1 | //===-- AArch64AdvSIMDScalar.cpp - Replace dead defs w/ zero reg --===// |
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 | // When profitable, replace GPR targeting i64 instructions with their |
9 | // AdvSIMD scalar equivalents. Generally speaking, "profitable" is defined |
10 | // as minimizing the number of cross-class register copies. |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | //===----------------------------------------------------------------------===// |
14 | // TODO: Graph based predicate heuristics. |
15 | // Walking the instruction list linearly will get many, perhaps most, of |
16 | // the cases, but to do a truly thorough job of this, we need a more |
17 | // wholistic approach. |
18 | // |
19 | // This optimization is very similar in spirit to the register allocator's |
20 | // spill placement, only here we're determining where to place cross-class |
21 | // register copies rather than spills. As such, a similar approach is |
22 | // called for. |
23 | // |
24 | // We want to build up a set of graphs of all instructions which are candidates |
25 | // for transformation along with instructions which generate their inputs and |
26 | // consume their outputs. For each edge in the graph, we assign a weight |
27 | // based on whether there is a copy required there (weight zero if not) and |
28 | // the block frequency of the block containing the defining or using |
29 | // instruction, whichever is less. Our optimization is then a graph problem |
30 | // to minimize the total weight of all the graphs, then transform instructions |
31 | // and add or remove copy instructions as called for to implement the |
32 | // solution. |
33 | //===----------------------------------------------------------------------===// |
34 | |
35 | #include "AArch64.h" |
36 | #include "AArch64InstrInfo.h" |
37 | #include "AArch64RegisterInfo.h" |
38 | #include "llvm/ADT/Statistic.h" |
39 | #include "llvm/CodeGen/MachineFunction.h" |
40 | #include "llvm/CodeGen/MachineFunctionPass.h" |
41 | #include "llvm/CodeGen/MachineInstr.h" |
42 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
43 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
44 | #include "llvm/Support/CommandLine.h" |
45 | #include "llvm/Support/Debug.h" |
46 | #include "llvm/Support/raw_ostream.h" |
47 | using namespace llvm; |
48 | |
49 | #define DEBUG_TYPE "aarch64-simd-scalar" |
50 | |
51 | // Allow forcing all i64 operations with equivalent SIMD instructions to use |
52 | // them. For stress-testing the transformation function. |
53 | static cl::opt<bool> |
54 | TransformAll("aarch64-simd-scalar-force-all" , |
55 | cl::desc("Force use of AdvSIMD scalar instructions everywhere" ), |
56 | cl::init(Val: false), cl::Hidden); |
57 | |
58 | STATISTIC(NumScalarInsnsUsed, "Number of scalar instructions used" ); |
59 | STATISTIC(NumCopiesDeleted, "Number of cross-class copies deleted" ); |
60 | STATISTIC(NumCopiesInserted, "Number of cross-class copies inserted" ); |
61 | |
62 | #define AARCH64_ADVSIMD_NAME "AdvSIMD Scalar Operation Optimization" |
63 | |
64 | namespace { |
65 | class AArch64AdvSIMDScalar : public MachineFunctionPass { |
66 | MachineRegisterInfo *MRI; |
67 | const TargetInstrInfo *TII; |
68 | |
69 | private: |
70 | // isProfitableToTransform - Predicate function to determine whether an |
71 | // instruction should be transformed to its equivalent AdvSIMD scalar |
72 | // instruction. "add Xd, Xn, Xm" ==> "add Dd, Da, Db", for example. |
73 | bool isProfitableToTransform(const MachineInstr &MI) const; |
74 | |
75 | // transformInstruction - Perform the transformation of an instruction |
76 | // to its equivalant AdvSIMD scalar instruction. Update inputs and outputs |
77 | // to be the correct register class, minimizing cross-class copies. |
78 | void transformInstruction(MachineInstr &MI); |
79 | |
80 | // processMachineBasicBlock - Main optimzation loop. |
81 | bool processMachineBasicBlock(MachineBasicBlock *MBB); |
82 | |
83 | public: |
84 | static char ID; // Pass identification, replacement for typeid. |
85 | explicit AArch64AdvSIMDScalar() : MachineFunctionPass(ID) { |
86 | initializeAArch64AdvSIMDScalarPass(*PassRegistry::getPassRegistry()); |
87 | } |
88 | |
89 | bool runOnMachineFunction(MachineFunction &F) override; |
90 | |
91 | StringRef getPassName() const override { return AARCH64_ADVSIMD_NAME; } |
92 | |
93 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
94 | AU.setPreservesCFG(); |
95 | MachineFunctionPass::getAnalysisUsage(AU); |
96 | } |
97 | }; |
98 | char AArch64AdvSIMDScalar::ID = 0; |
99 | } // end anonymous namespace |
100 | |
101 | INITIALIZE_PASS(AArch64AdvSIMDScalar, "aarch64-simd-scalar" , |
102 | AARCH64_ADVSIMD_NAME, false, false) |
103 | |
104 | static bool isGPR64(unsigned Reg, unsigned SubReg, |
105 | const MachineRegisterInfo *MRI) { |
106 | if (SubReg) |
107 | return false; |
108 | if (Register::isVirtualRegister(Reg)) |
109 | return MRI->getRegClass(Reg)->hasSuperClassEq(RC: &AArch64::GPR64RegClass); |
110 | return AArch64::GPR64RegClass.contains(Reg); |
111 | } |
112 | |
113 | static bool isFPR64(unsigned Reg, unsigned SubReg, |
114 | const MachineRegisterInfo *MRI) { |
115 | if (Register::isVirtualRegister(Reg)) |
116 | return (MRI->getRegClass(Reg)->hasSuperClassEq(RC: &AArch64::FPR64RegClass) && |
117 | SubReg == 0) || |
118 | (MRI->getRegClass(Reg)->hasSuperClassEq(RC: &AArch64::FPR128RegClass) && |
119 | SubReg == AArch64::dsub); |
120 | // Physical register references just check the register class directly. |
121 | return (AArch64::FPR64RegClass.contains(Reg) && SubReg == 0) || |
122 | (AArch64::FPR128RegClass.contains(Reg) && SubReg == AArch64::dsub); |
123 | } |
124 | |
125 | // getSrcFromCopy - Get the original source register for a GPR64 <--> FPR64 |
126 | // copy instruction. Return nullptr if the instruction is not a copy. |
127 | static MachineOperand *getSrcFromCopy(MachineInstr *MI, |
128 | const MachineRegisterInfo *MRI, |
129 | unsigned &SubReg) { |
130 | SubReg = 0; |
131 | // The "FMOV Xd, Dn" instruction is the typical form. |
132 | if (MI->getOpcode() == AArch64::FMOVDXr || |
133 | MI->getOpcode() == AArch64::FMOVXDr) |
134 | return &MI->getOperand(i: 1); |
135 | // A lane zero extract "UMOV.d Xd, Vn[0]" is equivalent. We shouldn't see |
136 | // these at this stage, but it's easy to check for. |
137 | if (MI->getOpcode() == AArch64::UMOVvi64 && MI->getOperand(i: 2).getImm() == 0) { |
138 | SubReg = AArch64::dsub; |
139 | return &MI->getOperand(i: 1); |
140 | } |
141 | // Or just a plain COPY instruction. This can be directly to/from FPR64, |
142 | // or it can be a dsub subreg reference to an FPR128. |
143 | if (MI->getOpcode() == AArch64::COPY) { |
144 | if (isFPR64(Reg: MI->getOperand(i: 0).getReg(), SubReg: MI->getOperand(i: 0).getSubReg(), |
145 | MRI) && |
146 | isGPR64(Reg: MI->getOperand(i: 1).getReg(), SubReg: MI->getOperand(i: 1).getSubReg(), MRI)) |
147 | return &MI->getOperand(i: 1); |
148 | if (isGPR64(Reg: MI->getOperand(i: 0).getReg(), SubReg: MI->getOperand(i: 0).getSubReg(), |
149 | MRI) && |
150 | isFPR64(Reg: MI->getOperand(i: 1).getReg(), SubReg: MI->getOperand(i: 1).getSubReg(), |
151 | MRI)) { |
152 | SubReg = MI->getOperand(i: 1).getSubReg(); |
153 | return &MI->getOperand(i: 1); |
154 | } |
155 | } |
156 | |
157 | // Otherwise, this is some other kind of instruction. |
158 | return nullptr; |
159 | } |
160 | |
161 | // getTransformOpcode - For any opcode for which there is an AdvSIMD equivalent |
162 | // that we're considering transforming to, return that AdvSIMD opcode. For all |
163 | // others, return the original opcode. |
164 | static unsigned getTransformOpcode(unsigned Opc) { |
165 | switch (Opc) { |
166 | default: |
167 | break; |
168 | // FIXME: Lots more possibilities. |
169 | case AArch64::ADDXrr: |
170 | return AArch64::ADDv1i64; |
171 | case AArch64::SUBXrr: |
172 | return AArch64::SUBv1i64; |
173 | case AArch64::ANDXrr: |
174 | return AArch64::ANDv8i8; |
175 | case AArch64::EORXrr: |
176 | return AArch64::EORv8i8; |
177 | case AArch64::ORRXrr: |
178 | return AArch64::ORRv8i8; |
179 | } |
180 | // No AdvSIMD equivalent, so just return the original opcode. |
181 | return Opc; |
182 | } |
183 | |
184 | static bool isTransformable(const MachineInstr &MI) { |
185 | unsigned Opc = MI.getOpcode(); |
186 | return Opc != getTransformOpcode(Opc); |
187 | } |
188 | |
189 | // isProfitableToTransform - Predicate function to determine whether an |
190 | // instruction should be transformed to its equivalent AdvSIMD scalar |
191 | // instruction. "add Xd, Xn, Xm" ==> "add Dd, Da, Db", for example. |
192 | bool AArch64AdvSIMDScalar::isProfitableToTransform( |
193 | const MachineInstr &MI) const { |
194 | // If this instruction isn't eligible to be transformed (no SIMD equivalent), |
195 | // early exit since that's the common case. |
196 | if (!isTransformable(MI)) |
197 | return false; |
198 | |
199 | // Count the number of copies we'll need to add and approximate the number |
200 | // of copies that a transform will enable us to remove. |
201 | unsigned NumNewCopies = 3; |
202 | unsigned NumRemovableCopies = 0; |
203 | |
204 | Register OrigSrc0 = MI.getOperand(i: 1).getReg(); |
205 | Register OrigSrc1 = MI.getOperand(i: 2).getReg(); |
206 | unsigned SubReg0; |
207 | unsigned SubReg1; |
208 | if (!MRI->def_empty(RegNo: OrigSrc0)) { |
209 | MachineRegisterInfo::def_instr_iterator Def = |
210 | MRI->def_instr_begin(RegNo: OrigSrc0); |
211 | assert(std::next(Def) == MRI->def_instr_end() && "Multiple def in SSA!" ); |
212 | MachineOperand *MOSrc0 = getSrcFromCopy(MI: &*Def, MRI, SubReg&: SubReg0); |
213 | // If the source was from a copy, we don't need to insert a new copy. |
214 | if (MOSrc0) |
215 | --NumNewCopies; |
216 | // If there are no other users of the original source, we can delete |
217 | // that instruction. |
218 | if (MOSrc0 && MRI->hasOneNonDBGUse(RegNo: OrigSrc0)) |
219 | ++NumRemovableCopies; |
220 | } |
221 | if (!MRI->def_empty(RegNo: OrigSrc1)) { |
222 | MachineRegisterInfo::def_instr_iterator Def = |
223 | MRI->def_instr_begin(RegNo: OrigSrc1); |
224 | assert(std::next(Def) == MRI->def_instr_end() && "Multiple def in SSA!" ); |
225 | MachineOperand *MOSrc1 = getSrcFromCopy(MI: &*Def, MRI, SubReg&: SubReg1); |
226 | if (MOSrc1) |
227 | --NumNewCopies; |
228 | // If there are no other users of the original source, we can delete |
229 | // that instruction. |
230 | if (MOSrc1 && MRI->hasOneNonDBGUse(RegNo: OrigSrc1)) |
231 | ++NumRemovableCopies; |
232 | } |
233 | |
234 | // If any of the uses of the original instructions is a cross class copy, |
235 | // that's a copy that will be removable if we transform. Likewise, if |
236 | // any of the uses is a transformable instruction, it's likely the tranforms |
237 | // will chain, enabling us to save a copy there, too. This is an aggressive |
238 | // heuristic that approximates the graph based cost analysis described above. |
239 | Register Dst = MI.getOperand(i: 0).getReg(); |
240 | bool AllUsesAreCopies = true; |
241 | for (MachineRegisterInfo::use_instr_nodbg_iterator |
242 | Use = MRI->use_instr_nodbg_begin(RegNo: Dst), |
243 | E = MRI->use_instr_nodbg_end(); |
244 | Use != E; ++Use) { |
245 | unsigned SubReg; |
246 | if (getSrcFromCopy(MI: &*Use, MRI, SubReg) || isTransformable(MI: *Use)) |
247 | ++NumRemovableCopies; |
248 | // If the use is an INSERT_SUBREG, that's still something that can |
249 | // directly use the FPR64, so we don't invalidate AllUsesAreCopies. It's |
250 | // preferable to have it use the FPR64 in most cases, as if the source |
251 | // vector is an IMPLICIT_DEF, the INSERT_SUBREG just goes away entirely. |
252 | // Ditto for a lane insert. |
253 | else if (Use->getOpcode() == AArch64::INSERT_SUBREG || |
254 | Use->getOpcode() == AArch64::INSvi64gpr) |
255 | ; |
256 | else |
257 | AllUsesAreCopies = false; |
258 | } |
259 | // If all of the uses of the original destination register are copies to |
260 | // FPR64, then we won't end up having a new copy back to GPR64 either. |
261 | if (AllUsesAreCopies) |
262 | --NumNewCopies; |
263 | |
264 | // If a transform will not increase the number of cross-class copies required, |
265 | // return true. |
266 | if (NumNewCopies <= NumRemovableCopies) |
267 | return true; |
268 | |
269 | // Finally, even if we otherwise wouldn't transform, check if we're forcing |
270 | // transformation of everything. |
271 | return TransformAll; |
272 | } |
273 | |
274 | static MachineInstr *insertCopy(const TargetInstrInfo *TII, MachineInstr &MI, |
275 | unsigned Dst, unsigned Src, bool IsKill) { |
276 | MachineInstrBuilder MIB = BuildMI(BB&: *MI.getParent(), I&: MI, MIMD: MI.getDebugLoc(), |
277 | MCID: TII->get(Opcode: AArch64::COPY), DestReg: Dst) |
278 | .addReg(RegNo: Src, flags: getKillRegState(B: IsKill)); |
279 | LLVM_DEBUG(dbgs() << " adding copy: " << *MIB); |
280 | ++NumCopiesInserted; |
281 | return MIB; |
282 | } |
283 | |
284 | // transformInstruction - Perform the transformation of an instruction |
285 | // to its equivalant AdvSIMD scalar instruction. Update inputs and outputs |
286 | // to be the correct register class, minimizing cross-class copies. |
287 | void AArch64AdvSIMDScalar::transformInstruction(MachineInstr &MI) { |
288 | LLVM_DEBUG(dbgs() << "Scalar transform: " << MI); |
289 | |
290 | MachineBasicBlock *MBB = MI.getParent(); |
291 | unsigned OldOpc = MI.getOpcode(); |
292 | unsigned NewOpc = getTransformOpcode(Opc: OldOpc); |
293 | assert(OldOpc != NewOpc && "transform an instruction to itself?!" ); |
294 | |
295 | // Check if we need a copy for the source registers. |
296 | Register OrigSrc0 = MI.getOperand(i: 1).getReg(); |
297 | Register OrigSrc1 = MI.getOperand(i: 2).getReg(); |
298 | unsigned Src0 = 0, SubReg0; |
299 | unsigned Src1 = 0, SubReg1; |
300 | bool KillSrc0 = false, KillSrc1 = false; |
301 | if (!MRI->def_empty(RegNo: OrigSrc0)) { |
302 | MachineRegisterInfo::def_instr_iterator Def = |
303 | MRI->def_instr_begin(RegNo: OrigSrc0); |
304 | assert(std::next(Def) == MRI->def_instr_end() && "Multiple def in SSA!" ); |
305 | MachineOperand *MOSrc0 = getSrcFromCopy(MI: &*Def, MRI, SubReg&: SubReg0); |
306 | // If there are no other users of the original source, we can delete |
307 | // that instruction. |
308 | if (MOSrc0) { |
309 | Src0 = MOSrc0->getReg(); |
310 | KillSrc0 = MOSrc0->isKill(); |
311 | // Src0 is going to be reused, thus, it cannot be killed anymore. |
312 | MOSrc0->setIsKill(false); |
313 | if (MRI->hasOneNonDBGUse(RegNo: OrigSrc0)) { |
314 | assert(MOSrc0 && "Can't delete copy w/o a valid original source!" ); |
315 | Def->eraseFromParent(); |
316 | ++NumCopiesDeleted; |
317 | } |
318 | } |
319 | } |
320 | if (!MRI->def_empty(RegNo: OrigSrc1)) { |
321 | MachineRegisterInfo::def_instr_iterator Def = |
322 | MRI->def_instr_begin(RegNo: OrigSrc1); |
323 | assert(std::next(Def) == MRI->def_instr_end() && "Multiple def in SSA!" ); |
324 | MachineOperand *MOSrc1 = getSrcFromCopy(MI: &*Def, MRI, SubReg&: SubReg1); |
325 | // If there are no other users of the original source, we can delete |
326 | // that instruction. |
327 | if (MOSrc1) { |
328 | Src1 = MOSrc1->getReg(); |
329 | KillSrc1 = MOSrc1->isKill(); |
330 | // Src0 is going to be reused, thus, it cannot be killed anymore. |
331 | MOSrc1->setIsKill(false); |
332 | if (MRI->hasOneNonDBGUse(RegNo: OrigSrc1)) { |
333 | assert(MOSrc1 && "Can't delete copy w/o a valid original source!" ); |
334 | Def->eraseFromParent(); |
335 | ++NumCopiesDeleted; |
336 | } |
337 | } |
338 | } |
339 | // If we weren't able to reference the original source directly, create a |
340 | // copy. |
341 | if (!Src0) { |
342 | SubReg0 = 0; |
343 | Src0 = MRI->createVirtualRegister(RegClass: &AArch64::FPR64RegClass); |
344 | insertCopy(TII, MI, Dst: Src0, Src: OrigSrc0, IsKill: KillSrc0); |
345 | KillSrc0 = true; |
346 | } |
347 | if (!Src1) { |
348 | SubReg1 = 0; |
349 | Src1 = MRI->createVirtualRegister(RegClass: &AArch64::FPR64RegClass); |
350 | insertCopy(TII, MI, Dst: Src1, Src: OrigSrc1, IsKill: KillSrc1); |
351 | KillSrc1 = true; |
352 | } |
353 | |
354 | // Create a vreg for the destination. |
355 | // FIXME: No need to do this if the ultimate user expects an FPR64. |
356 | // Check for that and avoid the copy if possible. |
357 | Register Dst = MRI->createVirtualRegister(RegClass: &AArch64::FPR64RegClass); |
358 | |
359 | // For now, all of the new instructions have the same simple three-register |
360 | // form, so no need to special case based on what instruction we're |
361 | // building. |
362 | BuildMI(BB&: *MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg: Dst) |
363 | .addReg(RegNo: Src0, flags: getKillRegState(B: KillSrc0), SubReg: SubReg0) |
364 | .addReg(RegNo: Src1, flags: getKillRegState(B: KillSrc1), SubReg: SubReg1); |
365 | |
366 | // Now copy the result back out to a GPR. |
367 | // FIXME: Try to avoid this if all uses could actually just use the FPR64 |
368 | // directly. |
369 | insertCopy(TII, MI, Dst: MI.getOperand(i: 0).getReg(), Src: Dst, IsKill: true); |
370 | |
371 | // Erase the old instruction. |
372 | MI.eraseFromParent(); |
373 | |
374 | ++NumScalarInsnsUsed; |
375 | } |
376 | |
377 | // processMachineBasicBlock - Main optimzation loop. |
378 | bool AArch64AdvSIMDScalar::processMachineBasicBlock(MachineBasicBlock *MBB) { |
379 | bool Changed = false; |
380 | for (MachineInstr &MI : llvm::make_early_inc_range(Range&: *MBB)) { |
381 | if (isProfitableToTransform(MI)) { |
382 | transformInstruction(MI); |
383 | Changed = true; |
384 | } |
385 | } |
386 | return Changed; |
387 | } |
388 | |
389 | // runOnMachineFunction - Pass entry point from PassManager. |
390 | bool AArch64AdvSIMDScalar::runOnMachineFunction(MachineFunction &mf) { |
391 | bool Changed = false; |
392 | LLVM_DEBUG(dbgs() << "***** AArch64AdvSIMDScalar *****\n" ); |
393 | |
394 | if (skipFunction(F: mf.getFunction())) |
395 | return false; |
396 | |
397 | MRI = &mf.getRegInfo(); |
398 | TII = mf.getSubtarget().getInstrInfo(); |
399 | |
400 | // Just check things on a one-block-at-a-time basis. |
401 | for (MachineBasicBlock &MBB : mf) |
402 | if (processMachineBasicBlock(MBB: &MBB)) |
403 | Changed = true; |
404 | return Changed; |
405 | } |
406 | |
407 | // createAArch64AdvSIMDScalar - Factory function used by AArch64TargetMachine |
408 | // to add the pass to the PassManager. |
409 | FunctionPass *llvm::createAArch64AdvSIMDScalar() { |
410 | return new AArch64AdvSIMDScalar(); |
411 | } |
412 | |