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