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.getSizeInBits().getFixedSize());
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::SMULFIX, VT, Expand);
661     setOperationAction(ISD::SMULFIXSAT, VT, Expand);
662     setOperationAction(ISD::UMULFIX, VT, Expand);
663     setOperationAction(ISD::UMULFIXSAT, VT, Expand);
664     setOperationAction(ISD::SDIVFIX, VT, Expand);
665     setOperationAction(ISD::SDIVFIXSAT, VT, Expand);
666     setOperationAction(ISD::UDIVFIX, VT, Expand);
667     setOperationAction(ISD::UDIVFIXSAT, VT, Expand);
668 
669     // Overflow operations default to expand
670     setOperationAction(ISD::SADDO, VT, Expand);
671     setOperationAction(ISD::SSUBO, VT, Expand);
672     setOperationAction(ISD::UADDO, VT, Expand);
673     setOperationAction(ISD::USUBO, VT, Expand);
674     setOperationAction(ISD::SMULO, VT, Expand);
675     setOperationAction(ISD::UMULO, VT, Expand);
676 
677     // ADDCARRY operations default to expand
678     setOperationAction(ISD::ADDCARRY, VT, Expand);
679     setOperationAction(ISD::SUBCARRY, VT, Expand);
680     setOperationAction(ISD::SETCCCARRY, VT, Expand);
681 
682     // ADDC/ADDE/SUBC/SUBE default to expand.
683     setOperationAction(ISD::ADDC, VT, Expand);
684     setOperationAction(ISD::ADDE, VT, Expand);
685     setOperationAction(ISD::SUBC, VT, Expand);
686     setOperationAction(ISD::SUBE, VT, Expand);
687 
688     // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
689     setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
690     setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
691 
692     setOperationAction(ISD::BITREVERSE, VT, Expand);
693 
694     // These library functions default to expand.
695     setOperationAction(ISD::FROUND, VT, Expand);
696     setOperationAction(ISD::FROUNDEVEN, VT, Expand);
697     setOperationAction(ISD::FPOWI, VT, Expand);
698 
699     // These operations default to expand for vector types.
700     if (VT.isVector()) {
701       setOperationAction(ISD::FCOPYSIGN, VT, Expand);
702       setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
703       setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
704       setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
705       setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
706       setOperationAction(ISD::SPLAT_VECTOR, VT, Expand);
707     }
708 
709     // Constrained floating-point operations default to expand.
710 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN)               \
711     setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
712 #include "llvm/IR/ConstrainedOps.def"
713 
714     // For most targets @llvm.get.dynamic.area.offset just returns 0.
715     setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
716 
717     // Vector reduction default to expand.
718     setOperationAction(ISD::VECREDUCE_FADD, VT, Expand);
719     setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand);
720     setOperationAction(ISD::VECREDUCE_ADD, VT, Expand);
721     setOperationAction(ISD::VECREDUCE_MUL, VT, Expand);
722     setOperationAction(ISD::VECREDUCE_AND, VT, Expand);
723     setOperationAction(ISD::VECREDUCE_OR, VT, Expand);
724     setOperationAction(ISD::VECREDUCE_XOR, VT, Expand);
725     setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand);
726     setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand);
727     setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand);
728     setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand);
729     setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand);
730     setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand);
731   }
732 
733   // Most targets ignore the @llvm.prefetch intrinsic.
734   setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
735 
736   // Most targets also ignore the @llvm.readcyclecounter intrinsic.
737   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
738 
739   // ConstantFP nodes default to expand.  Targets can either change this to
740   // Legal, in which case all fp constants are legal, or use isFPImmLegal()
741   // to optimize expansions for certain constants.
742   setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
743   setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
744   setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
745   setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
746   setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
747 
748   // These library functions default to expand.
749   for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
750     setOperationAction(ISD::FCBRT,      VT, Expand);
751     setOperationAction(ISD::FLOG ,      VT, Expand);
752     setOperationAction(ISD::FLOG2,      VT, Expand);
753     setOperationAction(ISD::FLOG10,     VT, Expand);
754     setOperationAction(ISD::FEXP ,      VT, Expand);
755     setOperationAction(ISD::FEXP2,      VT, Expand);
756     setOperationAction(ISD::FFLOOR,     VT, Expand);
757     setOperationAction(ISD::FNEARBYINT, VT, Expand);
758     setOperationAction(ISD::FCEIL,      VT, Expand);
759     setOperationAction(ISD::FRINT,      VT, Expand);
760     setOperationAction(ISD::FTRUNC,     VT, Expand);
761     setOperationAction(ISD::FROUND,     VT, Expand);
762     setOperationAction(ISD::FROUNDEVEN, VT, Expand);
763     setOperationAction(ISD::LROUND,     VT, Expand);
764     setOperationAction(ISD::LLROUND,    VT, Expand);
765     setOperationAction(ISD::LRINT,      VT, Expand);
766     setOperationAction(ISD::LLRINT,     VT, Expand);
767   }
768 
769   // Default ISD::TRAP to expand (which turns it into abort).
770   setOperationAction(ISD::TRAP, MVT::Other, Expand);
771 
772   // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
773   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
774   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
775 }
776 
777 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
778                                                EVT) const {
779   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
780 }
781 
782 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
783                                          bool LegalTypes) const {
784   assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
785   if (LHSTy.isVector())
786     return LHSTy;
787   return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
788                     : getPointerTy(DL);
789 }
790 
791 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
792   assert(isTypeLegal(VT));
793   switch (Op) {
794   default:
795     return false;
796   case ISD::SDIV:
797   case ISD::UDIV:
798   case ISD::SREM:
799   case ISD::UREM:
800     return true;
801   }
802 }
803 
804 bool TargetLoweringBase::isFreeAddrSpaceCast(unsigned SrcAS,
805                                              unsigned DestAS) const {
806   return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
807 }
808 
809 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
810   // If the command-line option was specified, ignore this request.
811   if (!JumpIsExpensiveOverride.getNumOccurrences())
812     JumpIsExpensive = isExpensive;
813 }
814 
815 TargetLoweringBase::LegalizeKind
816 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
817   // If this is a simple type, use the ComputeRegisterProp mechanism.
818   if (VT.isSimple()) {
819     MVT SVT = VT.getSimpleVT();
820     assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
821     MVT NVT = TransformToType[SVT.SimpleTy];
822     LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
823 
824     assert((LA == TypeLegal || LA == TypeSoftenFloat ||
825             LA == TypeSoftPromoteHalf ||
826             (NVT.isVector() ||
827              ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
828            "Promote may not follow Expand or Promote");
829 
830     if (LA == TypeSplitVector)
831       return LegalizeKind(LA,
832                           EVT::getVectorVT(Context, SVT.getVectorElementType(),
833                                            SVT.getVectorElementCount() / 2));
834     if (LA == TypeScalarizeVector)
835       return LegalizeKind(LA, SVT.getVectorElementType());
836     return LegalizeKind(LA, NVT);
837   }
838 
839   // Handle Extended Scalar Types.
840   if (!VT.isVector()) {
841     assert(VT.isInteger() && "Float types must be simple");
842     unsigned BitSize = VT.getSizeInBits();
843     // First promote to a power-of-two size, then expand if necessary.
844     if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
845       EVT NVT = VT.getRoundIntegerType(Context);
846       assert(NVT != VT && "Unable to round integer VT");
847       LegalizeKind NextStep = getTypeConversion(Context, NVT);
848       // Avoid multi-step promotion.
849       if (NextStep.first == TypePromoteInteger)
850         return NextStep;
851       // Return rounded integer type.
852       return LegalizeKind(TypePromoteInteger, NVT);
853     }
854 
855     return LegalizeKind(TypeExpandInteger,
856                         EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
857   }
858 
859   // Handle vector types.
860   ElementCount NumElts = VT.getVectorElementCount();
861   EVT EltVT = VT.getVectorElementType();
862 
863   // Vectors with only one element are always scalarized.
864   if (NumElts == 1)
865     return LegalizeKind(TypeScalarizeVector, EltVT);
866 
867   if (VT.getVectorElementCount() == ElementCount(1, true))
868     report_fatal_error("Cannot legalize this vector");
869 
870   // Try to widen vector elements until the element type is a power of two and
871   // promote it to a legal type later on, for example:
872   // <3 x i8> -> <4 x i8> -> <4 x i32>
873   if (EltVT.isInteger()) {
874     // Vectors with a number of elements that is not a power of two are always
875     // widened, for example <3 x i8> -> <4 x i8>.
876     if (!VT.isPow2VectorType()) {
877       NumElts = NumElts.NextPowerOf2();
878       EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
879       return LegalizeKind(TypeWidenVector, NVT);
880     }
881 
882     // Examine the element type.
883     LegalizeKind LK = getTypeConversion(Context, EltVT);
884 
885     // If type is to be expanded, split the vector.
886     //  <4 x i140> -> <2 x i140>
887     if (LK.first == TypeExpandInteger)
888       return LegalizeKind(TypeSplitVector,
889                           EVT::getVectorVT(Context, EltVT, NumElts / 2));
890 
891     // Promote the integer element types until a legal vector type is found
892     // or until the element integer type is too big. If a legal type was not
893     // found, fallback to the usual mechanism of widening/splitting the
894     // vector.
895     EVT OldEltVT = EltVT;
896     while (true) {
897       // Increase the bitwidth of the element to the next pow-of-two
898       // (which is greater than 8 bits).
899       EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
900                   .getRoundIntegerType(Context);
901 
902       // Stop trying when getting a non-simple element type.
903       // Note that vector elements may be greater than legal vector element
904       // types. Example: X86 XMM registers hold 64bit element on 32bit
905       // systems.
906       if (!EltVT.isSimple())
907         break;
908 
909       // Build a new vector type and check if it is legal.
910       MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
911       // Found a legal promoted vector type.
912       if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
913         return LegalizeKind(TypePromoteInteger,
914                             EVT::getVectorVT(Context, EltVT, NumElts));
915     }
916 
917     // Reset the type to the unexpanded type if we did not find a legal vector
918     // type with a promoted vector element type.
919     EltVT = OldEltVT;
920   }
921 
922   // Try to widen the vector until a legal type is found.
923   // If there is no wider legal type, split the vector.
924   while (true) {
925     // Round up to the next power of 2.
926     NumElts = NumElts.NextPowerOf2();
927 
928     // If there is no simple vector type with this many elements then there
929     // cannot be a larger legal vector type.  Note that this assumes that
930     // there are no skipped intermediate vector types in the simple types.
931     if (!EltVT.isSimple())
932       break;
933     MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
934     if (LargerVector == MVT())
935       break;
936 
937     // If this type is legal then widen the vector.
938     if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
939       return LegalizeKind(TypeWidenVector, LargerVector);
940   }
941 
942   // Widen odd vectors to next power of two.
943   if (!VT.isPow2VectorType()) {
944     EVT NVT = VT.getPow2VectorType(Context);
945     return LegalizeKind(TypeWidenVector, NVT);
946   }
947 
948   // Vectors with illegal element types are expanded.
949   EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorElementCount() / 2);
950   return LegalizeKind(TypeSplitVector, NVT);
951 }
952 
953 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
954                                           unsigned &NumIntermediates,
955                                           MVT &RegisterVT,
956                                           TargetLoweringBase *TLI) {
957   // Figure out the right, legal destination reg to copy into.
958   ElementCount EC = VT.getVectorElementCount();
959   MVT EltTy = VT.getVectorElementType();
960 
961   unsigned NumVectorRegs = 1;
962 
963   // Scalable vectors cannot be scalarized, so splitting or widening is
964   // required.
965   if (VT.isScalableVector() && !isPowerOf2_32(EC.Min))
966     llvm_unreachable(
967         "Splitting or widening of non-power-of-2 MVTs is not implemented.");
968 
969   // FIXME: We don't support non-power-of-2-sized vectors for now.
970   // Ideally we could break down into LHS/RHS like LegalizeDAG does.
971   if (!isPowerOf2_32(EC.Min)) {
972     // Split EC to unit size (scalable property is preserved).
973     NumVectorRegs = EC.Min;
974     EC = EC / NumVectorRegs;
975   }
976 
977   // Divide the input until we get to a supported size. This will
978   // always end up with an EC that represent a scalar or a scalable
979   // scalar.
980   while (EC.Min > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
981     EC.Min >>= 1;
982     NumVectorRegs <<= 1;
983   }
984 
985   NumIntermediates = NumVectorRegs;
986 
987   MVT NewVT = MVT::getVectorVT(EltTy, EC);
988   if (!TLI->isTypeLegal(NewVT))
989     NewVT = EltTy;
990   IntermediateVT = NewVT;
991 
992   unsigned LaneSizeInBits = NewVT.getScalarSizeInBits().getFixedSize();
993 
994   // Convert sizes such as i33 to i64.
995   if (!isPowerOf2_32(LaneSizeInBits))
996     LaneSizeInBits = NextPowerOf2(LaneSizeInBits);
997 
998   MVT DestVT = TLI->getRegisterType(NewVT);
999   RegisterVT = DestVT;
1000   if (EVT(DestVT).bitsLT(NewVT))    // Value is expanded, e.g. i64 -> i16.
1001     return NumVectorRegs *
1002            (LaneSizeInBits / DestVT.getScalarSizeInBits().getFixedSize());
1003 
1004   // Otherwise, promotion or legal types use the same number of registers as
1005   // the vector decimated to the appropriate level.
1006   return NumVectorRegs;
1007 }
1008 
1009 /// isLegalRC - Return true if the value types that can be represented by the
1010 /// specified register class are all legal.
1011 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
1012                                    const TargetRegisterClass &RC) const {
1013   for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1014     if (isTypeLegal(*I))
1015       return true;
1016   return false;
1017 }
1018 
1019 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1020 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1021 MachineBasicBlock *
1022 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1023                                    MachineBasicBlock *MBB) const {
1024   MachineInstr *MI = &InitialMI;
1025   MachineFunction &MF = *MI->getMF();
1026   MachineFrameInfo &MFI = MF.getFrameInfo();
1027 
1028   // We're handling multiple types of operands here:
1029   // PATCHPOINT MetaArgs - live-in, read only, direct
1030   // STATEPOINT Deopt Spill - live-through, read only, indirect
1031   // STATEPOINT Deopt Alloca - live-through, read only, direct
1032   // (We're currently conservative and mark the deopt slots read/write in
1033   // practice.)
1034   // STATEPOINT GC Spill - live-through, read/write, indirect
1035   // STATEPOINT GC Alloca - live-through, read/write, direct
1036   // The live-in vs live-through is handled already (the live through ones are
1037   // all stack slots), but we need to handle the different type of stackmap
1038   // operands and memory effects here.
1039 
1040   if (!llvm::any_of(MI->operands(),
1041                     [](MachineOperand &Operand) { return Operand.isFI(); }))
1042     return MBB;
1043 
1044   MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1045 
1046   // Inherit previous memory operands.
1047   MIB.cloneMemRefs(*MI);
1048 
1049   for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1050     MachineOperand &MO = MI->getOperand(i);
1051     if (!MO.isFI()) {
1052       // Index of Def operand this Use it tied to.
1053       // Since Defs are coming before Uses, if Use is tied, then
1054       // index of Def must be smaller that index of that Use.
1055       // Also, Defs preserve their position in new MI.
1056       unsigned TiedTo = i;
1057       if (MO.isReg() && MO.isTied())
1058         TiedTo = MI->findTiedOperandIdx(i);
1059       MIB.add(MO);
1060       if (TiedTo < i)
1061         MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1062       continue;
1063     }
1064 
1065     // foldMemoryOperand builds a new MI after replacing a single FI operand
1066     // with the canonical set of five x86 addressing-mode operands.
1067     int FI = MO.getIndex();
1068 
1069     // Add frame index operands recognized by stackmaps.cpp
1070     if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
1071       // indirect-mem-ref tag, size, #FI, offset.
1072       // Used for spills inserted by StatepointLowering.  This codepath is not
1073       // used for patchpoints/stackmaps at all, for these spilling is done via
1074       // foldMemoryOperand callback only.
1075       assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1076       MIB.addImm(StackMaps::IndirectMemRefOp);
1077       MIB.addImm(MFI.getObjectSize(FI));
1078       MIB.add(MO);
1079       MIB.addImm(0);
1080     } else {
1081       // direct-mem-ref tag, #FI, offset.
1082       // Used by patchpoint, and direct alloca arguments to statepoints
1083       MIB.addImm(StackMaps::DirectMemRefOp);
1084       MIB.add(MO);
1085       MIB.addImm(0);
1086     }
1087 
1088     assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1089 
1090     // Add a new memory operand for this FI.
1091     assert(MFI.getObjectOffset(FI) != -1);
1092 
1093     // Note: STATEPOINT MMOs are added during SelectionDAG.  STACKMAP, and
1094     // PATCHPOINT should be updated to do the same. (TODO)
1095     if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1096       auto Flags = MachineMemOperand::MOLoad;
1097       MachineMemOperand *MMO = MF.getMachineMemOperand(
1098           MachinePointerInfo::getFixedStack(MF, FI), Flags,
1099           MF.getDataLayout().getPointerSize(), MFI.getObjectAlign(FI));
1100       MIB->addMemOperand(MF, MMO);
1101     }
1102   }
1103   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1104   MI->eraseFromParent();
1105   return MBB;
1106 }
1107 
1108 MachineBasicBlock *
1109 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
1110                                         MachineBasicBlock *MBB) const {
1111   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1112          "Called emitXRayCustomEvent on the wrong MI!");
1113   auto &MF = *MI.getMF();
1114   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1115   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1116     MIB.add(MI.getOperand(OpIdx));
1117 
1118   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1119   MI.eraseFromParent();
1120   return MBB;
1121 }
1122 
1123 MachineBasicBlock *
1124 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1125                                        MachineBasicBlock *MBB) const {
1126   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1127          "Called emitXRayTypedEvent on the wrong MI!");
1128   auto &MF = *MI.getMF();
1129   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1130   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1131     MIB.add(MI.getOperand(OpIdx));
1132 
1133   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1134   MI.eraseFromParent();
1135   return MBB;
1136 }
1137 
1138 /// findRepresentativeClass - Return the largest legal super-reg register class
1139 /// of the register class for the specified type and its associated "cost".
1140 // This function is in TargetLowering because it uses RegClassForVT which would
1141 // need to be moved to TargetRegisterInfo and would necessitate moving
1142 // isTypeLegal over as well - a massive change that would just require
1143 // TargetLowering having a TargetRegisterInfo class member that it would use.
1144 std::pair<const TargetRegisterClass *, uint8_t>
1145 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1146                                             MVT VT) const {
1147   const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1148   if (!RC)
1149     return std::make_pair(RC, 0);
1150 
1151   // Compute the set of all super-register classes.
1152   BitVector SuperRegRC(TRI->getNumRegClasses());
1153   for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1154     SuperRegRC.setBitsInMask(RCI.getMask());
1155 
1156   // Find the first legal register class with the largest spill size.
1157   const TargetRegisterClass *BestRC = RC;
1158   for (unsigned i : SuperRegRC.set_bits()) {
1159     const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1160     // We want the largest possible spill size.
1161     if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1162       continue;
1163     if (!isLegalRC(*TRI, *SuperRC))
1164       continue;
1165     BestRC = SuperRC;
1166   }
1167   return std::make_pair(BestRC, 1);
1168 }
1169 
1170 /// computeRegisterProperties - Once all of the register classes are added,
1171 /// this allows us to compute derived properties we expose.
1172 void TargetLoweringBase::computeRegisterProperties(
1173     const TargetRegisterInfo *TRI) {
1174   static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1175                 "Too many value types for ValueTypeActions to hold!");
1176 
1177   // Everything defaults to needing one register.
1178   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1179     NumRegistersForVT[i] = 1;
1180     RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1181   }
1182   // ...except isVoid, which doesn't need any registers.
1183   NumRegistersForVT[MVT::isVoid] = 0;
1184 
1185   // Find the largest integer register class.
1186   unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1187   for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1188     assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1189 
1190   // Every integer value type larger than this largest register takes twice as
1191   // many registers to represent as the previous ValueType.
1192   for (unsigned ExpandedReg = LargestIntReg + 1;
1193        ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1194     NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1195     RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1196     TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1197     ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1198                                    TypeExpandInteger);
1199   }
1200 
1201   // Inspect all of the ValueType's smaller than the largest integer
1202   // register to see which ones need promotion.
1203   unsigned LegalIntReg = LargestIntReg;
1204   for (unsigned IntReg = LargestIntReg - 1;
1205        IntReg >= (unsigned)MVT::i1; --IntReg) {
1206     MVT IVT = (MVT::SimpleValueType)IntReg;
1207     if (isTypeLegal(IVT)) {
1208       LegalIntReg = IntReg;
1209     } else {
1210       RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1211         (MVT::SimpleValueType)LegalIntReg;
1212       ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1213     }
1214   }
1215 
1216   // ppcf128 type is really two f64's.
1217   if (!isTypeLegal(MVT::ppcf128)) {
1218     if (isTypeLegal(MVT::f64)) {
1219       NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1220       RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1221       TransformToType[MVT::ppcf128] = MVT::f64;
1222       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1223     } else {
1224       NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1225       RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1226       TransformToType[MVT::ppcf128] = MVT::i128;
1227       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1228     }
1229   }
1230 
1231   // Decide how to handle f128. If the target does not have native f128 support,
1232   // expand it to i128 and we will be generating soft float library calls.
1233   if (!isTypeLegal(MVT::f128)) {
1234     NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1235     RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1236     TransformToType[MVT::f128] = MVT::i128;
1237     ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1238   }
1239 
1240   // Decide how to handle f64. If the target does not have native f64 support,
1241   // expand it to i64 and we will be generating soft float library calls.
1242   if (!isTypeLegal(MVT::f64)) {
1243     NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1244     RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1245     TransformToType[MVT::f64] = MVT::i64;
1246     ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1247   }
1248 
1249   // Decide how to handle f32. If the target does not have native f32 support,
1250   // expand it to i32 and we will be generating soft float library calls.
1251   if (!isTypeLegal(MVT::f32)) {
1252     NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1253     RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1254     TransformToType[MVT::f32] = MVT::i32;
1255     ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1256   }
1257 
1258   // Decide how to handle f16. If the target does not have native f16 support,
1259   // promote it to f32, because there are no f16 library calls (except for
1260   // conversions).
1261   if (!isTypeLegal(MVT::f16)) {
1262     // Allow targets to control how we legalize half.
1263     if (softPromoteHalfType()) {
1264       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1265       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1266       TransformToType[MVT::f16] = MVT::f32;
1267       ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1268     } else {
1269       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1270       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1271       TransformToType[MVT::f16] = MVT::f32;
1272       ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1273     }
1274   }
1275 
1276   // Loop over all of the vector value types to see which need transformations.
1277   for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1278        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1279     MVT VT = (MVT::SimpleValueType) i;
1280     if (isTypeLegal(VT))
1281       continue;
1282 
1283     MVT EltVT = VT.getVectorElementType();
1284     ElementCount EC = VT.getVectorElementCount();
1285     bool IsLegalWiderType = false;
1286     bool IsScalable = VT.isScalableVector();
1287     LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1288     switch (PreferredAction) {
1289     case TypePromoteInteger: {
1290       MVT::SimpleValueType EndVT = IsScalable ?
1291                                    MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1292                                    MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1293       // Try to promote the elements of integer vectors. If no legal
1294       // promotion was found, fall through to the widen-vector method.
1295       for (unsigned nVT = i + 1;
1296            (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1297         MVT SVT = (MVT::SimpleValueType) nVT;
1298         // Promote vectors of integers to vectors with the same number
1299         // of elements, with a wider element type.
1300         if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() &&
1301             SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1302           TransformToType[i] = SVT;
1303           RegisterTypeForVT[i] = SVT;
1304           NumRegistersForVT[i] = 1;
1305           ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1306           IsLegalWiderType = true;
1307           break;
1308         }
1309       }
1310       if (IsLegalWiderType)
1311         break;
1312       LLVM_FALLTHROUGH;
1313     }
1314 
1315     case TypeWidenVector:
1316       if (isPowerOf2_32(EC.Min)) {
1317         // Try to widen the vector.
1318         for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1319           MVT SVT = (MVT::SimpleValueType) nVT;
1320           if (SVT.getVectorElementType() == EltVT &&
1321               SVT.isScalableVector() == IsScalable &&
1322               SVT.getVectorElementCount().Min > EC.Min && isTypeLegal(SVT)) {
1323             TransformToType[i] = SVT;
1324             RegisterTypeForVT[i] = SVT;
1325             NumRegistersForVT[i] = 1;
1326             ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1327             IsLegalWiderType = true;
1328             break;
1329           }
1330         }
1331         if (IsLegalWiderType)
1332           break;
1333       } else {
1334         // Only widen to the next power of 2 to keep consistency with EVT.
1335         MVT NVT = VT.getPow2VectorType();
1336         if (isTypeLegal(NVT)) {
1337           TransformToType[i] = NVT;
1338           ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1339           RegisterTypeForVT[i] = NVT;
1340           NumRegistersForVT[i] = 1;
1341           break;
1342         }
1343       }
1344       LLVM_FALLTHROUGH;
1345 
1346     case TypeSplitVector:
1347     case TypeScalarizeVector: {
1348       MVT IntermediateVT;
1349       MVT RegisterVT;
1350       unsigned NumIntermediates;
1351       unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1352           NumIntermediates, RegisterVT, this);
1353       NumRegistersForVT[i] = NumRegisters;
1354       assert(NumRegistersForVT[i] == NumRegisters &&
1355              "NumRegistersForVT size cannot represent NumRegisters!");
1356       RegisterTypeForVT[i] = RegisterVT;
1357 
1358       MVT NVT = VT.getPow2VectorType();
1359       if (NVT == VT) {
1360         // Type is already a power of 2.  The default action is to split.
1361         TransformToType[i] = MVT::Other;
1362         if (PreferredAction == TypeScalarizeVector)
1363           ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1364         else if (PreferredAction == TypeSplitVector)
1365           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1366         else if (EC.Min > 1)
1367           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1368         else
1369           ValueTypeActions.setTypeAction(VT, EC.Scalable
1370                                                  ? TypeScalarizeScalableVector
1371                                                  : TypeScalarizeVector);
1372       } else {
1373         TransformToType[i] = NVT;
1374         ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1375       }
1376       break;
1377     }
1378     default:
1379       llvm_unreachable("Unknown vector legalization action!");
1380     }
1381   }
1382 
1383   // Determine the 'representative' register class for each value type.
1384   // An representative register class is the largest (meaning one which is
1385   // not a sub-register class / subreg register class) legal register class for
1386   // a group of value types. For example, on i386, i8, i16, and i32
1387   // representative would be GR32; while on x86_64 it's GR64.
1388   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1389     const TargetRegisterClass* RRC;
1390     uint8_t Cost;
1391     std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1392     RepRegClassForVT[i] = RRC;
1393     RepRegClassCostForVT[i] = Cost;
1394   }
1395 }
1396 
1397 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1398                                            EVT VT) const {
1399   assert(!VT.isVector() && "No default SetCC type for vectors!");
1400   return getPointerTy(DL).SimpleTy;
1401 }
1402 
1403 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1404   return MVT::i32; // return the default value
1405 }
1406 
1407 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1408 /// legal first class types.  For example, MVT::v8f32 maps to 2 MVT::v4f32
1409 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1410 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1411 ///
1412 /// This method returns the number of registers needed, and the VT for each
1413 /// register.  It also returns the VT and quantity of the intermediate values
1414 /// before they are promoted/expanded.
1415 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1416                                                 EVT &IntermediateVT,
1417                                                 unsigned &NumIntermediates,
1418                                                 MVT &RegisterVT) const {
1419   ElementCount EltCnt = VT.getVectorElementCount();
1420 
1421   // If there is a wider vector type with the same element type as this one,
1422   // or a promoted vector type that has the same number of elements which
1423   // are wider, then we should convert to that legal vector type.
1424   // This handles things like <2 x float> -> <4 x float> and
1425   // <4 x i1> -> <4 x i32>.
1426   LegalizeTypeAction TA = getTypeAction(Context, VT);
1427   if (EltCnt.Min != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1428     EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1429     if (isTypeLegal(RegisterEVT)) {
1430       IntermediateVT = RegisterEVT;
1431       RegisterVT = RegisterEVT.getSimpleVT();
1432       NumIntermediates = 1;
1433       return 1;
1434     }
1435   }
1436 
1437   // Figure out the right, legal destination reg to copy into.
1438   EVT EltTy = VT.getVectorElementType();
1439 
1440   unsigned NumVectorRegs = 1;
1441 
1442   // Scalable vectors cannot be scalarized, so handle the legalisation of the
1443   // types like done elsewhere in SelectionDAG.
1444   if (VT.isScalableVector() && !isPowerOf2_32(EltCnt.Min)) {
1445     LegalizeKind LK;
1446     EVT PartVT = VT;
1447     do {
1448       // Iterate until we've found a legal (part) type to hold VT.
1449       LK = getTypeConversion(Context, PartVT);
1450       PartVT = LK.second;
1451     } while (LK.first != TypeLegal);
1452 
1453     NumIntermediates =
1454         VT.getVectorElementCount().Min / PartVT.getVectorElementCount().Min;
1455 
1456     // FIXME: This code needs to be extended to handle more complex vector
1457     // breakdowns, like nxv7i64 -> nxv8i64 -> 4 x nxv2i64. Currently the only
1458     // supported cases are vectors that are broken down into equal parts
1459     // such as nxv6i64 -> 3 x nxv2i64.
1460     assert(NumIntermediates * PartVT.getVectorElementCount().Min ==
1461                VT.getVectorElementCount().Min &&
1462            "Expected an integer multiple of PartVT");
1463     IntermediateVT = PartVT;
1464     RegisterVT = getRegisterType(Context, IntermediateVT);
1465     return NumIntermediates;
1466   }
1467 
1468   // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally
1469   // we could break down into LHS/RHS like LegalizeDAG does.
1470   if (!isPowerOf2_32(EltCnt.Min)) {
1471     NumVectorRegs = EltCnt.Min;
1472     EltCnt.Min = 1;
1473   }
1474 
1475   // Divide the input until we get to a supported size.  This will always
1476   // end with a scalar if the target doesn't support vectors.
1477   while (EltCnt.Min > 1 &&
1478          !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1479     EltCnt.Min >>= 1;
1480     NumVectorRegs <<= 1;
1481   }
1482 
1483   NumIntermediates = NumVectorRegs;
1484 
1485   EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1486   if (!isTypeLegal(NewVT))
1487     NewVT = EltTy;
1488   IntermediateVT = NewVT;
1489 
1490   MVT DestVT = getRegisterType(Context, NewVT);
1491   RegisterVT = DestVT;
1492 
1493   if (EVT(DestVT).bitsLT(NewVT)) {  // Value is expanded, e.g. i64 -> i16.
1494     TypeSize NewVTSize = NewVT.getSizeInBits();
1495     // Convert sizes such as i33 to i64.
1496     if (!isPowerOf2_32(NewVTSize.getKnownMinSize()))
1497       NewVTSize = NewVTSize.NextPowerOf2();
1498     return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1499   }
1500 
1501   // Otherwise, promotion or legal types use the same number of registers as
1502   // the vector decimated to the appropriate level.
1503   return NumVectorRegs;
1504 }
1505 
1506 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI,
1507                                                 uint64_t NumCases,
1508                                                 uint64_t Range,
1509                                                 ProfileSummaryInfo *PSI,
1510                                                 BlockFrequencyInfo *BFI) const {
1511   // FIXME: This function check the maximum table size and density, but the
1512   // minimum size is not checked. It would be nice if the minimum size is
1513   // also combined within this function. Currently, the minimum size check is
1514   // performed in findJumpTable() in SelectionDAGBuiler and
1515   // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1516   const bool OptForSize =
1517       SI->getParent()->getParent()->hasOptSize() ||
1518       llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
1519   const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1520   const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1521 
1522   // Check whether the number of cases is small enough and
1523   // the range is dense enough for a jump table.
1524   return (OptForSize || Range <= MaxJumpTableSize) &&
1525          (NumCases * 100 >= Range * MinDensity);
1526 }
1527 
1528 /// Get the EVTs and ArgFlags collections that represent the legalized return
1529 /// type of the given function.  This does not require a DAG or a return value,
1530 /// and is suitable for use before any DAGs for the function are constructed.
1531 /// TODO: Move this out of TargetLowering.cpp.
1532 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1533                          AttributeList attr,
1534                          SmallVectorImpl<ISD::OutputArg> &Outs,
1535                          const TargetLowering &TLI, const DataLayout &DL) {
1536   SmallVector<EVT, 4> ValueVTs;
1537   ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1538   unsigned NumValues = ValueVTs.size();
1539   if (NumValues == 0) return;
1540 
1541   for (unsigned j = 0, f = NumValues; j != f; ++j) {
1542     EVT VT = ValueVTs[j];
1543     ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1544 
1545     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1546       ExtendKind = ISD::SIGN_EXTEND;
1547     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1548       ExtendKind = ISD::ZERO_EXTEND;
1549 
1550     // FIXME: C calling convention requires the return type to be promoted to
1551     // at least 32-bit. But this is not necessary for non-C calling
1552     // conventions. The frontend should mark functions whose return values
1553     // require promoting with signext or zeroext attributes.
1554     if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1555       MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1556       if (VT.bitsLT(MinVT))
1557         VT = MinVT;
1558     }
1559 
1560     unsigned NumParts =
1561         TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1562     MVT PartVT =
1563         TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1564 
1565     // 'inreg' on function refers to return value
1566     ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1567     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1568       Flags.setInReg();
1569 
1570     // Propagate extension type if any
1571     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1572       Flags.setSExt();
1573     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1574       Flags.setZExt();
1575 
1576     for (unsigned i = 0; i < NumParts; ++i)
1577       Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0));
1578   }
1579 }
1580 
1581 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1582 /// function arguments in the caller parameter area.  This is the actual
1583 /// alignment, not its logarithm.
1584 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1585                                                    const DataLayout &DL) const {
1586   return DL.getABITypeAlign(Ty).value();
1587 }
1588 
1589 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1590     LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1591     Align Alignment, MachineMemOperand::Flags Flags, bool *Fast) const {
1592   // Check if the specified alignment is sufficient based on the data layout.
1593   // TODO: While using the data layout works in practice, a better solution
1594   // would be to implement this check directly (make this a virtual function).
1595   // For example, the ABI alignment may change based on software platform while
1596   // this function should only be affected by hardware implementation.
1597   Type *Ty = VT.getTypeForEVT(Context);
1598   if (Alignment >= DL.getABITypeAlign(Ty)) {
1599     // Assume that an access that meets the ABI-specified alignment is fast.
1600     if (Fast != nullptr)
1601       *Fast = true;
1602     return true;
1603   }
1604 
1605   // This is a misaligned access.
1606   return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment.value(), Flags,
1607                                         Fast);
1608 }
1609 
1610 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1611     LLVMContext &Context, const DataLayout &DL, EVT VT,
1612     const MachineMemOperand &MMO, bool *Fast) const {
1613   return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
1614                                         MMO.getAlign(), MMO.getFlags(), Fast);
1615 }
1616 
1617 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1618                                             const DataLayout &DL, EVT VT,
1619                                             unsigned AddrSpace, Align Alignment,
1620                                             MachineMemOperand::Flags Flags,
1621                                             bool *Fast) const {
1622   return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
1623                                         Flags, Fast);
1624 }
1625 
1626 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1627                                             const DataLayout &DL, EVT VT,
1628                                             const MachineMemOperand &MMO,
1629                                             bool *Fast) const {
1630   return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
1631                             MMO.getFlags(), Fast);
1632 }
1633 
1634 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1635   return BranchProbability(MinPercentageForPredictableBranch, 100);
1636 }
1637 
1638 //===----------------------------------------------------------------------===//
1639 //  TargetTransformInfo Helpers
1640 //===----------------------------------------------------------------------===//
1641 
1642 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1643   enum InstructionOpcodes {
1644 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1645 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1646 #include "llvm/IR/Instruction.def"
1647   };
1648   switch (static_cast<InstructionOpcodes>(Opcode)) {
1649   case Ret:            return 0;
1650   case Br:             return 0;
1651   case Switch:         return 0;
1652   case IndirectBr:     return 0;
1653   case Invoke:         return 0;
1654   case CallBr:         return 0;
1655   case Resume:         return 0;
1656   case Unreachable:    return 0;
1657   case CleanupRet:     return 0;
1658   case CatchRet:       return 0;
1659   case CatchPad:       return 0;
1660   case CatchSwitch:    return 0;
1661   case CleanupPad:     return 0;
1662   case FNeg:           return ISD::FNEG;
1663   case Add:            return ISD::ADD;
1664   case FAdd:           return ISD::FADD;
1665   case Sub:            return ISD::SUB;
1666   case FSub:           return ISD::FSUB;
1667   case Mul:            return ISD::MUL;
1668   case FMul:           return ISD::FMUL;
1669   case UDiv:           return ISD::UDIV;
1670   case SDiv:           return ISD::SDIV;
1671   case FDiv:           return ISD::FDIV;
1672   case URem:           return ISD::UREM;
1673   case SRem:           return ISD::SREM;
1674   case FRem:           return ISD::FREM;
1675   case Shl:            return ISD::SHL;
1676   case LShr:           return ISD::SRL;
1677   case AShr:           return ISD::SRA;
1678   case And:            return ISD::AND;
1679   case Or:             return ISD::OR;
1680   case Xor:            return ISD::XOR;
1681   case Alloca:         return 0;
1682   case Load:           return ISD::LOAD;
1683   case Store:          return ISD::STORE;
1684   case GetElementPtr:  return 0;
1685   case Fence:          return 0;
1686   case AtomicCmpXchg:  return 0;
1687   case AtomicRMW:      return 0;
1688   case Trunc:          return ISD::TRUNCATE;
1689   case ZExt:           return ISD::ZERO_EXTEND;
1690   case SExt:           return ISD::SIGN_EXTEND;
1691   case FPToUI:         return ISD::FP_TO_UINT;
1692   case FPToSI:         return ISD::FP_TO_SINT;
1693   case UIToFP:         return ISD::UINT_TO_FP;
1694   case SIToFP:         return ISD::SINT_TO_FP;
1695   case FPTrunc:        return ISD::FP_ROUND;
1696   case FPExt:          return ISD::FP_EXTEND;
1697   case PtrToInt:       return ISD::BITCAST;
1698   case IntToPtr:       return ISD::BITCAST;
1699   case BitCast:        return ISD::BITCAST;
1700   case AddrSpaceCast:  return ISD::ADDRSPACECAST;
1701   case ICmp:           return ISD::SETCC;
1702   case FCmp:           return ISD::SETCC;
1703   case PHI:            return 0;
1704   case Call:           return 0;
1705   case Select:         return ISD::SELECT;
1706   case UserOp1:        return 0;
1707   case UserOp2:        return 0;
1708   case VAArg:          return 0;
1709   case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1710   case InsertElement:  return ISD::INSERT_VECTOR_ELT;
1711   case ShuffleVector:  return ISD::VECTOR_SHUFFLE;
1712   case ExtractValue:   return ISD::MERGE_VALUES;
1713   case InsertValue:    return ISD::MERGE_VALUES;
1714   case LandingPad:     return 0;
1715   case Freeze:         return ISD::FREEZE;
1716   }
1717 
1718   llvm_unreachable("Unknown instruction type encountered!");
1719 }
1720 
1721 std::pair<int, MVT>
1722 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1723                                             Type *Ty) const {
1724   LLVMContext &C = Ty->getContext();
1725   EVT MTy = getValueType(DL, Ty);
1726 
1727   int Cost = 1;
1728   // We keep legalizing the type until we find a legal kind. We assume that
1729   // the only operation that costs anything is the split. After splitting
1730   // we need to handle two types.
1731   while (true) {
1732     LegalizeKind LK = getTypeConversion(C, MTy);
1733 
1734     if (LK.first == TypeLegal)
1735       return std::make_pair(Cost, MTy.getSimpleVT());
1736 
1737     if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1738       Cost *= 2;
1739 
1740     // Do not loop with f128 type.
1741     if (MTy == LK.second)
1742       return std::make_pair(Cost, MTy.getSimpleVT());
1743 
1744     // Keep legalizing the type.
1745     MTy = LK.second;
1746   }
1747 }
1748 
1749 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1750                                                               bool UseTLS) const {
1751   // compiler-rt provides a variable with a magic name.  Targets that do not
1752   // link with compiler-rt may also provide such a variable.
1753   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1754   const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1755   auto UnsafeStackPtr =
1756       dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1757 
1758   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1759 
1760   if (!UnsafeStackPtr) {
1761     auto TLSModel = UseTLS ?
1762         GlobalValue::InitialExecTLSModel :
1763         GlobalValue::NotThreadLocal;
1764     // The global variable is not defined yet, define it ourselves.
1765     // We use the initial-exec TLS model because we do not support the
1766     // variable living anywhere other than in the main executable.
1767     UnsafeStackPtr = new GlobalVariable(
1768         *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1769         UnsafeStackPtrVar, nullptr, TLSModel);
1770   } else {
1771     // The variable exists, check its type and attributes.
1772     if (UnsafeStackPtr->getValueType() != StackPtrTy)
1773       report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1774     if (UseTLS != UnsafeStackPtr->isThreadLocal())
1775       report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1776                          (UseTLS ? "" : "not ") + "be thread-local");
1777   }
1778   return UnsafeStackPtr;
1779 }
1780 
1781 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1782   if (!TM.getTargetTriple().isAndroid())
1783     return getDefaultSafeStackPointerLocation(IRB, true);
1784 
1785   // Android provides a libc function to retrieve the address of the current
1786   // thread's unsafe stack pointer.
1787   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1788   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1789   FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address",
1790                                              StackPtrTy->getPointerTo(0));
1791   return IRB.CreateCall(Fn);
1792 }
1793 
1794 //===----------------------------------------------------------------------===//
1795 //  Loop Strength Reduction hooks
1796 //===----------------------------------------------------------------------===//
1797 
1798 /// isLegalAddressingMode - Return true if the addressing mode represented
1799 /// by AM is legal for this target, for a load/store of the specified type.
1800 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1801                                                const AddrMode &AM, Type *Ty,
1802                                                unsigned AS, Instruction *I) const {
1803   // The default implementation of this implements a conservative RISCy, r+r and
1804   // r+i addr mode.
1805 
1806   // Allows a sign-extended 16-bit immediate field.
1807   if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1808     return false;
1809 
1810   // No global is ever allowed as a base.
1811   if (AM.BaseGV)
1812     return false;
1813 
1814   // Only support r+r,
1815   switch (AM.Scale) {
1816   case 0:  // "r+i" or just "i", depending on HasBaseReg.
1817     break;
1818   case 1:
1819     if (AM.HasBaseReg && AM.BaseOffs)  // "r+r+i" is not allowed.
1820       return false;
1821     // Otherwise we have r+r or r+i.
1822     break;
1823   case 2:
1824     if (AM.HasBaseReg || AM.BaseOffs)  // 2*r+r  or  2*r+i is not allowed.
1825       return false;
1826     // Allow 2*r as r+r.
1827     break;
1828   default: // Don't allow n * r
1829     return false;
1830   }
1831 
1832   return true;
1833 }
1834 
1835 //===----------------------------------------------------------------------===//
1836 //  Stack Protector
1837 //===----------------------------------------------------------------------===//
1838 
1839 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1840 // so that SelectionDAG handle SSP.
1841 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1842   if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1843     Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1844     PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1845     return M.getOrInsertGlobal("__guard_local", PtrTy);
1846   }
1847   return nullptr;
1848 }
1849 
1850 // Currently only support "standard" __stack_chk_guard.
1851 // TODO: add LOAD_STACK_GUARD support.
1852 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1853   if (!M.getNamedValue("__stack_chk_guard"))
1854     new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1855                        GlobalVariable::ExternalLinkage,
1856                        nullptr, "__stack_chk_guard");
1857 }
1858 
1859 // Currently only support "standard" __stack_chk_guard.
1860 // TODO: add LOAD_STACK_GUARD support.
1861 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1862   return M.getNamedValue("__stack_chk_guard");
1863 }
1864 
1865 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1866   return nullptr;
1867 }
1868 
1869 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1870   return MinimumJumpTableEntries;
1871 }
1872 
1873 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1874   MinimumJumpTableEntries = Val;
1875 }
1876 
1877 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1878   return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1879 }
1880 
1881 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1882   return MaximumJumpTableSize;
1883 }
1884 
1885 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1886   MaximumJumpTableSize = Val;
1887 }
1888 
1889 bool TargetLoweringBase::isJumpTableRelative() const {
1890   return getTargetMachine().isPositionIndependent();
1891 }
1892 
1893 //===----------------------------------------------------------------------===//
1894 //  Reciprocal Estimates
1895 //===----------------------------------------------------------------------===//
1896 
1897 /// Get the reciprocal estimate attribute string for a function that will
1898 /// override the target defaults.
1899 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1900   const Function &F = MF.getFunction();
1901   return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1902 }
1903 
1904 /// Construct a string for the given reciprocal operation of the given type.
1905 /// This string should match the corresponding option to the front-end's
1906 /// "-mrecip" flag assuming those strings have been passed through in an
1907 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1908 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1909   std::string Name = VT.isVector() ? "vec-" : "";
1910 
1911   Name += IsSqrt ? "sqrt" : "div";
1912 
1913   // TODO: Handle "half" or other float types?
1914   if (VT.getScalarType() == MVT::f64) {
1915     Name += "d";
1916   } else {
1917     assert(VT.getScalarType() == MVT::f32 &&
1918            "Unexpected FP type for reciprocal estimate");
1919     Name += "f";
1920   }
1921 
1922   return Name;
1923 }
1924 
1925 /// Return the character position and value (a single numeric character) of a
1926 /// customized refinement operation in the input string if it exists. Return
1927 /// false if there is no customized refinement step count.
1928 static bool parseRefinementStep(StringRef In, size_t &Position,
1929                                 uint8_t &Value) {
1930   const char RefStepToken = ':';
1931   Position = In.find(RefStepToken);
1932   if (Position == StringRef::npos)
1933     return false;
1934 
1935   StringRef RefStepString = In.substr(Position + 1);
1936   // Allow exactly one numeric character for the additional refinement
1937   // step parameter.
1938   if (RefStepString.size() == 1) {
1939     char RefStepChar = RefStepString[0];
1940     if (RefStepChar >= '0' && RefStepChar <= '9') {
1941       Value = RefStepChar - '0';
1942       return true;
1943     }
1944   }
1945   report_fatal_error("Invalid refinement step for -recip.");
1946 }
1947 
1948 /// For the input attribute string, return one of the ReciprocalEstimate enum
1949 /// status values (enabled, disabled, or not specified) for this operation on
1950 /// the specified data type.
1951 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1952   if (Override.empty())
1953     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1954 
1955   SmallVector<StringRef, 4> OverrideVector;
1956   Override.split(OverrideVector, ',');
1957   unsigned NumArgs = OverrideVector.size();
1958 
1959   // Check if "all", "none", or "default" was specified.
1960   if (NumArgs == 1) {
1961     // Look for an optional setting of the number of refinement steps needed
1962     // for this type of reciprocal operation.
1963     size_t RefPos;
1964     uint8_t RefSteps;
1965     if (parseRefinementStep(Override, RefPos, RefSteps)) {
1966       // Split the string for further processing.
1967       Override = Override.substr(0, RefPos);
1968     }
1969 
1970     // All reciprocal types are enabled.
1971     if (Override == "all")
1972       return TargetLoweringBase::ReciprocalEstimate::Enabled;
1973 
1974     // All reciprocal types are disabled.
1975     if (Override == "none")
1976       return TargetLoweringBase::ReciprocalEstimate::Disabled;
1977 
1978     // Target defaults for enablement are used.
1979     if (Override == "default")
1980       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1981   }
1982 
1983   // The attribute string may omit the size suffix ('f'/'d').
1984   std::string VTName = getReciprocalOpName(IsSqrt, VT);
1985   std::string VTNameNoSize = VTName;
1986   VTNameNoSize.pop_back();
1987   static const char DisabledPrefix = '!';
1988 
1989   for (StringRef RecipType : OverrideVector) {
1990     size_t RefPos;
1991     uint8_t RefSteps;
1992     if (parseRefinementStep(RecipType, RefPos, RefSteps))
1993       RecipType = RecipType.substr(0, RefPos);
1994 
1995     // Ignore the disablement token for string matching.
1996     bool IsDisabled = RecipType[0] == DisabledPrefix;
1997     if (IsDisabled)
1998       RecipType = RecipType.substr(1);
1999 
2000     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2001       return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
2002                         : TargetLoweringBase::ReciprocalEstimate::Enabled;
2003   }
2004 
2005   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2006 }
2007 
2008 /// For the input attribute string, return the customized refinement step count
2009 /// for this operation on the specified data type. If the step count does not
2010 /// exist, return the ReciprocalEstimate enum value for unspecified.
2011 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2012   if (Override.empty())
2013     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2014 
2015   SmallVector<StringRef, 4> OverrideVector;
2016   Override.split(OverrideVector, ',');
2017   unsigned NumArgs = OverrideVector.size();
2018 
2019   // Check if "all", "default", or "none" was specified.
2020   if (NumArgs == 1) {
2021     // Look for an optional setting of the number of refinement steps needed
2022     // for this type of reciprocal operation.
2023     size_t RefPos;
2024     uint8_t RefSteps;
2025     if (!parseRefinementStep(Override, RefPos, RefSteps))
2026       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2027 
2028     // Split the string for further processing.
2029     Override = Override.substr(0, RefPos);
2030     assert(Override != "none" &&
2031            "Disabled reciprocals, but specifed refinement steps?");
2032 
2033     // If this is a general override, return the specified number of steps.
2034     if (Override == "all" || Override == "default")
2035       return RefSteps;
2036   }
2037 
2038   // The attribute string may omit the size suffix ('f'/'d').
2039   std::string VTName = getReciprocalOpName(IsSqrt, VT);
2040   std::string VTNameNoSize = VTName;
2041   VTNameNoSize.pop_back();
2042 
2043   for (StringRef RecipType : OverrideVector) {
2044     size_t RefPos;
2045     uint8_t RefSteps;
2046     if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2047       continue;
2048 
2049     RecipType = RecipType.substr(0, RefPos);
2050     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2051       return RefSteps;
2052   }
2053 
2054   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2055 }
2056 
2057 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
2058                                                     MachineFunction &MF) const {
2059   return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
2060 }
2061 
2062 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
2063                                                    MachineFunction &MF) const {
2064   return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
2065 }
2066 
2067 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
2068                                                MachineFunction &MF) const {
2069   return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
2070 }
2071 
2072 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
2073                                               MachineFunction &MF) const {
2074   return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
2075 }
2076 
2077 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
2078   MF.getRegInfo().freezeReservedRegs(MF);
2079 }
2080 
2081 MachineMemOperand::Flags
2082 TargetLoweringBase::getLoadMemOperandFlags(const LoadInst &LI,
2083                                            const DataLayout &DL) const {
2084   MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad;
2085   if (LI.isVolatile())
2086     Flags |= MachineMemOperand::MOVolatile;
2087 
2088   if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2089     Flags |= MachineMemOperand::MONonTemporal;
2090 
2091   if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2092     Flags |= MachineMemOperand::MOInvariant;
2093 
2094   if (isDereferenceablePointer(LI.getPointerOperand(), LI.getType(), DL))
2095     Flags |= MachineMemOperand::MODereferenceable;
2096 
2097   Flags |= getTargetMMOFlags(LI);
2098   return Flags;
2099 }
2100 
2101 MachineMemOperand::Flags
2102 TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI,
2103                                             const DataLayout &DL) const {
2104   MachineMemOperand::Flags Flags = MachineMemOperand::MOStore;
2105 
2106   if (SI.isVolatile())
2107     Flags |= MachineMemOperand::MOVolatile;
2108 
2109   if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2110     Flags |= MachineMemOperand::MONonTemporal;
2111 
2112   // FIXME: Not preserving dereferenceable
2113   Flags |= getTargetMMOFlags(SI);
2114   return Flags;
2115 }
2116 
2117 MachineMemOperand::Flags
2118 TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI,
2119                                              const DataLayout &DL) const {
2120   auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
2121 
2122   if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2123     if (RMW->isVolatile())
2124       Flags |= MachineMemOperand::MOVolatile;
2125   } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2126     if (CmpX->isVolatile())
2127       Flags |= MachineMemOperand::MOVolatile;
2128   } else
2129     llvm_unreachable("not an atomic instruction");
2130 
2131   // FIXME: Not preserving dereferenceable
2132   Flags |= getTargetMMOFlags(AI);
2133   return Flags;
2134 }
2135 
2136 //===----------------------------------------------------------------------===//
2137 //  GlobalISel Hooks
2138 //===----------------------------------------------------------------------===//
2139 
2140 bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI,
2141                                         const TargetTransformInfo *TTI) const {
2142   auto &MF = *MI.getMF();
2143   auto &MRI = MF.getRegInfo();
2144   // Assuming a spill and reload of a value has a cost of 1 instruction each,
2145   // this helper function computes the maximum number of uses we should consider
2146   // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2147   // break even in terms of code size when the original MI has 2 users vs
2148   // choosing to potentially spill. Any more than 2 users we we have a net code
2149   // size increase. This doesn't take into account register pressure though.
2150   auto maxUses = [](unsigned RematCost) {
2151     // A cost of 1 means remats are basically free.
2152     if (RematCost == 1)
2153       return UINT_MAX;
2154     if (RematCost == 2)
2155       return 2U;
2156 
2157     // Remat is too expensive, only sink if there's one user.
2158     if (RematCost > 2)
2159       return 1U;
2160     llvm_unreachable("Unexpected remat cost");
2161   };
2162 
2163   // Helper to walk through uses and terminate if we've reached a limit. Saves
2164   // us spending time traversing uses if all we want to know is if it's >= min.
2165   auto isUsesAtMost = [&](unsigned Reg, unsigned MaxUses) {
2166     unsigned NumUses = 0;
2167     auto UI = MRI.use_instr_nodbg_begin(Reg), UE = MRI.use_instr_nodbg_end();
2168     for (; UI != UE && NumUses < MaxUses; ++UI) {
2169       NumUses++;
2170     }
2171     // If we haven't reached the end yet then there are more than MaxUses users.
2172     return UI == UE;
2173   };
2174 
2175   switch (MI.getOpcode()) {
2176   default:
2177     return false;
2178   // Constants-like instructions should be close to their users.
2179   // We don't want long live-ranges for them.
2180   case TargetOpcode::G_CONSTANT:
2181   case TargetOpcode::G_FCONSTANT:
2182   case TargetOpcode::G_FRAME_INDEX:
2183   case TargetOpcode::G_INTTOPTR:
2184     return true;
2185   case TargetOpcode::G_GLOBAL_VALUE: {
2186     unsigned RematCost = TTI->getGISelRematGlobalCost();
2187     Register Reg = MI.getOperand(0).getReg();
2188     unsigned MaxUses = maxUses(RematCost);
2189     if (MaxUses == UINT_MAX)
2190       return true; // Remats are "free" so always localize.
2191     bool B = isUsesAtMost(Reg, MaxUses);
2192     return B;
2193   }
2194   }
2195 }
2196