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
22 #include "AMDGPUIntrinsicInfo.h"
23 #include "AMDGPUSubtarget.h"
24 #include "SIDefines.h"
25 #include "SIISelLowering.h"
26 #include "SIInstrInfo.h"
27 #include "SIMachineFunctionInfo.h"
28 #include "SIRegisterInfo.h"
29 #include "llvm/ADT/BitVector.h"
30 #include "llvm/ADT/StringSwitch.h"
31 #include "llvm/CodeGen/CallingConvLower.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/CodeGen/Analysis.h"
36 #include "llvm/IR/DiagnosticInfo.h"
37 #include "llvm/IR/Function.h"
41 static cl::opt<bool> EnableVGPRIndexMode(
42 "amdgpu-vgpr-index-mode",
43 cl::desc("Use GPR indexing mode instead of movrel for vector indexing"),
47 static unsigned findFirstFreeSGPR(CCState &CCInfo) {
48 unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
49 for (unsigned Reg = 0; Reg < NumSGPRs; ++Reg) {
50 if (!CCInfo.isAllocated(AMDGPU::SGPR0 + Reg)) {
51 return AMDGPU::SGPR0 + Reg;
54 llvm_unreachable("Cannot allocate sgpr");
57 SITargetLowering::SITargetLowering(const TargetMachine &TM,
58 const SISubtarget &STI)
59 : AMDGPUTargetLowering(TM, STI) {
60 addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
61 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
63 addRegisterClass(MVT::i32, &AMDGPU::SReg_32_XM0RegClass);
64 addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);
66 addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
67 addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
68 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
70 addRegisterClass(MVT::v2i64, &AMDGPU::SReg_128RegClass);
71 addRegisterClass(MVT::v2f64, &AMDGPU::SReg_128RegClass);
73 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
74 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
76 addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
77 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
79 addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass);
80 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
82 if (Subtarget->has16BitInsts()) {
83 addRegisterClass(MVT::i16, &AMDGPU::SReg_32_XM0RegClass);
84 addRegisterClass(MVT::f16, &AMDGPU::SReg_32_XM0RegClass);
87 computeRegisterProperties(STI.getRegisterInfo());
89 // We need to custom lower vector stores from local memory
90 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
91 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
92 setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
93 setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
94 setOperationAction(ISD::LOAD, MVT::i1, Custom);
96 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
97 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
98 setOperationAction(ISD::STORE, MVT::v8i32, Custom);
99 setOperationAction(ISD::STORE, MVT::v16i32, Custom);
100 setOperationAction(ISD::STORE, MVT::i1, Custom);
102 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
103 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
104 setOperationAction(ISD::ConstantPool, MVT::v2i64, Expand);
106 setOperationAction(ISD::SELECT, MVT::i1, Promote);
107 setOperationAction(ISD::SELECT, MVT::i64, Custom);
108 setOperationAction(ISD::SELECT, MVT::f64, Promote);
109 AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
111 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
112 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
113 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
114 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
115 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
117 setOperationAction(ISD::SETCC, MVT::i1, Promote);
118 setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
119 setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
120 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
122 setOperationAction(ISD::TRUNCATE, MVT::v2i32, Expand);
123 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
125 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
130 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
133 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
134 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
135 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
137 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
138 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
139 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
140 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
141 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
142 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
144 // We only support LOAD/STORE and vector manipulation ops for vectors
145 // with > 4 elements.
146 for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32, MVT::v2i64, MVT::v2f64}) {
147 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
151 case ISD::BUILD_VECTOR:
153 case ISD::EXTRACT_VECTOR_ELT:
154 case ISD::INSERT_VECTOR_ELT:
155 case ISD::INSERT_SUBVECTOR:
156 case ISD::EXTRACT_SUBVECTOR:
157 case ISD::SCALAR_TO_VECTOR:
159 case ISD::CONCAT_VECTORS:
160 setOperationAction(Op, VT, Custom);
163 setOperationAction(Op, VT, Expand);
169 // TODO: For dynamic 64-bit vector inserts/extracts, should emit a pseudo that
170 // is expanded to avoid having two separate loops in case the index is a VGPR.
172 // Most operations are naturally 32-bit vector operations. We only support
173 // load and store of i64 vectors, so promote v2i64 vector operations to v4i32.
174 for (MVT Vec64 : { MVT::v2i64, MVT::v2f64 }) {
175 setOperationAction(ISD::BUILD_VECTOR, Vec64, Promote);
176 AddPromotedToType(ISD::BUILD_VECTOR, Vec64, MVT::v4i32);
178 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Vec64, Promote);
179 AddPromotedToType(ISD::EXTRACT_VECTOR_ELT, Vec64, MVT::v4i32);
181 setOperationAction(ISD::INSERT_VECTOR_ELT, Vec64, Promote);
182 AddPromotedToType(ISD::INSERT_VECTOR_ELT, Vec64, MVT::v4i32);
184 setOperationAction(ISD::SCALAR_TO_VECTOR, Vec64, Promote);
185 AddPromotedToType(ISD::SCALAR_TO_VECTOR, Vec64, MVT::v4i32);
188 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
189 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
190 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
191 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
193 // BUFFER/FLAT_ATOMIC_CMP_SWAP on GCN GPUs needs input marshalling,
194 // and output demarshalling
195 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
196 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
198 // We can't return success/failure, only the old value,
199 // let LLVM add the comparison
200 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i32, Expand);
201 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i64, Expand);
203 if (getSubtarget()->hasFlatAddressSpace()) {
204 setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
205 setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom);
208 setOperationAction(ISD::BSWAP, MVT::i32, Legal);
209 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
211 // On SI this is s_memtime and s_memrealtime on VI.
212 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
213 setOperationAction(ISD::TRAP, MVT::Other, Custom);
215 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
216 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
218 if (Subtarget->getGeneration() >= SISubtarget::SEA_ISLANDS) {
219 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
220 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
221 setOperationAction(ISD::FRINT, MVT::f64, Legal);
224 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
226 setOperationAction(ISD::FSIN, MVT::f32, Custom);
227 setOperationAction(ISD::FCOS, MVT::f32, Custom);
228 setOperationAction(ISD::FDIV, MVT::f32, Custom);
229 setOperationAction(ISD::FDIV, MVT::f64, Custom);
231 if (Subtarget->has16BitInsts()) {
232 setOperationAction(ISD::Constant, MVT::i16, Legal);
234 setOperationAction(ISD::SMIN, MVT::i16, Legal);
235 setOperationAction(ISD::SMAX, MVT::i16, Legal);
237 setOperationAction(ISD::UMIN, MVT::i16, Legal);
238 setOperationAction(ISD::UMAX, MVT::i16, Legal);
240 setOperationAction(ISD::SIGN_EXTEND, MVT::i16, Promote);
241 AddPromotedToType(ISD::SIGN_EXTEND, MVT::i16, MVT::i32);
243 setOperationAction(ISD::ROTR, MVT::i16, Promote);
244 setOperationAction(ISD::ROTL, MVT::i16, Promote);
246 setOperationAction(ISD::SDIV, MVT::i16, Promote);
247 setOperationAction(ISD::UDIV, MVT::i16, Promote);
248 setOperationAction(ISD::SREM, MVT::i16, Promote);
249 setOperationAction(ISD::UREM, MVT::i16, Promote);
251 setOperationAction(ISD::BSWAP, MVT::i16, Promote);
252 setOperationAction(ISD::BITREVERSE, MVT::i16, Promote);
254 setOperationAction(ISD::CTTZ, MVT::i16, Promote);
255 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Promote);
256 setOperationAction(ISD::CTLZ, MVT::i16, Promote);
257 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Promote);
259 setOperationAction(ISD::SELECT_CC, MVT::i16, Expand);
261 setOperationAction(ISD::BR_CC, MVT::i16, Expand);
263 setOperationAction(ISD::LOAD, MVT::i16, Custom);
265 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
267 setOperationAction(ISD::FP16_TO_FP, MVT::i16, Promote);
268 AddPromotedToType(ISD::FP16_TO_FP, MVT::i16, MVT::i32);
269 setOperationAction(ISD::FP_TO_FP16, MVT::i16, Promote);
270 AddPromotedToType(ISD::FP_TO_FP16, MVT::i16, MVT::i32);
272 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
273 setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
274 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
275 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
277 // F16 - Constant Actions.
278 setOperationAction(ISD::ConstantFP, MVT::f16, Legal);
280 // F16 - Load/Store Actions.
281 setOperationAction(ISD::LOAD, MVT::f16, Promote);
282 AddPromotedToType(ISD::LOAD, MVT::f16, MVT::i16);
283 setOperationAction(ISD::STORE, MVT::f16, Promote);
284 AddPromotedToType(ISD::STORE, MVT::f16, MVT::i16);
286 // F16 - VOP1 Actions.
287 setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
288 setOperationAction(ISD::FCOS, MVT::f16, Promote);
289 setOperationAction(ISD::FSIN, MVT::f16, Promote);
290 setOperationAction(ISD::FP_TO_SINT, MVT::f16, Promote);
291 setOperationAction(ISD::FP_TO_UINT, MVT::f16, Promote);
292 setOperationAction(ISD::SINT_TO_FP, MVT::f16, Promote);
293 setOperationAction(ISD::UINT_TO_FP, MVT::f16, Promote);
295 // F16 - VOP2 Actions.
296 setOperationAction(ISD::BR_CC, MVT::f16, Expand);
297 setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
298 setOperationAction(ISD::FMAXNUM, MVT::f16, Legal);
299 setOperationAction(ISD::FMINNUM, MVT::f16, Legal);
300 setOperationAction(ISD::FDIV, MVT::f16, Custom);
302 // F16 - VOP3 Actions.
303 setOperationAction(ISD::FMA, MVT::f16, Legal);
304 if (!Subtarget->hasFP16Denormals())
305 setOperationAction(ISD::FMAD, MVT::f16, Legal);
308 setTargetDAGCombine(ISD::FADD);
309 setTargetDAGCombine(ISD::FSUB);
310 setTargetDAGCombine(ISD::FMINNUM);
311 setTargetDAGCombine(ISD::FMAXNUM);
312 setTargetDAGCombine(ISD::SMIN);
313 setTargetDAGCombine(ISD::SMAX);
314 setTargetDAGCombine(ISD::UMIN);
315 setTargetDAGCombine(ISD::UMAX);
316 setTargetDAGCombine(ISD::SETCC);
317 setTargetDAGCombine(ISD::AND);
318 setTargetDAGCombine(ISD::OR);
319 setTargetDAGCombine(ISD::XOR);
320 setTargetDAGCombine(ISD::SINT_TO_FP);
321 setTargetDAGCombine(ISD::UINT_TO_FP);
322 setTargetDAGCombine(ISD::FCANONICALIZE);
324 // All memory operations. Some folding on the pointer operand is done to help
325 // matching the constant offsets in the addressing modes.
326 setTargetDAGCombine(ISD::LOAD);
327 setTargetDAGCombine(ISD::STORE);
328 setTargetDAGCombine(ISD::ATOMIC_LOAD);
329 setTargetDAGCombine(ISD::ATOMIC_STORE);
330 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
331 setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
332 setTargetDAGCombine(ISD::ATOMIC_SWAP);
333 setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
334 setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
335 setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
336 setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
337 setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
338 setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
339 setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
340 setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
341 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
342 setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
344 setSchedulingPreference(Sched::RegPressure);
347 const SISubtarget *SITargetLowering::getSubtarget() const {
348 return static_cast<const SISubtarget *>(Subtarget);
351 //===----------------------------------------------------------------------===//
352 // TargetLowering queries
353 //===----------------------------------------------------------------------===//
355 bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
357 unsigned IntrID) const {
359 case Intrinsic::amdgcn_atomic_inc:
360 case Intrinsic::amdgcn_atomic_dec:
361 Info.opc = ISD::INTRINSIC_W_CHAIN;
362 Info.memVT = MVT::getVT(CI.getType());
363 Info.ptrVal = CI.getOperand(0);
367 Info.writeMem = true;
374 bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
376 // SI has some legal vector types, but no legal vector operations. Say no
377 // shuffles are legal in order to prefer scalarizing some vector operations.
381 bool SITargetLowering::isLegalFlatAddressingMode(const AddrMode &AM) const {
382 // Flat instructions do not have offsets, and only have the register
384 return AM.BaseOffs == 0 && (AM.Scale == 0 || AM.Scale == 1);
387 bool SITargetLowering::isLegalMUBUFAddressingMode(const AddrMode &AM) const {
388 // MUBUF / MTBUF instructions have a 12-bit unsigned byte offset, and
389 // additionally can do r + r + i with addr64. 32-bit has more addressing
390 // mode options. Depending on the resource constant, it can also do
391 // (i64 r0) + (i32 r1) * (i14 i).
393 // Private arrays end up using a scratch buffer most of the time, so also
394 // assume those use MUBUF instructions. Scratch loads / stores are currently
395 // implemented as mubuf instructions with offen bit set, so slightly
396 // different than the normal addr64.
397 if (!isUInt<12>(AM.BaseOffs))
400 // FIXME: Since we can split immediate into soffset and immediate offset,
401 // would it make sense to allow any immediate?
404 case 0: // r + i or just i, depending on HasBaseReg.
407 return true; // We have r + r or r + i.
414 // Allow 2 * r as r + r
415 // Or 2 * r + i is allowed as r + r + i.
417 default: // Don't allow n * r
422 bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL,
423 const AddrMode &AM, Type *Ty,
425 // No global is ever allowed as a base.
430 case AMDGPUAS::GLOBAL_ADDRESS: {
431 if (Subtarget->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
432 // Assume the we will use FLAT for all global memory accesses
434 // FIXME: This assumption is currently wrong. On VI we still use
435 // MUBUF instructions for the r + i addressing mode. As currently
436 // implemented, the MUBUF instructions only work on buffer < 4GB.
437 // It may be possible to support > 4GB buffers with MUBUF instructions,
438 // by setting the stride value in the resource descriptor which would
439 // increase the size limit to (stride * 4GB). However, this is risky,
440 // because it has never been validated.
441 return isLegalFlatAddressingMode(AM);
444 return isLegalMUBUFAddressingMode(AM);
446 case AMDGPUAS::CONSTANT_ADDRESS: {
447 // If the offset isn't a multiple of 4, it probably isn't going to be
448 // correctly aligned.
449 // FIXME: Can we get the real alignment here?
450 if (AM.BaseOffs % 4 != 0)
451 return isLegalMUBUFAddressingMode(AM);
453 // There are no SMRD extloads, so if we have to do a small type access we
454 // will use a MUBUF load.
455 // FIXME?: We also need to do this if unaligned, but we don't know the
457 if (DL.getTypeStoreSize(Ty) < 4)
458 return isLegalMUBUFAddressingMode(AM);
460 if (Subtarget->getGeneration() == SISubtarget::SOUTHERN_ISLANDS) {
461 // SMRD instructions have an 8-bit, dword offset on SI.
462 if (!isUInt<8>(AM.BaseOffs / 4))
464 } else if (Subtarget->getGeneration() == SISubtarget::SEA_ISLANDS) {
465 // On CI+, this can also be a 32-bit literal constant offset. If it fits
466 // in 8-bits, it can use a smaller encoding.
467 if (!isUInt<32>(AM.BaseOffs / 4))
469 } else if (Subtarget->getGeneration() == SISubtarget::VOLCANIC_ISLANDS) {
470 // On VI, these use the SMEM format and the offset is 20-bit in bytes.
471 if (!isUInt<20>(AM.BaseOffs))
474 llvm_unreachable("unhandled generation");
476 if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
479 if (AM.Scale == 1 && AM.HasBaseReg)
485 case AMDGPUAS::PRIVATE_ADDRESS:
486 return isLegalMUBUFAddressingMode(AM);
488 case AMDGPUAS::LOCAL_ADDRESS:
489 case AMDGPUAS::REGION_ADDRESS: {
490 // Basic, single offset DS instructions allow a 16-bit unsigned immediate
492 // XXX - If doing a 4-byte aligned 8-byte type access, we effectively have
493 // an 8-bit dword offset but we don't know the alignment here.
494 if (!isUInt<16>(AM.BaseOffs))
497 if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
500 if (AM.Scale == 1 && AM.HasBaseReg)
505 case AMDGPUAS::FLAT_ADDRESS:
506 case AMDGPUAS::UNKNOWN_ADDRESS_SPACE:
507 // For an unknown address space, this usually means that this is for some
508 // reason being used for pure arithmetic, and not based on some addressing
509 // computation. We don't have instructions that compute pointers with any
510 // addressing modes, so treat them as having no offset like flat
512 return isLegalFlatAddressingMode(AM);
515 llvm_unreachable("unhandled address space");
519 bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
522 bool *IsFast) const {
526 // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
527 // which isn't a simple VT.
528 // Until MVT is extended to handle this, simply check for the size and
529 // rely on the condition below: allow accesses if the size is a multiple of 4.
530 if (VT == MVT::Other || (VT != MVT::Other && VT.getSizeInBits() > 1024 &&
531 VT.getStoreSize() > 16)) {
535 if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS ||
536 AddrSpace == AMDGPUAS::REGION_ADDRESS) {
537 // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
538 // aligned, 8 byte access in a single operation using ds_read2/write2_b32
539 // with adjacent offsets.
540 bool AlignedBy4 = (Align % 4 == 0);
542 *IsFast = AlignedBy4;
547 // FIXME: We have to be conservative here and assume that flat operations
548 // will access scratch. If we had access to the IR function, then we
549 // could determine if any private memory was used in the function.
550 if (!Subtarget->hasUnalignedScratchAccess() &&
551 (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
552 AddrSpace == AMDGPUAS::FLAT_ADDRESS)) {
556 if (Subtarget->hasUnalignedBufferAccess()) {
557 // If we have an uniform constant load, it still requires using a slow
558 // buffer instruction if unaligned.
560 *IsFast = (AddrSpace == AMDGPUAS::CONSTANT_ADDRESS) ?
561 (Align % 4 == 0) : true;
567 // Smaller than dword value must be aligned.
568 if (VT.bitsLT(MVT::i32))
571 // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
572 // byte-address are ignored, thus forcing Dword alignment.
573 // This applies to private, global, and constant memory.
577 return VT.bitsGT(MVT::i32) && Align % 4 == 0;
580 EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
581 unsigned SrcAlign, bool IsMemset,
584 MachineFunction &MF) const {
585 // FIXME: Should account for address space here.
587 // The default fallback uses the private pointer size as a guess for a type to
588 // use. Make sure we switch these to 64-bit accesses.
590 if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
593 if (Size >= 8 && DstAlign >= 4)
600 static bool isFlatGlobalAddrSpace(unsigned AS) {
601 return AS == AMDGPUAS::GLOBAL_ADDRESS ||
602 AS == AMDGPUAS::FLAT_ADDRESS ||
603 AS == AMDGPUAS::CONSTANT_ADDRESS;
606 bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
607 unsigned DestAS) const {
608 return isFlatGlobalAddrSpace(SrcAS) && isFlatGlobalAddrSpace(DestAS);
611 bool SITargetLowering::isMemOpHasNoClobberedMemOperand(const SDNode *N) const {
612 const MemSDNode *MemNode = cast<MemSDNode>(N);
613 const Value *Ptr = MemNode->getMemOperand()->getValue();
614 const Instruction *I = dyn_cast<Instruction>(Ptr);
615 return I && I->getMetadata("amdgpu.noclobber");
618 bool SITargetLowering::isCheapAddrSpaceCast(unsigned SrcAS,
619 unsigned DestAS) const {
620 // Flat -> private/local is a simple truncate.
621 // Flat -> global is no-op
622 if (SrcAS == AMDGPUAS::FLAT_ADDRESS)
625 return isNoopAddrSpaceCast(SrcAS, DestAS);
628 bool SITargetLowering::isMemOpUniform(const SDNode *N) const {
629 const MemSDNode *MemNode = cast<MemSDNode>(N);
630 const Value *Ptr = MemNode->getMemOperand()->getValue();
632 // UndefValue means this is a load of a kernel input. These are uniform.
633 // Sometimes LDS instructions have constant pointers.
634 // If Ptr is null, then that means this mem operand contains a
635 // PseudoSourceValue like GOT.
636 if (!Ptr || isa<UndefValue>(Ptr) || isa<Argument>(Ptr) ||
637 isa<Constant>(Ptr) || isa<GlobalValue>(Ptr))
640 const Instruction *I = dyn_cast<Instruction>(Ptr);
641 return I && I->getMetadata("amdgpu.uniform");
644 TargetLoweringBase::LegalizeTypeAction
645 SITargetLowering::getPreferredVectorAction(EVT VT) const {
646 if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
647 return TypeSplitVector;
649 return TargetLoweringBase::getPreferredVectorAction(VT);
652 bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
654 // FIXME: Could be smarter if called for vector constants.
658 bool SITargetLowering::isTypeDesirableForOp(unsigned Op, EVT VT) const {
659 if (Subtarget->has16BitInsts() && VT == MVT::i16) {
664 // These operations are done with 32-bit instructions anyway.
676 // SimplifySetCC uses this function to determine whether or not it should
677 // create setcc with i1 operands. We don't have instructions for i1 setcc.
678 if (VT == MVT::i1 && Op == ISD::SETCC)
681 return TargetLowering::isTypeDesirableForOp(Op, VT);
684 SDValue SITargetLowering::LowerParameterPtr(SelectionDAG &DAG,
685 const SDLoc &SL, SDValue Chain,
686 unsigned Offset) const {
687 const DataLayout &DL = DAG.getDataLayout();
688 MachineFunction &MF = DAG.getMachineFunction();
689 const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
690 unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::KERNARG_SEGMENT_PTR);
692 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
693 MVT PtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
694 SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
695 MRI.getLiveInVirtReg(InputPtrReg), PtrVT);
696 return DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
697 DAG.getConstant(Offset, SL, PtrVT));
700 SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
701 const SDLoc &SL, SDValue Chain,
702 unsigned Offset, bool Signed) const {
703 const DataLayout &DL = DAG.getDataLayout();
704 Type *Ty = MemVT.getTypeForEVT(*DAG.getContext());
705 PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
706 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
708 unsigned Align = DL.getABITypeAlignment(Ty);
710 SDValue Ptr = LowerParameterPtr(DAG, SL, Chain, Offset);
711 SDValue Load = DAG.getLoad(MemVT, SL, Chain, Ptr, PtrInfo, Align,
712 MachineMemOperand::MONonTemporal |
713 MachineMemOperand::MODereferenceable |
714 MachineMemOperand::MOInvariant);
717 if (MemVT.isFloatingPoint())
718 Val = getFPExtOrFPTrunc(DAG, Load, SL, VT);
720 Val = DAG.getSExtOrTrunc(Load, SL, VT);
722 Val = DAG.getZExtOrTrunc(Load, SL, VT);
729 return DAG.getMergeValues(Ops, SL);
732 SDValue SITargetLowering::LowerFormalArguments(
733 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
734 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
735 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
736 const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
738 MachineFunction &MF = DAG.getMachineFunction();
739 FunctionType *FType = MF.getFunction()->getFunctionType();
740 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
741 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
743 if (Subtarget->isAmdHsaOS() && AMDGPU::isShader(CallConv)) {
744 const Function *Fn = MF.getFunction();
745 DiagnosticInfoUnsupported NoGraphicsHSA(
746 *Fn, "unsupported non-compute shaders with HSA", DL.getDebugLoc());
747 DAG.getContext()->diagnose(NoGraphicsHSA);
748 return DAG.getEntryNode();
751 // Create stack objects that are used for emitting debugger prologue if
752 // "amdgpu-debugger-emit-prologue" attribute was specified.
753 if (ST.debuggerEmitPrologue())
754 createDebuggerPrologueStackObjects(MF);
756 SmallVector<ISD::InputArg, 16> Splits;
757 BitVector Skipped(Ins.size());
759 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
760 const ISD::InputArg &Arg = Ins[i];
762 // First check if it's a PS input addr
763 if (CallConv == CallingConv::AMDGPU_PS && !Arg.Flags.isInReg() &&
764 !Arg.Flags.isByVal() && PSInputNum <= 15) {
766 if (!Arg.Used && !Info->isPSInputAllocated(PSInputNum)) {
767 // We can safely skip PS inputs
773 Info->markPSInputAllocated(PSInputNum);
775 Info->PSInputEna |= 1 << PSInputNum;
780 if (AMDGPU::isShader(CallConv)) {
781 // Second split vertices into their elements
782 if (Arg.VT.isVector()) {
783 ISD::InputArg NewArg = Arg;
784 NewArg.Flags.setSplit();
785 NewArg.VT = Arg.VT.getVectorElementType();
787 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
788 // three or five element vertex only needs three or five registers,
789 // NOT four or eight.
790 Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
791 unsigned NumElements = ParamType->getVectorNumElements();
793 for (unsigned j = 0; j != NumElements; ++j) {
794 Splits.push_back(NewArg);
795 NewArg.PartOffset += NewArg.VT.getStoreSize();
798 Splits.push_back(Arg);
803 SmallVector<CCValAssign, 16> ArgLocs;
804 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
807 // At least one interpolation mode must be enabled or else the GPU will hang.
809 // Check PSInputAddr instead of PSInputEna. The idea is that if the user set
810 // PSInputAddr, the user wants to enable some bits after the compilation
811 // based on run-time states. Since we can't know what the final PSInputEna
812 // will look like, so we shouldn't do anything here and the user should take
813 // responsibility for the correct programming.
815 // Otherwise, the following restrictions apply:
816 // - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled.
817 // - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be
819 if (CallConv == CallingConv::AMDGPU_PS &&
820 ((Info->getPSInputAddr() & 0x7F) == 0 ||
821 ((Info->getPSInputAddr() & 0xF) == 0 && Info->isPSInputAllocated(11)))) {
822 CCInfo.AllocateReg(AMDGPU::VGPR0);
823 CCInfo.AllocateReg(AMDGPU::VGPR1);
824 Info->markPSInputAllocated(0);
825 Info->PSInputEna |= 1;
828 if (!AMDGPU::isShader(CallConv)) {
829 assert(Info->hasWorkGroupIDX() && Info->hasWorkItemIDX());
831 assert(!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() &&
832 !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() &&
833 !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() &&
834 !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() &&
835 !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() &&
836 !Info->hasWorkItemIDZ());
839 // FIXME: How should these inputs interact with inreg / custom SGPR inputs?
840 if (Info->hasPrivateSegmentBuffer()) {
841 unsigned PrivateSegmentBufferReg = Info->addPrivateSegmentBuffer(*TRI);
842 MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SReg_128RegClass);
843 CCInfo.AllocateReg(PrivateSegmentBufferReg);
846 if (Info->hasDispatchPtr()) {
847 unsigned DispatchPtrReg = Info->addDispatchPtr(*TRI);
848 MF.addLiveIn(DispatchPtrReg, &AMDGPU::SGPR_64RegClass);
849 CCInfo.AllocateReg(DispatchPtrReg);
852 if (Info->hasQueuePtr()) {
853 unsigned QueuePtrReg = Info->addQueuePtr(*TRI);
854 MF.addLiveIn(QueuePtrReg, &AMDGPU::SGPR_64RegClass);
855 CCInfo.AllocateReg(QueuePtrReg);
858 if (Info->hasKernargSegmentPtr()) {
859 unsigned InputPtrReg = Info->addKernargSegmentPtr(*TRI);
860 MF.addLiveIn(InputPtrReg, &AMDGPU::SGPR_64RegClass);
861 CCInfo.AllocateReg(InputPtrReg);
864 if (Info->hasDispatchID()) {
865 unsigned DispatchIDReg = Info->addDispatchID(*TRI);
866 MF.addLiveIn(DispatchIDReg, &AMDGPU::SGPR_64RegClass);
867 CCInfo.AllocateReg(DispatchIDReg);
870 if (Info->hasFlatScratchInit()) {
871 unsigned FlatScratchInitReg = Info->addFlatScratchInit(*TRI);
872 MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SGPR_64RegClass);
873 CCInfo.AllocateReg(FlatScratchInitReg);
876 if (!AMDGPU::isShader(CallConv))
877 analyzeFormalArgumentsCompute(CCInfo, Ins);
879 AnalyzeFormalArguments(CCInfo, Splits);
881 SmallVector<SDValue, 16> Chains;
883 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
885 const ISD::InputArg &Arg = Ins[i];
887 InVals.push_back(DAG.getUNDEF(Arg.VT));
891 CCValAssign &VA = ArgLocs[ArgIdx++];
892 MVT VT = VA.getLocVT();
896 EVT MemVT = VA.getLocVT();
897 const unsigned Offset = Subtarget->getExplicitKernelArgOffset() +
898 VA.getLocMemOffset();
899 // The first 36 bytes of the input buffer contains information about
900 // thread group and global sizes.
901 SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, Chain,
902 Offset, Ins[i].Flags.isSExt());
903 Chains.push_back(Arg.getValue(1));
906 dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
907 if (Subtarget->getGeneration() == SISubtarget::SOUTHERN_ISLANDS &&
908 ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
909 // On SI local pointers are just offsets into LDS, so they are always
910 // less than 16-bits. On CI and newer they could potentially be
911 // real pointers, so we can't guarantee their size.
912 Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
913 DAG.getValueType(MVT::i16));
916 InVals.push_back(Arg);
917 Info->setABIArgOffset(Offset + MemVT.getStoreSize());
920 assert(VA.isRegLoc() && "Parameter must be in a register!");
922 unsigned Reg = VA.getLocReg();
924 if (VT == MVT::i64) {
925 // For now assume it is a pointer
926 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
927 &AMDGPU::SGPR_64RegClass);
928 Reg = MF.addLiveIn(Reg, &AMDGPU::SGPR_64RegClass);
929 SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
930 InVals.push_back(Copy);
934 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
936 Reg = MF.addLiveIn(Reg, RC);
937 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
939 if (Arg.VT.isVector()) {
941 // Build a vector from the registers
942 Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
943 unsigned NumElements = ParamType->getVectorNumElements();
945 SmallVector<SDValue, 4> Regs;
947 for (unsigned j = 1; j != NumElements; ++j) {
948 Reg = ArgLocs[ArgIdx++].getLocReg();
949 Reg = MF.addLiveIn(Reg, RC);
951 SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
952 Regs.push_back(Copy);
955 // Fill up the missing vector elements
956 NumElements = Arg.VT.getVectorNumElements() - NumElements;
957 Regs.append(NumElements, DAG.getUNDEF(VT));
959 InVals.push_back(DAG.getBuildVector(Arg.VT, DL, Regs));
963 InVals.push_back(Val);
966 // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
967 // these from the dispatch pointer.
969 // Start adding system SGPRs.
970 if (Info->hasWorkGroupIDX()) {
971 unsigned Reg = Info->addWorkGroupIDX();
972 MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
973 CCInfo.AllocateReg(Reg);
976 if (Info->hasWorkGroupIDY()) {
977 unsigned Reg = Info->addWorkGroupIDY();
978 MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
979 CCInfo.AllocateReg(Reg);
982 if (Info->hasWorkGroupIDZ()) {
983 unsigned Reg = Info->addWorkGroupIDZ();
984 MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
985 CCInfo.AllocateReg(Reg);
988 if (Info->hasWorkGroupInfo()) {
989 unsigned Reg = Info->addWorkGroupInfo();
990 MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
991 CCInfo.AllocateReg(Reg);
994 if (Info->hasPrivateSegmentWaveByteOffset()) {
995 // Scratch wave offset passed in system SGPR.
996 unsigned PrivateSegmentWaveByteOffsetReg;
998 if (AMDGPU::isShader(CallConv)) {
999 PrivateSegmentWaveByteOffsetReg = findFirstFreeSGPR(CCInfo);
1000 Info->setPrivateSegmentWaveByteOffset(PrivateSegmentWaveByteOffsetReg);
1002 PrivateSegmentWaveByteOffsetReg = Info->addPrivateSegmentWaveByteOffset();
1004 MF.addLiveIn(PrivateSegmentWaveByteOffsetReg, &AMDGPU::SGPR_32RegClass);
1005 CCInfo.AllocateReg(PrivateSegmentWaveByteOffsetReg);
1008 // Now that we've figured out where the scratch register inputs are, see if
1009 // should reserve the arguments and use them directly.
1010 bool HasStackObjects = MF.getFrameInfo().hasStackObjects();
1011 // Record that we know we have non-spill stack objects so we don't need to
1012 // check all stack objects later.
1013 if (HasStackObjects)
1014 Info->setHasNonSpillStackObjects(true);
1016 // Everything live out of a block is spilled with fast regalloc, so it's
1017 // almost certain that spilling will be required.
1018 if (getTargetMachine().getOptLevel() == CodeGenOpt::None)
1019 HasStackObjects = true;
1021 if (ST.isAmdCodeObjectV2()) {
1022 if (HasStackObjects) {
1023 // If we have stack objects, we unquestionably need the private buffer
1024 // resource. For the Code Object V2 ABI, this will be the first 4 user
1025 // SGPR inputs. We can reserve those and use them directly.
1027 unsigned PrivateSegmentBufferReg = TRI->getPreloadedValue(
1028 MF, SIRegisterInfo::PRIVATE_SEGMENT_BUFFER);
1029 Info->setScratchRSrcReg(PrivateSegmentBufferReg);
1031 unsigned PrivateSegmentWaveByteOffsetReg = TRI->getPreloadedValue(
1032 MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
1033 Info->setScratchWaveOffsetReg(PrivateSegmentWaveByteOffsetReg);
1035 unsigned ReservedBufferReg
1036 = TRI->reservedPrivateSegmentBufferReg(MF);
1037 unsigned ReservedOffsetReg
1038 = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
1040 // We tentatively reserve the last registers (skipping the last two
1041 // which may contain VCC). After register allocation, we'll replace
1042 // these with the ones immediately after those which were really
1043 // allocated. In the prologue copies will be inserted from the argument
1044 // to these reserved registers.
1045 Info->setScratchRSrcReg(ReservedBufferReg);
1046 Info->setScratchWaveOffsetReg(ReservedOffsetReg);
1049 unsigned ReservedBufferReg = TRI->reservedPrivateSegmentBufferReg(MF);
1051 // Without HSA, relocations are used for the scratch pointer and the
1052 // buffer resource setup is always inserted in the prologue. Scratch wave
1053 // offset is still in an input SGPR.
1054 Info->setScratchRSrcReg(ReservedBufferReg);
1056 if (HasStackObjects) {
1057 unsigned ScratchWaveOffsetReg = TRI->getPreloadedValue(
1058 MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
1059 Info->setScratchWaveOffsetReg(ScratchWaveOffsetReg);
1061 unsigned ReservedOffsetReg
1062 = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
1063 Info->setScratchWaveOffsetReg(ReservedOffsetReg);
1067 if (Info->hasWorkItemIDX()) {
1068 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X);
1069 MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
1070 CCInfo.AllocateReg(Reg);
1073 if (Info->hasWorkItemIDY()) {
1074 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y);
1075 MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
1076 CCInfo.AllocateReg(Reg);
1079 if (Info->hasWorkItemIDZ()) {
1080 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z);
1081 MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
1082 CCInfo.AllocateReg(Reg);
1088 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
1092 SITargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
1094 const SmallVectorImpl<ISD::OutputArg> &Outs,
1095 const SmallVectorImpl<SDValue> &OutVals,
1096 const SDLoc &DL, SelectionDAG &DAG) const {
1097 MachineFunction &MF = DAG.getMachineFunction();
1098 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1100 if (!AMDGPU::isShader(CallConv))
1101 return AMDGPUTargetLowering::LowerReturn(Chain, CallConv, isVarArg, Outs,
1104 Info->setIfReturnsVoid(Outs.size() == 0);
1106 SmallVector<ISD::OutputArg, 48> Splits;
1107 SmallVector<SDValue, 48> SplitVals;
1109 // Split vectors into their elements.
1110 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
1111 const ISD::OutputArg &Out = Outs[i];
1113 if (Out.VT.isVector()) {
1114 MVT VT = Out.VT.getVectorElementType();
1115 ISD::OutputArg NewOut = Out;
1116 NewOut.Flags.setSplit();
1119 // We want the original number of vector elements here, e.g.
1120 // three or five, not four or eight.
1121 unsigned NumElements = Out.ArgVT.getVectorNumElements();
1123 for (unsigned j = 0; j != NumElements; ++j) {
1124 SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, OutVals[i],
1125 DAG.getConstant(j, DL, MVT::i32));
1126 SplitVals.push_back(Elem);
1127 Splits.push_back(NewOut);
1128 NewOut.PartOffset += NewOut.VT.getStoreSize();
1131 SplitVals.push_back(OutVals[i]);
1132 Splits.push_back(Out);
1136 // CCValAssign - represent the assignment of the return value to a location.
1137 SmallVector<CCValAssign, 48> RVLocs;
1139 // CCState - Info about the registers and stack slots.
1140 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1143 // Analyze outgoing return values.
1144 AnalyzeReturn(CCInfo, Splits);
1147 SmallVector<SDValue, 48> RetOps;
1148 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1150 // Copy the result values into the output registers.
1151 for (unsigned i = 0, realRVLocIdx = 0;
1153 ++i, ++realRVLocIdx) {
1154 CCValAssign &VA = RVLocs[i];
1155 assert(VA.isRegLoc() && "Can only return in registers!");
1157 SDValue Arg = SplitVals[realRVLocIdx];
1159 // Copied from other backends.
1160 switch (VA.getLocInfo()) {
1161 default: llvm_unreachable("Unknown loc info!");
1162 case CCValAssign::Full:
1164 case CCValAssign::BCvt:
1165 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
1169 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
1170 Flag = Chain.getValue(1);
1171 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1174 // Update chain and glue.
1177 RetOps.push_back(Flag);
1179 unsigned Opc = Info->returnsVoid() ? AMDGPUISD::ENDPGM : AMDGPUISD::RETURN;
1180 return DAG.getNode(Opc, DL, MVT::Other, RetOps);
1183 unsigned SITargetLowering::getRegisterByName(const char* RegName, EVT VT,
1184 SelectionDAG &DAG) const {
1185 unsigned Reg = StringSwitch<unsigned>(RegName)
1186 .Case("m0", AMDGPU::M0)
1187 .Case("exec", AMDGPU::EXEC)
1188 .Case("exec_lo", AMDGPU::EXEC_LO)
1189 .Case("exec_hi", AMDGPU::EXEC_HI)
1190 .Case("flat_scratch", AMDGPU::FLAT_SCR)
1191 .Case("flat_scratch_lo", AMDGPU::FLAT_SCR_LO)
1192 .Case("flat_scratch_hi", AMDGPU::FLAT_SCR_HI)
1193 .Default(AMDGPU::NoRegister);
1195 if (Reg == AMDGPU::NoRegister) {
1196 report_fatal_error(Twine("invalid register name \""
1197 + StringRef(RegName) + "\"."));
1201 if (Subtarget->getGeneration() == SISubtarget::SOUTHERN_ISLANDS &&
1202 Subtarget->getRegisterInfo()->regsOverlap(Reg, AMDGPU::FLAT_SCR)) {
1203 report_fatal_error(Twine("invalid register \""
1204 + StringRef(RegName) + "\" for subtarget."));
1209 case AMDGPU::EXEC_LO:
1210 case AMDGPU::EXEC_HI:
1211 case AMDGPU::FLAT_SCR_LO:
1212 case AMDGPU::FLAT_SCR_HI:
1213 if (VT.getSizeInBits() == 32)
1217 case AMDGPU::FLAT_SCR:
1218 if (VT.getSizeInBits() == 64)
1222 llvm_unreachable("missing register type checking");
1225 report_fatal_error(Twine("invalid type for register \""
1226 + StringRef(RegName) + "\"."));
1229 // If kill is not the last instruction, split the block so kill is always a
1230 // proper terminator.
1231 MachineBasicBlock *SITargetLowering::splitKillBlock(MachineInstr &MI,
1232 MachineBasicBlock *BB) const {
1233 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
1235 MachineBasicBlock::iterator SplitPoint(&MI);
1238 if (SplitPoint == BB->end()) {
1239 // Don't bother with a new block.
1240 MI.setDesc(TII->get(AMDGPU::SI_KILL_TERMINATOR));
1244 MachineFunction *MF = BB->getParent();
1245 MachineBasicBlock *SplitBB
1246 = MF->CreateMachineBasicBlock(BB->getBasicBlock());
1248 MF->insert(++MachineFunction::iterator(BB), SplitBB);
1249 SplitBB->splice(SplitBB->begin(), BB, SplitPoint, BB->end());
1251 SplitBB->transferSuccessorsAndUpdatePHIs(BB);
1252 BB->addSuccessor(SplitBB);
1254 MI.setDesc(TII->get(AMDGPU::SI_KILL_TERMINATOR));
1258 // Do a v_movrels_b32 or v_movreld_b32 for each unique value of \p IdxReg in the
1259 // wavefront. If the value is uniform and just happens to be in a VGPR, this
1260 // will only do one iteration. In the worst case, this will loop 64 times.
1262 // TODO: Just use v_readlane_b32 if we know the VGPR has a uniform value.
1263 static MachineBasicBlock::iterator emitLoadM0FromVGPRLoop(
1264 const SIInstrInfo *TII,
1265 MachineRegisterInfo &MRI,
1266 MachineBasicBlock &OrigBB,
1267 MachineBasicBlock &LoopBB,
1269 const MachineOperand &IdxReg,
1273 unsigned InitSaveExecReg,
1275 bool UseGPRIdxMode) {
1276 MachineBasicBlock::iterator I = LoopBB.begin();
1278 unsigned PhiExec = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1279 unsigned NewExec = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1280 unsigned CurrentIdxReg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1281 unsigned CondReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1283 BuildMI(LoopBB, I, DL, TII->get(TargetOpcode::PHI), PhiReg)
1289 BuildMI(LoopBB, I, DL, TII->get(TargetOpcode::PHI), PhiExec)
1290 .addReg(InitSaveExecReg)
1295 // Read the next variant <- also loop target.
1296 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32), CurrentIdxReg)
1297 .addReg(IdxReg.getReg(), getUndefRegState(IdxReg.isUndef()));
1299 // Compare the just read M0 value to all possible Idx values.
1300 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e64), CondReg)
1301 .addReg(CurrentIdxReg)
1302 .addReg(IdxReg.getReg(), 0, IdxReg.getSubReg());
1304 if (UseGPRIdxMode) {
1307 IdxReg = CurrentIdxReg;
1309 IdxReg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1310 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_ADD_I32), IdxReg)
1311 .addReg(CurrentIdxReg, RegState::Kill)
1315 MachineInstr *SetIdx =
1316 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_IDX))
1317 .addReg(IdxReg, RegState::Kill);
1318 SetIdx->getOperand(2).setIsUndef();
1320 // Move index from VCC into M0
1322 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
1323 .addReg(CurrentIdxReg, RegState::Kill);
1325 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
1326 .addReg(CurrentIdxReg, RegState::Kill)
1331 // Update EXEC, save the original EXEC value to VCC.
1332 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), NewExec)
1333 .addReg(CondReg, RegState::Kill);
1335 MRI.setSimpleHint(NewExec, CondReg);
1337 // Update EXEC, switch all done bits to 0 and all todo bits to 1.
1338 MachineInstr *InsertPt =
1339 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
1340 .addReg(AMDGPU::EXEC)
1343 // XXX - s_xor_b64 sets scc to 1 if the result is nonzero, so can we use
1346 // Loop back to V_READFIRSTLANE_B32 if there are still variants to cover.
1347 BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
1350 return InsertPt->getIterator();
1353 // This has slightly sub-optimal regalloc when the source vector is killed by
1354 // the read. The register allocator does not understand that the kill is
1355 // per-workitem, so is kept alive for the whole loop so we end up not re-using a
1356 // subregister from it, using 1 more VGPR than necessary. This was saved when
1357 // this was expanded after register allocation.
1358 static MachineBasicBlock::iterator loadM0FromVGPR(const SIInstrInfo *TII,
1359 MachineBasicBlock &MBB,
1361 unsigned InitResultReg,
1364 bool UseGPRIdxMode) {
1365 MachineFunction *MF = MBB.getParent();
1366 MachineRegisterInfo &MRI = MF->getRegInfo();
1367 const DebugLoc &DL = MI.getDebugLoc();
1368 MachineBasicBlock::iterator I(&MI);
1370 unsigned DstReg = MI.getOperand(0).getReg();
1371 unsigned SaveExec = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1372 unsigned TmpExec = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1374 BuildMI(MBB, I, DL, TII->get(TargetOpcode::IMPLICIT_DEF), TmpExec);
1376 // Save the EXEC mask
1377 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B64), SaveExec)
1378 .addReg(AMDGPU::EXEC);
1380 // To insert the loop we need to split the block. Move everything after this
1381 // point to a new block, and insert a new empty block between the two.
1382 MachineBasicBlock *LoopBB = MF->CreateMachineBasicBlock();
1383 MachineBasicBlock *RemainderBB = MF->CreateMachineBasicBlock();
1384 MachineFunction::iterator MBBI(MBB);
1387 MF->insert(MBBI, LoopBB);
1388 MF->insert(MBBI, RemainderBB);
1390 LoopBB->addSuccessor(LoopBB);
1391 LoopBB->addSuccessor(RemainderBB);
1393 // Move the rest of the block into a new block.
1394 RemainderBB->transferSuccessorsAndUpdatePHIs(&MBB);
1395 RemainderBB->splice(RemainderBB->begin(), &MBB, I, MBB.end());
1397 MBB.addSuccessor(LoopBB);
1399 const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);
1401 auto InsPt = emitLoadM0FromVGPRLoop(TII, MRI, MBB, *LoopBB, DL, *Idx,
1402 InitResultReg, DstReg, PhiReg, TmpExec,
1403 Offset, UseGPRIdxMode);
1405 MachineBasicBlock::iterator First = RemainderBB->begin();
1406 BuildMI(*RemainderBB, First, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
1412 // Returns subreg index, offset
1413 static std::pair<unsigned, int>
1414 computeIndirectRegAndOffset(const SIRegisterInfo &TRI,
1415 const TargetRegisterClass *SuperRC,
1418 int NumElts = SuperRC->getSize() / 4;
1420 // Skip out of bounds offsets, or else we would end up using an undefined
1422 if (Offset >= NumElts || Offset < 0)
1423 return std::make_pair(AMDGPU::sub0, Offset);
1425 return std::make_pair(AMDGPU::sub0 + Offset, 0);
1428 // Return true if the index is an SGPR and was set.
1429 static bool setM0ToIndexFromSGPR(const SIInstrInfo *TII,
1430 MachineRegisterInfo &MRI,
1434 bool IsIndirectSrc) {
1435 MachineBasicBlock *MBB = MI.getParent();
1436 const DebugLoc &DL = MI.getDebugLoc();
1437 MachineBasicBlock::iterator I(&MI);
1439 const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);
1440 const TargetRegisterClass *IdxRC = MRI.getRegClass(Idx->getReg());
1442 assert(Idx->getReg() != AMDGPU::NoRegister);
1444 if (!TII->getRegisterInfo().isSGPRClass(IdxRC))
1447 if (UseGPRIdxMode) {
1448 unsigned IdxMode = IsIndirectSrc ?
1449 VGPRIndexMode::SRC0_ENABLE : VGPRIndexMode::DST_ENABLE;
1451 MachineInstr *SetOn =
1452 BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
1456 SetOn->getOperand(3).setIsUndef();
1458 unsigned Tmp = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
1459 BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_ADD_I32), Tmp)
1462 MachineInstr *SetOn =
1463 BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
1464 .addReg(Tmp, RegState::Kill)
1467 SetOn->getOperand(3).setIsUndef();
1474 BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
1477 BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
1485 // Control flow needs to be inserted if indexing with a VGPR.
1486 static MachineBasicBlock *emitIndirectSrc(MachineInstr &MI,
1487 MachineBasicBlock &MBB,
1488 const SISubtarget &ST) {
1489 const SIInstrInfo *TII = ST.getInstrInfo();
1490 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1491 MachineFunction *MF = MBB.getParent();
1492 MachineRegisterInfo &MRI = MF->getRegInfo();
1494 unsigned Dst = MI.getOperand(0).getReg();
1495 unsigned SrcReg = TII->getNamedOperand(MI, AMDGPU::OpName::src)->getReg();
1496 int Offset = TII->getNamedOperand(MI, AMDGPU::OpName::offset)->getImm();
1498 const TargetRegisterClass *VecRC = MRI.getRegClass(SrcReg);
1501 std::tie(SubReg, Offset)
1502 = computeIndirectRegAndOffset(TRI, VecRC, SrcReg, Offset);
1504 bool UseGPRIdxMode = ST.hasVGPRIndexMode() && EnableVGPRIndexMode;
1506 if (setM0ToIndexFromSGPR(TII, MRI, MI, Offset, UseGPRIdxMode, true)) {
1507 MachineBasicBlock::iterator I(&MI);
1508 const DebugLoc &DL = MI.getDebugLoc();
1510 if (UseGPRIdxMode) {
1511 // TODO: Look at the uses to avoid the copy. This may require rescheduling
1512 // to avoid interfering with other uses, so probably requires a new
1513 // optimization pass.
1514 BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOV_B32_e32), Dst)
1515 .addReg(SrcReg, RegState::Undef, SubReg)
1516 .addReg(SrcReg, RegState::Implicit)
1517 .addReg(AMDGPU::M0, RegState::Implicit);
1518 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
1520 BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
1521 .addReg(SrcReg, RegState::Undef, SubReg)
1522 .addReg(SrcReg, RegState::Implicit);
1525 MI.eraseFromParent();
1531 const DebugLoc &DL = MI.getDebugLoc();
1532 MachineBasicBlock::iterator I(&MI);
1534 unsigned PhiReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1535 unsigned InitReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1537 BuildMI(MBB, I, DL, TII->get(TargetOpcode::IMPLICIT_DEF), InitReg);
1539 if (UseGPRIdxMode) {
1540 MachineInstr *SetOn = BuildMI(MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
1541 .addImm(0) // Reset inside loop.
1542 .addImm(VGPRIndexMode::SRC0_ENABLE);
1543 SetOn->getOperand(3).setIsUndef();
1545 // Disable again after the loop.
1546 BuildMI(MBB, std::next(I), DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
1549 auto InsPt = loadM0FromVGPR(TII, MBB, MI, InitReg, PhiReg, Offset, UseGPRIdxMode);
1550 MachineBasicBlock *LoopBB = InsPt->getParent();
1552 if (UseGPRIdxMode) {
1553 BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::V_MOV_B32_e32), Dst)
1554 .addReg(SrcReg, RegState::Undef, SubReg)
1555 .addReg(SrcReg, RegState::Implicit)
1556 .addReg(AMDGPU::M0, RegState::Implicit);
1558 BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
1559 .addReg(SrcReg, RegState::Undef, SubReg)
1560 .addReg(SrcReg, RegState::Implicit);
1563 MI.eraseFromParent();
1568 static unsigned getMOVRELDPseudo(const TargetRegisterClass *VecRC) {
1569 switch (VecRC->getSize()) {
1571 return AMDGPU::V_MOVRELD_B32_V1;
1573 return AMDGPU::V_MOVRELD_B32_V2;
1575 return AMDGPU::V_MOVRELD_B32_V4;
1577 return AMDGPU::V_MOVRELD_B32_V8;
1579 return AMDGPU::V_MOVRELD_B32_V16;
1581 llvm_unreachable("unsupported size for MOVRELD pseudos");
1585 static MachineBasicBlock *emitIndirectDst(MachineInstr &MI,
1586 MachineBasicBlock &MBB,
1587 const SISubtarget &ST) {
1588 const SIInstrInfo *TII = ST.getInstrInfo();
1589 const SIRegisterInfo &TRI = TII->getRegisterInfo();
1590 MachineFunction *MF = MBB.getParent();
1591 MachineRegisterInfo &MRI = MF->getRegInfo();
1593 unsigned Dst = MI.getOperand(0).getReg();
1594 const MachineOperand *SrcVec = TII->getNamedOperand(MI, AMDGPU::OpName::src);
1595 const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);
1596 const MachineOperand *Val = TII->getNamedOperand(MI, AMDGPU::OpName::val);
1597 int Offset = TII->getNamedOperand(MI, AMDGPU::OpName::offset)->getImm();
1598 const TargetRegisterClass *VecRC = MRI.getRegClass(SrcVec->getReg());
1600 // This can be an immediate, but will be folded later.
1601 assert(Val->getReg());
1604 std::tie(SubReg, Offset) = computeIndirectRegAndOffset(TRI, VecRC,
1607 bool UseGPRIdxMode = ST.hasVGPRIndexMode() && EnableVGPRIndexMode;
1609 if (Idx->getReg() == AMDGPU::NoRegister) {
1610 MachineBasicBlock::iterator I(&MI);
1611 const DebugLoc &DL = MI.getDebugLoc();
1613 assert(Offset == 0);
1615 BuildMI(MBB, I, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dst)
1616 .addOperand(*SrcVec)
1620 MI.eraseFromParent();
1624 if (setM0ToIndexFromSGPR(TII, MRI, MI, Offset, UseGPRIdxMode, false)) {
1625 MachineBasicBlock::iterator I(&MI);
1626 const DebugLoc &DL = MI.getDebugLoc();
1628 if (UseGPRIdxMode) {
1629 BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOV_B32_indirect))
1630 .addReg(SrcVec->getReg(), RegState::Undef, SubReg) // vdst
1632 .addReg(Dst, RegState::ImplicitDefine)
1633 .addReg(SrcVec->getReg(), RegState::Implicit)
1634 .addReg(AMDGPU::M0, RegState::Implicit);
1636 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
1638 const MCInstrDesc &MovRelDesc = TII->get(getMOVRELDPseudo(VecRC));
1640 BuildMI(MBB, I, DL, MovRelDesc)
1641 .addReg(Dst, RegState::Define)
1642 .addReg(SrcVec->getReg())
1644 .addImm(SubReg - AMDGPU::sub0);
1647 MI.eraseFromParent();
1652 MRI.clearKillFlags(Val->getReg());
1654 const DebugLoc &DL = MI.getDebugLoc();
1656 if (UseGPRIdxMode) {
1657 MachineBasicBlock::iterator I(&MI);
1659 MachineInstr *SetOn = BuildMI(MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
1660 .addImm(0) // Reset inside loop.
1661 .addImm(VGPRIndexMode::DST_ENABLE);
1662 SetOn->getOperand(3).setIsUndef();
1664 // Disable again after the loop.
1665 BuildMI(MBB, std::next(I), DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
1668 unsigned PhiReg = MRI.createVirtualRegister(VecRC);
1670 auto InsPt = loadM0FromVGPR(TII, MBB, MI, SrcVec->getReg(), PhiReg,
1671 Offset, UseGPRIdxMode);
1672 MachineBasicBlock *LoopBB = InsPt->getParent();
1674 if (UseGPRIdxMode) {
1675 BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::V_MOV_B32_indirect))
1676 .addReg(PhiReg, RegState::Undef, SubReg) // vdst
1677 .addOperand(*Val) // src0
1678 .addReg(Dst, RegState::ImplicitDefine)
1679 .addReg(PhiReg, RegState::Implicit)
1680 .addReg(AMDGPU::M0, RegState::Implicit);
1682 const MCInstrDesc &MovRelDesc = TII->get(getMOVRELDPseudo(VecRC));
1684 BuildMI(*LoopBB, InsPt, DL, MovRelDesc)
1685 .addReg(Dst, RegState::Define)
1688 .addImm(SubReg - AMDGPU::sub0);
1691 MI.eraseFromParent();
1696 MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter(
1697 MachineInstr &MI, MachineBasicBlock *BB) const {
1699 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
1700 MachineFunction *MF = BB->getParent();
1701 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
1703 if (TII->isMIMG(MI)) {
1704 if (!MI.memoperands_empty())
1706 // Add a memoperand for mimg instructions so that they aren't assumed to
1707 // be ordered memory instuctions.
1709 MachinePointerInfo PtrInfo(MFI->getImagePSV());
1710 MachineMemOperand::Flags Flags = MachineMemOperand::MODereferenceable;
1712 Flags |= MachineMemOperand::MOStore;
1715 Flags |= MachineMemOperand::MOLoad;
1717 auto MMO = MF->getMachineMemOperand(PtrInfo, Flags, 0, 0);
1718 MI.addMemOperand(*MF, MMO);
1722 switch (MI.getOpcode()) {
1723 case AMDGPU::SI_INIT_M0: {
1724 BuildMI(*BB, MI.getIterator(), MI.getDebugLoc(),
1725 TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
1726 .addOperand(MI.getOperand(0));
1727 MI.eraseFromParent();
1730 case AMDGPU::GET_GROUPSTATICSIZE: {
1731 DebugLoc DL = MI.getDebugLoc();
1732 BuildMI(*BB, MI, DL, TII->get(AMDGPU::S_MOV_B32))
1733 .addOperand(MI.getOperand(0))
1734 .addImm(MFI->getLDSSize());
1735 MI.eraseFromParent();
1738 case AMDGPU::SI_INDIRECT_SRC_V1:
1739 case AMDGPU::SI_INDIRECT_SRC_V2:
1740 case AMDGPU::SI_INDIRECT_SRC_V4:
1741 case AMDGPU::SI_INDIRECT_SRC_V8:
1742 case AMDGPU::SI_INDIRECT_SRC_V16:
1743 return emitIndirectSrc(MI, *BB, *getSubtarget());
1744 case AMDGPU::SI_INDIRECT_DST_V1:
1745 case AMDGPU::SI_INDIRECT_DST_V2:
1746 case AMDGPU::SI_INDIRECT_DST_V4:
1747 case AMDGPU::SI_INDIRECT_DST_V8:
1748 case AMDGPU::SI_INDIRECT_DST_V16:
1749 return emitIndirectDst(MI, *BB, *getSubtarget());
1750 case AMDGPU::SI_KILL:
1751 return splitKillBlock(MI, BB);
1752 case AMDGPU::V_CNDMASK_B64_PSEUDO: {
1753 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
1755 unsigned Dst = MI.getOperand(0).getReg();
1756 unsigned Src0 = MI.getOperand(1).getReg();
1757 unsigned Src1 = MI.getOperand(2).getReg();
1758 const DebugLoc &DL = MI.getDebugLoc();
1759 unsigned SrcCond = MI.getOperand(3).getReg();
1761 unsigned DstLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1762 unsigned DstHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1764 BuildMI(*BB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstLo)
1765 .addReg(Src0, 0, AMDGPU::sub0)
1766 .addReg(Src1, 0, AMDGPU::sub0)
1768 BuildMI(*BB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstHi)
1769 .addReg(Src0, 0, AMDGPU::sub1)
1770 .addReg(Src1, 0, AMDGPU::sub1)
1773 BuildMI(*BB, MI, DL, TII->get(AMDGPU::REG_SEQUENCE), Dst)
1775 .addImm(AMDGPU::sub0)
1777 .addImm(AMDGPU::sub1);
1778 MI.eraseFromParent();
1781 case AMDGPU::SI_BR_UNDEF: {
1782 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
1783 const DebugLoc &DL = MI.getDebugLoc();
1784 MachineInstr *Br = BuildMI(*BB, MI, DL, TII->get(AMDGPU::S_CBRANCH_SCC1))
1785 .addOperand(MI.getOperand(0));
1786 Br->getOperand(1).setIsUndef(true); // read undef SCC
1787 MI.eraseFromParent();
1791 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
1795 bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
1796 // This currently forces unfolding various combinations of fsub into fma with
1797 // free fneg'd operands. As long as we have fast FMA (controlled by
1798 // isFMAFasterThanFMulAndFAdd), we should perform these.
1800 // When fma is quarter rate, for f64 where add / sub are at best half rate,
1801 // most of these combines appear to be cycle neutral but save on instruction
1802 // count / code size.
1806 EVT SITargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &Ctx,
1808 if (!VT.isVector()) {
1811 return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements());
1814 MVT SITargetLowering::getScalarShiftAmountTy(const DataLayout &, EVT VT) const {
1815 // TODO: Should i16 be used always if legal? For now it would force VALU
1817 return (VT == MVT::i16) ? MVT::i16 : MVT::i32;
1820 // Answering this is somewhat tricky and depends on the specific device which
1821 // have different rates for fma or all f64 operations.
1823 // v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
1824 // regardless of which device (although the number of cycles differs between
1825 // devices), so it is always profitable for f64.
1827 // v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
1828 // only on full rate devices. Normally, we should prefer selecting v_mad_f32
1829 // which we can always do even without fused FP ops since it returns the same
1830 // result as the separate operations and since it is always full
1831 // rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
1832 // however does not support denormals, so we do report fma as faster if we have
1833 // a fast fma device and require denormals.
1835 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
1836 VT = VT.getScalarType();
1841 switch (VT.getSimpleVT().SimpleTy) {
1843 // This is as fast on some subtargets. However, we always have full rate f32
1844 // mad available which returns the same result as the separate operations
1845 // which we should prefer over fma. We can't use this if we want to support
1846 // denormals, so only report this in these cases.
1847 return Subtarget->hasFP32Denormals() && Subtarget->hasFastFMAF32();
1851 return Subtarget->has16BitInsts() && Subtarget->hasFP16Denormals();
1859 //===----------------------------------------------------------------------===//
1860 // Custom DAG Lowering Operations
1861 //===----------------------------------------------------------------------===//
1863 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
1864 switch (Op.getOpcode()) {
1865 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
1866 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
1868 SDValue Result = LowerLOAD(Op, DAG);
1869 assert((!Result.getNode() ||
1870 Result.getNode()->getNumValues() == 2) &&
1871 "Load should return a value and a chain");
1877 return LowerTrig(Op, DAG);
1878 case ISD::SELECT: return LowerSELECT(Op, DAG);
1879 case ISD::FDIV: return LowerFDIV(Op, DAG);
1880 case ISD::ATOMIC_CMP_SWAP: return LowerATOMIC_CMP_SWAP(Op, DAG);
1881 case ISD::STORE: return LowerSTORE(Op, DAG);
1882 case ISD::GlobalAddress: {
1883 MachineFunction &MF = DAG.getMachineFunction();
1884 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1885 return LowerGlobalAddress(MFI, Op, DAG);
1887 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
1888 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
1889 case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
1890 case ISD::ADDRSPACECAST: return lowerADDRSPACECAST(Op, DAG);
1891 case ISD::TRAP: return lowerTRAP(Op, DAG);
1893 return lowerFP_ROUND(Op, DAG);
1898 /// \brief Helper function for LowerBRCOND
1899 static SDNode *findUser(SDValue Value, unsigned Opcode) {
1901 SDNode *Parent = Value.getNode();
1902 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
1905 if (I.getUse().get() != Value)
1908 if (I->getOpcode() == Opcode)
1914 bool SITargetLowering::isCFIntrinsic(const SDNode *Intr) const {
1915 if (Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
1916 switch (cast<ConstantSDNode>(Intr->getOperand(1))->getZExtValue()) {
1917 case AMDGPUIntrinsic::amdgcn_if:
1918 case AMDGPUIntrinsic::amdgcn_else:
1919 case AMDGPUIntrinsic::amdgcn_end_cf:
1920 case AMDGPUIntrinsic::amdgcn_loop:
1927 if (Intr->getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
1928 switch (cast<ConstantSDNode>(Intr->getOperand(0))->getZExtValue()) {
1929 case AMDGPUIntrinsic::amdgcn_break:
1930 case AMDGPUIntrinsic::amdgcn_if_break:
1931 case AMDGPUIntrinsic::amdgcn_else_break:
1941 void SITargetLowering::createDebuggerPrologueStackObjects(
1942 MachineFunction &MF) const {
1943 // Create stack objects that are used for emitting debugger prologue.
1945 // Debugger prologue writes work group IDs and work item IDs to scratch memory
1946 // at fixed location in the following format:
1947 // offset 0: work group ID x
1948 // offset 4: work group ID y
1949 // offset 8: work group ID z
1950 // offset 16: work item ID x
1951 // offset 20: work item ID y
1952 // offset 24: work item ID z
1953 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1956 // For each dimension:
1957 for (unsigned i = 0; i < 3; ++i) {
1958 // Create fixed stack object for work group ID.
1959 ObjectIdx = MF.getFrameInfo().CreateFixedObject(4, i * 4, true);
1960 Info->setDebuggerWorkGroupIDStackObjectIndex(i, ObjectIdx);
1961 // Create fixed stack object for work item ID.
1962 ObjectIdx = MF.getFrameInfo().CreateFixedObject(4, i * 4 + 16, true);
1963 Info->setDebuggerWorkItemIDStackObjectIndex(i, ObjectIdx);
1967 bool SITargetLowering::shouldEmitFixup(const GlobalValue *GV) const {
1968 const Triple &TT = getTargetMachine().getTargetTriple();
1969 return GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1970 AMDGPU::shouldEmitConstantsToTextSection(TT);
1973 bool SITargetLowering::shouldEmitGOTReloc(const GlobalValue *GV) const {
1974 return (GV->getType()->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS ||
1975 GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) &&
1976 !shouldEmitFixup(GV) &&
1977 !getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV);
1980 bool SITargetLowering::shouldEmitPCReloc(const GlobalValue *GV) const {
1981 return !shouldEmitFixup(GV) && !shouldEmitGOTReloc(GV);
1984 /// This transforms the control flow intrinsics to get the branch destination as
1985 /// last parameter, also switches branch target with BR if the need arise
1986 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
1987 SelectionDAG &DAG) const {
1991 SDNode *Intr = BRCOND.getOperand(1).getNode();
1992 SDValue Target = BRCOND.getOperand(2);
1993 SDNode *BR = nullptr;
1994 SDNode *SetCC = nullptr;
1996 if (Intr->getOpcode() == ISD::SETCC) {
1997 // As long as we negate the condition everything is fine
1999 Intr = SetCC->getOperand(0).getNode();
2002 // Get the target from BR if we don't negate the condition
2003 BR = findUser(BRCOND, ISD::BR);
2004 Target = BR->getOperand(1);
2007 // FIXME: This changes the types of the intrinsics instead of introducing new
2008 // nodes with the correct types.
2009 // e.g. llvm.amdgcn.loop
2011 // eg: i1,ch = llvm.amdgcn.loop t0, TargetConstant:i32<6271>, t3
2012 // => t9: ch = llvm.amdgcn.loop t0, TargetConstant:i32<6271>, t3, BasicBlock:ch<bb1 0x7fee5286d088>
2014 if (!isCFIntrinsic(Intr)) {
2015 // This is a uniform branch so we don't need to legalize.
2019 bool HaveChain = Intr->getOpcode() == ISD::INTRINSIC_VOID ||
2020 Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN;
2023 (SetCC->getConstantOperandVal(1) == 1 &&
2024 cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
2027 // operands of the new intrinsic call
2028 SmallVector<SDValue, 4> Ops;
2030 Ops.push_back(BRCOND.getOperand(0));
2032 Ops.append(Intr->op_begin() + (HaveChain ? 1 : 0), Intr->op_end());
2033 Ops.push_back(Target);
2035 ArrayRef<EVT> Res(Intr->value_begin() + 1, Intr->value_end());
2037 // build the new intrinsic call
2038 SDNode *Result = DAG.getNode(
2039 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
2040 DAG.getVTList(Res), Ops).getNode();
2045 BRCOND.getOperand(0)
2048 Result = DAG.getMergeValues(Ops, DL).getNode();
2052 // Give the branch instruction our target
2055 BRCOND.getOperand(2)
2057 SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops);
2058 DAG.ReplaceAllUsesWith(BR, NewBR.getNode());
2059 BR = NewBR.getNode();
2062 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
2064 // Copy the intrinsic results to registers
2065 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
2066 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
2070 Chain = DAG.getCopyToReg(
2072 CopyToReg->getOperand(1),
2073 SDValue(Result, i - 1),
2076 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
2079 // Remove the old intrinsic from the chain
2080 DAG.ReplaceAllUsesOfValueWith(
2081 SDValue(Intr, Intr->getNumValues() - 1),
2082 Intr->getOperand(0));
2087 SDValue SITargetLowering::getFPExtOrFPTrunc(SelectionDAG &DAG,
2091 return Op.getValueType().bitsLE(VT) ?
2092 DAG.getNode(ISD::FP_EXTEND, DL, VT, Op) :
2093 DAG.getNode(ISD::FTRUNC, DL, VT, Op);
2096 SDValue SITargetLowering::lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
2097 assert(Op.getValueType() == MVT::f16 &&
2098 "Do not know how to custom lower FP_ROUND for non-f16 type");
2100 SDValue Src = Op.getOperand(0);
2101 EVT SrcVT = Src.getValueType();
2102 if (SrcVT != MVT::f64)
2107 SDValue FpToFp16 = DAG.getNode(ISD::FP_TO_FP16, DL, MVT::i32, Src);
2108 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FpToFp16);
2109 return DAG.getNode(ISD::BITCAST, DL, MVT::f16, Trunc);;
2112 SDValue SITargetLowering::getSegmentAperture(unsigned AS,
2113 SelectionDAG &DAG) const {
2115 MachineFunction &MF = DAG.getMachineFunction();
2116 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
2117 unsigned UserSGPR = Info->getQueuePtrUserSGPR();
2118 assert(UserSGPR != AMDGPU::NoRegister);
2120 SDValue QueuePtr = CreateLiveInRegister(
2121 DAG, &AMDGPU::SReg_64RegClass, UserSGPR, MVT::i64);
2123 // Offset into amd_queue_t for group_segment_aperture_base_hi /
2124 // private_segment_aperture_base_hi.
2125 uint32_t StructOffset = (AS == AMDGPUAS::LOCAL_ADDRESS) ? 0x40 : 0x44;
2127 SDValue Ptr = DAG.getNode(ISD::ADD, SL, MVT::i64, QueuePtr,
2128 DAG.getConstant(StructOffset, SL, MVT::i64));
2130 // TODO: Use custom target PseudoSourceValue.
2131 // TODO: We should use the value from the IR intrinsic call, but it might not
2132 // be available and how do we get it?
2133 Value *V = UndefValue::get(PointerType::get(Type::getInt8Ty(*DAG.getContext()),
2134 AMDGPUAS::CONSTANT_ADDRESS));
2136 MachinePointerInfo PtrInfo(V, StructOffset);
2137 return DAG.getLoad(MVT::i32, SL, QueuePtr.getValue(1), Ptr, PtrInfo,
2138 MinAlign(64, StructOffset),
2139 MachineMemOperand::MODereferenceable |
2140 MachineMemOperand::MOInvariant);
2143 SDValue SITargetLowering::lowerADDRSPACECAST(SDValue Op,
2144 SelectionDAG &DAG) const {
2146 const AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(Op);
2148 SDValue Src = ASC->getOperand(0);
2150 // FIXME: Really support non-0 null pointers.
2151 SDValue SegmentNullPtr = DAG.getConstant(-1, SL, MVT::i32);
2152 SDValue FlatNullPtr = DAG.getConstant(0, SL, MVT::i64);
2154 // flat -> local/private
2155 if (ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS) {
2156 if (ASC->getDestAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
2157 ASC->getDestAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
2158 SDValue NonNull = DAG.getSetCC(SL, MVT::i1, Src, FlatNullPtr, ISD::SETNE);
2159 SDValue Ptr = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, Src);
2161 return DAG.getNode(ISD::SELECT, SL, MVT::i32,
2162 NonNull, Ptr, SegmentNullPtr);
2166 // local/private -> flat
2167 if (ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) {
2168 if (ASC->getSrcAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
2169 ASC->getSrcAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
2171 = DAG.getSetCC(SL, MVT::i1, Src, SegmentNullPtr, ISD::SETNE);
2173 SDValue Aperture = getSegmentAperture(ASC->getSrcAddressSpace(), DAG);
2175 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Src, Aperture);
2177 return DAG.getNode(ISD::SELECT, SL, MVT::i64, NonNull,
2178 DAG.getNode(ISD::BITCAST, SL, MVT::i64, CvtPtr),
2183 // global <-> flat are no-ops and never emitted.
2185 const MachineFunction &MF = DAG.getMachineFunction();
2186 DiagnosticInfoUnsupported InvalidAddrSpaceCast(
2187 *MF.getFunction(), "invalid addrspacecast", SL.getDebugLoc());
2188 DAG.getContext()->diagnose(InvalidAddrSpaceCast);
2190 return DAG.getUNDEF(ASC->getValueType(0));
2194 SITargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2195 // We can fold offsets for anything that doesn't require a GOT relocation.
2196 return (GA->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS ||
2197 GA->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) &&
2198 !shouldEmitGOTReloc(GA->getGlobal());
2201 static SDValue buildPCRelGlobalAddress(SelectionDAG &DAG, const GlobalValue *GV,
2202 SDLoc DL, unsigned Offset, EVT PtrVT,
2203 unsigned GAFlags = SIInstrInfo::MO_NONE) {
2204 // In order to support pc-relative addressing, the PC_ADD_REL_OFFSET SDNode is
2205 // lowered to the following code sequence:
2207 // For constant address space:
2208 // s_getpc_b64 s[0:1]
2209 // s_add_u32 s0, s0, $symbol
2210 // s_addc_u32 s1, s1, 0
2212 // s_getpc_b64 returns the address of the s_add_u32 instruction and then
2213 // a fixup or relocation is emitted to replace $symbol with a literal
2214 // constant, which is a pc-relative offset from the encoding of the $symbol
2215 // operand to the global variable.
2217 // For global address space:
2218 // s_getpc_b64 s[0:1]
2219 // s_add_u32 s0, s0, $symbol@{gotpc}rel32@lo
2220 // s_addc_u32 s1, s1, $symbol@{gotpc}rel32@hi
2222 // s_getpc_b64 returns the address of the s_add_u32 instruction and then
2223 // fixups or relocations are emitted to replace $symbol@*@lo and
2224 // $symbol@*@hi with lower 32 bits and higher 32 bits of a literal constant,
2225 // which is a 64-bit pc-relative offset from the encoding of the $symbol
2226 // operand to the global variable.
2228 // What we want here is an offset from the value returned by s_getpc
2229 // (which is the address of the s_add_u32 instruction) to the global
2230 // variable, but since the encoding of $symbol starts 4 bytes after the start
2231 // of the s_add_u32 instruction, we end up with an offset that is 4 bytes too
2232 // small. This requires us to add 4 to the global variable offset in order to
2233 // compute the correct address.
2234 SDValue PtrLo = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, Offset + 4,
2236 SDValue PtrHi = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, Offset + 4,
2237 GAFlags == SIInstrInfo::MO_NONE ?
2238 GAFlags : GAFlags + 1);
2239 return DAG.getNode(AMDGPUISD::PC_ADD_REL_OFFSET, DL, PtrVT, PtrLo, PtrHi);
2242 SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
2244 SelectionDAG &DAG) const {
2245 GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
2247 if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
2248 GSD->getAddressSpace() != AMDGPUAS::GLOBAL_ADDRESS)
2249 return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
2252 const GlobalValue *GV = GSD->getGlobal();
2253 EVT PtrVT = Op.getValueType();
2255 if (shouldEmitFixup(GV))
2256 return buildPCRelGlobalAddress(DAG, GV, DL, GSD->getOffset(), PtrVT);
2257 else if (shouldEmitPCReloc(GV))
2258 return buildPCRelGlobalAddress(DAG, GV, DL, GSD->getOffset(), PtrVT,
2259 SIInstrInfo::MO_REL32);
2261 SDValue GOTAddr = buildPCRelGlobalAddress(DAG, GV, DL, 0, PtrVT,
2262 SIInstrInfo::MO_GOTPCREL32);
2264 Type *Ty = PtrVT.getTypeForEVT(*DAG.getContext());
2265 PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
2266 const DataLayout &DataLayout = DAG.getDataLayout();
2267 unsigned Align = DataLayout.getABITypeAlignment(PtrTy);
2268 // FIXME: Use a PseudoSourceValue once those can be assigned an address space.
2269 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
2271 return DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), GOTAddr, PtrInfo, Align,
2272 MachineMemOperand::MODereferenceable |
2273 MachineMemOperand::MOInvariant);
2276 SDValue SITargetLowering::lowerTRAP(SDValue Op,
2277 SelectionDAG &DAG) const {
2278 const MachineFunction &MF = DAG.getMachineFunction();
2279 DiagnosticInfoUnsupported NoTrap(*MF.getFunction(),
2280 "trap handler not supported",
2283 DAG.getContext()->diagnose(NoTrap);
2287 // FIXME: This should really be selected to s_trap, but that requires
2288 // setting up the trap handler for it o do anything.
2289 return DAG.getNode(AMDGPUISD::ENDPGM, SDLoc(Op), MVT::Other,
2293 SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain,
2294 const SDLoc &DL, SDValue V) const {
2295 // We can't use S_MOV_B32 directly, because there is no way to specify m0 as
2296 // the destination register.
2298 // We can't use CopyToReg, because MachineCSE won't combine COPY instructions,
2299 // so we will end up with redundant moves to m0.
2301 // We use a pseudo to ensure we emit s_mov_b32 with m0 as the direct result.
2303 // A Null SDValue creates a glue result.
2304 SDNode *M0 = DAG.getMachineNode(AMDGPU::SI_INIT_M0, DL, MVT::Other, MVT::Glue,
2306 return SDValue(M0, 0);
2309 SDValue SITargetLowering::lowerImplicitZextParam(SelectionDAG &DAG,
2312 unsigned Offset) const {
2314 SDValue Param = LowerParameter(DAG, MVT::i32, MVT::i32, SL,
2315 DAG.getEntryNode(), Offset, false);
2316 // The local size values will have the hi 16-bits as zero.
2317 return DAG.getNode(ISD::AssertZext, SL, MVT::i32, Param,
2318 DAG.getValueType(VT));
2321 static SDValue emitNonHSAIntrinsicError(SelectionDAG& DAG, SDLoc DL, EVT VT) {
2322 DiagnosticInfoUnsupported BadIntrin(*DAG.getMachineFunction().getFunction(),
2323 "non-hsa intrinsic with hsa target",
2325 DAG.getContext()->diagnose(BadIntrin);
2326 return DAG.getUNDEF(VT);
2329 static SDValue emitRemovedIntrinsicError(SelectionDAG& DAG, SDLoc DL, EVT VT) {
2330 DiagnosticInfoUnsupported BadIntrin(*DAG.getMachineFunction().getFunction(),
2331 "intrinsic not supported on subtarget",
2333 DAG.getContext()->diagnose(BadIntrin);
2334 return DAG.getUNDEF(VT);
2337 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2338 SelectionDAG &DAG) const {
2339 MachineFunction &MF = DAG.getMachineFunction();
2340 auto MFI = MF.getInfo<SIMachineFunctionInfo>();
2341 const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
2343 EVT VT = Op.getValueType();
2345 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2347 // TODO: Should this propagate fast-math-flags?
2349 switch (IntrinsicID) {
2350 case Intrinsic::amdgcn_dispatch_ptr:
2351 case Intrinsic::amdgcn_queue_ptr: {
2352 if (!Subtarget->isAmdCodeObjectV2()) {
2353 DiagnosticInfoUnsupported BadIntrin(
2354 *MF.getFunction(), "unsupported hsa intrinsic without hsa target",
2356 DAG.getContext()->diagnose(BadIntrin);
2357 return DAG.getUNDEF(VT);
2360 auto Reg = IntrinsicID == Intrinsic::amdgcn_dispatch_ptr ?
2361 SIRegisterInfo::DISPATCH_PTR : SIRegisterInfo::QUEUE_PTR;
2362 return CreateLiveInRegister(DAG, &AMDGPU::SReg_64RegClass,
2363 TRI->getPreloadedValue(MF, Reg), VT);
2365 case Intrinsic::amdgcn_implicitarg_ptr: {
2366 unsigned offset = getImplicitParameterOffset(MFI, FIRST_IMPLICIT);
2367 return LowerParameterPtr(DAG, DL, DAG.getEntryNode(), offset);
2369 case Intrinsic::amdgcn_kernarg_segment_ptr: {
2371 = TRI->getPreloadedValue(MF, SIRegisterInfo::KERNARG_SEGMENT_PTR);
2372 return CreateLiveInRegister(DAG, &AMDGPU::SReg_64RegClass, Reg, VT);
2374 case Intrinsic::amdgcn_dispatch_id: {
2375 unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::DISPATCH_ID);
2376 return CreateLiveInRegister(DAG, &AMDGPU::SReg_64RegClass, Reg, VT);
2378 case Intrinsic::amdgcn_rcp:
2379 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
2380 case Intrinsic::amdgcn_rsq:
2381 case AMDGPUIntrinsic::AMDGPU_rsq: // Legacy name
2382 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
2383 case Intrinsic::amdgcn_rsq_legacy: {
2384 if (Subtarget->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
2385 return emitRemovedIntrinsicError(DAG, DL, VT);
2387 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
2389 case Intrinsic::amdgcn_rcp_legacy: {
2390 if (Subtarget->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
2391 return emitRemovedIntrinsicError(DAG, DL, VT);
2392 return DAG.getNode(AMDGPUISD::RCP_LEGACY, DL, VT, Op.getOperand(1));
2394 case Intrinsic::amdgcn_rsq_clamp: {
2395 if (Subtarget->getGeneration() < SISubtarget::VOLCANIC_ISLANDS)
2396 return DAG.getNode(AMDGPUISD::RSQ_CLAMP, DL, VT, Op.getOperand(1));
2398 Type *Type = VT.getTypeForEVT(*DAG.getContext());
2399 APFloat Max = APFloat::getLargest(Type->getFltSemantics());
2400 APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
2402 SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
2403 SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
2404 DAG.getConstantFP(Max, DL, VT));
2405 return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
2406 DAG.getConstantFP(Min, DL, VT));
2408 case Intrinsic::r600_read_ngroups_x:
2409 if (Subtarget->isAmdHsaOS())
2410 return emitNonHSAIntrinsicError(DAG, DL, VT);
2412 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
2413 SI::KernelInputOffsets::NGROUPS_X, false);
2414 case Intrinsic::r600_read_ngroups_y:
2415 if (Subtarget->isAmdHsaOS())
2416 return emitNonHSAIntrinsicError(DAG, DL, VT);
2418 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
2419 SI::KernelInputOffsets::NGROUPS_Y, false);
2420 case Intrinsic::r600_read_ngroups_z:
2421 if (Subtarget->isAmdHsaOS())
2422 return emitNonHSAIntrinsicError(DAG, DL, VT);
2424 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
2425 SI::KernelInputOffsets::NGROUPS_Z, false);
2426 case Intrinsic::r600_read_global_size_x:
2427 if (Subtarget->isAmdHsaOS())
2428 return emitNonHSAIntrinsicError(DAG, DL, VT);
2430 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
2431 SI::KernelInputOffsets::GLOBAL_SIZE_X, false);
2432 case Intrinsic::r600_read_global_size_y:
2433 if (Subtarget->isAmdHsaOS())
2434 return emitNonHSAIntrinsicError(DAG, DL, VT);
2436 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
2437 SI::KernelInputOffsets::GLOBAL_SIZE_Y, false);
2438 case Intrinsic::r600_read_global_size_z:
2439 if (Subtarget->isAmdHsaOS())
2440 return emitNonHSAIntrinsicError(DAG, DL, VT);
2442 return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
2443 SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
2444 case Intrinsic::r600_read_local_size_x:
2445 if (Subtarget->isAmdHsaOS())
2446 return emitNonHSAIntrinsicError(DAG, DL, VT);
2448 return lowerImplicitZextParam(DAG, Op, MVT::i16,
2449 SI::KernelInputOffsets::LOCAL_SIZE_X);
2450 case Intrinsic::r600_read_local_size_y:
2451 if (Subtarget->isAmdHsaOS())
2452 return emitNonHSAIntrinsicError(DAG, DL, VT);
2454 return lowerImplicitZextParam(DAG, Op, MVT::i16,
2455 SI::KernelInputOffsets::LOCAL_SIZE_Y);
2456 case Intrinsic::r600_read_local_size_z:
2457 if (Subtarget->isAmdHsaOS())
2458 return emitNonHSAIntrinsicError(DAG, DL, VT);
2460 return lowerImplicitZextParam(DAG, Op, MVT::i16,
2461 SI::KernelInputOffsets::LOCAL_SIZE_Z);
2462 case Intrinsic::amdgcn_workgroup_id_x:
2463 case Intrinsic::r600_read_tgid_x:
2464 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32_XM0RegClass,
2465 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_X), VT);
2466 case Intrinsic::amdgcn_workgroup_id_y:
2467 case Intrinsic::r600_read_tgid_y:
2468 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32_XM0RegClass,
2469 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Y), VT);
2470 case Intrinsic::amdgcn_workgroup_id_z:
2471 case Intrinsic::r600_read_tgid_z:
2472 return CreateLiveInRegister(DAG, &AMDGPU::SReg_32_XM0RegClass,
2473 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Z), VT);
2474 case Intrinsic::amdgcn_workitem_id_x:
2475 case Intrinsic::r600_read_tidig_x:
2476 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
2477 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X), VT);
2478 case Intrinsic::amdgcn_workitem_id_y:
2479 case Intrinsic::r600_read_tidig_y:
2480 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
2481 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y), VT);
2482 case Intrinsic::amdgcn_workitem_id_z:
2483 case Intrinsic::r600_read_tidig_z:
2484 return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
2485 TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z), VT);
2486 case AMDGPUIntrinsic::SI_load_const: {
2492 MachineMemOperand *MMO = MF.getMachineMemOperand(
2493 MachinePointerInfo(),
2494 MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
2495 MachineMemOperand::MOInvariant,
2496 VT.getStoreSize(), 4);
2497 return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
2498 Op->getVTList(), Ops, VT, MMO);
2500 case AMDGPUIntrinsic::amdgcn_fdiv_fast: {
2501 return lowerFDIV_FAST(Op, DAG);
2503 case AMDGPUIntrinsic::SI_vs_load_input:
2504 return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
2509 case AMDGPUIntrinsic::SI_fs_constant: {
2510 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
2511 SDValue Glue = M0.getValue(1);
2512 return DAG.getNode(AMDGPUISD::INTERP_MOV, DL, MVT::f32,
2513 DAG.getConstant(2, DL, MVT::i32), // P0
2514 Op.getOperand(1), Op.getOperand(2), Glue);
2516 case AMDGPUIntrinsic::SI_packf16:
2517 if (Op.getOperand(1).isUndef() && Op.getOperand(2).isUndef())
2518 return DAG.getUNDEF(MVT::i32);
2520 case AMDGPUIntrinsic::SI_fs_interp: {
2521 SDValue IJ = Op.getOperand(4);
2522 SDValue I = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
2523 DAG.getConstant(0, DL, MVT::i32));
2524 SDValue J = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
2525 DAG.getConstant(1, DL, MVT::i32));
2526 I = DAG.getNode(ISD::BITCAST, DL, MVT::f32, I);
2527 J = DAG.getNode(ISD::BITCAST, DL, MVT::f32, J);
2528 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
2529 SDValue Glue = M0.getValue(1);
2530 SDValue P1 = DAG.getNode(AMDGPUISD::INTERP_P1, DL,
2531 DAG.getVTList(MVT::f32, MVT::Glue),
2532 I, Op.getOperand(1), Op.getOperand(2), Glue);
2533 Glue = SDValue(P1.getNode(), 1);
2534 return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, P1, J,
2535 Op.getOperand(1), Op.getOperand(2), Glue);
2537 case Intrinsic::amdgcn_interp_mov: {
2538 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(4));
2539 SDValue Glue = M0.getValue(1);
2540 return DAG.getNode(AMDGPUISD::INTERP_MOV, DL, MVT::f32, Op.getOperand(1),
2541 Op.getOperand(2), Op.getOperand(3), Glue);
2543 case Intrinsic::amdgcn_interp_p1: {
2544 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(4));
2545 SDValue Glue = M0.getValue(1);
2546 return DAG.getNode(AMDGPUISD::INTERP_P1, DL, MVT::f32, Op.getOperand(1),
2547 Op.getOperand(2), Op.getOperand(3), Glue);
2549 case Intrinsic::amdgcn_interp_p2: {
2550 SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(5));
2551 SDValue Glue = SDValue(M0.getNode(), 1);
2552 return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, Op.getOperand(1),
2553 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4),
2556 case Intrinsic::amdgcn_sin:
2557 return DAG.getNode(AMDGPUISD::SIN_HW, DL, VT, Op.getOperand(1));
2559 case Intrinsic::amdgcn_cos:
2560 return DAG.getNode(AMDGPUISD::COS_HW, DL, VT, Op.getOperand(1));
2562 case Intrinsic::amdgcn_log_clamp: {
2563 if (Subtarget->getGeneration() < SISubtarget::VOLCANIC_ISLANDS)
2566 DiagnosticInfoUnsupported BadIntrin(
2567 *MF.getFunction(), "intrinsic not supported on subtarget",
2569 DAG.getContext()->diagnose(BadIntrin);
2570 return DAG.getUNDEF(VT);
2572 case Intrinsic::amdgcn_ldexp:
2573 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT,
2574 Op.getOperand(1), Op.getOperand(2));
2576 case Intrinsic::amdgcn_fract:
2577 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
2579 case Intrinsic::amdgcn_class:
2580 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
2581 Op.getOperand(1), Op.getOperand(2));
2582 case Intrinsic::amdgcn_div_fmas:
2583 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
2584 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3),
2587 case Intrinsic::amdgcn_div_fixup:
2588 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
2589 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
2591 case Intrinsic::amdgcn_trig_preop:
2592 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
2593 Op.getOperand(1), Op.getOperand(2));
2594 case Intrinsic::amdgcn_div_scale: {
2595 // 3rd parameter required to be a constant.
2596 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
2598 return DAG.getUNDEF(VT);
2600 // Translate to the operands expected by the machine instruction. The
2601 // first parameter must be the same as the first instruction.
2602 SDValue Numerator = Op.getOperand(1);
2603 SDValue Denominator = Op.getOperand(2);
2605 // Note this order is opposite of the machine instruction's operations,
2606 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
2607 // intrinsic has the numerator as the first operand to match a normal
2608 // division operation.
2610 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
2612 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
2613 Denominator, Numerator);
2615 case Intrinsic::amdgcn_icmp: {
2616 const auto *CD = dyn_cast<ConstantSDNode>(Op.getOperand(3));
2617 int CondCode = CD->getSExtValue();
2619 if (CondCode < ICmpInst::Predicate::FIRST_ICMP_PREDICATE ||
2620 CondCode >= ICmpInst::Predicate::BAD_ICMP_PREDICATE)
2621 return DAG.getUNDEF(VT);
2623 ICmpInst::Predicate IcInput = static_cast<ICmpInst::Predicate>(CondCode);
2624 ISD::CondCode CCOpcode = getICmpCondCode(IcInput);
2625 return DAG.getNode(AMDGPUISD::SETCC, DL, VT, Op.getOperand(1),
2626 Op.getOperand(2), DAG.getCondCode(CCOpcode));
2628 case Intrinsic::amdgcn_fcmp: {
2629 const auto *CD = dyn_cast<ConstantSDNode>(Op.getOperand(3));
2630 int CondCode = CD->getSExtValue();
2632 if (CondCode <= FCmpInst::Predicate::FCMP_FALSE ||
2633 CondCode >= FCmpInst::Predicate::FCMP_TRUE)
2634 return DAG.getUNDEF(VT);
2636 FCmpInst::Predicate IcInput = static_cast<FCmpInst::Predicate>(CondCode);
2637 ISD::CondCode CCOpcode = getFCmpCondCode(IcInput);
2638 return DAG.getNode(AMDGPUISD::SETCC, DL, VT, Op.getOperand(1),
2639 Op.getOperand(2), DAG.getCondCode(CCOpcode));
2641 case Intrinsic::amdgcn_fmul_legacy:
2642 return DAG.getNode(AMDGPUISD::FMUL_LEGACY, DL, VT,
2643 Op.getOperand(1), Op.getOperand(2));
2644 case Intrinsic::amdgcn_sffbh:
2645 case AMDGPUIntrinsic::AMDGPU_flbit_i32: // Legacy name.
2646 return DAG.getNode(AMDGPUISD::FFBH_I32, DL, VT, Op.getOperand(1));
2648 return AMDGPUTargetLowering::LowerOperation(Op, DAG);
2652 SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
2653 SelectionDAG &DAG) const {
2654 unsigned IntrID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2657 case Intrinsic::amdgcn_atomic_inc:
2658 case Intrinsic::amdgcn_atomic_dec: {
2659 MemSDNode *M = cast<MemSDNode>(Op);
2660 unsigned Opc = (IntrID == Intrinsic::amdgcn_atomic_inc) ?
2661 AMDGPUISD::ATOMIC_INC : AMDGPUISD::ATOMIC_DEC;
2663 M->getOperand(0), // Chain
2664 M->getOperand(2), // Ptr
2665 M->getOperand(3) // Value
2668 return DAG.getMemIntrinsicNode(Opc, SDLoc(Op), M->getVTList(), Ops,
2669 M->getMemoryVT(), M->getMemOperand());
2671 case Intrinsic::amdgcn_buffer_load:
2672 case Intrinsic::amdgcn_buffer_load_format: {
2674 Op.getOperand(0), // Chain
2675 Op.getOperand(2), // rsrc
2676 Op.getOperand(3), // vindex
2677 Op.getOperand(4), // offset
2678 Op.getOperand(5), // glc
2679 Op.getOperand(6) // slc
2681 MachineFunction &MF = DAG.getMachineFunction();
2682 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
2684 unsigned Opc = (IntrID == Intrinsic::amdgcn_buffer_load) ?
2685 AMDGPUISD::BUFFER_LOAD : AMDGPUISD::BUFFER_LOAD_FORMAT;
2686 EVT VT = Op.getValueType();
2687 EVT IntVT = VT.changeTypeToInteger();
2689 MachineMemOperand *MMO = MF.getMachineMemOperand(
2690 MachinePointerInfo(MFI->getBufferPSV()),
2691 MachineMemOperand::MOLoad,
2692 VT.getStoreSize(), VT.getStoreSize());
2694 return DAG.getMemIntrinsicNode(Opc, DL, Op->getVTList(), Ops, IntVT, MMO);
2701 SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
2702 SelectionDAG &DAG) const {
2703 MachineFunction &MF = DAG.getMachineFunction();
2705 SDValue Chain = Op.getOperand(0);
2706 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2708 switch (IntrinsicID) {
2709 case AMDGPUIntrinsic::SI_sendmsg:
2710 case Intrinsic::amdgcn_s_sendmsg: {
2711 Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3));
2712 SDValue Glue = Chain.getValue(1);
2713 return DAG.getNode(AMDGPUISD::SENDMSG, DL, MVT::Other, Chain,
2714 Op.getOperand(2), Glue);
2716 case Intrinsic::amdgcn_s_sendmsghalt: {
2717 Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3));
2718 SDValue Glue = Chain.getValue(1);
2719 return DAG.getNode(AMDGPUISD::SENDMSGHALT, DL, MVT::Other, Chain,
2720 Op.getOperand(2), Glue);
2722 case AMDGPUIntrinsic::SI_tbuffer_store: {
2740 EVT VT = Op.getOperand(3).getValueType();
2742 MachineMemOperand *MMO = MF.getMachineMemOperand(
2743 MachinePointerInfo(),
2744 MachineMemOperand::MOStore,
2745 VT.getStoreSize(), 4);
2746 return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
2747 Op->getVTList(), Ops, VT, MMO);
2749 case AMDGPUIntrinsic::AMDGPU_kill: {
2750 SDValue Src = Op.getOperand(2);
2751 if (const ConstantFPSDNode *K = dyn_cast<ConstantFPSDNode>(Src)) {
2752 if (!K->isNegative())
2755 SDValue NegOne = DAG.getTargetConstant(FloatToBits(-1.0f), DL, MVT::i32);
2756 return DAG.getNode(AMDGPUISD::KILL, DL, MVT::Other, Chain, NegOne);
2759 SDValue Cast = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Src);
2760 return DAG.getNode(AMDGPUISD::KILL, DL, MVT::Other, Chain, Cast);
2762 case AMDGPUIntrinsic::SI_export: {
2763 const ConstantSDNode *En = cast<ConstantSDNode>(Op.getOperand(2));
2764 const ConstantSDNode *VM = cast<ConstantSDNode>(Op.getOperand(3));
2765 const ConstantSDNode *Done = cast<ConstantSDNode>(Op.getOperand(4));
2766 const ConstantSDNode *Tgt = cast<ConstantSDNode>(Op.getOperand(5));
2767 const ConstantSDNode *Compr = cast<ConstantSDNode>(Op.getOperand(6));
2769 const SDValue Ops[] = {
2771 DAG.getTargetConstant(En->getZExtValue(), DL, MVT::i8),
2772 DAG.getTargetConstant(VM->getZExtValue(), DL, MVT::i1),
2773 DAG.getTargetConstant(Tgt->getZExtValue(), DL, MVT::i8),
2774 DAG.getTargetConstant(Compr->getZExtValue(), DL, MVT::i1),
2775 Op.getOperand(7), // src0
2776 Op.getOperand(8), // src1
2777 Op.getOperand(9), // src2
2778 Op.getOperand(10) // src3
2781 unsigned Opc = Done->isNullValue() ?
2782 AMDGPUISD::EXPORT : AMDGPUISD::EXPORT_DONE;
2783 return DAG.getNode(Opc, DL, Op->getVTList(), Ops);
2790 SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2792 LoadSDNode *Load = cast<LoadSDNode>(Op);
2793 ISD::LoadExtType ExtType = Load->getExtensionType();
2794 EVT MemVT = Load->getMemoryVT();
2796 if (ExtType == ISD::NON_EXTLOAD && MemVT.getSizeInBits() < 32) {
2797 // FIXME: Copied from PPC
2798 // First, load into 32 bits, then truncate to 1 bit.
2800 SDValue Chain = Load->getChain();
2801 SDValue BasePtr = Load->getBasePtr();
2802 MachineMemOperand *MMO = Load->getMemOperand();
2804 EVT RealMemVT = (MemVT == MVT::i1) ? MVT::i8 : MVT::i16;
2806 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
2807 BasePtr, RealMemVT, MMO);
2810 DAG.getNode(ISD::TRUNCATE, DL, MemVT, NewLD),
2814 return DAG.getMergeValues(Ops, DL);
2817 if (!MemVT.isVector())
2820 assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
2821 "Custom lowering for non-i32 vectors hasn't been implemented.");
2823 unsigned AS = Load->getAddressSpace();
2824 if (!allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), MemVT,
2825 AS, Load->getAlignment())) {
2827 std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(Load, DAG);
2828 return DAG.getMergeValues(Ops, DL);
2831 MachineFunction &MF = DAG.getMachineFunction();
2832 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
2833 // If there is a possibilty that flat instruction access scratch memory
2834 // then we need to use the same legalization rules we use for private.
2835 if (AS == AMDGPUAS::FLAT_ADDRESS)
2836 AS = MFI->hasFlatScratchInit() ?
2837 AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
2839 unsigned NumElements = MemVT.getVectorNumElements();
2841 case AMDGPUAS::CONSTANT_ADDRESS:
2842 if (isMemOpUniform(Load))
2844 // Non-uniform loads will be selected to MUBUF instructions, so they
2845 // have the same legalization requirements as global and private
2849 case AMDGPUAS::GLOBAL_ADDRESS: {
2850 if (Subtarget->getScalarizeGlobalBehavior() && isMemOpUniform(Load) &&
2851 isMemOpHasNoClobberedMemOperand(Load))
2853 // Non-uniform loads will be selected to MUBUF instructions, so they
2854 // have the same legalization requirements as global and private
2859 case AMDGPUAS::FLAT_ADDRESS:
2860 if (NumElements > 4)
2861 return SplitVectorLoad(Op, DAG);
2862 // v4 loads are supported for private and global memory.
2864 case AMDGPUAS::PRIVATE_ADDRESS: {
2865 // Depending on the setting of the private_element_size field in the
2866 // resource descriptor, we can only make private accesses up to a certain
2868 switch (Subtarget->getMaxPrivateElementSize()) {
2870 return scalarizeVectorLoad(Load, DAG);
2872 if (NumElements > 2)
2873 return SplitVectorLoad(Op, DAG);
2876 // Same as global/flat
2877 if (NumElements > 4)
2878 return SplitVectorLoad(Op, DAG);
2881 llvm_unreachable("unsupported private_element_size");
2884 case AMDGPUAS::LOCAL_ADDRESS: {
2885 if (NumElements > 2)
2886 return SplitVectorLoad(Op, DAG);
2888 if (NumElements == 2)
2891 // If properly aligned, if we split we might be able to use ds_read_b64.
2892 return SplitVectorLoad(Op, DAG);
2899 SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2900 if (Op.getValueType() != MVT::i64)
2904 SDValue Cond = Op.getOperand(0);
2906 SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
2907 SDValue One = DAG.getConstant(1, DL, MVT::i32);
2909 SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
2910 SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));
2912 SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
2913 SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);
2915 SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);
2917 SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
2918 SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);
2920 SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);
2922 SDValue Res = DAG.getBuildVector(MVT::v2i32, DL, {Lo, Hi});
2923 return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
2926 // Catch division cases where we can use shortcuts with rcp and rsq
2928 SDValue SITargetLowering::lowerFastUnsafeFDIV(SDValue Op,
2929 SelectionDAG &DAG) const {
2931 SDValue LHS = Op.getOperand(0);
2932 SDValue RHS = Op.getOperand(1);
2933 EVT VT = Op.getValueType();
2934 bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
2936 if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
2937 if (Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals()) ||
2939 if (CLHS->isExactlyValue(1.0)) {
2940 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
2941 // the CI documentation has a worst case error of 1 ulp.
2942 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
2943 // use it as long as we aren't trying to use denormals.
2945 // v_rcp_f16 and v_rsq_f16 DO support denormals.
2947 // 1.0 / sqrt(x) -> rsq(x)
2949 // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
2950 // error seems really high at 2^29 ULP.
2951 if (RHS.getOpcode() == ISD::FSQRT)
2952 return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
2954 // 1.0 / x -> rcp(x)
2955 return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
2958 // Same as for 1.0, but expand the sign out of the constant.
2959 if (CLHS->isExactlyValue(-1.0)) {
2960 // -1.0 / x -> rcp (fneg x)
2961 SDValue FNegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
2962 return DAG.getNode(AMDGPUISD::RCP, SL, VT, FNegRHS);
2967 const SDNodeFlags *Flags = Op->getFlags();
2969 if (Unsafe || Flags->hasAllowReciprocal()) {
2970 // Turn into multiply by the reciprocal.
2971 // x / y -> x * (1.0 / y)
2973 Flags.setUnsafeAlgebra(true);
2974 SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
2975 return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip, &Flags);
2981 static SDValue getFPBinOp(SelectionDAG &DAG, unsigned Opcode, const SDLoc &SL,
2982 EVT VT, SDValue A, SDValue B, SDValue GlueChain) {
2983 if (GlueChain->getNumValues() <= 1) {
2984 return DAG.getNode(Opcode, SL, VT, A, B);
2987 assert(GlueChain->getNumValues() == 3);
2989 SDVTList VTList = DAG.getVTList(VT, MVT::Other, MVT::Glue);
2991 default: llvm_unreachable("no chain equivalent for opcode");
2993 Opcode = AMDGPUISD::FMUL_W_CHAIN;
2997 return DAG.getNode(Opcode, SL, VTList, GlueChain.getValue(1), A, B,
2998 GlueChain.getValue(2));
3001 static SDValue getFPTernOp(SelectionDAG &DAG, unsigned Opcode, const SDLoc &SL,
3002 EVT VT, SDValue A, SDValue B, SDValue C,
3003 SDValue GlueChain) {
3004 if (GlueChain->getNumValues() <= 1) {
3005 return DAG.getNode(Opcode, SL, VT, A, B, C);
3008 assert(GlueChain->getNumValues() == 3);
3010 SDVTList VTList = DAG.getVTList(VT, MVT::Other, MVT::Glue);
3012 default: llvm_unreachable("no chain equivalent for opcode");
3014 Opcode = AMDGPUISD::FMA_W_CHAIN;
3018 return DAG.getNode(Opcode, SL, VTList, GlueChain.getValue(1), A, B, C,
3019 GlueChain.getValue(2));
3022 SDValue SITargetLowering::LowerFDIV16(SDValue Op, SelectionDAG &DAG) const {
3023 if (SDValue FastLowered = lowerFastUnsafeFDIV(Op, DAG))
3027 SDValue Src0 = Op.getOperand(0);
3028 SDValue Src1 = Op.getOperand(1);
3030 SDValue CvtSrc0 = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, Src0);
3031 SDValue CvtSrc1 = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, Src1);
3033 SDValue RcpSrc1 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, CvtSrc1);
3034 SDValue Quot = DAG.getNode(ISD::FMUL, SL, MVT::f32, CvtSrc0, RcpSrc1);
3036 SDValue FPRoundFlag = DAG.getTargetConstant(0, SL, MVT::i32);
3037 SDValue BestQuot = DAG.getNode(ISD::FP_ROUND, SL, MVT::f16, Quot, FPRoundFlag);
3039 return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f16, BestQuot, Src1, Src0);
3042 // Faster 2.5 ULP division that does not support denormals.
3043 SDValue SITargetLowering::lowerFDIV_FAST(SDValue Op, SelectionDAG &DAG) const {
3045 SDValue LHS = Op.getOperand(1);
3046 SDValue RHS = Op.getOperand(2);
3048 SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);
3050 const APFloat K0Val(BitsToFloat(0x6f800000));
3051 const SDValue K0 = DAG.getConstantFP(K0Val, SL, MVT::f32);
3053 const APFloat K1Val(BitsToFloat(0x2f800000));
3054 const SDValue K1 = DAG.getConstantFP(K1Val, SL, MVT::f32);
3056 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
3059 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f32);
3061 SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);
3063 SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);
3065 // TODO: Should this propagate fast-math-flags?
3066 r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);
3068 // rcp does not support denormals.
3069 SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);
3071 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);
3073 return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
3076 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
3077 if (SDValue FastLowered = lowerFastUnsafeFDIV(Op, DAG))
3081 SDValue LHS = Op.getOperand(0);
3082 SDValue RHS = Op.getOperand(1);
3084 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
3086 SDVTList ScaleVT = DAG.getVTList(MVT::f32, MVT::i1);
3088 SDValue DenominatorScaled = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT,
3090 SDValue NumeratorScaled = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT,
3093 // Denominator is scaled to not be denormal, so using rcp is ok.
3094 SDValue ApproxRcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32,
3096 SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f32,
3099 const unsigned Denorm32Reg = AMDGPU::Hwreg::ID_MODE |
3100 (4 << AMDGPU::Hwreg::OFFSET_SHIFT_) |
3101 (1 << AMDGPU::Hwreg::WIDTH_M1_SHIFT_);
3103 const SDValue BitField = DAG.getTargetConstant(Denorm32Reg, SL, MVT::i16);
3105 if (!Subtarget->hasFP32Denormals()) {
3106 SDVTList BindParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
3107 const SDValue EnableDenormValue = DAG.getConstant(FP_DENORM_FLUSH_NONE,
3109 SDValue EnableDenorm = DAG.getNode(AMDGPUISD::SETREG, SL, BindParamVTs,
3111 EnableDenormValue, BitField);
3114 EnableDenorm.getValue(0),
3115 EnableDenorm.getValue(1)
3118 NegDivScale0 = DAG.getMergeValues(Ops, SL);
3121 SDValue Fma0 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, NegDivScale0,
3122 ApproxRcp, One, NegDivScale0);
3124 SDValue Fma1 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, Fma0, ApproxRcp,
3127 SDValue Mul = getFPBinOp(DAG, ISD::FMUL, SL, MVT::f32, NumeratorScaled,
3130 SDValue Fma2 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, NegDivScale0, Mul,
3131 NumeratorScaled, Mul);
3133 SDValue Fma3 = getFPTernOp(DAG, ISD::FMA,SL, MVT::f32, Fma2, Fma1, Mul, Fma2);
3135 SDValue Fma4 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, NegDivScale0, Fma3,
3136 NumeratorScaled, Fma3);
3138 if (!Subtarget->hasFP32Denormals()) {
3139 const SDValue DisableDenormValue =
3140 DAG.getConstant(FP_DENORM_FLUSH_IN_FLUSH_OUT, SL, MVT::i32);
3141 SDValue DisableDenorm = DAG.getNode(AMDGPUISD::SETREG, SL, MVT::Other,
3147 SDValue OutputChain = DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
3148 DisableDenorm, DAG.getRoot());
3149 DAG.setRoot(OutputChain);
3152 SDValue Scale = NumeratorScaled.getValue(1);
3153 SDValue Fmas = DAG.getNode(AMDGPUISD::DIV_FMAS, SL, MVT::f32,
3154 Fma4, Fma1, Fma3, Scale);
3156 return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f32, Fmas, RHS, LHS);
3159 SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
3160 if (DAG.getTarget().Options.UnsafeFPMath)
3161 return lowerFastUnsafeFDIV(Op, DAG);
3164 SDValue X = Op.getOperand(0);
3165 SDValue Y = Op.getOperand(1);
3167 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
3169 SDVTList ScaleVT = DAG.getVTList(MVT::f64, MVT::i1);
3171 SDValue DivScale0 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, Y, Y, X);
3173 SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f64, DivScale0);
3175 SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f64, DivScale0);
3177 SDValue Fma0 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Rcp, One);
3179 SDValue Fma1 = DAG.getNode(ISD::FMA, SL, MVT::f64, Rcp, Fma0, Rcp);
3181 SDValue Fma2 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Fma1, One);
3183 SDValue DivScale1 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, X, Y, X);
3185 SDValue Fma3 = DAG.getNode(ISD::FMA, SL, MVT::f64, Fma1, Fma2, Fma1);
3186 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, DivScale1, Fma3);
3188 SDValue Fma4 = DAG.getNode(ISD::FMA, SL, MVT::f64,
3189 NegDivScale0, Mul, DivScale1);
3193 if (Subtarget->getGeneration() == SISubtarget::SOUTHERN_ISLANDS) {
3194 // Workaround a hardware bug on SI where the condition output from div_scale
3197 const SDValue Hi = DAG.getConstant(1, SL, MVT::i32);
3199 // Figure out if the scale to use for div_fmas.
3200 SDValue NumBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
3201 SDValue DenBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Y);
3202 SDValue Scale0BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale0);
3203 SDValue Scale1BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale1);
3205 SDValue NumHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, NumBC, Hi);
3206 SDValue DenHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, DenBC, Hi);
3209 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale0BC, Hi);
3211 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale1BC, Hi);
3213 SDValue CmpDen = DAG.getSetCC(SL, MVT::i1, DenHi, Scale0Hi, ISD::SETEQ);
3214 SDValue CmpNum = DAG.getSetCC(SL, MVT::i1, NumHi, Scale1Hi, ISD::SETEQ);
3215 Scale = DAG.getNode(ISD::XOR, SL, MVT::i1, CmpNum, CmpDen);
3217 Scale = DivScale1.getValue(1);
3220 SDValue Fmas = DAG.getNode(AMDGPUISD::DIV_FMAS, SL, MVT::f64,
3221 Fma4, Fma3, Mul, Scale);
3223 return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f64, Fmas, Y, X);
3226 SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
3227 EVT VT = Op.getValueType();
3230 return LowerFDIV32(Op, DAG);
3233 return LowerFDIV64(Op, DAG);
3236 return LowerFDIV16(Op, DAG);
3238 llvm_unreachable("Unexpected type for fdiv");
3241 SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
3243 StoreSDNode *Store = cast<StoreSDNode>(Op);
3244 EVT VT = Store->getMemoryVT();
3246 if (VT == MVT::i1) {
3247 return DAG.getTruncStore(Store->getChain(), DL,
3248 DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
3249 Store->getBasePtr(), MVT::i1, Store->getMemOperand());
3252 assert(VT.isVector() &&
3253 Store->getValue().getValueType().getScalarType() == MVT::i32);
3255 unsigned AS = Store->getAddressSpace();
3256 if (!allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
3257 AS, Store->getAlignment())) {
3258 return expandUnalignedStore(Store, DAG);
3261 MachineFunction &MF = DAG.getMachineFunction();
3262 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
3263 // If there is a possibilty that flat instruction access scratch memory
3264 // then we need to use the same legalization rules we use for private.
3265 if (AS == AMDGPUAS::FLAT_ADDRESS)
3266 AS = MFI->hasFlatScratchInit() ?
3267 AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
3269 unsigned NumElements = VT.getVectorNumElements();
3271 case AMDGPUAS::GLOBAL_ADDRESS:
3272 case AMDGPUAS::FLAT_ADDRESS:
3273 if (NumElements > 4)
3274 return SplitVectorStore(Op, DAG);
3276 case AMDGPUAS::PRIVATE_ADDRESS: {
3277 switch (Subtarget->getMaxPrivateElementSize()) {
3279 return scalarizeVectorStore(Store, DAG);
3281 if (NumElements > 2)
3282 return SplitVectorStore(Op, DAG);
3285 if (NumElements > 4)
3286 return SplitVectorStore(Op, DAG);
3289 llvm_unreachable("unsupported private_element_size");
3292 case AMDGPUAS::LOCAL_ADDRESS: {
3293 if (NumElements > 2)
3294 return SplitVectorStore(Op, DAG);
3296 if (NumElements == 2)
3299 // If properly aligned, if we split we might be able to use ds_write_b64.
3300 return SplitVectorStore(Op, DAG);
3303 llvm_unreachable("unhandled address space");
3307 SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
3309 EVT VT = Op.getValueType();
3310 SDValue Arg = Op.getOperand(0);
3311 // TODO: Should this propagate fast-math-flags?
3312 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT,
3313 DAG.getNode(ISD::FMUL, DL, VT, Arg,
3314 DAG.getConstantFP(0.5/M_PI, DL,
3317 switch (Op.getOpcode()) {
3319 return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
3321 return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
3323 llvm_unreachable("Wrong trig opcode");
3327 SDValue SITargetLowering::LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
3328 AtomicSDNode *AtomicNode = cast<AtomicSDNode>(Op);
3329 assert(AtomicNode->isCompareAndSwap());
3330 unsigned AS = AtomicNode->getAddressSpace();
3332 // No custom lowering required for local address space
3333 if (!isFlatGlobalAddrSpace(AS))
3336 // Non-local address space requires custom lowering for atomic compare
3337 // and swap; cmp and swap should be in a v2i32 or v2i64 in case of _X2
3339 SDValue ChainIn = Op.getOperand(0);
3340 SDValue Addr = Op.getOperand(1);
3341 SDValue Old = Op.getOperand(2);
3342 SDValue New = Op.getOperand(3);
3343 EVT VT = Op.getValueType();
3344 MVT SimpleVT = VT.getSimpleVT();
3345 MVT VecType = MVT::getVectorVT(SimpleVT, 2);
3347 SDValue NewOld = DAG.getBuildVector(VecType, DL, {New, Old});
3348 SDValue Ops[] = { ChainIn, Addr, NewOld };
3350 return DAG.getMemIntrinsicNode(AMDGPUISD::ATOMIC_CMP_SWAP, DL, Op->getVTList(),
3351 Ops, VT, AtomicNode->getMemOperand());
3354 //===----------------------------------------------------------------------===//
3355 // Custom DAG optimizations
3356 //===----------------------------------------------------------------------===//
3358 SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
3359 DAGCombinerInfo &DCI) const {
3360 EVT VT = N->getValueType(0);
3361 EVT ScalarVT = VT.getScalarType();
3362 if (ScalarVT != MVT::f32)
3365 SelectionDAG &DAG = DCI.DAG;
3368 SDValue Src = N->getOperand(0);
3369 EVT SrcVT = Src.getValueType();
3371 // TODO: We could try to match extracting the higher bytes, which would be
3372 // easier if i8 vectors weren't promoted to i32 vectors, particularly after
3373 // types are legalized. v4i8 -> v4f32 is probably the only case to worry
3374 // about in practice.
3375 if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
3376 if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
3377 SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
3378 DCI.AddToWorklist(Cvt.getNode());
3386 /// \brief Return true if the given offset Size in bytes can be folded into
3387 /// the immediate offsets of a memory instruction for the given address space.
3388 static bool canFoldOffset(unsigned OffsetSize, unsigned AS,
3389 const SISubtarget &STI) {
3391 case AMDGPUAS::GLOBAL_ADDRESS: {
3392 // MUBUF instructions a 12-bit offset in bytes.
3393 return isUInt<12>(OffsetSize);
3395 case AMDGPUAS::CONSTANT_ADDRESS: {
3396 // SMRD instructions have an 8-bit offset in dwords on SI and
3397 // a 20-bit offset in bytes on VI.
3398 if (STI.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
3399 return isUInt<20>(OffsetSize);
3401 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
3403 case AMDGPUAS::LOCAL_ADDRESS:
3404 case AMDGPUAS::REGION_ADDRESS: {
3405 // The single offset versions have a 16-bit offset in bytes.
3406 return isUInt<16>(OffsetSize);
3408 case AMDGPUAS::PRIVATE_ADDRESS:
3409 // Indirect register addressing does not use any offsets.
3415 // (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)
3417 // This is a variant of
3418 // (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
3420 // The normal DAG combiner will do this, but only if the add has one use since
3421 // that would increase the number of instructions.
3423 // This prevents us from seeing a constant offset that can be folded into a
3424 // memory instruction's addressing mode. If we know the resulting add offset of
3425 // a pointer can be folded into an addressing offset, we can replace the pointer
3426 // operand with the add of new constant offset. This eliminates one of the uses,
3427 // and may allow the remaining use to also be simplified.
3429 SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
3431 DAGCombinerInfo &DCI) const {
3432 SDValue N0 = N->getOperand(0);
3433 SDValue N1 = N->getOperand(1);
3435 if (N0.getOpcode() != ISD::ADD)
3438 const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
3442 const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3446 // If the resulting offset is too large, we can't fold it into the addressing
3448 APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
3449 if (!canFoldOffset(Offset.getZExtValue(), AddrSpace, *getSubtarget()))
3452 SelectionDAG &DAG = DCI.DAG;
3454 EVT VT = N->getValueType(0);
3456 SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
3457 SDValue COffset = DAG.getConstant(Offset, SL, MVT::i32);
3459 return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
3462 SDValue SITargetLowering::performMemSDNodeCombine(MemSDNode *N,
3463 DAGCombinerInfo &DCI) const {
3464 SDValue Ptr = N->getBasePtr();
3465 SelectionDAG &DAG = DCI.DAG;
3468 // TODO: We could also do this for multiplies.
3469 unsigned AS = N->getAddressSpace();
3470 if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
3471 SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
3473 SmallVector<SDValue, 8> NewOps(N->op_begin(), N->op_end());
3475 NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
3476 return SDValue(DAG.UpdateNodeOperands(N, NewOps), 0);
3483 static bool bitOpWithConstantIsReducible(unsigned Opc, uint32_t Val) {
3484 return (Opc == ISD::AND && (Val == 0 || Val == 0xffffffff)) ||
3485 (Opc == ISD::OR && (Val == 0xffffffff || Val == 0)) ||
3486 (Opc == ISD::XOR && Val == 0);
3489 // Break up 64-bit bit operation of a constant into two 32-bit and/or/xor. This
3490 // will typically happen anyway for a VALU 64-bit and. This exposes other 32-bit
3491 // integer combine opportunities since most 64-bit operations are decomposed
3492 // this way. TODO: We won't want this for SALU especially if it is an inline
3494 SDValue SITargetLowering::splitBinaryBitConstantOp(
3495 DAGCombinerInfo &DCI,
3497 unsigned Opc, SDValue LHS,
3498 const ConstantSDNode *CRHS) const {
3499 uint64_t Val = CRHS->getZExtValue();
3500 uint32_t ValLo = Lo_32(Val);
3501 uint32_t ValHi = Hi_32(Val);
3502 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
3504 if ((bitOpWithConstantIsReducible(Opc, ValLo) ||
3505 bitOpWithConstantIsReducible(Opc, ValHi)) ||
3506 (CRHS->hasOneUse() && !TII->isInlineConstant(CRHS->getAPIntValue()))) {
3507 // If we need to materialize a 64-bit immediate, it will be split up later
3508 // anyway. Avoid creating the harder to understand 64-bit immediate
3510 return splitBinaryBitConstantOpImpl(DCI, SL, Opc, LHS, ValLo, ValHi);
3516 SDValue SITargetLowering::performAndCombine(SDNode *N,
3517 DAGCombinerInfo &DCI) const {
3518 if (DCI.isBeforeLegalize())
3521 SelectionDAG &DAG = DCI.DAG;
3522 EVT VT = N->getValueType(0);
3523 SDValue LHS = N->getOperand(0);
3524 SDValue RHS = N->getOperand(1);
3527 if (VT == MVT::i64) {
3528 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS);
3531 = splitBinaryBitConstantOp(DCI, SDLoc(N), ISD::AND, LHS, CRHS))
3536 // (and (fcmp ord x, x), (fcmp une (fabs x), inf)) ->
3537 // fp_class x, ~(s_nan | q_nan | n_infinity | p_infinity)
3538 if (LHS.getOpcode() == ISD::SETCC && RHS.getOpcode() == ISD::SETCC) {
3539 ISD::CondCode LCC = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
3540 ISD::CondCode RCC = cast<CondCodeSDNode>(RHS.getOperand(2))->get();
3542 SDValue X = LHS.getOperand(0);
3543 SDValue Y = RHS.getOperand(0);
3544 if (Y.getOpcode() != ISD::FABS || Y.getOperand(0) != X)
3547 if (LCC == ISD::SETO) {
3548 if (X != LHS.getOperand(1))
3551 if (RCC == ISD::SETUNE) {
3552 const ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(RHS.getOperand(1));
3553 if (!C1 || !C1->isInfinity() || C1->isNegative())
3556 const uint32_t Mask = SIInstrFlags::N_NORMAL |
3557 SIInstrFlags::N_SUBNORMAL |
3558 SIInstrFlags::N_ZERO |
3559 SIInstrFlags::P_ZERO |
3560 SIInstrFlags::P_SUBNORMAL |
3561 SIInstrFlags::P_NORMAL;
3563 static_assert(((~(SIInstrFlags::S_NAN |
3564 SIInstrFlags::Q_NAN |
3565 SIInstrFlags::N_INFINITY |
3566 SIInstrFlags::P_INFINITY)) & 0x3ff) == Mask,
3570 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
3571 X, DAG.getConstant(Mask, DL, MVT::i32));
3579 SDValue SITargetLowering::performOrCombine(SDNode *N,
3580 DAGCombinerInfo &DCI) const {
3581 SelectionDAG &DAG = DCI.DAG;
3582 SDValue LHS = N->getOperand(0);
3583 SDValue RHS = N->getOperand(1);
3585 EVT VT = N->getValueType(0);
3586 if (VT == MVT::i1) {
3587 // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2)
3588 if (LHS.getOpcode() == AMDGPUISD::FP_CLASS &&
3589 RHS.getOpcode() == AMDGPUISD::FP_CLASS) {
3590 SDValue Src = LHS.getOperand(0);
3591 if (Src != RHS.getOperand(0))
3594 const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
3595 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1));
3599 // Only 10 bits are used.
3600 static const uint32_t MaxMask = 0x3ff;
3602 uint32_t NewMask = (CLHS->getZExtValue() | CRHS->getZExtValue()) & MaxMask;
3604 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
3605 Src, DAG.getConstant(NewMask, DL, MVT::i32));
3614 // TODO: This could be a generic combine with a predicate for extracting the
3615 // high half of an integer being free.
3617 // (or i64:x, (zero_extend i32:y)) ->
3618 // i64 (bitcast (v2i32 build_vector (or i32:y, lo_32(x)), hi_32(x)))
3619 if (LHS.getOpcode() == ISD::ZERO_EXTEND &&
3620 RHS.getOpcode() != ISD::ZERO_EXTEND)
3621 std::swap(LHS, RHS);
3623 if (RHS.getOpcode() == ISD::ZERO_EXTEND) {
3624 SDValue ExtSrc = RHS.getOperand(0);
3625 EVT SrcVT = ExtSrc.getValueType();
3626 if (SrcVT == MVT::i32) {
3628 SDValue LowLHS, HiBits;
3629 std::tie(LowLHS, HiBits) = split64BitValue(LHS, DAG);
3630 SDValue LowOr = DAG.getNode(ISD::OR, SL, MVT::i32, LowLHS, ExtSrc);
3632 DCI.AddToWorklist(LowOr.getNode());
3633 DCI.AddToWorklist(HiBits.getNode());
3635 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
3637 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
3641 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3644 = splitBinaryBitConstantOp(DCI, SDLoc(N), ISD::OR, LHS, CRHS))
3651 SDValue SITargetLowering::performXorCombine(SDNode *N,
3652 DAGCombinerInfo &DCI) const {
3653 EVT VT = N->getValueType(0);
3657 SDValue LHS = N->getOperand(0);
3658 SDValue RHS = N->getOperand(1);
3660 const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS);
3663 = splitBinaryBitConstantOp(DCI, SDLoc(N), ISD::XOR, LHS, CRHS))
3670 SDValue SITargetLowering::performClassCombine(SDNode *N,
3671 DAGCombinerInfo &DCI) const {
3672 SelectionDAG &DAG = DCI.DAG;
3673 SDValue Mask = N->getOperand(1);
3675 // fp_class x, 0 -> false
3676 if (const ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Mask)) {
3677 if (CMask->isNullValue())
3678 return DAG.getConstant(0, SDLoc(N), MVT::i1);
3681 if (N->getOperand(0).isUndef())
3682 return DAG.getUNDEF(MVT::i1);
3687 // Constant fold canonicalize.
3688 SDValue SITargetLowering::performFCanonicalizeCombine(
3690 DAGCombinerInfo &DCI) const {
3691 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
3695 SelectionDAG &DAG = DCI.DAG;
3696 const APFloat &C = CFP->getValueAPF();
3698 // Flush denormals to 0 if not enabled.
3699 if (C.isDenormal()) {
3700 EVT VT = N->getValueType(0);
3701 if (VT == MVT::f32 && !Subtarget->hasFP32Denormals())
3702 return DAG.getConstantFP(0.0, SDLoc(N), VT);
3704 if (VT == MVT::f64 && !Subtarget->hasFP64Denormals())
3705 return DAG.getConstantFP(0.0, SDLoc(N), VT);
3707 if (VT == MVT::f16 && !Subtarget->hasFP16Denormals())
3708 return DAG.getConstantFP(0.0, SDLoc(N), VT);
3712 EVT VT = N->getValueType(0);
3713 APFloat CanonicalQNaN = APFloat::getQNaN(C.getSemantics());
3714 if (C.isSignaling()) {
3715 // Quiet a signaling NaN.
3716 return DAG.getConstantFP(CanonicalQNaN, SDLoc(N), VT);
3719 // Make sure it is the canonical NaN bitpattern.
3721 // TODO: Can we use -1 as the canonical NaN value since it's an inline
3723 if (C.bitcastToAPInt() != CanonicalQNaN.bitcastToAPInt())
3724 return DAG.getConstantFP(CanonicalQNaN, SDLoc(N), VT);
3727 return SDValue(CFP, 0);
3730 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {
3733 return AMDGPUISD::FMAX3;
3735 return AMDGPUISD::SMAX3;
3737 return AMDGPUISD::UMAX3;
3739 return AMDGPUISD::FMIN3;
3741 return AMDGPUISD::SMIN3;
3743 return AMDGPUISD::UMIN3;
3745 llvm_unreachable("Not a min/max opcode");
3749 static SDValue performIntMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
3750 SDValue Op0, SDValue Op1, bool Signed) {
3751 ConstantSDNode *K1 = dyn_cast<ConstantSDNode>(Op1);
3755 ConstantSDNode *K0 = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
3760 if (K0->getAPIntValue().sge(K1->getAPIntValue()))
3763 if (K0->getAPIntValue().uge(K1->getAPIntValue()))
3767 EVT VT = K0->getValueType(0);
3770 unsigned ExtOp = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3772 SDValue Tmp1, Tmp2, Tmp3;
3773 Tmp1 = DAG.getNode(ExtOp, SL, NVT, Op0->getOperand(0));
3774 Tmp2 = DAG.getNode(ExtOp, SL, NVT, Op0->getOperand(1));
3775 Tmp3 = DAG.getNode(ExtOp, SL, NVT, Op1);
3777 if (VT == MVT::i16) {
3778 Tmp1 = DAG.getNode(Signed ? AMDGPUISD::SMED3 : AMDGPUISD::UMED3, SL, NVT,
3781 return DAG.getNode(ISD::TRUNCATE, SL, VT, Tmp1);
3783 return DAG.getNode(Signed ? AMDGPUISD::SMED3 : AMDGPUISD::UMED3, SL, VT,
3784 Op0.getOperand(0), SDValue(K0, 0), SDValue(K1, 0));
3787 static bool isKnownNeverSNan(SelectionDAG &DAG, SDValue Op) {
3788 if (!DAG.getTargetLoweringInfo().hasFloatingPointExceptions())
3791 return DAG.isKnownNeverNaN(Op);
3794 static SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
3795 SDValue Op0, SDValue Op1) {
3796 ConstantFPSDNode *K1 = dyn_cast<ConstantFPSDNode>(Op1);
3800 ConstantFPSDNode *K0 = dyn_cast<ConstantFPSDNode>(Op0.getOperand(1));
3804 // Ordered >= (although NaN inputs should have folded away by now).
3805 APFloat::cmpResult Cmp = K0->getValueAPF().compare(K1->getValueAPF());
3806 if (Cmp == APFloat::cmpGreaterThan)
3809 // This isn't safe with signaling NaNs because in IEEE mode, min/max on a
3810 // signaling NaN gives a quiet NaN. The quiet NaN input to the min would then
3811 // give the other result, which is different from med3 with a NaN input.
3812 SDValue Var = Op0.getOperand(0);
3813 if (!isKnownNeverSNan(DAG, Var))
3816 return DAG.getNode(AMDGPUISD::FMED3, SL, K0->getValueType(0),
3817 Var, SDValue(K0, 0), SDValue(K1, 0));
3820 SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
3821 DAGCombinerInfo &DCI) const {
3822 SelectionDAG &DAG = DCI.DAG;
3824 unsigned Opc = N->getOpcode();
3825 SDValue Op0 = N->getOperand(0);
3826 SDValue Op1 = N->getOperand(1);
3828 // Only do this if the inner op has one use since this will just increases
3829 // register pressure for no benefit.
3831 if (Opc != AMDGPUISD::FMIN_LEGACY && Opc != AMDGPUISD::FMAX_LEGACY) {
3832 // max(max(a, b), c) -> max3(a, b, c)
3833 // min(min(a, b), c) -> min3(a, b, c)
3834 if (Op0.getOpcode() == Opc && Op0.hasOneUse()) {
3836 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
3845 // max(a, max(b, c)) -> max3(a, b, c)
3846 // min(a, min(b, c)) -> min3(a, b, c)
3847 if (Op1.getOpcode() == Opc && Op1.hasOneUse()) {
3849 return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
3858 // min(max(x, K0), K1), K0 < K1 -> med3(x, K0, K1)
3859 if (Opc == ISD::SMIN && Op0.getOpcode() == ISD::SMAX && Op0.hasOneUse()) {
3860 if (SDValue Med3 = performIntMed3ImmCombine(DAG, SDLoc(N), Op0, Op1, true))
3864 if (Opc == ISD::UMIN && Op0.getOpcode() == ISD::UMAX && Op0.hasOneUse()) {
3865 if (SDValue Med3 = performIntMed3ImmCombine(DAG, SDLoc(N), Op0, Op1, false))
3869 // fminnum(fmaxnum(x, K0), K1), K0 < K1 && !is_snan(x) -> fmed3(x, K0, K1)
3870 if (((Opc == ISD::FMINNUM && Op0.getOpcode() == ISD::FMAXNUM) ||
3871 (Opc == AMDGPUISD::FMIN_LEGACY &&
3872 Op0.getOpcode() == AMDGPUISD::FMAX_LEGACY)) &&
3873 N->getValueType(0) == MVT::f32 && Op0.hasOneUse()) {
3874 if (SDValue Res = performFPMed3ImmCombine(DAG, SDLoc(N), Op0, Op1))
3881 unsigned SITargetLowering::getFusedOpcode(const SelectionDAG &DAG,
3883 const SDNode *N1) const {
3884 EVT VT = N0->getValueType(0);
3886 // Only do this if we are not trying to support denormals. v_mad_f32 does not
3887 // support denormals ever.
3888 if ((VT == MVT::f32 && !Subtarget->hasFP32Denormals()) ||
3889 (VT == MVT::f16 && !Subtarget->hasFP16Denormals()))
3892 const TargetOptions &Options = DAG.getTarget().Options;
3893 if ((Options.AllowFPOpFusion == FPOpFusion::Fast ||
3894 Options.UnsafeFPMath ||
3895 (cast<BinaryWithFlagsSDNode>(N0)->Flags.hasUnsafeAlgebra() &&
3896 cast<BinaryWithFlagsSDNode>(N1)->Flags.hasUnsafeAlgebra())) &&
3897 isFMAFasterThanFMulAndFAdd(VT)) {
3904 SDValue SITargetLowering::performFAddCombine(SDNode *N,
3905 DAGCombinerInfo &DCI) const {
3906 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3909 SelectionDAG &DAG = DCI.DAG;
3910 EVT VT = N->getValueType(0);
3911 assert(!VT.isVector());
3914 SDValue LHS = N->getOperand(0);
3915 SDValue RHS = N->getOperand(1);
3917 // These should really be instruction patterns, but writing patterns with
3918 // source modiifiers is a pain.
3920 // fadd (fadd (a, a), b) -> mad 2.0, a, b
3921 if (LHS.getOpcode() == ISD::FADD) {
3922 SDValue A = LHS.getOperand(0);
3923 if (A == LHS.getOperand(1)) {
3924 unsigned FusedOp = getFusedOpcode(DAG, N, LHS.getNode());
3926 const SDValue Two = DAG.getConstantFP(2.0, SL, VT);
3927 return DAG.getNode(FusedOp, SL, VT, A, Two, RHS);
3932 // fadd (b, fadd (a, a)) -> mad 2.0, a, b
3933 if (RHS.getOpcode() == ISD::FADD) {
3934 SDValue A = RHS.getOperand(0);
3935 if (A == RHS.getOperand(1)) {
3936 unsigned FusedOp = getFusedOpcode(DAG, N, RHS.getNode());
3938 const SDValue Two = DAG.getConstantFP(2.0, SL, VT);
3939 return DAG.getNode(FusedOp, SL, VT, A, Two, LHS);
3947 SDValue SITargetLowering::performFSubCombine(SDNode *N,
3948 DAGCombinerInfo &DCI) const {
3949 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3952 SelectionDAG &DAG = DCI.DAG;
3954 EVT VT = N->getValueType(0);
3955 assert(!VT.isVector());
3957 // Try to get the fneg to fold into the source modifier. This undoes generic
3958 // DAG combines and folds them into the mad.
3960 // Only do this if we are not trying to support denormals. v_mad_f32 does
3961 // not support denormals ever.
3962 SDValue LHS = N->getOperand(0);
3963 SDValue RHS = N->getOperand(1);
3964 if (LHS.getOpcode() == ISD::FADD) {
3965 // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c)
3966 SDValue A = LHS.getOperand(0);
3967 if (A == LHS.getOperand(1)) {
3968 unsigned FusedOp = getFusedOpcode(DAG, N, LHS.getNode());
3970 const SDValue Two = DAG.getConstantFP(2.0, SL, VT);
3971 SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3973 return DAG.getNode(FusedOp, SL, VT, A, Two, NegRHS);
3978 if (RHS.getOpcode() == ISD::FADD) {
3979 // (fsub c, (fadd a, a)) -> mad -2.0, a, c
3981 SDValue A = RHS.getOperand(0);
3982 if (A == RHS.getOperand(1)) {
3983 unsigned FusedOp = getFusedOpcode(DAG, N, RHS.getNode());
3985 const SDValue NegTwo = DAG.getConstantFP(-2.0, SL, VT);
3986 return DAG.getNode(FusedOp, SL, VT, A, NegTwo, LHS);
3994 SDValue SITargetLowering::performSetCCCombine(SDNode *N,
3995 DAGCombinerInfo &DCI) const {
3996 SelectionDAG &DAG = DCI.DAG;
3999 SDValue LHS = N->getOperand(0);
4000 SDValue RHS = N->getOperand(1);
4001 EVT VT = LHS.getValueType();
4003 if (VT != MVT::f32 && VT != MVT::f64 && (Subtarget->has16BitInsts() &&
4007 // Match isinf pattern
4008 // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity))
4009 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
4010 if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) {
4011 const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
4015 const APFloat &APF = CRHS->getValueAPF();
4016 if (APF.isInfinity() && !APF.isNegative()) {
4017 unsigned Mask = SIInstrFlags::P_INFINITY | SIInstrFlags::N_INFINITY;
4018 return DAG.getNode(AMDGPUISD::FP_CLASS, SL, MVT::i1, LHS.getOperand(0),
4019 DAG.getConstant(Mask, SL, MVT::i32));
4026 SDValue SITargetLowering::performCvtF32UByteNCombine(SDNode *N,
4027 DAGCombinerInfo &DCI) const {
4028 SelectionDAG &DAG = DCI.DAG;
4030 unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
4032 SDValue Src = N->getOperand(0);
4033 SDValue Srl = N->getOperand(0);
4034 if (Srl.getOpcode() == ISD::ZERO_EXTEND)
4035 Srl = Srl.getOperand(0);
4037 // TODO: Handle (or x, (srl y, 8)) pattern when known bits are zero.
4038 if (Srl.getOpcode() == ISD::SRL) {
4039 // cvt_f32_ubyte0 (srl x, 16) -> cvt_f32_ubyte2 x
4040 // cvt_f32_ubyte1 (srl x, 16) -> cvt_f32_ubyte3 x
4041 // cvt_f32_ubyte0 (srl x, 8) -> cvt_f32_ubyte1 x
4043 if (const ConstantSDNode *C =
4044 dyn_cast<ConstantSDNode>(Srl.getOperand(1))) {
4045 Srl = DAG.getZExtOrTrunc(Srl.getOperand(0), SDLoc(Srl.getOperand(0)),
4048 unsigned SrcOffset = C->getZExtValue() + 8 * Offset;
4049 if (SrcOffset < 32 && SrcOffset % 8 == 0) {
4050 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0 + SrcOffset / 8, SL,
4056 APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
4058 APInt KnownZero, KnownOne;
4059 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
4060 !DCI.isBeforeLegalizeOps());
4061 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4062 if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
4063 TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
4064 DCI.CommitTargetLoweringOpt(TLO);
4070 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
4071 DAGCombinerInfo &DCI) const {
4072 switch (N->getOpcode()) {
4074 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
4076 return performFAddCombine(N, DCI);
4078 return performFSubCombine(N, DCI);
4080 return performSetCCCombine(N, DCI);
4087 case AMDGPUISD::FMIN_LEGACY:
4088 case AMDGPUISD::FMAX_LEGACY: {
4089 if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
4090 N->getValueType(0) != MVT::f64 &&
4091 getTargetMachine().getOptLevel() > CodeGenOpt::None)
4092 return performMinMaxCombine(N, DCI);
4097 case ISD::ATOMIC_LOAD:
4098 case ISD::ATOMIC_STORE:
4099 case ISD::ATOMIC_CMP_SWAP:
4100 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
4101 case ISD::ATOMIC_SWAP:
4102 case ISD::ATOMIC_LOAD_ADD:
4103 case ISD::ATOMIC_LOAD_SUB:
4104 case ISD::ATOMIC_LOAD_AND:
4105 case ISD::ATOMIC_LOAD_OR:
4106 case ISD::ATOMIC_LOAD_XOR:
4107 case ISD::ATOMIC_LOAD_NAND:
4108 case ISD::ATOMIC_LOAD_MIN:
4109 case ISD::ATOMIC_LOAD_MAX:
4110 case ISD::ATOMIC_LOAD_UMIN:
4111 case ISD::ATOMIC_LOAD_UMAX:
4112 case AMDGPUISD::ATOMIC_INC:
4113 case AMDGPUISD::ATOMIC_DEC: { // TODO: Target mem intrinsics.
4114 if (DCI.isBeforeLegalize())
4116 return performMemSDNodeCombine(cast<MemSDNode>(N), DCI);
4119 return performAndCombine(N, DCI);
4121 return performOrCombine(N, DCI);
4123 return performXorCombine(N, DCI);
4124 case AMDGPUISD::FP_CLASS:
4125 return performClassCombine(N, DCI);
4126 case ISD::FCANONICALIZE:
4127 return performFCanonicalizeCombine(N, DCI);
4128 case AMDGPUISD::FRACT:
4129 case AMDGPUISD::RCP:
4130 case AMDGPUISD::RSQ:
4131 case AMDGPUISD::RCP_LEGACY:
4132 case AMDGPUISD::RSQ_LEGACY:
4133 case AMDGPUISD::RSQ_CLAMP:
4134 case AMDGPUISD::LDEXP: {
4135 SDValue Src = N->getOperand(0);
4140 case ISD::SINT_TO_FP:
4141 case ISD::UINT_TO_FP:
4142 return performUCharToFloatCombine(N, DCI);
4143 case AMDGPUISD::CVT_F32_UBYTE0:
4144 case AMDGPUISD::CVT_F32_UBYTE1:
4145 case AMDGPUISD::CVT_F32_UBYTE2:
4146 case AMDGPUISD::CVT_F32_UBYTE3:
4147 return performCvtF32UByteNCombine(N, DCI);
4149 return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
4152 /// \brief Helper function for adjustWritemask
4153 static unsigned SubIdx2Lane(unsigned Idx) {
4156 case AMDGPU::sub0: return 0;
4157 case AMDGPU::sub1: return 1;
4158 case AMDGPU::sub2: return 2;
4159 case AMDGPU::sub3: return 3;
4163 /// \brief Adjust the writemask of MIMG instructions
4164 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
4165 SelectionDAG &DAG) const {
4166 SDNode *Users[4] = { };
4168 unsigned DmaskIdx = (Node->getNumOperands() - Node->getNumValues() == 9) ? 2 : 3;
4169 unsigned OldDmask = Node->getConstantOperandVal(DmaskIdx);
4170 unsigned NewDmask = 0;
4172 // Try to figure out the used register components
4173 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
4176 // Abort if we can't understand the usage
4177 if (!I->isMachineOpcode() ||
4178 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
4181 // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
4182 // Note that subregs are packed, i.e. Lane==0 is the first bit set
4183 // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
4185 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
4187 // Set which texture component corresponds to the lane.
4189 for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
4191 Comp = countTrailingZeros(Dmask);
4192 Dmask &= ~(1 << Comp);
4195 // Abort if we have more than one user per component
4200 NewDmask |= 1 << Comp;
4203 // Abort if there's no change
4204 if (NewDmask == OldDmask)
4207 // Adjust the writemask in the node
4208 std::vector<SDValue> Ops;
4209 Ops.insert(Ops.end(), Node->op_begin(), Node->op_begin() + DmaskIdx);
4210 Ops.push_back(DAG.getTargetConstant(NewDmask, SDLoc(Node), MVT::i32));
4211 Ops.insert(Ops.end(), Node->op_begin() + DmaskIdx + 1, Node->op_end());
4212 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);
4214 // If we only got one lane, replace it with a copy
4215 // (if NewDmask has only one bit set...)
4216 if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
4217 SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, SDLoc(),
4219 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
4220 SDLoc(), Users[Lane]->getValueType(0),
4221 SDValue(Node, 0), RC);
4222 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
4226 // Update the users of the node with the new indices
4227 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
4229 SDNode *User = Users[i];
4233 SDValue Op = DAG.getTargetConstant(Idx, SDLoc(User), MVT::i32);
4234 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
4238 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
4239 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
4240 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
4245 static bool isFrameIndexOp(SDValue Op) {
4246 if (Op.getOpcode() == ISD::AssertZext)
4247 Op = Op.getOperand(0);
4249 return isa<FrameIndexSDNode>(Op);
4252 /// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
4253 /// with frame index operands.
4254 /// LLVM assumes that inputs are to these instructions are registers.
4255 void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
4256 SelectionDAG &DAG) const {
4258 SmallVector<SDValue, 8> Ops;
4259 for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
4260 if (!isFrameIndexOp(Node->getOperand(i))) {
4261 Ops.push_back(Node->getOperand(i));
4266 Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL,
4267 Node->getOperand(i).getValueType(),
4268 Node->getOperand(i)), 0));
4271 DAG.UpdateNodeOperands(Node, Ops);
4274 /// \brief Fold the instructions after selecting them.
4275 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
4276 SelectionDAG &DAG) const {
4277 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
4278 unsigned Opcode = Node->getMachineOpcode();
4280 if (TII->isMIMG(Opcode) && !TII->get(Opcode).mayStore() &&
4281 !TII->isGather4(Opcode))
4282 adjustWritemask(Node, DAG);
4284 if (Opcode == AMDGPU::INSERT_SUBREG ||
4285 Opcode == AMDGPU::REG_SEQUENCE) {
4286 legalizeTargetIndependentNode(Node, DAG);
4292 /// \brief Assign the register class depending on the number of
4293 /// bits set in the writemask
4294 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
4295 SDNode *Node) const {
4296 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
4298 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
4300 if (TII->isVOP3(MI.getOpcode())) {
4301 // Make sure constant bus requirements are respected.
4302 TII->legalizeOperandsVOP3(MRI, MI);
4306 if (TII->isMIMG(MI)) {
4307 unsigned VReg = MI.getOperand(0).getReg();
4308 const TargetRegisterClass *RC = MRI.getRegClass(VReg);
4309 // TODO: Need mapping tables to handle other cases (register classes).
4310 if (RC != &AMDGPU::VReg_128RegClass)
4313 unsigned DmaskIdx = MI.getNumOperands() == 12 ? 3 : 4;
4314 unsigned Writemask = MI.getOperand(DmaskIdx).getImm();
4315 unsigned BitsSet = 0;
4316 for (unsigned i = 0; i < 4; ++i)
4317 BitsSet += Writemask & (1 << i) ? 1 : 0;
4320 case 1: RC = &AMDGPU::VGPR_32RegClass; break;
4321 case 2: RC = &AMDGPU::VReg_64RegClass; break;
4322 case 3: RC = &AMDGPU::VReg_96RegClass; break;
4325 unsigned NewOpcode = TII->getMaskedMIMGOp(MI.getOpcode(), BitsSet);
4326 MI.setDesc(TII->get(NewOpcode));
4327 MRI.setRegClass(VReg, RC);
4331 // Replace unused atomics with the no return version.
4332 int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI.getOpcode());
4333 if (NoRetAtomicOp != -1) {
4334 if (!Node->hasAnyUseOfValue(0)) {
4335 MI.setDesc(TII->get(NoRetAtomicOp));
4336 MI.RemoveOperand(0);
4340 // For mubuf_atomic_cmpswap, we need to have tablegen use an extract_subreg
4341 // instruction, because the return type of these instructions is a vec2 of
4342 // the memory type, so it can be tied to the input operand.
4343 // This means these instructions always have a use, so we need to add a
4344 // special case to check if the atomic has only one extract_subreg use,
4345 // which itself has no uses.
4346 if ((Node->hasNUsesOfValue(1, 0) &&
4347 Node->use_begin()->isMachineOpcode() &&
4348 Node->use_begin()->getMachineOpcode() == AMDGPU::EXTRACT_SUBREG &&
4349 !Node->use_begin()->hasAnyUseOfValue(0))) {
4350 unsigned Def = MI.getOperand(0).getReg();
4352 // Change this into a noret atomic.
4353 MI.setDesc(TII->get(NoRetAtomicOp));
4354 MI.RemoveOperand(0);
4356 // If we only remove the def operand from the atomic instruction, the
4357 // extract_subreg will be left with a use of a vreg without a def.
4358 // So we need to insert an implicit_def to avoid machine verifier
4360 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(),
4361 TII->get(AMDGPU::IMPLICIT_DEF), Def);
4367 static SDValue buildSMovImm32(SelectionDAG &DAG, const SDLoc &DL,
4369 SDValue K = DAG.getTargetConstant(Val, DL, MVT::i32);
4370 return SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, K), 0);
4373 MachineSDNode *SITargetLowering::wrapAddr64Rsrc(SelectionDAG &DAG,
4375 SDValue Ptr) const {
4376 const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
4378 // Build the half of the subregister with the constants before building the
4379 // full 128-bit register. If we are building multiple resource descriptors,
4380 // this will allow CSEing of the 2-component register.
4381 const SDValue Ops0[] = {
4382 DAG.getTargetConstant(AMDGPU::SGPR_64RegClassID, DL, MVT::i32),
4383 buildSMovImm32(DAG, DL, 0),
4384 DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
4385 buildSMovImm32(DAG, DL, TII->getDefaultRsrcDataFormat() >> 32),
4386 DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
4389 SDValue SubRegHi = SDValue(DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL,
4390 MVT::v2i32, Ops0), 0);
4392 // Combine the constants and the pointer.
4393 const SDValue Ops1[] = {
4394 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
4396 DAG.getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32),
4398 DAG.getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32)
4401 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops1);
4404 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
4405 /// The TID (Thread ID) is multiplied by the stride value (bits [61:48]
4406 /// of the resource descriptor) to create an offset, which is added to
4407 /// the resource pointer.
4408 MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG, const SDLoc &DL,
4409 SDValue Ptr, uint32_t RsrcDword1,
4410 uint64_t RsrcDword2And3) const {
4411 SDValue PtrLo = DAG.getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
4412 SDValue PtrHi = DAG.getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
4414 PtrHi = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
4415 DAG.getConstant(RsrcDword1, DL, MVT::i32)),
4419 SDValue DataLo = buildSMovImm32(DAG, DL,
4420 RsrcDword2And3 & UINT64_C(0xFFFFFFFF));
4421 SDValue DataHi = buildSMovImm32(DAG, DL, RsrcDword2And3 >> 32);
4423 const SDValue Ops[] = {
4424 DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
4426 DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
4428 DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
4430 DAG.getTargetConstant(AMDGPU::sub2, DL, MVT::i32),
4432 DAG.getTargetConstant(AMDGPU::sub3, DL, MVT::i32)
4435 return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
4438 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
4439 const TargetRegisterClass *RC,
4440 unsigned Reg, EVT VT) const {
4441 SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);
4443 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
4444 cast<RegisterSDNode>(VReg)->getReg(), VT);
4447 //===----------------------------------------------------------------------===//
4448 // SI Inline Assembly Support
4449 //===----------------------------------------------------------------------===//
4451 std::pair<unsigned, const TargetRegisterClass *>
4452 SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
4453 StringRef Constraint,
4455 if (!isTypeLegal(VT))
4456 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4458 if (Constraint.size() == 1) {
4459 switch (Constraint[0]) {
4462 switch (VT.getSizeInBits()) {
4464 return std::make_pair(0U, nullptr);
4467 return std::make_pair(0U, &AMDGPU::SReg_32_XM0RegClass);
4469 return std::make_pair(0U, &AMDGPU::SGPR_64RegClass);
4471 return std::make_pair(0U, &AMDGPU::SReg_128RegClass);
4473 return std::make_pair(0U, &AMDGPU::SReg_256RegClass);
4477 switch (VT.getSizeInBits()) {
4479 return std::make_pair(0U, nullptr);
4482 return std::make_pair(0U, &AMDGPU::VGPR_32RegClass);
4484 return std::make_pair(0U, &AMDGPU::VReg_64RegClass);
4486 return std::make_pair(0U, &AMDGPU::VReg_96RegClass);
4488 return std::make_pair(0U, &AMDGPU::VReg_128RegClass);
4490 return std::make_pair(0U, &AMDGPU::VReg_256RegClass);
4492 return std::make_pair(0U, &AMDGPU::VReg_512RegClass);
4497 if (Constraint.size() > 1) {
4498 const TargetRegisterClass *RC = nullptr;
4499 if (Constraint[1] == 'v') {
4500 RC = &AMDGPU::VGPR_32RegClass;
4501 } else if (Constraint[1] == 's') {
4502 RC = &AMDGPU::SGPR_32RegClass;
4507 bool Failed = Constraint.substr(2).getAsInteger(10, Idx);
4508 if (!Failed && Idx < RC->getNumRegs())
4509 return std::make_pair(RC->getRegister(Idx), RC);
4512 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4515 SITargetLowering::ConstraintType
4516 SITargetLowering::getConstraintType(StringRef Constraint) const {
4517 if (Constraint.size() == 1) {
4518 switch (Constraint[0]) {
4522 return C_RegisterClass;
4525 return TargetLowering::getConstraintType(Constraint);