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"
23 #include "llvm/IR/Intrinsics.h"
28 #define DEBUG_TYPE "systemz-lower"
31 // Represents a sequence for extracting a 0/1 value from an IPM result:
32 // (((X ^ XORValue) + AddValue) >> Bit)
33 struct IPMConversion {
34 IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
35 : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
42 // Represents information about a comparison.
44 Comparison(SDValue Op0In, SDValue Op1In)
45 : Op0(Op0In), Op1(Op1In), Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
47 // The operands to the comparison.
50 // The opcode that should be used to compare Op0 and Op1.
53 // A SystemZICMP value. Only used for integer comparisons.
56 // The mask of CC values that Opcode can produce.
59 // The mask of CC values for which the original condition is true.
62 } // end anonymous namespace
64 // Classify VT as either 32 or 64 bit.
65 static bool is32Bit(EVT VT) {
66 switch (VT.getSimpleVT().SimpleTy) {
72 llvm_unreachable("Unsupported type");
76 // Return a version of MachineOperand that can be safely used before the
78 static MachineOperand earlyUseOperand(MachineOperand Op) {
84 SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &TM,
85 const SystemZSubtarget &STI)
86 : TargetLowering(TM), Subtarget(STI) {
87 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
89 // Set up the register classes.
90 if (Subtarget.hasHighWord())
91 addRegisterClass(MVT::i32, &SystemZ::GRX32BitRegClass);
93 addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
94 addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
95 if (Subtarget.hasVector()) {
96 addRegisterClass(MVT::f32, &SystemZ::VR32BitRegClass);
97 addRegisterClass(MVT::f64, &SystemZ::VR64BitRegClass);
99 addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
100 addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
102 addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
104 if (Subtarget.hasVector()) {
105 addRegisterClass(MVT::v16i8, &SystemZ::VR128BitRegClass);
106 addRegisterClass(MVT::v8i16, &SystemZ::VR128BitRegClass);
107 addRegisterClass(MVT::v4i32, &SystemZ::VR128BitRegClass);
108 addRegisterClass(MVT::v2i64, &SystemZ::VR128BitRegClass);
109 addRegisterClass(MVT::v4f32, &SystemZ::VR128BitRegClass);
110 addRegisterClass(MVT::v2f64, &SystemZ::VR128BitRegClass);
113 // Compute derived properties from the register classes
114 computeRegisterProperties(Subtarget.getRegisterInfo());
116 // Set up special registers.
117 setExceptionPointerRegister(SystemZ::R6D);
118 setExceptionSelectorRegister(SystemZ::R7D);
119 setStackPointerRegisterToSaveRestore(SystemZ::R15D);
121 // TODO: It may be better to default to latency-oriented scheduling, however
122 // LLVM's current latency-oriented scheduler can't handle physreg definitions
123 // such as SystemZ has with CC, so set this to the register-pressure
124 // scheduler, because it can.
125 setSchedulingPreference(Sched::RegPressure);
127 setBooleanContents(ZeroOrOneBooleanContent);
128 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
130 // Instructions are strings of 2-byte aligned 2-byte values.
131 setMinFunctionAlignment(2);
133 // Handle operations that are handled in a similar way for all types.
134 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
135 I <= MVT::LAST_FP_VALUETYPE;
137 MVT VT = MVT::SimpleValueType(I);
138 if (isTypeLegal(VT)) {
139 // Lower SET_CC into an IPM-based sequence.
140 setOperationAction(ISD::SETCC, VT, Custom);
142 // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
143 setOperationAction(ISD::SELECT, VT, Expand);
145 // Lower SELECT_CC and BR_CC into separate comparisons and branches.
146 setOperationAction(ISD::SELECT_CC, VT, Custom);
147 setOperationAction(ISD::BR_CC, VT, Custom);
151 // Expand jump table branches as address arithmetic followed by an
153 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
155 // Expand BRCOND into a BR_CC (see above).
156 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
158 // Handle integer types.
159 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
160 I <= MVT::LAST_INTEGER_VALUETYPE;
162 MVT VT = MVT::SimpleValueType(I);
163 if (isTypeLegal(VT)) {
164 // Expand individual DIV and REMs into DIVREMs.
165 setOperationAction(ISD::SDIV, VT, Expand);
166 setOperationAction(ISD::UDIV, VT, Expand);
167 setOperationAction(ISD::SREM, VT, Expand);
168 setOperationAction(ISD::UREM, VT, Expand);
169 setOperationAction(ISD::SDIVREM, VT, Custom);
170 setOperationAction(ISD::UDIVREM, VT, Custom);
172 // Lower ATOMIC_LOAD and ATOMIC_STORE into normal volatile loads and
173 // stores, putting a serialization instruction after the stores.
174 setOperationAction(ISD::ATOMIC_LOAD, VT, Custom);
175 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
177 // Lower ATOMIC_LOAD_SUB into ATOMIC_LOAD_ADD if LAA and LAAG are
178 // available, or if the operand is constant.
179 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
181 // Use POPCNT on z196 and above.
182 if (Subtarget.hasPopulationCount())
183 setOperationAction(ISD::CTPOP, VT, Custom);
185 setOperationAction(ISD::CTPOP, VT, Expand);
187 // No special instructions for these.
188 setOperationAction(ISD::CTTZ, VT, Expand);
189 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
190 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
191 setOperationAction(ISD::ROTR, VT, Expand);
193 // Use *MUL_LOHI where possible instead of MULH*.
194 setOperationAction(ISD::MULHS, VT, Expand);
195 setOperationAction(ISD::MULHU, VT, Expand);
196 setOperationAction(ISD::SMUL_LOHI, VT, Custom);
197 setOperationAction(ISD::UMUL_LOHI, VT, Custom);
199 // Only z196 and above have native support for conversions to unsigned.
200 if (!Subtarget.hasFPExtension())
201 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
205 // Type legalization will convert 8- and 16-bit atomic operations into
206 // forms that operate on i32s (but still keeping the original memory VT).
207 // Lower them into full i32 operations.
208 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
209 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
210 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
211 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
212 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
213 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
214 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
215 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
216 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
217 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
218 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
219 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
221 // z10 has instructions for signed but not unsigned FP conversion.
222 // Handle unsigned 32-bit types as signed 64-bit types.
223 if (!Subtarget.hasFPExtension()) {
224 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
225 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
228 // We have native support for a 64-bit CTLZ, via FLOGR.
229 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
230 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
232 // Give LowerOperation the chance to replace 64-bit ORs with subregs.
233 setOperationAction(ISD::OR, MVT::i64, Custom);
235 // FIXME: Can we support these natively?
236 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
237 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
238 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
240 // We have native instructions for i8, i16 and i32 extensions, but not i1.
241 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
242 for (MVT VT : MVT::integer_valuetypes()) {
243 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
244 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
245 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
248 // Handle the various types of symbolic address.
249 setOperationAction(ISD::ConstantPool, PtrVT, Custom);
250 setOperationAction(ISD::GlobalAddress, PtrVT, Custom);
251 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
252 setOperationAction(ISD::BlockAddress, PtrVT, Custom);
253 setOperationAction(ISD::JumpTable, PtrVT, Custom);
255 // We need to handle dynamic allocations specially because of the
256 // 160-byte area at the bottom of the stack.
257 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
259 // Use custom expanders so that we can force the function to use
261 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
262 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
264 // Handle prefetches with PFD or PFDRL.
265 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
267 for (MVT VT : MVT::vector_valuetypes()) {
268 // Assume by default that all vector operations need to be expanded.
269 for (unsigned Opcode = 0; Opcode < ISD::BUILTIN_OP_END; ++Opcode)
270 if (getOperationAction(Opcode, VT) == Legal)
271 setOperationAction(Opcode, VT, Expand);
273 // Likewise all truncating stores and extending loads.
274 for (MVT InnerVT : MVT::vector_valuetypes()) {
275 setTruncStoreAction(VT, InnerVT, Expand);
276 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
277 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
278 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
281 if (isTypeLegal(VT)) {
282 // These operations are legal for anything that can be stored in a
283 // vector register, even if there is no native support for the format
284 // as such. In particular, we can do these for v4f32 even though there
285 // are no specific instructions for that format.
286 setOperationAction(ISD::LOAD, VT, Legal);
287 setOperationAction(ISD::STORE, VT, Legal);
288 setOperationAction(ISD::VSELECT, VT, Legal);
289 setOperationAction(ISD::BITCAST, VT, Legal);
290 setOperationAction(ISD::UNDEF, VT, Legal);
292 // Likewise, except that we need to replace the nodes with something
294 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
295 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
299 // Handle integer vector types.
300 for (MVT VT : MVT::integer_vector_valuetypes()) {
301 if (isTypeLegal(VT)) {
302 // These operations have direct equivalents.
303 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Legal);
304 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Legal);
305 setOperationAction(ISD::ADD, VT, Legal);
306 setOperationAction(ISD::SUB, VT, Legal);
307 if (VT != MVT::v2i64)
308 setOperationAction(ISD::MUL, VT, Legal);
309 setOperationAction(ISD::AND, VT, Legal);
310 setOperationAction(ISD::OR, VT, Legal);
311 setOperationAction(ISD::XOR, VT, Legal);
312 setOperationAction(ISD::CTPOP, VT, Custom);
313 setOperationAction(ISD::CTTZ, VT, Legal);
314 setOperationAction(ISD::CTLZ, VT, Legal);
315 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Custom);
316 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Custom);
318 // Convert a GPR scalar to a vector by inserting it into element 0.
319 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
321 // Use a series of unpacks for extensions.
322 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
323 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
325 // Detect shifts by a scalar amount and convert them into
327 setOperationAction(ISD::SHL, VT, Custom);
328 setOperationAction(ISD::SRA, VT, Custom);
329 setOperationAction(ISD::SRL, VT, Custom);
331 // At present ROTL isn't matched by DAGCombiner. ROTR should be
332 // converted into ROTL.
333 setOperationAction(ISD::ROTL, VT, Expand);
334 setOperationAction(ISD::ROTR, VT, Expand);
336 // Map SETCCs onto one of VCE, VCH or VCHL, swapping the operands
337 // and inverting the result as necessary.
338 setOperationAction(ISD::SETCC, VT, Custom);
342 if (Subtarget.hasVector()) {
343 // There should be no need to check for float types other than v2f64
344 // since <2 x f32> isn't a legal type.
345 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
346 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
347 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
348 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
351 // Handle floating-point types.
352 for (unsigned I = MVT::FIRST_FP_VALUETYPE;
353 I <= MVT::LAST_FP_VALUETYPE;
355 MVT VT = MVT::SimpleValueType(I);
356 if (isTypeLegal(VT)) {
357 // We can use FI for FRINT.
358 setOperationAction(ISD::FRINT, VT, Legal);
360 // We can use the extended form of FI for other rounding operations.
361 if (Subtarget.hasFPExtension()) {
362 setOperationAction(ISD::FNEARBYINT, VT, Legal);
363 setOperationAction(ISD::FFLOOR, VT, Legal);
364 setOperationAction(ISD::FCEIL, VT, Legal);
365 setOperationAction(ISD::FTRUNC, VT, Legal);
366 setOperationAction(ISD::FROUND, VT, Legal);
369 // No special instructions for these.
370 setOperationAction(ISD::FSIN, VT, Expand);
371 setOperationAction(ISD::FCOS, VT, Expand);
372 setOperationAction(ISD::FREM, VT, Expand);
376 // Handle floating-point vector types.
377 if (Subtarget.hasVector()) {
378 // Scalar-to-vector conversion is just a subreg.
379 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal);
380 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
382 // Some insertions and extractions can be done directly but others
383 // need to go via integers.
384 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
385 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
386 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
387 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
389 // These operations have direct equivalents.
390 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
391 setOperationAction(ISD::FNEG, MVT::v2f64, Legal);
392 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
393 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
394 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
395 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
396 setOperationAction(ISD::FABS, MVT::v2f64, Legal);
397 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
398 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
399 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
400 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
401 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
402 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
403 setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
406 // We have fused multiply-addition for f32 and f64 but not f128.
407 setOperationAction(ISD::FMA, MVT::f32, Legal);
408 setOperationAction(ISD::FMA, MVT::f64, Legal);
409 setOperationAction(ISD::FMA, MVT::f128, Expand);
411 // Needed so that we don't try to implement f128 constant loads using
412 // a load-and-extend of a f80 constant (in cases where the constant
413 // would fit in an f80).
414 for (MVT VT : MVT::fp_valuetypes())
415 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
417 // Floating-point truncation and stores need to be done separately.
418 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
419 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
420 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
422 // We have 64-bit FPR<->GPR moves, but need special handling for
424 if (!Subtarget.hasVector()) {
425 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
426 setOperationAction(ISD::BITCAST, MVT::f32, Custom);
429 // VASTART and VACOPY need to deal with the SystemZ-specific varargs
430 // structure, but VAEND is a no-op.
431 setOperationAction(ISD::VASTART, MVT::Other, Custom);
432 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
433 setOperationAction(ISD::VAEND, MVT::Other, Expand);
435 // Codes for which we want to perform some z-specific combinations.
436 setTargetDAGCombine(ISD::SIGN_EXTEND);
437 setTargetDAGCombine(ISD::STORE);
438 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
439 setTargetDAGCombine(ISD::FP_ROUND);
441 // Handle intrinsics.
442 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
443 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
445 // We want to use MVC in preference to even a single load/store pair.
446 MaxStoresPerMemcpy = 0;
447 MaxStoresPerMemcpyOptSize = 0;
449 // The main memset sequence is a byte store followed by an MVC.
450 // Two STC or MV..I stores win over that, but the kind of fused stores
451 // generated by target-independent code don't when the byte value is
452 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better
453 // than "STC;MVC". Handle the choice in target-specific code instead.
454 MaxStoresPerMemset = 0;
455 MaxStoresPerMemsetOptSize = 0;
458 EVT SystemZTargetLowering::getSetCCResultType(const DataLayout &DL,
459 LLVMContext &, EVT VT) const {
462 return VT.changeVectorElementTypeToInteger();
465 bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
466 VT = VT.getScalarType();
471 switch (VT.getSimpleVT().SimpleTy) {
484 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
485 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
486 return Imm.isZero() || Imm.isNegZero();
489 bool SystemZTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
490 // We can use CGFI or CLGFI.
491 return isInt<32>(Imm) || isUInt<32>(Imm);
494 bool SystemZTargetLowering::isLegalAddImmediate(int64_t Imm) const {
495 // We can use ALGFI or SLGFI.
496 return isUInt<32>(Imm) || isUInt<32>(-Imm);
499 bool SystemZTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
503 // Unaligned accesses should never be slower than the expanded version.
504 // We check specifically for aligned accesses in the few cases where
505 // they are required.
511 bool SystemZTargetLowering::isLegalAddressingMode(const DataLayout &DL,
512 const AddrMode &AM, Type *Ty,
514 // Punt on globals for now, although they can be used in limited
515 // RELATIVE LONG cases.
519 // Require a 20-bit signed offset.
520 if (!isInt<20>(AM.BaseOffs))
523 // Indexing is OK but no scale factor can be applied.
524 return AM.Scale == 0 || AM.Scale == 1;
527 bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
528 if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
530 unsigned FromBits = FromType->getPrimitiveSizeInBits();
531 unsigned ToBits = ToType->getPrimitiveSizeInBits();
532 return FromBits > ToBits;
535 bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
536 if (!FromVT.isInteger() || !ToVT.isInteger())
538 unsigned FromBits = FromVT.getSizeInBits();
539 unsigned ToBits = ToVT.getSizeInBits();
540 return FromBits > ToBits;
543 //===----------------------------------------------------------------------===//
544 // Inline asm support
545 //===----------------------------------------------------------------------===//
547 TargetLowering::ConstraintType
548 SystemZTargetLowering::getConstraintType(StringRef Constraint) const {
549 if (Constraint.size() == 1) {
550 switch (Constraint[0]) {
551 case 'a': // Address register
552 case 'd': // Data register (equivalent to 'r')
553 case 'f': // Floating-point register
554 case 'h': // High-part register
555 case 'r': // General-purpose register
556 return C_RegisterClass;
558 case 'Q': // Memory with base and unsigned 12-bit displacement
559 case 'R': // Likewise, plus an index
560 case 'S': // Memory with base and signed 20-bit displacement
561 case 'T': // Likewise, plus an index
562 case 'm': // Equivalent to 'T'.
565 case 'I': // Unsigned 8-bit constant
566 case 'J': // Unsigned 12-bit constant
567 case 'K': // Signed 16-bit constant
568 case 'L': // Signed 20-bit displacement (on all targets we support)
569 case 'M': // 0x7fffffff
576 return TargetLowering::getConstraintType(Constraint);
579 TargetLowering::ConstraintWeight SystemZTargetLowering::
580 getSingleConstraintMatchWeight(AsmOperandInfo &info,
581 const char *constraint) const {
582 ConstraintWeight weight = CW_Invalid;
583 Value *CallOperandVal = info.CallOperandVal;
584 // If we don't have a value, we can't do a match,
585 // but allow it at the lowest weight.
588 Type *type = CallOperandVal->getType();
589 // Look at the constraint type.
590 switch (*constraint) {
592 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
595 case 'a': // Address register
596 case 'd': // Data register (equivalent to 'r')
597 case 'h': // High-part register
598 case 'r': // General-purpose register
599 if (CallOperandVal->getType()->isIntegerTy())
600 weight = CW_Register;
603 case 'f': // Floating-point register
604 if (type->isFloatingPointTy())
605 weight = CW_Register;
608 case 'I': // Unsigned 8-bit constant
609 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
610 if (isUInt<8>(C->getZExtValue()))
611 weight = CW_Constant;
614 case 'J': // Unsigned 12-bit constant
615 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
616 if (isUInt<12>(C->getZExtValue()))
617 weight = CW_Constant;
620 case 'K': // Signed 16-bit constant
621 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
622 if (isInt<16>(C->getSExtValue()))
623 weight = CW_Constant;
626 case 'L': // Signed 20-bit displacement (on all targets we support)
627 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
628 if (isInt<20>(C->getSExtValue()))
629 weight = CW_Constant;
632 case 'M': // 0x7fffffff
633 if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
634 if (C->getZExtValue() == 0x7fffffff)
635 weight = CW_Constant;
641 // Parse a "{tNNN}" register constraint for which the register type "t"
642 // has already been verified. MC is the class associated with "t" and
643 // Map maps 0-based register numbers to LLVM register numbers.
644 static std::pair<unsigned, const TargetRegisterClass *>
645 parseRegisterNumber(StringRef Constraint, const TargetRegisterClass *RC,
646 const unsigned *Map) {
647 assert(*(Constraint.end()-1) == '}' && "Missing '}'");
648 if (isdigit(Constraint[2])) {
651 Constraint.slice(2, Constraint.size() - 1).getAsInteger(10, Index);
652 if (!Failed && Index < 16 && Map[Index])
653 return std::make_pair(Map[Index], RC);
655 return std::make_pair(0U, nullptr);
658 std::pair<unsigned, const TargetRegisterClass *>
659 SystemZTargetLowering::getRegForInlineAsmConstraint(
660 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
661 if (Constraint.size() == 1) {
662 // GCC Constraint Letters
663 switch (Constraint[0]) {
665 case 'd': // Data register (equivalent to 'r')
666 case 'r': // General-purpose register
668 return std::make_pair(0U, &SystemZ::GR64BitRegClass);
669 else if (VT == MVT::i128)
670 return std::make_pair(0U, &SystemZ::GR128BitRegClass);
671 return std::make_pair(0U, &SystemZ::GR32BitRegClass);
673 case 'a': // Address register
675 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
676 else if (VT == MVT::i128)
677 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
678 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
680 case 'h': // High-part register (an LLVM extension)
681 return std::make_pair(0U, &SystemZ::GRH32BitRegClass);
683 case 'f': // Floating-point register
685 return std::make_pair(0U, &SystemZ::FP64BitRegClass);
686 else if (VT == MVT::f128)
687 return std::make_pair(0U, &SystemZ::FP128BitRegClass);
688 return std::make_pair(0U, &SystemZ::FP32BitRegClass);
691 if (Constraint.size() > 0 && Constraint[0] == '{') {
692 // We need to override the default register parsing for GPRs and FPRs
693 // because the interpretation depends on VT. The internal names of
694 // the registers are also different from the external names
695 // (F0D and F0S instead of F0, etc.).
696 if (Constraint[1] == 'r') {
698 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
699 SystemZMC::GR32Regs);
701 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
702 SystemZMC::GR128Regs);
703 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
704 SystemZMC::GR64Regs);
706 if (Constraint[1] == 'f') {
708 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
709 SystemZMC::FP32Regs);
711 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
712 SystemZMC::FP128Regs);
713 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
714 SystemZMC::FP64Regs);
717 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
720 void SystemZTargetLowering::
721 LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
722 std::vector<SDValue> &Ops,
723 SelectionDAG &DAG) const {
724 // Only support length 1 constraints for now.
725 if (Constraint.length() == 1) {
726 switch (Constraint[0]) {
727 case 'I': // Unsigned 8-bit constant
728 if (auto *C = dyn_cast<ConstantSDNode>(Op))
729 if (isUInt<8>(C->getZExtValue()))
730 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
734 case 'J': // Unsigned 12-bit constant
735 if (auto *C = dyn_cast<ConstantSDNode>(Op))
736 if (isUInt<12>(C->getZExtValue()))
737 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
741 case 'K': // Signed 16-bit constant
742 if (auto *C = dyn_cast<ConstantSDNode>(Op))
743 if (isInt<16>(C->getSExtValue()))
744 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
748 case 'L': // Signed 20-bit displacement (on all targets we support)
749 if (auto *C = dyn_cast<ConstantSDNode>(Op))
750 if (isInt<20>(C->getSExtValue()))
751 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
755 case 'M': // 0x7fffffff
756 if (auto *C = dyn_cast<ConstantSDNode>(Op))
757 if (C->getZExtValue() == 0x7fffffff)
758 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
763 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
766 //===----------------------------------------------------------------------===//
767 // Calling conventions
768 //===----------------------------------------------------------------------===//
770 #include "SystemZGenCallingConv.inc"
772 bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
773 Type *ToType) const {
774 return isTruncateFree(FromType, ToType);
777 bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
778 if (!CI->isTailCall())
783 // We do not yet support 128-bit single-element vector types. If the user
784 // attempts to use such types as function argument or return type, prefer
785 // to error out instead of emitting code violating the ABI.
786 static void VerifyVectorType(MVT VT, EVT ArgVT) {
787 if (ArgVT.isVector() && !VT.isVector())
788 report_fatal_error("Unsupported vector argument or return type");
791 static void VerifyVectorTypes(const SmallVectorImpl<ISD::InputArg> &Ins) {
792 for (unsigned i = 0; i < Ins.size(); ++i)
793 VerifyVectorType(Ins[i].VT, Ins[i].ArgVT);
796 static void VerifyVectorTypes(const SmallVectorImpl<ISD::OutputArg> &Outs) {
797 for (unsigned i = 0; i < Outs.size(); ++i)
798 VerifyVectorType(Outs[i].VT, Outs[i].ArgVT);
801 // Value is a value that has been passed to us in the location described by VA
802 // (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
803 // any loads onto Chain.
804 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL,
805 CCValAssign &VA, SDValue Chain,
807 // If the argument has been promoted from a smaller type, insert an
808 // assertion to capture this.
809 if (VA.getLocInfo() == CCValAssign::SExt)
810 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
811 DAG.getValueType(VA.getValVT()));
812 else if (VA.getLocInfo() == CCValAssign::ZExt)
813 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
814 DAG.getValueType(VA.getValVT()));
817 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
818 else if (VA.getLocInfo() == CCValAssign::Indirect)
819 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
820 MachinePointerInfo(), false, false, false, 0);
821 else if (VA.getLocInfo() == CCValAssign::BCvt) {
822 // If this is a short vector argument loaded from the stack,
823 // extend from i64 to full vector size and then bitcast.
824 assert(VA.getLocVT() == MVT::i64);
825 assert(VA.getValVT().isVector());
826 Value = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i64,
827 Value, DAG.getUNDEF(MVT::i64));
828 Value = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Value);
830 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
834 // Value is a value of type VA.getValVT() that we need to copy into
835 // the location described by VA. Return a copy of Value converted to
836 // VA.getValVT(). The caller is responsible for handling indirect values.
837 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL,
838 CCValAssign &VA, SDValue Value) {
839 switch (VA.getLocInfo()) {
840 case CCValAssign::SExt:
841 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
842 case CCValAssign::ZExt:
843 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
844 case CCValAssign::AExt:
845 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
846 case CCValAssign::BCvt:
847 // If this is a short vector argument to be stored to the stack,
848 // bitcast to v2i64 and then extract first element.
849 assert(VA.getLocVT() == MVT::i64);
850 assert(VA.getValVT().isVector());
851 Value = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Value);
852 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VA.getLocVT(), Value,
853 DAG.getConstant(0, DL, MVT::i32));
854 case CCValAssign::Full:
857 llvm_unreachable("Unhandled getLocInfo()");
861 SDValue SystemZTargetLowering::
862 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
863 const SmallVectorImpl<ISD::InputArg> &Ins,
864 SDLoc DL, SelectionDAG &DAG,
865 SmallVectorImpl<SDValue> &InVals) const {
866 MachineFunction &MF = DAG.getMachineFunction();
867 MachineFrameInfo *MFI = MF.getFrameInfo();
868 MachineRegisterInfo &MRI = MF.getRegInfo();
869 SystemZMachineFunctionInfo *FuncInfo =
870 MF.getInfo<SystemZMachineFunctionInfo>();
872 static_cast<const SystemZFrameLowering *>(Subtarget.getFrameLowering());
874 // Detect unsupported vector argument types.
875 if (Subtarget.hasVector())
876 VerifyVectorTypes(Ins);
878 // Assign locations to all of the incoming arguments.
879 SmallVector<CCValAssign, 16> ArgLocs;
880 SystemZCCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
881 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
883 unsigned NumFixedGPRs = 0;
884 unsigned NumFixedFPRs = 0;
885 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
887 CCValAssign &VA = ArgLocs[I];
888 EVT LocVT = VA.getLocVT();
890 // Arguments passed in registers
891 const TargetRegisterClass *RC;
892 switch (LocVT.getSimpleVT().SimpleTy) {
894 // Integers smaller than i64 should be promoted to i64.
895 llvm_unreachable("Unexpected argument type");
898 RC = &SystemZ::GR32BitRegClass;
902 RC = &SystemZ::GR64BitRegClass;
906 RC = &SystemZ::FP32BitRegClass;
910 RC = &SystemZ::FP64BitRegClass;
918 RC = &SystemZ::VR128BitRegClass;
922 unsigned VReg = MRI.createVirtualRegister(RC);
923 MRI.addLiveIn(VA.getLocReg(), VReg);
924 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
926 assert(VA.isMemLoc() && "Argument not register or memory");
928 // Create the frame index object for this incoming parameter.
929 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
930 VA.getLocMemOffset(), true);
932 // Create the SelectionDAG nodes corresponding to a load
933 // from this parameter. Unpromoted ints and floats are
934 // passed as right-justified 8-byte values.
935 EVT PtrVT = getPointerTy(DAG.getDataLayout());
936 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
937 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
938 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
939 DAG.getIntPtrConstant(4, DL));
940 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
941 MachinePointerInfo::getFixedStack(FI),
942 false, false, false, 0);
945 // Convert the value of the argument register into the value that's
947 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
951 // Save the number of non-varargs registers for later use by va_start, etc.
952 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
953 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
955 // Likewise the address (in the form of a frame index) of where the
956 // first stack vararg would be. The 1-byte size here is arbitrary.
957 int64_t StackSize = CCInfo.getNextStackOffset();
958 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true));
960 // ...and a similar frame index for the caller-allocated save area
961 // that will be used to store the incoming registers.
962 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea();
963 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true);
964 FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
966 // Store the FPR varargs in the reserved frame slots. (We store the
967 // GPRs as part of the prologue.)
968 if (NumFixedFPRs < SystemZ::NumArgFPRs) {
969 SDValue MemOps[SystemZ::NumArgFPRs];
970 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) {
971 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]);
972 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true);
973 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
974 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I],
975 &SystemZ::FP64BitRegClass);
976 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
977 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
978 MachinePointerInfo::getFixedStack(FI),
982 // Join the stores, which are independent of one another.
983 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
984 makeArrayRef(&MemOps[NumFixedFPRs],
985 SystemZ::NumArgFPRs-NumFixedFPRs));
992 static bool canUseSiblingCall(const CCState &ArgCCInfo,
993 SmallVectorImpl<CCValAssign> &ArgLocs) {
994 // Punt if there are any indirect or stack arguments, or if the call
995 // needs the call-saved argument register R6.
996 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
997 CCValAssign &VA = ArgLocs[I];
998 if (VA.getLocInfo() == CCValAssign::Indirect)
1002 unsigned Reg = VA.getLocReg();
1003 if (Reg == SystemZ::R6H || Reg == SystemZ::R6L || Reg == SystemZ::R6D)
1010 SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
1011 SmallVectorImpl<SDValue> &InVals) const {
1012 SelectionDAG &DAG = CLI.DAG;
1014 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1015 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1016 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1017 SDValue Chain = CLI.Chain;
1018 SDValue Callee = CLI.Callee;
1019 bool &IsTailCall = CLI.IsTailCall;
1020 CallingConv::ID CallConv = CLI.CallConv;
1021 bool IsVarArg = CLI.IsVarArg;
1022 MachineFunction &MF = DAG.getMachineFunction();
1023 EVT PtrVT = getPointerTy(MF.getDataLayout());
1025 // Detect unsupported vector argument and return types.
1026 if (Subtarget.hasVector()) {
1027 VerifyVectorTypes(Outs);
1028 VerifyVectorTypes(Ins);
1031 // Analyze the operands of the call, assigning locations to each operand.
1032 SmallVector<CCValAssign, 16> ArgLocs;
1033 SystemZCCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
1034 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
1036 // We don't support GuaranteedTailCallOpt, only automatically-detected
1038 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs))
1041 // Get a count of how many bytes are to be pushed on the stack.
1042 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
1044 // Mark the start of the call.
1046 Chain = DAG.getCALLSEQ_START(Chain,
1047 DAG.getConstant(NumBytes, DL, PtrVT, true),
1050 // Copy argument values to their designated locations.
1051 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
1052 SmallVector<SDValue, 8> MemOpChains;
1054 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
1055 CCValAssign &VA = ArgLocs[I];
1056 SDValue ArgValue = OutVals[I];
1058 if (VA.getLocInfo() == CCValAssign::Indirect) {
1059 // Store the argument in a stack slot and pass its address.
1060 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1061 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1062 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot,
1063 MachinePointerInfo::getFixedStack(FI),
1065 ArgValue = SpillSlot;
1067 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
1070 // Queue up the argument copies and emit them at the end.
1071 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
1073 assert(VA.isMemLoc() && "Argument not register or memory");
1075 // Work out the address of the stack slot. Unpromoted ints and
1076 // floats are passed as right-justified 8-byte values.
1077 if (!StackPtr.getNode())
1078 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT);
1079 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset();
1080 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
1082 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
1083 DAG.getIntPtrConstant(Offset, DL));
1086 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address,
1087 MachinePointerInfo(),
1092 // Join the stores, which are independent of one another.
1093 if (!MemOpChains.empty())
1094 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
1096 // Accept direct calls by converting symbolic call addresses to the
1097 // associated Target* opcodes. Force %r1 to be used for indirect
1100 if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1101 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
1102 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
1103 } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1104 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
1105 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
1106 } else if (IsTailCall) {
1107 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
1108 Glue = Chain.getValue(1);
1109 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
1112 // Build a sequence of copy-to-reg nodes, chained and glued together.
1113 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
1114 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
1115 RegsToPass[I].second, Glue);
1116 Glue = Chain.getValue(1);
1119 // The first call operand is the chain and the second is the target address.
1120 SmallVector<SDValue, 8> Ops;
1121 Ops.push_back(Chain);
1122 Ops.push_back(Callee);
1124 // Add argument registers to the end of the list so that they are
1125 // known live into the call.
1126 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
1127 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
1128 RegsToPass[I].second.getValueType()));
1130 // Add a register mask operand representing the call-preserved registers.
1131 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
1132 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
1133 assert(Mask && "Missing call preserved mask for calling convention");
1134 Ops.push_back(DAG.getRegisterMask(Mask));
1136 // Glue the call to the argument copies, if any.
1138 Ops.push_back(Glue);
1141 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1143 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, Ops);
1144 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, Ops);
1145 Glue = Chain.getValue(1);
1147 // Mark the end of the call, which is glued to the call itself.
1148 Chain = DAG.getCALLSEQ_END(Chain,
1149 DAG.getConstant(NumBytes, DL, PtrVT, true),
1150 DAG.getConstant(0, DL, PtrVT, true),
1152 Glue = Chain.getValue(1);
1154 // Assign locations to each value returned by this call.
1155 SmallVector<CCValAssign, 16> RetLocs;
1156 CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
1157 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
1159 // Copy all of the result registers out of their specified physreg.
1160 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
1161 CCValAssign &VA = RetLocs[I];
1163 // Copy the value out, gluing the copy to the end of the call sequence.
1164 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
1165 VA.getLocVT(), Glue);
1166 Chain = RetValue.getValue(1);
1167 Glue = RetValue.getValue(2);
1169 // Convert the value of the return register into the value that's
1171 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
1178 SystemZTargetLowering::LowerReturn(SDValue Chain,
1179 CallingConv::ID CallConv, bool IsVarArg,
1180 const SmallVectorImpl<ISD::OutputArg> &Outs,
1181 const SmallVectorImpl<SDValue> &OutVals,
1182 SDLoc DL, SelectionDAG &DAG) const {
1183 MachineFunction &MF = DAG.getMachineFunction();
1185 // Detect unsupported vector return types.
1186 if (Subtarget.hasVector())
1187 VerifyVectorTypes(Outs);
1189 // Assign locations to each returned value.
1190 SmallVector<CCValAssign, 16> RetLocs;
1191 CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
1192 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
1194 // Quick exit for void returns
1195 if (RetLocs.empty())
1196 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain);
1198 // Copy the result values into the output registers.
1200 SmallVector<SDValue, 4> RetOps;
1201 RetOps.push_back(Chain);
1202 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
1203 CCValAssign &VA = RetLocs[I];
1204 SDValue RetValue = OutVals[I];
1206 // Make the return register live on exit.
1207 assert(VA.isRegLoc() && "Can only return in registers!");
1209 // Promote the value as required.
1210 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
1212 // Chain and glue the copies together.
1213 unsigned Reg = VA.getLocReg();
1214 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
1215 Glue = Chain.getValue(1);
1216 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
1219 // Update chain and glue.
1222 RetOps.push_back(Glue);
1224 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, RetOps);
1227 SDValue SystemZTargetLowering::
1228 prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, SelectionDAG &DAG) const {
1229 return DAG.getNode(SystemZISD::SERIALIZE, DL, MVT::Other, Chain);
1232 // Return true if Op is an intrinsic node with chain that returns the CC value
1233 // as its only (other) argument. Provide the associated SystemZISD opcode and
1234 // the mask of valid CC values if so.
1235 static bool isIntrinsicWithCCAndChain(SDValue Op, unsigned &Opcode,
1236 unsigned &CCValid) {
1237 unsigned Id = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1239 case Intrinsic::s390_tbegin:
1240 Opcode = SystemZISD::TBEGIN;
1241 CCValid = SystemZ::CCMASK_TBEGIN;
1244 case Intrinsic::s390_tbegin_nofloat:
1245 Opcode = SystemZISD::TBEGIN_NOFLOAT;
1246 CCValid = SystemZ::CCMASK_TBEGIN;
1249 case Intrinsic::s390_tend:
1250 Opcode = SystemZISD::TEND;
1251 CCValid = SystemZ::CCMASK_TEND;
1259 // Return true if Op is an intrinsic node without chain that returns the
1260 // CC value as its final argument. Provide the associated SystemZISD
1261 // opcode and the mask of valid CC values if so.
1262 static bool isIntrinsicWithCC(SDValue Op, unsigned &Opcode, unsigned &CCValid) {
1263 unsigned Id = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1265 case Intrinsic::s390_vpkshs:
1266 case Intrinsic::s390_vpksfs:
1267 case Intrinsic::s390_vpksgs:
1268 Opcode = SystemZISD::PACKS_CC;
1269 CCValid = SystemZ::CCMASK_VCMP;
1272 case Intrinsic::s390_vpklshs:
1273 case Intrinsic::s390_vpklsfs:
1274 case Intrinsic::s390_vpklsgs:
1275 Opcode = SystemZISD::PACKLS_CC;
1276 CCValid = SystemZ::CCMASK_VCMP;
1279 case Intrinsic::s390_vceqbs:
1280 case Intrinsic::s390_vceqhs:
1281 case Intrinsic::s390_vceqfs:
1282 case Intrinsic::s390_vceqgs:
1283 Opcode = SystemZISD::VICMPES;
1284 CCValid = SystemZ::CCMASK_VCMP;
1287 case Intrinsic::s390_vchbs:
1288 case Intrinsic::s390_vchhs:
1289 case Intrinsic::s390_vchfs:
1290 case Intrinsic::s390_vchgs:
1291 Opcode = SystemZISD::VICMPHS;
1292 CCValid = SystemZ::CCMASK_VCMP;
1295 case Intrinsic::s390_vchlbs:
1296 case Intrinsic::s390_vchlhs:
1297 case Intrinsic::s390_vchlfs:
1298 case Intrinsic::s390_vchlgs:
1299 Opcode = SystemZISD::VICMPHLS;
1300 CCValid = SystemZ::CCMASK_VCMP;
1303 case Intrinsic::s390_vtm:
1304 Opcode = SystemZISD::VTM;
1305 CCValid = SystemZ::CCMASK_VCMP;
1308 case Intrinsic::s390_vfaebs:
1309 case Intrinsic::s390_vfaehs:
1310 case Intrinsic::s390_vfaefs:
1311 Opcode = SystemZISD::VFAE_CC;
1312 CCValid = SystemZ::CCMASK_ANY;
1315 case Intrinsic::s390_vfaezbs:
1316 case Intrinsic::s390_vfaezhs:
1317 case Intrinsic::s390_vfaezfs:
1318 Opcode = SystemZISD::VFAEZ_CC;
1319 CCValid = SystemZ::CCMASK_ANY;
1322 case Intrinsic::s390_vfeebs:
1323 case Intrinsic::s390_vfeehs:
1324 case Intrinsic::s390_vfeefs:
1325 Opcode = SystemZISD::VFEE_CC;
1326 CCValid = SystemZ::CCMASK_ANY;
1329 case Intrinsic::s390_vfeezbs:
1330 case Intrinsic::s390_vfeezhs:
1331 case Intrinsic::s390_vfeezfs:
1332 Opcode = SystemZISD::VFEEZ_CC;
1333 CCValid = SystemZ::CCMASK_ANY;
1336 case Intrinsic::s390_vfenebs:
1337 case Intrinsic::s390_vfenehs:
1338 case Intrinsic::s390_vfenefs:
1339 Opcode = SystemZISD::VFENE_CC;
1340 CCValid = SystemZ::CCMASK_ANY;
1343 case Intrinsic::s390_vfenezbs:
1344 case Intrinsic::s390_vfenezhs:
1345 case Intrinsic::s390_vfenezfs:
1346 Opcode = SystemZISD::VFENEZ_CC;
1347 CCValid = SystemZ::CCMASK_ANY;
1350 case Intrinsic::s390_vistrbs:
1351 case Intrinsic::s390_vistrhs:
1352 case Intrinsic::s390_vistrfs:
1353 Opcode = SystemZISD::VISTR_CC;
1354 CCValid = SystemZ::CCMASK_0 | SystemZ::CCMASK_3;
1357 case Intrinsic::s390_vstrcbs:
1358 case Intrinsic::s390_vstrchs:
1359 case Intrinsic::s390_vstrcfs:
1360 Opcode = SystemZISD::VSTRC_CC;
1361 CCValid = SystemZ::CCMASK_ANY;
1364 case Intrinsic::s390_vstrczbs:
1365 case Intrinsic::s390_vstrczhs:
1366 case Intrinsic::s390_vstrczfs:
1367 Opcode = SystemZISD::VSTRCZ_CC;
1368 CCValid = SystemZ::CCMASK_ANY;
1371 case Intrinsic::s390_vfcedbs:
1372 Opcode = SystemZISD::VFCMPES;
1373 CCValid = SystemZ::CCMASK_VCMP;
1376 case Intrinsic::s390_vfchdbs:
1377 Opcode = SystemZISD::VFCMPHS;
1378 CCValid = SystemZ::CCMASK_VCMP;
1381 case Intrinsic::s390_vfchedbs:
1382 Opcode = SystemZISD::VFCMPHES;
1383 CCValid = SystemZ::CCMASK_VCMP;
1386 case Intrinsic::s390_vftcidb:
1387 Opcode = SystemZISD::VFTCI;
1388 CCValid = SystemZ::CCMASK_VCMP;
1396 // Emit an intrinsic with chain with a glued value instead of its CC result.
1397 static SDValue emitIntrinsicWithChainAndGlue(SelectionDAG &DAG, SDValue Op,
1399 // Copy all operands except the intrinsic ID.
1400 unsigned NumOps = Op.getNumOperands();
1401 SmallVector<SDValue, 6> Ops;
1402 Ops.reserve(NumOps - 1);
1403 Ops.push_back(Op.getOperand(0));
1404 for (unsigned I = 2; I < NumOps; ++I)
1405 Ops.push_back(Op.getOperand(I));
1407 assert(Op->getNumValues() == 2 && "Expected only CC result and chain");
1408 SDVTList RawVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1409 SDValue Intr = DAG.getNode(Opcode, SDLoc(Op), RawVTs, Ops);
1410 SDValue OldChain = SDValue(Op.getNode(), 1);
1411 SDValue NewChain = SDValue(Intr.getNode(), 0);
1412 DAG.ReplaceAllUsesOfValueWith(OldChain, NewChain);
1416 // Emit an intrinsic with a glued value instead of its CC result.
1417 static SDValue emitIntrinsicWithGlue(SelectionDAG &DAG, SDValue Op,
1419 // Copy all operands except the intrinsic ID.
1420 unsigned NumOps = Op.getNumOperands();
1421 SmallVector<SDValue, 6> Ops;
1422 Ops.reserve(NumOps - 1);
1423 for (unsigned I = 1; I < NumOps; ++I)
1424 Ops.push_back(Op.getOperand(I));
1426 if (Op->getNumValues() == 1)
1427 return DAG.getNode(Opcode, SDLoc(Op), MVT::Glue, Ops);
1428 assert(Op->getNumValues() == 2 && "Expected exactly one non-CC result");
1429 SDVTList RawVTs = DAG.getVTList(Op->getValueType(0), MVT::Glue);
1430 return DAG.getNode(Opcode, SDLoc(Op), RawVTs, Ops);
1433 // CC is a comparison that will be implemented using an integer or
1434 // floating-point comparison. Return the condition code mask for
1435 // a branch on true. In the integer case, CCMASK_CMP_UO is set for
1436 // unsigned comparisons and clear for signed ones. In the floating-point
1437 // case, CCMASK_CMP_UO has its normal mask meaning (unordered).
1438 static unsigned CCMaskForCondCode(ISD::CondCode CC) {
1440 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
1441 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
1442 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
1446 llvm_unreachable("Invalid integer condition!");
1455 case ISD::SETO: return SystemZ::CCMASK_CMP_O;
1456 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
1461 // Return a sequence for getting a 1 from an IPM result when CC has a
1462 // value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
1463 // The handling of CC values outside CCValid doesn't matter.
1464 static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
1465 // Deal with cases where the result can be taken directly from a bit
1466 // of the IPM result.
1467 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
1468 return IPMConversion(0, 0, SystemZ::IPM_CC);
1469 if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
1470 return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
1472 // Deal with cases where we can add a value to force the sign bit
1473 // to contain the right value. Putting the bit in 31 means we can
1474 // use SRL rather than RISBG(L), and also makes it easier to get a
1475 // 0/-1 value, so it has priority over the other tests below.
1477 // These sequences rely on the fact that the upper two bits of the
1478 // IPM result are zero.
1479 uint64_t TopBit = uint64_t(1) << 31;
1480 if (CCMask == (CCValid & SystemZ::CCMASK_0))
1481 return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
1482 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
1483 return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
1484 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1486 | SystemZ::CCMASK_2)))
1487 return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
1488 if (CCMask == (CCValid & SystemZ::CCMASK_3))
1489 return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
1490 if (CCMask == (CCValid & (SystemZ::CCMASK_1
1492 | SystemZ::CCMASK_3)))
1493 return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
1495 // Next try inverting the value and testing a bit. 0/1 could be
1496 // handled this way too, but we dealt with that case above.
1497 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
1498 return IPMConversion(-1, 0, SystemZ::IPM_CC);
1500 // Handle cases where adding a value forces a non-sign bit to contain
1502 if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
1503 return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
1504 if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
1505 return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
1507 // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
1508 // can be done by inverting the low CC bit and applying one of the
1509 // sign-based extractions above.
1510 if (CCMask == (CCValid & SystemZ::CCMASK_1))
1511 return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
1512 if (CCMask == (CCValid & SystemZ::CCMASK_2))
1513 return IPMConversion(1 << SystemZ::IPM_CC,
1514 TopBit - (3 << SystemZ::IPM_CC), 31);
1515 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1517 | SystemZ::CCMASK_3)))
1518 return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
1519 if (CCMask == (CCValid & (SystemZ::CCMASK_0
1521 | SystemZ::CCMASK_3)))
1522 return IPMConversion(1 << SystemZ::IPM_CC,
1523 TopBit - (1 << SystemZ::IPM_CC), 31);
1525 llvm_unreachable("Unexpected CC combination");
1528 // If C can be converted to a comparison against zero, adjust the operands
1530 static void adjustZeroCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1531 if (C.ICmpType == SystemZICMP::UnsignedOnly)
1534 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
1538 int64_t Value = ConstOp1->getSExtValue();
1539 if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
1540 (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
1541 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
1542 (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
1543 C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1544 C.Op1 = DAG.getConstant(0, DL, C.Op1.getValueType());
1548 // If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
1549 // adjust the operands as necessary.
1550 static void adjustSubwordCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1551 // For us to make any changes, it must a comparison between a single-use
1552 // load and a constant.
1553 if (!C.Op0.hasOneUse() ||
1554 C.Op0.getOpcode() != ISD::LOAD ||
1555 C.Op1.getOpcode() != ISD::Constant)
1558 // We must have an 8- or 16-bit load.
1559 auto *Load = cast<LoadSDNode>(C.Op0);
1560 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
1561 if (NumBits != 8 && NumBits != 16)
1564 // The load must be an extending one and the constant must be within the
1565 // range of the unextended value.
1566 auto *ConstOp1 = cast<ConstantSDNode>(C.Op1);
1567 uint64_t Value = ConstOp1->getZExtValue();
1568 uint64_t Mask = (1 << NumBits) - 1;
1569 if (Load->getExtensionType() == ISD::SEXTLOAD) {
1570 // Make sure that ConstOp1 is in range of C.Op0.
1571 int64_t SignedValue = ConstOp1->getSExtValue();
1572 if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
1574 if (C.ICmpType != SystemZICMP::SignedOnly) {
1575 // Unsigned comparison between two sign-extended values is equivalent
1576 // to unsigned comparison between two zero-extended values.
1578 } else if (NumBits == 8) {
1579 // Try to treat the comparison as unsigned, so that we can use CLI.
1580 // Adjust CCMask and Value as necessary.
1581 if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
1582 // Test whether the high bit of the byte is set.
1583 Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
1584 else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
1585 // Test whether the high bit of the byte is clear.
1586 Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
1588 // No instruction exists for this combination.
1590 C.ICmpType = SystemZICMP::UnsignedOnly;
1592 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
1595 assert(C.ICmpType == SystemZICMP::Any &&
1596 "Signedness shouldn't matter here.");
1600 // Make sure that the first operand is an i32 of the right extension type.
1601 ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
1604 if (C.Op0.getValueType() != MVT::i32 ||
1605 Load->getExtensionType() != ExtType)
1606 C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
1607 Load->getChain(), Load->getBasePtr(),
1608 Load->getPointerInfo(), Load->getMemoryVT(),
1609 Load->isVolatile(), Load->isNonTemporal(),
1610 Load->isInvariant(), Load->getAlignment());
1612 // Make sure that the second operand is an i32 with the right value.
1613 if (C.Op1.getValueType() != MVT::i32 ||
1614 Value != ConstOp1->getZExtValue())
1615 C.Op1 = DAG.getConstant(Value, DL, MVT::i32);
1618 // Return true if Op is either an unextended load, or a load suitable
1619 // for integer register-memory comparisons of type ICmpType.
1620 static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
1621 auto *Load = dyn_cast<LoadSDNode>(Op.getNode());
1623 // There are no instructions to compare a register with a memory byte.
1624 if (Load->getMemoryVT() == MVT::i8)
1626 // Otherwise decide on extension type.
1627 switch (Load->getExtensionType()) {
1628 case ISD::NON_EXTLOAD:
1631 return ICmpType != SystemZICMP::UnsignedOnly;
1633 return ICmpType != SystemZICMP::SignedOnly;
1641 // Return true if it is better to swap the operands of C.
1642 static bool shouldSwapCmpOperands(const Comparison &C) {
1643 // Leave f128 comparisons alone, since they have no memory forms.
1644 if (C.Op0.getValueType() == MVT::f128)
1647 // Always keep a floating-point constant second, since comparisons with
1648 // zero can use LOAD TEST and comparisons with other constants make a
1649 // natural memory operand.
1650 if (isa<ConstantFPSDNode>(C.Op1))
1653 // Never swap comparisons with zero since there are many ways to optimize
1655 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1656 if (ConstOp1 && ConstOp1->getZExtValue() == 0)
1659 // Also keep natural memory operands second if the loaded value is
1660 // only used here. Several comparisons have memory forms.
1661 if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
1664 // Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
1665 // In that case we generally prefer the memory to be second.
1666 if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
1667 // The only exceptions are when the second operand is a constant and
1668 // we can use things like CHHSI.
1671 // The unsigned memory-immediate instructions can handle 16-bit
1672 // unsigned integers.
1673 if (C.ICmpType != SystemZICMP::SignedOnly &&
1674 isUInt<16>(ConstOp1->getZExtValue()))
1676 // The signed memory-immediate instructions can handle 16-bit
1678 if (C.ICmpType != SystemZICMP::UnsignedOnly &&
1679 isInt<16>(ConstOp1->getSExtValue()))
1684 // Try to promote the use of CGFR and CLGFR.
1685 unsigned Opcode0 = C.Op0.getOpcode();
1686 if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
1688 if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
1690 if (C.ICmpType != SystemZICMP::SignedOnly &&
1691 Opcode0 == ISD::AND &&
1692 C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
1693 cast<ConstantSDNode>(C.Op0.getOperand(1))->getZExtValue() == 0xffffffff)
1699 // Return a version of comparison CC mask CCMask in which the LT and GT
1700 // actions are swapped.
1701 static unsigned reverseCCMask(unsigned CCMask) {
1702 return ((CCMask & SystemZ::CCMASK_CMP_EQ) |
1703 (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
1704 (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
1705 (CCMask & SystemZ::CCMASK_CMP_UO));
1708 // Check whether C tests for equality between X and Y and whether X - Y
1709 // or Y - X is also computed. In that case it's better to compare the
1710 // result of the subtraction against zero.
1711 static void adjustForSubtraction(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1712 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
1713 C.CCMask == SystemZ::CCMASK_CMP_NE) {
1714 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1716 if (N->getOpcode() == ISD::SUB &&
1717 ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
1718 (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
1719 C.Op0 = SDValue(N, 0);
1720 C.Op1 = DAG.getConstant(0, DL, N->getValueType(0));
1727 // Check whether C compares a floating-point value with zero and if that
1728 // floating-point value is also negated. In this case we can use the
1729 // negation to set CC, so avoiding separate LOAD AND TEST and
1730 // LOAD (NEGATIVE/COMPLEMENT) instructions.
1731 static void adjustForFNeg(Comparison &C) {
1732 auto *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
1733 if (C1 && C1->isZero()) {
1734 for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
1736 if (N->getOpcode() == ISD::FNEG) {
1737 C.Op0 = SDValue(N, 0);
1738 C.CCMask = reverseCCMask(C.CCMask);
1745 // Check whether C compares (shl X, 32) with 0 and whether X is
1746 // also sign-extended. In that case it is better to test the result
1747 // of the sign extension using LTGFR.
1749 // This case is important because InstCombine transforms a comparison
1750 // with (sext (trunc X)) into a comparison with (shl X, 32).
1751 static void adjustForLTGFR(Comparison &C) {
1752 // Check for a comparison between (shl X, 32) and 0.
1753 if (C.Op0.getOpcode() == ISD::SHL &&
1754 C.Op0.getValueType() == MVT::i64 &&
1755 C.Op1.getOpcode() == ISD::Constant &&
1756 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1757 auto *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
1758 if (C1 && C1->getZExtValue() == 32) {
1759 SDValue ShlOp0 = C.Op0.getOperand(0);
1760 // See whether X has any SIGN_EXTEND_INREG uses.
1761 for (auto I = ShlOp0->use_begin(), E = ShlOp0->use_end(); I != E; ++I) {
1763 if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
1764 cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
1765 C.Op0 = SDValue(N, 0);
1773 // If C compares the truncation of an extending load, try to compare
1774 // the untruncated value instead. This exposes more opportunities to
1776 static void adjustICmpTruncate(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1777 if (C.Op0.getOpcode() == ISD::TRUNCATE &&
1778 C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
1779 C.Op1.getOpcode() == ISD::Constant &&
1780 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
1781 auto *L = cast<LoadSDNode>(C.Op0.getOperand(0));
1782 if (L->getMemoryVT().getStoreSizeInBits()
1783 <= C.Op0.getValueType().getSizeInBits()) {
1784 unsigned Type = L->getExtensionType();
1785 if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
1786 (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
1787 C.Op0 = C.Op0.getOperand(0);
1788 C.Op1 = DAG.getConstant(0, DL, C.Op0.getValueType());
1794 // Return true if shift operation N has an in-range constant shift value.
1795 // Store it in ShiftVal if so.
1796 static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
1797 auto *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
1801 uint64_t Amount = Shift->getZExtValue();
1802 if (Amount >= N.getValueType().getSizeInBits())
1809 // Check whether an AND with Mask is suitable for a TEST UNDER MASK
1810 // instruction and whether the CC value is descriptive enough to handle
1811 // a comparison of type Opcode between the AND result and CmpVal.
1812 // CCMask says which comparison result is being tested and BitSize is
1813 // the number of bits in the operands. If TEST UNDER MASK can be used,
1814 // return the corresponding CC mask, otherwise return 0.
1815 static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask,
1816 uint64_t Mask, uint64_t CmpVal,
1817 unsigned ICmpType) {
1818 assert(Mask != 0 && "ANDs with zero should have been removed by now");
1820 // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.
1821 if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&
1822 !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))
1825 // Work out the masks for the lowest and highest bits.
1826 unsigned HighShift = 63 - countLeadingZeros(Mask);
1827 uint64_t High = uint64_t(1) << HighShift;
1828 uint64_t Low = uint64_t(1) << countTrailingZeros(Mask);
1830 // Signed ordered comparisons are effectively unsigned if the sign
1832 bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly);
1834 // Check for equality comparisons with 0, or the equivalent.
1836 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1837 return SystemZ::CCMASK_TM_ALL_0;
1838 if (CCMask == SystemZ::CCMASK_CMP_NE)
1839 return SystemZ::CCMASK_TM_SOME_1;
1841 if (EffectivelyUnsigned && CmpVal <= Low) {
1842 if (CCMask == SystemZ::CCMASK_CMP_LT)
1843 return SystemZ::CCMASK_TM_ALL_0;
1844 if (CCMask == SystemZ::CCMASK_CMP_GE)
1845 return SystemZ::CCMASK_TM_SOME_1;
1847 if (EffectivelyUnsigned && CmpVal < Low) {
1848 if (CCMask == SystemZ::CCMASK_CMP_LE)
1849 return SystemZ::CCMASK_TM_ALL_0;
1850 if (CCMask == SystemZ::CCMASK_CMP_GT)
1851 return SystemZ::CCMASK_TM_SOME_1;
1854 // Check for equality comparisons with the mask, or the equivalent.
1855 if (CmpVal == Mask) {
1856 if (CCMask == SystemZ::CCMASK_CMP_EQ)
1857 return SystemZ::CCMASK_TM_ALL_1;
1858 if (CCMask == SystemZ::CCMASK_CMP_NE)
1859 return SystemZ::CCMASK_TM_SOME_0;
1861 if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) {
1862 if (CCMask == SystemZ::CCMASK_CMP_GT)
1863 return SystemZ::CCMASK_TM_ALL_1;
1864 if (CCMask == SystemZ::CCMASK_CMP_LE)
1865 return SystemZ::CCMASK_TM_SOME_0;
1867 if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) {
1868 if (CCMask == SystemZ::CCMASK_CMP_GE)
1869 return SystemZ::CCMASK_TM_ALL_1;
1870 if (CCMask == SystemZ::CCMASK_CMP_LT)
1871 return SystemZ::CCMASK_TM_SOME_0;
1874 // Check for ordered comparisons with the top bit.
1875 if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) {
1876 if (CCMask == SystemZ::CCMASK_CMP_LE)
1877 return SystemZ::CCMASK_TM_MSB_0;
1878 if (CCMask == SystemZ::CCMASK_CMP_GT)
1879 return SystemZ::CCMASK_TM_MSB_1;
1881 if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) {
1882 if (CCMask == SystemZ::CCMASK_CMP_LT)
1883 return SystemZ::CCMASK_TM_MSB_0;
1884 if (CCMask == SystemZ::CCMASK_CMP_GE)
1885 return SystemZ::CCMASK_TM_MSB_1;
1888 // If there are just two bits, we can do equality checks for Low and High
1890 if (Mask == Low + High) {
1891 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low)
1892 return SystemZ::CCMASK_TM_MIXED_MSB_0;
1893 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low)
1894 return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY;
1895 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High)
1896 return SystemZ::CCMASK_TM_MIXED_MSB_1;
1897 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High)
1898 return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY;
1901 // Looks like we've exhausted our options.
1905 // See whether C can be implemented as a TEST UNDER MASK instruction.
1906 // Update the arguments with the TM version if so.
1907 static void adjustForTestUnderMask(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
1908 // Check that we have a comparison with a constant.
1909 auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
1912 uint64_t CmpVal = ConstOp1->getZExtValue();
1914 // Check whether the nonconstant input is an AND with a constant mask.
1917 ConstantSDNode *Mask = nullptr;
1918 if (C.Op0.getOpcode() == ISD::AND) {
1919 NewC.Op0 = C.Op0.getOperand(0);
1920 NewC.Op1 = C.Op0.getOperand(1);
1921 Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
1924 MaskVal = Mask->getZExtValue();
1926 // There is no instruction to compare with a 64-bit immediate
1927 // so use TMHH instead if possible. We need an unsigned ordered
1928 // comparison with an i64 immediate.
1929 if (NewC.Op0.getValueType() != MVT::i64 ||
1930 NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
1931 NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
1932 NewC.ICmpType == SystemZICMP::SignedOnly)
1934 // Convert LE and GT comparisons into LT and GE.
1935 if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
1936 NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
1937 if (CmpVal == uint64_t(-1))
1940 NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
1942 // If the low N bits of Op1 are zero than the low N bits of Op0 can
1943 // be masked off without changing the result.
1944 MaskVal = -(CmpVal & -CmpVal);
1945 NewC.ICmpType = SystemZICMP::UnsignedOnly;
1950 // Check whether the combination of mask, comparison value and comparison
1951 // type are suitable.
1952 unsigned BitSize = NewC.Op0.getValueType().getSizeInBits();
1953 unsigned NewCCMask, ShiftVal;
1954 if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1955 NewC.Op0.getOpcode() == ISD::SHL &&
1956 isSimpleShift(NewC.Op0, ShiftVal) &&
1957 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1958 MaskVal >> ShiftVal,
1960 SystemZICMP::Any))) {
1961 NewC.Op0 = NewC.Op0.getOperand(0);
1962 MaskVal >>= ShiftVal;
1963 } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
1964 NewC.Op0.getOpcode() == ISD::SRL &&
1965 isSimpleShift(NewC.Op0, ShiftVal) &&
1966 (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
1967 MaskVal << ShiftVal,
1969 SystemZICMP::UnsignedOnly))) {
1970 NewC.Op0 = NewC.Op0.getOperand(0);
1971 MaskVal <<= ShiftVal;
1973 NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
1979 // Go ahead and make the change.
1980 C.Opcode = SystemZISD::TM;
1982 if (Mask && Mask->getZExtValue() == MaskVal)
1983 C.Op1 = SDValue(Mask, 0);
1985 C.Op1 = DAG.getConstant(MaskVal, DL, C.Op0.getValueType());
1986 C.CCValid = SystemZ::CCMASK_TM;
1987 C.CCMask = NewCCMask;
1990 // Return a Comparison that tests the condition-code result of intrinsic
1991 // node Call against constant integer CC using comparison code Cond.
1992 // Opcode is the opcode of the SystemZISD operation for the intrinsic
1993 // and CCValid is the set of possible condition-code results.
1994 static Comparison getIntrinsicCmp(SelectionDAG &DAG, unsigned Opcode,
1995 SDValue Call, unsigned CCValid, uint64_t CC,
1996 ISD::CondCode Cond) {
1997 Comparison C(Call, SDValue());
1999 C.CCValid = CCValid;
2000 if (Cond == ISD::SETEQ)
2001 // bit 3 for CC==0, bit 0 for CC==3, always false for CC>3.
2002 C.CCMask = CC < 4 ? 1 << (3 - CC) : 0;
2003 else if (Cond == ISD::SETNE)
2004 // ...and the inverse of that.
2005 C.CCMask = CC < 4 ? ~(1 << (3 - CC)) : -1;
2006 else if (Cond == ISD::SETLT || Cond == ISD::SETULT)
2007 // bits above bit 3 for CC==0 (always false), bits above bit 0 for CC==3,
2008 // always true for CC>3.
2009 C.CCMask = CC < 4 ? ~0U << (4 - CC) : -1;
2010 else if (Cond == ISD::SETGE || Cond == ISD::SETUGE)
2011 // ...and the inverse of that.
2012 C.CCMask = CC < 4 ? ~(~0U << (4 - CC)) : 0;
2013 else if (Cond == ISD::SETLE || Cond == ISD::SETULE)
2014 // bit 3 and above for CC==0, bit 0 and above for CC==3 (always true),
2015 // always true for CC>3.
2016 C.CCMask = CC < 4 ? ~0U << (3 - CC) : -1;
2017 else if (Cond == ISD::SETGT || Cond == ISD::SETUGT)
2018 // ...and the inverse of that.
2019 C.CCMask = CC < 4 ? ~(~0U << (3 - CC)) : 0;
2021 llvm_unreachable("Unexpected integer comparison type");
2022 C.CCMask &= CCValid;
2026 // Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
2027 static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
2028 ISD::CondCode Cond, SDLoc DL) {
2029 if (CmpOp1.getOpcode() == ISD::Constant) {
2030 uint64_t Constant = cast<ConstantSDNode>(CmpOp1)->getZExtValue();
2031 unsigned Opcode, CCValid;
2032 if (CmpOp0.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
2033 CmpOp0.getResNo() == 0 && CmpOp0->hasNUsesOfValue(1, 0) &&
2034 isIntrinsicWithCCAndChain(CmpOp0, Opcode, CCValid))
2035 return getIntrinsicCmp(DAG, Opcode, CmpOp0, CCValid, Constant, Cond);
2036 if (CmpOp0.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
2037 CmpOp0.getResNo() == CmpOp0->getNumValues() - 1 &&
2038 isIntrinsicWithCC(CmpOp0, Opcode, CCValid))
2039 return getIntrinsicCmp(DAG, Opcode, CmpOp0, CCValid, Constant, Cond);
2041 Comparison C(CmpOp0, CmpOp1);
2042 C.CCMask = CCMaskForCondCode(Cond);
2043 if (C.Op0.getValueType().isFloatingPoint()) {
2044 C.CCValid = SystemZ::CCMASK_FCMP;
2045 C.Opcode = SystemZISD::FCMP;
2048 C.CCValid = SystemZ::CCMASK_ICMP;
2049 C.Opcode = SystemZISD::ICMP;
2050 // Choose the type of comparison. Equality and inequality tests can
2051 // use either signed or unsigned comparisons. The choice also doesn't
2052 // matter if both sign bits are known to be clear. In those cases we
2053 // want to give the main isel code the freedom to choose whichever
2055 if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
2056 C.CCMask == SystemZ::CCMASK_CMP_NE ||
2057 (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
2058 C.ICmpType = SystemZICMP::Any;
2059 else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
2060 C.ICmpType = SystemZICMP::UnsignedOnly;
2062 C.ICmpType = SystemZICMP::SignedOnly;
2063 C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
2064 adjustZeroCmp(DAG, DL, C);
2065 adjustSubwordCmp(DAG, DL, C);
2066 adjustForSubtraction(DAG, DL, C);
2068 adjustICmpTruncate(DAG, DL, C);
2071 if (shouldSwapCmpOperands(C)) {
2072 std::swap(C.Op0, C.Op1);
2073 C.CCMask = reverseCCMask(C.CCMask);
2076 adjustForTestUnderMask(DAG, DL, C);
2080 // Emit the comparison instruction described by C.
2081 static SDValue emitCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
2082 if (!C.Op1.getNode()) {
2084 switch (C.Op0.getOpcode()) {
2085 case ISD::INTRINSIC_W_CHAIN:
2086 Op = emitIntrinsicWithChainAndGlue(DAG, C.Op0, C.Opcode);
2088 case ISD::INTRINSIC_WO_CHAIN:
2089 Op = emitIntrinsicWithGlue(DAG, C.Op0, C.Opcode);
2092 llvm_unreachable("Invalid comparison operands");
2094 return SDValue(Op.getNode(), Op->getNumValues() - 1);
2096 if (C.Opcode == SystemZISD::ICMP)
2097 return DAG.getNode(SystemZISD::ICMP, DL, MVT::Glue, C.Op0, C.Op1,
2098 DAG.getConstant(C.ICmpType, DL, MVT::i32));
2099 if (C.Opcode == SystemZISD::TM) {
2100 bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
2101 bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
2102 return DAG.getNode(SystemZISD::TM, DL, MVT::Glue, C.Op0, C.Op1,
2103 DAG.getConstant(RegisterOnly, DL, MVT::i32));
2105 return DAG.getNode(C.Opcode, DL, MVT::Glue, C.Op0, C.Op1);
2108 // Implement a 32-bit *MUL_LOHI operation by extending both operands to
2109 // 64 bits. Extend is the extension type to use. Store the high part
2110 // in Hi and the low part in Lo.
2111 static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL,
2112 unsigned Extend, SDValue Op0, SDValue Op1,
2113 SDValue &Hi, SDValue &Lo) {
2114 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
2115 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
2116 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
2117 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
2118 DAG.getConstant(32, DL, MVT::i64));
2119 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
2120 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
2123 // Lower a binary operation that produces two VT results, one in each
2124 // half of a GR128 pair. Op0 and Op1 are the VT operands to the operation,
2125 // Extend extends Op0 to a GR128, and Opcode performs the GR128 operation
2126 // on the extended Op0 and (unextended) Op1. Store the even register result
2127 // in Even and the odd register result in Odd.
2128 static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
2129 unsigned Extend, unsigned Opcode,
2130 SDValue Op0, SDValue Op1,
2131 SDValue &Even, SDValue &Odd) {
2132 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0);
2133 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped,
2134 SDValue(In128, 0), Op1);
2135 bool Is32Bit = is32Bit(VT);
2136 Even = DAG.getTargetExtractSubreg(SystemZ::even128(Is32Bit), DL, VT, Result);
2137 Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
2140 // Return an i32 value that is 1 if the CC value produced by Glue is
2141 // in the mask CCMask and 0 otherwise. CC is known to have a value
2142 // in CCValid, so other values can be ignored.
2143 static SDValue emitSETCC(SelectionDAG &DAG, SDLoc DL, SDValue Glue,
2144 unsigned CCValid, unsigned CCMask) {
2145 IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
2146 SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
2148 if (Conversion.XORValue)
2149 Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
2150 DAG.getConstant(Conversion.XORValue, DL, MVT::i32));
2152 if (Conversion.AddValue)
2153 Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
2154 DAG.getConstant(Conversion.AddValue, DL, MVT::i32));
2156 // The SHR/AND sequence should get optimized to an RISBG.
2157 Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
2158 DAG.getConstant(Conversion.Bit, DL, MVT::i32));
2159 if (Conversion.Bit != 31)
2160 Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
2161 DAG.getConstant(1, DL, MVT::i32));
2165 // Return the SystemISD vector comparison operation for CC, or 0 if it cannot
2166 // be done directly. IsFP is true if CC is for a floating-point rather than
2167 // integer comparison.
2168 static unsigned getVectorComparison(ISD::CondCode CC, bool IsFP) {
2172 return IsFP ? SystemZISD::VFCMPE : SystemZISD::VICMPE;
2176 return IsFP ? SystemZISD::VFCMPHE : static_cast<SystemZISD::NodeType>(0);
2180 return IsFP ? SystemZISD::VFCMPH : SystemZISD::VICMPH;
2183 return IsFP ? static_cast<SystemZISD::NodeType>(0) : SystemZISD::VICMPHL;
2190 // Return the SystemZISD vector comparison operation for CC or its inverse,
2191 // or 0 if neither can be done directly. Indicate in Invert whether the
2192 // result is for the inverse of CC. IsFP is true if CC is for a
2193 // floating-point rather than integer comparison.
2194 static unsigned getVectorComparisonOrInvert(ISD::CondCode CC, bool IsFP,
2196 if (unsigned Opcode = getVectorComparison(CC, IsFP)) {
2201 CC = ISD::getSetCCInverse(CC, !IsFP);
2202 if (unsigned Opcode = getVectorComparison(CC, IsFP)) {
2210 // Return a v2f64 that contains the extended form of elements Start and Start+1
2211 // of v4f32 value Op.
2212 static SDValue expandV4F32ToV2F64(SelectionDAG &DAG, int Start, SDLoc DL,
2214 int Mask[] = { Start, -1, Start + 1, -1 };
2215 Op = DAG.getVectorShuffle(MVT::v4f32, DL, Op, DAG.getUNDEF(MVT::v4f32), Mask);
2216 return DAG.getNode(SystemZISD::VEXTEND, DL, MVT::v2f64, Op);
2219 // Build a comparison of vectors CmpOp0 and CmpOp1 using opcode Opcode,
2220 // producing a result of type VT.
2221 static SDValue getVectorCmp(SelectionDAG &DAG, unsigned Opcode, SDLoc DL,
2222 EVT VT, SDValue CmpOp0, SDValue CmpOp1) {
2223 // There is no hardware support for v4f32, so extend the vector into
2224 // two v2f64s and compare those.
2225 if (CmpOp0.getValueType() == MVT::v4f32) {
2226 SDValue H0 = expandV4F32ToV2F64(DAG, 0, DL, CmpOp0);
2227 SDValue L0 = expandV4F32ToV2F64(DAG, 2, DL, CmpOp0);
2228 SDValue H1 = expandV4F32ToV2F64(DAG, 0, DL, CmpOp1);
2229 SDValue L1 = expandV4F32ToV2F64(DAG, 2, DL, CmpOp1);
2230 SDValue HRes = DAG.getNode(Opcode, DL, MVT::v2i64, H0, H1);
2231 SDValue LRes = DAG.getNode(Opcode, DL, MVT::v2i64, L0, L1);
2232 return DAG.getNode(SystemZISD::PACK, DL, VT, HRes, LRes);
2234 return DAG.getNode(Opcode, DL, VT, CmpOp0, CmpOp1);
2237 // Lower a vector comparison of type CC between CmpOp0 and CmpOp1, producing
2238 // an integer mask of type VT.
2239 static SDValue lowerVectorSETCC(SelectionDAG &DAG, SDLoc DL, EVT VT,
2240 ISD::CondCode CC, SDValue CmpOp0,
2242 bool IsFP = CmpOp0.getValueType().isFloatingPoint();
2243 bool Invert = false;
2246 // Handle tests for order using (or (ogt y x) (oge x y)).
2250 assert(IsFP && "Unexpected integer comparison");
2251 SDValue LT = getVectorCmp(DAG, SystemZISD::VFCMPH, DL, VT, CmpOp1, CmpOp0);
2252 SDValue GE = getVectorCmp(DAG, SystemZISD::VFCMPHE, DL, VT, CmpOp0, CmpOp1);
2253 Cmp = DAG.getNode(ISD::OR, DL, VT, LT, GE);
2257 // Handle <> tests using (or (ogt y x) (ogt x y)).
2261 assert(IsFP && "Unexpected integer comparison");
2262 SDValue LT = getVectorCmp(DAG, SystemZISD::VFCMPH, DL, VT, CmpOp1, CmpOp0);
2263 SDValue GT = getVectorCmp(DAG, SystemZISD::VFCMPH, DL, VT, CmpOp0, CmpOp1);
2264 Cmp = DAG.getNode(ISD::OR, DL, VT, LT, GT);
2268 // Otherwise a single comparison is enough. It doesn't really
2269 // matter whether we try the inversion or the swap first, since
2270 // there are no cases where both work.
2272 if (unsigned Opcode = getVectorComparisonOrInvert(CC, IsFP, Invert))
2273 Cmp = getVectorCmp(DAG, Opcode, DL, VT, CmpOp0, CmpOp1);
2275 CC = ISD::getSetCCSwappedOperands(CC);
2276 if (unsigned Opcode = getVectorComparisonOrInvert(CC, IsFP, Invert))
2277 Cmp = getVectorCmp(DAG, Opcode, DL, VT, CmpOp1, CmpOp0);
2279 llvm_unreachable("Unhandled comparison");
2284 SDValue Mask = DAG.getNode(SystemZISD::BYTE_MASK, DL, MVT::v16i8,
2285 DAG.getConstant(65535, DL, MVT::i32));
2286 Mask = DAG.getNode(ISD::BITCAST, DL, VT, Mask);
2287 Cmp = DAG.getNode(ISD::XOR, DL, VT, Cmp, Mask);
2292 SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
2293 SelectionDAG &DAG) const {
2294 SDValue CmpOp0 = Op.getOperand(0);
2295 SDValue CmpOp1 = Op.getOperand(1);
2296 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2298 EVT VT = Op.getValueType();
2300 return lowerVectorSETCC(DAG, DL, VT, CC, CmpOp0, CmpOp1);
2302 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL));
2303 SDValue Glue = emitCmp(DAG, DL, C);
2304 return emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
2307 SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2308 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2309 SDValue CmpOp0 = Op.getOperand(2);
2310 SDValue CmpOp1 = Op.getOperand(3);
2311 SDValue Dest = Op.getOperand(4);
2314 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL));
2315 SDValue Glue = emitCmp(DAG, DL, C);
2316 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
2317 Op.getOperand(0), DAG.getConstant(C.CCValid, DL, MVT::i32),
2318 DAG.getConstant(C.CCMask, DL, MVT::i32), Dest, Glue);
2321 // Return true if Pos is CmpOp and Neg is the negative of CmpOp,
2322 // allowing Pos and Neg to be wider than CmpOp.
2323 static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
2324 return (Neg.getOpcode() == ISD::SUB &&
2325 Neg.getOperand(0).getOpcode() == ISD::Constant &&
2326 cast<ConstantSDNode>(Neg.getOperand(0))->getZExtValue() == 0 &&
2327 Neg.getOperand(1) == Pos &&
2329 (Pos.getOpcode() == ISD::SIGN_EXTEND &&
2330 Pos.getOperand(0) == CmpOp)));
2333 // Return the absolute or negative absolute of Op; IsNegative decides which.
2334 static SDValue getAbsolute(SelectionDAG &DAG, SDLoc DL, SDValue Op,
2336 Op = DAG.getNode(SystemZISD::IABS, DL, Op.getValueType(), Op);
2338 Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
2339 DAG.getConstant(0, DL, Op.getValueType()), Op);
2343 SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
2344 SelectionDAG &DAG) const {
2345 SDValue CmpOp0 = Op.getOperand(0);
2346 SDValue CmpOp1 = Op.getOperand(1);
2347 SDValue TrueOp = Op.getOperand(2);
2348 SDValue FalseOp = Op.getOperand(3);
2349 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2352 Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL));
2354 // Check for absolute and negative-absolute selections, including those
2355 // where the comparison value is sign-extended (for LPGFR and LNGFR).
2356 // This check supplements the one in DAGCombiner.
2357 if (C.Opcode == SystemZISD::ICMP &&
2358 C.CCMask != SystemZ::CCMASK_CMP_EQ &&
2359 C.CCMask != SystemZ::CCMASK_CMP_NE &&
2360 C.Op1.getOpcode() == ISD::Constant &&
2361 cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
2362 if (isAbsolute(C.Op0, TrueOp, FalseOp))
2363 return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
2364 if (isAbsolute(C.Op0, FalseOp, TrueOp))
2365 return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
2368 SDValue Glue = emitCmp(DAG, DL, C);
2370 // Special case for handling -1/0 results. The shifts we use here
2371 // should get optimized with the IPM conversion sequence.
2372 auto *TrueC = dyn_cast<ConstantSDNode>(TrueOp);
2373 auto *FalseC = dyn_cast<ConstantSDNode>(FalseOp);
2374 if (TrueC && FalseC) {
2375 int64_t TrueVal = TrueC->getSExtValue();
2376 int64_t FalseVal = FalseC->getSExtValue();
2377 if ((TrueVal == -1 && FalseVal == 0) || (TrueVal == 0 && FalseVal == -1)) {
2378 // Invert the condition if we want -1 on false.
2380 C.CCMask ^= C.CCValid;
2381 SDValue Result = emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
2382 EVT VT = Op.getValueType();
2383 // Extend the result to VT. Upper bits are ignored.
2385 Result = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Result);
2386 // Sign-extend from the low bit.
2387 SDValue ShAmt = DAG.getConstant(VT.getSizeInBits() - 1, DL, MVT::i32);
2388 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Result, ShAmt);
2389 return DAG.getNode(ISD::SRA, DL, VT, Shl, ShAmt);
2393 SDValue Ops[] = {TrueOp, FalseOp, DAG.getConstant(C.CCValid, DL, MVT::i32),
2394 DAG.getConstant(C.CCMask, DL, MVT::i32), Glue};
2396 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
2397 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, Ops);
2400 SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
2401 SelectionDAG &DAG) const {
2403 const GlobalValue *GV = Node->getGlobal();
2404 int64_t Offset = Node->getOffset();
2405 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2406 Reloc::Model RM = DAG.getTarget().getRelocationModel();
2407 CodeModel::Model CM = DAG.getTarget().getCodeModel();
2410 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
2411 // Assign anchors at 1<<12 byte boundaries.
2412 uint64_t Anchor = Offset & ~uint64_t(0xfff);
2413 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor);
2414 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
2416 // The offset can be folded into the address if it is aligned to a halfword.
2418 if (Offset != 0 && (Offset & 1) == 0) {
2419 SDValue Full = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor + Offset);
2420 Result = DAG.getNode(SystemZISD::PCREL_OFFSET, DL, PtrVT, Full, Result);
2424 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
2425 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
2426 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2427 MachinePointerInfo::getGOT(), false, false, false, 0);
2430 // If there was a non-zero offset that we didn't fold, create an explicit
2433 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
2434 DAG.getConstant(Offset, DL, PtrVT));
2439 SDValue SystemZTargetLowering::lowerTLSGetOffset(GlobalAddressSDNode *Node,
2442 SDValue GOTOffset) const {
2444 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2445 SDValue Chain = DAG.getEntryNode();
2448 // __tls_get_offset takes the GOT offset in %r2 and the GOT in %r12.
2449 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2450 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R12D, GOT, Glue);
2451 Glue = Chain.getValue(1);
2452 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R2D, GOTOffset, Glue);
2453 Glue = Chain.getValue(1);
2455 // The first call operand is the chain and the second is the TLS symbol.
2456 SmallVector<SDValue, 8> Ops;
2457 Ops.push_back(Chain);
2458 Ops.push_back(DAG.getTargetGlobalAddress(Node->getGlobal(), DL,
2459 Node->getValueType(0),
2462 // Add argument registers to the end of the list so that they are
2463 // known live into the call.
2464 Ops.push_back(DAG.getRegister(SystemZ::R2D, PtrVT));
2465 Ops.push_back(DAG.getRegister(SystemZ::R12D, PtrVT));
2467 // Add a register mask operand representing the call-preserved registers.
2468 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
2469 const uint32_t *Mask =
2470 TRI->getCallPreservedMask(DAG.getMachineFunction(), CallingConv::C);
2471 assert(Mask && "Missing call preserved mask for calling convention");
2472 Ops.push_back(DAG.getRegisterMask(Mask));
2474 // Glue the call to the argument copies.
2475 Ops.push_back(Glue);
2478 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2479 Chain = DAG.getNode(Opcode, DL, NodeTys, Ops);
2480 Glue = Chain.getValue(1);
2482 // Copy the return value from %r2.
2483 return DAG.getCopyFromReg(Chain, DL, SystemZ::R2D, PtrVT, Glue);
2486 SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
2487 SelectionDAG &DAG) const {
2488 if (DAG.getTarget().Options.EmulatedTLS)
2489 return LowerToTLSEmulatedModel(Node, DAG);
2491 const GlobalValue *GV = Node->getGlobal();
2492 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2493 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
2495 // The high part of the thread pointer is in access register 0.
2496 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
2497 DAG.getConstant(0, DL, MVT::i32));
2498 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
2500 // The low part of the thread pointer is in access register 1.
2501 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
2502 DAG.getConstant(1, DL, MVT::i32));
2503 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
2505 // Merge them into a single 64-bit address.
2506 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
2507 DAG.getConstant(32, DL, PtrVT));
2508 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
2510 // Get the offset of GA from the thread pointer, based on the TLS model.
2513 case TLSModel::GeneralDynamic: {
2514 // Load the GOT offset of the tls_index (module ID / per-symbol offset).
2515 SystemZConstantPoolValue *CPV =
2516 SystemZConstantPoolValue::Create(GV, SystemZCP::TLSGD);
2518 Offset = DAG.getConstantPool(CPV, PtrVT, 8);
2519 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
2520 Offset, MachinePointerInfo::getConstantPool(),
2521 false, false, false, 0);
2523 // Call __tls_get_offset to retrieve the offset.
2524 Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_GDCALL, Offset);
2528 case TLSModel::LocalDynamic: {
2529 // Load the GOT offset of the module ID.
2530 SystemZConstantPoolValue *CPV =
2531 SystemZConstantPoolValue::Create(GV, SystemZCP::TLSLDM);
2533 Offset = DAG.getConstantPool(CPV, PtrVT, 8);
2534 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
2535 Offset, MachinePointerInfo::getConstantPool(),
2536 false, false, false, 0);
2538 // Call __tls_get_offset to retrieve the module base offset.
2539 Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_LDCALL, Offset);
2541 // Note: The SystemZLDCleanupPass will remove redundant computations
2542 // of the module base offset. Count total number of local-dynamic
2543 // accesses to trigger execution of that pass.
2544 SystemZMachineFunctionInfo* MFI =
2545 DAG.getMachineFunction().getInfo<SystemZMachineFunctionInfo>();
2546 MFI->incNumLocalDynamicTLSAccesses();
2548 // Add the per-symbol offset.
2549 CPV = SystemZConstantPoolValue::Create(GV, SystemZCP::DTPOFF);
2551 SDValue DTPOffset = DAG.getConstantPool(CPV, PtrVT, 8);
2552 DTPOffset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
2553 DTPOffset, MachinePointerInfo::getConstantPool(),
2554 false, false, false, 0);
2556 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Offset, DTPOffset);
2560 case TLSModel::InitialExec: {
2561 // Load the offset from the GOT.
2562 Offset = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2563 SystemZII::MO_INDNTPOFF);
2564 Offset = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Offset);
2565 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
2566 Offset, MachinePointerInfo::getGOT(),
2567 false, false, false, 0);
2571 case TLSModel::LocalExec: {
2572 // Force the offset into the constant pool and load it from there.
2573 SystemZConstantPoolValue *CPV =
2574 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
2576 Offset = DAG.getConstantPool(CPV, PtrVT, 8);
2577 Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
2578 Offset, MachinePointerInfo::getConstantPool(),
2579 false, false, false, 0);
2584 // Add the base and offset together.
2585 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
2588 SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
2589 SelectionDAG &DAG) const {
2591 const BlockAddress *BA = Node->getBlockAddress();
2592 int64_t Offset = Node->getOffset();
2593 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2595 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
2596 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
2600 SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
2601 SelectionDAG &DAG) const {
2603 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2604 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
2606 // Use LARL to load the address of the table.
2607 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
2610 SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
2611 SelectionDAG &DAG) const {
2613 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2616 if (CP->isMachineConstantPoolEntry())
2617 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2618 CP->getAlignment());
2620 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2621 CP->getAlignment(), CP->getOffset());
2623 // Use LARL to load the address of the constant pool entry.
2624 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
2627 SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
2628 SelectionDAG &DAG) const {
2630 SDValue In = Op.getOperand(0);
2631 EVT InVT = In.getValueType();
2632 EVT ResVT = Op.getValueType();
2634 // Convert loads directly. This is normally done by DAGCombiner,
2635 // but we need this case for bitcasts that are created during lowering
2636 // and which are then lowered themselves.
2637 if (auto *LoadN = dyn_cast<LoadSDNode>(In))
2638 return DAG.getLoad(ResVT, DL, LoadN->getChain(), LoadN->getBasePtr(),
2639 LoadN->getMemOperand());
2641 if (InVT == MVT::i32 && ResVT == MVT::f32) {
2643 if (Subtarget.hasHighWord()) {
2644 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL,
2646 In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
2647 MVT::i64, SDValue(U64, 0), In);
2649 In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
2650 In64 = DAG.getNode(ISD::SHL, DL, MVT::i64, In64,
2651 DAG.getConstant(32, DL, MVT::i64));
2653 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, In64);
2654 return DAG.getTargetExtractSubreg(SystemZ::subreg_r32,
2655 DL, MVT::f32, Out64);
2657 if (InVT == MVT::f32 && ResVT == MVT::i32) {
2658 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
2659 SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_r32, DL,
2660 MVT::f64, SDValue(U64, 0), In);
2661 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
2662 if (Subtarget.hasHighWord())
2663 return DAG.getTargetExtractSubreg(SystemZ::subreg_h32, DL,
2665 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64,
2666 DAG.getConstant(32, DL, MVT::i64));
2667 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
2669 llvm_unreachable("Unexpected bitcast combination");
2672 SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
2673 SelectionDAG &DAG) const {
2674 MachineFunction &MF = DAG.getMachineFunction();
2675 SystemZMachineFunctionInfo *FuncInfo =
2676 MF.getInfo<SystemZMachineFunctionInfo>();
2677 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2679 SDValue Chain = Op.getOperand(0);
2680 SDValue Addr = Op.getOperand(1);
2681 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2684 // The initial values of each field.
2685 const unsigned NumFields = 4;
2686 SDValue Fields[NumFields] = {
2687 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), DL, PtrVT),
2688 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), DL, PtrVT),
2689 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
2690 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
2693 // Store each field into its respective slot.
2694 SDValue MemOps[NumFields];
2695 unsigned Offset = 0;
2696 for (unsigned I = 0; I < NumFields; ++I) {
2697 SDValue FieldAddr = Addr;
2699 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
2700 DAG.getIntPtrConstant(Offset, DL));
2701 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
2702 MachinePointerInfo(SV, Offset),
2706 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2709 SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
2710 SelectionDAG &DAG) const {
2711 SDValue Chain = Op.getOperand(0);
2712 SDValue DstPtr = Op.getOperand(1);
2713 SDValue SrcPtr = Op.getOperand(2);
2714 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
2715 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
2718 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32, DL),
2719 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false,
2720 /*isTailCall*/false,
2721 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
2724 SDValue SystemZTargetLowering::
2725 lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
2726 SDValue Chain = Op.getOperand(0);
2727 SDValue Size = Op.getOperand(1);
2730 unsigned SPReg = getStackPointerRegisterToSaveRestore();
2732 // Get a reference to the stack pointer.
2733 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
2735 // Get the new stack pointer value.
2736 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size);
2738 // Copy the new stack pointer back.
2739 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
2741 // The allocated data lives above the 160 bytes allocated for the standard
2742 // frame, plus any outgoing stack arguments. We don't know how much that
2743 // amounts to yet, so emit a special ADJDYNALLOC placeholder.
2744 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
2745 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
2747 SDValue Ops[2] = { Result, Chain };
2748 return DAG.getMergeValues(Ops, DL);
2751 SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
2752 SelectionDAG &DAG) const {
2753 EVT VT = Op.getValueType();
2757 // Just do a normal 64-bit multiplication and extract the results.
2758 // We define this so that it can be used for constant division.
2759 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
2760 Op.getOperand(1), Ops[1], Ops[0]);
2762 // Do a full 128-bit multiplication based on UMUL_LOHI64:
2764 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
2766 // but using the fact that the upper halves are either all zeros
2769 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
2771 // and grouping the right terms together since they are quicker than the
2774 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
2775 SDValue C63 = DAG.getConstant(63, DL, MVT::i64);
2776 SDValue LL = Op.getOperand(0);
2777 SDValue RL = Op.getOperand(1);
2778 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
2779 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
2780 // UMUL_LOHI64 returns the low result in the odd register and the high
2781 // result in the even register. SMUL_LOHI is defined to return the
2782 // low half first, so the results are in reverse order.
2783 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2784 LL, RL, Ops[1], Ops[0]);
2785 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
2786 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
2787 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
2788 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
2790 return DAG.getMergeValues(Ops, DL);
2793 SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
2794 SelectionDAG &DAG) const {
2795 EVT VT = Op.getValueType();
2799 // Just do a normal 64-bit multiplication and extract the results.
2800 // We define this so that it can be used for constant division.
2801 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
2802 Op.getOperand(1), Ops[1], Ops[0]);
2804 // UMUL_LOHI64 returns the low result in the odd register and the high
2805 // result in the even register. UMUL_LOHI is defined to return the
2806 // low half first, so the results are in reverse order.
2807 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
2808 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2809 return DAG.getMergeValues(Ops, DL);
2812 SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
2813 SelectionDAG &DAG) const {
2814 SDValue Op0 = Op.getOperand(0);
2815 SDValue Op1 = Op.getOperand(1);
2816 EVT VT = Op.getValueType();
2820 // We use DSGF for 32-bit division.
2822 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
2823 Opcode = SystemZISD::SDIVREM32;
2824 } else if (DAG.ComputeNumSignBits(Op1) > 32) {
2825 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
2826 Opcode = SystemZISD::SDIVREM32;
2828 Opcode = SystemZISD::SDIVREM64;
2830 // DSG(F) takes a 64-bit dividend, so the even register in the GR128
2831 // input is "don't care". The instruction returns the remainder in
2832 // the even register and the quotient in the odd register.
2834 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
2835 Op0, Op1, Ops[1], Ops[0]);
2836 return DAG.getMergeValues(Ops, DL);
2839 SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
2840 SelectionDAG &DAG) const {
2841 EVT VT = Op.getValueType();
2844 // DL(G) uses a double-width dividend, so we need to clear the even
2845 // register in the GR128 input. The instruction returns the remainder
2846 // in the even register and the quotient in the odd register.
2849 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32,
2850 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2852 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
2853 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
2854 return DAG.getMergeValues(Ops, DL);
2857 SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
2858 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
2860 // Get the known-zero masks for each operand.
2861 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
2862 APInt KnownZero[2], KnownOne[2];
2863 DAG.computeKnownBits(Ops[0], KnownZero[0], KnownOne[0]);
2864 DAG.computeKnownBits(Ops[1], KnownZero[1], KnownOne[1]);
2866 // See if the upper 32 bits of one operand and the lower 32 bits of the
2867 // other are known zero. They are the low and high operands respectively.
2868 uint64_t Masks[] = { KnownZero[0].getZExtValue(),
2869 KnownZero[1].getZExtValue() };
2871 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
2873 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
2878 SDValue LowOp = Ops[Low];
2879 SDValue HighOp = Ops[High];
2881 // If the high part is a constant, we're better off using IILH.
2882 if (HighOp.getOpcode() == ISD::Constant)
2885 // If the low part is a constant that is outside the range of LHI,
2886 // then we're better off using IILF.
2887 if (LowOp.getOpcode() == ISD::Constant) {
2888 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue());
2889 if (!isInt<16>(Value))
2893 // Check whether the high part is an AND that doesn't change the
2894 // high 32 bits and just masks out low bits. We can skip it if so.
2895 if (HighOp.getOpcode() == ISD::AND &&
2896 HighOp.getOperand(1).getOpcode() == ISD::Constant) {
2897 SDValue HighOp0 = HighOp.getOperand(0);
2898 uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
2899 if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
2903 // Take advantage of the fact that all GR32 operations only change the
2904 // low 32 bits by truncating Low to an i32 and inserting it directly
2905 // using a subreg. The interesting cases are those where the truncation
2908 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
2909 return DAG.getTargetInsertSubreg(SystemZ::subreg_l32, DL,
2910 MVT::i64, HighOp, Low32);
2913 SDValue SystemZTargetLowering::lowerCTPOP(SDValue Op,
2914 SelectionDAG &DAG) const {
2915 EVT VT = Op.getValueType();
2917 Op = Op.getOperand(0);
2919 // Handle vector types via VPOPCT.
2920 if (VT.isVector()) {
2921 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Op);
2922 Op = DAG.getNode(SystemZISD::POPCNT, DL, MVT::v16i8, Op);
2923 switch (VT.getVectorElementType().getSizeInBits()) {
2927 Op = DAG.getNode(ISD::BITCAST, DL, VT, Op);
2928 SDValue Shift = DAG.getConstant(8, DL, MVT::i32);
2929 SDValue Tmp = DAG.getNode(SystemZISD::VSHL_BY_SCALAR, DL, VT, Op, Shift);
2930 Op = DAG.getNode(ISD::ADD, DL, VT, Op, Tmp);
2931 Op = DAG.getNode(SystemZISD::VSRL_BY_SCALAR, DL, VT, Op, Shift);
2935 SDValue Tmp = DAG.getNode(SystemZISD::BYTE_MASK, DL, MVT::v16i8,
2936 DAG.getConstant(0, DL, MVT::i32));
2937 Op = DAG.getNode(SystemZISD::VSUM, DL, VT, Op, Tmp);
2941 SDValue Tmp = DAG.getNode(SystemZISD::BYTE_MASK, DL, MVT::v16i8,
2942 DAG.getConstant(0, DL, MVT::i32));
2943 Op = DAG.getNode(SystemZISD::VSUM, DL, MVT::v4i32, Op, Tmp);
2944 Op = DAG.getNode(SystemZISD::VSUM, DL, VT, Op, Tmp);
2948 llvm_unreachable("Unexpected type");
2953 // Get the known-zero mask for the operand.
2954 APInt KnownZero, KnownOne;
2955 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2956 unsigned NumSignificantBits = (~KnownZero).getActiveBits();
2957 if (NumSignificantBits == 0)
2958 return DAG.getConstant(0, DL, VT);
2960 // Skip known-zero high parts of the operand.
2961 int64_t OrigBitSize = VT.getSizeInBits();
2962 int64_t BitSize = (int64_t)1 << Log2_32_Ceil(NumSignificantBits);
2963 BitSize = std::min(BitSize, OrigBitSize);
2965 // The POPCNT instruction counts the number of bits in each byte.
2966 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op);
2967 Op = DAG.getNode(SystemZISD::POPCNT, DL, MVT::i64, Op);
2968 Op = DAG.getNode(ISD::TRUNCATE, DL, VT, Op);
2970 // Add up per-byte counts in a binary tree. All bits of Op at
2971 // position larger than BitSize remain zero throughout.
2972 for (int64_t I = BitSize / 2; I >= 8; I = I / 2) {
2973 SDValue Tmp = DAG.getNode(ISD::SHL, DL, VT, Op, DAG.getConstant(I, DL, VT));
2974 if (BitSize != OrigBitSize)
2975 Tmp = DAG.getNode(ISD::AND, DL, VT, Tmp,
2976 DAG.getConstant(((uint64_t)1 << BitSize) - 1, DL, VT));
2977 Op = DAG.getNode(ISD::ADD, DL, VT, Op, Tmp);
2980 // Extract overall result from high byte.
2982 Op = DAG.getNode(ISD::SRL, DL, VT, Op,
2983 DAG.getConstant(BitSize - 8, DL, VT));
2988 // Op is an atomic load. Lower it into a normal volatile load.
2989 SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
2990 SelectionDAG &DAG) const {
2991 auto *Node = cast<AtomicSDNode>(Op.getNode());
2992 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), Op.getValueType(),
2993 Node->getChain(), Node->getBasePtr(),
2994 Node->getMemoryVT(), Node->getMemOperand());
2997 // Op is an atomic store. Lower it into a normal volatile store followed
2998 // by a serialization.
2999 SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
3000 SelectionDAG &DAG) const {
3001 auto *Node = cast<AtomicSDNode>(Op.getNode());
3002 SDValue Chain = DAG.getTruncStore(Node->getChain(), SDLoc(Op), Node->getVal(),
3003 Node->getBasePtr(), Node->getMemoryVT(),
3004 Node->getMemOperand());
3005 return SDValue(DAG.getMachineNode(SystemZ::Serialize, SDLoc(Op), MVT::Other,
3009 // Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
3010 // two into the fullword ATOMIC_LOADW_* operation given by Opcode.
3011 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
3013 unsigned Opcode) const {
3014 auto *Node = cast<AtomicSDNode>(Op.getNode());
3016 // 32-bit operations need no code outside the main loop.
3017 EVT NarrowVT = Node->getMemoryVT();
3018 EVT WideVT = MVT::i32;
3019 if (NarrowVT == WideVT)
3022 int64_t BitSize = NarrowVT.getSizeInBits();
3023 SDValue ChainIn = Node->getChain();
3024 SDValue Addr = Node->getBasePtr();
3025 SDValue Src2 = Node->getVal();
3026 MachineMemOperand *MMO = Node->getMemOperand();
3028 EVT PtrVT = Addr.getValueType();
3030 // Convert atomic subtracts of constants into additions.
3031 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
3032 if (auto *Const = dyn_cast<ConstantSDNode>(Src2)) {
3033 Opcode = SystemZISD::ATOMIC_LOADW_ADD;
3034 Src2 = DAG.getConstant(-Const->getSExtValue(), DL, Src2.getValueType());
3037 // Get the address of the containing word.
3038 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
3039 DAG.getConstant(-4, DL, PtrVT));
3041 // Get the number of bits that the word must be rotated left in order
3042 // to bring the field to the top bits of a GR32.
3043 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
3044 DAG.getConstant(3, DL, PtrVT));
3045 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
3047 // Get the complementing shift amount, for rotating a field in the top
3048 // bits back to its proper position.
3049 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
3050 DAG.getConstant(0, DL, WideVT), BitShift);
3052 // Extend the source operand to 32 bits and prepare it for the inner loop.
3053 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
3054 // operations require the source to be shifted in advance. (This shift
3055 // can be folded if the source is constant.) For AND and NAND, the lower
3056 // bits must be set, while for other opcodes they should be left clear.
3057 if (Opcode != SystemZISD::ATOMIC_SWAPW)
3058 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
3059 DAG.getConstant(32 - BitSize, DL, WideVT));
3060 if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
3061 Opcode == SystemZISD::ATOMIC_LOADW_NAND)
3062 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
3063 DAG.getConstant(uint32_t(-1) >> BitSize, DL, WideVT));
3065 // Construct the ATOMIC_LOADW_* node.
3066 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
3067 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
3068 DAG.getConstant(BitSize, DL, WideVT) };
3069 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
3072 // Rotate the result of the final CS so that the field is in the lower
3073 // bits of a GR32, then truncate it.
3074 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
3075 DAG.getConstant(BitSize, DL, WideVT));
3076 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
3078 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
3079 return DAG.getMergeValues(RetOps, DL);
3082 // Op is an ATOMIC_LOAD_SUB operation. Lower 8- and 16-bit operations
3083 // into ATOMIC_LOADW_SUBs and decide whether to convert 32- and 64-bit
3084 // operations into additions.
3085 SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
3086 SelectionDAG &DAG) const {
3087 auto *Node = cast<AtomicSDNode>(Op.getNode());
3088 EVT MemVT = Node->getMemoryVT();
3089 if (MemVT == MVT::i32 || MemVT == MVT::i64) {
3090 // A full-width operation.
3091 assert(Op.getValueType() == MemVT && "Mismatched VTs");
3092 SDValue Src2 = Node->getVal();
3096 if (auto *Op2 = dyn_cast<ConstantSDNode>(Src2)) {
3097 // Use an addition if the operand is constant and either LAA(G) is
3098 // available or the negative value is in the range of A(G)FHI.
3099 int64_t Value = (-Op2->getAPIntValue()).getSExtValue();
3100 if (isInt<32>(Value) || Subtarget.hasInterlockedAccess1())
3101 NegSrc2 = DAG.getConstant(Value, DL, MemVT);
3102 } else if (Subtarget.hasInterlockedAccess1())
3103 // Use LAA(G) if available.
3104 NegSrc2 = DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, DL, MemVT),
3107 if (NegSrc2.getNode())
3108 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
3109 Node->getChain(), Node->getBasePtr(), NegSrc2,
3110 Node->getMemOperand(), Node->getOrdering(),
3111 Node->getSynchScope());
3113 // Use the node as-is.
3117 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
3120 // Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
3121 // into a fullword ATOMIC_CMP_SWAPW operation.
3122 SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
3123 SelectionDAG &DAG) const {
3124 auto *Node = cast<AtomicSDNode>(Op.getNode());
3126 // We have native support for 32-bit compare and swap.
3127 EVT NarrowVT = Node->getMemoryVT();
3128 EVT WideVT = MVT::i32;
3129 if (NarrowVT == WideVT)
3132 int64_t BitSize = NarrowVT.getSizeInBits();
3133 SDValue ChainIn = Node->getOperand(0);
3134 SDValue Addr = Node->getOperand(1);
3135 SDValue CmpVal = Node->getOperand(2);
3136 SDValue SwapVal = Node->getOperand(3);
3137 MachineMemOperand *MMO = Node->getMemOperand();
3139 EVT PtrVT = Addr.getValueType();
3141 // Get the address of the containing word.
3142 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
3143 DAG.getConstant(-4, DL, PtrVT));
3145 // Get the number of bits that the word must be rotated left in order
3146 // to bring the field to the top bits of a GR32.
3147 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
3148 DAG.getConstant(3, DL, PtrVT));
3149 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
3151 // Get the complementing shift amount, for rotating a field in the top
3152 // bits back to its proper position.
3153 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
3154 DAG.getConstant(0, DL, WideVT), BitShift);
3156 // Construct the ATOMIC_CMP_SWAPW node.
3157 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
3158 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
3159 NegBitShift, DAG.getConstant(BitSize, DL, WideVT) };
3160 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
3161 VTList, Ops, NarrowVT, MMO);
3165 SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
3166 SelectionDAG &DAG) const {
3167 MachineFunction &MF = DAG.getMachineFunction();
3168 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
3169 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
3170 SystemZ::R15D, Op.getValueType());
3173 SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
3174 SelectionDAG &DAG) const {
3175 MachineFunction &MF = DAG.getMachineFunction();
3176 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true);
3177 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op),
3178 SystemZ::R15D, Op.getOperand(1));
3181 SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op,
3182 SelectionDAG &DAG) const {
3183 bool IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
3185 // Just preserve the chain.
3186 return Op.getOperand(0);
3189 bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
3190 unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
3191 auto *Node = cast<MemIntrinsicSDNode>(Op.getNode());
3194 DAG.getConstant(Code, DL, MVT::i32),
3197 return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, DL,
3198 Node->getVTList(), Ops,
3199 Node->getMemoryVT(), Node->getMemOperand());
3202 // Return an i32 that contains the value of CC immediately after After,
3203 // whose final operand must be MVT::Glue.
3204 static SDValue getCCResult(SelectionDAG &DAG, SDNode *After) {
3206 SDValue Glue = SDValue(After, After->getNumValues() - 1);
3207 SDValue IPM = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
3208 return DAG.getNode(ISD::SRL, DL, MVT::i32, IPM,
3209 DAG.getConstant(SystemZ::IPM_CC, DL, MVT::i32));
3213 SystemZTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
3214 SelectionDAG &DAG) const {
3215 unsigned Opcode, CCValid;
3216 if (isIntrinsicWithCCAndChain(Op, Opcode, CCValid)) {
3217 assert(Op->getNumValues() == 2 && "Expected only CC result and chain");
3218 SDValue Glued = emitIntrinsicWithChainAndGlue(DAG, Op, Opcode);
3219 SDValue CC = getCCResult(DAG, Glued.getNode());
3220 DAG.ReplaceAllUsesOfValueWith(SDValue(Op.getNode(), 0), CC);
3228 SystemZTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
3229 SelectionDAG &DAG) const {
3230 unsigned Opcode, CCValid;
3231 if (isIntrinsicWithCC(Op, Opcode, CCValid)) {
3232 SDValue Glued = emitIntrinsicWithGlue(DAG, Op, Opcode);
3233 SDValue CC = getCCResult(DAG, Glued.getNode());
3234 if (Op->getNumValues() == 1)
3236 assert(Op->getNumValues() == 2 && "Expected a CC and non-CC result");
3237 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), Op->getVTList(),
3241 unsigned Id = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3243 case Intrinsic::s390_vpdi:
3244 return DAG.getNode(SystemZISD::PERMUTE_DWORDS, SDLoc(Op), Op.getValueType(),
3245 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
3247 case Intrinsic::s390_vperm:
3248 return DAG.getNode(SystemZISD::PERMUTE, SDLoc(Op), Op.getValueType(),
3249 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
3251 case Intrinsic::s390_vuphb:
3252 case Intrinsic::s390_vuphh:
3253 case Intrinsic::s390_vuphf:
3254 return DAG.getNode(SystemZISD::UNPACK_HIGH, SDLoc(Op), Op.getValueType(),
3257 case Intrinsic::s390_vuplhb:
3258 case Intrinsic::s390_vuplhh:
3259 case Intrinsic::s390_vuplhf:
3260 return DAG.getNode(SystemZISD::UNPACKL_HIGH, SDLoc(Op), Op.getValueType(),
3263 case Intrinsic::s390_vuplb:
3264 case Intrinsic::s390_vuplhw:
3265 case Intrinsic::s390_vuplf:
3266 return DAG.getNode(SystemZISD::UNPACK_LOW, SDLoc(Op), Op.getValueType(),
3269 case Intrinsic::s390_vupllb:
3270 case Intrinsic::s390_vupllh:
3271 case Intrinsic::s390_vupllf:
3272 return DAG.getNode(SystemZISD::UNPACKL_LOW, SDLoc(Op), Op.getValueType(),
3275 case Intrinsic::s390_vsumb:
3276 case Intrinsic::s390_vsumh:
3277 case Intrinsic::s390_vsumgh:
3278 case Intrinsic::s390_vsumgf:
3279 case Intrinsic::s390_vsumqf:
3280 case Intrinsic::s390_vsumqg:
3281 return DAG.getNode(SystemZISD::VSUM, SDLoc(Op), Op.getValueType(),
3282 Op.getOperand(1), Op.getOperand(2));
3289 // Says that SystemZISD operation Opcode can be used to perform the equivalent
3290 // of a VPERM with permute vector Bytes. If Opcode takes three operands,
3291 // Operand is the constant third operand, otherwise it is the number of
3292 // bytes in each element of the result.
3296 unsigned char Bytes[SystemZ::VectorBytes];
3300 static const Permute PermuteForms[] = {
3302 { SystemZISD::MERGE_HIGH, 8,
3303 { 0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23 } },
3305 { SystemZISD::MERGE_HIGH, 4,
3306 { 0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23 } },
3308 { SystemZISD::MERGE_HIGH, 2,
3309 { 0, 1, 16, 17, 2, 3, 18, 19, 4, 5, 20, 21, 6, 7, 22, 23 } },
3311 { SystemZISD::MERGE_HIGH, 1,
3312 { 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23 } },
3314 { SystemZISD::MERGE_LOW, 8,
3315 { 8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31 } },
3317 { SystemZISD::MERGE_LOW, 4,
3318 { 8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31 } },
3320 { SystemZISD::MERGE_LOW, 2,
3321 { 8, 9, 24, 25, 10, 11, 26, 27, 12, 13, 28, 29, 14, 15, 30, 31 } },
3323 { SystemZISD::MERGE_LOW, 1,
3324 { 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31 } },
3326 { SystemZISD::PACK, 4,
3327 { 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 } },
3329 { SystemZISD::PACK, 2,
3330 { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 } },
3332 { SystemZISD::PACK, 1,
3333 { 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31 } },
3334 // VPDI V1, V2, 4 (low half of V1, high half of V2)
3335 { SystemZISD::PERMUTE_DWORDS, 4,
3336 { 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 } },
3337 // VPDI V1, V2, 1 (high half of V1, low half of V2)
3338 { SystemZISD::PERMUTE_DWORDS, 1,
3339 { 0, 1, 2, 3, 4, 5, 6, 7, 24, 25, 26, 27, 28, 29, 30, 31 } }
3342 // Called after matching a vector shuffle against a particular pattern.
3343 // Both the original shuffle and the pattern have two vector operands.
3344 // OpNos[0] is the operand of the original shuffle that should be used for
3345 // operand 0 of the pattern, or -1 if operand 0 of the pattern can be anything.
3346 // OpNos[1] is the same for operand 1 of the pattern. Resolve these -1s and
3347 // set OpNo0 and OpNo1 to the shuffle operands that should actually be used
3348 // for operands 0 and 1 of the pattern.
3349 static bool chooseShuffleOpNos(int *OpNos, unsigned &OpNo0, unsigned &OpNo1) {
3353 OpNo0 = OpNo1 = OpNos[1];
3354 } else if (OpNos[1] < 0) {
3355 OpNo0 = OpNo1 = OpNos[0];
3363 // Bytes is a VPERM-like permute vector, except that -1 is used for
3364 // undefined bytes. Return true if the VPERM can be implemented using P.
3365 // When returning true set OpNo0 to the VPERM operand that should be
3366 // used for operand 0 of P and likewise OpNo1 for operand 1 of P.
3368 // For example, if swapping the VPERM operands allows P to match, OpNo0
3369 // will be 1 and OpNo1 will be 0. If instead Bytes only refers to one
3370 // operand, but rewriting it to use two duplicated operands allows it to
3371 // match P, then OpNo0 and OpNo1 will be the same.
3372 static bool matchPermute(const SmallVectorImpl<int> &Bytes, const Permute &P,
3373 unsigned &OpNo0, unsigned &OpNo1) {
3374 int OpNos[] = { -1, -1 };
3375 for (unsigned I = 0; I < SystemZ::VectorBytes; ++I) {
3378 // Make sure that the two permute vectors use the same suboperand
3379 // byte number. Only the operand numbers (the high bits) are
3380 // allowed to differ.
3381 if ((Elt ^ P.Bytes[I]) & (SystemZ::VectorBytes - 1))
3383 int ModelOpNo = P.Bytes[I] / SystemZ::VectorBytes;
3384 int RealOpNo = unsigned(Elt) / SystemZ::VectorBytes;
3385 // Make sure that the operand mappings are consistent with previous
3387 if (OpNos[ModelOpNo] == 1 - RealOpNo)
3389 OpNos[ModelOpNo] = RealOpNo;
3392 return chooseShuffleOpNos(OpNos, OpNo0, OpNo1);
3395 // As above, but search for a matching permute.
3396 static const Permute *matchPermute(const SmallVectorImpl<int> &Bytes,
3397 unsigned &OpNo0, unsigned &OpNo1) {
3398 for (auto &P : PermuteForms)
3399 if (matchPermute(Bytes, P, OpNo0, OpNo1))
3404 // Bytes is a VPERM-like permute vector, except that -1 is used for
3405 // undefined bytes. This permute is an operand of an outer permute.
3406 // See whether redistributing the -1 bytes gives a shuffle that can be
3407 // implemented using P. If so, set Transform to a VPERM-like permute vector
3408 // that, when applied to the result of P, gives the original permute in Bytes.
3409 static bool matchDoublePermute(const SmallVectorImpl<int> &Bytes,
3411 SmallVectorImpl<int> &Transform) {
3413 for (unsigned From = 0; From < SystemZ::VectorBytes; ++From) {
3414 int Elt = Bytes[From];
3416 // Byte number From of the result is undefined.
3417 Transform[From] = -1;
3419 while (P.Bytes[To] != Elt) {
3421 if (To == SystemZ::VectorBytes)
3424 Transform[From] = To;
3430 // As above, but search for a matching permute.
3431 static const Permute *matchDoublePermute(const SmallVectorImpl<int> &Bytes,
3432 SmallVectorImpl<int> &Transform) {
3433 for (auto &P : PermuteForms)
3434 if (matchDoublePermute(Bytes, P, Transform))
3439 // Convert the mask of the given VECTOR_SHUFFLE into a byte-level mask,
3440 // as if it had type vNi8.
3441 static void getVPermMask(ShuffleVectorSDNode *VSN,
3442 SmallVectorImpl<int> &Bytes) {
3443 EVT VT = VSN->getValueType(0);
3444 unsigned NumElements = VT.getVectorNumElements();
3445 unsigned BytesPerElement = VT.getVectorElementType().getStoreSize();
3446 Bytes.resize(NumElements * BytesPerElement, -1);
3447 for (unsigned I = 0; I < NumElements; ++I) {
3448 int Index = VSN->getMaskElt(I);
3450 for (unsigned J = 0; J < BytesPerElement; ++J)
3451 Bytes[I * BytesPerElement + J] = Index * BytesPerElement + J;
3455 // Bytes is a VPERM-like permute vector, except that -1 is used for
3456 // undefined bytes. See whether bytes [Start, Start + BytesPerElement) of
3457 // the result come from a contiguous sequence of bytes from one input.
3458 // Set Base to the selector for the first byte if so.
3459 static bool getShuffleInput(const SmallVectorImpl<int> &Bytes, unsigned Start,
3460 unsigned BytesPerElement, int &Base) {
3462 for (unsigned I = 0; I < BytesPerElement; ++I) {
3463 if (Bytes[Start + I] >= 0) {
3464 unsigned Elem = Bytes[Start + I];
3467 // Make sure the bytes would come from one input operand.
3468 if (unsigned(Base) % Bytes.size() + BytesPerElement > Bytes.size())
3470 } else if (unsigned(Base) != Elem - I)
3477 // Bytes is a VPERM-like permute vector, except that -1 is used for
3478 // undefined bytes. Return true if it can be performed using VSLDI.
3479 // When returning true, set StartIndex to the shift amount and OpNo0
3480 // and OpNo1 to the VPERM operands that should be used as the first
3481 // and second shift operand respectively.
3482 static bool isShlDoublePermute(const SmallVectorImpl<int> &Bytes,
3483 unsigned &StartIndex, unsigned &OpNo0,
3485 int OpNos[] = { -1, -1 };
3487 for (unsigned I = 0; I < 16; ++I) {
3488 int Index = Bytes[I];
3490 int ExpectedShift = (Index - I) % SystemZ::VectorBytes;
3491 int ModelOpNo = unsigned(ExpectedShift + I) / SystemZ::VectorBytes;
3492 int RealOpNo = unsigned(Index) / SystemZ::VectorBytes;
3494 Shift = ExpectedShift;
3495 else if (Shift != ExpectedShift)
3497 // Make sure that the operand mappings are consistent with previous
3499 if (OpNos[ModelOpNo] == 1 - RealOpNo)
3501 OpNos[ModelOpNo] = RealOpNo;
3505 return chooseShuffleOpNos(OpNos, OpNo0, OpNo1);
3508 // Create a node that performs P on operands Op0 and Op1, casting the
3509 // operands to the appropriate type. The type of the result is determined by P.
3510 static SDValue getPermuteNode(SelectionDAG &DAG, SDLoc DL,
3511 const Permute &P, SDValue Op0, SDValue Op1) {
3512 // VPDI (PERMUTE_DWORDS) always operates on v2i64s. The input
3513 // elements of a PACK are twice as wide as the outputs.
3514 unsigned InBytes = (P.Opcode == SystemZISD::PERMUTE_DWORDS ? 8 :
3515 P.Opcode == SystemZISD::PACK ? P.Operand * 2 :
3517 // Cast both operands to the appropriate type.
3518 MVT InVT = MVT::getVectorVT(MVT::getIntegerVT(InBytes * 8),
3519 SystemZ::VectorBytes / InBytes);
3520 Op0 = DAG.getNode(ISD::BITCAST, DL, InVT, Op0);
3521 Op1 = DAG.getNode(ISD::BITCAST, DL, InVT, Op1);
3523 if (P.Opcode == SystemZISD::PERMUTE_DWORDS) {
3524 SDValue Op2 = DAG.getConstant(P.Operand, DL, MVT::i32);
3525 Op = DAG.getNode(SystemZISD::PERMUTE_DWORDS, DL, InVT, Op0, Op1, Op2);
3526 } else if (P.Opcode == SystemZISD::PACK) {
3527 MVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(P.Operand * 8),
3528 SystemZ::VectorBytes / P.Operand);
3529 Op = DAG.getNode(SystemZISD::PACK, DL, OutVT, Op0, Op1);
3531 Op = DAG.getNode(P.Opcode, DL, InVT, Op0, Op1);
3536 // Bytes is a VPERM-like permute vector, except that -1 is used for
3537 // undefined bytes. Implement it on operands Ops[0] and Ops[1] using
3539 static SDValue getGeneralPermuteNode(SelectionDAG &DAG, SDLoc DL, SDValue *Ops,
3540 const SmallVectorImpl<int> &Bytes) {
3541 for (unsigned I = 0; I < 2; ++I)
3542 Ops[I] = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Ops[I]);
3544 // First see whether VSLDI can be used.
3545 unsigned StartIndex, OpNo0, OpNo1;
3546 if (isShlDoublePermute(Bytes, StartIndex, OpNo0, OpNo1))
3547 return DAG.getNode(SystemZISD::SHL_DOUBLE, DL, MVT::v16i8, Ops[OpNo0],
3548 Ops[OpNo1], DAG.getConstant(StartIndex, DL, MVT::i32));
3550 // Fall back on VPERM. Construct an SDNode for the permute vector.
3551 SDValue IndexNodes[SystemZ::VectorBytes];
3552 for (unsigned I = 0; I < SystemZ::VectorBytes; ++I)
3554 IndexNodes[I] = DAG.getConstant(Bytes[I], DL, MVT::i32);
3556 IndexNodes[I] = DAG.getUNDEF(MVT::i32);
3557 SDValue Op2 = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, IndexNodes);
3558 return DAG.getNode(SystemZISD::PERMUTE, DL, MVT::v16i8, Ops[0], Ops[1], Op2);
3562 // Describes a general N-operand vector shuffle.
3563 struct GeneralShuffle {
3564 GeneralShuffle(EVT vt) : VT(vt) {}
3566 void add(SDValue, unsigned);
3567 SDValue getNode(SelectionDAG &, SDLoc);
3569 // The operands of the shuffle.
3570 SmallVector<SDValue, SystemZ::VectorBytes> Ops;
3572 // Index I is -1 if byte I of the result is undefined. Otherwise the
3573 // result comes from byte Bytes[I] % SystemZ::VectorBytes of operand
3574 // Bytes[I] / SystemZ::VectorBytes.
3575 SmallVector<int, SystemZ::VectorBytes> Bytes;
3577 // The type of the shuffle result.
3582 // Add an extra undefined element to the shuffle.
3583 void GeneralShuffle::addUndef() {
3584 unsigned BytesPerElement = VT.getVectorElementType().getStoreSize();
3585 for (unsigned I = 0; I < BytesPerElement; ++I)
3586 Bytes.push_back(-1);
3589 // Add an extra element to the shuffle, taking it from element Elem of Op.
3590 // A null Op indicates a vector input whose value will be calculated later;
3591 // there is at most one such input per shuffle and it always has the same
3592 // type as the result.
3593 void GeneralShuffle::add(SDValue Op, unsigned Elem) {
3594 unsigned BytesPerElement = VT.getVectorElementType().getStoreSize();
3596 // The source vector can have wider elements than the result,
3597 // either through an explicit TRUNCATE or because of type legalization.
3598 // We want the least significant part.
3599 EVT FromVT = Op.getNode() ? Op.getValueType() : VT;
3600 unsigned FromBytesPerElement = FromVT.getVectorElementType().getStoreSize();
3601 assert(FromBytesPerElement >= BytesPerElement &&
3602 "Invalid EXTRACT_VECTOR_ELT");
3603 unsigned Byte = ((Elem * FromBytesPerElement) % SystemZ::VectorBytes +
3604 (FromBytesPerElement - BytesPerElement));
3606 // Look through things like shuffles and bitcasts.
3607 while (Op.getNode()) {
3608 if (Op.getOpcode() == ISD::BITCAST)
3609 Op = Op.getOperand(0);
3610 else if (Op.getOpcode() == ISD::VECTOR_SHUFFLE && Op.hasOneUse()) {
3611 // See whether the bytes we need come from a contiguous part of one
3613 SmallVector<int, SystemZ::VectorBytes> OpBytes;
3614 getVPermMask(cast<ShuffleVectorSDNode>(Op), OpBytes);
3616 if (!getShuffleInput(OpBytes, Byte, BytesPerElement, NewByte))
3622 Op = Op.getOperand(unsigned(NewByte) / SystemZ::VectorBytes);
3623 Byte = unsigned(NewByte) % SystemZ::VectorBytes;
3624 } else if (Op.getOpcode() == ISD::UNDEF) {
3631 // Make sure that the source of the extraction is in Ops.
3633 for (; OpNo < Ops.size(); ++OpNo)
3634 if (Ops[OpNo] == Op)
3636 if (OpNo == Ops.size())
3639 // Add the element to Bytes.
3640 unsigned Base = OpNo * SystemZ::VectorBytes + Byte;
3641 for (unsigned I = 0; I < BytesPerElement; ++I)
3642 Bytes.push_back(Base + I);
3645 // Return SDNodes for the completed shuffle.
3646 SDValue GeneralShuffle::getNode(SelectionDAG &DAG, SDLoc DL) {
3647 assert(Bytes.size() == SystemZ::VectorBytes && "Incomplete vector");
3649 if (Ops.size() == 0)
3650 return DAG.getUNDEF(VT);
3652 // Make sure that there are at least two shuffle operands.
3653 if (Ops.size() == 1)
3654 Ops.push_back(DAG.getUNDEF(MVT::v16i8));
3656 // Create a tree of shuffles, deferring root node until after the loop.
3657 // Try to redistribute the undefined elements of non-root nodes so that
3658 // the non-root shuffles match something like a pack or merge, then adjust
3659 // the parent node's permute vector to compensate for the new order.
3660 // Among other things, this copes with vectors like <2 x i16> that were
3661 // padded with undefined elements during type legalization.
3663 // In the best case this redistribution will lead to the whole tree
3664 // using packs and merges. It should rarely be a loss in other cases.
3665 unsigned Stride = 1;
3666 for (; Stride * 2 < Ops.size(); Stride *= 2) {
3667 for (unsigned I = 0; I < Ops.size() - Stride; I += Stride * 2) {
3668 SDValue SubOps[] = { Ops[I], Ops[I + Stride] };
3670 // Create a mask for just these two operands.
3671 SmallVector<int, SystemZ::VectorBytes> NewBytes(SystemZ::VectorBytes);
3672 for (unsigned J = 0; J < SystemZ::VectorBytes; ++J) {
3673 unsigned OpNo = unsigned(Bytes[J]) / SystemZ::VectorBytes;
3674 unsigned Byte = unsigned(Bytes[J]) % SystemZ::VectorBytes;
3677 else if (OpNo == I + Stride)
3678 NewBytes[J] = SystemZ::VectorBytes + Byte;
3682 // See if it would be better to reorganize NewMask to avoid using VPERM.
3683 SmallVector<int, SystemZ::VectorBytes> NewBytesMap(SystemZ::VectorBytes);
3684 if (const Permute *P = matchDoublePermute(NewBytes, NewBytesMap)) {
3685 Ops[I] = getPermuteNode(DAG, DL, *P, SubOps[0], SubOps[1]);
3686 // Applying NewBytesMap to Ops[I] gets back to NewBytes.
3687 for (unsigned J = 0; J < SystemZ::VectorBytes; ++J) {
3688 if (NewBytes[J] >= 0) {
3689 assert(unsigned(NewBytesMap[J]) < SystemZ::VectorBytes &&
3690 "Invalid double permute");
3691 Bytes[J] = I * SystemZ::VectorBytes + NewBytesMap[J];
3693 assert(NewBytesMap[J] < 0 && "Invalid double permute");
3696 // Just use NewBytes on the operands.
3697 Ops[I] = getGeneralPermuteNode(DAG, DL, SubOps, NewBytes);
3698 for (unsigned J = 0; J < SystemZ::VectorBytes; ++J)
3699 if (NewBytes[J] >= 0)
3700 Bytes[J] = I * SystemZ::VectorBytes + J;
3705 // Now we just have 2 inputs. Put the second operand in Ops[1].
3707 Ops[1] = Ops[Stride];
3708 for (unsigned I = 0; I < SystemZ::VectorBytes; ++I)
3709 if (Bytes[I] >= int(SystemZ::VectorBytes))
3710 Bytes[I] -= (Stride - 1) * SystemZ::VectorBytes;
3713 // Look for an instruction that can do the permute without resorting
3715 unsigned OpNo0, OpNo1;
3717 if (const Permute *P = matchPermute(Bytes, OpNo0, OpNo1))
3718 Op = getPermuteNode(DAG, DL, *P, Ops[OpNo0], Ops[OpNo1]);
3720 Op = getGeneralPermuteNode(DAG, DL, &Ops[0], Bytes);
3721 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
3724 // Return true if the given BUILD_VECTOR is a scalar-to-vector conversion.
3725 static bool isScalarToVector(SDValue Op) {
3726 for (unsigned I = 1, E = Op.getNumOperands(); I != E; ++I)
3727 if (Op.getOperand(I).getOpcode() != ISD::UNDEF)
3732 // Return a vector of type VT that contains Value in the first element.
3733 // The other elements don't matter.
3734 static SDValue buildScalarToVector(SelectionDAG &DAG, SDLoc DL, EVT VT,
3736 // If we have a constant, replicate it to all elements and let the
3737 // BUILD_VECTOR lowering take care of it.
3738 if (Value.getOpcode() == ISD::Constant ||
3739 Value.getOpcode() == ISD::ConstantFP) {
3740 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Value);
3741 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
3743 if (Value.getOpcode() == ISD::UNDEF)
3744 return DAG.getUNDEF(VT);
3745 return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
3748 // Return a vector of type VT in which Op0 is in element 0 and Op1 is in
3749 // element 1. Used for cases in which replication is cheap.
3750 static SDValue buildMergeScalars(SelectionDAG &DAG, SDLoc DL, EVT VT,
3751 SDValue Op0, SDValue Op1) {
3752 if (Op0.getOpcode() == ISD::UNDEF) {
3753 if (Op1.getOpcode() == ISD::UNDEF)
3754 return DAG.getUNDEF(VT);
3755 return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op1);
3757 if (Op1.getOpcode() == ISD::UNDEF)
3758 return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op0);
3759 return DAG.getNode(SystemZISD::MERGE_HIGH, DL, VT,
3760 buildScalarToVector(DAG, DL, VT, Op0),
3761 buildScalarToVector(DAG, DL, VT, Op1));
3764 // Extend GPR scalars Op0 and Op1 to doublewords and return a v2i64
3766 static SDValue joinDwords(SelectionDAG &DAG, SDLoc DL, SDValue Op0,
3768 if (Op0.getOpcode() == ISD::UNDEF && Op1.getOpcode() == ISD::UNDEF)
3769 return DAG.getUNDEF(MVT::v2i64);
3770 // If one of the two inputs is undefined then replicate the other one,
3771 // in order to avoid using another register unnecessarily.
3772 if (Op0.getOpcode() == ISD::UNDEF)
3773 Op0 = Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op1);
3774 else if (Op1.getOpcode() == ISD::UNDEF)
3775 Op0 = Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
3777 Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
3778 Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op1);
3780 return DAG.getNode(SystemZISD::JOIN_DWORDS, DL, MVT::v2i64, Op0, Op1);
3783 // Try to represent constant BUILD_VECTOR node BVN using a
3784 // SystemZISD::BYTE_MASK-style mask. Store the mask value in Mask
3786 static bool tryBuildVectorByteMask(BuildVectorSDNode *BVN, uint64_t &Mask) {
3787 EVT ElemVT = BVN->getValueType(0).getVectorElementType();
3788 unsigned BytesPerElement = ElemVT.getStoreSize();
3789 for (unsigned I = 0, E = BVN->getNumOperands(); I != E; ++I) {
3790 SDValue Op = BVN->getOperand(I);
3791 if (Op.getOpcode() != ISD::UNDEF) {
3793 if (Op.getOpcode() == ISD::Constant)
3794 Value = dyn_cast<ConstantSDNode>(Op)->getZExtValue();
3795 else if (Op.getOpcode() == ISD::ConstantFP)
3796 Value = (dyn_cast<ConstantFPSDNode>(Op)->getValueAPF().bitcastToAPInt()
3800 for (unsigned J = 0; J < BytesPerElement; ++J) {
3801 uint64_t Byte = (Value >> (J * 8)) & 0xff;
3803 Mask |= 1ULL << ((E - I - 1) * BytesPerElement + J);
3812 // Try to load a vector constant in which BitsPerElement-bit value Value
3813 // is replicated to fill the vector. VT is the type of the resulting
3814 // constant, which may have elements of a different size from BitsPerElement.
3815 // Return the SDValue of the constant on success, otherwise return
3817 static SDValue tryBuildVectorReplicate(SelectionDAG &DAG,
3818 const SystemZInstrInfo *TII,
3819 SDLoc DL, EVT VT, uint64_t Value,
3820 unsigned BitsPerElement) {
3821 // Signed 16-bit values can be replicated using VREPI.
3822 int64_t SignedValue = SignExtend64(Value, BitsPerElement);
3823 if (isInt<16>(SignedValue)) {
3824 MVT VecVT = MVT::getVectorVT(MVT::getIntegerVT(BitsPerElement),
3825 SystemZ::VectorBits / BitsPerElement);
3826 SDValue Op = DAG.getNode(SystemZISD::REPLICATE, DL, VecVT,
3827 DAG.getConstant(SignedValue, DL, MVT::i32));
3828 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
3830 // See whether rotating the constant left some N places gives a value that
3831 // is one less than a power of 2 (i.e. all zeros followed by all ones).
3832 // If so we can use VGM.
3833 unsigned Start, End;
3834 if (TII->isRxSBGMask(Value, BitsPerElement, Start, End)) {
3835 // isRxSBGMask returns the bit numbers for a full 64-bit value,
3836 // with 0 denoting 1 << 63 and 63 denoting 1. Convert them to
3837 // bit numbers for an BitsPerElement value, so that 0 denotes
3838 // 1 << (BitsPerElement-1).
3839 Start -= 64 - BitsPerElement;
3840 End -= 64 - BitsPerElement;
3841 MVT VecVT = MVT::getVectorVT(MVT::getIntegerVT(BitsPerElement),
3842 SystemZ::VectorBits / BitsPerElement);
3843 SDValue Op = DAG.getNode(SystemZISD::ROTATE_MASK, DL, VecVT,
3844 DAG.getConstant(Start, DL, MVT::i32),
3845 DAG.getConstant(End, DL, MVT::i32));
3846 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
3851 // If a BUILD_VECTOR contains some EXTRACT_VECTOR_ELTs, it's usually
3852 // better to use VECTOR_SHUFFLEs on them, only using BUILD_VECTOR for
3853 // the non-EXTRACT_VECTOR_ELT elements. See if the given BUILD_VECTOR
3854 // would benefit from this representation and return it if so.
3855 static SDValue tryBuildVectorShuffle(SelectionDAG &DAG,
3856 BuildVectorSDNode *BVN) {
3857 EVT VT = BVN->getValueType(0);
3858 unsigned NumElements = VT.getVectorNumElements();
3860 // Represent the BUILD_VECTOR as an N-operand VECTOR_SHUFFLE-like operation
3861 // on byte vectors. If there are non-EXTRACT_VECTOR_ELT elements that still
3862 // need a BUILD_VECTOR, add an additional placeholder operand for that
3863 // BUILD_VECTOR and store its operands in ResidueOps.
3864 GeneralShuffle GS(VT);
3865 SmallVector<SDValue, SystemZ::VectorBytes> ResidueOps;
3866 bool FoundOne = false;
3867 for (unsigned I = 0; I < NumElements; ++I) {
3868 SDValue Op = BVN->getOperand(I);
3869 if (Op.getOpcode() == ISD::TRUNCATE)
3870 Op = Op.getOperand(0);
3871 if (Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
3872 Op.getOperand(1).getOpcode() == ISD::Constant) {
3873 unsigned Elem = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
3874 GS.add(Op.getOperand(0), Elem);
3876 } else if (Op.getOpcode() == ISD::UNDEF) {
3879 GS.add(SDValue(), ResidueOps.size());
3880 ResidueOps.push_back(Op);
3884 // Nothing to do if there are no EXTRACT_VECTOR_ELTs.
3888 // Create the BUILD_VECTOR for the remaining elements, if any.
3889 if (!ResidueOps.empty()) {
3890 while (ResidueOps.size() < NumElements)
3891 ResidueOps.push_back(DAG.getUNDEF(VT.getVectorElementType()));
3892 for (auto &Op : GS.Ops) {
3893 if (!Op.getNode()) {
3894 Op = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BVN), VT, ResidueOps);
3899 return GS.getNode(DAG, SDLoc(BVN));
3902 // Combine GPR scalar values Elems into a vector of type VT.
3903 static SDValue buildVector(SelectionDAG &DAG, SDLoc DL, EVT VT,
3904 SmallVectorImpl<SDValue> &Elems) {
3905 // See whether there is a single replicated value.
3907 unsigned int NumElements = Elems.size();
3908 unsigned int Count = 0;
3909 for (auto Elem : Elems) {
3910 if (Elem.getOpcode() != ISD::UNDEF) {
3911 if (!Single.getNode())
3913 else if (Elem != Single) {
3920 // There are three cases here:
3922 // - if the only defined element is a loaded one, the best sequence
3923 // is a replicating load.
3925 // - otherwise, if the only defined element is an i64 value, we will
3926 // end up with the same VLVGP sequence regardless of whether we short-cut
3927 // for replication or fall through to the later code.
3929 // - otherwise, if the only defined element is an i32 or smaller value,
3930 // we would need 2 instructions to replicate it: VLVGP followed by VREPx.
3931 // This is only a win if the single defined element is used more than once.
3932 // In other cases we're better off using a single VLVGx.
3933 if (Single.getNode() && (Count > 1 || Single.getOpcode() == ISD::LOAD))
3934 return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Single);
3936 // The best way of building a v2i64 from two i64s is to use VLVGP.
3937 if (VT == MVT::v2i64)
3938 return joinDwords(DAG, DL, Elems[0], Elems[1]);
3940 // Use a 64-bit merge high to combine two doubles.
3941 if (VT == MVT::v2f64)
3942 return buildMergeScalars(DAG, DL, VT, Elems[0], Elems[1]);
3944 // Build v4f32 values directly from the FPRs:
3946 // <Axxx> <Bxxx> <Cxxxx> <Dxxx>
3951 if (VT == MVT::v4f32) {
3952 SDValue Op01 = buildMergeScalars(DAG, DL, VT, Elems[0], Elems[1]);
3953 SDValue Op23 = buildMergeScalars(DAG, DL, VT, Elems[2], Elems[3]);
3954 // Avoid unnecessary undefs by reusing the other operand.
3955 if (Op01.getOpcode() == ISD::UNDEF)
3957 else if (Op23.getOpcode() == ISD::UNDEF)
3959 // Merging identical replications is a no-op.
3960 if (Op01.getOpcode() == SystemZISD::REPLICATE && Op01 == Op23)
3962 Op01 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Op01);
3963 Op23 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Op23);
3964 SDValue Op = DAG.getNode(SystemZISD::MERGE_HIGH,
3965 DL, MVT::v2i64, Op01, Op23);
3966 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
3969 // Collect the constant terms.
3970 SmallVector<SDValue, SystemZ::VectorBytes> Constants(NumElements, SDValue());
3971 SmallVector<bool, SystemZ::VectorBytes> Done(NumElements, false);
3973 unsigned NumConstants = 0;
3974 for (unsigned I = 0; I < NumElements; ++I) {
3975 SDValue Elem = Elems[I];
3976 if (Elem.getOpcode() == ISD::Constant ||
3977 Elem.getOpcode() == ISD::ConstantFP) {
3979 Constants[I] = Elem;
3983 // If there was at least one constant, fill in the other elements of
3984 // Constants with undefs to get a full vector constant and use that
3985 // as the starting point.
3987 if (NumConstants > 0) {
3988 for (unsigned I = 0; I < NumElements; ++I)
3989 if (!Constants[I].getNode())
3990 Constants[I] = DAG.getUNDEF(Elems[I].getValueType());
3991 Result = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Constants);
3993 // Otherwise try to use VLVGP to start the sequence in order to
3994 // avoid a false dependency on any previous contents of the vector
3995 // register. This only makes sense if one of the associated elements
3997 unsigned I1 = NumElements / 2 - 1;
3998 unsigned I2 = NumElements - 1;
3999 bool Def1 = (Elems[I1].getOpcode() != ISD::UNDEF);
4000 bool Def2 = (Elems[I2].getOpcode() != ISD::UNDEF);
4002 SDValue Elem1 = Elems[Def1 ? I1 : I2];
4003 SDValue Elem2 = Elems[Def2 ? I2 : I1];
4004 Result = DAG.getNode(ISD::BITCAST, DL, VT,
4005 joinDwords(DAG, DL, Elem1, Elem2));
4009 Result = DAG.getUNDEF(VT);
4012 // Use VLVGx to insert the other elements.
4013 for (unsigned I = 0; I < NumElements; ++I)
4014 if (!Done[I] && Elems[I].getOpcode() != ISD::UNDEF)
4015 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Result, Elems[I],
4016 DAG.getConstant(I, DL, MVT::i32));
4020 SDValue SystemZTargetLowering::lowerBUILD_VECTOR(SDValue Op,
4021 SelectionDAG &DAG) const {
4022 const SystemZInstrInfo *TII =
4023 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
4024 auto *BVN = cast<BuildVectorSDNode>(Op.getNode());
4026 EVT VT = Op.getValueType();
4028 if (BVN->isConstant()) {
4029 // Try using VECTOR GENERATE BYTE MASK. This is the architecturally-
4030 // preferred way of creating all-zero and all-one vectors so give it
4031 // priority over other methods below.
4033 if (tryBuildVectorByteMask(BVN, Mask)) {
4034 SDValue Op = DAG.getNode(SystemZISD::BYTE_MASK, DL, MVT::v16i8,
4035 DAG.getConstant(Mask, DL, MVT::i32));
4036 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
4039 // Try using some form of replication.
4040 APInt SplatBits, SplatUndef;
4041 unsigned SplatBitSize;
4043 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs,
4045 SplatBitSize <= 64) {
4046 // First try assuming that any undefined bits above the highest set bit
4047 // and below the lowest set bit are 1s. This increases the likelihood of
4048 // being able to use a sign-extended element value in VECTOR REPLICATE
4049 // IMMEDIATE or a wraparound mask in VECTOR GENERATE MASK.
4050 uint64_t SplatBitsZ = SplatBits.getZExtValue();
4051 uint64_t SplatUndefZ = SplatUndef.getZExtValue();
4052 uint64_t Lower = (SplatUndefZ
4053 & ((uint64_t(1) << findFirstSet(SplatBitsZ)) - 1));
4054 uint64_t Upper = (SplatUndefZ
4055 & ~((uint64_t(1) << findLastSet(SplatBitsZ)) - 1));
4056 uint64_t Value = SplatBitsZ | Upper | Lower;
4057 SDValue Op = tryBuildVectorReplicate(DAG, TII, DL, VT, Value,
4062 // Now try assuming that any undefined bits between the first and
4063 // last defined set bits are set. This increases the chances of
4064 // using a non-wraparound mask.
4065 uint64_t Middle = SplatUndefZ & ~Upper & ~Lower;
4066 Value = SplatBitsZ | Middle;
4067 Op = tryBuildVectorReplicate(DAG, TII, DL, VT, Value, SplatBitSize);
4072 // Fall back to loading it from memory.
4076 // See if we should use shuffles to construct the vector from other vectors.
4077 SDValue Res = tryBuildVectorShuffle(DAG, BVN);
4081 // Detect SCALAR_TO_VECTOR conversions.
4082 if (isOperationLegal(ISD::SCALAR_TO_VECTOR, VT) && isScalarToVector(Op))
4083 return buildScalarToVector(DAG, DL, VT, Op.getOperand(0));
4085 // Otherwise use buildVector to build the vector up from GPRs.
4086 unsigned NumElements = Op.getNumOperands();
4087 SmallVector<SDValue, SystemZ::VectorBytes> Ops(NumElements);
4088 for (unsigned I = 0; I < NumElements; ++I)
4089 Ops[I] = Op.getOperand(I);
4090 return buildVector(DAG, DL, VT, Ops);
4093 SDValue SystemZTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
4094 SelectionDAG &DAG) const {
4095 auto *VSN = cast<ShuffleVectorSDNode>(Op.getNode());
4097 EVT VT = Op.getValueType();
4098 unsigned NumElements = VT.getVectorNumElements();
4100 if (VSN->isSplat()) {
4101 SDValue Op0 = Op.getOperand(0);
4102 unsigned Index = VSN->getSplatIndex();
4103 assert(Index < VT.getVectorNumElements() &&
4104 "Splat index should be defined and in first operand");
4105 // See whether the value we're splatting is directly available as a scalar.
4106 if ((Index == 0 && Op0.getOpcode() == ISD::SCALAR_TO_VECTOR) ||
4107 Op0.getOpcode() == ISD::BUILD_VECTOR)
4108 return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op0.getOperand(Index));
4109 // Otherwise keep it as a vector-to-vector operation.
4110 return DAG.getNode(SystemZISD::SPLAT, DL, VT, Op.getOperand(0),
4111 DAG.getConstant(Index, DL, MVT::i32));
4114 GeneralShuffle GS(VT);
4115 for (unsigned I = 0; I < NumElements; ++I) {
4116 int Elt = VSN->getMaskElt(I);
4120 GS.add(Op.getOperand(unsigned(Elt) / NumElements),
4121 unsigned(Elt) % NumElements);
4123 return GS.getNode(DAG, SDLoc(VSN));
4126 SDValue SystemZTargetLowering::lowerSCALAR_TO_VECTOR(SDValue Op,
4127 SelectionDAG &DAG) const {
4129 // Just insert the scalar into element 0 of an undefined vector.
4130 return DAG.getNode(ISD::INSERT_VECTOR_ELT, DL,
4131 Op.getValueType(), DAG.getUNDEF(Op.getValueType()),
4132 Op.getOperand(0), DAG.getConstant(0, DL, MVT::i32));
4135 SDValue SystemZTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
4136 SelectionDAG &DAG) const {
4137 // Handle insertions of floating-point values.
4139 SDValue Op0 = Op.getOperand(0);
4140 SDValue Op1 = Op.getOperand(1);
4141 SDValue Op2 = Op.getOperand(2);
4142 EVT VT = Op.getValueType();
4144 // Insertions into constant indices of a v2f64 can be done using VPDI.
4145 // However, if the inserted value is a bitcast or a constant then it's
4146 // better to use GPRs, as below.
4147 if (VT == MVT::v2f64 &&
4148 Op1.getOpcode() != ISD::BITCAST &&
4149 Op1.getOpcode() != ISD::ConstantFP &&
4150 Op2.getOpcode() == ISD::Constant) {
4151 uint64_t Index = dyn_cast<ConstantSDNode>(Op2)->getZExtValue();
4152 unsigned Mask = VT.getVectorNumElements() - 1;
4157 // Otherwise bitcast to the equivalent integer form and insert via a GPR.
4158 MVT IntVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
4159 MVT IntVecVT = MVT::getVectorVT(IntVT, VT.getVectorNumElements());
4160 SDValue Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntVecVT,
4161 DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0),
4162 DAG.getNode(ISD::BITCAST, DL, IntVT, Op1), Op2);
4163 return DAG.getNode(ISD::BITCAST, DL, VT, Res);
4167 SystemZTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op,
4168 SelectionDAG &DAG) const {
4169 // Handle extractions of floating-point values.
4171 SDValue Op0 = Op.getOperand(0);
4172 SDValue Op1 = Op.getOperand(1);
4173 EVT VT = Op.getValueType();
4174 EVT VecVT = Op0.getValueType();
4176 // Extractions of constant indices can be done directly.
4177 if (auto *CIndexN = dyn_cast<ConstantSDNode>(Op1)) {
4178 uint64_t Index = CIndexN->getZExtValue();
4179 unsigned Mask = VecVT.getVectorNumElements() - 1;
4184 // Otherwise bitcast to the equivalent integer form and extract via a GPR.
4185 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits());
4186 MVT IntVecVT = MVT::getVectorVT(IntVT, VecVT.getVectorNumElements());
4187 SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, IntVT,
4188 DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0), Op1);
4189 return DAG.getNode(ISD::BITCAST, DL, VT, Res);
4193 SystemZTargetLowering::lowerExtendVectorInreg(SDValue Op, SelectionDAG &DAG,
4194 unsigned UnpackHigh) const {
4195 SDValue PackedOp = Op.getOperand(0);
4196 EVT OutVT = Op.getValueType();
4197 EVT InVT = PackedOp.getValueType();
4198 unsigned ToBits = OutVT.getVectorElementType().getSizeInBits();
4199 unsigned FromBits = InVT.getVectorElementType().getSizeInBits();
4202 EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(FromBits),
4203 SystemZ::VectorBits / FromBits);
4204 PackedOp = DAG.getNode(UnpackHigh, SDLoc(PackedOp), OutVT, PackedOp);
4205 } while (FromBits != ToBits);
4209 SDValue SystemZTargetLowering::lowerShift(SDValue Op, SelectionDAG &DAG,
4210 unsigned ByScalar) const {
4211 // Look for cases where a vector shift can use the *_BY_SCALAR form.
4212 SDValue Op0 = Op.getOperand(0);
4213 SDValue Op1 = Op.getOperand(1);
4215 EVT VT = Op.getValueType();
4216 unsigned ElemBitSize = VT.getVectorElementType().getSizeInBits();
4218 // See whether the shift vector is a splat represented as BUILD_VECTOR.
4219 if (auto *BVN = dyn_cast<BuildVectorSDNode>(Op1)) {
4220 APInt SplatBits, SplatUndef;
4221 unsigned SplatBitSize;
4223 // Check for constant splats. Use ElemBitSize as the minimum element
4224 // width and reject splats that need wider elements.
4225 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs,
4226 ElemBitSize, true) &&
4227 SplatBitSize == ElemBitSize) {
4228 SDValue Shift = DAG.getConstant(SplatBits.getZExtValue() & 0xfff,
4230 return DAG.getNode(ByScalar, DL, VT, Op0, Shift);
4232 // Check for variable splats.
4233 BitVector UndefElements;
4234 SDValue Splat = BVN->getSplatValue(&UndefElements);
4236 // Since i32 is the smallest legal type, we either need a no-op
4238 SDValue Shift = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Splat);
4239 return DAG.getNode(ByScalar, DL, VT, Op0, Shift);
4243 // See whether the shift vector is a splat represented as SHUFFLE_VECTOR,
4244 // and the shift amount is directly available in a GPR.
4245 if (auto *VSN = dyn_cast<ShuffleVectorSDNode>(Op1)) {
4246 if (VSN->isSplat()) {
4247 SDValue VSNOp0 = VSN->getOperand(0);
4248 unsigned Index = VSN->getSplatIndex();
4249 assert(Index < VT.getVectorNumElements() &&
4250 "Splat index should be defined and in first operand");
4251 if ((Index == 0 && VSNOp0.getOpcode() == ISD::SCALAR_TO_VECTOR) ||
4252 VSNOp0.getOpcode() == ISD::BUILD_VECTOR) {
4253 // Since i32 is the smallest legal type, we either need a no-op
4255 SDValue Shift = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32,
4256 VSNOp0.getOperand(Index));
4257 return DAG.getNode(ByScalar, DL, VT, Op0, Shift);
4262 // Otherwise just treat the current form as legal.
4266 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
4267 SelectionDAG &DAG) const {
4268 switch (Op.getOpcode()) {
4270 return lowerBR_CC(Op, DAG);
4271 case ISD::SELECT_CC:
4272 return lowerSELECT_CC(Op, DAG);
4274 return lowerSETCC(Op, DAG);
4275 case ISD::GlobalAddress:
4276 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
4277 case ISD::GlobalTLSAddress:
4278 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
4279 case ISD::BlockAddress:
4280 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
4281 case ISD::JumpTable:
4282 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
4283 case ISD::ConstantPool:
4284 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
4286 return lowerBITCAST(Op, DAG);
4288 return lowerVASTART(Op, DAG);
4290 return lowerVACOPY(Op, DAG);
4291 case ISD::DYNAMIC_STACKALLOC:
4292 return lowerDYNAMIC_STACKALLOC(Op, DAG);
4293 case ISD::SMUL_LOHI:
4294 return lowerSMUL_LOHI(Op, DAG);
4295 case ISD::UMUL_LOHI:
4296 return lowerUMUL_LOHI(Op, DAG);
4298 return lowerSDIVREM(Op, DAG);
4300 return lowerUDIVREM(Op, DAG);
4302 return lowerOR(Op, DAG);
4304 return lowerCTPOP(Op, DAG);
4305 case ISD::CTLZ_ZERO_UNDEF:
4306 return DAG.getNode(ISD::CTLZ, SDLoc(Op),
4307 Op.getValueType(), Op.getOperand(0));
4308 case ISD::CTTZ_ZERO_UNDEF:
4309 return DAG.getNode(ISD::CTTZ, SDLoc(Op),
4310 Op.getValueType(), Op.getOperand(0));
4311 case ISD::ATOMIC_SWAP:
4312 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
4313 case ISD::ATOMIC_STORE:
4314 return lowerATOMIC_STORE(Op, DAG);
4315 case ISD::ATOMIC_LOAD:
4316 return lowerATOMIC_LOAD(Op, DAG);
4317 case ISD::ATOMIC_LOAD_ADD:
4318 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
4319 case ISD::ATOMIC_LOAD_SUB:
4320 return lowerATOMIC_LOAD_SUB(Op, DAG);
4321 case ISD::ATOMIC_LOAD_AND:
4322 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
4323 case ISD::ATOMIC_LOAD_OR:
4324 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
4325 case ISD::ATOMIC_LOAD_XOR:
4326 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
4327 case ISD::ATOMIC_LOAD_NAND:
4328 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
4329 case ISD::ATOMIC_LOAD_MIN:
4330 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
4331 case ISD::ATOMIC_LOAD_MAX:
4332 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
4333 case ISD::ATOMIC_LOAD_UMIN:
4334 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
4335 case ISD::ATOMIC_LOAD_UMAX:
4336 return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
4337 case ISD::ATOMIC_CMP_SWAP:
4338 return lowerATOMIC_CMP_SWAP(Op, DAG);
4339 case ISD::STACKSAVE:
4340 return lowerSTACKSAVE(Op, DAG);
4341 case ISD::STACKRESTORE:
4342 return lowerSTACKRESTORE(Op, DAG);
4344 return lowerPREFETCH(Op, DAG);
4345 case ISD::INTRINSIC_W_CHAIN:
4346 return lowerINTRINSIC_W_CHAIN(Op, DAG);
4347 case ISD::INTRINSIC_WO_CHAIN:
4348 return lowerINTRINSIC_WO_CHAIN(Op, DAG);
4349 case ISD::BUILD_VECTOR:
4350 return lowerBUILD_VECTOR(Op, DAG);
4351 case ISD::VECTOR_SHUFFLE:
4352 return lowerVECTOR_SHUFFLE(Op, DAG);
4353 case ISD::SCALAR_TO_VECTOR:
4354 return lowerSCALAR_TO_VECTOR(Op, DAG);
4355 case ISD::INSERT_VECTOR_ELT:
4356 return lowerINSERT_VECTOR_ELT(Op, DAG);
4357 case ISD::EXTRACT_VECTOR_ELT:
4358 return lowerEXTRACT_VECTOR_ELT(Op, DAG);
4359 case ISD::SIGN_EXTEND_VECTOR_INREG:
4360 return lowerExtendVectorInreg(Op, DAG, SystemZISD::UNPACK_HIGH);
4361 case ISD::ZERO_EXTEND_VECTOR_INREG:
4362 return lowerExtendVectorInreg(Op, DAG, SystemZISD::UNPACKL_HIGH);
4364 return lowerShift(Op, DAG, SystemZISD::VSHL_BY_SCALAR);
4366 return lowerShift(Op, DAG, SystemZISD::VSRL_BY_SCALAR);
4368 return lowerShift(Op, DAG, SystemZISD::VSRA_BY_SCALAR);
4370 llvm_unreachable("Unexpected node to lower");
4374 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
4375 #define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
4376 switch ((SystemZISD::NodeType)Opcode) {
4377 case SystemZISD::FIRST_NUMBER: break;
4383 OPCODE(PCREL_WRAPPER);
4384 OPCODE(PCREL_OFFSET);
4390 OPCODE(SELECT_CCMASK);
4391 OPCODE(ADJDYNALLOC);
4392 OPCODE(EXTRACT_ACCESS);
4394 OPCODE(UMUL_LOHI64);
4411 OPCODE(SEARCH_STRING);
4415 OPCODE(TBEGIN_NOFLOAT);
4418 OPCODE(ROTATE_MASK);
4420 OPCODE(JOIN_DWORDS);
4425 OPCODE(PERMUTE_DWORDS);
4430 OPCODE(UNPACK_HIGH);
4431 OPCODE(UNPACKL_HIGH);
4433 OPCODE(UNPACKL_LOW);
4434 OPCODE(VSHL_BY_SCALAR);
4435 OPCODE(VSRL_BY_SCALAR);
4436 OPCODE(VSRA_BY_SCALAR);
4463 OPCODE(ATOMIC_SWAPW);
4464 OPCODE(ATOMIC_LOADW_ADD);
4465 OPCODE(ATOMIC_LOADW_SUB);
4466 OPCODE(ATOMIC_LOADW_AND);
4467 OPCODE(ATOMIC_LOADW_OR);
4468 OPCODE(ATOMIC_LOADW_XOR);
4469 OPCODE(ATOMIC_LOADW_NAND);
4470 OPCODE(ATOMIC_LOADW_MIN);
4471 OPCODE(ATOMIC_LOADW_MAX);
4472 OPCODE(ATOMIC_LOADW_UMIN);
4473 OPCODE(ATOMIC_LOADW_UMAX);
4474 OPCODE(ATOMIC_CMP_SWAPW);
4481 // Return true if VT is a vector whose elements are a whole number of bytes
4483 static bool canTreatAsByteVector(EVT VT) {
4484 return VT.isVector() && VT.getVectorElementType().getSizeInBits() % 8 == 0;
4487 // Try to simplify an EXTRACT_VECTOR_ELT from a vector of type VecVT
4488 // producing a result of type ResVT. Op is a possibly bitcast version
4489 // of the input vector and Index is the index (based on type VecVT) that
4490 // should be extracted. Return the new extraction if a simplification
4491 // was possible or if Force is true.
4492 SDValue SystemZTargetLowering::combineExtract(SDLoc DL, EVT ResVT, EVT VecVT,
4493 SDValue Op, unsigned Index,
4494 DAGCombinerInfo &DCI,
4496 SelectionDAG &DAG = DCI.DAG;
4498 // The number of bytes being extracted.
4499 unsigned BytesPerElement = VecVT.getVectorElementType().getStoreSize();
4502 unsigned Opcode = Op.getOpcode();
4503 if (Opcode == ISD::BITCAST)
4504 // Look through bitcasts.
4505 Op = Op.getOperand(0);
4506 else if (Opcode == ISD::VECTOR_SHUFFLE &&
4507 canTreatAsByteVector(Op.getValueType())) {
4508 // Get a VPERM-like permute mask and see whether the bytes covered
4509 // by the extracted element are a contiguous sequence from one
4511 SmallVector<int, SystemZ::VectorBytes> Bytes;
4512 getVPermMask(cast<ShuffleVectorSDNode>(Op), Bytes);
4514 if (!getShuffleInput(Bytes, Index * BytesPerElement,
4515 BytesPerElement, First))
4518 return DAG.getUNDEF(ResVT);
4519 // Make sure the contiguous sequence starts at a multiple of the
4520 // original element size.
4521 unsigned Byte = unsigned(First) % Bytes.size();
4522 if (Byte % BytesPerElement != 0)
4524 // We can get the extracted value directly from an input.
4525 Index = Byte / BytesPerElement;
4526 Op = Op.getOperand(unsigned(First) / Bytes.size());
4528 } else if (Opcode == ISD::BUILD_VECTOR &&
4529 canTreatAsByteVector(Op.getValueType())) {
4530 // We can only optimize this case if the BUILD_VECTOR elements are
4531 // at least as wide as the extracted value.
4532 EVT OpVT = Op.getValueType();
4533 unsigned OpBytesPerElement = OpVT.getVectorElementType().getStoreSize();
4534 if (OpBytesPerElement < BytesPerElement)
4536 // Make sure that the least-significant bit of the extracted value
4537 // is the least significant bit of an input.
4538 unsigned End = (Index + 1) * BytesPerElement;
4539 if (End % OpBytesPerElement != 0)
4541 // We're extracting the low part of one operand of the BUILD_VECTOR.
4542 Op = Op.getOperand(End / OpBytesPerElement - 1);
4543 if (!Op.getValueType().isInteger()) {
4544 EVT VT = MVT::getIntegerVT(Op.getValueType().getSizeInBits());
4545 Op = DAG.getNode(ISD::BITCAST, DL, VT, Op);
4546 DCI.AddToWorklist(Op.getNode());
4548 EVT VT = MVT::getIntegerVT(ResVT.getSizeInBits());
4549 Op = DAG.getNode(ISD::TRUNCATE, DL, VT, Op);
4551 DCI.AddToWorklist(Op.getNode());
4552 Op = DAG.getNode(ISD::BITCAST, DL, ResVT, Op);
4555 } else if ((Opcode == ISD::SIGN_EXTEND_VECTOR_INREG ||
4556 Opcode == ISD::ZERO_EXTEND_VECTOR_INREG ||
4557 Opcode == ISD::ANY_EXTEND_VECTOR_INREG) &&
4558 canTreatAsByteVector(Op.getValueType()) &&
4559 canTreatAsByteVector(Op.getOperand(0).getValueType())) {
4560 // Make sure that only the unextended bits are significant.
4561 EVT ExtVT = Op.getValueType();
4562 EVT OpVT = Op.getOperand(0).getValueType();
4563 unsigned ExtBytesPerElement = ExtVT.getVectorElementType().getStoreSize();
4564 unsigned OpBytesPerElement = OpVT.getVectorElementType().getStoreSize();
4565 unsigned Byte = Index * BytesPerElement;
4566 unsigned SubByte = Byte % ExtBytesPerElement;
4567 unsigned MinSubByte = ExtBytesPerElement - OpBytesPerElement;
4568 if (SubByte < MinSubByte ||
4569 SubByte + BytesPerElement > ExtBytesPerElement)
4571 // Get the byte offset of the unextended element
4572 Byte = Byte / ExtBytesPerElement * OpBytesPerElement;
4573 // ...then add the byte offset relative to that element.
4574 Byte += SubByte - MinSubByte;
4575 if (Byte % BytesPerElement != 0)
4577 Op = Op.getOperand(0);
4578 Index = Byte / BytesPerElement;
4584 if (Op.getValueType() != VecVT) {
4585 Op = DAG.getNode(ISD::BITCAST, DL, VecVT, Op);
4586 DCI.AddToWorklist(Op.getNode());
4588 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Op,
4589 DAG.getConstant(Index, DL, MVT::i32));
4594 // Optimize vector operations in scalar value Op on the basis that Op
4595 // is truncated to TruncVT.
4597 SystemZTargetLowering::combineTruncateExtract(SDLoc DL, EVT TruncVT, SDValue Op,
4598 DAGCombinerInfo &DCI) const {
4599 // If we have (trunc (extract_vector_elt X, Y)), try to turn it into
4600 // (extract_vector_elt (bitcast X), Y'), where (bitcast X) has elements
4602 if (Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4603 TruncVT.getSizeInBits() % 8 == 0) {
4604 SDValue Vec = Op.getOperand(0);
4605 EVT VecVT = Vec.getValueType();
4606 if (canTreatAsByteVector(VecVT)) {
4607 if (auto *IndexN = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
4608 unsigned BytesPerElement = VecVT.getVectorElementType().getStoreSize();
4609 unsigned TruncBytes = TruncVT.getStoreSize();
4610 if (BytesPerElement % TruncBytes == 0) {
4611 // Calculate the value of Y' in the above description. We are
4612 // splitting the original elements into Scale equal-sized pieces
4613 // and for truncation purposes want the last (least-significant)
4614 // of these pieces for IndexN. This is easiest to do by calculating
4615 // the start index of the following element and then subtracting 1.
4616 unsigned Scale = BytesPerElement / TruncBytes;
4617 unsigned NewIndex = (IndexN->getZExtValue() + 1) * Scale - 1;
4619 // Defer the creation of the bitcast from X to combineExtract,
4620 // which might be able to optimize the extraction.
4621 VecVT = MVT::getVectorVT(MVT::getIntegerVT(TruncBytes * 8),
4622 VecVT.getStoreSize() / TruncBytes);
4623 EVT ResVT = (TruncBytes < 4 ? MVT::i32 : TruncVT);
4624 return combineExtract(DL, ResVT, VecVT, Vec, NewIndex, DCI, true);
4632 SDValue SystemZTargetLowering::PerformDAGCombine(SDNode *N,
4633 DAGCombinerInfo &DCI) const {
4634 SelectionDAG &DAG = DCI.DAG;
4635 unsigned Opcode = N->getOpcode();
4636 if (Opcode == ISD::SIGN_EXTEND) {
4637 // Convert (sext (ashr (shl X, C1), C2)) to
4638 // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
4639 // cheap as narrower ones.
4640 SDValue N0 = N->getOperand(0);
4641 EVT VT = N->getValueType(0);
4642 if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
4643 auto *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4644 SDValue Inner = N0.getOperand(0);
4645 if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
4646 if (auto *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1))) {
4647 unsigned Extra = (VT.getSizeInBits() -
4648 N0.getValueType().getSizeInBits());
4649 unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
4650 unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
4651 EVT ShiftVT = N0.getOperand(1).getValueType();
4652 SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
4653 Inner.getOperand(0));
4654 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
4655 DAG.getConstant(NewShlAmt, SDLoc(Inner),
4657 return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
4658 DAG.getConstant(NewSraAmt, SDLoc(N0), ShiftVT));
4663 if (Opcode == SystemZISD::MERGE_HIGH ||
4664 Opcode == SystemZISD::MERGE_LOW) {
4665 SDValue Op0 = N->getOperand(0);
4666 SDValue Op1 = N->getOperand(1);
4667 if (Op0.getOpcode() == ISD::BITCAST)
4668 Op0 = Op0.getOperand(0);
4669 if (Op0.getOpcode() == SystemZISD::BYTE_MASK &&
4670 cast<ConstantSDNode>(Op0.getOperand(0))->getZExtValue() == 0) {
4671 // (z_merge_* 0, 0) -> 0. This is mostly useful for using VLLEZF
4673 if (Op1 == N->getOperand(0))
4675 // (z_merge_? 0, X) -> (z_unpackl_? 0, X).
4676 EVT VT = Op1.getValueType();
4677 unsigned ElemBytes = VT.getVectorElementType().getStoreSize();
4678 if (ElemBytes <= 4) {
4679 Opcode = (Opcode == SystemZISD::MERGE_HIGH ?
4680 SystemZISD::UNPACKL_HIGH : SystemZISD::UNPACKL_LOW);
4681 EVT InVT = VT.changeVectorElementTypeToInteger();
4682 EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(ElemBytes * 16),
4683 SystemZ::VectorBytes / ElemBytes / 2);
4685 Op1 = DAG.getNode(ISD::BITCAST, SDLoc(N), InVT, Op1);
4686 DCI.AddToWorklist(Op1.getNode());
4688 SDValue Op = DAG.getNode(Opcode, SDLoc(N), OutVT, Op1);
4689 DCI.AddToWorklist(Op.getNode());
4690 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
4694 // If we have (truncstoreiN (extract_vector_elt X, Y), Z) then it is better
4695 // for the extraction to be done on a vMiN value, so that we can use VSTE.
4696 // If X has wider elements then convert it to:
4697 // (truncstoreiN (extract_vector_elt (bitcast X), Y2), Z).
4698 if (Opcode == ISD::STORE) {
4699 auto *SN = cast<StoreSDNode>(N);
4700 EVT MemVT = SN->getMemoryVT();
4701 if (MemVT.isInteger()) {
4702 SDValue Value = combineTruncateExtract(SDLoc(N), MemVT,
4703 SN->getValue(), DCI);
4704 if (Value.getNode()) {
4705 DCI.AddToWorklist(Value.getNode());
4707 // Rewrite the store with the new form of stored value.
4708 return DAG.getTruncStore(SN->getChain(), SDLoc(SN), Value,
4709 SN->getBasePtr(), SN->getMemoryVT(),
4710 SN->getMemOperand());
4714 // Try to simplify a vector extraction.
4715 if (Opcode == ISD::EXTRACT_VECTOR_ELT) {
4716 if (auto *IndexN = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
4717 SDValue Op0 = N->getOperand(0);
4718 EVT VecVT = Op0.getValueType();
4719 return combineExtract(SDLoc(N), N->getValueType(0), VecVT, Op0,
4720 IndexN->getZExtValue(), DCI, false);
4723 // (join_dwords X, X) == (replicate X)
4724 if (Opcode == SystemZISD::JOIN_DWORDS &&
4725 N->getOperand(0) == N->getOperand(1))
4726 return DAG.getNode(SystemZISD::REPLICATE, SDLoc(N), N->getValueType(0),
4728 // (fround (extract_vector_elt X 0))
4729 // (fround (extract_vector_elt X 1)) ->
4730 // (extract_vector_elt (VROUND X) 0)
4731 // (extract_vector_elt (VROUND X) 1)
4733 // This is a special case since the target doesn't really support v2f32s.
4734 if (Opcode == ISD::FP_ROUND) {
4735 SDValue Op0 = N->getOperand(0);
4736 if (N->getValueType(0) == MVT::f32 &&
4738 Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4739 Op0.getOperand(0).getValueType() == MVT::v2f64 &&
4740 Op0.getOperand(1).getOpcode() == ISD::Constant &&
4741 cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue() == 0) {
4742 SDValue Vec = Op0.getOperand(0);
4743 for (auto *U : Vec->uses()) {
4744 if (U != Op0.getNode() &&
4746 U->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4747 U->getOperand(0) == Vec &&
4748 U->getOperand(1).getOpcode() == ISD::Constant &&
4749 cast<ConstantSDNode>(U->getOperand(1))->getZExtValue() == 1) {
4750 SDValue OtherRound = SDValue(*U->use_begin(), 0);
4751 if (OtherRound.getOpcode() == ISD::FP_ROUND &&
4752 OtherRound.getOperand(0) == SDValue(U, 0) &&
4753 OtherRound.getValueType() == MVT::f32) {
4754 SDValue VRound = DAG.getNode(SystemZISD::VROUND, SDLoc(N),
4756 DCI.AddToWorklist(VRound.getNode());
4758 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(U), MVT::f32,
4759 VRound, DAG.getConstant(2, SDLoc(U), MVT::i32));
4760 DCI.AddToWorklist(Extract1.getNode());
4761 DAG.ReplaceAllUsesOfValueWith(OtherRound, Extract1);
4763 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op0), MVT::f32,
4764 VRound, DAG.getConstant(0, SDLoc(Op0), MVT::i32));
4774 //===----------------------------------------------------------------------===//
4776 //===----------------------------------------------------------------------===//
4778 // Create a new basic block after MBB.
4779 static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
4780 MachineFunction &MF = *MBB->getParent();
4781 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
4782 MF.insert(std::next(MachineFunction::iterator(MBB)), NewMBB);
4786 // Split MBB after MI and return the new block (the one that contains
4787 // instructions after MI).
4788 static MachineBasicBlock *splitBlockAfter(MachineInstr *MI,
4789 MachineBasicBlock *MBB) {
4790 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
4791 NewMBB->splice(NewMBB->begin(), MBB,
4792 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
4793 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
4797 // Split MBB before MI and return the new block (the one that contains MI).
4798 static MachineBasicBlock *splitBlockBefore(MachineInstr *MI,
4799 MachineBasicBlock *MBB) {
4800 MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
4801 NewMBB->splice(NewMBB->begin(), MBB, MI, MBB->end());
4802 NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
4806 // Force base value Base into a register before MI. Return the register.
4807 static unsigned forceReg(MachineInstr *MI, MachineOperand &Base,
4808 const SystemZInstrInfo *TII) {
4810 return Base.getReg();
4812 MachineBasicBlock *MBB = MI->getParent();
4813 MachineFunction &MF = *MBB->getParent();
4814 MachineRegisterInfo &MRI = MF.getRegInfo();
4816 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
4817 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LA), Reg)
4818 .addOperand(Base).addImm(0).addReg(0);
4822 // Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI.
4824 SystemZTargetLowering::emitSelect(MachineInstr *MI,
4825 MachineBasicBlock *MBB) const {
4826 const SystemZInstrInfo *TII =
4827 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
4829 unsigned DestReg = MI->getOperand(0).getReg();
4830 unsigned TrueReg = MI->getOperand(1).getReg();
4831 unsigned FalseReg = MI->getOperand(2).getReg();
4832 unsigned CCValid = MI->getOperand(3).getImm();
4833 unsigned CCMask = MI->getOperand(4).getImm();
4834 DebugLoc DL = MI->getDebugLoc();
4836 MachineBasicBlock *StartMBB = MBB;
4837 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
4838 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
4841 // BRC CCMask, JoinMBB
4842 // # fallthrough to FalseMBB
4844 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
4845 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
4846 MBB->addSuccessor(JoinMBB);
4847 MBB->addSuccessor(FalseMBB);
4850 // # fallthrough to JoinMBB
4852 MBB->addSuccessor(JoinMBB);
4855 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ]
4858 BuildMI(*MBB, MI, DL, TII->get(SystemZ::PHI), DestReg)
4859 .addReg(TrueReg).addMBB(StartMBB)
4860 .addReg(FalseReg).addMBB(FalseMBB);
4862 MI->eraseFromParent();
4866 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
4867 // StoreOpcode is the store to use and Invert says whether the store should
4868 // happen when the condition is false rather than true. If a STORE ON
4869 // CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
4871 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
4872 MachineBasicBlock *MBB,
4873 unsigned StoreOpcode, unsigned STOCOpcode,
4874 bool Invert) const {
4875 const SystemZInstrInfo *TII =
4876 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
4878 unsigned SrcReg = MI->getOperand(0).getReg();
4879 MachineOperand Base = MI->getOperand(1);
4880 int64_t Disp = MI->getOperand(2).getImm();
4881 unsigned IndexReg = MI->getOperand(3).getReg();
4882 unsigned CCValid = MI->getOperand(4).getImm();
4883 unsigned CCMask = MI->getOperand(5).getImm();
4884 DebugLoc DL = MI->getDebugLoc();
4886 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
4888 // Use STOCOpcode if possible. We could use different store patterns in
4889 // order to avoid matching the index register, but the performance trade-offs
4890 // might be more complicated in that case.
4891 if (STOCOpcode && !IndexReg && Subtarget.hasLoadStoreOnCond()) {
4894 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
4895 .addReg(SrcReg).addOperand(Base).addImm(Disp)
4896 .addImm(CCValid).addImm(CCMask);
4897 MI->eraseFromParent();
4901 // Get the condition needed to branch around the store.
4905 MachineBasicBlock *StartMBB = MBB;
4906 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB);
4907 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB);
4910 // BRC CCMask, JoinMBB
4911 // # fallthrough to FalseMBB
4913 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
4914 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB);
4915 MBB->addSuccessor(JoinMBB);
4916 MBB->addSuccessor(FalseMBB);
4919 // store %SrcReg, %Disp(%Index,%Base)
4920 // # fallthrough to JoinMBB
4922 BuildMI(MBB, DL, TII->get(StoreOpcode))
4923 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg);
4924 MBB->addSuccessor(JoinMBB);
4926 MI->eraseFromParent();
4930 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_*
4931 // or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that
4932 // performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}.
4933 // BitSize is the width of the field in bits, or 0 if this is a partword
4934 // ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize
4935 // is one of the operands. Invert says whether the field should be
4936 // inverted after performing BinOpcode (e.g. for NAND).
4938 SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI,
4939 MachineBasicBlock *MBB,
4942 bool Invert) const {
4943 MachineFunction &MF = *MBB->getParent();
4944 const SystemZInstrInfo *TII =
4945 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
4946 MachineRegisterInfo &MRI = MF.getRegInfo();
4947 bool IsSubWord = (BitSize < 32);
4949 // Extract the operands. Base can be a register or a frame index.
4950 // Src2 can be a register or immediate.
4951 unsigned Dest = MI->getOperand(0).getReg();
4952 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
4953 int64_t Disp = MI->getOperand(2).getImm();
4954 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3));
4955 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
4956 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
4957 DebugLoc DL = MI->getDebugLoc();
4959 BitSize = MI->getOperand(6).getImm();
4961 // Subword operations use 32-bit registers.
4962 const TargetRegisterClass *RC = (BitSize <= 32 ?
4963 &SystemZ::GR32BitRegClass :
4964 &SystemZ::GR64BitRegClass);
4965 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
4966 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
4968 // Get the right opcodes for the displacement.
4969 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
4970 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
4971 assert(LOpcode && CSOpcode && "Displacement out of range");
4973 // Create virtual registers for temporary results.
4974 unsigned OrigVal = MRI.createVirtualRegister(RC);
4975 unsigned OldVal = MRI.createVirtualRegister(RC);
4976 unsigned NewVal = (BinOpcode || IsSubWord ?
4977 MRI.createVirtualRegister(RC) : Src2.getReg());
4978 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
4979 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
4981 // Insert a basic block for the main loop.
4982 MachineBasicBlock *StartMBB = MBB;
4983 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
4984 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
4988 // %OrigVal = L Disp(%Base)
4989 // # fall through to LoopMMB
4991 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
4992 .addOperand(Base).addImm(Disp).addReg(0);
4993 MBB->addSuccessor(LoopMBB);
4996 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ]
4997 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
4998 // %RotatedNewVal = OP %RotatedOldVal, %Src2
4999 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
5000 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
5002 // # fall through to DoneMMB
5004 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
5005 .addReg(OrigVal).addMBB(StartMBB)
5006 .addReg(Dest).addMBB(LoopMBB);
5008 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
5009 .addReg(OldVal).addReg(BitShift).addImm(0);
5011 // Perform the operation normally and then invert every bit of the field.
5012 unsigned Tmp = MRI.createVirtualRegister(RC);
5013 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
5014 .addReg(RotatedOldVal).addOperand(Src2);
5016 // XILF with the upper BitSize bits set.
5017 BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
5018 .addReg(Tmp).addImm(-1U << (32 - BitSize));
5020 // Use LCGR and add -1 to the result, which is more compact than
5021 // an XILF, XILH pair.
5022 unsigned Tmp2 = MRI.createVirtualRegister(RC);
5023 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp);
5024 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal)
5025 .addReg(Tmp2).addImm(-1);
5027 } else if (BinOpcode)
5028 // A simply binary operation.
5029 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal)
5030 .addReg(RotatedOldVal).addOperand(Src2);
5032 // Use RISBG to rotate Src2 into position and use it to replace the
5033 // field in RotatedOldVal.
5034 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal)
5035 .addReg(RotatedOldVal).addReg(Src2.getReg())
5036 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize);
5038 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
5039 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
5040 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
5041 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
5042 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5043 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
5044 MBB->addSuccessor(LoopMBB);
5045 MBB->addSuccessor(DoneMBB);
5047 MI->eraseFromParent();
5051 // Implement EmitInstrWithCustomInserter for pseudo
5052 // ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the
5053 // instruction that should be used to compare the current field with the
5054 // minimum or maximum value. KeepOldMask is the BRC condition-code mask
5055 // for when the current field should be kept. BitSize is the width of
5056 // the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction.
5058 SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI,
5059 MachineBasicBlock *MBB,
5060 unsigned CompareOpcode,
5061 unsigned KeepOldMask,
5062 unsigned BitSize) const {
5063 MachineFunction &MF = *MBB->getParent();
5064 const SystemZInstrInfo *TII =
5065 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
5066 MachineRegisterInfo &MRI = MF.getRegInfo();
5067 bool IsSubWord = (BitSize < 32);
5069 // Extract the operands. Base can be a register or a frame index.
5070 unsigned Dest = MI->getOperand(0).getReg();
5071 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
5072 int64_t Disp = MI->getOperand(2).getImm();
5073 unsigned Src2 = MI->getOperand(3).getReg();
5074 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0);
5075 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0);
5076 DebugLoc DL = MI->getDebugLoc();
5078 BitSize = MI->getOperand(6).getImm();
5080 // Subword operations use 32-bit registers.
5081 const TargetRegisterClass *RC = (BitSize <= 32 ?
5082 &SystemZ::GR32BitRegClass :
5083 &SystemZ::GR64BitRegClass);
5084 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG;
5085 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG;
5087 // Get the right opcodes for the displacement.
5088 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp);
5089 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp);
5090 assert(LOpcode && CSOpcode && "Displacement out of range");
5092 // Create virtual registers for temporary results.
5093 unsigned OrigVal = MRI.createVirtualRegister(RC);
5094 unsigned OldVal = MRI.createVirtualRegister(RC);
5095 unsigned NewVal = MRI.createVirtualRegister(RC);
5096 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal);
5097 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2);
5098 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal);
5100 // Insert 3 basic blocks for the loop.
5101 MachineBasicBlock *StartMBB = MBB;
5102 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
5103 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
5104 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB);
5105 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB);
5109 // %OrigVal = L Disp(%Base)
5110 // # fall through to LoopMMB
5112 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal)
5113 .addOperand(Base).addImm(Disp).addReg(0);
5114 MBB->addSuccessor(LoopMBB);
5117 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ]
5118 // %RotatedOldVal = RLL %OldVal, 0(%BitShift)
5119 // CompareOpcode %RotatedOldVal, %Src2
5120 // BRC KeepOldMask, UpdateMBB
5122 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
5123 .addReg(OrigVal).addMBB(StartMBB)
5124 .addReg(Dest).addMBB(UpdateMBB);
5126 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal)
5127 .addReg(OldVal).addReg(BitShift).addImm(0);
5128 BuildMI(MBB, DL, TII->get(CompareOpcode))
5129 .addReg(RotatedOldVal).addReg(Src2);
5130 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5131 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB);
5132 MBB->addSuccessor(UpdateMBB);
5133 MBB->addSuccessor(UseAltMBB);
5136 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0
5137 // # fall through to UpdateMMB
5140 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal)
5141 .addReg(RotatedOldVal).addReg(Src2)
5142 .addImm(32).addImm(31 + BitSize).addImm(0);
5143 MBB->addSuccessor(UpdateMBB);
5146 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ],
5147 // [ %RotatedAltVal, UseAltMBB ]
5148 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift)
5149 // %Dest = CS %OldVal, %NewVal, Disp(%Base)
5151 // # fall through to DoneMMB
5153 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal)
5154 .addReg(RotatedOldVal).addMBB(LoopMBB)
5155 .addReg(RotatedAltVal).addMBB(UseAltMBB);
5157 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal)
5158 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0);
5159 BuildMI(MBB, DL, TII->get(CSOpcode), Dest)
5160 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp);
5161 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5162 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
5163 MBB->addSuccessor(LoopMBB);
5164 MBB->addSuccessor(DoneMBB);
5166 MI->eraseFromParent();
5170 // Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW
5173 SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
5174 MachineBasicBlock *MBB) const {
5175 MachineFunction &MF = *MBB->getParent();
5176 const SystemZInstrInfo *TII =
5177 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
5178 MachineRegisterInfo &MRI = MF.getRegInfo();
5180 // Extract the operands. Base can be a register or a frame index.
5181 unsigned Dest = MI->getOperand(0).getReg();
5182 MachineOperand Base = earlyUseOperand(MI->getOperand(1));
5183 int64_t Disp = MI->getOperand(2).getImm();
5184 unsigned OrigCmpVal = MI->getOperand(3).getReg();
5185 unsigned OrigSwapVal = MI->getOperand(4).getReg();
5186 unsigned BitShift = MI->getOperand(5).getReg();
5187 unsigned NegBitShift = MI->getOperand(6).getReg();
5188 int64_t BitSize = MI->getOperand(7).getImm();
5189 DebugLoc DL = MI->getDebugLoc();
5191 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass;
5193 // Get the right opcodes for the displacement.
5194 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp);
5195 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp);
5196 assert(LOpcode && CSOpcode && "Displacement out of range");
5198 // Create virtual registers for temporary results.
5199 unsigned OrigOldVal = MRI.createVirtualRegister(RC);
5200 unsigned OldVal = MRI.createVirtualRegister(RC);
5201 unsigned CmpVal = MRI.createVirtualRegister(RC);
5202 unsigned SwapVal = MRI.createVirtualRegister(RC);
5203 unsigned StoreVal = MRI.createVirtualRegister(RC);
5204 unsigned RetryOldVal = MRI.createVirtualRegister(RC);
5205 unsigned RetryCmpVal = MRI.createVirtualRegister(RC);
5206 unsigned RetrySwapVal = MRI.createVirtualRegister(RC);
5208 // Insert 2 basic blocks for the loop.
5209 MachineBasicBlock *StartMBB = MBB;
5210 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
5211 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
5212 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB);
5216 // %OrigOldVal = L Disp(%Base)
5217 // # fall through to LoopMMB
5219 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal)
5220 .addOperand(Base).addImm(Disp).addReg(0);
5221 MBB->addSuccessor(LoopMBB);
5224 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ]
5225 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ]
5226 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ]
5227 // %Dest = RLL %OldVal, BitSize(%BitShift)
5228 // ^^ The low BitSize bits contain the field
5230 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0
5231 // ^^ Replace the upper 32-BitSize bits of the
5232 // comparison value with those that we loaded,
5233 // so that we can use a full word comparison.
5234 // CR %Dest, %RetryCmpVal
5236 // # Fall through to SetMBB
5238 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal)
5239 .addReg(OrigOldVal).addMBB(StartMBB)
5240 .addReg(RetryOldVal).addMBB(SetMBB);
5241 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal)
5242 .addReg(OrigCmpVal).addMBB(StartMBB)
5243 .addReg(RetryCmpVal).addMBB(SetMBB);
5244 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal)
5245 .addReg(OrigSwapVal).addMBB(StartMBB)
5246 .addReg(RetrySwapVal).addMBB(SetMBB);
5247 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest)
5248 .addReg(OldVal).addReg(BitShift).addImm(BitSize);
5249 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal)
5250 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
5251 BuildMI(MBB, DL, TII->get(SystemZ::CR))
5252 .addReg(Dest).addReg(RetryCmpVal);
5253 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5254 .addImm(SystemZ::CCMASK_ICMP)
5255 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB);
5256 MBB->addSuccessor(DoneMBB);
5257 MBB->addSuccessor(SetMBB);
5260 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0
5261 // ^^ Replace the upper 32-BitSize bits of the new
5262 // value with those that we loaded.
5263 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift)
5264 // ^^ Rotate the new field to its proper position.
5265 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base)
5267 // # fall through to ExitMMB
5269 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal)
5270 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0);
5271 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal)
5272 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize);
5273 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal)
5274 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp);
5275 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5276 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB);
5277 MBB->addSuccessor(LoopMBB);
5278 MBB->addSuccessor(DoneMBB);
5280 MI->eraseFromParent();
5284 // Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true
5285 // if the high register of the GR128 value must be cleared or false if
5286 // it's "don't care". SubReg is subreg_l32 when extending a GR32
5287 // and subreg_l64 when extending a GR64.
5289 SystemZTargetLowering::emitExt128(MachineInstr *MI,
5290 MachineBasicBlock *MBB,
5291 bool ClearEven, unsigned SubReg) const {
5292 MachineFunction &MF = *MBB->getParent();
5293 const SystemZInstrInfo *TII =
5294 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
5295 MachineRegisterInfo &MRI = MF.getRegInfo();
5296 DebugLoc DL = MI->getDebugLoc();
5298 unsigned Dest = MI->getOperand(0).getReg();
5299 unsigned Src = MI->getOperand(1).getReg();
5300 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
5302 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128);
5304 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass);
5305 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass);
5307 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64)
5309 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128)
5310 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_h64);
5313 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest)
5314 .addReg(In128).addReg(Src).addImm(SubReg);
5316 MI->eraseFromParent();
5321 SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
5322 MachineBasicBlock *MBB,
5323 unsigned Opcode) const {
5324 MachineFunction &MF = *MBB->getParent();
5325 const SystemZInstrInfo *TII =
5326 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
5327 MachineRegisterInfo &MRI = MF.getRegInfo();
5328 DebugLoc DL = MI->getDebugLoc();
5330 MachineOperand DestBase = earlyUseOperand(MI->getOperand(0));
5331 uint64_t DestDisp = MI->getOperand(1).getImm();
5332 MachineOperand SrcBase = earlyUseOperand(MI->getOperand(2));
5333 uint64_t SrcDisp = MI->getOperand(3).getImm();
5334 uint64_t Length = MI->getOperand(4).getImm();
5336 // When generating more than one CLC, all but the last will need to
5337 // branch to the end when a difference is found.
5338 MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ?
5339 splitBlockAfter(MI, MBB) : nullptr);
5341 // Check for the loop form, in which operand 5 is the trip count.
5342 if (MI->getNumExplicitOperands() > 5) {
5343 bool HaveSingleBase = DestBase.isIdenticalTo(SrcBase);
5345 uint64_t StartCountReg = MI->getOperand(5).getReg();
5346 uint64_t StartSrcReg = forceReg(MI, SrcBase, TII);
5347 uint64_t StartDestReg = (HaveSingleBase ? StartSrcReg :
5348 forceReg(MI, DestBase, TII));
5350 const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass;
5351 uint64_t ThisSrcReg = MRI.createVirtualRegister(RC);
5352 uint64_t ThisDestReg = (HaveSingleBase ? ThisSrcReg :
5353 MRI.createVirtualRegister(RC));
5354 uint64_t NextSrcReg = MRI.createVirtualRegister(RC);
5355 uint64_t NextDestReg = (HaveSingleBase ? NextSrcReg :
5356 MRI.createVirtualRegister(RC));
5358 RC = &SystemZ::GR64BitRegClass;
5359 uint64_t ThisCountReg = MRI.createVirtualRegister(RC);
5360 uint64_t NextCountReg = MRI.createVirtualRegister(RC);
5362 MachineBasicBlock *StartMBB = MBB;
5363 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
5364 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
5365 MachineBasicBlock *NextMBB = (EndMBB ? emitBlockAfter(LoopMBB) : LoopMBB);
5368 // # fall through to LoopMMB
5369 MBB->addSuccessor(LoopMBB);
5372 // %ThisDestReg = phi [ %StartDestReg, StartMBB ],
5373 // [ %NextDestReg, NextMBB ]
5374 // %ThisSrcReg = phi [ %StartSrcReg, StartMBB ],
5375 // [ %NextSrcReg, NextMBB ]
5376 // %ThisCountReg = phi [ %StartCountReg, StartMBB ],
5377 // [ %NextCountReg, NextMBB ]
5378 // ( PFD 2, 768+DestDisp(%ThisDestReg) )
5379 // Opcode DestDisp(256,%ThisDestReg), SrcDisp(%ThisSrcReg)
5382 // The prefetch is used only for MVC. The JLH is used only for CLC.
5385 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisDestReg)
5386 .addReg(StartDestReg).addMBB(StartMBB)
5387 .addReg(NextDestReg).addMBB(NextMBB);
5388 if (!HaveSingleBase)
5389 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisSrcReg)
5390 .addReg(StartSrcReg).addMBB(StartMBB)
5391 .addReg(NextSrcReg).addMBB(NextMBB);
5392 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisCountReg)
5393 .addReg(StartCountReg).addMBB(StartMBB)
5394 .addReg(NextCountReg).addMBB(NextMBB);
5395 if (Opcode == SystemZ::MVC)
5396 BuildMI(MBB, DL, TII->get(SystemZ::PFD))
5397 .addImm(SystemZ::PFD_WRITE)
5398 .addReg(ThisDestReg).addImm(DestDisp + 768).addReg(0);
5399 BuildMI(MBB, DL, TII->get(Opcode))
5400 .addReg(ThisDestReg).addImm(DestDisp).addImm(256)
5401 .addReg(ThisSrcReg).addImm(SrcDisp);
5403 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5404 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
5406 MBB->addSuccessor(EndMBB);
5407 MBB->addSuccessor(NextMBB);
5411 // %NextDestReg = LA 256(%ThisDestReg)
5412 // %NextSrcReg = LA 256(%ThisSrcReg)
5413 // %NextCountReg = AGHI %ThisCountReg, -1
5414 // CGHI %NextCountReg, 0
5416 // # fall through to DoneMMB
5418 // The AGHI, CGHI and JLH should be converted to BRCTG by later passes.
5421 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextDestReg)
5422 .addReg(ThisDestReg).addImm(256).addReg(0);
5423 if (!HaveSingleBase)
5424 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextSrcReg)
5425 .addReg(ThisSrcReg).addImm(256).addReg(0);
5426 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), NextCountReg)
5427 .addReg(ThisCountReg).addImm(-1);
5428 BuildMI(MBB, DL, TII->get(SystemZ::CGHI))
5429 .addReg(NextCountReg).addImm(0);
5430 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5431 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
5433 MBB->addSuccessor(LoopMBB);
5434 MBB->addSuccessor(DoneMBB);
5436 DestBase = MachineOperand::CreateReg(NextDestReg, false);
5437 SrcBase = MachineOperand::CreateReg(NextSrcReg, false);
5441 // Handle any remaining bytes with straight-line code.
5442 while (Length > 0) {
5443 uint64_t ThisLength = std::min(Length, uint64_t(256));
5444 // The previous iteration might have created out-of-range displacements.
5445 // Apply them using LAY if so.
5446 if (!isUInt<12>(DestDisp)) {
5447 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
5448 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
5449 .addOperand(DestBase).addImm(DestDisp).addReg(0);
5450 DestBase = MachineOperand::CreateReg(Reg, false);
5453 if (!isUInt<12>(SrcDisp)) {
5454 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass);
5455 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg)
5456 .addOperand(SrcBase).addImm(SrcDisp).addReg(0);
5457 SrcBase = MachineOperand::CreateReg(Reg, false);
5460 BuildMI(*MBB, MI, DL, TII->get(Opcode))
5461 .addOperand(DestBase).addImm(DestDisp).addImm(ThisLength)
5462 .addOperand(SrcBase).addImm(SrcDisp);
5463 DestDisp += ThisLength;
5464 SrcDisp += ThisLength;
5465 Length -= ThisLength;
5466 // If there's another CLC to go, branch to the end if a difference
5468 if (EndMBB && Length > 0) {
5469 MachineBasicBlock *NextMBB = splitBlockBefore(MI, MBB);
5470 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5471 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE)
5473 MBB->addSuccessor(EndMBB);
5474 MBB->addSuccessor(NextMBB);
5479 MBB->addSuccessor(EndMBB);
5481 MBB->addLiveIn(SystemZ::CC);
5484 MI->eraseFromParent();
5488 // Decompose string pseudo-instruction MI into a loop that continually performs
5489 // Opcode until CC != 3.
5491 SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
5492 MachineBasicBlock *MBB,
5493 unsigned Opcode) const {
5494 MachineFunction &MF = *MBB->getParent();
5495 const SystemZInstrInfo *TII =
5496 static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
5497 MachineRegisterInfo &MRI = MF.getRegInfo();
5498 DebugLoc DL = MI->getDebugLoc();
5500 uint64_t End1Reg = MI->getOperand(0).getReg();
5501 uint64_t Start1Reg = MI->getOperand(1).getReg();
5502 uint64_t Start2Reg = MI->getOperand(2).getReg();
5503 uint64_t CharReg = MI->getOperand(3).getReg();
5505 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass;
5506 uint64_t This1Reg = MRI.createVirtualRegister(RC);
5507 uint64_t This2Reg = MRI.createVirtualRegister(RC);
5508 uint64_t End2Reg = MRI.createVirtualRegister(RC);
5510 MachineBasicBlock *StartMBB = MBB;
5511 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB);
5512 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB);
5515 // # fall through to LoopMMB
5516 MBB->addSuccessor(LoopMBB);
5519 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ]
5520 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ]
5522 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0L
5524 // # fall through to DoneMMB
5526 // The load of R0L can be hoisted by post-RA LICM.
5529 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg)
5530 .addReg(Start1Reg).addMBB(StartMBB)
5531 .addReg(End1Reg).addMBB(LoopMBB);
5532 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg)
5533 .addReg(Start2Reg).addMBB(StartMBB)
5534 .addReg(End2Reg).addMBB(LoopMBB);
5535 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0L).addReg(CharReg);
5536 BuildMI(MBB, DL, TII->get(Opcode))
5537 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define)
5538 .addReg(This1Reg).addReg(This2Reg);
5539 BuildMI(MBB, DL, TII->get(SystemZ::BRC))
5540 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB);
5541 MBB->addSuccessor(LoopMBB);
5542 MBB->addSuccessor(DoneMBB);
5544 DoneMBB->addLiveIn(SystemZ::CC);
5546 MI->eraseFromParent();
5550 // Update TBEGIN instruction with final opcode and register clobbers.
5552 SystemZTargetLowering::emitTransactionBegin(MachineInstr *MI,
5553 MachineBasicBlock *MBB,
5555 bool NoFloat) const {
5556 MachineFunction &MF = *MBB->getParent();
5557 const TargetFrameLowering *TFI = Subtarget.getFrameLowering();
5558 const SystemZInstrInfo *TII = Subtarget.getInstrInfo();
5561 MI->setDesc(TII->get(Opcode));
5563 // We cannot handle a TBEGIN that clobbers the stack or frame pointer.
5564 // Make sure to add the corresponding GRSM bits if they are missing.
5565 uint64_t Control = MI->getOperand(2).getImm();
5566 static const unsigned GPRControlBit[16] = {
5567 0x8000, 0x8000, 0x4000, 0x4000, 0x2000, 0x2000, 0x1000, 0x1000,
5568 0x0800, 0x0800, 0x0400, 0x0400, 0x0200, 0x0200, 0x0100, 0x0100
5570 Control |= GPRControlBit[15];
5572 Control |= GPRControlBit[11];
5573 MI->getOperand(2).setImm(Control);
5575 // Add GPR clobbers.
5576 for (int I = 0; I < 16; I++) {
5577 if ((Control & GPRControlBit[I]) == 0) {
5578 unsigned Reg = SystemZMC::GR64Regs[I];
5579 MI->addOperand(MachineOperand::CreateReg(Reg, true, true));
5583 // Add FPR/VR clobbers.
5584 if (!NoFloat && (Control & 4) != 0) {
5585 if (Subtarget.hasVector()) {
5586 for (int I = 0; I < 32; I++) {
5587 unsigned Reg = SystemZMC::VR128Regs[I];
5588 MI->addOperand(MachineOperand::CreateReg(Reg, true, true));
5591 for (int I = 0; I < 16; I++) {
5592 unsigned Reg = SystemZMC::FP64Regs[I];
5593 MI->addOperand(MachineOperand::CreateReg(Reg, true, true));
5601 MachineBasicBlock *SystemZTargetLowering::
5602 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
5603 switch (MI->getOpcode()) {
5604 case SystemZ::Select32Mux:
5605 case SystemZ::Select32:
5606 case SystemZ::SelectF32:
5607 case SystemZ::Select64:
5608 case SystemZ::SelectF64:
5609 case SystemZ::SelectF128:
5610 return emitSelect(MI, MBB);
5612 case SystemZ::CondStore8Mux:
5613 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, false);
5614 case SystemZ::CondStore8MuxInv:
5615 return emitCondStore(MI, MBB, SystemZ::STCMux, 0, true);
5616 case SystemZ::CondStore16Mux:
5617 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, false);
5618 case SystemZ::CondStore16MuxInv:
5619 return emitCondStore(MI, MBB, SystemZ::STHMux, 0, true);
5620 case SystemZ::CondStore8:
5621 return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
5622 case SystemZ::CondStore8Inv:
5623 return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
5624 case SystemZ::CondStore16:
5625 return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
5626 case SystemZ::CondStore16Inv:
5627 return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
5628 case SystemZ::CondStore32:
5629 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
5630 case SystemZ::CondStore32Inv:
5631 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
5632 case SystemZ::CondStore64:
5633 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
5634 case SystemZ::CondStore64Inv:
5635 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
5636 case SystemZ::CondStoreF32:
5637 return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
5638 case SystemZ::CondStoreF32Inv:
5639 return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
5640 case SystemZ::CondStoreF64:
5641 return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
5642 case SystemZ::CondStoreF64Inv:
5643 return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
5645 case SystemZ::AEXT128_64:
5646 return emitExt128(MI, MBB, false, SystemZ::subreg_l64);
5647 case SystemZ::ZEXT128_32:
5648 return emitExt128(MI, MBB, true, SystemZ::subreg_l32);
5649 case SystemZ::ZEXT128_64:
5650 return emitExt128(MI, MBB, true, SystemZ::subreg_l64);
5652 case SystemZ::ATOMIC_SWAPW:
5653 return emitAtomicLoadBinary(MI, MBB, 0, 0);
5654 case SystemZ::ATOMIC_SWAP_32:
5655 return emitAtomicLoadBinary(MI, MBB, 0, 32);
5656 case SystemZ::ATOMIC_SWAP_64:
5657 return emitAtomicLoadBinary(MI, MBB, 0, 64);
5659 case SystemZ::ATOMIC_LOADW_AR:
5660 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0);
5661 case SystemZ::ATOMIC_LOADW_AFI:
5662 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0);
5663 case SystemZ::ATOMIC_LOAD_AR:
5664 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32);
5665 case SystemZ::ATOMIC_LOAD_AHI:
5666 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32);
5667 case SystemZ::ATOMIC_LOAD_AFI:
5668 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32);
5669 case SystemZ::ATOMIC_LOAD_AGR:
5670 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64);
5671 case SystemZ::ATOMIC_LOAD_AGHI:
5672 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64);
5673 case SystemZ::ATOMIC_LOAD_AGFI:
5674 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64);
5676 case SystemZ::ATOMIC_LOADW_SR:
5677 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0);
5678 case SystemZ::ATOMIC_LOAD_SR:
5679 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32);
5680 case SystemZ::ATOMIC_LOAD_SGR:
5681 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64);
5683 case SystemZ::ATOMIC_LOADW_NR:
5684 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0);
5685 case SystemZ::ATOMIC_LOADW_NILH:
5686 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0);
5687 case SystemZ::ATOMIC_LOAD_NR:
5688 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32);
5689 case SystemZ::ATOMIC_LOAD_NILL:
5690 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32);
5691 case SystemZ::ATOMIC_LOAD_NILH:
5692 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32);
5693 case SystemZ::ATOMIC_LOAD_NILF:
5694 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32);
5695 case SystemZ::ATOMIC_LOAD_NGR:
5696 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64);
5697 case SystemZ::ATOMIC_LOAD_NILL64:
5698 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64);
5699 case SystemZ::ATOMIC_LOAD_NILH64:
5700 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64);
5701 case SystemZ::ATOMIC_LOAD_NIHL64:
5702 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64);
5703 case SystemZ::ATOMIC_LOAD_NIHH64:
5704 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64);
5705 case SystemZ::ATOMIC_LOAD_NILF64:
5706 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64);
5707 case SystemZ::ATOMIC_LOAD_NIHF64:
5708 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64);
5710 case SystemZ::ATOMIC_LOADW_OR:
5711 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0);
5712 case SystemZ::ATOMIC_LOADW_OILH:
5713 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 0);
5714 case SystemZ::ATOMIC_LOAD_OR:
5715 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32);
5716 case SystemZ::ATOMIC_LOAD_OILL:
5717 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 32);
5718 case SystemZ::ATOMIC_LOAD_OILH:
5719 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 32);
5720 case SystemZ::ATOMIC_LOAD_OILF:
5721 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 32);
5722 case SystemZ::ATOMIC_LOAD_OGR:
5723 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64);
5724 case SystemZ::ATOMIC_LOAD_OILL64:
5725 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL64, 64);
5726 case SystemZ::ATOMIC_LOAD_OILH64:
5727 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH64, 64);
5728 case SystemZ::ATOMIC_LOAD_OIHL64:
5729 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL64, 64);
5730 case SystemZ::ATOMIC_LOAD_OIHH64:
5731 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH64, 64);
5732 case SystemZ::ATOMIC_LOAD_OILF64:
5733 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF64, 64);
5734 case SystemZ::ATOMIC_LOAD_OIHF64:
5735 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF64, 64);
5737 case SystemZ::ATOMIC_LOADW_XR:
5738 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0);
5739 case SystemZ::ATOMIC_LOADW_XILF:
5740 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 0);
5741 case SystemZ::ATOMIC_LOAD_XR:
5742 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32);
5743 case SystemZ::ATOMIC_LOAD_XILF:
5744 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 32);
5745 case SystemZ::ATOMIC_LOAD_XGR:
5746 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64);
5747 case SystemZ::ATOMIC_LOAD_XILF64:
5748 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF64, 64);
5749 case SystemZ::ATOMIC_LOAD_XIHF64:
5750 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF64, 64);
5752 case SystemZ::ATOMIC_LOADW_NRi:
5753 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true);
5754 case SystemZ::ATOMIC_LOADW_NILHi:
5755 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 0, true);
5756 case SystemZ::ATOMIC_LOAD_NRi:
5757 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true);
5758 case SystemZ::ATOMIC_LOAD_NILLi:
5759 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 32, true);
5760 case SystemZ::ATOMIC_LOAD_NILHi:
5761 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 32, true);
5762 case SystemZ::ATOMIC_LOAD_NILFi:
5763 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 32, true);
5764 case SystemZ::ATOMIC_LOAD_NGRi:
5765 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true);
5766 case SystemZ::ATOMIC_LOAD_NILL64i:
5767 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL64, 64, true);
5768 case SystemZ::ATOMIC_LOAD_NILH64i:
5769 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH64, 64, true);
5770 case SystemZ::ATOMIC_LOAD_NIHL64i:
5771 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL64, 64, true);
5772 case SystemZ::ATOMIC_LOAD_NIHH64i:
5773 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH64, 64, true);
5774 case SystemZ::ATOMIC_LOAD_NILF64i:
5775 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF64, 64, true);
5776 case SystemZ::ATOMIC_LOAD_NIHF64i:
5777 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF64, 64, true);
5779 case SystemZ::ATOMIC_LOADW_MIN:
5780 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
5781 SystemZ::CCMASK_CMP_LE, 0);
5782 case SystemZ::ATOMIC_LOAD_MIN_32:
5783 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
5784 SystemZ::CCMASK_CMP_LE, 32);
5785 case SystemZ::ATOMIC_LOAD_MIN_64:
5786 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
5787 SystemZ::CCMASK_CMP_LE, 64);
5789 case SystemZ::ATOMIC_LOADW_MAX:
5790 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
5791 SystemZ::CCMASK_CMP_GE, 0);
5792 case SystemZ::ATOMIC_LOAD_MAX_32:
5793 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR,
5794 SystemZ::CCMASK_CMP_GE, 32);
5795 case SystemZ::ATOMIC_LOAD_MAX_64:
5796 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR,
5797 SystemZ::CCMASK_CMP_GE, 64);
5799 case SystemZ::ATOMIC_LOADW_UMIN:
5800 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
5801 SystemZ::CCMASK_CMP_LE, 0);
5802 case SystemZ::ATOMIC_LOAD_UMIN_32:
5803 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
5804 SystemZ::CCMASK_CMP_LE, 32);
5805 case SystemZ::ATOMIC_LOAD_UMIN_64:
5806 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
5807 SystemZ::CCMASK_CMP_LE, 64);
5809 case SystemZ::ATOMIC_LOADW_UMAX:
5810 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
5811 SystemZ::CCMASK_CMP_GE, 0);
5812 case SystemZ::ATOMIC_LOAD_UMAX_32:
5813 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR,
5814 SystemZ::CCMASK_CMP_GE, 32);
5815 case SystemZ::ATOMIC_LOAD_UMAX_64:
5816 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR,
5817 SystemZ::CCMASK_CMP_GE, 64);
5819 case SystemZ::ATOMIC_CMP_SWAPW:
5820 return emitAtomicCmpSwapW(MI, MBB);
5821 case SystemZ::MVCSequence:
5822 case SystemZ::MVCLoop:
5823 return emitMemMemWrapper(MI, MBB, SystemZ::MVC);
5824 case SystemZ::NCSequence:
5825 case SystemZ::NCLoop:
5826 return emitMemMemWrapper(MI, MBB, SystemZ::NC);
5827 case SystemZ::OCSequence:
5828 case SystemZ::OCLoop:
5829 return emitMemMemWrapper(MI, MBB, SystemZ::OC);
5830 case SystemZ::XCSequence:
5831 case SystemZ::XCLoop:
5832 return emitMemMemWrapper(MI, MBB, SystemZ::XC);
5833 case SystemZ::CLCSequence:
5834 case SystemZ::CLCLoop:
5835 return emitMemMemWrapper(MI, MBB, SystemZ::CLC);
5836 case SystemZ::CLSTLoop:
5837 return emitStringWrapper(MI, MBB, SystemZ::CLST);
5838 case SystemZ::MVSTLoop:
5839 return emitStringWrapper(MI, MBB, SystemZ::MVST);
5840 case SystemZ::SRSTLoop:
5841 return emitStringWrapper(MI, MBB, SystemZ::SRST);
5842 case SystemZ::TBEGIN:
5843 return emitTransactionBegin(MI, MBB, SystemZ::TBEGIN, false);
5844 case SystemZ::TBEGIN_nofloat:
5845 return emitTransactionBegin(MI, MBB, SystemZ::TBEGIN, true);
5846 case SystemZ::TBEGINC:
5847 return emitTransactionBegin(MI, MBB, SystemZ::TBEGINC, true);
5849 llvm_unreachable("Unexpected instr type to insert");