1 //===-- X86ShuffleDecode.cpp - X86 shuffle decode logic -------------------===//
\r
3 // The LLVM Compiler Infrastructure
\r
5 // This file is distributed under the University of Illinois Open Source
\r
6 // License. See LICENSE.TXT for details.
\r
8 //===----------------------------------------------------------------------===//
\r
10 // Define several functions to decode x86 specific shuffle semantics into a
\r
11 // generic vector mask.
\r
13 //===----------------------------------------------------------------------===//
\r
15 #include "X86ShuffleDecode.h"
\r
16 #include "llvm/IR/Constants.h"
\r
17 #include "llvm/CodeGen/MachineValueType.h"
\r
19 //===----------------------------------------------------------------------===//
\r
20 // Vector Mask Decoding
\r
21 //===----------------------------------------------------------------------===//
\r
25 void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
26 // Defaults the copying the dest value.
\r
27 ShuffleMask.push_back(0);
\r
28 ShuffleMask.push_back(1);
\r
29 ShuffleMask.push_back(2);
\r
30 ShuffleMask.push_back(3);
\r
32 // Decode the immediate.
\r
33 unsigned ZMask = Imm & 15;
\r
34 unsigned CountD = (Imm >> 4) & 3;
\r
35 unsigned CountS = (Imm >> 6) & 3;
\r
37 // CountS selects which input element to use.
\r
38 unsigned InVal = 4+CountS;
\r
39 // CountD specifies which element of destination to update.
\r
40 ShuffleMask[CountD] = InVal;
\r
41 // ZMask zaps values, potentially overriding the CountD elt.
\r
42 if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero;
\r
43 if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero;
\r
44 if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero;
\r
45 if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero;
\r
48 // <3,1> or <6,7,2,3>
\r
49 void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
\r
50 for (unsigned i = NElts/2; i != NElts; ++i)
\r
51 ShuffleMask.push_back(NElts+i);
\r
53 for (unsigned i = NElts/2; i != NElts; ++i)
\r
54 ShuffleMask.push_back(i);
\r
57 // <0,2> or <0,1,4,5>
\r
58 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
\r
59 for (unsigned i = 0; i != NElts/2; ++i)
\r
60 ShuffleMask.push_back(i);
\r
62 for (unsigned i = 0; i != NElts/2; ++i)
\r
63 ShuffleMask.push_back(NElts+i);
\r
66 void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
\r
67 unsigned NumElts = VT.getVectorNumElements();
\r
68 for (int i = 0, e = NumElts / 2; i < e; ++i) {
\r
69 ShuffleMask.push_back(2 * i);
\r
70 ShuffleMask.push_back(2 * i);
\r
74 void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
\r
75 unsigned NumElts = VT.getVectorNumElements();
\r
76 for (int i = 0, e = NumElts / 2; i < e; ++i) {
\r
77 ShuffleMask.push_back(2 * i + 1);
\r
78 ShuffleMask.push_back(2 * i + 1);
\r
82 void DecodeMOVDDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
\r
83 unsigned VectorSizeInBits = VT.getSizeInBits();
\r
84 unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
\r
85 unsigned NumElts = VT.getVectorNumElements();
\r
86 unsigned NumLanes = VectorSizeInBits / 128;
\r
87 unsigned NumLaneElts = NumElts / NumLanes;
\r
88 unsigned NumLaneSubElts = 64 / ScalarSizeInBits;
\r
90 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
\r
91 for (unsigned i = 0; i < NumLaneElts; i += NumLaneSubElts)
\r
92 for (unsigned s = 0; s != NumLaneSubElts; s++)
\r
93 ShuffleMask.push_back(l + s);
\r
96 void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
97 unsigned VectorSizeInBits = VT.getSizeInBits();
\r
98 unsigned NumElts = VectorSizeInBits / 8;
\r
99 unsigned NumLanes = VectorSizeInBits / 128;
\r
100 unsigned NumLaneElts = NumElts / NumLanes;
\r
102 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
\r
103 for (unsigned i = 0; i < NumLaneElts; ++i) {
\r
104 int M = SM_SentinelZero;
\r
105 if (i >= Imm) M = i - Imm + l;
\r
106 ShuffleMask.push_back(M);
\r
110 void DecodePSRLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
111 unsigned VectorSizeInBits = VT.getSizeInBits();
\r
112 unsigned NumElts = VectorSizeInBits / 8;
\r
113 unsigned NumLanes = VectorSizeInBits / 128;
\r
114 unsigned NumLaneElts = NumElts / NumLanes;
\r
116 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
\r
117 for (unsigned i = 0; i < NumLaneElts; ++i) {
\r
118 unsigned Base = i + Imm;
\r
120 if (Base >= NumLaneElts) M = SM_SentinelZero;
\r
121 ShuffleMask.push_back(M);
\r
125 void DecodePALIGNRMask(MVT VT, unsigned Imm,
\r
126 SmallVectorImpl<int> &ShuffleMask) {
\r
127 unsigned NumElts = VT.getVectorNumElements();
\r
128 unsigned Offset = Imm * (VT.getVectorElementType().getSizeInBits() / 8);
\r
130 unsigned NumLanes = VT.getSizeInBits() / 128;
\r
131 unsigned NumLaneElts = NumElts / NumLanes;
\r
133 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
\r
134 for (unsigned i = 0; i != NumLaneElts; ++i) {
\r
135 unsigned Base = i + Offset;
\r
136 // if i+offset is out of this lane then we actually need the other source
\r
137 if (Base >= NumLaneElts) Base += NumElts - NumLaneElts;
\r
138 ShuffleMask.push_back(Base + l);
\r
143 /// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*.
\r
144 /// VT indicates the type of the vector allowing it to handle different
\r
145 /// datatypes and vector widths.
\r
146 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
147 unsigned NumElts = VT.getVectorNumElements();
\r
149 unsigned NumLanes = VT.getSizeInBits() / 128;
\r
150 unsigned NumLaneElts = NumElts / NumLanes;
\r
152 unsigned NewImm = Imm;
\r
153 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
\r
154 for (unsigned i = 0; i != NumLaneElts; ++i) {
\r
155 ShuffleMask.push_back(NewImm % NumLaneElts + l);
\r
156 NewImm /= NumLaneElts;
\r
158 if (NumLaneElts == 4) NewImm = Imm; // reload imm
\r
162 void DecodePSHUFHWMask(MVT VT, unsigned Imm,
\r
163 SmallVectorImpl<int> &ShuffleMask) {
\r
164 unsigned NumElts = VT.getVectorNumElements();
\r
166 for (unsigned l = 0; l != NumElts; l += 8) {
\r
167 unsigned NewImm = Imm;
\r
168 for (unsigned i = 0, e = 4; i != e; ++i) {
\r
169 ShuffleMask.push_back(l + i);
\r
171 for (unsigned i = 4, e = 8; i != e; ++i) {
\r
172 ShuffleMask.push_back(l + 4 + (NewImm & 3));
\r
178 void DecodePSHUFLWMask(MVT VT, unsigned Imm,
\r
179 SmallVectorImpl<int> &ShuffleMask) {
\r
180 unsigned NumElts = VT.getVectorNumElements();
\r
182 for (unsigned l = 0; l != NumElts; l += 8) {
\r
183 unsigned NewImm = Imm;
\r
184 for (unsigned i = 0, e = 4; i != e; ++i) {
\r
185 ShuffleMask.push_back(l + (NewImm & 3));
\r
188 for (unsigned i = 4, e = 8; i != e; ++i) {
\r
189 ShuffleMask.push_back(l + i);
\r
194 /// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates
\r
195 /// the type of the vector allowing it to handle different datatypes and vector
\r
197 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
198 unsigned NumElts = VT.getVectorNumElements();
\r
200 unsigned NumLanes = VT.getSizeInBits() / 128;
\r
201 unsigned NumLaneElts = NumElts / NumLanes;
\r
203 unsigned NewImm = Imm;
\r
204 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
\r
205 // each half of a lane comes from different source
\r
206 for (unsigned s = 0; s != NumElts*2; s += NumElts) {
\r
207 for (unsigned i = 0; i != NumLaneElts/2; ++i) {
\r
208 ShuffleMask.push_back(NewImm % NumLaneElts + s + l);
\r
209 NewImm /= NumLaneElts;
\r
212 if (NumLaneElts == 4) NewImm = Imm; // reload imm
\r
216 /// DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd
\r
217 /// and punpckh*. VT indicates the type of the vector allowing it to handle
\r
218 /// different datatypes and vector widths.
\r
219 void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
\r
220 unsigned NumElts = VT.getVectorNumElements();
\r
222 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
\r
223 // independently on 128-bit lanes.
\r
224 unsigned NumLanes = VT.getSizeInBits() / 128;
\r
225 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
\r
226 unsigned NumLaneElts = NumElts / NumLanes;
\r
228 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
\r
229 for (unsigned i = l + NumLaneElts/2, e = l + NumLaneElts; i != e; ++i) {
\r
230 ShuffleMask.push_back(i); // Reads from dest/src1
\r
231 ShuffleMask.push_back(i+NumElts); // Reads from src/src2
\r
236 /// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
\r
237 /// and punpckl*. VT indicates the type of the vector allowing it to handle
\r
238 /// different datatypes and vector widths.
\r
239 void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
\r
240 unsigned NumElts = VT.getVectorNumElements();
\r
242 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
\r
243 // independently on 128-bit lanes.
\r
244 unsigned NumLanes = VT.getSizeInBits() / 128;
\r
245 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
\r
246 unsigned NumLaneElts = NumElts / NumLanes;
\r
248 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
\r
249 for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) {
\r
250 ShuffleMask.push_back(i); // Reads from dest/src1
\r
251 ShuffleMask.push_back(i+NumElts); // Reads from src/src2
\r
256 void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
\r
257 SmallVectorImpl<int> &ShuffleMask) {
\r
259 return; // Not a shuffle
\r
261 unsigned HalfSize = VT.getVectorNumElements()/2;
\r
263 for (unsigned l = 0; l != 2; ++l) {
\r
264 unsigned HalfBegin = ((Imm >> (l*4)) & 0x3) * HalfSize;
\r
265 for (unsigned i = HalfBegin, e = HalfBegin+HalfSize; i != e; ++i)
\r
266 ShuffleMask.push_back(i);
\r
270 void DecodePSHUFBMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
\r
271 Type *MaskTy = C->getType();
\r
272 // It is not an error for the PSHUFB mask to not be a vector of i8 because the
\r
273 // constant pool uniques constants by their bit representation.
\r
274 // e.g. the following take up the same space in the constant pool:
\r
275 // i128 -170141183420855150465331762880109871104
\r
277 // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160>
\r
279 // <4 x i32> <i32 -2147483648, i32 -2147483648,
\r
280 // i32 -2147483648, i32 -2147483648>
\r
282 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
\r
284 if (MaskTySize != 128 && MaskTySize != 256) // FIXME: Add support for AVX-512.
\r
287 // This is a straightforward byte vector.
\r
288 if (MaskTy->isVectorTy() && MaskTy->getVectorElementType()->isIntegerTy(8)) {
\r
289 int NumElements = MaskTy->getVectorNumElements();
\r
290 ShuffleMask.reserve(NumElements);
\r
292 for (int i = 0; i < NumElements; ++i) {
\r
293 // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte
\r
294 // lane of the vector we're inside.
\r
295 int Base = i < 16 ? 0 : 16;
\r
296 Constant *COp = C->getAggregateElement(i);
\r
298 ShuffleMask.clear();
\r
300 } else if (isa<UndefValue>(COp)) {
\r
301 ShuffleMask.push_back(SM_SentinelUndef);
\r
304 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
\r
305 // If the high bit (7) of the byte is set, the element is zeroed.
\r
306 if (Element & (1 << 7))
\r
307 ShuffleMask.push_back(SM_SentinelZero);
\r
309 // Only the least significant 4 bits of the byte are used.
\r
310 int Index = Base + (Element & 0xf);
\r
311 ShuffleMask.push_back(Index);
\r
315 // TODO: Handle funny-looking vectors too.
\r
318 void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask,
\r
319 SmallVectorImpl<int> &ShuffleMask) {
\r
320 for (int i = 0, e = RawMask.size(); i < e; ++i) {
\r
321 uint64_t M = RawMask[i];
\r
322 if (M == (uint64_t)SM_SentinelUndef) {
\r
323 ShuffleMask.push_back(M);
\r
326 // For AVX vectors with 32 bytes the base of the shuffle is the half of
\r
327 // the vector we're inside.
\r
328 int Base = i < 16 ? 0 : 16;
\r
329 // If the high bit (7) of the byte is set, the element is zeroed.
\r
331 ShuffleMask.push_back(SM_SentinelZero);
\r
333 // Only the least significant 4 bits of the byte are used.
\r
334 int Index = Base + (M & 0xf);
\r
335 ShuffleMask.push_back(Index);
\r
340 void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
341 int ElementBits = VT.getScalarSizeInBits();
\r
342 int NumElements = VT.getVectorNumElements();
\r
343 for (int i = 0; i < NumElements; ++i) {
\r
344 // If there are more than 8 elements in the vector, then any immediate blend
\r
345 // mask applies to each 128-bit lane. There can never be more than
\r
346 // 8 elements in a 128-bit lane with an immediate blend.
\r
347 int Bit = NumElements > 8 ? i % (128 / ElementBits) : i;
\r
349 "Immediate blends only operate over 8 elements at a time!");
\r
350 ShuffleMask.push_back(((Imm >> Bit) & 1) ? NumElements + i : i);
\r
354 /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
\r
355 /// No VT provided since it only works on 256-bit, 4 element vectors.
\r
356 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
\r
357 for (unsigned i = 0; i != 4; ++i) {
\r
358 ShuffleMask.push_back((Imm >> (2*i)) & 3);
\r
362 void DecodeVPERMILPMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
\r
363 Type *MaskTy = C->getType();
\r
364 assert(MaskTy->isVectorTy() && "Expected a vector constant mask!");
\r
365 assert(MaskTy->getVectorElementType()->isIntegerTy() &&
\r
366 "Expected integer constant mask elements!");
\r
367 int ElementBits = MaskTy->getScalarSizeInBits();
\r
368 int NumElements = MaskTy->getVectorNumElements();
\r
369 assert((NumElements == 2 || NumElements == 4 || NumElements == 8) &&
\r
370 "Unexpected number of vector elements.");
\r
371 ShuffleMask.reserve(NumElements);
\r
372 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
\r
373 assert((unsigned)NumElements == CDS->getNumElements() &&
\r
374 "Constant mask has a different number of elements!");
\r
376 for (int i = 0; i < NumElements; ++i) {
\r
377 int Base = (i * ElementBits / 128) * (128 / ElementBits);
\r
378 uint64_t Element = CDS->getElementAsInteger(i);
\r
379 // Only the least significant 2 bits of the integer are used.
\r
380 int Index = Base + (Element & 0x3);
\r
381 ShuffleMask.push_back(Index);
\r
383 } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
\r
384 assert((unsigned)NumElements == C->getNumOperands() &&
\r
385 "Constant mask has a different number of elements!");
\r
387 for (int i = 0; i < NumElements; ++i) {
\r
388 int Base = (i * ElementBits / 128) * (128 / ElementBits);
\r
389 Constant *COp = CV->getOperand(i);
\r
390 if (isa<UndefValue>(COp)) {
\r
391 ShuffleMask.push_back(SM_SentinelUndef);
\r
394 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
\r
395 // Only the least significant 2 bits of the integer are used.
\r
396 int Index = Base + (Element & 0x3);
\r
397 ShuffleMask.push_back(Index);
\r
402 void DecodeZeroExtendMask(MVT SrcVT, MVT DstVT, SmallVectorImpl<int> &Mask) {
\r
403 unsigned NumDstElts = DstVT.getVectorNumElements();
\r
404 unsigned SrcScalarBits = SrcVT.getScalarSizeInBits();
\r
405 unsigned DstScalarBits = DstVT.getScalarSizeInBits();
\r
406 unsigned Scale = DstScalarBits / SrcScalarBits;
\r
407 assert(SrcScalarBits < DstScalarBits &&
\r
408 "Expected zero extension mask to increase scalar size");
\r
409 assert(SrcVT.getVectorNumElements() >= NumDstElts &&
\r
410 "Too many zero extension lanes");
\r
412 for (unsigned i = 0; i != NumDstElts; i++) {
\r
414 for (unsigned j = 1; j != Scale; j++)
\r
415 Mask.push_back(SM_SentinelZero);
\r
419 void DecodeZeroMoveLowMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
\r
420 unsigned NumElts = VT.getVectorNumElements();
\r
421 ShuffleMask.push_back(0);
\r
422 for (unsigned i = 1; i < NumElts; i++)
\r
423 ShuffleMask.push_back(SM_SentinelZero);
\r
426 void DecodeScalarMoveMask(MVT VT, bool IsLoad, SmallVectorImpl<int> &Mask) {
\r
427 // First element comes from the first element of second source.
\r
428 // Remaining elements: Load zero extends / Move copies from first source.
\r
429 unsigned NumElts = VT.getVectorNumElements();
\r
430 Mask.push_back(NumElts);
\r
431 for (unsigned i = 1; i < NumElts; i++)
\r
432 Mask.push_back(IsLoad ? static_cast<int>(SM_SentinelZero) : i);
\r
434 } // llvm namespace
\r