1 | //=== AArch64PostSelectOptimize.cpp ---------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This pass does post-instruction-selection optimizations in the GlobalISel |
10 | // pipeline, before the rest of codegen runs. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "AArch64.h" |
15 | #include "AArch64TargetMachine.h" |
16 | #include "MCTargetDesc/AArch64MCTargetDesc.h" |
17 | #include "llvm/ADT/STLExtras.h" |
18 | #include "llvm/CodeGen/GlobalISel/Utils.h" |
19 | #include "llvm/CodeGen/MachineBasicBlock.h" |
20 | #include "llvm/CodeGen/MachineFunctionPass.h" |
21 | #include "llvm/CodeGen/MachineInstr.h" |
22 | #include "llvm/CodeGen/MachineOperand.h" |
23 | #include "llvm/CodeGen/TargetPassConfig.h" |
24 | #include "llvm/Support/Debug.h" |
25 | #include "llvm/Support/ErrorHandling.h" |
26 | |
27 | #define DEBUG_TYPE "aarch64-post-select-optimize" |
28 | |
29 | using namespace llvm; |
30 | |
31 | namespace { |
32 | class AArch64PostSelectOptimize : public MachineFunctionPass { |
33 | public: |
34 | static char ID; |
35 | |
36 | AArch64PostSelectOptimize() : MachineFunctionPass(ID) {} |
37 | |
38 | StringRef getPassName() const override { |
39 | return "AArch64 Post Select Optimizer" ; |
40 | } |
41 | |
42 | bool runOnMachineFunction(MachineFunction &MF) override; |
43 | |
44 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
45 | |
46 | private: |
47 | bool optimizeNZCVDefs(MachineBasicBlock &MBB); |
48 | bool doPeepholeOpts(MachineBasicBlock &MBB); |
49 | /// Look for cross regclass copies that can be trivially eliminated. |
50 | bool foldSimpleCrossClassCopies(MachineInstr &MI); |
51 | bool foldCopyDup(MachineInstr &MI); |
52 | }; |
53 | } // end anonymous namespace |
54 | |
55 | void AArch64PostSelectOptimize::getAnalysisUsage(AnalysisUsage &AU) const { |
56 | AU.addRequired<TargetPassConfig>(); |
57 | AU.setPreservesCFG(); |
58 | getSelectionDAGFallbackAnalysisUsage(AU); |
59 | MachineFunctionPass::getAnalysisUsage(AU); |
60 | } |
61 | |
62 | unsigned getNonFlagSettingVariant(unsigned Opc) { |
63 | switch (Opc) { |
64 | default: |
65 | return 0; |
66 | case AArch64::SUBSXrr: |
67 | return AArch64::SUBXrr; |
68 | case AArch64::SUBSWrr: |
69 | return AArch64::SUBWrr; |
70 | case AArch64::SUBSXrs: |
71 | return AArch64::SUBXrs; |
72 | case AArch64::SUBSWrs: |
73 | return AArch64::SUBWrs; |
74 | case AArch64::SUBSXri: |
75 | return AArch64::SUBXri; |
76 | case AArch64::SUBSWri: |
77 | return AArch64::SUBWri; |
78 | case AArch64::ADDSXrr: |
79 | return AArch64::ADDXrr; |
80 | case AArch64::ADDSWrr: |
81 | return AArch64::ADDWrr; |
82 | case AArch64::ADDSXrs: |
83 | return AArch64::ADDXrs; |
84 | case AArch64::ADDSWrs: |
85 | return AArch64::ADDWrs; |
86 | case AArch64::ADDSXri: |
87 | return AArch64::ADDXri; |
88 | case AArch64::ADDSWri: |
89 | return AArch64::ADDWri; |
90 | case AArch64::SBCSXr: |
91 | return AArch64::SBCXr; |
92 | case AArch64::SBCSWr: |
93 | return AArch64::SBCWr; |
94 | case AArch64::ADCSXr: |
95 | return AArch64::ADCXr; |
96 | case AArch64::ADCSWr: |
97 | return AArch64::ADCWr; |
98 | } |
99 | } |
100 | |
101 | bool AArch64PostSelectOptimize::doPeepholeOpts(MachineBasicBlock &MBB) { |
102 | bool Changed = false; |
103 | for (auto &MI : make_early_inc_range(Range&: MBB)) { |
104 | bool CurrentIterChanged = foldSimpleCrossClassCopies(MI); |
105 | if (!CurrentIterChanged) |
106 | CurrentIterChanged |= foldCopyDup(MI); |
107 | Changed |= CurrentIterChanged; |
108 | } |
109 | return Changed; |
110 | } |
111 | |
112 | bool AArch64PostSelectOptimize::foldSimpleCrossClassCopies(MachineInstr &MI) { |
113 | auto *MF = MI.getMF(); |
114 | auto &MRI = MF->getRegInfo(); |
115 | |
116 | if (!MI.isCopy()) |
117 | return false; |
118 | |
119 | if (MI.getOperand(i: 1).getSubReg()) |
120 | return false; // Don't deal with subreg copies |
121 | |
122 | Register Src = MI.getOperand(i: 1).getReg(); |
123 | Register Dst = MI.getOperand(i: 0).getReg(); |
124 | |
125 | if (Src.isPhysical() || Dst.isPhysical()) |
126 | return false; |
127 | |
128 | const TargetRegisterClass *SrcRC = MRI.getRegClass(Reg: Src); |
129 | const TargetRegisterClass *DstRC = MRI.getRegClass(Reg: Dst); |
130 | |
131 | if (SrcRC == DstRC) |
132 | return false; |
133 | |
134 | |
135 | if (SrcRC->hasSubClass(RC: DstRC)) { |
136 | // This is the case where the source class is a superclass of the dest, so |
137 | // if the copy is the only user of the source, we can just constrain the |
138 | // source reg to the dest class. |
139 | |
140 | if (!MRI.hasOneNonDBGUse(RegNo: Src)) |
141 | return false; // Only constrain single uses of the source. |
142 | |
143 | // Constrain to dst reg class as long as it's not a weird class that only |
144 | // has a few registers. |
145 | if (!MRI.constrainRegClass(Reg: Src, RC: DstRC, /* MinNumRegs */ 25)) |
146 | return false; |
147 | } else if (DstRC->hasSubClass(RC: SrcRC)) { |
148 | // This is the inverse case, where the destination class is a superclass of |
149 | // the source. Here, if the copy is the only user, we can just constrain |
150 | // the user of the copy to use the smaller class of the source. |
151 | } else { |
152 | return false; |
153 | } |
154 | |
155 | MRI.replaceRegWith(FromReg: Dst, ToReg: Src); |
156 | MI.eraseFromParent(); |
157 | return true; |
158 | } |
159 | |
160 | bool AArch64PostSelectOptimize::foldCopyDup(MachineInstr &MI) { |
161 | if (!MI.isCopy()) |
162 | return false; |
163 | |
164 | auto *MF = MI.getMF(); |
165 | auto &MRI = MF->getRegInfo(); |
166 | auto *TII = MF->getSubtarget().getInstrInfo(); |
167 | |
168 | // Optimize COPY(y:GPR, DUP(x:FPR, i)) -> UMOV(y:GPR, x:FPR, i). |
169 | // Here Dst is y and Src is the result of DUP. |
170 | Register Dst = MI.getOperand(i: 0).getReg(); |
171 | Register Src = MI.getOperand(i: 1).getReg(); |
172 | |
173 | if (!Dst.isVirtual() || !Src.isVirtual()) |
174 | return false; |
175 | |
176 | auto TryMatchDUP = [&](const TargetRegisterClass *GPRRegClass, |
177 | const TargetRegisterClass *FPRRegClass, unsigned DUP, |
178 | unsigned UMOV) { |
179 | if (MRI.getRegClassOrNull(Reg: Dst) != GPRRegClass || |
180 | MRI.getRegClassOrNull(Reg: Src) != FPRRegClass) |
181 | return false; |
182 | |
183 | // There is a special case when one of the uses is COPY(z:FPR, y:GPR). |
184 | // In this case, we get COPY(z:FPR, COPY(y:GPR, DUP(x:FPR, i))), which can |
185 | // be folded by peephole-opt into just DUP(z:FPR, i), so this transform is |
186 | // not worthwhile in that case. |
187 | for (auto &Use : MRI.use_nodbg_instructions(Reg: Dst)) { |
188 | if (!Use.isCopy()) |
189 | continue; |
190 | |
191 | Register UseOp0 = Use.getOperand(i: 0).getReg(); |
192 | Register UseOp1 = Use.getOperand(i: 1).getReg(); |
193 | if (UseOp0.isPhysical() || UseOp1.isPhysical()) |
194 | return false; |
195 | |
196 | if (MRI.getRegClassOrNull(Reg: UseOp0) == FPRRegClass && |
197 | MRI.getRegClassOrNull(Reg: UseOp1) == GPRRegClass) |
198 | return false; |
199 | } |
200 | |
201 | MachineInstr *SrcMI = MRI.getUniqueVRegDef(Reg: Src); |
202 | if (!SrcMI || SrcMI->getOpcode() != DUP || !MRI.hasOneNonDBGUse(RegNo: Src)) |
203 | return false; |
204 | |
205 | Register DupSrc = SrcMI->getOperand(i: 1).getReg(); |
206 | int64_t DupImm = SrcMI->getOperand(i: 2).getImm(); |
207 | |
208 | BuildMI(BB&: *MI.getParent(), I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: UMOV), DestReg: Dst) |
209 | .addReg(RegNo: DupSrc) |
210 | .addImm(Val: DupImm); |
211 | SrcMI->eraseFromParent(); |
212 | MI.eraseFromParent(); |
213 | return true; |
214 | }; |
215 | |
216 | return TryMatchDUP(&AArch64::GPR32RegClass, &AArch64::FPR32RegClass, |
217 | AArch64::DUPi32, AArch64::UMOVvi32) || |
218 | TryMatchDUP(&AArch64::GPR64RegClass, &AArch64::FPR64RegClass, |
219 | AArch64::DUPi64, AArch64::UMOVvi64); |
220 | } |
221 | |
222 | bool AArch64PostSelectOptimize::optimizeNZCVDefs(MachineBasicBlock &MBB) { |
223 | // If we find a dead NZCV implicit-def, we |
224 | // - try to convert the operation to a non-flag-setting equivalent |
225 | // - or mark the def as dead to aid later peephole optimizations. |
226 | |
227 | // Use cases: |
228 | // 1) |
229 | // Consider the following code: |
230 | // FCMPSrr %0, %1, implicit-def $nzcv |
231 | // %sel1:gpr32 = CSELWr %_, %_, 12, implicit $nzcv |
232 | // %sub:gpr32 = SUBSWrr %_, %_, implicit-def $nzcv |
233 | // FCMPSrr %0, %1, implicit-def $nzcv |
234 | // %sel2:gpr32 = CSELWr %_, %_, 12, implicit $nzcv |
235 | // This kind of code where we have 2 FCMPs each feeding a CSEL can happen |
236 | // when we have a single IR fcmp being used by two selects. During selection, |
237 | // to ensure that there can be no clobbering of nzcv between the fcmp and the |
238 | // csel, we have to generate an fcmp immediately before each csel is |
239 | // selected. |
240 | // However, often we can essentially CSE these together later in MachineCSE. |
241 | // This doesn't work though if there are unrelated flag-setting instructions |
242 | // in between the two FCMPs. In this case, the SUBS defines NZCV |
243 | // but it doesn't have any users, being overwritten by the second FCMP. |
244 | // |
245 | // 2) |
246 | // The instruction selector always emits the flag-setting variant of ADC/SBC |
247 | // while selecting G_UADDE/G_SADDE/G_USUBE/G_SSUBE. If the carry-out of these |
248 | // instructions is never used, we can switch to the non-flag-setting variant. |
249 | |
250 | bool Changed = false; |
251 | auto &MF = *MBB.getParent(); |
252 | auto &Subtarget = MF.getSubtarget(); |
253 | const auto &TII = Subtarget.getInstrInfo(); |
254 | auto TRI = Subtarget.getRegisterInfo(); |
255 | auto RBI = Subtarget.getRegBankInfo(); |
256 | auto &MRI = MF.getRegInfo(); |
257 | |
258 | LiveRegUnits LRU(*MBB.getParent()->getSubtarget().getRegisterInfo()); |
259 | LRU.addLiveOuts(MBB); |
260 | |
261 | for (auto &II : instructionsWithoutDebug(It: MBB.rbegin(), End: MBB.rend())) { |
262 | bool NZCVDead = LRU.available(Reg: AArch64::NZCV); |
263 | if (NZCVDead && II.definesRegister(Reg: AArch64::NZCV, /*TRI=*/nullptr)) { |
264 | // The instruction defines NZCV, but NZCV is dead. |
265 | unsigned NewOpc = getNonFlagSettingVariant(Opc: II.getOpcode()); |
266 | int DeadNZCVIdx = |
267 | II.findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr); |
268 | if (DeadNZCVIdx != -1) { |
269 | if (NewOpc) { |
270 | // If there is an equivalent non-flag-setting op, we convert. |
271 | LLVM_DEBUG(dbgs() << "Post-select optimizer: converting flag-setting " |
272 | "op: " |
273 | << II); |
274 | II.setDesc(TII->get(Opcode: NewOpc)); |
275 | II.removeOperand(OpNo: DeadNZCVIdx); |
276 | // Changing the opcode can result in differing regclass requirements, |
277 | // e.g. SUBSWri uses gpr32 for the dest, whereas SUBWri uses gpr32sp. |
278 | // Constrain the regclasses, possibly introducing a copy. |
279 | constrainOperandRegClass(MF, TRI: *TRI, MRI, TII: *TII, RBI: *RBI, InsertPt&: II, II: II.getDesc(), |
280 | RegMO&: II.getOperand(i: 0), OpIdx: 0); |
281 | Changed |= true; |
282 | } else { |
283 | // Otherwise, we just set the nzcv imp-def operand to be dead, so the |
284 | // peephole optimizations can optimize them further. |
285 | II.getOperand(i: DeadNZCVIdx).setIsDead(); |
286 | } |
287 | } |
288 | } |
289 | LRU.stepBackward(MI: II); |
290 | } |
291 | return Changed; |
292 | } |
293 | |
294 | bool AArch64PostSelectOptimize::runOnMachineFunction(MachineFunction &MF) { |
295 | if (MF.getProperties().hasFailedISel()) |
296 | return false; |
297 | assert(MF.getProperties().hasSelected() && "Expected a selected MF" ); |
298 | |
299 | bool Changed = false; |
300 | for (auto &BB : MF) { |
301 | Changed |= optimizeNZCVDefs(MBB&: BB); |
302 | Changed |= doPeepholeOpts(MBB&: BB); |
303 | } |
304 | return Changed; |
305 | } |
306 | |
307 | char AArch64PostSelectOptimize::ID = 0; |
308 | INITIALIZE_PASS_BEGIN(AArch64PostSelectOptimize, DEBUG_TYPE, |
309 | "Optimize AArch64 selected instructions" , |
310 | false, false) |
311 | INITIALIZE_PASS_END(AArch64PostSelectOptimize, DEBUG_TYPE, |
312 | "Optimize AArch64 selected instructions" , false, |
313 | false) |
314 | |
315 | namespace llvm { |
316 | FunctionPass *createAArch64PostSelectOptimize() { |
317 | return new AArch64PostSelectOptimize(); |
318 | } |
319 | } // end namespace llvm |
320 | |