LLVM  6.0.0svn
X86ShuffleDecode.cpp
Go to the documentation of this file.
1 //===-- X86ShuffleDecode.cpp - X86 shuffle decode logic -------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Define several functions to decode x86 specific shuffle semantics into a
11 // generic vector mask.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "X86ShuffleDecode.h"
16 #include "llvm/ADT/ArrayRef.h"
18 
19 //===----------------------------------------------------------------------===//
20 // Vector Mask Decoding
21 //===----------------------------------------------------------------------===//
22 
23 namespace llvm {
24 
25 void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
26  // Defaults the copying the dest value.
27  ShuffleMask.push_back(0);
28  ShuffleMask.push_back(1);
29  ShuffleMask.push_back(2);
30  ShuffleMask.push_back(3);
31 
32  // Decode the immediate.
33  unsigned ZMask = Imm & 15;
34  unsigned CountD = (Imm >> 4) & 3;
35  unsigned CountS = (Imm >> 6) & 3;
36 
37  // CountS selects which input element to use.
38  unsigned InVal = 4 + CountS;
39  // CountD specifies which element of destination to update.
40  ShuffleMask[CountD] = InVal;
41  // ZMask zaps values, potentially overriding the CountD elt.
42  if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero;
43  if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero;
44  if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero;
45  if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero;
46 }
47 
48 void DecodeInsertElementMask(MVT VT, unsigned Idx, unsigned Len,
49  SmallVectorImpl<int> &ShuffleMask) {
50  unsigned NumElts = VT.getVectorNumElements();
51  assert((Idx + Len) <= NumElts && "Insertion out of range");
52 
53  for (unsigned i = 0; i != NumElts; ++i)
54  ShuffleMask.push_back(i);
55  for (unsigned i = 0; i != Len; ++i)
56  ShuffleMask[Idx + i] = NumElts + i;
57 }
58 
59 // <3,1> or <6,7,2,3>
60 void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
61  for (unsigned i = NElts / 2; i != NElts; ++i)
62  ShuffleMask.push_back(NElts + i);
63 
64  for (unsigned i = NElts / 2; i != NElts; ++i)
65  ShuffleMask.push_back(i);
66 }
67 
68 // <0,2> or <0,1,4,5>
69 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
70  for (unsigned i = 0; i != NElts / 2; ++i)
71  ShuffleMask.push_back(i);
72 
73  for (unsigned i = 0; i != NElts / 2; ++i)
74  ShuffleMask.push_back(NElts + i);
75 }
76 
78  unsigned NumElts = VT.getVectorNumElements();
79  for (int i = 0, e = NumElts / 2; i < e; ++i) {
80  ShuffleMask.push_back(2 * i);
81  ShuffleMask.push_back(2 * i);
82  }
83 }
84 
86  unsigned NumElts = VT.getVectorNumElements();
87  for (int i = 0, e = NumElts / 2; i < e; ++i) {
88  ShuffleMask.push_back(2 * i + 1);
89  ShuffleMask.push_back(2 * i + 1);
90  }
91 }
92 
93 void DecodeMOVDDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
94  unsigned VectorSizeInBits = VT.getSizeInBits();
95  unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
96  unsigned NumElts = VT.getVectorNumElements();
97  unsigned NumLanes = VectorSizeInBits / 128;
98  unsigned NumLaneElts = NumElts / NumLanes;
99  unsigned NumLaneSubElts = 64 / ScalarSizeInBits;
100 
101  for (unsigned l = 0; l < NumElts; l += NumLaneElts)
102  for (unsigned i = 0; i < NumLaneElts; i += NumLaneSubElts)
103  for (unsigned s = 0; s != NumLaneSubElts; s++)
104  ShuffleMask.push_back(l + s);
105 }
106 
107 void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
108  unsigned VectorSizeInBits = VT.getSizeInBits();
109  unsigned NumElts = VectorSizeInBits / 8;
110  unsigned NumLanes = VectorSizeInBits / 128;
111  unsigned NumLaneElts = NumElts / NumLanes;
112 
113  for (unsigned l = 0; l < NumElts; l += NumLaneElts)
114  for (unsigned i = 0; i < NumLaneElts; ++i) {
115  int M = SM_SentinelZero;
116  if (i >= Imm) M = i - Imm + l;
117  ShuffleMask.push_back(M);
118  }
119 }
120 
121 void DecodePSRLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
122  unsigned VectorSizeInBits = VT.getSizeInBits();
123  unsigned NumElts = VectorSizeInBits / 8;
124  unsigned NumLanes = VectorSizeInBits / 128;
125  unsigned NumLaneElts = NumElts / NumLanes;
126 
127  for (unsigned l = 0; l < NumElts; l += NumLaneElts)
128  for (unsigned i = 0; i < NumLaneElts; ++i) {
129  unsigned Base = i + Imm;
130  int M = Base + l;
131  if (Base >= NumLaneElts) M = SM_SentinelZero;
132  ShuffleMask.push_back(M);
133  }
134 }
135 
136 void DecodePALIGNRMask(MVT VT, unsigned Imm,
137  SmallVectorImpl<int> &ShuffleMask) {
138  unsigned NumElts = VT.getVectorNumElements();
139  unsigned Offset = Imm * (VT.getScalarSizeInBits() / 8);
140 
141  unsigned NumLanes = VT.getSizeInBits() / 128;
142  unsigned NumLaneElts = NumElts / NumLanes;
143 
144  for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
145  for (unsigned i = 0; i != NumLaneElts; ++i) {
146  unsigned Base = i + Offset;
147  // if i+offset is out of this lane then we actually need the other source
148  if (Base >= NumLaneElts) Base += NumElts - NumLaneElts;
149  ShuffleMask.push_back(Base + l);
150  }
151  }
152 }
153 
154 void DecodeVALIGNMask(MVT VT, unsigned Imm,
155  SmallVectorImpl<int> &ShuffleMask) {
156  int NumElts = VT.getVectorNumElements();
157  // Not all bits of the immediate are used so mask it.
158  assert(isPowerOf2_32(NumElts) && "NumElts should be power of 2");
159  Imm = Imm & (NumElts - 1);
160  for (int i = 0; i != NumElts; ++i)
161  ShuffleMask.push_back(i + Imm);
162 }
163 
164 /// DecodePSHUFMask - This decodes the shuffle masks for pshufw, pshufd, and vpermilp*.
165 /// VT indicates the type of the vector allowing it to handle different
166 /// datatypes and vector widths.
167 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
168  unsigned NumElts = VT.getVectorNumElements();
169 
170  unsigned NumLanes = VT.getSizeInBits() / 128;
171  if (NumLanes == 0) NumLanes = 1; // Handle MMX
172  unsigned NumLaneElts = NumElts / NumLanes;
173 
174  unsigned NewImm = Imm;
175  for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
176  for (unsigned i = 0; i != NumLaneElts; ++i) {
177  ShuffleMask.push_back(NewImm % NumLaneElts + l);
178  NewImm /= NumLaneElts;
179  }
180  if (NumLaneElts == 4) NewImm = Imm; // reload imm
181  }
182 }
183 
184 void DecodePSHUFHWMask(MVT VT, unsigned Imm,
185  SmallVectorImpl<int> &ShuffleMask) {
186  unsigned NumElts = VT.getVectorNumElements();
187 
188  for (unsigned l = 0; l != NumElts; l += 8) {
189  unsigned NewImm = Imm;
190  for (unsigned i = 0, e = 4; i != e; ++i) {
191  ShuffleMask.push_back(l + i);
192  }
193  for (unsigned i = 4, e = 8; i != e; ++i) {
194  ShuffleMask.push_back(l + 4 + (NewImm & 3));
195  NewImm >>= 2;
196  }
197  }
198 }
199 
200 void DecodePSHUFLWMask(MVT VT, unsigned Imm,
201  SmallVectorImpl<int> &ShuffleMask) {
202  unsigned NumElts = VT.getVectorNumElements();
203 
204  for (unsigned l = 0; l != NumElts; l += 8) {
205  unsigned NewImm = Imm;
206  for (unsigned i = 0, e = 4; i != e; ++i) {
207  ShuffleMask.push_back(l + (NewImm & 3));
208  NewImm >>= 2;
209  }
210  for (unsigned i = 4, e = 8; i != e; ++i) {
211  ShuffleMask.push_back(l + i);
212  }
213  }
214 }
215 
216 void DecodePSWAPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
217  unsigned NumElts = VT.getVectorNumElements();
218  unsigned NumHalfElts = NumElts / 2;
219 
220  for (unsigned l = 0; l != NumHalfElts; ++l)
221  ShuffleMask.push_back(l + NumHalfElts);
222  for (unsigned h = 0; h != NumHalfElts; ++h)
223  ShuffleMask.push_back(h);
224 }
225 
226 /// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates
227 /// the type of the vector allowing it to handle different datatypes and vector
228 /// widths.
229 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
230  unsigned NumElts = VT.getVectorNumElements();
231 
232  unsigned NumLanes = VT.getSizeInBits() / 128;
233  unsigned NumLaneElts = NumElts / NumLanes;
234 
235  unsigned NewImm = Imm;
236  for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
237  // each half of a lane comes from different source
238  for (unsigned s = 0; s != NumElts * 2; s += NumElts) {
239  for (unsigned i = 0; i != NumLaneElts / 2; ++i) {
240  ShuffleMask.push_back(NewImm % NumLaneElts + s + l);
241  NewImm /= NumLaneElts;
242  }
243  }
244  if (NumLaneElts == 4) NewImm = Imm; // reload imm
245  }
246 }
247 
248 /// DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd
249 /// and punpckh*. VT indicates the type of the vector allowing it to handle
250 /// different datatypes and vector widths.
251 void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
252  unsigned NumElts = VT.getVectorNumElements();
253 
254  // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
255  // independently on 128-bit lanes.
256  unsigned NumLanes = VT.getSizeInBits() / 128;
257  if (NumLanes == 0) NumLanes = 1; // Handle MMX
258  unsigned NumLaneElts = NumElts / NumLanes;
259 
260  for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
261  for (unsigned i = l + NumLaneElts / 2, e = l + NumLaneElts; i != e; ++i) {
262  ShuffleMask.push_back(i); // Reads from dest/src1
263  ShuffleMask.push_back(i + NumElts); // Reads from src/src2
264  }
265  }
266 }
267 
268 /// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
269 /// and punpckl*. VT indicates the type of the vector allowing it to handle
270 /// different datatypes and vector widths.
271 void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
272  unsigned NumElts = VT.getVectorNumElements();
273 
274  // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
275  // independently on 128-bit lanes.
276  unsigned NumLanes = VT.getSizeInBits() / 128;
277  if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
278  unsigned NumLaneElts = NumElts / NumLanes;
279 
280  for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
281  for (unsigned i = l, e = l + NumLaneElts / 2; i != e; ++i) {
282  ShuffleMask.push_back(i); // Reads from dest/src1
283  ShuffleMask.push_back(i + NumElts); // Reads from src/src2
284  }
285  }
286 }
287 
288 /// Decodes a broadcast of the first element of a vector.
289 void DecodeVectorBroadcast(MVT DstVT, SmallVectorImpl<int> &ShuffleMask) {
290  unsigned NumElts = DstVT.getVectorNumElements();
291  ShuffleMask.append(NumElts, 0);
292 }
293 
294 /// Decodes a broadcast of a subvector to a larger vector type.
295 void DecodeSubVectorBroadcast(MVT DstVT, MVT SrcVT,
296  SmallVectorImpl<int> &ShuffleMask) {
297  assert(SrcVT.getScalarType() == DstVT.getScalarType() &&
298  "Non matching vector element types");
299  unsigned NumElts = SrcVT.getVectorNumElements();
300  unsigned Scale = DstVT.getSizeInBits() / SrcVT.getSizeInBits();
301 
302  for (unsigned i = 0; i != Scale; ++i)
303  for (unsigned j = 0; j != NumElts; ++j)
304  ShuffleMask.push_back(j);
305 }
306 
307 /// \brief Decode a shuffle packed values at 128-bit granularity
308 /// (SHUFF32x4/SHUFF64x2/SHUFI32x4/SHUFI64x2)
309 /// immediate mask into a shuffle mask.
310 void decodeVSHUF64x2FamilyMask(MVT VT, unsigned Imm,
311  SmallVectorImpl<int> &ShuffleMask) {
312  unsigned NumLanes = VT.getSizeInBits() / 128;
313  unsigned NumElementsInLane = 128 / VT.getScalarSizeInBits();
314  unsigned ControlBitsMask = NumLanes - 1;
315  unsigned NumControlBits = NumLanes / 2;
316 
317  for (unsigned l = 0; l != NumLanes; ++l) {
318  unsigned LaneMask = (Imm >> (l * NumControlBits)) & ControlBitsMask;
319  // We actually need the other source.
320  if (l >= NumLanes / 2)
321  LaneMask += NumLanes;
322  for (unsigned i = 0; i != NumElementsInLane; ++i)
323  ShuffleMask.push_back(LaneMask * NumElementsInLane + i);
324  }
325 }
326 
327 void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
328  SmallVectorImpl<int> &ShuffleMask) {
329  unsigned HalfSize = VT.getVectorNumElements() / 2;
330 
331  for (unsigned l = 0; l != 2; ++l) {
332  unsigned HalfMask = Imm >> (l * 4);
333  unsigned HalfBegin = (HalfMask & 0x3) * HalfSize;
334  for (unsigned i = HalfBegin, e = HalfBegin + HalfSize; i != e; ++i)
335  ShuffleMask.push_back(HalfMask & 8 ? SM_SentinelZero : (int)i);
336  }
337 }
338 
340  SmallVectorImpl<int> &ShuffleMask) {
341  for (int i = 0, e = RawMask.size(); i < e; ++i) {
342  uint64_t M = RawMask[i];
343  if (M == (uint64_t)SM_SentinelUndef) {
344  ShuffleMask.push_back(M);
345  continue;
346  }
347  // For 256/512-bit vectors the base of the shuffle is the 128-bit
348  // subvector we're inside.
349  int Base = (i / 16) * 16;
350  // If the high bit (7) of the byte is set, the element is zeroed.
351  if (M & (1 << 7))
352  ShuffleMask.push_back(SM_SentinelZero);
353  else {
354  // Only the least significant 4 bits of the byte are used.
355  int Index = Base + (M & 0xf);
356  ShuffleMask.push_back(Index);
357  }
358  }
359 }
360 
361 void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
362  int ElementBits = VT.getScalarSizeInBits();
363  int NumElements = VT.getVectorNumElements();
364  for (int i = 0; i < NumElements; ++i) {
365  // If there are more than 8 elements in the vector, then any immediate blend
366  // mask applies to each 128-bit lane. There can never be more than
367  // 8 elements in a 128-bit lane with an immediate blend.
368  int Bit = NumElements > 8 ? i % (128 / ElementBits) : i;
369  assert(Bit < 8 &&
370  "Immediate blends only operate over 8 elements at a time!");
371  ShuffleMask.push_back(((Imm >> Bit) & 1) ? NumElements + i : i);
372  }
373 }
374 
376  SmallVectorImpl<int> &ShuffleMask) {
377  assert(RawMask.size() == 16 && "Illegal VPPERM shuffle mask size");
378 
379  // VPPERM Operation
380  // Bits[4:0] - Byte Index (0 - 31)
381  // Bits[7:5] - Permute Operation
382  //
383  // Permute Operation:
384  // 0 - Source byte (no logical operation).
385  // 1 - Invert source byte.
386  // 2 - Bit reverse of source byte.
387  // 3 - Bit reverse of inverted source byte.
388  // 4 - 00h (zero - fill).
389  // 5 - FFh (ones - fill).
390  // 6 - Most significant bit of source byte replicated in all bit positions.
391  // 7 - Invert most significant bit of source byte and replicate in all bit positions.
392  for (int i = 0, e = RawMask.size(); i < e; ++i) {
393  uint64_t M = RawMask[i];
394  if (M == (uint64_t)SM_SentinelUndef) {
395  ShuffleMask.push_back(M);
396  continue;
397  }
398 
399  uint64_t PermuteOp = (M >> 5) & 0x7;
400  if (PermuteOp == 4) {
401  ShuffleMask.push_back(SM_SentinelZero);
402  continue;
403  }
404  if (PermuteOp != 0) {
405  ShuffleMask.clear();
406  return;
407  }
408 
409  uint64_t Index = M & 0x1F;
410  ShuffleMask.push_back((int)Index);
411  }
412 }
413 
414 /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
415 void DecodeVPERMMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
416  assert((VT.is256BitVector() || VT.is512BitVector()) &&
417  (VT.getScalarSizeInBits() == 64) && "Unexpected vector value type");
418  unsigned NumElts = VT.getVectorNumElements();
419  for (unsigned l = 0; l != NumElts; l += 4)
420  for (unsigned i = 0; i != 4; ++i)
421  ShuffleMask.push_back(l + ((Imm >> (2 * i)) & 3));
422 }
423 
425  unsigned NumDstElts = DstVT.getVectorNumElements();
426  unsigned SrcScalarBits = SrcScalarVT.getSizeInBits();
427  unsigned DstScalarBits = DstVT.getScalarSizeInBits();
428  unsigned Scale = DstScalarBits / SrcScalarBits;
429  assert(SrcScalarBits < DstScalarBits &&
430  "Expected zero extension mask to increase scalar size");
431 
432  for (unsigned i = 0; i != NumDstElts; i++) {
433  Mask.push_back(i);
434  for (unsigned j = 1; j != Scale; j++)
436  }
437 }
438 
440  unsigned NumElts = VT.getVectorNumElements();
441  ShuffleMask.push_back(0);
442  for (unsigned i = 1; i < NumElts; i++)
443  ShuffleMask.push_back(SM_SentinelZero);
444 }
445 
447  // First element comes from the first element of second source.
448  // Remaining elements: Load zero extends / Move copies from first source.
449  unsigned NumElts = VT.getVectorNumElements();
450  Mask.push_back(NumElts);
451  for (unsigned i = 1; i < NumElts; i++)
452  Mask.push_back(IsLoad ? static_cast<int>(SM_SentinelZero) : i);
453 }
454 
455 void DecodeEXTRQIMask(MVT VT, int Len, int Idx,
456  SmallVectorImpl<int> &ShuffleMask) {
457  assert(VT.is128BitVector() && "Expected 128-bit vector");
458  unsigned NumElts = VT.getVectorNumElements();
459  unsigned EltSize = VT.getScalarSizeInBits();
460  unsigned HalfElts = NumElts / 2;
461 
462  // Only the bottom 6 bits are valid for each immediate.
463  Len &= 0x3F;
464  Idx &= 0x3F;
465 
466  // We can only decode this bit extraction instruction as a shuffle if both the
467  // length and index work with whole elements.
468  if (0 != (Len % EltSize) || 0 != (Idx % EltSize))
469  return;
470 
471  // A length of zero is equivalent to a bit length of 64.
472  if (Len == 0)
473  Len = 64;
474 
475  // If the length + index exceeds the bottom 64 bits the result is undefined.
476  if ((Len + Idx) > 64) {
477  ShuffleMask.append(NumElts, SM_SentinelUndef);
478  return;
479  }
480 
481  // Convert index and index to work with elements.
482  Len /= EltSize;
483  Idx /= EltSize;
484 
485  // EXTRQ: Extract Len elements starting from Idx. Zero pad the remaining
486  // elements of the lower 64-bits. The upper 64-bits are undefined.
487  for (int i = 0; i != Len; ++i)
488  ShuffleMask.push_back(i + Idx);
489  for (int i = Len; i != (int)HalfElts; ++i)
490  ShuffleMask.push_back(SM_SentinelZero);
491  for (int i = HalfElts; i != (int)NumElts; ++i)
492  ShuffleMask.push_back(SM_SentinelUndef);
493 }
494 
495 void DecodeINSERTQIMask(MVT VT, int Len, int Idx,
496  SmallVectorImpl<int> &ShuffleMask) {
497  assert(VT.is128BitVector() && "Expected 128-bit vector");
498  unsigned NumElts = VT.getVectorNumElements();
499  unsigned EltSize = VT.getScalarSizeInBits();
500  unsigned HalfElts = NumElts / 2;
501 
502  // Only the bottom 6 bits are valid for each immediate.
503  Len &= 0x3F;
504  Idx &= 0x3F;
505 
506  // We can only decode this bit insertion instruction as a shuffle if both the
507  // length and index work with whole elements.
508  if (0 != (Len % EltSize) || 0 != (Idx % EltSize))
509  return;
510 
511  // A length of zero is equivalent to a bit length of 64.
512  if (Len == 0)
513  Len = 64;
514 
515  // If the length + index exceeds the bottom 64 bits the result is undefined.
516  if ((Len + Idx) > 64) {
517  ShuffleMask.append(NumElts, SM_SentinelUndef);
518  return;
519  }
520 
521  // Convert index and index to work with elements.
522  Len /= EltSize;
523  Idx /= EltSize;
524 
525  // INSERTQ: Extract lowest Len elements from lower half of second source and
526  // insert over first source starting at Idx element. The upper 64-bits are
527  // undefined.
528  for (int i = 0; i != Idx; ++i)
529  ShuffleMask.push_back(i);
530  for (int i = 0; i != Len; ++i)
531  ShuffleMask.push_back(i + NumElts);
532  for (int i = Idx + Len; i != (int)HalfElts; ++i)
533  ShuffleMask.push_back(i);
534  for (int i = HalfElts; i != (int)NumElts; ++i)
535  ShuffleMask.push_back(SM_SentinelUndef);
536 }
537 
539  SmallVectorImpl<int> &ShuffleMask) {
540  unsigned VecSize = VT.getSizeInBits();
541  unsigned EltSize = VT.getScalarSizeInBits();
542  unsigned NumLanes = VecSize / 128;
543  unsigned NumEltsPerLane = VT.getVectorNumElements() / NumLanes;
544  assert((VecSize == 128 || VecSize == 256 || VecSize == 512) &&
545  "Unexpected vector size");
546  assert((EltSize == 32 || EltSize == 64) && "Unexpected element size");
547 
548  for (unsigned i = 0, e = RawMask.size(); i < e; ++i) {
549  uint64_t M = RawMask[i];
550  M = (EltSize == 64 ? ((M >> 1) & 0x1) : (M & 0x3));
551  unsigned LaneOffset = i & ~(NumEltsPerLane - 1);
552  ShuffleMask.push_back((int)(LaneOffset + M));
553  }
554 }
555 
556 void DecodeVPERMIL2PMask(MVT VT, unsigned M2Z, ArrayRef<uint64_t> RawMask,
557  SmallVectorImpl<int> &ShuffleMask) {
558  unsigned VecSize = VT.getSizeInBits();
559  unsigned EltSize = VT.getScalarSizeInBits();
560  unsigned NumLanes = VecSize / 128;
561  unsigned NumElts = VT.getVectorNumElements();
562  unsigned NumEltsPerLane = NumElts / NumLanes;
563  assert((VecSize == 128 || VecSize == 256) && "Unexpected vector size");
564  assert((EltSize == 32 || EltSize == 64) && "Unexpected element size");
565  assert((NumElts == RawMask.size()) && "Unexpected mask size");
566 
567  for (unsigned i = 0, e = RawMask.size(); i < e; ++i) {
568  // VPERMIL2 Operation.
569  // Bits[3] - Match Bit.
570  // Bits[2:1] - (Per Lane) PD Shuffle Mask.
571  // Bits[2:0] - (Per Lane) PS Shuffle Mask.
572  uint64_t Selector = RawMask[i];
573  unsigned MatchBit = (Selector >> 3) & 0x1;
574 
575  // M2Z[0:1] MatchBit
576  // 0Xb X Source selected by Selector index.
577  // 10b 0 Source selected by Selector index.
578  // 10b 1 Zero.
579  // 11b 0 Zero.
580  // 11b 1 Source selected by Selector index.
581  if ((M2Z & 0x2) != 0 && MatchBit != (M2Z & 0x1)) {
582  ShuffleMask.push_back(SM_SentinelZero);
583  continue;
584  }
585 
586  int Index = i & ~(NumEltsPerLane - 1);
587  if (EltSize == 64)
588  Index += (Selector >> 1) & 0x1;
589  else
590  Index += Selector & 0x3;
591 
592  int Src = (Selector >> 2) & 0x1;
593  Index += Src * NumElts;
594  ShuffleMask.push_back(Index);
595  }
596 }
597 
599  SmallVectorImpl<int> &ShuffleMask) {
600  uint64_t EltMaskSize = RawMask.size() - 1;
601  for (auto M : RawMask) {
602  M &= EltMaskSize;
603  ShuffleMask.push_back((int)M);
604  }
605 }
606 
608  SmallVectorImpl<int> &ShuffleMask) {
609  uint64_t EltMaskSize = (RawMask.size() * 2) - 1;
610  for (auto M : RawMask) {
611  M &= EltMaskSize;
612  ShuffleMask.push_back((int)M);
613  }
614 }
615 
616 } // llvm namespace
void push_back(const T &Elt)
Definition: SmallVector.h:212
void DecodeVPERMILPMask(MVT VT, ArrayRef< uint64_t > RawMask, SmallVectorImpl< int > &ShuffleMask)
Decode a VPERMILPD/VPERMILPS variable mask from a raw array of constants.
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
void DecodeScalarMoveMask(MVT VT, bool IsLoad, SmallVectorImpl< int > &Mask)
Decode a scalar float move instruction as a shuffle mask.
void DecodePALIGNRMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
bool is256BitVector() const
Return true if this is a 256-bit vector type.
unsigned getVectorNumElements() const
void DecodeUNPCKLMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd and punpckl*.
void DecodeINSERTQIMask(MVT VT, int Len, int Idx, SmallVectorImpl< int > &ShuffleMask)
Decode a SSE4A INSERTQ instruction as a shuffle mask.
void DecodeEXTRQIMask(MVT VT, int Len, int Idx, SmallVectorImpl< int > &ShuffleMask)
Decode a SSE4A EXTRQ instruction as a shuffle mask.
void DecodeMOVDDUPMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
void DecodeSubVectorBroadcast(MVT DstVT, MVT SrcVT, SmallVectorImpl< int > &ShuffleMask)
Decodes a broadcast of a subvector to a larger vector type.
unsigned getSizeInBits() const
void DecodeZeroMoveLowMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
Decode a move lower and zero upper instruction as a shuffle mask.
void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
Decode a BLEND immediate mask into a shuffle mask.
bool is128BitVector() const
Return true if this is a 128-bit vector type.
void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl< int > &ShuffleMask)
Decode a MOVLHPS instruction as a v2f64/v4f32 shuffle mask.
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:421
Machine Value Type.
void DecodeVPPERMMask(ArrayRef< uint64_t > RawMask, SmallVectorImpl< int > &ShuffleMask)
Decode a VPPERM mask from a raw array of constants such as from BUILD_VECTOR.
unsigned getScalarSizeInBits() const
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:149
void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
DecodeSHUFPMask - This decodes the shuffle masks for shufp*.
void DecodePSHUFLWMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
Decodes the shuffle masks for pshuflw.
void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
Decode a 128-bit INSERTPS instruction as a v4f32 shuffle mask.
void DecodeZeroExtendMask(MVT SrcScalarVT, MVT DstVT, SmallVectorImpl< int > &Mask)
Decode a zero extension instruction as a shuffle mask.
void DecodeVectorBroadcast(MVT DstVT, SmallVectorImpl< int > &ShuffleMask)
Decodes a broadcast of the first element of a vector.
void DecodeVPERMVMask(ArrayRef< uint64_t > RawMask, SmallVectorImpl< int > &ShuffleMask)
Decode a VPERM W/D/Q/PS/PD mask from a raw array of constants.
void DecodeVPERMIL2PMask(MVT VT, unsigned M2Z, ArrayRef< uint64_t > RawMask, SmallVectorImpl< int > &ShuffleMask)
Decode a VPERMIL2PD/VPERMIL2PS variable mask from a raw array of constants.
void DecodePSRLDQMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
void DecodeVALIGNMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
DecodePSHUFMask - This decodes the shuffle masks for pshufw, pshufd, and vpermilp*.
void DecodeVPERMV3Mask(ArrayRef< uint64_t > RawMask, SmallVectorImpl< int > &ShuffleMask)
Decode a VPERMT2 W/D/Q/PS/PD mask from a raw array of constants.
MVT getScalarType() const
If this is a vector, return the element type, otherwise return this.
void DecodePSWAPMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
Decodes a PSWAPD 3DNow! instruction.
void DecodeUNPCKHMask(MVT VT, SmallVectorImpl< int > &ShuffleMask)
DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd and punpckh*.
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:398
void DecodeVPERM2X128Mask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
void DecodeInsertElementMask(MVT VT, unsigned Idx, unsigned Len, SmallVectorImpl< int > &ShuffleMask)
void decodeVSHUF64x2FamilyMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
Decode a shuffle packed values at 128-bit granularity (SHUFF32x4/SHUFF64x2/SHUFI32x4/SHUFI64x2) immed...
void DecodePSHUFHWMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
Decodes the shuffle masks for pshufhw.
void DecodeVPERMMask(MVT VT, unsigned Imm, SmallVectorImpl< int > &ShuffleMask)
DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
bool is512BitVector() const
Return true if this is a 512-bit vector type.
void DecodePSHUFBMask(ArrayRef< uint64_t > RawMask, SmallVectorImpl< int > &ShuffleMask)
Decode a PSHUFB mask from a raw array of constants such as from BUILD_VECTOR.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
std::underlying_type< E >::type Mask()
Get a bitmask with 1s in all places up to the high-order bit of E&#39;s largest value.
Definition: BitmaskEnum.h:81
void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl< int > &ShuffleMask)
Decode a MOVHLPS instruction as a v2f64/v4f32 shuffle mask.