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