459 case SystemZ::ATOMIC_LOAD_OR:
3460 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
3461 case SystemZ::ATOMIC_LOAD_OILL:
3462 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 32);
3463 case SystemZ::ATOMIC_LOAD_OILH:
3464 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 32);
3465 case SystemZ::ATOMIC_LOAD_OILF:
3466 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 32);
3467 case SystemZ::ATOMIC_LOAD_OGR:
3468 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
3469 case SystemZ::ATOMIC_LOAD_OILL64:
3470 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL64, 64);
3471 case SystemZ::ATOMIC_LOAD_OILH64:
3472 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH64, 64);
3473 case SystemZ::ATOMIC_LOAD_OIHL64:
3474 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL64, 64);
3475 case SystemZ::ATOMIC_LOAD_OIHH64:
3476 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH64, 64);
3477 case SystemZ::ATOMIC_LOAD_OILF64:
3478 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF64, 64);
3479 case SystemZ::ATOMIC_LOAD_OIHF64:
3480 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF64, 64);
3482 case SystemZ::ATOMIC_LOADW_XR:
3483 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
3484 case SystemZ::ATOMIC_LOADW_XILF:
3485 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 0);
3486 case SystemZ::ATOMIC_LOAD_XR:
3487 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
3488 case SystemZ::ATOMIC_LOAD_XILF:
3489 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 32);
3490 case SystemZ::ATOMIC_LOAD_XGR:
3491 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
3492 case SystemZ::ATOMIC_LOAD_XILF64:
3493 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF64, 64);
3494 case SystemZ::ATOMIC_LOAD_XIHF64:
3495 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF64, 64);
3497 case SystemZ::ATOMIC_LOADW_NRi:
3498 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
3499 case SystemZ::ATOMIC_LOADW_NILHi:
3500 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0, true);
3501 case SystemZ::ATOMIC_LOAD_NRi:
3502 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
3503 case SystemZ::ATOMIC_LOAD_NILLi:
3504 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32, true);
3505 case SystemZ::ATOMIC_LOAD_NILHi:
3506 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32, true);
3507 case SystemZ::ATOMIC_LOAD_NILFi:
3508 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32, true);
3509 case SystemZ::ATOMIC_LOAD_NGRi:
3510 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
3511 case SystemZ::ATOMIC_LOAD_NILL64i:
3512 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64, true);
3513 case SystemZ::ATOMIC_LOAD_NILH64i:
3514 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64, true);
3515 case SystemZ::ATOMIC_LOAD_NIHL64i:
3516 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64, true);
3517 case SystemZ::ATOMIC_LOAD_NIHH64i:
3518 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64, true);
3519 case SystemZ::ATOMIC_LOAD_NILF64i:
3520 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64, true);
3521 case SystemZ::ATOMIC_LOAD_NIHF64i:
3522 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64, true);
3524 case SystemZ::ATOMIC_LOADW_MIN:
3525 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3526 SystemZ::CCMASK_CMP_LE, 0);
3527 case SystemZ::ATOMIC_LOAD_MIN_32:
3528 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3529 SystemZ::CCMASK_CMP_LE, 32);
3530 case SystemZ::ATOMIC_LOAD_MIN_64:
3531 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3532 SystemZ::CCMASK_CMP_LE, 64);
3534 case SystemZ::ATOMIC_LOADW_MAX:
3535 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3536 SystemZ::CCMASK_CMP_GE, 0);
3537 case SystemZ::ATOMIC_LOAD_MAX_32:
3538 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
3539 SystemZ::CCMASK_CMP_GE, 32);
3540 case SystemZ::ATOMIC_LOAD_MAX_64:
3541 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
3542 SystemZ::CCMASK_CMP_GE, 64);
3544 case SystemZ::ATOMIC_LOADW_UMIN:
3545 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3546 SystemZ::CCMASK_CMP_LE, 0);
3547 case SystemZ::ATOMIC_LOAD_UMIN_32:
3548 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3549 SystemZ::CCMASK_CMP_LE, 32);
3550 case SystemZ::ATOMIC_LOAD_UMIN_64:
3551 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3552 SystemZ::CCMASK_CMP_LE, 64);
3554 case SystemZ::ATOMIC_LOADW_UMAX:
3555 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3556 SystemZ::CCMASK_CMP_GE, 0);
3557 case SystemZ::ATOMIC_LOAD_UMAX_32:
3558 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
3559 SystemZ::CCMASK_CMP_GE, 32);
3560 case SystemZ::ATOMIC_LOAD_UMAX_64:
3561 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
3562 SystemZ::CCMASK_CMP_GE, 64);
3564 case SystemZ::ATOMIC_CMP_SWAPW:
3565 return emitAtomicCmpSwapW(MI, MBB);
3566 case SystemZ::MVCSequence:
3567 case SystemZ::MVCLoop:
3568 return emitMemMemWrapper(MI, MBB, SystemZ::MVC);
3569 case SystemZ::NCSequence:
3570 case SystemZ::NCLoop:
3571 return emitMemMemWrapper(MI, MBB, SystemZ::NC);
3572 case SystemZ::OCSequence:
3573 case SystemZ::OCLoop:
3574 return emitMemMemWrapper(MI, MBB, SystemZ::OC);
3575 case SystemZ::XCSequence:
3576 case SystemZ::XCLoop:
3577 return emitMemMemWrapper(MI, MBB, SystemZ::XC);
3578 case SystemZ::CLCSequence:
3579 case SystemZ::CLCLoop:
3580 return emitMemMemWrapper(MI, MBB, SystemZ::CLC);
3581 case SystemZ::CLSTLoop:
3582 return emitStringWrapper(MI, MBB, SystemZ::CLST);
3583 case SystemZ::MVSTLoop:
3584 return emitStringWrapper(MI, MBB, SystemZ::MVST);
3585 case SystemZ::SRSTLoop:
3586 return emitStringWrapper(MI, MBB, SystemZ::SRST);
3588 llvm_unreachable("Unexpected instr type to insert");