1 //===-- ARMTargetTransformInfo.cpp - ARM 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 /// ARM 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 //===----------------------------------------------------------------------===//
17 #define DEBUG_TYPE "armtti"
19 #include "ARMTargetMachine.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Target/TargetLowering.h"
23 #include "llvm/Target/CostTable.h"
26 // Declare the pass initialization routine locally as target-specific passes
27 // don't havve a target-wide initialization entry point, and so we rely on the
28 // pass constructor initialization.
30 void initializeARMTTIPass(PassRegistry &);
35 class ARMTTI : public ImmutablePass, public TargetTransformInfo {
36 const ARMBaseTargetMachine *TM;
37 const ARMSubtarget *ST;
38 const ARMTargetLowering *TLI;
40 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
41 /// are set if the result needs to be inserted and/or extracted from vectors.
42 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
45 ARMTTI() : ImmutablePass(ID), TM(0), ST(0), TLI(0) {
46 llvm_unreachable("This pass cannot be directly constructed");
49 ARMTTI(const ARMBaseTargetMachine *TM)
50 : ImmutablePass(ID), TM(TM), ST(TM->getSubtargetImpl()),
51 TLI(TM->getTargetLowering()) {
52 initializeARMTTIPass(*PassRegistry::getPassRegistry());
55 virtual void initializePass() {
59 virtual void finalizePass() {
63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 TargetTransformInfo::getAnalysisUsage(AU);
67 /// Pass identification.
70 /// Provide necessary pointer adjustments for the two base classes.
71 virtual void *getAdjustedAnalysisPointer(const void *ID) {
72 if (ID == &TargetTransformInfo::ID)
73 return (TargetTransformInfo*)this;
77 /// \name Scalar TTI Implementations
80 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
85 /// \name Vector TTI Implementations
88 unsigned getNumberOfRegisters(bool Vector) const {
95 if (ST->isThumb1Only())
100 unsigned getRegisterBitWidth(bool Vector) const {
110 unsigned getMaximumUnrollFactor() const {
111 // These are out of order CPUs:
112 if (ST->isCortexA15() || ST->isSwift())
117 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
120 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) const;
124 } // end anonymous namespace
126 INITIALIZE_AG_PASS(ARMTTI, TargetTransformInfo, "armtti",
127 "ARM Target Transform Info", true, true, false)
131 llvm::createARMTargetTransformInfoPass(const ARMBaseTargetMachine *TM) {
132 return new ARMTTI(TM);
136 unsigned ARMTTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
137 assert(Ty->isIntegerTy());
139 unsigned Bits = Ty->getPrimitiveSizeInBits();
140 if (Bits == 0 || Bits > 32)
143 int32_t SImmVal = Imm.getSExtValue();
144 uint32_t ZImmVal = Imm.getZExtValue();
145 if (!ST->isThumb()) {
146 if ((SImmVal >= 0 && SImmVal < 65536) ||
147 (ARM_AM::getSOImmVal(ZImmVal) != -1) ||
148 (ARM_AM::getSOImmVal(~ZImmVal) != -1))
150 return ST->hasV6T2Ops() ? 2 : 3;
151 } else if (ST->isThumb2()) {
152 if ((SImmVal >= 0 && SImmVal < 65536) ||
153 (ARM_AM::getT2SOImmVal(ZImmVal) != -1) ||
154 (ARM_AM::getT2SOImmVal(~ZImmVal) != -1))
156 return ST->hasV6T2Ops() ? 2 : 3;
158 if (SImmVal >= 0 && SImmVal < 256)
160 if ((~ZImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal))
162 // Load from constantpool.
168 unsigned ARMTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
170 int ISD = TLI->InstructionOpcodeToISD(Opcode);
171 assert(ISD && "Invalid opcode");
173 EVT SrcTy = TLI->getValueType(Src);
174 EVT DstTy = TLI->getValueType(Dst);
176 if (!SrcTy.isSimple() || !DstTy.isSimple())
177 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
179 // Some arithmetic, load and store operations have specific instructions
180 // to cast up/down their types automatically at no extra cost.
181 // TODO: Get these tables to know at least what the related operations are.
182 static const TypeConversionCostTblEntry<MVT> NEONVectorConversionTbl[] = {
183 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0 },
184 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0 },
185 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
186 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
187 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 0 },
188 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 },
190 // Vector float <-> i32 conversions.
191 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
192 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
193 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 },
194 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 },
196 // Vector double <-> i32 conversions.
197 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
198 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
199 { ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 2 },
200 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 2 }
203 if (SrcTy.isVector() && ST->hasNEON()) {
204 int Idx = ConvertCostTableLookup<MVT>(NEONVectorConversionTbl,
205 array_lengthof(NEONVectorConversionTbl),
206 ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT());
208 return NEONVectorConversionTbl[Idx].Cost;
211 // Scalar float to integer conversions.
212 static const TypeConversionCostTblEntry<MVT> NEONFloatConversionTbl[] = {
213 { ISD::FP_TO_SINT, MVT::i1, MVT::f32, 2 },
214 { ISD::FP_TO_UINT, MVT::i1, MVT::f32, 2 },
215 { ISD::FP_TO_SINT, MVT::i1, MVT::f64, 2 },
216 { ISD::FP_TO_UINT, MVT::i1, MVT::f64, 2 },
217 { ISD::FP_TO_SINT, MVT::i8, MVT::f32, 2 },
218 { ISD::FP_TO_UINT, MVT::i8, MVT::f32, 2 },
219 { ISD::FP_TO_SINT, MVT::i8, MVT::f64, 2 },
220 { ISD::FP_TO_UINT, MVT::i8, MVT::f64, 2 },
221 { ISD::FP_TO_SINT, MVT::i16, MVT::f32, 2 },
222 { ISD::FP_TO_UINT, MVT::i16, MVT::f32, 2 },
223 { ISD::FP_TO_SINT, MVT::i16, MVT::f64, 2 },
224 { ISD::FP_TO_UINT, MVT::i16, MVT::f64, 2 },
225 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 2 },
226 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 2 },
227 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 2 },
228 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 2 },
229 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 10 },
230 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 10 },
231 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 10 },
232 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 10 }
234 if (SrcTy.isFloatingPoint() && ST->hasNEON()) {
235 int Idx = ConvertCostTableLookup<MVT>(NEONFloatConversionTbl,
236 array_lengthof(NEONFloatConversionTbl),
237 ISD, DstTy.getSimpleVT(),
238 SrcTy.getSimpleVT());
240 return NEONFloatConversionTbl[Idx].Cost;
244 // Scalar integer to float conversions.
245 static const TypeConversionCostTblEntry<MVT> NEONIntegerConversionTbl[] = {
246 { ISD::SINT_TO_FP, MVT::f32, MVT::i1, 2 },
247 { ISD::UINT_TO_FP, MVT::f32, MVT::i1, 2 },
248 { ISD::SINT_TO_FP, MVT::f64, MVT::i1, 2 },
249 { ISD::UINT_TO_FP, MVT::f64, MVT::i1, 2 },
250 { ISD::SINT_TO_FP, MVT::f32, MVT::i8, 2 },
251 { ISD::UINT_TO_FP, MVT::f32, MVT::i8, 2 },
252 { ISD::SINT_TO_FP, MVT::f64, MVT::i8, 2 },
253 { ISD::UINT_TO_FP, MVT::f64, MVT::i8, 2 },
254 { ISD::SINT_TO_FP, MVT::f32, MVT::i16, 2 },
255 { ISD::UINT_TO_FP, MVT::f32, MVT::i16, 2 },
256 { ISD::SINT_TO_FP, MVT::f64, MVT::i16, 2 },
257 { ISD::UINT_TO_FP, MVT::f64, MVT::i16, 2 },
258 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 2 },
259 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 2 },
260 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 2 },
261 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 2 },
262 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 10 },
263 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 10 },
264 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 10 },
265 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 10 }
268 if (SrcTy.isInteger() && ST->hasNEON()) {
269 int Idx = ConvertCostTableLookup<MVT>(NEONIntegerConversionTbl,
270 array_lengthof(NEONIntegerConversionTbl),
271 ISD, DstTy.getSimpleVT(),
272 SrcTy.getSimpleVT());
274 return NEONIntegerConversionTbl[Idx].Cost;
277 // Scalar integer conversion costs.
278 static const TypeConversionCostTblEntry<MVT> ARMIntegerConversionTbl[] = {
279 // i16 -> i64 requires two dependent operations.
280 { ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 },
282 // Truncates on i64 are assumed to be free.
283 { ISD::TRUNCATE, MVT::i32, MVT::i64, 0 },
284 { ISD::TRUNCATE, MVT::i16, MVT::i64, 0 },
285 { ISD::TRUNCATE, MVT::i8, MVT::i64, 0 },
286 { ISD::TRUNCATE, MVT::i1, MVT::i64, 0 }
289 if (SrcTy.isInteger()) {
291 ConvertCostTableLookup<MVT>(ARMIntegerConversionTbl,
292 array_lengthof(ARMIntegerConversionTbl),
293 ISD, DstTy.getSimpleVT(),
294 SrcTy.getSimpleVT());
296 return ARMIntegerConversionTbl[Idx].Cost;
300 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
303 unsigned ARMTTI::getVectorInstrCost(unsigned Opcode, Type *ValTy,
304 unsigned Index) const {
305 // Penalize inserting into an D-subregister.
307 Opcode == Instruction::InsertElement &&
308 ValTy->isVectorTy() &&
309 ValTy->getScalarSizeInBits() <= 32)
312 return TargetTransformInfo::getVectorInstrCost(Opcode, ValTy, Index);