Bug Summary

File:include/llvm/CodeGen/BasicTTIImpl.h
Warning:line 418, column 36
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name PPCTargetTransformInfo.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-9/lib/clang/9.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/lib/Target/PowerPC -I /build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC -I /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/include -I /build/llvm-toolchain-snapshot-9~svn362543/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/9.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-9/lib/clang/9.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/lib/Target/PowerPC -fdebug-prefix-map=/build/llvm-toolchain-snapshot-9~svn362543=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2019-06-05-060531-1271-1 -x c++ /build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp -faddrsig

/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp

1//===-- PPCTargetTransformInfo.cpp - PPC specific TTI ---------------------===//
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#include "PPCTargetTransformInfo.h"
10#include "llvm/Analysis/TargetTransformInfo.h"
11#include "llvm/CodeGen/BasicTTIImpl.h"
12#include "llvm/CodeGen/CostTable.h"
13#include "llvm/CodeGen/TargetLowering.h"
14#include "llvm/Support/CommandLine.h"
15#include "llvm/Support/Debug.h"
16using namespace llvm;
17
18#define DEBUG_TYPE"ppctti" "ppctti"
19
20static cl::opt<bool> DisablePPCConstHoist("disable-ppc-constant-hoisting",
21cl::desc("disable constant hoisting on PPC"), cl::init(false), cl::Hidden);
22
23// This is currently only used for the data prefetch pass which is only enabled
24// for BG/Q by default.
25static cl::opt<unsigned>
26CacheLineSize("ppc-loop-prefetch-cache-line", cl::Hidden, cl::init(64),
27 cl::desc("The loop prefetch cache line size"));
28
29static cl::opt<bool>
30EnablePPCColdCC("ppc-enable-coldcc", cl::Hidden, cl::init(false),
31 cl::desc("Enable using coldcc calling conv for cold "
32 "internal functions"));
33
34//===----------------------------------------------------------------------===//
35//
36// PPC cost model.
37//
38//===----------------------------------------------------------------------===//
39
40TargetTransformInfo::PopcntSupportKind
41PPCTTIImpl::getPopcntSupport(unsigned TyWidth) {
42 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2")((isPowerOf2_32(TyWidth) && "Ty width must be power of 2"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(TyWidth) && \"Ty width must be power of 2\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 42, __PRETTY_FUNCTION__))
;
43 if (ST->hasPOPCNTD() != PPCSubtarget::POPCNTD_Unavailable && TyWidth <= 64)
44 return ST->hasPOPCNTD() == PPCSubtarget::POPCNTD_Slow ?
45 TTI::PSK_SlowHardware : TTI::PSK_FastHardware;
46 return TTI::PSK_Software;
47}
48
49int PPCTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
50 if (DisablePPCConstHoist)
51 return BaseT::getIntImmCost(Imm, Ty);
52
53 assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("Ty->isIntegerTy()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 53, __PRETTY_FUNCTION__))
;
54
55 unsigned BitSize = Ty->getPrimitiveSizeInBits();
56 if (BitSize == 0)
57 return ~0U;
58
59 if (Imm == 0)
60 return TTI::TCC_Free;
61
62 if (Imm.getBitWidth() <= 64) {
63 if (isInt<16>(Imm.getSExtValue()))
64 return TTI::TCC_Basic;
65
66 if (isInt<32>(Imm.getSExtValue())) {
67 // A constant that can be materialized using lis.
68 if ((Imm.getZExtValue() & 0xFFFF) == 0)
69 return TTI::TCC_Basic;
70
71 return 2 * TTI::TCC_Basic;
72 }
73 }
74
75 return 4 * TTI::TCC_Basic;
76}
77
78int PPCTTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
79 Type *Ty) {
80 if (DisablePPCConstHoist)
81 return BaseT::getIntImmCost(IID, Idx, Imm, Ty);
82
83 assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("Ty->isIntegerTy()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 83, __PRETTY_FUNCTION__))
;
84
85 unsigned BitSize = Ty->getPrimitiveSizeInBits();
86 if (BitSize == 0)
87 return ~0U;
88
89 switch (IID) {
90 default:
91 return TTI::TCC_Free;
92 case Intrinsic::sadd_with_overflow:
93 case Intrinsic::uadd_with_overflow:
94 case Intrinsic::ssub_with_overflow:
95 case Intrinsic::usub_with_overflow:
96 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<16>(Imm.getSExtValue()))
97 return TTI::TCC_Free;
98 break;
99 case Intrinsic::experimental_stackmap:
100 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
101 return TTI::TCC_Free;
102 break;
103 case Intrinsic::experimental_patchpoint_void:
104 case Intrinsic::experimental_patchpoint_i64:
105 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
106 return TTI::TCC_Free;
107 break;
108 }
109 return PPCTTIImpl::getIntImmCost(Imm, Ty);
110}
111
112int PPCTTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
113 Type *Ty) {
114 if (DisablePPCConstHoist)
115 return BaseT::getIntImmCost(Opcode, Idx, Imm, Ty);
116
117 assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail
("Ty->isIntegerTy()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 117, __PRETTY_FUNCTION__))
;
118
119 unsigned BitSize = Ty->getPrimitiveSizeInBits();
120 if (BitSize == 0)
121 return ~0U;
122
123 unsigned ImmIdx = ~0U;
124 bool ShiftedFree = false, RunFree = false, UnsignedFree = false,
125 ZeroFree = false;
126 switch (Opcode) {
127 default:
128 return TTI::TCC_Free;
129 case Instruction::GetElementPtr:
130 // Always hoist the base address of a GetElementPtr. This prevents the
131 // creation of new constants for every base constant that gets constant
132 // folded with the offset.
133 if (Idx == 0)
134 return 2 * TTI::TCC_Basic;
135 return TTI::TCC_Free;
136 case Instruction::And:
137 RunFree = true; // (for the rotate-and-mask instructions)
138 LLVM_FALLTHROUGH[[clang::fallthrough]];
139 case Instruction::Add:
140 case Instruction::Or:
141 case Instruction::Xor:
142 ShiftedFree = true;
143 LLVM_FALLTHROUGH[[clang::fallthrough]];
144 case Instruction::Sub:
145 case Instruction::Mul:
146 case Instruction::Shl:
147 case Instruction::LShr:
148 case Instruction::AShr:
149 ImmIdx = 1;
150 break;
151 case Instruction::ICmp:
152 UnsignedFree = true;
153 ImmIdx = 1;
154 // Zero comparisons can use record-form instructions.
155 LLVM_FALLTHROUGH[[clang::fallthrough]];
156 case Instruction::Select:
157 ZeroFree = true;
158 break;
159 case Instruction::PHI:
160 case Instruction::Call:
161 case Instruction::Ret:
162 case Instruction::Load:
163 case Instruction::Store:
164 break;
165 }
166
167 if (ZeroFree && Imm == 0)
168 return TTI::TCC_Free;
169
170 if (Idx == ImmIdx && Imm.getBitWidth() <= 64) {
171 if (isInt<16>(Imm.getSExtValue()))
172 return TTI::TCC_Free;
173
174 if (RunFree) {
175 if (Imm.getBitWidth() <= 32 &&
176 (isShiftedMask_32(Imm.getZExtValue()) ||
177 isShiftedMask_32(~Imm.getZExtValue())))
178 return TTI::TCC_Free;
179
180 if (ST->isPPC64() &&
181 (isShiftedMask_64(Imm.getZExtValue()) ||
182 isShiftedMask_64(~Imm.getZExtValue())))
183 return TTI::TCC_Free;
184 }
185
186 if (UnsignedFree && isUInt<16>(Imm.getZExtValue()))
187 return TTI::TCC_Free;
188
189 if (ShiftedFree && (Imm.getZExtValue() & 0xFFFF) == 0)
190 return TTI::TCC_Free;
191 }
192
193 return PPCTTIImpl::getIntImmCost(Imm, Ty);
194}
195
196unsigned PPCTTIImpl::getUserCost(const User *U,
197 ArrayRef<const Value *> Operands) {
198 if (U->getType()->isVectorTy()) {
1
Taking false branch
199 // Instructions that need to be split should cost more.
200 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, U->getType());
201 return LT.first * BaseT::getUserCost(U, Operands);
202 }
203
204 return BaseT::getUserCost(U, Operands);
2
Calling 'TargetTransformInfoImplCRTPBase::getUserCost'
205}
206
207void PPCTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
208 TTI::UnrollingPreferences &UP) {
209 if (ST->getDarwinDirective() == PPC::DIR_A2) {
210 // The A2 is in-order with a deep pipeline, and concatenation unrolling
211 // helps expose latency-hiding opportunities to the instruction scheduler.
212 UP.Partial = UP.Runtime = true;
213
214 // We unroll a lot on the A2 (hundreds of instructions), and the benefits
215 // often outweigh the cost of a division to compute the trip count.
216 UP.AllowExpensiveTripCount = true;
217 }
218
219 BaseT::getUnrollingPreferences(L, SE, UP);
220}
221
222// This function returns true to allow using coldcc calling convention.
223// Returning true results in coldcc being used for functions which are cold at
224// all call sites when the callers of the functions are not calling any other
225// non coldcc functions.
226bool PPCTTIImpl::useColdCCForColdCall(Function &F) {
227 return EnablePPCColdCC;
228}
229
230bool PPCTTIImpl::enableAggressiveInterleaving(bool LoopHasReductions) {
231 // On the A2, always unroll aggressively. For QPX unaligned loads, we depend
232 // on combining the loads generated for consecutive accesses, and failure to
233 // do so is particularly expensive. This makes it much more likely (compared
234 // to only using concatenation unrolling).
235 if (ST->getDarwinDirective() == PPC::DIR_A2)
236 return true;
237
238 return LoopHasReductions;
239}
240
241const PPCTTIImpl::TTI::MemCmpExpansionOptions *
242PPCTTIImpl::enableMemCmpExpansion(bool IsZeroCmp) const {
243 static const auto Options = []() {
244 TTI::MemCmpExpansionOptions Options;
245 Options.LoadSizes.push_back(8);
246 Options.LoadSizes.push_back(4);
247 Options.LoadSizes.push_back(2);
248 Options.LoadSizes.push_back(1);
249 return Options;
250 }();
251 return &Options;
252}
253
254bool PPCTTIImpl::enableInterleavedAccessVectorization() {
255 return true;
256}
257
258unsigned PPCTTIImpl::getNumberOfRegisters(bool Vector) {
259 if (Vector && !ST->hasAltivec() && !ST->hasQPX())
260 return 0;
261 return ST->hasVSX() ? 64 : 32;
262}
263
264unsigned PPCTTIImpl::getRegisterBitWidth(bool Vector) const {
265 if (Vector) {
266 if (ST->hasQPX()) return 256;
267 if (ST->hasAltivec()) return 128;
268 return 0;
269 }
270
271 if (ST->isPPC64())
272 return 64;
273 return 32;
274
275}
276
277unsigned PPCTTIImpl::getCacheLineSize() {
278 // Check first if the user specified a custom line size.
279 if (CacheLineSize.getNumOccurrences() > 0)
280 return CacheLineSize;
281
282 // On P7, P8 or P9 we have a cache line size of 128.
283 unsigned Directive = ST->getDarwinDirective();
284 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8 ||
285 Directive == PPC::DIR_PWR9)
286 return 128;
287
288 // On other processors return a default of 64 bytes.
289 return 64;
290}
291
292unsigned PPCTTIImpl::getPrefetchDistance() {
293 // This seems like a reasonable default for the BG/Q (this pass is enabled, by
294 // default, only on the BG/Q).
295 return 300;
296}
297
298unsigned PPCTTIImpl::getMaxInterleaveFactor(unsigned VF) {
299 unsigned Directive = ST->getDarwinDirective();
300 // The 440 has no SIMD support, but floating-point instructions
301 // have a 5-cycle latency, so unroll by 5x for latency hiding.
302 if (Directive == PPC::DIR_440)
303 return 5;
304
305 // The A2 has no SIMD support, but floating-point instructions
306 // have a 6-cycle latency, so unroll by 6x for latency hiding.
307 if (Directive == PPC::DIR_A2)
308 return 6;
309
310 // FIXME: For lack of any better information, do no harm...
311 if (Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500)
312 return 1;
313
314 // For P7 and P8, floating-point instructions have a 6-cycle latency and
315 // there are two execution units, so unroll by 12x for latency hiding.
316 // FIXME: the same for P9 as previous gen until POWER9 scheduling is ready
317 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8 ||
318 Directive == PPC::DIR_PWR9)
319 return 12;
320
321 // For most things, modern systems have two execution units (and
322 // out-of-order execution).
323 return 2;
324}
325
326// Adjust the cost of vector instructions on targets which there is overlap
327// between the vector and scalar units, thereby reducing the overall throughput
328// of vector code wrt. scalar code.
329int PPCTTIImpl::vectorCostAdjustment(int Cost, unsigned Opcode, Type *Ty1,
330 Type *Ty2) {
331 if (!ST->vectorsUseTwoUnits() || !Ty1->isVectorTy())
332 return Cost;
333
334 std::pair<int, MVT> LT1 = TLI->getTypeLegalizationCost(DL, Ty1);
335 // If type legalization involves splitting the vector, we don't want to
336 // double the cost at every step - only the last step.
337 if (LT1.first != 1 || !LT1.second.isVector())
338 return Cost;
339
340 int ISD = TLI->InstructionOpcodeToISD(Opcode);
341 if (TLI->isOperationExpand(ISD, LT1.second))
342 return Cost;
343
344 if (Ty2) {
345 std::pair<int, MVT> LT2 = TLI->getTypeLegalizationCost(DL, Ty2);
346 if (LT2.first != 1 || !LT2.second.isVector())
347 return Cost;
348 }
349
350 return Cost * 2;
351}
352
353int PPCTTIImpl::getArithmeticInstrCost(
354 unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info,
355 TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo,
356 TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args) {
357 assert(TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode")((TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode"
) ? static_cast<void> (0) : __assert_fail ("TLI->InstructionOpcodeToISD(Opcode) && \"Invalid opcode\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 357, __PRETTY_FUNCTION__))
;
358
359 // Fallback to the default implementation.
360 int Cost = BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info,
361 Opd1PropInfo, Opd2PropInfo);
362 return vectorCostAdjustment(Cost, Opcode, Ty, nullptr);
363}
364
365int PPCTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
366 Type *SubTp) {
367 // Legalize the type.
368 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
369
370 // PPC, for both Altivec/VSX and QPX, support cheap arbitrary permutations
371 // (at least in the sense that there need only be one non-loop-invariant
372 // instruction). We need one such shuffle instruction for each actual
373 // register (this is not true for arbitrary shuffles, but is true for the
374 // structured types of shuffles covered by TTI::ShuffleKind).
375 return vectorCostAdjustment(LT.first, Instruction::ShuffleVector, Tp,
376 nullptr);
377}
378
379int PPCTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
380 const Instruction *I) {
381 assert(TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode")((TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode"
) ? static_cast<void> (0) : __assert_fail ("TLI->InstructionOpcodeToISD(Opcode) && \"Invalid opcode\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 381, __PRETTY_FUNCTION__))
;
382
383 int Cost = BaseT::getCastInstrCost(Opcode, Dst, Src);
384 return vectorCostAdjustment(Cost, Opcode, Dst, Src);
385}
386
387int PPCTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
388 const Instruction *I) {
389 int Cost = BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
390 return vectorCostAdjustment(Cost, Opcode, ValTy, nullptr);
391}
392
393int PPCTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
394 assert(Val->isVectorTy() && "This must be a vector type")((Val->isVectorTy() && "This must be a vector type"
) ? static_cast<void> (0) : __assert_fail ("Val->isVectorTy() && \"This must be a vector type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 394, __PRETTY_FUNCTION__))
;
395
396 int ISD = TLI->InstructionOpcodeToISD(Opcode);
397 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 397, __PRETTY_FUNCTION__))
;
398
399 int Cost = BaseT::getVectorInstrCost(Opcode, Val, Index);
400 Cost = vectorCostAdjustment(Cost, Opcode, Val, nullptr);
401
402 if (ST->hasVSX() && Val->getScalarType()->isDoubleTy()) {
403 // Double-precision scalars are already located in index #0 (or #1 if LE).
404 if (ISD == ISD::EXTRACT_VECTOR_ELT &&
405 Index == (ST->isLittleEndian() ? 1 : 0))
406 return 0;
407
408 return Cost;
409
410 } else if (ST->hasQPX() && Val->getScalarType()->isFloatingPointTy()) {
411 // Floating point scalars are already located in index #0.
412 if (Index == 0)
413 return 0;
414
415 return Cost;
416 }
417
418 // Estimated cost of a load-hit-store delay. This was obtained
419 // experimentally as a minimum needed to prevent unprofitable
420 // vectorization for the paq8p benchmark. It may need to be
421 // raised further if other unprofitable cases remain.
422 unsigned LHSPenalty = 2;
423 if (ISD == ISD::INSERT_VECTOR_ELT)
424 LHSPenalty += 7;
425
426 // Vector element insert/extract with Altivec is very expensive,
427 // because they require store and reload with the attendant
428 // processor stall for load-hit-store. Until VSX is available,
429 // these need to be estimated as very costly.
430 if (ISD == ISD::EXTRACT_VECTOR_ELT ||
431 ISD == ISD::INSERT_VECTOR_ELT)
432 return LHSPenalty + Cost;
433
434 return Cost;
435}
436
437int PPCTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
438 unsigned AddressSpace, const Instruction *I) {
439 // Legalize the type.
440 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
441 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&(((Opcode == Instruction::Load || Opcode == Instruction::Store
) && "Invalid Opcode") ? static_cast<void> (0) :
__assert_fail ("(Opcode == Instruction::Load || Opcode == Instruction::Store) && \"Invalid Opcode\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 442, __PRETTY_FUNCTION__))
442 "Invalid Opcode")(((Opcode == Instruction::Load || Opcode == Instruction::Store
) && "Invalid Opcode") ? static_cast<void> (0) :
__assert_fail ("(Opcode == Instruction::Load || Opcode == Instruction::Store) && \"Invalid Opcode\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 442, __PRETTY_FUNCTION__))
;
443
444 int Cost = BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
445 Cost = vectorCostAdjustment(Cost, Opcode, Src, nullptr);
446
447 bool IsAltivecType = ST->hasAltivec() &&
448 (LT.second == MVT::v16i8 || LT.second == MVT::v8i16 ||
449 LT.second == MVT::v4i32 || LT.second == MVT::v4f32);
450 bool IsVSXType = ST->hasVSX() &&
451 (LT.second == MVT::v2f64 || LT.second == MVT::v2i64);
452 bool IsQPXType = ST->hasQPX() &&
453 (LT.second == MVT::v4f64 || LT.second == MVT::v4f32);
454
455 // VSX has 32b/64b load instructions. Legalization can handle loading of
456 // 32b/64b to VSR correctly and cheaply. But BaseT::getMemoryOpCost and
457 // PPCTargetLowering can't compute the cost appropriately. So here we
458 // explicitly check this case.
459 unsigned MemBytes = Src->getPrimitiveSizeInBits();
460 if (Opcode == Instruction::Load && ST->hasVSX() && IsAltivecType &&
461 (MemBytes == 64 || (ST->hasP8Vector() && MemBytes == 32)))
462 return 1;
463
464 // Aligned loads and stores are easy.
465 unsigned SrcBytes = LT.second.getStoreSize();
466 if (!SrcBytes || !Alignment || Alignment >= SrcBytes)
467 return Cost;
468
469 // If we can use the permutation-based load sequence, then this is also
470 // relatively cheap (not counting loop-invariant instructions): one load plus
471 // one permute (the last load in a series has extra cost, but we're
472 // neglecting that here). Note that on the P7, we could do unaligned loads
473 // for Altivec types using the VSX instructions, but that's more expensive
474 // than using the permutation-based load sequence. On the P8, that's no
475 // longer true.
476 if (Opcode == Instruction::Load &&
477 ((!ST->hasP8Vector() && IsAltivecType) || IsQPXType) &&
478 Alignment >= LT.second.getScalarType().getStoreSize())
479 return Cost + LT.first; // Add the cost of the permutations.
480
481 // For VSX, we can do unaligned loads and stores on Altivec/VSX types. On the
482 // P7, unaligned vector loads are more expensive than the permutation-based
483 // load sequence, so that might be used instead, but regardless, the net cost
484 // is about the same (not counting loop-invariant instructions).
485 if (IsVSXType || (ST->hasVSX() && IsAltivecType))
486 return Cost;
487
488 // Newer PPC supports unaligned memory access.
489 if (TLI->allowsMisalignedMemoryAccesses(LT.second, 0))
490 return Cost;
491
492 // PPC in general does not support unaligned loads and stores. They'll need
493 // to be decomposed based on the alignment factor.
494
495 // Add the cost of each scalar load or store.
496 Cost += LT.first*(SrcBytes/Alignment-1);
497
498 // For a vector type, there is also scalarization overhead (only for
499 // stores, loads are expanded using the vector-load + permutation sequence,
500 // which is much less expensive).
501 if (Src->isVectorTy() && Opcode == Instruction::Store)
502 for (int i = 0, e = Src->getVectorNumElements(); i < e; ++i)
503 Cost += getVectorInstrCost(Instruction::ExtractElement, Src, i);
504
505 return Cost;
506}
507
508int PPCTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
509 unsigned Factor,
510 ArrayRef<unsigned> Indices,
511 unsigned Alignment,
512 unsigned AddressSpace,
513 bool UseMaskForCond,
514 bool UseMaskForGaps) {
515 if (UseMaskForCond || UseMaskForGaps)
516 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
517 Alignment, AddressSpace,
518 UseMaskForCond, UseMaskForGaps);
519
520 assert(isa<VectorType>(VecTy) &&((isa<VectorType>(VecTy) && "Expect a vector type for interleaved memory op"
) ? static_cast<void> (0) : __assert_fail ("isa<VectorType>(VecTy) && \"Expect a vector type for interleaved memory op\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 521, __PRETTY_FUNCTION__))
521 "Expect a vector type for interleaved memory op")((isa<VectorType>(VecTy) && "Expect a vector type for interleaved memory op"
) ? static_cast<void> (0) : __assert_fail ("isa<VectorType>(VecTy) && \"Expect a vector type for interleaved memory op\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Target/PowerPC/PPCTargetTransformInfo.cpp"
, 521, __PRETTY_FUNCTION__))
;
522
523 // Legalize the type.
524 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, VecTy);
525
526 // Firstly, the cost of load/store operation.
527 int Cost = getMemoryOpCost(Opcode, VecTy, Alignment, AddressSpace);
528
529 // PPC, for both Altivec/VSX and QPX, support cheap arbitrary permutations
530 // (at least in the sense that there need only be one non-loop-invariant
531 // instruction). For each result vector, we need one shuffle per incoming
532 // vector (except that the first shuffle can take two incoming vectors
533 // because it does not need to take itself).
534 Cost += Factor*(LT.first-1);
535
536 return Cost;
537}
538

/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h

1//===- TargetTransformInfoImpl.h --------------------------------*- 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/// \file
9/// This file provides helpers for the implementation of
10/// a TargetTransformInfo-conforming class.
11///
12//===----------------------------------------------------------------------===//
13
14#ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
15#define LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
16
17#include "llvm/Analysis/ScalarEvolutionExpressions.h"
18#include "llvm/Analysis/TargetTransformInfo.h"
19#include "llvm/Analysis/VectorUtils.h"
20#include "llvm/IR/CallSite.h"
21#include "llvm/IR/DataLayout.h"
22#include "llvm/IR/Function.h"
23#include "llvm/IR/GetElementPtrTypeIterator.h"
24#include "llvm/IR/Operator.h"
25#include "llvm/IR/Type.h"
26
27namespace llvm {
28
29/// Base class for use as a mix-in that aids implementing
30/// a TargetTransformInfo-compatible class.
31class TargetTransformInfoImplBase {
32protected:
33 typedef TargetTransformInfo TTI;
34
35 const DataLayout &DL;
36
37 explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {}
38
39public:
40 // Provide value semantics. MSVC requires that we spell all of these out.
41 TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg)
42 : DL(Arg.DL) {}
43 TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {}
44
45 const DataLayout &getDataLayout() const { return DL; }
46
47 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
48 switch (Opcode) {
49 default:
50 // By default, just classify everything as 'basic'.
51 return TTI::TCC_Basic;
52
53 case Instruction::GetElementPtr:
54 llvm_unreachable("Use getGEPCost for GEP operations!")::llvm::llvm_unreachable_internal("Use getGEPCost for GEP operations!"
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 54)
;
55
56 case Instruction::BitCast:
57 assert(OpTy && "Cast instructions must provide the operand type")((OpTy && "Cast instructions must provide the operand type"
) ? static_cast<void> (0) : __assert_fail ("OpTy && \"Cast instructions must provide the operand type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 57, __PRETTY_FUNCTION__))
;
58 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
59 // Identity and pointer-to-pointer casts are free.
60 return TTI::TCC_Free;
61
62 // Otherwise, the default basic cost is used.
63 return TTI::TCC_Basic;
64
65 case Instruction::FDiv:
66 case Instruction::FRem:
67 case Instruction::SDiv:
68 case Instruction::SRem:
69 case Instruction::UDiv:
70 case Instruction::URem:
71 return TTI::TCC_Expensive;
72
73 case Instruction::IntToPtr: {
74 // An inttoptr cast is free so long as the input is a legal integer type
75 // which doesn't contain values outside the range of a pointer.
76 unsigned OpSize = OpTy->getScalarSizeInBits();
77 if (DL.isLegalInteger(OpSize) &&
78 OpSize <= DL.getPointerTypeSizeInBits(Ty))
79 return TTI::TCC_Free;
80
81 // Otherwise it's not a no-op.
82 return TTI::TCC_Basic;
83 }
84 case Instruction::PtrToInt: {
85 // A ptrtoint cast is free so long as the result is large enough to store
86 // the pointer, and a legal integer type.
87 unsigned DestSize = Ty->getScalarSizeInBits();
88 if (DL.isLegalInteger(DestSize) &&
89 DestSize >= DL.getPointerTypeSizeInBits(OpTy))
90 return TTI::TCC_Free;
91
92 // Otherwise it's not a no-op.
93 return TTI::TCC_Basic;
94 }
95 case Instruction::Trunc:
96 // trunc to a native type is free (assuming the target has compare and
97 // shift-right of the same width).
98 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty)))
99 return TTI::TCC_Free;
100
101 return TTI::TCC_Basic;
102 }
103 }
104
105 int getGEPCost(Type *PointeeType, const Value *Ptr,
106 ArrayRef<const Value *> Operands) {
107 // In the basic model, we just assume that all-constant GEPs will be folded
108 // into their uses via addressing modes.
109 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
110 if (!isa<Constant>(Operands[Idx]))
111 return TTI::TCC_Basic;
112
113 return TTI::TCC_Free;
114 }
115
116 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
117 unsigned &JTSize) {
118 JTSize = 0;
119 return SI.getNumCases();
120 }
121
122 int getExtCost(const Instruction *I, const Value *Src) {
123 return TTI::TCC_Basic;
124 }
125
126 unsigned getCallCost(FunctionType *FTy, int NumArgs, const User *U) {
127 assert(FTy && "FunctionType must be provided to this routine.")((FTy && "FunctionType must be provided to this routine."
) ? static_cast<void> (0) : __assert_fail ("FTy && \"FunctionType must be provided to this routine.\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 127, __PRETTY_FUNCTION__))
;
128
129 // The target-independent implementation just measures the size of the
130 // function by approximating that each argument will take on average one
131 // instruction to prepare.
132
133 if (NumArgs < 0)
134 // Set the argument number to the number of explicit arguments in the
135 // function.
136 NumArgs = FTy->getNumParams();
137
138 return TTI::TCC_Basic * (NumArgs + 1);
139 }
140
141 unsigned getInliningThresholdMultiplier() { return 1; }
142
143 unsigned getMemcpyCost(const Instruction *I) {
144 return TTI::TCC_Expensive;
145 }
146
147 bool hasBranchDivergence() { return false; }
148
149 bool isSourceOfDivergence(const Value *V) { return false; }
150
151 bool isAlwaysUniform(const Value *V) { return false; }
152
153 unsigned getFlatAddressSpace () {
154 return -1;
155 }
156
157 bool isLoweredToCall(const Function *F) {
158 assert(F && "A concrete function must be provided to this routine.")((F && "A concrete function must be provided to this routine."
) ? static_cast<void> (0) : __assert_fail ("F && \"A concrete function must be provided to this routine.\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 158, __PRETTY_FUNCTION__))
;
159
160 // FIXME: These should almost certainly not be handled here, and instead
161 // handled with the help of TLI or the target itself. This was largely
162 // ported from existing analysis heuristics here so that such refactorings
163 // can take place in the future.
164
165 if (F->isIntrinsic())
166 return false;
167
168 if (F->hasLocalLinkage() || !F->hasName())
169 return true;
170
171 StringRef Name = F->getName();
172
173 // These will all likely lower to a single selection DAG node.
174 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
175 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
176 Name == "fmin" || Name == "fminf" || Name == "fminl" ||
177 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
178 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
179 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
180 return false;
181
182 // These are all likely to be optimized into something smaller.
183 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
184 Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
185 Name == "floorf" || Name == "ceil" || Name == "round" ||
186 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
187 Name == "llabs")
188 return false;
189
190 return true;
191 }
192
193 void getUnrollingPreferences(Loop *, ScalarEvolution &,
194 TTI::UnrollingPreferences &) {}
195
196 bool isLegalAddImmediate(int64_t Imm) { return false; }
197
198 bool isLegalICmpImmediate(int64_t Imm) { return false; }
199
200 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
201 bool HasBaseReg, int64_t Scale,
202 unsigned AddrSpace, Instruction *I = nullptr) {
203 // Guess that only reg and reg+reg addressing is allowed. This heuristic is
204 // taken from the implementation of LSR.
205 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
206 }
207
208 bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2) {
209 return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds,
210 C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
211 std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds,
212 C2.ScaleCost, C2.ImmCost, C2.SetupCost);
213 }
214
215 bool canMacroFuseCmp() { return false; }
216
217 bool shouldFavorPostInc() const { return false; }
218
219 bool shouldFavorBackedgeIndex(const Loop *L) const { return false; }
220
221 bool isLegalMaskedStore(Type *DataType) { return false; }
222
223 bool isLegalMaskedLoad(Type *DataType) { return false; }
224
225 bool isLegalMaskedScatter(Type *DataType) { return false; }
226
227 bool isLegalMaskedGather(Type *DataType) { return false; }
228
229 bool isLegalMaskedCompressStore(Type *DataType) { return false; }
230
231 bool isLegalMaskedExpandLoad(Type *DataType) { return false; }
232
233 bool hasDivRemOp(Type *DataType, bool IsSigned) { return false; }
234
235 bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { return false; }
236
237 bool prefersVectorizedAddressing() { return true; }
238
239 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
240 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
241 // Guess that all legal addressing mode are free.
242 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
243 Scale, AddrSpace))
244 return 0;
245 return -1;
246 }
247
248 bool LSRWithInstrQueries() { return false; }
249
250 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
251
252 bool isProfitableToHoist(Instruction *I) { return true; }
253
254 bool useAA() { return false; }
255
256 bool isTypeLegal(Type *Ty) { return false; }
257
258 unsigned getJumpBufAlignment() { return 0; }
259
260 unsigned getJumpBufSize() { return 0; }
261
262 bool shouldBuildLookupTables() { return true; }
263 bool shouldBuildLookupTablesForConstant(Constant *C) { return true; }
264
265 bool useColdCCForColdCall(Function &F) { return false; }
266
267 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
268 return 0;
269 }
270
271 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
272 unsigned VF) { return 0; }
273
274 bool supportsEfficientVectorElementLoadStore() { return false; }
275
276 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
277
278 const TTI::MemCmpExpansionOptions *enableMemCmpExpansion(
279 bool IsZeroCmp) const {
280 return nullptr;
281 }
282
283 bool enableInterleavedAccessVectorization() { return false; }
284
285 bool enableMaskedInterleavedAccessVectorization() { return false; }
286
287 bool isFPVectorizationPotentiallyUnsafe() { return false; }
288
289 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
290 unsigned BitWidth,
291 unsigned AddressSpace,
292 unsigned Alignment,
293 bool *Fast) { return false; }
294
295 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
296 return TTI::PSK_Software;
297 }
298
299 bool haveFastSqrt(Type *Ty) { return false; }
300
301 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; }
302
303 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
304
305 int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
306 Type *Ty) {
307 return 0;
308 }
309
310 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
311
312 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
313 Type *Ty) {
314 return TTI::TCC_Free;
315 }
316
317 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
318 Type *Ty) {
319 return TTI::TCC_Free;
320 }
321
322 unsigned getNumberOfRegisters(bool Vector) { return 8; }
323
324 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
325
326 unsigned getMinVectorRegisterBitWidth() { return 128; }
327
328 bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
329
330 unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
331
332 bool
333 shouldConsiderAddressTypePromotion(const Instruction &I,
334 bool &AllowPromotionWithoutCommonHeader) {
335 AllowPromotionWithoutCommonHeader = false;
336 return false;
337 }
338
339 unsigned getCacheLineSize() { return 0; }
340
341 llvm::Optional<unsigned> getCacheSize(TargetTransformInfo::CacheLevel Level) {
342 switch (Level) {
343 case TargetTransformInfo::CacheLevel::L1D:
344 LLVM_FALLTHROUGH[[clang::fallthrough]];
345 case TargetTransformInfo::CacheLevel::L2D:
346 return llvm::Optional<unsigned>();
347 }
348
349 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel")::llvm::llvm_unreachable_internal("Unknown TargetTransformInfo::CacheLevel"
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 349)
;
350 }
351
352 llvm::Optional<unsigned> getCacheAssociativity(
353 TargetTransformInfo::CacheLevel Level) {
354 switch (Level) {
355 case TargetTransformInfo::CacheLevel::L1D:
356 LLVM_FALLTHROUGH[[clang::fallthrough]];
357 case TargetTransformInfo::CacheLevel::L2D:
358 return llvm::Optional<unsigned>();
359 }
360
361 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel")::llvm::llvm_unreachable_internal("Unknown TargetTransformInfo::CacheLevel"
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 361)
;
362 }
363
364 unsigned getPrefetchDistance() { return 0; }
365
366 unsigned getMinPrefetchStride() { return 1; }
367
368 unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX(2147483647 *2U +1U); }
369
370 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
371
372 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
373 TTI::OperandValueKind Opd1Info,
374 TTI::OperandValueKind Opd2Info,
375 TTI::OperandValueProperties Opd1PropInfo,
376 TTI::OperandValueProperties Opd2PropInfo,
377 ArrayRef<const Value *> Args) {
378 return 1;
379 }
380
381 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
382 Type *SubTp) {
383 return 1;
384 }
385
386 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
387 const Instruction *I) { return 1; }
388
389 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
390 VectorType *VecTy, unsigned Index) {
391 return 1;
392 }
393
394 unsigned getCFInstrCost(unsigned Opcode) { return 1; }
395
396 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
397 const Instruction *I) {
398 return 1;
399 }
400
401 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
402 return 1;
403 }
404
405 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
406 unsigned AddressSpace, const Instruction *I) {
407 return 1;
408 }
409
410 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
411 unsigned AddressSpace) {
412 return 1;
413 }
414
415 unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
416 bool VariableMask,
417 unsigned Alignment) {
418 return 1;
419 }
420
421 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
422 unsigned Factor,
423 ArrayRef<unsigned> Indices,
424 unsigned Alignment, unsigned AddressSpace,
425 bool UseMaskForCond = false,
426 bool UseMaskForGaps = false) {
427 return 1;
428 }
429
430 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
431 ArrayRef<Type *> Tys, FastMathFlags FMF,
432 unsigned ScalarizationCostPassed) {
433 return 1;
434 }
435 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
436 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
437 return 1;
438 }
439
440 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
441 return 1;
442 }
443
444 unsigned getNumberOfParts(Type *Tp) { return 0; }
445
446 unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *,
447 const SCEV *) {
448 return 0;
449 }
450
451 unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; }
452
453 unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; }
454
455 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
456
457 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
458 return false;
459 }
460
461 unsigned getAtomicMemIntrinsicMaxElementSize() const {
462 // Note for overrides: You must ensure for all element unordered-atomic
463 // memory intrinsics that all power-of-2 element sizes up to, and
464 // including, the return value of this method have a corresponding
465 // runtime lib call. These runtime lib call definitions can be found
466 // in RuntimeLibcalls.h
467 return 0;
468 }
469
470 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
471 Type *ExpectedType) {
472 return nullptr;
473 }
474
475 Type *getMemcpyLoopLoweringType(LLVMContext &Context, Value *Length,
476 unsigned SrcAlign, unsigned DestAlign) const {
477 return Type::getInt8Ty(Context);
478 }
479
480 void getMemcpyLoopResidualLoweringType(SmallVectorImpl<Type *> &OpsOut,
481 LLVMContext &Context,
482 unsigned RemainingBytes,
483 unsigned SrcAlign,
484 unsigned DestAlign) const {
485 for (unsigned i = 0; i != RemainingBytes; ++i)
486 OpsOut.push_back(Type::getInt8Ty(Context));
487 }
488
489 bool areInlineCompatible(const Function *Caller,
490 const Function *Callee) const {
491 return (Caller->getFnAttribute("target-cpu") ==
492 Callee->getFnAttribute("target-cpu")) &&
493 (Caller->getFnAttribute("target-features") ==
494 Callee->getFnAttribute("target-features"));
495 }
496
497 bool areFunctionArgsABICompatible(const Function *Caller, const Function *Callee,
498 SmallPtrSetImpl<Argument *> &Args) const {
499 return (Caller->getFnAttribute("target-cpu") ==
500 Callee->getFnAttribute("target-cpu")) &&
501 (Caller->getFnAttribute("target-features") ==
502 Callee->getFnAttribute("target-features"));
503 }
504
505 bool isIndexedLoadLegal(TTI::MemIndexedMode Mode, Type *Ty,
506 const DataLayout &DL) const {
507 return false;
508 }
509
510 bool isIndexedStoreLegal(TTI::MemIndexedMode Mode, Type *Ty,
511 const DataLayout &DL) const {
512 return false;
513 }
514
515 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; }
516
517 bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; }
518
519 bool isLegalToVectorizeStore(StoreInst *SI) const { return true; }
520
521 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
522 unsigned Alignment,
523 unsigned AddrSpace) const {
524 return true;
525 }
526
527 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
528 unsigned Alignment,
529 unsigned AddrSpace) const {
530 return true;
531 }
532
533 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
534 unsigned ChainSizeInBytes,
535 VectorType *VecTy) const {
536 return VF;
537 }
538
539 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
540 unsigned ChainSizeInBytes,
541 VectorType *VecTy) const {
542 return VF;
543 }
544
545 bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
546 TTI::ReductionFlags Flags) const {
547 return false;
548 }
549
550 bool shouldExpandReduction(const IntrinsicInst *II) const {
551 return true;
552 }
553
554protected:
555 // Obtain the minimum required size to hold the value (without the sign)
556 // In case of a vector it returns the min required size for one element.
557 unsigned minRequiredElementSize(const Value* Val, bool &isSigned) {
558 if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) {
559 const auto* VectorValue = cast<Constant>(Val);
560
561 // In case of a vector need to pick the max between the min
562 // required size for each element
563 auto *VT = cast<VectorType>(Val->getType());
564
565 // Assume unsigned elements
566 isSigned = false;
567
568 // The max required size is the total vector width divided by num
569 // of elements in the vector
570 unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements();
571
572 unsigned MinRequiredSize = 0;
573 for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) {
574 if (auto* IntElement =
575 dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) {
576 bool signedElement = IntElement->getValue().isNegative();
577 // Get the element min required size.
578 unsigned ElementMinRequiredSize =
579 IntElement->getValue().getMinSignedBits() - 1;
580 // In case one element is signed then all the vector is signed.
581 isSigned |= signedElement;
582 // Save the max required bit size between all the elements.
583 MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize);
584 }
585 else {
586 // not an int constant element
587 return MaxRequiredSize;
588 }
589 }
590 return MinRequiredSize;
591 }
592
593 if (const auto* CI = dyn_cast<ConstantInt>(Val)) {
594 isSigned = CI->getValue().isNegative();
595 return CI->getValue().getMinSignedBits() - 1;
596 }
597
598 if (const auto* Cast = dyn_cast<SExtInst>(Val)) {
599 isSigned = true;
600 return Cast->getSrcTy()->getScalarSizeInBits() - 1;
601 }
602
603 if (const auto* Cast = dyn_cast<ZExtInst>(Val)) {
604 isSigned = false;
605 return Cast->getSrcTy()->getScalarSizeInBits();
606 }
607
608 isSigned = false;
609 return Val->getType()->getScalarSizeInBits();
610 }
611
612 bool isStridedAccess(const SCEV *Ptr) {
613 return Ptr && isa<SCEVAddRecExpr>(Ptr);
614 }
615
616 const SCEVConstant *getConstantStrideStep(ScalarEvolution *SE,
617 const SCEV *Ptr) {
618 if (!isStridedAccess(Ptr))
619 return nullptr;
620 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr);
621 return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE));
622 }
623
624 bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr,
625 int64_t MergeDistance) {
626 const SCEVConstant *Step = getConstantStrideStep(SE, Ptr);
627 if (!Step)
628 return false;
629 APInt StrideVal = Step->getAPInt();
630 if (StrideVal.getBitWidth() > 64)
631 return false;
632 // FIXME: Need to take absolute value for negative stride case.
633 return StrideVal.getSExtValue() < MergeDistance;
634 }
635};
636
637/// CRTP base class for use as a mix-in that aids implementing
638/// a TargetTransformInfo-compatible class.
639template <typename T>
640class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
641private:
642 typedef TargetTransformInfoImplBase BaseT;
643
644protected:
645 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
646
647public:
648 using BaseT::getCallCost;
649
650 unsigned getCallCost(const Function *F, int NumArgs, const User *U) {
651 assert(F && "A concrete function must be provided to this routine.")((F && "A concrete function must be provided to this routine."
) ? static_cast<void> (0) : __assert_fail ("F && \"A concrete function must be provided to this routine.\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 651, __PRETTY_FUNCTION__))
;
652
653 if (NumArgs < 0)
654 // Set the argument number to the number of explicit arguments in the
655 // function.
656 NumArgs = F->arg_size();
657
658 if (Intrinsic::ID IID = F->getIntrinsicID()) {
659 FunctionType *FTy = F->getFunctionType();
660 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
661 return static_cast<T *>(this)
662 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys, U);
663 }
664
665 if (!static_cast<T *>(this)->isLoweredToCall(F))
666 return TTI::TCC_Basic; // Give a basic cost if it will be lowered
667 // directly.
668
669 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs, U);
670 }
671
672 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments,
673 const User *U) {
674 // Simply delegate to generic handling of the call.
675 // FIXME: We should use instsimplify or something else to catch calls which
676 // will constant fold with these arguments.
677 return static_cast<T *>(this)->getCallCost(F, Arguments.size(), U);
678 }
679
680 using BaseT::getGEPCost;
681
682 int getGEPCost(Type *PointeeType, const Value *Ptr,
683 ArrayRef<const Value *> Operands) {
684 const GlobalValue *BaseGV = nullptr;
685 if (Ptr != nullptr) {
686 // TODO: will remove this when pointers have an opaque type.
687 assert(Ptr->getType()->getScalarType()->getPointerElementType() ==((Ptr->getType()->getScalarType()->getPointerElementType
() == PointeeType && "explicit pointee type doesn't match operand's pointee type"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->getScalarType()->getPointerElementType() == PointeeType && \"explicit pointee type doesn't match operand's pointee type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 689, __PRETTY_FUNCTION__))
688 PointeeType &&((Ptr->getType()->getScalarType()->getPointerElementType
() == PointeeType && "explicit pointee type doesn't match operand's pointee type"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->getScalarType()->getPointerElementType() == PointeeType && \"explicit pointee type doesn't match operand's pointee type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 689, __PRETTY_FUNCTION__))
689 "explicit pointee type doesn't match operand's pointee type")((Ptr->getType()->getScalarType()->getPointerElementType
() == PointeeType && "explicit pointee type doesn't match operand's pointee type"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->getScalarType()->getPointerElementType() == PointeeType && \"explicit pointee type doesn't match operand's pointee type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 689, __PRETTY_FUNCTION__))
;
690 BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
691 }
692 bool HasBaseReg = (BaseGV == nullptr);
693
694 auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType());
695 APInt BaseOffset(PtrSizeBits, 0);
696 int64_t Scale = 0;
697
698 auto GTI = gep_type_begin(PointeeType, Operands);
699 Type *TargetType = nullptr;
700
701 // Handle the case where the GEP instruction has a single operand,
702 // the basis, therefore TargetType is a nullptr.
703 if (Operands.empty())
704 return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic;
705
706 for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
707 TargetType = GTI.getIndexedType();
708 // We assume that the cost of Scalar GEP with constant index and the
709 // cost of Vector GEP with splat constant index are the same.
710 const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
711 if (!ConstIdx)
712 if (auto Splat = getSplatValue(*I))
713 ConstIdx = dyn_cast<ConstantInt>(Splat);
714 if (StructType *STy = GTI.getStructTypeOrNull()) {
715 // For structures the index is always splat or scalar constant
716 assert(ConstIdx && "Unexpected GEP index")((ConstIdx && "Unexpected GEP index") ? static_cast<
void> (0) : __assert_fail ("ConstIdx && \"Unexpected GEP index\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 716, __PRETTY_FUNCTION__))
;
717 uint64_t Field = ConstIdx->getZExtValue();
718 BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
719 } else {
720 int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
721 if (ConstIdx) {
722 BaseOffset +=
723 ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize;
724 } else {
725 // Needs scale register.
726 if (Scale != 0)
727 // No addressing mode takes two scale registers.
728 return TTI::TCC_Basic;
729 Scale = ElementSize;
730 }
731 }
732 }
733
734 // Assumes the address space is 0 when Ptr is nullptr.
735 unsigned AS =
736 (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
737
738 if (static_cast<T *>(this)->isLegalAddressingMode(
739 TargetType, const_cast<GlobalValue *>(BaseGV),
740 BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale, AS))
741 return TTI::TCC_Free;
742 return TTI::TCC_Basic;
743 }
744
745 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
746 ArrayRef<Type *> ParamTys, const User *U) {
747 switch (IID) {
748 default:
749 // Intrinsics rarely (if ever) have normal argument setup constraints.
750 // Model them as having a basic instruction cost.
751 return TTI::TCC_Basic;
752
753 // TODO: other libc intrinsics.
754 case Intrinsic::memcpy:
755 return static_cast<T *>(this)->getMemcpyCost(dyn_cast<Instruction>(U));
756
757 case Intrinsic::annotation:
758 case Intrinsic::assume:
759 case Intrinsic::sideeffect:
760 case Intrinsic::dbg_declare:
761 case Intrinsic::dbg_value:
762 case Intrinsic::dbg_label:
763 case Intrinsic::invariant_start:
764 case Intrinsic::invariant_end:
765 case Intrinsic::launder_invariant_group:
766 case Intrinsic::strip_invariant_group:
767 case Intrinsic::is_constant:
768 case Intrinsic::lifetime_start:
769 case Intrinsic::lifetime_end:
770 case Intrinsic::objectsize:
771 case Intrinsic::ptr_annotation:
772 case Intrinsic::var_annotation:
773 case Intrinsic::experimental_gc_result:
774 case Intrinsic::experimental_gc_relocate:
775 case Intrinsic::coro_alloc:
776 case Intrinsic::coro_begin:
777 case Intrinsic::coro_free:
778 case Intrinsic::coro_end:
779 case Intrinsic::coro_frame:
780 case Intrinsic::coro_size:
781 case Intrinsic::coro_suspend:
782 case Intrinsic::coro_param:
783 case Intrinsic::coro_subfn_addr:
784 // These intrinsics don't actually represent code after lowering.
785 return TTI::TCC_Free;
786 }
787 }
788
789 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
790 ArrayRef<const Value *> Arguments, const User *U) {
791 // Delegate to the generic intrinsic handling code. This mostly provides an
792 // opportunity for targets to (for example) special case the cost of
793 // certain intrinsics based on constants used as arguments.
794 SmallVector<Type *, 8> ParamTys;
795 ParamTys.reserve(Arguments.size());
796 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
797 ParamTys.push_back(Arguments[Idx]->getType());
798 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys, U);
799 }
800
801 unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) {
802 if (isa<PHINode>(U))
3
Taking false branch
803 return TTI::TCC_Free; // Model all PHI nodes as free.
804
805 // Static alloca doesn't generate target instructions.
806 if (auto *A = dyn_cast<AllocaInst>(U))
4
Taking false branch
807 if (A->isStaticAlloca())
808 return TTI::TCC_Free;
809
810 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
5
Taking false branch
811 return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(),
812 GEP->getPointerOperand(),
813 Operands.drop_front());
814 }
815
816 if (auto CS = ImmutableCallSite(U)) {
6
Taking false branch
817 const Function *F = CS.getCalledFunction();
818 if (!F) {
819 // Just use the called value type.
820 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
821 return static_cast<T *>(this)
822 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size(), U);
823 }
824
825 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
826 return static_cast<T *>(this)->getCallCost(F, Arguments, U);
827 }
828
829 if (isa<SExtInst>(U) || isa<ZExtInst>(U) || isa<FPExtInst>(U))
7
Taking false branch
830 // The old behaviour of generally treating extensions of icmp to be free
831 // has been removed. A target that needs it should override getUserCost().
832 return static_cast<T *>(this)->getExtCost(cast<Instruction>(U),
833 Operands.back());
834
835 return static_cast<T *>(this)->getOperationCost(
11
Calling 'BasicTTIImplBase::getOperationCost'
836 Operator::getOpcode(U), U->getType(),
837 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
8
Assuming the condition is false
9
'?' condition is false
10
Passing null pointer value via 3rd parameter 'OpTy'
838 }
839
840 int getInstructionLatency(const Instruction *I) {
841 SmallVector<const Value *, 4> Operands(I->value_op_begin(),
842 I->value_op_end());
843 if (getUserCost(I, Operands) == TTI::TCC_Free)
844 return 0;
845
846 if (isa<LoadInst>(I))
847 return 4;
848
849 Type *DstTy = I->getType();
850
851 // Usually an intrinsic is a simple instruction.
852 // A real function call is much slower.
853 if (auto *CI = dyn_cast<CallInst>(I)) {
854 const Function *F = CI->getCalledFunction();
855 if (!F || static_cast<T *>(this)->isLoweredToCall(F))
856 return 40;
857 // Some intrinsics return a value and a flag, we use the value type
858 // to decide its latency.
859 if (StructType* StructTy = dyn_cast<StructType>(DstTy))
860 DstTy = StructTy->getElementType(0);
861 // Fall through to simple instructions.
862 }
863
864 if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
865 DstTy = VectorTy->getElementType();
866 if (DstTy->isFloatingPointTy())
867 return 3;
868
869 return 1;
870 }
871};
872}
873
874#endif

/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h

1//===- BasicTTIImpl.h -------------------------------------------*- 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 provides a helper that implements much of the TTI interface in
11/// terms of the target-independent code generator and TargetLowering
12/// interfaces.
13//
14//===----------------------------------------------------------------------===//
15
16#ifndef LLVM_CODEGEN_BASICTTIIMPL_H
17#define LLVM_CODEGEN_BASICTTIIMPL_H
18
19#include "llvm/ADT/APInt.h"
20#include "llvm/ADT/ArrayRef.h"
21#include "llvm/ADT/BitVector.h"
22#include "llvm/ADT/SmallPtrSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/Analysis/LoopInfo.h"
25#include "llvm/Analysis/TargetTransformInfo.h"
26#include "llvm/Analysis/TargetTransformInfoImpl.h"
27#include "llvm/CodeGen/ISDOpcodes.h"
28#include "llvm/CodeGen/TargetLowering.h"
29#include "llvm/CodeGen/TargetSubtargetInfo.h"
30#include "llvm/CodeGen/ValueTypes.h"
31#include "llvm/IR/BasicBlock.h"
32#include "llvm/IR/CallSite.h"
33#include "llvm/IR/Constant.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DataLayout.h"
36#include "llvm/IR/DerivedTypes.h"
37#include "llvm/IR/InstrTypes.h"
38#include "llvm/IR/Instruction.h"
39#include "llvm/IR/Instructions.h"
40#include "llvm/IR/Intrinsics.h"
41#include "llvm/IR/Operator.h"
42#include "llvm/IR/Type.h"
43#include "llvm/IR/Value.h"
44#include "llvm/MC/MCSchedule.h"
45#include "llvm/Support/Casting.h"
46#include "llvm/Support/CommandLine.h"
47#include "llvm/Support/ErrorHandling.h"
48#include "llvm/Support/MachineValueType.h"
49#include "llvm/Support/MathExtras.h"
50#include <algorithm>
51#include <cassert>
52#include <cstdint>
53#include <limits>
54#include <utility>
55
56namespace llvm {
57
58class Function;
59class GlobalValue;
60class LLVMContext;
61class ScalarEvolution;
62class SCEV;
63class TargetMachine;
64
65extern cl::opt<unsigned> PartialUnrollingThreshold;
66
67/// Base class which can be used to help build a TTI implementation.
68///
69/// This class provides as much implementation of the TTI interface as is
70/// possible using the target independent parts of the code generator.
71///
72/// In order to subclass it, your class must implement a getST() method to
73/// return the subtarget, and a getTLI() method to return the target lowering.
74/// We need these methods implemented in the derived class so that this class
75/// doesn't have to duplicate storage for them.
76template <typename T>
77class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
78private:
79 using BaseT = TargetTransformInfoImplCRTPBase<T>;
80 using TTI = TargetTransformInfo;
81
82 /// Estimate a cost of Broadcast as an extract and sequence of insert
83 /// operations.
84 unsigned getBroadcastShuffleOverhead(Type *Ty) {
85 assert(Ty->isVectorTy() && "Can only shuffle vectors")((Ty->isVectorTy() && "Can only shuffle vectors") ?
static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only shuffle vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 85, __PRETTY_FUNCTION__))
;
86 unsigned Cost = 0;
87 // Broadcast cost is equal to the cost of extracting the zero'th element
88 // plus the cost of inserting it into every element of the result vector.
89 Cost += static_cast<T *>(this)->getVectorInstrCost(
90 Instruction::ExtractElement, Ty, 0);
91
92 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
93 Cost += static_cast<T *>(this)->getVectorInstrCost(
94 Instruction::InsertElement, Ty, i);
95 }
96 return Cost;
97 }
98
99 /// Estimate a cost of shuffle as a sequence of extract and insert
100 /// operations.
101 unsigned getPermuteShuffleOverhead(Type *Ty) {
102 assert(Ty->isVectorTy() && "Can only shuffle vectors")((Ty->isVectorTy() && "Can only shuffle vectors") ?
static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only shuffle vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 102, __PRETTY_FUNCTION__))
;
103 unsigned Cost = 0;
104 // Shuffle cost is equal to the cost of extracting element from its argument
105 // plus the cost of inserting them onto the result vector.
106
107 // e.g. <4 x float> has a mask of <0,5,2,7> i.e we need to extract from
108 // index 0 of first vector, index 1 of second vector,index 2 of first
109 // vector and finally index 3 of second vector and insert them at index
110 // <0,1,2,3> of result vector.
111 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
112 Cost += static_cast<T *>(this)
113 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
114 Cost += static_cast<T *>(this)
115 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
116 }
117 return Cost;
118 }
119
120 /// Estimate a cost of subvector extraction as a sequence of extract and
121 /// insert operations.
122 unsigned getExtractSubvectorOverhead(Type *Ty, int Index, Type *SubTy) {
123 assert(Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() &&((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only extract subvectors from vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only extract subvectors from vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 124, __PRETTY_FUNCTION__))
124 "Can only extract subvectors from vectors")((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only extract subvectors from vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only extract subvectors from vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 124, __PRETTY_FUNCTION__))
;
125 int NumSubElts = SubTy->getVectorNumElements();
126 assert((Index + NumSubElts) <= (int)Ty->getVectorNumElements() &&(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_ExtractSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_ExtractSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 127, __PRETTY_FUNCTION__))
127 "SK_ExtractSubvector index out of range")(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_ExtractSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_ExtractSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 127, __PRETTY_FUNCTION__))
;
128
129 unsigned Cost = 0;
130 // Subvector extraction cost is equal to the cost of extracting element from
131 // the source type plus the cost of inserting them into the result vector
132 // type.
133 for (int i = 0; i != NumSubElts; ++i) {
134 Cost += static_cast<T *>(this)->getVectorInstrCost(
135 Instruction::ExtractElement, Ty, i + Index);
136 Cost += static_cast<T *>(this)->getVectorInstrCost(
137 Instruction::InsertElement, SubTy, i);
138 }
139 return Cost;
140 }
141
142 /// Estimate a cost of subvector insertion as a sequence of extract and
143 /// insert operations.
144 unsigned getInsertSubvectorOverhead(Type *Ty, int Index, Type *SubTy) {
145 assert(Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() &&((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only insert subvectors into vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only insert subvectors into vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 146, __PRETTY_FUNCTION__))
146 "Can only insert subvectors into vectors")((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only insert subvectors into vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only insert subvectors into vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 146, __PRETTY_FUNCTION__))
;
147 int NumSubElts = SubTy->getVectorNumElements();
148 assert((Index + NumSubElts) <= (int)Ty->getVectorNumElements() &&(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_InsertSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_InsertSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 149, __PRETTY_FUNCTION__))
149 "SK_InsertSubvector index out of range")(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_InsertSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_InsertSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 149, __PRETTY_FUNCTION__))
;
150
151 unsigned Cost = 0;
152 // Subvector insertion cost is equal to the cost of extracting element from
153 // the source type plus the cost of inserting them into the result vector
154 // type.
155 for (int i = 0; i != NumSubElts; ++i) {
156 Cost += static_cast<T *>(this)->getVectorInstrCost(
157 Instruction::ExtractElement, SubTy, i);
158 Cost += static_cast<T *>(this)->getVectorInstrCost(
159 Instruction::InsertElement, Ty, i + Index);
160 }
161 return Cost;
162 }
163
164 /// Local query method delegates up to T which *must* implement this!
165 const TargetSubtargetInfo *getST() const {
166 return static_cast<const T *>(this)->getST();
167 }
168
169 /// Local query method delegates up to T which *must* implement this!
170 const TargetLoweringBase *getTLI() const {
171 return static_cast<const T *>(this)->getTLI();
172 }
173
174 static ISD::MemIndexedMode getISDIndexedMode(TTI::MemIndexedMode M) {
175 switch (M) {
176 case TTI::MIM_Unindexed:
177 return ISD::UNINDEXED;
178 case TTI::MIM_PreInc:
179 return ISD::PRE_INC;
180 case TTI::MIM_PreDec:
181 return ISD::PRE_DEC;
182 case TTI::MIM_PostInc:
183 return ISD::POST_INC;
184 case TTI::MIM_PostDec:
185 return ISD::POST_DEC;
186 }
187 llvm_unreachable("Unexpected MemIndexedMode")::llvm::llvm_unreachable_internal("Unexpected MemIndexedMode"
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 187)
;
188 }
189
190protected:
191 explicit BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
192 : BaseT(DL) {}
193
194 using TargetTransformInfoImplBase::DL;
195
196public:
197 /// \name Scalar TTI Implementations
198 /// @{
199 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
200 unsigned BitWidth, unsigned AddressSpace,
201 unsigned Alignment, bool *Fast) const {
202 EVT E = EVT::getIntegerVT(Context, BitWidth);
203 return getTLI()->allowsMisalignedMemoryAccesses(E, AddressSpace, Alignment, Fast);
204 }
205
206 bool hasBranchDivergence() { return false; }
207
208 bool isSourceOfDivergence(const Value *V) { return false; }
209
210 bool isAlwaysUniform(const Value *V) { return false; }
211
212 unsigned getFlatAddressSpace() {
213 // Return an invalid address space.
214 return -1;
215 }
216
217 bool isLegalAddImmediate(int64_t imm) {
218 return getTLI()->isLegalAddImmediate(imm);
219 }
220
221 bool isLegalICmpImmediate(int64_t imm) {
222 return getTLI()->isLegalICmpImmediate(imm);
223 }
224
225 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
226 bool HasBaseReg, int64_t Scale,
227 unsigned AddrSpace, Instruction *I = nullptr) {
228 TargetLoweringBase::AddrMode AM;
229 AM.BaseGV = BaseGV;
230 AM.BaseOffs = BaseOffset;
231 AM.HasBaseReg = HasBaseReg;
232 AM.Scale = Scale;
233 return getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace, I);
234 }
235
236 bool isIndexedLoadLegal(TTI::MemIndexedMode M, Type *Ty,
237 const DataLayout &DL) const {
238 EVT VT = getTLI()->getValueType(DL, Ty);
239 return getTLI()->isIndexedLoadLegal(getISDIndexedMode(M), VT);
240 }
241
242 bool isIndexedStoreLegal(TTI::MemIndexedMode M, Type *Ty,
243 const DataLayout &DL) const {
244 EVT VT = getTLI()->getValueType(DL, Ty);
245 return getTLI()->isIndexedStoreLegal(getISDIndexedMode(M), VT);
246 }
247
248 bool isLSRCostLess(TTI::LSRCost C1, TTI::LSRCost C2) {
249 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
250 }
251
252 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
253 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
254 TargetLoweringBase::AddrMode AM;
255 AM.BaseGV = BaseGV;
256 AM.BaseOffs = BaseOffset;
257 AM.HasBaseReg = HasBaseReg;
258 AM.Scale = Scale;
259 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
260 }
261
262 bool isTruncateFree(Type *Ty1, Type *Ty2) {
263 return getTLI()->isTruncateFree(Ty1, Ty2);
264 }
265
266 bool isProfitableToHoist(Instruction *I) {
267 return getTLI()->isProfitableToHoist(I);
268 }
269
270 bool useAA() const { return getST()->useAA(); }
271
272 bool isTypeLegal(Type *Ty) {
273 EVT VT = getTLI()->getValueType(DL, Ty);
274 return getTLI()->isTypeLegal(VT);
275 }
276
277 int getGEPCost(Type *PointeeType, const Value *Ptr,
278 ArrayRef<const Value *> Operands) {
279 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
280 }
281
282 int getExtCost(const Instruction *I, const Value *Src) {
283 if (getTLI()->isExtFree(I))
284 return TargetTransformInfo::TCC_Free;
285
286 if (isa<ZExtInst>(I) || isa<SExtInst>(I))
287 if (const LoadInst *LI = dyn_cast<LoadInst>(Src))
288 if (getTLI()->isExtLoad(LI, I, DL))
289 return TargetTransformInfo::TCC_Free;
290
291 return TargetTransformInfo::TCC_Basic;
292 }
293
294 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
295 ArrayRef<const Value *> Arguments, const User *U) {
296 return BaseT::getIntrinsicCost(IID, RetTy, Arguments, U);
297 }
298
299 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
300 ArrayRef<Type *> ParamTys, const User *U) {
301 if (IID == Intrinsic::cttz) {
302 if (getTLI()->isCheapToSpeculateCttz())
303 return TargetTransformInfo::TCC_Basic;
304 return TargetTransformInfo::TCC_Expensive;
305 }
306
307 if (IID == Intrinsic::ctlz) {
308 if (getTLI()->isCheapToSpeculateCtlz())
309 return TargetTransformInfo::TCC_Basic;
310 return TargetTransformInfo::TCC_Expensive;
311 }
312
313 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys, U);
314 }
315
316 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
317 unsigned &JumpTableSize) {
318 /// Try to find the estimated number of clusters. Note that the number of
319 /// clusters identified in this function could be different from the actural
320 /// numbers found in lowering. This function ignore switches that are
321 /// lowered with a mix of jump table / bit test / BTree. This function was
322 /// initially intended to be used when estimating the cost of switch in
323 /// inline cost heuristic, but it's a generic cost model to be used in other
324 /// places (e.g., in loop unrolling).
325 unsigned N = SI.getNumCases();
326 const TargetLoweringBase *TLI = getTLI();
327 const DataLayout &DL = this->getDataLayout();
328
329 JumpTableSize = 0;
330 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
331
332 // Early exit if both a jump table and bit test are not allowed.
333 if (N < 1 || (!IsJTAllowed && DL.getIndexSizeInBits(0u) < N))
334 return N;
335
336 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
337 APInt MinCaseVal = MaxCaseVal;
338 for (auto CI : SI.cases()) {
339 const APInt &CaseVal = CI.getCaseValue()->getValue();
340 if (CaseVal.sgt(MaxCaseVal))
341 MaxCaseVal = CaseVal;
342 if (CaseVal.slt(MinCaseVal))
343 MinCaseVal = CaseVal;
344 }
345
346 // Check if suitable for a bit test
347 if (N <= DL.getIndexSizeInBits(0u)) {
348 SmallPtrSet<const BasicBlock *, 4> Dests;
349 for (auto I : SI.cases())
350 Dests.insert(I.getCaseSuccessor());
351
352 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
353 DL))
354 return 1;
355 }
356
357 // Check if suitable for a jump table.
358 if (IsJTAllowed) {
359 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
360 return N;
361 uint64_t Range =
362 (MaxCaseVal - MinCaseVal)
363 .getLimitedValue(std::numeric_limits<uint64_t>::max() - 1) + 1;
364 // Check whether a range of clusters is dense enough for a jump table
365 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
366 JumpTableSize = Range;
367 return 1;
368 }
369 }
370 return N;
371 }
372
373 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
374
375 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
376
377 bool shouldBuildLookupTables() {
378 const TargetLoweringBase *TLI = getTLI();
379 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
380 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
381 }
382
383 bool haveFastSqrt(Type *Ty) {
384 const TargetLoweringBase *TLI = getTLI();
385 EVT VT = TLI->getValueType(DL, Ty);
386 return TLI->isTypeLegal(VT) &&
387 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
388 }
389
390 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
391 return true;
392 }
393
394 unsigned getFPOpCost(Type *Ty) {
395 // Check whether FADD is available, as a proxy for floating-point in
396 // general.
397 const TargetLoweringBase *TLI = getTLI();
398 EVT VT = TLI->getValueType(DL, Ty);
399 if (TLI->isOperationLegalOrCustomOrPromote(ISD::FADD, VT))
400 return TargetTransformInfo::TCC_Basic;
401 return TargetTransformInfo::TCC_Expensive;
402 }
403
404 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
405 const TargetLoweringBase *TLI = getTLI();
406 switch (Opcode) {
12
Control jumps to 'case AddrSpaceCast:' at line 417
407 default: break;
408 case Instruction::Trunc:
409 if (TLI->isTruncateFree(OpTy, Ty))
410 return TargetTransformInfo::TCC_Free;
411 return TargetTransformInfo::TCC_Basic;
412 case Instruction::ZExt:
413 if (TLI->isZExtFree(OpTy, Ty))
414 return TargetTransformInfo::TCC_Free;
415 return TargetTransformInfo::TCC_Basic;
416
417 case Instruction::AddrSpaceCast:
418 if (TLI->isFreeAddrSpaceCast(OpTy->getPointerAddressSpace(),
13
Called C++ object pointer is null
419 Ty->getPointerAddressSpace()))
420 return TargetTransformInfo::TCC_Free;
421 return TargetTransformInfo::TCC_Basic;
422 }
423
424 return BaseT::getOperationCost(Opcode, Ty, OpTy);
425 }
426
427 unsigned getInliningThresholdMultiplier() { return 1; }
428
429 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
430 TTI::UnrollingPreferences &UP) {
431 // This unrolling functionality is target independent, but to provide some
432 // motivation for its intended use, for x86:
433
434 // According to the Intel 64 and IA-32 Architectures Optimization Reference
435 // Manual, Intel Core models and later have a loop stream detector (and
436 // associated uop queue) that can benefit from partial unrolling.
437 // The relevant requirements are:
438 // - The loop must have no more than 4 (8 for Nehalem and later) branches
439 // taken, and none of them may be calls.
440 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
441
442 // According to the Software Optimization Guide for AMD Family 15h
443 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
444 // and loop buffer which can benefit from partial unrolling.
445 // The relevant requirements are:
446 // - The loop must have fewer than 16 branches
447 // - The loop must have less than 40 uops in all executed loop branches
448
449 // The number of taken branches in a loop is hard to estimate here, and
450 // benchmarking has revealed that it is better not to be conservative when
451 // estimating the branch count. As a result, we'll ignore the branch limits
452 // until someone finds a case where it matters in practice.
453
454 unsigned MaxOps;
455 const TargetSubtargetInfo *ST = getST();
456 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
457 MaxOps = PartialUnrollingThreshold;
458 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
459 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
460 else
461 return;
462
463 // Scan the loop: don't unroll loops with calls.
464 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
465 ++I) {
466 BasicBlock *BB = *I;
467
468 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
469 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
470 ImmutableCallSite CS(&*J);
471 if (const Function *F = CS.getCalledFunction()) {
472 if (!static_cast<T *>(this)->isLoweredToCall(F))
473 continue;
474 }
475
476 return;
477 }
478 }
479
480 // Enable runtime and partial unrolling up to the specified size.
481 // Enable using trip count upper bound to unroll loops.
482 UP.Partial = UP.Runtime = UP.UpperBound = true;
483 UP.PartialThreshold = MaxOps;
484
485 // Avoid unrolling when optimizing for size.
486 UP.OptSizeThreshold = 0;
487 UP.PartialOptSizeThreshold = 0;
488
489 // Set number of instructions optimized when "back edge"
490 // becomes "fall through" to default value of 2.
491 UP.BEInsns = 2;
492 }
493
494 int getInstructionLatency(const Instruction *I) {
495 if (isa<LoadInst>(I))
496 return getST()->getSchedModel().DefaultLoadLatency;
497
498 return BaseT::getInstructionLatency(I);
499 }
500
501 /// @}
502
503 /// \name Vector TTI Implementations
504 /// @{
505
506 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
507
508 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
509
510 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
511 /// are set if the result needs to be inserted and/or extracted from vectors.
512 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
513 assert(Ty->isVectorTy() && "Can only scalarize vectors")((Ty->isVectorTy() && "Can only scalarize vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only scalarize vectors\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 513, __PRETTY_FUNCTION__))
;
514 unsigned Cost = 0;
515
516 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
517 if (Insert)
518 Cost += static_cast<T *>(this)
519 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
520 if (Extract)
521 Cost += static_cast<T *>(this)
522 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
523 }
524
525 return Cost;
526 }
527
528 /// Estimate the overhead of scalarizing an instructions unique
529 /// non-constant operands. The types of the arguments are ordinarily
530 /// scalar, in which case the costs are multiplied with VF.
531 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
532 unsigned VF) {
533 unsigned Cost = 0;
534 SmallPtrSet<const Value*, 4> UniqueOperands;
535 for (const Value *A : Args) {
536 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
537 Type *VecTy = nullptr;
538 if (A->getType()->isVectorTy()) {
539 VecTy = A->getType();
540 // If A is a vector operand, VF should be 1 or correspond to A.
541 assert((VF == 1 || VF == VecTy->getVectorNumElements()) &&(((VF == 1 || VF == VecTy->getVectorNumElements()) &&
"Vector argument does not match VF") ? static_cast<void>
(0) : __assert_fail ("(VF == 1 || VF == VecTy->getVectorNumElements()) && \"Vector argument does not match VF\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 542, __PRETTY_FUNCTION__))
542 "Vector argument does not match VF")(((VF == 1 || VF == VecTy->getVectorNumElements()) &&
"Vector argument does not match VF") ? static_cast<void>
(0) : __assert_fail ("(VF == 1 || VF == VecTy->getVectorNumElements()) && \"Vector argument does not match VF\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 542, __PRETTY_FUNCTION__))
;
543 }
544 else
545 VecTy = VectorType::get(A->getType(), VF);
546
547 Cost += getScalarizationOverhead(VecTy, false, true);
548 }
549 }
550
551 return Cost;
552 }
553
554 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
555 assert(VecTy->isVectorTy())((VecTy->isVectorTy()) ? static_cast<void> (0) : __assert_fail
("VecTy->isVectorTy()", "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 555, __PRETTY_FUNCTION__))
;
556
557 unsigned Cost = 0;
558
559 Cost += getScalarizationOverhead(VecTy, true, false);
560 if (!Args.empty())
561 Cost += getOperandsScalarizationOverhead(Args,
562 VecTy->getVectorNumElements());
563 else
564 // When no information on arguments is provided, we add the cost
565 // associated with one argument as a heuristic.
566 Cost += getScalarizationOverhead(VecTy, false, true);
567
568 return Cost;
569 }
570
571 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
572
573 unsigned getArithmeticInstrCost(
574 unsigned Opcode, Type *Ty,
575 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
576 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
577 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
578 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
579 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
580 // Check if any of the operands are vector operands.
581 const TargetLoweringBase *TLI = getTLI();
582 int ISD = TLI->InstructionOpcodeToISD(Opcode);
583 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 583, __PRETTY_FUNCTION__))
;
584
585 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
586
587 bool IsFloat = Ty->isFPOrFPVectorTy();
588 // Assume that floating point arithmetic operations cost twice as much as
589 // integer operations.
590 unsigned OpCost = (IsFloat ? 2 : 1);
591
592 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
593 // The operation is legal. Assume it costs 1.
594 // TODO: Once we have extract/insert subvector cost we need to use them.
595 return LT.first * OpCost;
596 }
597
598 if (!TLI->isOperationExpand(ISD, LT.second)) {
599 // If the operation is custom lowered, then assume that the code is twice
600 // as expensive.
601 return LT.first * 2 * OpCost;
602 }
603
604 // Else, assume that we need to scalarize this op.
605 // TODO: If one of the types get legalized by splitting, handle this
606 // similarly to what getCastInstrCost() does.
607 if (Ty->isVectorTy()) {
608 unsigned Num = Ty->getVectorNumElements();
609 unsigned Cost = static_cast<T *>(this)
610 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
611 // Return the cost of multiple scalar invocation plus the cost of
612 // inserting and extracting the values.
613 return getScalarizationOverhead(Ty, Args) + Num * Cost;
614 }
615
616 // We don't know anything about this scalar instruction.
617 return OpCost;
618 }
619
620 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
621 Type *SubTp) {
622 switch (Kind) {
623 case TTI::SK_Broadcast:
624 return getBroadcastShuffleOverhead(Tp);
625 case TTI::SK_Select:
626 case TTI::SK_Reverse:
627 case TTI::SK_Transpose:
628 case TTI::SK_PermuteSingleSrc:
629 case TTI::SK_PermuteTwoSrc:
630 return getPermuteShuffleOverhead(Tp);
631 case TTI::SK_ExtractSubvector:
632 return getExtractSubvectorOverhead(Tp, Index, SubTp);
633 case TTI::SK_InsertSubvector:
634 return getInsertSubvectorOverhead(Tp, Index, SubTp);
635 }
636 llvm_unreachable("Unknown TTI::ShuffleKind")::llvm::llvm_unreachable_internal("Unknown TTI::ShuffleKind",
"/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 636)
;
637 }
638
639 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
640 const Instruction *I = nullptr) {
641 const TargetLoweringBase *TLI = getTLI();
642 int ISD = TLI->InstructionOpcodeToISD(Opcode);
643 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 643, __PRETTY_FUNCTION__))
;
644 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
645 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
646
647 // Check for NOOP conversions.
648 if (SrcLT.first == DstLT.first &&
649 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
650
651 // Bitcast between types that are legalized to the same type are free.
652 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
653 return 0;
654 }
655
656 if (Opcode == Instruction::Trunc &&
657 TLI->isTruncateFree(SrcLT.second, DstLT.second))
658 return 0;
659
660 if (Opcode == Instruction::ZExt &&
661 TLI->isZExtFree(SrcLT.second, DstLT.second))
662 return 0;
663
664 if (Opcode == Instruction::AddrSpaceCast &&
665 TLI->isFreeAddrSpaceCast(Src->getPointerAddressSpace(),
666 Dst->getPointerAddressSpace()))
667 return 0;
668
669 // If this is a zext/sext of a load, return 0 if the corresponding
670 // extending load exists on target.
671 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
672 I && isa<LoadInst>(I->getOperand(0))) {
673 EVT ExtVT = EVT::getEVT(Dst);
674 EVT LoadVT = EVT::getEVT(Src);
675 unsigned LType =
676 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
677 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
678 return 0;
679 }
680
681 // If the cast is marked as legal (or promote) then assume low cost.
682 if (SrcLT.first == DstLT.first &&
683 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
684 return 1;
685
686 // Handle scalar conversions.
687 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
688 // Scalar bitcasts are usually free.
689 if (Opcode == Instruction::BitCast)
690 return 0;
691
692 // Just check the op cost. If the operation is legal then assume it costs
693 // 1.
694 if (!TLI->isOperationExpand(ISD, DstLT.second))
695 return 1;
696
697 // Assume that illegal scalar instruction are expensive.
698 return 4;
699 }
700
701 // Check vector-to-vector casts.
702 if (Dst->isVectorTy() && Src->isVectorTy()) {
703 // If the cast is between same-sized registers, then the check is simple.
704 if (SrcLT.first == DstLT.first &&
705 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
706
707 // Assume that Zext is done using AND.
708 if (Opcode == Instruction::ZExt)
709 return 1;
710
711 // Assume that sext is done using SHL and SRA.
712 if (Opcode == Instruction::SExt)
713 return 2;
714
715 // Just check the op cost. If the operation is legal then assume it
716 // costs
717 // 1 and multiply by the type-legalization overhead.
718 if (!TLI->isOperationExpand(ISD, DstLT.second))
719 return SrcLT.first * 1;
720 }
721
722 // If we are legalizing by splitting, query the concrete TTI for the cost
723 // of casting the original vector twice. We also need to factor in the
724 // cost of the split itself. Count that as 1, to be consistent with
725 // TLI->getTypeLegalizationCost().
726 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
727 TargetLowering::TypeSplitVector) ||
728 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
729 TargetLowering::TypeSplitVector)) {
730 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
731 Dst->getVectorNumElements() / 2);
732 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
733 Src->getVectorNumElements() / 2);
734 T *TTI = static_cast<T *>(this);
735 return TTI->getVectorSplitCost() +
736 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
737 }
738
739 // In other cases where the source or destination are illegal, assume
740 // the operation will get scalarized.
741 unsigned Num = Dst->getVectorNumElements();
742 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
743 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
744
745 // Return the cost of multiple scalar invocation plus the cost of
746 // inserting and extracting the values.
747 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
748 }
749
750 // We already handled vector-to-vector and scalar-to-scalar conversions.
751 // This
752 // is where we handle bitcast between vectors and scalars. We need to assume
753 // that the conversion is scalarized in one way or another.
754 if (Opcode == Instruction::BitCast)
755 // Illegal bitcasts are done by storing and loading from a stack slot.
756 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
757 : 0) +
758 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
759 : 0);
760
761 llvm_unreachable("Unhandled cast")::llvm::llvm_unreachable_internal("Unhandled cast", "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 761)
;
762 }
763
764 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
765 VectorType *VecTy, unsigned Index) {
766 return static_cast<T *>(this)->getVectorInstrCost(
767 Instruction::ExtractElement, VecTy, Index) +
768 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
769 VecTy->getElementType());
770 }
771
772 unsigned getCFInstrCost(unsigned Opcode) {
773 // Branches are assumed to be predicted.
774 return 0;
775 }
776
777 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
778 const Instruction *I) {
779 const TargetLoweringBase *TLI = getTLI();
780 int ISD = TLI->InstructionOpcodeToISD(Opcode);
781 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 781, __PRETTY_FUNCTION__))
;
782
783 // Selects on vectors are actually vector selects.
784 if (ISD == ISD::SELECT) {
785 assert(CondTy && "CondTy must exist")((CondTy && "CondTy must exist") ? static_cast<void
> (0) : __assert_fail ("CondTy && \"CondTy must exist\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 785, __PRETTY_FUNCTION__))
;
786 if (CondTy->isVectorTy())
787 ISD = ISD::VSELECT;
788 }
789 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
790
791 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
792 !TLI->isOperationExpand(ISD, LT.second)) {
793 // The operation is legal. Assume it costs 1. Multiply
794 // by the type-legalization overhead.
795 return LT.first * 1;
796 }
797
798 // Otherwise, assume that the cast is scalarized.
799 // TODO: If one of the types get legalized by splitting, handle this
800 // similarly to what getCastInstrCost() does.
801 if (ValTy->isVectorTy()) {
802 unsigned Num = ValTy->getVectorNumElements();
803 if (CondTy)
804 CondTy = CondTy->getScalarType();
805 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
806 Opcode, ValTy->getScalarType(), CondTy, I);
807
808 // Return the cost of multiple scalar invocation plus the cost of
809 // inserting and extracting the values.
810 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
811 }
812
813 // Unknown scalar opcode.
814 return 1;
815 }
816
817 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
818 std::pair<unsigned, MVT> LT =
819 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
820
821 return LT.first;
822 }
823
824 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
825 unsigned AddressSpace, const Instruction *I = nullptr) {
826 assert(!Src->isVoidTy() && "Invalid type")((!Src->isVoidTy() && "Invalid type") ? static_cast
<void> (0) : __assert_fail ("!Src->isVoidTy() && \"Invalid type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 826, __PRETTY_FUNCTION__))
;
827 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
828
829 // Assuming that all loads of legal types cost 1.
830 unsigned Cost = LT.first;
831
832 if (Src->isVectorTy() &&
833 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
834 // This is a vector load that legalizes to a larger type than the vector
835 // itself. Unless the corresponding extending load or truncating store is
836 // legal, then this will scalarize.
837 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
838 EVT MemVT = getTLI()->getValueType(DL, Src);
839 if (Opcode == Instruction::Store)
840 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
841 else
842 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
843
844 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
845 // This is a vector load/store for some illegal type that is scalarized.
846 // We must account for the cost of building or decomposing the vector.
847 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
848 Opcode == Instruction::Store);
849 }
850 }
851
852 return Cost;
853 }
854
855 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
856 unsigned Factor,
857 ArrayRef<unsigned> Indices,
858 unsigned Alignment, unsigned AddressSpace,
859 bool UseMaskForCond = false,
860 bool UseMaskForGaps = false) {
861 VectorType *VT = dyn_cast<VectorType>(VecTy);
862 assert(VT && "Expect a vector type for interleaved memory op")((VT && "Expect a vector type for interleaved memory op"
) ? static_cast<void> (0) : __assert_fail ("VT && \"Expect a vector type for interleaved memory op\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 862, __PRETTY_FUNCTION__))
;
863
864 unsigned NumElts = VT->getNumElements();
865 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor")((Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor"
) ? static_cast<void> (0) : __assert_fail ("Factor > 1 && NumElts % Factor == 0 && \"Invalid interleave factor\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 865, __PRETTY_FUNCTION__))
;
866
867 unsigned NumSubElts = NumElts / Factor;
868 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
869
870 // Firstly, the cost of load/store operation.
871 unsigned Cost;
872 if (UseMaskForCond || UseMaskForGaps)
873 Cost = static_cast<T *>(this)->getMaskedMemoryOpCost(
874 Opcode, VecTy, Alignment, AddressSpace);
875 else
876 Cost = static_cast<T *>(this)->getMemoryOpCost(Opcode, VecTy, Alignment,
877 AddressSpace);
878
879 // Legalize the vector type, and get the legalized and unlegalized type
880 // sizes.
881 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
882 unsigned VecTySize =
883 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
884 unsigned VecTyLTSize = VecTyLT.getStoreSize();
885
886 // Return the ceiling of dividing A by B.
887 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
888
889 // Scale the cost of the memory operation by the fraction of legalized
890 // instructions that will actually be used. We shouldn't account for the
891 // cost of dead instructions since they will be removed.
892 //
893 // E.g., An interleaved load of factor 8:
894 // %vec = load <16 x i64>, <16 x i64>* %ptr
895 // %v0 = shufflevector %vec, undef, <0, 8>
896 //
897 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
898 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
899 // type). The other loads are unused.
900 //
901 // We only scale the cost of loads since interleaved store groups aren't
902 // allowed to have gaps.
903 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
904 // The number of loads of a legal type it will take to represent a load
905 // of the unlegalized vector type.
906 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
907
908 // The number of elements of the unlegalized type that correspond to a
909 // single legal instruction.
910 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
911
912 // Determine which legal instructions will be used.
913 BitVector UsedInsts(NumLegalInsts, false);
914 for (unsigned Index : Indices)
915 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
916 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
917
918 // Scale the cost of the load by the fraction of legal instructions that
919 // will be used.
920 Cost *= UsedInsts.count() / NumLegalInsts;
921 }
922
923 // Then plus the cost of interleave operation.
924 if (Opcode == Instruction::Load) {
925 // The interleave cost is similar to extract sub vectors' elements
926 // from the wide vector, and insert them into sub vectors.
927 //
928 // E.g. An interleaved load of factor 2 (with one member of index 0):
929 // %vec = load <8 x i32>, <8 x i32>* %ptr
930 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
931 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
932 // <8 x i32> vector and insert them into a <4 x i32> vector.
933
934 assert(Indices.size() <= Factor &&((Indices.size() <= Factor && "Interleaved memory op has too many members"
) ? static_cast<void> (0) : __assert_fail ("Indices.size() <= Factor && \"Interleaved memory op has too many members\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 935, __PRETTY_FUNCTION__))
935 "Interleaved memory op has too many members")((Indices.size() <= Factor && "Interleaved memory op has too many members"
) ? static_cast<void> (0) : __assert_fail ("Indices.size() <= Factor && \"Interleaved memory op has too many members\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 935, __PRETTY_FUNCTION__))
;
936
937 for (unsigned Index : Indices) {
938 assert(Index < Factor && "Invalid index for interleaved memory op")((Index < Factor && "Invalid index for interleaved memory op"
) ? static_cast<void> (0) : __assert_fail ("Index < Factor && \"Invalid index for interleaved memory op\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 938, __PRETTY_FUNCTION__))
;
939
940 // Extract elements from loaded vector for each sub vector.
941 for (unsigned i = 0; i < NumSubElts; i++)
942 Cost += static_cast<T *>(this)->getVectorInstrCost(
943 Instruction::ExtractElement, VT, Index + i * Factor);
944 }
945
946 unsigned InsSubCost = 0;
947 for (unsigned i = 0; i < NumSubElts; i++)
948 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
949 Instruction::InsertElement, SubVT, i);
950
951 Cost += Indices.size() * InsSubCost;
952 } else {
953 // The interleave cost is extract all elements from sub vectors, and
954 // insert them into the wide vector.
955 //
956 // E.g. An interleaved store of factor 2:
957 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
958 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
959 // The cost is estimated as extract all elements from both <4 x i32>
960 // vectors and insert into the <8 x i32> vector.
961
962 unsigned ExtSubCost = 0;
963 for (unsigned i = 0; i < NumSubElts; i++)
964 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
965 Instruction::ExtractElement, SubVT, i);
966 Cost += ExtSubCost * Factor;
967
968 for (unsigned i = 0; i < NumElts; i++)
969 Cost += static_cast<T *>(this)
970 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
971 }
972
973 if (!UseMaskForCond)
974 return Cost;
975
976 Type *I8Type = Type::getInt8Ty(VT->getContext());
977 VectorType *MaskVT = VectorType::get(I8Type, NumElts);
978 SubVT = VectorType::get(I8Type, NumSubElts);
979
980 // The Mask shuffling cost is extract all the elements of the Mask
981 // and insert each of them Factor times into the wide vector:
982 //
983 // E.g. an interleaved group with factor 3:
984 // %mask = icmp ult <8 x i32> %vec1, %vec2
985 // %interleaved.mask = shufflevector <8 x i1> %mask, <8 x i1> undef,
986 // <24 x i32> <0,0,0,1,1,1,2,2,2,3,3,3,4,4,4,5,5,5,6,6,6,7,7,7>
987 // The cost is estimated as extract all mask elements from the <8xi1> mask
988 // vector and insert them factor times into the <24xi1> shuffled mask
989 // vector.
990 for (unsigned i = 0; i < NumSubElts; i++)
991 Cost += static_cast<T *>(this)->getVectorInstrCost(
992 Instruction::ExtractElement, SubVT, i);
993
994 for (unsigned i = 0; i < NumElts; i++)
995 Cost += static_cast<T *>(this)->getVectorInstrCost(
996 Instruction::InsertElement, MaskVT, i);
997
998 // The Gaps mask is invariant and created outside the loop, therefore the
999 // cost of creating it is not accounted for here. However if we have both
1000 // a MaskForGaps and some other mask that guards the execution of the
1001 // memory access, we need to account for the cost of And-ing the two masks
1002 // inside the loop.
1003 if (UseMaskForGaps)
1004 Cost += static_cast<T *>(this)->getArithmeticInstrCost(
1005 BinaryOperator::And, MaskVT);
1006
1007 return Cost;
1008 }
1009
1010 /// Get intrinsic cost based on arguments.
1011 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1012 ArrayRef<Value *> Args, FastMathFlags FMF,
1013 unsigned VF = 1) {
1014 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
1015 assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type")(((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type"
) ? static_cast<void> (0) : __assert_fail ("(RetVF == 1 || VF == 1) && \"VF > 1 and RetVF is a vector type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1015, __PRETTY_FUNCTION__))
;
1016 auto *ConcreteTTI = static_cast<T *>(this);
1017
1018 switch (IID) {
1019 default: {
1020 // Assume that we need to scalarize this intrinsic.
1021 SmallVector<Type *, 4> Types;
1022 for (Value *Op : Args) {
1023 Type *OpTy = Op->getType();
1024 assert(VF == 1 || !OpTy->isVectorTy())((VF == 1 || !OpTy->isVectorTy()) ? static_cast<void>
(0) : __assert_fail ("VF == 1 || !OpTy->isVectorTy()", "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1024, __PRETTY_FUNCTION__))
;
1025 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
1026 }
1027
1028 if (VF > 1 && !RetTy->isVoidTy())
1029 RetTy = VectorType::get(RetTy, VF);
1030
1031 // Compute the scalarization overhead based on Args for a vector
1032 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
1033 // CostModel will pass a vector RetTy and VF is 1.
1034 unsigned ScalarizationCost = std::numeric_limits<unsigned>::max();
1035 if (RetVF > 1 || VF > 1) {
1036 ScalarizationCost = 0;
1037 if (!RetTy->isVoidTy())
1038 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
1039 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
1040 }
1041
1042 return ConcreteTTI->getIntrinsicInstrCost(IID, RetTy, Types, FMF,
1043 ScalarizationCost);
1044 }
1045 case Intrinsic::masked_scatter: {
1046 assert(VF == 1 && "Can't vectorize types here.")((VF == 1 && "Can't vectorize types here.") ? static_cast
<void> (0) : __assert_fail ("VF == 1 && \"Can't vectorize types here.\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1046, __PRETTY_FUNCTION__))
;
1047 Value *Mask = Args[3];
1048 bool VarMask = !isa<Constant>(Mask);
1049 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
1050 return ConcreteTTI->getGatherScatterOpCost(
1051 Instruction::Store, Args[0]->getType(), Args[1], VarMask, Alignment);
1052 }
1053 case Intrinsic::masked_gather: {
1054 assert(VF == 1 && "Can't vectorize types here.")((VF == 1 && "Can't vectorize types here.") ? static_cast
<void> (0) : __assert_fail ("VF == 1 && \"Can't vectorize types here.\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1054, __PRETTY_FUNCTION__))
;
1055 Value *Mask = Args[2];
1056 bool VarMask = !isa<Constant>(Mask);
1057 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
1058 return ConcreteTTI->getGatherScatterOpCost(Instruction::Load, RetTy,
1059 Args[0], VarMask, Alignment);
1060 }
1061 case Intrinsic::experimental_vector_reduce_add:
1062 case Intrinsic::experimental_vector_reduce_mul:
1063 case Intrinsic::experimental_vector_reduce_and:
1064 case Intrinsic::experimental_vector_reduce_or:
1065 case Intrinsic::experimental_vector_reduce_xor:
1066 case Intrinsic::experimental_vector_reduce_fadd:
1067 case Intrinsic::experimental_vector_reduce_fmul:
1068 case Intrinsic::experimental_vector_reduce_smax:
1069 case Intrinsic::experimental_vector_reduce_smin:
1070 case Intrinsic::experimental_vector_reduce_fmax:
1071 case Intrinsic::experimental_vector_reduce_fmin:
1072 case Intrinsic::experimental_vector_reduce_umax:
1073 case Intrinsic::experimental_vector_reduce_umin:
1074 return getIntrinsicInstrCost(IID, RetTy, Args[0]->getType(), FMF);
1075 case Intrinsic::fshl:
1076 case Intrinsic::fshr: {
1077 Value *X = Args[0];
1078 Value *Y = Args[1];
1079 Value *Z = Args[2];
1080 TTI::OperandValueProperties OpPropsX, OpPropsY, OpPropsZ, OpPropsBW;
1081 TTI::OperandValueKind OpKindX = TTI::getOperandInfo(X, OpPropsX);
1082 TTI::OperandValueKind OpKindY = TTI::getOperandInfo(Y, OpPropsY);
1083 TTI::OperandValueKind OpKindZ = TTI::getOperandInfo(Z, OpPropsZ);
1084 TTI::OperandValueKind OpKindBW = TTI::OK_UniformConstantValue;
1085 OpPropsBW = isPowerOf2_32(RetTy->getScalarSizeInBits()) ? TTI::OP_PowerOf2
1086 : TTI::OP_None;
1087 // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
1088 // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW))
1089 unsigned Cost = 0;
1090 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Or, RetTy);
1091 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Sub, RetTy);
1092 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Shl, RetTy,
1093 OpKindX, OpKindZ, OpPropsX);
1094 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::LShr, RetTy,
1095 OpKindY, OpKindZ, OpPropsY);
1096 // Non-constant shift amounts requires a modulo.
1097 if (OpKindZ != TTI::OK_UniformConstantValue &&
1098 OpKindZ != TTI::OK_NonUniformConstantValue)
1099 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::URem, RetTy,
1100 OpKindZ, OpKindBW, OpPropsZ,
1101 OpPropsBW);
1102 // For non-rotates (X != Y) we must add shift-by-zero handling costs.
1103 if (X != Y) {
1104 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1105 if (RetVF > 1)
1106 CondTy = VectorType::get(CondTy, RetVF);
1107 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, RetTy,
1108 CondTy, nullptr);
1109 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1110 CondTy, nullptr);
1111 }
1112 return Cost;
1113 }
1114 }
1115 }
1116
1117 /// Get intrinsic cost based on argument types.
1118 /// If ScalarizationCostPassed is std::numeric_limits<unsigned>::max(), the
1119 /// cost of scalarizing the arguments and the return value will be computed
1120 /// based on types.
1121 unsigned getIntrinsicInstrCost(
1122 Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> Tys, FastMathFlags FMF,
1123 unsigned ScalarizationCostPassed = std::numeric_limits<unsigned>::max()) {
1124 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
1125 auto *ConcreteTTI = static_cast<T *>(this);
1126
1127 SmallVector<unsigned, 2> ISDs;
1128 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
1129 switch (IID) {
1130 default: {
1131 // Assume that we need to scalarize this intrinsic.
1132 unsigned ScalarizationCost = ScalarizationCostPassed;
1133 unsigned ScalarCalls = 1;
1134 Type *ScalarRetTy = RetTy;
1135 if (RetTy->isVectorTy()) {
1136 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1137 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
1138 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
1139 ScalarRetTy = RetTy->getScalarType();
1140 }
1141 SmallVector<Type *, 4> ScalarTys;
1142 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1143 Type *Ty = Tys[i];
1144 if (Ty->isVectorTy()) {
1145 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1146 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
1147 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
1148 Ty = Ty->getScalarType();
1149 }
1150 ScalarTys.push_back(Ty);
1151 }
1152 if (ScalarCalls == 1)
1153 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
1154
1155 unsigned ScalarCost =
1156 ConcreteTTI->getIntrinsicInstrCost(IID, ScalarRetTy, ScalarTys, FMF);
1157
1158 return ScalarCalls * ScalarCost + ScalarizationCost;
1159 }
1160 // Look for intrinsics that can be lowered directly or turned into a scalar
1161 // intrinsic call.
1162 case Intrinsic::sqrt:
1163 ISDs.push_back(ISD::FSQRT);
1164 break;
1165 case Intrinsic::sin:
1166 ISDs.push_back(ISD::FSIN);
1167 break;
1168 case Intrinsic::cos:
1169 ISDs.push_back(ISD::FCOS);
1170 break;
1171 case Intrinsic::exp:
1172 ISDs.push_back(ISD::FEXP);
1173 break;
1174 case Intrinsic::exp2:
1175 ISDs.push_back(ISD::FEXP2);
1176 break;
1177 case Intrinsic::log:
1178 ISDs.push_back(ISD::FLOG);
1179 break;
1180 case Intrinsic::log10:
1181 ISDs.push_back(ISD::FLOG10);
1182 break;
1183 case Intrinsic::log2:
1184 ISDs.push_back(ISD::FLOG2);
1185 break;
1186 case Intrinsic::fabs:
1187 ISDs.push_back(ISD::FABS);
1188 break;
1189 case Intrinsic::canonicalize:
1190 ISDs.push_back(ISD::FCANONICALIZE);
1191 break;
1192 case Intrinsic::minnum:
1193 ISDs.push_back(ISD::FMINNUM);
1194 if (FMF.noNaNs())
1195 ISDs.push_back(ISD::FMINIMUM);
1196 break;
1197 case Intrinsic::maxnum:
1198 ISDs.push_back(ISD::FMAXNUM);
1199 if (FMF.noNaNs())
1200 ISDs.push_back(ISD::FMAXIMUM);
1201 break;
1202 case Intrinsic::copysign:
1203 ISDs.push_back(ISD::FCOPYSIGN);
1204 break;
1205 case Intrinsic::floor:
1206 ISDs.push_back(ISD::FFLOOR);
1207 break;
1208 case Intrinsic::ceil:
1209 ISDs.push_back(ISD::FCEIL);
1210 break;
1211 case Intrinsic::trunc:
1212 ISDs.push_back(ISD::FTRUNC);
1213 break;
1214 case Intrinsic::nearbyint:
1215 ISDs.push_back(ISD::FNEARBYINT);
1216 break;
1217 case Intrinsic::rint:
1218 ISDs.push_back(ISD::FRINT);
1219 break;
1220 case Intrinsic::round:
1221 ISDs.push_back(ISD::FROUND);
1222 break;
1223 case Intrinsic::pow:
1224 ISDs.push_back(ISD::FPOW);
1225 break;
1226 case Intrinsic::fma:
1227 ISDs.push_back(ISD::FMA);
1228 break;
1229 case Intrinsic::fmuladd:
1230 ISDs.push_back(ISD::FMA);
1231 break;
1232 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
1233 case Intrinsic::lifetime_start:
1234 case Intrinsic::lifetime_end:
1235 case Intrinsic::sideeffect:
1236 return 0;
1237 case Intrinsic::masked_store:
1238 return ConcreteTTI->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0,
1239 0);
1240 case Intrinsic::masked_load:
1241 return ConcreteTTI->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
1242 case Intrinsic::experimental_vector_reduce_add:
1243 return ConcreteTTI->getArithmeticReductionCost(Instruction::Add, Tys[0],
1244 /*IsPairwiseForm=*/false);
1245 case Intrinsic::experimental_vector_reduce_mul:
1246 return ConcreteTTI->getArithmeticReductionCost(Instruction::Mul, Tys[0],
1247 /*IsPairwiseForm=*/false);
1248 case Intrinsic::experimental_vector_reduce_and:
1249 return ConcreteTTI->getArithmeticReductionCost(Instruction::And, Tys[0],
1250 /*IsPairwiseForm=*/false);
1251 case Intrinsic::experimental_vector_reduce_or:
1252 return ConcreteTTI->getArithmeticReductionCost(Instruction::Or, Tys[0],
1253 /*IsPairwiseForm=*/false);
1254 case Intrinsic::experimental_vector_reduce_xor:
1255 return ConcreteTTI->getArithmeticReductionCost(Instruction::Xor, Tys[0],
1256 /*IsPairwiseForm=*/false);
1257 case Intrinsic::experimental_vector_reduce_fadd:
1258 return ConcreteTTI->getArithmeticReductionCost(Instruction::FAdd, Tys[0],
1259 /*IsPairwiseForm=*/false);
1260 case Intrinsic::experimental_vector_reduce_fmul:
1261 return ConcreteTTI->getArithmeticReductionCost(Instruction::FMul, Tys[0],
1262 /*IsPairwiseForm=*/false);
1263 case Intrinsic::experimental_vector_reduce_smax:
1264 case Intrinsic::experimental_vector_reduce_smin:
1265 case Intrinsic::experimental_vector_reduce_fmax:
1266 case Intrinsic::experimental_vector_reduce_fmin:
1267 return ConcreteTTI->getMinMaxReductionCost(
1268 Tys[0], CmpInst::makeCmpResultType(Tys[0]), /*IsPairwiseForm=*/false,
1269 /*IsSigned=*/true);
1270 case Intrinsic::experimental_vector_reduce_umax:
1271 case Intrinsic::experimental_vector_reduce_umin:
1272 return ConcreteTTI->getMinMaxReductionCost(
1273 Tys[0], CmpInst::makeCmpResultType(Tys[0]), /*IsPairwiseForm=*/false,
1274 /*IsSigned=*/false);
1275 case Intrinsic::sadd_sat:
1276 case Intrinsic::ssub_sat: {
1277 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1278 if (RetVF > 1)
1279 CondTy = VectorType::get(CondTy, RetVF);
1280
1281 Type *OpTy = StructType::create({RetTy, CondTy});
1282 Intrinsic::ID OverflowOp = IID == Intrinsic::sadd_sat
1283 ? Intrinsic::sadd_with_overflow
1284 : Intrinsic::ssub_with_overflow;
1285
1286 // SatMax -> Overflow && SumDiff < 0
1287 // SatMin -> Overflow && SumDiff >= 0
1288 unsigned Cost = 0;
1289 Cost += ConcreteTTI->getIntrinsicInstrCost(
1290 OverflowOp, OpTy, {RetTy, RetTy}, FMF, ScalarizationCostPassed);
1291 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, RetTy,
1292 CondTy, nullptr);
1293 Cost += 2 * ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1294 CondTy, nullptr);
1295 return Cost;
1296 }
1297 case Intrinsic::uadd_sat:
1298 case Intrinsic::usub_sat: {
1299 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1300 if (RetVF > 1)
1301 CondTy = VectorType::get(CondTy, RetVF);
1302
1303 Type *OpTy = StructType::create({RetTy, CondTy});
1304 Intrinsic::ID OverflowOp = IID == Intrinsic::uadd_sat
1305 ? Intrinsic::uadd_with_overflow
1306 : Intrinsic::usub_with_overflow;
1307
1308 unsigned Cost = 0;
1309 Cost += ConcreteTTI->getIntrinsicInstrCost(
1310 OverflowOp, OpTy, {RetTy, RetTy}, FMF, ScalarizationCostPassed);
1311 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1312 CondTy, nullptr);
1313 return Cost;
1314 }
1315 case Intrinsic::smul_fix:
1316 case Intrinsic::umul_fix: {
1317 unsigned ExtSize = RetTy->getScalarSizeInBits() * 2;
1318 Type *ExtTy = Type::getIntNTy(RetTy->getContext(), ExtSize);
1319 if (RetVF > 1)
1320 ExtTy = VectorType::get(ExtTy, RetVF);
1321
1322 unsigned ExtOp =
1323 IID == Intrinsic::smul_fix ? Instruction::SExt : Instruction::ZExt;
1324
1325 unsigned Cost = 0;
1326 Cost += 2 * ConcreteTTI->getCastInstrCost(ExtOp, ExtTy, RetTy);
1327 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Mul, ExtTy);
1328 Cost +=
1329 2 * ConcreteTTI->getCastInstrCost(Instruction::Trunc, RetTy, ExtTy);
1330 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::LShr, RetTy,
1331 TTI::OK_AnyValue,
1332 TTI::OK_UniformConstantValue);
1333 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Shl, RetTy,
1334 TTI::OK_AnyValue,
1335 TTI::OK_UniformConstantValue);
1336 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Or, RetTy);
1337 return Cost;
1338 }
1339 case Intrinsic::sadd_with_overflow:
1340 case Intrinsic::ssub_with_overflow: {
1341 Type *SumTy = RetTy->getContainedType(0);
1342 Type *OverflowTy = RetTy->getContainedType(1);
1343 unsigned Opcode = IID == Intrinsic::sadd_with_overflow
1344 ? BinaryOperator::Add
1345 : BinaryOperator::Sub;
1346
1347 // LHSSign -> LHS >= 0
1348 // RHSSign -> RHS >= 0
1349 // SumSign -> Sum >= 0
1350 //
1351 // Add:
1352 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
1353 // Sub:
1354 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
1355 unsigned Cost = 0;
1356 Cost += ConcreteTTI->getArithmeticInstrCost(Opcode, SumTy);
1357 Cost += 3 * ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, SumTy,
1358 OverflowTy, nullptr);
1359 Cost += 2 * ConcreteTTI->getCmpSelInstrCost(
1360 BinaryOperator::ICmp, OverflowTy, OverflowTy, nullptr);
1361 Cost +=
1362 ConcreteTTI->getArithmeticInstrCost(BinaryOperator::And, OverflowTy);
1363 return Cost;
1364 }
1365 case Intrinsic::uadd_with_overflow:
1366 case Intrinsic::usub_with_overflow: {
1367 Type *SumTy = RetTy->getContainedType(0);
1368 Type *OverflowTy = RetTy->getContainedType(1);
1369 unsigned Opcode = IID == Intrinsic::uadd_with_overflow
1370 ? BinaryOperator::Add
1371 : BinaryOperator::Sub;
1372
1373 unsigned Cost = 0;
1374 Cost += ConcreteTTI->getArithmeticInstrCost(Opcode, SumTy);
1375 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, SumTy,
1376 OverflowTy, nullptr);
1377 return Cost;
1378 }
1379 case Intrinsic::smul_with_overflow:
1380 case Intrinsic::umul_with_overflow: {
1381 Type *MulTy = RetTy->getContainedType(0);
1382 Type *OverflowTy = RetTy->getContainedType(1);
1383 unsigned ExtSize = MulTy->getScalarSizeInBits() * 2;
1384 Type *ExtTy = Type::getIntNTy(RetTy->getContext(), ExtSize);
1385 if (MulTy->isVectorTy())
1386 ExtTy = VectorType::get(ExtTy, MulTy->getVectorNumElements() );
1387
1388 unsigned ExtOp =
1389 IID == Intrinsic::smul_fix ? Instruction::SExt : Instruction::ZExt;
1390
1391 unsigned Cost = 0;
1392 Cost += 2 * ConcreteTTI->getCastInstrCost(ExtOp, ExtTy, MulTy);
1393 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Mul, ExtTy);
1394 Cost +=
1395 2 * ConcreteTTI->getCastInstrCost(Instruction::Trunc, MulTy, ExtTy);
1396 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::LShr, MulTy,
1397 TTI::OK_AnyValue,
1398 TTI::OK_UniformConstantValue);
1399
1400 if (IID == Intrinsic::smul_with_overflow)
1401 Cost += ConcreteTTI->getArithmeticInstrCost(
1402 Instruction::AShr, MulTy, TTI::OK_AnyValue,
1403 TTI::OK_UniformConstantValue);
1404
1405 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, MulTy,
1406 OverflowTy, nullptr);
1407 return Cost;
1408 }
1409 case Intrinsic::ctpop:
1410 ISDs.push_back(ISD::CTPOP);
1411 // In case of legalization use TCC_Expensive. This is cheaper than a
1412 // library call but still not a cheap instruction.
1413 SingleCallCost = TargetTransformInfo::TCC_Expensive;
1414 break;
1415 // FIXME: ctlz, cttz, ...
1416 }
1417
1418 const TargetLoweringBase *TLI = getTLI();
1419 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1420
1421 SmallVector<unsigned, 2> LegalCost;
1422 SmallVector<unsigned, 2> CustomCost;
1423 for (unsigned ISD : ISDs) {
1424 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
1425 if (IID == Intrinsic::fabs && LT.second.isFloatingPoint() &&
1426 TLI->isFAbsFree(LT.second)) {
1427 return 0;
1428 }
1429
1430 // The operation is legal. Assume it costs 1.
1431 // If the type is split to multiple registers, assume that there is some
1432 // overhead to this.
1433 // TODO: Once we have extract/insert subvector cost we need to use them.
1434 if (LT.first > 1)
1435 LegalCost.push_back(LT.first * 2);
1436 else
1437 LegalCost.push_back(LT.first * 1);
1438 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1439 // If the operation is custom lowered then assume
1440 // that the code is twice as expensive.
1441 CustomCost.push_back(LT.first * 2);
1442 }
1443 }
1444
1445 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1446 if (MinLegalCostI != LegalCost.end())
1447 return *MinLegalCostI;
1448
1449 auto MinCustomCostI =
1450 std::min_element(CustomCost.begin(), CustomCost.end());
1451 if (MinCustomCostI != CustomCost.end())
1452 return *MinCustomCostI;
1453
1454 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1455 // point mul followed by an add.
1456 if (IID == Intrinsic::fmuladd)
1457 return ConcreteTTI->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1458 ConcreteTTI->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1459
1460 // Else, assume that we need to scalarize this intrinsic. For math builtins
1461 // this will emit a costly libcall, adding call overhead and spills. Make it
1462 // very expensive.
1463 if (RetTy->isVectorTy()) {
1464 unsigned ScalarizationCost =
1465 ((ScalarizationCostPassed != std::numeric_limits<unsigned>::max())
1466 ? ScalarizationCostPassed
1467 : getScalarizationOverhead(RetTy, true, false));
1468 unsigned ScalarCalls = RetTy->getVectorNumElements();
1469 SmallVector<Type *, 4> ScalarTys;
1470 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1471 Type *Ty = Tys[i];
1472 if (Ty->isVectorTy())
1473 Ty = Ty->getScalarType();
1474 ScalarTys.push_back(Ty);
1475 }
1476 unsigned ScalarCost = ConcreteTTI->getIntrinsicInstrCost(
1477 IID, RetTy->getScalarType(), ScalarTys, FMF);
1478 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1479 if (Tys[i]->isVectorTy()) {
1480 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1481 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1482 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1483 }
1484 }
1485
1486 return ScalarCalls * ScalarCost + ScalarizationCost;
1487 }
1488
1489 // This is going to be turned into a library call, make it expensive.
1490 return SingleCallCost;
1491 }
1492
1493 /// Compute a cost of the given call instruction.
1494 ///
1495 /// Compute the cost of calling function F with return type RetTy and
1496 /// argument types Tys. F might be nullptr, in this case the cost of an
1497 /// arbitrary call with the specified signature will be returned.
1498 /// This is used, for instance, when we estimate call of a vector
1499 /// counterpart of the given function.
1500 /// \param F Called function, might be nullptr.
1501 /// \param RetTy Return value types.
1502 /// \param Tys Argument types.
1503 /// \returns The cost of Call instruction.
1504 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1505 return 10;
1506 }
1507
1508 unsigned getNumberOfParts(Type *Tp) {
1509 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1510 return LT.first;
1511 }
1512
1513 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1514 const SCEV *) {
1515 return 0;
1516 }
1517
1518 /// Try to calculate arithmetic and shuffle op costs for reduction operations.
1519 /// We're assuming that reduction operation are performing the following way:
1520 /// 1. Non-pairwise reduction
1521 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1522 /// <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1523 /// \----------------v-------------/ \----------v------------/
1524 /// n/2 elements n/2 elements
1525 /// %red1 = op <n x t> %val, <n x t> val1
1526 /// After this operation we have a vector %red1 where only the first n/2
1527 /// elements are meaningful, the second n/2 elements are undefined and can be
1528 /// dropped. All other operations are actually working with the vector of
1529 /// length n/2, not n, though the real vector length is still n.
1530 /// %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1531 /// <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1532 /// \----------------v-------------/ \----------v------------/
1533 /// n/4 elements 3*n/4 elements
1534 /// %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1535 /// length n/2, the resulting vector has length n/4 etc.
1536 /// 2. Pairwise reduction:
1537 /// Everything is the same except for an additional shuffle operation which
1538 /// is used to produce operands for pairwise kind of reductions.
1539 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1540 /// <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1541 /// \-------------v----------/ \----------v------------/
1542 /// n/2 elements n/2 elements
1543 /// %val2 = shufflevector<n x t> %val, <n x t> %undef,
1544 /// <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1545 /// \-------------v----------/ \----------v------------/
1546 /// n/2 elements n/2 elements
1547 /// %red1 = op <n x t> %val1, <n x t> val2
1548 /// Again, the operation is performed on <n x t> vector, but the resulting
1549 /// vector %red1 is <n/2 x t> vector.
1550 ///
1551 /// The cost model should take into account that the actual length of the
1552 /// vector is reduced on each iteration.
1553 unsigned getArithmeticReductionCost(unsigned Opcode, Type *Ty,
1554 bool IsPairwise) {
1555 assert(Ty->isVectorTy() && "Expect a vector type")((Ty->isVectorTy() && "Expect a vector type") ? static_cast
<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Expect a vector type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1555, __PRETTY_FUNCTION__))
;
1556 Type *ScalarTy = Ty->getVectorElementType();
1557 unsigned NumVecElts = Ty->getVectorNumElements();
1558 unsigned NumReduxLevels = Log2_32(NumVecElts);
1559 unsigned ArithCost = 0;
1560 unsigned ShuffleCost = 0;
1561 auto *ConcreteTTI = static_cast<T *>(this);
1562 std::pair<unsigned, MVT> LT =
1563 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1564 unsigned LongVectorCount = 0;
1565 unsigned MVTLen =
1566 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1567 while (NumVecElts > MVTLen) {
1568 NumVecElts /= 2;
1569 Type *SubTy = VectorType::get(ScalarTy, NumVecElts);
1570 // Assume the pairwise shuffles add a cost.
1571 ShuffleCost += (IsPairwise + 1) *
1572 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1573 NumVecElts, SubTy);
1574 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, SubTy);
1575 Ty = SubTy;
1576 ++LongVectorCount;
1577 }
1578
1579 NumReduxLevels -= LongVectorCount;
1580
1581 // The minimal length of the vector is limited by the real length of vector
1582 // operations performed on the current platform. That's why several final
1583 // reduction operations are performed on the vectors with the same
1584 // architecture-dependent length.
1585
1586 // Non pairwise reductions need one shuffle per reduction level. Pairwise
1587 // reductions need two shuffles on every level, but the last one. On that
1588 // level one of the shuffles is <0, u, u, ...> which is identity.
1589 unsigned NumShuffles = NumReduxLevels;
1590 if (IsPairwise && NumReduxLevels >= 1)
1591 NumShuffles += NumReduxLevels - 1;
1592 ShuffleCost += NumShuffles *
1593 ConcreteTTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty,
1594 0, Ty);
1595 ArithCost += NumReduxLevels *
1596 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1597 return ShuffleCost + ArithCost +
1598 ConcreteTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
1599 }
1600
1601 /// Try to calculate op costs for min/max reduction operations.
1602 /// \param CondTy Conditional type for the Select instruction.
1603 unsigned getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwise,
1604 bool) {
1605 assert(Ty->isVectorTy() && "Expect a vector type")((Ty->isVectorTy() && "Expect a vector type") ? static_cast
<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Expect a vector type\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1605, __PRETTY_FUNCTION__))
;
1606 Type *ScalarTy = Ty->getVectorElementType();
1607 Type *ScalarCondTy = CondTy->getVectorElementType();
1608 unsigned NumVecElts = Ty->getVectorNumElements();
1609 unsigned NumReduxLevels = Log2_32(NumVecElts);
1610 unsigned CmpOpcode;
1611 if (Ty->isFPOrFPVectorTy()) {
1612 CmpOpcode = Instruction::FCmp;
1613 } else {
1614 assert(Ty->isIntOrIntVectorTy() &&((Ty->isIntOrIntVectorTy() && "expecting floating point or integer type for min/max reduction"
) ? static_cast<void> (0) : __assert_fail ("Ty->isIntOrIntVectorTy() && \"expecting floating point or integer type for min/max reduction\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1615, __PRETTY_FUNCTION__))
1615 "expecting floating point or integer type for min/max reduction")((Ty->isIntOrIntVectorTy() && "expecting floating point or integer type for min/max reduction"
) ? static_cast<void> (0) : __assert_fail ("Ty->isIntOrIntVectorTy() && \"expecting floating point or integer type for min/max reduction\""
, "/build/llvm-toolchain-snapshot-9~svn362543/include/llvm/CodeGen/BasicTTIImpl.h"
, 1615, __PRETTY_FUNCTION__))
;
1616 CmpOpcode = Instruction::ICmp;
1617 }
1618 unsigned MinMaxCost = 0;
1619 unsigned ShuffleCost = 0;
1620 auto *ConcreteTTI = static_cast<T *>(this);
1621 std::pair<unsigned, MVT> LT =
1622 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1623 unsigned LongVectorCount = 0;
1624 unsigned MVTLen =
1625 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1626 while (NumVecElts > MVTLen) {
1627 NumVecElts /= 2;
1628 Type *SubTy = VectorType::get(ScalarTy, NumVecElts);
1629 CondTy = VectorType::get(ScalarCondTy, NumVecElts);
1630
1631 // Assume the pairwise shuffles add a cost.
1632 ShuffleCost += (IsPairwise + 1) *
1633 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1634 NumVecElts, SubTy);
1635 MinMaxCost +=
1636 ConcreteTTI->getCmpSelInstrCost(CmpOpcode, SubTy, CondTy, nullptr) +
1637 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, SubTy, CondTy,
1638 nullptr);
1639 Ty = SubTy;
1640 ++LongVectorCount;
1641 }
1642
1643 NumReduxLevels -= LongVectorCount;
1644
1645 // The minimal length of the vector is limited by the real length of vector
1646 // operations performed on the current platform. That's why several final
1647 // reduction opertions are perfomed on the vectors with the same
1648 // architecture-dependent length.
1649
1650 // Non pairwise reductions need one shuffle per reduction level. Pairwise
1651 // reductions need two shuffles on every level, but the last one. On that
1652 // level one of the shuffles is <0, u, u, ...> which is identity.
1653 unsigned NumShuffles = NumReduxLevels;
1654 if (IsPairwise && NumReduxLevels >= 1)
1655 NumShuffles += NumReduxLevels - 1;
1656 ShuffleCost += NumShuffles *
1657 ConcreteTTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty,
1658 0, Ty);
1659 MinMaxCost +=
1660 NumReduxLevels *
1661 (ConcreteTTI->getCmpSelInstrCost(CmpOpcode, Ty, CondTy, nullptr) +
1662 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
1663 nullptr));
1664 // The last min/max should be in vector registers and we counted it above.
1665 // So just need a single extractelement.
1666 return ShuffleCost + MinMaxCost +
1667 ConcreteTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
1668 }
1669
1670 unsigned getVectorSplitCost() { return 1; }
1671
1672 /// @}
1673};
1674
1675/// Concrete BasicTTIImpl that can be used if no further customization
1676/// is needed.
1677class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1678 using BaseT = BasicTTIImplBase<BasicTTIImpl>;
1679
1680 friend class BasicTTIImplBase<BasicTTIImpl>;
1681
1682 const TargetSubtargetInfo *ST;
1683 const TargetLoweringBase *TLI;
1684
1685 const TargetSubtargetInfo *getST() const { return ST; }
1686 const TargetLoweringBase *getTLI() const { return TLI; }
1687
1688public:
1689 explicit BasicTTIImpl(const TargetMachine *TM, const Function &F);
1690};
1691
1692} // end namespace llvm
1693
1694#endif // LLVM_CODEGEN_BASICTTIIMPL_H