Bug Summary

File:include/llvm/MC/LaneBitmask.h
Warning:line 86, column 34
The result of the left shift is undefined due to shifting by '4294967295', which is greater or equal to the width of type 'llvm::LaneBitmask::Type'

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CodeGenRegisters.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/utils/TableGen -I /build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn345461/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/utils/TableGen -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-10-27-211344-32123-1 -x c++ /build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp -faddrsig

/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp

1//===- CodeGenRegisters.cpp - Register and RegisterClass Info -------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines structures to encapsulate information gleaned from the
11// target register and register class definitions.
12//
13//===----------------------------------------------------------------------===//
14
15#include "CodeGenRegisters.h"
16#include "CodeGenTarget.h"
17#include "llvm/ADT/ArrayRef.h"
18#include "llvm/ADT/BitVector.h"
19#include "llvm/ADT/DenseMap.h"
20#include "llvm/ADT/IntEqClasses.h"
21#include "llvm/ADT/SetVector.h"
22#include "llvm/ADT/SmallPtrSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/STLExtras.h"
25#include "llvm/ADT/StringExtras.h"
26#include "llvm/ADT/StringRef.h"
27#include "llvm/ADT/Twine.h"
28#include "llvm/Support/Debug.h"
29#include "llvm/Support/MathExtras.h"
30#include "llvm/Support/raw_ostream.h"
31#include "llvm/TableGen/Error.h"
32#include "llvm/TableGen/Record.h"
33#include <algorithm>
34#include <cassert>
35#include <cstdint>
36#include <iterator>
37#include <map>
38#include <queue>
39#include <set>
40#include <string>
41#include <tuple>
42#include <utility>
43#include <vector>
44
45using namespace llvm;
46
47#define DEBUG_TYPE"regalloc-emitter" "regalloc-emitter"
48
49//===----------------------------------------------------------------------===//
50// CodeGenSubRegIndex
51//===----------------------------------------------------------------------===//
52
53CodeGenSubRegIndex::CodeGenSubRegIndex(Record *R, unsigned Enum)
54 : TheDef(R), EnumValue(Enum), AllSuperRegsCovered(true), Artificial(true) {
55 Name = R->getName();
56 if (R->getValue("Namespace"))
57 Namespace = R->getValueAsString("Namespace");
58 Size = R->getValueAsInt("Size");
59 Offset = R->getValueAsInt("Offset");
60}
61
62CodeGenSubRegIndex::CodeGenSubRegIndex(StringRef N, StringRef Nspace,
63 unsigned Enum)
64 : TheDef(nullptr), Name(N), Namespace(Nspace), Size(-1), Offset(-1),
65 EnumValue(Enum), AllSuperRegsCovered(true), Artificial(true) {
66}
67
68std::string CodeGenSubRegIndex::getQualifiedName() const {
69 std::string N = getNamespace();
70 if (!N.empty())
71 N += "::";
72 N += getName();
73 return N;
74}
75
76void CodeGenSubRegIndex::updateComponents(CodeGenRegBank &RegBank) {
77 if (!TheDef)
78 return;
79
80 std::vector<Record*> Comps = TheDef->getValueAsListOfDefs("ComposedOf");
81 if (!Comps.empty()) {
82 if (Comps.size() != 2)
83 PrintFatalError(TheDef->getLoc(),
84 "ComposedOf must have exactly two entries");
85 CodeGenSubRegIndex *A = RegBank.getSubRegIdx(Comps[0]);
86 CodeGenSubRegIndex *B = RegBank.getSubRegIdx(Comps[1]);
87 CodeGenSubRegIndex *X = A->addComposite(B, this);
88 if (X)
89 PrintFatalError(TheDef->getLoc(), "Ambiguous ComposedOf entries");
90 }
91
92 std::vector<Record*> Parts =
93 TheDef->getValueAsListOfDefs("CoveringSubRegIndices");
94 if (!Parts.empty()) {
95 if (Parts.size() < 2)
96 PrintFatalError(TheDef->getLoc(),
97 "CoveredBySubRegs must have two or more entries");
98 SmallVector<CodeGenSubRegIndex*, 8> IdxParts;
99 for (Record *Part : Parts)
100 IdxParts.push_back(RegBank.getSubRegIdx(Part));
101 setConcatenationOf(IdxParts);
102 }
103}
104
105LaneBitmask CodeGenSubRegIndex::computeLaneMask() const {
106 // Already computed?
107 if (LaneMask.any())
108 return LaneMask;
109
110 // Recursion guard, shouldn't be required.
111 LaneMask = LaneBitmask::getAll();
112
113 // The lane mask is simply the union of all sub-indices.
114 LaneBitmask M;
115 for (const auto &C : Composed)
116 M |= C.second->computeLaneMask();
117 assert(M.any() && "Missing lane mask, sub-register cycle?")((M.any() && "Missing lane mask, sub-register cycle?"
) ? static_cast<void> (0) : __assert_fail ("M.any() && \"Missing lane mask, sub-register cycle?\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 117, __PRETTY_FUNCTION__))
;
118 LaneMask = M;
119 return LaneMask;
120}
121
122void CodeGenSubRegIndex::setConcatenationOf(
123 ArrayRef<CodeGenSubRegIndex*> Parts) {
124 if (ConcatenationOf.empty())
125 ConcatenationOf.assign(Parts.begin(), Parts.end());
126 else
127 assert(std::equal(Parts.begin(), Parts.end(),((std::equal(Parts.begin(), Parts.end(), ConcatenationOf.begin
()) && "parts consistent") ? static_cast<void> (
0) : __assert_fail ("std::equal(Parts.begin(), Parts.end(), ConcatenationOf.begin()) && \"parts consistent\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 128, __PRETTY_FUNCTION__))
128 ConcatenationOf.begin()) && "parts consistent")((std::equal(Parts.begin(), Parts.end(), ConcatenationOf.begin
()) && "parts consistent") ? static_cast<void> (
0) : __assert_fail ("std::equal(Parts.begin(), Parts.end(), ConcatenationOf.begin()) && \"parts consistent\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 128, __PRETTY_FUNCTION__))
;
129}
130
131void CodeGenSubRegIndex::computeConcatTransitiveClosure() {
132 for (SmallVectorImpl<CodeGenSubRegIndex*>::iterator
133 I = ConcatenationOf.begin(); I != ConcatenationOf.end(); /*empty*/) {
134 CodeGenSubRegIndex *SubIdx = *I;
135 SubIdx->computeConcatTransitiveClosure();
136#ifndef NDEBUG
137 for (CodeGenSubRegIndex *SRI : SubIdx->ConcatenationOf)
138 assert(SRI->ConcatenationOf.empty() && "No transitive closure?")((SRI->ConcatenationOf.empty() && "No transitive closure?"
) ? static_cast<void> (0) : __assert_fail ("SRI->ConcatenationOf.empty() && \"No transitive closure?\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 138, __PRETTY_FUNCTION__))
;
139#endif
140
141 if (SubIdx->ConcatenationOf.empty()) {
142 ++I;
143 } else {
144 I = ConcatenationOf.erase(I);
145 I = ConcatenationOf.insert(I, SubIdx->ConcatenationOf.begin(),
146 SubIdx->ConcatenationOf.end());
147 I += SubIdx->ConcatenationOf.size();
148 }
149 }
150}
151
152//===----------------------------------------------------------------------===//
153// CodeGenRegister
154//===----------------------------------------------------------------------===//
155
156CodeGenRegister::CodeGenRegister(Record *R, unsigned Enum)
157 : TheDef(R),
158 EnumValue(Enum),
159 CostPerUse(R->getValueAsInt("CostPerUse")),
160 CoveredBySubRegs(R->getValueAsBit("CoveredBySubRegs")),
161 HasDisjunctSubRegs(false),
162 SubRegsComplete(false),
163 SuperRegsComplete(false),
164 TopoSig(~0u) {
165 Artificial = R->getValueAsBit("isArtificial");
166}
167
168void CodeGenRegister::buildObjectGraph(CodeGenRegBank &RegBank) {
169 std::vector<Record*> SRIs = TheDef->getValueAsListOfDefs("SubRegIndices");
170 std::vector<Record*> SRs = TheDef->getValueAsListOfDefs("SubRegs");
171
172 if (SRIs.size() != SRs.size())
173 PrintFatalError(TheDef->getLoc(),
174 "SubRegs and SubRegIndices must have the same size");
175
176 for (unsigned i = 0, e = SRIs.size(); i != e; ++i) {
177 ExplicitSubRegIndices.push_back(RegBank.getSubRegIdx(SRIs[i]));
178 ExplicitSubRegs.push_back(RegBank.getReg(SRs[i]));
179 }
180
181 // Also compute leading super-registers. Each register has a list of
182 // covered-by-subregs super-registers where it appears as the first explicit
183 // sub-register.
184 //
185 // This is used by computeSecondarySubRegs() to find candidates.
186 if (CoveredBySubRegs && !ExplicitSubRegs.empty())
187 ExplicitSubRegs.front()->LeadingSuperRegs.push_back(this);
188
189 // Add ad hoc alias links. This is a symmetric relationship between two
190 // registers, so build a symmetric graph by adding links in both ends.
191 std::vector<Record*> Aliases = TheDef->getValueAsListOfDefs("Aliases");
192 for (Record *Alias : Aliases) {
193 CodeGenRegister *Reg = RegBank.getReg(Alias);
194 ExplicitAliases.push_back(Reg);
195 Reg->ExplicitAliases.push_back(this);
196 }
197}
198
199const StringRef CodeGenRegister::getName() const {
200 assert(TheDef && "no def")((TheDef && "no def") ? static_cast<void> (0) :
__assert_fail ("TheDef && \"no def\"", "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 200, __PRETTY_FUNCTION__))
;
201 return TheDef->getName();
202}
203
204namespace {
205
206// Iterate over all register units in a set of registers.
207class RegUnitIterator {
208 CodeGenRegister::Vec::const_iterator RegI, RegE;
209 CodeGenRegister::RegUnitList::iterator UnitI, UnitE;
210
211public:
212 RegUnitIterator(const CodeGenRegister::Vec &Regs):
213 RegI(Regs.begin()), RegE(Regs.end()) {
214
215 if (RegI != RegE) {
216 UnitI = (*RegI)->getRegUnits().begin();
217 UnitE = (*RegI)->getRegUnits().end();
218 advance();
219 }
220 }
221
222 bool isValid() const { return UnitI != UnitE; }
223
224 unsigned operator* () const { assert(isValid())((isValid()) ? static_cast<void> (0) : __assert_fail ("isValid()"
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 224, __PRETTY_FUNCTION__))
; return *UnitI; }
225
226 const CodeGenRegister *getReg() const { assert(isValid())((isValid()) ? static_cast<void> (0) : __assert_fail ("isValid()"
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 226, __PRETTY_FUNCTION__))
; return *RegI; }
227
228 /// Preincrement. Move to the next unit.
229 void operator++() {
230 assert(isValid() && "Cannot advance beyond the last operand")((isValid() && "Cannot advance beyond the last operand"
) ? static_cast<void> (0) : __assert_fail ("isValid() && \"Cannot advance beyond the last operand\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 230, __PRETTY_FUNCTION__))
;
231 ++UnitI;
232 advance();
233 }
234
235protected:
236 void advance() {
237 while (UnitI == UnitE) {
238 if (++RegI == RegE)
239 break;
240 UnitI = (*RegI)->getRegUnits().begin();
241 UnitE = (*RegI)->getRegUnits().end();
242 }
243 }
244};
245
246} // end anonymous namespace
247
248// Return true of this unit appears in RegUnits.
249static bool hasRegUnit(CodeGenRegister::RegUnitList &RegUnits, unsigned Unit) {
250 return RegUnits.test(Unit);
251}
252
253// Inherit register units from subregisters.
254// Return true if the RegUnits changed.
255bool CodeGenRegister::inheritRegUnits(CodeGenRegBank &RegBank) {
256 bool changed = false;
257 for (const auto &SubReg : SubRegs) {
258 CodeGenRegister *SR = SubReg.second;
259 // Merge the subregister's units into this register's RegUnits.
260 changed |= (RegUnits |= SR->RegUnits);
261 }
262
263 return changed;
264}
265
266const CodeGenRegister::SubRegMap &
267CodeGenRegister::computeSubRegs(CodeGenRegBank &RegBank) {
268 // Only compute this map once.
269 if (SubRegsComplete)
270 return SubRegs;
271 SubRegsComplete = true;
272
273 HasDisjunctSubRegs = ExplicitSubRegs.size() > 1;
274
275 // First insert the explicit subregs and make sure they are fully indexed.
276 for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
277 CodeGenRegister *SR = ExplicitSubRegs[i];
278 CodeGenSubRegIndex *Idx = ExplicitSubRegIndices[i];
279 if (!SR->Artificial)
280 Idx->Artificial = false;
281 if (!SubRegs.insert(std::make_pair(Idx, SR)).second)
282 PrintFatalError(TheDef->getLoc(), "SubRegIndex " + Idx->getName() +
283 " appears twice in Register " + getName());
284 // Map explicit sub-registers first, so the names take precedence.
285 // The inherited sub-registers are mapped below.
286 SubReg2Idx.insert(std::make_pair(SR, Idx));
287 }
288
289 // Keep track of inherited subregs and how they can be reached.
290 SmallPtrSet<CodeGenRegister*, 8> Orphans;
291
292 // Clone inherited subregs and place duplicate entries in Orphans.
293 // Here the order is important - earlier subregs take precedence.
294 for (CodeGenRegister *ESR : ExplicitSubRegs) {
295 const SubRegMap &Map = ESR->computeSubRegs(RegBank);
296 HasDisjunctSubRegs |= ESR->HasDisjunctSubRegs;
297
298 for (const auto &SR : Map) {
299 if (!SubRegs.insert(SR).second)
300 Orphans.insert(SR.second);
301 }
302 }
303
304 // Expand any composed subreg indices.
305 // If dsub_2 has ComposedOf = [qsub_1, dsub_0], and this register has a
306 // qsub_1 subreg, add a dsub_2 subreg. Keep growing Indices and process
307 // expanded subreg indices recursively.
308 SmallVector<CodeGenSubRegIndex*, 8> Indices = ExplicitSubRegIndices;
309 for (unsigned i = 0; i != Indices.size(); ++i) {
310 CodeGenSubRegIndex *Idx = Indices[i];
311 const CodeGenSubRegIndex::CompMap &Comps = Idx->getComposites();
312 CodeGenRegister *SR = SubRegs[Idx];
313 const SubRegMap &Map = SR->computeSubRegs(RegBank);
314
315 // Look at the possible compositions of Idx.
316 // They may not all be supported by SR.
317 for (CodeGenSubRegIndex::CompMap::const_iterator I = Comps.begin(),
318 E = Comps.end(); I != E; ++I) {
319 SubRegMap::const_iterator SRI = Map.find(I->first);
320 if (SRI == Map.end())
321 continue; // Idx + I->first doesn't exist in SR.
322 // Add I->second as a name for the subreg SRI->second, assuming it is
323 // orphaned, and the name isn't already used for something else.
324 if (SubRegs.count(I->second) || !Orphans.erase(SRI->second))
325 continue;
326 // We found a new name for the orphaned sub-register.
327 SubRegs.insert(std::make_pair(I->second, SRI->second));
328 Indices.push_back(I->second);
329 }
330 }
331
332 // Now Orphans contains the inherited subregisters without a direct index.
333 // Create inferred indexes for all missing entries.
334 // Work backwards in the Indices vector in order to compose subregs bottom-up.
335 // Consider this subreg sequence:
336 //
337 // qsub_1 -> dsub_0 -> ssub_0
338 //
339 // The qsub_1 -> dsub_0 composition becomes dsub_2, so the ssub_0 register
340 // can be reached in two different ways:
341 //
342 // qsub_1 -> ssub_0
343 // dsub_2 -> ssub_0
344 //
345 // We pick the latter composition because another register may have [dsub_0,
346 // dsub_1, dsub_2] subregs without necessarily having a qsub_1 subreg. The
347 // dsub_2 -> ssub_0 composition can be shared.
348 while (!Indices.empty() && !Orphans.empty()) {
349 CodeGenSubRegIndex *Idx = Indices.pop_back_val();
350 CodeGenRegister *SR = SubRegs[Idx];
351 const SubRegMap &Map = SR->computeSubRegs(RegBank);
352 for (const auto &SubReg : Map)
353 if (Orphans.erase(SubReg.second))
354 SubRegs[RegBank.getCompositeSubRegIndex(Idx, SubReg.first)] = SubReg.second;
355 }
356
357 // Compute the inverse SubReg -> Idx map.
358 for (const auto &SubReg : SubRegs) {
359 if (SubReg.second == this) {
360 ArrayRef<SMLoc> Loc;
361 if (TheDef)
362 Loc = TheDef->getLoc();
363 PrintFatalError(Loc, "Register " + getName() +
364 " has itself as a sub-register");
365 }
366
367 // Compute AllSuperRegsCovered.
368 if (!CoveredBySubRegs)
369 SubReg.first->AllSuperRegsCovered = false;
370
371 // Ensure that every sub-register has a unique name.
372 DenseMap<const CodeGenRegister*, CodeGenSubRegIndex*>::iterator Ins =
373 SubReg2Idx.insert(std::make_pair(SubReg.second, SubReg.first)).first;
374 if (Ins->second == SubReg.first)
375 continue;
376 // Trouble: Two different names for SubReg.second.
377 ArrayRef<SMLoc> Loc;
378 if (TheDef)
379 Loc = TheDef->getLoc();
380 PrintFatalError(Loc, "Sub-register can't have two names: " +
381 SubReg.second->getName() + " available as " +
382 SubReg.first->getName() + " and " + Ins->second->getName());
383 }
384
385 // Derive possible names for sub-register concatenations from any explicit
386 // sub-registers. By doing this before computeSecondarySubRegs(), we ensure
387 // that getConcatSubRegIndex() won't invent any concatenated indices that the
388 // user already specified.
389 for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
390 CodeGenRegister *SR = ExplicitSubRegs[i];
391 if (!SR->CoveredBySubRegs || SR->ExplicitSubRegs.size() <= 1 ||
392 SR->Artificial)
393 continue;
394
395 // SR is composed of multiple sub-regs. Find their names in this register.
396 SmallVector<CodeGenSubRegIndex*, 8> Parts;
397 for (unsigned j = 0, e = SR->ExplicitSubRegs.size(); j != e; ++j) {
398 CodeGenSubRegIndex &I = *SR->ExplicitSubRegIndices[j];
399 if (!I.Artificial)
400 Parts.push_back(getSubRegIndex(SR->ExplicitSubRegs[j]));
401 }
402
403 // Offer this as an existing spelling for the concatenation of Parts.
404 CodeGenSubRegIndex &Idx = *ExplicitSubRegIndices[i];
405 Idx.setConcatenationOf(Parts);
406 }
407
408 // Initialize RegUnitList. Because getSubRegs is called recursively, this
409 // processes the register hierarchy in postorder.
410 //
411 // Inherit all sub-register units. It is good enough to look at the explicit
412 // sub-registers, the other registers won't contribute any more units.
413 for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
414 CodeGenRegister *SR = ExplicitSubRegs[i];
415 RegUnits |= SR->RegUnits;
416 }
417
418 // Absent any ad hoc aliasing, we create one register unit per leaf register.
419 // These units correspond to the maximal cliques in the register overlap
420 // graph which is optimal.
421 //
422 // When there is ad hoc aliasing, we simply create one unit per edge in the
423 // undirected ad hoc aliasing graph. Technically, we could do better by
424 // identifying maximal cliques in the ad hoc graph, but cliques larger than 2
425 // are extremely rare anyway (I've never seen one), so we don't bother with
426 // the added complexity.
427 for (unsigned i = 0, e = ExplicitAliases.size(); i != e; ++i) {
428 CodeGenRegister *AR = ExplicitAliases[i];
429 // Only visit each edge once.
430 if (AR->SubRegsComplete)
431 continue;
432 // Create a RegUnit representing this alias edge, and add it to both
433 // registers.
434 unsigned Unit = RegBank.newRegUnit(this, AR);
435 RegUnits.set(Unit);
436 AR->RegUnits.set(Unit);
437 }
438
439 // Finally, create units for leaf registers without ad hoc aliases. Note that
440 // a leaf register with ad hoc aliases doesn't get its own unit - it isn't
441 // necessary. This means the aliasing leaf registers can share a single unit.
442 if (RegUnits.empty())
443 RegUnits.set(RegBank.newRegUnit(this));
444
445 // We have now computed the native register units. More may be adopted later
446 // for balancing purposes.
447 NativeRegUnits = RegUnits;
448
449 return SubRegs;
450}
451
452// In a register that is covered by its sub-registers, try to find redundant
453// sub-registers. For example:
454//
455// QQ0 = {Q0, Q1}
456// Q0 = {D0, D1}
457// Q1 = {D2, D3}
458//
459// We can infer that D1_D2 is also a sub-register, even if it wasn't named in
460// the register definition.
461//
462// The explicitly specified registers form a tree. This function discovers
463// sub-register relationships that would force a DAG.
464//
465void CodeGenRegister::computeSecondarySubRegs(CodeGenRegBank &RegBank) {
466 SmallVector<SubRegMap::value_type, 8> NewSubRegs;
467
468 std::queue<std::pair<CodeGenSubRegIndex*,CodeGenRegister*>> SubRegQueue;
469 for (std::pair<CodeGenSubRegIndex*,CodeGenRegister*> P : SubRegs)
470 SubRegQueue.push(P);
471
472 // Look at the leading super-registers of each sub-register. Those are the
473 // candidates for new sub-registers, assuming they are fully contained in
474 // this register.
475 while (!SubRegQueue.empty()) {
476 CodeGenSubRegIndex *SubRegIdx;
477 const CodeGenRegister *SubReg;
478 std::tie(SubRegIdx, SubReg) = SubRegQueue.front();
479 SubRegQueue.pop();
480
481 const CodeGenRegister::SuperRegList &Leads = SubReg->LeadingSuperRegs;
482 for (unsigned i = 0, e = Leads.size(); i != e; ++i) {
483 CodeGenRegister *Cand = const_cast<CodeGenRegister*>(Leads[i]);
484 // Already got this sub-register?
485 if (Cand == this || getSubRegIndex(Cand))
486 continue;
487 // Check if each component of Cand is already a sub-register.
488 assert(!Cand->ExplicitSubRegs.empty() &&((!Cand->ExplicitSubRegs.empty() && "Super-register has no sub-registers"
) ? static_cast<void> (0) : __assert_fail ("!Cand->ExplicitSubRegs.empty() && \"Super-register has no sub-registers\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 489, __PRETTY_FUNCTION__))
489 "Super-register has no sub-registers")((!Cand->ExplicitSubRegs.empty() && "Super-register has no sub-registers"
) ? static_cast<void> (0) : __assert_fail ("!Cand->ExplicitSubRegs.empty() && \"Super-register has no sub-registers\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 489, __PRETTY_FUNCTION__))
;
490 if (Cand->ExplicitSubRegs.size() == 1)
491 continue;
492 SmallVector<CodeGenSubRegIndex*, 8> Parts;
493 // We know that the first component is (SubRegIdx,SubReg). However we
494 // may still need to split it into smaller subregister parts.
495 assert(Cand->ExplicitSubRegs[0] == SubReg && "LeadingSuperRegs correct")((Cand->ExplicitSubRegs[0] == SubReg && "LeadingSuperRegs correct"
) ? static_cast<void> (0) : __assert_fail ("Cand->ExplicitSubRegs[0] == SubReg && \"LeadingSuperRegs correct\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 495, __PRETTY_FUNCTION__))
;
496 assert(getSubRegIndex(SubReg) == SubRegIdx && "LeadingSuperRegs correct")((getSubRegIndex(SubReg) == SubRegIdx && "LeadingSuperRegs correct"
) ? static_cast<void> (0) : __assert_fail ("getSubRegIndex(SubReg) == SubRegIdx && \"LeadingSuperRegs correct\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 496, __PRETTY_FUNCTION__))
;
497 for (CodeGenRegister *SubReg : Cand->ExplicitSubRegs) {
498 if (CodeGenSubRegIndex *SubRegIdx = getSubRegIndex(SubReg)) {
499 if (SubRegIdx->ConcatenationOf.empty()) {
500 Parts.push_back(SubRegIdx);
501 } else
502 for (CodeGenSubRegIndex *SubIdx : SubRegIdx->ConcatenationOf)
503 Parts.push_back(SubIdx);
504 } else {
505 // Sub-register doesn't exist.
506 Parts.clear();
507 break;
508 }
509 }
510 // There is nothing to do if some Cand sub-register is not part of this
511 // register.
512 if (Parts.empty())
513 continue;
514
515 // Each part of Cand is a sub-register of this. Make the full Cand also
516 // a sub-register with a concatenated sub-register index.
517 CodeGenSubRegIndex *Concat = RegBank.getConcatSubRegIndex(Parts);
518 std::pair<CodeGenSubRegIndex*,CodeGenRegister*> NewSubReg =
519 std::make_pair(Concat, Cand);
520
521 if (!SubRegs.insert(NewSubReg).second)
522 continue;
523
524 // We inserted a new subregister.
525 NewSubRegs.push_back(NewSubReg);
526 SubRegQueue.push(NewSubReg);
527 SubReg2Idx.insert(std::make_pair(Cand, Concat));
528 }
529 }
530
531 // Create sub-register index composition maps for the synthesized indices.
532 for (unsigned i = 0, e = NewSubRegs.size(); i != e; ++i) {
533 CodeGenSubRegIndex *NewIdx = NewSubRegs[i].first;
534 CodeGenRegister *NewSubReg = NewSubRegs[i].second;
535 for (SubRegMap::const_iterator SI = NewSubReg->SubRegs.begin(),
536 SE = NewSubReg->SubRegs.end(); SI != SE; ++SI) {
537 CodeGenSubRegIndex *SubIdx = getSubRegIndex(SI->second);
538 if (!SubIdx)
539 PrintFatalError(TheDef->getLoc(), "No SubRegIndex for " +
540 SI->second->getName() + " in " + getName());
541 NewIdx->addComposite(SI->first, SubIdx);
542 }
543 }
544}
545
546void CodeGenRegister::computeSuperRegs(CodeGenRegBank &RegBank) {
547 // Only visit each register once.
548 if (SuperRegsComplete)
549 return;
550 SuperRegsComplete = true;
551
552 // Make sure all sub-registers have been visited first, so the super-reg
553 // lists will be topologically ordered.
554 for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
555 I != E; ++I)
556 I->second->computeSuperRegs(RegBank);
557
558 // Now add this as a super-register on all sub-registers.
559 // Also compute the TopoSigId in post-order.
560 TopoSigId Id;
561 for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
562 I != E; ++I) {
563 // Topological signature computed from SubIdx, TopoId(SubReg).
564 // Loops and idempotent indices have TopoSig = ~0u.
565 Id.push_back(I->first->EnumValue);
566 Id.push_back(I->second->TopoSig);
567
568 // Don't add duplicate entries.
569 if (!I->second->SuperRegs.empty() && I->second->SuperRegs.back() == this)
570 continue;
571 I->second->SuperRegs.push_back(this);
572 }
573 TopoSig = RegBank.getTopoSig(Id);
574}
575
576void
577CodeGenRegister::addSubRegsPreOrder(SetVector<const CodeGenRegister*> &OSet,
578 CodeGenRegBank &RegBank) const {
579 assert(SubRegsComplete && "Must precompute sub-registers")((SubRegsComplete && "Must precompute sub-registers")
? static_cast<void> (0) : __assert_fail ("SubRegsComplete && \"Must precompute sub-registers\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 579, __PRETTY_FUNCTION__))
;
580 for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
581 CodeGenRegister *SR = ExplicitSubRegs[i];
582 if (OSet.insert(SR))
583 SR->addSubRegsPreOrder(OSet, RegBank);
584 }
585 // Add any secondary sub-registers that weren't part of the explicit tree.
586 for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
587 I != E; ++I)
588 OSet.insert(I->second);
589}
590
591// Get the sum of this register's unit weights.
592unsigned CodeGenRegister::getWeight(const CodeGenRegBank &RegBank) const {
593 unsigned Weight = 0;
594 for (RegUnitList::iterator I = RegUnits.begin(), E = RegUnits.end();
595 I != E; ++I) {
596 Weight += RegBank.getRegUnit(*I).Weight;
597 }
598 return Weight;
599}
600
601//===----------------------------------------------------------------------===//
602// RegisterTuples
603//===----------------------------------------------------------------------===//
604
605// A RegisterTuples def is used to generate pseudo-registers from lists of
606// sub-registers. We provide a SetTheory expander class that returns the new
607// registers.
608namespace {
609
610struct TupleExpander : SetTheory::Expander {
611 // Reference to SynthDefs in the containing CodeGenRegBank, to keep track of
612 // the synthesized definitions for their lifetime.
613 std::vector<std::unique_ptr<Record>> &SynthDefs;
614
615 TupleExpander(std::vector<std::unique_ptr<Record>> &SynthDefs)
616 : SynthDefs(SynthDefs) {}
617
618 void expand(SetTheory &ST, Record *Def, SetTheory::RecSet &Elts) override {
619 std::vector<Record*> Indices = Def->getValueAsListOfDefs("SubRegIndices");
620 unsigned Dim = Indices.size();
621 ListInit *SubRegs = Def->getValueAsListInit("SubRegs");
622 if (Dim != SubRegs->size())
623 PrintFatalError(Def->getLoc(), "SubRegIndices and SubRegs size mismatch");
624 if (Dim < 2)
625 PrintFatalError(Def->getLoc(),
626 "Tuples must have at least 2 sub-registers");
627
628 // Evaluate the sub-register lists to be zipped.
629 unsigned Length = ~0u;
630 SmallVector<SetTheory::RecSet, 4> Lists(Dim);
631 for (unsigned i = 0; i != Dim; ++i) {
632 ST.evaluate(SubRegs->getElement(i), Lists[i], Def->getLoc());
633 Length = std::min(Length, unsigned(Lists[i].size()));
634 }
635
636 if (Length == 0)
637 return;
638
639 // Precompute some types.
640 Record *RegisterCl = Def->getRecords().getClass("Register");
641 RecTy *RegisterRecTy = RecordRecTy::get(RegisterCl);
642 StringInit *BlankName = StringInit::get("");
643
644 // Zip them up.
645 for (unsigned n = 0; n != Length; ++n) {
646 std::string Name;
647 Record *Proto = Lists[0][n];
648 std::vector<Init*> Tuple;
649 unsigned CostPerUse = 0;
650 for (unsigned i = 0; i != Dim; ++i) {
651 Record *Reg = Lists[i][n];
652 if (i) Name += '_';
653 Name += Reg->getName();
654 Tuple.push_back(DefInit::get(Reg));
655 CostPerUse = std::max(CostPerUse,
656 unsigned(Reg->getValueAsInt("CostPerUse")));
657 }
658
659 // Create a new Record representing the synthesized register. This record
660 // is only for consumption by CodeGenRegister, it is not added to the
661 // RecordKeeper.
662 SynthDefs.emplace_back(
663 llvm::make_unique<Record>(Name, Def->getLoc(), Def->getRecords()));
664 Record *NewReg = SynthDefs.back().get();
665 Elts.insert(NewReg);
666
667 // Copy Proto super-classes.
668 ArrayRef<std::pair<Record *, SMRange>> Supers = Proto->getSuperClasses();
669 for (const auto &SuperPair : Supers)
670 NewReg->addSuperClass(SuperPair.first, SuperPair.second);
671
672 // Copy Proto fields.
673 for (unsigned i = 0, e = Proto->getValues().size(); i != e; ++i) {
674 RecordVal RV = Proto->getValues()[i];
675
676 // Skip existing fields, like NAME.
677 if (NewReg->getValue(RV.getNameInit()))
678 continue;
679
680 StringRef Field = RV.getName();
681
682 // Replace the sub-register list with Tuple.
683 if (Field == "SubRegs")
684 RV.setValue(ListInit::get(Tuple, RegisterRecTy));
685
686 // Provide a blank AsmName. MC hacks are required anyway.
687 if (Field == "AsmName")
688 RV.setValue(BlankName);
689
690 // CostPerUse is aggregated from all Tuple members.
691 if (Field == "CostPerUse")
692 RV.setValue(IntInit::get(CostPerUse));
693
694 // Composite registers are always covered by sub-registers.
695 if (Field == "CoveredBySubRegs")
696 RV.setValue(BitInit::get(true));
697
698 // Copy fields from the RegisterTuples def.
699 if (Field == "SubRegIndices" ||
700 Field == "CompositeIndices") {
701 NewReg->addValue(*Def->getValue(Field));
702 continue;
703 }
704
705 // Some fields get their default uninitialized value.
706 if (Field == "DwarfNumbers" ||
707 Field == "DwarfAlias" ||
708 Field == "Aliases") {
709 if (const RecordVal *DefRV = RegisterCl->getValue(Field))
710 NewReg->addValue(*DefRV);
711 continue;
712 }
713
714 // Everything else is copied from Proto.
715 NewReg->addValue(RV);
716 }
717 }
718 }
719};
720
721} // end anonymous namespace
722
723//===----------------------------------------------------------------------===//
724// CodeGenRegisterClass
725//===----------------------------------------------------------------------===//
726
727static void sortAndUniqueRegisters(CodeGenRegister::Vec &M) {
728 llvm::sort(M, deref<llvm::less>());
729 M.erase(std::unique(M.begin(), M.end(), deref<llvm::equal>()), M.end());
730}
731
732CodeGenRegisterClass::CodeGenRegisterClass(CodeGenRegBank &RegBank, Record *R)
733 : TheDef(R),
734 Name(R->getName()),
735 TopoSigs(RegBank.getNumTopoSigs()),
736 EnumValue(-1) {
737
738 std::vector<Record*> TypeList = R->getValueAsListOfDefs("RegTypes");
739 for (unsigned i = 0, e = TypeList.size(); i != e; ++i) {
740 Record *Type = TypeList[i];
741 if (!Type->isSubClassOf("ValueType"))
742 PrintFatalError("RegTypes list member '" + Type->getName() +
743 "' does not derive from the ValueType class!");
744 VTs.push_back(getValueTypeByHwMode(Type, RegBank.getHwModes()));
745 }
746 assert(!VTs.empty() && "RegisterClass must contain at least one ValueType!")((!VTs.empty() && "RegisterClass must contain at least one ValueType!"
) ? static_cast<void> (0) : __assert_fail ("!VTs.empty() && \"RegisterClass must contain at least one ValueType!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 746, __PRETTY_FUNCTION__))
;
747
748 // Allocation order 0 is the full set. AltOrders provides others.
749 const SetTheory::RecVec *Elements = RegBank.getSets().expand(R);
750 ListInit *AltOrders = R->getValueAsListInit("AltOrders");
751 Orders.resize(1 + AltOrders->size());
752
753 // Default allocation order always contains all registers.
754 Artificial = true;
755 for (unsigned i = 0, e = Elements->size(); i != e; ++i) {
756 Orders[0].push_back((*Elements)[i]);
757 const CodeGenRegister *Reg = RegBank.getReg((*Elements)[i]);
758 Members.push_back(Reg);
759 Artificial &= Reg->Artificial;
760 TopoSigs.set(Reg->getTopoSig());
761 }
762 sortAndUniqueRegisters(Members);
763
764 // Alternative allocation orders may be subsets.
765 SetTheory::RecSet Order;
766 for (unsigned i = 0, e = AltOrders->size(); i != e; ++i) {
767 RegBank.getSets().evaluate(AltOrders->getElement(i), Order, R->getLoc());
768 Orders[1 + i].append(Order.begin(), Order.end());
769 // Verify that all altorder members are regclass members.
770 while (!Order.empty()) {
771 CodeGenRegister *Reg = RegBank.getReg(Order.back());
772 Order.pop_back();
773 if (!contains(Reg))
774 PrintFatalError(R->getLoc(), " AltOrder register " + Reg->getName() +
775 " is not a class member");
776 }
777 }
778
779 Namespace = R->getValueAsString("Namespace");
780
781 if (const RecordVal *RV = R->getValue("RegInfos"))
782 if (DefInit *DI = dyn_cast_or_null<DefInit>(RV->getValue()))
783 RSI = RegSizeInfoByHwMode(DI->getDef(), RegBank.getHwModes());
784 unsigned Size = R->getValueAsInt("Size");
785 assert((RSI.hasDefault() || Size != 0 || VTs[0].isSimple()) &&(((RSI.hasDefault() || Size != 0 || VTs[0].isSimple()) &&
"Impossible to determine register size") ? static_cast<void
> (0) : __assert_fail ("(RSI.hasDefault() || Size != 0 || VTs[0].isSimple()) && \"Impossible to determine register size\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 786, __PRETTY_FUNCTION__))
786 "Impossible to determine register size")(((RSI.hasDefault() || Size != 0 || VTs[0].isSimple()) &&
"Impossible to determine register size") ? static_cast<void
> (0) : __assert_fail ("(RSI.hasDefault() || Size != 0 || VTs[0].isSimple()) && \"Impossible to determine register size\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 786, __PRETTY_FUNCTION__))
;
787 if (!RSI.hasDefault()) {
788 RegSizeInfo RI;
789 RI.RegSize = RI.SpillSize = Size ? Size
790 : VTs[0].getSimple().getSizeInBits();
791 RI.SpillAlignment = R->getValueAsInt("Alignment");
792 RSI.Map.insert({DefaultMode, RI});
793 }
794
795 CopyCost = R->getValueAsInt("CopyCost");
796 Allocatable = R->getValueAsBit("isAllocatable");
797 AltOrderSelect = R->getValueAsString("AltOrderSelect");
798 int AllocationPriority = R->getValueAsInt("AllocationPriority");
799 if (AllocationPriority < 0 || AllocationPriority > 63)
800 PrintFatalError(R->getLoc(), "AllocationPriority out of range [0,63]");
801 this->AllocationPriority = AllocationPriority;
802}
803
804// Create an inferred register class that was missing from the .td files.
805// Most properties will be inherited from the closest super-class after the
806// class structure has been computed.
807CodeGenRegisterClass::CodeGenRegisterClass(CodeGenRegBank &RegBank,
808 StringRef Name, Key Props)
809 : Members(*Props.Members),
810 TheDef(nullptr),
811 Name(Name),
812 TopoSigs(RegBank.getNumTopoSigs()),
813 EnumValue(-1),
814 RSI(Props.RSI),
815 CopyCost(0),
816 Allocatable(true),
817 AllocationPriority(0) {
818 Artificial = true;
819 for (const auto R : Members) {
820 TopoSigs.set(R->getTopoSig());
821 Artificial &= R->Artificial;
822 }
823}
824
825// Compute inherited propertied for a synthesized register class.
826void CodeGenRegisterClass::inheritProperties(CodeGenRegBank &RegBank) {
827 assert(!getDef() && "Only synthesized classes can inherit properties")((!getDef() && "Only synthesized classes can inherit properties"
) ? static_cast<void> (0) : __assert_fail ("!getDef() && \"Only synthesized classes can inherit properties\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 827, __PRETTY_FUNCTION__))
;
828 assert(!SuperClasses.empty() && "Synthesized class without super class")((!SuperClasses.empty() && "Synthesized class without super class"
) ? static_cast<void> (0) : __assert_fail ("!SuperClasses.empty() && \"Synthesized class without super class\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 828, __PRETTY_FUNCTION__))
;
829
830 // The last super-class is the smallest one.
831 CodeGenRegisterClass &Super = *SuperClasses.back();
832
833 // Most properties are copied directly.
834 // Exceptions are members, size, and alignment
835 Namespace = Super.Namespace;
836 VTs = Super.VTs;
837 CopyCost = Super.CopyCost;
838 Allocatable = Super.Allocatable;
839 AltOrderSelect = Super.AltOrderSelect;
840 AllocationPriority = Super.AllocationPriority;
841
842 // Copy all allocation orders, filter out foreign registers from the larger
843 // super-class.
844 Orders.resize(Super.Orders.size());
845 for (unsigned i = 0, ie = Super.Orders.size(); i != ie; ++i)
846 for (unsigned j = 0, je = Super.Orders[i].size(); j != je; ++j)
847 if (contains(RegBank.getReg(Super.Orders[i][j])))
848 Orders[i].push_back(Super.Orders[i][j]);
849}
850
851bool CodeGenRegisterClass::contains(const CodeGenRegister *Reg) const {
852 return std::binary_search(Members.begin(), Members.end(), Reg,
853 deref<llvm::less>());
854}
855
856namespace llvm {
857
858 raw_ostream &operator<<(raw_ostream &OS, const CodeGenRegisterClass::Key &K) {
859 OS << "{ " << K.RSI;
860 for (const auto R : *K.Members)
861 OS << ", " << R->getName();
862 return OS << " }";
863 }
864
865} // end namespace llvm
866
867// This is a simple lexicographical order that can be used to search for sets.
868// It is not the same as the topological order provided by TopoOrderRC.
869bool CodeGenRegisterClass::Key::
870operator<(const CodeGenRegisterClass::Key &B) const {
871 assert(Members && B.Members)((Members && B.Members) ? static_cast<void> (0)
: __assert_fail ("Members && B.Members", "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 871, __PRETTY_FUNCTION__))
;
872 return std::tie(*Members, RSI) < std::tie(*B.Members, B.RSI);
873}
874
875// Returns true if RC is a strict subclass.
876// RC is a sub-class of this class if it is a valid replacement for any
877// instruction operand where a register of this classis required. It must
878// satisfy these conditions:
879//
880// 1. All RC registers are also in this.
881// 2. The RC spill size must not be smaller than our spill size.
882// 3. RC spill alignment must be compatible with ours.
883//
884static bool testSubClass(const CodeGenRegisterClass *A,
885 const CodeGenRegisterClass *B) {
886 return A->RSI.isSubClassOf(B->RSI) &&
887 std::includes(A->getMembers().begin(), A->getMembers().end(),
888 B->getMembers().begin(), B->getMembers().end(),
889 deref<llvm::less>());
890}
891
892/// Sorting predicate for register classes. This provides a topological
893/// ordering that arranges all register classes before their sub-classes.
894///
895/// Register classes with the same registers, spill size, and alignment form a
896/// clique. They will be ordered alphabetically.
897///
898static bool TopoOrderRC(const CodeGenRegisterClass &PA,
899 const CodeGenRegisterClass &PB) {
900 auto *A = &PA;
901 auto *B = &PB;
902 if (A == B)
903 return false;
904
905 if (A->RSI < B->RSI)
906 return true;
907 if (A->RSI != B->RSI)
908 return false;
909
910 // Order by descending set size. Note that the classes' allocation order may
911 // not have been computed yet. The Members set is always vaild.
912 if (A->getMembers().size() > B->getMembers().size())
913 return true;
914 if (A->getMembers().size() < B->getMembers().size())
915 return false;
916
917 // Finally order by name as a tie breaker.
918 return StringRef(A->getName()) < B->getName();
919}
920
921std::string CodeGenRegisterClass::getQualifiedName() const {
922 if (Namespace.empty())
923 return getName();
924 else
925 return (Namespace + "::" + getName()).str();
926}
927
928// Compute sub-classes of all register classes.
929// Assume the classes are ordered topologically.
930void CodeGenRegisterClass::computeSubClasses(CodeGenRegBank &RegBank) {
931 auto &RegClasses = RegBank.getRegClasses();
932
933 // Visit backwards so sub-classes are seen first.
934 for (auto I = RegClasses.rbegin(), E = RegClasses.rend(); I != E; ++I) {
935 CodeGenRegisterClass &RC = *I;
936 RC.SubClasses.resize(RegClasses.size());
937 RC.SubClasses.set(RC.EnumValue);
938 if (RC.Artificial)
939 continue;
940
941 // Normally, all subclasses have IDs >= rci, unless RC is part of a clique.
942 for (auto I2 = I.base(), E2 = RegClasses.end(); I2 != E2; ++I2) {
943 CodeGenRegisterClass &SubRC = *I2;
944 if (RC.SubClasses.test(SubRC.EnumValue))
945 continue;
946 if (!testSubClass(&RC, &SubRC))
947 continue;
948 // SubRC is a sub-class. Grap all its sub-classes so we won't have to
949 // check them again.
950 RC.SubClasses |= SubRC.SubClasses;
951 }
952
953 // Sweep up missed clique members. They will be immediately preceding RC.
954 for (auto I2 = std::next(I); I2 != E && testSubClass(&RC, &*I2); ++I2)
955 RC.SubClasses.set(I2->EnumValue);
956 }
957
958 // Compute the SuperClasses lists from the SubClasses vectors.
959 for (auto &RC : RegClasses) {
960 const BitVector &SC = RC.getSubClasses();
961 auto I = RegClasses.begin();
962 for (int s = 0, next_s = SC.find_first(); next_s != -1;
963 next_s = SC.find_next(s)) {
964 std::advance(I, next_s - s);
965 s = next_s;
966 if (&*I == &RC)
967 continue;
968 I->SuperClasses.push_back(&RC);
969 }
970 }
971
972 // With the class hierarchy in place, let synthesized register classes inherit
973 // properties from their closest super-class. The iteration order here can
974 // propagate properties down multiple levels.
975 for (auto &RC : RegClasses)
976 if (!RC.getDef())
977 RC.inheritProperties(RegBank);
978}
979
980Optional<std::pair<CodeGenRegisterClass *, CodeGenRegisterClass *>>
981CodeGenRegisterClass::getMatchingSubClassWithSubRegs(
982 CodeGenRegBank &RegBank, const CodeGenSubRegIndex *SubIdx) const {
983 auto SizeOrder = [](const CodeGenRegisterClass *A,
984 const CodeGenRegisterClass *B) {
985 return A->getMembers().size() > B->getMembers().size();
986 };
987
988 auto &RegClasses = RegBank.getRegClasses();
989
990 // Find all the subclasses of this one that fully support the sub-register
991 // index and order them by size. BiggestSuperRC should always be first.
992 CodeGenRegisterClass *BiggestSuperRegRC = getSubClassWithSubReg(SubIdx);
993 if (!BiggestSuperRegRC)
994 return None;
995 BitVector SuperRegRCsBV = BiggestSuperRegRC->getSubClasses();
996 std::vector<CodeGenRegisterClass *> SuperRegRCs;
997 for (auto &RC : RegClasses)
998 if (SuperRegRCsBV[RC.EnumValue])
999 SuperRegRCs.emplace_back(&RC);
1000 llvm::sort(SuperRegRCs, SizeOrder);
1001 assert(SuperRegRCs.front() == BiggestSuperRegRC && "Biggest class wasn't first")((SuperRegRCs.front() == BiggestSuperRegRC && "Biggest class wasn't first"
) ? static_cast<void> (0) : __assert_fail ("SuperRegRCs.front() == BiggestSuperRegRC && \"Biggest class wasn't first\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1001, __PRETTY_FUNCTION__))
;
1002
1003 // Find all the subreg classes and order them by size too.
1004 std::vector<std::pair<CodeGenRegisterClass *, BitVector>> SuperRegClasses;
1005 for (auto &RC: RegClasses) {
1006 BitVector SuperRegClassesBV(RegClasses.size());
1007 RC.getSuperRegClasses(SubIdx, SuperRegClassesBV);
1008 if (SuperRegClassesBV.any())
1009 SuperRegClasses.push_back(std::make_pair(&RC, SuperRegClassesBV));
1010 }
1011 llvm::sort(SuperRegClasses,
1012 [&](const std::pair<CodeGenRegisterClass *, BitVector> &A,
1013 const std::pair<CodeGenRegisterClass *, BitVector> &B) {
1014 return SizeOrder(A.first, B.first);
1015 });
1016
1017 // Find the biggest subclass and subreg class such that R:subidx is in the
1018 // subreg class for all R in subclass.
1019 //
1020 // For example:
1021 // All registers in X86's GR64 have a sub_32bit subregister but no class
1022 // exists that contains all the 32-bit subregisters because GR64 contains RIP
1023 // but GR32 does not contain EIP. Instead, we constrain SuperRegRC to
1024 // GR32_with_sub_8bit (which is identical to GR32_with_sub_32bit) and then,
1025 // having excluded RIP, we are able to find a SubRegRC (GR32).
1026 CodeGenRegisterClass *ChosenSuperRegClass = nullptr;
1027 CodeGenRegisterClass *SubRegRC = nullptr;
1028 for (auto *SuperRegRC : SuperRegRCs) {
1029 for (const auto &SuperRegClassPair : SuperRegClasses) {
1030 const BitVector &SuperRegClassBV = SuperRegClassPair.second;
1031 if (SuperRegClassBV[SuperRegRC->EnumValue]) {
1032 SubRegRC = SuperRegClassPair.first;
1033 ChosenSuperRegClass = SuperRegRC;
1034
1035 // If SubRegRC is bigger than SuperRegRC then there are members of
1036 // SubRegRC that don't have super registers via SubIdx. Keep looking to
1037 // find a better fit and fall back on this one if there isn't one.
1038 //
1039 // This is intended to prevent X86 from making odd choices such as
1040 // picking LOW32_ADDR_ACCESS_RBP instead of GR32 in the example above.
1041 // LOW32_ADDR_ACCESS_RBP is a valid choice but contains registers that
1042 // aren't subregisters of SuperRegRC whereas GR32 has a direct 1:1
1043 // mapping.
1044 if (SuperRegRC->getMembers().size() >= SubRegRC->getMembers().size())
1045 return std::make_pair(ChosenSuperRegClass, SubRegRC);
1046 }
1047 }
1048
1049 // If we found a fit but it wasn't quite ideal because SubRegRC had excess
1050 // registers, then we're done.
1051 if (ChosenSuperRegClass)
1052 return std::make_pair(ChosenSuperRegClass, SubRegRC);
1053 }
1054
1055 return None;
1056}
1057
1058void CodeGenRegisterClass::getSuperRegClasses(const CodeGenSubRegIndex *SubIdx,
1059 BitVector &Out) const {
1060 auto FindI = SuperRegClasses.find(SubIdx);
1061 if (FindI == SuperRegClasses.end())
1062 return;
1063 for (CodeGenRegisterClass *RC : FindI->second)
1064 Out.set(RC->EnumValue);
1065}
1066
1067// Populate a unique sorted list of units from a register set.
1068void CodeGenRegisterClass::buildRegUnitSet(const CodeGenRegBank &RegBank,
1069 std::vector<unsigned> &RegUnits) const {
1070 std::vector<unsigned> TmpUnits;
1071 for (RegUnitIterator UnitI(Members); UnitI.isValid(); ++UnitI) {
1072 const RegUnit &RU = RegBank.getRegUnit(*UnitI);
1073 if (!RU.Artificial)
1074 TmpUnits.push_back(*UnitI);
1075 }
1076 llvm::sort(TmpUnits);
1077 std::unique_copy(TmpUnits.begin(), TmpUnits.end(),
1078 std::back_inserter(RegUnits));
1079}
1080
1081//===----------------------------------------------------------------------===//
1082// CodeGenRegBank
1083//===----------------------------------------------------------------------===//
1084
1085CodeGenRegBank::CodeGenRegBank(RecordKeeper &Records,
1086 const CodeGenHwModes &Modes) : CGH(Modes) {
1087 // Configure register Sets to understand register classes and tuples.
1088 Sets.addFieldExpander("RegisterClass", "MemberList");
1089 Sets.addFieldExpander("CalleeSavedRegs", "SaveList");
1090 Sets.addExpander("RegisterTuples",
1091 llvm::make_unique<TupleExpander>(SynthDefs));
1092
1093 // Read in the user-defined (named) sub-register indices.
1094 // More indices will be synthesized later.
1095 std::vector<Record*> SRIs = Records.getAllDerivedDefinitions("SubRegIndex");
1096 llvm::sort(SRIs, LessRecord());
1097 for (unsigned i = 0, e = SRIs.size(); i != e; ++i)
1098 getSubRegIdx(SRIs[i]);
1099 // Build composite maps from ComposedOf fields.
1100 for (auto &Idx : SubRegIndices)
1101 Idx.updateComponents(*this);
1102
1103 // Read in the register definitions.
1104 std::vector<Record*> Regs = Records.getAllDerivedDefinitions("Register");
1105 llvm::sort(Regs, LessRecordRegister());
1106 // Assign the enumeration values.
1107 for (unsigned i = 0, e = Regs.size(); i != e; ++i)
1108 getReg(Regs[i]);
1109
1110 // Expand tuples and number the new registers.
1111 std::vector<Record*> Tups =
1112 Records.getAllDerivedDefinitions("RegisterTuples");
1113
1114 for (Record *R : Tups) {
1115 std::vector<Record *> TupRegs = *Sets.expand(R);
1116 llvm::sort(TupRegs, LessRecordRegister());
1117 for (Record *RC : TupRegs)
1118 getReg(RC);
1119 }
1120
1121 // Now all the registers are known. Build the object graph of explicit
1122 // register-register references.
1123 for (auto &Reg : Registers)
1124 Reg.buildObjectGraph(*this);
1125
1126 // Compute register name map.
1127 for (auto &Reg : Registers)
1128 // FIXME: This could just be RegistersByName[name] = register, except that
1129 // causes some failures in MIPS - perhaps they have duplicate register name
1130 // entries? (or maybe there's a reason for it - I don't know much about this
1131 // code, just drive-by refactoring)
1132 RegistersByName.insert(
1133 std::make_pair(Reg.TheDef->getValueAsString("AsmName"), &Reg));
1134
1135 // Precompute all sub-register maps.
1136 // This will create Composite entries for all inferred sub-register indices.
1137 for (auto &Reg : Registers)
1138 Reg.computeSubRegs(*this);
1139
1140 // Compute transitive closure of subregister index ConcatenationOf vectors
1141 // and initialize ConcatIdx map.
1142 for (CodeGenSubRegIndex &SRI : SubRegIndices) {
1143 SRI.computeConcatTransitiveClosure();
1144 if (!SRI.ConcatenationOf.empty())
1145 ConcatIdx.insert(std::make_pair(
1146 SmallVector<CodeGenSubRegIndex*,8>(SRI.ConcatenationOf.begin(),
1147 SRI.ConcatenationOf.end()), &SRI));
1148 }
1149
1150 // Infer even more sub-registers by combining leading super-registers.
1151 for (auto &Reg : Registers)
1152 if (Reg.CoveredBySubRegs)
1153 Reg.computeSecondarySubRegs(*this);
1154
1155 // After the sub-register graph is complete, compute the topologically
1156 // ordered SuperRegs list.
1157 for (auto &Reg : Registers)
1158 Reg.computeSuperRegs(*this);
1159
1160 // For each pair of Reg:SR, if both are non-artificial, mark the
1161 // corresponding sub-register index as non-artificial.
1162 for (auto &Reg : Registers) {
1163 if (Reg.Artificial)
1164 continue;
1165 for (auto P : Reg.getSubRegs()) {
1166 const CodeGenRegister *SR = P.second;
1167 if (!SR->Artificial)
1168 P.first->Artificial = false;
1169 }
1170 }
1171
1172 // Native register units are associated with a leaf register. They've all been
1173 // discovered now.
1174 NumNativeRegUnits = RegUnits.size();
1175
1176 // Read in register class definitions.
1177 std::vector<Record*> RCs = Records.getAllDerivedDefinitions("RegisterClass");
1178 if (RCs.empty())
1179 PrintFatalError("No 'RegisterClass' subclasses defined!");
1180
1181 // Allocate user-defined register classes.
1182 for (auto *R : RCs) {
1183 RegClasses.emplace_back(*this, R);
1184 CodeGenRegisterClass &RC = RegClasses.back();
1185 if (!RC.Artificial)
1186 addToMaps(&RC);
1187 }
1188
1189 // Infer missing classes to create a full algebra.
1190 computeInferredRegisterClasses();
1191
1192 // Order register classes topologically and assign enum values.
1193 RegClasses.sort(TopoOrderRC);
1194 unsigned i = 0;
1195 for (auto &RC : RegClasses)
1196 RC.EnumValue = i++;
1197 CodeGenRegisterClass::computeSubClasses(*this);
1198}
1199
1200// Create a synthetic CodeGenSubRegIndex without a corresponding Record.
1201CodeGenSubRegIndex*
1202CodeGenRegBank::createSubRegIndex(StringRef Name, StringRef Namespace) {
1203 SubRegIndices.emplace_back(Name, Namespace, SubRegIndices.size() + 1);
1204 return &SubRegIndices.back();
1205}
1206
1207CodeGenSubRegIndex *CodeGenRegBank::getSubRegIdx(Record *Def) {
1208 CodeGenSubRegIndex *&Idx = Def2SubRegIdx[Def];
1209 if (Idx)
1210 return Idx;
1211 SubRegIndices.emplace_back(Def, SubRegIndices.size() + 1);
1212 Idx = &SubRegIndices.back();
1213 return Idx;
1214}
1215
1216CodeGenRegister *CodeGenRegBank::getReg(Record *Def) {
1217 CodeGenRegister *&Reg = Def2Reg[Def];
1218 if (Reg)
1219 return Reg;
1220 Registers.emplace_back(Def, Registers.size() + 1);
1221 Reg = &Registers.back();
1222 return Reg;
1223}
1224
1225void CodeGenRegBank::addToMaps(CodeGenRegisterClass *RC) {
1226 if (Record *Def = RC->getDef())
1227 Def2RC.insert(std::make_pair(Def, RC));
1228
1229 // Duplicate classes are rejected by insert().
1230 // That's OK, we only care about the properties handled by CGRC::Key.
1231 CodeGenRegisterClass::Key K(*RC);
1232 Key2RC.insert(std::make_pair(K, RC));
1233}
1234
1235// Create a synthetic sub-class if it is missing.
1236CodeGenRegisterClass*
1237CodeGenRegBank::getOrCreateSubClass(const CodeGenRegisterClass *RC,
1238 const CodeGenRegister::Vec *Members,
1239 StringRef Name) {
1240 // Synthetic sub-class has the same size and alignment as RC.
1241 CodeGenRegisterClass::Key K(Members, RC->RSI);
1242 RCKeyMap::const_iterator FoundI = Key2RC.find(K);
1243 if (FoundI != Key2RC.end())
1244 return FoundI->second;
1245
1246 // Sub-class doesn't exist, create a new one.
1247 RegClasses.emplace_back(*this, Name, K);
1248 addToMaps(&RegClasses.back());
1249 return &RegClasses.back();
1250}
1251
1252CodeGenRegisterClass *CodeGenRegBank::getRegClass(Record *Def) {
1253 if (CodeGenRegisterClass *RC = Def2RC[Def])
1254 return RC;
1255
1256 PrintFatalError(Def->getLoc(), "Not a known RegisterClass!");
1257}
1258
1259CodeGenSubRegIndex*
1260CodeGenRegBank::getCompositeSubRegIndex(CodeGenSubRegIndex *A,
1261 CodeGenSubRegIndex *B) {
1262 // Look for an existing entry.
1263 CodeGenSubRegIndex *Comp = A->compose(B);
1264 if (Comp)
1265 return Comp;
1266
1267 // None exists, synthesize one.
1268 std::string Name = A->getName() + "_then_" + B->getName();
1269 Comp = createSubRegIndex(Name, A->getNamespace());
1270 A->addComposite(B, Comp);
1271 return Comp;
1272}
1273
1274CodeGenSubRegIndex *CodeGenRegBank::
1275getConcatSubRegIndex(const SmallVector<CodeGenSubRegIndex *, 8> &Parts) {
1276 assert(Parts.size() > 1 && "Need two parts to concatenate")((Parts.size() > 1 && "Need two parts to concatenate"
) ? static_cast<void> (0) : __assert_fail ("Parts.size() > 1 && \"Need two parts to concatenate\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1276, __PRETTY_FUNCTION__))
;
1277#ifndef NDEBUG
1278 for (CodeGenSubRegIndex *Idx : Parts) {
1279 assert(Idx->ConcatenationOf.empty() && "No transitive closure?")((Idx->ConcatenationOf.empty() && "No transitive closure?"
) ? static_cast<void> (0) : __assert_fail ("Idx->ConcatenationOf.empty() && \"No transitive closure?\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1279, __PRETTY_FUNCTION__))
;
1280 }
1281#endif
1282
1283 // Look for an existing entry.
1284 CodeGenSubRegIndex *&Idx = ConcatIdx[Parts];
1285 if (Idx)
1286 return Idx;
1287
1288 // None exists, synthesize one.
1289 std::string Name = Parts.front()->getName();
1290 // Determine whether all parts are contiguous.
1291 bool isContinuous = true;
1292 unsigned Size = Parts.front()->Size;
1293 unsigned LastOffset = Parts.front()->Offset;
1294 unsigned LastSize = Parts.front()->Size;
1295 for (unsigned i = 1, e = Parts.size(); i != e; ++i) {
1296 Name += '_';
1297 Name += Parts[i]->getName();
1298 Size += Parts[i]->Size;
1299 if (Parts[i]->Offset != (LastOffset + LastSize))
1300 isContinuous = false;
1301 LastOffset = Parts[i]->Offset;
1302 LastSize = Parts[i]->Size;
1303 }
1304 Idx = createSubRegIndex(Name, Parts.front()->getNamespace());
1305 Idx->Size = Size;
1306 Idx->Offset = isContinuous ? Parts.front()->Offset : -1;
1307 Idx->ConcatenationOf.assign(Parts.begin(), Parts.end());
1308 return Idx;
1309}
1310
1311void CodeGenRegBank::computeComposites() {
1312 // Keep track of TopoSigs visited. We only need to visit each TopoSig once,
1313 // and many registers will share TopoSigs on regular architectures.
1314 BitVector TopoSigs(getNumTopoSigs());
1315
1316 for (const auto &Reg1 : Registers) {
1317 // Skip identical subreg structures already processed.
1318 if (TopoSigs.test(Reg1.getTopoSig()))
1319 continue;
1320 TopoSigs.set(Reg1.getTopoSig());
1321
1322 const CodeGenRegister::SubRegMap &SRM1 = Reg1.getSubRegs();
1323 for (CodeGenRegister::SubRegMap::const_iterator i1 = SRM1.begin(),
1324 e1 = SRM1.end(); i1 != e1; ++i1) {
1325 CodeGenSubRegIndex *Idx1 = i1->first;
1326 CodeGenRegister *Reg2 = i1->second;
1327 // Ignore identity compositions.
1328 if (&Reg1 == Reg2)
1329 continue;
1330 const CodeGenRegister::SubRegMap &SRM2 = Reg2->getSubRegs();
1331 // Try composing Idx1 with another SubRegIndex.
1332 for (CodeGenRegister::SubRegMap::const_iterator i2 = SRM2.begin(),
1333 e2 = SRM2.end(); i2 != e2; ++i2) {
1334 CodeGenSubRegIndex *Idx2 = i2->first;
1335 CodeGenRegister *Reg3 = i2->second;
1336 // Ignore identity compositions.
1337 if (Reg2 == Reg3)
1338 continue;
1339 // OK Reg1:IdxPair == Reg3. Find the index with Reg:Idx == Reg3.
1340 CodeGenSubRegIndex *Idx3 = Reg1.getSubRegIndex(Reg3);
1341 assert(Idx3 && "Sub-register doesn't have an index")((Idx3 && "Sub-register doesn't have an index") ? static_cast
<void> (0) : __assert_fail ("Idx3 && \"Sub-register doesn't have an index\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1341, __PRETTY_FUNCTION__))
;
1342
1343 // Conflicting composition? Emit a warning but allow it.
1344 if (CodeGenSubRegIndex *Prev = Idx1->addComposite(Idx2, Idx3))
1345 PrintWarning(Twine("SubRegIndex ") + Idx1->getQualifiedName() +
1346 " and " + Idx2->getQualifiedName() +
1347 " compose ambiguously as " + Prev->getQualifiedName() +
1348 " or " + Idx3->getQualifiedName());
1349 }
1350 }
1351 }
1352}
1353
1354// Compute lane masks. This is similar to register units, but at the
1355// sub-register index level. Each bit in the lane mask is like a register unit
1356// class, and two lane masks will have a bit in common if two sub-register
1357// indices overlap in some register.
1358//
1359// Conservatively share a lane mask bit if two sub-register indices overlap in
1360// some registers, but not in others. That shouldn't happen a lot.
1361void CodeGenRegBank::computeSubRegLaneMasks() {
1362 // First assign individual bits to all the leaf indices.
1363 unsigned Bit = 0;
1364 // Determine mask of lanes that cover their registers.
1365 CoveringLanes = LaneBitmask::getAll();
1366 for (auto &Idx : SubRegIndices) {
1367 if (Idx.getComposites().empty()) {
1368 if (Bit > LaneBitmask::BitWidth) {
1369 PrintFatalError(
1370 Twine("Ran out of lanemask bits to represent subregister ")
1371 + Idx.getName());
1372 }
1373 Idx.LaneMask = LaneBitmask::getLane(Bit);
1374 ++Bit;
1375 } else {
1376 Idx.LaneMask = LaneBitmask::getNone();
1377 }
1378 }
1379
1380 // Compute transformation sequences for composeSubRegIndexLaneMask. The idea
1381 // here is that for each possible target subregister we look at the leafs
1382 // in the subregister graph that compose for this target and create
1383 // transformation sequences for the lanemasks. Each step in the sequence
1384 // consists of a bitmask and a bitrotate operation. As the rotation amounts
1385 // are usually the same for many subregisters we can easily combine the steps
1386 // by combining the masks.
1387 for (const auto &Idx : SubRegIndices) {
1388 const auto &Composites = Idx.getComposites();
1389 auto &LaneTransforms = Idx.CompositionLaneMaskTransform;
1390
1391 if (Composites.empty()) {
2
Assuming the condition is true
3
Taking true branch
1392 // Moving from a class with no subregisters we just had a single lane:
1393 // The subregister must be a leaf subregister and only occupies 1 bit.
1394 // Move the bit from the class without subregisters into that position.
1395 unsigned DstBit = Idx.LaneMask.getHighestLane();
4
Calling 'LaneBitmask::getHighestLane'
9
Returning from 'LaneBitmask::getHighestLane'
10
'DstBit' initialized to 4294967295
1396 assert(Idx.LaneMask == LaneBitmask::getLane(DstBit) &&((Idx.LaneMask == LaneBitmask::getLane(DstBit) && "Must be a leaf subregister"
) ? static_cast<void> (0) : __assert_fail ("Idx.LaneMask == LaneBitmask::getLane(DstBit) && \"Must be a leaf subregister\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1397, __PRETTY_FUNCTION__))
11
Within the expansion of the macro 'assert':
a
Passing the value 4294967295 via 1st parameter 'Lane'
b
Calling 'LaneBitmask::getLane'
1397 "Must be a leaf subregister")((Idx.LaneMask == LaneBitmask::getLane(DstBit) && "Must be a leaf subregister"
) ? static_cast<void> (0) : __assert_fail ("Idx.LaneMask == LaneBitmask::getLane(DstBit) && \"Must be a leaf subregister\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1397, __PRETTY_FUNCTION__))
;
1398 MaskRolPair MaskRol = { LaneBitmask::getLane(0), (uint8_t)DstBit };
1399 LaneTransforms.push_back(MaskRol);
1400 } else {
1401 // Go through all leaf subregisters and find the ones that compose with
1402 // Idx. These make out all possible valid bits in the lane mask we want to
1403 // transform. Looking only at the leafs ensure that only a single bit in
1404 // the mask is set.
1405 unsigned NextBit = 0;
1406 for (auto &Idx2 : SubRegIndices) {
1407 // Skip non-leaf subregisters.
1408 if (!Idx2.getComposites().empty())
1409 continue;
1410 // Replicate the behaviour from the lane mask generation loop above.
1411 unsigned SrcBit = NextBit;
1412 LaneBitmask SrcMask = LaneBitmask::getLane(SrcBit);
1413 if (NextBit < LaneBitmask::BitWidth-1)
1414 ++NextBit;
1415 assert(Idx2.LaneMask == SrcMask)((Idx2.LaneMask == SrcMask) ? static_cast<void> (0) : __assert_fail
("Idx2.LaneMask == SrcMask", "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1415, __PRETTY_FUNCTION__))
;
1416
1417 // Get the composed subregister if there is any.
1418 auto C = Composites.find(&Idx2);
1419 if (C == Composites.end())
1420 continue;
1421 const CodeGenSubRegIndex *Composite = C->second;
1422 // The Composed subreg should be a leaf subreg too
1423 assert(Composite->getComposites().empty())((Composite->getComposites().empty()) ? static_cast<void
> (0) : __assert_fail ("Composite->getComposites().empty()"
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1423, __PRETTY_FUNCTION__))
;
1424
1425 // Create Mask+Rotate operation and merge with existing ops if possible.
1426 unsigned DstBit = Composite->LaneMask.getHighestLane();
1427 int Shift = DstBit - SrcBit;
1428 uint8_t RotateLeft = Shift >= 0 ? (uint8_t)Shift
1429 : LaneBitmask::BitWidth + Shift;
1430 for (auto &I : LaneTransforms) {
1431 if (I.RotateLeft == RotateLeft) {
1432 I.Mask |= SrcMask;
1433 SrcMask = LaneBitmask::getNone();
1434 }
1435 }
1436 if (SrcMask.any()) {
1437 MaskRolPair MaskRol = { SrcMask, RotateLeft };
1438 LaneTransforms.push_back(MaskRol);
1439 }
1440 }
1441 }
1442
1443 // Optimize if the transformation consists of one step only: Set mask to
1444 // 0xffffffff (including some irrelevant invalid bits) so that it should
1445 // merge with more entries later while compressing the table.
1446 if (LaneTransforms.size() == 1)
1447 LaneTransforms[0].Mask = LaneBitmask::getAll();
1448
1449 // Further compression optimization: For invalid compositions resulting
1450 // in a sequence with 0 entries we can just pick any other. Choose
1451 // Mask 0xffffffff with Rotation 0.
1452 if (LaneTransforms.size() == 0) {
1453 MaskRolPair P = { LaneBitmask::getAll(), 0 };
1454 LaneTransforms.push_back(P);
1455 }
1456 }
1457
1458 // FIXME: What if ad-hoc aliasing introduces overlaps that aren't represented
1459 // by the sub-register graph? This doesn't occur in any known targets.
1460
1461 // Inherit lanes from composites.
1462 for (const auto &Idx : SubRegIndices) {
1463 LaneBitmask Mask = Idx.computeLaneMask();
1464 // If some super-registers without CoveredBySubRegs use this index, we can
1465 // no longer assume that the lanes are covering their registers.
1466 if (!Idx.AllSuperRegsCovered)
1467 CoveringLanes &= ~Mask;
1468 }
1469
1470 // Compute lane mask combinations for register classes.
1471 for (auto &RegClass : RegClasses) {
1472 LaneBitmask LaneMask;
1473 for (const auto &SubRegIndex : SubRegIndices) {
1474 if (RegClass.getSubClassWithSubReg(&SubRegIndex) == nullptr)
1475 continue;
1476 LaneMask |= SubRegIndex.LaneMask;
1477 }
1478
1479 // For classes without any subregisters set LaneMask to 1 instead of 0.
1480 // This makes it easier for client code to handle classes uniformly.
1481 if (LaneMask.none())
1482 LaneMask = LaneBitmask::getLane(0);
1483
1484 RegClass.LaneMask = LaneMask;
1485 }
1486}
1487
1488namespace {
1489
1490// UberRegSet is a helper class for computeRegUnitWeights. Each UberRegSet is
1491// the transitive closure of the union of overlapping register
1492// classes. Together, the UberRegSets form a partition of the registers. If we
1493// consider overlapping register classes to be connected, then each UberRegSet
1494// is a set of connected components.
1495//
1496// An UberRegSet will likely be a horizontal slice of register names of
1497// the same width. Nontrivial subregisters should then be in a separate
1498// UberRegSet. But this property isn't required for valid computation of
1499// register unit weights.
1500//
1501// A Weight field caches the max per-register unit weight in each UberRegSet.
1502//
1503// A set of SingularDeterminants flags single units of some register in this set
1504// for which the unit weight equals the set weight. These units should not have
1505// their weight increased.
1506struct UberRegSet {
1507 CodeGenRegister::Vec Regs;
1508 unsigned Weight = 0;
1509 CodeGenRegister::RegUnitList SingularDeterminants;
1510
1511 UberRegSet() = default;
1512};
1513
1514} // end anonymous namespace
1515
1516// Partition registers into UberRegSets, where each set is the transitive
1517// closure of the union of overlapping register classes.
1518//
1519// UberRegSets[0] is a special non-allocatable set.
1520static void computeUberSets(std::vector<UberRegSet> &UberSets,
1521 std::vector<UberRegSet*> &RegSets,
1522 CodeGenRegBank &RegBank) {
1523 const auto &Registers = RegBank.getRegisters();
1524
1525 // The Register EnumValue is one greater than its index into Registers.
1526 assert(Registers.size() == Registers.back().EnumValue &&((Registers.size() == Registers.back().EnumValue && "register enum value mismatch"
) ? static_cast<void> (0) : __assert_fail ("Registers.size() == Registers.back().EnumValue && \"register enum value mismatch\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1527, __PRETTY_FUNCTION__))
1527 "register enum value mismatch")((Registers.size() == Registers.back().EnumValue && "register enum value mismatch"
) ? static_cast<void> (0) : __assert_fail ("Registers.size() == Registers.back().EnumValue && \"register enum value mismatch\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1527, __PRETTY_FUNCTION__))
;
1528
1529 // For simplicitly make the SetID the same as EnumValue.
1530 IntEqClasses UberSetIDs(Registers.size()+1);
1531 std::set<unsigned> AllocatableRegs;
1532 for (auto &RegClass : RegBank.getRegClasses()) {
1533 if (!RegClass.Allocatable)
1534 continue;
1535
1536 const CodeGenRegister::Vec &Regs = RegClass.getMembers();
1537 if (Regs.empty())
1538 continue;
1539
1540 unsigned USetID = UberSetIDs.findLeader((*Regs.begin())->EnumValue);
1541 assert(USetID && "register number 0 is invalid")((USetID && "register number 0 is invalid") ? static_cast
<void> (0) : __assert_fail ("USetID && \"register number 0 is invalid\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1541, __PRETTY_FUNCTION__))
;
1542
1543 AllocatableRegs.insert((*Regs.begin())->EnumValue);
1544 for (auto I = std::next(Regs.begin()), E = Regs.end(); I != E; ++I) {
1545 AllocatableRegs.insert((*I)->EnumValue);
1546 UberSetIDs.join(USetID, (*I)->EnumValue);
1547 }
1548 }
1549 // Combine non-allocatable regs.
1550 for (const auto &Reg : Registers) {
1551 unsigned RegNum = Reg.EnumValue;
1552 if (AllocatableRegs.count(RegNum))
1553 continue;
1554
1555 UberSetIDs.join(0, RegNum);
1556 }
1557 UberSetIDs.compress();
1558
1559 // Make the first UberSet a special unallocatable set.
1560 unsigned ZeroID = UberSetIDs[0];
1561
1562 // Insert Registers into the UberSets formed by union-find.
1563 // Do not resize after this.
1564 UberSets.resize(UberSetIDs.getNumClasses());
1565 unsigned i = 0;
1566 for (const CodeGenRegister &Reg : Registers) {
1567 unsigned USetID = UberSetIDs[Reg.EnumValue];
1568 if (!USetID)
1569 USetID = ZeroID;
1570 else if (USetID == ZeroID)
1571 USetID = 0;
1572
1573 UberRegSet *USet = &UberSets[USetID];
1574 USet->Regs.push_back(&Reg);
1575 sortAndUniqueRegisters(USet->Regs);
1576 RegSets[i++] = USet;
1577 }
1578}
1579
1580// Recompute each UberSet weight after changing unit weights.
1581static void computeUberWeights(std::vector<UberRegSet> &UberSets,
1582 CodeGenRegBank &RegBank) {
1583 // Skip the first unallocatable set.
1584 for (std::vector<UberRegSet>::iterator I = std::next(UberSets.begin()),
1585 E = UberSets.end(); I != E; ++I) {
1586
1587 // Initialize all unit weights in this set, and remember the max units/reg.
1588 const CodeGenRegister *Reg = nullptr;
1589 unsigned MaxWeight = 0, Weight = 0;
1590 for (RegUnitIterator UnitI(I->Regs); UnitI.isValid(); ++UnitI) {
1591 if (Reg != UnitI.getReg()) {
1592 if (Weight > MaxWeight)
1593 MaxWeight = Weight;
1594 Reg = UnitI.getReg();
1595 Weight = 0;
1596 }
1597 if (!RegBank.getRegUnit(*UnitI).Artificial) {
1598 unsigned UWeight = RegBank.getRegUnit(*UnitI).Weight;
1599 if (!UWeight) {
1600 UWeight = 1;
1601 RegBank.increaseRegUnitWeight(*UnitI, UWeight);
1602 }
1603 Weight += UWeight;
1604 }
1605 }
1606 if (Weight > MaxWeight)
1607 MaxWeight = Weight;
1608 if (I->Weight != MaxWeight) {
1609 LLVM_DEBUG(dbgs() << "UberSet " << I - UberSets.begin() << " Weight "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UberSet " << I
- UberSets.begin() << " Weight " << MaxWeight; for
(auto &Unit : I->Regs) dbgs() << " " << Unit
->getName(); dbgs() << "\n"; } } while (false)
1610 << MaxWeight;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UberSet " << I
- UberSets.begin() << " Weight " << MaxWeight; for
(auto &Unit : I->Regs) dbgs() << " " << Unit
->getName(); dbgs() << "\n"; } } while (false)
1611 for (auto &Unitdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UberSet " << I
- UberSets.begin() << " Weight " << MaxWeight; for
(auto &Unit : I->Regs) dbgs() << " " << Unit
->getName(); dbgs() << "\n"; } } while (false)
1612 : I->Regs) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UberSet " << I
- UberSets.begin() << " Weight " << MaxWeight; for
(auto &Unit : I->Regs) dbgs() << " " << Unit
->getName(); dbgs() << "\n"; } } while (false)
1613 << " " << Unit->getName();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UberSet " << I
- UberSets.begin() << " Weight " << MaxWeight; for
(auto &Unit : I->Regs) dbgs() << " " << Unit
->getName(); dbgs() << "\n"; } } while (false)
1614 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UberSet " << I
- UberSets.begin() << " Weight " << MaxWeight; for
(auto &Unit : I->Regs) dbgs() << " " << Unit
->getName(); dbgs() << "\n"; } } while (false)
;
1615 // Update the set weight.
1616 I->Weight = MaxWeight;
1617 }
1618
1619 // Find singular determinants.
1620 for (const auto R : I->Regs) {
1621 if (R->getRegUnits().count() == 1 && R->getWeight(RegBank) == I->Weight) {
1622 I->SingularDeterminants |= R->getRegUnits();
1623 }
1624 }
1625 }
1626}
1627
1628// normalizeWeight is a computeRegUnitWeights helper that adjusts the weight of
1629// a register and its subregisters so that they have the same weight as their
1630// UberSet. Self-recursion processes the subregister tree in postorder so
1631// subregisters are normalized first.
1632//
1633// Side effects:
1634// - creates new adopted register units
1635// - causes superregisters to inherit adopted units
1636// - increases the weight of "singular" units
1637// - induces recomputation of UberWeights.
1638static bool normalizeWeight(CodeGenRegister *Reg,
1639 std::vector<UberRegSet> &UberSets,
1640 std::vector<UberRegSet*> &RegSets,
1641 BitVector &NormalRegs,
1642 CodeGenRegister::RegUnitList &NormalUnits,
1643 CodeGenRegBank &RegBank) {
1644 NormalRegs.resize(std::max(Reg->EnumValue + 1, NormalRegs.size()));
1645 if (NormalRegs.test(Reg->EnumValue))
1646 return false;
1647 NormalRegs.set(Reg->EnumValue);
1648
1649 bool Changed = false;
1650 const CodeGenRegister::SubRegMap &SRM = Reg->getSubRegs();
1651 for (CodeGenRegister::SubRegMap::const_iterator SRI = SRM.begin(),
1652 SRE = SRM.end(); SRI != SRE; ++SRI) {
1653 if (SRI->second == Reg)
1654 continue; // self-cycles happen
1655
1656 Changed |= normalizeWeight(SRI->second, UberSets, RegSets,
1657 NormalRegs, NormalUnits, RegBank);
1658 }
1659 // Postorder register normalization.
1660
1661 // Inherit register units newly adopted by subregisters.
1662 if (Reg->inheritRegUnits(RegBank))
1663 computeUberWeights(UberSets, RegBank);
1664
1665 // Check if this register is too skinny for its UberRegSet.
1666 UberRegSet *UberSet = RegSets[RegBank.getRegIndex(Reg)];
1667
1668 unsigned RegWeight = Reg->getWeight(RegBank);
1669 if (UberSet->Weight > RegWeight) {
1670 // A register unit's weight can be adjusted only if it is the singular unit
1671 // for this register, has not been used to normalize a subregister's set,
1672 // and has not already been used to singularly determine this UberRegSet.
1673 unsigned AdjustUnit = *Reg->getRegUnits().begin();
1674 if (Reg->getRegUnits().count() != 1
1675 || hasRegUnit(NormalUnits, AdjustUnit)
1676 || hasRegUnit(UberSet->SingularDeterminants, AdjustUnit)) {
1677 // We don't have an adjustable unit, so adopt a new one.
1678 AdjustUnit = RegBank.newRegUnit(UberSet->Weight - RegWeight);
1679 Reg->adoptRegUnit(AdjustUnit);
1680 // Adopting a unit does not immediately require recomputing set weights.
1681 }
1682 else {
1683 // Adjust the existing single unit.
1684 if (!RegBank.getRegUnit(AdjustUnit).Artificial)
1685 RegBank.increaseRegUnitWeight(AdjustUnit, UberSet->Weight - RegWeight);
1686 // The unit may be shared among sets and registers within this set.
1687 computeUberWeights(UberSets, RegBank);
1688 }
1689 Changed = true;
1690 }
1691
1692 // Mark these units normalized so superregisters can't change their weights.
1693 NormalUnits |= Reg->getRegUnits();
1694
1695 return Changed;
1696}
1697
1698// Compute a weight for each register unit created during getSubRegs.
1699//
1700// The goal is that two registers in the same class will have the same weight,
1701// where each register's weight is defined as sum of its units' weights.
1702void CodeGenRegBank::computeRegUnitWeights() {
1703 std::vector<UberRegSet> UberSets;
1704 std::vector<UberRegSet*> RegSets(Registers.size());
1705 computeUberSets(UberSets, RegSets, *this);
1706 // UberSets and RegSets are now immutable.
1707
1708 computeUberWeights(UberSets, *this);
1709
1710 // Iterate over each Register, normalizing the unit weights until reaching
1711 // a fix point.
1712 unsigned NumIters = 0;
1713 for (bool Changed = true; Changed; ++NumIters) {
1714 assert(NumIters <= NumNativeRegUnits && "Runaway register unit weights")((NumIters <= NumNativeRegUnits && "Runaway register unit weights"
) ? static_cast<void> (0) : __assert_fail ("NumIters <= NumNativeRegUnits && \"Runaway register unit weights\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1714, __PRETTY_FUNCTION__))
;
1715 Changed = false;
1716 for (auto &Reg : Registers) {
1717 CodeGenRegister::RegUnitList NormalUnits;
1718 BitVector NormalRegs;
1719 Changed |= normalizeWeight(&Reg, UberSets, RegSets, NormalRegs,
1720 NormalUnits, *this);
1721 }
1722 }
1723}
1724
1725// Find a set in UniqueSets with the same elements as Set.
1726// Return an iterator into UniqueSets.
1727static std::vector<RegUnitSet>::const_iterator
1728findRegUnitSet(const std::vector<RegUnitSet> &UniqueSets,
1729 const RegUnitSet &Set) {
1730 std::vector<RegUnitSet>::const_iterator
1731 I = UniqueSets.begin(), E = UniqueSets.end();
1732 for(;I != E; ++I) {
1733 if (I->Units == Set.Units)
1734 break;
1735 }
1736 return I;
1737}
1738
1739// Return true if the RUSubSet is a subset of RUSuperSet.
1740static bool isRegUnitSubSet(const std::vector<unsigned> &RUSubSet,
1741 const std::vector<unsigned> &RUSuperSet) {
1742 return std::includes(RUSuperSet.begin(), RUSuperSet.end(),
1743 RUSubSet.begin(), RUSubSet.end());
1744}
1745
1746/// Iteratively prune unit sets. Prune subsets that are close to the superset,
1747/// but with one or two registers removed. We occasionally have registers like
1748/// APSR and PC thrown in with the general registers. We also see many
1749/// special-purpose register subsets, such as tail-call and Thumb
1750/// encodings. Generating all possible overlapping sets is combinatorial and
1751/// overkill for modeling pressure. Ideally we could fix this statically in
1752/// tablegen by (1) having the target define register classes that only include
1753/// the allocatable registers and marking other classes as non-allocatable and
1754/// (2) having a way to mark special purpose classes as "don't-care" classes for
1755/// the purpose of pressure. However, we make an attempt to handle targets that
1756/// are not nicely defined by merging nearly identical register unit sets
1757/// statically. This generates smaller tables. Then, dynamically, we adjust the
1758/// set limit by filtering the reserved registers.
1759///
1760/// Merge sets only if the units have the same weight. For example, on ARM,
1761/// Q-tuples with ssub index 0 include all S regs but also include D16+. We
1762/// should not expand the S set to include D regs.
1763void CodeGenRegBank::pruneUnitSets() {
1764 assert(RegClassUnitSets.empty() && "this invalidates RegClassUnitSets")((RegClassUnitSets.empty() && "this invalidates RegClassUnitSets"
) ? static_cast<void> (0) : __assert_fail ("RegClassUnitSets.empty() && \"this invalidates RegClassUnitSets\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1764, __PRETTY_FUNCTION__))
;
1765
1766 // Form an equivalence class of UnitSets with no significant difference.
1767 std::vector<unsigned> SuperSetIDs;
1768 for (unsigned SubIdx = 0, EndIdx = RegUnitSets.size();
1769 SubIdx != EndIdx; ++SubIdx) {
1770 const RegUnitSet &SubSet = RegUnitSets[SubIdx];
1771 unsigned SuperIdx = 0;
1772 for (; SuperIdx != EndIdx; ++SuperIdx) {
1773 if (SuperIdx == SubIdx)
1774 continue;
1775
1776 unsigned UnitWeight = RegUnits[SubSet.Units[0]].Weight;
1777 const RegUnitSet &SuperSet = RegUnitSets[SuperIdx];
1778 if (isRegUnitSubSet(SubSet.Units, SuperSet.Units)
1779 && (SubSet.Units.size() + 3 > SuperSet.Units.size())
1780 && UnitWeight == RegUnits[SuperSet.Units[0]].Weight
1781 && UnitWeight == RegUnits[SuperSet.Units.back()].Weight) {
1782 LLVM_DEBUG(dbgs() << "UnitSet " << SubIdx << " subsumed by " << SuperIdxdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << SubIdx
<< " subsumed by " << SuperIdx << "\n"; } }
while (false)
1783 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << SubIdx
<< " subsumed by " << SuperIdx << "\n"; } }
while (false)
;
1784 // We can pick any of the set names for the merged set. Go for the
1785 // shortest one to avoid picking the name of one of the classes that are
1786 // artificially created by tablegen. So "FPR128_lo" instead of
1787 // "QQQQ_with_qsub3_in_FPR128_lo".
1788 if (RegUnitSets[SubIdx].Name.size() < RegUnitSets[SuperIdx].Name.size())
1789 RegUnitSets[SuperIdx].Name = RegUnitSets[SubIdx].Name;
1790 break;
1791 }
1792 }
1793 if (SuperIdx == EndIdx)
1794 SuperSetIDs.push_back(SubIdx);
1795 }
1796 // Populate PrunedUnitSets with each equivalence class's superset.
1797 std::vector<RegUnitSet> PrunedUnitSets(SuperSetIDs.size());
1798 for (unsigned i = 0, e = SuperSetIDs.size(); i != e; ++i) {
1799 unsigned SuperIdx = SuperSetIDs[i];
1800 PrunedUnitSets[i].Name = RegUnitSets[SuperIdx].Name;
1801 PrunedUnitSets[i].Units.swap(RegUnitSets[SuperIdx].Units);
1802 }
1803 RegUnitSets.swap(PrunedUnitSets);
1804}
1805
1806// Create a RegUnitSet for each RegClass that contains all units in the class
1807// including adopted units that are necessary to model register pressure. Then
1808// iteratively compute RegUnitSets such that the union of any two overlapping
1809// RegUnitSets is repreresented.
1810//
1811// RegisterInfoEmitter will map each RegClass to its RegUnitClass and any
1812// RegUnitSet that is a superset of that RegUnitClass.
1813void CodeGenRegBank::computeRegUnitSets() {
1814 assert(RegUnitSets.empty() && "dirty RegUnitSets")((RegUnitSets.empty() && "dirty RegUnitSets") ? static_cast
<void> (0) : __assert_fail ("RegUnitSets.empty() && \"dirty RegUnitSets\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1814, __PRETTY_FUNCTION__))
;
1815
1816 // Compute a unique RegUnitSet for each RegClass.
1817 auto &RegClasses = getRegClasses();
1818 for (auto &RC : RegClasses) {
1819 if (!RC.Allocatable || RC.Artificial)
1820 continue;
1821
1822 // Speculatively grow the RegUnitSets to hold the new set.
1823 RegUnitSets.resize(RegUnitSets.size() + 1);
1824 RegUnitSets.back().Name = RC.getName();
1825
1826 // Compute a sorted list of units in this class.
1827 RC.buildRegUnitSet(*this, RegUnitSets.back().Units);
1828
1829 // Find an existing RegUnitSet.
1830 std::vector<RegUnitSet>::const_iterator SetI =
1831 findRegUnitSet(RegUnitSets, RegUnitSets.back());
1832 if (SetI != std::prev(RegUnitSets.end()))
1833 RegUnitSets.pop_back();
1834 }
1835
1836 LLVM_DEBUG(dbgs() << "\nBefore pruning:\n"; for (unsigned USIdx = 0,do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1837 USEnd = RegUnitSets.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1838 USIdx < USEnd; ++USIdx) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1839 dbgs() << "UnitSet " << USIdx << " " << RegUnitSets[USIdx].Name << ":";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1840 for (auto &U : RegUnitSets[USIdx].Units)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1841 printRegUnitName(U);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1842 dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
1843 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore pruning:\n"
; for (unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx <
USEnd; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; }; } } while (false)
;
1844
1845 // Iteratively prune unit sets.
1846 pruneUnitSets();
1847
1848 LLVM_DEBUG(dbgs() << "\nBefore union:\n"; for (unsigned USIdx = 0,do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1849 USEnd = RegUnitSets.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1850 USIdx < USEnd; ++USIdx) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1851 dbgs() << "UnitSet " << USIdx << " " << RegUnitSets[USIdx].Name << ":";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1852 for (auto &U : RegUnitSets[USIdx].Units)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1853 printRegUnitName(U);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1854 dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
1855 } dbgs() << "\nUnion sets:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\nBefore union:\n"; for
(unsigned USIdx = 0, USEnd = RegUnitSets.size(); USIdx < USEnd
; ++USIdx) { dbgs() << "UnitSet " << USIdx <<
" " << RegUnitSets[USIdx].Name << ":"; for (auto
&U : RegUnitSets[USIdx].Units) printRegUnitName(U); dbgs
() << "\n"; } dbgs() << "\nUnion sets:\n"; } } while
(false)
;
1856
1857 // Iterate over all unit sets, including new ones added by this loop.
1858 unsigned NumRegUnitSubSets = RegUnitSets.size();
1859 for (unsigned Idx = 0, EndIdx = RegUnitSets.size(); Idx != EndIdx; ++Idx) {
1860 // In theory, this is combinatorial. In practice, it needs to be bounded
1861 // by a small number of sets for regpressure to be efficient.
1862 // If the assert is hit, we need to implement pruning.
1863 assert(Idx < (2*NumRegUnitSubSets) && "runaway unit set inference")((Idx < (2*NumRegUnitSubSets) && "runaway unit set inference"
) ? static_cast<void> (0) : __assert_fail ("Idx < (2*NumRegUnitSubSets) && \"runaway unit set inference\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1863, __PRETTY_FUNCTION__))
;
1864
1865 // Compare new sets with all original classes.
1866 for (unsigned SearchIdx = (Idx >= NumRegUnitSubSets) ? 0 : Idx+1;
1867 SearchIdx != EndIdx; ++SearchIdx) {
1868 std::set<unsigned> Intersection;
1869 std::set_intersection(RegUnitSets[Idx].Units.begin(),
1870 RegUnitSets[Idx].Units.end(),
1871 RegUnitSets[SearchIdx].Units.begin(),
1872 RegUnitSets[SearchIdx].Units.end(),
1873 std::inserter(Intersection, Intersection.begin()));
1874 if (Intersection.empty())
1875 continue;
1876
1877 // Speculatively grow the RegUnitSets to hold the new set.
1878 RegUnitSets.resize(RegUnitSets.size() + 1);
1879 RegUnitSets.back().Name =
1880 RegUnitSets[Idx].Name + "+" + RegUnitSets[SearchIdx].Name;
1881
1882 std::set_union(RegUnitSets[Idx].Units.begin(),
1883 RegUnitSets[Idx].Units.end(),
1884 RegUnitSets[SearchIdx].Units.begin(),
1885 RegUnitSets[SearchIdx].Units.end(),
1886 std::inserter(RegUnitSets.back().Units,
1887 RegUnitSets.back().Units.begin()));
1888
1889 // Find an existing RegUnitSet, or add the union to the unique sets.
1890 std::vector<RegUnitSet>::const_iterator SetI =
1891 findRegUnitSet(RegUnitSets, RegUnitSets.back());
1892 if (SetI != std::prev(RegUnitSets.end()))
1893 RegUnitSets.pop_back();
1894 else {
1895 LLVM_DEBUG(dbgs() << "UnitSet " << RegUnitSets.size() - 1 << " "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << RegUnitSets
.size() - 1 << " " << RegUnitSets.back().Name <<
":"; for (auto &U : RegUnitSets.back().Units) printRegUnitName
(U); dbgs() << "\n";; } } while (false)
1896 << RegUnitSets.back().Name << ":";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << RegUnitSets
.size() - 1 << " " << RegUnitSets.back().Name <<
":"; for (auto &U : RegUnitSets.back().Units) printRegUnitName
(U); dbgs() << "\n";; } } while (false)
1897 for (auto &Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << RegUnitSets
.size() - 1 << " " << RegUnitSets.back().Name <<
":"; for (auto &U : RegUnitSets.back().Units) printRegUnitName
(U); dbgs() << "\n";; } } while (false)
1898 : RegUnitSets.back().Units) printRegUnitName(U);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << RegUnitSets
.size() - 1 << " " << RegUnitSets.back().Name <<
":"; for (auto &U : RegUnitSets.back().Units) printRegUnitName
(U); dbgs() << "\n";; } } while (false)
1899 dbgs() << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "UnitSet " << RegUnitSets
.size() - 1 << " " << RegUnitSets.back().Name <<
":"; for (auto &U : RegUnitSets.back().Units) printRegUnitName
(U); dbgs() << "\n";; } } while (false)
;
1900 }
1901 }
1902 }
1903
1904 // Iteratively prune unit sets after inferring supersets.
1905 pruneUnitSets();
1906
1907 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1908 dbgs() << "\n"; for (unsigned USIdx = 0, USEnd = RegUnitSets.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1909 USIdx < USEnd; ++USIdx) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1910 dbgs() << "UnitSet " << USIdx << " " << RegUnitSets[USIdx].Name << ":";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1911 for (auto &U : RegUnitSets[USIdx].Units)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1912 printRegUnitName(U);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1913 dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
1914 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; for (unsigned USIdx
= 0, USEnd = RegUnitSets.size(); USIdx < USEnd; ++USIdx) {
dbgs() << "UnitSet " << USIdx << " " <<
RegUnitSets[USIdx].Name << ":"; for (auto &U : RegUnitSets
[USIdx].Units) printRegUnitName(U); dbgs() << "\n"; }; }
} while (false)
;
1915
1916 // For each register class, list the UnitSets that are supersets.
1917 RegClassUnitSets.resize(RegClasses.size());
1918 int RCIdx = -1;
1919 for (auto &RC : RegClasses) {
1920 ++RCIdx;
1921 if (!RC.Allocatable)
1922 continue;
1923
1924 // Recompute the sorted list of units in this class.
1925 std::vector<unsigned> RCRegUnits;
1926 RC.buildRegUnitSet(*this, RCRegUnits);
1927
1928 // Don't increase pressure for unallocatable regclasses.
1929 if (RCRegUnits.empty())
1930 continue;
1931
1932 LLVM_DEBUG(dbgs() << "RC " << RC.getName() << " Units: \n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "RC " << RC.getName
() << " Units: \n"; for (auto U : RCRegUnits) printRegUnitName
(U); dbgs() << "\n UnitSetIDs:"; } } while (false)
1933 for (auto Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "RC " << RC.getName
() << " Units: \n"; for (auto U : RCRegUnits) printRegUnitName
(U); dbgs() << "\n UnitSetIDs:"; } } while (false)
1934 : RCRegUnits) printRegUnitName(U);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "RC " << RC.getName
() << " Units: \n"; for (auto U : RCRegUnits) printRegUnitName
(U); dbgs() << "\n UnitSetIDs:"; } } while (false)
1935 dbgs() << "\n UnitSetIDs:")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "RC " << RC.getName
() << " Units: \n"; for (auto U : RCRegUnits) printRegUnitName
(U); dbgs() << "\n UnitSetIDs:"; } } while (false)
;
1936
1937 // Find all supersets.
1938 for (unsigned USIdx = 0, USEnd = RegUnitSets.size();
1939 USIdx != USEnd; ++USIdx) {
1940 if (isRegUnitSubSet(RCRegUnits, RegUnitSets[USIdx].Units)) {
1941 LLVM_DEBUG(dbgs() << " " << USIdx)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << " " << USIdx; }
} while (false)
;
1942 RegClassUnitSets[RCIdx].push_back(USIdx);
1943 }
1944 }
1945 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("regalloc-emitter")) { dbgs() << "\n"; } } while (false
)
;
1946 assert(!RegClassUnitSets[RCIdx].empty() && "missing unit set for regclass")((!RegClassUnitSets[RCIdx].empty() && "missing unit set for regclass"
) ? static_cast<void> (0) : __assert_fail ("!RegClassUnitSets[RCIdx].empty() && \"missing unit set for regclass\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 1946, __PRETTY_FUNCTION__))
;
1947 }
1948
1949 // For each register unit, ensure that we have the list of UnitSets that
1950 // contain the unit. Normally, this matches an existing list of UnitSets for a
1951 // register class. If not, we create a new entry in RegClassUnitSets as a
1952 // "fake" register class.
1953 for (unsigned UnitIdx = 0, UnitEnd = NumNativeRegUnits;
1954 UnitIdx < UnitEnd; ++UnitIdx) {
1955 std::vector<unsigned> RUSets;
1956 for (unsigned i = 0, e = RegUnitSets.size(); i != e; ++i) {
1957 RegUnitSet &RUSet = RegUnitSets[i];
1958 if (!is_contained(RUSet.Units, UnitIdx))
1959 continue;
1960 RUSets.push_back(i);
1961 }
1962 unsigned RCUnitSetsIdx = 0;
1963 for (unsigned e = RegClassUnitSets.size();
1964 RCUnitSetsIdx != e; ++RCUnitSetsIdx) {
1965 if (RegClassUnitSets[RCUnitSetsIdx] == RUSets) {
1966 break;
1967 }
1968 }
1969 RegUnits[UnitIdx].RegClassUnitSetsIdx = RCUnitSetsIdx;
1970 if (RCUnitSetsIdx == RegClassUnitSets.size()) {
1971 // Create a new list of UnitSets as a "fake" register class.
1972 RegClassUnitSets.resize(RCUnitSetsIdx + 1);
1973 RegClassUnitSets[RCUnitSetsIdx].swap(RUSets);
1974 }
1975 }
1976}
1977
1978void CodeGenRegBank::computeRegUnitLaneMasks() {
1979 for (auto &Register : Registers) {
1980 // Create an initial lane mask for all register units.
1981 const auto &RegUnits = Register.getRegUnits();
1982 CodeGenRegister::RegUnitLaneMaskList
1983 RegUnitLaneMasks(RegUnits.count(), LaneBitmask::getNone());
1984 // Iterate through SubRegisters.
1985 typedef CodeGenRegister::SubRegMap SubRegMap;
1986 const SubRegMap &SubRegs = Register.getSubRegs();
1987 for (SubRegMap::const_iterator S = SubRegs.begin(),
1988 SE = SubRegs.end(); S != SE; ++S) {
1989 CodeGenRegister *SubReg = S->second;
1990 // Ignore non-leaf subregisters, their lane masks are fully covered by
1991 // the leaf subregisters anyway.
1992 if (!SubReg->getSubRegs().empty())
1993 continue;
1994 CodeGenSubRegIndex *SubRegIndex = S->first;
1995 const CodeGenRegister *SubRegister = S->second;
1996 LaneBitmask LaneMask = SubRegIndex->LaneMask;
1997 // Distribute LaneMask to Register Units touched.
1998 for (unsigned SUI : SubRegister->getRegUnits()) {
1999 bool Found = false;
2000 unsigned u = 0;
2001 for (unsigned RU : RegUnits) {
2002 if (SUI == RU) {
2003 RegUnitLaneMasks[u] |= LaneMask;
2004 assert(!Found)((!Found) ? static_cast<void> (0) : __assert_fail ("!Found"
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 2004, __PRETTY_FUNCTION__))
;
2005 Found = true;
2006 }
2007 ++u;
2008 }
2009 (void)Found;
2010 assert(Found)((Found) ? static_cast<void> (0) : __assert_fail ("Found"
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 2010, __PRETTY_FUNCTION__))
;
2011 }
2012 }
2013 Register.setRegUnitLaneMasks(RegUnitLaneMasks);
2014 }
2015}
2016
2017void CodeGenRegBank::computeDerivedInfo() {
2018 computeComposites();
2019 computeSubRegLaneMasks();
1
Calling 'CodeGenRegBank::computeSubRegLaneMasks'
2020
2021 // Compute a weight for each register unit created during getSubRegs.
2022 // This may create adopted register units (with unit # >= NumNativeRegUnits).
2023 computeRegUnitWeights();
2024
2025 // Compute a unique set of RegUnitSets. One for each RegClass and inferred
2026 // supersets for the union of overlapping sets.
2027 computeRegUnitSets();
2028
2029 computeRegUnitLaneMasks();
2030
2031 // Compute register class HasDisjunctSubRegs/CoveredBySubRegs flag.
2032 for (CodeGenRegisterClass &RC : RegClasses) {
2033 RC.HasDisjunctSubRegs = false;
2034 RC.CoveredBySubRegs = true;
2035 for (const CodeGenRegister *Reg : RC.getMembers()) {
2036 RC.HasDisjunctSubRegs |= Reg->HasDisjunctSubRegs;
2037 RC.CoveredBySubRegs &= Reg->CoveredBySubRegs;
2038 }
2039 }
2040
2041 // Get the weight of each set.
2042 for (unsigned Idx = 0, EndIdx = RegUnitSets.size(); Idx != EndIdx; ++Idx)
2043 RegUnitSets[Idx].Weight = getRegUnitSetWeight(RegUnitSets[Idx].Units);
2044
2045 // Find the order of each set.
2046 RegUnitSetOrder.reserve(RegUnitSets.size());
2047 for (unsigned Idx = 0, EndIdx = RegUnitSets.size(); Idx != EndIdx; ++Idx)
2048 RegUnitSetOrder.push_back(Idx);
2049
2050 std::stable_sort(RegUnitSetOrder.begin(), RegUnitSetOrder.end(),
2051 [this](unsigned ID1, unsigned ID2) {
2052 return getRegPressureSet(ID1).Units.size() <
2053 getRegPressureSet(ID2).Units.size();
2054 });
2055 for (unsigned Idx = 0, EndIdx = RegUnitSets.size(); Idx != EndIdx; ++Idx) {
2056 RegUnitSets[RegUnitSetOrder[Idx]].Order = Idx;
2057 }
2058}
2059
2060//
2061// Synthesize missing register class intersections.
2062//
2063// Make sure that sub-classes of RC exists such that getCommonSubClass(RC, X)
2064// returns a maximal register class for all X.
2065//
2066void CodeGenRegBank::inferCommonSubClass(CodeGenRegisterClass *RC) {
2067 assert(!RegClasses.empty())((!RegClasses.empty()) ? static_cast<void> (0) : __assert_fail
("!RegClasses.empty()", "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 2067, __PRETTY_FUNCTION__))
;
2068 // Stash the iterator to the last element so that this loop doesn't visit
2069 // elements added by the getOrCreateSubClass call within it.
2070 for (auto I = RegClasses.begin(), E = std::prev(RegClasses.end());
2071 I != std::next(E); ++I) {
2072 CodeGenRegisterClass *RC1 = RC;
2073 CodeGenRegisterClass *RC2 = &*I;
2074 if (RC1 == RC2)
2075 continue;
2076
2077 // Compute the set intersection of RC1 and RC2.
2078 const CodeGenRegister::Vec &Memb1 = RC1->getMembers();
2079 const CodeGenRegister::Vec &Memb2 = RC2->getMembers();
2080 CodeGenRegister::Vec Intersection;
2081 std::set_intersection(
2082 Memb1.begin(), Memb1.end(), Memb2.begin(), Memb2.end(),
2083 std::inserter(Intersection, Intersection.begin()), deref<llvm::less>());
2084
2085 // Skip disjoint class pairs.
2086 if (Intersection.empty())
2087 continue;
2088
2089 // If RC1 and RC2 have different spill sizes or alignments, use the
2090 // stricter one for sub-classing. If they are equal, prefer RC1.
2091 if (RC2->RSI.hasStricterSpillThan(RC1->RSI))
2092 std::swap(RC1, RC2);
2093
2094 getOrCreateSubClass(RC1, &Intersection,
2095 RC1->getName() + "_and_" + RC2->getName());
2096 }
2097}
2098
2099//
2100// Synthesize missing sub-classes for getSubClassWithSubReg().
2101//
2102// Make sure that the set of registers in RC with a given SubIdx sub-register
2103// form a register class. Update RC->SubClassWithSubReg.
2104//
2105void CodeGenRegBank::inferSubClassWithSubReg(CodeGenRegisterClass *RC) {
2106 // Map SubRegIndex to set of registers in RC supporting that SubRegIndex.
2107 typedef std::map<const CodeGenSubRegIndex *, CodeGenRegister::Vec,
2108 deref<llvm::less>> SubReg2SetMap;
2109
2110 // Compute the set of registers supporting each SubRegIndex.
2111 SubReg2SetMap SRSets;
2112 for (const auto R : RC->getMembers()) {
2113 if (R->Artificial)
2114 continue;
2115 const CodeGenRegister::SubRegMap &SRM = R->getSubRegs();
2116 for (CodeGenRegister::SubRegMap::const_iterator I = SRM.begin(),
2117 E = SRM.end(); I != E; ++I) {
2118 if (!I->first->Artificial)
2119 SRSets[I->first].push_back(R);
2120 }
2121 }
2122
2123 for (auto I : SRSets)
2124 sortAndUniqueRegisters(I.second);
2125
2126 // Find matching classes for all SRSets entries. Iterate in SubRegIndex
2127 // numerical order to visit synthetic indices last.
2128 for (const auto &SubIdx : SubRegIndices) {
2129 if (SubIdx.Artificial)
2130 continue;
2131 SubReg2SetMap::const_iterator I = SRSets.find(&SubIdx);
2132 // Unsupported SubRegIndex. Skip it.
2133 if (I == SRSets.end())
2134 continue;
2135 // In most cases, all RC registers support the SubRegIndex.
2136 if (I->second.size() == RC->getMembers().size()) {
2137 RC->setSubClassWithSubReg(&SubIdx, RC);
2138 continue;
2139 }
2140 // This is a real subset. See if we have a matching class.
2141 CodeGenRegisterClass *SubRC =
2142 getOrCreateSubClass(RC, &I->second,
2143 RC->getName() + "_with_" + I->first->getName());
2144 RC->setSubClassWithSubReg(&SubIdx, SubRC);
2145 }
2146}
2147
2148//
2149// Synthesize missing sub-classes of RC for getMatchingSuperRegClass().
2150//
2151// Create sub-classes of RC such that getMatchingSuperRegClass(RC, SubIdx, X)
2152// has a maximal result for any SubIdx and any X >= FirstSubRegRC.
2153//
2154
2155void CodeGenRegBank::inferMatchingSuperRegClass(CodeGenRegisterClass *RC,
2156 std::list<CodeGenRegisterClass>::iterator FirstSubRegRC) {
2157 SmallVector<std::pair<const CodeGenRegister*,
2158 const CodeGenRegister*>, 16> SSPairs;
2159 BitVector TopoSigs(getNumTopoSigs());
2160
2161 // Iterate in SubRegIndex numerical order to visit synthetic indices last.
2162 for (auto &SubIdx : SubRegIndices) {
2163 // Skip indexes that aren't fully supported by RC's registers. This was
2164 // computed by inferSubClassWithSubReg() above which should have been
2165 // called first.
2166 if (RC->getSubClassWithSubReg(&SubIdx) != RC)
2167 continue;
2168
2169 // Build list of (Super, Sub) pairs for this SubIdx.
2170 SSPairs.clear();
2171 TopoSigs.reset();
2172 for (const auto Super : RC->getMembers()) {
2173 const CodeGenRegister *Sub = Super->getSubRegs().find(&SubIdx)->second;
2174 assert(Sub && "Missing sub-register")((Sub && "Missing sub-register") ? static_cast<void
> (0) : __assert_fail ("Sub && \"Missing sub-register\""
, "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 2174, __PRETTY_FUNCTION__))
;
2175 SSPairs.push_back(std::make_pair(Super, Sub));
2176 TopoSigs.set(Sub->getTopoSig());
2177 }
2178
2179 // Iterate over sub-register class candidates. Ignore classes created by
2180 // this loop. They will never be useful.
2181 // Store an iterator to the last element (not end) so that this loop doesn't
2182 // visit newly inserted elements.
2183 assert(!RegClasses.empty())((!RegClasses.empty()) ? static_cast<void> (0) : __assert_fail
("!RegClasses.empty()", "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 2183, __PRETTY_FUNCTION__))
;
2184 for (auto I = FirstSubRegRC, E = std::prev(RegClasses.end());
2185 I != std::next(E); ++I) {
2186 CodeGenRegisterClass &SubRC = *I;
2187 if (SubRC.Artificial)
2188 continue;
2189 // Topological shortcut: SubRC members have the wrong shape.
2190 if (!TopoSigs.anyCommon(SubRC.getTopoSigs()))
2191 continue;
2192 // Compute the subset of RC that maps into SubRC.
2193 CodeGenRegister::Vec SubSetVec;
2194 for (unsigned i = 0, e = SSPairs.size(); i != e; ++i)
2195 if (SubRC.contains(SSPairs[i].second))
2196 SubSetVec.push_back(SSPairs[i].first);
2197
2198 if (SubSetVec.empty())
2199 continue;
2200
2201 // RC injects completely into SubRC.
2202 sortAndUniqueRegisters(SubSetVec);
2203 if (SubSetVec.size() == SSPairs.size()) {
2204 SubRC.addSuperRegClass(&SubIdx, RC);
2205 continue;
2206 }
2207
2208 // Only a subset of RC maps into SubRC. Make sure it is represented by a
2209 // class.
2210 getOrCreateSubClass(RC, &SubSetVec, RC->getName() + "_with_" +
2211 SubIdx.getName() + "_in_" +
2212 SubRC.getName());
2213 }
2214 }
2215}
2216
2217//
2218// Infer missing register classes.
2219//
2220void CodeGenRegBank::computeInferredRegisterClasses() {
2221 assert(!RegClasses.empty())((!RegClasses.empty()) ? static_cast<void> (0) : __assert_fail
("!RegClasses.empty()", "/build/llvm-toolchain-snapshot-8~svn345461/utils/TableGen/CodeGenRegisters.cpp"
, 2221, __PRETTY_FUNCTION__))
;
2222 // When this function is called, the register classes have not been sorted
2223 // and assigned EnumValues yet. That means getSubClasses(),
2224 // getSuperClasses(), and hasSubClass() functions are defunct.
2225
2226 // Use one-before-the-end so it doesn't move forward when new elements are
2227 // added.
2228 auto FirstNewRC = std::prev(RegClasses.end());
2229
2230 // Visit all register classes, including the ones being added by the loop.
2231 // Watch out for iterator invalidation here.
2232 for (auto I = RegClasses.begin(), E = RegClasses.end(); I != E; ++I) {
2233 CodeGenRegisterClass *RC = &*I;
2234 if (RC->Artificial)
2235 continue;
2236
2237 // Synthesize answers for getSubClassWithSubReg().
2238 inferSubClassWithSubReg(RC);
2239
2240 // Synthesize answers for getCommonSubClass().
2241 inferCommonSubClass(RC);
2242
2243 // Synthesize answers for getMatchingSuperRegClass().
2244 inferMatchingSuperRegClass(RC);
2245
2246 // New register classes are created while this loop is running, and we need
2247 // to visit all of them. I particular, inferMatchingSuperRegClass needs
2248 // to match old super-register classes with sub-register classes created
2249 // after inferMatchingSuperRegClass was called. At this point,
2250 // inferMatchingSuperRegClass has checked SuperRC = [0..rci] with SubRC =
2251 // [0..FirstNewRC). We need to cover SubRC = [FirstNewRC..rci].
2252 if (I == FirstNewRC) {
2253 auto NextNewRC = std::prev(RegClasses.end());
2254 for (auto I2 = RegClasses.begin(), E2 = std::next(FirstNewRC); I2 != E2;
2255 ++I2)
2256 inferMatchingSuperRegClass(&*I2, E2);
2257 FirstNewRC = NextNewRC;
2258 }
2259 }
2260}
2261
2262/// getRegisterClassForRegister - Find the register class that contains the
2263/// specified physical register. If the register is not in a register class,
2264/// return null. If the register is in multiple classes, and the classes have a
2265/// superset-subset relationship and the same set of types, return the
2266/// superclass. Otherwise return null.
2267const CodeGenRegisterClass*
2268CodeGenRegBank::getRegClassForRegister(Record *R) {
2269 const CodeGenRegister *Reg = getReg(R);
2270 const CodeGenRegisterClass *FoundRC = nullptr;
2271 for (const auto &RC : getRegClasses()) {
2272 if (!RC.contains(Reg))
2273 continue;
2274
2275 // If this is the first class that contains the register,
2276 // make a note of it and go on to the next class.
2277 if (!FoundRC) {
2278 FoundRC = &RC;
2279 continue;
2280 }
2281
2282 // If a register's classes have different types, return null.
2283 if (RC.getValueTypes() != FoundRC->getValueTypes())
2284 return nullptr;
2285
2286 // Check to see if the previously found class that contains
2287 // the register is a subclass of the current class. If so,
2288 // prefer the superclass.
2289 if (RC.hasSubClass(FoundRC)) {
2290 FoundRC = &RC;
2291 continue;
2292 }
2293
2294 // Check to see if the previously found class that contains
2295 // the register is a superclass of the current class. If so,
2296 // prefer the superclass.
2297 if (FoundRC->hasSubClass(&RC))
2298 continue;
2299
2300 // Multiple classes, and neither is a superclass of the other.
2301 // Return null.
2302 return nullptr;
2303 }
2304 return FoundRC;
2305}
2306
2307BitVector CodeGenRegBank::computeCoveredRegisters(ArrayRef<Record*> Regs) {
2308 SetVector<const CodeGenRegister*> Set;
2309
2310 // First add Regs with all sub-registers.
2311 for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
2312 CodeGenRegister *Reg = getReg(Regs[i]);
2313 if (Set.insert(Reg))
2314 // Reg is new, add all sub-registers.
2315 // The pre-ordering is not important here.
2316 Reg->addSubRegsPreOrder(Set, *this);
2317 }
2318
2319 // Second, find all super-registers that are completely covered by the set.
2320 for (unsigned i = 0; i != Set.size(); ++i) {
2321 const CodeGenRegister::SuperRegList &SR = Set[i]->getSuperRegs();
2322 for (unsigned j = 0, e = SR.size(); j != e; ++j) {
2323 const CodeGenRegister *Super = SR[j];
2324 if (!Super->CoveredBySubRegs || Set.count(Super))
2325 continue;
2326 // This new super-register is covered by its sub-registers.
2327 bool AllSubsInSet = true;
2328 const CodeGenRegister::SubRegMap &SRM = Super->getSubRegs();
2329 for (CodeGenRegister::SubRegMap::const_iterator I = SRM.begin(),
2330 E = SRM.end(); I != E; ++I)
2331 if (!Set.count(I->second)) {
2332 AllSubsInSet = false;
2333 break;
2334 }
2335 // All sub-registers in Set, add Super as well.
2336 // We will visit Super later to recheck its super-registers.
2337 if (AllSubsInSet)
2338 Set.insert(Super);
2339 }
2340 }
2341
2342 // Convert to BitVector.
2343 BitVector BV(Registers.size() + 1);
2344 for (unsigned i = 0, e = Set.size(); i != e; ++i)
2345 BV.set(Set[i]->EnumValue);
2346 return BV;
2347}
2348
2349void CodeGenRegBank::printRegUnitName(unsigned Unit) const {
2350 if (Unit < NumNativeRegUnits)
2351 dbgs() << ' ' << RegUnits[Unit].Roots[0]->getName();
2352 else
2353 dbgs() << " #" << Unit;
2354}

/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/MC/LaneBitmask.h

1//===- llvm/MC/LaneBitmask.h ------------------------------------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9///
10/// \file
11/// A common definition of LaneBitmask for use in TableGen and CodeGen.
12///
13/// A lane mask is a bitmask representing the covering of a register with
14/// sub-registers.
15///
16/// This is typically used to track liveness at sub-register granularity.
17/// Lane masks for sub-register indices are similar to register units for
18/// physical registers. The individual bits in a lane mask can't be assigned
19/// any specific meaning. They can be used to check if two sub-register
20/// indices overlap.
21///
22/// Iff the target has a register such that:
23///
24/// getSubReg(Reg, A) overlaps getSubReg(Reg, B)
25///
26/// then:
27///
28/// (getSubRegIndexLaneMask(A) & getSubRegIndexLaneMask(B)) != 0
29
30#ifndef LLVM_MC_LANEBITMASK_H
31#define LLVM_MC_LANEBITMASK_H
32
33#include "llvm/Support/Compiler.h"
34#include "llvm/Support/Format.h"
35#include "llvm/Support/Printable.h"
36#include "llvm/Support/raw_ostream.h"
37
38namespace llvm {
39
40 struct LaneBitmask {
41 // When changing the underlying type, change the format string as well.
42 using Type = unsigned;
43 enum : unsigned { BitWidth = 8*sizeof(Type) };
44 constexpr static const char *const FormatStr = "%08X";
45
46 constexpr LaneBitmask() = default;
47 explicit constexpr LaneBitmask(Type V) : Mask(V) {}
48
49 constexpr bool operator== (LaneBitmask M) const { return Mask == M.Mask; }
50 constexpr bool operator!= (LaneBitmask M) const { return Mask != M.Mask; }
51 constexpr bool operator< (LaneBitmask M) const { return Mask < M.Mask; }
52 constexpr bool none() const { return Mask == 0; }
53 constexpr bool any() const { return Mask != 0; }
54 constexpr bool all() const { return ~Mask == 0; }
55
56 constexpr LaneBitmask operator~() const {
57 return LaneBitmask(~Mask);
58 }
59 constexpr LaneBitmask operator|(LaneBitmask M) const {
60 return LaneBitmask(Mask | M.Mask);
61 }
62 constexpr LaneBitmask operator&(LaneBitmask M) const {
63 return LaneBitmask(Mask & M.Mask);
64 }
65 LaneBitmask &operator|=(LaneBitmask M) {
66 Mask |= M.Mask;
67 return *this;
68 }
69 LaneBitmask &operator&=(LaneBitmask M) {
70 Mask &= M.Mask;
71 return *this;
72 }
73
74 constexpr Type getAsInteger() const { return Mask; }
75
76 unsigned getNumLanes() const {
77 return countPopulation(Mask);
78 }
79 unsigned getHighestLane() const {
80 return Log2_32(Mask);
5
Calling 'Log2_32'
7
Returning from 'Log2_32'
8
Returning the value 4294967295
81 }
82
83 static constexpr LaneBitmask getNone() { return LaneBitmask(0); }
84 static constexpr LaneBitmask getAll() { return ~LaneBitmask(0); }
85 static constexpr LaneBitmask getLane(unsigned Lane) {
86 return LaneBitmask(Type(1) << Lane);
12
The result of the left shift is undefined due to shifting by '4294967295', which is greater or equal to the width of type 'llvm::LaneBitmask::Type'
87 }
88
89 private:
90 Type Mask = 0;
91 };
92
93 /// Create Printable object to print LaneBitmasks on a \ref raw_ostream.
94 inline Printable PrintLaneMask(LaneBitmask LaneMask) {
95 return Printable([LaneMask](raw_ostream &OS) {
96 OS << format(LaneBitmask::FormatStr, LaneMask.getAsInteger());
97 });
98 }
99
100} // end namespace llvm
101
102#endif // LLVM_MC_LANEBITMASK_H

/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h

1//===-- llvm/Support/MathExtras.h - Useful math functions -------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file contains some functions that are useful for math stuff.
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef LLVM_SUPPORT_MATHEXTRAS_H
15#define LLVM_SUPPORT_MATHEXTRAS_H
16
17#include "llvm/Support/Compiler.h"
18#include "llvm/Support/SwapByteOrder.h"
19#include <algorithm>
20#include <cassert>
21#include <climits>
22#include <cstring>
23#include <limits>
24#include <type_traits>
25
26#ifdef __ANDROID_NDK__
27#include <android/api-level.h>
28#endif
29
30#ifdef _MSC_VER
31// Declare these intrinsics manually rather including intrin.h. It's very
32// expensive, and MathExtras.h is popular.
33// #include <intrin.h>
34extern "C" {
35unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);
36unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);
37unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);
38unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);
39}
40#endif
41
42namespace llvm {
43/// The behavior an operation has on an input of 0.
44enum ZeroBehavior {
45 /// The returned value is undefined.
46 ZB_Undefined,
47 /// The returned value is numeric_limits<T>::max()
48 ZB_Max,
49 /// The returned value is numeric_limits<T>::digits
50 ZB_Width
51};
52
53namespace detail {
54template <typename T, std::size_t SizeOfT> struct TrailingZerosCounter {
55 static std::size_t count(T Val, ZeroBehavior) {
56 if (!Val)
57 return std::numeric_limits<T>::digits;
58 if (Val & 0x1)
59 return 0;
60
61 // Bisection method.
62 std::size_t ZeroBits = 0;
63 T Shift = std::numeric_limits<T>::digits >> 1;
64 T Mask = std::numeric_limits<T>::max() >> Shift;
65 while (Shift) {
66 if ((Val & Mask) == 0) {
67 Val >>= Shift;
68 ZeroBits |= Shift;
69 }
70 Shift >>= 1;
71 Mask >>= Shift;
72 }
73 return ZeroBits;
74 }
75};
76
77#if __GNUC__4 >= 4 || defined(_MSC_VER)
78template <typename T> struct TrailingZerosCounter<T, 4> {
79 static std::size_t count(T Val, ZeroBehavior ZB) {
80 if (ZB != ZB_Undefined && Val == 0)
81 return 32;
82
83#if __has_builtin(__builtin_ctz)1 || LLVM_GNUC_PREREQ(4, 0, 0)((4 << 20) + (2 << 10) + 1 >= ((4) << 20
) + ((0) << 10) + (0))
84 return __builtin_ctz(Val);
85#elif defined(_MSC_VER)
86 unsigned long Index;
87 _BitScanForward(&Index, Val);
88 return Index;
89#endif
90 }
91};
92
93#if !defined(_MSC_VER) || defined(_M_X64)
94template <typename T> struct TrailingZerosCounter<T, 8> {
95 static std::size_t count(T Val, ZeroBehavior ZB) {
96 if (ZB != ZB_Undefined && Val == 0)
97 return 64;
98
99#if __has_builtin(__builtin_ctzll)1 || LLVM_GNUC_PREREQ(4, 0, 0)((4 << 20) + (2 << 10) + 1 >= ((4) << 20
) + ((0) << 10) + (0))
100 return __builtin_ctzll(Val);
101#elif defined(_MSC_VER)
102 unsigned long Index;
103 _BitScanForward64(&Index, Val);
104 return Index;
105#endif
106 }
107};
108#endif
109#endif
110} // namespace detail
111
112/// Count number of 0's from the least significant bit to the most
113/// stopping at the first 1.
114///
115/// Only unsigned integral types are allowed.
116///
117/// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are
118/// valid arguments.
119template <typename T>
120std::size_t countTrailingZeros(T Val, ZeroBehavior ZB = ZB_Width) {
121 static_assert(std::numeric_limits<T>::is_integer &&
122 !std::numeric_limits<T>::is_signed,
123 "Only unsigned integral types are allowed.");
124 return llvm::detail::TrailingZerosCounter<T, sizeof(T)>::count(Val, ZB);
125}
126
127namespace detail {
128template <typename T, std::size_t SizeOfT> struct LeadingZerosCounter {
129 static std::size_t count(T Val, ZeroBehavior) {
130 if (!Val)
131 return std::numeric_limits<T>::digits;
132
133 // Bisection method.
134 std::size_t ZeroBits = 0;
135 for (T Shift = std::numeric_limits<T>::digits >> 1; Shift; Shift >>= 1) {
136 T Tmp = Val >> Shift;
137 if (Tmp)
138 Val = Tmp;
139 else
140 ZeroBits |= Shift;
141 }
142 return ZeroBits;
143 }
144};
145
146#if __GNUC__4 >= 4 || defined(_MSC_VER)
147template <typename T> struct LeadingZerosCounter<T, 4> {
148 static std::size_t count(T Val, ZeroBehavior ZB) {
149 if (ZB != ZB_Undefined && Val == 0)
150 return 32;
151
152#if __has_builtin(__builtin_clz)1 || LLVM_GNUC_PREREQ(4, 0, 0)((4 << 20) + (2 << 10) + 1 >= ((4) << 20
) + ((0) << 10) + (0))
153 return __builtin_clz(Val);
154#elif defined(_MSC_VER)
155 unsigned long Index;
156 _BitScanReverse(&Index, Val);
157 return Index ^ 31;
158#endif
159 }
160};
161
162#if !defined(_MSC_VER) || defined(_M_X64)
163template <typename T> struct LeadingZerosCounter<T, 8> {
164 static std::size_t count(T Val, ZeroBehavior ZB) {
165 if (ZB != ZB_Undefined && Val == 0)
166 return 64;
167
168#if __has_builtin(__builtin_clzll)1 || LLVM_GNUC_PREREQ(4, 0, 0)((4 << 20) + (2 << 10) + 1 >= ((4) << 20
) + ((0) << 10) + (0))
169 return __builtin_clzll(Val);
170#elif defined(_MSC_VER)
171 unsigned long Index;
172 _BitScanReverse64(&Index, Val);
173 return Index ^ 63;
174#endif
175 }
176};
177#endif
178#endif
179} // namespace detail
180
181/// Count number of 0's from the most significant bit to the least
182/// stopping at the first 1.
183///
184/// Only unsigned integral types are allowed.
185///
186/// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are
187/// valid arguments.
188template <typename T>
189std::size_t countLeadingZeros(T Val, ZeroBehavior ZB = ZB_Width) {
190 static_assert(std::numeric_limits<T>::is_integer &&
191 !std::numeric_limits<T>::is_signed,
192 "Only unsigned integral types are allowed.");
193 return llvm::detail::LeadingZerosCounter<T, sizeof(T)>::count(Val, ZB);
194}
195
196/// Get the index of the first set bit starting from the least
197/// significant bit.
198///
199/// Only unsigned integral types are allowed.
200///
201/// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are
202/// valid arguments.
203template <typename T> T findFirstSet(T Val, ZeroBehavior ZB = ZB_Max) {
204 if (ZB == ZB_Max && Val == 0)
205 return std::numeric_limits<T>::max();
206
207 return countTrailingZeros(Val, ZB_Undefined);
208}
209
210/// Create a bitmask with the N right-most bits set to 1, and all other
211/// bits set to 0. Only unsigned types are allowed.
212template <typename T> T maskTrailingOnes(unsigned N) {
213 static_assert(std::is_unsigned<T>::value, "Invalid type!");
214 const unsigned Bits = CHAR_BIT8 * sizeof(T);
215 assert(N <= Bits && "Invalid bit index")((N <= Bits && "Invalid bit index") ? static_cast<
void> (0) : __assert_fail ("N <= Bits && \"Invalid bit index\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 215, __PRETTY_FUNCTION__))
;
216 return N == 0 ? 0 : (T(-1) >> (Bits - N));
217}
218
219/// Create a bitmask with the N left-most bits set to 1, and all other
220/// bits set to 0. Only unsigned types are allowed.
221template <typename T> T maskLeadingOnes(unsigned N) {
222 return ~maskTrailingOnes<T>(CHAR_BIT8 * sizeof(T) - N);
223}
224
225/// Create a bitmask with the N right-most bits set to 0, and all other
226/// bits set to 1. Only unsigned types are allowed.
227template <typename T> T maskTrailingZeros(unsigned N) {
228 return maskLeadingOnes<T>(CHAR_BIT8 * sizeof(T) - N);
229}
230
231/// Create a bitmask with the N left-most bits set to 0, and all other
232/// bits set to 1. Only unsigned types are allowed.
233template <typename T> T maskLeadingZeros(unsigned N) {
234 return maskTrailingOnes<T>(CHAR_BIT8 * sizeof(T) - N);
235}
236
237/// Get the index of the last set bit starting from the least
238/// significant bit.
239///
240/// Only unsigned integral types are allowed.
241///
242/// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are
243/// valid arguments.
244template <typename T> T findLastSet(T Val, ZeroBehavior ZB = ZB_Max) {
245 if (ZB == ZB_Max && Val == 0)
246 return std::numeric_limits<T>::max();
247
248 // Use ^ instead of - because both gcc and llvm can remove the associated ^
249 // in the __builtin_clz intrinsic on x86.
250 return countLeadingZeros(Val, ZB_Undefined) ^
251 (std::numeric_limits<T>::digits - 1);
252}
253
254/// Macro compressed bit reversal table for 256 bits.
255///
256/// http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
257static const unsigned char BitReverseTable256[256] = {
258#define R2(n) n, n + 2 * 64, n + 1 * 64, n + 3 * 64
259#define R4(n) R2(n), R2(n + 2 * 16), R2(n + 1 * 16), R2(n + 3 * 16)
260#define R6(n) R4(n), R4(n + 2 * 4), R4(n + 1 * 4), R4(n + 3 * 4)
261 R6(0), R6(2), R6(1), R6(3)
262#undef R2
263#undef R4
264#undef R6
265};
266
267/// Reverse the bits in \p Val.
268template <typename T>
269T reverseBits(T Val) {
270 unsigned char in[sizeof(Val)];
271 unsigned char out[sizeof(Val)];
272 std::memcpy(in, &Val, sizeof(Val));
273 for (unsigned i = 0; i < sizeof(Val); ++i)
274 out[(sizeof(Val) - i) - 1] = BitReverseTable256[in[i]];
275 std::memcpy(&Val, out, sizeof(Val));
276 return Val;
277}
278
279// NOTE: The following support functions use the _32/_64 extensions instead of
280// type overloading so that signed and unsigned integers can be used without
281// ambiguity.
282
283/// Return the high 32 bits of a 64 bit value.
284constexpr inline uint32_t Hi_32(uint64_t Value) {
285 return static_cast<uint32_t>(Value >> 32);
286}
287
288/// Return the low 32 bits of a 64 bit value.
289constexpr inline uint32_t Lo_32(uint64_t Value) {
290 return static_cast<uint32_t>(Value);
291}
292
293/// Make a 64-bit integer from a high / low pair of 32-bit integers.
294constexpr inline uint64_t Make_64(uint32_t High, uint32_t Low) {
295 return ((uint64_t)High << 32) | (uint64_t)Low;
296}
297
298/// Checks if an integer fits into the given bit width.
299template <unsigned N> constexpr inline bool isInt(int64_t x) {
300 return N >= 64 || (-(INT64_C(1)1L<<(N-1)) <= x && x < (INT64_C(1)1L<<(N-1)));
301}
302// Template specializations to get better code for common cases.
303template <> constexpr inline bool isInt<8>(int64_t x) {
304 return static_cast<int8_t>(x) == x;
305}
306template <> constexpr inline bool isInt<16>(int64_t x) {
307 return static_cast<int16_t>(x) == x;
308}
309template <> constexpr inline bool isInt<32>(int64_t x) {
310 return static_cast<int32_t>(x) == x;
311}
312
313/// Checks if a signed integer is an N bit number shifted left by S.
314template <unsigned N, unsigned S>
315constexpr inline bool isShiftedInt(int64_t x) {
316 static_assert(
317 N > 0, "isShiftedInt<0> doesn't make sense (refers to a 0-bit number.");
318 static_assert(N + S <= 64, "isShiftedInt<N, S> with N + S > 64 is too wide.");
319 return isInt<N + S>(x) && (x % (UINT64_C(1)1UL << S) == 0);
320}
321
322/// Checks if an unsigned integer fits into the given bit width.
323///
324/// This is written as two functions rather than as simply
325///
326/// return N >= 64 || X < (UINT64_C(1) << N);
327///
328/// to keep MSVC from (incorrectly) warning on isUInt<64> that we're shifting
329/// left too many places.
330template <unsigned N>
331constexpr inline typename std::enable_if<(N < 64), bool>::type
332isUInt(uint64_t X) {
333 static_assert(N > 0, "isUInt<0> doesn't make sense");
334 return X < (UINT64_C(1)1UL << (N));
335}
336template <unsigned N>
337constexpr inline typename std::enable_if<N >= 64, bool>::type
338isUInt(uint64_t X) {
339 return true;
340}
341
342// Template specializations to get better code for common cases.
343template <> constexpr inline bool isUInt<8>(uint64_t x) {
344 return static_cast<uint8_t>(x) == x;
345}
346template <> constexpr inline bool isUInt<16>(uint64_t x) {
347 return static_cast<uint16_t>(x) == x;
348}
349template <> constexpr inline bool isUInt<32>(uint64_t x) {
350 return static_cast<uint32_t>(x) == x;
351}
352
353/// Checks if a unsigned integer is an N bit number shifted left by S.
354template <unsigned N, unsigned S>
355constexpr inline bool isShiftedUInt(uint64_t x) {
356 static_assert(
357 N > 0, "isShiftedUInt<0> doesn't make sense (refers to a 0-bit number)");
358 static_assert(N + S <= 64,
359 "isShiftedUInt<N, S> with N + S > 64 is too wide.");
360 // Per the two static_asserts above, S must be strictly less than 64. So
361 // 1 << S is not undefined behavior.
362 return isUInt<N + S>(x) && (x % (UINT64_C(1)1UL << S) == 0);
363}
364
365/// Gets the maximum value for a N-bit unsigned integer.
366inline uint64_t maxUIntN(uint64_t N) {
367 assert(N > 0 && N <= 64 && "integer width out of range")((N > 0 && N <= 64 && "integer width out of range"
) ? static_cast<void> (0) : __assert_fail ("N > 0 && N <= 64 && \"integer width out of range\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 367, __PRETTY_FUNCTION__))
;
368
369 // uint64_t(1) << 64 is undefined behavior, so we can't do
370 // (uint64_t(1) << N) - 1
371 // without checking first that N != 64. But this works and doesn't have a
372 // branch.
373 return UINT64_MAX(18446744073709551615UL) >> (64 - N);
374}
375
376/// Gets the minimum value for a N-bit signed integer.
377inline int64_t minIntN(int64_t N) {
378 assert(N > 0 && N <= 64 && "integer width out of range")((N > 0 && N <= 64 && "integer width out of range"
) ? static_cast<void> (0) : __assert_fail ("N > 0 && N <= 64 && \"integer width out of range\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 378, __PRETTY_FUNCTION__))
;
379
380 return -(UINT64_C(1)1UL<<(N-1));
381}
382
383/// Gets the maximum value for a N-bit signed integer.
384inline int64_t maxIntN(int64_t N) {
385 assert(N > 0 && N <= 64 && "integer width out of range")((N > 0 && N <= 64 && "integer width out of range"
) ? static_cast<void> (0) : __assert_fail ("N > 0 && N <= 64 && \"integer width out of range\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 385, __PRETTY_FUNCTION__))
;
386
387 // This relies on two's complement wraparound when N == 64, so we convert to
388 // int64_t only at the very end to avoid UB.
389 return (UINT64_C(1)1UL << (N - 1)) - 1;
390}
391
392/// Checks if an unsigned integer fits into the given (dynamic) bit width.
393inline bool isUIntN(unsigned N, uint64_t x) {
394 return N >= 64 || x <= maxUIntN(N);
395}
396
397/// Checks if an signed integer fits into the given (dynamic) bit width.
398inline bool isIntN(unsigned N, int64_t x) {
399 return N >= 64 || (minIntN(N) <= x && x <= maxIntN(N));
400}
401
402/// Return true if the argument is a non-empty sequence of ones starting at the
403/// least significant bit with the remainder zero (32 bit version).
404/// Ex. isMask_32(0x0000FFFFU) == true.
405constexpr inline bool isMask_32(uint32_t Value) {
406 return Value && ((Value + 1) & Value) == 0;
407}
408
409/// Return true if the argument is a non-empty sequence of ones starting at the
410/// least significant bit with the remainder zero (64 bit version).
411constexpr inline bool isMask_64(uint64_t Value) {
412 return Value && ((Value + 1) & Value) == 0;
413}
414
415/// Return true if the argument contains a non-empty sequence of ones with the
416/// remainder zero (32 bit version.) Ex. isShiftedMask_32(0x0000FF00U) == true.
417constexpr inline bool isShiftedMask_32(uint32_t Value) {
418 return Value && isMask_32((Value - 1) | Value);
419}
420
421/// Return true if the argument contains a non-empty sequence of ones with the
422/// remainder zero (64 bit version.)
423constexpr inline bool isShiftedMask_64(uint64_t Value) {
424 return Value && isMask_64((Value - 1) | Value);
425}
426
427/// Return true if the argument is a power of two > 0.
428/// Ex. isPowerOf2_32(0x00100000U) == true (32 bit edition.)
429constexpr inline bool isPowerOf2_32(uint32_t Value) {
430 return Value && !(Value & (Value - 1));
431}
432
433/// Return true if the argument is a power of two > 0 (64 bit edition.)
434constexpr inline bool isPowerOf2_64(uint64_t Value) {
435 return Value && !(Value & (Value - 1));
436}
437
438/// Return a byte-swapped representation of the 16-bit argument.
439inline uint16_t ByteSwap_16(uint16_t Value) {
440 return sys::SwapByteOrder_16(Value);
441}
442
443/// Return a byte-swapped representation of the 32-bit argument.
444inline uint32_t ByteSwap_32(uint32_t Value) {
445 return sys::SwapByteOrder_32(Value);
446}
447
448/// Return a byte-swapped representation of the 64-bit argument.
449inline uint64_t ByteSwap_64(uint64_t Value) {
450 return sys::SwapByteOrder_64(Value);
451}
452
453/// Count the number of ones from the most significant bit to the first
454/// zero bit.
455///
456/// Ex. countLeadingOnes(0xFF0FFF00) == 8.
457/// Only unsigned integral types are allowed.
458///
459/// \param ZB the behavior on an input of all ones. Only ZB_Width and
460/// ZB_Undefined are valid arguments.
461template <typename T>
462std::size_t countLeadingOnes(T Value, ZeroBehavior ZB = ZB_Width) {
463 static_assert(std::numeric_limits<T>::is_integer &&
464 !std::numeric_limits<T>::is_signed,
465 "Only unsigned integral types are allowed.");
466 return countLeadingZeros<T>(~Value, ZB);
467}
468
469/// Count the number of ones from the least significant bit to the first
470/// zero bit.
471///
472/// Ex. countTrailingOnes(0x00FF00FF) == 8.
473/// Only unsigned integral types are allowed.
474///
475/// \param ZB the behavior on an input of all ones. Only ZB_Width and
476/// ZB_Undefined are valid arguments.
477template <typename T>
478std::size_t countTrailingOnes(T Value, ZeroBehavior ZB = ZB_Width) {
479 static_assert(std::numeric_limits<T>::is_integer &&
480 !std::numeric_limits<T>::is_signed,
481 "Only unsigned integral types are allowed.");
482 return countTrailingZeros<T>(~Value, ZB);
483}
484
485namespace detail {
486template <typename T, std::size_t SizeOfT> struct PopulationCounter {
487 static unsigned count(T Value) {
488 // Generic version, forward to 32 bits.
489 static_assert(SizeOfT <= 4, "Not implemented!");
490#if __GNUC__4 >= 4
491 return __builtin_popcount(Value);
492#else
493 uint32_t v = Value;
494 v = v - ((v >> 1) & 0x55555555);
495 v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
496 return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
497#endif
498 }
499};
500
501template <typename T> struct PopulationCounter<T, 8> {
502 static unsigned count(T Value) {
503#if __GNUC__4 >= 4
504 return __builtin_popcountll(Value);
505#else
506 uint64_t v = Value;
507 v = v - ((v >> 1) & 0x5555555555555555ULL);
508 v = (v & 0x3333333333333333ULL) + ((v >> 2) & 0x3333333333333333ULL);
509 v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0FULL;
510 return unsigned((uint64_t)(v * 0x0101010101010101ULL) >> 56);
511#endif
512 }
513};
514} // namespace detail
515
516/// Count the number of set bits in a value.
517/// Ex. countPopulation(0xF000F000) = 8
518/// Returns 0 if the word is zero.
519template <typename T>
520inline unsigned countPopulation(T Value) {
521 static_assert(std::numeric_limits<T>::is_integer &&
522 !std::numeric_limits<T>::is_signed,
523 "Only unsigned integral types are allowed.");
524 return detail::PopulationCounter<T, sizeof(T)>::count(Value);
525}
526
527/// Return the log base 2 of the specified value.
528inline double Log2(double Value) {
529#if defined(__ANDROID_API__) && __ANDROID_API__ < 18
530 return __builtin_log(Value) / __builtin_log(2.0);
531#else
532 return log2(Value);
533#endif
534}
535
536/// Return the floor log base 2 of the specified value, -1 if the value is zero.
537/// (32 bit edition.)
538/// Ex. Log2_32(32) == 5, Log2_32(1) == 0, Log2_32(0) == -1, Log2_32(6) == 2
539inline unsigned Log2_32(uint32_t Value) {
540 return 31 - countLeadingZeros(Value);
6
Returning the value 4294967295
541}
542
543/// Return the floor log base 2 of the specified value, -1 if the value is zero.
544/// (64 bit edition.)
545inline unsigned Log2_64(uint64_t Value) {
546 return 63 - countLeadingZeros(Value);
547}
548
549/// Return the ceil log base 2 of the specified value, 32 if the value is zero.
550/// (32 bit edition).
551/// Ex. Log2_32_Ceil(32) == 5, Log2_32_Ceil(1) == 0, Log2_32_Ceil(6) == 3
552inline unsigned Log2_32_Ceil(uint32_t Value) {
553 return 32 - countLeadingZeros(Value - 1);
554}
555
556/// Return the ceil log base 2 of the specified value, 64 if the value is zero.
557/// (64 bit edition.)
558inline unsigned Log2_64_Ceil(uint64_t Value) {
559 return 64 - countLeadingZeros(Value - 1);
560}
561
562/// Return the greatest common divisor of the values using Euclid's algorithm.
563inline uint64_t GreatestCommonDivisor64(uint64_t A, uint64_t B) {
564 while (B) {
565 uint64_t T = B;
566 B = A % B;
567 A = T;
568 }
569 return A;
570}
571
572/// This function takes a 64-bit integer and returns the bit equivalent double.
573inline double BitsToDouble(uint64_t Bits) {
574 double D;
575 static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes");
576 memcpy(&D, &Bits, sizeof(Bits));
577 return D;
578}
579
580/// This function takes a 32-bit integer and returns the bit equivalent float.
581inline float BitsToFloat(uint32_t Bits) {
582 float F;
583 static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes");
584 memcpy(&F, &Bits, sizeof(Bits));
585 return F;
586}
587
588/// This function takes a double and returns the bit equivalent 64-bit integer.
589/// Note that copying doubles around changes the bits of NaNs on some hosts,
590/// notably x86, so this routine cannot be used if these bits are needed.
591inline uint64_t DoubleToBits(double Double) {
592 uint64_t Bits;
593 static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes");
594 memcpy(&Bits, &Double, sizeof(Double));
595 return Bits;
596}
597
598/// This function takes a float and returns the bit equivalent 32-bit integer.
599/// Note that copying floats around changes the bits of NaNs on some hosts,
600/// notably x86, so this routine cannot be used if these bits are needed.
601inline uint32_t FloatToBits(float Float) {
602 uint32_t Bits;
603 static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes");
604 memcpy(&Bits, &Float, sizeof(Float));
605 return Bits;
606}
607
608/// A and B are either alignments or offsets. Return the minimum alignment that
609/// may be assumed after adding the two together.
610constexpr inline uint64_t MinAlign(uint64_t A, uint64_t B) {
611 // The largest power of 2 that divides both A and B.
612 //
613 // Replace "-Value" by "1+~Value" in the following commented code to avoid
614 // MSVC warning C4146
615 // return (A | B) & -(A | B);
616 return (A | B) & (1 + ~(A | B));
617}
618
619/// Aligns \c Addr to \c Alignment bytes, rounding up.
620///
621/// Alignment should be a power of two. This method rounds up, so
622/// alignAddr(7, 4) == 8 and alignAddr(8, 4) == 8.
623inline uintptr_t alignAddr(const void *Addr, size_t Alignment) {
624 assert(Alignment && isPowerOf2_64((uint64_t)Alignment) &&((Alignment && isPowerOf2_64((uint64_t)Alignment) &&
"Alignment is not a power of two!") ? static_cast<void>
(0) : __assert_fail ("Alignment && isPowerOf2_64((uint64_t)Alignment) && \"Alignment is not a power of two!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 625, __PRETTY_FUNCTION__))
625 "Alignment is not a power of two!")((Alignment && isPowerOf2_64((uint64_t)Alignment) &&
"Alignment is not a power of two!") ? static_cast<void>
(0) : __assert_fail ("Alignment && isPowerOf2_64((uint64_t)Alignment) && \"Alignment is not a power of two!\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 625, __PRETTY_FUNCTION__))
;
626
627 assert((uintptr_t)Addr + Alignment - 1 >= (uintptr_t)Addr)(((uintptr_t)Addr + Alignment - 1 >= (uintptr_t)Addr) ? static_cast
<void> (0) : __assert_fail ("(uintptr_t)Addr + Alignment - 1 >= (uintptr_t)Addr"
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 627, __PRETTY_FUNCTION__))
;
628
629 return (((uintptr_t)Addr + Alignment - 1) & ~(uintptr_t)(Alignment - 1));
630}
631
632/// Returns the necessary adjustment for aligning \c Ptr to \c Alignment
633/// bytes, rounding up.
634inline size_t alignmentAdjustment(const void *Ptr, size_t Alignment) {
635 return alignAddr(Ptr, Alignment) - (uintptr_t)Ptr;
636}
637
638/// Returns the next power of two (in 64-bits) that is strictly greater than A.
639/// Returns zero on overflow.
640inline uint64_t NextPowerOf2(uint64_t A) {
641 A |= (A >> 1);
642 A |= (A >> 2);
643 A |= (A >> 4);
644 A |= (A >> 8);
645 A |= (A >> 16);
646 A |= (A >> 32);
647 return A + 1;
648}
649
650/// Returns the power of two which is less than or equal to the given value.
651/// Essentially, it is a floor operation across the domain of powers of two.
652inline uint64_t PowerOf2Floor(uint64_t A) {
653 if (!A) return 0;
654 return 1ull << (63 - countLeadingZeros(A, ZB_Undefined));
655}
656
657/// Returns the power of two which is greater than or equal to the given value.
658/// Essentially, it is a ceil operation across the domain of powers of two.
659inline uint64_t PowerOf2Ceil(uint64_t A) {
660 if (!A)
661 return 0;
662 return NextPowerOf2(A - 1);
663}
664
665/// Returns the next integer (mod 2**64) that is greater than or equal to
666/// \p Value and is a multiple of \p Align. \p Align must be non-zero.
667///
668/// If non-zero \p Skew is specified, the return value will be a minimal
669/// integer that is greater than or equal to \p Value and equal to
670/// \p Align * N + \p Skew for some integer N. If \p Skew is larger than
671/// \p Align, its value is adjusted to '\p Skew mod \p Align'.
672///
673/// Examples:
674/// \code
675/// alignTo(5, 8) = 8
676/// alignTo(17, 8) = 24
677/// alignTo(~0LL, 8) = 0
678/// alignTo(321, 255) = 510
679///
680/// alignTo(5, 8, 7) = 7
681/// alignTo(17, 8, 1) = 17
682/// alignTo(~0LL, 8, 3) = 3
683/// alignTo(321, 255, 42) = 552
684/// \endcode
685inline uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew = 0) {
686 assert(Align != 0u && "Align can't be 0.")((Align != 0u && "Align can't be 0.") ? static_cast<
void> (0) : __assert_fail ("Align != 0u && \"Align can't be 0.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 686, __PRETTY_FUNCTION__))
;
687 Skew %= Align;
688 return (Value + Align - 1 - Skew) / Align * Align + Skew;
689}
690
691/// Returns the next integer (mod 2**64) that is greater than or equal to
692/// \p Value and is a multiple of \c Align. \c Align must be non-zero.
693template <uint64_t Align> constexpr inline uint64_t alignTo(uint64_t Value) {
694 static_assert(Align != 0u, "Align must be non-zero");
695 return (Value + Align - 1) / Align * Align;
696}
697
698/// Returns the integer ceil(Numerator / Denominator).
699inline uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator) {
700 return alignTo(Numerator, Denominator) / Denominator;
701}
702
703/// \c alignTo for contexts where a constant expression is required.
704/// \sa alignTo
705///
706/// \todo FIXME: remove when \c constexpr becomes really \c constexpr
707template <uint64_t Align>
708struct AlignTo {
709 static_assert(Align != 0u, "Align must be non-zero");
710 template <uint64_t Value>
711 struct from_value {
712 static const uint64_t value = (Value + Align - 1) / Align * Align;
713 };
714};
715
716/// Returns the largest uint64_t less than or equal to \p Value and is
717/// \p Skew mod \p Align. \p Align must be non-zero
718inline uint64_t alignDown(uint64_t Value, uint64_t Align, uint64_t Skew = 0) {
719 assert(Align != 0u && "Align can't be 0.")((Align != 0u && "Align can't be 0.") ? static_cast<
void> (0) : __assert_fail ("Align != 0u && \"Align can't be 0.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 719, __PRETTY_FUNCTION__))
;
720 Skew %= Align;
721 return (Value - Skew) / Align * Align + Skew;
722}
723
724/// Returns the offset to the next integer (mod 2**64) that is greater than
725/// or equal to \p Value and is a multiple of \p Align. \p Align must be
726/// non-zero.
727inline uint64_t OffsetToAlignment(uint64_t Value, uint64_t Align) {
728 return alignTo(Value, Align) - Value;
729}
730
731/// Sign-extend the number in the bottom B bits of X to a 32-bit integer.
732/// Requires 0 < B <= 32.
733template <unsigned B> constexpr inline int32_t SignExtend32(uint32_t X) {
734 static_assert(B > 0, "Bit width can't be 0.");
735 static_assert(B <= 32, "Bit width out of range.");
736 return int32_t(X << (32 - B)) >> (32 - B);
737}
738
739/// Sign-extend the number in the bottom B bits of X to a 32-bit integer.
740/// Requires 0 < B < 32.
741inline int32_t SignExtend32(uint32_t X, unsigned B) {
742 assert(B > 0 && "Bit width can't be 0.")((B > 0 && "Bit width can't be 0.") ? static_cast<
void> (0) : __assert_fail ("B > 0 && \"Bit width can't be 0.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 742, __PRETTY_FUNCTION__))
;
743 assert(B <= 32 && "Bit width out of range.")((B <= 32 && "Bit width out of range.") ? static_cast
<void> (0) : __assert_fail ("B <= 32 && \"Bit width out of range.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 743, __PRETTY_FUNCTION__))
;
744 return int32_t(X << (32 - B)) >> (32 - B);
745}
746
747/// Sign-extend the number in the bottom B bits of X to a 64-bit integer.
748/// Requires 0 < B < 64.
749template <unsigned B> constexpr inline int64_t SignExtend64(uint64_t x) {
750 static_assert(B > 0, "Bit width can't be 0.");
751 static_assert(B <= 64, "Bit width out of range.");
752 return int64_t(x << (64 - B)) >> (64 - B);
753}
754
755/// Sign-extend the number in the bottom B bits of X to a 64-bit integer.
756/// Requires 0 < B < 64.
757inline int64_t SignExtend64(uint64_t X, unsigned B) {
758 assert(B > 0 && "Bit width can't be 0.")((B > 0 && "Bit width can't be 0.") ? static_cast<
void> (0) : __assert_fail ("B > 0 && \"Bit width can't be 0.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 758, __PRETTY_FUNCTION__))
;
759 assert(B <= 64 && "Bit width out of range.")((B <= 64 && "Bit width out of range.") ? static_cast
<void> (0) : __assert_fail ("B <= 64 && \"Bit width out of range.\""
, "/build/llvm-toolchain-snapshot-8~svn345461/include/llvm/Support/MathExtras.h"
, 759, __PRETTY_FUNCTION__))
;
760 return int64_t(X << (64 - B)) >> (64 - B);
761}
762
763/// Subtract two unsigned integers, X and Y, of type T and return the absolute
764/// value of the result.
765template <typename T>
766typename std::enable_if<std::is_unsigned<T>::value, T>::type
767AbsoluteDifference(T X, T Y) {
768 return std::max(X, Y) - std::min(X, Y);
769}
770
771/// Add two unsigned integers, X and Y, of type T. Clamp the result to the
772/// maximum representable value of T on overflow. ResultOverflowed indicates if
773/// the result is larger than the maximum representable value of type T.
774template <typename T>
775typename std::enable_if<std::is_unsigned<T>::value, T>::type
776SaturatingAdd(T X, T Y, bool *ResultOverflowed = nullptr) {
777 bool Dummy;
778 bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy;
779 // Hacker's Delight, p. 29
780 T Z = X + Y;
781 Overflowed = (Z < X || Z < Y);
782 if (Overflowed)
783 return std::numeric_limits<T>::max();
784 else
785 return Z;
786}
787
788/// Multiply two unsigned integers, X and Y, of type T. Clamp the result to the
789/// maximum representable value of T on overflow. ResultOverflowed indicates if
790/// the result is larger than the maximum representable value of type T.
791template <typename T>
792typename std::enable_if<std::is_unsigned<T>::value, T>::type
793SaturatingMultiply(T X, T Y, bool *ResultOverflowed = nullptr) {
794 bool Dummy;
795 bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy;
796
797 // Hacker's Delight, p. 30 has a different algorithm, but we don't use that
798 // because it fails for uint16_t (where multiplication can have undefined
799 // behavior due to promotion to int), and requires a division in addition
800 // to the multiplication.
801
802 Overflowed = false;
803
804 // Log2(Z) would be either Log2Z or Log2Z + 1.
805 // Special case: if X or Y is 0, Log2_64 gives -1, and Log2Z
806 // will necessarily be less than Log2Max as desired.
807 int Log2Z = Log2_64(X) + Log2_64(Y);
808 const T Max = std::numeric_limits<T>::max();
809 int Log2Max = Log2_64(Max);
810 if (Log2Z < Log2Max) {
811 return X * Y;
812 }
813 if (Log2Z > Log2Max) {
814 Overflowed = true;
815 return Max;
816 }
817
818 // We're going to use the top bit, and maybe overflow one
819 // bit past it. Multiply all but the bottom bit then add
820 // that on at the end.
821 T Z = (X >> 1) * Y;
822 if (Z & ~(Max >> 1)) {
823 Overflowed = true;
824 return Max;
825 }
826 Z <<= 1;
827 if (X & 1)
828 return SaturatingAdd(Z, Y, ResultOverflowed);
829
830 return Z;
831}
832
833/// Multiply two unsigned integers, X and Y, and add the unsigned integer, A to
834/// the product. Clamp the result to the maximum representable value of T on
835/// overflow. ResultOverflowed indicates if the result is larger than the
836/// maximum representable value of type T.
837template <typename T>
838typename std::enable_if<std::is_unsigned<T>::value, T>::type
839SaturatingMultiplyAdd(T X, T Y, T A, bool *ResultOverflowed = nullptr) {
840 bool Dummy;
841 bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy;
842
843 T Product = SaturatingMultiply(X, Y, &Overflowed);
844 if (Overflowed)
845 return Product;
846
847 return SaturatingAdd(A, Product, &Overflowed);
848}
849
850/// Use this rather than HUGE_VALF; the latter causes warnings on MSVC.
851extern const float huge_valf;
852} // End llvm namespace
853
854#endif