1 //===-- X86ShuffleDecode.cpp - X86 shuffle decode logic -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Define several functions to decode x86 specific shuffle semantics into a
11 // generic vector mask.
13 //===----------------------------------------------------------------------===//
15 #include "X86ShuffleDecode.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/InstrTypes.h"
20 #include "llvm/CodeGen/MachineValueType.h"
22 //===----------------------------------------------------------------------===//
23 // Vector Mask Decoding
24 //===----------------------------------------------------------------------===//
28 void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
29 // Defaults the copying the dest value.
30 ShuffleMask.push_back(0);
31 ShuffleMask.push_back(1);
32 ShuffleMask.push_back(2);
33 ShuffleMask.push_back(3);
35 // Decode the immediate.
36 unsigned ZMask = Imm & 15;
37 unsigned CountD = (Imm >> 4) & 3;
38 unsigned CountS = (Imm >> 6) & 3;
40 // CountS selects which input element to use.
41 unsigned InVal = 4+CountS;
42 // CountD specifies which element of destination to update.
43 ShuffleMask[CountD] = InVal;
44 // ZMask zaps values, potentially overriding the CountD elt.
45 if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero;
46 if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero;
47 if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero;
48 if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero;
52 void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
53 for (unsigned i = NElts/2; i != NElts; ++i)
54 ShuffleMask.push_back(NElts+i);
56 for (unsigned i = NElts/2; i != NElts; ++i)
57 ShuffleMask.push_back(i);
61 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
62 for (unsigned i = 0; i != NElts/2; ++i)
63 ShuffleMask.push_back(i);
65 for (unsigned i = 0; i != NElts/2; ++i)
66 ShuffleMask.push_back(NElts+i);
69 void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
70 unsigned NumElts = VT.getVectorNumElements();
71 for (int i = 0, e = NumElts / 2; i < e; ++i) {
72 ShuffleMask.push_back(2 * i);
73 ShuffleMask.push_back(2 * i);
77 void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
78 unsigned NumElts = VT.getVectorNumElements();
79 for (int i = 0, e = NumElts / 2; i < e; ++i) {
80 ShuffleMask.push_back(2 * i + 1);
81 ShuffleMask.push_back(2 * i + 1);
85 void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
86 unsigned VectorSizeInBits = VT.getSizeInBits();
87 unsigned NumElts = VectorSizeInBits / 8;
88 unsigned NumLanes = VectorSizeInBits / 128;
89 unsigned NumLaneElts = NumElts / NumLanes;
91 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
92 for (unsigned i = 0; i < NumLaneElts; ++i) {
93 int M = SM_SentinelZero;
94 if (i >= Imm) M = i - Imm + l;
95 ShuffleMask.push_back(M);
99 void DecodePSRLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
100 unsigned VectorSizeInBits = VT.getSizeInBits();
101 unsigned NumElts = VectorSizeInBits / 8;
102 unsigned NumLanes = VectorSizeInBits / 128;
103 unsigned NumLaneElts = NumElts / NumLanes;
105 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
106 for (unsigned i = 0; i < NumLaneElts; ++i) {
107 unsigned Base = i + Imm;
109 if (Base >= NumLaneElts) M = SM_SentinelZero;
110 ShuffleMask.push_back(M);
114 void DecodePALIGNRMask(MVT VT, unsigned Imm,
115 SmallVectorImpl<int> &ShuffleMask) {
116 unsigned NumElts = VT.getVectorNumElements();
117 unsigned Offset = Imm * (VT.getVectorElementType().getSizeInBits() / 8);
119 unsigned NumLanes = VT.getSizeInBits() / 128;
120 unsigned NumLaneElts = NumElts / NumLanes;
122 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
123 for (unsigned i = 0; i != NumLaneElts; ++i) {
124 unsigned Base = i + Offset;
125 // if i+offset is out of this lane then we actually need the other source
126 if (Base >= NumLaneElts) Base += NumElts - NumLaneElts;
127 ShuffleMask.push_back(Base + l);
132 /// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*.
133 /// VT indicates the type of the vector allowing it to handle different
134 /// datatypes and vector widths.
135 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
136 unsigned NumElts = VT.getVectorNumElements();
138 unsigned NumLanes = VT.getSizeInBits() / 128;
139 unsigned NumLaneElts = NumElts / NumLanes;
141 unsigned NewImm = Imm;
142 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
143 for (unsigned i = 0; i != NumLaneElts; ++i) {
144 ShuffleMask.push_back(NewImm % NumLaneElts + l);
145 NewImm /= NumLaneElts;
147 if (NumLaneElts == 4) NewImm = Imm; // reload imm
151 void DecodePSHUFHWMask(MVT VT, unsigned Imm,
152 SmallVectorImpl<int> &ShuffleMask) {
153 unsigned NumElts = VT.getVectorNumElements();
155 for (unsigned l = 0; l != NumElts; l += 8) {
156 unsigned NewImm = Imm;
157 for (unsigned i = 0, e = 4; i != e; ++i) {
158 ShuffleMask.push_back(l + i);
160 for (unsigned i = 4, e = 8; i != e; ++i) {
161 ShuffleMask.push_back(l + 4 + (NewImm & 3));
167 void DecodePSHUFLWMask(MVT VT, unsigned Imm,
168 SmallVectorImpl<int> &ShuffleMask) {
169 unsigned NumElts = VT.getVectorNumElements();
171 for (unsigned l = 0; l != NumElts; l += 8) {
172 unsigned NewImm = Imm;
173 for (unsigned i = 0, e = 4; i != e; ++i) {
174 ShuffleMask.push_back(l + (NewImm & 3));
177 for (unsigned i = 4, e = 8; i != e; ++i) {
178 ShuffleMask.push_back(l + i);
183 /// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates
184 /// the type of the vector allowing it to handle different datatypes and vector
186 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
187 unsigned NumElts = VT.getVectorNumElements();
189 unsigned NumLanes = VT.getSizeInBits() / 128;
190 unsigned NumLaneElts = NumElts / NumLanes;
192 unsigned NewImm = Imm;
193 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
194 // each half of a lane comes from different source
195 for (unsigned s = 0; s != NumElts*2; s += NumElts) {
196 for (unsigned i = 0; i != NumLaneElts/2; ++i) {
197 ShuffleMask.push_back(NewImm % NumLaneElts + s + l);
198 NewImm /= NumLaneElts;
201 if (NumLaneElts == 4) NewImm = Imm; // reload imm
205 /// DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd
206 /// and punpckh*. VT indicates the type of the vector allowing it to handle
207 /// different datatypes and vector widths.
208 void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
209 unsigned NumElts = VT.getVectorNumElements();
211 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
212 // independently on 128-bit lanes.
213 unsigned NumLanes = VT.getSizeInBits() / 128;
214 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
215 unsigned NumLaneElts = NumElts / NumLanes;
217 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
218 for (unsigned i = l + NumLaneElts/2, e = l + NumLaneElts; i != e; ++i) {
219 ShuffleMask.push_back(i); // Reads from dest/src1
220 ShuffleMask.push_back(i+NumElts); // Reads from src/src2
225 /// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
226 /// and punpckl*. VT indicates the type of the vector allowing it to handle
227 /// different datatypes and vector widths.
228 void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
229 unsigned NumElts = VT.getVectorNumElements();
231 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
232 // independently on 128-bit lanes.
233 unsigned NumLanes = VT.getSizeInBits() / 128;
234 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
235 unsigned NumLaneElts = NumElts / NumLanes;
237 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
238 for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) {
239 ShuffleMask.push_back(i); // Reads from dest/src1
240 ShuffleMask.push_back(i+NumElts); // Reads from src/src2
245 void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
246 SmallVectorImpl<int> &ShuffleMask) {
248 return; // Not a shuffle
250 unsigned HalfSize = VT.getVectorNumElements()/2;
252 for (unsigned l = 0; l != 2; ++l) {
253 unsigned HalfBegin = ((Imm >> (l*4)) & 0x3) * HalfSize;
254 for (unsigned i = HalfBegin, e = HalfBegin+HalfSize; i != e; ++i)
255 ShuffleMask.push_back(i);
259 void DecodePSHUFBMask(const Constant *C, const DataLayout *TD,
260 SmallVectorImpl<int> &ShuffleMask) {
261 Type *MaskTy = C->getType();
262 // It is not an error for the PSHUFB mask to not be a vector of i8 because the
263 // constant pool uniques constants by their bit representation.
264 // e.g. the following take up the same space in the constant pool:
265 // i128 -170141183420855150465331762880109871104
267 // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160>
269 // <4 x i32> <i32 -2147483648, i32 -2147483648, i32 -2147483648, i32
272 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
274 VectorType *DestTy = nullptr;
275 if (MaskTySize == 128)
276 DestTy = VectorType::get(Type::getInt8Ty(MaskTy->getContext()), 16);
277 else if (MaskTySize == 256)
278 DestTy = VectorType::get(Type::getInt8Ty(MaskTy->getContext()), 32);
280 // FIXME: Add support for AVX-512.
284 if (DestTy != MaskTy) {
285 if (!CastInst::castIsValid(Instruction::BitCast, const_cast<Constant *>(C),
289 C = ConstantFoldInstOperands(Instruction::BitCast, DestTy,
290 const_cast<Constant *>(C), TD);
294 int NumElements = MaskTy->getVectorNumElements();
295 ShuffleMask.reserve(NumElements);
297 for (int i = 0; i < NumElements; ++i) {
298 // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte
299 // lane of the vector we're inside.
300 int Base = i < 16 ? 0 : 16;
301 Constant *COp = C->getAggregateElement(i);
305 } else if (isa<UndefValue>(COp)) {
306 ShuffleMask.push_back(SM_SentinelUndef);
309 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
310 // If the high bit (7) of the byte is set, the element is zeroed.
311 if (Element & (1 << 7))
312 ShuffleMask.push_back(SM_SentinelZero);
314 // Only the least significant 4 bits of the byte are used.
315 int Index = Base + (Element & 0xf);
316 ShuffleMask.push_back(Index);
321 void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask,
322 SmallVectorImpl<int> &ShuffleMask) {
323 for (int i = 0, e = RawMask.size(); i < e; ++i) {
324 uint64_t M = RawMask[i];
325 if (M == (uint64_t)SM_SentinelUndef) {
326 ShuffleMask.push_back(M);
329 // For AVX vectors with 32 bytes the base of the shuffle is the half of
330 // the vector we're inside.
331 int Base = i < 16 ? 0 : 16;
332 // If the high bit (7) of the byte is set, the element is zeroed.
334 ShuffleMask.push_back(SM_SentinelZero);
336 // Only the least significant 4 bits of the byte are used.
337 int Index = Base + (M & 0xf);
338 ShuffleMask.push_back(Index);
343 void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
344 int ElementBits = VT.getScalarSizeInBits();
345 int NumElements = VT.getVectorNumElements();
346 for (int i = 0; i < NumElements; ++i) {
347 // If there are more than 8 elements in the vector, then any immediate blend
348 // mask applies to each 128-bit lane. There can never be more than
349 // 8 elements in a 128-bit lane with an immediate blend.
350 int Bit = NumElements > 8 ? i % (128 / ElementBits) : i;
352 "Immediate blends only operate over 8 elements at a time!");
353 ShuffleMask.push_back(((Imm >> Bit) & 1) ? NumElements + i : i);
357 /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
358 /// No VT provided since it only works on 256-bit, 4 element vectors.
359 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
360 for (unsigned i = 0; i != 4; ++i) {
361 ShuffleMask.push_back((Imm >> (2*i)) & 3);
365 void DecodeVPERMILPMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
366 Type *MaskTy = C->getType();
367 assert(MaskTy->isVectorTy() && "Expected a vector constant mask!");
368 assert(MaskTy->getVectorElementType()->isIntegerTy() &&
369 "Expected integer constant mask elements!");
370 int ElementBits = MaskTy->getScalarSizeInBits();
371 int NumElements = MaskTy->getVectorNumElements();
372 assert((NumElements == 2 || NumElements == 4 || NumElements == 8) &&
373 "Unexpected number of vector elements.");
374 ShuffleMask.reserve(NumElements);
375 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
376 assert((unsigned)NumElements == CDS->getNumElements() &&
377 "Constant mask has a different number of elements!");
379 for (int i = 0; i < NumElements; ++i) {
380 int Base = (i * ElementBits / 128) * (128 / ElementBits);
381 uint64_t Element = CDS->getElementAsInteger(i);
382 // Only the least significant 2 bits of the integer are used.
383 int Index = Base + (Element & 0x3);
384 ShuffleMask.push_back(Index);
386 } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
387 assert((unsigned)NumElements == C->getNumOperands() &&
388 "Constant mask has a different number of elements!");
390 for (int i = 0; i < NumElements; ++i) {
391 int Base = (i * ElementBits / 128) * (128 / ElementBits);
392 Constant *COp = CV->getOperand(i);
393 if (isa<UndefValue>(COp)) {
394 ShuffleMask.push_back(SM_SentinelUndef);
397 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
398 // Only the least significant 2 bits of the integer are used.
399 int Index = Base + (Element & 0x3);
400 ShuffleMask.push_back(Index);