AArch64ExpandImm.cpp revision 351344
1//===- AArch64ExpandImm.h - AArch64 Immediate Expansion -------------------===//
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//
9// This file implements the AArch64ExpandImm stuff.
10//
11//===----------------------------------------------------------------------===//
12
13#include "AArch64.h"
14#include "AArch64ExpandImm.h"
15#include "MCTargetDesc/AArch64AddressingModes.h"
16
17namespace llvm {
18
19namespace AArch64_IMM {
20
21/// Helper function which extracts the specified 16-bit chunk from a
22/// 64-bit value.
23static uint64_t getChunk(uint64_t Imm, unsigned ChunkIdx) {
24  assert(ChunkIdx < 4 && "Out of range chunk index specified!");
25
26  return (Imm >> (ChunkIdx * 16)) & 0xFFFF;
27}
28
29/// Check whether the given 16-bit chunk replicated to full 64-bit width
30/// can be materialized with an ORR instruction.
31static bool canUseOrr(uint64_t Chunk, uint64_t &Encoding) {
32  Chunk = (Chunk << 48) | (Chunk << 32) | (Chunk << 16) | Chunk;
33
34  return AArch64_AM::processLogicalImmediate(Chunk, 64, Encoding);
35}
36
37/// Check for identical 16-bit chunks within the constant and if so
38/// materialize them with a single ORR instruction. The remaining one or two
39/// 16-bit chunks will be materialized with MOVK instructions.
40///
41/// This allows us to materialize constants like |A|B|A|A| or |A|B|C|A| (order
42/// of the chunks doesn't matter), assuming |A|A|A|A| can be materialized with
43/// an ORR instruction.
44static bool tryToreplicateChunks(uint64_t UImm,
45				 SmallVectorImpl<ImmInsnModel> &Insn) {
46  using CountMap = DenseMap<uint64_t, unsigned>;
47
48  CountMap Counts;
49
50  // Scan the constant and count how often every chunk occurs.
51  for (unsigned Idx = 0; Idx < 4; ++Idx)
52    ++Counts[getChunk(UImm, Idx)];
53
54  // Traverse the chunks to find one which occurs more than once.
55  for (CountMap::const_iterator Chunk = Counts.begin(), End = Counts.end();
56       Chunk != End; ++Chunk) {
57    const uint64_t ChunkVal = Chunk->first;
58    const unsigned Count = Chunk->second;
59
60    uint64_t Encoding = 0;
61
62    // We are looking for chunks which have two or three instances and can be
63    // materialized with an ORR instruction.
64    if ((Count != 2 && Count != 3) || !canUseOrr(ChunkVal, Encoding))
65      continue;
66
67    const bool CountThree = Count == 3;
68
69    Insn.push_back({ AArch64::ORRXri, 0, Encoding });
70
71    unsigned ShiftAmt = 0;
72    uint64_t Imm16 = 0;
73    // Find the first chunk not materialized with the ORR instruction.
74    for (; ShiftAmt < 64; ShiftAmt += 16) {
75      Imm16 = (UImm >> ShiftAmt) & 0xFFFF;
76
77      if (Imm16 != ChunkVal)
78        break;
79    }
80
81    // Create the first MOVK instruction.
82    Insn.push_back({ AArch64::MOVKXi, Imm16,
83		     AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftAmt) });
84
85    // In case we have three instances the whole constant is now materialized
86    // and we can exit.
87    if (CountThree)
88      return true;
89
90    // Find the remaining chunk which needs to be materialized.
91    for (ShiftAmt += 16; ShiftAmt < 64; ShiftAmt += 16) {
92      Imm16 = (UImm >> ShiftAmt) & 0xFFFF;
93
94      if (Imm16 != ChunkVal)
95        break;
96    }
97    Insn.push_back({ AArch64::MOVKXi, Imm16,
98                     AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftAmt) });
99    return true;
100  }
101
102  return false;
103}
104
105/// Check whether this chunk matches the pattern '1...0...'. This pattern
106/// starts a contiguous sequence of ones if we look at the bits from the LSB
107/// towards the MSB.
108static bool isStartChunk(uint64_t Chunk) {
109  if (Chunk == 0 || Chunk == std::numeric_limits<uint64_t>::max())
110    return false;
111
112  return isMask_64(~Chunk);
113}
114
115/// Check whether this chunk matches the pattern '0...1...' This pattern
116/// ends a contiguous sequence of ones if we look at the bits from the LSB
117/// towards the MSB.
118static bool isEndChunk(uint64_t Chunk) {
119  if (Chunk == 0 || Chunk == std::numeric_limits<uint64_t>::max())
120    return false;
121
122  return isMask_64(Chunk);
123}
124
125/// Clear or set all bits in the chunk at the given index.
126static uint64_t updateImm(uint64_t Imm, unsigned Idx, bool Clear) {
127  const uint64_t Mask = 0xFFFF;
128
129  if (Clear)
130    // Clear chunk in the immediate.
131    Imm &= ~(Mask << (Idx * 16));
132  else
133    // Set all bits in the immediate for the particular chunk.
134    Imm |= Mask << (Idx * 16);
135
136  return Imm;
137}
138
139/// Check whether the constant contains a sequence of contiguous ones,
140/// which might be interrupted by one or two chunks. If so, materialize the
141/// sequence of contiguous ones with an ORR instruction.
142/// Materialize the chunks which are either interrupting the sequence or outside
143/// of the sequence with a MOVK instruction.
144///
145/// Assuming S is a chunk which starts the sequence (1...0...), E is a chunk
146/// which ends the sequence (0...1...). Then we are looking for constants which
147/// contain at least one S and E chunk.
148/// E.g. |E|A|B|S|, |A|E|B|S| or |A|B|E|S|.
149///
150/// We are also looking for constants like |S|A|B|E| where the contiguous
151/// sequence of ones wraps around the MSB into the LSB.
152static bool trySequenceOfOnes(uint64_t UImm,
153                              SmallVectorImpl<ImmInsnModel> &Insn) {
154  const int NotSet = -1;
155  const uint64_t Mask = 0xFFFF;
156
157  int StartIdx = NotSet;
158  int EndIdx = NotSet;
159  // Try to find the chunks which start/end a contiguous sequence of ones.
160  for (int Idx = 0; Idx < 4; ++Idx) {
161    int64_t Chunk = getChunk(UImm, Idx);
162    // Sign extend the 16-bit chunk to 64-bit.
163    Chunk = (Chunk << 48) >> 48;
164
165    if (isStartChunk(Chunk))
166      StartIdx = Idx;
167    else if (isEndChunk(Chunk))
168      EndIdx = Idx;
169  }
170
171  // Early exit in case we can't find a start/end chunk.
172  if (StartIdx == NotSet || EndIdx == NotSet)
173    return false;
174
175  // Outside of the contiguous sequence of ones everything needs to be zero.
176  uint64_t Outside = 0;
177  // Chunks between the start and end chunk need to have all their bits set.
178  uint64_t Inside = Mask;
179
180  // If our contiguous sequence of ones wraps around from the MSB into the LSB,
181  // just swap indices and pretend we are materializing a contiguous sequence
182  // of zeros surrounded by a contiguous sequence of ones.
183  if (StartIdx > EndIdx) {
184    std::swap(StartIdx, EndIdx);
185    std::swap(Outside, Inside);
186  }
187
188  uint64_t OrrImm = UImm;
189  int FirstMovkIdx = NotSet;
190  int SecondMovkIdx = NotSet;
191
192  // Find out which chunks we need to patch up to obtain a contiguous sequence
193  // of ones.
194  for (int Idx = 0; Idx < 4; ++Idx) {
195    const uint64_t Chunk = getChunk(UImm, Idx);
196
197    // Check whether we are looking at a chunk which is not part of the
198    // contiguous sequence of ones.
199    if ((Idx < StartIdx || EndIdx < Idx) && Chunk != Outside) {
200      OrrImm = updateImm(OrrImm, Idx, Outside == 0);
201
202      // Remember the index we need to patch.
203      if (FirstMovkIdx == NotSet)
204        FirstMovkIdx = Idx;
205      else
206        SecondMovkIdx = Idx;
207
208      // Check whether we are looking a chunk which is part of the contiguous
209      // sequence of ones.
210    } else if (Idx > StartIdx && Idx < EndIdx && Chunk != Inside) {
211      OrrImm = updateImm(OrrImm, Idx, Inside != Mask);
212
213      // Remember the index we need to patch.
214      if (FirstMovkIdx == NotSet)
215        FirstMovkIdx = Idx;
216      else
217        SecondMovkIdx = Idx;
218    }
219  }
220  assert(FirstMovkIdx != NotSet && "Constant materializable with single ORR!");
221
222  // Create the ORR-immediate instruction.
223  uint64_t Encoding = 0;
224  AArch64_AM::processLogicalImmediate(OrrImm, 64, Encoding);
225  Insn.push_back({ AArch64::ORRXri, 0, Encoding });
226
227  const bool SingleMovk = SecondMovkIdx == NotSet;
228  Insn.push_back({ AArch64::MOVKXi, getChunk(UImm, FirstMovkIdx),
229                   AArch64_AM::getShifterImm(AArch64_AM::LSL,
230                                             FirstMovkIdx * 16) });
231
232  // Early exit in case we only need to emit a single MOVK instruction.
233  if (SingleMovk)
234    return true;
235
236  // Create the second MOVK instruction.
237  Insn.push_back({ AArch64::MOVKXi, getChunk(UImm, SecondMovkIdx),
238	           AArch64_AM::getShifterImm(AArch64_AM::LSL,
239                                             SecondMovkIdx * 16) });
240
241  return true;
242}
243
244/// \brief Expand a MOVi32imm or MOVi64imm pseudo instruction to a
245/// MOVZ or MOVN of width BitSize followed by up to 3 MOVK instructions.
246static inline void expandMOVImmSimple(uint64_t Imm, unsigned BitSize,
247				      unsigned OneChunks, unsigned ZeroChunks,
248				      SmallVectorImpl<ImmInsnModel> &Insn) {
249  const unsigned Mask = 0xFFFF;
250
251  // Use a MOVZ or MOVN instruction to set the high bits, followed by one or
252  // more MOVK instructions to insert additional 16-bit portions into the
253  // lower bits.
254  bool isNeg = false;
255
256  // Use MOVN to materialize the high bits if we have more all one chunks
257  // than all zero chunks.
258  if (OneChunks > ZeroChunks) {
259    isNeg = true;
260    Imm = ~Imm;
261  }
262
263  unsigned FirstOpc;
264  if (BitSize == 32) {
265    Imm &= (1LL << 32) - 1;
266    FirstOpc = (isNeg ? AArch64::MOVNWi : AArch64::MOVZWi);
267  } else {
268    FirstOpc = (isNeg ? AArch64::MOVNXi : AArch64::MOVZXi);
269  }
270  unsigned Shift = 0;     // LSL amount for high bits with MOVZ/MOVN
271  unsigned LastShift = 0; // LSL amount for last MOVK
272  if (Imm != 0) {
273    unsigned LZ = countLeadingZeros(Imm);
274    unsigned TZ = countTrailingZeros(Imm);
275    Shift = (TZ / 16) * 16;
276    LastShift = ((63 - LZ) / 16) * 16;
277  }
278  unsigned Imm16 = (Imm >> Shift) & Mask;
279
280  Insn.push_back({ FirstOpc, Imm16,
281                   AArch64_AM::getShifterImm(AArch64_AM::LSL, Shift) });
282
283  if (Shift == LastShift)
284    return;
285
286  // If a MOVN was used for the high bits of a negative value, flip the rest
287  // of the bits back for use with MOVK.
288  if (isNeg)
289    Imm = ~Imm;
290
291  unsigned Opc = (BitSize == 32 ? AArch64::MOVKWi : AArch64::MOVKXi);
292  while (Shift < LastShift) {
293    Shift += 16;
294    Imm16 = (Imm >> Shift) & Mask;
295    if (Imm16 == (isNeg ? Mask : 0))
296      continue; // This 16-bit portion is already set correctly.
297
298    Insn.push_back({ Opc, Imm16,
299                     AArch64_AM::getShifterImm(AArch64_AM::LSL, Shift) });
300  }
301}
302
303/// Expand a MOVi32imm or MOVi64imm pseudo instruction to one or more
304/// real move-immediate instructions to synthesize the immediate.
305void expandMOVImm(uint64_t Imm, unsigned BitSize,
306		  SmallVectorImpl<ImmInsnModel> &Insn) {
307  const unsigned Mask = 0xFFFF;
308
309  // Scan the immediate and count the number of 16-bit chunks which are either
310  // all ones or all zeros.
311  unsigned OneChunks = 0;
312  unsigned ZeroChunks = 0;
313  for (unsigned Shift = 0; Shift < BitSize; Shift += 16) {
314    const unsigned Chunk = (Imm >> Shift) & Mask;
315    if (Chunk == Mask)
316      OneChunks++;
317    else if (Chunk == 0)
318      ZeroChunks++;
319  }
320
321  // Prefer MOVZ/MOVN over ORR because of the rules for the "mov" alias.
322  if ((BitSize / 16) - OneChunks <= 1 || (BitSize / 16) - ZeroChunks <= 1) {
323    expandMOVImmSimple(Imm, BitSize, OneChunks, ZeroChunks, Insn);
324    return;
325  }
326
327  // Try a single ORR.
328  uint64_t UImm = Imm << (64 - BitSize) >> (64 - BitSize);
329  uint64_t Encoding;
330  if (AArch64_AM::processLogicalImmediate(UImm, BitSize, Encoding)) {
331    unsigned Opc = (BitSize == 32 ? AArch64::ORRWri : AArch64::ORRXri);
332    Insn.push_back({ Opc, 0, Encoding });
333    return;
334  }
335
336  // One to up three instruction sequences.
337  //
338  // Prefer MOVZ/MOVN followed by MOVK; it's more readable, and possibly the
339  // fastest sequence with fast literal generation.
340  if (OneChunks >= (BitSize / 16) - 2 || ZeroChunks >= (BitSize / 16) - 2) {
341    expandMOVImmSimple(Imm, BitSize, OneChunks, ZeroChunks, Insn);
342    return;
343  }
344
345  assert(BitSize == 64 && "All 32-bit immediates can be expanded with a"
346                          "MOVZ/MOVK pair");
347
348  // Try other two-instruction sequences.
349
350  // 64-bit ORR followed by MOVK.
351  // We try to construct the ORR immediate in three different ways: either we
352  // zero out the chunk which will be replaced, we fill the chunk which will
353  // be replaced with ones, or we take the bit pattern from the other half of
354  // the 64-bit immediate. This is comprehensive because of the way ORR
355  // immediates are constructed.
356  for (unsigned Shift = 0; Shift < BitSize; Shift += 16) {
357    uint64_t ShiftedMask = (0xFFFFULL << Shift);
358    uint64_t ZeroChunk = UImm & ~ShiftedMask;
359    uint64_t OneChunk = UImm | ShiftedMask;
360    uint64_t RotatedImm = (UImm << 32) | (UImm >> 32);
361    uint64_t ReplicateChunk = ZeroChunk | (RotatedImm & ShiftedMask);
362    if (AArch64_AM::processLogicalImmediate(ZeroChunk, BitSize, Encoding) ||
363        AArch64_AM::processLogicalImmediate(OneChunk, BitSize, Encoding) ||
364        AArch64_AM::processLogicalImmediate(ReplicateChunk, BitSize,
365                                            Encoding)) {
366      // Create the ORR-immediate instruction.
367      Insn.push_back({ AArch64::ORRXri, 0, Encoding });
368
369      // Create the MOVK instruction.
370      const unsigned Imm16 = getChunk(UImm, Shift / 16);
371      Insn.push_back({ AArch64::MOVKXi, Imm16,
372		       AArch64_AM::getShifterImm(AArch64_AM::LSL, Shift) });
373      return;
374    }
375  }
376
377  // FIXME: Add more two-instruction sequences.
378
379  // Three instruction sequences.
380  //
381  // Prefer MOVZ/MOVN followed by two MOVK; it's more readable, and possibly
382  // the fastest sequence with fast literal generation. (If neither MOVK is
383  // part of a fast literal generation pair, it could be slower than the
384  // four-instruction sequence, but we won't worry about that for now.)
385  if (OneChunks || ZeroChunks) {
386    expandMOVImmSimple(Imm, BitSize, OneChunks, ZeroChunks, Insn);
387    return;
388  }
389
390  // Check for identical 16-bit chunks within the constant and if so materialize
391  // them with a single ORR instruction. The remaining one or two 16-bit chunks
392  // will be materialized with MOVK instructions.
393  if (BitSize == 64 && tryToreplicateChunks(UImm, Insn))
394    return;
395
396  // Check whether the constant contains a sequence of contiguous ones, which
397  // might be interrupted by one or two chunks. If so, materialize the sequence
398  // of contiguous ones with an ORR instruction. Materialize the chunks which
399  // are either interrupting the sequence or outside of the sequence with a
400  // MOVK instruction.
401  if (BitSize == 64 && trySequenceOfOnes(UImm, Insn))
402    return;
403
404  // We found no possible two or three instruction sequence; use the general
405  // four-instruction sequence.
406  expandMOVImmSimple(Imm, BitSize, OneChunks, ZeroChunks, Insn);
407}
408
409} // end namespace AArch64_AM
410
411} // end namespace llvm
412