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/Analysis/TargetTransformInfo.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/ISDOpcodes.h"
24 #include "llvm/CodeGen/MachineBasicBlock.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineOperand.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/RuntimeLibcalls.h"
33 #include "llvm/CodeGen/StackMaps.h"
34 #include "llvm/CodeGen/TargetLowering.h"
35 #include "llvm/CodeGen/TargetOpcodes.h"
36 #include "llvm/CodeGen/TargetRegisterInfo.h"
37 #include "llvm/CodeGen/ValueTypes.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalValue.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/IRBuilder.h"
46 #include "llvm/IR/Module.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/BranchProbability.h"
49 #include "llvm/Support/Casting.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MachineValueType.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Target/TargetMachine.h"
56 #include "llvm/Transforms/Utils/SizeOpts.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <cstddef>
60 #include <cstdint>
61 #include <cstring>
62 #include <iterator>
63 #include <string>
64 #include <tuple>
65 #include <utility>
66 
67 using namespace llvm;
68 
69 static cl::opt<bool> JumpIsExpensiveOverride(
70     "jump-is-expensive", cl::init(false),
71     cl::desc("Do not create extra branches to split comparison logic."),
72     cl::Hidden);
73 
74 static cl::opt<unsigned> MinimumJumpTableEntries
75   ("min-jump-table-entries", cl::init(4), cl::Hidden,
76    cl::desc("Set minimum number of entries to use a jump table."));
77 
78 static cl::opt<unsigned> MaximumJumpTableSize
79   ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden,
80    cl::desc("Set maximum size of jump tables."));
81 
82 /// Minimum jump table density for normal functions.
83 static cl::opt<unsigned>
84     JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden,
85                      cl::desc("Minimum density for building a jump table in "
86                               "a normal function"));
87 
88 /// Minimum jump table density for -Os or -Oz functions.
89 static cl::opt<unsigned> OptsizeJumpTableDensity(
90     "optsize-jump-table-density", cl::init(40), cl::Hidden,
91     cl::desc("Minimum density for building a jump table in "
92              "an optsize function"));
93 
94 // FIXME: This option is only to test if the strict fp operation processed
95 // correctly by preventing mutating strict fp operation to normal fp operation
96 // during development. When the backend supports strict float operation, this
97 // option will be meaningless.
98 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
99        cl::desc("Don't mutate strict-float node to a legalize node"),
100        cl::init(false), cl::Hidden);
101 
102 static bool darwinHasSinCos(const Triple &TT) {
103   assert(TT.isOSDarwin() && "should be called with darwin triple");
104   // Don't bother with 32 bit x86.
105   if (TT.getArch() == Triple::x86)
106     return false;
107   // Macos < 10.9 has no sincos_stret.
108   if (TT.isMacOSX())
109     return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
110   // iOS < 7.0 has no sincos_stret.
111   if (TT.isiOS())
112     return !TT.isOSVersionLT(7, 0);
113   // Any other darwin such as WatchOS/TvOS is new enough.
114   return true;
115 }
116 
117 // Although this default value is arbitrary, it is not random. It is assumed
118 // that a condition that evaluates the same way by a higher percentage than this
119 // is best represented as control flow. Therefore, the default value N should be
120 // set such that the win from N% correct executions is greater than the loss
121 // from (100 - N)% mispredicted executions for the majority of intended targets.
122 static cl::opt<int> MinPercentageForPredictableBranch(
123     "min-predictable-branch", cl::init(99),
124     cl::desc("Minimum percentage (0-100) that a condition must be either true "
125              "or false to assume that the condition is predictable"),
126     cl::Hidden);
127 
128 void TargetLoweringBase::InitLibcalls(const Triple &TT) {
129 #define HANDLE_LIBCALL(code, name) \
130   setLibcallName(RTLIB::code, name);
131 #include "llvm/IR/RuntimeLibcalls.def"
132 #undef HANDLE_LIBCALL
133   // Initialize calling conventions to their default.
134   for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC)
135     setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C);
136 
137   // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf".
138   if (TT.getArch() == Triple::ppc || TT.isPPC64()) {
139     setLibcallName(RTLIB::ADD_F128, "__addkf3");
140     setLibcallName(RTLIB::SUB_F128, "__subkf3");
141     setLibcallName(RTLIB::MUL_F128, "__mulkf3");
142     setLibcallName(RTLIB::DIV_F128, "__divkf3");
143     setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2");
144     setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2");
145     setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2");
146     setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2");
147     setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi");
148     setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi");
149     setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi");
150     setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi");
151     setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf");
152     setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf");
153     setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf");
154     setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf");
155     setLibcallName(RTLIB::OEQ_F128, "__eqkf2");
156     setLibcallName(RTLIB::UNE_F128, "__nekf2");
157     setLibcallName(RTLIB::OGE_F128, "__gekf2");
158     setLibcallName(RTLIB::OLT_F128, "__ltkf2");
159     setLibcallName(RTLIB::OLE_F128, "__lekf2");
160     setLibcallName(RTLIB::OGT_F128, "__gtkf2");
161     setLibcallName(RTLIB::UO_F128, "__unordkf2");
162   }
163 
164   // A few names are different on particular architectures or environments.
165   if (TT.isOSDarwin()) {
166     // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
167     // of the gnueabi-style __gnu_*_ieee.
168     // FIXME: What about other targets?
169     setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
170     setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
171 
172     // Some darwins have an optimized __bzero/bzero function.
173     switch (TT.getArch()) {
174     case Triple::x86:
175     case Triple::x86_64:
176       if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6))
177         setLibcallName(RTLIB::BZERO, "__bzero");
178       break;
179     case Triple::aarch64:
180     case Triple::aarch64_32:
181       setLibcallName(RTLIB::BZERO, "bzero");
182       break;
183     default:
184       break;
185     }
186 
187     if (darwinHasSinCos(TT)) {
188       setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret");
189       setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret");
190       if (TT.isWatchABI()) {
191         setLibcallCallingConv(RTLIB::SINCOS_STRET_F32,
192                               CallingConv::ARM_AAPCS_VFP);
193         setLibcallCallingConv(RTLIB::SINCOS_STRET_F64,
194                               CallingConv::ARM_AAPCS_VFP);
195       }
196     }
197   } else {
198     setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee");
199     setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee");
200   }
201 
202   if (TT.isGNUEnvironment() || TT.isOSFuchsia() ||
203       (TT.isAndroid() && !TT.isAndroidVersionLT(9))) {
204     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
205     setLibcallName(RTLIB::SINCOS_F64, "sincos");
206     setLibcallName(RTLIB::SINCOS_F80, "sincosl");
207     setLibcallName(RTLIB::SINCOS_F128, "sincosl");
208     setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
209   }
210 
211   if (TT.isPS4CPU()) {
212     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
213     setLibcallName(RTLIB::SINCOS_F64, "sincos");
214   }
215 
216   if (TT.isOSOpenBSD()) {
217     setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
218   }
219 }
220 
221 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
222 /// UNKNOWN_LIBCALL if there is none.
223 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
224   if (OpVT == MVT::f16) {
225     if (RetVT == MVT::f32)
226       return FPEXT_F16_F32;
227   } else if (OpVT == MVT::f32) {
228     if (RetVT == MVT::f64)
229       return FPEXT_F32_F64;
230     if (RetVT == MVT::f128)
231       return FPEXT_F32_F128;
232     if (RetVT == MVT::ppcf128)
233       return FPEXT_F32_PPCF128;
234   } else if (OpVT == MVT::f64) {
235     if (RetVT == MVT::f128)
236       return FPEXT_F64_F128;
237     else if (RetVT == MVT::ppcf128)
238       return FPEXT_F64_PPCF128;
239   } else if (OpVT == MVT::f80) {
240     if (RetVT == MVT::f128)
241       return FPEXT_F80_F128;
242   }
243 
244   return UNKNOWN_LIBCALL;
245 }
246 
247 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
248 /// UNKNOWN_LIBCALL if there is none.
249 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
250   if (RetVT == MVT::f16) {
251     if (OpVT == MVT::f32)
252       return FPROUND_F32_F16;
253     if (OpVT == MVT::f64)
254       return FPROUND_F64_F16;
255     if (OpVT == MVT::f80)
256       return FPROUND_F80_F16;
257     if (OpVT == MVT::f128)
258       return FPROUND_F128_F16;
259     if (OpVT == MVT::ppcf128)
260       return FPROUND_PPCF128_F16;
261   } else if (RetVT == MVT::f32) {
262     if (OpVT == MVT::f64)
263       return FPROUND_F64_F32;
264     if (OpVT == MVT::f80)
265       return FPROUND_F80_F32;
266     if (OpVT == MVT::f128)
267       return FPROUND_F128_F32;
268     if (OpVT == MVT::ppcf128)
269       return FPROUND_PPCF128_F32;
270   } else if (RetVT == MVT::f64) {
271     if (OpVT == MVT::f80)
272       return FPROUND_F80_F64;
273     if (OpVT == MVT::f128)
274       return FPROUND_F128_F64;
275     if (OpVT == MVT::ppcf128)
276       return FPROUND_PPCF128_F64;
277   } else if (RetVT == MVT::f80) {
278     if (OpVT == MVT::f128)
279       return FPROUND_F128_F80;
280   }
281 
282   return UNKNOWN_LIBCALL;
283 }
284 
285 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
286 /// UNKNOWN_LIBCALL if there is none.
287 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
288   if (OpVT == MVT::f32) {
289     if (RetVT == MVT::i32)
290       return FPTOSINT_F32_I32;
291     if (RetVT == MVT::i64)
292       return FPTOSINT_F32_I64;
293     if (RetVT == MVT::i128)
294       return FPTOSINT_F32_I128;
295   } else if (OpVT == MVT::f64) {
296     if (RetVT == MVT::i32)
297       return FPTOSINT_F64_I32;
298     if (RetVT == MVT::i64)
299       return FPTOSINT_F64_I64;
300     if (RetVT == MVT::i128)
301       return FPTOSINT_F64_I128;
302   } else if (OpVT == MVT::f80) {
303     if (RetVT == MVT::i32)
304       return FPTOSINT_F80_I32;
305     if (RetVT == MVT::i64)
306       return FPTOSINT_F80_I64;
307     if (RetVT == MVT::i128)
308       return FPTOSINT_F80_I128;
309   } else if (OpVT == MVT::f128) {
310     if (RetVT == MVT::i32)
311       return FPTOSINT_F128_I32;
312     if (RetVT == MVT::i64)
313       return FPTOSINT_F128_I64;
314     if (RetVT == MVT::i128)
315       return FPTOSINT_F128_I128;
316   } else if (OpVT == MVT::ppcf128) {
317     if (RetVT == MVT::i32)
318       return FPTOSINT_PPCF128_I32;
319     if (RetVT == MVT::i64)
320       return FPTOSINT_PPCF128_I64;
321     if (RetVT == MVT::i128)
322       return FPTOSINT_PPCF128_I128;
323   }
324   return UNKNOWN_LIBCALL;
325 }
326 
327 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
328 /// UNKNOWN_LIBCALL if there is none.
329 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
330   if (OpVT == MVT::f32) {
331     if (RetVT == MVT::i32)
332       return FPTOUINT_F32_I32;
333     if (RetVT == MVT::i64)
334       return FPTOUINT_F32_I64;
335     if (RetVT == MVT::i128)
336       return FPTOUINT_F32_I128;
337   } else if (OpVT == MVT::f64) {
338     if (RetVT == MVT::i32)
339       return FPTOUINT_F64_I32;
340     if (RetVT == MVT::i64)
341       return FPTOUINT_F64_I64;
342     if (RetVT == MVT::i128)
343       return FPTOUINT_F64_I128;
344   } else if (OpVT == MVT::f80) {
345     if (RetVT == MVT::i32)
346       return FPTOUINT_F80_I32;
347     if (RetVT == MVT::i64)
348       return FPTOUINT_F80_I64;
349     if (RetVT == MVT::i128)
350       return FPTOUINT_F80_I128;
351   } else if (OpVT == MVT::f128) {
352     if (RetVT == MVT::i32)
353       return FPTOUINT_F128_I32;
354     if (RetVT == MVT::i64)
355       return FPTOUINT_F128_I64;
356     if (RetVT == MVT::i128)
357       return FPTOUINT_F128_I128;
358   } else if (OpVT == MVT::ppcf128) {
359     if (RetVT == MVT::i32)
360       return FPTOUINT_PPCF128_I32;
361     if (RetVT == MVT::i64)
362       return FPTOUINT_PPCF128_I64;
363     if (RetVT == MVT::i128)
364       return FPTOUINT_PPCF128_I128;
365   }
366   return UNKNOWN_LIBCALL;
367 }
368 
369 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
370 /// UNKNOWN_LIBCALL if there is none.
371 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
372   if (OpVT == MVT::i32) {
373     if (RetVT == MVT::f32)
374       return SINTTOFP_I32_F32;
375     if (RetVT == MVT::f64)
376       return SINTTOFP_I32_F64;
377     if (RetVT == MVT::f80)
378       return SINTTOFP_I32_F80;
379     if (RetVT == MVT::f128)
380       return SINTTOFP_I32_F128;
381     if (RetVT == MVT::ppcf128)
382       return SINTTOFP_I32_PPCF128;
383   } else if (OpVT == MVT::i64) {
384     if (RetVT == MVT::f32)
385       return SINTTOFP_I64_F32;
386     if (RetVT == MVT::f64)
387       return SINTTOFP_I64_F64;
388     if (RetVT == MVT::f80)
389       return SINTTOFP_I64_F80;
390     if (RetVT == MVT::f128)
391       return SINTTOFP_I64_F128;
392     if (RetVT == MVT::ppcf128)
393       return SINTTOFP_I64_PPCF128;
394   } else if (OpVT == MVT::i128) {
395     if (RetVT == MVT::f32)
396       return SINTTOFP_I128_F32;
397     if (RetVT == MVT::f64)
398       return SINTTOFP_I128_F64;
399     if (RetVT == MVT::f80)
400       return SINTTOFP_I128_F80;
401     if (RetVT == MVT::f128)
402       return SINTTOFP_I128_F128;
403     if (RetVT == MVT::ppcf128)
404       return SINTTOFP_I128_PPCF128;
405   }
406   return UNKNOWN_LIBCALL;
407 }
408 
409 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
410 /// UNKNOWN_LIBCALL if there is none.
411 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
412   if (OpVT == MVT::i32) {
413     if (RetVT == MVT::f32)
414       return UINTTOFP_I32_F32;
415     if (RetVT == MVT::f64)
416       return UINTTOFP_I32_F64;
417     if (RetVT == MVT::f80)
418       return UINTTOFP_I32_F80;
419     if (RetVT == MVT::f128)
420       return UINTTOFP_I32_F128;
421     if (RetVT == MVT::ppcf128)
422       return UINTTOFP_I32_PPCF128;
423   } else if (OpVT == MVT::i64) {
424     if (RetVT == MVT::f32)
425       return UINTTOFP_I64_F32;
426     if (RetVT == MVT::f64)
427       return UINTTOFP_I64_F64;
428     if (RetVT == MVT::f80)
429       return UINTTOFP_I64_F80;
430     if (RetVT == MVT::f128)
431       return UINTTOFP_I64_F128;
432     if (RetVT == MVT::ppcf128)
433       return UINTTOFP_I64_PPCF128;
434   } else if (OpVT == MVT::i128) {
435     if (RetVT == MVT::f32)
436       return UINTTOFP_I128_F32;
437     if (RetVT == MVT::f64)
438       return UINTTOFP_I128_F64;
439     if (RetVT == MVT::f80)
440       return UINTTOFP_I128_F80;
441     if (RetVT == MVT::f128)
442       return UINTTOFP_I128_F128;
443     if (RetVT == MVT::ppcf128)
444       return UINTTOFP_I128_PPCF128;
445   }
446   return UNKNOWN_LIBCALL;
447 }
448 
449 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
450 #define OP_TO_LIBCALL(Name, Enum)                                              \
451   case Name:                                                                   \
452     switch (VT.SimpleTy) {                                                     \
453     default:                                                                   \
454       return UNKNOWN_LIBCALL;                                                  \
455     case MVT::i8:                                                              \
456       return Enum##_1;                                                         \
457     case MVT::i16:                                                             \
458       return Enum##_2;                                                         \
459     case MVT::i32:                                                             \
460       return Enum##_4;                                                         \
461     case MVT::i64:                                                             \
462       return Enum##_8;                                                         \
463     case MVT::i128:                                                            \
464       return Enum##_16;                                                        \
465     }
466 
467   switch (Opc) {
468     OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
469     OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
470     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
471     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
472     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
473     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
474     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
475     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
476     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
477     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
478     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
479     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
480   }
481 
482 #undef OP_TO_LIBCALL
483 
484   return UNKNOWN_LIBCALL;
485 }
486 
487 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
488   switch (ElementSize) {
489   case 1:
490     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
491   case 2:
492     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
493   case 4:
494     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
495   case 8:
496     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
497   case 16:
498     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
499   default:
500     return UNKNOWN_LIBCALL;
501   }
502 }
503 
504 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
505   switch (ElementSize) {
506   case 1:
507     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
508   case 2:
509     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
510   case 4:
511     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
512   case 8:
513     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
514   case 16:
515     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
516   default:
517     return UNKNOWN_LIBCALL;
518   }
519 }
520 
521 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
522   switch (ElementSize) {
523   case 1:
524     return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
525   case 2:
526     return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
527   case 4:
528     return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
529   case 8:
530     return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
531   case 16:
532     return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
533   default:
534     return UNKNOWN_LIBCALL;
535   }
536 }
537 
538 /// InitCmpLibcallCCs - Set default comparison libcall CC.
539 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
540   memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
541   CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
542   CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
543   CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
544   CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
545   CCs[RTLIB::UNE_F32] = ISD::SETNE;
546   CCs[RTLIB::UNE_F64] = ISD::SETNE;
547   CCs[RTLIB::UNE_F128] = ISD::SETNE;
548   CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
549   CCs[RTLIB::OGE_F32] = ISD::SETGE;
550   CCs[RTLIB::OGE_F64] = ISD::SETGE;
551   CCs[RTLIB::OGE_F128] = ISD::SETGE;
552   CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
553   CCs[RTLIB::OLT_F32] = ISD::SETLT;
554   CCs[RTLIB::OLT_F64] = ISD::SETLT;
555   CCs[RTLIB::OLT_F128] = ISD::SETLT;
556   CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
557   CCs[RTLIB::OLE_F32] = ISD::SETLE;
558   CCs[RTLIB::OLE_F64] = ISD::SETLE;
559   CCs[RTLIB::OLE_F128] = ISD::SETLE;
560   CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
561   CCs[RTLIB::OGT_F32] = ISD::SETGT;
562   CCs[RTLIB::OGT_F64] = ISD::SETGT;
563   CCs[RTLIB::OGT_F128] = ISD::SETGT;
564   CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
565   CCs[RTLIB::UO_F32] = ISD::SETNE;
566   CCs[RTLIB::UO_F64] = ISD::SETNE;
567   CCs[RTLIB::UO_F128] = ISD::SETNE;
568   CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
569 }
570 
571 /// NOTE: The TargetMachine owns TLOF.
572 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
573   initActions();
574 
575   // Perform these initializations only once.
576   MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
577       MaxLoadsPerMemcmp = 8;
578   MaxGluedStoresPerMemcpy = 0;
579   MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
580       MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
581   HasMultipleConditionRegisters = false;
582   HasExtractBitsInsn = false;
583   JumpIsExpensive = JumpIsExpensiveOverride;
584   PredictableSelectIsExpensive = false;
585   EnableExtLdPromotion = false;
586   StackPointerRegisterToSaveRestore = 0;
587   BooleanContents = UndefinedBooleanContent;
588   BooleanFloatContents = UndefinedBooleanContent;
589   BooleanVectorContents = UndefinedBooleanContent;
590   SchedPreferenceInfo = Sched::ILP;
591   GatherAllAliasesMaxDepth = 18;
592   IsStrictFPEnabled = DisableStrictNodeMutation;
593   // TODO: the default will be switched to 0 in the next commit, along
594   // with the Target-specific changes necessary.
595   MaxAtomicSizeInBitsSupported = 1024;
596 
597   MinCmpXchgSizeInBits = 0;
598   SupportsUnalignedAtomics = false;
599 
600   std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
601 
602   InitLibcalls(TM.getTargetTriple());
603   InitCmpLibcallCCs(CmpLibcallCCs);
604 }
605 
606 void TargetLoweringBase::initActions() {
607   // All operations default to being supported.
608   memset(OpActions, 0, sizeof(OpActions));
609   memset(LoadExtActions, 0, sizeof(LoadExtActions));
610   memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
611   memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
612   memset(CondCodeActions, 0, sizeof(CondCodeActions));
613   std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
614   std::fill(std::begin(TargetDAGCombineArray),
615             std::end(TargetDAGCombineArray), 0);
616 
617   for (MVT VT : MVT::fp_valuetypes()) {
618     MVT IntVT = MVT::getIntegerVT(VT.getFixedSizeInBits());
619     if (IntVT.isValid()) {
620       setOperationAction(ISD::ATOMIC_SWAP, VT, Promote);
621       AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT);
622     }
623   }
624 
625   // Set default actions for various operations.
626   for (MVT VT : MVT::all_valuetypes()) {
627     // Default all indexed load / store to expand.
628     for (unsigned IM = (unsigned)ISD::PRE_INC;
629          IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
630       setIndexedLoadAction(IM, VT, Expand);
631       setIndexedStoreAction(IM, VT, Expand);
632       setIndexedMaskedLoadAction(IM, VT, Expand);
633       setIndexedMaskedStoreAction(IM, VT, Expand);
634     }
635 
636     // Most backends expect to see the node which just returns the value loaded.
637     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
638 
639     // These operations default to expand.
640     setOperationAction(ISD::FGETSIGN, VT, Expand);
641     setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
642     setOperationAction(ISD::FMINNUM, VT, Expand);
643     setOperationAction(ISD::FMAXNUM, VT, Expand);
644     setOperationAction(ISD::FMINNUM_IEEE, VT, Expand);
645     setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand);
646     setOperationAction(ISD::FMINIMUM, VT, Expand);
647     setOperationAction(ISD::FMAXIMUM, VT, Expand);
648     setOperationAction(ISD::FMAD, VT, Expand);
649     setOperationAction(ISD::SMIN, VT, Expand);
650     setOperationAction(ISD::SMAX, VT, Expand);
651     setOperationAction(ISD::UMIN, VT, Expand);
652     setOperationAction(ISD::UMAX, VT, Expand);
653     setOperationAction(ISD::ABS, VT, Expand);
654     setOperationAction(ISD::FSHL, VT, Expand);
655     setOperationAction(ISD::FSHR, VT, Expand);
656     setOperationAction(ISD::SADDSAT, VT, Expand);
657     setOperationAction(ISD::UADDSAT, VT, Expand);
658     setOperationAction(ISD::SSUBSAT, VT, Expand);
659     setOperationAction(ISD::USUBSAT, VT, Expand);
660     setOperationAction(ISD::SSHLSAT, VT, Expand);
661     setOperationAction(ISD::USHLSAT, VT, Expand);
662     setOperationAction(ISD::SMULFIX, VT, Expand);
663     setOperationAction(ISD::SMULFIXSAT, VT, Expand);
664     setOperationAction(ISD::UMULFIX, VT, Expand);
665     setOperationAction(ISD::UMULFIXSAT, VT, Expand);
666     setOperationAction(ISD::SDIVFIX, VT, Expand);
667     setOperationAction(ISD::SDIVFIXSAT, VT, Expand);
668     setOperationAction(ISD::UDIVFIX, VT, Expand);
669     setOperationAction(ISD::UDIVFIXSAT, VT, Expand);
670 
671     // Overflow operations default to expand
672     setOperationAction(ISD::SADDO, VT, Expand);
673     setOperationAction(ISD::SSUBO, VT, Expand);
674     setOperationAction(ISD::UADDO, VT, Expand);
675     setOperationAction(ISD::USUBO, VT, Expand);
676     setOperationAction(ISD::SMULO, VT, Expand);
677     setOperationAction(ISD::UMULO, VT, Expand);
678 
679     // ADDCARRY operations default to expand
680     setOperationAction(ISD::ADDCARRY, VT, Expand);
681     setOperationAction(ISD::SUBCARRY, VT, Expand);
682     setOperationAction(ISD::SETCCCARRY, VT, Expand);
683     setOperationAction(ISD::SADDO_CARRY, VT, Expand);
684     setOperationAction(ISD::SSUBO_CARRY, VT, Expand);
685 
686     // ADDC/ADDE/SUBC/SUBE default to expand.
687     setOperationAction(ISD::ADDC, VT, Expand);
688     setOperationAction(ISD::ADDE, VT, Expand);
689     setOperationAction(ISD::SUBC, VT, Expand);
690     setOperationAction(ISD::SUBE, VT, Expand);
691 
692     // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
693     setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
694     setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
695 
696     setOperationAction(ISD::BITREVERSE, VT, Expand);
697     setOperationAction(ISD::PARITY, VT, Expand);
698 
699     // These library functions default to expand.
700     setOperationAction(ISD::FROUND, VT, Expand);
701     setOperationAction(ISD::FROUNDEVEN, VT, Expand);
702     setOperationAction(ISD::FPOWI, VT, Expand);
703 
704     // These operations default to expand for vector types.
705     if (VT.isVector()) {
706       setOperationAction(ISD::FCOPYSIGN, VT, Expand);
707       setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
708       setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
709       setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
710       setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
711       setOperationAction(ISD::SPLAT_VECTOR, VT, Expand);
712     }
713 
714     // Constrained floating-point operations default to expand.
715 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN)               \
716     setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
717 #include "llvm/IR/ConstrainedOps.def"
718 
719     // For most targets @llvm.get.dynamic.area.offset just returns 0.
720     setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
721 
722     // Vector reduction default to expand.
723     setOperationAction(ISD::VECREDUCE_FADD, VT, Expand);
724     setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand);
725     setOperationAction(ISD::VECREDUCE_ADD, VT, Expand);
726     setOperationAction(ISD::VECREDUCE_MUL, VT, Expand);
727     setOperationAction(ISD::VECREDUCE_AND, VT, Expand);
728     setOperationAction(ISD::VECREDUCE_OR, VT, Expand);
729     setOperationAction(ISD::VECREDUCE_XOR, VT, Expand);
730     setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand);
731     setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand);
732     setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand);
733     setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand);
734     setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand);
735     setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand);
736   }
737 
738   // Most targets ignore the @llvm.prefetch intrinsic.
739   setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
740 
741   // Most targets also ignore the @llvm.readcyclecounter intrinsic.
742   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
743 
744   // ConstantFP nodes default to expand.  Targets can either change this to
745   // Legal, in which case all fp constants are legal, or use isFPImmLegal()
746   // to optimize expansions for certain constants.
747   setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
748   setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
749   setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
750   setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
751   setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
752 
753   // These library functions default to expand.
754   for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
755     setOperationAction(ISD::FCBRT,      VT, Expand);
756     setOperationAction(ISD::FLOG ,      VT, Expand);
757     setOperationAction(ISD::FLOG2,      VT, Expand);
758     setOperationAction(ISD::FLOG10,     VT, Expand);
759     setOperationAction(ISD::FEXP ,      VT, Expand);
760     setOperationAction(ISD::FEXP2,      VT, Expand);
761     setOperationAction(ISD::FFLOOR,     VT, Expand);
762     setOperationAction(ISD::FNEARBYINT, VT, Expand);
763     setOperationAction(ISD::FCEIL,      VT, Expand);
764     setOperationAction(ISD::FRINT,      VT, Expand);
765     setOperationAction(ISD::FTRUNC,     VT, Expand);
766     setOperationAction(ISD::FROUND,     VT, Expand);
767     setOperationAction(ISD::FROUNDEVEN, VT, Expand);
768     setOperationAction(ISD::LROUND,     VT, Expand);
769     setOperationAction(ISD::LLROUND,    VT, Expand);
770     setOperationAction(ISD::LRINT,      VT, Expand);
771     setOperationAction(ISD::LLRINT,     VT, Expand);
772   }
773 
774   // Default ISD::TRAP to expand (which turns it into abort).
775   setOperationAction(ISD::TRAP, MVT::Other, Expand);
776 
777   // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
778   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
779   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
780 }
781 
782 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
783                                                EVT) const {
784   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
785 }
786 
787 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
788                                          bool LegalTypes) const {
789   assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
790   if (LHSTy.isVector())
791     return LHSTy;
792   return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
793                     : getPointerTy(DL);
794 }
795 
796 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
797   assert(isTypeLegal(VT));
798   switch (Op) {
799   default:
800     return false;
801   case ISD::SDIV:
802   case ISD::UDIV:
803   case ISD::SREM:
804   case ISD::UREM:
805     return true;
806   }
807 }
808 
809 bool TargetLoweringBase::isFreeAddrSpaceCast(unsigned SrcAS,
810                                              unsigned DestAS) const {
811   return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
812 }
813 
814 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
815   // If the command-line option was specified, ignore this request.
816   if (!JumpIsExpensiveOverride.getNumOccurrences())
817     JumpIsExpensive = isExpensive;
818 }
819 
820 TargetLoweringBase::LegalizeKind
821 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
822   // If this is a simple type, use the ComputeRegisterProp mechanism.
823   if (VT.isSimple()) {
824     MVT SVT = VT.getSimpleVT();
825     assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
826     MVT NVT = TransformToType[SVT.SimpleTy];
827     LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
828 
829     assert((LA == TypeLegal || LA == TypeSoftenFloat ||
830             LA == TypeSoftPromoteHalf ||
831             (NVT.isVector() ||
832              ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
833            "Promote may not follow Expand or Promote");
834 
835     if (LA == TypeSplitVector)
836       return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context));
837     if (LA == TypeScalarizeVector)
838       return LegalizeKind(LA, SVT.getVectorElementType());
839     return LegalizeKind(LA, NVT);
840   }
841 
842   // Handle Extended Scalar Types.
843   if (!VT.isVector()) {
844     assert(VT.isInteger() && "Float types must be simple");
845     unsigned BitSize = VT.getSizeInBits();
846     // First promote to a power-of-two size, then expand if necessary.
847     if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
848       EVT NVT = VT.getRoundIntegerType(Context);
849       assert(NVT != VT && "Unable to round integer VT");
850       LegalizeKind NextStep = getTypeConversion(Context, NVT);
851       // Avoid multi-step promotion.
852       if (NextStep.first == TypePromoteInteger)
853         return NextStep;
854       // Return rounded integer type.
855       return LegalizeKind(TypePromoteInteger, NVT);
856     }
857 
858     return LegalizeKind(TypeExpandInteger,
859                         EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
860   }
861 
862   // Handle vector types.
863   ElementCount NumElts = VT.getVectorElementCount();
864   EVT EltVT = VT.getVectorElementType();
865 
866   // Vectors with only one element are always scalarized.
867   if (NumElts.isScalar())
868     return LegalizeKind(TypeScalarizeVector, EltVT);
869 
870   if (VT.getVectorElementCount() == ElementCount::getScalable(1))
871     report_fatal_error("Cannot legalize this vector");
872 
873   // Try to widen vector elements until the element type is a power of two and
874   // promote it to a legal type later on, for example:
875   // <3 x i8> -> <4 x i8> -> <4 x i32>
876   if (EltVT.isInteger()) {
877     // Vectors with a number of elements that is not a power of two are always
878     // widened, for example <3 x i8> -> <4 x i8>.
879     if (!VT.isPow2VectorType()) {
880       NumElts = NumElts.coefficientNextPowerOf2();
881       EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
882       return LegalizeKind(TypeWidenVector, NVT);
883     }
884 
885     // Examine the element type.
886     LegalizeKind LK = getTypeConversion(Context, EltVT);
887 
888     // If type is to be expanded, split the vector.
889     //  <4 x i140> -> <2 x i140>
890     if (LK.first == TypeExpandInteger)
891       return LegalizeKind(TypeSplitVector,
892                           VT.getHalfNumVectorElementsVT(Context));
893 
894     // Promote the integer element types until a legal vector type is found
895     // or until the element integer type is too big. If a legal type was not
896     // found, fallback to the usual mechanism of widening/splitting the
897     // vector.
898     EVT OldEltVT = EltVT;
899     while (true) {
900       // Increase the bitwidth of the element to the next pow-of-two
901       // (which is greater than 8 bits).
902       EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
903                   .getRoundIntegerType(Context);
904 
905       // Stop trying when getting a non-simple element type.
906       // Note that vector elements may be greater than legal vector element
907       // types. Example: X86 XMM registers hold 64bit element on 32bit
908       // systems.
909       if (!EltVT.isSimple())
910         break;
911 
912       // Build a new vector type and check if it is legal.
913       MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
914       // Found a legal promoted vector type.
915       if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
916         return LegalizeKind(TypePromoteInteger,
917                             EVT::getVectorVT(Context, EltVT, NumElts));
918     }
919 
920     // Reset the type to the unexpanded type if we did not find a legal vector
921     // type with a promoted vector element type.
922     EltVT = OldEltVT;
923   }
924 
925   // Try to widen the vector until a legal type is found.
926   // If there is no wider legal type, split the vector.
927   while (true) {
928     // Round up to the next power of 2.
929     NumElts = NumElts.coefficientNextPowerOf2();
930 
931     // If there is no simple vector type with this many elements then there
932     // cannot be a larger legal vector type.  Note that this assumes that
933     // there are no skipped intermediate vector types in the simple types.
934     if (!EltVT.isSimple())
935       break;
936     MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
937     if (LargerVector == MVT())
938       break;
939 
940     // If this type is legal then widen the vector.
941     if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
942       return LegalizeKind(TypeWidenVector, LargerVector);
943   }
944 
945   // Widen odd vectors to next power of two.
946   if (!VT.isPow2VectorType()) {
947     EVT NVT = VT.getPow2VectorType(Context);
948     return LegalizeKind(TypeWidenVector, NVT);
949   }
950 
951   // Vectors with illegal element types are expanded.
952   EVT NVT = EVT::getVectorVT(Context, EltVT,
953                              VT.getVectorElementCount().divideCoefficientBy(2));
954   return LegalizeKind(TypeSplitVector, NVT);
955 }
956 
957 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
958                                           unsigned &NumIntermediates,
959                                           MVT &RegisterVT,
960                                           TargetLoweringBase *TLI) {
961   // Figure out the right, legal destination reg to copy into.
962   ElementCount EC = VT.getVectorElementCount();
963   MVT EltTy = VT.getVectorElementType();
964 
965   unsigned NumVectorRegs = 1;
966 
967   // Scalable vectors cannot be scalarized, so splitting or widening is
968   // required.
969   if (VT.isScalableVector() && !isPowerOf2_32(EC.getKnownMinValue()))
970     llvm_unreachable(
971         "Splitting or widening of non-power-of-2 MVTs is not implemented.");
972 
973   // FIXME: We don't support non-power-of-2-sized vectors for now.
974   // Ideally we could break down into LHS/RHS like LegalizeDAG does.
975   if (!isPowerOf2_32(EC.getKnownMinValue())) {
976     // Split EC to unit size (scalable property is preserved).
977     NumVectorRegs = EC.getKnownMinValue();
978     EC = ElementCount::getFixed(1);
979   }
980 
981   // Divide the input until we get to a supported size. This will
982   // always end up with an EC that represent a scalar or a scalable
983   // scalar.
984   while (EC.getKnownMinValue() > 1 &&
985          !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
986     EC = EC.divideCoefficientBy(2);
987     NumVectorRegs <<= 1;
988   }
989 
990   NumIntermediates = NumVectorRegs;
991 
992   MVT NewVT = MVT::getVectorVT(EltTy, EC);
993   if (!TLI->isTypeLegal(NewVT))
994     NewVT = EltTy;
995   IntermediateVT = NewVT;
996 
997   unsigned LaneSizeInBits = NewVT.getScalarSizeInBits();
998 
999   // Convert sizes such as i33 to i64.
1000   if (!isPowerOf2_32(LaneSizeInBits))
1001     LaneSizeInBits = NextPowerOf2(LaneSizeInBits);
1002 
1003   MVT DestVT = TLI->getRegisterType(NewVT);
1004   RegisterVT = DestVT;
1005   if (EVT(DestVT).bitsLT(NewVT))    // Value is expanded, e.g. i64 -> i16.
1006     return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits());
1007 
1008   // Otherwise, promotion or legal types use the same number of registers as
1009   // the vector decimated to the appropriate level.
1010   return NumVectorRegs;
1011 }
1012 
1013 /// isLegalRC - Return true if the value types that can be represented by the
1014 /// specified register class are all legal.
1015 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
1016                                    const TargetRegisterClass &RC) const {
1017   for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1018     if (isTypeLegal(*I))
1019       return true;
1020   return false;
1021 }
1022 
1023 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1024 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1025 MachineBasicBlock *
1026 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1027                                    MachineBasicBlock *MBB) const {
1028   MachineInstr *MI = &InitialMI;
1029   MachineFunction &MF = *MI->getMF();
1030   MachineFrameInfo &MFI = MF.getFrameInfo();
1031 
1032   // We're handling multiple types of operands here:
1033   // PATCHPOINT MetaArgs - live-in, read only, direct
1034   // STATEPOINT Deopt Spill - live-through, read only, indirect
1035   // STATEPOINT Deopt Alloca - live-through, read only, direct
1036   // (We're currently conservative and mark the deopt slots read/write in
1037   // practice.)
1038   // STATEPOINT GC Spill - live-through, read/write, indirect
1039   // STATEPOINT GC Alloca - live-through, read/write, direct
1040   // The live-in vs live-through is handled already (the live through ones are
1041   // all stack slots), but we need to handle the different type of stackmap
1042   // operands and memory effects here.
1043 
1044   if (!llvm::any_of(MI->operands(),
1045                     [](MachineOperand &Operand) { return Operand.isFI(); }))
1046     return MBB;
1047 
1048   MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1049 
1050   // Inherit previous memory operands.
1051   MIB.cloneMemRefs(*MI);
1052 
1053   for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1054     MachineOperand &MO = MI->getOperand(i);
1055     if (!MO.isFI()) {
1056       // Index of Def operand this Use it tied to.
1057       // Since Defs are coming before Uses, if Use is tied, then
1058       // index of Def must be smaller that index of that Use.
1059       // Also, Defs preserve their position in new MI.
1060       unsigned TiedTo = i;
1061       if (MO.isReg() && MO.isTied())
1062         TiedTo = MI->findTiedOperandIdx(i);
1063       MIB.add(MO);
1064       if (TiedTo < i)
1065         MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1066       continue;
1067     }
1068 
1069     // foldMemoryOperand builds a new MI after replacing a single FI operand
1070     // with the canonical set of five x86 addressing-mode operands.
1071     int FI = MO.getIndex();
1072 
1073     // Add frame index operands recognized by stackmaps.cpp
1074     if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
1075       // indirect-mem-ref tag, size, #FI, offset.
1076       // Used for spills inserted by StatepointLowering.  This codepath is not
1077       // used for patchpoints/stackmaps at all, for these spilling is done via
1078       // foldMemoryOperand callback only.
1079       assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1080       MIB.addImm(StackMaps::IndirectMemRefOp);
1081       MIB.addImm(MFI.getObjectSize(FI));
1082       MIB.add(MO);
1083       MIB.addImm(0);
1084     } else {
1085       // direct-mem-ref tag, #FI, offset.
1086       // Used by patchpoint, and direct alloca arguments to statepoints
1087       MIB.addImm(StackMaps::DirectMemRefOp);
1088       MIB.add(MO);
1089       MIB.addImm(0);
1090     }
1091 
1092     assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1093 
1094     // Add a new memory operand for this FI.
1095     assert(MFI.getObjectOffset(FI) != -1);
1096 
1097     // Note: STATEPOINT MMOs are added during SelectionDAG.  STACKMAP, and
1098     // PATCHPOINT should be updated to do the same. (TODO)
1099     if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1100       auto Flags = MachineMemOperand::MOLoad;
1101       MachineMemOperand *MMO = MF.getMachineMemOperand(
1102           MachinePointerInfo::getFixedStack(MF, FI), Flags,
1103           MF.getDataLayout().getPointerSize(), MFI.getObjectAlign(FI));
1104       MIB->addMemOperand(MF, MMO);
1105     }
1106   }
1107   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1108   MI->eraseFromParent();
1109   return MBB;
1110 }
1111 
1112 MachineBasicBlock *
1113 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
1114                                         MachineBasicBlock *MBB) const {
1115   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1116          "Called emitXRayCustomEvent on the wrong MI!");
1117   auto &MF = *MI.getMF();
1118   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1119   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1120     MIB.add(MI.getOperand(OpIdx));
1121 
1122   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1123   MI.eraseFromParent();
1124   return MBB;
1125 }
1126 
1127 MachineBasicBlock *
1128 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1129                                        MachineBasicBlock *MBB) const {
1130   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1131          "Called emitXRayTypedEvent on the wrong MI!");
1132   auto &MF = *MI.getMF();
1133   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1134   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1135     MIB.add(MI.getOperand(OpIdx));
1136 
1137   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1138   MI.eraseFromParent();
1139   return MBB;
1140 }
1141 
1142 /// findRepresentativeClass - Return the largest legal super-reg register class
1143 /// of the register class for the specified type and its associated "cost".
1144 // This function is in TargetLowering because it uses RegClassForVT which would
1145 // need to be moved to TargetRegisterInfo and would necessitate moving
1146 // isTypeLegal over as well - a massive change that would just require
1147 // TargetLowering having a TargetRegisterInfo class member that it would use.
1148 std::pair<const TargetRegisterClass *, uint8_t>
1149 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1150                                             MVT VT) const {
1151   const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1152   if (!RC)
1153     return std::make_pair(RC, 0);
1154 
1155   // Compute the set of all super-register classes.
1156   BitVector SuperRegRC(TRI->getNumRegClasses());
1157   for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1158     SuperRegRC.setBitsInMask(RCI.getMask());
1159 
1160   // Find the first legal register class with the largest spill size.
1161   const TargetRegisterClass *BestRC = RC;
1162   for (unsigned i : SuperRegRC.set_bits()) {
1163     const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1164     // We want the largest possible spill size.
1165     if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1166       continue;
1167     if (!isLegalRC(*TRI, *SuperRC))
1168       continue;
1169     BestRC = SuperRC;
1170   }
1171   return std::make_pair(BestRC, 1);
1172 }
1173 
1174 /// computeRegisterProperties - Once all of the register classes are added,
1175 /// this allows us to compute derived properties we expose.
1176 void TargetLoweringBase::computeRegisterProperties(
1177     const TargetRegisterInfo *TRI) {
1178   static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1179                 "Too many value types for ValueTypeActions to hold!");
1180 
1181   // Everything defaults to needing one register.
1182   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1183     NumRegistersForVT[i] = 1;
1184     RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1185   }
1186   // ...except isVoid, which doesn't need any registers.
1187   NumRegistersForVT[MVT::isVoid] = 0;
1188 
1189   // Find the largest integer register class.
1190   unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1191   for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1192     assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1193 
1194   // Every integer value type larger than this largest register takes twice as
1195   // many registers to represent as the previous ValueType.
1196   for (unsigned ExpandedReg = LargestIntReg + 1;
1197        ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1198     NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1199     RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1200     TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1201     ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1202                                    TypeExpandInteger);
1203   }
1204 
1205   // Inspect all of the ValueType's smaller than the largest integer
1206   // register to see which ones need promotion.
1207   unsigned LegalIntReg = LargestIntReg;
1208   for (unsigned IntReg = LargestIntReg - 1;
1209        IntReg >= (unsigned)MVT::i1; --IntReg) {
1210     MVT IVT = (MVT::SimpleValueType)IntReg;
1211     if (isTypeLegal(IVT)) {
1212       LegalIntReg = IntReg;
1213     } else {
1214       RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1215         (MVT::SimpleValueType)LegalIntReg;
1216       ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1217     }
1218   }
1219 
1220   // ppcf128 type is really two f64's.
1221   if (!isTypeLegal(MVT::ppcf128)) {
1222     if (isTypeLegal(MVT::f64)) {
1223       NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1224       RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1225       TransformToType[MVT::ppcf128] = MVT::f64;
1226       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1227     } else {
1228       NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1229       RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1230       TransformToType[MVT::ppcf128] = MVT::i128;
1231       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1232     }
1233   }
1234 
1235   // Decide how to handle f128. If the target does not have native f128 support,
1236   // expand it to i128 and we will be generating soft float library calls.
1237   if (!isTypeLegal(MVT::f128)) {
1238     NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1239     RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1240     TransformToType[MVT::f128] = MVT::i128;
1241     ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1242   }
1243 
1244   // Decide how to handle f64. If the target does not have native f64 support,
1245   // expand it to i64 and we will be generating soft float library calls.
1246   if (!isTypeLegal(MVT::f64)) {
1247     NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1248     RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1249     TransformToType[MVT::f64] = MVT::i64;
1250     ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1251   }
1252 
1253   // Decide how to handle f32. If the target does not have native f32 support,
1254   // expand it to i32 and we will be generating soft float library calls.
1255   if (!isTypeLegal(MVT::f32)) {
1256     NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1257     RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1258     TransformToType[MVT::f32] = MVT::i32;
1259     ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1260   }
1261 
1262   // Decide how to handle f16. If the target does not have native f16 support,
1263   // promote it to f32, because there are no f16 library calls (except for
1264   // conversions).
1265   if (!isTypeLegal(MVT::f16)) {
1266     // Allow targets to control how we legalize half.
1267     if (softPromoteHalfType()) {
1268       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1269       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1270       TransformToType[MVT::f16] = MVT::f32;
1271       ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1272     } else {
1273       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1274       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1275       TransformToType[MVT::f16] = MVT::f32;
1276       ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1277     }
1278   }
1279 
1280   // Loop over all of the vector value types to see which need transformations.
1281   for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1282        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1283     MVT VT = (MVT::SimpleValueType) i;
1284     if (isTypeLegal(VT))
1285       continue;
1286 
1287     MVT EltVT = VT.getVectorElementType();
1288     ElementCount EC = VT.getVectorElementCount();
1289     bool IsLegalWiderType = false;
1290     bool IsScalable = VT.isScalableVector();
1291     LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1292     switch (PreferredAction) {
1293     case TypePromoteInteger: {
1294       MVT::SimpleValueType EndVT = IsScalable ?
1295                                    MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1296                                    MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1297       // Try to promote the elements of integer vectors. If no legal
1298       // promotion was found, fall through to the widen-vector method.
1299       for (unsigned nVT = i + 1;
1300            (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1301         MVT SVT = (MVT::SimpleValueType) nVT;
1302         // Promote vectors of integers to vectors with the same number
1303         // of elements, with a wider element type.
1304         if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() &&
1305             SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1306           TransformToType[i] = SVT;
1307           RegisterTypeForVT[i] = SVT;
1308           NumRegistersForVT[i] = 1;
1309           ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1310           IsLegalWiderType = true;
1311           break;
1312         }
1313       }
1314       if (IsLegalWiderType)
1315         break;
1316       LLVM_FALLTHROUGH;
1317     }
1318 
1319     case TypeWidenVector:
1320       if (isPowerOf2_32(EC.getKnownMinValue())) {
1321         // Try to widen the vector.
1322         for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1323           MVT SVT = (MVT::SimpleValueType) nVT;
1324           if (SVT.getVectorElementType() == EltVT &&
1325               SVT.isScalableVector() == IsScalable &&
1326               SVT.getVectorElementCount().getKnownMinValue() >
1327                   EC.getKnownMinValue() &&
1328               isTypeLegal(SVT)) {
1329             TransformToType[i] = SVT;
1330             RegisterTypeForVT[i] = SVT;
1331             NumRegistersForVT[i] = 1;
1332             ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1333             IsLegalWiderType = true;
1334             break;
1335           }
1336         }
1337         if (IsLegalWiderType)
1338           break;
1339       } else {
1340         // Only widen to the next power of 2 to keep consistency with EVT.
1341         MVT NVT = VT.getPow2VectorType();
1342         if (isTypeLegal(NVT)) {
1343           TransformToType[i] = NVT;
1344           ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1345           RegisterTypeForVT[i] = NVT;
1346           NumRegistersForVT[i] = 1;
1347           break;
1348         }
1349       }
1350       LLVM_FALLTHROUGH;
1351 
1352     case TypeSplitVector:
1353     case TypeScalarizeVector: {
1354       MVT IntermediateVT;
1355       MVT RegisterVT;
1356       unsigned NumIntermediates;
1357       unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1358           NumIntermediates, RegisterVT, this);
1359       NumRegistersForVT[i] = NumRegisters;
1360       assert(NumRegistersForVT[i] == NumRegisters &&
1361              "NumRegistersForVT size cannot represent NumRegisters!");
1362       RegisterTypeForVT[i] = RegisterVT;
1363 
1364       MVT NVT = VT.getPow2VectorType();
1365       if (NVT == VT) {
1366         // Type is already a power of 2.  The default action is to split.
1367         TransformToType[i] = MVT::Other;
1368         if (PreferredAction == TypeScalarizeVector)
1369           ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1370         else if (PreferredAction == TypeSplitVector)
1371           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1372         else if (EC.getKnownMinValue() > 1)
1373           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1374         else
1375           ValueTypeActions.setTypeAction(VT, EC.isScalable()
1376                                                  ? TypeScalarizeScalableVector
1377                                                  : TypeScalarizeVector);
1378       } else {
1379         TransformToType[i] = NVT;
1380         ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1381       }
1382       break;
1383     }
1384     default:
1385       llvm_unreachable("Unknown vector legalization action!");
1386     }
1387   }
1388 
1389   // Determine the 'representative' register class for each value type.
1390   // An representative register class is the largest (meaning one which is
1391   // not a sub-register class / subreg register class) legal register class for
1392   // a group of value types. For example, on i386, i8, i16, and i32
1393   // representative would be GR32; while on x86_64 it's GR64.
1394   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1395     const TargetRegisterClass* RRC;
1396     uint8_t Cost;
1397     std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1398     RepRegClassForVT[i] = RRC;
1399     RepRegClassCostForVT[i] = Cost;
1400   }
1401 }
1402 
1403 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1404                                            EVT VT) const {
1405   assert(!VT.isVector() && "No default SetCC type for vectors!");
1406   return getPointerTy(DL).SimpleTy;
1407 }
1408 
1409 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1410   return MVT::i32; // return the default value
1411 }
1412 
1413 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1414 /// legal first class types.  For example, MVT::v8f32 maps to 2 MVT::v4f32
1415 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1416 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1417 ///
1418 /// This method returns the number of registers needed, and the VT for each
1419 /// register.  It also returns the VT and quantity of the intermediate values
1420 /// before they are promoted/expanded.
1421 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1422                                                 EVT &IntermediateVT,
1423                                                 unsigned &NumIntermediates,
1424                                                 MVT &RegisterVT) const {
1425   ElementCount EltCnt = VT.getVectorElementCount();
1426 
1427   // If there is a wider vector type with the same element type as this one,
1428   // or a promoted vector type that has the same number of elements which
1429   // are wider, then we should convert to that legal vector type.
1430   // This handles things like <2 x float> -> <4 x float> and
1431   // <4 x i1> -> <4 x i32>.
1432   LegalizeTypeAction TA = getTypeAction(Context, VT);
1433   if (EltCnt.getKnownMinValue() != 1 &&
1434       (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1435     EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1436     if (isTypeLegal(RegisterEVT)) {
1437       IntermediateVT = RegisterEVT;
1438       RegisterVT = RegisterEVT.getSimpleVT();
1439       NumIntermediates = 1;
1440       return 1;
1441     }
1442   }
1443 
1444   // Figure out the right, legal destination reg to copy into.
1445   EVT EltTy = VT.getVectorElementType();
1446 
1447   unsigned NumVectorRegs = 1;
1448 
1449   // Scalable vectors cannot be scalarized, so handle the legalisation of the
1450   // types like done elsewhere in SelectionDAG.
1451   if (VT.isScalableVector() && !isPowerOf2_32(EltCnt.getKnownMinValue())) {
1452     LegalizeKind LK;
1453     EVT PartVT = VT;
1454     do {
1455       // Iterate until we've found a legal (part) type to hold VT.
1456       LK = getTypeConversion(Context, PartVT);
1457       PartVT = LK.second;
1458     } while (LK.first != TypeLegal);
1459 
1460     NumIntermediates = VT.getVectorElementCount().getKnownMinValue() /
1461                        PartVT.getVectorElementCount().getKnownMinValue();
1462 
1463     // FIXME: This code needs to be extended to handle more complex vector
1464     // breakdowns, like nxv7i64 -> nxv8i64 -> 4 x nxv2i64. Currently the only
1465     // supported cases are vectors that are broken down into equal parts
1466     // such as nxv6i64 -> 3 x nxv2i64.
1467     assert((PartVT.getVectorElementCount() * NumIntermediates) ==
1468                VT.getVectorElementCount() &&
1469            "Expected an integer multiple of PartVT");
1470     IntermediateVT = PartVT;
1471     RegisterVT = getRegisterType(Context, IntermediateVT);
1472     return NumIntermediates;
1473   }
1474 
1475   // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally
1476   // we could break down into LHS/RHS like LegalizeDAG does.
1477   if (!isPowerOf2_32(EltCnt.getKnownMinValue())) {
1478     NumVectorRegs = EltCnt.getKnownMinValue();
1479     EltCnt = ElementCount::getFixed(1);
1480   }
1481 
1482   // Divide the input until we get to a supported size.  This will always
1483   // end with a scalar if the target doesn't support vectors.
1484   while (EltCnt.getKnownMinValue() > 1 &&
1485          !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1486     EltCnt = EltCnt.divideCoefficientBy(2);
1487     NumVectorRegs <<= 1;
1488   }
1489 
1490   NumIntermediates = NumVectorRegs;
1491 
1492   EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1493   if (!isTypeLegal(NewVT))
1494     NewVT = EltTy;
1495   IntermediateVT = NewVT;
1496 
1497   MVT DestVT = getRegisterType(Context, NewVT);
1498   RegisterVT = DestVT;
1499 
1500   if (EVT(DestVT).bitsLT(NewVT)) {  // Value is expanded, e.g. i64 -> i16.
1501     TypeSize NewVTSize = NewVT.getSizeInBits();
1502     // Convert sizes such as i33 to i64.
1503     if (!isPowerOf2_32(NewVTSize.getKnownMinSize()))
1504       NewVTSize = NewVTSize.coefficientNextPowerOf2();
1505     return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1506   }
1507 
1508   // Otherwise, promotion or legal types use the same number of registers as
1509   // the vector decimated to the appropriate level.
1510   return NumVectorRegs;
1511 }
1512 
1513 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI,
1514                                                 uint64_t NumCases,
1515                                                 uint64_t Range,
1516                                                 ProfileSummaryInfo *PSI,
1517                                                 BlockFrequencyInfo *BFI) const {
1518   // FIXME: This function check the maximum table size and density, but the
1519   // minimum size is not checked. It would be nice if the minimum size is
1520   // also combined within this function. Currently, the minimum size check is
1521   // performed in findJumpTable() in SelectionDAGBuiler and
1522   // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1523   const bool OptForSize =
1524       SI->getParent()->getParent()->hasOptSize() ||
1525       llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
1526   const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1527   const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1528 
1529   // Check whether the number of cases is small enough and
1530   // the range is dense enough for a jump table.
1531   return (OptForSize || Range <= MaxJumpTableSize) &&
1532          (NumCases * 100 >= Range * MinDensity);
1533 }
1534 
1535 /// Get the EVTs and ArgFlags collections that represent the legalized return
1536 /// type of the given function.  This does not require a DAG or a return value,
1537 /// and is suitable for use before any DAGs for the function are constructed.
1538 /// TODO: Move this out of TargetLowering.cpp.
1539 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1540                          AttributeList attr,
1541                          SmallVectorImpl<ISD::OutputArg> &Outs,
1542                          const TargetLowering &TLI, const DataLayout &DL) {
1543   SmallVector<EVT, 4> ValueVTs;
1544   ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1545   unsigned NumValues = ValueVTs.size();
1546   if (NumValues == 0) return;
1547 
1548   for (unsigned j = 0, f = NumValues; j != f; ++j) {
1549     EVT VT = ValueVTs[j];
1550     ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1551 
1552     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1553       ExtendKind = ISD::SIGN_EXTEND;
1554     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1555       ExtendKind = ISD::ZERO_EXTEND;
1556 
1557     // FIXME: C calling convention requires the return type to be promoted to
1558     // at least 32-bit. But this is not necessary for non-C calling
1559     // conventions. The frontend should mark functions whose return values
1560     // require promoting with signext or zeroext attributes.
1561     if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1562       MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1563       if (VT.bitsLT(MinVT))
1564         VT = MinVT;
1565     }
1566 
1567     unsigned NumParts =
1568         TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1569     MVT PartVT =
1570         TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1571 
1572     // 'inreg' on function refers to return value
1573     ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1574     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1575       Flags.setInReg();
1576 
1577     // Propagate extension type if any
1578     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1579       Flags.setSExt();
1580     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1581       Flags.setZExt();
1582 
1583     for (unsigned i = 0; i < NumParts; ++i)
1584       Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0));
1585   }
1586 }
1587 
1588 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1589 /// function arguments in the caller parameter area.  This is the actual
1590 /// alignment, not its logarithm.
1591 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1592                                                    const DataLayout &DL) const {
1593   return DL.getABITypeAlign(Ty).value();
1594 }
1595 
1596 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1597     LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1598     Align Alignment, MachineMemOperand::Flags Flags, bool *Fast) const {
1599   // Check if the specified alignment is sufficient based on the data layout.
1600   // TODO: While using the data layout works in practice, a better solution
1601   // would be to implement this check directly (make this a virtual function).
1602   // For example, the ABI alignment may change based on software platform while
1603   // this function should only be affected by hardware implementation.
1604   Type *Ty = VT.getTypeForEVT(Context);
1605   if (Alignment >= DL.getABITypeAlign(Ty)) {
1606     // Assume that an access that meets the ABI-specified alignment is fast.
1607     if (Fast != nullptr)
1608       *Fast = true;
1609     return true;
1610   }
1611 
1612   // This is a misaligned access.
1613   return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment.value(), Flags,
1614                                         Fast);
1615 }
1616 
1617 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1618     LLVMContext &Context, const DataLayout &DL, EVT VT,
1619     const MachineMemOperand &MMO, bool *Fast) const {
1620   return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
1621                                         MMO.getAlign(), MMO.getFlags(), Fast);
1622 }
1623 
1624 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1625                                             const DataLayout &DL, EVT VT,
1626                                             unsigned AddrSpace, Align Alignment,
1627                                             MachineMemOperand::Flags Flags,
1628                                             bool *Fast) const {
1629   return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
1630                                         Flags, Fast);
1631 }
1632 
1633 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1634                                             const DataLayout &DL, EVT VT,
1635                                             const MachineMemOperand &MMO,
1636                                             bool *Fast) const {
1637   return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
1638                             MMO.getFlags(), Fast);
1639 }
1640 
1641 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1642   return BranchProbability(MinPercentageForPredictableBranch, 100);
1643 }
1644 
1645 //===----------------------------------------------------------------------===//
1646 //  TargetTransformInfo Helpers
1647 //===----------------------------------------------------------------------===//
1648 
1649 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1650   enum InstructionOpcodes {
1651 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1652 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1653 #include "llvm/IR/Instruction.def"
1654   };
1655   switch (static_cast<InstructionOpcodes>(Opcode)) {
1656   case Ret:            return 0;
1657   case Br:             return 0;
1658   case Switch:         return 0;
1659   case IndirectBr:     return 0;
1660   case Invoke:         return 0;
1661   case CallBr:         return 0;
1662   case Resume:         return 0;
1663   case Unreachable:    return 0;
1664   case CleanupRet:     return 0;
1665   case CatchRet:       return 0;
1666   case CatchPad:       return 0;
1667   case CatchSwitch:    return 0;
1668   case CleanupPad:     return 0;
1669   case FNeg:           return ISD::FNEG;
1670   case Add:            return ISD::ADD;
1671   case FAdd:           return ISD::FADD;
1672   case Sub:            return ISD::SUB;
1673   case FSub:           return ISD::FSUB;
1674   case Mul:            return ISD::MUL;
1675   case FMul:           return ISD::FMUL;
1676   case UDiv:           return ISD::UDIV;
1677   case SDiv:           return ISD::SDIV;
1678   case FDiv:           return ISD::FDIV;
1679   case URem:           return ISD::UREM;
1680   case SRem:           return ISD::SREM;
1681   case FRem:           return ISD::FREM;
1682   case Shl:            return ISD::SHL;
1683   case LShr:           return ISD::SRL;
1684   case AShr:           return ISD::SRA;
1685   case And:            return ISD::AND;
1686   case Or:             return ISD::OR;
1687   case Xor:            return ISD::XOR;
1688   case Alloca:         return 0;
1689   case Load:           return ISD::LOAD;
1690   case Store:          return ISD::STORE;
1691   case GetElementPtr:  return 0;
1692   case Fence:          return 0;
1693   case AtomicCmpXchg:  return 0;
1694   case AtomicRMW:      return 0;
1695   case Trunc:          return ISD::TRUNCATE;
1696   case ZExt:           return ISD::ZERO_EXTEND;
1697   case SExt:           return ISD::SIGN_EXTEND;
1698   case FPToUI:         return ISD::FP_TO_UINT;
1699   case FPToSI:         return ISD::FP_TO_SINT;
1700   case UIToFP:         return ISD::UINT_TO_FP;
1701   case SIToFP:         return ISD::SINT_TO_FP;
1702   case FPTrunc:        return ISD::FP_ROUND;
1703   case FPExt:          return ISD::FP_EXTEND;
1704   case PtrToInt:       return ISD::BITCAST;
1705   case IntToPtr:       return ISD::BITCAST;
1706   case BitCast:        return ISD::BITCAST;
1707   case AddrSpaceCast:  return ISD::ADDRSPACECAST;
1708   case ICmp:           return ISD::SETCC;
1709   case FCmp:           return ISD::SETCC;
1710   case PHI:            return 0;
1711   case Call:           return 0;
1712   case Select:         return ISD::SELECT;
1713   case UserOp1:        return 0;
1714   case UserOp2:        return 0;
1715   case VAArg:          return 0;
1716   case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1717   case InsertElement:  return ISD::INSERT_VECTOR_ELT;
1718   case ShuffleVector:  return ISD::VECTOR_SHUFFLE;
1719   case ExtractValue:   return ISD::MERGE_VALUES;
1720   case InsertValue:    return ISD::MERGE_VALUES;
1721   case LandingPad:     return 0;
1722   case Freeze:         return ISD::FREEZE;
1723   }
1724 
1725   llvm_unreachable("Unknown instruction type encountered!");
1726 }
1727 
1728 std::pair<int, MVT>
1729 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1730                                             Type *Ty) const {
1731   LLVMContext &C = Ty->getContext();
1732   EVT MTy = getValueType(DL, Ty);
1733 
1734   int Cost = 1;
1735   // We keep legalizing the type until we find a legal kind. We assume that
1736   // the only operation that costs anything is the split. After splitting
1737   // we need to handle two types.
1738   while (true) {
1739     LegalizeKind LK = getTypeConversion(C, MTy);
1740 
1741     if (LK.first == TypeLegal)
1742       return std::make_pair(Cost, MTy.getSimpleVT());
1743 
1744     if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1745       Cost *= 2;
1746 
1747     // Do not loop with f128 type.
1748     if (MTy == LK.second)
1749       return std::make_pair(Cost, MTy.getSimpleVT());
1750 
1751     // Keep legalizing the type.
1752     MTy = LK.second;
1753   }
1754 }
1755 
1756 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1757                                                               bool UseTLS) const {
1758   // compiler-rt provides a variable with a magic name.  Targets that do not
1759   // link with compiler-rt may also provide such a variable.
1760   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1761   const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1762   auto UnsafeStackPtr =
1763       dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1764 
1765   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1766 
1767   if (!UnsafeStackPtr) {
1768     auto TLSModel = UseTLS ?
1769         GlobalValue::InitialExecTLSModel :
1770         GlobalValue::NotThreadLocal;
1771     // The global variable is not defined yet, define it ourselves.
1772     // We use the initial-exec TLS model because we do not support the
1773     // variable living anywhere other than in the main executable.
1774     UnsafeStackPtr = new GlobalVariable(
1775         *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1776         UnsafeStackPtrVar, nullptr, TLSModel);
1777   } else {
1778     // The variable exists, check its type and attributes.
1779     if (UnsafeStackPtr->getValueType() != StackPtrTy)
1780       report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1781     if (UseTLS != UnsafeStackPtr->isThreadLocal())
1782       report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1783                          (UseTLS ? "" : "not ") + "be thread-local");
1784   }
1785   return UnsafeStackPtr;
1786 }
1787 
1788 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1789   if (!TM.getTargetTriple().isAndroid())
1790     return getDefaultSafeStackPointerLocation(IRB, true);
1791 
1792   // Android provides a libc function to retrieve the address of the current
1793   // thread's unsafe stack pointer.
1794   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1795   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1796   FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address",
1797                                              StackPtrTy->getPointerTo(0));
1798   return IRB.CreateCall(Fn);
1799 }
1800 
1801 //===----------------------------------------------------------------------===//
1802 //  Loop Strength Reduction hooks
1803 //===----------------------------------------------------------------------===//
1804 
1805 /// isLegalAddressingMode - Return true if the addressing mode represented
1806 /// by AM is legal for this target, for a load/store of the specified type.
1807 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1808                                                const AddrMode &AM, Type *Ty,
1809                                                unsigned AS, Instruction *I) const {
1810   // The default implementation of this implements a conservative RISCy, r+r and
1811   // r+i addr mode.
1812 
1813   // Allows a sign-extended 16-bit immediate field.
1814   if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1815     return false;
1816 
1817   // No global is ever allowed as a base.
1818   if (AM.BaseGV)
1819     return false;
1820 
1821   // Only support r+r,
1822   switch (AM.Scale) {
1823   case 0:  // "r+i" or just "i", depending on HasBaseReg.
1824     break;
1825   case 1:
1826     if (AM.HasBaseReg && AM.BaseOffs)  // "r+r+i" is not allowed.
1827       return false;
1828     // Otherwise we have r+r or r+i.
1829     break;
1830   case 2:
1831     if (AM.HasBaseReg || AM.BaseOffs)  // 2*r+r  or  2*r+i is not allowed.
1832       return false;
1833     // Allow 2*r as r+r.
1834     break;
1835   default: // Don't allow n * r
1836     return false;
1837   }
1838 
1839   return true;
1840 }
1841 
1842 //===----------------------------------------------------------------------===//
1843 //  Stack Protector
1844 //===----------------------------------------------------------------------===//
1845 
1846 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1847 // so that SelectionDAG handle SSP.
1848 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1849   if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1850     Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1851     PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1852     Constant *C = M.getOrInsertGlobal("__guard_local", PtrTy);
1853     if (GlobalVariable *G = dyn_cast_or_null<GlobalVariable>(C))
1854       G->setVisibility(GlobalValue::HiddenVisibility);
1855     return C;
1856   }
1857   return nullptr;
1858 }
1859 
1860 // Currently only support "standard" __stack_chk_guard.
1861 // TODO: add LOAD_STACK_GUARD support.
1862 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1863   if (!M.getNamedValue("__stack_chk_guard"))
1864     new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1865                        GlobalVariable::ExternalLinkage,
1866                        nullptr, "__stack_chk_guard");
1867 }
1868 
1869 // Currently only support "standard" __stack_chk_guard.
1870 // TODO: add LOAD_STACK_GUARD support.
1871 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1872   return M.getNamedValue("__stack_chk_guard");
1873 }
1874 
1875 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1876   return nullptr;
1877 }
1878 
1879 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1880   return MinimumJumpTableEntries;
1881 }
1882 
1883 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1884   MinimumJumpTableEntries = Val;
1885 }
1886 
1887 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1888   return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1889 }
1890 
1891 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1892   return MaximumJumpTableSize;
1893 }
1894 
1895 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1896   MaximumJumpTableSize = Val;
1897 }
1898 
1899 bool TargetLoweringBase::isJumpTableRelative() const {
1900   return getTargetMachine().isPositionIndependent();
1901 }
1902 
1903 //===----------------------------------------------------------------------===//
1904 //  Reciprocal Estimates
1905 //===----------------------------------------------------------------------===//
1906 
1907 /// Get the reciprocal estimate attribute string for a function that will
1908 /// override the target defaults.
1909 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1910   const Function &F = MF.getFunction();
1911   return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1912 }
1913 
1914 /// Construct a string for the given reciprocal operation of the given type.
1915 /// This string should match the corresponding option to the front-end's
1916 /// "-mrecip" flag assuming those strings have been passed through in an
1917 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1918 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1919   std::string Name = VT.isVector() ? "vec-" : "";
1920 
1921   Name += IsSqrt ? "sqrt" : "div";
1922 
1923   // TODO: Handle "half" or other float types?
1924   if (VT.getScalarType() == MVT::f64) {
1925     Name += "d";
1926   } else {
1927     assert(VT.getScalarType() == MVT::f32 &&
1928            "Unexpected FP type for reciprocal estimate");
1929     Name += "f";
1930   }
1931 
1932   return Name;
1933 }
1934 
1935 /// Return the character position and value (a single numeric character) of a
1936 /// customized refinement operation in the input string if it exists. Return
1937 /// false if there is no customized refinement step count.
1938 static bool parseRefinementStep(StringRef In, size_t &Position,
1939                                 uint8_t &Value) {
1940   const char RefStepToken = ':';
1941   Position = In.find(RefStepToken);
1942   if (Position == StringRef::npos)
1943     return false;
1944 
1945   StringRef RefStepString = In.substr(Position + 1);
1946   // Allow exactly one numeric character for the additional refinement
1947   // step parameter.
1948   if (RefStepString.size() == 1) {
1949     char RefStepChar = RefStepString[0];
1950     if (RefStepChar >= '0' && RefStepChar <= '9') {
1951       Value = RefStepChar - '0';
1952       return true;
1953     }
1954   }
1955   report_fatal_error("Invalid refinement step for -recip.");
1956 }
1957 
1958 /// For the input attribute string, return one of the ReciprocalEstimate enum
1959 /// status values (enabled, disabled, or not specified) for this operation on
1960 /// the specified data type.
1961 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1962   if (Override.empty())
1963     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1964 
1965   SmallVector<StringRef, 4> OverrideVector;
1966   Override.split(OverrideVector, ',');
1967   unsigned NumArgs = OverrideVector.size();
1968 
1969   // Check if "all", "none", or "default" was specified.
1970   if (NumArgs == 1) {
1971     // Look for an optional setting of the number of refinement steps needed
1972     // for this type of reciprocal operation.
1973     size_t RefPos;
1974     uint8_t RefSteps;
1975     if (parseRefinementStep(Override, RefPos, RefSteps)) {
1976       // Split the string for further processing.
1977       Override = Override.substr(0, RefPos);
1978     }
1979 
1980     // All reciprocal types are enabled.
1981     if (Override == "all")
1982       return TargetLoweringBase::ReciprocalEstimate::Enabled;
1983 
1984     // All reciprocal types are disabled.
1985     if (Override == "none")
1986       return TargetLoweringBase::ReciprocalEstimate::Disabled;
1987 
1988     // Target defaults for enablement are used.
1989     if (Override == "default")
1990       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1991   }
1992 
1993   // The attribute string may omit the size suffix ('f'/'d').
1994   std::string VTName = getReciprocalOpName(IsSqrt, VT);
1995   std::string VTNameNoSize = VTName;
1996   VTNameNoSize.pop_back();
1997   static const char DisabledPrefix = '!';
1998 
1999   for (StringRef RecipType : OverrideVector) {
2000     size_t RefPos;
2001     uint8_t RefSteps;
2002     if (parseRefinementStep(RecipType, RefPos, RefSteps))
2003       RecipType = RecipType.substr(0, RefPos);
2004 
2005     // Ignore the disablement token for string matching.
2006     bool IsDisabled = RecipType[0] == DisabledPrefix;
2007     if (IsDisabled)
2008       RecipType = RecipType.substr(1);
2009 
2010     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2011       return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
2012                         : TargetLoweringBase::ReciprocalEstimate::Enabled;
2013   }
2014 
2015   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2016 }
2017 
2018 /// For the input attribute string, return the customized refinement step count
2019 /// for this operation on the specified data type. If the step count does not
2020 /// exist, return the ReciprocalEstimate enum value for unspecified.
2021 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2022   if (Override.empty())
2023     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2024 
2025   SmallVector<StringRef, 4> OverrideVector;
2026   Override.split(OverrideVector, ',');
2027   unsigned NumArgs = OverrideVector.size();
2028 
2029   // Check if "all", "default", or "none" was specified.
2030   if (NumArgs == 1) {
2031     // Look for an optional setting of the number of refinement steps needed
2032     // for this type of reciprocal operation.
2033     size_t RefPos;
2034     uint8_t RefSteps;
2035     if (!parseRefinementStep(Override, RefPos, RefSteps))
2036       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2037 
2038     // Split the string for further processing.
2039     Override = Override.substr(0, RefPos);
2040     assert(Override != "none" &&
2041            "Disabled reciprocals, but specifed refinement steps?");
2042 
2043     // If this is a general override, return the specified number of steps.
2044     if (Override == "all" || Override == "default")
2045       return RefSteps;
2046   }
2047 
2048   // The attribute string may omit the size suffix ('f'/'d').
2049   std::string VTName = getReciprocalOpName(IsSqrt, VT);
2050   std::string VTNameNoSize = VTName;
2051   VTNameNoSize.pop_back();
2052 
2053   for (StringRef RecipType : OverrideVector) {
2054     size_t RefPos;
2055     uint8_t RefSteps;
2056     if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2057       continue;
2058 
2059     RecipType = RecipType.substr(0, RefPos);
2060     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2061       return RefSteps;
2062   }
2063 
2064   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2065 }
2066 
2067 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
2068                                                     MachineFunction &MF) const {
2069   return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
2070 }
2071 
2072 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
2073                                                    MachineFunction &MF) const {
2074   return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
2075 }
2076 
2077 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
2078                                                MachineFunction &MF) const {
2079   return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
2080 }
2081 
2082 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
2083                                               MachineFunction &MF) const {
2084   return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
2085 }
2086 
2087 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
2088   MF.getRegInfo().freezeReservedRegs(MF);
2089 }
2090 
2091 MachineMemOperand::Flags
2092 TargetLoweringBase::getLoadMemOperandFlags(const LoadInst &LI,
2093                                            const DataLayout &DL) const {
2094   MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad;
2095   if (LI.isVolatile())
2096     Flags |= MachineMemOperand::MOVolatile;
2097 
2098   if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2099     Flags |= MachineMemOperand::MONonTemporal;
2100 
2101   if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2102     Flags |= MachineMemOperand::MOInvariant;
2103 
2104   if (isDereferenceablePointer(LI.getPointerOperand(), LI.getType(), DL))
2105     Flags |= MachineMemOperand::MODereferenceable;
2106 
2107   Flags |= getTargetMMOFlags(LI);
2108   return Flags;
2109 }
2110 
2111 MachineMemOperand::Flags
2112 TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI,
2113                                             const DataLayout &DL) const {
2114   MachineMemOperand::Flags Flags = MachineMemOperand::MOStore;
2115 
2116   if (SI.isVolatile())
2117     Flags |= MachineMemOperand::MOVolatile;
2118 
2119   if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2120     Flags |= MachineMemOperand::MONonTemporal;
2121 
2122   // FIXME: Not preserving dereferenceable
2123   Flags |= getTargetMMOFlags(SI);
2124   return Flags;
2125 }
2126 
2127 MachineMemOperand::Flags
2128 TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI,
2129                                              const DataLayout &DL) const {
2130   auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
2131 
2132   if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2133     if (RMW->isVolatile())
2134       Flags |= MachineMemOperand::MOVolatile;
2135   } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2136     if (CmpX->isVolatile())
2137       Flags |= MachineMemOperand::MOVolatile;
2138   } else
2139     llvm_unreachable("not an atomic instruction");
2140 
2141   // FIXME: Not preserving dereferenceable
2142   Flags |= getTargetMMOFlags(AI);
2143   return Flags;
2144 }
2145 
2146 //===----------------------------------------------------------------------===//
2147 //  GlobalISel Hooks
2148 //===----------------------------------------------------------------------===//
2149 
2150 bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI,
2151                                         const TargetTransformInfo *TTI) const {
2152   auto &MF = *MI.getMF();
2153   auto &MRI = MF.getRegInfo();
2154   // Assuming a spill and reload of a value has a cost of 1 instruction each,
2155   // this helper function computes the maximum number of uses we should consider
2156   // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2157   // break even in terms of code size when the original MI has 2 users vs
2158   // choosing to potentially spill. Any more than 2 users we we have a net code
2159   // size increase. This doesn't take into account register pressure though.
2160   auto maxUses = [](unsigned RematCost) {
2161     // A cost of 1 means remats are basically free.
2162     if (RematCost == 1)
2163       return UINT_MAX;
2164     if (RematCost == 2)
2165       return 2U;
2166 
2167     // Remat is too expensive, only sink if there's one user.
2168     if (RematCost > 2)
2169       return 1U;
2170     llvm_unreachable("Unexpected remat cost");
2171   };
2172 
2173   // Helper to walk through uses and terminate if we've reached a limit. Saves
2174   // us spending time traversing uses if all we want to know is if it's >= min.
2175   auto isUsesAtMost = [&](unsigned Reg, unsigned MaxUses) {
2176     unsigned NumUses = 0;
2177     auto UI = MRI.use_instr_nodbg_begin(Reg), UE = MRI.use_instr_nodbg_end();
2178     for (; UI != UE && NumUses < MaxUses; ++UI) {
2179       NumUses++;
2180     }
2181     // If we haven't reached the end yet then there are more than MaxUses users.
2182     return UI == UE;
2183   };
2184 
2185   switch (MI.getOpcode()) {
2186   default:
2187     return false;
2188   // Constants-like instructions should be close to their users.
2189   // We don't want long live-ranges for them.
2190   case TargetOpcode::G_CONSTANT:
2191   case TargetOpcode::G_FCONSTANT:
2192   case TargetOpcode::G_FRAME_INDEX:
2193   case TargetOpcode::G_INTTOPTR:
2194     return true;
2195   case TargetOpcode::G_GLOBAL_VALUE: {
2196     unsigned RematCost = TTI->getGISelRematGlobalCost();
2197     Register Reg = MI.getOperand(0).getReg();
2198     unsigned MaxUses = maxUses(RematCost);
2199     if (MaxUses == UINT_MAX)
2200       return true; // Remats are "free" so always localize.
2201     bool B = isUsesAtMost(Reg, MaxUses);
2202     return B;
2203   }
2204   }
2205 }
2206