1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
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 This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUDiagnosticInfoUnsupported.h"
19 #include "AMDGPUFrameLowering.h"
20 #include "AMDGPUIntrinsicInfo.h"
21 #include "AMDGPURegisterInfo.h"
22 #include "AMDGPUSubtarget.h"
23 #include "R600MachineFunctionInfo.h"
24 #include "SIMachineFunctionInfo.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
34 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
35 CCValAssign::LocInfo LocInfo,
36 ISD::ArgFlagsTy ArgFlags, CCState &State) {
37 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
38 ArgFlags.getOrigAlign());
39 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
44 #include "AMDGPUGenCallingConv.inc"
46 // Find a larger type to do a load / store of a vector with.
47 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
48 unsigned StoreSize = VT.getStoreSizeInBits();
50 return EVT::getIntegerVT(Ctx, StoreSize);
52 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
53 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
56 // Type for a vector that will be loaded to.
57 EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
58 unsigned StoreSize = VT.getStoreSizeInBits();
60 return EVT::getIntegerVT(Ctx, 32);
62 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
65 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM,
66 const AMDGPUSubtarget &STI)
67 : TargetLowering(TM), Subtarget(&STI) {
68 setOperationAction(ISD::Constant, MVT::i32, Legal);
69 setOperationAction(ISD::Constant, MVT::i64, Legal);
70 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
71 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
73 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
74 setOperationAction(ISD::BRIND, MVT::Other, Expand);
76 // This is totally unsupported, just custom lower to produce an error.
77 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
79 // We need to custom lower some of the intrinsics
80 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
82 // Library functions. These default to Expand, but we have instructions
84 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
85 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
86 setOperationAction(ISD::FPOW, MVT::f32, Legal);
87 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
88 setOperationAction(ISD::FABS, MVT::f32, Legal);
89 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
90 setOperationAction(ISD::FRINT, MVT::f32, Legal);
91 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
92 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
93 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
95 setOperationAction(ISD::FROUND, MVT::f32, Custom);
96 setOperationAction(ISD::FROUND, MVT::f64, Custom);
98 setOperationAction(ISD::FREM, MVT::f32, Custom);
99 setOperationAction(ISD::FREM, MVT::f64, Custom);
101 // v_mad_f32 does not support denormals according to some sources.
102 if (!Subtarget->hasFP32Denormals())
103 setOperationAction(ISD::FMAD, MVT::f32, Legal);
105 // Expand to fneg + fadd.
106 setOperationAction(ISD::FSUB, MVT::f64, Expand);
108 // Lower floating point store/load to integer store/load to reduce the number
109 // of patterns in tablegen.
110 setOperationAction(ISD::STORE, MVT::f32, Promote);
111 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
113 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
114 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
116 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
117 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
119 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
120 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
122 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
123 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
125 setOperationAction(ISD::STORE, MVT::f64, Promote);
126 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
128 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
129 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
131 // Custom lowering of vector stores is required for local address space
133 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
135 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
136 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
137 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
139 // XXX: This can be change to Custom, once ExpandVectorStores can
140 // handle 64-bit stores.
141 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
143 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
146 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
147 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
150 setOperationAction(ISD::LOAD, MVT::f32, Promote);
151 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
153 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
154 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
156 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
157 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
159 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
160 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
162 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
163 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
165 setOperationAction(ISD::LOAD, MVT::f64, Promote);
166 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
168 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
169 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
171 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
172 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
173 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
174 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
175 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
176 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
177 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
178 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
179 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
180 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
182 // There are no 64-bit extloads. These should be done as a 32-bit extload and
183 // an extension to 64-bit.
184 for (MVT VT : MVT::integer_valuetypes()) {
185 setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
186 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
187 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
190 for (MVT VT : MVT::integer_vector_valuetypes()) {
191 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
192 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
193 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
194 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
195 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
196 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
197 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
198 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
199 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
200 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
201 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
202 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
205 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
207 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
208 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
209 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
210 setOperationAction(ISD::FRINT, MVT::f64, Custom);
211 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
214 if (!Subtarget->hasBFI()) {
215 // fcopysign can be done in a single instruction with BFI.
216 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
217 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
220 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
222 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
223 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
224 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
225 setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
227 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
228 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
229 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
230 setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
232 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
233 setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
234 setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
235 setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
237 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
238 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
240 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
241 for (MVT VT : ScalarIntVTs) {
242 setOperationAction(ISD::SREM, VT, Expand);
243 setOperationAction(ISD::SDIV, VT, Expand);
245 // GPU does not have divrem function for signed or unsigned.
246 setOperationAction(ISD::SDIVREM, VT, Custom);
247 setOperationAction(ISD::UDIVREM, VT, Custom);
249 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
250 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
251 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
253 setOperationAction(ISD::BSWAP, VT, Expand);
254 setOperationAction(ISD::CTTZ, VT, Expand);
255 setOperationAction(ISD::CTLZ, VT, Expand);
258 if (!Subtarget->hasBCNT(32))
259 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
261 if (!Subtarget->hasBCNT(64))
262 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
264 // The hardware supports 32-bit ROTR, but not ROTL.
265 setOperationAction(ISD::ROTL, MVT::i32, Expand);
266 setOperationAction(ISD::ROTL, MVT::i64, Expand);
267 setOperationAction(ISD::ROTR, MVT::i64, Expand);
269 setOperationAction(ISD::MUL, MVT::i64, Expand);
270 setOperationAction(ISD::MULHU, MVT::i64, Expand);
271 setOperationAction(ISD::MULHS, MVT::i64, Expand);
272 setOperationAction(ISD::UDIV, MVT::i32, Expand);
273 setOperationAction(ISD::UREM, MVT::i32, Expand);
274 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
275 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
276 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
277 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
278 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
280 setOperationAction(ISD::SMIN, MVT::i32, Legal);
281 setOperationAction(ISD::UMIN, MVT::i32, Legal);
282 setOperationAction(ISD::SMAX, MVT::i32, Legal);
283 setOperationAction(ISD::UMAX, MVT::i32, Legal);
285 if (!Subtarget->hasFFBH())
286 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
288 if (!Subtarget->hasFFBL())
289 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
291 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
292 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
294 static const MVT::SimpleValueType VectorIntTypes[] = {
295 MVT::v2i32, MVT::v4i32
298 for (MVT VT : VectorIntTypes) {
299 // Expand the following operations for the current type by default.
300 setOperationAction(ISD::ADD, VT, Expand);
301 setOperationAction(ISD::AND, VT, Expand);
302 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
303 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
304 setOperationAction(ISD::MUL, VT, Expand);
305 setOperationAction(ISD::OR, VT, Expand);
306 setOperationAction(ISD::SHL, VT, Expand);
307 setOperationAction(ISD::SRA, VT, Expand);
308 setOperationAction(ISD::SRL, VT, Expand);
309 setOperationAction(ISD::ROTL, VT, Expand);
310 setOperationAction(ISD::ROTR, VT, Expand);
311 setOperationAction(ISD::SUB, VT, Expand);
312 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
313 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
314 setOperationAction(ISD::SDIV, VT, Expand);
315 setOperationAction(ISD::UDIV, VT, Expand);
316 setOperationAction(ISD::SREM, VT, Expand);
317 setOperationAction(ISD::UREM, VT, Expand);
318 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
319 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
320 setOperationAction(ISD::SDIVREM, VT, Custom);
321 setOperationAction(ISD::UDIVREM, VT, Expand);
322 setOperationAction(ISD::ADDC, VT, Expand);
323 setOperationAction(ISD::SUBC, VT, Expand);
324 setOperationAction(ISD::ADDE, VT, Expand);
325 setOperationAction(ISD::SUBE, VT, Expand);
326 setOperationAction(ISD::SELECT, VT, Expand);
327 setOperationAction(ISD::VSELECT, VT, Expand);
328 setOperationAction(ISD::SELECT_CC, VT, Expand);
329 setOperationAction(ISD::XOR, VT, Expand);
330 setOperationAction(ISD::BSWAP, VT, Expand);
331 setOperationAction(ISD::CTPOP, VT, Expand);
332 setOperationAction(ISD::CTTZ, VT, Expand);
333 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
334 setOperationAction(ISD::CTLZ, VT, Expand);
335 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
336 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
339 static const MVT::SimpleValueType FloatVectorTypes[] = {
340 MVT::v2f32, MVT::v4f32
343 for (MVT VT : FloatVectorTypes) {
344 setOperationAction(ISD::FABS, VT, Expand);
345 setOperationAction(ISD::FMINNUM, VT, Expand);
346 setOperationAction(ISD::FMAXNUM, VT, Expand);
347 setOperationAction(ISD::FADD, VT, Expand);
348 setOperationAction(ISD::FCEIL, VT, Expand);
349 setOperationAction(ISD::FCOS, VT, Expand);
350 setOperationAction(ISD::FDIV, VT, Expand);
351 setOperationAction(ISD::FEXP2, VT, Expand);
352 setOperationAction(ISD::FLOG2, VT, Expand);
353 setOperationAction(ISD::FREM, VT, Expand);
354 setOperationAction(ISD::FPOW, VT, Expand);
355 setOperationAction(ISD::FFLOOR, VT, Expand);
356 setOperationAction(ISD::FTRUNC, VT, Expand);
357 setOperationAction(ISD::FMUL, VT, Expand);
358 setOperationAction(ISD::FMA, VT, Expand);
359 setOperationAction(ISD::FRINT, VT, Expand);
360 setOperationAction(ISD::FNEARBYINT, VT, Expand);
361 setOperationAction(ISD::FSQRT, VT, Expand);
362 setOperationAction(ISD::FSIN, VT, Expand);
363 setOperationAction(ISD::FSUB, VT, Expand);
364 setOperationAction(ISD::FNEG, VT, Expand);
365 setOperationAction(ISD::SELECT, VT, Expand);
366 setOperationAction(ISD::VSELECT, VT, Expand);
367 setOperationAction(ISD::SELECT_CC, VT, Expand);
368 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
369 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
372 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
373 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
375 setTargetDAGCombine(ISD::SHL);
376 setTargetDAGCombine(ISD::MUL);
377 setTargetDAGCombine(ISD::SELECT);
378 setTargetDAGCombine(ISD::SELECT_CC);
379 setTargetDAGCombine(ISD::STORE);
381 setTargetDAGCombine(ISD::FADD);
382 setTargetDAGCombine(ISD::FSUB);
384 setBooleanContents(ZeroOrNegativeOneBooleanContent);
385 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
387 setSchedulingPreference(Sched::RegPressure);
388 setJumpIsExpensive(true);
390 // SI at least has hardware support for floating point exceptions, but no way
391 // of using or handling them is implemented. They are also optional in OpenCL
393 setHasFloatingPointExceptions(false);
395 setSelectIsExpensive(false);
396 PredictableSelectIsExpensive = false;
398 setFsqrtIsCheap(true);
400 // We want to find all load dependencies for long chains of stores to enable
401 // merging into very wide vectors. The problem is with vectors with > 4
402 // elements. MergeConsecutiveStores will attempt to merge these because x8/x16
403 // vectors are a legal type, even though we have to split the loads
404 // usually. When we can more precisely specify load legality per address
405 // space, we should be able to make FindBetterChain/MergeConsecutiveStores
406 // smarter so that they can figure out what to do in 2 iterations without all
407 // N > 4 stores on the same chain.
408 GatherAllAliasesMaxDepth = 16;
410 // FIXME: Need to really handle these.
411 MaxStoresPerMemcpy = 4096;
412 MaxStoresPerMemmove = 4096;
413 MaxStoresPerMemset = 4096;
416 //===----------------------------------------------------------------------===//
417 // Target Information
418 //===----------------------------------------------------------------------===//
420 MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
424 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
428 // The backend supports 32 and 64 bit floating point immediates.
429 // FIXME: Why are we reporting vectors of FP immediates as legal?
430 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
431 EVT ScalarVT = VT.getScalarType();
432 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
435 // We don't want to shrink f64 / f32 constants.
436 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
437 EVT ScalarVT = VT.getScalarType();
438 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
441 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
445 unsigned NewSize = NewVT.getStoreSizeInBits();
447 // If we are reducing to a 32-bit load, this is always better.
451 EVT OldVT = N->getValueType(0);
452 unsigned OldSize = OldVT.getStoreSizeInBits();
454 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
455 // extloads, so doing one requires using a buffer_load. In cases where we
456 // still couldn't use a scalar load, using the wider load shouldn't really
459 // If the old size already had to be an extload, there's no harm in continuing
460 // to reduce the width.
461 return (OldSize < 32);
464 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
466 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
469 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
470 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
472 return ((LScalarSize <= CastScalarSize) ||
473 (CastScalarSize >= 32) ||
477 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
478 // profitable with the expansion for 64-bit since it's generally good to
480 // FIXME: These should really have the size as a parameter.
481 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
485 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
489 //===---------------------------------------------------------------------===//
491 //===---------------------------------------------------------------------===//
493 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
494 assert(VT.isFloatingPoint());
495 return VT == MVT::f32 || VT == MVT::f64;
498 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
499 assert(VT.isFloatingPoint());
500 return VT == MVT::f32 || VT == MVT::f64;
503 bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
509 bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
510 // There are few operations which truly have vector input operands. Any vector
511 // operation is going to involve operations on each component, and a
512 // build_vector will be a copy per element, so it always makes sense to use a
513 // build_vector input in place of the extracted element to avoid a copy into a
516 // We should probably only do this if all users are extracts only, but this
517 // should be the common case.
521 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
522 // Truncate is just accessing a subregister.
523 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
526 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
527 // Truncate is just accessing a subregister.
528 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
529 (Dest->getPrimitiveSizeInBits() % 32 == 0);
532 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
533 unsigned SrcSize = Src->getScalarSizeInBits();
534 unsigned DestSize = Dest->getScalarSizeInBits();
536 return SrcSize == 32 && DestSize == 64;
539 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
540 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
541 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
542 // this will enable reducing 64-bit operations the 32-bit, which is always
544 return Src == MVT::i32 && Dest == MVT::i64;
547 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
548 return isZExtFree(Val.getValueType(), VT2);
551 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
552 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
553 // limited number of native 64-bit operations. Shrinking an operation to fit
554 // in a single 32-bit register should always be helpful. As currently used,
555 // this is much less general than the name suggests, and is only used in
556 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
557 // not profitable, and may actually be harmful.
558 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
561 //===---------------------------------------------------------------------===//
562 // TargetLowering Callbacks
563 //===---------------------------------------------------------------------===//
565 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
566 const SmallVectorImpl<ISD::InputArg> &Ins) const {
568 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
571 SDValue AMDGPUTargetLowering::LowerReturn(
573 CallingConv::ID CallConv,
575 const SmallVectorImpl<ISD::OutputArg> &Outs,
576 const SmallVectorImpl<SDValue> &OutVals,
577 SDLoc DL, SelectionDAG &DAG) const {
578 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
581 //===---------------------------------------------------------------------===//
582 // Target specific lowering
583 //===---------------------------------------------------------------------===//
585 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
586 SmallVectorImpl<SDValue> &InVals) const {
587 SDValue Callee = CLI.Callee;
588 SelectionDAG &DAG = CLI.DAG;
590 const Function &Fn = *DAG.getMachineFunction().getFunction();
592 StringRef FuncName("<unknown>");
594 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
595 FuncName = G->getSymbol();
596 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
597 FuncName = G->getGlobal()->getName();
599 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
600 DAG.getContext()->diagnose(NoCalls);
604 SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
605 SelectionDAG &DAG) const {
606 const Function &Fn = *DAG.getMachineFunction().getFunction();
608 DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "dynamic alloca");
609 DAG.getContext()->diagnose(NoDynamicAlloca);
613 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
614 SelectionDAG &DAG) const {
615 switch (Op.getOpcode()) {
617 Op.getNode()->dump();
618 llvm_unreachable("Custom lowering code for this"
619 "instruction is not implemented yet!");
621 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
622 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
623 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
624 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
625 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
626 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
627 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
628 case ISD::FREM: return LowerFREM(Op, DAG);
629 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
630 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
631 case ISD::FRINT: return LowerFRINT(Op, DAG);
632 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
633 case ISD::FROUND: return LowerFROUND(Op, DAG);
634 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
635 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
636 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
637 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
638 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
639 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
644 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
645 SmallVectorImpl<SDValue> &Results,
646 SelectionDAG &DAG) const {
647 switch (N->getOpcode()) {
648 case ISD::SIGN_EXTEND_INREG:
649 // Different parts of legalization seem to interpret which type of
650 // sign_extend_inreg is the one to check for custom lowering. The extended
651 // from type is what really matters, but some places check for custom
652 // lowering of the result type. This results in trying to use
653 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
654 // nothing here and let the illegal result integer be handled normally.
657 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
661 Results.push_back(SDValue(Node, 0));
662 Results.push_back(SDValue(Node, 1));
663 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
665 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
669 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
670 if (Lowered.getNode())
671 Results.push_back(Lowered);
679 // FIXME: This implements accesses to initialized globals in the constant
680 // address space by copying them to private and accessing that. It does not
681 // properly handle illegal types or vectors. The private vector loads are not
682 // scalarized, and the illegal scalars hit an assertion. This technique will not
683 // work well with large initializers, and this should eventually be
684 // removed. Initialized globals should be placed into a data section that the
685 // runtime will load into a buffer before the kernel is executed. Uses of the
686 // global need to be replaced with a pointer loaded from an implicit kernel
687 // argument into this buffer holding the copy of the data, which will remove the
688 // need for any of this.
689 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
690 const GlobalValue *GV,
691 const SDValue &InitPtr,
693 SelectionDAG &DAG) const {
694 const DataLayout &TD = DAG.getDataLayout();
696 Type *InitTy = Init->getType();
698 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
699 EVT VT = EVT::getEVT(InitTy);
700 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
701 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, DL, VT), InitPtr,
702 MachinePointerInfo(UndefValue::get(PtrTy)), false,
703 false, TD.getPrefTypeAlignment(InitTy));
706 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
707 EVT VT = EVT::getEVT(CFP->getType());
708 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
709 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, DL, VT), InitPtr,
710 MachinePointerInfo(UndefValue::get(PtrTy)), false,
711 false, TD.getPrefTypeAlignment(CFP->getType()));
714 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
715 const StructLayout *SL = TD.getStructLayout(ST);
717 EVT PtrVT = InitPtr.getValueType();
718 SmallVector<SDValue, 8> Chains;
720 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
721 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), DL, PtrVT);
722 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
724 Constant *Elt = Init->getAggregateElement(I);
725 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
728 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
731 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
732 EVT PtrVT = InitPtr.getValueType();
734 unsigned NumElements;
735 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
736 NumElements = AT->getNumElements();
737 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
738 NumElements = VT->getNumElements();
740 llvm_unreachable("Unexpected type");
742 unsigned EltSize = TD.getTypeAllocSize(SeqTy->getElementType());
743 SmallVector<SDValue, 8> Chains;
744 for (unsigned i = 0; i < NumElements; ++i) {
745 SDValue Offset = DAG.getConstant(i * EltSize, DL, PtrVT);
746 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
748 Constant *Elt = Init->getAggregateElement(i);
749 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
752 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
755 if (isa<UndefValue>(Init)) {
756 EVT VT = EVT::getEVT(InitTy);
757 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
758 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
759 MachinePointerInfo(UndefValue::get(PtrTy)), false,
760 false, TD.getPrefTypeAlignment(InitTy));
764 llvm_unreachable("Unhandled constant initializer");
767 static bool hasDefinedInitializer(const GlobalValue *GV) {
768 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
769 if (!GVar || !GVar->hasInitializer())
772 if (isa<UndefValue>(GVar->getInitializer()))
778 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
780 SelectionDAG &DAG) const {
782 const DataLayout &DL = DAG.getDataLayout();
783 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
784 const GlobalValue *GV = G->getGlobal();
786 switch (G->getAddressSpace()) {
787 case AMDGPUAS::LOCAL_ADDRESS: {
788 // XXX: What does the value of G->getOffset() mean?
789 assert(G->getOffset() == 0 &&
790 "Do not know what to do with an non-zero offset");
792 // TODO: We could emit code to handle the initialization somewhere.
793 if (hasDefinedInitializer(GV))
797 if (MFI->LocalMemoryObjects.count(GV) == 0) {
798 uint64_t Size = DL.getTypeAllocSize(GV->getType()->getElementType());
799 Offset = MFI->LDSSize;
800 MFI->LocalMemoryObjects[GV] = Offset;
801 // XXX: Account for alignment?
802 MFI->LDSSize += Size;
804 Offset = MFI->LocalMemoryObjects[GV];
807 return DAG.getConstant(Offset, SDLoc(Op),
808 getPointerTy(DL, AMDGPUAS::LOCAL_ADDRESS));
810 case AMDGPUAS::CONSTANT_ADDRESS: {
811 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
812 Type *EltType = GV->getType()->getElementType();
813 unsigned Size = DL.getTypeAllocSize(EltType);
814 unsigned Alignment = DL.getPrefTypeAlignment(EltType);
816 MVT PrivPtrVT = getPointerTy(DL, AMDGPUAS::PRIVATE_ADDRESS);
817 MVT ConstPtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
819 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
820 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
822 const GlobalVariable *Var = cast<GlobalVariable>(GV);
823 if (!Var->hasInitializer()) {
824 // This has no use, but bugpoint will hit it.
825 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
828 const Constant *Init = Var->getInitializer();
829 SmallVector<SDNode*, 8> WorkList;
831 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
832 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
833 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
835 WorkList.push_back(*I);
837 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
838 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
839 E = WorkList.end(); I != E; ++I) {
840 SmallVector<SDValue, 8> Ops;
841 Ops.push_back(Chain);
842 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
843 Ops.push_back((*I)->getOperand(i));
845 DAG.UpdateNodeOperands(*I, Ops);
847 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
851 const Function &Fn = *DAG.getMachineFunction().getFunction();
852 DiagnosticInfoUnsupported BadInit(Fn,
853 "initializer for address space");
854 DAG.getContext()->diagnose(BadInit);
858 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
859 SelectionDAG &DAG) const {
860 SmallVector<SDValue, 8> Args;
862 for (const SDUse &U : Op->ops())
863 DAG.ExtractVectorElements(U.get(), Args);
865 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
868 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
869 SelectionDAG &DAG) const {
871 SmallVector<SDValue, 8> Args;
872 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
873 EVT VT = Op.getValueType();
874 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
875 VT.getVectorNumElements());
877 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
880 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
881 SelectionDAG &DAG) const {
883 MachineFunction &MF = DAG.getMachineFunction();
884 const AMDGPUFrameLowering *TFL = Subtarget->getFrameLowering();
886 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
888 unsigned FrameIndex = FIN->getIndex();
889 unsigned IgnoredFrameReg;
891 TFL->getFrameIndexReference(MF, FrameIndex, IgnoredFrameReg);
892 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), SDLoc(Op),
896 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
897 SelectionDAG &DAG) const {
898 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
900 EVT VT = Op.getValueType();
902 switch (IntrinsicID) {
904 case AMDGPUIntrinsic::AMDGPU_abs:
905 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
906 return LowerIntrinsicIABS(Op, DAG);
907 case AMDGPUIntrinsic::AMDGPU_lrp:
908 return LowerIntrinsicLRP(Op, DAG);
910 case AMDGPUIntrinsic::AMDGPU_clamp:
911 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
912 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
913 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
915 case Intrinsic::AMDGPU_div_scale: {
916 // 3rd parameter required to be a constant.
917 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
919 return DAG.getUNDEF(VT);
921 // Translate to the operands expected by the machine instruction. The
922 // first parameter must be the same as the first instruction.
923 SDValue Numerator = Op.getOperand(1);
924 SDValue Denominator = Op.getOperand(2);
926 // Note this order is opposite of the machine instruction's operations,
927 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
928 // intrinsic has the numerator as the first operand to match a normal
929 // division operation.
931 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
933 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
934 Denominator, Numerator);
937 case Intrinsic::AMDGPU_div_fmas:
938 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
939 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3),
942 case Intrinsic::AMDGPU_div_fixup:
943 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
944 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
946 case Intrinsic::AMDGPU_trig_preop:
947 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
948 Op.getOperand(1), Op.getOperand(2));
950 case Intrinsic::AMDGPU_rcp:
951 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
953 case Intrinsic::AMDGPU_rsq:
954 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
956 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
957 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
959 case Intrinsic::AMDGPU_rsq_clamped:
960 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
961 Type *Type = VT.getTypeForEVT(*DAG.getContext());
962 APFloat Max = APFloat::getLargest(Type->getFltSemantics());
963 APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
965 SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
966 SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
967 DAG.getConstantFP(Max, DL, VT));
968 return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
969 DAG.getConstantFP(Min, DL, VT));
971 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
974 case Intrinsic::AMDGPU_ldexp:
975 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
978 case AMDGPUIntrinsic::AMDGPU_imax:
979 return DAG.getNode(ISD::SMAX, DL, VT, Op.getOperand(1),
981 case AMDGPUIntrinsic::AMDGPU_umax:
982 return DAG.getNode(ISD::UMAX, DL, VT, Op.getOperand(1),
984 case AMDGPUIntrinsic::AMDGPU_imin:
985 return DAG.getNode(ISD::SMIN, DL, VT, Op.getOperand(1),
987 case AMDGPUIntrinsic::AMDGPU_umin:
988 return DAG.getNode(ISD::UMIN, DL, VT, Op.getOperand(1),
991 case AMDGPUIntrinsic::AMDGPU_umul24:
992 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
993 Op.getOperand(1), Op.getOperand(2));
995 case AMDGPUIntrinsic::AMDGPU_imul24:
996 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
997 Op.getOperand(1), Op.getOperand(2));
999 case AMDGPUIntrinsic::AMDGPU_umad24:
1000 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
1001 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
1003 case AMDGPUIntrinsic::AMDGPU_imad24:
1004 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
1005 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
1007 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
1008 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
1010 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
1011 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
1013 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
1014 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
1016 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
1017 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
1019 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
1020 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
1025 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
1026 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
1031 case AMDGPUIntrinsic::AMDGPU_bfi:
1032 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
1037 case AMDGPUIntrinsic::AMDGPU_bfm:
1038 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
1042 case Intrinsic::AMDGPU_class:
1043 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
1044 Op.getOperand(1), Op.getOperand(2));
1046 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
1047 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
1049 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
1050 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
1051 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
1052 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
1053 case AMDGPUIntrinsic::AMDGPU_brev: // Legacy name
1054 return DAG.getNode(ISD::BITREVERSE, DL, VT, Op.getOperand(1));
1058 ///IABS(a) = SMAX(sub(0, a), a)
1059 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
1060 SelectionDAG &DAG) const {
1062 EVT VT = Op.getValueType();
1063 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1066 return DAG.getNode(ISD::SMAX, DL, VT, Neg, Op.getOperand(1));
1069 /// Linear Interpolation
1070 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
1071 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
1072 SelectionDAG &DAG) const {
1074 EVT VT = Op.getValueType();
1075 // TODO: Should this propagate fast-math-flags?
1076 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
1077 DAG.getConstantFP(1.0f, DL, MVT::f32),
1079 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1081 return DAG.getNode(ISD::FADD, DL, VT,
1082 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1086 /// \brief Generate Min/Max node
1087 SDValue AMDGPUTargetLowering::CombineFMinMaxLegacy(SDLoc DL,
1094 DAGCombinerInfo &DCI) const {
1095 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1098 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1101 SelectionDAG &DAG = DCI.DAG;
1102 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1111 case ISD::SETFALSE2:
1120 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1121 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1127 // Ordered. Assume ordered for undefined.
1129 // Only do this after legalization to avoid interfering with other combines
1130 // which might occur.
1131 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1132 !DCI.isCalledByLegalizer())
1135 // We need to permute the operands to get the correct NaN behavior. The
1136 // selected operand is the second one based on the failing compare with NaN,
1137 // so permute it based on the compare type the hardware uses.
1139 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1140 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1145 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1146 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1152 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1153 !DCI.isCalledByLegalizer())
1157 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1158 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1160 case ISD::SETCC_INVALID:
1161 llvm_unreachable("Invalid setcc condcode!");
1166 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1167 SelectionDAG &DAG) const {
1168 LoadSDNode *Load = cast<LoadSDNode>(Op);
1169 EVT MemVT = Load->getMemoryVT();
1170 EVT MemEltVT = MemVT.getVectorElementType();
1172 EVT LoadVT = Op.getValueType();
1173 EVT EltVT = LoadVT.getVectorElementType();
1174 EVT PtrVT = Load->getBasePtr().getValueType();
1176 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1177 SmallVector<SDValue, 8> Loads;
1178 SmallVector<SDValue, 8> Chains;
1181 unsigned MemEltSize = MemEltVT.getStoreSize();
1182 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1184 for (unsigned i = 0; i < NumElts; ++i) {
1185 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1186 DAG.getConstant(i * MemEltSize, SL, PtrVT));
1189 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1190 Load->getChain(), Ptr,
1191 SrcValue.getWithOffset(i * MemEltSize),
1192 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1193 Load->isInvariant(), Load->getAlignment());
1194 Loads.push_back(NewLoad.getValue(0));
1195 Chains.push_back(NewLoad.getValue(1));
1199 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1200 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1203 return DAG.getMergeValues(Ops, SL);
1206 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1207 SelectionDAG &DAG) const {
1208 EVT VT = Op.getValueType();
1210 // If this is a 2 element vector, we really want to scalarize and not create
1211 // weird 1 element vectors.
1212 if (VT.getVectorNumElements() == 2)
1213 return ScalarizeVectorLoad(Op, DAG);
1215 LoadSDNode *Load = cast<LoadSDNode>(Op);
1216 SDValue BasePtr = Load->getBasePtr();
1217 EVT PtrVT = BasePtr.getValueType();
1218 EVT MemVT = Load->getMemoryVT();
1221 const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
1224 EVT LoMemVT, HiMemVT;
1227 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1228 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1229 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1231 unsigned Size = LoMemVT.getStoreSize();
1232 unsigned BaseAlign = Load->getAlignment();
1233 unsigned HiAlign = MinAlign(BaseAlign, Size);
1236 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1237 Load->getChain(), BasePtr,
1239 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1240 Load->isInvariant(), BaseAlign);
1242 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1243 DAG.getConstant(Size, SL, PtrVT));
1246 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1247 Load->getChain(), HiPtr,
1248 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1249 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1250 Load->isInvariant(), HiAlign);
1253 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1254 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1255 LoLoad.getValue(1), HiLoad.getValue(1))
1258 return DAG.getMergeValues(Ops, SL);
1261 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1262 SelectionDAG &DAG) const {
1263 StoreSDNode *Store = cast<StoreSDNode>(Op);
1264 EVT MemVT = Store->getMemoryVT();
1265 unsigned MemBits = MemVT.getSizeInBits();
1267 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1268 // truncating store into an i32 store.
1269 // XXX: We could also handle optimize other vector bitwidths.
1270 if (!MemVT.isVector() || MemBits > 32) {
1275 SDValue Value = Store->getValue();
1276 EVT VT = Value.getValueType();
1277 EVT ElemVT = VT.getVectorElementType();
1278 SDValue Ptr = Store->getBasePtr();
1279 EVT MemEltVT = MemVT.getVectorElementType();
1280 unsigned MemEltBits = MemEltVT.getSizeInBits();
1281 unsigned MemNumElements = MemVT.getVectorNumElements();
1282 unsigned PackedSize = MemVT.getStoreSizeInBits();
1283 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, DL, MVT::i32);
1285 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1287 SDValue PackedValue;
1288 for (unsigned i = 0; i < MemNumElements; ++i) {
1289 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1290 DAG.getConstant(i, DL, MVT::i32));
1291 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1292 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1294 SDValue Shift = DAG.getConstant(MemEltBits * i, DL, MVT::i32);
1295 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1300 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1304 if (PackedSize < 32) {
1305 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1306 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1307 Store->getMemOperand()->getPointerInfo(),
1309 Store->isNonTemporal(), Store->isVolatile(),
1310 Store->getAlignment());
1313 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1314 Store->getMemOperand()->getPointerInfo(),
1315 Store->isVolatile(), Store->isNonTemporal(),
1316 Store->getAlignment());
1319 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1320 SelectionDAG &DAG) const {
1321 StoreSDNode *Store = cast<StoreSDNode>(Op);
1322 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1323 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1324 EVT PtrVT = Store->getBasePtr().getValueType();
1325 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1328 SmallVector<SDValue, 8> Chains;
1330 unsigned EltSize = MemEltVT.getStoreSize();
1331 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1333 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1334 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1336 DAG.getConstant(i, SL, MVT::i32));
1338 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), SL, PtrVT);
1339 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1341 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1342 SrcValue.getWithOffset(i * EltSize),
1343 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1344 Store->getAlignment());
1345 Chains.push_back(NewStore);
1348 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1351 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1352 SelectionDAG &DAG) const {
1353 StoreSDNode *Store = cast<StoreSDNode>(Op);
1354 SDValue Val = Store->getValue();
1355 EVT VT = Val.getValueType();
1357 // If this is a 2 element vector, we really want to scalarize and not create
1358 // weird 1 element vectors.
1359 if (VT.getVectorNumElements() == 2)
1360 return ScalarizeVectorStore(Op, DAG);
1362 EVT MemVT = Store->getMemoryVT();
1363 SDValue Chain = Store->getChain();
1364 SDValue BasePtr = Store->getBasePtr();
1368 EVT LoMemVT, HiMemVT;
1371 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1372 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1373 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1375 EVT PtrVT = BasePtr.getValueType();
1376 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1377 DAG.getConstant(LoMemVT.getStoreSize(), SL,
1380 const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
1381 unsigned BaseAlign = Store->getAlignment();
1382 unsigned Size = LoMemVT.getStoreSize();
1383 unsigned HiAlign = MinAlign(BaseAlign, Size);
1386 = DAG.getTruncStore(Chain, SL, Lo,
1390 Store->isNonTemporal(),
1391 Store->isVolatile(),
1394 = DAG.getTruncStore(Chain, SL, Hi,
1396 SrcValue.getWithOffset(Size),
1398 Store->isNonTemporal(),
1399 Store->isVolatile(),
1402 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1406 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1408 LoadSDNode *Load = cast<LoadSDNode>(Op);
1409 ISD::LoadExtType ExtType = Load->getExtensionType();
1410 EVT VT = Op.getValueType();
1411 EVT MemVT = Load->getMemoryVT();
1413 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1414 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1415 // FIXME: Copied from PPC
1416 // First, load into 32 bits, then truncate to 1 bit.
1418 SDValue Chain = Load->getChain();
1419 SDValue BasePtr = Load->getBasePtr();
1420 MachineMemOperand *MMO = Load->getMemOperand();
1422 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1423 BasePtr, MVT::i8, MMO);
1426 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1430 return DAG.getMergeValues(Ops, DL);
1433 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1434 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1435 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1438 // <SI && AS=PRIVATE && EXTLOAD && size < 32bit,
1439 // register (2-)byte extract.
1441 // Get Register holding the target.
1442 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1443 DAG.getConstant(2, DL, MVT::i32));
1444 // Load the Register.
1445 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1446 Load->getChain(), Ptr,
1447 DAG.getTargetConstant(0, DL, MVT::i32),
1450 // Get offset within the register.
1451 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1453 DAG.getConstant(0x3, DL, MVT::i32));
1455 // Bit offset of target byte (byteIdx * 8).
1456 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1457 DAG.getConstant(3, DL, MVT::i32));
1459 // Shift to the right.
1460 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1462 // Eliminate the upper bits by setting them to ...
1463 EVT MemEltVT = MemVT.getScalarType();
1466 if (ExtType == ISD::SEXTLOAD) {
1467 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1470 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1474 return DAG.getMergeValues(Ops, DL);
1479 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1483 return DAG.getMergeValues(Ops, DL);
1486 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1488 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1489 if (Result.getNode()) {
1493 StoreSDNode *Store = cast<StoreSDNode>(Op);
1494 SDValue Chain = Store->getChain();
1495 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1496 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1497 Store->getValue().getValueType().isVector()) {
1498 return SplitVectorStore(Op, DAG);
1501 EVT MemVT = Store->getMemoryVT();
1502 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1503 MemVT.bitsLT(MVT::i32)) {
1505 if (Store->getMemoryVT() == MVT::i8) {
1507 } else if (Store->getMemoryVT() == MVT::i16) {
1510 SDValue BasePtr = Store->getBasePtr();
1511 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1512 DAG.getConstant(2, DL, MVT::i32));
1513 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1515 DAG.getTargetConstant(0, DL, MVT::i32));
1517 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1518 DAG.getConstant(0x3, DL, MVT::i32));
1520 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1521 DAG.getConstant(3, DL, MVT::i32));
1523 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1526 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1528 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1529 MaskedValue, ShiftAmt);
1531 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32,
1532 DAG.getConstant(Mask, DL, MVT::i32),
1534 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1535 DAG.getConstant(0xffffffff, DL, MVT::i32));
1536 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1538 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1539 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1541 DAG.getTargetConstant(0, DL, MVT::i32));
1546 // This is a shortcut for integer division because we have fast i32<->f32
1547 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1548 // float is enough to accurately represent up to a 24-bit integer.
1549 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1551 EVT VT = Op.getValueType();
1552 SDValue LHS = Op.getOperand(0);
1553 SDValue RHS = Op.getOperand(1);
1554 MVT IntVT = MVT::i32;
1555 MVT FltVT = MVT::f32;
1557 ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1558 ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1560 if (VT.isVector()) {
1561 unsigned NElts = VT.getVectorNumElements();
1562 IntVT = MVT::getVectorVT(MVT::i32, NElts);
1563 FltVT = MVT::getVectorVT(MVT::f32, NElts);
1566 unsigned BitSize = VT.getScalarType().getSizeInBits();
1568 SDValue jq = DAG.getConstant(1, DL, IntVT);
1571 // char|short jq = ia ^ ib;
1572 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1574 // jq = jq >> (bitsize - 2)
1575 jq = DAG.getNode(ISD::SRA, DL, VT, jq,
1576 DAG.getConstant(BitSize - 2, DL, VT));
1579 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
1582 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1585 // int ia = (int)LHS;
1587 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1589 // int ib, (int)RHS;
1591 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1593 // float fa = (float)ia;
1594 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1596 // float fb = (float)ib;
1597 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1599 // TODO: Should this propagate fast-math-flags?
1600 // float fq = native_divide(fa, fb);
1601 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1602 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1605 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1607 // float fqneg = -fq;
1608 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1610 // float fr = mad(fqneg, fb, fa);
1611 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1612 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1614 // int iq = (int)fq;
1615 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1618 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1621 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1623 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1625 // int cv = fr >= fb;
1626 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1628 // jq = (cv ? jq : 0);
1629 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
1631 // dst = trunc/extend to legal type
1632 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1635 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1637 // Rem needs compensation, it's easier to recompute it
1638 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1639 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1645 return DAG.getMergeValues(Res, DL);
1648 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1650 SmallVectorImpl<SDValue> &Results) const {
1651 assert(Op.getValueType() == MVT::i64);
1654 EVT VT = Op.getValueType();
1655 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1657 SDValue one = DAG.getConstant(1, DL, HalfVT);
1658 SDValue zero = DAG.getConstant(0, DL, HalfVT);
1661 SDValue LHS = Op.getOperand(0);
1662 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1663 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1665 SDValue RHS = Op.getOperand(1);
1666 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1667 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1669 if (VT == MVT::i64 &&
1670 DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
1671 DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
1673 SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1676 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(0), zero);
1677 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(1), zero);
1678 Results.push_back(DIV);
1679 Results.push_back(REM);
1683 // Get Speculative values
1684 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1685 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1687 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1688 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, zero);
1690 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1691 SDValue DIV_Lo = zero;
1693 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1695 for (unsigned i = 0; i < halfBitWidth; ++i) {
1696 const unsigned bitPos = halfBitWidth - i - 1;
1697 SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
1698 // Get value of high bit
1699 SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1700 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1701 HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
1704 REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
1706 REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
1708 SDValue BIT = DAG.getConstant(1 << bitPos, DL, HalfVT);
1709 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1711 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1714 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1715 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1718 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1719 Results.push_back(DIV);
1720 Results.push_back(REM);
1723 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1724 SelectionDAG &DAG) const {
1726 EVT VT = Op.getValueType();
1728 if (VT == MVT::i64) {
1729 SmallVector<SDValue, 2> Results;
1730 LowerUDIVREM64(Op, DAG, Results);
1731 return DAG.getMergeValues(Results, DL);
1734 SDValue Num = Op.getOperand(0);
1735 SDValue Den = Op.getOperand(1);
1737 if (VT == MVT::i32) {
1738 if (DAG.MaskedValueIsZero(Num, APInt::getHighBitsSet(32, 8)) &&
1739 DAG.MaskedValueIsZero(Den, APInt::getHighBitsSet(32, 8))) {
1740 // TODO: We technically could do this for i64, but shouldn't that just be
1741 // handled by something generally reducing 64-bit division on 32-bit
1742 // values to 32-bit?
1743 return LowerDIVREM24(Op, DAG, false);
1747 // RCP = URECIP(Den) = 2^32 / Den + e
1748 // e is rounding error.
1749 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1751 // RCP_LO = mul(RCP, Den) */
1752 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1754 // RCP_HI = mulhu (RCP, Den) */
1755 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1757 // NEG_RCP_LO = -RCP_LO
1758 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1761 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1762 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1765 // Calculate the rounding error from the URECIP instruction
1766 // E = mulhu(ABS_RCP_LO, RCP)
1767 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1769 // RCP_A_E = RCP + E
1770 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1772 // RCP_S_E = RCP - E
1773 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1775 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1776 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1779 // Quotient = mulhu(Tmp0, Num)
1780 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1782 // Num_S_Remainder = Quotient * Den
1783 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1785 // Remainder = Num - Num_S_Remainder
1786 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1788 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1789 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1790 DAG.getConstant(-1, DL, VT),
1791 DAG.getConstant(0, DL, VT),
1793 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1794 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1796 DAG.getConstant(-1, DL, VT),
1797 DAG.getConstant(0, DL, VT),
1799 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1800 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1803 // Calculate Division result:
1805 // Quotient_A_One = Quotient + 1
1806 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1807 DAG.getConstant(1, DL, VT));
1809 // Quotient_S_One = Quotient - 1
1810 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1811 DAG.getConstant(1, DL, VT));
1813 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1814 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1815 Quotient, Quotient_A_One, ISD::SETEQ);
1817 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1818 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1819 Quotient_S_One, Div, ISD::SETEQ);
1821 // Calculate Rem result:
1823 // Remainder_S_Den = Remainder - Den
1824 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1826 // Remainder_A_Den = Remainder + Den
1827 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1829 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1830 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1831 Remainder, Remainder_S_Den, ISD::SETEQ);
1833 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1834 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1835 Remainder_A_Den, Rem, ISD::SETEQ);
1840 return DAG.getMergeValues(Ops, DL);
1843 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1844 SelectionDAG &DAG) const {
1846 EVT VT = Op.getValueType();
1848 SDValue LHS = Op.getOperand(0);
1849 SDValue RHS = Op.getOperand(1);
1851 SDValue Zero = DAG.getConstant(0, DL, VT);
1852 SDValue NegOne = DAG.getConstant(-1, DL, VT);
1854 if (VT == MVT::i32 &&
1855 DAG.ComputeNumSignBits(LHS) > 8 &&
1856 DAG.ComputeNumSignBits(RHS) > 8) {
1857 return LowerDIVREM24(Op, DAG, true);
1859 if (VT == MVT::i64 &&
1860 DAG.ComputeNumSignBits(LHS) > 32 &&
1861 DAG.ComputeNumSignBits(RHS) > 32) {
1862 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1865 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1866 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1867 SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1870 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
1871 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
1873 return DAG.getMergeValues(Res, DL);
1876 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1877 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1878 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1879 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1881 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1882 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1884 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1885 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1887 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1888 SDValue Rem = Div.getValue(1);
1890 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1891 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1893 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1894 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1900 return DAG.getMergeValues(Res, DL);
1903 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1904 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1906 EVT VT = Op.getValueType();
1907 SDValue X = Op.getOperand(0);
1908 SDValue Y = Op.getOperand(1);
1910 // TODO: Should this propagate fast-math-flags?
1912 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1913 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1914 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1916 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1919 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1921 SDValue Src = Op.getOperand(0);
1923 // result = trunc(src)
1924 // if (src > 0.0 && src != result)
1927 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1929 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
1930 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
1933 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
1935 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1936 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1937 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1939 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1940 // TODO: Should this propagate fast-math-flags?
1941 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1944 static SDValue extractF64Exponent(SDValue Hi, SDLoc SL, SelectionDAG &DAG) {
1945 const unsigned FractBits = 52;
1946 const unsigned ExpBits = 11;
1948 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1950 DAG.getConstant(FractBits - 32, SL, MVT::i32),
1951 DAG.getConstant(ExpBits, SL, MVT::i32));
1952 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1953 DAG.getConstant(1023, SL, MVT::i32));
1958 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1960 SDValue Src = Op.getOperand(0);
1962 assert(Op.getValueType() == MVT::f64);
1964 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1965 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1967 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1969 // Extract the upper half, since this is where we will find the sign and
1971 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1973 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
1975 const unsigned FractBits = 52;
1977 // Extract the sign bit.
1978 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
1979 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1981 // Extend back to to 64-bits.
1982 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1984 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1986 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1987 const SDValue FractMask
1988 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
1990 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1991 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1992 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1995 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
1997 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
1999 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2000 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2002 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
2003 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
2005 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
2008 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
2010 SDValue Src = Op.getOperand(0);
2012 assert(Op.getValueType() == MVT::f64);
2014 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
2015 SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
2016 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
2018 // TODO: Should this propagate fast-math-flags?
2020 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
2021 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
2023 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
2025 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
2026 SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
2029 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2030 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
2032 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
2035 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
2036 // FNEARBYINT and FRINT are the same, except in their handling of FP
2037 // exceptions. Those aren't really meaningful for us, and OpenCL only has
2038 // rint, so just treat them as equivalent.
2039 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
2042 // XXX - May require not supporting f32 denormals?
2043 SDValue AMDGPUTargetLowering::LowerFROUND32(SDValue Op, SelectionDAG &DAG) const {
2045 SDValue X = Op.getOperand(0);
2047 SDValue T = DAG.getNode(ISD::FTRUNC, SL, MVT::f32, X);
2049 // TODO: Should this propagate fast-math-flags?
2051 SDValue Diff = DAG.getNode(ISD::FSUB, SL, MVT::f32, X, T);
2053 SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, MVT::f32, Diff);
2055 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f32);
2056 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
2057 const SDValue Half = DAG.getConstantFP(0.5, SL, MVT::f32);
2059 SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f32, One, X);
2062 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f32);
2064 SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
2066 SDValue Sel = DAG.getNode(ISD::SELECT, SL, MVT::f32, Cmp, SignOne, Zero);
2068 return DAG.getNode(ISD::FADD, SL, MVT::f32, T, Sel);
2071 SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
2073 SDValue X = Op.getOperand(0);
2075 SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
2077 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2078 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
2079 const SDValue NegOne = DAG.getConstant(-1, SL, MVT::i32);
2080 const SDValue FiftyOne = DAG.getConstant(51, SL, MVT::i32);
2082 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
2084 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
2086 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
2088 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2090 const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), SL,
2093 SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
2094 SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
2095 DAG.getConstant(INT64_C(0x0008000000000000), SL,
2099 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
2100 SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
2101 DAG.getConstant(0, SL, MVT::i64), Tmp0,
2104 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
2105 D, DAG.getConstant(0, SL, MVT::i64));
2106 SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
2108 K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
2109 K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
2111 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2112 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2113 SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
2115 SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
2117 DAG.getConstantFP(1.0, SL, MVT::f64),
2118 DAG.getConstantFP(0.0, SL, MVT::f64));
2120 SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
2122 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
2123 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
2128 SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2129 EVT VT = Op.getValueType();
2132 return LowerFROUND32(Op, DAG);
2135 return LowerFROUND64(Op, DAG);
2137 llvm_unreachable("unhandled type");
2140 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2142 SDValue Src = Op.getOperand(0);
2144 // result = trunc(src);
2145 // if (src < 0.0 && src != result)
2148 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2150 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2151 const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
2154 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2156 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2157 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2158 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2160 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2161 // TODO: Should this propagate fast-math-flags?
2162 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2165 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2166 bool Signed) const {
2168 SDValue Src = Op.getOperand(0);
2170 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2172 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2173 DAG.getConstant(0, SL, MVT::i32));
2174 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2175 DAG.getConstant(1, SL, MVT::i32));
2177 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2180 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2182 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2183 DAG.getConstant(32, SL, MVT::i32));
2184 // TODO: Should this propagate fast-math-flags?
2185 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2188 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2189 SelectionDAG &DAG) const {
2190 SDValue S0 = Op.getOperand(0);
2191 if (S0.getValueType() != MVT::i64)
2194 EVT DestVT = Op.getValueType();
2195 if (DestVT == MVT::f64)
2196 return LowerINT_TO_FP64(Op, DAG, false);
2198 assert(DestVT == MVT::f32);
2202 // f32 uint_to_fp i64
2203 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2204 DAG.getConstant(0, DL, MVT::i32));
2205 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2206 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2207 DAG.getConstant(1, DL, MVT::i32));
2208 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2209 // TODO: Should this propagate fast-math-flags?
2210 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2211 DAG.getConstantFP(4294967296.0f, DL, MVT::f32)); // 2^32
2212 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2215 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2216 SelectionDAG &DAG) const {
2217 SDValue Src = Op.getOperand(0);
2218 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2219 return LowerINT_TO_FP64(Op, DAG, true);
2224 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2225 bool Signed) const {
2228 SDValue Src = Op.getOperand(0);
2230 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2232 SDValue K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), SL,
2234 SDValue K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), SL,
2236 // TODO: Should this propagate fast-math-flags?
2237 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2239 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2242 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2244 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2245 MVT::i32, FloorMul);
2246 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2248 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2250 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2253 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2254 SelectionDAG &DAG) const {
2255 SDValue Src = Op.getOperand(0);
2257 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2258 return LowerFP64_TO_INT(Op, DAG, true);
2263 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2264 SelectionDAG &DAG) const {
2265 SDValue Src = Op.getOperand(0);
2267 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2268 return LowerFP64_TO_INT(Op, DAG, false);
2273 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2274 SelectionDAG &DAG) const {
2275 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2276 MVT VT = Op.getSimpleValueType();
2277 MVT ScalarVT = VT.getScalarType();
2282 SDValue Src = Op.getOperand(0);
2285 // TODO: Don't scalarize on Evergreen?
2286 unsigned NElts = VT.getVectorNumElements();
2287 SmallVector<SDValue, 8> Args;
2288 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2290 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2291 for (unsigned I = 0; I < NElts; ++I)
2292 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2294 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2297 //===----------------------------------------------------------------------===//
2298 // Custom DAG optimizations
2299 //===----------------------------------------------------------------------===//
2301 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2302 APInt KnownZero, KnownOne;
2303 EVT VT = Op.getValueType();
2304 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2306 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2309 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2310 EVT VT = Op.getValueType();
2312 // In order for this to be a signed 24-bit value, bit 23, must
2314 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2315 // as unsigned 24-bit values.
2316 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2319 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2321 SelectionDAG &DAG = DCI.DAG;
2322 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2323 EVT VT = Op.getValueType();
2325 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2326 APInt KnownZero, KnownOne;
2327 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2328 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2329 DCI.CommitTargetLoweringOpt(TLO);
2332 template <typename IntTy>
2333 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2334 uint32_t Offset, uint32_t Width, SDLoc DL) {
2335 if (Width + Offset < 32) {
2336 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2337 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2338 return DAG.getConstant(Result, DL, MVT::i32);
2341 return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
2344 static bool usesAllNormalStores(SDNode *LoadVal) {
2345 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2346 if (!ISD::isNormalStore(*I))
2353 // If we have a copy of an illegal type, replace it with a load / store of an
2354 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2355 // instructions emitted when handling extloads and truncstores. Ideally we could
2356 // recognize the pack / unpack pattern to eliminate it.
2357 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2358 DAGCombinerInfo &DCI) const {
2359 if (!DCI.isBeforeLegalize())
2362 StoreSDNode *SN = cast<StoreSDNode>(N);
2363 SDValue Value = SN->getValue();
2364 EVT VT = Value.getValueType();
2366 if (isTypeLegal(VT) || SN->isVolatile() ||
2367 !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2370 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2371 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2374 EVT MemVT = LoadVal->getMemoryVT();
2377 SelectionDAG &DAG = DCI.DAG;
2378 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2380 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2382 LoadVal->getChain(),
2383 LoadVal->getBasePtr(),
2384 LoadVal->getOffset(),
2386 LoadVal->getMemOperand());
2388 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2389 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2391 return DAG.getStore(SN->getChain(), SL, NewLoad,
2392 SN->getBasePtr(), SN->getMemOperand());
2395 SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
2396 DAGCombinerInfo &DCI) const {
2397 if (N->getValueType(0) != MVT::i64)
2400 // i64 (shl x, 32) -> (build_pair 0, x)
2402 // Doing this with moves theoretically helps MI optimizations that understand
2403 // copies. 2 v_mov_b32_e32 will have the same code size / cycle count as
2404 // v_lshl_b64. In the SALU case, I think this is slightly worse since it
2405 // doubles the code size and I'm unsure about cycle count.
2406 const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
2407 if (!RHS || RHS->getZExtValue() != 32)
2410 SDValue LHS = N->getOperand(0);
2413 SelectionDAG &DAG = DCI.DAG;
2415 // Extract low 32-bits.
2416 SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
2418 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2419 return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, Zero, Lo);
2422 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2423 DAGCombinerInfo &DCI) const {
2424 EVT VT = N->getValueType(0);
2426 if (VT.isVector() || VT.getSizeInBits() > 32)
2429 SelectionDAG &DAG = DCI.DAG;
2432 SDValue N0 = N->getOperand(0);
2433 SDValue N1 = N->getOperand(1);
2436 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2437 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2438 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2439 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2440 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2441 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2442 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2443 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2448 // We need to use sext even for MUL_U24, because MUL_U24 is used
2449 // for signed multiply of 8 and 16-bit types.
2450 return DAG.getSExtOrTrunc(Mul, DL, VT);
2453 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2454 DAGCombinerInfo &DCI) const {
2455 SelectionDAG &DAG = DCI.DAG;
2458 switch(N->getOpcode()) {
2462 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
2465 return performShlCombine(N, DCI);
2468 return performMulCombine(N, DCI);
2469 case AMDGPUISD::MUL_I24:
2470 case AMDGPUISD::MUL_U24: {
2471 SDValue N0 = N->getOperand(0);
2472 SDValue N1 = N->getOperand(1);
2473 simplifyI24(N0, DCI);
2474 simplifyI24(N1, DCI);
2478 SDValue Cond = N->getOperand(0);
2479 if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
2480 EVT VT = N->getValueType(0);
2481 SDValue LHS = Cond.getOperand(0);
2482 SDValue RHS = Cond.getOperand(1);
2483 SDValue CC = Cond.getOperand(2);
2485 SDValue True = N->getOperand(1);
2486 SDValue False = N->getOperand(2);
2489 return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
2494 case AMDGPUISD::BFE_I32:
2495 case AMDGPUISD::BFE_U32: {
2496 assert(!N->getValueType(0).isVector() &&
2497 "Vector handling of BFE not implemented");
2498 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2502 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2504 return DAG.getConstant(0, DL, MVT::i32);
2506 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2510 SDValue BitsFrom = N->getOperand(0);
2511 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2513 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2515 if (OffsetVal == 0) {
2516 // This is already sign / zero extended, so try to fold away extra BFEs.
2517 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2519 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2520 if (OpSignBits >= SignBits)
2523 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2525 // This is a sign_extend_inreg. Replace it to take advantage of existing
2526 // DAG Combines. If not eliminated, we will match back to BFE during
2529 // TODO: The sext_inreg of extended types ends, although we can could
2530 // handle them in a single BFE.
2531 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2532 DAG.getValueType(SmallVT));
2535 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2538 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2540 return constantFoldBFE<int32_t>(DAG,
2541 CVal->getSExtValue(),
2547 return constantFoldBFE<uint32_t>(DAG,
2548 CVal->getZExtValue(),
2554 if ((OffsetVal + WidthVal) >= 32) {
2555 SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
2556 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2557 BitsFrom, ShiftVal);
2560 if (BitsFrom.hasOneUse()) {
2561 APInt Demanded = APInt::getBitsSet(32,
2563 OffsetVal + WidthVal);
2565 APInt KnownZero, KnownOne;
2566 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2567 !DCI.isBeforeLegalizeOps());
2568 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2569 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2570 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2571 KnownZero, KnownOne, TLO)) {
2572 DCI.CommitTargetLoweringOpt(TLO);
2580 return performStoreCombine(N, DCI);
2585 //===----------------------------------------------------------------------===//
2587 //===----------------------------------------------------------------------===//
2589 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2592 const SmallVectorImpl<ISD::InputArg> &Ins,
2593 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2595 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2596 if (Ins[i].ArgVT == Ins[i].VT) {
2597 OrigIns.push_back(Ins[i]);
2602 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2603 // Vector has been split into scalars.
2604 VT = Ins[i].ArgVT.getVectorElementType();
2605 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2606 Ins[i].ArgVT.getVectorElementType() !=
2607 Ins[i].VT.getVectorElementType()) {
2608 // Vector elements have been promoted
2611 // Vector has been spilt into smaller vectors.
2615 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2616 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2617 OrigIns.push_back(Arg);
2621 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2622 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2623 return CFP->isExactlyValue(1.0);
2625 return isAllOnesConstant(Op);
2628 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2629 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2630 return CFP->getValueAPF().isZero();
2632 return isNullConstant(Op);
2635 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2636 const TargetRegisterClass *RC,
2637 unsigned Reg, EVT VT) const {
2638 MachineFunction &MF = DAG.getMachineFunction();
2639 MachineRegisterInfo &MRI = MF.getRegInfo();
2640 unsigned VirtualRegister;
2641 if (!MRI.isLiveIn(Reg)) {
2642 VirtualRegister = MRI.createVirtualRegister(RC);
2643 MRI.addLiveIn(Reg, VirtualRegister);
2645 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2647 return DAG.getRegister(VirtualRegister, VT);
2650 uint32_t AMDGPUTargetLowering::getImplicitParameterOffset(
2651 const AMDGPUMachineFunction *MFI, const ImplicitParameter Param) const {
2652 uint64_t ArgOffset = MFI->ABIArgOffset;
2657 return ArgOffset + 4;
2659 llvm_unreachable("unexpected implicit parameter type");
2662 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2664 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2665 switch ((AMDGPUISD::NodeType)Opcode) {
2666 case AMDGPUISD::FIRST_NUMBER: break;
2668 NODE_NAME_CASE(CALL);
2669 NODE_NAME_CASE(UMUL);
2670 NODE_NAME_CASE(RET_FLAG);
2671 NODE_NAME_CASE(BRANCH_COND);
2674 NODE_NAME_CASE(DWORDADDR)
2675 NODE_NAME_CASE(FRACT)
2676 NODE_NAME_CASE(CLAMP)
2677 NODE_NAME_CASE(COS_HW)
2678 NODE_NAME_CASE(SIN_HW)
2679 NODE_NAME_CASE(FMAX_LEGACY)
2680 NODE_NAME_CASE(FMIN_LEGACY)
2681 NODE_NAME_CASE(FMAX3)
2682 NODE_NAME_CASE(SMAX3)
2683 NODE_NAME_CASE(UMAX3)
2684 NODE_NAME_CASE(FMIN3)
2685 NODE_NAME_CASE(SMIN3)
2686 NODE_NAME_CASE(UMIN3)
2687 NODE_NAME_CASE(URECIP)
2688 NODE_NAME_CASE(DIV_SCALE)
2689 NODE_NAME_CASE(DIV_FMAS)
2690 NODE_NAME_CASE(DIV_FIXUP)
2691 NODE_NAME_CASE(TRIG_PREOP)
2694 NODE_NAME_CASE(RSQ_LEGACY)
2695 NODE_NAME_CASE(RSQ_CLAMPED)
2696 NODE_NAME_CASE(LDEXP)
2697 NODE_NAME_CASE(FP_CLASS)
2698 NODE_NAME_CASE(DOT4)
2699 NODE_NAME_CASE(CARRY)
2700 NODE_NAME_CASE(BORROW)
2701 NODE_NAME_CASE(BFE_U32)
2702 NODE_NAME_CASE(BFE_I32)
2705 NODE_NAME_CASE(MUL_U24)
2706 NODE_NAME_CASE(MUL_I24)
2707 NODE_NAME_CASE(MAD_U24)
2708 NODE_NAME_CASE(MAD_I24)
2709 NODE_NAME_CASE(TEXTURE_FETCH)
2710 NODE_NAME_CASE(EXPORT)
2711 NODE_NAME_CASE(CONST_ADDRESS)
2712 NODE_NAME_CASE(REGISTER_LOAD)
2713 NODE_NAME_CASE(REGISTER_STORE)
2714 NODE_NAME_CASE(LOAD_CONSTANT)
2715 NODE_NAME_CASE(LOAD_INPUT)
2716 NODE_NAME_CASE(SAMPLE)
2717 NODE_NAME_CASE(SAMPLEB)
2718 NODE_NAME_CASE(SAMPLED)
2719 NODE_NAME_CASE(SAMPLEL)
2720 NODE_NAME_CASE(CVT_F32_UBYTE0)
2721 NODE_NAME_CASE(CVT_F32_UBYTE1)
2722 NODE_NAME_CASE(CVT_F32_UBYTE2)
2723 NODE_NAME_CASE(CVT_F32_UBYTE3)
2724 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2725 NODE_NAME_CASE(CONST_DATA_PTR)
2726 case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
2727 NODE_NAME_CASE(SENDMSG)
2728 NODE_NAME_CASE(INTERP_MOV)
2729 NODE_NAME_CASE(INTERP_P1)
2730 NODE_NAME_CASE(INTERP_P2)
2731 NODE_NAME_CASE(STORE_MSKOR)
2732 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2733 case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
2738 SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
2739 DAGCombinerInfo &DCI,
2740 unsigned &RefinementSteps,
2741 bool &UseOneConstNR) const {
2742 SelectionDAG &DAG = DCI.DAG;
2743 EVT VT = Operand.getValueType();
2745 if (VT == MVT::f32) {
2746 RefinementSteps = 0;
2747 return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
2750 // TODO: There is also f64 rsq instruction, but the documentation is less
2751 // clear on its precision.
2756 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
2757 DAGCombinerInfo &DCI,
2758 unsigned &RefinementSteps) const {
2759 SelectionDAG &DAG = DCI.DAG;
2760 EVT VT = Operand.getValueType();
2762 if (VT == MVT::f32) {
2763 // Reciprocal, < 1 ulp error.
2765 // This reciprocal approximation converges to < 0.5 ulp error with one
2766 // newton rhapson performed with two fused multiple adds (FMAs).
2768 RefinementSteps = 0;
2769 return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
2772 // TODO: There is also f64 rcp instruction, but the documentation is less
2773 // clear on its precision.
2778 static void computeKnownBitsForMinMax(const SDValue Op0,
2782 const SelectionDAG &DAG,
2784 APInt Op0Zero, Op0One;
2785 APInt Op1Zero, Op1One;
2786 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2787 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2789 KnownZero = Op0Zero & Op1Zero;
2790 KnownOne = Op0One & Op1One;
2793 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2797 const SelectionDAG &DAG,
2798 unsigned Depth) const {
2800 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2804 unsigned Opc = Op.getOpcode();
2809 case ISD::INTRINSIC_WO_CHAIN: {
2810 // FIXME: The intrinsic should just use the node.
2811 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2812 case AMDGPUIntrinsic::AMDGPU_imax:
2813 case AMDGPUIntrinsic::AMDGPU_umax:
2814 case AMDGPUIntrinsic::AMDGPU_imin:
2815 case AMDGPUIntrinsic::AMDGPU_umin:
2816 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2817 KnownZero, KnownOne, DAG, Depth);
2825 case AMDGPUISD::CARRY:
2826 case AMDGPUISD::BORROW: {
2827 KnownZero = APInt::getHighBitsSet(32, 31);
2831 case AMDGPUISD::BFE_I32:
2832 case AMDGPUISD::BFE_U32: {
2833 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2837 unsigned BitWidth = 32;
2838 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2840 if (Opc == AMDGPUISD::BFE_U32)
2841 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2848 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2850 const SelectionDAG &DAG,
2851 unsigned Depth) const {
2852 switch (Op.getOpcode()) {
2853 case AMDGPUISD::BFE_I32: {
2854 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2858 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2859 if (!isNullConstant(Op.getOperand(1)))
2862 // TODO: Could probably figure something out with non-0 offsets.
2863 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2864 return std::max(SignBits, Op0SignBits);
2867 case AMDGPUISD::BFE_U32: {
2868 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2869 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
2872 case AMDGPUISD::CARRY:
2873 case AMDGPUISD::BORROW: