1 //===- RISCVMatInt.cpp - Immediate materialisation -------------*- C++ -*--===//
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 #include "RISCVMatInt.h"
10 #include "MCTargetDesc/RISCVMCTargetDesc.h"
11 #include "llvm/ADT/APInt.h"
12 #include "llvm/Support/MathExtras.h"
13 using namespace llvm;
14 
15 static int getInstSeqCost(RISCVMatInt::InstSeq &Res, bool HasRVC) {
16   if (!HasRVC)
17     return Res.size();
18 
19   int Cost = 0;
20   for (auto Instr : Res) {
21     // Assume instructions that aren't listed aren't compressible.
22     bool Compressed = false;
23     switch (Instr.Opc) {
24     case RISCV::SLLI:
25     case RISCV::SRLI:
26       Compressed = true;
27       break;
28     case RISCV::ADDI:
29     case RISCV::ADDIW:
30     case RISCV::LUI:
31       Compressed = isInt<6>(Instr.Imm);
32       break;
33     }
34     // Two RVC instructions take the same space as one RVI instruction, but
35     // can take longer to execute than the single RVI instruction. Thus, we
36     // consider that two RVC instruction are slightly more costly than one
37     // RVI instruction. For longer sequences of RVC instructions the space
38     // savings can be worth it, though. The costs below try to model that.
39     if (!Compressed)
40       Cost += 100; // Baseline cost of one RVI instruction: 100%.
41     else
42       Cost += 70; // 70% cost of baseline.
43   }
44   return Cost;
45 }
46 
47 // Recursively generate a sequence for materializing an integer.
48 static void generateInstSeqImpl(int64_t Val,
49                                 const FeatureBitset &ActiveFeatures,
50                                 RISCVMatInt::InstSeq &Res) {
51   bool IsRV64 = ActiveFeatures[RISCV::Feature64Bit];
52 
53   if (isInt<32>(Val)) {
54     // Depending on the active bits in the immediate Value v, the following
55     // instruction sequences are emitted:
56     //
57     // v == 0                        : ADDI
58     // v[0,12) != 0 && v[12,32) == 0 : ADDI
59     // v[0,12) == 0 && v[12,32) != 0 : LUI
60     // v[0,32) != 0                  : LUI+ADDI(W)
61     int64_t Hi20 = ((Val + 0x800) >> 12) & 0xFFFFF;
62     int64_t Lo12 = SignExtend64<12>(Val);
63 
64     if (Hi20)
65       Res.push_back(RISCVMatInt::Inst(RISCV::LUI, Hi20));
66 
67     if (Lo12 || Hi20 == 0) {
68       unsigned AddiOpc = (IsRV64 && Hi20) ? RISCV::ADDIW : RISCV::ADDI;
69       Res.push_back(RISCVMatInt::Inst(AddiOpc, Lo12));
70     }
71     return;
72   }
73 
74   assert(IsRV64 && "Can't emit >32-bit imm for non-RV64 target");
75 
76   // Use BSETI for a single bit.
77   if (ActiveFeatures[RISCV::FeatureStdExtZbs] && isPowerOf2_64(Val)) {
78     Res.push_back(RISCVMatInt::Inst(RISCV::BSETI, Log2_64(Val)));
79     return;
80   }
81 
82   // In the worst case, for a full 64-bit constant, a sequence of 8 instructions
83   // (i.e., LUI+ADDIW+SLLI+ADDI+SLLI+ADDI+SLLI+ADDI) has to be emitted. Note
84   // that the first two instructions (LUI+ADDIW) can contribute up to 32 bits
85   // while the following ADDI instructions contribute up to 12 bits each.
86   //
87   // On the first glance, implementing this seems to be possible by simply
88   // emitting the most significant 32 bits (LUI+ADDIW) followed by as many left
89   // shift (SLLI) and immediate additions (ADDI) as needed. However, due to the
90   // fact that ADDI performs a sign extended addition, doing it like that would
91   // only be possible when at most 11 bits of the ADDI instructions are used.
92   // Using all 12 bits of the ADDI instructions, like done by GAS, actually
93   // requires that the constant is processed starting with the least significant
94   // bit.
95   //
96   // In the following, constants are processed from LSB to MSB but instruction
97   // emission is performed from MSB to LSB by recursively calling
98   // generateInstSeq. In each recursion, first the lowest 12 bits are removed
99   // from the constant and the optimal shift amount, which can be greater than
100   // 12 bits if the constant is sparse, is determined. Then, the shifted
101   // remaining constant is processed recursively and gets emitted as soon as it
102   // fits into 32 bits. The emission of the shifts and additions is subsequently
103   // performed when the recursion returns.
104 
105   int64_t Lo12 = SignExtend64<12>(Val);
106   Val = (uint64_t)Val - (uint64_t)Lo12;
107 
108   int ShiftAmount = 0;
109   bool Unsigned = false;
110 
111   // Val might now be valid for LUI without needing a shift.
112   if (!isInt<32>(Val)) {
113     ShiftAmount = findFirstSet((uint64_t)Val);
114     Val >>= ShiftAmount;
115 
116     // If the remaining bits don't fit in 12 bits, we might be able to reduce the
117     // shift amount in order to use LUI which will zero the lower 12 bits.
118     if (ShiftAmount > 12 && !isInt<12>(Val)) {
119       if (isInt<32>((uint64_t)Val << 12)) {
120         // Reduce the shift amount and add zeros to the LSBs so it will match LUI.
121         ShiftAmount -= 12;
122         Val = (uint64_t)Val << 12;
123       } else if (isUInt<32>((uint64_t)Val << 12) &&
124                  ActiveFeatures[RISCV::FeatureStdExtZba]) {
125         // Reduce the shift amount and add zeros to the LSBs so it will match
126         // LUI, then shift left with SLLI.UW to clear the upper 32 set bits.
127         ShiftAmount -= 12;
128         Val = ((uint64_t)Val << 12) | (0xffffffffull << 32);
129         Unsigned = true;
130       }
131     }
132 
133     // Try to use SLLI_UW for Val when it is uint32 but not int32.
134     if (isUInt<32>((uint64_t)Val) && !isInt<32>((uint64_t)Val) &&
135         ActiveFeatures[RISCV::FeatureStdExtZba]) {
136       // Use LUI+ADDI or LUI to compose, then clear the upper 32 bits with
137       // SLLI_UW.
138       Val = ((uint64_t)Val) | (0xffffffffull << 32);
139       Unsigned = true;
140     }
141   }
142 
143   generateInstSeqImpl(Val, ActiveFeatures, Res);
144 
145   // Skip shift if we were able to use LUI directly.
146   if (ShiftAmount) {
147     if (Unsigned)
148       Res.push_back(RISCVMatInt::Inst(RISCV::SLLI_UW, ShiftAmount));
149     else
150       Res.push_back(RISCVMatInt::Inst(RISCV::SLLI, ShiftAmount));
151   }
152 
153   if (Lo12)
154     Res.push_back(RISCVMatInt::Inst(RISCV::ADDI, Lo12));
155 }
156 
157 static unsigned extractRotateInfo(int64_t Val) {
158   // for case: 0b111..1..xxxxxx1..1..
159   unsigned LeadingOnes = countLeadingOnes((uint64_t)Val);
160   unsigned TrailingOnes = countTrailingOnes((uint64_t)Val);
161   if (TrailingOnes > 0 && TrailingOnes < 64 &&
162       (LeadingOnes + TrailingOnes) > (64 - 12))
163     return 64 - TrailingOnes;
164 
165   // for case: 0bxxx1..1..1...xxx
166   unsigned UpperTrailingOnes = countTrailingOnes(Hi_32(Val));
167   unsigned LowerLeadingOnes = countLeadingOnes(Lo_32(Val));
168   if (UpperTrailingOnes < 32 &&
169       (UpperTrailingOnes + LowerLeadingOnes) > (64 - 12))
170     return 32 - UpperTrailingOnes;
171 
172   return 0;
173 }
174 
175 namespace llvm {
176 namespace RISCVMatInt {
177 InstSeq generateInstSeq(int64_t Val, const FeatureBitset &ActiveFeatures) {
178   RISCVMatInt::InstSeq Res;
179   generateInstSeqImpl(Val, ActiveFeatures, Res);
180 
181   // If there are trailing zeros, try generating a sign extended constant with
182   // no trailing zeros and use a final SLLI to restore them.
183   if ((Val & 1) == 0 && Res.size() > 2) {
184     unsigned TrailingZeros = countTrailingZeros((uint64_t)Val);
185     int64_t ShiftedVal = Val >> TrailingZeros;
186     RISCVMatInt::InstSeq TmpSeq;
187     generateInstSeqImpl(ShiftedVal, ActiveFeatures, TmpSeq);
188     TmpSeq.push_back(RISCVMatInt::Inst(RISCV::SLLI, TrailingZeros));
189 
190     // Keep the new sequence if it is an improvement.
191     if (TmpSeq.size() < Res.size()) {
192       Res = TmpSeq;
193       // A 2 instruction sequence is the best we can do.
194       if (Res.size() <= 2)
195         return Res;
196     }
197   }
198 
199   // If the constant is positive we might be able to generate a shifted constant
200   // with no leading zeros and use a final SRLI to restore them.
201   if (Val > 0 && Res.size() > 2) {
202     assert(ActiveFeatures[RISCV::Feature64Bit] &&
203            "Expected RV32 to only need 2 instructions");
204     unsigned LeadingZeros = countLeadingZeros((uint64_t)Val);
205     uint64_t ShiftedVal = (uint64_t)Val << LeadingZeros;
206     // Fill in the bits that will be shifted out with 1s. An example where this
207     // helps is trailing one masks with 32 or more ones. This will generate
208     // ADDI -1 and an SRLI.
209     ShiftedVal |= maskTrailingOnes<uint64_t>(LeadingZeros);
210 
211     RISCVMatInt::InstSeq TmpSeq;
212     generateInstSeqImpl(ShiftedVal, ActiveFeatures, TmpSeq);
213     TmpSeq.push_back(RISCVMatInt::Inst(RISCV::SRLI, LeadingZeros));
214 
215     // Keep the new sequence if it is an improvement.
216     if (TmpSeq.size() < Res.size()) {
217       Res = TmpSeq;
218       // A 2 instruction sequence is the best we can do.
219       if (Res.size() <= 2)
220         return Res;
221     }
222 
223     // Some cases can benefit from filling the lower bits with zeros instead.
224     ShiftedVal &= maskTrailingZeros<uint64_t>(LeadingZeros);
225     TmpSeq.clear();
226     generateInstSeqImpl(ShiftedVal, ActiveFeatures, TmpSeq);
227     TmpSeq.push_back(RISCVMatInt::Inst(RISCV::SRLI, LeadingZeros));
228 
229     // Keep the new sequence if it is an improvement.
230     if (TmpSeq.size() < Res.size()) {
231       Res = TmpSeq;
232       // A 2 instruction sequence is the best we can do.
233       if (Res.size() <= 2)
234         return Res;
235     }
236 
237     // If we have exactly 32 leading zeros and Zba, we can try using zext.w at
238     // the end of the sequence.
239     if (LeadingZeros == 32 && ActiveFeatures[RISCV::FeatureStdExtZba]) {
240       // Try replacing upper bits with 1.
241       uint64_t LeadingOnesVal = Val | maskLeadingOnes<uint64_t>(LeadingZeros);
242       TmpSeq.clear();
243       generateInstSeqImpl(LeadingOnesVal, ActiveFeatures, TmpSeq);
244       TmpSeq.push_back(RISCVMatInt::Inst(RISCV::ADD_UW, 0));
245 
246       // Keep the new sequence if it is an improvement.
247       if (TmpSeq.size() < Res.size()) {
248         Res = TmpSeq;
249         // A 2 instruction sequence is the best we can do.
250         if (Res.size() <= 2)
251           return Res;
252       }
253     }
254   }
255 
256   // Perform optimization with BCLRI/BSETI in the Zbs extension.
257   if (Res.size() > 2 && ActiveFeatures[RISCV::FeatureStdExtZbs]) {
258     assert(ActiveFeatures[RISCV::Feature64Bit] &&
259            "Expected RV32 to only need 2 instructions");
260 
261     // 1. For values in range 0xffffffff 7fffffff ~ 0xffffffff 00000000,
262     //    call generateInstSeqImpl with Val|0x80000000 (which is expected be
263     //    an int32), then emit (BCLRI r, 31).
264     // 2. For values in range 0x80000000 ~ 0xffffffff, call generateInstSeqImpl
265     //    with Val&~0x80000000 (which is expected to be an int32), then
266     //    emit (BSETI r, 31).
267     int64_t NewVal;
268     unsigned Opc;
269     if (Val < 0) {
270       Opc = RISCV::BCLRI;
271       NewVal = Val | 0x80000000ll;
272     } else {
273       Opc = RISCV::BSETI;
274       NewVal = Val & ~0x80000000ll;
275     }
276     if (isInt<32>(NewVal)) {
277       RISCVMatInt::InstSeq TmpSeq;
278       generateInstSeqImpl(NewVal, ActiveFeatures, TmpSeq);
279       TmpSeq.push_back(RISCVMatInt::Inst(Opc, 31));
280       if (TmpSeq.size() < Res.size())
281         Res = TmpSeq;
282     }
283 
284     // Try to use BCLRI for upper 32 bits if the original lower 32 bits are
285     // negative int32, or use BSETI for upper 32 bits if the original lower
286     // 32 bits are positive int32.
287     int32_t Lo = Val;
288     uint32_t Hi = Val >> 32;
289     Opc = 0;
290     RISCVMatInt::InstSeq TmpSeq;
291     generateInstSeqImpl(Lo, ActiveFeatures, TmpSeq);
292     // Check if it is profitable to use BCLRI/BSETI.
293     if (Lo > 0 && TmpSeq.size() + countPopulation(Hi) < Res.size()) {
294       Opc = RISCV::BSETI;
295     } else if (Lo < 0 && TmpSeq.size() + countPopulation(~Hi) < Res.size()) {
296       Opc = RISCV::BCLRI;
297       Hi = ~Hi;
298     }
299     // Search for each bit and build corresponding BCLRI/BSETI.
300     if (Opc > 0) {
301       while (Hi != 0) {
302         unsigned Bit = countTrailingZeros(Hi);
303         TmpSeq.push_back(RISCVMatInt::Inst(Opc, Bit + 32));
304         Hi &= ~(1 << Bit);
305       }
306       if (TmpSeq.size() < Res.size())
307         Res = TmpSeq;
308     }
309   }
310 
311   // Perform optimization with SH*ADD in the Zba extension.
312   if (Res.size() > 2 && ActiveFeatures[RISCV::FeatureStdExtZba]) {
313     assert(ActiveFeatures[RISCV::Feature64Bit] &&
314            "Expected RV32 to only need 2 instructions");
315     int64_t Div = 0;
316     unsigned Opc = 0;
317     RISCVMatInt::InstSeq TmpSeq;
318     // Select the opcode and divisor.
319     if ((Val % 3) == 0 && isInt<32>(Val / 3)) {
320       Div = 3;
321       Opc = RISCV::SH1ADD;
322     } else if ((Val % 5) == 0 && isInt<32>(Val / 5)) {
323       Div = 5;
324       Opc = RISCV::SH2ADD;
325     } else if ((Val % 9) == 0 && isInt<32>(Val / 9)) {
326       Div = 9;
327       Opc = RISCV::SH3ADD;
328     }
329     // Build the new instruction sequence.
330     if (Div > 0) {
331       generateInstSeqImpl(Val / Div, ActiveFeatures, TmpSeq);
332       TmpSeq.push_back(RISCVMatInt::Inst(Opc, 0));
333       if (TmpSeq.size() < Res.size())
334         Res = TmpSeq;
335     } else {
336       // Try to use LUI+SH*ADD+ADDI.
337       int64_t Hi52 = ((uint64_t)Val + 0x800ull) & ~0xfffull;
338       int64_t Lo12 = SignExtend64<12>(Val);
339       Div = 0;
340       if (isInt<32>(Hi52 / 3) && (Hi52 % 3) == 0) {
341         Div = 3;
342         Opc = RISCV::SH1ADD;
343       } else if (isInt<32>(Hi52 / 5) && (Hi52 % 5) == 0) {
344         Div = 5;
345         Opc = RISCV::SH2ADD;
346       } else if (isInt<32>(Hi52 / 9) && (Hi52 % 9) == 0) {
347         Div = 9;
348         Opc = RISCV::SH3ADD;
349       }
350       // Build the new instruction sequence.
351       if (Div > 0) {
352         // For Val that has zero Lo12 (implies Val equals to Hi52) should has
353         // already been processed to LUI+SH*ADD by previous optimization.
354         assert(Lo12 != 0 &&
355                "unexpected instruction sequence for immediate materialisation");
356         assert(TmpSeq.empty() && "Expected empty TmpSeq");
357         generateInstSeqImpl(Hi52 / Div, ActiveFeatures, TmpSeq);
358         TmpSeq.push_back(RISCVMatInt::Inst(Opc, 0));
359         TmpSeq.push_back(RISCVMatInt::Inst(RISCV::ADDI, Lo12));
360         if (TmpSeq.size() < Res.size())
361           Res = TmpSeq;
362       }
363     }
364   }
365 
366   // Perform optimization with rori in the Zbb extension.
367   if (Res.size() > 2 && ActiveFeatures[RISCV::FeatureStdExtZbb]) {
368     if (unsigned Rotate = extractRotateInfo(Val)) {
369       RISCVMatInt::InstSeq TmpSeq;
370       uint64_t NegImm12 =
371           ((uint64_t)Val >> (64 - Rotate)) | ((uint64_t)Val << Rotate);
372       assert(isInt<12>(NegImm12));
373       TmpSeq.push_back(RISCVMatInt::Inst(RISCV::ADDI, NegImm12));
374       TmpSeq.push_back(RISCVMatInt::Inst(RISCV::RORI, Rotate));
375       Res = TmpSeq;
376     }
377   }
378   return Res;
379 }
380 
381 int getIntMatCost(const APInt &Val, unsigned Size,
382                   const FeatureBitset &ActiveFeatures, bool CompressionCost) {
383   bool IsRV64 = ActiveFeatures[RISCV::Feature64Bit];
384   bool HasRVC = CompressionCost && ActiveFeatures[RISCV::FeatureStdExtC];
385   int PlatRegSize = IsRV64 ? 64 : 32;
386 
387   // Split the constant into platform register sized chunks, and calculate cost
388   // of each chunk.
389   int Cost = 0;
390   for (unsigned ShiftVal = 0; ShiftVal < Size; ShiftVal += PlatRegSize) {
391     APInt Chunk = Val.ashr(ShiftVal).sextOrTrunc(PlatRegSize);
392     InstSeq MatSeq = generateInstSeq(Chunk.getSExtValue(), ActiveFeatures);
393     Cost += getInstSeqCost(MatSeq, HasRVC);
394   }
395   return std::max(1, Cost);
396 }
397 
398 OpndKind Inst::getOpndKind() const {
399   switch (Opc) {
400   default:
401     llvm_unreachable("Unexpected opcode!");
402   case RISCV::LUI:
403     return RISCVMatInt::Imm;
404   case RISCV::ADD_UW:
405     return RISCVMatInt::RegX0;
406   case RISCV::SH1ADD:
407   case RISCV::SH2ADD:
408   case RISCV::SH3ADD:
409     return RISCVMatInt::RegReg;
410   case RISCV::ADDI:
411   case RISCV::ADDIW:
412   case RISCV::SLLI:
413   case RISCV::SRLI:
414   case RISCV::SLLI_UW:
415   case RISCV::RORI:
416   case RISCV::BSETI:
417   case RISCV::BCLRI:
418     return RISCVMatInt::RegImm;
419   }
420 }
421 
422 } // namespace RISCVMatInt
423 } // namespace llvm
424