1 | //===--------------- PPCVSXFMAMutate.cpp - VSX FMA Mutation ---------------===// |
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 mutates the form of VSX FMA instructions to avoid unnecessary |
10 | // copies. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "MCTargetDesc/PPCPredicates.h" |
15 | #include "PPC.h" |
16 | #include "PPCInstrBuilder.h" |
17 | #include "PPCInstrInfo.h" |
18 | #include "PPCMachineFunctionInfo.h" |
19 | #include "PPCTargetMachine.h" |
20 | #include "llvm/ADT/STLExtras.h" |
21 | #include "llvm/ADT/Statistic.h" |
22 | #include "llvm/CodeGen/LiveIntervals.h" |
23 | #include "llvm/CodeGen/MachineDominators.h" |
24 | #include "llvm/CodeGen/MachineFrameInfo.h" |
25 | #include "llvm/CodeGen/MachineFunctionPass.h" |
26 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
27 | #include "llvm/CodeGen/MachineMemOperand.h" |
28 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
29 | #include "llvm/CodeGen/PseudoSourceValue.h" |
30 | #include "llvm/CodeGen/ScheduleDAG.h" |
31 | #include "llvm/CodeGen/SlotIndexes.h" |
32 | #include "llvm/InitializePasses.h" |
33 | #include "llvm/MC/MCAsmInfo.h" |
34 | #include "llvm/MC/TargetRegistry.h" |
35 | #include "llvm/Support/CommandLine.h" |
36 | #include "llvm/Support/Debug.h" |
37 | #include "llvm/Support/ErrorHandling.h" |
38 | #include "llvm/Support/raw_ostream.h" |
39 | |
40 | using namespace llvm; |
41 | |
42 | // Temporarily disable FMA mutation by default, since it doesn't handle |
43 | // cross-basic-block intervals well. |
44 | // See: http://lists.llvm.org/pipermail/llvm-dev/2016-February/095669.html |
45 | // http://reviews.llvm.org/D17087 |
46 | static cl::opt<bool> DisableVSXFMAMutate( |
47 | "disable-ppc-vsx-fma-mutation" , |
48 | cl::desc("Disable VSX FMA instruction mutation" ), cl::init(Val: true), |
49 | cl::Hidden); |
50 | |
51 | #define DEBUG_TYPE "ppc-vsx-fma-mutate" |
52 | |
53 | namespace llvm { namespace PPC { |
54 | int getAltVSXFMAOpcode(uint16_t Opcode); |
55 | } } |
56 | |
57 | namespace { |
58 | // PPCVSXFMAMutate pass - For copies between VSX registers and non-VSX registers |
59 | // (Altivec and scalar floating-point registers), we need to transform the |
60 | // copies into subregister copies with other restrictions. |
61 | struct PPCVSXFMAMutate : public MachineFunctionPass { |
62 | static char ID; |
63 | PPCVSXFMAMutate() : MachineFunctionPass(ID) { |
64 | initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry()); |
65 | } |
66 | |
67 | LiveIntervals *LIS; |
68 | const PPCInstrInfo *TII; |
69 | |
70 | protected: |
71 | bool processBlock(MachineBasicBlock &MBB) { |
72 | bool Changed = false; |
73 | |
74 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
75 | const TargetRegisterInfo *TRI = &TII->getRegisterInfo(); |
76 | for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); |
77 | I != IE; ++I) { |
78 | MachineInstr &MI = *I; |
79 | |
80 | // The default (A-type) VSX FMA form kills the addend (it is taken from |
81 | // the target register, which is then updated to reflect the result of |
82 | // the FMA). If the instruction, however, kills one of the registers |
83 | // used for the product, then we can use the M-form instruction (which |
84 | // will take that value from the to-be-defined register). |
85 | |
86 | int AltOpc = PPC::getAltVSXFMAOpcode(Opcode: MI.getOpcode()); |
87 | if (AltOpc == -1) |
88 | continue; |
89 | |
90 | // This pass is run after register coalescing, and so we're looking for |
91 | // a situation like this: |
92 | // ... |
93 | // %5 = COPY %9; VSLRC:%5,%9 |
94 | // %5<def,tied1> = XSMADDADP %5<tied0>, %17, %16, |
95 | // implicit %rm; VSLRC:%5,%17,%16 |
96 | // ... |
97 | // %9<def,tied1> = XSMADDADP %9<tied0>, %17, %19, |
98 | // implicit %rm; VSLRC:%9,%17,%19 |
99 | // ... |
100 | // Where we can eliminate the copy by changing from the A-type to the |
101 | // M-type instruction. Specifically, for this example, this means: |
102 | // %5<def,tied1> = XSMADDADP %5<tied0>, %17, %16, |
103 | // implicit %rm; VSLRC:%5,%17,%16 |
104 | // is replaced by: |
105 | // %16<def,tied1> = XSMADDMDP %16<tied0>, %18, %9, |
106 | // implicit %rm; VSLRC:%16,%18,%9 |
107 | // and we remove: %5 = COPY %9; VSLRC:%5,%9 |
108 | |
109 | SlotIndex FMAIdx = LIS->getInstructionIndex(Instr: MI); |
110 | |
111 | VNInfo *AddendValNo = |
112 | LIS->getInterval(Reg: MI.getOperand(i: 1).getReg()).Query(Idx: FMAIdx).valueIn(); |
113 | |
114 | // This can be null if the register is undef. |
115 | if (!AddendValNo) |
116 | continue; |
117 | |
118 | MachineInstr *AddendMI = LIS->getInstructionFromIndex(index: AddendValNo->def); |
119 | |
120 | // The addend and this instruction must be in the same block. |
121 | |
122 | if (!AddendMI || AddendMI->getParent() != MI.getParent()) |
123 | continue; |
124 | |
125 | // The addend must be a full copy within the same register class. |
126 | |
127 | if (!AddendMI->isFullCopy()) |
128 | continue; |
129 | |
130 | Register AddendSrcReg = AddendMI->getOperand(i: 1).getReg(); |
131 | if (AddendSrcReg.isVirtual()) { |
132 | if (MRI.getRegClass(Reg: AddendMI->getOperand(i: 0).getReg()) != |
133 | MRI.getRegClass(Reg: AddendSrcReg)) |
134 | continue; |
135 | } else { |
136 | // If AddendSrcReg is a physical register, make sure the destination |
137 | // register class contains it. |
138 | if (!MRI.getRegClass(Reg: AddendMI->getOperand(i: 0).getReg()) |
139 | ->contains(Reg: AddendSrcReg)) |
140 | continue; |
141 | } |
142 | |
143 | // In theory, there could be other uses of the addend copy before this |
144 | // fma. We could deal with this, but that would require additional |
145 | // logic below and I suspect it will not occur in any relevant |
146 | // situations. Additionally, check whether the copy source is killed |
147 | // prior to the fma. In order to replace the addend here with the |
148 | // source of the copy, it must still be live here. We can't use |
149 | // interval testing for a physical register, so as long as we're |
150 | // walking the MIs we may as well test liveness here. |
151 | // |
152 | // FIXME: There is a case that occurs in practice, like this: |
153 | // %9 = COPY %f1; VSSRC:%9 |
154 | // ... |
155 | // %6 = COPY %9; VSSRC:%6,%9 |
156 | // %7 = COPY %9; VSSRC:%7,%9 |
157 | // %9<def,tied1> = XSMADDASP %9<tied0>, %1, %4; VSSRC: |
158 | // %6<def,tied1> = XSMADDASP %6<tied0>, %1, %2; VSSRC: |
159 | // %7<def,tied1> = XSMADDASP %7<tied0>, %1, %3; VSSRC: |
160 | // which prevents an otherwise-profitable transformation. |
161 | bool OtherUsers = false, KillsAddendSrc = false; |
162 | for (auto J = std::prev(x: I), JE = MachineBasicBlock::iterator(AddendMI); |
163 | J != JE; --J) { |
164 | if (J->readsVirtualRegister(Reg: AddendMI->getOperand(i: 0).getReg())) { |
165 | OtherUsers = true; |
166 | break; |
167 | } |
168 | if (J->modifiesRegister(Reg: AddendSrcReg, TRI) || |
169 | J->killsRegister(Reg: AddendSrcReg, TRI)) { |
170 | KillsAddendSrc = true; |
171 | break; |
172 | } |
173 | } |
174 | |
175 | if (OtherUsers || KillsAddendSrc) |
176 | continue; |
177 | |
178 | |
179 | // The transformation doesn't work well with things like: |
180 | // %5 = A-form-op %5, %11, %5; |
181 | // unless %11 is also a kill, so skip when it is not, |
182 | // and check operand 3 to see it is also a kill to handle the case: |
183 | // %5 = A-form-op %5, %5, %11; |
184 | // where %5 and %11 are both kills. This case would be skipped |
185 | // otherwise. |
186 | Register OldFMAReg = MI.getOperand(i: 0).getReg(); |
187 | |
188 | // Find one of the product operands that is killed by this instruction. |
189 | unsigned KilledProdOp = 0, OtherProdOp = 0; |
190 | Register Reg2 = MI.getOperand(i: 2).getReg(); |
191 | Register Reg3 = MI.getOperand(i: 3).getReg(); |
192 | if (LIS->getInterval(Reg: Reg2).Query(Idx: FMAIdx).isKill() |
193 | && Reg2 != OldFMAReg) { |
194 | KilledProdOp = 2; |
195 | OtherProdOp = 3; |
196 | } else if (LIS->getInterval(Reg: Reg3).Query(Idx: FMAIdx).isKill() |
197 | && Reg3 != OldFMAReg) { |
198 | KilledProdOp = 3; |
199 | OtherProdOp = 2; |
200 | } |
201 | |
202 | // If there are no usable killed product operands, then this |
203 | // transformation is likely not profitable. |
204 | if (!KilledProdOp) |
205 | continue; |
206 | |
207 | // If the addend copy is used only by this MI, then the addend source |
208 | // register is likely not live here. This could be fixed (based on the |
209 | // legality checks above, the live range for the addend source register |
210 | // could be extended), but it seems likely that such a trivial copy can |
211 | // be coalesced away later, and thus is not worth the effort. |
212 | if (AddendSrcReg.isVirtual() && |
213 | !LIS->getInterval(Reg: AddendSrcReg).liveAt(index: FMAIdx)) |
214 | continue; |
215 | |
216 | // Transform: (O2 * O3) + O1 -> (O2 * O1) + O3. |
217 | |
218 | Register KilledProdReg = MI.getOperand(i: KilledProdOp).getReg(); |
219 | Register OtherProdReg = MI.getOperand(i: OtherProdOp).getReg(); |
220 | |
221 | unsigned AddSubReg = AddendMI->getOperand(i: 1).getSubReg(); |
222 | unsigned KilledProdSubReg = MI.getOperand(i: KilledProdOp).getSubReg(); |
223 | unsigned OtherProdSubReg = MI.getOperand(i: OtherProdOp).getSubReg(); |
224 | |
225 | bool AddRegKill = AddendMI->getOperand(i: 1).isKill(); |
226 | bool KilledProdRegKill = MI.getOperand(i: KilledProdOp).isKill(); |
227 | bool OtherProdRegKill = MI.getOperand(i: OtherProdOp).isKill(); |
228 | |
229 | bool AddRegUndef = AddendMI->getOperand(i: 1).isUndef(); |
230 | bool KilledProdRegUndef = MI.getOperand(i: KilledProdOp).isUndef(); |
231 | bool OtherProdRegUndef = MI.getOperand(i: OtherProdOp).isUndef(); |
232 | |
233 | // If there isn't a class that fits, we can't perform the transform. |
234 | // This is needed for correctness with a mixture of VSX and Altivec |
235 | // instructions to make sure that a low VSX register is not assigned to |
236 | // the Altivec instruction. |
237 | if (!MRI.constrainRegClass(Reg: KilledProdReg, |
238 | RC: MRI.getRegClass(Reg: OldFMAReg))) |
239 | continue; |
240 | |
241 | assert(OldFMAReg == AddendMI->getOperand(0).getReg() && |
242 | "Addend copy not tied to old FMA output!" ); |
243 | |
244 | LLVM_DEBUG(dbgs() << "VSX FMA Mutation:\n " << MI); |
245 | |
246 | MI.getOperand(i: 0).setReg(KilledProdReg); |
247 | MI.getOperand(i: 1).setReg(KilledProdReg); |
248 | MI.getOperand(i: 3).setReg(AddendSrcReg); |
249 | |
250 | MI.getOperand(i: 0).setSubReg(KilledProdSubReg); |
251 | MI.getOperand(i: 1).setSubReg(KilledProdSubReg); |
252 | MI.getOperand(i: 3).setSubReg(AddSubReg); |
253 | |
254 | MI.getOperand(i: 1).setIsKill(KilledProdRegKill); |
255 | MI.getOperand(i: 3).setIsKill(AddRegKill); |
256 | |
257 | MI.getOperand(i: 1).setIsUndef(KilledProdRegUndef); |
258 | MI.getOperand(i: 3).setIsUndef(AddRegUndef); |
259 | |
260 | MI.setDesc(TII->get(Opcode: AltOpc)); |
261 | |
262 | // If the addend is also a multiplicand, replace it with the addend |
263 | // source in both places. |
264 | if (OtherProdReg == AddendMI->getOperand(i: 0).getReg()) { |
265 | MI.getOperand(i: 2).setReg(AddendSrcReg); |
266 | MI.getOperand(i: 2).setSubReg(AddSubReg); |
267 | MI.getOperand(i: 2).setIsKill(AddRegKill); |
268 | MI.getOperand(i: 2).setIsUndef(AddRegUndef); |
269 | } else { |
270 | MI.getOperand(i: 2).setReg(OtherProdReg); |
271 | MI.getOperand(i: 2).setSubReg(OtherProdSubReg); |
272 | MI.getOperand(i: 2).setIsKill(OtherProdRegKill); |
273 | MI.getOperand(i: 2).setIsUndef(OtherProdRegUndef); |
274 | } |
275 | |
276 | LLVM_DEBUG(dbgs() << " -> " << MI); |
277 | |
278 | // The killed product operand was killed here, so we can reuse it now |
279 | // for the result of the fma. |
280 | |
281 | LiveInterval &FMAInt = LIS->getInterval(Reg: OldFMAReg); |
282 | VNInfo *FMAValNo = FMAInt.getVNInfoAt(Idx: FMAIdx.getRegSlot()); |
283 | for (auto UI = MRI.reg_nodbg_begin(RegNo: OldFMAReg), UE = MRI.reg_nodbg_end(); |
284 | UI != UE;) { |
285 | MachineOperand &UseMO = *UI; |
286 | MachineInstr *UseMI = UseMO.getParent(); |
287 | ++UI; |
288 | |
289 | // Don't replace the result register of the copy we're about to erase. |
290 | if (UseMI == AddendMI) |
291 | continue; |
292 | |
293 | UseMO.substVirtReg(Reg: KilledProdReg, SubIdx: KilledProdSubReg, *TRI); |
294 | } |
295 | |
296 | // Extend the live intervals of the killed product operand to hold the |
297 | // fma result. |
298 | |
299 | LiveInterval &NewFMAInt = LIS->getInterval(Reg: KilledProdReg); |
300 | for (auto &AI : FMAInt) { |
301 | // Don't add the segment that corresponds to the original copy. |
302 | if (AI.valno == AddendValNo) |
303 | continue; |
304 | |
305 | VNInfo *NewFMAValNo = |
306 | NewFMAInt.getNextValue(Def: AI.start, VNInfoAllocator&: LIS->getVNInfoAllocator()); |
307 | |
308 | NewFMAInt.addSegment( |
309 | S: LiveInterval::Segment(AI.start, AI.end, NewFMAValNo)); |
310 | } |
311 | LLVM_DEBUG(dbgs() << " extended: " << NewFMAInt << '\n'); |
312 | |
313 | // Extend the live interval of the addend source (it might end at the |
314 | // copy to be removed, or somewhere in between there and here). This |
315 | // is necessary only if it is a physical register. |
316 | if (!AddendSrcReg.isVirtual()) |
317 | for (MCRegUnit Unit : TRI->regunits(Reg: AddendSrcReg.asMCReg())) { |
318 | LiveRange &AddendSrcRange = LIS->getRegUnit(Unit); |
319 | AddendSrcRange.extendInBlock(StartIdx: LIS->getMBBStartIdx(mbb: &MBB), |
320 | Kill: FMAIdx.getRegSlot()); |
321 | LLVM_DEBUG(dbgs() << " extended: " << AddendSrcRange << '\n'); |
322 | } |
323 | |
324 | FMAInt.removeValNo(ValNo: FMAValNo); |
325 | LLVM_DEBUG(dbgs() << " trimmed: " << FMAInt << '\n'); |
326 | |
327 | // Remove the (now unused) copy. |
328 | |
329 | LLVM_DEBUG(dbgs() << " removing: " << *AddendMI << '\n'); |
330 | LIS->RemoveMachineInstrFromMaps(MI&: *AddendMI); |
331 | AddendMI->eraseFromParent(); |
332 | |
333 | Changed = true; |
334 | } |
335 | |
336 | return Changed; |
337 | } |
338 | |
339 | public: |
340 | bool runOnMachineFunction(MachineFunction &MF) override { |
341 | if (skipFunction(F: MF.getFunction())) |
342 | return false; |
343 | |
344 | // If we don't have VSX then go ahead and return without doing |
345 | // anything. |
346 | const PPCSubtarget &STI = MF.getSubtarget<PPCSubtarget>(); |
347 | if (!STI.hasVSX()) |
348 | return false; |
349 | |
350 | LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
351 | |
352 | TII = STI.getInstrInfo(); |
353 | |
354 | bool Changed = false; |
355 | |
356 | if (DisableVSXFMAMutate) |
357 | return Changed; |
358 | |
359 | for (MachineBasicBlock &B : llvm::make_early_inc_range(Range&: MF)) |
360 | if (processBlock(MBB&: B)) |
361 | Changed = true; |
362 | |
363 | return Changed; |
364 | } |
365 | |
366 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
367 | AU.addRequired<LiveIntervalsWrapperPass>(); |
368 | AU.addPreserved<LiveIntervalsWrapperPass>(); |
369 | AU.addRequired<SlotIndexesWrapperPass>(); |
370 | AU.addPreserved<SlotIndexesWrapperPass>(); |
371 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
372 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
373 | MachineFunctionPass::getAnalysisUsage(AU); |
374 | } |
375 | }; |
376 | } |
377 | |
378 | INITIALIZE_PASS_BEGIN(PPCVSXFMAMutate, DEBUG_TYPE, |
379 | "PowerPC VSX FMA Mutation" , false, false) |
380 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
381 | INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass) |
382 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
383 | INITIALIZE_PASS_END(PPCVSXFMAMutate, DEBUG_TYPE, |
384 | "PowerPC VSX FMA Mutation" , false, false) |
385 | |
386 | char &llvm::PPCVSXFMAMutateID = PPCVSXFMAMutate::ID; |
387 | |
388 | char PPCVSXFMAMutate::ID = 0; |
389 | FunctionPass *llvm::createPPCVSXFMAMutatePass() { |
390 | return new PPCVSXFMAMutate(); |
391 | } |
392 | |