xref: /llvm-project/llvm/lib/Target/ARM/ARMLegalizerInfo.cpp (revision b1fd7849361f16b9059ddd6956c0f3f2eca784b3)
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/ValueTypes.h"
21 #include "llvm/IR/DerivedTypes.h"
22 #include "llvm/IR/Type.h"
23 #include "llvm/Target/TargetOpcodes.h"
24 
25 using namespace llvm;
26 
27 #ifndef LLVM_BUILD_GLOBAL_ISEL
28 #error "You shouldn't build this"
29 #endif
30 
31 static bool AEABI(const ARMSubtarget &ST) {
32   return ST.isTargetAEABI() || ST.isTargetGNUAEABI() || ST.isTargetMuslAEABI();
33 }
34 
35 ARMLegalizerInfo::ARMLegalizerInfo(const ARMSubtarget &ST) {
36   using namespace TargetOpcode;
37 
38   const LLT p0 = LLT::pointer(0, 32);
39 
40   const LLT s1 = LLT::scalar(1);
41   const LLT s8 = LLT::scalar(8);
42   const LLT s16 = LLT::scalar(16);
43   const LLT s32 = LLT::scalar(32);
44   const LLT s64 = LLT::scalar(64);
45 
46   setAction({G_GLOBAL_VALUE, p0}, Legal);
47   setAction({G_FRAME_INDEX, p0}, Legal);
48 
49   for (unsigned Op : {G_LOAD, G_STORE}) {
50     for (auto Ty : {s1, s8, s16, s32, p0})
51       setAction({Op, Ty}, Legal);
52     setAction({Op, 1, p0}, Legal);
53   }
54 
55   for (unsigned Op : {G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR}) {
56     for (auto Ty : {s1, s8, s16})
57       setAction({Op, Ty}, WidenScalar);
58     setAction({Op, s32}, Legal);
59   }
60 
61   for (unsigned Op : {G_SDIV, G_UDIV}) {
62     for (auto Ty : {s8, s16})
63       setAction({Op, Ty}, WidenScalar);
64     if (ST.hasDivideInARMMode())
65       setAction({Op, s32}, Legal);
66     else
67       setAction({Op, s32}, Libcall);
68   }
69 
70   for (unsigned Op : {G_SREM, G_UREM}) {
71     for (auto Ty : {s8, s16})
72       setAction({Op, Ty}, WidenScalar);
73     if (ST.hasDivideInARMMode())
74       setAction({Op, s32}, Lower);
75     else if (AEABI(ST))
76       setAction({Op, s32}, Custom);
77     else
78       setAction({Op, s32}, Libcall);
79   }
80 
81   for (unsigned Op : {G_SEXT, G_ZEXT}) {
82     setAction({Op, s32}, Legal);
83     for (auto Ty : {s1, s8, s16})
84       setAction({Op, 1, Ty}, Legal);
85   }
86 
87   setAction({G_GEP, p0}, Legal);
88   setAction({G_GEP, 1, s32}, Legal);
89 
90   setAction({G_SELECT, s32}, Legal);
91   setAction({G_SELECT, p0}, Legal);
92   setAction({G_SELECT, 1, s1}, Legal);
93 
94   setAction({G_BRCOND, s1}, Legal);
95 
96   setAction({G_CONSTANT, s32}, Legal);
97   for (auto Ty : {s1, s8, s16})
98     setAction({G_CONSTANT, Ty}, WidenScalar);
99 
100   setAction({G_ICMP, s1}, Legal);
101   for (auto Ty : {s8, s16})
102     setAction({G_ICMP, 1, Ty}, WidenScalar);
103   for (auto Ty : {s32, p0})
104     setAction({G_ICMP, 1, Ty}, Legal);
105 
106   if (!ST.useSoftFloat() && ST.hasVFP2()) {
107     setAction({G_FADD, s32}, Legal);
108     setAction({G_FADD, s64}, Legal);
109 
110     setAction({G_LOAD, s64}, Legal);
111     setAction({G_STORE, s64}, Legal);
112 
113     setAction({G_FCMP, s1}, Legal);
114     setAction({G_FCMP, 1, s32}, Legal);
115     setAction({G_FCMP, 1, s64}, Legal);
116   } else {
117     for (auto Ty : {s32, s64})
118       setAction({G_FADD, Ty}, Libcall);
119 
120     setAction({G_FCMP, s1}, Legal);
121     setAction({G_FCMP, 1, s32}, Custom);
122     setAction({G_FCMP, 1, s64}, Custom);
123 
124     if (AEABI(ST))
125       setFCmpLibcallsAEABI();
126     else
127       setFCmpLibcallsGNU();
128   }
129 
130   for (unsigned Op : {G_FREM, G_FPOW})
131     for (auto Ty : {s32, s64})
132       setAction({Op, Ty}, Libcall);
133 
134   computeTables();
135 }
136 
137 void ARMLegalizerInfo::setFCmpLibcallsAEABI() {
138   // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be
139   // default-initialized.
140   FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
141   FCmp32Libcalls[CmpInst::FCMP_OEQ] = {
142       {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE}};
143   FCmp32Libcalls[CmpInst::FCMP_OGE] = {
144       {RTLIB::OGE_F32, CmpInst::BAD_ICMP_PREDICATE}};
145   FCmp32Libcalls[CmpInst::FCMP_OGT] = {
146       {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE}};
147   FCmp32Libcalls[CmpInst::FCMP_OLE] = {
148       {RTLIB::OLE_F32, CmpInst::BAD_ICMP_PREDICATE}};
149   FCmp32Libcalls[CmpInst::FCMP_OLT] = {
150       {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}};
151   FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}};
152   FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_EQ}};
153   FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_EQ}};
154   FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_EQ}};
155   FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_EQ}};
156   FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_EQ}};
157   FCmp32Libcalls[CmpInst::FCMP_UNO] = {
158       {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}};
159   FCmp32Libcalls[CmpInst::FCMP_ONE] = {
160       {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE},
161       {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}};
162   FCmp32Libcalls[CmpInst::FCMP_UEQ] = {
163       {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE},
164       {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}};
165 
166   FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
167   FCmp64Libcalls[CmpInst::FCMP_OEQ] = {
168       {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE}};
169   FCmp64Libcalls[CmpInst::FCMP_OGE] = {
170       {RTLIB::OGE_F64, CmpInst::BAD_ICMP_PREDICATE}};
171   FCmp64Libcalls[CmpInst::FCMP_OGT] = {
172       {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE}};
173   FCmp64Libcalls[CmpInst::FCMP_OLE] = {
174       {RTLIB::OLE_F64, CmpInst::BAD_ICMP_PREDICATE}};
175   FCmp64Libcalls[CmpInst::FCMP_OLT] = {
176       {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}};
177   FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}};
178   FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_EQ}};
179   FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_EQ}};
180   FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_EQ}};
181   FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_EQ}};
182   FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_EQ}};
183   FCmp64Libcalls[CmpInst::FCMP_UNO] = {
184       {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}};
185   FCmp64Libcalls[CmpInst::FCMP_ONE] = {
186       {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE},
187       {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}};
188   FCmp64Libcalls[CmpInst::FCMP_UEQ] = {
189       {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE},
190       {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}};
191 }
192 
193 void ARMLegalizerInfo::setFCmpLibcallsGNU() {
194   // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be
195   // default-initialized.
196   FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
197   FCmp32Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ}};
198   FCmp32Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F32, CmpInst::ICMP_SGE}};
199   FCmp32Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT}};
200   FCmp32Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F32, CmpInst::ICMP_SLE}};
201   FCmp32Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F32, CmpInst::ICMP_SLT}};
202   FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}};
203   FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_SGE}};
204   FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_SGT}};
205   FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SLE}};
206   FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_SLT}};
207   FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_NE}};
208   FCmp32Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F32, CmpInst::ICMP_NE}};
209   FCmp32Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT},
210                                        {RTLIB::OLT_F32, CmpInst::ICMP_SLT}};
211   FCmp32Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ},
212                                        {RTLIB::UO_F32, CmpInst::ICMP_NE}};
213 
214   FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1);
215   FCmp64Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ}};
216   FCmp64Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F64, CmpInst::ICMP_SGE}};
217   FCmp64Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT}};
218   FCmp64Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F64, CmpInst::ICMP_SLE}};
219   FCmp64Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F64, CmpInst::ICMP_SLT}};
220   FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}};
221   FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_SGE}};
222   FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_SGT}};
223   FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SLE}};
224   FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_SLT}};
225   FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_NE}};
226   FCmp64Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F64, CmpInst::ICMP_NE}};
227   FCmp64Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT},
228                                        {RTLIB::OLT_F64, CmpInst::ICMP_SLT}};
229   FCmp64Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ},
230                                        {RTLIB::UO_F64, CmpInst::ICMP_NE}};
231 }
232 
233 ARMLegalizerInfo::FCmpLibcallsList
234 ARMLegalizerInfo::getFCmpLibcalls(CmpInst::Predicate Predicate,
235                                   unsigned Size) const {
236   assert(CmpInst::isFPPredicate(Predicate) && "Unsupported FCmp predicate");
237   if (Size == 32)
238     return FCmp32Libcalls[Predicate];
239   if (Size == 64)
240     return FCmp64Libcalls[Predicate];
241   llvm_unreachable("Unsupported size for FCmp predicate");
242 }
243 
244 bool ARMLegalizerInfo::legalizeCustom(MachineInstr &MI,
245                                       MachineRegisterInfo &MRI,
246                                       MachineIRBuilder &MIRBuilder) const {
247   using namespace TargetOpcode;
248 
249   MIRBuilder.setInstr(MI);
250 
251   switch (MI.getOpcode()) {
252   default:
253     return false;
254   case G_SREM:
255   case G_UREM: {
256     unsigned OriginalResult = MI.getOperand(0).getReg();
257     auto Size = MRI.getType(OriginalResult).getSizeInBits();
258     if (Size != 32)
259       return false;
260 
261     auto Libcall =
262         MI.getOpcode() == G_SREM ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32;
263 
264     // Our divmod libcalls return a struct containing the quotient and the
265     // remainder. We need to create a virtual register for it.
266     auto &Ctx = MIRBuilder.getMF().getFunction()->getContext();
267     Type *ArgTy = Type::getInt32Ty(Ctx);
268     StructType *RetTy = StructType::get(Ctx, {ArgTy, ArgTy}, /* Packed */ true);
269     auto RetVal = MRI.createGenericVirtualRegister(
270         getLLTForType(*RetTy, MIRBuilder.getMF().getDataLayout()));
271 
272     auto Status = createLibcall(MIRBuilder, Libcall, {RetVal, RetTy},
273                                 {{MI.getOperand(1).getReg(), ArgTy},
274                                  {MI.getOperand(2).getReg(), ArgTy}});
275     if (Status != LegalizerHelper::Legalized)
276       return false;
277 
278     // The remainder is the second result of divmod. Split the return value into
279     // a new, unused register for the quotient and the destination of the
280     // original instruction for the remainder.
281     MIRBuilder.buildUnmerge(
282         {MRI.createGenericVirtualRegister(LLT::scalar(32)), OriginalResult},
283         RetVal);
284     break;
285   }
286   case G_FCMP: {
287     assert(MRI.getType(MI.getOperand(2).getReg()) ==
288                MRI.getType(MI.getOperand(3).getReg()) &&
289            "Mismatched operands for G_FCMP");
290     auto OpSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits();
291 
292     auto OriginalResult = MI.getOperand(0).getReg();
293     auto Predicate =
294         static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
295     auto Libcalls = getFCmpLibcalls(Predicate, OpSize);
296 
297     if (Libcalls.empty()) {
298       assert((Predicate == CmpInst::FCMP_TRUE ||
299               Predicate == CmpInst::FCMP_FALSE) &&
300              "Predicate needs libcalls, but none specified");
301       MIRBuilder.buildConstant(OriginalResult,
302                                Predicate == CmpInst::FCMP_TRUE ? 1 : 0);
303       MI.eraseFromParent();
304       return true;
305     }
306 
307     auto &Ctx = MIRBuilder.getMF().getFunction()->getContext();
308     assert((OpSize == 32 || OpSize == 64) && "Unsupported operand size");
309     auto *ArgTy = OpSize == 32 ? Type::getFloatTy(Ctx) : Type::getDoubleTy(Ctx);
310     auto *RetTy = Type::getInt32Ty(Ctx);
311 
312     SmallVector<unsigned, 2> Results;
313     for (auto Libcall : Libcalls) {
314       auto LibcallResult = MRI.createGenericVirtualRegister(LLT::scalar(32));
315       auto Status =
316           createLibcall(MIRBuilder, Libcall.LibcallID, {LibcallResult, RetTy},
317                         {{MI.getOperand(2).getReg(), ArgTy},
318                          {MI.getOperand(3).getReg(), ArgTy}});
319 
320       if (Status != LegalizerHelper::Legalized)
321         return false;
322 
323       auto ProcessedResult =
324           Libcalls.size() == 1
325               ? OriginalResult
326               : MRI.createGenericVirtualRegister(MRI.getType(OriginalResult));
327 
328       // We have a result, but we need to transform it into a proper 1-bit 0 or
329       // 1, taking into account the different peculiarities of the values
330       // returned by the comparison functions.
331       CmpInst::Predicate ResultPred = Libcall.Predicate;
332       if (ResultPred == CmpInst::BAD_ICMP_PREDICATE) {
333         // We have a nice 0 or 1, and we just need to truncate it back to 1 bit
334         // to keep the types consistent.
335         MIRBuilder.buildTrunc(ProcessedResult, LibcallResult);
336       } else {
337         // We need to compare against 0.
338         assert(CmpInst::isIntPredicate(ResultPred) && "Unsupported predicate");
339         auto Zero = MRI.createGenericVirtualRegister(LLT::scalar(32));
340         MIRBuilder.buildConstant(Zero, 0);
341         MIRBuilder.buildICmp(ResultPred, ProcessedResult, LibcallResult, Zero);
342       }
343       Results.push_back(ProcessedResult);
344     }
345 
346     if (Results.size() != 1) {
347       assert(Results.size() == 2 && "Unexpected number of results");
348       MIRBuilder.buildOr(OriginalResult, Results[0], Results[1]);
349     }
350     break;
351   }
352   }
353 
354   MI.eraseFromParent();
355   return true;
356 }
357