xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
1 //===- AMDGPUTargetTransformInfo.h - AMDGPU specific TTI --------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file a TargetTransformInfo::Concept conforming object specific to the
11 /// AMDGPU target machine. It uses the target's detailed information to
12 /// provide more precise answers to certain TTI queries, while letting the
13 /// target independent and default TTI implementations handle the rest.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H
18 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H
19 
20 #include "AMDGPU.h"
21 #include "AMDGPUSubtarget.h"
22 #include "AMDGPUTargetMachine.h"
23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
24 #include "Utils/AMDGPUBaseInfo.h"
25 #include "llvm/ADT/ArrayRef.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/CodeGen/BasicTTIImpl.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/MC/SubtargetFeature.h"
30 #include "llvm/Support/MathExtras.h"
31 #include <cassert>
32 
33 namespace llvm {
34 
35 class AMDGPUTargetLowering;
36 class Loop;
37 class ScalarEvolution;
38 class Type;
39 class Value;
40 
41 class AMDGPUTTIImpl final : public BasicTTIImplBase<AMDGPUTTIImpl> {
42   using BaseT = BasicTTIImplBase<AMDGPUTTIImpl>;
43   using TTI = TargetTransformInfo;
44 
45   friend BaseT;
46 
47   Triple TargetTriple;
48 
49   const TargetSubtargetInfo *ST;
50   const TargetLoweringBase *TLI;
51 
52   const TargetSubtargetInfo *getST() const { return ST; }
53   const TargetLoweringBase *getTLI() const { return TLI; }
54 
55 public:
56   explicit AMDGPUTTIImpl(const AMDGPUTargetMachine *TM, const Function &F)
57       : BaseT(TM, F.getParent()->getDataLayout()),
58         TargetTriple(TM->getTargetTriple()),
59         ST(static_cast<const GCNSubtarget *>(TM->getSubtargetImpl(F))),
60         TLI(ST->getTargetLowering()) {}
61 
62   void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
63                                TTI::UnrollingPreferences &UP);
64 };
65 
66 class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> {
67   using BaseT = BasicTTIImplBase<GCNTTIImpl>;
68   using TTI = TargetTransformInfo;
69 
70   friend BaseT;
71 
72   const GCNSubtarget *ST;
73   const AMDGPUTargetLowering *TLI;
74   AMDGPUTTIImpl CommonTTI;
75   bool IsGraphicsShader;
76 
77   const FeatureBitset InlineFeatureIgnoreList = {
78     // Codegen control options which don't matter.
79     AMDGPU::FeatureEnableLoadStoreOpt,
80     AMDGPU::FeatureEnableSIScheduler,
81     AMDGPU::FeatureEnableUnsafeDSOffsetFolding,
82     AMDGPU::FeatureFlatForGlobal,
83     AMDGPU::FeaturePromoteAlloca,
84     AMDGPU::FeatureUnalignedBufferAccess,
85     AMDGPU::FeatureUnalignedScratchAccess,
86 
87     AMDGPU::FeatureAutoWaitcntBeforeBarrier,
88 
89     // Property of the kernel/environment which can't actually differ.
90     AMDGPU::FeatureSGPRInitBug,
91     AMDGPU::FeatureXNACK,
92     AMDGPU::FeatureTrapHandler,
93     AMDGPU::FeatureCodeObjectV3,
94 
95     // The default assumption needs to be ecc is enabled, but no directly
96     // exposed operations depend on it, so it can be safely inlined.
97     AMDGPU::FeatureSRAMECC,
98 
99     // Perf-tuning features
100     AMDGPU::FeatureFastFMAF32,
101     AMDGPU::HalfRate64Ops
102   };
103 
104   const GCNSubtarget *getST() const { return ST; }
105   const AMDGPUTargetLowering *getTLI() const { return TLI; }
106 
107   static inline int getFullRateInstrCost() {
108     return TargetTransformInfo::TCC_Basic;
109   }
110 
111   static inline int getHalfRateInstrCost() {
112     return 2 * TargetTransformInfo::TCC_Basic;
113   }
114 
115   // TODO: The size is usually 8 bytes, but takes 4x as many cycles. Maybe
116   // should be 2 or 4.
117   static inline int getQuarterRateInstrCost() {
118     return 3 * TargetTransformInfo::TCC_Basic;
119   }
120 
121    // On some parts, normal fp64 operations are half rate, and others
122    // quarter. This also applies to some integer operations.
123   inline int get64BitInstrCost() const {
124     return ST->hasHalfRate64Ops() ?
125       getHalfRateInstrCost() : getQuarterRateInstrCost();
126   }
127 
128 public:
129   explicit GCNTTIImpl(const AMDGPUTargetMachine *TM, const Function &F)
130     : BaseT(TM, F.getParent()->getDataLayout()),
131       ST(static_cast<const GCNSubtarget*>(TM->getSubtargetImpl(F))),
132       TLI(ST->getTargetLowering()),
133       CommonTTI(TM, F),
134       IsGraphicsShader(AMDGPU::isShader(F.getCallingConv())) {}
135 
136   bool hasBranchDivergence() { return true; }
137 
138   void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
139                                TTI::UnrollingPreferences &UP);
140 
141   TTI::PopcntSupportKind getPopcntSupport(unsigned TyWidth) {
142     assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
143     return TTI::PSK_FastHardware;
144   }
145 
146   unsigned getHardwareNumberOfRegisters(bool Vector) const;
147   unsigned getNumberOfRegisters(bool Vector) const;
148   unsigned getRegisterBitWidth(bool Vector) const;
149   unsigned getMinVectorRegisterBitWidth() const;
150   unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
151                                unsigned ChainSizeInBytes,
152                                VectorType *VecTy) const;
153   unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
154                                 unsigned ChainSizeInBytes,
155                                 VectorType *VecTy) const;
156   unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const;
157 
158   bool isLegalToVectorizeMemChain(unsigned ChainSizeInBytes,
159                                   unsigned Alignment,
160                                   unsigned AddrSpace) const;
161   bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
162                                    unsigned Alignment,
163                                    unsigned AddrSpace) const;
164   bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
165                                     unsigned Alignment,
166                                     unsigned AddrSpace) const;
167 
168   unsigned getMaxInterleaveFactor(unsigned VF);
169 
170   bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
171 
172   int getArithmeticInstrCost(
173     unsigned Opcode, Type *Ty,
174     TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
175     TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
176     TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
177     TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
178     ArrayRef<const Value *> Args = ArrayRef<const Value *>());
179 
180   unsigned getCFInstrCost(unsigned Opcode);
181 
182   int getVectorInstrCost(unsigned Opcode, Type *ValTy, unsigned Index);
183   bool isSourceOfDivergence(const Value *V) const;
184   bool isAlwaysUniform(const Value *V) const;
185 
186   unsigned getFlatAddressSpace() const {
187     // Don't bother running InferAddressSpaces pass on graphics shaders which
188     // don't use flat addressing.
189     if (IsGraphicsShader)
190       return -1;
191     return AMDGPUAS::FLAT_ADDRESS;
192   }
193 
194   bool collectFlatAddressOperands(SmallVectorImpl<int> &OpIndexes,
195                                   Intrinsic::ID IID) const;
196   bool rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
197                                         Value *OldV, Value *NewV) const;
198 
199   unsigned getVectorSplitCost() { return 0; }
200 
201   unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
202                           Type *SubTp);
203 
204   bool areInlineCompatible(const Function *Caller,
205                            const Function *Callee) const;
206 
207   unsigned getInliningThresholdMultiplier() { return 9; }
208 
209   int getInlinerVectorBonusPercent() { return 0; }
210 
211   int getArithmeticReductionCost(unsigned Opcode,
212                                  Type *Ty,
213                                  bool IsPairwise);
214   int getMinMaxReductionCost(Type *Ty, Type *CondTy,
215                              bool IsPairwiseForm,
216                              bool IsUnsigned);
217   unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands);
218 };
219 
220 class R600TTIImpl final : public BasicTTIImplBase<R600TTIImpl> {
221   using BaseT = BasicTTIImplBase<R600TTIImpl>;
222   using TTI = TargetTransformInfo;
223 
224   friend BaseT;
225 
226   const R600Subtarget *ST;
227   const AMDGPUTargetLowering *TLI;
228   AMDGPUTTIImpl CommonTTI;
229 
230 public:
231   explicit R600TTIImpl(const AMDGPUTargetMachine *TM, const Function &F)
232     : BaseT(TM, F.getParent()->getDataLayout()),
233       ST(static_cast<const R600Subtarget*>(TM->getSubtargetImpl(F))),
234       TLI(ST->getTargetLowering()),
235       CommonTTI(TM, F)	{}
236 
237   const R600Subtarget *getST() const { return ST; }
238   const AMDGPUTargetLowering *getTLI() const { return TLI; }
239 
240   void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
241                                TTI::UnrollingPreferences &UP);
242   unsigned getHardwareNumberOfRegisters(bool Vec) const;
243   unsigned getNumberOfRegisters(bool Vec) const;
244   unsigned getRegisterBitWidth(bool Vector) const;
245   unsigned getMinVectorRegisterBitWidth() const;
246   unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const;
247   bool isLegalToVectorizeMemChain(unsigned ChainSizeInBytes, unsigned Alignment,
248                                   unsigned AddrSpace) const;
249   bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
250 		                   unsigned Alignment,
251                                    unsigned AddrSpace) const;
252   bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
253                                     unsigned Alignment,
254                                     unsigned AddrSpace) const;
255   unsigned getMaxInterleaveFactor(unsigned VF);
256   unsigned getCFInstrCost(unsigned Opcode);
257   int getVectorInstrCost(unsigned Opcode, Type *ValTy, unsigned Index);
258 };
259 
260 } // end namespace llvm
261 
262 #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H
263