1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file implements a TargetTransformInfo analysis pass specific to the
11 /// X86 target machine. It uses the target's detailed information to provide
12 /// more precise answers to certain TTI queries, while letting the target
13 /// independent and default TTI implementations handle the rest.
15 //===----------------------------------------------------------------------===//
18 #include "X86TargetMachine.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Target/CostTable.h"
23 #include "llvm/Target/TargetLowering.h"
26 #define DEBUG_TYPE "x86tti"
28 // Declare the pass initialization routine locally as target-specific passes
29 // don't have a target-wide initialization entry point, and so we rely on the
30 // pass constructor initialization.
32 void initializeX86TTIPass(PassRegistry &);
37 class X86TTI final : public ImmutablePass, public TargetTransformInfo {
38 const X86Subtarget *ST;
39 const X86TargetLowering *TLI;
41 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
42 /// are set if the result needs to be inserted and/or extracted from vectors.
43 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
46 X86TTI() : ImmutablePass(ID), ST(nullptr), TLI(nullptr) {
47 llvm_unreachable("This pass cannot be directly constructed");
50 X86TTI(const X86TargetMachine *TM)
51 : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
52 TLI(TM->getTargetLowering()) {
53 initializeX86TTIPass(*PassRegistry::getPassRegistry());
56 void initializePass() override {
60 void getAnalysisUsage(AnalysisUsage &AU) const override {
61 TargetTransformInfo::getAnalysisUsage(AU);
64 /// Pass identification.
67 /// Provide necessary pointer adjustments for the two base classes.
68 void *getAdjustedAnalysisPointer(const void *ID) override {
69 if (ID == &TargetTransformInfo::ID)
70 return (TargetTransformInfo*)this;
74 /// \name Scalar TTI Implementations
76 PopcntSupportKind getPopcntSupport(unsigned TyWidth) const override;
80 /// \name Vector TTI Implementations
83 unsigned getNumberOfRegisters(bool Vector) const override;
84 unsigned getRegisterBitWidth(bool Vector) const override;
85 unsigned getMaximumUnrollFactor() const override;
86 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
87 OperandValueKind) const override;
88 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
89 int Index, Type *SubTp) const override;
90 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
91 Type *Src) const override;
92 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
93 Type *CondTy) const override;
94 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
95 unsigned Index) const override;
96 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
97 unsigned AddressSpace) const override;
99 unsigned getAddressComputationCost(Type *PtrTy,
100 bool IsComplex) const override;
102 unsigned getReductionCost(unsigned Opcode, Type *Ty,
103 bool IsPairwiseForm) const override;
105 unsigned getIntImmCost(int64_t) const;
107 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override;
109 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
110 Type *Ty) const override;
111 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
112 Type *Ty) const override;
117 } // end anonymous namespace
119 INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti",
120 "X86 Target Transform Info", true, true, false)
124 llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) {
125 return new X86TTI(TM);
129 //===----------------------------------------------------------------------===//
133 //===----------------------------------------------------------------------===//
135 X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const {
136 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
137 // TODO: Currently the __builtin_popcount() implementation using SSE3
138 // instructions is inefficient. Once the problem is fixed, we should
139 // call ST->hasSSE3() instead of ST->hasPOPCNT().
140 return ST->hasPOPCNT() ? PSK_FastHardware : PSK_Software;
143 unsigned X86TTI::getNumberOfRegisters(bool Vector) const {
144 if (Vector && !ST->hasSSE1())
152 unsigned X86TTI::getRegisterBitWidth(bool Vector) const {
154 if (ST->hasAVX()) return 256;
155 if (ST->hasSSE1()) return 128;
165 unsigned X86TTI::getMaximumUnrollFactor() const {
169 // Sandybridge and Haswell have multiple execution ports and pipelined
177 unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
178 OperandValueKind Op1Info,
179 OperandValueKind Op2Info) const {
180 // Legalize the type.
181 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
183 int ISD = TLI->InstructionOpcodeToISD(Opcode);
184 assert(ISD && "Invalid opcode");
186 static const CostTblEntry<MVT::SimpleValueType>
187 AVX2UniformConstCostTable[] = {
188 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
189 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
190 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
191 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
194 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
196 int Idx = CostTableLookup(AVX2UniformConstCostTable, ISD, LT.second);
198 return LT.first * AVX2UniformConstCostTable[Idx].Cost;
201 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTable[] = {
202 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
203 // customize them to detect the cases where shift amount is a scalar one.
204 { ISD::SHL, MVT::v4i32, 1 },
205 { ISD::SRL, MVT::v4i32, 1 },
206 { ISD::SRA, MVT::v4i32, 1 },
207 { ISD::SHL, MVT::v8i32, 1 },
208 { ISD::SRL, MVT::v8i32, 1 },
209 { ISD::SRA, MVT::v8i32, 1 },
210 { ISD::SHL, MVT::v2i64, 1 },
211 { ISD::SRL, MVT::v2i64, 1 },
212 { ISD::SHL, MVT::v4i64, 1 },
213 { ISD::SRL, MVT::v4i64, 1 },
215 { ISD::SHL, MVT::v32i8, 42 }, // cmpeqb sequence.
216 { ISD::SHL, MVT::v16i16, 16*10 }, // Scalarized.
218 { ISD::SRL, MVT::v32i8, 32*10 }, // Scalarized.
219 { ISD::SRL, MVT::v16i16, 8*10 }, // Scalarized.
221 { ISD::SRA, MVT::v32i8, 32*10 }, // Scalarized.
222 { ISD::SRA, MVT::v16i16, 16*10 }, // Scalarized.
223 { ISD::SRA, MVT::v4i64, 4*10 }, // Scalarized.
225 // Vectorizing division is a bad idea. See the SSE2 table for more comments.
226 { ISD::SDIV, MVT::v32i8, 32*20 },
227 { ISD::SDIV, MVT::v16i16, 16*20 },
228 { ISD::SDIV, MVT::v8i32, 8*20 },
229 { ISD::SDIV, MVT::v4i64, 4*20 },
230 { ISD::UDIV, MVT::v32i8, 32*20 },
231 { ISD::UDIV, MVT::v16i16, 16*20 },
232 { ISD::UDIV, MVT::v8i32, 8*20 },
233 { ISD::UDIV, MVT::v4i64, 4*20 },
236 // Look for AVX2 lowering tricks.
238 if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
239 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
240 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
241 // On AVX2, a packed v16i16 shift left by a constant build_vector
242 // is lowered into a vector multiply (vpmullw).
245 int Idx = CostTableLookup(AVX2CostTable, ISD, LT.second);
247 return LT.first * AVX2CostTable[Idx].Cost;
250 static const CostTblEntry<MVT::SimpleValueType>
251 SSE2UniformConstCostTable[] = {
252 // We don't correctly identify costs of casts because they are marked as
254 // Constant splats are cheaper for the following instructions.
255 { ISD::SHL, MVT::v16i8, 1 }, // psllw.
256 { ISD::SHL, MVT::v8i16, 1 }, // psllw.
257 { ISD::SHL, MVT::v4i32, 1 }, // pslld
258 { ISD::SHL, MVT::v2i64, 1 }, // psllq.
260 { ISD::SRL, MVT::v16i8, 1 }, // psrlw.
261 { ISD::SRL, MVT::v8i16, 1 }, // psrlw.
262 { ISD::SRL, MVT::v4i32, 1 }, // psrld.
263 { ISD::SRL, MVT::v2i64, 1 }, // psrlq.
265 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
266 { ISD::SRA, MVT::v8i16, 1 }, // psraw.
267 { ISD::SRA, MVT::v4i32, 1 }, // psrad.
269 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
270 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
271 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
272 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
275 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
278 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
279 return LT.first * 15;
281 int Idx = CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second);
283 return LT.first * SSE2UniformConstCostTable[Idx].Cost;
286 if (ISD == ISD::SHL &&
287 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
289 if ((VT == MVT::v8i16 && ST->hasSSE2()) ||
290 (VT == MVT::v4i32 && ST->hasSSE41()))
291 // Vector shift left by non uniform constant can be lowered
292 // into vector multiply (pmullw/pmulld).
294 if (VT == MVT::v4i32 && ST->hasSSE2())
295 // A vector shift left by non uniform constant is converted
296 // into a vector multiply; the new multiply is eventually
297 // lowered into a sequence of shuffles and 2 x pmuludq.
301 static const CostTblEntry<MVT::SimpleValueType> SSE2CostTable[] = {
302 // We don't correctly identify costs of casts because they are marked as
304 // For some cases, where the shift amount is a scalar we would be able
305 // to generate better code. Unfortunately, when this is the case the value
306 // (the splat) will get hoisted out of the loop, thereby making it invisible
307 // to ISel. The cost model must return worst case assumptions because it is
308 // used for vectorization and we don't want to make vectorized code worse
310 { ISD::SHL, MVT::v16i8, 30 }, // cmpeqb sequence.
311 { ISD::SHL, MVT::v8i16, 8*10 }, // Scalarized.
312 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
313 { ISD::SHL, MVT::v2i64, 2*10 }, // Scalarized.
314 { ISD::SHL, MVT::v4i64, 4*10 }, // Scalarized.
316 { ISD::SRL, MVT::v16i8, 16*10 }, // Scalarized.
317 { ISD::SRL, MVT::v8i16, 8*10 }, // Scalarized.
318 { ISD::SRL, MVT::v4i32, 4*10 }, // Scalarized.
319 { ISD::SRL, MVT::v2i64, 2*10 }, // Scalarized.
321 { ISD::SRA, MVT::v16i8, 16*10 }, // Scalarized.
322 { ISD::SRA, MVT::v8i16, 8*10 }, // Scalarized.
323 { ISD::SRA, MVT::v4i32, 4*10 }, // Scalarized.
324 { ISD::SRA, MVT::v2i64, 2*10 }, // Scalarized.
326 // It is not a good idea to vectorize division. We have to scalarize it and
327 // in the process we will often end up having to spilling regular
328 // registers. The overhead of division is going to dominate most kernels
329 // anyways so try hard to prevent vectorization of division - it is
330 // generally a bad idea. Assume somewhat arbitrarily that we have to be able
331 // to hide "20 cycles" for each lane.
332 { ISD::SDIV, MVT::v16i8, 16*20 },
333 { ISD::SDIV, MVT::v8i16, 8*20 },
334 { ISD::SDIV, MVT::v4i32, 4*20 },
335 { ISD::SDIV, MVT::v2i64, 2*20 },
336 { ISD::UDIV, MVT::v16i8, 16*20 },
337 { ISD::UDIV, MVT::v8i16, 8*20 },
338 { ISD::UDIV, MVT::v4i32, 4*20 },
339 { ISD::UDIV, MVT::v2i64, 2*20 },
343 int Idx = CostTableLookup(SSE2CostTable, ISD, LT.second);
345 return LT.first * SSE2CostTable[Idx].Cost;
348 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTable[] = {
349 // We don't have to scalarize unsupported ops. We can issue two half-sized
350 // operations and we only need to extract the upper YMM half.
351 // Two ops + 1 extract + 1 insert = 4.
352 { ISD::MUL, MVT::v16i16, 4 },
353 { ISD::MUL, MVT::v8i32, 4 },
354 { ISD::SUB, MVT::v8i32, 4 },
355 { ISD::ADD, MVT::v8i32, 4 },
356 { ISD::SUB, MVT::v4i64, 4 },
357 { ISD::ADD, MVT::v4i64, 4 },
358 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
359 // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
360 // Because we believe v4i64 to be a legal type, we must also include the
361 // split factor of two in the cost table. Therefore, the cost here is 18
363 { ISD::MUL, MVT::v4i64, 18 },
366 // Look for AVX1 lowering tricks.
367 if (ST->hasAVX() && !ST->hasAVX2()) {
370 // v16i16 and v8i32 shifts by non-uniform constants are lowered into a
371 // sequence of extract + two vector multiply + insert.
372 if (ISD == ISD::SHL && (VT == MVT::v8i32 || VT == MVT::v16i16) &&
373 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)
376 int Idx = CostTableLookup(AVX1CostTable, ISD, VT);
378 return LT.first * AVX1CostTable[Idx].Cost;
381 // Custom lowering of vectors.
382 static const CostTblEntry<MVT::SimpleValueType> CustomLowered[] = {
383 // A v2i64/v4i64 and multiply is custom lowered as a series of long
384 // multiplies(3), shifts(4) and adds(2).
385 { ISD::MUL, MVT::v2i64, 9 },
386 { ISD::MUL, MVT::v4i64, 9 },
388 int Idx = CostTableLookup(CustomLowered, ISD, LT.second);
390 return LT.first * CustomLowered[Idx].Cost;
392 // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
393 // 2x pmuludq, 2x shuffle.
394 if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
398 // Fallback to the default implementation.
399 return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
403 unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
405 // We only estimate the cost of reverse and alternate shuffles.
406 if (Kind != SK_Reverse && Kind != SK_Alternate)
407 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
409 if (Kind == SK_Reverse) {
410 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
412 if (LT.second.getSizeInBits() > 128)
413 Cost = 3; // Extract + insert + copy.
415 // Multiple by the number of parts.
416 return Cost * LT.first;
419 if (Kind == SK_Alternate) {
420 // 64-bit packed float vectors (v2f32) are widened to type v4f32.
421 // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
422 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
424 // The backend knows how to generate a single VEX.256 version of
425 // instruction VPBLENDW if the target supports AVX2.
426 if (ST->hasAVX2() && LT.second == MVT::v16i16)
429 static const CostTblEntry<MVT::SimpleValueType> AVXAltShuffleTbl[] = {
430 {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1}, // vblendpd
431 {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1}, // vblendpd
433 {ISD::VECTOR_SHUFFLE, MVT::v8i32, 1}, // vblendps
434 {ISD::VECTOR_SHUFFLE, MVT::v8f32, 1}, // vblendps
436 // This shuffle is custom lowered into a sequence of:
437 // 2x vextractf128 , 2x vpblendw , 1x vinsertf128
438 {ISD::VECTOR_SHUFFLE, MVT::v16i16, 5},
440 // This shuffle is custom lowered into a long sequence of:
441 // 2x vextractf128 , 4x vpshufb , 2x vpor , 1x vinsertf128
442 {ISD::VECTOR_SHUFFLE, MVT::v32i8, 9}
446 int Idx = CostTableLookup(AVXAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
448 return LT.first * AVXAltShuffleTbl[Idx].Cost;
451 static const CostTblEntry<MVT::SimpleValueType> SSE41AltShuffleTbl[] = {
452 // These are lowered into movsd.
453 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
454 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
456 // packed float vectors with four elements are lowered into BLENDI dag
457 // nodes. A v4i32/v4f32 BLENDI generates a single 'blendps'/'blendpd'.
458 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
459 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
461 // This shuffle generates a single pshufw.
462 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
464 // There is no instruction that matches a v16i8 alternate shuffle.
465 // The backend will expand it into the sequence 'pshufb + pshufb + or'.
466 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}
469 if (ST->hasSSE41()) {
470 int Idx = CostTableLookup(SSE41AltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
472 return LT.first * SSE41AltShuffleTbl[Idx].Cost;
475 static const CostTblEntry<MVT::SimpleValueType> SSSE3AltShuffleTbl[] = {
476 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd
477 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd
479 // SSE3 doesn't have 'blendps'. The following shuffles are expanded into
480 // the sequence 'shufps + pshufd'
481 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
482 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
484 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 3}, // pshufb + pshufb + or
485 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3} // pshufb + pshufb + or
488 if (ST->hasSSSE3()) {
489 int Idx = CostTableLookup(SSSE3AltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
491 return LT.first * SSSE3AltShuffleTbl[Idx].Cost;
494 static const CostTblEntry<MVT::SimpleValueType> SSEAltShuffleTbl[] = {
495 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd
496 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd
498 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, // shufps + pshufd
499 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, // shufps + pshufd
501 // This is expanded into a long sequence of four extract + four insert.
502 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 8}, // 4 x pextrw + 4 pinsrw.
504 // 8 x (pinsrw + pextrw + and + movb + movzb + or)
505 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 48}
508 // Fall-back (SSE3 and SSE2).
509 int Idx = CostTableLookup(SSEAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
511 return LT.first * SSEAltShuffleTbl[Idx].Cost;
512 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
515 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
518 unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
519 int ISD = TLI->InstructionOpcodeToISD(Opcode);
520 assert(ISD && "Invalid opcode");
522 std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
523 std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
525 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
527 // These are somewhat magic numbers justified by looking at the output of
528 // Intel's IACA, running some kernels and making sure when we take
529 // legalization into account the throughput will be overestimated.
530 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
531 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
532 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
533 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
534 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
535 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
536 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
537 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
538 // There are faster sequences for float conversions.
539 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
540 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
541 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
542 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
543 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
544 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
545 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
546 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
549 if (ST->hasSSE2() && !ST->hasAVX()) {
551 ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
553 return LTSrc.first * SSE2ConvTbl[Idx].Cost;
556 EVT SrcTy = TLI->getValueType(Src);
557 EVT DstTy = TLI->getValueType(Dst);
559 // The function getSimpleVT only handles simple value types.
560 if (!SrcTy.isSimple() || !DstTy.isSimple())
561 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
563 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
564 AVX2ConversionTbl[] = {
565 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
566 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
567 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
568 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
569 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
570 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
571 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
572 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
573 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
574 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
575 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
576 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
577 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
578 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
579 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
580 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
582 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 },
583 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 },
584 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 },
585 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 },
586 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 },
587 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 },
590 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
591 AVXConversionTbl[] = {
592 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
593 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
594 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 },
595 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 },
596 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 7 },
597 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 },
598 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
599 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
600 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 },
601 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 },
602 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 6 },
603 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 },
604 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 },
605 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
606 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
607 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
609 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 },
610 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 },
611 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 4 },
612 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 },
613 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 },
614 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 },
615 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 9 },
617 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 },
618 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 },
619 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
620 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
621 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
622 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
623 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 },
624 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
625 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 },
626 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 },
627 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 },
628 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
630 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 },
631 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 },
632 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
633 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 },
634 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 },
635 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 },
636 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
637 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 },
638 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 },
639 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 },
640 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 },
641 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 },
642 // The generic code to compute the scalar overhead is currently broken.
643 // Workaround this limitation by estimating the scalarization overhead
644 // here. We have roughly 10 instructions per scalar element.
645 // Multiply that by the vector width.
646 // FIXME: remove that when PR19268 is fixed.
647 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
648 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 4*10 },
650 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 7 },
651 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
652 // This node is expanded into scalarized operations but BasicTTI is overly
653 // optimistic estimating its cost. It computes 3 per element (one
654 // vector-extract, one scalar conversion and one vector-insert). The
655 // problem is that the inserts form a read-modify-write chain so latency
656 // should be factored in too. Inflating the cost per element by 1.
657 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 },
658 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 },
662 int Idx = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
663 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
665 return AVX2ConversionTbl[Idx].Cost;
669 int Idx = ConvertCostTableLookup(AVXConversionTbl, ISD, DstTy.getSimpleVT(),
670 SrcTy.getSimpleVT());
672 return AVXConversionTbl[Idx].Cost;
675 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
678 unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
679 Type *CondTy) const {
680 // Legalize the type.
681 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
685 int ISD = TLI->InstructionOpcodeToISD(Opcode);
686 assert(ISD && "Invalid opcode");
688 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTbl[] = {
689 { ISD::SETCC, MVT::v2f64, 1 },
690 { ISD::SETCC, MVT::v4f32, 1 },
691 { ISD::SETCC, MVT::v2i64, 1 },
692 { ISD::SETCC, MVT::v4i32, 1 },
693 { ISD::SETCC, MVT::v8i16, 1 },
694 { ISD::SETCC, MVT::v16i8, 1 },
697 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTbl[] = {
698 { ISD::SETCC, MVT::v4f64, 1 },
699 { ISD::SETCC, MVT::v8f32, 1 },
700 // AVX1 does not support 8-wide integer compare.
701 { ISD::SETCC, MVT::v4i64, 4 },
702 { ISD::SETCC, MVT::v8i32, 4 },
703 { ISD::SETCC, MVT::v16i16, 4 },
704 { ISD::SETCC, MVT::v32i8, 4 },
707 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTbl[] = {
708 { ISD::SETCC, MVT::v4i64, 1 },
709 { ISD::SETCC, MVT::v8i32, 1 },
710 { ISD::SETCC, MVT::v16i16, 1 },
711 { ISD::SETCC, MVT::v32i8, 1 },
715 int Idx = CostTableLookup(AVX2CostTbl, ISD, MTy);
717 return LT.first * AVX2CostTbl[Idx].Cost;
721 int Idx = CostTableLookup(AVX1CostTbl, ISD, MTy);
723 return LT.first * AVX1CostTbl[Idx].Cost;
726 if (ST->hasSSE42()) {
727 int Idx = CostTableLookup(SSE42CostTbl, ISD, MTy);
729 return LT.first * SSE42CostTbl[Idx].Cost;
732 return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
735 unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val,
736 unsigned Index) const {
737 assert(Val->isVectorTy() && "This must be a vector type");
740 // Legalize the type.
741 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val);
743 // This type is legalized to a scalar type.
744 if (!LT.second.isVector())
747 // The type may be split. Normalize the index to the new type.
748 unsigned Width = LT.second.getVectorNumElements();
749 Index = Index % Width;
751 // Floating point scalars are already located in index #0.
752 if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
756 return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index);
759 unsigned X86TTI::getScalarizationOverhead(Type *Ty, bool Insert,
760 bool Extract) const {
761 assert (Ty->isVectorTy() && "Can only scalarize vectors");
764 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
766 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
768 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
774 unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
775 unsigned AddressSpace) const {
776 // Handle non-power-of-two vectors such as <3 x float>
777 if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
778 unsigned NumElem = VTy->getVectorNumElements();
780 // Handle a few common cases:
782 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
783 // Cost = 64 bit store + extract + 32 bit store.
787 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
788 // Cost = 128 bit store + unpack + 64 bit store.
791 // Assume that all other non-power-of-two numbers are scalarized.
792 if (!isPowerOf2_32(NumElem)) {
793 unsigned Cost = TargetTransformInfo::getMemoryOpCost(Opcode,
794 VTy->getScalarType(),
797 unsigned SplitCost = getScalarizationOverhead(Src,
798 Opcode == Instruction::Load,
799 Opcode==Instruction::Store);
800 return NumElem * Cost + SplitCost;
804 // Legalize the type.
805 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
806 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
809 // Each load/store unit costs 1.
810 unsigned Cost = LT.first * 1;
812 // On Sandybridge 256bit load/stores are double pumped
813 // (but not on Haswell).
814 if (LT.second.getSizeInBits() > 128 && !ST->hasAVX2())
820 unsigned X86TTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
821 // Address computations in vectorized code with non-consecutive addresses will
822 // likely result in more instructions compared to scalar code where the
823 // computation can more often be merged into the index mode. The resulting
824 // extra micro-ops can significantly decrease throughput.
825 unsigned NumVectorInstToHideOverhead = 10;
827 if (Ty->isVectorTy() && IsComplex)
828 return NumVectorInstToHideOverhead;
830 return TargetTransformInfo::getAddressComputationCost(Ty, IsComplex);
833 unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
834 bool IsPairwise) const {
836 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
840 int ISD = TLI->InstructionOpcodeToISD(Opcode);
841 assert(ISD && "Invalid opcode");
843 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
844 // and make it as the cost.
846 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblPairWise[] = {
847 { ISD::FADD, MVT::v2f64, 2 },
848 { ISD::FADD, MVT::v4f32, 4 },
849 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
850 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
851 { ISD::ADD, MVT::v8i16, 5 },
854 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblPairWise[] = {
855 { ISD::FADD, MVT::v4f32, 4 },
856 { ISD::FADD, MVT::v4f64, 5 },
857 { ISD::FADD, MVT::v8f32, 7 },
858 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
859 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
860 { ISD::ADD, MVT::v4i64, 5 }, // The data reported by the IACA tool is "4.8".
861 { ISD::ADD, MVT::v8i16, 5 },
862 { ISD::ADD, MVT::v8i32, 5 },
865 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblNoPairWise[] = {
866 { ISD::FADD, MVT::v2f64, 2 },
867 { ISD::FADD, MVT::v4f32, 4 },
868 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
869 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3".
870 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3".
873 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblNoPairWise[] = {
874 { ISD::FADD, MVT::v4f32, 3 },
875 { ISD::FADD, MVT::v4f64, 3 },
876 { ISD::FADD, MVT::v8f32, 4 },
877 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
878 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "2.8".
879 { ISD::ADD, MVT::v4i64, 3 },
880 { ISD::ADD, MVT::v8i16, 4 },
881 { ISD::ADD, MVT::v8i32, 5 },
886 int Idx = CostTableLookup(AVX1CostTblPairWise, ISD, MTy);
888 return LT.first * AVX1CostTblPairWise[Idx].Cost;
891 if (ST->hasSSE42()) {
892 int Idx = CostTableLookup(SSE42CostTblPairWise, ISD, MTy);
894 return LT.first * SSE42CostTblPairWise[Idx].Cost;
898 int Idx = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy);
900 return LT.first * AVX1CostTblNoPairWise[Idx].Cost;
903 if (ST->hasSSE42()) {
904 int Idx = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy);
906 return LT.first * SSE42CostTblNoPairWise[Idx].Cost;
910 return TargetTransformInfo::getReductionCost(Opcode, ValTy, IsPairwise);
913 /// \brief Calculate the cost of materializing a 64-bit value. This helper
914 /// method might only calculate a fraction of a larger immediate. Therefore it
915 /// is valid to return a cost of ZERO.
916 unsigned X86TTI::getIntImmCost(int64_t Val) const {
923 return 2 * TCC_Basic;
926 unsigned X86TTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
927 assert(Ty->isIntegerTy());
929 unsigned BitSize = Ty->getPrimitiveSizeInBits();
933 // Never hoist constants larger than 128bit, because this might lead to
934 // incorrect code generation or assertions in codegen.
935 // Fixme: Create a cost model for types larger than i128 once the codegen
936 // issues have been fixed.
943 // Sign-extend all constants to a multiple of 64-bit.
946 ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
948 // Split the constant into 64-bit chunks and calculate the cost for each
951 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
952 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
953 int64_t Val = Tmp.getSExtValue();
954 Cost += getIntImmCost(Val);
956 // We need at least one instruction to materialze the constant.
957 return std::max(1U, Cost);
960 unsigned X86TTI::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
962 assert(Ty->isIntegerTy());
964 unsigned BitSize = Ty->getPrimitiveSizeInBits();
965 // There is no cost model for constants with a bit size of 0. Return TCC_Free
966 // here, so that constant hoisting will ignore this constant.
970 unsigned ImmIdx = ~0U;
972 default: return TCC_Free;
973 case Instruction::GetElementPtr:
974 // Always hoist the base address of a GetElementPtr. This prevents the
975 // creation of new constants for every base constant that gets constant
976 // folded with the offset.
978 return 2 * TCC_Basic;
980 case Instruction::Store:
983 case Instruction::Add:
984 case Instruction::Sub:
985 case Instruction::Mul:
986 case Instruction::UDiv:
987 case Instruction::SDiv:
988 case Instruction::URem:
989 case Instruction::SRem:
990 case Instruction::And:
991 case Instruction::Or:
992 case Instruction::Xor:
993 case Instruction::ICmp:
996 // Always return TCC_Free for the shift value of a shift instruction.
997 case Instruction::Shl:
998 case Instruction::LShr:
999 case Instruction::AShr:
1003 case Instruction::Trunc:
1004 case Instruction::ZExt:
1005 case Instruction::SExt:
1006 case Instruction::IntToPtr:
1007 case Instruction::PtrToInt:
1008 case Instruction::BitCast:
1009 case Instruction::PHI:
1010 case Instruction::Call:
1011 case Instruction::Select:
1012 case Instruction::Ret:
1013 case Instruction::Load:
1017 if (Idx == ImmIdx) {
1018 unsigned NumConstants = (BitSize + 63) / 64;
1019 unsigned Cost = X86TTI::getIntImmCost(Imm, Ty);
1020 return (Cost <= NumConstants * TCC_Basic)
1021 ? static_cast<unsigned>(TCC_Free)
1025 return X86TTI::getIntImmCost(Imm, Ty);
1028 unsigned X86TTI::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
1029 const APInt &Imm, Type *Ty) const {
1030 assert(Ty->isIntegerTy());
1032 unsigned BitSize = Ty->getPrimitiveSizeInBits();
1033 // There is no cost model for constants with a bit size of 0. Return TCC_Free
1034 // here, so that constant hoisting will ignore this constant.
1039 default: return TCC_Free;
1040 case Intrinsic::sadd_with_overflow:
1041 case Intrinsic::uadd_with_overflow:
1042 case Intrinsic::ssub_with_overflow:
1043 case Intrinsic::usub_with_overflow:
1044 case Intrinsic::smul_with_overflow:
1045 case Intrinsic::umul_with_overflow:
1046 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
1049 case Intrinsic::experimental_stackmap:
1050 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1053 case Intrinsic::experimental_patchpoint_void:
1054 case Intrinsic::experimental_patchpoint_i64:
1055 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1059 return X86TTI::getIntImmCost(Imm, Ty);