1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
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 implements the TargetLoweringBase class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/BitVector.h"
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/Analysis/Loads.h"
21 #include "llvm/CodeGen/Analysis.h"
22 #include "llvm/CodeGen/ISDOpcodes.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstr.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineOperand.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/RuntimeLibcalls.h"
32 #include "llvm/CodeGen/StackMaps.h"
33 #include "llvm/CodeGen/TargetLowering.h"
34 #include "llvm/CodeGen/TargetOpcodes.h"
35 #include "llvm/CodeGen/TargetRegisterInfo.h"
36 #include "llvm/CodeGen/ValueTypes.h"
37 #include "llvm/IR/Attributes.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/Support/BranchProbability.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MachineValueType.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetMachine.h"
55 #include <algorithm>
56 #include <cassert>
57 #include <cstddef>
58 #include <cstdint>
59 #include <cstring>
60 #include <iterator>
61 #include <string>
62 #include <tuple>
63 #include <utility>
64 
65 using namespace llvm;
66 
67 static cl::opt<bool> JumpIsExpensiveOverride(
68     "jump-is-expensive", cl::init(false),
69     cl::desc("Do not create extra branches to split comparison logic."),
70     cl::Hidden);
71 
72 static cl::opt<unsigned> MinimumJumpTableEntries
73   ("min-jump-table-entries", cl::init(4), cl::Hidden,
74    cl::desc("Set minimum number of entries to use a jump table."));
75 
76 static cl::opt<unsigned> MaximumJumpTableSize
77   ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden,
78    cl::desc("Set maximum size of jump tables."));
79 
80 /// Minimum jump table density for normal functions.
81 static cl::opt<unsigned>
82     JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden,
83                      cl::desc("Minimum density for building a jump table in "
84                               "a normal function"));
85 
86 /// Minimum jump table density for -Os or -Oz functions.
87 static cl::opt<unsigned> OptsizeJumpTableDensity(
88     "optsize-jump-table-density", cl::init(40), cl::Hidden,
89     cl::desc("Minimum density for building a jump table in "
90              "an optsize function"));
91 
92 // FIXME: This option is only to test if the strict fp operation processed
93 // correctly by preventing mutating strict fp operation to normal fp operation
94 // during development. When the backend supports strict float operation, this
95 // option will be meaningless.
96 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
97        cl::desc("Don't mutate strict-float node to a legalize node"),
98        cl::init(false), cl::Hidden);
99 
100 static bool darwinHasSinCos(const Triple &TT) {
101   assert(TT.isOSDarwin() && "should be called with darwin triple");
102   // Don't bother with 32 bit x86.
103   if (TT.getArch() == Triple::x86)
104     return false;
105   // Macos < 10.9 has no sincos_stret.
106   if (TT.isMacOSX())
107     return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
108   // iOS < 7.0 has no sincos_stret.
109   if (TT.isiOS())
110     return !TT.isOSVersionLT(7, 0);
111   // Any other darwin such as WatchOS/TvOS is new enough.
112   return true;
113 }
114 
115 // Although this default value is arbitrary, it is not random. It is assumed
116 // that a condition that evaluates the same way by a higher percentage than this
117 // is best represented as control flow. Therefore, the default value N should be
118 // set such that the win from N% correct executions is greater than the loss
119 // from (100 - N)% mispredicted executions for the majority of intended targets.
120 static cl::opt<int> MinPercentageForPredictableBranch(
121     "min-predictable-branch", cl::init(99),
122     cl::desc("Minimum percentage (0-100) that a condition must be either true "
123              "or false to assume that the condition is predictable"),
124     cl::Hidden);
125 
126 void TargetLoweringBase::InitLibcalls(const Triple &TT) {
127 #define HANDLE_LIBCALL(code, name) \
128   setLibcallName(RTLIB::code, name);
129 #include "llvm/IR/RuntimeLibcalls.def"
130 #undef HANDLE_LIBCALL
131   // Initialize calling conventions to their default.
132   for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC)
133     setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C);
134 
135   // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf".
136   if (TT.getArch() == Triple::ppc || TT.isPPC64()) {
137     setLibcallName(RTLIB::ADD_F128, "__addkf3");
138     setLibcallName(RTLIB::SUB_F128, "__subkf3");
139     setLibcallName(RTLIB::MUL_F128, "__mulkf3");
140     setLibcallName(RTLIB::DIV_F128, "__divkf3");
141     setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2");
142     setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2");
143     setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2");
144     setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2");
145     setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi");
146     setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi");
147     setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi");
148     setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi");
149     setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf");
150     setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf");
151     setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf");
152     setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf");
153     setLibcallName(RTLIB::OEQ_F128, "__eqkf2");
154     setLibcallName(RTLIB::UNE_F128, "__nekf2");
155     setLibcallName(RTLIB::OGE_F128, "__gekf2");
156     setLibcallName(RTLIB::OLT_F128, "__ltkf2");
157     setLibcallName(RTLIB::OLE_F128, "__lekf2");
158     setLibcallName(RTLIB::OGT_F128, "__gtkf2");
159     setLibcallName(RTLIB::UO_F128, "__unordkf2");
160   }
161 
162   // A few names are different on particular architectures or environments.
163   if (TT.isOSDarwin()) {
164     // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
165     // of the gnueabi-style __gnu_*_ieee.
166     // FIXME: What about other targets?
167     setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
168     setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
169 
170     // Some darwins have an optimized __bzero/bzero function.
171     switch (TT.getArch()) {
172     case Triple::x86:
173     case Triple::x86_64:
174       if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6))
175         setLibcallName(RTLIB::BZERO, "__bzero");
176       break;
177     case Triple::aarch64:
178     case Triple::aarch64_32:
179       setLibcallName(RTLIB::BZERO, "bzero");
180       break;
181     default:
182       break;
183     }
184 
185     if (darwinHasSinCos(TT)) {
186       setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret");
187       setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret");
188       if (TT.isWatchABI()) {
189         setLibcallCallingConv(RTLIB::SINCOS_STRET_F32,
190                               CallingConv::ARM_AAPCS_VFP);
191         setLibcallCallingConv(RTLIB::SINCOS_STRET_F64,
192                               CallingConv::ARM_AAPCS_VFP);
193       }
194     }
195   } else {
196     setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee");
197     setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee");
198   }
199 
200   if (TT.isGNUEnvironment() || TT.isOSFuchsia() ||
201       (TT.isAndroid() && !TT.isAndroidVersionLT(9))) {
202     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
203     setLibcallName(RTLIB::SINCOS_F64, "sincos");
204     setLibcallName(RTLIB::SINCOS_F80, "sincosl");
205     setLibcallName(RTLIB::SINCOS_F128, "sincosl");
206     setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
207   }
208 
209   if (TT.isPS4CPU()) {
210     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
211     setLibcallName(RTLIB::SINCOS_F64, "sincos");
212   }
213 
214   if (TT.isOSOpenBSD()) {
215     setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
216   }
217 }
218 
219 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
220 /// UNKNOWN_LIBCALL if there is none.
221 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
222   if (OpVT == MVT::f16) {
223     if (RetVT == MVT::f32)
224       return FPEXT_F16_F32;
225   } else if (OpVT == MVT::f32) {
226     if (RetVT == MVT::f64)
227       return FPEXT_F32_F64;
228     if (RetVT == MVT::f128)
229       return FPEXT_F32_F128;
230     if (RetVT == MVT::ppcf128)
231       return FPEXT_F32_PPCF128;
232   } else if (OpVT == MVT::f64) {
233     if (RetVT == MVT::f128)
234       return FPEXT_F64_F128;
235     else if (RetVT == MVT::ppcf128)
236       return FPEXT_F64_PPCF128;
237   } else if (OpVT == MVT::f80) {
238     if (RetVT == MVT::f128)
239       return FPEXT_F80_F128;
240   }
241 
242   return UNKNOWN_LIBCALL;
243 }
244 
245 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
246 /// UNKNOWN_LIBCALL if there is none.
247 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
248   if (RetVT == MVT::f16) {
249     if (OpVT == MVT::f32)
250       return FPROUND_F32_F16;
251     if (OpVT == MVT::f64)
252       return FPROUND_F64_F16;
253     if (OpVT == MVT::f80)
254       return FPROUND_F80_F16;
255     if (OpVT == MVT::f128)
256       return FPROUND_F128_F16;
257     if (OpVT == MVT::ppcf128)
258       return FPROUND_PPCF128_F16;
259   } else if (RetVT == MVT::f32) {
260     if (OpVT == MVT::f64)
261       return FPROUND_F64_F32;
262     if (OpVT == MVT::f80)
263       return FPROUND_F80_F32;
264     if (OpVT == MVT::f128)
265       return FPROUND_F128_F32;
266     if (OpVT == MVT::ppcf128)
267       return FPROUND_PPCF128_F32;
268   } else if (RetVT == MVT::f64) {
269     if (OpVT == MVT::f80)
270       return FPROUND_F80_F64;
271     if (OpVT == MVT::f128)
272       return FPROUND_F128_F64;
273     if (OpVT == MVT::ppcf128)
274       return FPROUND_PPCF128_F64;
275   } else if (RetVT == MVT::f80) {
276     if (OpVT == MVT::f128)
277       return FPROUND_F128_F80;
278   }
279 
280   return UNKNOWN_LIBCALL;
281 }
282 
283 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
284 /// UNKNOWN_LIBCALL if there is none.
285 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
286   if (OpVT == MVT::f32) {
287     if (RetVT == MVT::i32)
288       return FPTOSINT_F32_I32;
289     if (RetVT == MVT::i64)
290       return FPTOSINT_F32_I64;
291     if (RetVT == MVT::i128)
292       return FPTOSINT_F32_I128;
293   } else if (OpVT == MVT::f64) {
294     if (RetVT == MVT::i32)
295       return FPTOSINT_F64_I32;
296     if (RetVT == MVT::i64)
297       return FPTOSINT_F64_I64;
298     if (RetVT == MVT::i128)
299       return FPTOSINT_F64_I128;
300   } else if (OpVT == MVT::f80) {
301     if (RetVT == MVT::i32)
302       return FPTOSINT_F80_I32;
303     if (RetVT == MVT::i64)
304       return FPTOSINT_F80_I64;
305     if (RetVT == MVT::i128)
306       return FPTOSINT_F80_I128;
307   } else if (OpVT == MVT::f128) {
308     if (RetVT == MVT::i32)
309       return FPTOSINT_F128_I32;
310     if (RetVT == MVT::i64)
311       return FPTOSINT_F128_I64;
312     if (RetVT == MVT::i128)
313       return FPTOSINT_F128_I128;
314   } else if (OpVT == MVT::ppcf128) {
315     if (RetVT == MVT::i32)
316       return FPTOSINT_PPCF128_I32;
317     if (RetVT == MVT::i64)
318       return FPTOSINT_PPCF128_I64;
319     if (RetVT == MVT::i128)
320       return FPTOSINT_PPCF128_I128;
321   }
322   return UNKNOWN_LIBCALL;
323 }
324 
325 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
326 /// UNKNOWN_LIBCALL if there is none.
327 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
328   if (OpVT == MVT::f32) {
329     if (RetVT == MVT::i32)
330       return FPTOUINT_F32_I32;
331     if (RetVT == MVT::i64)
332       return FPTOUINT_F32_I64;
333     if (RetVT == MVT::i128)
334       return FPTOUINT_F32_I128;
335   } else if (OpVT == MVT::f64) {
336     if (RetVT == MVT::i32)
337       return FPTOUINT_F64_I32;
338     if (RetVT == MVT::i64)
339       return FPTOUINT_F64_I64;
340     if (RetVT == MVT::i128)
341       return FPTOUINT_F64_I128;
342   } else if (OpVT == MVT::f80) {
343     if (RetVT == MVT::i32)
344       return FPTOUINT_F80_I32;
345     if (RetVT == MVT::i64)
346       return FPTOUINT_F80_I64;
347     if (RetVT == MVT::i128)
348       return FPTOUINT_F80_I128;
349   } else if (OpVT == MVT::f128) {
350     if (RetVT == MVT::i32)
351       return FPTOUINT_F128_I32;
352     if (RetVT == MVT::i64)
353       return FPTOUINT_F128_I64;
354     if (RetVT == MVT::i128)
355       return FPTOUINT_F128_I128;
356   } else if (OpVT == MVT::ppcf128) {
357     if (RetVT == MVT::i32)
358       return FPTOUINT_PPCF128_I32;
359     if (RetVT == MVT::i64)
360       return FPTOUINT_PPCF128_I64;
361     if (RetVT == MVT::i128)
362       return FPTOUINT_PPCF128_I128;
363   }
364   return UNKNOWN_LIBCALL;
365 }
366 
367 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
368 /// UNKNOWN_LIBCALL if there is none.
369 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
370   if (OpVT == MVT::i32) {
371     if (RetVT == MVT::f32)
372       return SINTTOFP_I32_F32;
373     if (RetVT == MVT::f64)
374       return SINTTOFP_I32_F64;
375     if (RetVT == MVT::f80)
376       return SINTTOFP_I32_F80;
377     if (RetVT == MVT::f128)
378       return SINTTOFP_I32_F128;
379     if (RetVT == MVT::ppcf128)
380       return SINTTOFP_I32_PPCF128;
381   } else if (OpVT == MVT::i64) {
382     if (RetVT == MVT::f32)
383       return SINTTOFP_I64_F32;
384     if (RetVT == MVT::f64)
385       return SINTTOFP_I64_F64;
386     if (RetVT == MVT::f80)
387       return SINTTOFP_I64_F80;
388     if (RetVT == MVT::f128)
389       return SINTTOFP_I64_F128;
390     if (RetVT == MVT::ppcf128)
391       return SINTTOFP_I64_PPCF128;
392   } else if (OpVT == MVT::i128) {
393     if (RetVT == MVT::f32)
394       return SINTTOFP_I128_F32;
395     if (RetVT == MVT::f64)
396       return SINTTOFP_I128_F64;
397     if (RetVT == MVT::f80)
398       return SINTTOFP_I128_F80;
399     if (RetVT == MVT::f128)
400       return SINTTOFP_I128_F128;
401     if (RetVT == MVT::ppcf128)
402       return SINTTOFP_I128_PPCF128;
403   }
404   return UNKNOWN_LIBCALL;
405 }
406 
407 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
408 /// UNKNOWN_LIBCALL if there is none.
409 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
410   if (OpVT == MVT::i32) {
411     if (RetVT == MVT::f32)
412       return UINTTOFP_I32_F32;
413     if (RetVT == MVT::f64)
414       return UINTTOFP_I32_F64;
415     if (RetVT == MVT::f80)
416       return UINTTOFP_I32_F80;
417     if (RetVT == MVT::f128)
418       return UINTTOFP_I32_F128;
419     if (RetVT == MVT::ppcf128)
420       return UINTTOFP_I32_PPCF128;
421   } else if (OpVT == MVT::i64) {
422     if (RetVT == MVT::f32)
423       return UINTTOFP_I64_F32;
424     if (RetVT == MVT::f64)
425       return UINTTOFP_I64_F64;
426     if (RetVT == MVT::f80)
427       return UINTTOFP_I64_F80;
428     if (RetVT == MVT::f128)
429       return UINTTOFP_I64_F128;
430     if (RetVT == MVT::ppcf128)
431       return UINTTOFP_I64_PPCF128;
432   } else if (OpVT == MVT::i128) {
433     if (RetVT == MVT::f32)
434       return UINTTOFP_I128_F32;
435     if (RetVT == MVT::f64)
436       return UINTTOFP_I128_F64;
437     if (RetVT == MVT::f80)
438       return UINTTOFP_I128_F80;
439     if (RetVT == MVT::f128)
440       return UINTTOFP_I128_F128;
441     if (RetVT == MVT::ppcf128)
442       return UINTTOFP_I128_PPCF128;
443   }
444   return UNKNOWN_LIBCALL;
445 }
446 
447 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
448 #define OP_TO_LIBCALL(Name, Enum)                                              \
449   case Name:                                                                   \
450     switch (VT.SimpleTy) {                                                     \
451     default:                                                                   \
452       return UNKNOWN_LIBCALL;                                                  \
453     case MVT::i8:                                                              \
454       return Enum##_1;                                                         \
455     case MVT::i16:                                                             \
456       return Enum##_2;                                                         \
457     case MVT::i32:                                                             \
458       return Enum##_4;                                                         \
459     case MVT::i64:                                                             \
460       return Enum##_8;                                                         \
461     case MVT::i128:                                                            \
462       return Enum##_16;                                                        \
463     }
464 
465   switch (Opc) {
466     OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
467     OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
468     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
469     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
470     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
471     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
472     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
473     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
474     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
475     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
476     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
477     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
478   }
479 
480 #undef OP_TO_LIBCALL
481 
482   return UNKNOWN_LIBCALL;
483 }
484 
485 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
486   switch (ElementSize) {
487   case 1:
488     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
489   case 2:
490     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
491   case 4:
492     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
493   case 8:
494     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
495   case 16:
496     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
497   default:
498     return UNKNOWN_LIBCALL;
499   }
500 }
501 
502 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
503   switch (ElementSize) {
504   case 1:
505     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
506   case 2:
507     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
508   case 4:
509     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
510   case 8:
511     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
512   case 16:
513     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
514   default:
515     return UNKNOWN_LIBCALL;
516   }
517 }
518 
519 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
520   switch (ElementSize) {
521   case 1:
522     return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
523   case 2:
524     return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
525   case 4:
526     return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
527   case 8:
528     return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
529   case 16:
530     return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
531   default:
532     return UNKNOWN_LIBCALL;
533   }
534 }
535 
536 /// InitCmpLibcallCCs - Set default comparison libcall CC.
537 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
538   memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
539   CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
540   CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
541   CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
542   CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
543   CCs[RTLIB::UNE_F32] = ISD::SETNE;
544   CCs[RTLIB::UNE_F64] = ISD::SETNE;
545   CCs[RTLIB::UNE_F128] = ISD::SETNE;
546   CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
547   CCs[RTLIB::OGE_F32] = ISD::SETGE;
548   CCs[RTLIB::OGE_F64] = ISD::SETGE;
549   CCs[RTLIB::OGE_F128] = ISD::SETGE;
550   CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
551   CCs[RTLIB::OLT_F32] = ISD::SETLT;
552   CCs[RTLIB::OLT_F64] = ISD::SETLT;
553   CCs[RTLIB::OLT_F128] = ISD::SETLT;
554   CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
555   CCs[RTLIB::OLE_F32] = ISD::SETLE;
556   CCs[RTLIB::OLE_F64] = ISD::SETLE;
557   CCs[RTLIB::OLE_F128] = ISD::SETLE;
558   CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
559   CCs[RTLIB::OGT_F32] = ISD::SETGT;
560   CCs[RTLIB::OGT_F64] = ISD::SETGT;
561   CCs[RTLIB::OGT_F128] = ISD::SETGT;
562   CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
563   CCs[RTLIB::UO_F32] = ISD::SETNE;
564   CCs[RTLIB::UO_F64] = ISD::SETNE;
565   CCs[RTLIB::UO_F128] = ISD::SETNE;
566   CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
567 }
568 
569 /// NOTE: The TargetMachine owns TLOF.
570 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
571   initActions();
572 
573   // Perform these initializations only once.
574   MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
575       MaxLoadsPerMemcmp = 8;
576   MaxGluedStoresPerMemcpy = 0;
577   MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
578       MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
579   HasMultipleConditionRegisters = false;
580   HasExtractBitsInsn = false;
581   JumpIsExpensive = JumpIsExpensiveOverride;
582   PredictableSelectIsExpensive = false;
583   EnableExtLdPromotion = false;
584   StackPointerRegisterToSaveRestore = 0;
585   BooleanContents = UndefinedBooleanContent;
586   BooleanFloatContents = UndefinedBooleanContent;
587   BooleanVectorContents = UndefinedBooleanContent;
588   SchedPreferenceInfo = Sched::ILP;
589   GatherAllAliasesMaxDepth = 18;
590   IsStrictFPEnabled = DisableStrictNodeMutation;
591   // TODO: the default will be switched to 0 in the next commit, along
592   // with the Target-specific changes necessary.
593   MaxAtomicSizeInBitsSupported = 1024;
594 
595   MinCmpXchgSizeInBits = 0;
596   SupportsUnalignedAtomics = false;
597 
598   std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
599 
600   InitLibcalls(TM.getTargetTriple());
601   InitCmpLibcallCCs(CmpLibcallCCs);
602 }
603 
604 void TargetLoweringBase::initActions() {
605   // All operations default to being supported.
606   memset(OpActions, 0, sizeof(OpActions));
607   memset(LoadExtActions, 0, sizeof(LoadExtActions));
608   memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
609   memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
610   memset(CondCodeActions, 0, sizeof(CondCodeActions));
611   std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
612   std::fill(std::begin(TargetDAGCombineArray),
613             std::end(TargetDAGCombineArray), 0);
614 
615   for (MVT VT : MVT::fp_valuetypes()) {
616     MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits());
617     if (IntVT.isValid()) {
618       setOperationAction(ISD::ATOMIC_SWAP, VT, Promote);
619       AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT);
620     }
621   }
622 
623   // Set default actions for various operations.
624   for (MVT VT : MVT::all_valuetypes()) {
625     // Default all indexed load / store to expand.
626     for (unsigned IM = (unsigned)ISD::PRE_INC;
627          IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
628       setIndexedLoadAction(IM, VT, Expand);
629       setIndexedStoreAction(IM, VT, Expand);
630       setIndexedMaskedLoadAction(IM, VT, Expand);
631       setIndexedMaskedStoreAction(IM, VT, Expand);
632     }
633 
634     // Most backends expect to see the node which just returns the value loaded.
635     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
636 
637     // These operations default to expand.
638     setOperationAction(ISD::FGETSIGN, VT, Expand);
639     setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
640     setOperationAction(ISD::FMINNUM, VT, Expand);
641     setOperationAction(ISD::FMAXNUM, VT, Expand);
642     setOperationAction(ISD::FMINNUM_IEEE, VT, Expand);
643     setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand);
644     setOperationAction(ISD::FMINIMUM, VT, Expand);
645     setOperationAction(ISD::FMAXIMUM, VT, Expand);
646     setOperationAction(ISD::FMAD, VT, Expand);
647     setOperationAction(ISD::SMIN, VT, Expand);
648     setOperationAction(ISD::SMAX, VT, Expand);
649     setOperationAction(ISD::UMIN, VT, Expand);
650     setOperationAction(ISD::UMAX, VT, Expand);
651     setOperationAction(ISD::ABS, VT, Expand);
652     setOperationAction(ISD::FSHL, VT, Expand);
653     setOperationAction(ISD::FSHR, VT, Expand);
654     setOperationAction(ISD::SADDSAT, VT, Expand);
655     setOperationAction(ISD::UADDSAT, VT, Expand);
656     setOperationAction(ISD::SSUBSAT, VT, Expand);
657     setOperationAction(ISD::USUBSAT, VT, Expand);
658     setOperationAction(ISD::SMULFIX, VT, Expand);
659     setOperationAction(ISD::SMULFIXSAT, VT, Expand);
660     setOperationAction(ISD::UMULFIX, VT, Expand);
661     setOperationAction(ISD::UMULFIXSAT, VT, Expand);
662     setOperationAction(ISD::SDIVFIX, VT, Expand);
663     setOperationAction(ISD::UDIVFIX, VT, Expand);
664 
665     // Overflow operations default to expand
666     setOperationAction(ISD::SADDO, VT, Expand);
667     setOperationAction(ISD::SSUBO, VT, Expand);
668     setOperationAction(ISD::UADDO, VT, Expand);
669     setOperationAction(ISD::USUBO, VT, Expand);
670     setOperationAction(ISD::SMULO, VT, Expand);
671     setOperationAction(ISD::UMULO, VT, Expand);
672 
673     // ADDCARRY operations default to expand
674     setOperationAction(ISD::ADDCARRY, VT, Expand);
675     setOperationAction(ISD::SUBCARRY, VT, Expand);
676     setOperationAction(ISD::SETCCCARRY, VT, Expand);
677 
678     // ADDC/ADDE/SUBC/SUBE default to expand.
679     setOperationAction(ISD::ADDC, VT, Expand);
680     setOperationAction(ISD::ADDE, VT, Expand);
681     setOperationAction(ISD::SUBC, VT, Expand);
682     setOperationAction(ISD::SUBE, VT, Expand);
683 
684     // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
685     setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
686     setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
687 
688     setOperationAction(ISD::BITREVERSE, VT, Expand);
689 
690     // These library functions default to expand.
691     setOperationAction(ISD::FROUND, VT, Expand);
692     setOperationAction(ISD::FPOWI, VT, Expand);
693 
694     // These operations default to expand for vector types.
695     if (VT.isVector()) {
696       setOperationAction(ISD::FCOPYSIGN, VT, Expand);
697       setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
698       setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
699       setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
700       setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
701       setOperationAction(ISD::SPLAT_VECTOR, VT, Expand);
702     }
703 
704     // Constrained floating-point operations default to expand.
705 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN)               \
706     setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
707 #include "llvm/IR/ConstrainedOps.def"
708 
709     // For most targets @llvm.get.dynamic.area.offset just returns 0.
710     setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
711 
712     // Vector reduction default to expand.
713     setOperationAction(ISD::VECREDUCE_FADD, VT, Expand);
714     setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand);
715     setOperationAction(ISD::VECREDUCE_ADD, VT, Expand);
716     setOperationAction(ISD::VECREDUCE_MUL, VT, Expand);
717     setOperationAction(ISD::VECREDUCE_AND, VT, Expand);
718     setOperationAction(ISD::VECREDUCE_OR, VT, Expand);
719     setOperationAction(ISD::VECREDUCE_XOR, VT, Expand);
720     setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand);
721     setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand);
722     setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand);
723     setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand);
724     setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand);
725     setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand);
726   }
727 
728   // Most targets ignore the @llvm.prefetch intrinsic.
729   setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
730 
731   // Most targets also ignore the @llvm.readcyclecounter intrinsic.
732   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
733 
734   // ConstantFP nodes default to expand.  Targets can either change this to
735   // Legal, in which case all fp constants are legal, or use isFPImmLegal()
736   // to optimize expansions for certain constants.
737   setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
738   setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
739   setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
740   setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
741   setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
742 
743   // These library functions default to expand.
744   for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
745     setOperationAction(ISD::FCBRT,      VT, Expand);
746     setOperationAction(ISD::FLOG ,      VT, Expand);
747     setOperationAction(ISD::FLOG2,      VT, Expand);
748     setOperationAction(ISD::FLOG10,     VT, Expand);
749     setOperationAction(ISD::FEXP ,      VT, Expand);
750     setOperationAction(ISD::FEXP2,      VT, Expand);
751     setOperationAction(ISD::FFLOOR,     VT, Expand);
752     setOperationAction(ISD::FNEARBYINT, VT, Expand);
753     setOperationAction(ISD::FCEIL,      VT, Expand);
754     setOperationAction(ISD::FRINT,      VT, Expand);
755     setOperationAction(ISD::FTRUNC,     VT, Expand);
756     setOperationAction(ISD::FROUND,     VT, Expand);
757     setOperationAction(ISD::LROUND,     VT, Expand);
758     setOperationAction(ISD::LLROUND,    VT, Expand);
759     setOperationAction(ISD::LRINT,      VT, Expand);
760     setOperationAction(ISD::LLRINT,     VT, Expand);
761   }
762 
763   // Default ISD::TRAP to expand (which turns it into abort).
764   setOperationAction(ISD::TRAP, MVT::Other, Expand);
765 
766   // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
767   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
768   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
769 }
770 
771 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
772                                                EVT) const {
773   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
774 }
775 
776 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
777                                          bool LegalTypes) const {
778   assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
779   if (LHSTy.isVector())
780     return LHSTy;
781   return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
782                     : getPointerTy(DL);
783 }
784 
785 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
786   assert(isTypeLegal(VT));
787   switch (Op) {
788   default:
789     return false;
790   case ISD::SDIV:
791   case ISD::UDIV:
792   case ISD::SREM:
793   case ISD::UREM:
794     return true;
795   }
796 }
797 
798 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
799   // If the command-line option was specified, ignore this request.
800   if (!JumpIsExpensiveOverride.getNumOccurrences())
801     JumpIsExpensive = isExpensive;
802 }
803 
804 TargetLoweringBase::LegalizeKind
805 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
806   // If this is a simple type, use the ComputeRegisterProp mechanism.
807   if (VT.isSimple()) {
808     MVT SVT = VT.getSimpleVT();
809     assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
810     MVT NVT = TransformToType[SVT.SimpleTy];
811     LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
812 
813     assert((LA == TypeLegal || LA == TypeSoftenFloat ||
814             LA == TypeSoftPromoteHalf ||
815             (NVT.isVector() ||
816              ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
817            "Promote may not follow Expand or Promote");
818 
819     if (LA == TypeSplitVector)
820       return LegalizeKind(LA,
821                           EVT::getVectorVT(Context, SVT.getVectorElementType(),
822                                            SVT.getVectorNumElements() / 2));
823     if (LA == TypeScalarizeVector)
824       return LegalizeKind(LA, SVT.getVectorElementType());
825     return LegalizeKind(LA, NVT);
826   }
827 
828   // Handle Extended Scalar Types.
829   if (!VT.isVector()) {
830     assert(VT.isInteger() && "Float types must be simple");
831     unsigned BitSize = VT.getSizeInBits();
832     // First promote to a power-of-two size, then expand if necessary.
833     if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
834       EVT NVT = VT.getRoundIntegerType(Context);
835       assert(NVT != VT && "Unable to round integer VT");
836       LegalizeKind NextStep = getTypeConversion(Context, NVT);
837       // Avoid multi-step promotion.
838       if (NextStep.first == TypePromoteInteger)
839         return NextStep;
840       // Return rounded integer type.
841       return LegalizeKind(TypePromoteInteger, NVT);
842     }
843 
844     return LegalizeKind(TypeExpandInteger,
845                         EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
846   }
847 
848   // Handle vector types.
849   unsigned NumElts = VT.getVectorNumElements();
850   EVT EltVT = VT.getVectorElementType();
851 
852   // Vectors with only one element are always scalarized.
853   if (NumElts == 1)
854     return LegalizeKind(TypeScalarizeVector, EltVT);
855 
856   // Try to widen vector elements until the element type is a power of two and
857   // promote it to a legal type later on, for example:
858   // <3 x i8> -> <4 x i8> -> <4 x i32>
859   if (EltVT.isInteger()) {
860     // Vectors with a number of elements that is not a power of two are always
861     // widened, for example <3 x i8> -> <4 x i8>.
862     if (!VT.isPow2VectorType()) {
863       NumElts = (unsigned)NextPowerOf2(NumElts);
864       EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
865       return LegalizeKind(TypeWidenVector, NVT);
866     }
867 
868     // Examine the element type.
869     LegalizeKind LK = getTypeConversion(Context, EltVT);
870 
871     // If type is to be expanded, split the vector.
872     //  <4 x i140> -> <2 x i140>
873     if (LK.first == TypeExpandInteger)
874       return LegalizeKind(TypeSplitVector,
875                           EVT::getVectorVT(Context, EltVT, NumElts / 2));
876 
877     // Promote the integer element types until a legal vector type is found
878     // or until the element integer type is too big. If a legal type was not
879     // found, fallback to the usual mechanism of widening/splitting the
880     // vector.
881     EVT OldEltVT = EltVT;
882     while (true) {
883       // Increase the bitwidth of the element to the next pow-of-two
884       // (which is greater than 8 bits).
885       EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
886                   .getRoundIntegerType(Context);
887 
888       // Stop trying when getting a non-simple element type.
889       // Note that vector elements may be greater than legal vector element
890       // types. Example: X86 XMM registers hold 64bit element on 32bit
891       // systems.
892       if (!EltVT.isSimple())
893         break;
894 
895       // Build a new vector type and check if it is legal.
896       MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
897       // Found a legal promoted vector type.
898       if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
899         return LegalizeKind(TypePromoteInteger,
900                             EVT::getVectorVT(Context, EltVT, NumElts));
901     }
902 
903     // Reset the type to the unexpanded type if we did not find a legal vector
904     // type with a promoted vector element type.
905     EltVT = OldEltVT;
906   }
907 
908   // Try to widen the vector until a legal type is found.
909   // If there is no wider legal type, split the vector.
910   while (true) {
911     // Round up to the next power of 2.
912     NumElts = (unsigned)NextPowerOf2(NumElts);
913 
914     // If there is no simple vector type with this many elements then there
915     // cannot be a larger legal vector type.  Note that this assumes that
916     // there are no skipped intermediate vector types in the simple types.
917     if (!EltVT.isSimple())
918       break;
919     MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
920     if (LargerVector == MVT())
921       break;
922 
923     // If this type is legal then widen the vector.
924     if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
925       return LegalizeKind(TypeWidenVector, LargerVector);
926   }
927 
928   // Widen odd vectors to next power of two.
929   if (!VT.isPow2VectorType()) {
930     EVT NVT = VT.getPow2VectorType(Context);
931     return LegalizeKind(TypeWidenVector, NVT);
932   }
933 
934   // Vectors with illegal element types are expanded.
935   EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
936   return LegalizeKind(TypeSplitVector, NVT);
937 }
938 
939 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
940                                           unsigned &NumIntermediates,
941                                           MVT &RegisterVT,
942                                           TargetLoweringBase *TLI) {
943   // Figure out the right, legal destination reg to copy into.
944   unsigned NumElts = VT.getVectorNumElements();
945   MVT EltTy = VT.getVectorElementType();
946 
947   unsigned NumVectorRegs = 1;
948 
949   // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally we
950   // could break down into LHS/RHS like LegalizeDAG does.
951   if (!isPowerOf2_32(NumElts)) {
952     NumVectorRegs = NumElts;
953     NumElts = 1;
954   }
955 
956   // Divide the input until we get to a supported size.  This will always
957   // end with a scalar if the target doesn't support vectors.
958   while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
959     NumElts >>= 1;
960     NumVectorRegs <<= 1;
961   }
962 
963   NumIntermediates = NumVectorRegs;
964 
965   MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
966   if (!TLI->isTypeLegal(NewVT))
967     NewVT = EltTy;
968   IntermediateVT = NewVT;
969 
970   unsigned NewVTSize = NewVT.getSizeInBits();
971 
972   // Convert sizes such as i33 to i64.
973   if (!isPowerOf2_32(NewVTSize))
974     NewVTSize = NextPowerOf2(NewVTSize);
975 
976   MVT DestVT = TLI->getRegisterType(NewVT);
977   RegisterVT = DestVT;
978   if (EVT(DestVT).bitsLT(NewVT))    // Value is expanded, e.g. i64 -> i16.
979     return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
980 
981   // Otherwise, promotion or legal types use the same number of registers as
982   // the vector decimated to the appropriate level.
983   return NumVectorRegs;
984 }
985 
986 /// isLegalRC - Return true if the value types that can be represented by the
987 /// specified register class are all legal.
988 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
989                                    const TargetRegisterClass &RC) const {
990   for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
991     if (isTypeLegal(*I))
992       return true;
993   return false;
994 }
995 
996 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
997 /// sequence of memory operands that is recognized by PrologEpilogInserter.
998 MachineBasicBlock *
999 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1000                                    MachineBasicBlock *MBB) const {
1001   MachineInstr *MI = &InitialMI;
1002   MachineFunction &MF = *MI->getMF();
1003   MachineFrameInfo &MFI = MF.getFrameInfo();
1004 
1005   // We're handling multiple types of operands here:
1006   // PATCHPOINT MetaArgs - live-in, read only, direct
1007   // STATEPOINT Deopt Spill - live-through, read only, indirect
1008   // STATEPOINT Deopt Alloca - live-through, read only, direct
1009   // (We're currently conservative and mark the deopt slots read/write in
1010   // practice.)
1011   // STATEPOINT GC Spill - live-through, read/write, indirect
1012   // STATEPOINT GC Alloca - live-through, read/write, direct
1013   // The live-in vs live-through is handled already (the live through ones are
1014   // all stack slots), but we need to handle the different type of stackmap
1015   // operands and memory effects here.
1016 
1017   // MI changes inside this loop as we grow operands.
1018   for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) {
1019     MachineOperand &MO = MI->getOperand(OperIdx);
1020     if (!MO.isFI())
1021       continue;
1022 
1023     // foldMemoryOperand builds a new MI after replacing a single FI operand
1024     // with the canonical set of five x86 addressing-mode operands.
1025     int FI = MO.getIndex();
1026     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1027 
1028     // Copy operands before the frame-index.
1029     for (unsigned i = 0; i < OperIdx; ++i)
1030       MIB.add(MI->getOperand(i));
1031     // Add frame index operands recognized by stackmaps.cpp
1032     if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
1033       // indirect-mem-ref tag, size, #FI, offset.
1034       // Used for spills inserted by StatepointLowering.  This codepath is not
1035       // used for patchpoints/stackmaps at all, for these spilling is done via
1036       // foldMemoryOperand callback only.
1037       assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1038       MIB.addImm(StackMaps::IndirectMemRefOp);
1039       MIB.addImm(MFI.getObjectSize(FI));
1040       MIB.add(MI->getOperand(OperIdx));
1041       MIB.addImm(0);
1042     } else {
1043       // direct-mem-ref tag, #FI, offset.
1044       // Used by patchpoint, and direct alloca arguments to statepoints
1045       MIB.addImm(StackMaps::DirectMemRefOp);
1046       MIB.add(MI->getOperand(OperIdx));
1047       MIB.addImm(0);
1048     }
1049     // Copy the operands after the frame index.
1050     for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i)
1051       MIB.add(MI->getOperand(i));
1052 
1053     // Inherit previous memory operands.
1054     MIB.cloneMemRefs(*MI);
1055     assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1056 
1057     // Add a new memory operand for this FI.
1058     assert(MFI.getObjectOffset(FI) != -1);
1059 
1060     // Note: STATEPOINT MMOs are added during SelectionDAG.  STACKMAP, and
1061     // PATCHPOINT should be updated to do the same. (TODO)
1062     if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1063       auto Flags = MachineMemOperand::MOLoad;
1064       MachineMemOperand *MMO = MF.getMachineMemOperand(
1065           MachinePointerInfo::getFixedStack(MF, FI), Flags,
1066           MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI));
1067       MIB->addMemOperand(MF, MMO);
1068     }
1069 
1070     // Replace the instruction and update the operand index.
1071     MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1072     OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1;
1073     MI->eraseFromParent();
1074     MI = MIB;
1075   }
1076   return MBB;
1077 }
1078 
1079 MachineBasicBlock *
1080 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
1081                                         MachineBasicBlock *MBB) const {
1082   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1083          "Called emitXRayCustomEvent on the wrong MI!");
1084   auto &MF = *MI.getMF();
1085   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1086   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1087     MIB.add(MI.getOperand(OpIdx));
1088 
1089   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1090   MI.eraseFromParent();
1091   return MBB;
1092 }
1093 
1094 MachineBasicBlock *
1095 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1096                                        MachineBasicBlock *MBB) const {
1097   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1098          "Called emitXRayTypedEvent on the wrong MI!");
1099   auto &MF = *MI.getMF();
1100   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1101   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1102     MIB.add(MI.getOperand(OpIdx));
1103 
1104   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1105   MI.eraseFromParent();
1106   return MBB;
1107 }
1108 
1109 /// findRepresentativeClass - Return the largest legal super-reg register class
1110 /// of the register class for the specified type and its associated "cost".
1111 // This function is in TargetLowering because it uses RegClassForVT which would
1112 // need to be moved to TargetRegisterInfo and would necessitate moving
1113 // isTypeLegal over as well - a massive change that would just require
1114 // TargetLowering having a TargetRegisterInfo class member that it would use.
1115 std::pair<const TargetRegisterClass *, uint8_t>
1116 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1117                                             MVT VT) const {
1118   const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1119   if (!RC)
1120     return std::make_pair(RC, 0);
1121 
1122   // Compute the set of all super-register classes.
1123   BitVector SuperRegRC(TRI->getNumRegClasses());
1124   for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1125     SuperRegRC.setBitsInMask(RCI.getMask());
1126 
1127   // Find the first legal register class with the largest spill size.
1128   const TargetRegisterClass *BestRC = RC;
1129   for (unsigned i : SuperRegRC.set_bits()) {
1130     const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1131     // We want the largest possible spill size.
1132     if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1133       continue;
1134     if (!isLegalRC(*TRI, *SuperRC))
1135       continue;
1136     BestRC = SuperRC;
1137   }
1138   return std::make_pair(BestRC, 1);
1139 }
1140 
1141 /// computeRegisterProperties - Once all of the register classes are added,
1142 /// this allows us to compute derived properties we expose.
1143 void TargetLoweringBase::computeRegisterProperties(
1144     const TargetRegisterInfo *TRI) {
1145   static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1146                 "Too many value types for ValueTypeActions to hold!");
1147 
1148   // Everything defaults to needing one register.
1149   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1150     NumRegistersForVT[i] = 1;
1151     RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1152   }
1153   // ...except isVoid, which doesn't need any registers.
1154   NumRegistersForVT[MVT::isVoid] = 0;
1155 
1156   // Find the largest integer register class.
1157   unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1158   for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1159     assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1160 
1161   // Every integer value type larger than this largest register takes twice as
1162   // many registers to represent as the previous ValueType.
1163   for (unsigned ExpandedReg = LargestIntReg + 1;
1164        ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1165     NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1166     RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1167     TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1168     ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1169                                    TypeExpandInteger);
1170   }
1171 
1172   // Inspect all of the ValueType's smaller than the largest integer
1173   // register to see which ones need promotion.
1174   unsigned LegalIntReg = LargestIntReg;
1175   for (unsigned IntReg = LargestIntReg - 1;
1176        IntReg >= (unsigned)MVT::i1; --IntReg) {
1177     MVT IVT = (MVT::SimpleValueType)IntReg;
1178     if (isTypeLegal(IVT)) {
1179       LegalIntReg = IntReg;
1180     } else {
1181       RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1182         (MVT::SimpleValueType)LegalIntReg;
1183       ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1184     }
1185   }
1186 
1187   // ppcf128 type is really two f64's.
1188   if (!isTypeLegal(MVT::ppcf128)) {
1189     if (isTypeLegal(MVT::f64)) {
1190       NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1191       RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1192       TransformToType[MVT::ppcf128] = MVT::f64;
1193       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1194     } else {
1195       NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1196       RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1197       TransformToType[MVT::ppcf128] = MVT::i128;
1198       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1199     }
1200   }
1201 
1202   // Decide how to handle f128. If the target does not have native f128 support,
1203   // expand it to i128 and we will be generating soft float library calls.
1204   if (!isTypeLegal(MVT::f128)) {
1205     NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1206     RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1207     TransformToType[MVT::f128] = MVT::i128;
1208     ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1209   }
1210 
1211   // Decide how to handle f64. If the target does not have native f64 support,
1212   // expand it to i64 and we will be generating soft float library calls.
1213   if (!isTypeLegal(MVT::f64)) {
1214     NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1215     RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1216     TransformToType[MVT::f64] = MVT::i64;
1217     ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1218   }
1219 
1220   // Decide how to handle f32. If the target does not have native f32 support,
1221   // expand it to i32 and we will be generating soft float library calls.
1222   if (!isTypeLegal(MVT::f32)) {
1223     NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1224     RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1225     TransformToType[MVT::f32] = MVT::i32;
1226     ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1227   }
1228 
1229   // Decide how to handle f16. If the target does not have native f16 support,
1230   // promote it to f32, because there are no f16 library calls (except for
1231   // conversions).
1232   if (!isTypeLegal(MVT::f16)) {
1233     // Allow targets to control how we legalize half.
1234     if (softPromoteHalfType()) {
1235       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1236       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1237       TransformToType[MVT::f16] = MVT::f32;
1238       ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1239     } else {
1240       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1241       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1242       TransformToType[MVT::f16] = MVT::f32;
1243       ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1244     }
1245   }
1246 
1247   // Loop over all of the vector value types to see which need transformations.
1248   for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1249        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1250     MVT VT = (MVT::SimpleValueType) i;
1251     if (isTypeLegal(VT))
1252       continue;
1253 
1254     MVT EltVT = VT.getVectorElementType();
1255     unsigned NElts = VT.getVectorNumElements();
1256     bool IsLegalWiderType = false;
1257     bool IsScalable = VT.isScalableVector();
1258     LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1259     switch (PreferredAction) {
1260     case TypePromoteInteger: {
1261       MVT::SimpleValueType EndVT = IsScalable ?
1262                                    MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1263                                    MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1264       // Try to promote the elements of integer vectors. If no legal
1265       // promotion was found, fall through to the widen-vector method.
1266       for (unsigned nVT = i + 1;
1267            (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1268         MVT SVT = (MVT::SimpleValueType) nVT;
1269         // Promote vectors of integers to vectors with the same number
1270         // of elements, with a wider element type.
1271         if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() &&
1272             SVT.getVectorNumElements() == NElts &&
1273             SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) {
1274           TransformToType[i] = SVT;
1275           RegisterTypeForVT[i] = SVT;
1276           NumRegistersForVT[i] = 1;
1277           ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1278           IsLegalWiderType = true;
1279           break;
1280         }
1281       }
1282       if (IsLegalWiderType)
1283         break;
1284       LLVM_FALLTHROUGH;
1285     }
1286 
1287     case TypeWidenVector:
1288       if (isPowerOf2_32(NElts)) {
1289         // Try to widen the vector.
1290         for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1291           MVT SVT = (MVT::SimpleValueType) nVT;
1292           if (SVT.getVectorElementType() == EltVT
1293               && SVT.getVectorNumElements() > NElts
1294               && SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) {
1295             TransformToType[i] = SVT;
1296             RegisterTypeForVT[i] = SVT;
1297             NumRegistersForVT[i] = 1;
1298             ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1299             IsLegalWiderType = true;
1300             break;
1301           }
1302         }
1303         if (IsLegalWiderType)
1304           break;
1305       } else {
1306         // Only widen to the next power of 2 to keep consistency with EVT.
1307         MVT NVT = VT.getPow2VectorType();
1308         if (isTypeLegal(NVT)) {
1309           TransformToType[i] = NVT;
1310           ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1311           RegisterTypeForVT[i] = NVT;
1312           NumRegistersForVT[i] = 1;
1313           break;
1314         }
1315       }
1316       LLVM_FALLTHROUGH;
1317 
1318     case TypeSplitVector:
1319     case TypeScalarizeVector: {
1320       MVT IntermediateVT;
1321       MVT RegisterVT;
1322       unsigned NumIntermediates;
1323       unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1324           NumIntermediates, RegisterVT, this);
1325       NumRegistersForVT[i] = NumRegisters;
1326       assert(NumRegistersForVT[i] == NumRegisters &&
1327              "NumRegistersForVT size cannot represent NumRegisters!");
1328       RegisterTypeForVT[i] = RegisterVT;
1329 
1330       MVT NVT = VT.getPow2VectorType();
1331       if (NVT == VT) {
1332         // Type is already a power of 2.  The default action is to split.
1333         TransformToType[i] = MVT::Other;
1334         if (PreferredAction == TypeScalarizeVector)
1335           ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1336         else if (PreferredAction == TypeSplitVector)
1337           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1338         else
1339           // Set type action according to the number of elements.
1340           ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector
1341                                                         : TypeSplitVector);
1342       } else {
1343         TransformToType[i] = NVT;
1344         ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1345       }
1346       break;
1347     }
1348     default:
1349       llvm_unreachable("Unknown vector legalization action!");
1350     }
1351   }
1352 
1353   // Determine the 'representative' register class for each value type.
1354   // An representative register class is the largest (meaning one which is
1355   // not a sub-register class / subreg register class) legal register class for
1356   // a group of value types. For example, on i386, i8, i16, and i32
1357   // representative would be GR32; while on x86_64 it's GR64.
1358   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1359     const TargetRegisterClass* RRC;
1360     uint8_t Cost;
1361     std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1362     RepRegClassForVT[i] = RRC;
1363     RepRegClassCostForVT[i] = Cost;
1364   }
1365 }
1366 
1367 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1368                                            EVT VT) const {
1369   assert(!VT.isVector() && "No default SetCC type for vectors!");
1370   return getPointerTy(DL).SimpleTy;
1371 }
1372 
1373 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1374   return MVT::i32; // return the default value
1375 }
1376 
1377 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1378 /// legal first class types.  For example, MVT::v8f32 maps to 2 MVT::v4f32
1379 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1380 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1381 ///
1382 /// This method returns the number of registers needed, and the VT for each
1383 /// register.  It also returns the VT and quantity of the intermediate values
1384 /// before they are promoted/expanded.
1385 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1386                                                 EVT &IntermediateVT,
1387                                                 unsigned &NumIntermediates,
1388                                                 MVT &RegisterVT) const {
1389   unsigned NumElts = VT.getVectorNumElements();
1390 
1391   // If there is a wider vector type with the same element type as this one,
1392   // or a promoted vector type that has the same number of elements which
1393   // are wider, then we should convert to that legal vector type.
1394   // This handles things like <2 x float> -> <4 x float> and
1395   // <4 x i1> -> <4 x i32>.
1396   LegalizeTypeAction TA = getTypeAction(Context, VT);
1397   if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1398     EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1399     if (isTypeLegal(RegisterEVT)) {
1400       IntermediateVT = RegisterEVT;
1401       RegisterVT = RegisterEVT.getSimpleVT();
1402       NumIntermediates = 1;
1403       return 1;
1404     }
1405   }
1406 
1407   // Figure out the right, legal destination reg to copy into.
1408   EVT EltTy = VT.getVectorElementType();
1409 
1410   unsigned NumVectorRegs = 1;
1411 
1412   // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally we
1413   // could break down into LHS/RHS like LegalizeDAG does.
1414   if (!isPowerOf2_32(NumElts)) {
1415     NumVectorRegs = NumElts;
1416     NumElts = 1;
1417   }
1418 
1419   // Divide the input until we get to a supported size.  This will always
1420   // end with a scalar if the target doesn't support vectors.
1421   while (NumElts > 1 && !isTypeLegal(
1422                                    EVT::getVectorVT(Context, EltTy, NumElts))) {
1423     NumElts >>= 1;
1424     NumVectorRegs <<= 1;
1425   }
1426 
1427   NumIntermediates = NumVectorRegs;
1428 
1429   EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1430   if (!isTypeLegal(NewVT))
1431     NewVT = EltTy;
1432   IntermediateVT = NewVT;
1433 
1434   MVT DestVT = getRegisterType(Context, NewVT);
1435   RegisterVT = DestVT;
1436   unsigned NewVTSize = NewVT.getSizeInBits();
1437 
1438   // Convert sizes such as i33 to i64.
1439   if (!isPowerOf2_32(NewVTSize))
1440     NewVTSize = NextPowerOf2(NewVTSize);
1441 
1442   if (EVT(DestVT).bitsLT(NewVT))   // Value is expanded, e.g. i64 -> i16.
1443     return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1444 
1445   // Otherwise, promotion or legal types use the same number of registers as
1446   // the vector decimated to the appropriate level.
1447   return NumVectorRegs;
1448 }
1449 
1450 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI,
1451                                                 uint64_t NumCases,
1452                                                 uint64_t Range,
1453                                                 ProfileSummaryInfo *PSI,
1454                                                 BlockFrequencyInfo *BFI) const {
1455   // FIXME: This function check the maximum table size and density, but the
1456   // minimum size is not checked. It would be nice if the minimum size is
1457   // also combined within this function. Currently, the minimum size check is
1458   // performed in findJumpTable() in SelectionDAGBuiler and
1459   // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1460   const bool OptForSize =
1461       SI->getParent()->getParent()->hasOptSize() ||
1462       llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
1463   const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1464   const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1465 
1466   // Check whether the number of cases is small enough and
1467   // the range is dense enough for a jump table.
1468   return (OptForSize || Range <= MaxJumpTableSize) &&
1469          (NumCases * 100 >= Range * MinDensity);
1470 }
1471 
1472 /// Get the EVTs and ArgFlags collections that represent the legalized return
1473 /// type of the given function.  This does not require a DAG or a return value,
1474 /// and is suitable for use before any DAGs for the function are constructed.
1475 /// TODO: Move this out of TargetLowering.cpp.
1476 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1477                          AttributeList attr,
1478                          SmallVectorImpl<ISD::OutputArg> &Outs,
1479                          const TargetLowering &TLI, const DataLayout &DL) {
1480   SmallVector<EVT, 4> ValueVTs;
1481   ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1482   unsigned NumValues = ValueVTs.size();
1483   if (NumValues == 0) return;
1484 
1485   for (unsigned j = 0, f = NumValues; j != f; ++j) {
1486     EVT VT = ValueVTs[j];
1487     ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1488 
1489     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1490       ExtendKind = ISD::SIGN_EXTEND;
1491     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1492       ExtendKind = ISD::ZERO_EXTEND;
1493 
1494     // FIXME: C calling convention requires the return type to be promoted to
1495     // at least 32-bit. But this is not necessary for non-C calling
1496     // conventions. The frontend should mark functions whose return values
1497     // require promoting with signext or zeroext attributes.
1498     if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1499       MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1500       if (VT.bitsLT(MinVT))
1501         VT = MinVT;
1502     }
1503 
1504     unsigned NumParts =
1505         TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1506     MVT PartVT =
1507         TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1508 
1509     // 'inreg' on function refers to return value
1510     ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1511     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1512       Flags.setInReg();
1513 
1514     // Propagate extension type if any
1515     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1516       Flags.setSExt();
1517     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1518       Flags.setZExt();
1519 
1520     for (unsigned i = 0; i < NumParts; ++i)
1521       Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0));
1522   }
1523 }
1524 
1525 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1526 /// function arguments in the caller parameter area.  This is the actual
1527 /// alignment, not its logarithm.
1528 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1529                                                    const DataLayout &DL) const {
1530   return DL.getABITypeAlignment(Ty);
1531 }
1532 
1533 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1534     LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1535     unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const {
1536   // Check if the specified alignment is sufficient based on the data layout.
1537   // TODO: While using the data layout works in practice, a better solution
1538   // would be to implement this check directly (make this a virtual function).
1539   // For example, the ABI alignment may change based on software platform while
1540   // this function should only be affected by hardware implementation.
1541   Type *Ty = VT.getTypeForEVT(Context);
1542   if (Alignment >= DL.getABITypeAlignment(Ty)) {
1543     // Assume that an access that meets the ABI-specified alignment is fast.
1544     if (Fast != nullptr)
1545       *Fast = true;
1546     return true;
1547   }
1548 
1549   // This is a misaligned access.
1550   return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast);
1551 }
1552 
1553 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1554     LLVMContext &Context, const DataLayout &DL, EVT VT,
1555     const MachineMemOperand &MMO, bool *Fast) const {
1556   return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
1557                                         MMO.getAlignment(), MMO.getFlags(),
1558                                         Fast);
1559 }
1560 
1561 bool TargetLoweringBase::allowsMemoryAccess(
1562     LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1563     unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const {
1564   return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
1565                                         Flags, Fast);
1566 }
1567 
1568 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1569                                             const DataLayout &DL, EVT VT,
1570                                             const MachineMemOperand &MMO,
1571                                             bool *Fast) const {
1572   return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(),
1573                             MMO.getAlignment(), MMO.getFlags(), Fast);
1574 }
1575 
1576 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1577   return BranchProbability(MinPercentageForPredictableBranch, 100);
1578 }
1579 
1580 //===----------------------------------------------------------------------===//
1581 //  TargetTransformInfo Helpers
1582 //===----------------------------------------------------------------------===//
1583 
1584 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1585   enum InstructionOpcodes {
1586 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1587 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1588 #include "llvm/IR/Instruction.def"
1589   };
1590   switch (static_cast<InstructionOpcodes>(Opcode)) {
1591   case Ret:            return 0;
1592   case Br:             return 0;
1593   case Switch:         return 0;
1594   case IndirectBr:     return 0;
1595   case Invoke:         return 0;
1596   case CallBr:         return 0;
1597   case Resume:         return 0;
1598   case Unreachable:    return 0;
1599   case CleanupRet:     return 0;
1600   case CatchRet:       return 0;
1601   case CatchPad:       return 0;
1602   case CatchSwitch:    return 0;
1603   case CleanupPad:     return 0;
1604   case FNeg:           return ISD::FNEG;
1605   case Add:            return ISD::ADD;
1606   case FAdd:           return ISD::FADD;
1607   case Sub:            return ISD::SUB;
1608   case FSub:           return ISD::FSUB;
1609   case Mul:            return ISD::MUL;
1610   case FMul:           return ISD::FMUL;
1611   case UDiv:           return ISD::UDIV;
1612   case SDiv:           return ISD::SDIV;
1613   case FDiv:           return ISD::FDIV;
1614   case URem:           return ISD::UREM;
1615   case SRem:           return ISD::SREM;
1616   case FRem:           return ISD::FREM;
1617   case Shl:            return ISD::SHL;
1618   case LShr:           return ISD::SRL;
1619   case AShr:           return ISD::SRA;
1620   case And:            return ISD::AND;
1621   case Or:             return ISD::OR;
1622   case Xor:            return ISD::XOR;
1623   case Alloca:         return 0;
1624   case Load:           return ISD::LOAD;
1625   case Store:          return ISD::STORE;
1626   case GetElementPtr:  return 0;
1627   case Fence:          return 0;
1628   case AtomicCmpXchg:  return 0;
1629   case AtomicRMW:      return 0;
1630   case Trunc:          return ISD::TRUNCATE;
1631   case ZExt:           return ISD::ZERO_EXTEND;
1632   case SExt:           return ISD::SIGN_EXTEND;
1633   case FPToUI:         return ISD::FP_TO_UINT;
1634   case FPToSI:         return ISD::FP_TO_SINT;
1635   case UIToFP:         return ISD::UINT_TO_FP;
1636   case SIToFP:         return ISD::SINT_TO_FP;
1637   case FPTrunc:        return ISD::FP_ROUND;
1638   case FPExt:          return ISD::FP_EXTEND;
1639   case PtrToInt:       return ISD::BITCAST;
1640   case IntToPtr:       return ISD::BITCAST;
1641   case BitCast:        return ISD::BITCAST;
1642   case AddrSpaceCast:  return ISD::ADDRSPACECAST;
1643   case ICmp:           return ISD::SETCC;
1644   case FCmp:           return ISD::SETCC;
1645   case PHI:            return 0;
1646   case Call:           return 0;
1647   case Select:         return ISD::SELECT;
1648   case UserOp1:        return 0;
1649   case UserOp2:        return 0;
1650   case VAArg:          return 0;
1651   case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1652   case InsertElement:  return ISD::INSERT_VECTOR_ELT;
1653   case ShuffleVector:  return ISD::VECTOR_SHUFFLE;
1654   case ExtractValue:   return ISD::MERGE_VALUES;
1655   case InsertValue:    return ISD::MERGE_VALUES;
1656   case LandingPad:     return 0;
1657   case Freeze:         return 0;
1658   }
1659 
1660   llvm_unreachable("Unknown instruction type encountered!");
1661 }
1662 
1663 std::pair<int, MVT>
1664 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1665                                             Type *Ty) const {
1666   LLVMContext &C = Ty->getContext();
1667   EVT MTy = getValueType(DL, Ty);
1668 
1669   int Cost = 1;
1670   // We keep legalizing the type until we find a legal kind. We assume that
1671   // the only operation that costs anything is the split. After splitting
1672   // we need to handle two types.
1673   while (true) {
1674     LegalizeKind LK = getTypeConversion(C, MTy);
1675 
1676     if (LK.first == TypeLegal)
1677       return std::make_pair(Cost, MTy.getSimpleVT());
1678 
1679     if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1680       Cost *= 2;
1681 
1682     // Do not loop with f128 type.
1683     if (MTy == LK.second)
1684       return std::make_pair(Cost, MTy.getSimpleVT());
1685 
1686     // Keep legalizing the type.
1687     MTy = LK.second;
1688   }
1689 }
1690 
1691 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1692                                                               bool UseTLS) const {
1693   // compiler-rt provides a variable with a magic name.  Targets that do not
1694   // link with compiler-rt may also provide such a variable.
1695   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1696   const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1697   auto UnsafeStackPtr =
1698       dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1699 
1700   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1701 
1702   if (!UnsafeStackPtr) {
1703     auto TLSModel = UseTLS ?
1704         GlobalValue::InitialExecTLSModel :
1705         GlobalValue::NotThreadLocal;
1706     // The global variable is not defined yet, define it ourselves.
1707     // We use the initial-exec TLS model because we do not support the
1708     // variable living anywhere other than in the main executable.
1709     UnsafeStackPtr = new GlobalVariable(
1710         *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1711         UnsafeStackPtrVar, nullptr, TLSModel);
1712   } else {
1713     // The variable exists, check its type and attributes.
1714     if (UnsafeStackPtr->getValueType() != StackPtrTy)
1715       report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1716     if (UseTLS != UnsafeStackPtr->isThreadLocal())
1717       report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1718                          (UseTLS ? "" : "not ") + "be thread-local");
1719   }
1720   return UnsafeStackPtr;
1721 }
1722 
1723 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1724   if (!TM.getTargetTriple().isAndroid())
1725     return getDefaultSafeStackPointerLocation(IRB, true);
1726 
1727   // Android provides a libc function to retrieve the address of the current
1728   // thread's unsafe stack pointer.
1729   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1730   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1731   FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address",
1732                                              StackPtrTy->getPointerTo(0));
1733   return IRB.CreateCall(Fn);
1734 }
1735 
1736 //===----------------------------------------------------------------------===//
1737 //  Loop Strength Reduction hooks
1738 //===----------------------------------------------------------------------===//
1739 
1740 /// isLegalAddressingMode - Return true if the addressing mode represented
1741 /// by AM is legal for this target, for a load/store of the specified type.
1742 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1743                                                const AddrMode &AM, Type *Ty,
1744                                                unsigned AS, Instruction *I) const {
1745   // The default implementation of this implements a conservative RISCy, r+r and
1746   // r+i addr mode.
1747 
1748   // Allows a sign-extended 16-bit immediate field.
1749   if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1750     return false;
1751 
1752   // No global is ever allowed as a base.
1753   if (AM.BaseGV)
1754     return false;
1755 
1756   // Only support r+r,
1757   switch (AM.Scale) {
1758   case 0:  // "r+i" or just "i", depending on HasBaseReg.
1759     break;
1760   case 1:
1761     if (AM.HasBaseReg && AM.BaseOffs)  // "r+r+i" is not allowed.
1762       return false;
1763     // Otherwise we have r+r or r+i.
1764     break;
1765   case 2:
1766     if (AM.HasBaseReg || AM.BaseOffs)  // 2*r+r  or  2*r+i is not allowed.
1767       return false;
1768     // Allow 2*r as r+r.
1769     break;
1770   default: // Don't allow n * r
1771     return false;
1772   }
1773 
1774   return true;
1775 }
1776 
1777 //===----------------------------------------------------------------------===//
1778 //  Stack Protector
1779 //===----------------------------------------------------------------------===//
1780 
1781 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1782 // so that SelectionDAG handle SSP.
1783 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1784   if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1785     Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1786     PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1787     return M.getOrInsertGlobal("__guard_local", PtrTy);
1788   }
1789   return nullptr;
1790 }
1791 
1792 // Currently only support "standard" __stack_chk_guard.
1793 // TODO: add LOAD_STACK_GUARD support.
1794 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1795   if (!M.getNamedValue("__stack_chk_guard"))
1796     new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1797                        GlobalVariable::ExternalLinkage,
1798                        nullptr, "__stack_chk_guard");
1799 }
1800 
1801 // Currently only support "standard" __stack_chk_guard.
1802 // TODO: add LOAD_STACK_GUARD support.
1803 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1804   return M.getNamedValue("__stack_chk_guard");
1805 }
1806 
1807 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1808   return nullptr;
1809 }
1810 
1811 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1812   return MinimumJumpTableEntries;
1813 }
1814 
1815 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1816   MinimumJumpTableEntries = Val;
1817 }
1818 
1819 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1820   return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1821 }
1822 
1823 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1824   return MaximumJumpTableSize;
1825 }
1826 
1827 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1828   MaximumJumpTableSize = Val;
1829 }
1830 
1831 //===----------------------------------------------------------------------===//
1832 //  Reciprocal Estimates
1833 //===----------------------------------------------------------------------===//
1834 
1835 /// Get the reciprocal estimate attribute string for a function that will
1836 /// override the target defaults.
1837 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1838   const Function &F = MF.getFunction();
1839   return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1840 }
1841 
1842 /// Construct a string for the given reciprocal operation of the given type.
1843 /// This string should match the corresponding option to the front-end's
1844 /// "-mrecip" flag assuming those strings have been passed through in an
1845 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1846 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1847   std::string Name = VT.isVector() ? "vec-" : "";
1848 
1849   Name += IsSqrt ? "sqrt" : "div";
1850 
1851   // TODO: Handle "half" or other float types?
1852   if (VT.getScalarType() == MVT::f64) {
1853     Name += "d";
1854   } else {
1855     assert(VT.getScalarType() == MVT::f32 &&
1856            "Unexpected FP type for reciprocal estimate");
1857     Name += "f";
1858   }
1859 
1860   return Name;
1861 }
1862 
1863 /// Return the character position and value (a single numeric character) of a
1864 /// customized refinement operation in the input string if it exists. Return
1865 /// false if there is no customized refinement step count.
1866 static bool parseRefinementStep(StringRef In, size_t &Position,
1867                                 uint8_t &Value) {
1868   const char RefStepToken = ':';
1869   Position = In.find(RefStepToken);
1870   if (Position == StringRef::npos)
1871     return false;
1872 
1873   StringRef RefStepString = In.substr(Position + 1);
1874   // Allow exactly one numeric character for the additional refinement
1875   // step parameter.
1876   if (RefStepString.size() == 1) {
1877     char RefStepChar = RefStepString[0];
1878     if (RefStepChar >= '0' && RefStepChar <= '9') {
1879       Value = RefStepChar - '0';
1880       return true;
1881     }
1882   }
1883   report_fatal_error("Invalid refinement step for -recip.");
1884 }
1885 
1886 /// For the input attribute string, return one of the ReciprocalEstimate enum
1887 /// status values (enabled, disabled, or not specified) for this operation on
1888 /// the specified data type.
1889 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1890   if (Override.empty())
1891     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1892 
1893   SmallVector<StringRef, 4> OverrideVector;
1894   Override.split(OverrideVector, ',');
1895   unsigned NumArgs = OverrideVector.size();
1896 
1897   // Check if "all", "none", or "default" was specified.
1898   if (NumArgs == 1) {
1899     // Look for an optional setting of the number of refinement steps needed
1900     // for this type of reciprocal operation.
1901     size_t RefPos;
1902     uint8_t RefSteps;
1903     if (parseRefinementStep(Override, RefPos, RefSteps)) {
1904       // Split the string for further processing.
1905       Override = Override.substr(0, RefPos);
1906     }
1907 
1908     // All reciprocal types are enabled.
1909     if (Override == "all")
1910       return TargetLoweringBase::ReciprocalEstimate::Enabled;
1911 
1912     // All reciprocal types are disabled.
1913     if (Override == "none")
1914       return TargetLoweringBase::ReciprocalEstimate::Disabled;
1915 
1916     // Target defaults for enablement are used.
1917     if (Override == "default")
1918       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1919   }
1920 
1921   // The attribute string may omit the size suffix ('f'/'d').
1922   std::string VTName = getReciprocalOpName(IsSqrt, VT);
1923   std::string VTNameNoSize = VTName;
1924   VTNameNoSize.pop_back();
1925   static const char DisabledPrefix = '!';
1926 
1927   for (StringRef RecipType : OverrideVector) {
1928     size_t RefPos;
1929     uint8_t RefSteps;
1930     if (parseRefinementStep(RecipType, RefPos, RefSteps))
1931       RecipType = RecipType.substr(0, RefPos);
1932 
1933     // Ignore the disablement token for string matching.
1934     bool IsDisabled = RecipType[0] == DisabledPrefix;
1935     if (IsDisabled)
1936       RecipType = RecipType.substr(1);
1937 
1938     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1939       return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
1940                         : TargetLoweringBase::ReciprocalEstimate::Enabled;
1941   }
1942 
1943   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1944 }
1945 
1946 /// For the input attribute string, return the customized refinement step count
1947 /// for this operation on the specified data type. If the step count does not
1948 /// exist, return the ReciprocalEstimate enum value for unspecified.
1949 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
1950   if (Override.empty())
1951     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1952 
1953   SmallVector<StringRef, 4> OverrideVector;
1954   Override.split(OverrideVector, ',');
1955   unsigned NumArgs = OverrideVector.size();
1956 
1957   // Check if "all", "default", or "none" was specified.
1958   if (NumArgs == 1) {
1959     // Look for an optional setting of the number of refinement steps needed
1960     // for this type of reciprocal operation.
1961     size_t RefPos;
1962     uint8_t RefSteps;
1963     if (!parseRefinementStep(Override, RefPos, RefSteps))
1964       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1965 
1966     // Split the string for further processing.
1967     Override = Override.substr(0, RefPos);
1968     assert(Override != "none" &&
1969            "Disabled reciprocals, but specifed refinement steps?");
1970 
1971     // If this is a general override, return the specified number of steps.
1972     if (Override == "all" || Override == "default")
1973       return RefSteps;
1974   }
1975 
1976   // The attribute string may omit the size suffix ('f'/'d').
1977   std::string VTName = getReciprocalOpName(IsSqrt, VT);
1978   std::string VTNameNoSize = VTName;
1979   VTNameNoSize.pop_back();
1980 
1981   for (StringRef RecipType : OverrideVector) {
1982     size_t RefPos;
1983     uint8_t RefSteps;
1984     if (!parseRefinementStep(RecipType, RefPos, RefSteps))
1985       continue;
1986 
1987     RecipType = RecipType.substr(0, RefPos);
1988     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1989       return RefSteps;
1990   }
1991 
1992   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1993 }
1994 
1995 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
1996                                                     MachineFunction &MF) const {
1997   return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
1998 }
1999 
2000 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
2001                                                    MachineFunction &MF) const {
2002   return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
2003 }
2004 
2005 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
2006                                                MachineFunction &MF) const {
2007   return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
2008 }
2009 
2010 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
2011                                               MachineFunction &MF) const {
2012   return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
2013 }
2014 
2015 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
2016   MF.getRegInfo().freezeReservedRegs(MF);
2017 }
2018 
2019 MachineMemOperand::Flags
2020 TargetLoweringBase::getLoadMemOperandFlags(const LoadInst &LI,
2021                                            const DataLayout &DL) const {
2022   MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad;
2023   if (LI.isVolatile())
2024     Flags |= MachineMemOperand::MOVolatile;
2025 
2026   if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2027     Flags |= MachineMemOperand::MONonTemporal;
2028 
2029   if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2030     Flags |= MachineMemOperand::MOInvariant;
2031 
2032   if (isDereferenceablePointer(LI.getPointerOperand(), LI.getType(), DL))
2033     Flags |= MachineMemOperand::MODereferenceable;
2034 
2035   Flags |= getTargetMMOFlags(LI);
2036   return Flags;
2037 }
2038 
2039 MachineMemOperand::Flags
2040 TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI,
2041                                             const DataLayout &DL) const {
2042   MachineMemOperand::Flags Flags = MachineMemOperand::MOStore;
2043 
2044   if (SI.isVolatile())
2045     Flags |= MachineMemOperand::MOVolatile;
2046 
2047   if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2048     Flags |= MachineMemOperand::MONonTemporal;
2049 
2050   // FIXME: Not preserving dereferenceable
2051   Flags |= getTargetMMOFlags(SI);
2052   return Flags;
2053 }
2054 
2055 MachineMemOperand::Flags
2056 TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI,
2057                                              const DataLayout &DL) const {
2058   auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
2059 
2060   if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2061     if (RMW->isVolatile())
2062       Flags |= MachineMemOperand::MOVolatile;
2063   } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2064     if (CmpX->isVolatile())
2065       Flags |= MachineMemOperand::MOVolatile;
2066   } else
2067     llvm_unreachable("not an atomic instruction");
2068 
2069   // FIXME: Not preserving dereferenceable
2070   Flags |= getTargetMMOFlags(AI);
2071   return Flags;
2072 }
2073