1 | //===-- SystemZShortenInst.cpp - Instruction-shortening pass --------------===// |
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 tries to replace instructions with shorter forms. For example, |
10 | // IILF can be replaced with LLILL or LLILH if the constant fits and if the |
11 | // other 32 bits of the GR64 destination are not live. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "SystemZTargetMachine.h" |
16 | #include "llvm/CodeGen/LiveRegUnits.h" |
17 | #include "llvm/CodeGen/MachineFunctionPass.h" |
18 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
19 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
20 | |
21 | using namespace llvm; |
22 | |
23 | #define DEBUG_TYPE "systemz-shorten-inst" |
24 | |
25 | namespace { |
26 | class SystemZShortenInst : public MachineFunctionPass { |
27 | public: |
28 | static char ID; |
29 | SystemZShortenInst(); |
30 | |
31 | bool processBlock(MachineBasicBlock &MBB); |
32 | bool runOnMachineFunction(MachineFunction &F) override; |
33 | MachineFunctionProperties getRequiredProperties() const override { |
34 | return MachineFunctionProperties().setNoVRegs(); |
35 | } |
36 | |
37 | private: |
38 | bool shortenIIF(MachineInstr &MI, unsigned LLIxL, unsigned LLIxH); |
39 | bool shortenOn0(MachineInstr &MI, unsigned Opcode); |
40 | bool shortenOn01(MachineInstr &MI, unsigned Opcode); |
41 | bool shortenOn001(MachineInstr &MI, unsigned Opcode); |
42 | bool shortenOn001AddCC(MachineInstr &MI, unsigned Opcode); |
43 | bool shortenFPConv(MachineInstr &MI, unsigned Opcode); |
44 | bool shortenFusedFPOp(MachineInstr &MI, unsigned Opcode); |
45 | |
46 | const SystemZInstrInfo *TII; |
47 | const TargetRegisterInfo *TRI; |
48 | LiveRegUnits LiveRegs; |
49 | }; |
50 | |
51 | char SystemZShortenInst::ID = 0; |
52 | } // end anonymous namespace |
53 | |
54 | INITIALIZE_PASS(SystemZShortenInst, DEBUG_TYPE, |
55 | "SystemZ Instruction Shortening" , false, false) |
56 | |
57 | FunctionPass *llvm::createSystemZShortenInstPass(SystemZTargetMachine &TM) { |
58 | return new SystemZShortenInst(); |
59 | } |
60 | |
61 | SystemZShortenInst::SystemZShortenInst() |
62 | : MachineFunctionPass(ID), TII(nullptr) {} |
63 | |
64 | // Tie operands if MI has become a two-address instruction. |
65 | static void tieOpsIfNeeded(MachineInstr &MI) { |
66 | if (MI.getDesc().getOperandConstraint(OpNum: 1, Constraint: MCOI::TIED_TO) == 0 && |
67 | !MI.getOperand(i: 0).isTied()) |
68 | MI.tieOperands(DefIdx: 0, UseIdx: 1); |
69 | } |
70 | |
71 | // MI loads one word of a GPR using an IIxF instruction and LLIxL and LLIxH |
72 | // are the halfword immediate loads for the same word. Try to use one of them |
73 | // instead of IIxF. |
74 | bool SystemZShortenInst::shortenIIF(MachineInstr &MI, unsigned LLIxL, |
75 | unsigned LLIxH) { |
76 | Register Reg = MI.getOperand(i: 0).getReg(); |
77 | // The new opcode will clear the other half of the GR64 reg, so |
78 | // cancel if that is live. |
79 | unsigned thisSubRegIdx = |
80 | (SystemZ::GRH32BitRegClass.contains(Reg) ? SystemZ::subreg_h32 |
81 | : SystemZ::subreg_l32); |
82 | unsigned otherSubRegIdx = |
83 | (thisSubRegIdx == SystemZ::subreg_l32 ? SystemZ::subreg_h32 |
84 | : SystemZ::subreg_l32); |
85 | unsigned GR64BitReg = |
86 | TRI->getMatchingSuperReg(Reg, SubIdx: thisSubRegIdx, RC: &SystemZ::GR64BitRegClass); |
87 | Register OtherReg = TRI->getSubReg(Reg: GR64BitReg, Idx: otherSubRegIdx); |
88 | if (!LiveRegs.available(Reg: OtherReg)) |
89 | return false; |
90 | |
91 | uint64_t Imm = MI.getOperand(i: 1).getImm(); |
92 | if (SystemZ::isImmLL(Val: Imm)) { |
93 | MI.setDesc(TII->get(Opcode: LLIxL)); |
94 | MI.getOperand(i: 0).setReg(SystemZMC::getRegAsGR64(Reg)); |
95 | return true; |
96 | } |
97 | if (SystemZ::isImmLH(Val: Imm)) { |
98 | MI.setDesc(TII->get(Opcode: LLIxH)); |
99 | MI.getOperand(i: 0).setReg(SystemZMC::getRegAsGR64(Reg)); |
100 | MI.getOperand(i: 1).setImm(Imm >> 16); |
101 | return true; |
102 | } |
103 | return false; |
104 | } |
105 | |
106 | // Change MI's opcode to Opcode if register operand 0 has a 4-bit encoding. |
107 | bool SystemZShortenInst::shortenOn0(MachineInstr &MI, unsigned Opcode) { |
108 | if (SystemZMC::getFirstReg(Reg: MI.getOperand(i: 0).getReg()) < 16) { |
109 | MI.setDesc(TII->get(Opcode)); |
110 | return true; |
111 | } |
112 | return false; |
113 | } |
114 | |
115 | // Change MI's opcode to Opcode if register operands 0 and 1 have a |
116 | // 4-bit encoding. |
117 | bool SystemZShortenInst::shortenOn01(MachineInstr &MI, unsigned Opcode) { |
118 | if (SystemZMC::getFirstReg(Reg: MI.getOperand(i: 0).getReg()) < 16 && |
119 | SystemZMC::getFirstReg(Reg: MI.getOperand(i: 1).getReg()) < 16) { |
120 | MI.setDesc(TII->get(Opcode)); |
121 | return true; |
122 | } |
123 | return false; |
124 | } |
125 | |
126 | // Change MI's opcode to Opcode if register operands 0, 1 and 2 have a |
127 | // 4-bit encoding and if operands 0 and 1 are tied. Also ties op 0 |
128 | // with op 1, if MI becomes 2-address. |
129 | bool SystemZShortenInst::shortenOn001(MachineInstr &MI, unsigned Opcode) { |
130 | if (SystemZMC::getFirstReg(Reg: MI.getOperand(i: 0).getReg()) < 16 && |
131 | MI.getOperand(i: 1).getReg() == MI.getOperand(i: 0).getReg() && |
132 | SystemZMC::getFirstReg(Reg: MI.getOperand(i: 2).getReg()) < 16) { |
133 | MI.setDesc(TII->get(Opcode)); |
134 | tieOpsIfNeeded(MI); |
135 | return true; |
136 | } |
137 | return false; |
138 | } |
139 | |
140 | // Calls shortenOn001 if CCLive is false. CC def operand is added in |
141 | // case of success. |
142 | bool SystemZShortenInst::shortenOn001AddCC(MachineInstr &MI, unsigned Opcode) { |
143 | if (LiveRegs.available(Reg: SystemZ::CC) && shortenOn001(MI, Opcode)) { |
144 | MachineInstrBuilder(*MI.getParent()->getParent(), &MI) |
145 | .addReg(RegNo: SystemZ::CC, flags: RegState::ImplicitDefine | RegState::Dead); |
146 | return true; |
147 | } |
148 | return false; |
149 | } |
150 | |
151 | // MI is a vector-style conversion instruction with the operand order: |
152 | // destination, source, exact-suppress, rounding-mode. If both registers |
153 | // have a 4-bit encoding then change it to Opcode, which has operand order: |
154 | // destination, rouding-mode, source, exact-suppress. |
155 | bool SystemZShortenInst::shortenFPConv(MachineInstr &MI, unsigned Opcode) { |
156 | if (SystemZMC::getFirstReg(Reg: MI.getOperand(i: 0).getReg()) < 16 && |
157 | SystemZMC::getFirstReg(Reg: MI.getOperand(i: 1).getReg()) < 16) { |
158 | MachineOperand Dest(MI.getOperand(i: 0)); |
159 | MachineOperand Src(MI.getOperand(i: 1)); |
160 | MachineOperand Suppress(MI.getOperand(i: 2)); |
161 | MachineOperand Mode(MI.getOperand(i: 3)); |
162 | MI.removeOperand(OpNo: 3); |
163 | MI.removeOperand(OpNo: 2); |
164 | MI.removeOperand(OpNo: 1); |
165 | MI.removeOperand(OpNo: 0); |
166 | MI.setDesc(TII->get(Opcode)); |
167 | MachineInstrBuilder(*MI.getParent()->getParent(), &MI) |
168 | .add(MO: Dest) |
169 | .add(MO: Mode) |
170 | .add(MO: Src) |
171 | .add(MO: Suppress); |
172 | return true; |
173 | } |
174 | return false; |
175 | } |
176 | |
177 | bool SystemZShortenInst::shortenFusedFPOp(MachineInstr &MI, unsigned Opcode) { |
178 | MachineOperand &DstMO = MI.getOperand(i: 0); |
179 | MachineOperand &LHSMO = MI.getOperand(i: 1); |
180 | MachineOperand &RHSMO = MI.getOperand(i: 2); |
181 | MachineOperand &AccMO = MI.getOperand(i: 3); |
182 | if (SystemZMC::getFirstReg(Reg: DstMO.getReg()) < 16 && |
183 | SystemZMC::getFirstReg(Reg: LHSMO.getReg()) < 16 && |
184 | SystemZMC::getFirstReg(Reg: RHSMO.getReg()) < 16 && |
185 | SystemZMC::getFirstReg(Reg: AccMO.getReg()) < 16 && |
186 | DstMO.getReg() == AccMO.getReg()) { |
187 | MachineOperand Lhs(LHSMO); |
188 | MachineOperand Rhs(RHSMO); |
189 | MachineOperand Src(AccMO); |
190 | MI.removeOperand(OpNo: 3); |
191 | MI.removeOperand(OpNo: 2); |
192 | MI.removeOperand(OpNo: 1); |
193 | MI.setDesc(TII->get(Opcode)); |
194 | MachineInstrBuilder(*MI.getParent()->getParent(), &MI) |
195 | .add(MO: Src) |
196 | .add(MO: Lhs) |
197 | .add(MO: Rhs); |
198 | return true; |
199 | } |
200 | return false; |
201 | } |
202 | |
203 | // Process all instructions in MBB. Return true if something changed. |
204 | bool SystemZShortenInst::processBlock(MachineBasicBlock &MBB) { |
205 | bool Changed = false; |
206 | |
207 | // Set up the set of live registers at the end of MBB (live out) |
208 | LiveRegs.clear(); |
209 | LiveRegs.addLiveOuts(MBB); |
210 | |
211 | // Iterate backwards through the block looking for instructions to change. |
212 | for (MachineInstr &MI : llvm::reverse(C&: MBB)) { |
213 | switch (MI.getOpcode()) { |
214 | case SystemZ::IILF: |
215 | Changed |= shortenIIF(MI, LLIxL: SystemZ::LLILL, LLIxH: SystemZ::LLILH); |
216 | break; |
217 | |
218 | case SystemZ::IIHF: |
219 | Changed |= shortenIIF(MI, LLIxL: SystemZ::LLIHL, LLIxH: SystemZ::LLIHH); |
220 | break; |
221 | |
222 | case SystemZ::WFADB: |
223 | Changed |= shortenOn001AddCC(MI, Opcode: SystemZ::ADBR); |
224 | break; |
225 | |
226 | case SystemZ::WFASB: |
227 | Changed |= shortenOn001AddCC(MI, Opcode: SystemZ::AEBR); |
228 | break; |
229 | |
230 | case SystemZ::WFDDB: |
231 | Changed |= shortenOn001(MI, Opcode: SystemZ::DDBR); |
232 | break; |
233 | |
234 | case SystemZ::WFDSB: |
235 | Changed |= shortenOn001(MI, Opcode: SystemZ::DEBR); |
236 | break; |
237 | |
238 | case SystemZ::WFIDB: |
239 | Changed |= shortenFPConv(MI, Opcode: SystemZ::FIDBRA); |
240 | break; |
241 | |
242 | case SystemZ::WFISB: |
243 | Changed |= shortenFPConv(MI, Opcode: SystemZ::FIEBRA); |
244 | break; |
245 | |
246 | case SystemZ::WLDEB: |
247 | Changed |= shortenOn01(MI, Opcode: SystemZ::LDEBR); |
248 | break; |
249 | |
250 | case SystemZ::WLEDB: |
251 | Changed |= shortenFPConv(MI, Opcode: SystemZ::LEDBRA); |
252 | break; |
253 | |
254 | case SystemZ::WFMDB: |
255 | Changed |= shortenOn001(MI, Opcode: SystemZ::MDBR); |
256 | break; |
257 | |
258 | case SystemZ::WFMSB: |
259 | Changed |= shortenOn001(MI, Opcode: SystemZ::MEEBR); |
260 | break; |
261 | |
262 | case SystemZ::WFMADB: |
263 | Changed |= shortenFusedFPOp(MI, Opcode: SystemZ::MADBR); |
264 | break; |
265 | |
266 | case SystemZ::WFMASB: |
267 | Changed |= shortenFusedFPOp(MI, Opcode: SystemZ::MAEBR); |
268 | break; |
269 | |
270 | case SystemZ::WFMSDB: |
271 | Changed |= shortenFusedFPOp(MI, Opcode: SystemZ::MSDBR); |
272 | break; |
273 | |
274 | case SystemZ::WFMSSB: |
275 | Changed |= shortenFusedFPOp(MI, Opcode: SystemZ::MSEBR); |
276 | break; |
277 | |
278 | case SystemZ::WFLCDB: |
279 | Changed |= shortenOn01(MI, Opcode: SystemZ::LCDFR); |
280 | break; |
281 | |
282 | case SystemZ::WFLCSB: |
283 | Changed |= shortenOn01(MI, Opcode: SystemZ::LCDFR_32); |
284 | break; |
285 | |
286 | case SystemZ::WFLNDB: |
287 | Changed |= shortenOn01(MI, Opcode: SystemZ::LNDFR); |
288 | break; |
289 | |
290 | case SystemZ::WFLNSB: |
291 | Changed |= shortenOn01(MI, Opcode: SystemZ::LNDFR_32); |
292 | break; |
293 | |
294 | case SystemZ::WFLPDB: |
295 | Changed |= shortenOn01(MI, Opcode: SystemZ::LPDFR); |
296 | break; |
297 | |
298 | case SystemZ::WFLPSB: |
299 | Changed |= shortenOn01(MI, Opcode: SystemZ::LPDFR_32); |
300 | break; |
301 | |
302 | case SystemZ::WFSQDB: |
303 | Changed |= shortenOn01(MI, Opcode: SystemZ::SQDBR); |
304 | break; |
305 | |
306 | case SystemZ::WFSQSB: |
307 | Changed |= shortenOn01(MI, Opcode: SystemZ::SQEBR); |
308 | break; |
309 | |
310 | case SystemZ::WFSDB: |
311 | Changed |= shortenOn001AddCC(MI, Opcode: SystemZ::SDBR); |
312 | break; |
313 | |
314 | case SystemZ::WFSSB: |
315 | Changed |= shortenOn001AddCC(MI, Opcode: SystemZ::SEBR); |
316 | break; |
317 | |
318 | case SystemZ::WFCDB: |
319 | Changed |= shortenOn01(MI, Opcode: SystemZ::CDBR); |
320 | break; |
321 | |
322 | case SystemZ::WFCSB: |
323 | Changed |= shortenOn01(MI, Opcode: SystemZ::CEBR); |
324 | break; |
325 | |
326 | case SystemZ::WFKDB: |
327 | Changed |= shortenOn01(MI, Opcode: SystemZ::KDBR); |
328 | break; |
329 | |
330 | case SystemZ::WFKSB: |
331 | Changed |= shortenOn01(MI, Opcode: SystemZ::KEBR); |
332 | break; |
333 | |
334 | case SystemZ::VL32: |
335 | // For z13 we prefer LDE over LE to avoid partial register dependencies. |
336 | Changed |= shortenOn0(MI, Opcode: SystemZ::LDE32); |
337 | break; |
338 | |
339 | case SystemZ::VST32: |
340 | Changed |= shortenOn0(MI, Opcode: SystemZ::STE); |
341 | break; |
342 | |
343 | case SystemZ::VL64: |
344 | Changed |= shortenOn0(MI, Opcode: SystemZ::LD); |
345 | break; |
346 | |
347 | case SystemZ::VST64: |
348 | Changed |= shortenOn0(MI, Opcode: SystemZ::STD); |
349 | break; |
350 | |
351 | default: { |
352 | int TwoOperandOpcode = SystemZ::getTwoOperandOpcode(Opcode: MI.getOpcode()); |
353 | if (TwoOperandOpcode == -1) |
354 | break; |
355 | |
356 | if ((MI.getOperand(i: 0).getReg() != MI.getOperand(i: 1).getReg()) && |
357 | (!MI.isCommutable() || |
358 | MI.getOperand(i: 0).getReg() != MI.getOperand(i: 2).getReg() || |
359 | !TII->commuteInstruction(MI, NewMI: false, OpIdx1: 1, OpIdx2: 2))) |
360 | break; |
361 | |
362 | MI.setDesc(TII->get(Opcode: TwoOperandOpcode)); |
363 | MI.tieOperands(DefIdx: 0, UseIdx: 1); |
364 | if (TwoOperandOpcode == SystemZ::SLL || |
365 | TwoOperandOpcode == SystemZ::SLA || |
366 | TwoOperandOpcode == SystemZ::SRL || |
367 | TwoOperandOpcode == SystemZ::SRA) { |
368 | // These shifts only use the low 6 bits of the shift count. |
369 | MachineOperand &ImmMO = MI.getOperand(i: 3); |
370 | ImmMO.setImm(ImmMO.getImm() & 0xfff); |
371 | } |
372 | Changed = true; |
373 | break; |
374 | } |
375 | } |
376 | |
377 | LiveRegs.stepBackward(MI); |
378 | } |
379 | |
380 | return Changed; |
381 | } |
382 | |
383 | bool SystemZShortenInst::runOnMachineFunction(MachineFunction &F) { |
384 | if (skipFunction(F: F.getFunction())) |
385 | return false; |
386 | |
387 | const SystemZSubtarget &ST = F.getSubtarget<SystemZSubtarget>(); |
388 | TII = ST.getInstrInfo(); |
389 | TRI = ST.getRegisterInfo(); |
390 | LiveRegs.init(TRI: *TRI); |
391 | |
392 | bool Changed = false; |
393 | for (auto &MBB : F) |
394 | Changed |= processBlock(MBB); |
395 | |
396 | return Changed; |
397 | } |
398 | |