1 | //===-- HexagonVectorCombine.cpp ------------------------------------------===// |
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 | // HexagonVectorCombine is a utility class implementing a variety of functions |
9 | // that assist in vector-based optimizations. |
10 | // |
11 | // AlignVectors: replace unaligned vector loads and stores with aligned ones. |
12 | // HvxIdioms: recognize various opportunities to generate HVX intrinsic code. |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/ADT/APInt.h" |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/DenseMap.h" |
18 | #include "llvm/ADT/STLExtras.h" |
19 | #include "llvm/ADT/SmallVector.h" |
20 | #include "llvm/Analysis/AliasAnalysis.h" |
21 | #include "llvm/Analysis/AssumptionCache.h" |
22 | #include "llvm/Analysis/InstSimplifyFolder.h" |
23 | #include "llvm/Analysis/InstructionSimplify.h" |
24 | #include "llvm/Analysis/ScalarEvolution.h" |
25 | #include "llvm/Analysis/TargetLibraryInfo.h" |
26 | #include "llvm/Analysis/ValueTracking.h" |
27 | #include "llvm/Analysis/VectorUtils.h" |
28 | #include "llvm/CodeGen/TargetPassConfig.h" |
29 | #include "llvm/CodeGen/ValueTypes.h" |
30 | #include "llvm/IR/Dominators.h" |
31 | #include "llvm/IR/IRBuilder.h" |
32 | #include "llvm/IR/IntrinsicInst.h" |
33 | #include "llvm/IR/Intrinsics.h" |
34 | #include "llvm/IR/IntrinsicsHexagon.h" |
35 | #include "llvm/IR/Metadata.h" |
36 | #include "llvm/IR/PatternMatch.h" |
37 | #include "llvm/InitializePasses.h" |
38 | #include "llvm/Pass.h" |
39 | #include "llvm/Support/CommandLine.h" |
40 | #include "llvm/Support/KnownBits.h" |
41 | #include "llvm/Support/MathExtras.h" |
42 | #include "llvm/Support/raw_ostream.h" |
43 | #include "llvm/Target/TargetMachine.h" |
44 | #include "llvm/Transforms/Utils/Local.h" |
45 | |
46 | #include "HexagonSubtarget.h" |
47 | #include "HexagonTargetMachine.h" |
48 | |
49 | #include <algorithm> |
50 | #include <deque> |
51 | #include <map> |
52 | #include <optional> |
53 | #include <set> |
54 | #include <utility> |
55 | #include <vector> |
56 | |
57 | #define DEBUG_TYPE "hexagon-vc" |
58 | |
59 | using namespace llvm; |
60 | |
61 | namespace { |
62 | cl::opt<bool> DumpModule("hvc-dump-module" , cl::Hidden); |
63 | cl::opt<bool> VAEnabled("hvc-va" , cl::Hidden, cl::init(Val: true)); // Align |
64 | cl::opt<bool> VIEnabled("hvc-vi" , cl::Hidden, cl::init(Val: true)); // Idioms |
65 | cl::opt<bool> VADoFullStores("hvc-va-full-stores" , cl::Hidden); |
66 | |
67 | cl::opt<unsigned> VAGroupCountLimit("hvc-va-group-count-limit" , cl::Hidden, |
68 | cl::init(Val: ~0)); |
69 | cl::opt<unsigned> VAGroupSizeLimit("hvc-va-group-size-limit" , cl::Hidden, |
70 | cl::init(Val: ~0)); |
71 | |
72 | class HexagonVectorCombine { |
73 | public: |
74 | HexagonVectorCombine(Function &F_, AliasAnalysis &AA_, AssumptionCache &AC_, |
75 | DominatorTree &DT_, ScalarEvolution &SE_, |
76 | TargetLibraryInfo &TLI_, const TargetMachine &TM_) |
77 | : F(F_), DL(F.getDataLayout()), AA(AA_), AC(AC_), DT(DT_), |
78 | SE(SE_), TLI(TLI_), |
79 | HST(static_cast<const HexagonSubtarget &>(*TM_.getSubtargetImpl(F))) {} |
80 | |
81 | bool run(); |
82 | |
83 | // Common integer type. |
84 | IntegerType *getIntTy(unsigned Width = 32) const; |
85 | // Byte type: either scalar (when Length = 0), or vector with given |
86 | // element count. |
87 | Type *getByteTy(int ElemCount = 0) const; |
88 | // Boolean type: either scalar (when Length = 0), or vector with given |
89 | // element count. |
90 | Type *getBoolTy(int ElemCount = 0) const; |
91 | // Create a ConstantInt of type returned by getIntTy with the value Val. |
92 | ConstantInt *getConstInt(int Val, unsigned Width = 32) const; |
93 | // Get the integer value of V, if it exists. |
94 | std::optional<APInt> getIntValue(const Value *Val) const; |
95 | // Is Val a constant 0, or a vector of 0s? |
96 | bool isZero(const Value *Val) const; |
97 | // Is Val an undef value? |
98 | bool isUndef(const Value *Val) const; |
99 | // Is Val a scalar (i1 true) or a vector of (i1 true)? |
100 | bool isTrue(const Value *Val) const; |
101 | // Is Val a scalar (i1 false) or a vector of (i1 false)? |
102 | bool isFalse(const Value *Val) const; |
103 | |
104 | // Get HVX vector type with the given element type. |
105 | VectorType *getHvxTy(Type *ElemTy, bool Pair = false) const; |
106 | |
107 | enum SizeKind { |
108 | Store, // Store size |
109 | Alloc, // Alloc size |
110 | }; |
111 | int getSizeOf(const Value *Val, SizeKind Kind = Store) const; |
112 | int getSizeOf(const Type *Ty, SizeKind Kind = Store) const; |
113 | int getTypeAlignment(Type *Ty) const; |
114 | size_t length(Value *Val) const; |
115 | size_t length(Type *Ty) const; |
116 | |
117 | Constant *getNullValue(Type *Ty) const; |
118 | Constant *getFullValue(Type *Ty) const; |
119 | Constant *getConstSplat(Type *Ty, int Val) const; |
120 | |
121 | Value *simplify(Value *Val) const; |
122 | |
123 | Value *insertb(IRBuilderBase &Builder, Value *Dest, Value *Src, int Start, |
124 | int Length, int Where) const; |
125 | Value *vlalignb(IRBuilderBase &Builder, Value *Lo, Value *Hi, |
126 | Value *Amt) const; |
127 | Value *vralignb(IRBuilderBase &Builder, Value *Lo, Value *Hi, |
128 | Value *Amt) const; |
129 | Value *concat(IRBuilderBase &Builder, ArrayRef<Value *> Vecs) const; |
130 | Value *vresize(IRBuilderBase &Builder, Value *Val, int NewSize, |
131 | Value *Pad) const; |
132 | Value *rescale(IRBuilderBase &Builder, Value *Mask, Type *FromTy, |
133 | Type *ToTy) const; |
134 | Value *vlsb(IRBuilderBase &Builder, Value *Val) const; |
135 | Value *vbytes(IRBuilderBase &Builder, Value *Val) const; |
136 | Value *subvector(IRBuilderBase &Builder, Value *Val, unsigned Start, |
137 | unsigned Length) const; |
138 | Value *sublo(IRBuilderBase &Builder, Value *Val) const; |
139 | Value *subhi(IRBuilderBase &Builder, Value *Val) const; |
140 | Value *vdeal(IRBuilderBase &Builder, Value *Val0, Value *Val1) const; |
141 | Value *vshuff(IRBuilderBase &Builder, Value *Val0, Value *Val1) const; |
142 | |
143 | Value *createHvxIntrinsic(IRBuilderBase &Builder, Intrinsic::ID IntID, |
144 | Type *RetTy, ArrayRef<Value *> Args, |
145 | ArrayRef<Type *> ArgTys = std::nullopt, |
146 | ArrayRef<Value *> MDSources = std::nullopt) const; |
147 | SmallVector<Value *> splitVectorElements(IRBuilderBase &Builder, Value *Vec, |
148 | unsigned ToWidth) const; |
149 | Value *joinVectorElements(IRBuilderBase &Builder, ArrayRef<Value *> Values, |
150 | VectorType *ToType) const; |
151 | |
152 | std::optional<int> calculatePointerDifference(Value *Ptr0, Value *Ptr1) const; |
153 | |
154 | unsigned getNumSignificantBits(const Value *V, |
155 | const Instruction *CtxI = nullptr) const; |
156 | KnownBits getKnownBits(const Value *V, |
157 | const Instruction *CtxI = nullptr) const; |
158 | |
159 | bool isSafeToClone(const Instruction &In) const; |
160 | |
161 | template <typename T = std::vector<Instruction *>> |
162 | bool isSafeToMoveBeforeInBB(const Instruction &In, |
163 | BasicBlock::const_iterator To, |
164 | const T &IgnoreInsts = {}) const; |
165 | |
166 | // This function is only used for assertions at the moment. |
167 | [[maybe_unused]] bool isByteVecTy(Type *Ty) const; |
168 | |
169 | Function &F; |
170 | const DataLayout &DL; |
171 | AliasAnalysis &AA; |
172 | AssumptionCache &AC; |
173 | DominatorTree &DT; |
174 | ScalarEvolution &SE; |
175 | TargetLibraryInfo &TLI; |
176 | const HexagonSubtarget &HST; |
177 | |
178 | private: |
179 | Value *getElementRange(IRBuilderBase &Builder, Value *Lo, Value *Hi, |
180 | int Start, int Length) const; |
181 | }; |
182 | |
183 | class AlignVectors { |
184 | // This code tries to replace unaligned vector loads/stores with aligned |
185 | // ones. |
186 | // Consider unaligned load: |
187 | // %v = original_load %some_addr, align <bad> |
188 | // %user = %v |
189 | // It will generate |
190 | // = load ..., align <good> |
191 | // = load ..., align <good> |
192 | // = valign |
193 | // etc. |
194 | // %synthesize = combine/shuffle the loaded data so that it looks |
195 | // exactly like what "original_load" has loaded. |
196 | // %user = %synthesize |
197 | // Similarly for stores. |
198 | public: |
199 | AlignVectors(const HexagonVectorCombine &HVC_) : HVC(HVC_) {} |
200 | |
201 | bool run(); |
202 | |
203 | private: |
204 | using InstList = std::vector<Instruction *>; |
205 | using InstMap = DenseMap<Instruction *, Instruction *>; |
206 | |
207 | struct AddrInfo { |
208 | AddrInfo(const AddrInfo &) = default; |
209 | AddrInfo(const HexagonVectorCombine &HVC, Instruction *I, Value *A, Type *T, |
210 | Align H) |
211 | : Inst(I), Addr(A), ValTy(T), HaveAlign(H), |
212 | NeedAlign(HVC.getTypeAlignment(Ty: ValTy)) {} |
213 | AddrInfo &operator=(const AddrInfo &) = default; |
214 | |
215 | // XXX: add Size member? |
216 | Instruction *Inst; |
217 | Value *Addr; |
218 | Type *ValTy; |
219 | Align HaveAlign; |
220 | Align NeedAlign; |
221 | int Offset = 0; // Offset (in bytes) from the first member of the |
222 | // containing AddrList. |
223 | }; |
224 | using AddrList = std::vector<AddrInfo>; |
225 | |
226 | struct InstrLess { |
227 | bool operator()(const Instruction *A, const Instruction *B) const { |
228 | return A->comesBefore(Other: B); |
229 | } |
230 | }; |
231 | using DepList = std::set<Instruction *, InstrLess>; |
232 | |
233 | struct MoveGroup { |
234 | MoveGroup(const AddrInfo &AI, Instruction *B, bool Hvx, bool Load) |
235 | : Base(B), Main{AI.Inst}, Clones{}, IsHvx(Hvx), IsLoad(Load) {} |
236 | MoveGroup() = default; |
237 | Instruction *Base; // Base instruction of the parent address group. |
238 | InstList Main; // Main group of instructions. |
239 | InstList Deps; // List of dependencies. |
240 | InstMap Clones; // Map from original Deps to cloned ones. |
241 | bool IsHvx; // Is this group of HVX instructions? |
242 | bool IsLoad; // Is this a load group? |
243 | }; |
244 | using MoveList = std::vector<MoveGroup>; |
245 | |
246 | struct ByteSpan { |
247 | // A representation of "interesting" bytes within a given span of memory. |
248 | // These bytes are those that are loaded or stored, and they don't have |
249 | // to cover the entire span of memory. |
250 | // |
251 | // The representation works by picking a contiguous sequence of bytes |
252 | // from somewhere within a llvm::Value, and placing it at a given offset |
253 | // within the span. |
254 | // |
255 | // The sequence of bytes from llvm:Value is represented by Segment. |
256 | // Block is Segment, plus where it goes in the span. |
257 | // |
258 | // An important feature of ByteSpan is being able to make a "section", |
259 | // i.e. creating another ByteSpan corresponding to a range of offsets |
260 | // relative to the source span. |
261 | |
262 | struct Segment { |
263 | // Segment of a Value: 'Len' bytes starting at byte 'Begin'. |
264 | Segment(Value *Val, int Begin, int Len) |
265 | : Val(Val), Start(Begin), Size(Len) {} |
266 | Segment(const Segment &Seg) = default; |
267 | Segment &operator=(const Segment &Seg) = default; |
268 | Value *Val; // Value representable as a sequence of bytes. |
269 | int Start; // First byte of the value that belongs to the segment. |
270 | int Size; // Number of bytes in the segment. |
271 | }; |
272 | |
273 | struct Block { |
274 | Block(Value *Val, int Len, int Pos) : Seg(Val, 0, Len), Pos(Pos) {} |
275 | Block(Value *Val, int Off, int Len, int Pos) |
276 | : Seg(Val, Off, Len), Pos(Pos) {} |
277 | Block(const Block &Blk) = default; |
278 | Block &operator=(const Block &Blk) = default; |
279 | Segment Seg; // Value segment. |
280 | int Pos; // Position (offset) of the block in the span. |
281 | }; |
282 | |
283 | int extent() const; |
284 | ByteSpan section(int Start, int Length) const; |
285 | ByteSpan &shift(int Offset); |
286 | SmallVector<Value *, 8> values() const; |
287 | |
288 | int size() const { return Blocks.size(); } |
289 | Block &operator[](int i) { return Blocks[i]; } |
290 | const Block &operator[](int i) const { return Blocks[i]; } |
291 | |
292 | std::vector<Block> Blocks; |
293 | |
294 | using iterator = decltype(Blocks)::iterator; |
295 | iterator begin() { return Blocks.begin(); } |
296 | iterator end() { return Blocks.end(); } |
297 | using const_iterator = decltype(Blocks)::const_iterator; |
298 | const_iterator begin() const { return Blocks.begin(); } |
299 | const_iterator end() const { return Blocks.end(); } |
300 | }; |
301 | |
302 | Align getAlignFromValue(const Value *V) const; |
303 | std::optional<AddrInfo> getAddrInfo(Instruction &In) const; |
304 | bool isHvx(const AddrInfo &AI) const; |
305 | // This function is only used for assertions at the moment. |
306 | [[maybe_unused]] bool isSectorTy(Type *Ty) const; |
307 | |
308 | Value *getPayload(Value *Val) const; |
309 | Value *getMask(Value *Val) const; |
310 | Value *getPassThrough(Value *Val) const; |
311 | |
312 | Value *createAdjustedPointer(IRBuilderBase &Builder, Value *Ptr, Type *ValTy, |
313 | int Adjust, |
314 | const InstMap &CloneMap = InstMap()) const; |
315 | Value *createAlignedPointer(IRBuilderBase &Builder, Value *Ptr, Type *ValTy, |
316 | int Alignment, |
317 | const InstMap &CloneMap = InstMap()) const; |
318 | |
319 | Value *createLoad(IRBuilderBase &Builder, Type *ValTy, Value *Ptr, |
320 | Value *Predicate, int Alignment, Value *Mask, |
321 | Value *PassThru, |
322 | ArrayRef<Value *> MDSources = std::nullopt) const; |
323 | Value *createSimpleLoad(IRBuilderBase &Builder, Type *ValTy, Value *Ptr, |
324 | int Alignment, |
325 | ArrayRef<Value *> MDSources = std::nullopt) const; |
326 | |
327 | Value *createStore(IRBuilderBase &Builder, Value *Val, Value *Ptr, |
328 | Value *Predicate, int Alignment, Value *Mask, |
329 | ArrayRef<Value *> MDSources = std ::nullopt) const; |
330 | Value *createSimpleStore(IRBuilderBase &Builder, Value *Val, Value *Ptr, |
331 | int Alignment, |
332 | ArrayRef<Value *> MDSources = std ::nullopt) const; |
333 | |
334 | Value *createPredicatedLoad(IRBuilderBase &Builder, Type *ValTy, Value *Ptr, |
335 | Value *Predicate, int Alignment, |
336 | ArrayRef<Value *> MDSources = std::nullopt) const; |
337 | Value * |
338 | createPredicatedStore(IRBuilderBase &Builder, Value *Val, Value *Ptr, |
339 | Value *Predicate, int Alignment, |
340 | ArrayRef<Value *> MDSources = std::nullopt) const; |
341 | |
342 | DepList getUpwardDeps(Instruction *In, Instruction *Base) const; |
343 | bool createAddressGroups(); |
344 | MoveList createLoadGroups(const AddrList &Group) const; |
345 | MoveList createStoreGroups(const AddrList &Group) const; |
346 | bool moveTogether(MoveGroup &Move) const; |
347 | template <typename T> InstMap cloneBefore(Instruction *To, T &&Insts) const; |
348 | |
349 | void realignLoadGroup(IRBuilderBase &Builder, const ByteSpan &VSpan, |
350 | int ScLen, Value *AlignVal, Value *AlignAddr) const; |
351 | void realignStoreGroup(IRBuilderBase &Builder, const ByteSpan &VSpan, |
352 | int ScLen, Value *AlignVal, Value *AlignAddr) const; |
353 | bool realignGroup(const MoveGroup &Move) const; |
354 | |
355 | Value *makeTestIfUnaligned(IRBuilderBase &Builder, Value *AlignVal, |
356 | int Alignment) const; |
357 | |
358 | friend raw_ostream &operator<<(raw_ostream &OS, const AddrInfo &AI); |
359 | friend raw_ostream &operator<<(raw_ostream &OS, const MoveGroup &MG); |
360 | friend raw_ostream &operator<<(raw_ostream &OS, const ByteSpan::Block &B); |
361 | friend raw_ostream &operator<<(raw_ostream &OS, const ByteSpan &BS); |
362 | |
363 | std::map<Instruction *, AddrList> AddrGroups; |
364 | const HexagonVectorCombine &HVC; |
365 | }; |
366 | |
367 | LLVM_ATTRIBUTE_UNUSED |
368 | raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::AddrInfo &AI) { |
369 | OS << "Inst: " << AI.Inst << " " << *AI.Inst << '\n'; |
370 | OS << "Addr: " << *AI.Addr << '\n'; |
371 | OS << "Type: " << *AI.ValTy << '\n'; |
372 | OS << "HaveAlign: " << AI.HaveAlign.value() << '\n'; |
373 | OS << "NeedAlign: " << AI.NeedAlign.value() << '\n'; |
374 | OS << "Offset: " << AI.Offset; |
375 | return OS; |
376 | } |
377 | |
378 | LLVM_ATTRIBUTE_UNUSED |
379 | raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::MoveGroup &MG) { |
380 | OS << "IsLoad:" << (MG.IsLoad ? "yes" : "no" ); |
381 | OS << ", IsHvx:" << (MG.IsHvx ? "yes" : "no" ) << '\n'; |
382 | OS << "Main\n" ; |
383 | for (Instruction *I : MG.Main) |
384 | OS << " " << *I << '\n'; |
385 | OS << "Deps\n" ; |
386 | for (Instruction *I : MG.Deps) |
387 | OS << " " << *I << '\n'; |
388 | OS << "Clones\n" ; |
389 | for (auto [K, V] : MG.Clones) { |
390 | OS << " " ; |
391 | K->printAsOperand(O&: OS, PrintType: false); |
392 | OS << "\t-> " << *V << '\n'; |
393 | } |
394 | return OS; |
395 | } |
396 | |
397 | LLVM_ATTRIBUTE_UNUSED |
398 | raw_ostream &operator<<(raw_ostream &OS, |
399 | const AlignVectors::ByteSpan::Block &B) { |
400 | OS << " @" << B.Pos << " [" << B.Seg.Start << ',' << B.Seg.Size << "] " ; |
401 | if (B.Seg.Val == reinterpret_cast<const Value *>(&B)) { |
402 | OS << "(self:" << B.Seg.Val << ')'; |
403 | } else if (B.Seg.Val != nullptr) { |
404 | OS << *B.Seg.Val; |
405 | } else { |
406 | OS << "(null)" ; |
407 | } |
408 | return OS; |
409 | } |
410 | |
411 | LLVM_ATTRIBUTE_UNUSED |
412 | raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::ByteSpan &BS) { |
413 | OS << "ByteSpan[size=" << BS.size() << ", extent=" << BS.extent() << '\n'; |
414 | for (const AlignVectors::ByteSpan::Block &B : BS) |
415 | OS << B << '\n'; |
416 | OS << ']'; |
417 | return OS; |
418 | } |
419 | |
420 | class HvxIdioms { |
421 | public: |
422 | HvxIdioms(const HexagonVectorCombine &HVC_) : HVC(HVC_) { |
423 | auto *Int32Ty = HVC.getIntTy(Width: 32); |
424 | HvxI32Ty = HVC.getHvxTy(ElemTy: Int32Ty, /*Pair=*/false); |
425 | HvxP32Ty = HVC.getHvxTy(ElemTy: Int32Ty, /*Pair=*/true); |
426 | } |
427 | |
428 | bool run(); |
429 | |
430 | private: |
431 | enum Signedness { Positive, Signed, Unsigned }; |
432 | |
433 | // Value + sign |
434 | // This is to keep track of whether the value should be treated as signed |
435 | // or unsigned, or is known to be positive. |
436 | struct SValue { |
437 | Value *Val; |
438 | Signedness Sgn; |
439 | }; |
440 | |
441 | struct FxpOp { |
442 | unsigned Opcode; |
443 | unsigned Frac; // Number of fraction bits |
444 | SValue X, Y; |
445 | // If present, add 1 << RoundAt before shift: |
446 | std::optional<unsigned> RoundAt; |
447 | VectorType *ResTy; |
448 | }; |
449 | |
450 | auto getNumSignificantBits(Value *V, Instruction *In) const |
451 | -> std::pair<unsigned, Signedness>; |
452 | auto canonSgn(SValue X, SValue Y) const -> std::pair<SValue, SValue>; |
453 | |
454 | auto matchFxpMul(Instruction &In) const -> std::optional<FxpOp>; |
455 | auto processFxpMul(Instruction &In, const FxpOp &Op) const -> Value *; |
456 | |
457 | auto processFxpMulChopped(IRBuilderBase &Builder, Instruction &In, |
458 | const FxpOp &Op) const -> Value *; |
459 | auto createMulQ15(IRBuilderBase &Builder, SValue X, SValue Y, |
460 | bool Rounding) const -> Value *; |
461 | auto createMulQ31(IRBuilderBase &Builder, SValue X, SValue Y, |
462 | bool Rounding) const -> Value *; |
463 | // Return {Result, Carry}, where Carry is a vector predicate. |
464 | auto createAddCarry(IRBuilderBase &Builder, Value *X, Value *Y, |
465 | Value *CarryIn = nullptr) const |
466 | -> std::pair<Value *, Value *>; |
467 | auto createMul16(IRBuilderBase &Builder, SValue X, SValue Y) const -> Value *; |
468 | auto createMulH16(IRBuilderBase &Builder, SValue X, SValue Y) const |
469 | -> Value *; |
470 | auto createMul32(IRBuilderBase &Builder, SValue X, SValue Y) const |
471 | -> std::pair<Value *, Value *>; |
472 | auto createAddLong(IRBuilderBase &Builder, ArrayRef<Value *> WordX, |
473 | ArrayRef<Value *> WordY) const -> SmallVector<Value *>; |
474 | auto createMulLong(IRBuilderBase &Builder, ArrayRef<Value *> WordX, |
475 | Signedness SgnX, ArrayRef<Value *> WordY, |
476 | Signedness SgnY) const -> SmallVector<Value *>; |
477 | |
478 | VectorType *HvxI32Ty; |
479 | VectorType *HvxP32Ty; |
480 | const HexagonVectorCombine &HVC; |
481 | |
482 | friend raw_ostream &operator<<(raw_ostream &, const FxpOp &); |
483 | }; |
484 | |
485 | [[maybe_unused]] raw_ostream &operator<<(raw_ostream &OS, |
486 | const HvxIdioms::FxpOp &Op) { |
487 | static const char *SgnNames[] = {"Positive" , "Signed" , "Unsigned" }; |
488 | OS << Instruction::getOpcodeName(Opcode: Op.Opcode) << '.' << Op.Frac; |
489 | if (Op.RoundAt.has_value()) { |
490 | if (Op.Frac != 0 && *Op.RoundAt == Op.Frac - 1) { |
491 | OS << ":rnd" ; |
492 | } else { |
493 | OS << " + 1<<" << *Op.RoundAt; |
494 | } |
495 | } |
496 | OS << "\n X:(" << SgnNames[Op.X.Sgn] << ") " << *Op.X.Val << "\n" |
497 | << " Y:(" << SgnNames[Op.Y.Sgn] << ") " << *Op.Y.Val; |
498 | return OS; |
499 | } |
500 | |
501 | } // namespace |
502 | |
503 | namespace { |
504 | |
505 | template <typename T> T *getIfUnordered(T *MaybeT) { |
506 | return MaybeT && MaybeT->isUnordered() ? MaybeT : nullptr; |
507 | } |
508 | template <typename T> T *isCandidate(Instruction *In) { |
509 | return dyn_cast<T>(In); |
510 | } |
511 | template <> LoadInst *isCandidate<LoadInst>(Instruction *In) { |
512 | return getIfUnordered(MaybeT: dyn_cast<LoadInst>(Val: In)); |
513 | } |
514 | template <> StoreInst *isCandidate<StoreInst>(Instruction *In) { |
515 | return getIfUnordered(MaybeT: dyn_cast<StoreInst>(Val: In)); |
516 | } |
517 | |
518 | #if !defined(_MSC_VER) || _MSC_VER >= 1926 |
519 | // VS2017 and some versions of VS2019 have trouble compiling this: |
520 | // error C2976: 'std::map': too few template arguments |
521 | // VS 2019 16.x is known to work, except for 16.4/16.5 (MSC_VER 1924/1925) |
522 | template <typename Pred, typename... Ts> |
523 | void erase_if(std::map<Ts...> &map, Pred p) |
524 | #else |
525 | template <typename Pred, typename T, typename U> |
526 | void erase_if(std::map<T, U> &map, Pred p) |
527 | #endif |
528 | { |
529 | for (auto i = map.begin(), e = map.end(); i != e;) { |
530 | if (p(*i)) |
531 | i = map.erase(i); |
532 | else |
533 | i = std::next(i); |
534 | } |
535 | } |
536 | |
537 | // Forward other erase_ifs to the LLVM implementations. |
538 | template <typename Pred, typename T> void erase_if(T &&container, Pred p) { |
539 | llvm::erase_if(std::forward<T>(container), p); |
540 | } |
541 | |
542 | } // namespace |
543 | |
544 | // --- Begin AlignVectors |
545 | |
546 | // For brevity, only consider loads. We identify a group of loads where we |
547 | // know the relative differences between their addresses, so we know how they |
548 | // are laid out in memory (relative to one another). These loads can overlap, |
549 | // can be shorter or longer than the desired vector length. |
550 | // Ultimately we want to generate a sequence of aligned loads that will load |
551 | // every byte that the original loads loaded, and have the program use these |
552 | // loaded values instead of the original loads. |
553 | // We consider the contiguous memory area spanned by all these loads. |
554 | // |
555 | // Let's say that a single aligned vector load can load 16 bytes at a time. |
556 | // If the program wanted to use a byte at offset 13 from the beginning of the |
557 | // original span, it will be a byte at offset 13+x in the aligned data for |
558 | // some x>=0. This may happen to be in the first aligned load, or in the load |
559 | // following it. Since we generally don't know what the that alignment value |
560 | // is at compile time, we proactively do valigns on the aligned loads, so that |
561 | // byte that was at offset 13 is still at offset 13 after the valigns. |
562 | // |
563 | // This will be the starting point for making the rest of the program use the |
564 | // data loaded by the new loads. |
565 | // For each original load, and its users: |
566 | // %v = load ... |
567 | // ... = %v |
568 | // ... = %v |
569 | // we create |
570 | // %new_v = extract/combine/shuffle data from loaded/valigned vectors so |
571 | // it contains the same value as %v did before |
572 | // then replace all users of %v with %new_v. |
573 | // ... = %new_v |
574 | // ... = %new_v |
575 | |
576 | auto AlignVectors::ByteSpan::extent() const -> int { |
577 | if (size() == 0) |
578 | return 0; |
579 | int Min = Blocks[0].Pos; |
580 | int Max = Blocks[0].Pos + Blocks[0].Seg.Size; |
581 | for (int i = 1, e = size(); i != e; ++i) { |
582 | Min = std::min(a: Min, b: Blocks[i].Pos); |
583 | Max = std::max(a: Max, b: Blocks[i].Pos + Blocks[i].Seg.Size); |
584 | } |
585 | return Max - Min; |
586 | } |
587 | |
588 | auto AlignVectors::ByteSpan::section(int Start, int Length) const -> ByteSpan { |
589 | ByteSpan Section; |
590 | for (const ByteSpan::Block &B : Blocks) { |
591 | int L = std::max(a: B.Pos, b: Start); // Left end. |
592 | int R = std::min(a: B.Pos + B.Seg.Size, b: Start + Length); // Right end+1. |
593 | if (L < R) { |
594 | // How much to chop off the beginning of the segment: |
595 | int Off = L > B.Pos ? L - B.Pos : 0; |
596 | Section.Blocks.emplace_back(args: B.Seg.Val, args: B.Seg.Start + Off, args: R - L, args&: L); |
597 | } |
598 | } |
599 | return Section; |
600 | } |
601 | |
602 | auto AlignVectors::ByteSpan::shift(int Offset) -> ByteSpan & { |
603 | for (Block &B : Blocks) |
604 | B.Pos += Offset; |
605 | return *this; |
606 | } |
607 | |
608 | auto AlignVectors::ByteSpan::values() const -> SmallVector<Value *, 8> { |
609 | SmallVector<Value *, 8> Values(Blocks.size()); |
610 | for (int i = 0, e = Blocks.size(); i != e; ++i) |
611 | Values[i] = Blocks[i].Seg.Val; |
612 | return Values; |
613 | } |
614 | |
615 | auto AlignVectors::getAlignFromValue(const Value *V) const -> Align { |
616 | const auto *C = dyn_cast<ConstantInt>(Val: V); |
617 | assert(C && "Alignment must be a compile-time constant integer" ); |
618 | return C->getAlignValue(); |
619 | } |
620 | |
621 | auto AlignVectors::getAddrInfo(Instruction &In) const |
622 | -> std::optional<AddrInfo> { |
623 | if (auto *L = isCandidate<LoadInst>(In: &In)) |
624 | return AddrInfo(HVC, L, L->getPointerOperand(), L->getType(), |
625 | L->getAlign()); |
626 | if (auto *S = isCandidate<StoreInst>(In: &In)) |
627 | return AddrInfo(HVC, S, S->getPointerOperand(), |
628 | S->getValueOperand()->getType(), S->getAlign()); |
629 | if (auto *II = isCandidate<IntrinsicInst>(In: &In)) { |
630 | Intrinsic::ID ID = II->getIntrinsicID(); |
631 | switch (ID) { |
632 | case Intrinsic::masked_load: |
633 | return AddrInfo(HVC, II, II->getArgOperand(i: 0), II->getType(), |
634 | getAlignFromValue(V: II->getArgOperand(i: 1))); |
635 | case Intrinsic::masked_store: |
636 | return AddrInfo(HVC, II, II->getArgOperand(i: 1), |
637 | II->getArgOperand(i: 0)->getType(), |
638 | getAlignFromValue(V: II->getArgOperand(i: 2))); |
639 | } |
640 | } |
641 | return std::nullopt; |
642 | } |
643 | |
644 | auto AlignVectors::isHvx(const AddrInfo &AI) const -> bool { |
645 | return HVC.HST.isTypeForHVX(VecTy: AI.ValTy); |
646 | } |
647 | |
648 | auto AlignVectors::getPayload(Value *Val) const -> Value * { |
649 | if (auto *In = dyn_cast<Instruction>(Val)) { |
650 | Intrinsic::ID ID = 0; |
651 | if (auto *II = dyn_cast<IntrinsicInst>(Val: In)) |
652 | ID = II->getIntrinsicID(); |
653 | if (isa<StoreInst>(Val: In) || ID == Intrinsic::masked_store) |
654 | return In->getOperand(i: 0); |
655 | } |
656 | return Val; |
657 | } |
658 | |
659 | auto AlignVectors::getMask(Value *Val) const -> Value * { |
660 | if (auto *II = dyn_cast<IntrinsicInst>(Val)) { |
661 | switch (II->getIntrinsicID()) { |
662 | case Intrinsic::masked_load: |
663 | return II->getArgOperand(i: 2); |
664 | case Intrinsic::masked_store: |
665 | return II->getArgOperand(i: 3); |
666 | } |
667 | } |
668 | |
669 | Type *ValTy = getPayload(Val)->getType(); |
670 | if (auto *VecTy = dyn_cast<VectorType>(Val: ValTy)) |
671 | return HVC.getFullValue(Ty: HVC.getBoolTy(ElemCount: HVC.length(Ty: VecTy))); |
672 | return HVC.getFullValue(Ty: HVC.getBoolTy()); |
673 | } |
674 | |
675 | auto AlignVectors::getPassThrough(Value *Val) const -> Value * { |
676 | if (auto *II = dyn_cast<IntrinsicInst>(Val)) { |
677 | if (II->getIntrinsicID() == Intrinsic::masked_load) |
678 | return II->getArgOperand(i: 3); |
679 | } |
680 | return UndefValue::get(T: getPayload(Val)->getType()); |
681 | } |
682 | |
683 | auto AlignVectors::createAdjustedPointer(IRBuilderBase &Builder, Value *Ptr, |
684 | Type *ValTy, int Adjust, |
685 | const InstMap &CloneMap) const |
686 | -> Value * { |
687 | if (auto *I = dyn_cast<Instruction>(Val: Ptr)) |
688 | if (Instruction *New = CloneMap.lookup(Val: I)) |
689 | Ptr = New; |
690 | return Builder.CreatePtrAdd(Ptr, Offset: HVC.getConstInt(Val: Adjust), Name: "gep" ); |
691 | } |
692 | |
693 | auto AlignVectors::createAlignedPointer(IRBuilderBase &Builder, Value *Ptr, |
694 | Type *ValTy, int Alignment, |
695 | const InstMap &CloneMap) const |
696 | -> Value * { |
697 | auto remap = [&](Value *V) -> Value * { |
698 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
699 | for (auto [Old, New] : CloneMap) |
700 | I->replaceUsesOfWith(From: Old, To: New); |
701 | return I; |
702 | } |
703 | return V; |
704 | }; |
705 | Value *AsInt = Builder.CreatePtrToInt(V: Ptr, DestTy: HVC.getIntTy(), Name: "pti" ); |
706 | Value *Mask = HVC.getConstInt(Val: -Alignment); |
707 | Value *And = Builder.CreateAnd(LHS: remap(AsInt), RHS: Mask, Name: "and" ); |
708 | return Builder.CreateIntToPtr( |
709 | V: And, DestTy: PointerType::getUnqual(C&: ValTy->getContext()), Name: "itp" ); |
710 | } |
711 | |
712 | auto AlignVectors::createLoad(IRBuilderBase &Builder, Type *ValTy, Value *Ptr, |
713 | Value *Predicate, int Alignment, Value *Mask, |
714 | Value *PassThru, |
715 | ArrayRef<Value *> MDSources) const -> Value * { |
716 | bool HvxHasPredLoad = HVC.HST.useHVXV62Ops(); |
717 | // Predicate is nullptr if not creating predicated load |
718 | if (Predicate) { |
719 | assert(!Predicate->getType()->isVectorTy() && |
720 | "Expectning scalar predicate" ); |
721 | if (HVC.isFalse(Val: Predicate)) |
722 | return UndefValue::get(T: ValTy); |
723 | if (!HVC.isTrue(Val: Predicate) && HvxHasPredLoad) { |
724 | Value *Load = createPredicatedLoad(Builder, ValTy, Ptr, Predicate, |
725 | Alignment, MDSources); |
726 | return Builder.CreateSelect(C: Mask, True: Load, False: PassThru); |
727 | } |
728 | // Predicate == true here. |
729 | } |
730 | assert(!HVC.isUndef(Mask)); // Should this be allowed? |
731 | if (HVC.isZero(Val: Mask)) |
732 | return PassThru; |
733 | if (HVC.isTrue(Val: Mask)) |
734 | return createSimpleLoad(Builder, ValTy, Ptr, Alignment, MDSources); |
735 | |
736 | Instruction *Load = Builder.CreateMaskedLoad(Ty: ValTy, Ptr, Alignment: Align(Alignment), |
737 | Mask, PassThru, Name: "mld" ); |
738 | propagateMetadata(I: Load, VL: MDSources); |
739 | return Load; |
740 | } |
741 | |
742 | auto AlignVectors::createSimpleLoad(IRBuilderBase &Builder, Type *ValTy, |
743 | Value *Ptr, int Alignment, |
744 | ArrayRef<Value *> MDSources) const |
745 | -> Value * { |
746 | Instruction *Load = |
747 | Builder.CreateAlignedLoad(Ty: ValTy, Ptr, Align: Align(Alignment), Name: "ald" ); |
748 | propagateMetadata(I: Load, VL: MDSources); |
749 | return Load; |
750 | } |
751 | |
752 | auto AlignVectors::createPredicatedLoad(IRBuilderBase &Builder, Type *ValTy, |
753 | Value *Ptr, Value *Predicate, |
754 | int Alignment, |
755 | ArrayRef<Value *> MDSources) const |
756 | -> Value * { |
757 | assert(HVC.HST.isTypeForHVX(ValTy) && |
758 | "Predicates 'scalar' vector loads not yet supported" ); |
759 | assert(Predicate); |
760 | assert(!Predicate->getType()->isVectorTy() && "Expectning scalar predicate" ); |
761 | assert(HVC.getSizeOf(ValTy, HVC.Alloc) % Alignment == 0); |
762 | if (HVC.isFalse(Val: Predicate)) |
763 | return UndefValue::get(T: ValTy); |
764 | if (HVC.isTrue(Val: Predicate)) |
765 | return createSimpleLoad(Builder, ValTy, Ptr, Alignment, MDSources); |
766 | |
767 | auto V6_vL32b_pred_ai = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vL32b_pred_ai); |
768 | // FIXME: This may not put the offset from Ptr into the vmem offset. |
769 | return HVC.createHvxIntrinsic(Builder, IntID: V6_vL32b_pred_ai, RetTy: ValTy, |
770 | Args: {Predicate, Ptr, HVC.getConstInt(Val: 0)}, |
771 | ArgTys: std::nullopt, MDSources); |
772 | } |
773 | |
774 | auto AlignVectors::createStore(IRBuilderBase &Builder, Value *Val, Value *Ptr, |
775 | Value *Predicate, int Alignment, Value *Mask, |
776 | ArrayRef<Value *> MDSources) const -> Value * { |
777 | if (HVC.isZero(Val: Mask) || HVC.isUndef(Val) || HVC.isUndef(Val: Mask)) |
778 | return UndefValue::get(T: Val->getType()); |
779 | assert(!Predicate || (!Predicate->getType()->isVectorTy() && |
780 | "Expectning scalar predicate" )); |
781 | if (Predicate) { |
782 | if (HVC.isFalse(Val: Predicate)) |
783 | return UndefValue::get(T: Val->getType()); |
784 | if (HVC.isTrue(Val: Predicate)) |
785 | Predicate = nullptr; |
786 | } |
787 | // Here both Predicate and Mask are true or unknown. |
788 | |
789 | if (HVC.isTrue(Val: Mask)) { |
790 | if (Predicate) { // Predicate unknown |
791 | return createPredicatedStore(Builder, Val, Ptr, Predicate, Alignment, |
792 | MDSources); |
793 | } |
794 | // Predicate is true: |
795 | return createSimpleStore(Builder, Val, Ptr, Alignment, MDSources); |
796 | } |
797 | |
798 | // Mask is unknown |
799 | if (!Predicate) { |
800 | Instruction *Store = |
801 | Builder.CreateMaskedStore(Val, Ptr, Alignment: Align(Alignment), Mask); |
802 | propagateMetadata(I: Store, VL: MDSources); |
803 | return Store; |
804 | } |
805 | |
806 | // Both Predicate and Mask are unknown. |
807 | // Emulate masked store with predicated-load + mux + predicated-store. |
808 | Value *PredLoad = createPredicatedLoad(Builder, ValTy: Val->getType(), Ptr, |
809 | Predicate, Alignment, MDSources); |
810 | Value *Mux = Builder.CreateSelect(C: Mask, True: Val, False: PredLoad); |
811 | return createPredicatedStore(Builder, Val: Mux, Ptr, Predicate, Alignment, |
812 | MDSources); |
813 | } |
814 | |
815 | auto AlignVectors::createSimpleStore(IRBuilderBase &Builder, Value *Val, |
816 | Value *Ptr, int Alignment, |
817 | ArrayRef<Value *> MDSources) const |
818 | -> Value * { |
819 | Instruction *Store = Builder.CreateAlignedStore(Val, Ptr, Align: Align(Alignment)); |
820 | propagateMetadata(I: Store, VL: MDSources); |
821 | return Store; |
822 | } |
823 | |
824 | auto AlignVectors::createPredicatedStore(IRBuilderBase &Builder, Value *Val, |
825 | Value *Ptr, Value *Predicate, |
826 | int Alignment, |
827 | ArrayRef<Value *> MDSources) const |
828 | -> Value * { |
829 | assert(HVC.HST.isTypeForHVX(Val->getType()) && |
830 | "Predicates 'scalar' vector stores not yet supported" ); |
831 | assert(Predicate); |
832 | if (HVC.isFalse(Val: Predicate)) |
833 | return UndefValue::get(T: Val->getType()); |
834 | if (HVC.isTrue(Val: Predicate)) |
835 | return createSimpleStore(Builder, Val, Ptr, Alignment, MDSources); |
836 | |
837 | assert(HVC.getSizeOf(Val, HVC.Alloc) % Alignment == 0); |
838 | auto V6_vS32b_pred_ai = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vS32b_pred_ai); |
839 | // FIXME: This may not put the offset from Ptr into the vmem offset. |
840 | return HVC.createHvxIntrinsic(Builder, IntID: V6_vS32b_pred_ai, RetTy: nullptr, |
841 | Args: {Predicate, Ptr, HVC.getConstInt(Val: 0), Val}, |
842 | ArgTys: std::nullopt, MDSources); |
843 | } |
844 | |
845 | auto AlignVectors::getUpwardDeps(Instruction *In, Instruction *Base) const |
846 | -> DepList { |
847 | BasicBlock *Parent = Base->getParent(); |
848 | assert(In->getParent() == Parent && |
849 | "Base and In should be in the same block" ); |
850 | assert(Base->comesBefore(In) && "Base should come before In" ); |
851 | |
852 | DepList Deps; |
853 | std::deque<Instruction *> WorkQ = {In}; |
854 | while (!WorkQ.empty()) { |
855 | Instruction *D = WorkQ.front(); |
856 | WorkQ.pop_front(); |
857 | if (D != In) |
858 | Deps.insert(x: D); |
859 | for (Value *Op : D->operands()) { |
860 | if (auto *I = dyn_cast<Instruction>(Val: Op)) { |
861 | if (I->getParent() == Parent && Base->comesBefore(Other: I)) |
862 | WorkQ.push_back(x: I); |
863 | } |
864 | } |
865 | } |
866 | return Deps; |
867 | } |
868 | |
869 | auto AlignVectors::createAddressGroups() -> bool { |
870 | // An address group created here may contain instructions spanning |
871 | // multiple basic blocks. |
872 | AddrList WorkStack; |
873 | |
874 | auto findBaseAndOffset = [&](AddrInfo &AI) -> std::pair<Instruction *, int> { |
875 | for (AddrInfo &W : WorkStack) { |
876 | if (auto D = HVC.calculatePointerDifference(Ptr0: AI.Addr, Ptr1: W.Addr)) |
877 | return std::make_pair(x&: W.Inst, y&: *D); |
878 | } |
879 | return std::make_pair(x: nullptr, y: 0); |
880 | }; |
881 | |
882 | auto traverseBlock = [&](DomTreeNode *DomN, auto Visit) -> void { |
883 | BasicBlock &Block = *DomN->getBlock(); |
884 | for (Instruction &I : Block) { |
885 | auto AI = this->getAddrInfo(In&: I); // Use this-> for gcc6. |
886 | if (!AI) |
887 | continue; |
888 | auto F = findBaseAndOffset(*AI); |
889 | Instruction *GroupInst; |
890 | if (Instruction *BI = F.first) { |
891 | AI->Offset = F.second; |
892 | GroupInst = BI; |
893 | } else { |
894 | WorkStack.push_back(x: *AI); |
895 | GroupInst = AI->Inst; |
896 | } |
897 | AddrGroups[GroupInst].push_back(x: *AI); |
898 | } |
899 | |
900 | for (DomTreeNode *C : DomN->children()) |
901 | Visit(C, Visit); |
902 | |
903 | while (!WorkStack.empty() && WorkStack.back().Inst->getParent() == &Block) |
904 | WorkStack.pop_back(); |
905 | }; |
906 | |
907 | traverseBlock(HVC.DT.getRootNode(), traverseBlock); |
908 | assert(WorkStack.empty()); |
909 | |
910 | // AddrGroups are formed. |
911 | |
912 | // Remove groups of size 1. |
913 | erase_if(map&: AddrGroups, p: [](auto &G) { return G.second.size() == 1; }); |
914 | // Remove groups that don't use HVX types. |
915 | erase_if(map&: AddrGroups, p: [&](auto &G) { |
916 | return llvm::none_of( |
917 | G.second, [&](auto &I) { return HVC.HST.isTypeForHVX(VecTy: I.ValTy); }); |
918 | }); |
919 | |
920 | return !AddrGroups.empty(); |
921 | } |
922 | |
923 | auto AlignVectors::createLoadGroups(const AddrList &Group) const -> MoveList { |
924 | // Form load groups. |
925 | // To avoid complications with moving code across basic blocks, only form |
926 | // groups that are contained within a single basic block. |
927 | unsigned SizeLimit = VAGroupSizeLimit; |
928 | if (SizeLimit == 0) |
929 | return {}; |
930 | |
931 | auto tryAddTo = [&](const AddrInfo &Info, MoveGroup &Move) { |
932 | assert(!Move.Main.empty() && "Move group should have non-empty Main" ); |
933 | if (Move.Main.size() >= SizeLimit) |
934 | return false; |
935 | // Don't mix HVX and non-HVX instructions. |
936 | if (Move.IsHvx != isHvx(AI: Info)) |
937 | return false; |
938 | // Leading instruction in the load group. |
939 | Instruction *Base = Move.Main.front(); |
940 | if (Base->getParent() != Info.Inst->getParent()) |
941 | return false; |
942 | // Check if it's safe to move the load. |
943 | if (!HVC.isSafeToMoveBeforeInBB(In: *Info.Inst, To: Base->getIterator())) |
944 | return false; |
945 | // And if it's safe to clone the dependencies. |
946 | auto isSafeToCopyAtBase = [&](const Instruction *I) { |
947 | return HVC.isSafeToMoveBeforeInBB(In: *I, To: Base->getIterator()) && |
948 | HVC.isSafeToClone(In: *I); |
949 | }; |
950 | DepList Deps = getUpwardDeps(In: Info.Inst, Base); |
951 | if (!llvm::all_of(Range&: Deps, P: isSafeToCopyAtBase)) |
952 | return false; |
953 | |
954 | Move.Main.push_back(x: Info.Inst); |
955 | llvm::append_range(C&: Move.Deps, R&: Deps); |
956 | return true; |
957 | }; |
958 | |
959 | MoveList LoadGroups; |
960 | |
961 | for (const AddrInfo &Info : Group) { |
962 | if (!Info.Inst->mayReadFromMemory()) |
963 | continue; |
964 | if (LoadGroups.empty() || !tryAddTo(Info, LoadGroups.back())) |
965 | LoadGroups.emplace_back(args: Info, args: Group.front().Inst, args: isHvx(AI: Info), args: true); |
966 | } |
967 | |
968 | // Erase singleton groups. |
969 | erase_if(container&: LoadGroups, p: [](const MoveGroup &G) { return G.Main.size() <= 1; }); |
970 | |
971 | // Erase HVX groups on targets < HvxV62 (due to lack of predicated loads). |
972 | if (!HVC.HST.useHVXV62Ops()) |
973 | erase_if(container&: LoadGroups, p: [](const MoveGroup &G) { return G.IsHvx; }); |
974 | |
975 | return LoadGroups; |
976 | } |
977 | |
978 | auto AlignVectors::createStoreGroups(const AddrList &Group) const -> MoveList { |
979 | // Form store groups. |
980 | // To avoid complications with moving code across basic blocks, only form |
981 | // groups that are contained within a single basic block. |
982 | unsigned SizeLimit = VAGroupSizeLimit; |
983 | if (SizeLimit == 0) |
984 | return {}; |
985 | |
986 | auto tryAddTo = [&](const AddrInfo &Info, MoveGroup &Move) { |
987 | assert(!Move.Main.empty() && "Move group should have non-empty Main" ); |
988 | if (Move.Main.size() >= SizeLimit) |
989 | return false; |
990 | // For stores with return values we'd have to collect downward depenencies. |
991 | // There are no such stores that we handle at the moment, so omit that. |
992 | assert(Info.Inst->getType()->isVoidTy() && |
993 | "Not handling stores with return values" ); |
994 | // Don't mix HVX and non-HVX instructions. |
995 | if (Move.IsHvx != isHvx(AI: Info)) |
996 | return false; |
997 | // For stores we need to be careful whether it's safe to move them. |
998 | // Stores that are otherwise safe to move together may not appear safe |
999 | // to move over one another (i.e. isSafeToMoveBefore may return false). |
1000 | Instruction *Base = Move.Main.front(); |
1001 | if (Base->getParent() != Info.Inst->getParent()) |
1002 | return false; |
1003 | if (!HVC.isSafeToMoveBeforeInBB(In: *Info.Inst, To: Base->getIterator(), IgnoreInsts: Move.Main)) |
1004 | return false; |
1005 | Move.Main.push_back(x: Info.Inst); |
1006 | return true; |
1007 | }; |
1008 | |
1009 | MoveList StoreGroups; |
1010 | |
1011 | for (auto I = Group.rbegin(), E = Group.rend(); I != E; ++I) { |
1012 | const AddrInfo &Info = *I; |
1013 | if (!Info.Inst->mayWriteToMemory()) |
1014 | continue; |
1015 | if (StoreGroups.empty() || !tryAddTo(Info, StoreGroups.back())) |
1016 | StoreGroups.emplace_back(args: Info, args: Group.front().Inst, args: isHvx(AI: Info), args: false); |
1017 | } |
1018 | |
1019 | // Erase singleton groups. |
1020 | erase_if(container&: StoreGroups, p: [](const MoveGroup &G) { return G.Main.size() <= 1; }); |
1021 | |
1022 | // Erase HVX groups on targets < HvxV62 (due to lack of predicated loads). |
1023 | if (!HVC.HST.useHVXV62Ops()) |
1024 | erase_if(container&: StoreGroups, p: [](const MoveGroup &G) { return G.IsHvx; }); |
1025 | |
1026 | // Erase groups where every store is a full HVX vector. The reason is that |
1027 | // aligning predicated stores generates complex code that may be less |
1028 | // efficient than a sequence of unaligned vector stores. |
1029 | if (!VADoFullStores) { |
1030 | erase_if(container&: StoreGroups, p: [this](const MoveGroup &G) { |
1031 | return G.IsHvx && llvm::all_of(Range: G.Main, P: [this](Instruction *S) { |
1032 | auto MaybeInfo = this->getAddrInfo(In&: *S); |
1033 | assert(MaybeInfo.has_value()); |
1034 | return HVC.HST.isHVXVectorType( |
1035 | VecTy: EVT::getEVT(Ty: MaybeInfo->ValTy, HandleUnknown: false)); |
1036 | }); |
1037 | }); |
1038 | } |
1039 | |
1040 | return StoreGroups; |
1041 | } |
1042 | |
1043 | auto AlignVectors::moveTogether(MoveGroup &Move) const -> bool { |
1044 | // Move all instructions to be adjacent. |
1045 | assert(!Move.Main.empty() && "Move group should have non-empty Main" ); |
1046 | Instruction *Where = Move.Main.front(); |
1047 | |
1048 | if (Move.IsLoad) { |
1049 | // Move all the loads (and dependencies) to where the first load is. |
1050 | // Clone all deps to before Where, keeping order. |
1051 | Move.Clones = cloneBefore(To: Where, Insts&: Move.Deps); |
1052 | // Move all main instructions to after Where, keeping order. |
1053 | ArrayRef<Instruction *> Main(Move.Main); |
1054 | for (Instruction *M : Main) { |
1055 | if (M != Where) |
1056 | M->moveAfter(MovePos: Where); |
1057 | for (auto [Old, New] : Move.Clones) |
1058 | M->replaceUsesOfWith(From: Old, To: New); |
1059 | Where = M; |
1060 | } |
1061 | // Replace Deps with the clones. |
1062 | for (int i = 0, e = Move.Deps.size(); i != e; ++i) |
1063 | Move.Deps[i] = Move.Clones[Move.Deps[i]]; |
1064 | } else { |
1065 | // Move all the stores to where the last store is. |
1066 | // NOTE: Deps are empty for "store" groups. If they need to be |
1067 | // non-empty, decide on the order. |
1068 | assert(Move.Deps.empty()); |
1069 | // Move all main instructions to before Where, inverting order. |
1070 | ArrayRef<Instruction *> Main(Move.Main); |
1071 | for (Instruction *M : Main.drop_front(N: 1)) { |
1072 | M->moveBefore(MovePos: Where); |
1073 | Where = M; |
1074 | } |
1075 | } |
1076 | |
1077 | return Move.Main.size() + Move.Deps.size() > 1; |
1078 | } |
1079 | |
1080 | template <typename T> |
1081 | auto AlignVectors::cloneBefore(Instruction *To, T &&Insts) const -> InstMap { |
1082 | InstMap Map; |
1083 | |
1084 | for (Instruction *I : Insts) { |
1085 | assert(HVC.isSafeToClone(*I)); |
1086 | Instruction *C = I->clone(); |
1087 | C->setName(Twine("c." ) + I->getName() + "." ); |
1088 | C->insertBefore(InsertPos: To); |
1089 | |
1090 | for (auto [Old, New] : Map) |
1091 | C->replaceUsesOfWith(From: Old, To: New); |
1092 | Map.insert(KV: std::make_pair(x&: I, y&: C)); |
1093 | } |
1094 | return Map; |
1095 | } |
1096 | |
1097 | auto AlignVectors::realignLoadGroup(IRBuilderBase &Builder, |
1098 | const ByteSpan &VSpan, int ScLen, |
1099 | Value *AlignVal, Value *AlignAddr) const |
1100 | -> void { |
1101 | LLVM_DEBUG(dbgs() << __func__ << "\n" ); |
1102 | |
1103 | Type *SecTy = HVC.getByteTy(ElemCount: ScLen); |
1104 | int NumSectors = (VSpan.extent() + ScLen - 1) / ScLen; |
1105 | bool DoAlign = !HVC.isZero(Val: AlignVal); |
1106 | BasicBlock::iterator BasePos = Builder.GetInsertPoint(); |
1107 | BasicBlock *BaseBlock = Builder.GetInsertBlock(); |
1108 | |
1109 | ByteSpan ASpan; |
1110 | auto *True = HVC.getFullValue(Ty: HVC.getBoolTy(ElemCount: ScLen)); |
1111 | auto *Undef = UndefValue::get(T: SecTy); |
1112 | |
1113 | // Created load does not have to be "Instruction" (e.g. "undef"). |
1114 | SmallVector<Value *> Loads(NumSectors + DoAlign, nullptr); |
1115 | |
1116 | // We could create all of the aligned loads, and generate the valigns |
1117 | // at the location of the first load, but for large load groups, this |
1118 | // could create highly suboptimal code (there have been groups of 140+ |
1119 | // loads in real code). |
1120 | // Instead, place the loads/valigns as close to the users as possible. |
1121 | // In any case we need to have a mapping from the blocks of VSpan (the |
1122 | // span covered by the pre-existing loads) to ASpan (the span covered |
1123 | // by the aligned loads). There is a small problem, though: ASpan needs |
1124 | // to have pointers to the loads/valigns, but we don't have these loads |
1125 | // because we don't know where to put them yet. We find out by creating |
1126 | // a section of ASpan that corresponds to values (blocks) from VSpan, |
1127 | // and checking where the new load should be placed. We need to attach |
1128 | // this location information to each block in ASpan somehow, so we put |
1129 | // distincts values for Seg.Val in each ASpan.Blocks[i], and use a map |
1130 | // to store the location for each Seg.Val. |
1131 | // The distinct values happen to be Blocks[i].Seg.Val = &Blocks[i], |
1132 | // which helps with printing ByteSpans without crashing when printing |
1133 | // Segments with these temporary identifiers in place of Val. |
1134 | |
1135 | // Populate the blocks first, to avoid reallocations of the vector |
1136 | // interfering with generating the placeholder addresses. |
1137 | for (int Index = 0; Index != NumSectors; ++Index) |
1138 | ASpan.Blocks.emplace_back(args: nullptr, args&: ScLen, args: Index * ScLen); |
1139 | for (int Index = 0; Index != NumSectors; ++Index) { |
1140 | ASpan.Blocks[Index].Seg.Val = |
1141 | reinterpret_cast<Value *>(&ASpan.Blocks[Index]); |
1142 | } |
1143 | |
1144 | // Multiple values from VSpan can map to the same value in ASpan. Since we |
1145 | // try to create loads lazily, we need to find the earliest use for each |
1146 | // value from ASpan. |
1147 | DenseMap<void *, Instruction *> EarliestUser; |
1148 | auto isEarlier = [](Instruction *A, Instruction *B) { |
1149 | if (B == nullptr) |
1150 | return true; |
1151 | if (A == nullptr) |
1152 | return false; |
1153 | assert(A->getParent() == B->getParent()); |
1154 | return A->comesBefore(Other: B); |
1155 | }; |
1156 | auto earliestUser = [&](const auto &Uses) { |
1157 | Instruction *User = nullptr; |
1158 | for (const Use &U : Uses) { |
1159 | auto *I = dyn_cast<Instruction>(Val: U.getUser()); |
1160 | assert(I != nullptr && "Load used in a non-instruction?" ); |
1161 | // Make sure we only consider users in this block, but we need |
1162 | // to remember if there were users outside the block too. This is |
1163 | // because if no users are found, aligned loads will not be created. |
1164 | if (I->getParent() == BaseBlock) { |
1165 | if (!isa<PHINode>(Val: I)) |
1166 | User = std::min(a: User, b: I, comp: isEarlier); |
1167 | } else { |
1168 | User = std::min(a: User, b: BaseBlock->getTerminator(), comp: isEarlier); |
1169 | } |
1170 | } |
1171 | return User; |
1172 | }; |
1173 | |
1174 | for (const ByteSpan::Block &B : VSpan) { |
1175 | ByteSpan ASection = ASpan.section(Start: B.Pos, Length: B.Seg.Size); |
1176 | for (const ByteSpan::Block &S : ASection) { |
1177 | EarliestUser[S.Seg.Val] = std::min( |
1178 | a: EarliestUser[S.Seg.Val], b: earliestUser(B.Seg.Val->uses()), comp: isEarlier); |
1179 | } |
1180 | } |
1181 | |
1182 | LLVM_DEBUG({ |
1183 | dbgs() << "ASpan:\n" << ASpan << '\n'; |
1184 | dbgs() << "Earliest users of ASpan:\n" ; |
1185 | for (auto &[Val, User] : EarliestUser) { |
1186 | dbgs() << Val << "\n ->" << *User << '\n'; |
1187 | } |
1188 | }); |
1189 | |
1190 | auto createLoad = [&](IRBuilderBase &Builder, const ByteSpan &VSpan, |
1191 | int Index, bool MakePred) { |
1192 | Value *Ptr = |
1193 | createAdjustedPointer(Builder, Ptr: AlignAddr, ValTy: SecTy, Adjust: Index * ScLen); |
1194 | Value *Predicate = |
1195 | MakePred ? makeTestIfUnaligned(Builder, AlignVal, Alignment: ScLen) : nullptr; |
1196 | |
1197 | // If vector shifting is potentially needed, accumulate metadata |
1198 | // from source sections of twice the load width. |
1199 | int Start = (Index - DoAlign) * ScLen; |
1200 | int Width = (1 + DoAlign) * ScLen; |
1201 | return this->createLoad(Builder, ValTy: SecTy, Ptr, Predicate, Alignment: ScLen, Mask: True, PassThru: Undef, |
1202 | MDSources: VSpan.section(Start, Length: Width).values()); |
1203 | }; |
1204 | |
1205 | auto moveBefore = [this](Instruction *In, Instruction *To) { |
1206 | // Move In and its upward dependencies to before To. |
1207 | assert(In->getParent() == To->getParent()); |
1208 | DepList Deps = getUpwardDeps(In, Base: To); |
1209 | In->moveBefore(MovePos: To); |
1210 | // DepList is sorted with respect to positions in the basic block. |
1211 | InstMap Map = cloneBefore(To: In, Insts&: Deps); |
1212 | for (auto [Old, New] : Map) |
1213 | In->replaceUsesOfWith(From: Old, To: New); |
1214 | }; |
1215 | |
1216 | // Generate necessary loads at appropriate locations. |
1217 | LLVM_DEBUG(dbgs() << "Creating loads for ASpan sectors\n" ); |
1218 | for (int Index = 0; Index != NumSectors + 1; ++Index) { |
1219 | // In ASpan, each block will be either a single aligned load, or a |
1220 | // valign of a pair of loads. In the latter case, an aligned load j |
1221 | // will belong to the current valign, and the one in the previous |
1222 | // block (for j > 0). |
1223 | // Place the load at a location which will dominate the valign, assuming |
1224 | // the valign will be placed right before the earliest user. |
1225 | Instruction *PrevAt = |
1226 | DoAlign && Index > 0 ? EarliestUser[&ASpan[Index - 1]] : nullptr; |
1227 | Instruction *ThisAt = |
1228 | Index < NumSectors ? EarliestUser[&ASpan[Index]] : nullptr; |
1229 | if (auto *Where = std::min(a: PrevAt, b: ThisAt, comp: isEarlier)) { |
1230 | Builder.SetInsertPoint(Where); |
1231 | Loads[Index] = |
1232 | createLoad(Builder, VSpan, Index, DoAlign && Index == NumSectors); |
1233 | // We know it's safe to put the load at BasePos, but we'd prefer to put |
1234 | // it at "Where". To see if the load is safe to be placed at Where, put |
1235 | // it there first and then check if it's safe to move it to BasePos. |
1236 | // If not, then the load needs to be placed at BasePos. |
1237 | // We can't do this check proactively because we need the load to exist |
1238 | // in order to check legality. |
1239 | if (auto *Load = dyn_cast<Instruction>(Val: Loads[Index])) { |
1240 | if (!HVC.isSafeToMoveBeforeInBB(In: *Load, To: BasePos)) |
1241 | moveBefore(Load, &*BasePos); |
1242 | } |
1243 | LLVM_DEBUG(dbgs() << "Loads[" << Index << "]:" << *Loads[Index] << '\n'); |
1244 | } |
1245 | } |
1246 | |
1247 | // Generate valigns if needed, and fill in proper values in ASpan |
1248 | LLVM_DEBUG(dbgs() << "Creating values for ASpan sectors\n" ); |
1249 | for (int Index = 0; Index != NumSectors; ++Index) { |
1250 | ASpan[Index].Seg.Val = nullptr; |
1251 | if (auto *Where = EarliestUser[&ASpan[Index]]) { |
1252 | Builder.SetInsertPoint(Where); |
1253 | Value *Val = Loads[Index]; |
1254 | assert(Val != nullptr); |
1255 | if (DoAlign) { |
1256 | Value *NextLoad = Loads[Index + 1]; |
1257 | assert(NextLoad != nullptr); |
1258 | Val = HVC.vralignb(Builder, Lo: Val, Hi: NextLoad, Amt: AlignVal); |
1259 | } |
1260 | ASpan[Index].Seg.Val = Val; |
1261 | LLVM_DEBUG(dbgs() << "ASpan[" << Index << "]:" << *Val << '\n'); |
1262 | } |
1263 | } |
1264 | |
1265 | for (const ByteSpan::Block &B : VSpan) { |
1266 | ByteSpan ASection = ASpan.section(Start: B.Pos, Length: B.Seg.Size).shift(Offset: -B.Pos); |
1267 | Value *Accum = UndefValue::get(T: HVC.getByteTy(ElemCount: B.Seg.Size)); |
1268 | Builder.SetInsertPoint(cast<Instruction>(Val: B.Seg.Val)); |
1269 | |
1270 | // We're generating a reduction, where each instruction depends on |
1271 | // the previous one, so we need to order them according to the position |
1272 | // of their inputs in the code. |
1273 | std::vector<ByteSpan::Block *> ABlocks; |
1274 | for (ByteSpan::Block &S : ASection) { |
1275 | if (S.Seg.Val != nullptr) |
1276 | ABlocks.push_back(x: &S); |
1277 | } |
1278 | llvm::sort(C&: ABlocks, |
1279 | Comp: [&](const ByteSpan::Block *A, const ByteSpan::Block *B) { |
1280 | return isEarlier(cast<Instruction>(Val: A->Seg.Val), |
1281 | cast<Instruction>(Val: B->Seg.Val)); |
1282 | }); |
1283 | for (ByteSpan::Block *S : ABlocks) { |
1284 | // The processing of the data loaded by the aligned loads |
1285 | // needs to be inserted after the data is available. |
1286 | Instruction *SegI = cast<Instruction>(Val: S->Seg.Val); |
1287 | Builder.SetInsertPoint(&*std::next(x: SegI->getIterator())); |
1288 | Value *Pay = HVC.vbytes(Builder, Val: getPayload(Val: S->Seg.Val)); |
1289 | Accum = |
1290 | HVC.insertb(Builder, Dest: Accum, Src: Pay, Start: S->Seg.Start, Length: S->Seg.Size, Where: S->Pos); |
1291 | } |
1292 | // Instead of casting everything to bytes for the vselect, cast to the |
1293 | // original value type. This will avoid complications with casting masks. |
1294 | // For example, in cases when the original mask applied to i32, it could |
1295 | // be converted to a mask applicable to i8 via pred_typecast intrinsic, |
1296 | // but if the mask is not exactly of HVX length, extra handling would be |
1297 | // needed to make it work. |
1298 | Type *ValTy = getPayload(Val: B.Seg.Val)->getType(); |
1299 | Value *Cast = Builder.CreateBitCast(V: Accum, DestTy: ValTy, Name: "cst" ); |
1300 | Value *Sel = Builder.CreateSelect(C: getMask(Val: B.Seg.Val), True: Cast, |
1301 | False: getPassThrough(Val: B.Seg.Val), Name: "sel" ); |
1302 | B.Seg.Val->replaceAllUsesWith(V: Sel); |
1303 | } |
1304 | } |
1305 | |
1306 | auto AlignVectors::realignStoreGroup(IRBuilderBase &Builder, |
1307 | const ByteSpan &VSpan, int ScLen, |
1308 | Value *AlignVal, Value *AlignAddr) const |
1309 | -> void { |
1310 | LLVM_DEBUG(dbgs() << __func__ << "\n" ); |
1311 | |
1312 | Type *SecTy = HVC.getByteTy(ElemCount: ScLen); |
1313 | int NumSectors = (VSpan.extent() + ScLen - 1) / ScLen; |
1314 | bool DoAlign = !HVC.isZero(Val: AlignVal); |
1315 | |
1316 | // Stores. |
1317 | ByteSpan ASpanV, ASpanM; |
1318 | |
1319 | // Return a vector value corresponding to the input value Val: |
1320 | // either <1 x Val> for scalar Val, or Val itself for vector Val. |
1321 | auto MakeVec = [](IRBuilderBase &Builder, Value *Val) -> Value * { |
1322 | Type *Ty = Val->getType(); |
1323 | if (Ty->isVectorTy()) |
1324 | return Val; |
1325 | auto *VecTy = VectorType::get(ElementType: Ty, NumElements: 1, /*Scalable=*/false); |
1326 | return Builder.CreateBitCast(V: Val, DestTy: VecTy, Name: "cst" ); |
1327 | }; |
1328 | |
1329 | // Create an extra "undef" sector at the beginning and at the end. |
1330 | // They will be used as the left/right filler in the vlalign step. |
1331 | for (int Index = (DoAlign ? -1 : 0); Index != NumSectors + DoAlign; ++Index) { |
1332 | // For stores, the size of each section is an aligned vector length. |
1333 | // Adjust the store offsets relative to the section start offset. |
1334 | ByteSpan VSection = |
1335 | VSpan.section(Start: Index * ScLen, Length: ScLen).shift(Offset: -Index * ScLen); |
1336 | Value *Undef = UndefValue::get(T: SecTy); |
1337 | Value *Zero = HVC.getNullValue(Ty: SecTy); |
1338 | Value *AccumV = Undef; |
1339 | Value *AccumM = Zero; |
1340 | for (ByteSpan::Block &S : VSection) { |
1341 | Value *Pay = getPayload(Val: S.Seg.Val); |
1342 | Value *Mask = HVC.rescale(Builder, Mask: MakeVec(Builder, getMask(Val: S.Seg.Val)), |
1343 | FromTy: Pay->getType(), ToTy: HVC.getByteTy()); |
1344 | Value *PartM = HVC.insertb(Builder, Dest: Zero, Src: HVC.vbytes(Builder, Val: Mask), |
1345 | Start: S.Seg.Start, Length: S.Seg.Size, Where: S.Pos); |
1346 | AccumM = Builder.CreateOr(LHS: AccumM, RHS: PartM); |
1347 | |
1348 | Value *PartV = HVC.insertb(Builder, Dest: Undef, Src: HVC.vbytes(Builder, Val: Pay), |
1349 | Start: S.Seg.Start, Length: S.Seg.Size, Where: S.Pos); |
1350 | |
1351 | AccumV = Builder.CreateSelect( |
1352 | C: Builder.CreateICmp(P: CmpInst::ICMP_NE, LHS: PartM, RHS: Zero), True: PartV, False: AccumV); |
1353 | } |
1354 | ASpanV.Blocks.emplace_back(args&: AccumV, args&: ScLen, args: Index * ScLen); |
1355 | ASpanM.Blocks.emplace_back(args&: AccumM, args&: ScLen, args: Index * ScLen); |
1356 | } |
1357 | |
1358 | LLVM_DEBUG({ |
1359 | dbgs() << "ASpanV before vlalign:\n" << ASpanV << '\n'; |
1360 | dbgs() << "ASpanM before vlalign:\n" << ASpanM << '\n'; |
1361 | }); |
1362 | |
1363 | // vlalign |
1364 | if (DoAlign) { |
1365 | for (int Index = 1; Index != NumSectors + 2; ++Index) { |
1366 | Value *PrevV = ASpanV[Index - 1].Seg.Val, *ThisV = ASpanV[Index].Seg.Val; |
1367 | Value *PrevM = ASpanM[Index - 1].Seg.Val, *ThisM = ASpanM[Index].Seg.Val; |
1368 | assert(isSectorTy(PrevV->getType()) && isSectorTy(PrevM->getType())); |
1369 | ASpanV[Index - 1].Seg.Val = HVC.vlalignb(Builder, Lo: PrevV, Hi: ThisV, Amt: AlignVal); |
1370 | ASpanM[Index - 1].Seg.Val = HVC.vlalignb(Builder, Lo: PrevM, Hi: ThisM, Amt: AlignVal); |
1371 | } |
1372 | } |
1373 | |
1374 | LLVM_DEBUG({ |
1375 | dbgs() << "ASpanV after vlalign:\n" << ASpanV << '\n'; |
1376 | dbgs() << "ASpanM after vlalign:\n" << ASpanM << '\n'; |
1377 | }); |
1378 | |
1379 | auto createStore = [&](IRBuilderBase &Builder, const ByteSpan &ASpanV, |
1380 | const ByteSpan &ASpanM, int Index, bool MakePred) { |
1381 | Value *Val = ASpanV[Index].Seg.Val; |
1382 | Value *Mask = ASpanM[Index].Seg.Val; // bytes |
1383 | if (HVC.isUndef(Val) || HVC.isZero(Val: Mask)) |
1384 | return; |
1385 | Value *Ptr = |
1386 | createAdjustedPointer(Builder, Ptr: AlignAddr, ValTy: SecTy, Adjust: Index * ScLen); |
1387 | Value *Predicate = |
1388 | MakePred ? makeTestIfUnaligned(Builder, AlignVal, Alignment: ScLen) : nullptr; |
1389 | |
1390 | // If vector shifting is potentially needed, accumulate metadata |
1391 | // from source sections of twice the store width. |
1392 | int Start = (Index - DoAlign) * ScLen; |
1393 | int Width = (1 + DoAlign) * ScLen; |
1394 | this->createStore(Builder, Val, Ptr, Predicate, Alignment: ScLen, |
1395 | Mask: HVC.vlsb(Builder, Val: Mask), |
1396 | MDSources: VSpan.section(Start, Length: Width).values()); |
1397 | }; |
1398 | |
1399 | for (int Index = 0; Index != NumSectors + DoAlign; ++Index) { |
1400 | createStore(Builder, ASpanV, ASpanM, Index, DoAlign && Index == NumSectors); |
1401 | } |
1402 | } |
1403 | |
1404 | auto AlignVectors::realignGroup(const MoveGroup &Move) const -> bool { |
1405 | LLVM_DEBUG(dbgs() << "Realigning group:\n" << Move << '\n'); |
1406 | |
1407 | // TODO: Needs support for masked loads/stores of "scalar" vectors. |
1408 | if (!Move.IsHvx) |
1409 | return false; |
1410 | |
1411 | // Return the element with the maximum alignment from Range, |
1412 | // where GetValue obtains the value to compare from an element. |
1413 | auto getMaxOf = [](auto Range, auto GetValue) { |
1414 | return *llvm::max_element(Range, [&GetValue](auto &A, auto &B) { |
1415 | return GetValue(A) < GetValue(B); |
1416 | }); |
1417 | }; |
1418 | |
1419 | const AddrList &BaseInfos = AddrGroups.at(k: Move.Base); |
1420 | |
1421 | // Conceptually, there is a vector of N bytes covering the addresses |
1422 | // starting from the minimum offset (i.e. Base.Addr+Start). This vector |
1423 | // represents a contiguous memory region that spans all accessed memory |
1424 | // locations. |
1425 | // The correspondence between loaded or stored values will be expressed |
1426 | // in terms of this vector. For example, the 0th element of the vector |
1427 | // from the Base address info will start at byte Start from the beginning |
1428 | // of this conceptual vector. |
1429 | // |
1430 | // This vector will be loaded/stored starting at the nearest down-aligned |
1431 | // address and the amount od the down-alignment will be AlignVal: |
1432 | // valign(load_vector(align_down(Base+Start)), AlignVal) |
1433 | |
1434 | std::set<Instruction *> TestSet(Move.Main.begin(), Move.Main.end()); |
1435 | AddrList MoveInfos; |
1436 | llvm::copy_if( |
1437 | Range: BaseInfos, Out: std::back_inserter(x&: MoveInfos), |
1438 | P: [&TestSet](const AddrInfo &AI) { return TestSet.count(x: AI.Inst); }); |
1439 | |
1440 | // Maximum alignment present in the whole address group. |
1441 | const AddrInfo &WithMaxAlign = |
1442 | getMaxOf(MoveInfos, [](const AddrInfo &AI) { return AI.HaveAlign; }); |
1443 | Align MaxGiven = WithMaxAlign.HaveAlign; |
1444 | |
1445 | // Minimum alignment present in the move address group. |
1446 | const AddrInfo &WithMinOffset = |
1447 | getMaxOf(MoveInfos, [](const AddrInfo &AI) { return -AI.Offset; }); |
1448 | |
1449 | const AddrInfo &WithMaxNeeded = |
1450 | getMaxOf(MoveInfos, [](const AddrInfo &AI) { return AI.NeedAlign; }); |
1451 | Align MinNeeded = WithMaxNeeded.NeedAlign; |
1452 | |
1453 | // Set the builder's insertion point right before the load group, or |
1454 | // immediately after the store group. (Instructions in a store group are |
1455 | // listed in reverse order.) |
1456 | Instruction *InsertAt = Move.Main.front(); |
1457 | if (!Move.IsLoad) { |
1458 | // There should be a terminator (which store isn't, but check anyways). |
1459 | assert(InsertAt->getIterator() != InsertAt->getParent()->end()); |
1460 | InsertAt = &*std::next(x: InsertAt->getIterator()); |
1461 | } |
1462 | |
1463 | IRBuilder Builder(InsertAt->getParent(), InsertAt->getIterator(), |
1464 | InstSimplifyFolder(HVC.DL)); |
1465 | Value *AlignAddr = nullptr; // Actual aligned address. |
1466 | Value *AlignVal = nullptr; // Right-shift amount (for valign). |
1467 | |
1468 | if (MinNeeded <= MaxGiven) { |
1469 | int Start = WithMinOffset.Offset; |
1470 | int OffAtMax = WithMaxAlign.Offset; |
1471 | // Shift the offset of the maximally aligned instruction (OffAtMax) |
1472 | // back by just enough multiples of the required alignment to cover the |
1473 | // distance from Start to OffAtMax. |
1474 | // Calculate the address adjustment amount based on the address with the |
1475 | // maximum alignment. This is to allow a simple gep instruction instead |
1476 | // of potential bitcasts to i8*. |
1477 | int Adjust = -alignTo(Value: OffAtMax - Start, Align: MinNeeded.value()); |
1478 | AlignAddr = createAdjustedPointer(Builder, Ptr: WithMaxAlign.Addr, |
1479 | ValTy: WithMaxAlign.ValTy, Adjust, CloneMap: Move.Clones); |
1480 | int Diff = Start - (OffAtMax + Adjust); |
1481 | AlignVal = HVC.getConstInt(Val: Diff); |
1482 | assert(Diff >= 0); |
1483 | assert(static_cast<decltype(MinNeeded.value())>(Diff) < MinNeeded.value()); |
1484 | } else { |
1485 | // WithMinOffset is the lowest address in the group, |
1486 | // WithMinOffset.Addr = Base+Start. |
1487 | // Align instructions for both HVX (V6_valign) and scalar (S2_valignrb) |
1488 | // mask off unnecessary bits, so it's ok to just the original pointer as |
1489 | // the alignment amount. |
1490 | // Do an explicit down-alignment of the address to avoid creating an |
1491 | // aligned instruction with an address that is not really aligned. |
1492 | AlignAddr = |
1493 | createAlignedPointer(Builder, Ptr: WithMinOffset.Addr, ValTy: WithMinOffset.ValTy, |
1494 | Alignment: MinNeeded.value(), CloneMap: Move.Clones); |
1495 | AlignVal = |
1496 | Builder.CreatePtrToInt(V: WithMinOffset.Addr, DestTy: HVC.getIntTy(), Name: "pti" ); |
1497 | if (auto *I = dyn_cast<Instruction>(Val: AlignVal)) { |
1498 | for (auto [Old, New] : Move.Clones) |
1499 | I->replaceUsesOfWith(From: Old, To: New); |
1500 | } |
1501 | } |
1502 | |
1503 | ByteSpan VSpan; |
1504 | for (const AddrInfo &AI : MoveInfos) { |
1505 | VSpan.Blocks.emplace_back(args: AI.Inst, args: HVC.getSizeOf(Ty: AI.ValTy), |
1506 | args: AI.Offset - WithMinOffset.Offset); |
1507 | } |
1508 | |
1509 | // The aligned loads/stores will use blocks that are either scalars, |
1510 | // or HVX vectors. Let "sector" be the unified term for such a block. |
1511 | // blend(scalar, vector) -> sector... |
1512 | int ScLen = Move.IsHvx ? HVC.HST.getVectorLength() |
1513 | : std::max<int>(a: MinNeeded.value(), b: 4); |
1514 | assert(!Move.IsHvx || ScLen == 64 || ScLen == 128); |
1515 | assert(Move.IsHvx || ScLen == 4 || ScLen == 8); |
1516 | |
1517 | LLVM_DEBUG({ |
1518 | dbgs() << "ScLen: " << ScLen << "\n" ; |
1519 | dbgs() << "AlignVal:" << *AlignVal << "\n" ; |
1520 | dbgs() << "AlignAddr:" << *AlignAddr << "\n" ; |
1521 | dbgs() << "VSpan:\n" << VSpan << '\n'; |
1522 | }); |
1523 | |
1524 | if (Move.IsLoad) |
1525 | realignLoadGroup(Builder, VSpan, ScLen, AlignVal, AlignAddr); |
1526 | else |
1527 | realignStoreGroup(Builder, VSpan, ScLen, AlignVal, AlignAddr); |
1528 | |
1529 | for (auto *Inst : Move.Main) |
1530 | Inst->eraseFromParent(); |
1531 | |
1532 | return true; |
1533 | } |
1534 | |
1535 | auto AlignVectors::makeTestIfUnaligned(IRBuilderBase &Builder, Value *AlignVal, |
1536 | int Alignment) const -> Value * { |
1537 | auto *AlignTy = AlignVal->getType(); |
1538 | Value *And = Builder.CreateAnd( |
1539 | LHS: AlignVal, RHS: ConstantInt::get(Ty: AlignTy, V: Alignment - 1), Name: "and" ); |
1540 | Value *Zero = ConstantInt::get(Ty: AlignTy, V: 0); |
1541 | return Builder.CreateICmpNE(LHS: And, RHS: Zero, Name: "isz" ); |
1542 | } |
1543 | |
1544 | auto AlignVectors::isSectorTy(Type *Ty) const -> bool { |
1545 | if (!HVC.isByteVecTy(Ty)) |
1546 | return false; |
1547 | int Size = HVC.getSizeOf(Ty); |
1548 | if (HVC.HST.isTypeForHVX(VecTy: Ty)) |
1549 | return Size == static_cast<int>(HVC.HST.getVectorLength()); |
1550 | return Size == 4 || Size == 8; |
1551 | } |
1552 | |
1553 | auto AlignVectors::run() -> bool { |
1554 | LLVM_DEBUG(dbgs() << "Running HVC::AlignVectors on " << HVC.F.getName() |
1555 | << '\n'); |
1556 | if (!createAddressGroups()) |
1557 | return false; |
1558 | |
1559 | LLVM_DEBUG({ |
1560 | dbgs() << "Address groups(" << AddrGroups.size() << "):\n" ; |
1561 | for (auto &[In, AL] : AddrGroups) { |
1562 | for (const AddrInfo &AI : AL) |
1563 | dbgs() << "---\n" << AI << '\n'; |
1564 | } |
1565 | }); |
1566 | |
1567 | bool Changed = false; |
1568 | MoveList LoadGroups, StoreGroups; |
1569 | |
1570 | for (auto &G : AddrGroups) { |
1571 | llvm::append_range(C&: LoadGroups, R: createLoadGroups(Group: G.second)); |
1572 | llvm::append_range(C&: StoreGroups, R: createStoreGroups(Group: G.second)); |
1573 | } |
1574 | |
1575 | LLVM_DEBUG({ |
1576 | dbgs() << "\nLoad groups(" << LoadGroups.size() << "):\n" ; |
1577 | for (const MoveGroup &G : LoadGroups) |
1578 | dbgs() << G << "\n" ; |
1579 | dbgs() << "Store groups(" << StoreGroups.size() << "):\n" ; |
1580 | for (const MoveGroup &G : StoreGroups) |
1581 | dbgs() << G << "\n" ; |
1582 | }); |
1583 | |
1584 | // Cumulative limit on the number of groups. |
1585 | unsigned CountLimit = VAGroupCountLimit; |
1586 | if (CountLimit == 0) |
1587 | return false; |
1588 | |
1589 | if (LoadGroups.size() > CountLimit) { |
1590 | LoadGroups.resize(new_size: CountLimit); |
1591 | StoreGroups.clear(); |
1592 | } else { |
1593 | unsigned StoreLimit = CountLimit - LoadGroups.size(); |
1594 | if (StoreGroups.size() > StoreLimit) |
1595 | StoreGroups.resize(new_size: StoreLimit); |
1596 | } |
1597 | |
1598 | for (auto &M : LoadGroups) |
1599 | Changed |= moveTogether(Move&: M); |
1600 | for (auto &M : StoreGroups) |
1601 | Changed |= moveTogether(Move&: M); |
1602 | |
1603 | LLVM_DEBUG(dbgs() << "After moveTogether:\n" << HVC.F); |
1604 | |
1605 | for (auto &M : LoadGroups) |
1606 | Changed |= realignGroup(Move: M); |
1607 | for (auto &M : StoreGroups) |
1608 | Changed |= realignGroup(Move: M); |
1609 | |
1610 | return Changed; |
1611 | } |
1612 | |
1613 | // --- End AlignVectors |
1614 | |
1615 | // --- Begin HvxIdioms |
1616 | |
1617 | auto HvxIdioms::getNumSignificantBits(Value *V, Instruction *In) const |
1618 | -> std::pair<unsigned, Signedness> { |
1619 | unsigned Bits = HVC.getNumSignificantBits(V, CtxI: In); |
1620 | // The significant bits are calculated including the sign bit. This may |
1621 | // add an extra bit for zero-extended values, e.g. (zext i32 to i64) may |
1622 | // result in 33 significant bits. To avoid extra words, skip the extra |
1623 | // sign bit, but keep information that the value is to be treated as |
1624 | // unsigned. |
1625 | KnownBits Known = HVC.getKnownBits(V, CtxI: In); |
1626 | Signedness Sign = Signed; |
1627 | unsigned NumToTest = 0; // Number of bits used in test for unsignedness. |
1628 | if (isPowerOf2_32(Value: Bits)) |
1629 | NumToTest = Bits; |
1630 | else if (Bits > 1 && isPowerOf2_32(Value: Bits - 1)) |
1631 | NumToTest = Bits - 1; |
1632 | |
1633 | if (NumToTest != 0 && Known.Zero.ashr(ShiftAmt: NumToTest).isAllOnes()) { |
1634 | Sign = Unsigned; |
1635 | Bits = NumToTest; |
1636 | } |
1637 | |
1638 | // If the top bit of the nearest power-of-2 is zero, this value is |
1639 | // positive. It could be treated as either signed or unsigned. |
1640 | if (unsigned Pow2 = PowerOf2Ceil(A: Bits); Pow2 != Bits) { |
1641 | if (Known.Zero.ashr(ShiftAmt: Pow2 - 1).isAllOnes()) |
1642 | Sign = Positive; |
1643 | } |
1644 | return {Bits, Sign}; |
1645 | } |
1646 | |
1647 | auto HvxIdioms::canonSgn(SValue X, SValue Y) const |
1648 | -> std::pair<SValue, SValue> { |
1649 | // Canonicalize the signedness of X and Y, so that the result is one of: |
1650 | // S, S |
1651 | // U/P, S |
1652 | // U/P, U/P |
1653 | if (X.Sgn == Signed && Y.Sgn != Signed) |
1654 | std::swap(a&: X, b&: Y); |
1655 | return {X, Y}; |
1656 | } |
1657 | |
1658 | // Match |
1659 | // (X * Y) [>> N], or |
1660 | // ((X * Y) + (1 << M)) >> N |
1661 | auto HvxIdioms::matchFxpMul(Instruction &In) const -> std::optional<FxpOp> { |
1662 | using namespace PatternMatch; |
1663 | auto *Ty = In.getType(); |
1664 | |
1665 | if (!Ty->isVectorTy() || !Ty->getScalarType()->isIntegerTy()) |
1666 | return std::nullopt; |
1667 | |
1668 | unsigned Width = cast<IntegerType>(Val: Ty->getScalarType())->getBitWidth(); |
1669 | |
1670 | FxpOp Op; |
1671 | Value *Exp = &In; |
1672 | |
1673 | // Fixed-point multiplication is always shifted right (except when the |
1674 | // fraction is 0 bits). |
1675 | auto m_Shr = [](auto &&V, auto &&S) { |
1676 | return m_CombineOr(m_LShr(V, S), m_AShr(V, S)); |
1677 | }; |
1678 | |
1679 | const APInt *Qn = nullptr; |
1680 | if (Value * T; match(V: Exp, P: m_Shr(m_Value(V&: T), m_APInt(Res&: Qn)))) { |
1681 | Op.Frac = Qn->getZExtValue(); |
1682 | Exp = T; |
1683 | } else { |
1684 | Op.Frac = 0; |
1685 | } |
1686 | |
1687 | if (Op.Frac > Width) |
1688 | return std::nullopt; |
1689 | |
1690 | // Check if there is rounding added. |
1691 | const APInt *C = nullptr; |
1692 | if (Value * T; Op.Frac > 0 && match(V: Exp, P: m_Add(L: m_Value(V&: T), R: m_APInt(Res&: C)))) { |
1693 | uint64_t CV = C->getZExtValue(); |
1694 | if (CV != 0 && !isPowerOf2_64(Value: CV)) |
1695 | return std::nullopt; |
1696 | if (CV != 0) |
1697 | Op.RoundAt = Log2_64(Value: CV); |
1698 | Exp = T; |
1699 | } |
1700 | |
1701 | // Check if the rest is a multiplication. |
1702 | if (match(V: Exp, P: m_Mul(L: m_Value(V&: Op.X.Val), R: m_Value(V&: Op.Y.Val)))) { |
1703 | Op.Opcode = Instruction::Mul; |
1704 | // FIXME: The information below is recomputed. |
1705 | Op.X.Sgn = getNumSignificantBits(V: Op.X.Val, In: &In).second; |
1706 | Op.Y.Sgn = getNumSignificantBits(V: Op.Y.Val, In: &In).second; |
1707 | Op.ResTy = cast<VectorType>(Val: Ty); |
1708 | return Op; |
1709 | } |
1710 | |
1711 | return std::nullopt; |
1712 | } |
1713 | |
1714 | auto HvxIdioms::processFxpMul(Instruction &In, const FxpOp &Op) const |
1715 | -> Value * { |
1716 | assert(Op.X.Val->getType() == Op.Y.Val->getType()); |
1717 | |
1718 | auto *VecTy = dyn_cast<VectorType>(Val: Op.X.Val->getType()); |
1719 | if (VecTy == nullptr) |
1720 | return nullptr; |
1721 | auto *ElemTy = cast<IntegerType>(Val: VecTy->getElementType()); |
1722 | unsigned ElemWidth = ElemTy->getBitWidth(); |
1723 | |
1724 | // TODO: This can be relaxed after legalization is done pre-isel. |
1725 | if ((HVC.length(Ty: VecTy) * ElemWidth) % (8 * HVC.HST.getVectorLength()) != 0) |
1726 | return nullptr; |
1727 | |
1728 | // There are no special intrinsics that should be used for multiplying |
1729 | // signed 8-bit values, so just skip them. Normal codegen should handle |
1730 | // this just fine. |
1731 | if (ElemWidth <= 8) |
1732 | return nullptr; |
1733 | // Similarly, if this is just a multiplication that can be handled without |
1734 | // intervention, then leave it alone. |
1735 | if (ElemWidth <= 32 && Op.Frac == 0) |
1736 | return nullptr; |
1737 | |
1738 | auto [BitsX, SignX] = getNumSignificantBits(V: Op.X.Val, In: &In); |
1739 | auto [BitsY, SignY] = getNumSignificantBits(V: Op.Y.Val, In: &In); |
1740 | |
1741 | // TODO: Add multiplication of vectors by scalar registers (up to 4 bytes). |
1742 | |
1743 | Value *X = Op.X.Val, *Y = Op.Y.Val; |
1744 | IRBuilder Builder(In.getParent(), In.getIterator(), |
1745 | InstSimplifyFolder(HVC.DL)); |
1746 | |
1747 | auto roundUpWidth = [](unsigned Width) -> unsigned { |
1748 | if (Width <= 32 && !isPowerOf2_32(Value: Width)) { |
1749 | // If the element width is not a power of 2, round it up |
1750 | // to the next one. Do this for widths not exceeding 32. |
1751 | return PowerOf2Ceil(A: Width); |
1752 | } |
1753 | if (Width > 32 && Width % 32 != 0) { |
1754 | // For wider elements, round it up to the multiple of 32. |
1755 | return alignTo(Value: Width, Align: 32u); |
1756 | } |
1757 | return Width; |
1758 | }; |
1759 | |
1760 | BitsX = roundUpWidth(BitsX); |
1761 | BitsY = roundUpWidth(BitsY); |
1762 | |
1763 | // For elementwise multiplication vectors must have the same lengths, so |
1764 | // resize the elements of both inputs to the same width, the max of the |
1765 | // calculated significant bits. |
1766 | unsigned Width = std::max(a: BitsX, b: BitsY); |
1767 | |
1768 | auto *ResizeTy = VectorType::get(ElementType: HVC.getIntTy(Width), Other: VecTy); |
1769 | if (Width < ElemWidth) { |
1770 | X = Builder.CreateTrunc(V: X, DestTy: ResizeTy, Name: "trn" ); |
1771 | Y = Builder.CreateTrunc(V: Y, DestTy: ResizeTy, Name: "trn" ); |
1772 | } else if (Width > ElemWidth) { |
1773 | X = SignX == Signed ? Builder.CreateSExt(V: X, DestTy: ResizeTy, Name: "sxt" ) |
1774 | : Builder.CreateZExt(V: X, DestTy: ResizeTy, Name: "zxt" ); |
1775 | Y = SignY == Signed ? Builder.CreateSExt(V: Y, DestTy: ResizeTy, Name: "sxt" ) |
1776 | : Builder.CreateZExt(V: Y, DestTy: ResizeTy, Name: "zxt" ); |
1777 | }; |
1778 | |
1779 | assert(X->getType() == Y->getType() && X->getType() == ResizeTy); |
1780 | |
1781 | unsigned VecLen = HVC.length(Ty: ResizeTy); |
1782 | unsigned ChopLen = (8 * HVC.HST.getVectorLength()) / std::min(a: Width, b: 32u); |
1783 | |
1784 | SmallVector<Value *> Results; |
1785 | FxpOp ChopOp = Op; |
1786 | ChopOp.ResTy = VectorType::get(ElementType: Op.ResTy->getElementType(), NumElements: ChopLen, Scalable: false); |
1787 | |
1788 | for (unsigned V = 0; V != VecLen / ChopLen; ++V) { |
1789 | ChopOp.X.Val = HVC.subvector(Builder, Val: X, Start: V * ChopLen, Length: ChopLen); |
1790 | ChopOp.Y.Val = HVC.subvector(Builder, Val: Y, Start: V * ChopLen, Length: ChopLen); |
1791 | Results.push_back(Elt: processFxpMulChopped(Builder, In, Op: ChopOp)); |
1792 | if (Results.back() == nullptr) |
1793 | break; |
1794 | } |
1795 | |
1796 | if (Results.empty() || Results.back() == nullptr) |
1797 | return nullptr; |
1798 | |
1799 | Value *Cat = HVC.concat(Builder, Vecs: Results); |
1800 | Value *Ext = SignX == Signed || SignY == Signed |
1801 | ? Builder.CreateSExt(V: Cat, DestTy: VecTy, Name: "sxt" ) |
1802 | : Builder.CreateZExt(V: Cat, DestTy: VecTy, Name: "zxt" ); |
1803 | return Ext; |
1804 | } |
1805 | |
1806 | auto HvxIdioms::processFxpMulChopped(IRBuilderBase &Builder, Instruction &In, |
1807 | const FxpOp &Op) const -> Value * { |
1808 | assert(Op.X.Val->getType() == Op.Y.Val->getType()); |
1809 | auto *InpTy = cast<VectorType>(Val: Op.X.Val->getType()); |
1810 | unsigned Width = InpTy->getScalarSizeInBits(); |
1811 | bool Rounding = Op.RoundAt.has_value(); |
1812 | |
1813 | if (!Op.RoundAt || *Op.RoundAt == Op.Frac - 1) { |
1814 | // The fixed-point intrinsics do signed multiplication. |
1815 | if (Width == Op.Frac + 1 && Op.X.Sgn != Unsigned && Op.Y.Sgn != Unsigned) { |
1816 | Value *QMul = nullptr; |
1817 | if (Width == 16) { |
1818 | QMul = createMulQ15(Builder, X: Op.X, Y: Op.Y, Rounding); |
1819 | } else if (Width == 32) { |
1820 | QMul = createMulQ31(Builder, X: Op.X, Y: Op.Y, Rounding); |
1821 | } |
1822 | if (QMul != nullptr) |
1823 | return QMul; |
1824 | } |
1825 | } |
1826 | |
1827 | assert(Width >= 32 || isPowerOf2_32(Width)); // Width <= 32 => Width is 2^n |
1828 | assert(Width < 32 || Width % 32 == 0); // Width > 32 => Width is 32*k |
1829 | |
1830 | // If Width < 32, then it should really be 16. |
1831 | if (Width < 32) { |
1832 | if (Width < 16) |
1833 | return nullptr; |
1834 | // Getting here with Op.Frac == 0 isn't wrong, but suboptimal: here we |
1835 | // generate a full precision products, which is unnecessary if there is |
1836 | // no shift. |
1837 | assert(Width == 16); |
1838 | assert(Op.Frac != 0 && "Unshifted mul should have been skipped" ); |
1839 | if (Op.Frac == 16) { |
1840 | // Multiply high |
1841 | if (Value *MulH = createMulH16(Builder, X: Op.X, Y: Op.Y)) |
1842 | return MulH; |
1843 | } |
1844 | // Do full-precision multiply and shift. |
1845 | Value *Prod32 = createMul16(Builder, X: Op.X, Y: Op.Y); |
1846 | if (Rounding) { |
1847 | Value *RoundVal = HVC.getConstSplat(Ty: Prod32->getType(), Val: 1 << *Op.RoundAt); |
1848 | Prod32 = Builder.CreateAdd(LHS: Prod32, RHS: RoundVal, Name: "add" ); |
1849 | } |
1850 | |
1851 | Value *ShiftAmt = HVC.getConstSplat(Ty: Prod32->getType(), Val: Op.Frac); |
1852 | Value *Shifted = Op.X.Sgn == Signed || Op.Y.Sgn == Signed |
1853 | ? Builder.CreateAShr(LHS: Prod32, RHS: ShiftAmt, Name: "asr" ) |
1854 | : Builder.CreateLShr(LHS: Prod32, RHS: ShiftAmt, Name: "lsr" ); |
1855 | return Builder.CreateTrunc(V: Shifted, DestTy: InpTy, Name: "trn" ); |
1856 | } |
1857 | |
1858 | // Width >= 32 |
1859 | |
1860 | // Break up the arguments Op.X and Op.Y into vectors of smaller widths |
1861 | // in preparation of doing the multiplication by 32-bit parts. |
1862 | auto WordX = HVC.splitVectorElements(Builder, Vec: Op.X.Val, /*ToWidth=*/32); |
1863 | auto WordY = HVC.splitVectorElements(Builder, Vec: Op.Y.Val, /*ToWidth=*/32); |
1864 | auto WordP = createMulLong(Builder, WordX, SgnX: Op.X.Sgn, WordY, SgnY: Op.Y.Sgn); |
1865 | |
1866 | auto *HvxWordTy = cast<VectorType>(Val: WordP.front()->getType()); |
1867 | |
1868 | // Add the optional rounding to the proper word. |
1869 | if (Op.RoundAt.has_value()) { |
1870 | Value *Zero = HVC.getNullValue(Ty: WordX[0]->getType()); |
1871 | SmallVector<Value *> RoundV(WordP.size(), Zero); |
1872 | RoundV[*Op.RoundAt / 32] = |
1873 | HVC.getConstSplat(Ty: HvxWordTy, Val: 1 << (*Op.RoundAt % 32)); |
1874 | WordP = createAddLong(Builder, WordX: WordP, WordY: RoundV); |
1875 | } |
1876 | |
1877 | // createRightShiftLong? |
1878 | |
1879 | // Shift all products right by Op.Frac. |
1880 | unsigned SkipWords = Op.Frac / 32; |
1881 | Constant *ShiftAmt = HVC.getConstSplat(Ty: HvxWordTy, Val: Op.Frac % 32); |
1882 | |
1883 | for (int Dst = 0, End = WordP.size() - SkipWords; Dst != End; ++Dst) { |
1884 | int Src = Dst + SkipWords; |
1885 | Value *Lo = WordP[Src]; |
1886 | if (Src + 1 < End) { |
1887 | Value *Hi = WordP[Src + 1]; |
1888 | WordP[Dst] = Builder.CreateIntrinsic(RetTy: HvxWordTy, ID: Intrinsic::fshr, |
1889 | Args: {Hi, Lo, ShiftAmt}, |
1890 | /*FMFSource*/ nullptr, Name: "int" ); |
1891 | } else { |
1892 | // The shift of the most significant word. |
1893 | WordP[Dst] = Builder.CreateAShr(LHS: Lo, RHS: ShiftAmt, Name: "asr" ); |
1894 | } |
1895 | } |
1896 | if (SkipWords != 0) |
1897 | WordP.resize(N: WordP.size() - SkipWords); |
1898 | |
1899 | return HVC.joinVectorElements(Builder, Values: WordP, ToType: Op.ResTy); |
1900 | } |
1901 | |
1902 | auto HvxIdioms::createMulQ15(IRBuilderBase &Builder, SValue X, SValue Y, |
1903 | bool Rounding) const -> Value * { |
1904 | assert(X.Val->getType() == Y.Val->getType()); |
1905 | assert(X.Val->getType()->getScalarType() == HVC.getIntTy(16)); |
1906 | assert(HVC.HST.isHVXVectorType(EVT::getEVT(X.Val->getType(), false))); |
1907 | |
1908 | // There is no non-rounding intrinsic for i16. |
1909 | if (!Rounding || X.Sgn == Unsigned || Y.Sgn == Unsigned) |
1910 | return nullptr; |
1911 | |
1912 | auto V6_vmpyhvsrs = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyhvsrs); |
1913 | return HVC.createHvxIntrinsic(Builder, IntID: V6_vmpyhvsrs, RetTy: X.Val->getType(), |
1914 | Args: {X.Val, Y.Val}); |
1915 | } |
1916 | |
1917 | auto HvxIdioms::createMulQ31(IRBuilderBase &Builder, SValue X, SValue Y, |
1918 | bool Rounding) const -> Value * { |
1919 | Type *InpTy = X.Val->getType(); |
1920 | assert(InpTy == Y.Val->getType()); |
1921 | assert(InpTy->getScalarType() == HVC.getIntTy(32)); |
1922 | assert(HVC.HST.isHVXVectorType(EVT::getEVT(InpTy, false))); |
1923 | |
1924 | if (X.Sgn == Unsigned || Y.Sgn == Unsigned) |
1925 | return nullptr; |
1926 | |
1927 | auto V6_vmpyewuh = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyewuh); |
1928 | auto V6_vmpyo_acc = Rounding |
1929 | ? HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyowh_rnd_sacc) |
1930 | : HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyowh_sacc); |
1931 | Value *V1 = |
1932 | HVC.createHvxIntrinsic(Builder, IntID: V6_vmpyewuh, RetTy: InpTy, Args: {X.Val, Y.Val}); |
1933 | return HVC.createHvxIntrinsic(Builder, IntID: V6_vmpyo_acc, RetTy: InpTy, |
1934 | Args: {V1, X.Val, Y.Val}); |
1935 | } |
1936 | |
1937 | auto HvxIdioms::createAddCarry(IRBuilderBase &Builder, Value *X, Value *Y, |
1938 | Value *CarryIn) const |
1939 | -> std::pair<Value *, Value *> { |
1940 | assert(X->getType() == Y->getType()); |
1941 | auto VecTy = cast<VectorType>(Val: X->getType()); |
1942 | if (VecTy == HvxI32Ty && HVC.HST.useHVXV62Ops()) { |
1943 | SmallVector<Value *> Args = {X, Y}; |
1944 | Intrinsic::ID AddCarry; |
1945 | if (CarryIn == nullptr && HVC.HST.useHVXV66Ops()) { |
1946 | AddCarry = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vaddcarryo); |
1947 | } else { |
1948 | AddCarry = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vaddcarry); |
1949 | if (CarryIn == nullptr) |
1950 | CarryIn = HVC.getNullValue(Ty: HVC.getBoolTy(ElemCount: HVC.length(Ty: VecTy))); |
1951 | Args.push_back(Elt: CarryIn); |
1952 | } |
1953 | Value *Ret = HVC.createHvxIntrinsic(Builder, IntID: AddCarry, |
1954 | /*RetTy=*/nullptr, Args); |
1955 | Value *Result = Builder.CreateExtractValue(Agg: Ret, Idxs: {0}, Name: "ext" ); |
1956 | Value *CarryOut = Builder.CreateExtractValue(Agg: Ret, Idxs: {1}, Name: "ext" ); |
1957 | return {Result, CarryOut}; |
1958 | } |
1959 | |
1960 | // In other cases, do a regular add, and unsigned compare-less-than. |
1961 | // The carry-out can originate in two places: adding the carry-in or adding |
1962 | // the two input values. |
1963 | Value *Result1 = X; // Result1 = X + CarryIn |
1964 | if (CarryIn != nullptr) { |
1965 | unsigned Width = VecTy->getScalarSizeInBits(); |
1966 | uint32_t Mask = 1; |
1967 | if (Width < 32) { |
1968 | for (unsigned i = 0, e = 32 / Width; i != e; ++i) |
1969 | Mask = (Mask << Width) | 1; |
1970 | } |
1971 | auto V6_vandqrt = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vandqrt); |
1972 | Value *ValueIn = |
1973 | HVC.createHvxIntrinsic(Builder, IntID: V6_vandqrt, /*RetTy=*/nullptr, |
1974 | Args: {CarryIn, HVC.getConstInt(Val: Mask)}); |
1975 | Result1 = Builder.CreateAdd(LHS: X, RHS: ValueIn, Name: "add" ); |
1976 | } |
1977 | |
1978 | Value *CarryOut1 = Builder.CreateCmp(Pred: CmpInst::ICMP_ULT, LHS: Result1, RHS: X, Name: "cmp" ); |
1979 | Value *Result2 = Builder.CreateAdd(LHS: Result1, RHS: Y, Name: "add" ); |
1980 | Value *CarryOut2 = Builder.CreateCmp(Pred: CmpInst::ICMP_ULT, LHS: Result2, RHS: Y, Name: "cmp" ); |
1981 | return {Result2, Builder.CreateOr(LHS: CarryOut1, RHS: CarryOut2, Name: "orb" )}; |
1982 | } |
1983 | |
1984 | auto HvxIdioms::createMul16(IRBuilderBase &Builder, SValue X, SValue Y) const |
1985 | -> Value * { |
1986 | Intrinsic::ID V6_vmpyh = 0; |
1987 | std::tie(args&: X, args&: Y) = canonSgn(X, Y); |
1988 | |
1989 | if (X.Sgn == Signed) { |
1990 | V6_vmpyh = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyhv); |
1991 | } else if (Y.Sgn == Signed) { |
1992 | // In vmpyhus the second operand is unsigned |
1993 | V6_vmpyh = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyhus); |
1994 | } else { |
1995 | V6_vmpyh = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyuhv); |
1996 | } |
1997 | |
1998 | // i16*i16 -> i32 / interleaved |
1999 | Value *P = |
2000 | HVC.createHvxIntrinsic(Builder, IntID: V6_vmpyh, RetTy: HvxP32Ty, Args: {Y.Val, X.Val}); |
2001 | // Deinterleave |
2002 | return HVC.vshuff(Builder, Val0: HVC.sublo(Builder, Val: P), Val1: HVC.subhi(Builder, Val: P)); |
2003 | } |
2004 | |
2005 | auto HvxIdioms::createMulH16(IRBuilderBase &Builder, SValue X, SValue Y) const |
2006 | -> Value * { |
2007 | Type *HvxI16Ty = HVC.getHvxTy(ElemTy: HVC.getIntTy(Width: 16), /*Pair=*/false); |
2008 | |
2009 | if (HVC.HST.useHVXV69Ops()) { |
2010 | if (X.Sgn != Signed && Y.Sgn != Signed) { |
2011 | auto V6_vmpyuhvs = HVC.HST.getIntrinsicId(Opc: Hexagon::V6_vmpyuhvs); |
2012 | return HVC.createHvxIntrinsic(Builder, IntID: V6_vmpyuhvs, RetTy: HvxI16Ty, |
2013 | Args: {X.Val, Y.Val}); |
2014 | } |
2015 | } |
2016 | |
2017 | Type *HvxP16Ty = HVC.getHvxTy(ElemTy: HVC.getIntTy(Width: 16), /*Pair=*/true); |
2018 | Value *Pair16 = |
2019 | Builder.CreateBitCast(V: createMul16(Builder, X, Y), DestTy: HvxP16Ty, Name: "cst" ); |
2020 | unsigned Len = HVC.length(Ty: HvxP16Ty) / 2; |
2021 | |
2022 | SmallVector<int, 128> PickOdd(Len); |
2023 | for (int i = 0; i != static_cast<int>(Len); ++i) |
2024 | PickOdd[i] = 2 * i + 1; |
2025 | |
2026 | return Builder.CreateShuffleVector( |
2027 | V1: HVC.sublo(Builder, Val: Pair16), V2: HVC.subhi(Builder, Val: Pair16), Mask: PickOdd, Name: "shf" ); |
2028 | } |
2029 | |
2030 | auto HvxIdioms::createMul32(IRBuilderBase &Builder, SValue X, SValue Y) const |
2031 | -> std::pair<Value *, Value *> { |
2032 | assert(X.Val->getType() == Y.Val->getType()); |
2033 | assert(X.Val->getType() == HvxI32Ty); |
2034 | |
2035 | Intrinsic::ID V6_vmpy_parts; |
2036 | std::tie(args&: X, args&: Y) = canonSgn(X, Y); |
2037 | |
2038 | if (X.Sgn == Signed) { |
2039 | V6_vmpy_parts = Intrinsic::hexagon_V6_vmpyss_parts; |
2040 | } else if (Y.Sgn == Signed) { |
2041 | V6_vmpy_parts = Intrinsic::hexagon_V6_vmpyus_parts; |
2042 | } else { |
2043 | V6_vmpy_parts = Intrinsic::hexagon_V6_vmpyuu_parts; |
2044 | } |
2045 | |
2046 | Value *Parts = HVC.createHvxIntrinsic(Builder, IntID: V6_vmpy_parts, RetTy: nullptr, |
2047 | Args: {X.Val, Y.Val}, ArgTys: {HvxI32Ty}); |
2048 | Value *Hi = Builder.CreateExtractValue(Agg: Parts, Idxs: {0}, Name: "ext" ); |
2049 | Value *Lo = Builder.CreateExtractValue(Agg: Parts, Idxs: {1}, Name: "ext" ); |
2050 | return {Lo, Hi}; |
2051 | } |
2052 | |
2053 | auto HvxIdioms::createAddLong(IRBuilderBase &Builder, ArrayRef<Value *> WordX, |
2054 | ArrayRef<Value *> WordY) const |
2055 | -> SmallVector<Value *> { |
2056 | assert(WordX.size() == WordY.size()); |
2057 | unsigned Idx = 0, Length = WordX.size(); |
2058 | SmallVector<Value *> Sum(Length); |
2059 | |
2060 | while (Idx != Length) { |
2061 | if (HVC.isZero(Val: WordX[Idx])) |
2062 | Sum[Idx] = WordY[Idx]; |
2063 | else if (HVC.isZero(Val: WordY[Idx])) |
2064 | Sum[Idx] = WordX[Idx]; |
2065 | else |
2066 | break; |
2067 | ++Idx; |
2068 | } |
2069 | |
2070 | Value *Carry = nullptr; |
2071 | for (; Idx != Length; ++Idx) { |
2072 | std::tie(args&: Sum[Idx], args&: Carry) = |
2073 | createAddCarry(Builder, X: WordX[Idx], Y: WordY[Idx], CarryIn: Carry); |
2074 | } |
2075 | |
2076 | // This drops the final carry beyond the highest word. |
2077 | return Sum; |
2078 | } |
2079 | |
2080 | auto HvxIdioms::createMulLong(IRBuilderBase &Builder, ArrayRef<Value *> WordX, |
2081 | Signedness SgnX, ArrayRef<Value *> WordY, |
2082 | Signedness SgnY) const -> SmallVector<Value *> { |
2083 | SmallVector<SmallVector<Value *>> Products(WordX.size() + WordY.size()); |
2084 | |
2085 | // WordX[i] * WordY[j] produces words i+j and i+j+1 of the results, |
2086 | // that is halves 2(i+j), 2(i+j)+1, 2(i+j)+2, 2(i+j)+3. |
2087 | for (int i = 0, e = WordX.size(); i != e; ++i) { |
2088 | for (int j = 0, f = WordY.size(); j != f; ++j) { |
2089 | // Check the 4 halves that this multiplication can generate. |
2090 | Signedness SX = (i + 1 == e) ? SgnX : Unsigned; |
2091 | Signedness SY = (j + 1 == f) ? SgnY : Unsigned; |
2092 | auto [Lo, Hi] = createMul32(Builder, X: {.Val: WordX[i], .Sgn: SX}, Y: {.Val: WordY[j], .Sgn: SY}); |
2093 | Products[i + j + 0].push_back(Elt: Lo); |
2094 | Products[i + j + 1].push_back(Elt: Hi); |
2095 | } |
2096 | } |
2097 | |
2098 | Value *Zero = HVC.getNullValue(Ty: WordX[0]->getType()); |
2099 | |
2100 | auto pop_back_or_zero = [Zero](auto &Vector) -> Value * { |
2101 | if (Vector.empty()) |
2102 | return Zero; |
2103 | auto Last = Vector.back(); |
2104 | Vector.pop_back(); |
2105 | return Last; |
2106 | }; |
2107 | |
2108 | for (int i = 0, e = Products.size(); i != e; ++i) { |
2109 | while (Products[i].size() > 1) { |
2110 | Value *Carry = nullptr; // no carry-in |
2111 | for (int j = i; j != e; ++j) { |
2112 | auto &ProdJ = Products[j]; |
2113 | auto [Sum, CarryOut] = createAddCarry(Builder, X: pop_back_or_zero(ProdJ), |
2114 | Y: pop_back_or_zero(ProdJ), CarryIn: Carry); |
2115 | ProdJ.insert(I: ProdJ.begin(), Elt: Sum); |
2116 | Carry = CarryOut; |
2117 | } |
2118 | } |
2119 | } |
2120 | |
2121 | SmallVector<Value *> WordP; |
2122 | for (auto &P : Products) { |
2123 | assert(P.size() == 1 && "Should have been added together" ); |
2124 | WordP.push_back(Elt: P.front()); |
2125 | } |
2126 | |
2127 | return WordP; |
2128 | } |
2129 | |
2130 | auto HvxIdioms::run() -> bool { |
2131 | bool Changed = false; |
2132 | |
2133 | for (BasicBlock &B : HVC.F) { |
2134 | for (auto It = B.rbegin(); It != B.rend(); ++It) { |
2135 | if (auto Fxm = matchFxpMul(In&: *It)) { |
2136 | Value *New = processFxpMul(In&: *It, Op: *Fxm); |
2137 | // Always report "changed" for now. |
2138 | Changed = true; |
2139 | if (!New) |
2140 | continue; |
2141 | bool StartOver = !isa<Instruction>(Val: New); |
2142 | It->replaceAllUsesWith(V: New); |
2143 | RecursivelyDeleteTriviallyDeadInstructions(V: &*It, TLI: &HVC.TLI); |
2144 | It = StartOver ? B.rbegin() |
2145 | : cast<Instruction>(Val: New)->getReverseIterator(); |
2146 | Changed = true; |
2147 | } |
2148 | } |
2149 | } |
2150 | |
2151 | return Changed; |
2152 | } |
2153 | |
2154 | // --- End HvxIdioms |
2155 | |
2156 | auto HexagonVectorCombine::run() -> bool { |
2157 | if (DumpModule) |
2158 | dbgs() << "Module before HexagonVectorCombine\n" << *F.getParent(); |
2159 | |
2160 | bool Changed = false; |
2161 | if (HST.useHVXOps()) { |
2162 | if (VAEnabled) |
2163 | Changed |= AlignVectors(*this).run(); |
2164 | if (VIEnabled) |
2165 | Changed |= HvxIdioms(*this).run(); |
2166 | } |
2167 | |
2168 | if (DumpModule) { |
2169 | dbgs() << "Module " << (Changed ? "(modified)" : "(unchanged)" ) |
2170 | << " after HexagonVectorCombine\n" |
2171 | << *F.getParent(); |
2172 | } |
2173 | return Changed; |
2174 | } |
2175 | |
2176 | auto HexagonVectorCombine::getIntTy(unsigned Width) const -> IntegerType * { |
2177 | return IntegerType::get(C&: F.getContext(), NumBits: Width); |
2178 | } |
2179 | |
2180 | auto HexagonVectorCombine::getByteTy(int ElemCount) const -> Type * { |
2181 | assert(ElemCount >= 0); |
2182 | IntegerType *ByteTy = Type::getInt8Ty(C&: F.getContext()); |
2183 | if (ElemCount == 0) |
2184 | return ByteTy; |
2185 | return VectorType::get(ElementType: ByteTy, NumElements: ElemCount, /*Scalable=*/false); |
2186 | } |
2187 | |
2188 | auto HexagonVectorCombine::getBoolTy(int ElemCount) const -> Type * { |
2189 | assert(ElemCount >= 0); |
2190 | IntegerType *BoolTy = Type::getInt1Ty(C&: F.getContext()); |
2191 | if (ElemCount == 0) |
2192 | return BoolTy; |
2193 | return VectorType::get(ElementType: BoolTy, NumElements: ElemCount, /*Scalable=*/false); |
2194 | } |
2195 | |
2196 | auto HexagonVectorCombine::getConstInt(int Val, unsigned Width) const |
2197 | -> ConstantInt * { |
2198 | return ConstantInt::getSigned(Ty: getIntTy(Width), V: Val); |
2199 | } |
2200 | |
2201 | auto HexagonVectorCombine::isZero(const Value *Val) const -> bool { |
2202 | if (auto *C = dyn_cast<Constant>(Val)) |
2203 | return C->isZeroValue(); |
2204 | return false; |
2205 | } |
2206 | |
2207 | auto HexagonVectorCombine::getIntValue(const Value *Val) const |
2208 | -> std::optional<APInt> { |
2209 | if (auto *CI = dyn_cast<ConstantInt>(Val)) |
2210 | return CI->getValue(); |
2211 | return std::nullopt; |
2212 | } |
2213 | |
2214 | auto HexagonVectorCombine::isUndef(const Value *Val) const -> bool { |
2215 | return isa<UndefValue>(Val); |
2216 | } |
2217 | |
2218 | auto HexagonVectorCombine::isTrue(const Value *Val) const -> bool { |
2219 | return Val == ConstantInt::getTrue(Ty: Val->getType()); |
2220 | } |
2221 | |
2222 | auto HexagonVectorCombine::isFalse(const Value *Val) const -> bool { |
2223 | return isZero(Val); |
2224 | } |
2225 | |
2226 | auto HexagonVectorCombine::getHvxTy(Type *ElemTy, bool Pair) const |
2227 | -> VectorType * { |
2228 | EVT ETy = EVT::getEVT(Ty: ElemTy, HandleUnknown: false); |
2229 | assert(ETy.isSimple() && "Invalid HVX element type" ); |
2230 | // Do not allow boolean types here: they don't have a fixed length. |
2231 | assert(HST.isHVXElementType(ETy.getSimpleVT(), /*IncludeBool=*/false) && |
2232 | "Invalid HVX element type" ); |
2233 | unsigned HwLen = HST.getVectorLength(); |
2234 | unsigned NumElems = (8 * HwLen) / ETy.getSizeInBits(); |
2235 | return VectorType::get(ElementType: ElemTy, NumElements: Pair ? 2 * NumElems : NumElems, |
2236 | /*Scalable=*/false); |
2237 | } |
2238 | |
2239 | auto HexagonVectorCombine::getSizeOf(const Value *Val, SizeKind Kind) const |
2240 | -> int { |
2241 | return getSizeOf(Ty: Val->getType(), Kind); |
2242 | } |
2243 | |
2244 | auto HexagonVectorCombine::getSizeOf(const Type *Ty, SizeKind Kind) const |
2245 | -> int { |
2246 | auto *NcTy = const_cast<Type *>(Ty); |
2247 | switch (Kind) { |
2248 | case Store: |
2249 | return DL.getTypeStoreSize(Ty: NcTy).getFixedValue(); |
2250 | case Alloc: |
2251 | return DL.getTypeAllocSize(Ty: NcTy).getFixedValue(); |
2252 | } |
2253 | llvm_unreachable("Unhandled SizeKind enum" ); |
2254 | } |
2255 | |
2256 | auto HexagonVectorCombine::getTypeAlignment(Type *Ty) const -> int { |
2257 | // The actual type may be shorter than the HVX vector, so determine |
2258 | // the alignment based on subtarget info. |
2259 | if (HST.isTypeForHVX(VecTy: Ty)) |
2260 | return HST.getVectorLength(); |
2261 | return DL.getABITypeAlign(Ty).value(); |
2262 | } |
2263 | |
2264 | auto HexagonVectorCombine::length(Value *Val) const -> size_t { |
2265 | return length(Ty: Val->getType()); |
2266 | } |
2267 | |
2268 | auto HexagonVectorCombine::length(Type *Ty) const -> size_t { |
2269 | auto *VecTy = dyn_cast<VectorType>(Val: Ty); |
2270 | assert(VecTy && "Must be a vector type" ); |
2271 | return VecTy->getElementCount().getFixedValue(); |
2272 | } |
2273 | |
2274 | auto HexagonVectorCombine::getNullValue(Type *Ty) const -> Constant * { |
2275 | assert(Ty->isIntOrIntVectorTy()); |
2276 | auto Zero = ConstantInt::get(Ty: Ty->getScalarType(), V: 0); |
2277 | if (auto *VecTy = dyn_cast<VectorType>(Val: Ty)) |
2278 | return ConstantVector::getSplat(EC: VecTy->getElementCount(), Elt: Zero); |
2279 | return Zero; |
2280 | } |
2281 | |
2282 | auto HexagonVectorCombine::getFullValue(Type *Ty) const -> Constant * { |
2283 | assert(Ty->isIntOrIntVectorTy()); |
2284 | auto Minus1 = ConstantInt::get(Ty: Ty->getScalarType(), V: -1); |
2285 | if (auto *VecTy = dyn_cast<VectorType>(Val: Ty)) |
2286 | return ConstantVector::getSplat(EC: VecTy->getElementCount(), Elt: Minus1); |
2287 | return Minus1; |
2288 | } |
2289 | |
2290 | auto HexagonVectorCombine::getConstSplat(Type *Ty, int Val) const |
2291 | -> Constant * { |
2292 | assert(Ty->isVectorTy()); |
2293 | auto VecTy = cast<VectorType>(Val: Ty); |
2294 | Type *ElemTy = VecTy->getElementType(); |
2295 | // Add support for floats if needed. |
2296 | auto *Splat = ConstantVector::getSplat(EC: VecTy->getElementCount(), |
2297 | Elt: ConstantInt::get(Ty: ElemTy, V: Val)); |
2298 | return Splat; |
2299 | } |
2300 | |
2301 | auto HexagonVectorCombine::simplify(Value *V) const -> Value * { |
2302 | if (auto *In = dyn_cast<Instruction>(Val: V)) { |
2303 | SimplifyQuery Q(DL, &TLI, &DT, &AC, In); |
2304 | return simplifyInstruction(I: In, Q); |
2305 | } |
2306 | return nullptr; |
2307 | } |
2308 | |
2309 | // Insert bytes [Start..Start+Length) of Src into Dst at byte Where. |
2310 | auto HexagonVectorCombine::insertb(IRBuilderBase &Builder, Value *Dst, |
2311 | Value *Src, int Start, int Length, |
2312 | int Where) const -> Value * { |
2313 | assert(isByteVecTy(Dst->getType()) && isByteVecTy(Src->getType())); |
2314 | int SrcLen = getSizeOf(Val: Src); |
2315 | int DstLen = getSizeOf(Val: Dst); |
2316 | assert(0 <= Start && Start + Length <= SrcLen); |
2317 | assert(0 <= Where && Where + Length <= DstLen); |
2318 | |
2319 | int P2Len = PowerOf2Ceil(A: SrcLen | DstLen); |
2320 | auto *Undef = UndefValue::get(T: getByteTy()); |
2321 | Value *P2Src = vresize(Builder, Val: Src, NewSize: P2Len, Pad: Undef); |
2322 | Value *P2Dst = vresize(Builder, Val: Dst, NewSize: P2Len, Pad: Undef); |
2323 | |
2324 | SmallVector<int, 256> SMask(P2Len); |
2325 | for (int i = 0; i != P2Len; ++i) { |
2326 | // If i is in [Where, Where+Length), pick Src[Start+(i-Where)]. |
2327 | // Otherwise, pick Dst[i]; |
2328 | SMask[i] = |
2329 | (Where <= i && i < Where + Length) ? P2Len + Start + (i - Where) : i; |
2330 | } |
2331 | |
2332 | Value *P2Insert = Builder.CreateShuffleVector(V1: P2Dst, V2: P2Src, Mask: SMask, Name: "shf" ); |
2333 | return vresize(Builder, Val: P2Insert, NewSize: DstLen, Pad: Undef); |
2334 | } |
2335 | |
2336 | auto HexagonVectorCombine::vlalignb(IRBuilderBase &Builder, Value *Lo, |
2337 | Value *Hi, Value *Amt) const -> Value * { |
2338 | assert(Lo->getType() == Hi->getType() && "Argument type mismatch" ); |
2339 | if (isZero(Val: Amt)) |
2340 | return Hi; |
2341 | int VecLen = getSizeOf(Val: Hi); |
2342 | if (auto IntAmt = getIntValue(Val: Amt)) |
2343 | return getElementRange(Builder, Lo, Hi, Start: VecLen - IntAmt->getSExtValue(), |
2344 | Length: VecLen); |
2345 | |
2346 | if (HST.isTypeForHVX(VecTy: Hi->getType())) { |
2347 | assert(static_cast<unsigned>(VecLen) == HST.getVectorLength() && |
2348 | "Expecting an exact HVX type" ); |
2349 | return createHvxIntrinsic(Builder, IntID: HST.getIntrinsicId(Opc: Hexagon::V6_vlalignb), |
2350 | RetTy: Hi->getType(), Args: {Hi, Lo, Amt}); |
2351 | } |
2352 | |
2353 | if (VecLen == 4) { |
2354 | Value *Pair = concat(Builder, Vecs: {Lo, Hi}); |
2355 | Value *Shift = |
2356 | Builder.CreateLShr(LHS: Builder.CreateShl(LHS: Pair, RHS: Amt, Name: "shl" ), RHS: 32, Name: "lsr" ); |
2357 | Value *Trunc = |
2358 | Builder.CreateTrunc(V: Shift, DestTy: Type::getInt32Ty(C&: F.getContext()), Name: "trn" ); |
2359 | return Builder.CreateBitCast(V: Trunc, DestTy: Hi->getType(), Name: "cst" ); |
2360 | } |
2361 | if (VecLen == 8) { |
2362 | Value *Sub = Builder.CreateSub(LHS: getConstInt(Val: VecLen), RHS: Amt, Name: "sub" ); |
2363 | return vralignb(Builder, Lo, Hi, Amt: Sub); |
2364 | } |
2365 | llvm_unreachable("Unexpected vector length" ); |
2366 | } |
2367 | |
2368 | auto HexagonVectorCombine::vralignb(IRBuilderBase &Builder, Value *Lo, |
2369 | Value *Hi, Value *Amt) const -> Value * { |
2370 | assert(Lo->getType() == Hi->getType() && "Argument type mismatch" ); |
2371 | if (isZero(Val: Amt)) |
2372 | return Lo; |
2373 | int VecLen = getSizeOf(Val: Lo); |
2374 | if (auto IntAmt = getIntValue(Val: Amt)) |
2375 | return getElementRange(Builder, Lo, Hi, Start: IntAmt->getSExtValue(), Length: VecLen); |
2376 | |
2377 | if (HST.isTypeForHVX(VecTy: Lo->getType())) { |
2378 | assert(static_cast<unsigned>(VecLen) == HST.getVectorLength() && |
2379 | "Expecting an exact HVX type" ); |
2380 | return createHvxIntrinsic(Builder, IntID: HST.getIntrinsicId(Opc: Hexagon::V6_valignb), |
2381 | RetTy: Lo->getType(), Args: {Hi, Lo, Amt}); |
2382 | } |
2383 | |
2384 | if (VecLen == 4) { |
2385 | Value *Pair = concat(Builder, Vecs: {Lo, Hi}); |
2386 | Value *Shift = Builder.CreateLShr(LHS: Pair, RHS: Amt, Name: "lsr" ); |
2387 | Value *Trunc = |
2388 | Builder.CreateTrunc(V: Shift, DestTy: Type::getInt32Ty(C&: F.getContext()), Name: "trn" ); |
2389 | return Builder.CreateBitCast(V: Trunc, DestTy: Lo->getType(), Name: "cst" ); |
2390 | } |
2391 | if (VecLen == 8) { |
2392 | Type *Int64Ty = Type::getInt64Ty(C&: F.getContext()); |
2393 | Value *Lo64 = Builder.CreateBitCast(V: Lo, DestTy: Int64Ty, Name: "cst" ); |
2394 | Value *Hi64 = Builder.CreateBitCast(V: Hi, DestTy: Int64Ty, Name: "cst" ); |
2395 | Function *FI = Intrinsic::getDeclaration(M: F.getParent(), |
2396 | id: Intrinsic::hexagon_S2_valignrb); |
2397 | Value *Call = Builder.CreateCall(Callee: FI, Args: {Hi64, Lo64, Amt}, Name: "cup" ); |
2398 | return Builder.CreateBitCast(V: Call, DestTy: Lo->getType(), Name: "cst" ); |
2399 | } |
2400 | llvm_unreachable("Unexpected vector length" ); |
2401 | } |
2402 | |
2403 | // Concatenates a sequence of vectors of the same type. |
2404 | auto HexagonVectorCombine::concat(IRBuilderBase &Builder, |
2405 | ArrayRef<Value *> Vecs) const -> Value * { |
2406 | assert(!Vecs.empty()); |
2407 | SmallVector<int, 256> SMask; |
2408 | std::vector<Value *> Work[2]; |
2409 | int ThisW = 0, OtherW = 1; |
2410 | |
2411 | Work[ThisW].assign(first: Vecs.begin(), last: Vecs.end()); |
2412 | while (Work[ThisW].size() > 1) { |
2413 | auto *Ty = cast<VectorType>(Val: Work[ThisW].front()->getType()); |
2414 | SMask.resize(N: length(Ty) * 2); |
2415 | std::iota(first: SMask.begin(), last: SMask.end(), value: 0); |
2416 | |
2417 | Work[OtherW].clear(); |
2418 | if (Work[ThisW].size() % 2 != 0) |
2419 | Work[ThisW].push_back(x: UndefValue::get(T: Ty)); |
2420 | for (int i = 0, e = Work[ThisW].size(); i < e; i += 2) { |
2421 | Value *Joined = Builder.CreateShuffleVector( |
2422 | V1: Work[ThisW][i], V2: Work[ThisW][i + 1], Mask: SMask, Name: "shf" ); |
2423 | Work[OtherW].push_back(x: Joined); |
2424 | } |
2425 | std::swap(a&: ThisW, b&: OtherW); |
2426 | } |
2427 | |
2428 | // Since there may have been some undefs appended to make shuffle operands |
2429 | // have the same type, perform the last shuffle to only pick the original |
2430 | // elements. |
2431 | SMask.resize(N: Vecs.size() * length(Ty: Vecs.front()->getType())); |
2432 | std::iota(first: SMask.begin(), last: SMask.end(), value: 0); |
2433 | Value *Total = Work[ThisW].front(); |
2434 | return Builder.CreateShuffleVector(V: Total, Mask: SMask, Name: "shf" ); |
2435 | } |
2436 | |
2437 | auto HexagonVectorCombine::vresize(IRBuilderBase &Builder, Value *Val, |
2438 | int NewSize, Value *Pad) const -> Value * { |
2439 | assert(isa<VectorType>(Val->getType())); |
2440 | auto *ValTy = cast<VectorType>(Val: Val->getType()); |
2441 | assert(ValTy->getElementType() == Pad->getType()); |
2442 | |
2443 | int CurSize = length(Ty: ValTy); |
2444 | if (CurSize == NewSize) |
2445 | return Val; |
2446 | // Truncate? |
2447 | if (CurSize > NewSize) |
2448 | return getElementRange(Builder, Lo: Val, /*Ignored*/ Hi: Val, Start: 0, Length: NewSize); |
2449 | // Extend. |
2450 | SmallVector<int, 128> SMask(NewSize); |
2451 | std::iota(first: SMask.begin(), last: SMask.begin() + CurSize, value: 0); |
2452 | std::fill(SMask.begin() + CurSize, SMask.end(), CurSize); |
2453 | Value *PadVec = Builder.CreateVectorSplat(NumElts: CurSize, V: Pad, Name: "spt" ); |
2454 | return Builder.CreateShuffleVector(V1: Val, V2: PadVec, Mask: SMask, Name: "shf" ); |
2455 | } |
2456 | |
2457 | auto HexagonVectorCombine::rescale(IRBuilderBase &Builder, Value *Mask, |
2458 | Type *FromTy, Type *ToTy) const -> Value * { |
2459 | // Mask is a vector <N x i1>, where each element corresponds to an |
2460 | // element of FromTy. Remap it so that each element will correspond |
2461 | // to an element of ToTy. |
2462 | assert(isa<VectorType>(Mask->getType())); |
2463 | |
2464 | Type *FromSTy = FromTy->getScalarType(); |
2465 | Type *ToSTy = ToTy->getScalarType(); |
2466 | if (FromSTy == ToSTy) |
2467 | return Mask; |
2468 | |
2469 | int FromSize = getSizeOf(Ty: FromSTy); |
2470 | int ToSize = getSizeOf(Ty: ToSTy); |
2471 | assert(FromSize % ToSize == 0 || ToSize % FromSize == 0); |
2472 | |
2473 | auto *MaskTy = cast<VectorType>(Val: Mask->getType()); |
2474 | int FromCount = length(Ty: MaskTy); |
2475 | int ToCount = (FromCount * FromSize) / ToSize; |
2476 | assert((FromCount * FromSize) % ToSize == 0); |
2477 | |
2478 | auto *FromITy = getIntTy(Width: FromSize * 8); |
2479 | auto *ToITy = getIntTy(Width: ToSize * 8); |
2480 | |
2481 | // Mask <N x i1> -> sext to <N x FromTy> -> bitcast to <M x ToTy> -> |
2482 | // -> trunc to <M x i1>. |
2483 | Value *Ext = Builder.CreateSExt( |
2484 | V: Mask, DestTy: VectorType::get(ElementType: FromITy, NumElements: FromCount, /*Scalable=*/false), Name: "sxt" ); |
2485 | Value *Cast = Builder.CreateBitCast( |
2486 | V: Ext, DestTy: VectorType::get(ElementType: ToITy, NumElements: ToCount, /*Scalable=*/false), Name: "cst" ); |
2487 | return Builder.CreateTrunc( |
2488 | V: Cast, DestTy: VectorType::get(ElementType: getBoolTy(), NumElements: ToCount, /*Scalable=*/false), Name: "trn" ); |
2489 | } |
2490 | |
2491 | // Bitcast to bytes, and return least significant bits. |
2492 | auto HexagonVectorCombine::vlsb(IRBuilderBase &Builder, Value *Val) const |
2493 | -> Value * { |
2494 | Type *ScalarTy = Val->getType()->getScalarType(); |
2495 | if (ScalarTy == getBoolTy()) |
2496 | return Val; |
2497 | |
2498 | Value *Bytes = vbytes(Builder, Val); |
2499 | if (auto *VecTy = dyn_cast<VectorType>(Val: Bytes->getType())) |
2500 | return Builder.CreateTrunc(V: Bytes, DestTy: getBoolTy(ElemCount: getSizeOf(Ty: VecTy)), Name: "trn" ); |
2501 | // If Bytes is a scalar (i.e. Val was a scalar byte), return i1, not |
2502 | // <1 x i1>. |
2503 | return Builder.CreateTrunc(V: Bytes, DestTy: getBoolTy(), Name: "trn" ); |
2504 | } |
2505 | |
2506 | // Bitcast to bytes for non-bool. For bool, convert i1 -> i8. |
2507 | auto HexagonVectorCombine::vbytes(IRBuilderBase &Builder, Value *Val) const |
2508 | -> Value * { |
2509 | Type *ScalarTy = Val->getType()->getScalarType(); |
2510 | if (ScalarTy == getByteTy()) |
2511 | return Val; |
2512 | |
2513 | if (ScalarTy != getBoolTy()) |
2514 | return Builder.CreateBitCast(V: Val, DestTy: getByteTy(ElemCount: getSizeOf(Val)), Name: "cst" ); |
2515 | // For bool, return a sext from i1 to i8. |
2516 | if (auto *VecTy = dyn_cast<VectorType>(Val: Val->getType())) |
2517 | return Builder.CreateSExt(V: Val, DestTy: VectorType::get(ElementType: getByteTy(), Other: VecTy), Name: "sxt" ); |
2518 | return Builder.CreateSExt(V: Val, DestTy: getByteTy(), Name: "sxt" ); |
2519 | } |
2520 | |
2521 | auto HexagonVectorCombine::subvector(IRBuilderBase &Builder, Value *Val, |
2522 | unsigned Start, unsigned Length) const |
2523 | -> Value * { |
2524 | assert(Start + Length <= length(Val)); |
2525 | return getElementRange(Builder, Lo: Val, /*Ignored*/ Hi: Val, Start, Length); |
2526 | } |
2527 | |
2528 | auto HexagonVectorCombine::sublo(IRBuilderBase &Builder, Value *Val) const |
2529 | -> Value * { |
2530 | size_t Len = length(Val); |
2531 | assert(Len % 2 == 0 && "Length should be even" ); |
2532 | return subvector(Builder, Val, Start: 0, Length: Len / 2); |
2533 | } |
2534 | |
2535 | auto HexagonVectorCombine::subhi(IRBuilderBase &Builder, Value *Val) const |
2536 | -> Value * { |
2537 | size_t Len = length(Val); |
2538 | assert(Len % 2 == 0 && "Length should be even" ); |
2539 | return subvector(Builder, Val, Start: Len / 2, Length: Len / 2); |
2540 | } |
2541 | |
2542 | auto HexagonVectorCombine::vdeal(IRBuilderBase &Builder, Value *Val0, |
2543 | Value *Val1) const -> Value * { |
2544 | assert(Val0->getType() == Val1->getType()); |
2545 | int Len = length(Val: Val0); |
2546 | SmallVector<int, 128> Mask(2 * Len); |
2547 | |
2548 | for (int i = 0; i != Len; ++i) { |
2549 | Mask[i] = 2 * i; // Even |
2550 | Mask[i + Len] = 2 * i + 1; // Odd |
2551 | } |
2552 | return Builder.CreateShuffleVector(V1: Val0, V2: Val1, Mask, Name: "shf" ); |
2553 | } |
2554 | |
2555 | auto HexagonVectorCombine::vshuff(IRBuilderBase &Builder, Value *Val0, |
2556 | Value *Val1) const -> Value * { // |
2557 | assert(Val0->getType() == Val1->getType()); |
2558 | int Len = length(Val: Val0); |
2559 | SmallVector<int, 128> Mask(2 * Len); |
2560 | |
2561 | for (int i = 0; i != Len; ++i) { |
2562 | Mask[2 * i + 0] = i; // Val0 |
2563 | Mask[2 * i + 1] = i + Len; // Val1 |
2564 | } |
2565 | return Builder.CreateShuffleVector(V1: Val0, V2: Val1, Mask, Name: "shf" ); |
2566 | } |
2567 | |
2568 | auto HexagonVectorCombine::createHvxIntrinsic(IRBuilderBase &Builder, |
2569 | Intrinsic::ID IntID, Type *RetTy, |
2570 | ArrayRef<Value *> Args, |
2571 | ArrayRef<Type *> ArgTys, |
2572 | ArrayRef<Value *> MDSources) const |
2573 | -> Value * { |
2574 | auto getCast = [&](IRBuilderBase &Builder, Value *Val, |
2575 | Type *DestTy) -> Value * { |
2576 | Type *SrcTy = Val->getType(); |
2577 | if (SrcTy == DestTy) |
2578 | return Val; |
2579 | |
2580 | // Non-HVX type. It should be a scalar, and it should already have |
2581 | // a valid type. |
2582 | assert(HST.isTypeForHVX(SrcTy, /*IncludeBool=*/true)); |
2583 | |
2584 | Type *BoolTy = Type::getInt1Ty(C&: F.getContext()); |
2585 | if (cast<VectorType>(Val: SrcTy)->getElementType() != BoolTy) |
2586 | return Builder.CreateBitCast(V: Val, DestTy, Name: "cst" ); |
2587 | |
2588 | // Predicate HVX vector. |
2589 | unsigned HwLen = HST.getVectorLength(); |
2590 | Intrinsic::ID TC = HwLen == 64 ? Intrinsic::hexagon_V6_pred_typecast |
2591 | : Intrinsic::hexagon_V6_pred_typecast_128B; |
2592 | Function *FI = |
2593 | Intrinsic::getDeclaration(M: F.getParent(), id: TC, Tys: {DestTy, Val->getType()}); |
2594 | return Builder.CreateCall(Callee: FI, Args: {Val}, Name: "cup" ); |
2595 | }; |
2596 | |
2597 | Function *IntrFn = Intrinsic::getDeclaration(M: F.getParent(), id: IntID, Tys: ArgTys); |
2598 | FunctionType *IntrTy = IntrFn->getFunctionType(); |
2599 | |
2600 | SmallVector<Value *, 4> IntrArgs; |
2601 | for (int i = 0, e = Args.size(); i != e; ++i) { |
2602 | Value *A = Args[i]; |
2603 | Type *T = IntrTy->getParamType(i); |
2604 | if (A->getType() != T) { |
2605 | IntrArgs.push_back(Elt: getCast(Builder, A, T)); |
2606 | } else { |
2607 | IntrArgs.push_back(Elt: A); |
2608 | } |
2609 | } |
2610 | StringRef MaybeName = !IntrTy->getReturnType()->isVoidTy() ? "cup" : "" ; |
2611 | CallInst *Call = Builder.CreateCall(Callee: IntrFn, Args: IntrArgs, Name: MaybeName); |
2612 | |
2613 | MemoryEffects ME = Call->getAttributes().getMemoryEffects(); |
2614 | if (!ME.doesNotAccessMemory() && !ME.onlyAccessesInaccessibleMem()) |
2615 | propagateMetadata(I: Call, VL: MDSources); |
2616 | |
2617 | Type *CallTy = Call->getType(); |
2618 | if (RetTy == nullptr || CallTy == RetTy) |
2619 | return Call; |
2620 | // Scalar types should have RetTy matching the call return type. |
2621 | assert(HST.isTypeForHVX(CallTy, /*IncludeBool=*/true)); |
2622 | return getCast(Builder, Call, RetTy); |
2623 | } |
2624 | |
2625 | auto HexagonVectorCombine::splitVectorElements(IRBuilderBase &Builder, |
2626 | Value *Vec, |
2627 | unsigned ToWidth) const |
2628 | -> SmallVector<Value *> { |
2629 | // Break a vector of wide elements into a series of vectors with narrow |
2630 | // elements: |
2631 | // (...c0:b0:a0, ...c1:b1:a1, ...c2:b2:a2, ...) |
2632 | // --> |
2633 | // (a0, a1, a2, ...) // lowest "ToWidth" bits |
2634 | // (b0, b1, b2, ...) // the next lowest... |
2635 | // (c0, c1, c2, ...) // ... |
2636 | // ... |
2637 | // |
2638 | // The number of elements in each resulting vector is the same as |
2639 | // in the original vector. |
2640 | |
2641 | auto *VecTy = cast<VectorType>(Val: Vec->getType()); |
2642 | assert(VecTy->getElementType()->isIntegerTy()); |
2643 | unsigned FromWidth = VecTy->getScalarSizeInBits(); |
2644 | assert(isPowerOf2_32(ToWidth) && isPowerOf2_32(FromWidth)); |
2645 | assert(ToWidth <= FromWidth && "Breaking up into wider elements?" ); |
2646 | unsigned NumResults = FromWidth / ToWidth; |
2647 | |
2648 | SmallVector<Value *> Results(NumResults); |
2649 | Results[0] = Vec; |
2650 | unsigned Length = length(Ty: VecTy); |
2651 | |
2652 | // Do it by splitting in half, since those operations correspond to deal |
2653 | // instructions. |
2654 | auto splitInHalf = [&](unsigned Begin, unsigned End, auto splitFunc) -> void { |
2655 | // Take V = Results[Begin], split it in L, H. |
2656 | // Store Results[Begin] = L, Results[(Begin+End)/2] = H |
2657 | // Call itself recursively split(Begin, Half), split(Half+1, End) |
2658 | if (Begin + 1 == End) |
2659 | return; |
2660 | |
2661 | Value *Val = Results[Begin]; |
2662 | unsigned Width = Val->getType()->getScalarSizeInBits(); |
2663 | |
2664 | auto *VTy = VectorType::get(ElementType: getIntTy(Width: Width / 2), NumElements: 2 * Length, Scalable: false); |
2665 | Value *VVal = Builder.CreateBitCast(V: Val, DestTy: VTy, Name: "cst" ); |
2666 | |
2667 | Value *Res = vdeal(Builder, Val0: sublo(Builder, Val: VVal), Val1: subhi(Builder, Val: VVal)); |
2668 | |
2669 | unsigned Half = (Begin + End) / 2; |
2670 | Results[Begin] = sublo(Builder, Val: Res); |
2671 | Results[Half] = subhi(Builder, Val: Res); |
2672 | |
2673 | splitFunc(Begin, Half, splitFunc); |
2674 | splitFunc(Half, End, splitFunc); |
2675 | }; |
2676 | |
2677 | splitInHalf(0, NumResults, splitInHalf); |
2678 | return Results; |
2679 | } |
2680 | |
2681 | auto HexagonVectorCombine::joinVectorElements(IRBuilderBase &Builder, |
2682 | ArrayRef<Value *> Values, |
2683 | VectorType *ToType) const |
2684 | -> Value * { |
2685 | assert(ToType->getElementType()->isIntegerTy()); |
2686 | |
2687 | // If the list of values does not have power-of-2 elements, append copies |
2688 | // of the sign bit to it, to make the size be 2^n. |
2689 | // The reason for this is that the values will be joined in pairs, because |
2690 | // otherwise the shuffles will result in convoluted code. With pairwise |
2691 | // joins, the shuffles will hopefully be folded into a perfect shuffle. |
2692 | // The output will need to be sign-extended to a type with element width |
2693 | // being a power-of-2 anyways. |
2694 | SmallVector<Value *> Inputs(Values.begin(), Values.end()); |
2695 | |
2696 | unsigned ToWidth = ToType->getScalarSizeInBits(); |
2697 | unsigned Width = Inputs.front()->getType()->getScalarSizeInBits(); |
2698 | assert(Width <= ToWidth); |
2699 | assert(isPowerOf2_32(Width) && isPowerOf2_32(ToWidth)); |
2700 | unsigned Length = length(Ty: Inputs.front()->getType()); |
2701 | |
2702 | unsigned NeedInputs = ToWidth / Width; |
2703 | if (Inputs.size() != NeedInputs) { |
2704 | // Having too many inputs is ok: drop the high bits (usual wrap-around). |
2705 | // If there are too few, fill them with the sign bit. |
2706 | Value *Last = Inputs.back(); |
2707 | Value *Sign = Builder.CreateAShr( |
2708 | LHS: Last, RHS: getConstSplat(Ty: Last->getType(), Val: Width - 1), Name: "asr" ); |
2709 | Inputs.resize(N: NeedInputs, NV: Sign); |
2710 | } |
2711 | |
2712 | while (Inputs.size() > 1) { |
2713 | Width *= 2; |
2714 | auto *VTy = VectorType::get(ElementType: getIntTy(Width), NumElements: Length, Scalable: false); |
2715 | for (int i = 0, e = Inputs.size(); i < e; i += 2) { |
2716 | Value *Res = vshuff(Builder, Val0: Inputs[i], Val1: Inputs[i + 1]); |
2717 | Inputs[i / 2] = Builder.CreateBitCast(V: Res, DestTy: VTy, Name: "cst" ); |
2718 | } |
2719 | Inputs.resize(N: Inputs.size() / 2); |
2720 | } |
2721 | |
2722 | assert(Inputs.front()->getType() == ToType); |
2723 | return Inputs.front(); |
2724 | } |
2725 | |
2726 | auto HexagonVectorCombine::calculatePointerDifference(Value *Ptr0, |
2727 | Value *Ptr1) const |
2728 | -> std::optional<int> { |
2729 | // Try SCEV first. |
2730 | const SCEV *Scev0 = SE.getSCEV(V: Ptr0); |
2731 | const SCEV *Scev1 = SE.getSCEV(V: Ptr1); |
2732 | const SCEV *ScevDiff = SE.getMinusSCEV(LHS: Scev0, RHS: Scev1); |
2733 | if (auto *Const = dyn_cast<SCEVConstant>(Val: ScevDiff)) { |
2734 | APInt V = Const->getAPInt(); |
2735 | if (V.isSignedIntN(N: 8 * sizeof(int))) |
2736 | return static_cast<int>(V.getSExtValue()); |
2737 | } |
2738 | |
2739 | struct Builder : IRBuilder<> { |
2740 | Builder(BasicBlock *B) : IRBuilder<>(B->getTerminator()) {} |
2741 | ~Builder() { |
2742 | for (Instruction *I : llvm::reverse(C&: ToErase)) |
2743 | I->eraseFromParent(); |
2744 | } |
2745 | SmallVector<Instruction *, 8> ToErase; |
2746 | }; |
2747 | |
2748 | #define CallBuilder(B, F) \ |
2749 | [&](auto &B_) { \ |
2750 | Value *V = B_.F; \ |
2751 | if (auto *I = dyn_cast<Instruction>(V)) \ |
2752 | B_.ToErase.push_back(I); \ |
2753 | return V; \ |
2754 | }(B) |
2755 | |
2756 | auto Simplify = [this](Value *V) { |
2757 | if (Value *S = simplify(V)) |
2758 | return S; |
2759 | return V; |
2760 | }; |
2761 | |
2762 | auto StripBitCast = [](Value *V) { |
2763 | while (auto *C = dyn_cast<BitCastInst>(Val: V)) |
2764 | V = C->getOperand(i_nocapture: 0); |
2765 | return V; |
2766 | }; |
2767 | |
2768 | Ptr0 = StripBitCast(Ptr0); |
2769 | Ptr1 = StripBitCast(Ptr1); |
2770 | if (!isa<GetElementPtrInst>(Val: Ptr0) || !isa<GetElementPtrInst>(Val: Ptr1)) |
2771 | return std::nullopt; |
2772 | |
2773 | auto *Gep0 = cast<GetElementPtrInst>(Val: Ptr0); |
2774 | auto *Gep1 = cast<GetElementPtrInst>(Val: Ptr1); |
2775 | if (Gep0->getPointerOperand() != Gep1->getPointerOperand()) |
2776 | return std::nullopt; |
2777 | if (Gep0->getSourceElementType() != Gep1->getSourceElementType()) |
2778 | return std::nullopt; |
2779 | |
2780 | Builder B(Gep0->getParent()); |
2781 | int Scale = getSizeOf(Ty: Gep0->getSourceElementType(), Kind: Alloc); |
2782 | |
2783 | // FIXME: for now only check GEPs with a single index. |
2784 | if (Gep0->getNumOperands() != 2 || Gep1->getNumOperands() != 2) |
2785 | return std::nullopt; |
2786 | |
2787 | Value *Idx0 = Gep0->getOperand(i_nocapture: 1); |
2788 | Value *Idx1 = Gep1->getOperand(i_nocapture: 1); |
2789 | |
2790 | // First, try to simplify the subtraction directly. |
2791 | if (auto *Diff = dyn_cast<ConstantInt>( |
2792 | Val: Simplify(CallBuilder(B, CreateSub(Idx0, Idx1))))) |
2793 | return Diff->getSExtValue() * Scale; |
2794 | |
2795 | KnownBits Known0 = getKnownBits(V: Idx0, CtxI: Gep0); |
2796 | KnownBits Known1 = getKnownBits(V: Idx1, CtxI: Gep1); |
2797 | APInt Unknown = ~(Known0.Zero | Known0.One) | ~(Known1.Zero | Known1.One); |
2798 | if (Unknown.isAllOnes()) |
2799 | return std::nullopt; |
2800 | |
2801 | Value *MaskU = ConstantInt::get(Ty: Idx0->getType(), V: Unknown); |
2802 | Value *AndU0 = Simplify(CallBuilder(B, CreateAnd(Idx0, MaskU))); |
2803 | Value *AndU1 = Simplify(CallBuilder(B, CreateAnd(Idx1, MaskU))); |
2804 | Value *SubU = Simplify(CallBuilder(B, CreateSub(AndU0, AndU1))); |
2805 | int Diff0 = 0; |
2806 | if (auto *C = dyn_cast<ConstantInt>(Val: SubU)) { |
2807 | Diff0 = C->getSExtValue(); |
2808 | } else { |
2809 | return std::nullopt; |
2810 | } |
2811 | |
2812 | Value *MaskK = ConstantInt::get(Ty: MaskU->getType(), V: ~Unknown); |
2813 | Value *AndK0 = Simplify(CallBuilder(B, CreateAnd(Idx0, MaskK))); |
2814 | Value *AndK1 = Simplify(CallBuilder(B, CreateAnd(Idx1, MaskK))); |
2815 | Value *SubK = Simplify(CallBuilder(B, CreateSub(AndK0, AndK1))); |
2816 | int Diff1 = 0; |
2817 | if (auto *C = dyn_cast<ConstantInt>(Val: SubK)) { |
2818 | Diff1 = C->getSExtValue(); |
2819 | } else { |
2820 | return std::nullopt; |
2821 | } |
2822 | |
2823 | return (Diff0 + Diff1) * Scale; |
2824 | |
2825 | #undef CallBuilder |
2826 | } |
2827 | |
2828 | auto HexagonVectorCombine::getNumSignificantBits(const Value *V, |
2829 | const Instruction *CtxI) const |
2830 | -> unsigned { |
2831 | return ComputeMaxSignificantBits(Op: V, DL, /*Depth=*/0, AC: &AC, CxtI: CtxI, DT: &DT); |
2832 | } |
2833 | |
2834 | auto HexagonVectorCombine::getKnownBits(const Value *V, |
2835 | const Instruction *CtxI) const |
2836 | -> KnownBits { |
2837 | return computeKnownBits(V, DL, /*Depth=*/0, AC: &AC, CxtI: CtxI, DT: &DT); |
2838 | } |
2839 | |
2840 | auto HexagonVectorCombine::isSafeToClone(const Instruction &In) const -> bool { |
2841 | if (In.mayHaveSideEffects() || In.isAtomic() || In.isVolatile() || |
2842 | In.isFenceLike() || In.mayReadOrWriteMemory()) { |
2843 | return false; |
2844 | } |
2845 | if (isa<CallBase>(Val: In) || isa<AllocaInst>(Val: In)) |
2846 | return false; |
2847 | return true; |
2848 | } |
2849 | |
2850 | template <typename T> |
2851 | auto HexagonVectorCombine::isSafeToMoveBeforeInBB(const Instruction &In, |
2852 | BasicBlock::const_iterator To, |
2853 | const T &IgnoreInsts) const |
2854 | -> bool { |
2855 | auto getLocOrNone = |
2856 | [this](const Instruction &I) -> std::optional<MemoryLocation> { |
2857 | if (const auto *II = dyn_cast<IntrinsicInst>(Val: &I)) { |
2858 | switch (II->getIntrinsicID()) { |
2859 | case Intrinsic::masked_load: |
2860 | return MemoryLocation::getForArgument(Call: II, ArgIdx: 0, TLI); |
2861 | case Intrinsic::masked_store: |
2862 | return MemoryLocation::getForArgument(Call: II, ArgIdx: 1, TLI); |
2863 | } |
2864 | } |
2865 | return MemoryLocation::getOrNone(Inst: &I); |
2866 | }; |
2867 | |
2868 | // The source and the destination must be in the same basic block. |
2869 | const BasicBlock &Block = *In.getParent(); |
2870 | assert(Block.begin() == To || Block.end() == To || To->getParent() == &Block); |
2871 | // No PHIs. |
2872 | if (isa<PHINode>(Val: In) || (To != Block.end() && isa<PHINode>(Val: *To))) |
2873 | return false; |
2874 | |
2875 | if (!mayHaveNonDefUseDependency(I: In)) |
2876 | return true; |
2877 | bool MayWrite = In.mayWriteToMemory(); |
2878 | auto MaybeLoc = getLocOrNone(In); |
2879 | |
2880 | auto From = In.getIterator(); |
2881 | if (From == To) |
2882 | return true; |
2883 | bool MoveUp = (To != Block.end() && To->comesBefore(Other: &In)); |
2884 | auto Range = |
2885 | MoveUp ? std::make_pair(x&: To, y&: From) : std::make_pair(x: std::next(x: From), y&: To); |
2886 | for (auto It = Range.first; It != Range.second; ++It) { |
2887 | const Instruction &I = *It; |
2888 | if (llvm::is_contained(IgnoreInsts, &I)) |
2889 | continue; |
2890 | // assume intrinsic can be ignored |
2891 | if (auto *II = dyn_cast<IntrinsicInst>(Val: &I)) { |
2892 | if (II->getIntrinsicID() == Intrinsic::assume) |
2893 | continue; |
2894 | } |
2895 | // Parts based on isSafeToMoveBefore from CoveMoverUtils.cpp. |
2896 | if (I.mayThrow()) |
2897 | return false; |
2898 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
2899 | if (!CB->hasFnAttr(Kind: Attribute::WillReturn)) |
2900 | return false; |
2901 | if (!CB->hasFnAttr(Kind: Attribute::NoSync)) |
2902 | return false; |
2903 | } |
2904 | if (I.mayReadOrWriteMemory()) { |
2905 | auto MaybeLocI = getLocOrNone(I); |
2906 | if (MayWrite || I.mayWriteToMemory()) { |
2907 | if (!MaybeLoc || !MaybeLocI) |
2908 | return false; |
2909 | if (!AA.isNoAlias(*MaybeLoc, *MaybeLocI)) |
2910 | return false; |
2911 | } |
2912 | } |
2913 | } |
2914 | return true; |
2915 | } |
2916 | |
2917 | auto HexagonVectorCombine::isByteVecTy(Type *Ty) const -> bool { |
2918 | if (auto *VecTy = dyn_cast<VectorType>(Val: Ty)) |
2919 | return VecTy->getElementType() == getByteTy(); |
2920 | return false; |
2921 | } |
2922 | |
2923 | auto HexagonVectorCombine::getElementRange(IRBuilderBase &Builder, Value *Lo, |
2924 | Value *Hi, int Start, |
2925 | int Length) const -> Value * { |
2926 | assert(0 <= Start && size_t(Start + Length) < length(Lo) + length(Hi)); |
2927 | SmallVector<int, 128> SMask(Length); |
2928 | std::iota(first: SMask.begin(), last: SMask.end(), value: Start); |
2929 | return Builder.CreateShuffleVector(V1: Lo, V2: Hi, Mask: SMask, Name: "shf" ); |
2930 | } |
2931 | |
2932 | // Pass management. |
2933 | |
2934 | namespace llvm { |
2935 | void initializeHexagonVectorCombineLegacyPass(PassRegistry &); |
2936 | FunctionPass *createHexagonVectorCombineLegacyPass(); |
2937 | } // namespace llvm |
2938 | |
2939 | namespace { |
2940 | class HexagonVectorCombineLegacy : public FunctionPass { |
2941 | public: |
2942 | static char ID; |
2943 | |
2944 | HexagonVectorCombineLegacy() : FunctionPass(ID) {} |
2945 | |
2946 | StringRef getPassName() const override { return "Hexagon Vector Combine" ; } |
2947 | |
2948 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
2949 | AU.setPreservesCFG(); |
2950 | AU.addRequired<AAResultsWrapperPass>(); |
2951 | AU.addRequired<AssumptionCacheTracker>(); |
2952 | AU.addRequired<DominatorTreeWrapperPass>(); |
2953 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
2954 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
2955 | AU.addRequired<TargetPassConfig>(); |
2956 | FunctionPass::getAnalysisUsage(AU); |
2957 | } |
2958 | |
2959 | bool runOnFunction(Function &F) override { |
2960 | if (skipFunction(F)) |
2961 | return false; |
2962 | AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
2963 | AssumptionCache &AC = |
2964 | getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
2965 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
2966 | ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
2967 | TargetLibraryInfo &TLI = |
2968 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
2969 | auto &TM = getAnalysis<TargetPassConfig>().getTM<HexagonTargetMachine>(); |
2970 | HexagonVectorCombine HVC(F, AA, AC, DT, SE, TLI, TM); |
2971 | return HVC.run(); |
2972 | } |
2973 | }; |
2974 | } // namespace |
2975 | |
2976 | char HexagonVectorCombineLegacy::ID = 0; |
2977 | |
2978 | INITIALIZE_PASS_BEGIN(HexagonVectorCombineLegacy, DEBUG_TYPE, |
2979 | "Hexagon Vector Combine" , false, false) |
2980 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
2981 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
2982 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
2983 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) |
2984 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
2985 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
2986 | INITIALIZE_PASS_END(HexagonVectorCombineLegacy, DEBUG_TYPE, |
2987 | "Hexagon Vector Combine" , false, false) |
2988 | |
2989 | FunctionPass *llvm::createHexagonVectorCombineLegacyPass() { |
2990 | return new HexagonVectorCombineLegacy(); |
2991 | } |
2992 | |