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/IR/Constants.h"
17 #include "llvm/CodeGen/MachineValueType.h"
19 //===----------------------------------------------------------------------===//
20 // Vector Mask Decoding
21 //===----------------------------------------------------------------------===//
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);
32 // Decode the immediate.
33 unsigned ZMask = Imm & 15;
34 unsigned CountD = (Imm >> 4) & 3;
35 unsigned CountS = (Imm >> 6) & 3;
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;
49 void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
50 for (unsigned i = NElts / 2; i != NElts; ++i)
51 ShuffleMask.push_back(NElts + i);
53 for (unsigned i = NElts / 2; i != NElts; ++i)
54 ShuffleMask.push_back(i);
58 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
59 for (unsigned i = 0; i != NElts / 2; ++i)
60 ShuffleMask.push_back(i);
62 for (unsigned i = 0; i != NElts / 2; ++i)
63 ShuffleMask.push_back(NElts + i);
66 void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
67 unsigned NumElts = VT.getVectorNumElements();
68 for (int i = 0, e = NumElts / 2; i < e; ++i) {
69 ShuffleMask.push_back(2 * i);
70 ShuffleMask.push_back(2 * i);
74 void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
75 unsigned NumElts = VT.getVectorNumElements();
76 for (int i = 0, e = NumElts / 2; i < e; ++i) {
77 ShuffleMask.push_back(2 * i + 1);
78 ShuffleMask.push_back(2 * i + 1);
82 void DecodeMOVDDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
83 unsigned VectorSizeInBits = VT.getSizeInBits();
84 unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
85 unsigned NumElts = VT.getVectorNumElements();
86 unsigned NumLanes = VectorSizeInBits / 128;
87 unsigned NumLaneElts = NumElts / NumLanes;
88 unsigned NumLaneSubElts = 64 / ScalarSizeInBits;
90 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
91 for (unsigned i = 0; i < NumLaneElts; i += NumLaneSubElts)
92 for (unsigned s = 0; s != NumLaneSubElts; s++)
93 ShuffleMask.push_back(l + s);
96 void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
97 unsigned VectorSizeInBits = VT.getSizeInBits();
98 unsigned NumElts = VectorSizeInBits / 8;
99 unsigned NumLanes = VectorSizeInBits / 128;
100 unsigned NumLaneElts = NumElts / NumLanes;
102 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
103 for (unsigned i = 0; i < NumLaneElts; ++i) {
104 int M = SM_SentinelZero;
105 if (i >= Imm) M = i - Imm + l;
106 ShuffleMask.push_back(M);
110 void DecodePSRLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
111 unsigned VectorSizeInBits = VT.getSizeInBits();
112 unsigned NumElts = VectorSizeInBits / 8;
113 unsigned NumLanes = VectorSizeInBits / 128;
114 unsigned NumLaneElts = NumElts / NumLanes;
116 for (unsigned l = 0; l < NumElts; l += NumLaneElts)
117 for (unsigned i = 0; i < NumLaneElts; ++i) {
118 unsigned Base = i + Imm;
120 if (Base >= NumLaneElts) M = SM_SentinelZero;
121 ShuffleMask.push_back(M);
125 void DecodePALIGNRMask(MVT VT, unsigned Imm,
126 SmallVectorImpl<int> &ShuffleMask) {
127 unsigned NumElts = VT.getVectorNumElements();
128 unsigned Offset = Imm * (VT.getVectorElementType().getSizeInBits() / 8);
130 unsigned NumLanes = VT.getSizeInBits() / 128;
131 unsigned NumLaneElts = NumElts / NumLanes;
133 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
134 for (unsigned i = 0; i != NumLaneElts; ++i) {
135 unsigned Base = i + Offset;
136 // if i+offset is out of this lane then we actually need the other source
137 if (Base >= NumLaneElts) Base += NumElts - NumLaneElts;
138 ShuffleMask.push_back(Base + l);
143 /// DecodePSHUFMask - This decodes the shuffle masks for pshufw, pshufd, and vpermilp*.
144 /// VT indicates the type of the vector allowing it to handle different
145 /// datatypes and vector widths.
146 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
147 unsigned NumElts = VT.getVectorNumElements();
149 unsigned NumLanes = VT.getSizeInBits() / 128;
150 if (NumLanes == 0) NumLanes = 1; // Handle MMX
151 unsigned NumLaneElts = NumElts / NumLanes;
153 unsigned NewImm = Imm;
154 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
155 for (unsigned i = 0; i != NumLaneElts; ++i) {
156 ShuffleMask.push_back(NewImm % NumLaneElts + l);
157 NewImm /= NumLaneElts;
159 if (NumLaneElts == 4) NewImm = Imm; // reload imm
163 void DecodePSHUFHWMask(MVT VT, unsigned Imm,
164 SmallVectorImpl<int> &ShuffleMask) {
165 unsigned NumElts = VT.getVectorNumElements();
167 for (unsigned l = 0; l != NumElts; l += 8) {
168 unsigned NewImm = Imm;
169 for (unsigned i = 0, e = 4; i != e; ++i) {
170 ShuffleMask.push_back(l + i);
172 for (unsigned i = 4, e = 8; i != e; ++i) {
173 ShuffleMask.push_back(l + 4 + (NewImm & 3));
179 void DecodePSHUFLWMask(MVT VT, unsigned Imm,
180 SmallVectorImpl<int> &ShuffleMask) {
181 unsigned NumElts = VT.getVectorNumElements();
183 for (unsigned l = 0; l != NumElts; l += 8) {
184 unsigned NewImm = Imm;
185 for (unsigned i = 0, e = 4; i != e; ++i) {
186 ShuffleMask.push_back(l + (NewImm & 3));
189 for (unsigned i = 4, e = 8; i != e; ++i) {
190 ShuffleMask.push_back(l + i);
195 void DecodePSWAPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
196 unsigned NumElts = VT.getVectorNumElements();
197 unsigned NumHalfElts = NumElts / 2;
199 for (unsigned l = 0; l != NumHalfElts; ++l)
200 ShuffleMask.push_back(l + NumHalfElts);
201 for (unsigned h = 0; h != NumHalfElts; ++h)
202 ShuffleMask.push_back(h);
205 /// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates
206 /// the type of the vector allowing it to handle different datatypes and vector
208 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
209 unsigned NumElts = VT.getVectorNumElements();
211 unsigned NumLanes = VT.getSizeInBits() / 128;
212 unsigned NumLaneElts = NumElts / NumLanes;
214 unsigned NewImm = Imm;
215 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
216 // each half of a lane comes from different source
217 for (unsigned s = 0; s != NumElts * 2; s += NumElts) {
218 for (unsigned i = 0; i != NumLaneElts / 2; ++i) {
219 ShuffleMask.push_back(NewImm % NumLaneElts + s + l);
220 NewImm /= NumLaneElts;
223 if (NumLaneElts == 4) NewImm = Imm; // reload imm
227 /// DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd
228 /// and punpckh*. VT indicates the type of the vector allowing it to handle
229 /// different datatypes and vector widths.
230 void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
231 unsigned NumElts = VT.getVectorNumElements();
233 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
234 // independently on 128-bit lanes.
235 unsigned NumLanes = VT.getSizeInBits() / 128;
236 if (NumLanes == 0) NumLanes = 1; // Handle MMX
237 unsigned NumLaneElts = NumElts / NumLanes;
239 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
240 for (unsigned i = l + NumLaneElts / 2, e = l + NumLaneElts; i != e; ++i) {
241 ShuffleMask.push_back(i); // Reads from dest/src1
242 ShuffleMask.push_back(i + NumElts); // Reads from src/src2
247 /// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
248 /// and punpckl*. VT indicates the type of the vector allowing it to handle
249 /// different datatypes and vector widths.
250 void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
251 unsigned NumElts = VT.getVectorNumElements();
253 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
254 // independently on 128-bit lanes.
255 unsigned NumLanes = VT.getSizeInBits() / 128;
256 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
257 unsigned NumLaneElts = NumElts / NumLanes;
259 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
260 for (unsigned i = l, e = l + NumLaneElts / 2; i != e; ++i) {
261 ShuffleMask.push_back(i); // Reads from dest/src1
262 ShuffleMask.push_back(i + NumElts); // Reads from src/src2
267 void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
268 SmallVectorImpl<int> &ShuffleMask) {
269 unsigned HalfSize = VT.getVectorNumElements() / 2;
271 for (unsigned l = 0; l != 2; ++l) {
272 unsigned HalfMask = Imm >> (l * 4);
273 unsigned HalfBegin = (HalfMask & 0x3) * HalfSize;
274 for (unsigned i = HalfBegin, e = HalfBegin + HalfSize; i != e; ++i)
275 ShuffleMask.push_back(HalfMask & 8 ? SM_SentinelZero : (int)i);
279 void DecodePSHUFBMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
280 Type *MaskTy = C->getType();
281 // It is not an error for the PSHUFB mask to not be a vector of i8 because the
282 // constant pool uniques constants by their bit representation.
283 // e.g. the following take up the same space in the constant pool:
284 // i128 -170141183420855150465331762880109871104
286 // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160>
288 // <4 x i32> <i32 -2147483648, i32 -2147483648,
289 // i32 -2147483648, i32 -2147483648>
291 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
293 if (MaskTySize != 128 && MaskTySize != 256) // FIXME: Add support for AVX-512.
296 // This is a straightforward byte vector.
297 if (MaskTy->isVectorTy() && MaskTy->getVectorElementType()->isIntegerTy(8)) {
298 int NumElements = MaskTy->getVectorNumElements();
299 ShuffleMask.reserve(NumElements);
301 for (int i = 0; i < NumElements; ++i) {
302 // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte
303 // lane of the vector we're inside.
304 int Base = i < 16 ? 0 : 16;
305 Constant *COp = C->getAggregateElement(i);
309 } else if (isa<UndefValue>(COp)) {
310 ShuffleMask.push_back(SM_SentinelUndef);
313 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
314 // If the high bit (7) of the byte is set, the element is zeroed.
315 if (Element & (1 << 7))
316 ShuffleMask.push_back(SM_SentinelZero);
318 // Only the least significant 4 bits of the byte are used.
319 int Index = Base + (Element & 0xf);
320 ShuffleMask.push_back(Index);
324 // TODO: Handle funny-looking vectors too.
327 void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask,
328 SmallVectorImpl<int> &ShuffleMask) {
329 for (int i = 0, e = RawMask.size(); i < e; ++i) {
330 uint64_t M = RawMask[i];
331 if (M == (uint64_t)SM_SentinelUndef) {
332 ShuffleMask.push_back(M);
335 // For AVX vectors with 32 bytes the base of the shuffle is the half of
336 // the vector we're inside.
337 int Base = i < 16 ? 0 : 16;
338 // If the high bit (7) of the byte is set, the element is zeroed.
340 ShuffleMask.push_back(SM_SentinelZero);
342 // Only the least significant 4 bits of the byte are used.
343 int Index = Base + (M & 0xf);
344 ShuffleMask.push_back(Index);
349 void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
350 int ElementBits = VT.getScalarSizeInBits();
351 int NumElements = VT.getVectorNumElements();
352 for (int i = 0; i < NumElements; ++i) {
353 // If there are more than 8 elements in the vector, then any immediate blend
354 // mask applies to each 128-bit lane. There can never be more than
355 // 8 elements in a 128-bit lane with an immediate blend.
356 int Bit = NumElements > 8 ? i % (128 / ElementBits) : i;
358 "Immediate blends only operate over 8 elements at a time!");
359 ShuffleMask.push_back(((Imm >> Bit) & 1) ? NumElements + i : i);
363 /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
364 /// No VT provided since it only works on 256-bit, 4 element vectors.
365 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
366 for (unsigned i = 0; i != 4; ++i) {
367 ShuffleMask.push_back((Imm >> (2 * i)) & 3);
371 void DecodeVPERMILPMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
372 Type *MaskTy = C->getType();
373 assert(MaskTy->isVectorTy() && "Expected a vector constant mask!");
374 assert(MaskTy->getVectorElementType()->isIntegerTy() &&
375 "Expected integer constant mask elements!");
376 int ElementBits = MaskTy->getScalarSizeInBits();
377 int NumElements = MaskTy->getVectorNumElements();
378 assert((NumElements == 2 || NumElements == 4 || NumElements == 8) &&
379 "Unexpected number of vector elements.");
380 ShuffleMask.reserve(NumElements);
381 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
382 assert((unsigned)NumElements == CDS->getNumElements() &&
383 "Constant mask has a different number of elements!");
385 for (int i = 0; i < NumElements; ++i) {
386 int Base = (i * ElementBits / 128) * (128 / ElementBits);
387 uint64_t Element = CDS->getElementAsInteger(i);
388 // Only the least significant 2 bits of the integer are used.
389 int Index = Base + (Element & 0x3);
390 ShuffleMask.push_back(Index);
392 } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
393 assert((unsigned)NumElements == C->getNumOperands() &&
394 "Constant mask has a different number of elements!");
396 for (int i = 0; i < NumElements; ++i) {
397 int Base = (i * ElementBits / 128) * (128 / ElementBits);
398 Constant *COp = CV->getOperand(i);
399 if (isa<UndefValue>(COp)) {
400 ShuffleMask.push_back(SM_SentinelUndef);
403 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
404 // Only the least significant 2 bits of the integer are used.
405 int Index = Base + (Element & 0x3);
406 ShuffleMask.push_back(Index);
411 void DecodeZeroExtendMask(MVT SrcVT, MVT DstVT, SmallVectorImpl<int> &Mask) {
412 unsigned NumDstElts = DstVT.getVectorNumElements();
413 unsigned SrcScalarBits = SrcVT.getScalarSizeInBits();
414 unsigned DstScalarBits = DstVT.getScalarSizeInBits();
415 unsigned Scale = DstScalarBits / SrcScalarBits;
416 assert(SrcScalarBits < DstScalarBits &&
417 "Expected zero extension mask to increase scalar size");
418 assert(SrcVT.getVectorNumElements() >= NumDstElts &&
419 "Too many zero extension lanes");
421 for (unsigned i = 0; i != NumDstElts; i++) {
423 for (unsigned j = 1; j != Scale; j++)
424 Mask.push_back(SM_SentinelZero);
428 void DecodeZeroMoveLowMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
429 unsigned NumElts = VT.getVectorNumElements();
430 ShuffleMask.push_back(0);
431 for (unsigned i = 1; i < NumElts; i++)
432 ShuffleMask.push_back(SM_SentinelZero);
435 void DecodeScalarMoveMask(MVT VT, bool IsLoad, SmallVectorImpl<int> &Mask) {
436 // First element comes from the first element of second source.
437 // Remaining elements: Load zero extends / Move copies from first source.
438 unsigned NumElts = VT.getVectorNumElements();
439 Mask.push_back(NumElts);
440 for (unsigned i = 1; i < NumElts; i++)
441 Mask.push_back(IsLoad ? static_cast<int>(SM_SentinelZero) : i);
444 void DecodeEXTRQIMask(int Len, int Idx,
445 SmallVectorImpl<int> &ShuffleMask) {
446 // Only the bottom 6 bits are valid for each immediate.
450 // We can only decode this bit extraction instruction as a shuffle if both the
451 // length and index work with whole bytes.
452 if (0 != (Len % 8) || 0 != (Idx % 8))
455 // A length of zero is equivalent to a bit length of 64.
459 // If the length + index exceeds the bottom 64 bits the result is undefined.
460 if ((Len + Idx) > 64) {
461 ShuffleMask.append(16, SM_SentinelUndef);
465 // Convert index and index to work with bytes.
469 // EXTRQ: Extract Len bytes starting from Idx. Zero pad the remaining bytes
470 // of the lower 64-bits. The upper 64-bits are undefined.
471 for (int i = 0; i != Len; ++i)
472 ShuffleMask.push_back(i + Idx);
473 for (int i = Len; i != 8; ++i)
474 ShuffleMask.push_back(SM_SentinelZero);
475 for (int i = 8; i != 16; ++i)
476 ShuffleMask.push_back(SM_SentinelUndef);
479 void DecodeINSERTQIMask(int Len, int Idx,
480 SmallVectorImpl<int> &ShuffleMask) {
481 // Only the bottom 6 bits are valid for each immediate.
485 // We can only decode this bit insertion instruction as a shuffle if both the
486 // length and index work with whole bytes.
487 if (0 != (Len % 8) || 0 != (Idx % 8))
490 // A length of zero is equivalent to a bit length of 64.
494 // If the length + index exceeds the bottom 64 bits the result is undefined.
495 if ((Len + Idx) > 64) {
496 ShuffleMask.append(16, SM_SentinelUndef);
500 // Convert index and index to work with bytes.
504 // INSERTQ: Extract lowest Len bytes from lower half of second source and
505 // insert over first source starting at Idx byte. The upper 64-bits are
507 for (int i = 0; i != Idx; ++i)
508 ShuffleMask.push_back(i);
509 for (int i = 0; i != Len; ++i)
510 ShuffleMask.push_back(i + 16);
511 for (int i = Idx + Len; i != 8; ++i)
512 ShuffleMask.push_back(i);
513 for (int i = 8; i != 16; ++i)
514 ShuffleMask.push_back(SM_SentinelUndef);
517 void DecodeVPERMVMask(ArrayRef<uint64_t> RawMask,
518 SmallVectorImpl<int> &ShuffleMask) {
519 for (int i = 0, e = RawMask.size(); i < e; ++i) {
520 uint64_t M = RawMask[i];
521 ShuffleMask.push_back((int)M);
525 void DecodeVPERMV3Mask(ArrayRef<uint64_t> RawMask,
526 SmallVectorImpl<int> &ShuffleMask) {
527 for (int i = 0, e = RawMask.size(); i < e; ++i) {
528 uint64_t M = RawMask[i];
529 ShuffleMask.push_back((int)M);
533 void DecodeVPERMVMask(const Constant *C, MVT VT,
534 SmallVectorImpl<int> &ShuffleMask) {
535 Type *MaskTy = C->getType();
536 if (MaskTy->isVectorTy()) {
537 unsigned NumElements = MaskTy->getVectorNumElements();
538 if (NumElements == VT.getVectorNumElements()) {
539 for (unsigned i = 0; i < NumElements; ++i) {
540 Constant *COp = C->getAggregateElement(i);
541 if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) {
545 if (isa<UndefValue>(COp))
546 ShuffleMask.push_back(SM_SentinelUndef);
548 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
549 Element &= (1 << NumElements) - 1;
550 ShuffleMask.push_back(Element);
556 // Scalar value; just broadcast it
557 if (!isa<ConstantInt>(C))
559 uint64_t Element = cast<ConstantInt>(C)->getZExtValue();
560 int NumElements = VT.getVectorNumElements();
561 Element &= (1 << NumElements) - 1;
562 for (int i = 0; i < NumElements; ++i)
563 ShuffleMask.push_back(Element);
566 void DecodeVPERMV3Mask(const Constant *C, MVT VT,
567 SmallVectorImpl<int> &ShuffleMask) {
568 Type *MaskTy = C->getType();
569 unsigned NumElements = MaskTy->getVectorNumElements();
570 if (NumElements == VT.getVectorNumElements()) {
571 for (unsigned i = 0; i < NumElements; ++i) {
572 Constant *COp = C->getAggregateElement(i);
577 if (isa<UndefValue>(COp))
578 ShuffleMask.push_back(SM_SentinelUndef);
580 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
581 Element &= (1 << NumElements*2) - 1;
582 ShuffleMask.push_back(Element);