1 //===- StackMaps.cpp ------------------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "llvm/ADT/DenseMapInfo.h"
11 #include "llvm/ADT/STLExtras.h"
12 #include "llvm/ADT/Twine.h"
13 #include "llvm/CodeGen/AsmPrinter.h"
14 #include "llvm/CodeGen/MachineFrameInfo.h"
15 #include "llvm/CodeGen/MachineFunction.h"
16 #include "llvm/CodeGen/MachineInstr.h"
17 #include "llvm/CodeGen/MachineOperand.h"
18 #include "llvm/CodeGen/StackMaps.h"
19 #include "llvm/IR/DataLayout.h"
20 #include "llvm/MC/MCContext.h"
21 #include "llvm/MC/MCExpr.h"
22 #include "llvm/MC/MCObjectFileInfo.h"
23 #include "llvm/MC/MCRegisterInfo.h"
24 #include "llvm/MC/MCStreamer.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/MathExtras.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Target/TargetOpcodes.h"
31 #include "llvm/Target/TargetRegisterInfo.h"
32 #include "llvm/Target/TargetSubtargetInfo.h"
41 #define DEBUG_TYPE "stackmaps"
43 static cl::opt<int> StackMapVersion(
44 "stackmap-version", cl::init(2),
45 cl::desc("Specify the stackmap encoding version (default = 2)"));
47 const char *StackMaps::WSMP = "Stack Maps: ";
49 StackMapOpers::StackMapOpers(const MachineInstr *MI)
51 assert(getVarIdx() <= MI->getNumOperands() &&
52 "invalid stackmap definition");
55 PatchPointOpers::PatchPointOpers(const MachineInstr *MI)
56 : MI(MI), HasDef(MI->getOperand(0).isReg() && MI->getOperand(0).isDef() &&
57 !MI->getOperand(0).isImplicit()) {
59 unsigned CheckStartIdx = 0, e = MI->getNumOperands();
60 while (CheckStartIdx < e && MI->getOperand(CheckStartIdx).isReg() &&
61 MI->getOperand(CheckStartIdx).isDef() &&
62 !MI->getOperand(CheckStartIdx).isImplicit())
65 assert(getMetaIdx() == CheckStartIdx &&
66 "Unexpected additional definition in Patchpoint intrinsic.");
70 unsigned PatchPointOpers::getNextScratchIdx(unsigned StartIdx) const {
72 StartIdx = getVarIdx();
74 // Find the next scratch register (implicit def and early clobber)
75 unsigned ScratchIdx = StartIdx, e = MI->getNumOperands();
76 while (ScratchIdx < e &&
77 !(MI->getOperand(ScratchIdx).isReg() &&
78 MI->getOperand(ScratchIdx).isDef() &&
79 MI->getOperand(ScratchIdx).isImplicit() &&
80 MI->getOperand(ScratchIdx).isEarlyClobber()))
83 assert(ScratchIdx != e && "No scratch register available");
87 StackMaps::StackMaps(AsmPrinter &AP) : AP(AP) {
88 if (StackMapVersion != 2)
89 llvm_unreachable("Unsupported stackmap version!");
92 /// Go up the super-register chain until we hit a valid dwarf register number.
93 static unsigned getDwarfRegNum(unsigned Reg, const TargetRegisterInfo *TRI) {
94 int RegNum = TRI->getDwarfRegNum(Reg, false);
95 for (MCSuperRegIterator SR(Reg, TRI); SR.isValid() && RegNum < 0; ++SR)
96 RegNum = TRI->getDwarfRegNum(*SR, false);
98 assert(RegNum >= 0 && "Invalid Dwarf register number.");
99 return (unsigned)RegNum;
102 MachineInstr::const_mop_iterator
103 StackMaps::parseOperand(MachineInstr::const_mop_iterator MOI,
104 MachineInstr::const_mop_iterator MOE, LocationVec &Locs,
105 LiveOutVec &LiveOuts) const {
106 const TargetRegisterInfo *TRI = AP.MF->getSubtarget().getRegisterInfo();
108 switch (MOI->getImm()) {
110 llvm_unreachable("Unrecognized operand type.");
111 case StackMaps::DirectMemRefOp: {
112 auto &DL = AP.MF->getDataLayout();
114 unsigned Size = DL.getPointerSizeInBits();
115 assert((Size % 8) == 0 && "Need pointer size in bytes.");
117 unsigned Reg = (++MOI)->getReg();
118 int64_t Imm = (++MOI)->getImm();
119 Locs.emplace_back(StackMaps::Location::Direct, Size,
120 getDwarfRegNum(Reg, TRI), Imm);
123 case StackMaps::IndirectMemRefOp: {
124 int64_t Size = (++MOI)->getImm();
125 assert(Size > 0 && "Need a valid size for indirect memory locations.");
126 unsigned Reg = (++MOI)->getReg();
127 int64_t Imm = (++MOI)->getImm();
128 Locs.emplace_back(StackMaps::Location::Indirect, Size,
129 getDwarfRegNum(Reg, TRI), Imm);
132 case StackMaps::ConstantOp: {
134 assert(MOI->isImm() && "Expected constant operand.");
135 int64_t Imm = MOI->getImm();
136 Locs.emplace_back(Location::Constant, sizeof(int64_t), 0, Imm);
143 // The physical register number will ultimately be encoded as a DWARF regno.
144 // The stack map also records the size of a spill slot that can hold the
145 // register content. (The runtime can track the actual size of the data type
148 // Skip implicit registers (this includes our scratch registers)
149 if (MOI->isImplicit())
152 assert(TargetRegisterInfo::isPhysicalRegister(MOI->getReg()) &&
153 "Virtreg operands should have been rewritten before now.");
154 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(MOI->getReg());
155 assert(!MOI->getSubReg() && "Physical subreg still around.");
158 unsigned DwarfRegNum = getDwarfRegNum(MOI->getReg(), TRI);
159 unsigned LLVMRegNum = TRI->getLLVMRegNum(DwarfRegNum, false);
160 unsigned SubRegIdx = TRI->getSubRegIndex(LLVMRegNum, MOI->getReg());
162 Offset = TRI->getSubRegIdxOffset(SubRegIdx);
164 Locs.emplace_back(Location::Register, TRI->getSpillSize(*RC),
165 DwarfRegNum, Offset);
169 if (MOI->isRegLiveOut())
170 LiveOuts = parseRegisterLiveOutMask(MOI->getRegLiveOut());
175 void StackMaps::print(raw_ostream &OS) {
176 const TargetRegisterInfo *TRI =
177 AP.MF ? AP.MF->getSubtarget().getRegisterInfo() : nullptr;
178 OS << WSMP << "callsites:\n";
179 for (const auto &CSI : CSInfos) {
180 const LocationVec &CSLocs = CSI.Locations;
181 const LiveOutVec &LiveOuts = CSI.LiveOuts;
183 OS << WSMP << "callsite " << CSI.ID << "\n";
184 OS << WSMP << " has " << CSLocs.size() << " locations\n";
187 for (const auto &Loc : CSLocs) {
188 OS << WSMP << "\t\tLoc " << Idx << ": ";
190 case Location::Unprocessed:
191 OS << "<Unprocessed operand>";
193 case Location::Register:
196 OS << TRI->getName(Loc.Reg);
200 case Location::Direct:
203 OS << TRI->getName(Loc.Reg);
207 OS << " + " << Loc.Offset;
209 case Location::Indirect:
212 OS << TRI->getName(Loc.Reg);
215 OS << "+" << Loc.Offset;
217 case Location::Constant:
218 OS << "Constant " << Loc.Offset;
220 case Location::ConstantIndex:
221 OS << "Constant Index " << Loc.Offset;
224 OS << "\t[encoding: .byte " << Loc.Type << ", .byte " << Loc.Size
225 << ", .short " << Loc.Reg << ", .int " << Loc.Offset << "]\n";
229 OS << WSMP << "\thas " << LiveOuts.size() << " live-out registers\n";
232 for (const auto &LO : LiveOuts) {
233 OS << WSMP << "\t\tLO " << Idx << ": ";
235 OS << TRI->getName(LO.Reg);
238 OS << "\t[encoding: .short " << LO.DwarfRegNum << ", .byte 0, .byte "
245 /// Create a live-out register record for the given register Reg.
246 StackMaps::LiveOutReg
247 StackMaps::createLiveOutReg(unsigned Reg, const TargetRegisterInfo *TRI) const {
248 unsigned DwarfRegNum = getDwarfRegNum(Reg, TRI);
249 unsigned Size = TRI->getSpillSize(*TRI->getMinimalPhysRegClass(Reg));
250 return LiveOutReg(Reg, DwarfRegNum, Size);
253 /// Parse the register live-out mask and return a vector of live-out registers
254 /// that need to be recorded in the stackmap.
255 StackMaps::LiveOutVec
256 StackMaps::parseRegisterLiveOutMask(const uint32_t *Mask) const {
257 assert(Mask && "No register mask specified");
258 const TargetRegisterInfo *TRI = AP.MF->getSubtarget().getRegisterInfo();
261 // Create a LiveOutReg for each bit that is set in the register mask.
262 for (unsigned Reg = 0, NumRegs = TRI->getNumRegs(); Reg != NumRegs; ++Reg)
263 if ((Mask[Reg / 32] >> Reg % 32) & 1)
264 LiveOuts.push_back(createLiveOutReg(Reg, TRI));
266 // We don't need to keep track of a register if its super-register is already
267 // in the list. Merge entries that refer to the same dwarf register and use
268 // the maximum size that needs to be spilled.
270 std::sort(LiveOuts.begin(), LiveOuts.end(),
271 [](const LiveOutReg &LHS, const LiveOutReg &RHS) {
272 // Only sort by the dwarf register number.
273 return LHS.DwarfRegNum < RHS.DwarfRegNum;
276 for (auto I = LiveOuts.begin(), E = LiveOuts.end(); I != E; ++I) {
277 for (auto II = std::next(I); II != E; ++II) {
278 if (I->DwarfRegNum != II->DwarfRegNum) {
279 // Skip all the now invalid entries.
283 I->Size = std::max(I->Size, II->Size);
284 if (TRI->isSuperRegister(I->Reg, II->Reg))
286 II->Reg = 0; // mark for deletion.
291 llvm::remove_if(LiveOuts,
292 [](const LiveOutReg &LO) { return LO.Reg == 0; }),
298 void StackMaps::recordStackMapOpers(const MachineInstr &MI, uint64_t ID,
299 MachineInstr::const_mop_iterator MOI,
300 MachineInstr::const_mop_iterator MOE,
302 MCContext &OutContext = AP.OutStreamer->getContext();
303 MCSymbol *MILabel = OutContext.createTempSymbol();
304 AP.OutStreamer->EmitLabel(MILabel);
306 LocationVec Locations;
310 assert(PatchPointOpers(&MI).hasDef() && "Stackmap has no return value.");
311 parseOperand(MI.operands_begin(), std::next(MI.operands_begin()), Locations,
317 MOI = parseOperand(MOI, MOE, Locations, LiveOuts);
320 // Move large constants into the constant pool.
321 for (auto &Loc : Locations) {
322 // Constants are encoded as sign-extended integers.
323 // -1 is directly encoded as .long 0xFFFFFFFF with no constant pool.
324 if (Loc.Type == Location::Constant && !isInt<32>(Loc.Offset)) {
325 Loc.Type = Location::ConstantIndex;
326 // ConstPool is intentionally a MapVector of 'uint64_t's (as
327 // opposed to 'int64_t's). We should never be in a situation
328 // where we have to insert either the tombstone or the empty
329 // keys into a map, and for a DenseMap<uint64_t, T> these are
330 // (uint64_t)0 and (uint64_t)-1. They can be and are
331 // represented using 32 bit integers.
332 assert((uint64_t)Loc.Offset != DenseMapInfo<uint64_t>::getEmptyKey() &&
333 (uint64_t)Loc.Offset !=
334 DenseMapInfo<uint64_t>::getTombstoneKey() &&
335 "empty and tombstone keys should fit in 32 bits!");
336 auto Result = ConstPool.insert(std::make_pair(Loc.Offset, Loc.Offset));
337 Loc.Offset = Result.first - ConstPool.begin();
341 // Create an expression to calculate the offset of the callsite from function
343 const MCExpr *CSOffsetExpr = MCBinaryExpr::createSub(
344 MCSymbolRefExpr::create(MILabel, OutContext),
345 MCSymbolRefExpr::create(AP.CurrentFnSymForSize, OutContext), OutContext);
347 CSInfos.emplace_back(CSOffsetExpr, ID, std::move(Locations),
348 std::move(LiveOuts));
350 // Record the stack size of the current function and update callsite count.
351 const MachineFrameInfo &MFI = AP.MF->getFrameInfo();
352 const TargetRegisterInfo *RegInfo = AP.MF->getSubtarget().getRegisterInfo();
353 bool HasDynamicFrameSize =
354 MFI.hasVarSizedObjects() || RegInfo->needsStackRealignment(*(AP.MF));
355 uint64_t FrameSize = HasDynamicFrameSize ? UINT64_MAX : MFI.getStackSize();
357 auto CurrentIt = FnInfos.find(AP.CurrentFnSym);
358 if (CurrentIt != FnInfos.end())
359 CurrentIt->second.RecordCount++;
361 FnInfos.insert(std::make_pair(AP.CurrentFnSym, FunctionInfo(FrameSize)));
364 void StackMaps::recordStackMap(const MachineInstr &MI) {
365 assert(MI.getOpcode() == TargetOpcode::STACKMAP && "expected stackmap");
367 StackMapOpers opers(&MI);
368 const int64_t ID = MI.getOperand(PatchPointOpers::IDPos).getImm();
369 recordStackMapOpers(MI, ID, std::next(MI.operands_begin(), opers.getVarIdx()),
373 void StackMaps::recordPatchPoint(const MachineInstr &MI) {
374 assert(MI.getOpcode() == TargetOpcode::PATCHPOINT && "expected patchpoint");
376 PatchPointOpers opers(&MI);
377 const int64_t ID = opers.getID();
378 auto MOI = std::next(MI.operands_begin(), opers.getStackMapStartIdx());
379 recordStackMapOpers(MI, ID, MOI, MI.operands_end(),
380 opers.isAnyReg() && opers.hasDef());
384 auto &Locations = CSInfos.back().Locations;
385 if (opers.isAnyReg()) {
386 unsigned NArgs = opers.getNumCallArgs();
387 for (unsigned i = 0, e = (opers.hasDef() ? NArgs + 1 : NArgs); i != e; ++i)
388 assert(Locations[i].Type == Location::Register &&
389 "anyreg arg must be in reg.");
394 void StackMaps::recordStatepoint(const MachineInstr &MI) {
395 assert(MI.getOpcode() == TargetOpcode::STATEPOINT && "expected statepoint");
397 StatepointOpers opers(&MI);
398 // Record all the deopt and gc operands (they're contiguous and run from the
399 // initial index to the end of the operand list)
400 const unsigned StartIdx = opers.getVarIdx();
401 recordStackMapOpers(MI, opers.getID(), MI.operands_begin() + StartIdx,
402 MI.operands_end(), false);
405 /// Emit the stackmap header.
408 /// uint8 : Stack Map Version (currently 2)
409 /// uint8 : Reserved (expected to be 0)
410 /// uint16 : Reserved (expected to be 0)
412 /// uint32 : NumFunctions
413 /// uint32 : NumConstants
414 /// uint32 : NumRecords
415 void StackMaps::emitStackmapHeader(MCStreamer &OS) {
417 OS.EmitIntValue(StackMapVersion, 1); // Version.
418 OS.EmitIntValue(0, 1); // Reserved.
419 OS.EmitIntValue(0, 2); // Reserved.
422 DEBUG(dbgs() << WSMP << "#functions = " << FnInfos.size() << '\n');
423 OS.EmitIntValue(FnInfos.size(), 4);
425 DEBUG(dbgs() << WSMP << "#constants = " << ConstPool.size() << '\n');
426 OS.EmitIntValue(ConstPool.size(), 4);
428 DEBUG(dbgs() << WSMP << "#callsites = " << CSInfos.size() << '\n');
429 OS.EmitIntValue(CSInfos.size(), 4);
432 /// Emit the function frame record for each function.
434 /// StkSizeRecord[NumFunctions] {
435 /// uint64 : Function Address
436 /// uint64 : Stack Size
437 /// uint64 : Record Count
439 void StackMaps::emitFunctionFrameRecords(MCStreamer &OS) {
440 // Function Frame records.
441 DEBUG(dbgs() << WSMP << "functions:\n");
442 for (auto const &FR : FnInfos) {
443 DEBUG(dbgs() << WSMP << "function addr: " << FR.first
444 << " frame size: " << FR.second.StackSize
445 << " callsite count: " << FR.second.RecordCount << '\n');
446 OS.EmitSymbolValue(FR.first, 8);
447 OS.EmitIntValue(FR.second.StackSize, 8);
448 OS.EmitIntValue(FR.second.RecordCount, 8);
452 /// Emit the constant pool.
454 /// int64 : Constants[NumConstants]
455 void StackMaps::emitConstantPoolEntries(MCStreamer &OS) {
456 // Constant pool entries.
457 DEBUG(dbgs() << WSMP << "constants:\n");
458 for (const auto &ConstEntry : ConstPool) {
459 DEBUG(dbgs() << WSMP << ConstEntry.second << '\n');
460 OS.EmitIntValue(ConstEntry.second, 8);
464 /// Emit the callsite info for each callsite.
466 /// StkMapRecord[NumRecords] {
467 /// uint64 : PatchPoint ID
468 /// uint32 : Instruction Offset
469 /// uint16 : Reserved (record flags)
470 /// uint16 : NumLocations
471 /// Location[NumLocations] {
472 /// uint8 : Register | Direct | Indirect | Constant | ConstantIndex
473 /// uint8 : Size in Bytes
474 /// uint16 : Dwarf RegNum
478 /// uint16 : NumLiveOuts
479 /// LiveOuts[NumLiveOuts] {
480 /// uint16 : Dwarf RegNum
482 /// uint8 : Size in Bytes
484 /// uint32 : Padding (only if required to align to 8 byte)
487 /// Location Encoding, Type, Value:
488 /// 0x1, Register, Reg (value in register)
489 /// 0x2, Direct, Reg + Offset (frame index)
490 /// 0x3, Indirect, [Reg + Offset] (spilled value)
491 /// 0x4, Constant, Offset (small constant)
492 /// 0x5, ConstIndex, Constants[Offset] (large constant)
493 void StackMaps::emitCallsiteEntries(MCStreamer &OS) {
494 DEBUG(print(dbgs()));
496 for (const auto &CSI : CSInfos) {
497 const LocationVec &CSLocs = CSI.Locations;
498 const LiveOutVec &LiveOuts = CSI.LiveOuts;
500 // Verify stack map entry. It's better to communicate a problem to the
501 // runtime than crash in case of in-process compilation. Currently, we do
502 // simple overflow checks, but we may eventually communicate other
503 // compilation errors this way.
504 if (CSLocs.size() > UINT16_MAX || LiveOuts.size() > UINT16_MAX) {
505 OS.EmitIntValue(UINT64_MAX, 8); // Invalid ID.
506 OS.EmitValue(CSI.CSOffsetExpr, 4);
507 OS.EmitIntValue(0, 2); // Reserved.
508 OS.EmitIntValue(0, 2); // 0 locations.
509 OS.EmitIntValue(0, 2); // padding.
510 OS.EmitIntValue(0, 2); // 0 live-out registers.
511 OS.EmitIntValue(0, 4); // padding.
515 OS.EmitIntValue(CSI.ID, 8);
516 OS.EmitValue(CSI.CSOffsetExpr, 4);
518 // Reserved for flags.
519 OS.EmitIntValue(0, 2);
520 OS.EmitIntValue(CSLocs.size(), 2);
522 for (const auto &Loc : CSLocs) {
523 OS.EmitIntValue(Loc.Type, 1);
524 OS.EmitIntValue(Loc.Size, 1);
525 OS.EmitIntValue(Loc.Reg, 2);
526 OS.EmitIntValue(Loc.Offset, 4);
529 // Num live-out registers and padding to align to 4 byte.
530 OS.EmitIntValue(0, 2);
531 OS.EmitIntValue(LiveOuts.size(), 2);
533 for (const auto &LO : LiveOuts) {
534 OS.EmitIntValue(LO.DwarfRegNum, 2);
535 OS.EmitIntValue(0, 1);
536 OS.EmitIntValue(LO.Size, 1);
538 // Emit alignment to 8 byte.
539 OS.EmitValueToAlignment(8);
543 /// Serialize the stackmap data.
544 void StackMaps::serializeToStackMapSection() {
546 // Bail out if there's no stack map data.
547 assert((!CSInfos.empty() || ConstPool.empty()) &&
548 "Expected empty constant pool too!");
549 assert((!CSInfos.empty() || FnInfos.empty()) &&
550 "Expected empty function record too!");
554 MCContext &OutContext = AP.OutStreamer->getContext();
555 MCStreamer &OS = *AP.OutStreamer;
557 // Create the section.
558 MCSection *StackMapSection =
559 OutContext.getObjectFileInfo()->getStackMapSection();
560 OS.SwitchSection(StackMapSection);
562 // Emit a dummy symbol to force section inclusion.
563 OS.EmitLabel(OutContext.getOrCreateSymbol(Twine("__LLVM_StackMaps")));
566 DEBUG(dbgs() << "********** Stack Map Output **********\n");
567 emitStackmapHeader(OS);
568 emitFunctionFrameRecords(OS);
569 emitConstantPoolEntries(OS);
570 emitCallsiteEntries(OS);