xref: /llvm-project/llvm/lib/Target/ARM/ARMLegalizerInfo.cpp (revision b3bde2ea50decc455f3b2fff02e49351e4209d92)
1 //===- ARMLegalizerInfo.cpp --------------------------------------*- C++ -*-==//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
10 /// This file implements the targeting of the Machinelegalizer class for ARM.
11 /// \todo This should be generated by TableGen.
12 //===----------------------------------------------------------------------===//
13 
14 #include "ARMLegalizerInfo.h"
15 #include "ARMCallLowering.h"
16 #include "ARMSubtarget.h"
17 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
18 #include "llvm/CodeGen/LowLevelType.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/CodeGen/TargetOpcodes.h"
21 #include "llvm/CodeGen/ValueTypes.h"
22 #include "llvm/IR/DerivedTypes.h"
23 #include "llvm/IR/Type.h"
24 
25 using namespace llvm;
26 
27 /// FIXME: The following static functions are SizeChangeStrategy functions
28 /// that are meant to temporarily mimic the behaviour of the old legalization
29 /// based on doubling/halving non-legal types as closely as possible. This is
30 /// not entirly possible as only legalizing the types that are exactly a power
31 /// of 2 times the size of the legal types would require specifying all those
32 /// sizes explicitly.
33 /// In practice, not specifying those isn't a problem, and the below functions
34 /// should disappear quickly as we add support for legalizing non-power-of-2
35 /// sized types further.
36 static void
37 addAndInterleaveWithUnsupported(LegalizerInfo::SizeAndActionsVec &result,
38                                 const LegalizerInfo::SizeAndActionsVec &v) {
39   for (unsigned i = 0; i < v.size(); ++i) {
40     result.push_back(v[i]);
41     if (i + 1 < v[i].first && i + 1 < v.size() &&
42         v[i + 1].first != v[i].first + 1)
43       result.push_back({v[i].first + 1, LegalizerInfo::Unsupported});
44   }
45 }
46 
47 static LegalizerInfo::SizeAndActionsVec
48 widen_8_16(const LegalizerInfo::SizeAndActionsVec &v) {
49   assert(v.size() >= 1);
50   assert(v[0].first > 17);
51   LegalizerInfo::SizeAndActionsVec result = {
52       {1, LegalizerInfo::Unsupported},
53       {8, LegalizerInfo::WidenScalar},  {9, LegalizerInfo::Unsupported},
54       {16, LegalizerInfo::WidenScalar}, {17, LegalizerInfo::Unsupported}};
55   addAndInterleaveWithUnsupported(result, v);
56   auto Largest = result.back().first;
57   result.push_back({Largest + 1, LegalizerInfo::Unsupported});
58   return result;
59 }
60 
61 static LegalizerInfo::SizeAndActionsVec
62 widen_1_8_16(const LegalizerInfo::SizeAndActionsVec &v) {
63   assert(v.size() >= 1);
64   assert(v[0].first > 17);
65   LegalizerInfo::SizeAndActionsVec result = {
66       {1, LegalizerInfo::WidenScalar},  {2, LegalizerInfo::Unsupported},
67       {8, LegalizerInfo::WidenScalar},  {9, LegalizerInfo::Unsupported},
68       {16, LegalizerInfo::WidenScalar}, {17, LegalizerInfo::Unsupported}};
69   addAndInterleaveWithUnsupported(result, v);
70   auto Largest = result.back().first;
71   result.push_back({Largest + 1, LegalizerInfo::Unsupported});
72   return result;
73 }
74 
75 static bool AEABI(const ARMSubtarget &ST) {
76   return ST.isTargetAEABI() || ST.isTargetGNUAEABI() || ST.isTargetMuslAEABI();
77 }
78 
79 ARMLegalizerInfo::ARMLegalizerInfo(const ARMSubtarget &ST) {
80   using namespace TargetOpcode;
81 
82   const LLT p0 = LLT::pointer(0, 32);
83 
84   const LLT s1 = LLT::scalar(1);
85   const LLT s8 = LLT::scalar(8);
86   const LLT s16 = LLT::scalar(16);
87   const LLT s32 = LLT::scalar(32);
88   const LLT s64 = LLT::scalar(64);
89 
90   setAction({G_GLOBAL_VALUE, p0}, Legal);
91   setAction({G_FRAME_INDEX, p0}, Legal);
92 
93   for (unsigned Op : {G_LOAD, G_STORE}) {
94     for (auto Ty : {s1, s8, s16, s32, p0})
95       setAction({Op, Ty}, Legal);
96     setAction({Op, 1, p0}, Legal);
97   }
98 
99   for (unsigned Op : {G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR}) {
100     if (Op != G_ADD)
101       setLegalizeScalarToDifferentSizeStrategy(
102           Op, 0, widenToLargerTypesUnsupportedOtherwise);
103     setAction({Op, s32}, Legal);
104   }
105 
106   for (unsigned Op : {G_SDIV, G_UDIV}) {
107     setLegalizeScalarToDifferentSizeStrategy(Op, 0,
108         widenToLargerTypesUnsupportedOtherwise);
109     if (ST.hasDivideInARMMode())
110       setAction({Op, s32}, Legal);
111     else
112       setAction({Op, s32}, Libcall);
113   }
114 
115   for (unsigned Op : {G_SREM, G_UREM}) {
116     setLegalizeScalarToDifferentSizeStrategy(Op, 0, widen_8_16);
117     if (ST.hasDivideInARMMode())
118       setAction({Op, s32}, Lower);
119     else if (AEABI(ST))
120       setAction({Op, s32}, Custom);
121     else
122       setAction({Op, s32}, Libcall);
123   }
124 
125   for (unsigned Op : {G_SEXT, G_ZEXT, G_ANYEXT}) {
126     setAction({Op, s32}, Legal);
127   }
128 
129   for (unsigned Op : {G_ASHR, G_LSHR, G_SHL})
130     setAction({Op, s32}, Legal);
131 
132   setAction({G_GEP, p0}, Legal);
133   setAction({G_GEP, 1, s32}, Legal);
134 
135   setAction({G_SELECT, s32}, Legal);
136   setAction({G_SELECT, p0}, Legal);
137   setAction({G_SELECT, 1, s1}, Legal);
138 
139   setAction({G_BRCOND, s1}, Legal);
140 
141   setAction({G_CONSTANT, s32}, Legal);
142   setLegalizeScalarToDifferentSizeStrategy(G_CONSTANT, 0, widen_1_8_16);
143 
144   setAction({G_ICMP, s1}, Legal);
145   setLegalizeScalarToDifferentSizeStrategy(G_ICMP, 1,
146       widenToLargerTypesUnsupportedOtherwise);
147   for (auto Ty : {s32, p0})
148     setAction({G_ICMP, 1, Ty}, Legal);
149 
150   if (!ST.useSoftFloat() && ST.hasVFP2()) {
151     for (unsigned BinOp : {G_FADD, G_FSUB})
152       for (auto Ty : {s32, s64})
153         setAction({BinOp, Ty}, Legal);
154 
155     setAction({G_LOAD, s64}, Legal);
156     setAction({G_STORE, s64}, Legal);
157 
158     setAction({G_FCMP, s1}, Legal);
159     setAction({G_FCMP, 1, s32}, Legal);
160     setAction({G_FCMP, 1, s64}, Legal);
161   } else {
162     for (unsigned BinOp : {G_FADD, G_FSUB})
163       for (auto Ty : {s32, s64})
164         setAction({BinOp, Ty}, Libcall);
165 
166     setAction({G_FCMP, s1}, Legal);
167     setAction({G_FCMP, 1, s32}, Custom);
168     setAction({G_FCMP, 1, s64}, Custom);
169 
170     if (AEABI(ST))
171       setFCmpLibcallsAEABI();
172     else
173       setFCmpLibcallsGNU();
174   }
175 
176   for (unsigned Op : {G_FREM, G_FPOW})
177     for (auto Ty : {s32, s64})
178       setAction({Op, Ty}, Libcall);
179 
180   computeTables();
181 }
182 
183 void ARMLegalizerInfo::setFCmpLibcallsAEABI() {
184   // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be
185   // default-initialized.
186   FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
187   FCmp32Libcalls[CmpInst::FCMP_OEQ] = {
188       {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE}};
189   FCmp32Libcalls[CmpInst::FCMP_OGE] = {
190       {RTLIB::OGE_F32, CmpInst::BAD_ICMP_PREDICATE}};
191   FCmp32Libcalls[CmpInst::FCMP_OGT] = {
192       {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE}};
193   FCmp32Libcalls[CmpInst::FCMP_OLE] = {
194       {RTLIB::OLE_F32, CmpInst::BAD_ICMP_PREDICATE}};
195   FCmp32Libcalls[CmpInst::FCMP_OLT] = {
196       {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}};
197   FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}};
198   FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_EQ}};
199   FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_EQ}};
200   FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_EQ}};
201   FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_EQ}};
202   FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_EQ}};
203   FCmp32Libcalls[CmpInst::FCMP_UNO] = {
204       {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}};
205   FCmp32Libcalls[CmpInst::FCMP_ONE] = {
206       {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE},
207       {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}};
208   FCmp32Libcalls[CmpInst::FCMP_UEQ] = {
209       {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE},
210       {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}};
211 
212   FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
213   FCmp64Libcalls[CmpInst::FCMP_OEQ] = {
214       {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE}};
215   FCmp64Libcalls[CmpInst::FCMP_OGE] = {
216       {RTLIB::OGE_F64, CmpInst::BAD_ICMP_PREDICATE}};
217   FCmp64Libcalls[CmpInst::FCMP_OGT] = {
218       {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE}};
219   FCmp64Libcalls[CmpInst::FCMP_OLE] = {
220       {RTLIB::OLE_F64, CmpInst::BAD_ICMP_PREDICATE}};
221   FCmp64Libcalls[CmpInst::FCMP_OLT] = {
222       {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}};
223   FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}};
224   FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_EQ}};
225   FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_EQ}};
226   FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_EQ}};
227   FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_EQ}};
228   FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_EQ}};
229   FCmp64Libcalls[CmpInst::FCMP_UNO] = {
230       {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}};
231   FCmp64Libcalls[CmpInst::FCMP_ONE] = {
232       {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE},
233       {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}};
234   FCmp64Libcalls[CmpInst::FCMP_UEQ] = {
235       {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE},
236       {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}};
237 }
238 
239 void ARMLegalizerInfo::setFCmpLibcallsGNU() {
240   // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be
241   // default-initialized.
242   FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
243   FCmp32Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ}};
244   FCmp32Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F32, CmpInst::ICMP_SGE}};
245   FCmp32Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT}};
246   FCmp32Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F32, CmpInst::ICMP_SLE}};
247   FCmp32Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F32, CmpInst::ICMP_SLT}};
248   FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}};
249   FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_SGE}};
250   FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_SGT}};
251   FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SLE}};
252   FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_SLT}};
253   FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_NE}};
254   FCmp32Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F32, CmpInst::ICMP_NE}};
255   FCmp32Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT},
256                                        {RTLIB::OLT_F32, CmpInst::ICMP_SLT}};
257   FCmp32Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ},
258                                        {RTLIB::UO_F32, CmpInst::ICMP_NE}};
259 
260   FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
261   FCmp64Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ}};
262   FCmp64Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F64, CmpInst::ICMP_SGE}};
263   FCmp64Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT}};
264   FCmp64Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F64, CmpInst::ICMP_SLE}};
265   FCmp64Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F64, CmpInst::ICMP_SLT}};
266   FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}};
267   FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_SGE}};
268   FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_SGT}};
269   FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SLE}};
270   FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_SLT}};
271   FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_NE}};
272   FCmp64Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F64, CmpInst::ICMP_NE}};
273   FCmp64Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT},
274                                        {RTLIB::OLT_F64, CmpInst::ICMP_SLT}};
275   FCmp64Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ},
276                                        {RTLIB::UO_F64, CmpInst::ICMP_NE}};
277 }
278 
279 ARMLegalizerInfo::FCmpLibcallsList
280 ARMLegalizerInfo::getFCmpLibcalls(CmpInst::Predicate Predicate,
281                                   unsigned Size) const {
282   assert(CmpInst::isFPPredicate(Predicate) && "Unsupported FCmp predicate");
283   if (Size == 32)
284     return FCmp32Libcalls[Predicate];
285   if (Size == 64)
286     return FCmp64Libcalls[Predicate];
287   llvm_unreachable("Unsupported size for FCmp predicate");
288 }
289 
290 bool ARMLegalizerInfo::legalizeCustom(MachineInstr &MI,
291                                       MachineRegisterInfo &MRI,
292                                       MachineIRBuilder &MIRBuilder) const {
293   using namespace TargetOpcode;
294 
295   MIRBuilder.setInstr(MI);
296 
297   switch (MI.getOpcode()) {
298   default:
299     return false;
300   case G_SREM:
301   case G_UREM: {
302     unsigned OriginalResult = MI.getOperand(0).getReg();
303     auto Size = MRI.getType(OriginalResult).getSizeInBits();
304     if (Size != 32)
305       return false;
306 
307     auto Libcall =
308         MI.getOpcode() == G_SREM ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32;
309 
310     // Our divmod libcalls return a struct containing the quotient and the
311     // remainder. We need to create a virtual register for it.
312     auto &Ctx = MIRBuilder.getMF().getFunction()->getContext();
313     Type *ArgTy = Type::getInt32Ty(Ctx);
314     StructType *RetTy = StructType::get(Ctx, {ArgTy, ArgTy}, /* Packed */ true);
315     auto RetVal = MRI.createGenericVirtualRegister(
316         getLLTForType(*RetTy, MIRBuilder.getMF().getDataLayout()));
317 
318     auto Status = createLibcall(MIRBuilder, Libcall, {RetVal, RetTy},
319                                 {{MI.getOperand(1).getReg(), ArgTy},
320                                  {MI.getOperand(2).getReg(), ArgTy}});
321     if (Status != LegalizerHelper::Legalized)
322       return false;
323 
324     // The remainder is the second result of divmod. Split the return value into
325     // a new, unused register for the quotient and the destination of the
326     // original instruction for the remainder.
327     MIRBuilder.buildUnmerge(
328         {MRI.createGenericVirtualRegister(LLT::scalar(32)), OriginalResult},
329         RetVal);
330     break;
331   }
332   case G_FCMP: {
333     assert(MRI.getType(MI.getOperand(2).getReg()) ==
334                MRI.getType(MI.getOperand(3).getReg()) &&
335            "Mismatched operands for G_FCMP");
336     auto OpSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits();
337 
338     auto OriginalResult = MI.getOperand(0).getReg();
339     auto Predicate =
340         static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
341     auto Libcalls = getFCmpLibcalls(Predicate, OpSize);
342 
343     if (Libcalls.empty()) {
344       assert((Predicate == CmpInst::FCMP_TRUE ||
345               Predicate == CmpInst::FCMP_FALSE) &&
346              "Predicate needs libcalls, but none specified");
347       MIRBuilder.buildConstant(OriginalResult,
348                                Predicate == CmpInst::FCMP_TRUE ? 1 : 0);
349       MI.eraseFromParent();
350       return true;
351     }
352 
353     auto &Ctx = MIRBuilder.getMF().getFunction()->getContext();
354     assert((OpSize == 32 || OpSize == 64) && "Unsupported operand size");
355     auto *ArgTy = OpSize == 32 ? Type::getFloatTy(Ctx) : Type::getDoubleTy(Ctx);
356     auto *RetTy = Type::getInt32Ty(Ctx);
357 
358     SmallVector<unsigned, 2> Results;
359     for (auto Libcall : Libcalls) {
360       auto LibcallResult = MRI.createGenericVirtualRegister(LLT::scalar(32));
361       auto Status =
362           createLibcall(MIRBuilder, Libcall.LibcallID, {LibcallResult, RetTy},
363                         {{MI.getOperand(2).getReg(), ArgTy},
364                          {MI.getOperand(3).getReg(), ArgTy}});
365 
366       if (Status != LegalizerHelper::Legalized)
367         return false;
368 
369       auto ProcessedResult =
370           Libcalls.size() == 1
371               ? OriginalResult
372               : MRI.createGenericVirtualRegister(MRI.getType(OriginalResult));
373 
374       // We have a result, but we need to transform it into a proper 1-bit 0 or
375       // 1, taking into account the different peculiarities of the values
376       // returned by the comparison functions.
377       CmpInst::Predicate ResultPred = Libcall.Predicate;
378       if (ResultPred == CmpInst::BAD_ICMP_PREDICATE) {
379         // We have a nice 0 or 1, and we just need to truncate it back to 1 bit
380         // to keep the types consistent.
381         MIRBuilder.buildTrunc(ProcessedResult, LibcallResult);
382       } else {
383         // We need to compare against 0.
384         assert(CmpInst::isIntPredicate(ResultPred) && "Unsupported predicate");
385         auto Zero = MRI.createGenericVirtualRegister(LLT::scalar(32));
386         MIRBuilder.buildConstant(Zero, 0);
387         MIRBuilder.buildICmp(ResultPred, ProcessedResult, LibcallResult, Zero);
388       }
389       Results.push_back(ProcessedResult);
390     }
391 
392     if (Results.size() != 1) {
393       assert(Results.size() == 2 && "Unexpected number of results");
394       MIRBuilder.buildOr(OriginalResult, Results[0], Results[1]);
395     }
396     break;
397   }
398   }
399 
400   MI.eraseFromParent();
401   return true;
402 }
403