1 //===-- SIISelLowering.cpp - SI 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 //===----------------------------------------------------------------------===//
11 /// \brief Custom DAG lowering for SI
13 //===----------------------------------------------------------------------===//
17 #define _USE_MATH_DEFINES
21 #include "SIISelLowering.h"
23 #include "AMDGPUDiagnosticInfoUnsupported.h"
24 #include "AMDGPUIntrinsicInfo.h"
25 #include "AMDGPUSubtarget.h"
26 #include "SIInstrInfo.h"
27 #include "SIMachineFunctionInfo.h"
28 #include "SIRegisterInfo.h"
29 #include "llvm/ADT/BitVector.h"
30 #include "llvm/CodeGen/CallingConvLower.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAG.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/ADT/SmallString.h"
39 SITargetLowering::SITargetLowering(TargetMachine &TM,
40 const AMDGPUSubtarget &STI)
41 : AMDGPUTargetLowering(TM, STI) {
42 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
43 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
45 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
46 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
48 addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
49 addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);
51 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
52 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
53 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
55 addRegisterClass(MVT::v2i64, &AMDGPU::SReg_128RegClass);
56 addRegisterClass(MVT::v2f64, &AMDGPU::SReg_128RegClass);
58 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
59 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
61 addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
62 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
64 addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass);
65 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
67 computeRegisterProperties(STI.getRegisterInfo());
69 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
70 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
71 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
72 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
74 setOperationAction(ISD::ADD, MVT::i32, Legal);
75 setOperationAction(ISD::ADDC, MVT::i32, Legal);
76 setOperationAction(ISD::ADDE, MVT::i32, Legal);
77 setOperationAction(ISD::SUBC, MVT::i32, Legal);
78 setOperationAction(ISD::SUBE, MVT::i32, Legal);
80 setOperationAction(ISD::FSIN, MVT::f32, Custom);
81 setOperationAction(ISD::FCOS, MVT::f32, Custom);
83 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
84 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
86 // We need to custom lower vector stores from local memory
87 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
88 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
89 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
91 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
92 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
94 setOperationAction(ISD::STORE, MVT::i1, Custom);
95 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
97 setOperationAction(ISD::SELECT, MVT::i64, Custom);
98 setOperationAction(ISD::SELECT, MVT::f64, Promote);
99 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
101 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
102 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
103 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
104 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
106 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
107 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
109 setOperationAction(ISD::BSWAP, MVT::i32, Legal);
110 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
112 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
113 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
114 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
116 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
117 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
118 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
120 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
124 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
127 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
128 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
129 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
130 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
132 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
133 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
135 for (MVT VT : MVT::integer_valuetypes()) {
139 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
141 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
142 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
144 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
145 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
146 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
147 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
149 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
150 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
151 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
152 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
155 for (MVT VT : MVT::integer_vector_valuetypes()) {
156 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i16, Expand);
157 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v16i16, Expand);
160 for (MVT VT : MVT::fp_valuetypes())
161 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
163 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
164 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
166 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
167 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
168 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Expand);
169 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
172 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);
174 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
175 setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
177 setOperationAction(ISD::LOAD, MVT::i1, Custom);
179 setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
180 AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);
182 setOperationAction(ISD::STORE, MVT::v2i64, Promote);
183 AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);
185 setOperationAction(ISD::ConstantPool, MVT::v2i64, Expand);
187 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
188 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
189 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
191 // These should use UDIVREM, so set them to expand
192 setOperationAction(ISD::UDIV, MVT::i64, Expand);
193 setOperationAction(ISD::UREM, MVT::i64, Expand);
195 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
196 setOperationAction(ISD::SELECT, MVT::i1, Promote);
198 setOperationAction(ISD::TRUNCATE, MVT::v2i32, Expand);
201 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
203 // We only support LOAD/STORE and vector manipulation ops for vectors
204 // with > 4 elements.
205 for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32, MVT::v2i64, MVT::v2f64}) {
206 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
210 case ISD::BUILD_VECTOR:
212 case ISD::EXTRACT_VECTOR_ELT:
213 case ISD::INSERT_VECTOR_ELT:
214 case ISD::INSERT_SUBVECTOR:
215 case ISD::EXTRACT_SUBVECTOR:
216 case ISD::SCALAR_TO_VECTOR:
218 case ISD::CONCAT_VECTORS:
219 setOperationAction(Op, VT, Custom);
222 setOperationAction(Op, VT, Expand);
228 // Most operations are naturally 32-bit vector operations. We only support
229 // load and store of i64 vectors, so promote v2i64 vector operations to v4i32.
230 for (MVT Vec64 : { MVT::v2i64, MVT::v2f64 }) {
231 setOperationAction(ISD::BUILD_VECTOR, Vec64, Promote);
232 AddPromotedToType(ISD::BUILD_VECTOR, Vec64, MVT::v4i32);
234 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Vec64, Promote);
235 AddPromotedToType(ISD::EXTRACT_VECTOR_ELT, Vec64, MVT::v4i32);
237 setOperationAction(ISD::INSERT_VECTOR_ELT, Vec64, Promote);
238 AddPromotedToType(ISD::INSERT_VECTOR_ELT, Vec64, MVT::v4i32);
240 setOperationAction(ISD::SCALAR_TO_VECTOR, Vec64, Promote);
241 AddPromotedToType(ISD::SCALAR_TO_VECTOR, Vec64, MVT::v4i32);
244 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
245 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
246 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
247 setOperationAction(ISD::FRINT, MVT::f64, Legal);
250 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
251 setOperationAction(ISD::FDIV, MVT::f32, Custom);
252 setOperationAction(ISD::FDIV, MVT::f64, Custom);
254 setTargetDAGCombine(ISD::FADD);
255 setTargetDAGCombine(ISD::FSUB);
256 setTargetDAGCombine(ISD::FMINNUM);
257 setTargetDAGCombine(ISD::FMAXNUM);
258 setTargetDAGCombine(ISD::SMIN);
259 setTargetDAGCombine(ISD::SMAX);
260 setTargetDAGCombine(ISD::UMIN);
261 setTargetDAGCombine(ISD::UMAX);
262 setTargetDAGCombine(ISD::SETCC);
263 setTargetDAGCombine(ISD::AND);
264 setTargetDAGCombine(ISD::OR);
265 setTargetDAGCombine(ISD::UINT_TO_FP);
267 // All memory operations. Some folding on the pointer operand is done to help
268 // matching the constant offsets in the addressing modes.
269 setTargetDAGCombine(ISD::LOAD);
270 setTargetDAGCombine(ISD::STORE);
271 setTargetDAGCombine(ISD::ATOMIC_LOAD);
272 setTargetDAGCombine(ISD::ATOMIC_STORE);
273 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
274 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
275 setTargetDAGCombine(ISD::ATOMIC_SWAP);
276 setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
277 setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
278 setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
279 setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
280 setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
281 setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
282 setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
283 setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
284 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
285 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
287 setSchedulingPreference(Sched::RegPressure);
290 //===----------------------------------------------------------------------===//
291 // TargetLowering queries
292 //===----------------------------------------------------------------------===//
294 bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
296 // SI has some legal vector types, but no legal vector operations. Say no
297 // shuffles are legal in order to prefer scalarizing some vector operations.
301 bool SITargetLowering::isLegalFlatAddressingMode(const AddrMode &AM) const {
302 // Flat instructions do not have offsets, and only have the register
304 return AM.BaseOffs == 0 && (AM.Scale == 0 || AM.Scale == 1);
307 bool SITargetLowering::isLegalMUBUFAddressingMode(const AddrMode &AM) const {
308 // MUBUF / MTBUF instructions have a 12-bit unsigned byte offset, and
309 // additionally can do r + r + i with addr64. 32-bit has more addressing
310 // mode options. Depending on the resource constant, it can also do
311 // (i64 r0) + (i32 r1) * (i14 i).
313 // Private arrays end up using a scratch buffer most of the time, so also
314 // assume those use MUBUF instructions. Scratch loads / stores are currently
315 // implemented as mubuf instructions with offen bit set, so slightly
316 // different than the normal addr64.
317 if (!isUInt<12>(AM.BaseOffs))
320 // FIXME: Since we can split immediate into soffset and immediate offset,
321 // would it make sense to allow any immediate?
324 case 0: // r + i or just i, depending on HasBaseReg.
327 return true; // We have r + r or r + i.
334 // Allow 2 * r as r + r
335 // Or 2 * r + i is allowed as r + r + i.
337 default: // Don't allow n * r
342 bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL,
343 const AddrMode &AM, Type *Ty,
345 // No global is ever allowed as a base.
350 case AMDGPUAS::GLOBAL_ADDRESS: {
351 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
352 // Assume the we will use FLAT for all global memory accesses
354 // FIXME: This assumption is currently wrong. On VI we still use
355 // MUBUF instructions for the r + i addressing mode. As currently
356 // implemented, the MUBUF instructions only work on buffer < 4GB.
357 // It may be possible to support > 4GB buffers with MUBUF instructions,
358 // by setting the stride value in the resource descriptor which would
359 // increase the size limit to (stride * 4GB). However, this is risky,
360 // because it has never been validated.
361 return isLegalFlatAddressingMode(AM);
364 return isLegalMUBUFAddressingMode(AM);
366 case AMDGPUAS::CONSTANT_ADDRESS: {
367 // If the offset isn't a multiple of 4, it probably isn't going to be
368 // correctly aligned.
369 if (AM.BaseOffs % 4 != 0)
370 return isLegalMUBUFAddressingMode(AM);
372 // There are no SMRD extloads, so if we have to do a small type access we
373 // will use a MUBUF load.
374 // FIXME?: We also need to do this if unaligned, but we don't know the
376 if (DL.getTypeStoreSize(Ty) < 4)
377 return isLegalMUBUFAddressingMode(AM);
379 if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) {
380 // SMRD instructions have an 8-bit, dword offset on SI.
381 if (!isUInt<8>(AM.BaseOffs / 4))
383 } else if (Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS) {
384 // On CI+, this can also be a 32-bit literal constant offset. If it fits
385 // in 8-bits, it can use a smaller encoding.
386 if (!isUInt<32>(AM.BaseOffs / 4))
388 } else if (Subtarget->getGeneration() == AMDGPUSubtarget::VOLCANIC_ISLANDS) {
389 // On VI, these use the SMEM format and the offset is 20-bit in bytes.
390 if (!isUInt<20>(AM.BaseOffs))
393 llvm_unreachable("unhandled generation");
395 if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
398 if (AM.Scale == 1 && AM.HasBaseReg)
404 case AMDGPUAS::PRIVATE_ADDRESS:
405 case AMDGPUAS::UNKNOWN_ADDRESS_SPACE:
406 return isLegalMUBUFAddressingMode(AM);
408 case AMDGPUAS::LOCAL_ADDRESS:
409 case AMDGPUAS::REGION_ADDRESS: {
410 // Basic, single offset DS instructions allow a 16-bit unsigned immediate
412 // XXX - If doing a 4-byte aligned 8-byte type access, we effectively have
413 // an 8-bit dword offset but we don't know the alignment here.
414 if (!isUInt<16>(AM.BaseOffs))
417 if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
420 if (AM.Scale == 1 && AM.HasBaseReg)
425 case AMDGPUAS::FLAT_ADDRESS:
426 return isLegalFlatAddressingMode(AM);
429 llvm_unreachable("unhandled address space");
433 bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
436 bool *IsFast) const {
440 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
441 // which isn't a simple VT.
442 if (!VT.isSimple() || VT == MVT::Other)
445 // TODO - CI+ supports unaligned memory accesses, but this requires driver
448 // XXX - The only mention I see of this in the ISA manual is for LDS direct
449 // reads the "byte address and must be dword aligned". Is it also true for the
450 // normal loads and stores?
451 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS) {
452 // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
453 // aligned, 8 byte access in a single operation using ds_read2/write2_b32
454 // with adjacent offsets.
455 bool AlignedBy4 = (Align % 4 == 0);
457 *IsFast = AlignedBy4;
461 // Smaller than dword value must be aligned.
462 // FIXME: This should be allowed on CI+
463 if (VT.bitsLT(MVT::i32))
466 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
467 // byte-address are ignored, thus forcing Dword alignment.
468 // This applies to private, global, and constant memory.
472 return VT.bitsGT(MVT::i32) && Align % 4 == 0;
475 EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
476 unsigned SrcAlign, bool IsMemset,
479 MachineFunction &MF) const {
480 // FIXME: Should account for address space here.
482 // The default fallback uses the private pointer size as a guess for a type to
483 // use. Make sure we switch these to 64-bit accesses.
485 if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
488 if (Size >= 8 && DstAlign >= 4)
495 static bool isFlatGlobalAddrSpace(unsigned AS) {
496 return AS == AMDGPUAS::GLOBAL_ADDRESS ||
497 AS == AMDGPUAS::FLAT_ADDRESS ||
498 AS == AMDGPUAS::CONSTANT_ADDRESS;
501 bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
502 unsigned DestAS) const {
503 return isFlatGlobalAddrSpace(SrcAS) && isFlatGlobalAddrSpace(DestAS);
507 bool SITargetLowering::isMemOpUniform(const SDNode *N) const {
508 const MemSDNode *MemNode = cast<MemSDNode>(N);
509 const Value *Ptr = MemNode->getMemOperand()->getValue();
511 // UndefValue means this is a load of a kernel input. These are uniform.
512 // Sometimes LDS instructions have constant pointers
513 if (isa<UndefValue>(Ptr) || isa<Argument>(Ptr) || isa<Constant>(Ptr) ||
514 isa<GlobalValue>(Ptr))
517 const Instruction *I = dyn_cast_or_null<Instruction>(Ptr);
518 return I && I->getMetadata("amdgpu.uniform");
521 TargetLoweringBase::LegalizeTypeAction
522 SITargetLowering::getPreferredVectorAction(EVT VT) const {
523 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
524 return TypeSplitVector;
526 return TargetLoweringBase::getPreferredVectorAction(VT);
529 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
531 const SIInstrInfo *TII =
532 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
533 return TII->isInlineConstant(Imm);
536 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
537 SDLoc SL, SDValue Chain,
538 unsigned Offset, bool Signed) const {
539 const DataLayout &DL = DAG.getDataLayout();
540 MachineFunction &MF = DAG.getMachineFunction();
541 const SIRegisterInfo *TRI =
542 static_cast<const SIRegisterInfo*>(Subtarget->getRegisterInfo());
543 unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::KERNARG_SEGMENT_PTR);
545 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
547 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
548 MVT PtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
549 PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
550 SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
551 MRI.getLiveInVirtReg(InputPtrReg), PtrVT);
552 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
553 DAG.getConstant(Offset, SL, PtrVT));
554 SDValue PtrOffset = DAG.getUNDEF(PtrVT);
555 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
557 unsigned Align = DL.getABITypeAlignment(Ty);
559 ISD::LoadExtType ExtTy = Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
560 if (MemVT.isFloatingPoint())
561 ExtTy = ISD::EXTLOAD;
563 return DAG.getLoad(ISD::UNINDEXED, ExtTy,
564 VT, SL, Chain, Ptr, PtrOffset, PtrInfo, MemVT,
566 true, // isNonTemporal
571 SDValue SITargetLowering::LowerFormalArguments(
572 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
573 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
574 SmallVectorImpl<SDValue> &InVals) const {
575 const SIRegisterInfo *TRI =
576 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
578 MachineFunction &MF = DAG.getMachineFunction();
579 FunctionType *FType = MF.getFunction()->getFunctionType();
580 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
581 const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
583 if (Subtarget->isAmdHsaOS() && Info->getShaderType() != ShaderType::COMPUTE) {
584 const Function *Fn = MF.getFunction();
585 DiagnosticInfoUnsupported NoGraphicsHSA(*Fn, "non-compute shaders with HSA");
586 DAG.getContext()->diagnose(NoGraphicsHSA);
590 // FIXME: We currently assume all calling conventions are kernels.
592 SmallVector<ISD::InputArg, 16> Splits;
593 BitVector Skipped(Ins.size());
595 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
596 const ISD::InputArg &Arg = Ins[i];
598 // First check if it's a PS input addr
599 if (Info->getShaderType() == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
600 !Arg.Flags.isByVal()) {
602 assert((PSInputNum <= 15) && "Too many PS inputs!");
605 // We can safely skip PS inputs
611 Info->PSInputAddr |= 1 << PSInputNum++;
614 // Second split vertices into their elements
615 if (Info->getShaderType() != ShaderType::COMPUTE && Arg.VT.isVector()) {
616 ISD::InputArg NewArg = Arg;
617 NewArg.Flags.setSplit();
618 NewArg.VT = Arg.VT.getVectorElementType();
620 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
621 // three or five element vertex only needs three or five registers,
622 // NOT four or eight.
623 Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
624 unsigned NumElements = ParamType->getVectorNumElements();
626 for (unsigned j = 0; j != NumElements; ++j) {
627 Splits.push_back(NewArg);
628 NewArg.PartOffset += NewArg.VT.getStoreSize();
631 } else if (Info->getShaderType() != ShaderType::COMPUTE) {
632 Splits.push_back(Arg);
636 SmallVector<CCValAssign, 16> ArgLocs;
637 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
640 // At least one interpolation mode must be enabled or else the GPU will hang.
641 if (Info->getShaderType() == ShaderType::PIXEL &&
642 (Info->PSInputAddr & 0x7F) == 0) {
643 Info->PSInputAddr |= 1;
644 CCInfo.AllocateReg(AMDGPU::VGPR0);
645 CCInfo.AllocateReg(AMDGPU::VGPR1);
648 if (Info->getShaderType() == ShaderType::COMPUTE) {
649 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
653 // FIXME: How should these inputs interact with inreg / custom SGPR inputs?
654 if (Info->hasPrivateSegmentBuffer()) {
655 unsigned PrivateSegmentBufferReg = Info->addPrivateSegmentBuffer(*TRI);
656 MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SReg_128RegClass);
657 CCInfo.AllocateReg(PrivateSegmentBufferReg);
660 if (Info->hasDispatchPtr()) {
661 unsigned DispatchPtrReg = Info->addDispatchPtr(*TRI);
662 MF.addLiveIn(DispatchPtrReg, &AMDGPU::SReg_64RegClass);
663 CCInfo.AllocateReg(DispatchPtrReg);
666 if (Info->hasKernargSegmentPtr()) {
667 unsigned InputPtrReg = Info->addKernargSegmentPtr(*TRI);
668 MF.addLiveIn(InputPtrReg, &AMDGPU::SReg_64RegClass);
669 CCInfo.AllocateReg(InputPtrReg);
672 AnalyzeFormalArguments(CCInfo, Splits);
674 SmallVector<SDValue, 16> Chains;
676 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
678 const ISD::InputArg &Arg = Ins[i];
680 InVals.push_back(DAG.getUNDEF(Arg.VT));
684 CCValAssign &VA = ArgLocs[ArgIdx++];
685 MVT VT = VA.getLocVT();
689 EVT MemVT = Splits[i].VT;
690 const unsigned Offset = Subtarget->getExplicitKernelArgOffset() +
691 VA.getLocMemOffset();
692 // The first 36 bytes of the input buffer contains information about
693 // thread group and global sizes.
694 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, Chain,
695 Offset, Ins[i].Flags.isSExt());
696 Chains.push_back(Arg.getValue(1));
699 dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
700 if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
701 ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
702 // On SI local pointers are just offsets into LDS, so they are always
703 // less than 16-bits. On CI and newer they could potentially be
704 // real pointers, so we can't guarantee their size.
705 Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
706 DAG.getValueType(MVT::i16));
709 InVals.push_back(Arg);
710 Info->ABIArgOffset = Offset + MemVT.getStoreSize();
713 assert(VA.isRegLoc() && "Parameter must be in a register!");
715 unsigned Reg = VA.getLocReg();
717 if (VT == MVT::i64) {
718 // For now assume it is a pointer
719 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
720 &AMDGPU::SReg_64RegClass);
721 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
722 SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
723 InVals.push_back(Copy);
727 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
729 Reg = MF.addLiveIn(Reg, RC);
730 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
732 if (Arg.VT.isVector()) {
734 // Build a vector from the registers
735 Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
736 unsigned NumElements = ParamType->getVectorNumElements();
738 SmallVector<SDValue, 4> Regs;
740 for (unsigned j = 1; j != NumElements; ++j) {
741 Reg = ArgLocs[ArgIdx++].getLocReg();
742 Reg = MF.addLiveIn(Reg, RC);
744 SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
745 Regs.push_back(Copy);
748 // Fill up the missing vector elements
749 NumElements = Arg.VT.getVectorNumElements() - NumElements;
750 Regs.append(NumElements, DAG.getUNDEF(VT));
752 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
756 InVals.push_back(Val);
759 // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
760 // these from the dispatch pointer.
762 // Start adding system SGPRs.
763 if (Info->hasWorkGroupIDX()) {
764 unsigned Reg = Info->addWorkGroupIDX();
765 MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
766 CCInfo.AllocateReg(Reg);
768 llvm_unreachable("work group id x is always enabled");
770 if (Info->hasWorkGroupIDY()) {
771 unsigned Reg = Info->addWorkGroupIDY();
772 MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
773 CCInfo.AllocateReg(Reg);
776 if (Info->hasWorkGroupIDZ()) {
777 unsigned Reg = Info->addWorkGroupIDZ();
778 MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
779 CCInfo.AllocateReg(Reg);
782 if (Info->hasWorkGroupInfo()) {
783 unsigned Reg = Info->addWorkGroupInfo();
784 MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
785 CCInfo.AllocateReg(Reg);
788 if (Info->hasPrivateSegmentWaveByteOffset()) {
789 // Scratch wave offset passed in system SGPR.
790 unsigned PrivateSegmentWaveByteOffsetReg
791 = Info->addPrivateSegmentWaveByteOffset();
793 MF.addLiveIn(PrivateSegmentWaveByteOffsetReg, &AMDGPU::SGPR_32RegClass);
794 CCInfo.AllocateReg(PrivateSegmentWaveByteOffsetReg);
797 // Now that we've figured out where the scratch register inputs are, see if
798 // should reserve the arguments and use them directly.
800 bool HasStackObjects = MF.getFrameInfo()->hasStackObjects();
802 if (ST.isAmdHsaOS()) {
803 // TODO: Assume we will spill without optimizations.
804 if (HasStackObjects) {
805 // If we have stack objects, we unquestionably need the private buffer
806 // resource. For the HSA ABI, this will be the first 4 user SGPR
807 // inputs. We can reserve those and use them directly.
809 unsigned PrivateSegmentBufferReg = TRI->getPreloadedValue(
810 MF, SIRegisterInfo::PRIVATE_SEGMENT_BUFFER);
811 Info->setScratchRSrcReg(PrivateSegmentBufferReg);
813 unsigned PrivateSegmentWaveByteOffsetReg = TRI->getPreloadedValue(
814 MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
815 Info->setScratchWaveOffsetReg(PrivateSegmentWaveByteOffsetReg);
817 unsigned ReservedBufferReg
818 = TRI->reservedPrivateSegmentBufferReg(MF);
819 unsigned ReservedOffsetReg
820 = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
822 // We tentatively reserve the last registers (skipping the last two
823 // which may contain VCC). After register allocation, we'll replace
824 // these with the ones immediately after those which were really
825 // allocated. In the prologue copies will be inserted from the argument
826 // to these reserved registers.
827 Info->setScratchRSrcReg(ReservedBufferReg);
828 Info->setScratchWaveOffsetReg(ReservedOffsetReg);
831 unsigned ReservedBufferReg = TRI->reservedPrivateSegmentBufferReg(MF);
833 // Without HSA, relocations are used for the scratch pointer and the
834 // buffer resource setup is always inserted in the prologue. Scratch wave
835 // offset is still in an input SGPR.
836 Info->setScratchRSrcReg(ReservedBufferReg);
838 if (HasStackObjects) {
839 unsigned ScratchWaveOffsetReg = TRI->getPreloadedValue(
840 MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
841 Info->setScratchWaveOffsetReg(ScratchWaveOffsetReg);
843 unsigned ReservedOffsetReg
844 = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
845 Info->setScratchWaveOffsetReg(ReservedOffsetReg);
849 if (Info->hasWorkItemIDX()) {
850 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X);
851 MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
852 CCInfo.AllocateReg(Reg);
854 llvm_unreachable("workitem id x should always be enabled");
856 if (Info->hasWorkItemIDY()) {
857 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y);
858 MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
859 CCInfo.AllocateReg(Reg);
862 if (Info->hasWorkItemIDZ()) {
863 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z);
864 MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
865 CCInfo.AllocateReg(Reg);
871 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
874 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
875 MachineInstr * MI, MachineBasicBlock * BB) const {
877 switch (MI->getOpcode()) {
879 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
886 bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
887 // This currently forces unfolding various combinations of fsub into fma with
888 // free fneg'd operands. As long as we have fast FMA (controlled by
889 // isFMAFasterThanFMulAndFAdd), we should perform these.
891 // When fma is quarter rate, for f64 where add / sub are at best half rate,
892 // most of these combines appear to be cycle neutral but save on instruction
893 // count / code size.
897 EVT SITargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &Ctx,
899 if (!VT.isVector()) {
902 return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements());
905 MVT SITargetLowering::getScalarShiftAmountTy(const DataLayout &, EVT) const {
909 // Answering this is somewhat tricky and depends on the specific device which
910 // have different rates for fma or all f64 operations.
912 // v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
913 // regardless of which device (although the number of cycles differs between
914 // devices), so it is always profitable for f64.
916 // v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
917 // only on full rate devices. Normally, we should prefer selecting v_mad_f32
918 // which we can always do even without fused FP ops since it returns the same
919 // result as the separate operations and since it is always full
920 // rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
921 // however does not support denormals, so we do report fma as faster if we have
922 // a fast fma device and require denormals.
924 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
925 VT = VT.getScalarType();
930 switch (VT.getSimpleVT().SimpleTy) {
932 // This is as fast on some subtargets. However, we always have full rate f32
933 // mad available which returns the same result as the separate operations
934 // which we should prefer over fma. We can't use this if we want to support
935 // denormals, so only report this in these cases.
936 return Subtarget->hasFP32Denormals() && Subtarget->hasFastFMAF32();
946 //===----------------------------------------------------------------------===//
947 // Custom DAG Lowering Operations
948 //===----------------------------------------------------------------------===//
950 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
951 switch (Op.getOpcode()) {
952 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
953 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
954 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
956 SDValue Result = LowerLOAD(Op, DAG);
957 assert((!Result.getNode() ||
958 Result.getNode()->getNumValues() == 2) &&
959 "Load should return a value and a chain");
965 return LowerTrig(Op, DAG);
966 case ISD::SELECT: return LowerSELECT(Op, DAG);
967 case ISD::FDIV: return LowerFDIV(Op, DAG);
968 case ISD::STORE: return LowerSTORE(Op, DAG);
969 case ISD::GlobalAddress: {
970 MachineFunction &MF = DAG.getMachineFunction();
971 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
972 return LowerGlobalAddress(MFI, Op, DAG);
974 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
975 case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
980 /// \brief Helper function for LowerBRCOND
981 static SDNode *findUser(SDValue Value, unsigned Opcode) {
983 SDNode *Parent = Value.getNode();
984 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
987 if (I.getUse().get() != Value)
990 if (I->getOpcode() == Opcode)
996 SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
999 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op);
1000 unsigned FrameIndex = FINode->getIndex();
1002 // A FrameIndex node represents a 32-bit offset into scratch memory. If
1003 // the high bit of a frame index offset were to be set, this would mean
1004 // that it represented an offset of ~2GB * 64 = ~128GB from the start of the
1005 // scratch buffer, with 64 being the number of threads per wave.
1007 // If we know the machine uses less than 128GB of scratch, then we can
1008 // amrk the high bit of the FrameIndex node as known zero,
1009 // which is important, because it means in most situations we can
1010 // prove that values derived from FrameIndex nodes are non-negative.
1011 // This enables us to take advantage of more addressing modes when
1012 // accessing scratch buffers, since for scratch reads/writes, the register
1013 // offset must always be positive.
1015 SDValue TFI = DAG.getTargetFrameIndex(FrameIndex, MVT::i32);
1016 if (Subtarget->enableHugeScratchBuffer())
1019 return DAG.getNode(ISD::AssertZext, SL, MVT::i32, TFI,
1020 DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), 31)));
1023 /// This transforms the control flow intrinsics to get the branch destination as
1024 /// last parameter, also switches branch target with BR if the need arise
1025 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
1026 SelectionDAG &DAG) const {
1030 SDNode *Intr = BRCOND.getOperand(1).getNode();
1031 SDValue Target = BRCOND.getOperand(2);
1032 SDNode *BR = nullptr;
1034 if (Intr->getOpcode() == ISD::SETCC) {
1035 // As long as we negate the condition everything is fine
1036 SDNode *SetCC = Intr;
1037 assert(SetCC->getConstantOperandVal(1) == 1);
1038 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
1040 Intr = SetCC->getOperand(0).getNode();
1043 // Get the target from BR if we don't negate the condition
1044 BR = findUser(BRCOND, ISD::BR);
1045 Target = BR->getOperand(1);
1048 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
1050 // Build the result and
1051 ArrayRef<EVT> Res(Intr->value_begin() + 1, Intr->value_end());
1053 // operands of the new intrinsic call
1054 SmallVector<SDValue, 4> Ops;
1055 Ops.push_back(BRCOND.getOperand(0));
1056 Ops.append(Intr->op_begin() + 1, Intr->op_end());
1057 Ops.push_back(Target);
1059 // build the new intrinsic call
1060 SDNode *Result = DAG.getNode(
1061 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
1062 DAG.getVTList(Res), Ops).getNode();
1065 // Give the branch instruction our target
1068 BRCOND.getOperand(2)
1070 SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops);
1071 DAG.ReplaceAllUsesWith(BR, NewBR.getNode());
1072 BR = NewBR.getNode();
1075 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
1077 // Copy the intrinsic results to registers
1078 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
1079 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
1083 Chain = DAG.getCopyToReg(
1085 CopyToReg->getOperand(1),
1086 SDValue(Result, i - 1),
1089 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
1092 // Remove the old intrinsic from the chain
1093 DAG.ReplaceAllUsesOfValueWith(
1094 SDValue(Intr, Intr->getNumValues() - 1),
1095 Intr->getOperand(0));
1100 SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
1102 SelectionDAG &DAG) const {
1103 GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
1105 if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
1106 return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
1109 const GlobalValue *GV = GSD->getGlobal();
1110 MVT PtrVT = getPointerTy(DAG.getDataLayout(), GSD->getAddressSpace());
1112 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32);
1113 return DAG.getNode(AMDGPUISD::CONST_DATA_PTR, DL, PtrVT, GA);
1116 SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
1118 // We can't use CopyToReg, because MachineCSE won't combine COPY instructions,
1119 // so we will end up with redundant moves to m0.
1121 // We can't use S_MOV_B32, because there is no way to specify m0 as the
1122 // destination register.
1124 // We have to use them both. Machine cse will combine all the S_MOV_B32
1125 // instructions and the register coalescer eliminate the extra copies.
1126 SDNode *M0 = DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, V.getValueType(), V);
1127 return DAG.getCopyToReg(Chain, DL, DAG.getRegister(AMDGPU::M0, MVT::i32),
1128 SDValue(M0, 0), SDValue()); // Glue
1129 // A Null SDValue creates
1133 SDValue SITargetLowering::lowerImplicitZextParam(SelectionDAG &DAG,
1136 unsigned Offset) const {
1138 SDValue Param = LowerParameter(DAG, MVT::i32, MVT::i32, SL,
1139 DAG.getEntryNode(), Offset, false);
1140 // The local size values will have the hi 16-bits as zero.
1141 return DAG.getNode(ISD::AssertZext, SL, MVT::i32, Param,
1142 DAG.getValueType(VT));
1145 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
1146 SelectionDAG &DAG) const {
1147 MachineFunction &MF = DAG.getMachineFunction();
1148 auto MFI = MF.getInfo<SIMachineFunctionInfo>();
1149 const SIRegisterInfo *TRI =
1150 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
1152 EVT VT = Op.getValueType();
1154 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1156 // TODO: Should this propagate fast-math-flags?
1158 switch (IntrinsicID) {
1159 case Intrinsic::amdgcn_dispatch_ptr:
1160 if (!Subtarget->isAmdHsaOS()) {
1161 DiagnosticInfoUnsupported BadIntrin(*MF.getFunction(),
1162 "hsa intrinsic without hsa target");
1163 DAG.getContext()->diagnose(BadIntrin);
1164 return DAG.getUNDEF(VT);
1167 return CreateLiveInRegister(DAG, &AMDGPU::SReg_64RegClass,
1168 TRI->getPreloadedValue(MF, SIRegisterInfo::DISPATCH_PTR), VT);
1170 case Intrinsic::r600_read_ngroups_x:
1171 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
1172 SI::KernelInputOffsets::NGROUPS_X, false);
1173 case Intrinsic::r600_read_ngroups_y:
1174 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
1175 SI::KernelInputOffsets::NGROUPS_Y, false);
1176 case Intrinsic::r600_read_ngroups_z:
1177 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
1178 SI::KernelInputOffsets::NGROUPS_Z, false);
1179 case Intrinsic::r600_read_global_size_x:
1180 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
1181 SI::KernelInputOffsets::GLOBAL_SIZE_X, false);
1182 case Intrinsic::r600_read_global_size_y:
1183 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
1184 SI::KernelInputOffsets::GLOBAL_SIZE_Y, false);
1185 case Intrinsic::r600_read_global_size_z:
1186 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
1187 SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
1188 case Intrinsic::r600_read_local_size_x:
1189 return lowerImplicitZextParam(DAG, Op, MVT::i16,
1190 SI::KernelInputOffsets::LOCAL_SIZE_X);
1191 case Intrinsic::r600_read_local_size_y:
1192 return lowerImplicitZextParam(DAG, Op, MVT::i16,
1193 SI::KernelInputOffsets::LOCAL_SIZE_Y);
1194 case Intrinsic::r600_read_local_size_z:
1195 return lowerImplicitZextParam(DAG, Op, MVT::i16,
1196 SI::KernelInputOffsets::LOCAL_SIZE_Z);
1197 case Intrinsic::AMDGPU_read_workdim:
1198 // Really only 2 bits.
1199 return lowerImplicitZextParam(DAG, Op, MVT::i8,
1200 getImplicitParameterOffset(MFI, GRID_DIM));
1201 case Intrinsic::r600_read_tgid_x:
1202 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
1203 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_X), VT);
1204 case Intrinsic::r600_read_tgid_y:
1205 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
1206 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Y), VT);
1207 case Intrinsic::r600_read_tgid_z:
1208 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
1209 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Z), VT);
1210 case Intrinsic::r600_read_tidig_x:
1211 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
1212 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X), VT);
1213 case Intrinsic::r600_read_tidig_y:
1214 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
1215 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y), VT);
1216 case Intrinsic::r600_read_tidig_z:
1217 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
1218 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z), VT);
1219 case AMDGPUIntrinsic::SI_load_const: {
1225 MachineMemOperand *MMO = MF.getMachineMemOperand(
1226 MachinePointerInfo(),
1227 MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
1228 VT.getStoreSize(), 4);
1229 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
1230 Op->getVTList(), Ops, VT, MMO);
1232 case AMDGPUIntrinsic::SI_sample:
1233 return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
1234 case AMDGPUIntrinsic::SI_sampleb:
1235 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
1236 case AMDGPUIntrinsic::SI_sampled:
1237 return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
1238 case AMDGPUIntrinsic::SI_samplel:
1239 return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
1240 case AMDGPUIntrinsic::SI_vs_load_input:
1241 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
1246 case AMDGPUIntrinsic::AMDGPU_fract:
1247 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
1248 return DAG.getNode(ISD::FSUB, DL, VT, Op.getOperand(1),
1249 DAG.getNode(ISD::FFLOOR, DL, VT, Op.getOperand(1)));
1250 case AMDGPUIntrinsic::SI_fs_constant: {
1251 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
1252 SDValue Glue = M0.getValue(1);
1253 return DAG.getNode(AMDGPUISD::INTERP_MOV, DL, MVT::f32,
1254 DAG.getConstant(2, DL, MVT::i32), // P0
1255 Op.getOperand(1), Op.getOperand(2), Glue);
1257 case AMDGPUIntrinsic::SI_packf16:
1258 if (Op.getOperand(1).isUndef() && Op.getOperand(2).isUndef())
1259 return DAG.getUNDEF(MVT::i32);
1261 case AMDGPUIntrinsic::SI_fs_interp: {
1262 SDValue IJ = Op.getOperand(4);
1263 SDValue I = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
1264 DAG.getConstant(0, DL, MVT::i32));
1265 SDValue J = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
1266 DAG.getConstant(1, DL, MVT::i32));
1267 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
1268 SDValue Glue = M0.getValue(1);
1269 SDValue P1 = DAG.getNode(AMDGPUISD::INTERP_P1, DL,
1270 DAG.getVTList(MVT::f32, MVT::Glue),
1271 I, Op.getOperand(1), Op.getOperand(2), Glue);
1272 Glue = SDValue(P1.getNode(), 1);
1273 return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, P1, J,
1274 Op.getOperand(1), Op.getOperand(2), Glue);
1276 case Intrinsic::amdgcn_interp_p1: {
1277 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(4));
1278 SDValue Glue = M0.getValue(1);
1279 return DAG.getNode(AMDGPUISD::INTERP_P1, DL, MVT::f32, Op.getOperand(1),
1280 Op.getOperand(2), Op.getOperand(3), Glue);
1282 case Intrinsic::amdgcn_interp_p2: {
1283 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(5));
1284 SDValue Glue = SDValue(M0.getNode(), 1);
1285 return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, Op.getOperand(1),
1286 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4),
1290 return AMDGPUTargetLowering::LowerOperation(Op, DAG);
1294 SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
1295 SelectionDAG &DAG) const {
1296 MachineFunction &MF = DAG.getMachineFunction();
1298 SDValue Chain = Op.getOperand(0);
1299 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1301 switch (IntrinsicID) {
1302 case AMDGPUIntrinsic::SI_sendmsg: {
1303 Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3));
1304 SDValue Glue = Chain.getValue(1);
1305 return DAG.getNode(AMDGPUISD::SENDMSG, DL, MVT::Other, Chain,
1306 Op.getOperand(2), Glue);
1308 case AMDGPUIntrinsic::SI_tbuffer_store: {
1326 EVT VT = Op.getOperand(3).getValueType();
1328 MachineMemOperand *MMO = MF.getMachineMemOperand(
1329 MachinePointerInfo(),
1330 MachineMemOperand::MOStore,
1331 VT.getStoreSize(), 4);
1332 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
1333 Op->getVTList(), Ops, VT, MMO);
1340 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1342 LoadSDNode *Load = cast<LoadSDNode>(Op);
1344 if (Op.getValueType().isVector()) {
1345 assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
1346 "Custom lowering for non-i32 vectors hasn't been implemented.");
1347 unsigned NumElements = Op.getValueType().getVectorNumElements();
1348 assert(NumElements != 2 && "v2 loads are supported for all address spaces.");
1350 switch (Load->getAddressSpace()) {
1352 case AMDGPUAS::CONSTANT_ADDRESS:
1353 if (isMemOpUniform(Load))
1355 // Non-uniform loads will be selected to MUBUF instructions, so they
1356 // have the same legalization requires ments as global and private
1360 case AMDGPUAS::GLOBAL_ADDRESS:
1361 case AMDGPUAS::PRIVATE_ADDRESS:
1362 if (NumElements >= 8)
1363 return SplitVectorLoad(Op, DAG);
1365 // v4 loads are supported for private and global memory.
1366 if (NumElements <= 4)
1369 case AMDGPUAS::LOCAL_ADDRESS:
1370 // If properly aligned, if we split we might be able to use ds_read_b64.
1371 return SplitVectorLoad(Op, DAG);
1375 return AMDGPUTargetLowering::LowerLOAD(Op, DAG);
1378 SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
1380 SelectionDAG &DAG) const {
1381 return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
1387 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
1388 if (Op.getValueType() != MVT::i64)
1392 SDValue Cond = Op.getOperand(0);
1394 SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
1395 SDValue One = DAG.getConstant(1, DL, MVT::i32);
1397 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
1398 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
1400 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
1401 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
1403 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
1405 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
1406 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
1408 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
1410 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
1411 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
1414 // Catch division cases where we can use shortcuts with rcp and rsq
1416 SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
1418 SDValue LHS = Op.getOperand(0);
1419 SDValue RHS = Op.getOperand(1);
1420 EVT VT = Op.getValueType();
1421 bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
1423 if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
1424 if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
1425 CLHS->isExactlyValue(1.0)) {
1426 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
1427 // the CI documentation has a worst case error of 1 ulp.
1428 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
1429 // use it as long as we aren't trying to use denormals.
1431 // 1.0 / sqrt(x) -> rsq(x)
1433 // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
1434 // error seems really high at 2^29 ULP.
1435 if (RHS.getOpcode() == ISD::FSQRT)
1436 return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
1438 // 1.0 / x -> rcp(x)
1439 return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1444 // Turn into multiply by the reciprocal.
1445 // x / y -> x * (1.0 / y)
1447 Flags.setUnsafeAlgebra(true);
1448 SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
1449 return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip, &Flags);
1455 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
1456 SDValue FastLowered = LowerFastFDIV(Op, DAG);
1457 if (FastLowered.getNode())
1460 // This uses v_rcp_f32 which does not handle denormals. Let this hit a
1461 // selection error for now rather than do something incorrect.
1462 if (Subtarget->hasFP32Denormals())
1466 SDValue LHS = Op.getOperand(0);
1467 SDValue RHS = Op.getOperand(1);
1469 SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);
1471 const APFloat K0Val(BitsToFloat(0x6f800000));
1472 const SDValue K0 = DAG.getConstantFP(K0Val, SL, MVT::f32);
1474 const APFloat K1Val(BitsToFloat(0x2f800000));
1475 const SDValue K1 = DAG.getConstantFP(K1Val, SL, MVT::f32);
1477 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
1480 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f32);
1482 SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);
1484 SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);
1486 // TODO: Should this propagate fast-math-flags?
1488 r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);
1490 SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);
1492 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);
1494 return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
1497 SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
1498 if (DAG.getTarget().Options.UnsafeFPMath)
1499 return LowerFastFDIV(Op, DAG);
1502 SDValue X = Op.getOperand(0);
1503 SDValue Y = Op.getOperand(1);
1505 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
1507 SDVTList ScaleVT = DAG.getVTList(MVT::f64, MVT::i1);
1509 SDValue DivScale0 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, Y, Y, X);
1511 SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f64, DivScale0);
1513 SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f64, DivScale0);
1515 SDValue Fma0 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Rcp, One);
1517 SDValue Fma1 = DAG.getNode(ISD::FMA, SL, MVT::f64, Rcp, Fma0, Rcp);
1519 SDValue Fma2 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Fma1, One);
1521 SDValue DivScale1 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, X, Y, X);
1523 SDValue Fma3 = DAG.getNode(ISD::FMA, SL, MVT::f64, Fma1, Fma2, Fma1);
1524 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, DivScale1, Fma3);
1526 SDValue Fma4 = DAG.getNode(ISD::FMA, SL, MVT::f64,
1527 NegDivScale0, Mul, DivScale1);
1531 if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) {
1532 // Workaround a hardware bug on SI where the condition output from div_scale
1535 const SDValue Hi = DAG.getConstant(1, SL, MVT::i32);
1537 // Figure out if the scale to use for div_fmas.
1538 SDValue NumBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
1539 SDValue DenBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Y);
1540 SDValue Scale0BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale0);
1541 SDValue Scale1BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale1);
1543 SDValue NumHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, NumBC, Hi);
1544 SDValue DenHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, DenBC, Hi);
1547 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale0BC, Hi);
1549 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale1BC, Hi);
1551 SDValue CmpDen = DAG.getSetCC(SL, MVT::i1, DenHi, Scale0Hi, ISD::SETEQ);
1552 SDValue CmpNum = DAG.getSetCC(SL, MVT::i1, NumHi, Scale1Hi, ISD::SETEQ);
1553 Scale = DAG.getNode(ISD::XOR, SL, MVT::i1, CmpNum, CmpDen);
1555 Scale = DivScale1.getValue(1);
1558 SDValue Fmas = DAG.getNode(AMDGPUISD::DIV_FMAS, SL, MVT::f64,
1559 Fma4, Fma3, Mul, Scale);
1561 return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f64, Fmas, Y, X);
1564 SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
1565 EVT VT = Op.getValueType();
1568 return LowerFDIV32(Op, DAG);
1571 return LowerFDIV64(Op, DAG);
1573 llvm_unreachable("Unexpected type for fdiv");
1576 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1578 StoreSDNode *Store = cast<StoreSDNode>(Op);
1579 EVT VT = Store->getMemoryVT();
1581 // These stores are legal.
1582 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
1583 if (VT.isVector() && VT.getVectorNumElements() > 4)
1584 return ScalarizeVectorStore(Op, DAG);
1588 SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
1592 if (VT.isVector() && VT.getVectorNumElements() >= 8)
1593 return SplitVectorStore(Op, DAG);
1596 return DAG.getTruncStore(Store->getChain(), DL,
1597 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
1598 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
1603 SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
1605 EVT VT = Op.getValueType();
1606 SDValue Arg = Op.getOperand(0);
1607 // TODO: Should this propagate fast-math-flags?
1608 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT,
1609 DAG.getNode(ISD::FMUL, DL, VT, Arg,
1610 DAG.getConstantFP(0.5/M_PI, DL,
1613 switch (Op.getOpcode()) {
1615 return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
1617 return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
1619 llvm_unreachable("Wrong trig opcode");
1623 //===----------------------------------------------------------------------===//
1624 // Custom DAG optimizations
1625 //===----------------------------------------------------------------------===//
1627 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
1628 DAGCombinerInfo &DCI) const {
1629 EVT VT = N->getValueType(0);
1630 EVT ScalarVT = VT.getScalarType();
1631 if (ScalarVT != MVT::f32)
1634 SelectionDAG &DAG = DCI.DAG;
1637 SDValue Src = N->getOperand(0);
1638 EVT SrcVT = Src.getValueType();
1640 // TODO: We could try to match extracting the higher bytes, which would be
1641 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
1642 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
1643 // about in practice.
1644 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
1645 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
1646 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
1647 DCI.AddToWorklist(Cvt.getNode());
1652 // We are primarily trying to catch operations on illegal vector types
1653 // before they are expanded.
1654 // For scalars, we can use the more flexible method of checking masked bits
1655 // after legalization.
1656 if (!DCI.isBeforeLegalize() ||
1657 !SrcVT.isVector() ||
1658 SrcVT.getVectorElementType() != MVT::i8) {
1662 assert(DCI.isBeforeLegalize() && "Unexpected legal type");
1664 // Weird sized vectors are a pain to handle, but we know 3 is really the same
1666 unsigned NElts = SrcVT.getVectorNumElements();
1667 if (!SrcVT.isSimple() && NElts != 3)
1670 // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
1671 // prevent a mess from expanding to v4i32 and repacking.
1672 if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
1673 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
1674 EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
1675 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
1676 LoadSDNode *Load = cast<LoadSDNode>(Src);
1678 unsigned AS = Load->getAddressSpace();
1679 unsigned Align = Load->getAlignment();
1680 Type *Ty = LoadVT.getTypeForEVT(*DAG.getContext());
1681 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
1683 // Don't try to replace the load if we have to expand it due to alignment
1684 // problems. Otherwise we will end up scalarizing the load, and trying to
1685 // repack into the vector for no real reason.
1686 if (Align < ABIAlignment &&
1687 !allowsMisalignedMemoryAccesses(LoadVT, AS, Align, nullptr)) {
1691 SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
1695 Load->getMemOperand());
1697 // Make sure successors of the original load stay after it by updating
1698 // them to use the new Chain.
1699 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
1701 SmallVector<SDValue, 4> Elts;
1702 if (RegVT.isVector())
1703 DAG.ExtractVectorElements(NewLoad, Elts);
1705 Elts.push_back(NewLoad);
1707 SmallVector<SDValue, 4> Ops;
1709 unsigned EltIdx = 0;
1710 for (SDValue Elt : Elts) {
1711 unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
1712 for (unsigned I = 0; I < ComponentsInElt; ++I) {
1713 unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
1714 SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
1715 DCI.AddToWorklist(Cvt.getNode());
1722 assert(Ops.size() == NElts);
1724 return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
1730 /// \brief Return true if the given offset Size in bytes can be folded into
1731 /// the immediate offsets of a memory instruction for the given address space.
1732 static bool canFoldOffset(unsigned OffsetSize, unsigned AS,
1733 const AMDGPUSubtarget &STI) {
1735 case AMDGPUAS::GLOBAL_ADDRESS: {
1736 // MUBUF instructions a 12-bit offset in bytes.
1737 return isUInt<12>(OffsetSize);
1739 case AMDGPUAS::CONSTANT_ADDRESS: {
1740 // SMRD instructions have an 8-bit offset in dwords on SI and
1741 // a 20-bit offset in bytes on VI.
1742 if (STI.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1743 return isUInt<20>(OffsetSize);
1745 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
1747 case AMDGPUAS::LOCAL_ADDRESS:
1748 case AMDGPUAS::REGION_ADDRESS: {
1749 // The single offset versions have a 16-bit offset in bytes.
1750 return isUInt<16>(OffsetSize);
1752 case AMDGPUAS::PRIVATE_ADDRESS:
1753 // Indirect register addressing does not use any offsets.
1759 // (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)
1761 // This is a variant of
1762 // (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
1764 // The normal DAG combiner will do this, but only if the add has one use since
1765 // that would increase the number of instructions.
1767 // This prevents us from seeing a constant offset that can be folded into a
1768 // memory instruction's addressing mode. If we know the resulting add offset of
1769 // a pointer can be folded into an addressing offset, we can replace the pointer
1770 // operand with the add of new constant offset. This eliminates one of the uses,
1771 // and may allow the remaining use to also be simplified.
1773 SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
1775 DAGCombinerInfo &DCI) const {
1776 SDValue N0 = N->getOperand(0);
1777 SDValue N1 = N->getOperand(1);
1779 if (N0.getOpcode() != ISD::ADD)
1782 const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
1786 const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1790 // If the resulting offset is too large, we can't fold it into the addressing
1792 APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
1793 if (!canFoldOffset(Offset.getZExtValue(), AddrSpace, *Subtarget))
1796 SelectionDAG &DAG = DCI.DAG;
1798 EVT VT = N->getValueType(0);
1800 SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
1801 SDValue COffset = DAG.getConstant(Offset, SL, MVT::i32);
1803 return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
1806 SDValue SITargetLowering::performAndCombine(SDNode *N,
1807 DAGCombinerInfo &DCI) const {
1808 if (DCI.isBeforeLegalize())
1811 SelectionDAG &DAG = DCI.DAG;
1813 // (and (fcmp ord x, x), (fcmp une (fabs x), inf)) ->
1814 // fp_class x, ~(s_nan | q_nan | n_infinity | p_infinity)
1815 SDValue LHS = N->getOperand(0);
1816 SDValue RHS = N->getOperand(1);
1818 if (LHS.getOpcode() == ISD::SETCC &&
1819 RHS.getOpcode() == ISD::SETCC) {
1820 ISD::CondCode LCC = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
1821 ISD::CondCode RCC = cast<CondCodeSDNode>(RHS.getOperand(2))->get();
1823 SDValue X = LHS.getOperand(0);
1824 SDValue Y = RHS.getOperand(0);
1825 if (Y.getOpcode() != ISD::FABS || Y.getOperand(0) != X)
1828 if (LCC == ISD::SETO) {
1829 if (X != LHS.getOperand(1))
1832 if (RCC == ISD::SETUNE) {
1833 const ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(RHS.getOperand(1));
1834 if (!C1 || !C1->isInfinity() || C1->isNegative())
1837 const uint32_t Mask = SIInstrFlags::N_NORMAL |
1838 SIInstrFlags::N_SUBNORMAL |
1839 SIInstrFlags::N_ZERO |
1840 SIInstrFlags::P_ZERO |
1841 SIInstrFlags::P_SUBNORMAL |
1842 SIInstrFlags::P_NORMAL;
1844 static_assert(((~(SIInstrFlags::S_NAN |
1845 SIInstrFlags::Q_NAN |
1846 SIInstrFlags::N_INFINITY |
1847 SIInstrFlags::P_INFINITY)) & 0x3ff) == Mask,
1851 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
1852 X, DAG.getConstant(Mask, DL, MVT::i32));
1860 SDValue SITargetLowering::performOrCombine(SDNode *N,
1861 DAGCombinerInfo &DCI) const {
1862 SelectionDAG &DAG = DCI.DAG;
1863 SDValue LHS = N->getOperand(0);
1864 SDValue RHS = N->getOperand(1);
1866 // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2)
1867 if (LHS.getOpcode() == AMDGPUISD::FP_CLASS &&
1868 RHS.getOpcode() == AMDGPUISD::FP_CLASS) {
1869 SDValue Src = LHS.getOperand(0);
1870 if (Src != RHS.getOperand(0))
1873 const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
1874 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1));
1878 // Only 10 bits are used.
1879 static const uint32_t MaxMask = 0x3ff;
1881 uint32_t NewMask = (CLHS->getZExtValue() | CRHS->getZExtValue()) & MaxMask;
1883 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
1884 Src, DAG.getConstant(NewMask, DL, MVT::i32));
1890 SDValue SITargetLowering::performClassCombine(SDNode *N,
1891 DAGCombinerInfo &DCI) const {
1892 SelectionDAG &DAG = DCI.DAG;
1893 SDValue Mask = N->getOperand(1);
1895 // fp_class x, 0 -> false
1896 if (const ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Mask)) {
1897 if (CMask->isNullValue())
1898 return DAG.getConstant(0, SDLoc(N), MVT::i1);
1904 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {
1907 return AMDGPUISD::FMAX3;
1909 return AMDGPUISD::SMAX3;
1911 return AMDGPUISD::UMAX3;
1913 return AMDGPUISD::FMIN3;
1915 return AMDGPUISD::SMIN3;
1917 return AMDGPUISD::UMIN3;
1919 llvm_unreachable("Not a min/max opcode");
1923 SDValue SITargetLowering::performMin3Max3Combine(SDNode *N,
1924 DAGCombinerInfo &DCI) const {
1925 SelectionDAG &DAG = DCI.DAG;
1927 unsigned Opc = N->getOpcode();
1928 SDValue Op0 = N->getOperand(0);
1929 SDValue Op1 = N->getOperand(1);
1931 // Only do this if the inner op has one use since this will just increases
1932 // register pressure for no benefit.
1934 // max(max(a, b), c)
1935 if (Op0.getOpcode() == Opc && Op0.hasOneUse()) {
1937 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
1945 // max(a, max(b, c))
1946 if (Op1.getOpcode() == Opc && Op1.hasOneUse()) {
1948 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
1959 SDValue SITargetLowering::performSetCCCombine(SDNode *N,
1960 DAGCombinerInfo &DCI) const {
1961 SelectionDAG &DAG = DCI.DAG;
1964 SDValue LHS = N->getOperand(0);
1965 SDValue RHS = N->getOperand(1);
1966 EVT VT = LHS.getValueType();
1968 if (VT != MVT::f32 && VT != MVT::f64)
1971 // Match isinf pattern
1972 // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity))
1973 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
1974 if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) {
1975 const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
1979 const APFloat &APF = CRHS->getValueAPF();
1980 if (APF.isInfinity() && !APF.isNegative()) {
1981 unsigned Mask = SIInstrFlags::P_INFINITY | SIInstrFlags::N_INFINITY;
1982 return DAG.getNode(AMDGPUISD::FP_CLASS, SL, MVT::i1, LHS.getOperand(0),
1983 DAG.getConstant(Mask, SL, MVT::i32));
1990 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
1991 DAGCombinerInfo &DCI) const {
1992 SelectionDAG &DAG = DCI.DAG;
1995 switch (N->getOpcode()) {
1997 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1999 return performSetCCCombine(N, DCI);
2000 case ISD::FMAXNUM: // TODO: What about fmax_legacy?
2006 if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
2007 N->getValueType(0) != MVT::f64 &&
2008 getTargetMachine().getOptLevel() > CodeGenOpt::None)
2009 return performMin3Max3Combine(N, DCI);
2013 case AMDGPUISD::CVT_F32_UBYTE0:
2014 case AMDGPUISD::CVT_F32_UBYTE1:
2015 case AMDGPUISD::CVT_F32_UBYTE2:
2016 case AMDGPUISD::CVT_F32_UBYTE3: {
2017 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
2019 SDValue Src = N->getOperand(0);
2020 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
2022 APInt KnownZero, KnownOne;
2023 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2024 !DCI.isBeforeLegalizeOps());
2025 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2026 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
2027 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
2028 DCI.CommitTargetLoweringOpt(TLO);
2034 case ISD::UINT_TO_FP: {
2035 return performUCharToFloatCombine(N, DCI);
2038 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
2041 EVT VT = N->getValueType(0);
2045 // Only do this if we are not trying to support denormals. v_mad_f32 does
2046 // not support denormals ever.
2047 if (Subtarget->hasFP32Denormals())
2050 SDValue LHS = N->getOperand(0);
2051 SDValue RHS = N->getOperand(1);
2053 // These should really be instruction patterns, but writing patterns with
2054 // source modiifiers is a pain.
2056 // fadd (fadd (a, a), b) -> mad 2.0, a, b
2057 if (LHS.getOpcode() == ISD::FADD) {
2058 SDValue A = LHS.getOperand(0);
2059 if (A == LHS.getOperand(1)) {
2060 const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
2061 return DAG.getNode(ISD::FMAD, DL, VT, Two, A, RHS);
2065 // fadd (b, fadd (a, a)) -> mad 2.0, a, b
2066 if (RHS.getOpcode() == ISD::FADD) {
2067 SDValue A = RHS.getOperand(0);
2068 if (A == RHS.getOperand(1)) {
2069 const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
2070 return DAG.getNode(ISD::FMAD, DL, VT, Two, A, LHS);
2077 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
2080 EVT VT = N->getValueType(0);
2082 // Try to get the fneg to fold into the source modifier. This undoes generic
2083 // DAG combines and folds them into the mad.
2085 // Only do this if we are not trying to support denormals. v_mad_f32 does
2086 // not support denormals ever.
2087 if (VT == MVT::f32 &&
2088 !Subtarget->hasFP32Denormals()) {
2089 SDValue LHS = N->getOperand(0);
2090 SDValue RHS = N->getOperand(1);
2091 if (LHS.getOpcode() == ISD::FADD) {
2092 // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c)
2094 SDValue A = LHS.getOperand(0);
2095 if (A == LHS.getOperand(1)) {
2096 const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
2097 SDValue NegRHS = DAG.getNode(ISD::FNEG, DL, VT, RHS);
2099 return DAG.getNode(ISD::FMAD, DL, VT, Two, A, NegRHS);
2103 if (RHS.getOpcode() == ISD::FADD) {
2104 // (fsub c, (fadd a, a)) -> mad -2.0, a, c
2106 SDValue A = RHS.getOperand(0);
2107 if (A == RHS.getOperand(1)) {
2108 const SDValue NegTwo = DAG.getConstantFP(-2.0, DL, MVT::f32);
2109 return DAG.getNode(ISD::FMAD, DL, VT, NegTwo, A, LHS);
2120 case ISD::ATOMIC_LOAD:
2121 case ISD::ATOMIC_STORE:
2122 case ISD::ATOMIC_CMP_SWAP:
2123 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
2124 case ISD::ATOMIC_SWAP:
2125 case ISD::ATOMIC_LOAD_ADD:
2126 case ISD::ATOMIC_LOAD_SUB:
2127 case ISD::ATOMIC_LOAD_AND:
2128 case ISD::ATOMIC_LOAD_OR:
2129 case ISD::ATOMIC_LOAD_XOR:
2130 case ISD::ATOMIC_LOAD_NAND:
2131 case ISD::ATOMIC_LOAD_MIN:
2132 case ISD::ATOMIC_LOAD_MAX:
2133 case ISD::ATOMIC_LOAD_UMIN:
2134 case ISD::ATOMIC_LOAD_UMAX: { // TODO: Target mem intrinsics.
2135 if (DCI.isBeforeLegalize())
2138 MemSDNode *MemNode = cast<MemSDNode>(N);
2139 SDValue Ptr = MemNode->getBasePtr();
2141 // TODO: We could also do this for multiplies.
2142 unsigned AS = MemNode->getAddressSpace();
2143 if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
2144 SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
2146 SmallVector<SDValue, 8> NewOps(MemNode->op_begin(), MemNode->op_end());
2148 NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
2149 return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0);
2155 return performAndCombine(N, DCI);
2157 return performOrCombine(N, DCI);
2158 case AMDGPUISD::FP_CLASS:
2159 return performClassCombine(N, DCI);
2161 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
2164 /// \brief Analyze the possible immediate value Op
2166 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
2167 /// and the immediate value if it's a literal immediate
2168 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
2170 const SIInstrInfo *TII =
2171 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2173 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
2174 if (TII->isInlineConstant(Node->getAPIntValue()))
2177 uint64_t Val = Node->getZExtValue();
2178 return isUInt<32>(Val) ? Val : -1;
2181 if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
2182 if (TII->isInlineConstant(Node->getValueAPF().bitcastToAPInt()))
2185 if (Node->getValueType(0) == MVT::f32)
2186 return FloatToBits(Node->getValueAPF().convertToFloat());
2194 /// \brief Helper function for adjustWritemask
2195 static unsigned SubIdx2Lane(unsigned Idx) {
2198 case AMDGPU::sub0: return 0;
2199 case AMDGPU::sub1: return 1;
2200 case AMDGPU::sub2: return 2;
2201 case AMDGPU::sub3: return 3;
2205 /// \brief Adjust the writemask of MIMG instructions
2206 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
2207 SelectionDAG &DAG) const {
2208 SDNode *Users[4] = { };
2210 unsigned OldDmask = Node->getConstantOperandVal(0);
2211 unsigned NewDmask = 0;
2213 // Try to figure out the used register components
2214 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
2217 // Abort if we can't understand the usage
2218 if (!I->isMachineOpcode() ||
2219 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
2222 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
2223 // Note that subregs are packed, i.e. Lane==0 is the first bit set
2224 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
2226 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
2228 // Set which texture component corresponds to the lane.
2230 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
2232 Comp = countTrailingZeros(Dmask);
2233 Dmask &= ~(1 << Comp);
2236 // Abort if we have more than one user per component
2241 NewDmask |= 1 << Comp;
2244 // Abort if there's no change
2245 if (NewDmask == OldDmask)
2248 // Adjust the writemask in the node
2249 std::vector<SDValue> Ops;
2250 Ops.push_back(DAG.getTargetConstant(NewDmask, SDLoc(Node), MVT::i32));
2251 Ops.insert(Ops.end(), Node->op_begin() + 1, Node->op_end());
2252 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
2254 // If we only got one lane, replace it with a copy
2255 // (if NewDmask has only one bit set...)
2256 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
2257 SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, SDLoc(),
2259 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
2260 SDLoc(), Users[Lane]->getValueType(0),
2261 SDValue(Node, 0), RC);
2262 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
2266 // Update the users of the node with the new indices
2267 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
2269 SDNode *User = Users[i];
2273 SDValue Op = DAG.getTargetConstant(Idx, SDLoc(User), MVT::i32);
2274 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
2278 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
2279 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
2280 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
2285 static bool isFrameIndexOp(SDValue Op) {
2286 if (Op.getOpcode() == ISD::AssertZext)
2287 Op = Op.getOperand(0);
2289 return isa<FrameIndexSDNode>(Op);
2292 /// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
2293 /// with frame index operands.
2294 /// LLVM assumes that inputs are to these instructions are registers.
2295 void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
2296 SelectionDAG &DAG) const {
2298 SmallVector<SDValue, 8> Ops;
2299 for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
2300 if (!isFrameIndexOp(Node->getOperand(i))) {
2301 Ops.push_back(Node->getOperand(i));
2306 Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL,
2307 Node->getOperand(i).getValueType(),
2308 Node->getOperand(i)), 0));
2311 DAG.UpdateNodeOperands(Node, Ops);
2314 /// \brief Fold the instructions after selecting them.
2315 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
2316 SelectionDAG &DAG) const {
2317 const SIInstrInfo *TII =
2318 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2320 if (TII->isMIMG(Node->getMachineOpcode()))
2321 adjustWritemask(Node, DAG);
2323 if (Node->getMachineOpcode() == AMDGPU::INSERT_SUBREG ||
2324 Node->getMachineOpcode() == AMDGPU::REG_SEQUENCE) {
2325 legalizeTargetIndependentNode(Node, DAG);
2331 /// \brief Assign the register class depending on the number of
2332 /// bits set in the writemask
2333 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
2334 SDNode *Node) const {
2335 const SIInstrInfo *TII =
2336 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2338 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2340 if (TII->isVOP3(MI->getOpcode())) {
2341 // Make sure constant bus requirements are respected.
2342 TII->legalizeOperandsVOP3(MRI, MI);
2346 if (TII->isMIMG(*MI)) {
2347 unsigned VReg = MI->getOperand(0).getReg();
2348 unsigned Writemask = MI->getOperand(1).getImm();
2349 unsigned BitsSet = 0;
2350 for (unsigned i = 0; i < 4; ++i)
2351 BitsSet += Writemask & (1 << i) ? 1 : 0;
2353 const TargetRegisterClass *RC;
2356 case 1: RC = &AMDGPU::VGPR_32RegClass; break;
2357 case 2: RC = &AMDGPU::VReg_64RegClass; break;
2358 case 3: RC = &AMDGPU::VReg_96RegClass; break;
2361 unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
2362 MI->setDesc(TII->get(NewOpcode));
2363 MRI.setRegClass(VReg, RC);
2367 // Replace unused atomics with the no return version.
2368 int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI->getOpcode());
2369 if (NoRetAtomicOp != -1) {
2370 if (!Node->hasAnyUseOfValue(0)) {
2371 MI->setDesc(TII->get(NoRetAtomicOp));
2372 MI->RemoveOperand(0);
2379 static SDValue buildSMovImm32(SelectionDAG &DAG, SDLoc DL, uint64_t Val) {
2380 SDValue K = DAG.getTargetConstant(Val, DL, MVT::i32);
2381 return SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, K), 0);
2384 MachineSDNode *SITargetLowering::wrapAddr64Rsrc(SelectionDAG &DAG,
2386 SDValue Ptr) const {
2387 const SIInstrInfo *TII =
2388 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2390 // Build the half of the subregister with the constants before building the
2391 // full 128-bit register. If we are building multiple resource descriptors,
2392 // this will allow CSEing of the 2-component register.
2393 const SDValue Ops0[] = {
2394 DAG.getTargetConstant(AMDGPU::SGPR_64RegClassID, DL, MVT::i32),
2395 buildSMovImm32(DAG, DL, 0),
2396 DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
2397 buildSMovImm32(DAG, DL, TII->getDefaultRsrcDataFormat() >> 32),
2398 DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
2401 SDValue SubRegHi = SDValue(DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL,
2402 MVT::v2i32, Ops0), 0);
2404 // Combine the constants and the pointer.
2405 const SDValue Ops1[] = {
2406 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
2408 DAG.getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32),
2410 DAG.getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32)
2413 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops1);
2416 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
2417 /// The TID (Thread ID) is multiplied by the stride value (bits [61:48]
2418 /// of the resource descriptor) to create an offset, which is added to
2419 /// the resource pointer.
2420 MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG,
2423 uint32_t RsrcDword1,
2424 uint64_t RsrcDword2And3) const {
2425 SDValue PtrLo = DAG.getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
2426 SDValue PtrHi = DAG.getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
2428 PtrHi = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
2429 DAG.getConstant(RsrcDword1, DL, MVT::i32)),
2433 SDValue DataLo = buildSMovImm32(DAG, DL,
2434 RsrcDword2And3 & UINT64_C(0xFFFFFFFF));
2435 SDValue DataHi = buildSMovImm32(DAG, DL, RsrcDword2And3 >> 32);
2437 const SDValue Ops[] = {
2438 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
2440 DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
2442 DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
2444 DAG.getTargetConstant(AMDGPU::sub2, DL, MVT::i32),
2446 DAG.getTargetConstant(AMDGPU::sub3, DL, MVT::i32)
2449 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
2452 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2453 const TargetRegisterClass *RC,
2454 unsigned Reg, EVT VT) const {
2455 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
2457 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
2458 cast<RegisterSDNode>(VReg)->getReg(), VT);
2461 //===----------------------------------------------------------------------===//
2462 // SI Inline Assembly Support
2463 //===----------------------------------------------------------------------===//
2465 std::pair<unsigned, const TargetRegisterClass *>
2466 SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
2467 StringRef Constraint,
2470 if (Constraint.size() == 1) {
2471 switch (Constraint[0]) {
2474 switch (VT.getSizeInBits()) {
2476 return std::make_pair(0U, nullptr);
2478 return std::make_pair(0U, &AMDGPU::SGPR_32RegClass);
2480 return std::make_pair(0U, &AMDGPU::SGPR_64RegClass);
2482 return std::make_pair(0U, &AMDGPU::SReg_128RegClass);
2484 return std::make_pair(0U, &AMDGPU::SReg_256RegClass);
2488 switch (VT.getSizeInBits()) {
2490 return std::make_pair(0U, nullptr);
2492 return std::make_pair(0U, &AMDGPU::VGPR_32RegClass);
2494 return std::make_pair(0U, &AMDGPU::VReg_64RegClass);
2496 return std::make_pair(0U, &AMDGPU::VReg_96RegClass);
2498 return std::make_pair(0U, &AMDGPU::VReg_128RegClass);
2500 return std::make_pair(0U, &AMDGPU::VReg_256RegClass);
2502 return std::make_pair(0U, &AMDGPU::VReg_512RegClass);
2507 if (Constraint.size() > 1) {
2508 const TargetRegisterClass *RC = nullptr;
2509 if (Constraint[1] == 'v') {
2510 RC = &AMDGPU::VGPR_32RegClass;
2511 } else if (Constraint[1] == 's') {
2512 RC = &AMDGPU::SGPR_32RegClass;
2517 bool Failed = Constraint.substr(2).getAsInteger(10, Idx);
2518 if (!Failed && Idx < RC->getNumRegs())
2519 return std::make_pair(RC->getRegister(Idx), RC);
2522 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
2525 SITargetLowering::ConstraintType
2526 SITargetLowering::getConstraintType(StringRef Constraint) const {
2527 if (Constraint.size() == 1) {
2528 switch (Constraint[0]) {
2532 return C_RegisterClass;
2535 return TargetLowering::getConstraintType(Constraint);