1 | // |
2 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
3 | // See https://llvm.org/LICENSE.txt for license information. |
4 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
5 | // |
6 | //===----------------------------------------------------------------------===// |
7 | // |
8 | // This file contains a pass that performs optimization on SIMD instructions |
9 | // with high latency by splitting them into more efficient series of |
10 | // instructions. |
11 | // |
12 | // 1. Rewrite certain SIMD instructions with vector element due to their |
13 | // inefficiency on some targets. |
14 | // |
15 | // For example: |
16 | // fmla v0.4s, v1.4s, v2.s[1] |
17 | // |
18 | // Is rewritten into: |
19 | // dup v3.4s, v2.s[1] |
20 | // fmla v0.4s, v1.4s, v3.4s |
21 | // |
22 | // 2. Rewrite interleaved memory access instructions due to their |
23 | // inefficiency on some targets. |
24 | // |
25 | // For example: |
26 | // st2 {v0.4s, v1.4s}, addr |
27 | // |
28 | // Is rewritten into: |
29 | // zip1 v2.4s, v0.4s, v1.4s |
30 | // zip2 v3.4s, v0.4s, v1.4s |
31 | // stp q2, q3, addr |
32 | // |
33 | //===----------------------------------------------------------------------===// |
34 | |
35 | #include "AArch64InstrInfo.h" |
36 | #include "llvm/ADT/SmallVector.h" |
37 | #include "llvm/ADT/Statistic.h" |
38 | #include "llvm/ADT/StringRef.h" |
39 | #include "llvm/CodeGen/MachineBasicBlock.h" |
40 | #include "llvm/CodeGen/MachineFunction.h" |
41 | #include "llvm/CodeGen/MachineFunctionPass.h" |
42 | #include "llvm/CodeGen/MachineInstr.h" |
43 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
44 | #include "llvm/CodeGen/MachineOperand.h" |
45 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
46 | #include "llvm/CodeGen/TargetInstrInfo.h" |
47 | #include "llvm/CodeGen/TargetSchedule.h" |
48 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
49 | #include "llvm/MC/MCInstrDesc.h" |
50 | #include "llvm/MC/MCSchedule.h" |
51 | #include "llvm/Pass.h" |
52 | #include <unordered_map> |
53 | #include <map> |
54 | |
55 | using namespace llvm; |
56 | |
57 | #define DEBUG_TYPE "aarch64-simdinstr-opt" |
58 | |
59 | STATISTIC(NumModifiedInstr, |
60 | "Number of SIMD instructions modified" ); |
61 | |
62 | #define AARCH64_VECTOR_BY_ELEMENT_OPT_NAME \ |
63 | "AArch64 SIMD instructions optimization pass" |
64 | |
65 | namespace { |
66 | |
67 | struct AArch64SIMDInstrOpt : public MachineFunctionPass { |
68 | static char ID; |
69 | |
70 | const TargetInstrInfo *TII; |
71 | MachineRegisterInfo *MRI; |
72 | TargetSchedModel SchedModel; |
73 | |
74 | // The two maps below are used to cache decisions instead of recomputing: |
75 | // This is used to cache instruction replacement decisions within function |
76 | // units and across function units. |
77 | std::map<std::pair<unsigned, std::string>, bool> SIMDInstrTable; |
78 | // This is used to cache the decision of whether to leave the interleaved |
79 | // store instructions replacement pass early or not for a particular target. |
80 | std::unordered_map<std::string, bool> InterlEarlyExit; |
81 | |
82 | typedef enum { |
83 | VectorElem, |
84 | Interleave |
85 | } Subpass; |
86 | |
87 | // Instruction represented by OrigOpc is replaced by instructions in ReplOpc. |
88 | struct InstReplInfo { |
89 | unsigned OrigOpc; |
90 | std::vector<unsigned> ReplOpc; |
91 | const TargetRegisterClass RC; |
92 | }; |
93 | |
94 | #define RuleST2(OpcOrg, OpcR0, OpcR1, OpcR2, RC) \ |
95 | {OpcOrg, {OpcR0, OpcR1, OpcR2}, RC} |
96 | #define RuleST4(OpcOrg, OpcR0, OpcR1, OpcR2, OpcR3, OpcR4, OpcR5, OpcR6, \ |
97 | OpcR7, OpcR8, OpcR9, RC) \ |
98 | {OpcOrg, \ |
99 | {OpcR0, OpcR1, OpcR2, OpcR3, OpcR4, OpcR5, OpcR6, OpcR7, OpcR8, OpcR9}, RC} |
100 | |
101 | // The Instruction Replacement Table: |
102 | std::vector<InstReplInfo> IRT = { |
103 | // ST2 instructions |
104 | RuleST2(AArch64::ST2Twov2d, AArch64::ZIP1v2i64, AArch64::ZIP2v2i64, |
105 | AArch64::STPQi, AArch64::FPR128RegClass), |
106 | RuleST2(AArch64::ST2Twov4s, AArch64::ZIP1v4i32, AArch64::ZIP2v4i32, |
107 | AArch64::STPQi, AArch64::FPR128RegClass), |
108 | RuleST2(AArch64::ST2Twov2s, AArch64::ZIP1v2i32, AArch64::ZIP2v2i32, |
109 | AArch64::STPDi, AArch64::FPR64RegClass), |
110 | RuleST2(AArch64::ST2Twov8h, AArch64::ZIP1v8i16, AArch64::ZIP2v8i16, |
111 | AArch64::STPQi, AArch64::FPR128RegClass), |
112 | RuleST2(AArch64::ST2Twov4h, AArch64::ZIP1v4i16, AArch64::ZIP2v4i16, |
113 | AArch64::STPDi, AArch64::FPR64RegClass), |
114 | RuleST2(AArch64::ST2Twov16b, AArch64::ZIP1v16i8, AArch64::ZIP2v16i8, |
115 | AArch64::STPQi, AArch64::FPR128RegClass), |
116 | RuleST2(AArch64::ST2Twov8b, AArch64::ZIP1v8i8, AArch64::ZIP2v8i8, |
117 | AArch64::STPDi, AArch64::FPR64RegClass), |
118 | // ST4 instructions |
119 | RuleST4(AArch64::ST4Fourv2d, AArch64::ZIP1v2i64, AArch64::ZIP2v2i64, |
120 | AArch64::ZIP1v2i64, AArch64::ZIP2v2i64, AArch64::ZIP1v2i64, |
121 | AArch64::ZIP2v2i64, AArch64::ZIP1v2i64, AArch64::ZIP2v2i64, |
122 | AArch64::STPQi, AArch64::STPQi, AArch64::FPR128RegClass), |
123 | RuleST4(AArch64::ST4Fourv4s, AArch64::ZIP1v4i32, AArch64::ZIP2v4i32, |
124 | AArch64::ZIP1v4i32, AArch64::ZIP2v4i32, AArch64::ZIP1v4i32, |
125 | AArch64::ZIP2v4i32, AArch64::ZIP1v4i32, AArch64::ZIP2v4i32, |
126 | AArch64::STPQi, AArch64::STPQi, AArch64::FPR128RegClass), |
127 | RuleST4(AArch64::ST4Fourv2s, AArch64::ZIP1v2i32, AArch64::ZIP2v2i32, |
128 | AArch64::ZIP1v2i32, AArch64::ZIP2v2i32, AArch64::ZIP1v2i32, |
129 | AArch64::ZIP2v2i32, AArch64::ZIP1v2i32, AArch64::ZIP2v2i32, |
130 | AArch64::STPDi, AArch64::STPDi, AArch64::FPR64RegClass), |
131 | RuleST4(AArch64::ST4Fourv8h, AArch64::ZIP1v8i16, AArch64::ZIP2v8i16, |
132 | AArch64::ZIP1v8i16, AArch64::ZIP2v8i16, AArch64::ZIP1v8i16, |
133 | AArch64::ZIP2v8i16, AArch64::ZIP1v8i16, AArch64::ZIP2v8i16, |
134 | AArch64::STPQi, AArch64::STPQi, AArch64::FPR128RegClass), |
135 | RuleST4(AArch64::ST4Fourv4h, AArch64::ZIP1v4i16, AArch64::ZIP2v4i16, |
136 | AArch64::ZIP1v4i16, AArch64::ZIP2v4i16, AArch64::ZIP1v4i16, |
137 | AArch64::ZIP2v4i16, AArch64::ZIP1v4i16, AArch64::ZIP2v4i16, |
138 | AArch64::STPDi, AArch64::STPDi, AArch64::FPR64RegClass), |
139 | RuleST4(AArch64::ST4Fourv16b, AArch64::ZIP1v16i8, AArch64::ZIP2v16i8, |
140 | AArch64::ZIP1v16i8, AArch64::ZIP2v16i8, AArch64::ZIP1v16i8, |
141 | AArch64::ZIP2v16i8, AArch64::ZIP1v16i8, AArch64::ZIP2v16i8, |
142 | AArch64::STPQi, AArch64::STPQi, AArch64::FPR128RegClass), |
143 | RuleST4(AArch64::ST4Fourv8b, AArch64::ZIP1v8i8, AArch64::ZIP2v8i8, |
144 | AArch64::ZIP1v8i8, AArch64::ZIP2v8i8, AArch64::ZIP1v8i8, |
145 | AArch64::ZIP2v8i8, AArch64::ZIP1v8i8, AArch64::ZIP2v8i8, |
146 | AArch64::STPDi, AArch64::STPDi, AArch64::FPR64RegClass) |
147 | }; |
148 | |
149 | // A costly instruction is replaced in this work by N efficient instructions |
150 | // The maximum of N is curently 10 and it is for ST4 case. |
151 | static const unsigned MaxNumRepl = 10; |
152 | |
153 | AArch64SIMDInstrOpt() : MachineFunctionPass(ID) { |
154 | initializeAArch64SIMDInstrOptPass(*PassRegistry::getPassRegistry()); |
155 | } |
156 | |
157 | /// Based only on latency of instructions, determine if it is cost efficient |
158 | /// to replace the instruction InstDesc by the instructions stored in the |
159 | /// array InstDescRepl. |
160 | /// Return true if replacement is expected to be faster. |
161 | bool shouldReplaceInst(MachineFunction *MF, const MCInstrDesc *InstDesc, |
162 | SmallVectorImpl<const MCInstrDesc*> &ReplInstrMCID); |
163 | |
164 | /// Determine if we need to exit the instruction replacement optimization |
165 | /// passes early. This makes sure that no compile time is spent in this pass |
166 | /// for targets with no need for any of these optimizations. |
167 | /// Return true if early exit of the pass is recommended. |
168 | bool shouldExitEarly(MachineFunction *MF, Subpass SP); |
169 | |
170 | /// Check whether an equivalent DUP instruction has already been |
171 | /// created or not. |
172 | /// Return true when the DUP instruction already exists. In this case, |
173 | /// DestReg will point to the destination of the already created DUP. |
174 | bool reuseDUP(MachineInstr &MI, unsigned DupOpcode, unsigned SrcReg, |
175 | unsigned LaneNumber, unsigned *DestReg) const; |
176 | |
177 | /// Certain SIMD instructions with vector element operand are not efficient. |
178 | /// Rewrite them into SIMD instructions with vector operands. This rewrite |
179 | /// is driven by the latency of the instructions. |
180 | /// Return true if the SIMD instruction is modified. |
181 | bool optimizeVectElement(MachineInstr &MI); |
182 | |
183 | /// Process The REG_SEQUENCE instruction, and extract the source |
184 | /// operands of the ST2/4 instruction from it. |
185 | /// Example of such instructions. |
186 | /// %dest = REG_SEQUENCE %st2_src1, dsub0, %st2_src2, dsub1; |
187 | /// Return true when the instruction is processed successfully. |
188 | bool processSeqRegInst(MachineInstr *DefiningMI, unsigned* StReg, |
189 | unsigned* StRegKill, unsigned NumArg) const; |
190 | |
191 | /// Load/Store Interleaving instructions are not always beneficial. |
192 | /// Replace them by ZIP instructionand classical load/store. |
193 | /// Return true if the SIMD instruction is modified. |
194 | bool optimizeLdStInterleave(MachineInstr &MI); |
195 | |
196 | /// Return the number of useful source registers for this |
197 | /// instruction (2 for ST2 and 4 for ST4). |
198 | unsigned determineSrcReg(MachineInstr &MI) const; |
199 | |
200 | bool runOnMachineFunction(MachineFunction &Fn) override; |
201 | |
202 | StringRef getPassName() const override { |
203 | return AARCH64_VECTOR_BY_ELEMENT_OPT_NAME; |
204 | } |
205 | }; |
206 | |
207 | char AArch64SIMDInstrOpt::ID = 0; |
208 | |
209 | } // end anonymous namespace |
210 | |
211 | INITIALIZE_PASS(AArch64SIMDInstrOpt, "aarch64-simdinstr-opt" , |
212 | AARCH64_VECTOR_BY_ELEMENT_OPT_NAME, false, false) |
213 | |
214 | /// Based only on latency of instructions, determine if it is cost efficient |
215 | /// to replace the instruction InstDesc by the instructions stored in the |
216 | /// array InstDescRepl. |
217 | /// Return true if replacement is expected to be faster. |
218 | bool AArch64SIMDInstrOpt:: |
219 | shouldReplaceInst(MachineFunction *MF, const MCInstrDesc *InstDesc, |
220 | SmallVectorImpl<const MCInstrDesc*> &InstDescRepl) { |
221 | // Check if replacement decision is already available in the cached table. |
222 | // if so, return it. |
223 | std::string Subtarget = std::string(SchedModel.getSubtargetInfo()->getCPU()); |
224 | auto InstID = std::make_pair(x: InstDesc->getOpcode(), y&: Subtarget); |
225 | auto It = SIMDInstrTable.find(x: InstID); |
226 | if (It != SIMDInstrTable.end()) |
227 | return It->second; |
228 | |
229 | unsigned SCIdx = InstDesc->getSchedClass(); |
230 | const MCSchedClassDesc *SCDesc = |
231 | SchedModel.getMCSchedModel()->getSchedClassDesc(SchedClassIdx: SCIdx); |
232 | |
233 | // If a target does not define resources for the instructions |
234 | // of interest, then return false for no replacement. |
235 | const MCSchedClassDesc *SCDescRepl; |
236 | if (!SCDesc->isValid() || SCDesc->isVariant()) |
237 | { |
238 | SIMDInstrTable[InstID] = false; |
239 | return false; |
240 | } |
241 | for (const auto *IDesc : InstDescRepl) |
242 | { |
243 | SCDescRepl = SchedModel.getMCSchedModel()->getSchedClassDesc( |
244 | SchedClassIdx: IDesc->getSchedClass()); |
245 | if (!SCDescRepl->isValid() || SCDescRepl->isVariant()) |
246 | { |
247 | SIMDInstrTable[InstID] = false; |
248 | return false; |
249 | } |
250 | } |
251 | |
252 | // Replacement cost. |
253 | unsigned ReplCost = 0; |
254 | for (const auto *IDesc :InstDescRepl) |
255 | ReplCost += SchedModel.computeInstrLatency(Opcode: IDesc->getOpcode()); |
256 | |
257 | if (SchedModel.computeInstrLatency(Opcode: InstDesc->getOpcode()) > ReplCost) |
258 | { |
259 | SIMDInstrTable[InstID] = true; |
260 | return true; |
261 | } |
262 | else |
263 | { |
264 | SIMDInstrTable[InstID] = false; |
265 | return false; |
266 | } |
267 | } |
268 | |
269 | /// Determine if we need to exit this pass for a kind of instruction replacement |
270 | /// early. This makes sure that no compile time is spent in this pass for |
271 | /// targets with no need for any of these optimizations beyond performing this |
272 | /// check. |
273 | /// Return true if early exit of this pass for a kind of instruction |
274 | /// replacement is recommended for a target. |
275 | bool AArch64SIMDInstrOpt::shouldExitEarly(MachineFunction *MF, Subpass SP) { |
276 | const MCInstrDesc* OriginalMCID; |
277 | SmallVector<const MCInstrDesc*, MaxNumRepl> ReplInstrMCID; |
278 | |
279 | switch (SP) { |
280 | // For this optimization, check by comparing the latency of a representative |
281 | // instruction to that of the replacement instructions. |
282 | // TODO: check for all concerned instructions. |
283 | case VectorElem: |
284 | OriginalMCID = &TII->get(Opcode: AArch64::FMLAv4i32_indexed); |
285 | ReplInstrMCID.push_back(Elt: &TII->get(Opcode: AArch64::DUPv4i32lane)); |
286 | ReplInstrMCID.push_back(Elt: &TII->get(Opcode: AArch64::FMLAv4f32)); |
287 | if (shouldReplaceInst(MF, InstDesc: OriginalMCID, InstDescRepl&: ReplInstrMCID)) |
288 | return false; |
289 | break; |
290 | |
291 | // For this optimization, check for all concerned instructions. |
292 | case Interleave: |
293 | std::string Subtarget = |
294 | std::string(SchedModel.getSubtargetInfo()->getCPU()); |
295 | auto It = InterlEarlyExit.find(x: Subtarget); |
296 | if (It != InterlEarlyExit.end()) |
297 | return It->second; |
298 | |
299 | for (auto &I : IRT) { |
300 | OriginalMCID = &TII->get(Opcode: I.OrigOpc); |
301 | for (auto &Repl : I.ReplOpc) |
302 | ReplInstrMCID.push_back(Elt: &TII->get(Opcode: Repl)); |
303 | if (shouldReplaceInst(MF, InstDesc: OriginalMCID, InstDescRepl&: ReplInstrMCID)) { |
304 | InterlEarlyExit[Subtarget] = false; |
305 | return false; |
306 | } |
307 | ReplInstrMCID.clear(); |
308 | } |
309 | InterlEarlyExit[Subtarget] = true; |
310 | break; |
311 | } |
312 | |
313 | return true; |
314 | } |
315 | |
316 | /// Check whether an equivalent DUP instruction has already been |
317 | /// created or not. |
318 | /// Return true when the DUP instruction already exists. In this case, |
319 | /// DestReg will point to the destination of the already created DUP. |
320 | bool AArch64SIMDInstrOpt::reuseDUP(MachineInstr &MI, unsigned DupOpcode, |
321 | unsigned SrcReg, unsigned LaneNumber, |
322 | unsigned *DestReg) const { |
323 | for (MachineBasicBlock::iterator MII = MI, MIE = MI.getParent()->begin(); |
324 | MII != MIE;) { |
325 | MII--; |
326 | MachineInstr *CurrentMI = &*MII; |
327 | |
328 | if (CurrentMI->getOpcode() == DupOpcode && |
329 | CurrentMI->getNumOperands() == 3 && |
330 | CurrentMI->getOperand(i: 1).getReg() == SrcReg && |
331 | CurrentMI->getOperand(i: 2).getImm() == LaneNumber) { |
332 | *DestReg = CurrentMI->getOperand(i: 0).getReg(); |
333 | return true; |
334 | } |
335 | } |
336 | |
337 | return false; |
338 | } |
339 | |
340 | /// Certain SIMD instructions with vector element operand are not efficient. |
341 | /// Rewrite them into SIMD instructions with vector operands. This rewrite |
342 | /// is driven by the latency of the instructions. |
343 | /// The instruction of concerns are for the time being FMLA, FMLS, FMUL, |
344 | /// and FMULX and hence they are hardcoded. |
345 | /// |
346 | /// For example: |
347 | /// fmla v0.4s, v1.4s, v2.s[1] |
348 | /// |
349 | /// Is rewritten into |
350 | /// dup v3.4s, v2.s[1] // DUP not necessary if redundant |
351 | /// fmla v0.4s, v1.4s, v3.4s |
352 | /// |
353 | /// Return true if the SIMD instruction is modified. |
354 | bool AArch64SIMDInstrOpt::optimizeVectElement(MachineInstr &MI) { |
355 | const MCInstrDesc *MulMCID, *DupMCID; |
356 | const TargetRegisterClass *RC = &AArch64::FPR128RegClass; |
357 | |
358 | switch (MI.getOpcode()) { |
359 | default: |
360 | return false; |
361 | |
362 | // 4X32 instructions |
363 | case AArch64::FMLAv4i32_indexed: |
364 | DupMCID = &TII->get(Opcode: AArch64::DUPv4i32lane); |
365 | MulMCID = &TII->get(Opcode: AArch64::FMLAv4f32); |
366 | break; |
367 | case AArch64::FMLSv4i32_indexed: |
368 | DupMCID = &TII->get(Opcode: AArch64::DUPv4i32lane); |
369 | MulMCID = &TII->get(Opcode: AArch64::FMLSv4f32); |
370 | break; |
371 | case AArch64::FMULXv4i32_indexed: |
372 | DupMCID = &TII->get(Opcode: AArch64::DUPv4i32lane); |
373 | MulMCID = &TII->get(Opcode: AArch64::FMULXv4f32); |
374 | break; |
375 | case AArch64::FMULv4i32_indexed: |
376 | DupMCID = &TII->get(Opcode: AArch64::DUPv4i32lane); |
377 | MulMCID = &TII->get(Opcode: AArch64::FMULv4f32); |
378 | break; |
379 | |
380 | // 2X64 instructions |
381 | case AArch64::FMLAv2i64_indexed: |
382 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i64lane); |
383 | MulMCID = &TII->get(Opcode: AArch64::FMLAv2f64); |
384 | break; |
385 | case AArch64::FMLSv2i64_indexed: |
386 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i64lane); |
387 | MulMCID = &TII->get(Opcode: AArch64::FMLSv2f64); |
388 | break; |
389 | case AArch64::FMULXv2i64_indexed: |
390 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i64lane); |
391 | MulMCID = &TII->get(Opcode: AArch64::FMULXv2f64); |
392 | break; |
393 | case AArch64::FMULv2i64_indexed: |
394 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i64lane); |
395 | MulMCID = &TII->get(Opcode: AArch64::FMULv2f64); |
396 | break; |
397 | |
398 | // 2X32 instructions |
399 | case AArch64::FMLAv2i32_indexed: |
400 | RC = &AArch64::FPR64RegClass; |
401 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i32lane); |
402 | MulMCID = &TII->get(Opcode: AArch64::FMLAv2f32); |
403 | break; |
404 | case AArch64::FMLSv2i32_indexed: |
405 | RC = &AArch64::FPR64RegClass; |
406 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i32lane); |
407 | MulMCID = &TII->get(Opcode: AArch64::FMLSv2f32); |
408 | break; |
409 | case AArch64::FMULXv2i32_indexed: |
410 | RC = &AArch64::FPR64RegClass; |
411 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i32lane); |
412 | MulMCID = &TII->get(Opcode: AArch64::FMULXv2f32); |
413 | break; |
414 | case AArch64::FMULv2i32_indexed: |
415 | RC = &AArch64::FPR64RegClass; |
416 | DupMCID = &TII->get(Opcode: AArch64::DUPv2i32lane); |
417 | MulMCID = &TII->get(Opcode: AArch64::FMULv2f32); |
418 | break; |
419 | } |
420 | |
421 | SmallVector<const MCInstrDesc*, 2> ReplInstrMCID; |
422 | ReplInstrMCID.push_back(Elt: DupMCID); |
423 | ReplInstrMCID.push_back(Elt: MulMCID); |
424 | if (!shouldReplaceInst(MF: MI.getParent()->getParent(), InstDesc: &TII->get(Opcode: MI.getOpcode()), |
425 | InstDescRepl&: ReplInstrMCID)) |
426 | return false; |
427 | |
428 | const DebugLoc &DL = MI.getDebugLoc(); |
429 | MachineBasicBlock &MBB = *MI.getParent(); |
430 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
431 | |
432 | // Get the operands of the current SIMD arithmetic instruction. |
433 | Register MulDest = MI.getOperand(i: 0).getReg(); |
434 | Register SrcReg0 = MI.getOperand(i: 1).getReg(); |
435 | unsigned Src0IsKill = getKillRegState(B: MI.getOperand(i: 1).isKill()); |
436 | Register SrcReg1 = MI.getOperand(i: 2).getReg(); |
437 | unsigned Src1IsKill = getKillRegState(B: MI.getOperand(i: 2).isKill()); |
438 | unsigned DupDest; |
439 | |
440 | // Instructions of interest have either 4 or 5 operands. |
441 | if (MI.getNumOperands() == 5) { |
442 | Register SrcReg2 = MI.getOperand(i: 3).getReg(); |
443 | unsigned Src2IsKill = getKillRegState(B: MI.getOperand(i: 3).isKill()); |
444 | unsigned LaneNumber = MI.getOperand(i: 4).getImm(); |
445 | // Create a new DUP instruction. Note that if an equivalent DUP instruction |
446 | // has already been created before, then use that one instead of creating |
447 | // a new one. |
448 | if (!reuseDUP(MI, DupOpcode: DupMCID->getOpcode(), SrcReg: SrcReg2, LaneNumber, DestReg: &DupDest)) { |
449 | DupDest = MRI.createVirtualRegister(RegClass: RC); |
450 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *DupMCID, DestReg: DupDest) |
451 | .addReg(RegNo: SrcReg2, flags: Src2IsKill) |
452 | .addImm(Val: LaneNumber); |
453 | } |
454 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *MulMCID, DestReg: MulDest) |
455 | .addReg(RegNo: SrcReg0, flags: Src0IsKill) |
456 | .addReg(RegNo: SrcReg1, flags: Src1IsKill) |
457 | .addReg(RegNo: DupDest, flags: Src2IsKill); |
458 | } else if (MI.getNumOperands() == 4) { |
459 | unsigned LaneNumber = MI.getOperand(i: 3).getImm(); |
460 | if (!reuseDUP(MI, DupOpcode: DupMCID->getOpcode(), SrcReg: SrcReg1, LaneNumber, DestReg: &DupDest)) { |
461 | DupDest = MRI.createVirtualRegister(RegClass: RC); |
462 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *DupMCID, DestReg: DupDest) |
463 | .addReg(RegNo: SrcReg1, flags: Src1IsKill) |
464 | .addImm(Val: LaneNumber); |
465 | } |
466 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *MulMCID, DestReg: MulDest) |
467 | .addReg(RegNo: SrcReg0, flags: Src0IsKill) |
468 | .addReg(RegNo: DupDest, flags: Src1IsKill); |
469 | } else { |
470 | return false; |
471 | } |
472 | |
473 | ++NumModifiedInstr; |
474 | return true; |
475 | } |
476 | |
477 | /// Load/Store Interleaving instructions are not always beneficial. |
478 | /// Replace them by ZIP instructions and classical load/store. |
479 | /// |
480 | /// For example: |
481 | /// st2 {v0.4s, v1.4s}, addr |
482 | /// |
483 | /// Is rewritten into: |
484 | /// zip1 v2.4s, v0.4s, v1.4s |
485 | /// zip2 v3.4s, v0.4s, v1.4s |
486 | /// stp q2, q3, addr |
487 | // |
488 | /// For example: |
489 | /// st4 {v0.4s, v1.4s, v2.4s, v3.4s}, addr |
490 | /// |
491 | /// Is rewritten into: |
492 | /// zip1 v4.4s, v0.4s, v2.4s |
493 | /// zip2 v5.4s, v0.4s, v2.4s |
494 | /// zip1 v6.4s, v1.4s, v3.4s |
495 | /// zip2 v7.4s, v1.4s, v3.4s |
496 | /// zip1 v8.4s, v4.4s, v6.4s |
497 | /// zip2 v9.4s, v4.4s, v6.4s |
498 | /// zip1 v10.4s, v5.4s, v7.4s |
499 | /// zip2 v11.4s, v5.4s, v7.4s |
500 | /// stp q8, q9, addr |
501 | /// stp q10, q11, addr+32 |
502 | /// |
503 | /// Currently only instructions related to ST2 and ST4 are considered. |
504 | /// Other may be added later. |
505 | /// Return true if the SIMD instruction is modified. |
506 | bool AArch64SIMDInstrOpt::optimizeLdStInterleave(MachineInstr &MI) { |
507 | |
508 | unsigned SeqReg, AddrReg; |
509 | unsigned StReg[4], StRegKill[4]; |
510 | MachineInstr *DefiningMI; |
511 | const DebugLoc &DL = MI.getDebugLoc(); |
512 | MachineBasicBlock &MBB = *MI.getParent(); |
513 | SmallVector<unsigned, MaxNumRepl> ZipDest; |
514 | SmallVector<const MCInstrDesc*, MaxNumRepl> ReplInstrMCID; |
515 | |
516 | // If current instruction matches any of the rewriting rules, then |
517 | // gather information about parameters of the new instructions. |
518 | bool Match = false; |
519 | for (auto &I : IRT) { |
520 | if (MI.getOpcode() == I.OrigOpc) { |
521 | SeqReg = MI.getOperand(i: 0).getReg(); |
522 | AddrReg = MI.getOperand(i: 1).getReg(); |
523 | DefiningMI = MRI->getUniqueVRegDef(Reg: SeqReg); |
524 | unsigned NumReg = determineSrcReg(MI); |
525 | if (!processSeqRegInst(DefiningMI, StReg, StRegKill, NumArg: NumReg)) |
526 | return false; |
527 | |
528 | for (auto &Repl : I.ReplOpc) { |
529 | ReplInstrMCID.push_back(Elt: &TII->get(Opcode: Repl)); |
530 | // Generate destination registers but only for non-store instruction. |
531 | if (Repl != AArch64::STPQi && Repl != AArch64::STPDi) |
532 | ZipDest.push_back(Elt: MRI->createVirtualRegister(RegClass: &I.RC)); |
533 | } |
534 | Match = true; |
535 | break; |
536 | } |
537 | } |
538 | |
539 | if (!Match) |
540 | return false; |
541 | |
542 | // Determine if it is profitable to replace MI by the series of instructions |
543 | // represented in ReplInstrMCID. |
544 | if (!shouldReplaceInst(MF: MI.getParent()->getParent(), InstDesc: &TII->get(Opcode: MI.getOpcode()), |
545 | InstDescRepl&: ReplInstrMCID)) |
546 | return false; |
547 | |
548 | // Generate the replacement instructions composed of ZIP1, ZIP2, and STP (at |
549 | // this point, the code generation is hardcoded and does not rely on the IRT |
550 | // table used above given that code generation for ST2 replacement is somewhat |
551 | // different than for ST4 replacement. We could have added more info into the |
552 | // table related to how we build new instructions but we may be adding more |
553 | // complexity with that). |
554 | switch (MI.getOpcode()) { |
555 | default: |
556 | return false; |
557 | |
558 | case AArch64::ST2Twov16b: |
559 | case AArch64::ST2Twov8b: |
560 | case AArch64::ST2Twov8h: |
561 | case AArch64::ST2Twov4h: |
562 | case AArch64::ST2Twov4s: |
563 | case AArch64::ST2Twov2s: |
564 | case AArch64::ST2Twov2d: |
565 | // ZIP instructions |
566 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[0], DestReg: ZipDest[0]) |
567 | .addReg(RegNo: StReg[0]) |
568 | .addReg(RegNo: StReg[1]); |
569 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[1], DestReg: ZipDest[1]) |
570 | .addReg(RegNo: StReg[0], flags: StRegKill[0]) |
571 | .addReg(RegNo: StReg[1], flags: StRegKill[1]); |
572 | // STP instructions |
573 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[2]) |
574 | .addReg(RegNo: ZipDest[0]) |
575 | .addReg(RegNo: ZipDest[1]) |
576 | .addReg(RegNo: AddrReg) |
577 | .addImm(Val: 0); |
578 | break; |
579 | |
580 | case AArch64::ST4Fourv16b: |
581 | case AArch64::ST4Fourv8b: |
582 | case AArch64::ST4Fourv8h: |
583 | case AArch64::ST4Fourv4h: |
584 | case AArch64::ST4Fourv4s: |
585 | case AArch64::ST4Fourv2s: |
586 | case AArch64::ST4Fourv2d: |
587 | // ZIP instructions |
588 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[0], DestReg: ZipDest[0]) |
589 | .addReg(RegNo: StReg[0]) |
590 | .addReg(RegNo: StReg[2]); |
591 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[1], DestReg: ZipDest[1]) |
592 | .addReg(RegNo: StReg[0], flags: StRegKill[0]) |
593 | .addReg(RegNo: StReg[2], flags: StRegKill[2]); |
594 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[2], DestReg: ZipDest[2]) |
595 | .addReg(RegNo: StReg[1]) |
596 | .addReg(RegNo: StReg[3]); |
597 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[3], DestReg: ZipDest[3]) |
598 | .addReg(RegNo: StReg[1], flags: StRegKill[1]) |
599 | .addReg(RegNo: StReg[3], flags: StRegKill[3]); |
600 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[4], DestReg: ZipDest[4]) |
601 | .addReg(RegNo: ZipDest[0]) |
602 | .addReg(RegNo: ZipDest[2]); |
603 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[5], DestReg: ZipDest[5]) |
604 | .addReg(RegNo: ZipDest[0]) |
605 | .addReg(RegNo: ZipDest[2]); |
606 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[6], DestReg: ZipDest[6]) |
607 | .addReg(RegNo: ZipDest[1]) |
608 | .addReg(RegNo: ZipDest[3]); |
609 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[7], DestReg: ZipDest[7]) |
610 | .addReg(RegNo: ZipDest[1]) |
611 | .addReg(RegNo: ZipDest[3]); |
612 | // stp instructions |
613 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[8]) |
614 | .addReg(RegNo: ZipDest[4]) |
615 | .addReg(RegNo: ZipDest[5]) |
616 | .addReg(RegNo: AddrReg) |
617 | .addImm(Val: 0); |
618 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: *ReplInstrMCID[9]) |
619 | .addReg(RegNo: ZipDest[6]) |
620 | .addReg(RegNo: ZipDest[7]) |
621 | .addReg(RegNo: AddrReg) |
622 | .addImm(Val: 2); |
623 | break; |
624 | } |
625 | |
626 | ++NumModifiedInstr; |
627 | return true; |
628 | } |
629 | |
630 | /// Process The REG_SEQUENCE instruction, and extract the source |
631 | /// operands of the ST2/4 instruction from it. |
632 | /// Example of such instruction. |
633 | /// %dest = REG_SEQUENCE %st2_src1, dsub0, %st2_src2, dsub1; |
634 | /// Return true when the instruction is processed successfully. |
635 | bool AArch64SIMDInstrOpt::processSeqRegInst(MachineInstr *DefiningMI, |
636 | unsigned* StReg, unsigned* StRegKill, unsigned NumArg) const { |
637 | assert(DefiningMI != nullptr); |
638 | if (DefiningMI->getOpcode() != AArch64::REG_SEQUENCE) |
639 | return false; |
640 | |
641 | for (unsigned i=0; i<NumArg; i++) { |
642 | StReg[i] = DefiningMI->getOperand(i: 2*i+1).getReg(); |
643 | StRegKill[i] = getKillRegState(B: DefiningMI->getOperand(i: 2*i+1).isKill()); |
644 | |
645 | // Validation check for the other arguments. |
646 | if (DefiningMI->getOperand(i: 2*i+2).isImm()) { |
647 | switch (DefiningMI->getOperand(i: 2*i+2).getImm()) { |
648 | default: |
649 | return false; |
650 | |
651 | case AArch64::dsub0: |
652 | case AArch64::dsub1: |
653 | case AArch64::dsub2: |
654 | case AArch64::dsub3: |
655 | case AArch64::qsub0: |
656 | case AArch64::qsub1: |
657 | case AArch64::qsub2: |
658 | case AArch64::qsub3: |
659 | break; |
660 | } |
661 | } |
662 | else |
663 | return false; |
664 | } |
665 | return true; |
666 | } |
667 | |
668 | /// Return the number of useful source registers for this instruction |
669 | /// (2 for ST2 and 4 for ST4). |
670 | unsigned AArch64SIMDInstrOpt::determineSrcReg(MachineInstr &MI) const { |
671 | switch (MI.getOpcode()) { |
672 | default: |
673 | llvm_unreachable("Unsupported instruction for this pass" ); |
674 | |
675 | case AArch64::ST2Twov16b: |
676 | case AArch64::ST2Twov8b: |
677 | case AArch64::ST2Twov8h: |
678 | case AArch64::ST2Twov4h: |
679 | case AArch64::ST2Twov4s: |
680 | case AArch64::ST2Twov2s: |
681 | case AArch64::ST2Twov2d: |
682 | return 2; |
683 | |
684 | case AArch64::ST4Fourv16b: |
685 | case AArch64::ST4Fourv8b: |
686 | case AArch64::ST4Fourv8h: |
687 | case AArch64::ST4Fourv4h: |
688 | case AArch64::ST4Fourv4s: |
689 | case AArch64::ST4Fourv2s: |
690 | case AArch64::ST4Fourv2d: |
691 | return 4; |
692 | } |
693 | } |
694 | |
695 | bool AArch64SIMDInstrOpt::runOnMachineFunction(MachineFunction &MF) { |
696 | if (skipFunction(F: MF.getFunction())) |
697 | return false; |
698 | |
699 | TII = MF.getSubtarget().getInstrInfo(); |
700 | MRI = &MF.getRegInfo(); |
701 | const TargetSubtargetInfo &ST = MF.getSubtarget(); |
702 | const AArch64InstrInfo *AAII = |
703 | static_cast<const AArch64InstrInfo *>(ST.getInstrInfo()); |
704 | if (!AAII) |
705 | return false; |
706 | SchedModel.init(TSInfo: &ST); |
707 | if (!SchedModel.hasInstrSchedModel()) |
708 | return false; |
709 | |
710 | bool Changed = false; |
711 | for (auto OptimizationKind : {VectorElem, Interleave}) { |
712 | if (!shouldExitEarly(MF: &MF, SP: OptimizationKind)) { |
713 | SmallVector<MachineInstr *, 8> RemoveMIs; |
714 | for (MachineBasicBlock &MBB : MF) { |
715 | for (MachineInstr &MI : MBB) { |
716 | bool InstRewrite; |
717 | if (OptimizationKind == VectorElem) |
718 | InstRewrite = optimizeVectElement(MI) ; |
719 | else |
720 | InstRewrite = optimizeLdStInterleave(MI); |
721 | if (InstRewrite) { |
722 | // Add MI to the list of instructions to be removed given that it |
723 | // has been replaced. |
724 | RemoveMIs.push_back(Elt: &MI); |
725 | Changed = true; |
726 | } |
727 | } |
728 | } |
729 | for (MachineInstr *MI : RemoveMIs) |
730 | MI->eraseFromParent(); |
731 | } |
732 | } |
733 | |
734 | return Changed; |
735 | } |
736 | |
737 | /// Returns an instance of the high cost ASIMD instruction replacement |
738 | /// optimization pass. |
739 | FunctionPass *llvm::createAArch64SIMDInstrOptPass() { |
740 | return new AArch64SIMDInstrOpt(); |
741 | } |
742 | |