1 //===- StackMaps.cpp ------------------------------------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 #include "llvm/CodeGen/StackMaps.h"
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/TargetOpcodes.h"
19 #include "llvm/CodeGen/TargetRegisterInfo.h"
20 #include "llvm/CodeGen/TargetSubtargetInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCExpr.h"
24 #include "llvm/MC/MCObjectFileInfo.h"
25 #include "llvm/MC/MCRegisterInfo.h"
26 #include "llvm/MC/MCStreamer.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/MathExtras.h"
31 #include "llvm/Support/raw_ostream.h"
40 #define DEBUG_TYPE "stackmaps"
42 static cl::opt<int> StackMapVersion(
43 "stackmap-version", cl::init(3), cl::Hidden,
44 cl::desc("Specify the stackmap encoding version (default = 3)"));
46 const char *StackMaps::WSMP = "Stack Maps: ";
48 StackMapOpers::StackMapOpers(const MachineInstr *MI)
50 assert(getVarIdx() <= MI->getNumOperands() &&
51 "invalid stackmap definition");
54 PatchPointOpers::PatchPointOpers(const MachineInstr *MI)
55 : MI(MI), HasDef(MI->getOperand(0).isReg() && MI->getOperand(0).isDef() &&
56 !MI->getOperand(0).isImplicit()) {
58 unsigned CheckStartIdx = 0, e = MI->getNumOperands();
59 while (CheckStartIdx < e && MI->getOperand(CheckStartIdx).isReg() &&
60 MI->getOperand(CheckStartIdx).isDef() &&
61 !MI->getOperand(CheckStartIdx).isImplicit())
64 assert(getMetaIdx() == CheckStartIdx &&
65 "Unexpected additional definition in Patchpoint intrinsic.");
69 unsigned PatchPointOpers::getNextScratchIdx(unsigned StartIdx) const {
71 StartIdx = getVarIdx();
73 // Find the next scratch register (implicit def and early clobber)
74 unsigned ScratchIdx = StartIdx, e = MI->getNumOperands();
75 while (ScratchIdx < e &&
76 !(MI->getOperand(ScratchIdx).isReg() &&
77 MI->getOperand(ScratchIdx).isDef() &&
78 MI->getOperand(ScratchIdx).isImplicit() &&
79 MI->getOperand(ScratchIdx).isEarlyClobber()))
82 assert(ScratchIdx != e && "No scratch register available");
86 StackMaps::StackMaps(AsmPrinter &AP) : AP(AP) {
87 if (StackMapVersion != 3)
88 llvm_unreachable("Unsupported stackmap version!");
91 /// Go up the super-register chain until we hit a valid dwarf register number.
92 static unsigned getDwarfRegNum(unsigned Reg, const TargetRegisterInfo *TRI) {
93 int RegNum = TRI->getDwarfRegNum(Reg, false);
94 for (MCSuperRegIterator SR(Reg, TRI); SR.isValid() && RegNum < 0; ++SR)
95 RegNum = TRI->getDwarfRegNum(*SR, false);
97 assert(RegNum >= 0 && "Invalid Dwarf register number.");
98 return (unsigned)RegNum;
101 MachineInstr::const_mop_iterator
102 StackMaps::parseOperand(MachineInstr::const_mop_iterator MOI,
103 MachineInstr::const_mop_iterator MOE, LocationVec &Locs,
104 LiveOutVec &LiveOuts) const {
105 const TargetRegisterInfo *TRI = AP.MF->getSubtarget().getRegisterInfo();
107 switch (MOI->getImm()) {
109 llvm_unreachable("Unrecognized operand type.");
110 case StackMaps::DirectMemRefOp: {
111 auto &DL = AP.MF->getDataLayout();
113 unsigned Size = DL.getPointerSizeInBits();
114 assert((Size % 8) == 0 && "Need pointer size in bytes.");
116 Register Reg = (++MOI)->getReg();
117 int64_t Imm = (++MOI)->getImm();
118 Locs.emplace_back(StackMaps::Location::Direct, Size,
119 getDwarfRegNum(Reg, TRI), Imm);
122 case StackMaps::IndirectMemRefOp: {
123 int64_t Size = (++MOI)->getImm();
124 assert(Size > 0 && "Need a valid size for indirect memory locations.");
125 Register Reg = (++MOI)->getReg();
126 int64_t Imm = (++MOI)->getImm();
127 Locs.emplace_back(StackMaps::Location::Indirect, Size,
128 getDwarfRegNum(Reg, TRI), Imm);
131 case StackMaps::ConstantOp: {
133 assert(MOI->isImm() && "Expected constant operand.");
134 int64_t Imm = MOI->getImm();
135 Locs.emplace_back(Location::Constant, sizeof(int64_t), 0, Imm);
142 // The physical register number will ultimately be encoded as a DWARF regno.
143 // The stack map also records the size of a spill slot that can hold the
144 // register content. (The runtime can track the actual size of the data type
147 // Skip implicit registers (this includes our scratch registers)
148 if (MOI->isImplicit())
151 assert(Register::isPhysicalRegister(MOI->getReg()) &&
152 "Virtreg operands should have been rewritten before now.");
153 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(MOI->getReg());
154 assert(!MOI->getSubReg() && "Physical subreg still around.");
157 unsigned DwarfRegNum = getDwarfRegNum(MOI->getReg(), TRI);
158 unsigned LLVMRegNum = *TRI->getLLVMRegNum(DwarfRegNum, false);
159 unsigned SubRegIdx = TRI->getSubRegIndex(LLVMRegNum, MOI->getReg());
161 Offset = TRI->getSubRegIdxOffset(SubRegIdx);
163 Locs.emplace_back(Location::Register, TRI->getSpillSize(*RC),
164 DwarfRegNum, Offset);
168 if (MOI->isRegLiveOut())
169 LiveOuts = parseRegisterLiveOutMask(MOI->getRegLiveOut());
174 void StackMaps::print(raw_ostream &OS) {
175 const TargetRegisterInfo *TRI =
176 AP.MF ? AP.MF->getSubtarget().getRegisterInfo() : nullptr;
177 OS << WSMP << "callsites:\n";
178 for (const auto &CSI : CSInfos) {
179 const LocationVec &CSLocs = CSI.Locations;
180 const LiveOutVec &LiveOuts = CSI.LiveOuts;
182 OS << WSMP << "callsite " << CSI.ID << "\n";
183 OS << WSMP << " has " << CSLocs.size() << " locations\n";
186 for (const auto &Loc : CSLocs) {
187 OS << WSMP << "\t\tLoc " << Idx << ": ";
189 case Location::Unprocessed:
190 OS << "<Unprocessed operand>";
192 case Location::Register:
195 OS << printReg(Loc.Reg, TRI);
199 case Location::Direct:
202 OS << printReg(Loc.Reg, TRI);
206 OS << " + " << Loc.Offset;
208 case Location::Indirect:
211 OS << printReg(Loc.Reg, TRI);
214 OS << "+" << Loc.Offset;
216 case Location::Constant:
217 OS << "Constant " << Loc.Offset;
219 case Location::ConstantIndex:
220 OS << "Constant Index " << Loc.Offset;
223 OS << "\t[encoding: .byte " << Loc.Type << ", .byte 0"
224 << ", .short " << Loc.Size << ", .short " << Loc.Reg << ", .short 0"
225 << ", .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 << printReg(LO.Reg, TRI);
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 llvm::sort(LiveOuts, [](const LiveOutReg &LHS, const LiveOutReg &RHS) {
271 // Only sort by the dwarf register number.
272 return LHS.DwarfRegNum < RHS.DwarfRegNum;
275 for (auto I = LiveOuts.begin(), E = LiveOuts.end(); I != E; ++I) {
276 for (auto II = std::next(I); II != E; ++II) {
277 if (I->DwarfRegNum != II->DwarfRegNum) {
278 // Skip all the now invalid entries.
282 I->Size = std::max(I->Size, II->Size);
283 if (TRI->isSuperRegister(I->Reg, II->Reg))
285 II->Reg = 0; // mark for deletion.
290 llvm::remove_if(LiveOuts,
291 [](const LiveOutReg &LO) { return LO.Reg == 0; }),
297 void StackMaps::recordStackMapOpers(const MCSymbol &MILabel,
298 const MachineInstr &MI, uint64_t ID,
299 MachineInstr::const_mop_iterator MOI,
300 MachineInstr::const_mop_iterator MOE,
302 MCContext &OutContext = AP.OutStreamer->getContext();
304 LocationVec Locations;
308 assert(PatchPointOpers(&MI).hasDef() && "Stackmap has no return value.");
309 parseOperand(MI.operands_begin(), std::next(MI.operands_begin()), Locations,
315 MOI = parseOperand(MOI, MOE, Locations, LiveOuts);
318 // Move large constants into the constant pool.
319 for (auto &Loc : Locations) {
320 // Constants are encoded as sign-extended integers.
321 // -1 is directly encoded as .long 0xFFFFFFFF with no constant pool.
322 if (Loc.Type == Location::Constant && !isInt<32>(Loc.Offset)) {
323 Loc.Type = Location::ConstantIndex;
324 // ConstPool is intentionally a MapVector of 'uint64_t's (as
325 // opposed to 'int64_t's). We should never be in a situation
326 // where we have to insert either the tombstone or the empty
327 // keys into a map, and for a DenseMap<uint64_t, T> these are
328 // (uint64_t)0 and (uint64_t)-1. They can be and are
329 // represented using 32 bit integers.
330 assert((uint64_t)Loc.Offset != DenseMapInfo<uint64_t>::getEmptyKey() &&
331 (uint64_t)Loc.Offset !=
332 DenseMapInfo<uint64_t>::getTombstoneKey() &&
333 "empty and tombstone keys should fit in 32 bits!");
334 auto Result = ConstPool.insert(std::make_pair(Loc.Offset, Loc.Offset));
335 Loc.Offset = Result.first - ConstPool.begin();
339 // Create an expression to calculate the offset of the callsite from function
341 const MCExpr *CSOffsetExpr = MCBinaryExpr::createSub(
342 MCSymbolRefExpr::create(&MILabel, OutContext),
343 MCSymbolRefExpr::create(AP.CurrentFnSymForSize, OutContext), OutContext);
345 CSInfos.emplace_back(CSOffsetExpr, ID, std::move(Locations),
346 std::move(LiveOuts));
348 // Record the stack size of the current function and update callsite count.
349 const MachineFrameInfo &MFI = AP.MF->getFrameInfo();
350 const TargetRegisterInfo *RegInfo = AP.MF->getSubtarget().getRegisterInfo();
351 bool HasDynamicFrameSize =
352 MFI.hasVarSizedObjects() || RegInfo->needsStackRealignment(*(AP.MF));
353 uint64_t FrameSize = HasDynamicFrameSize ? UINT64_MAX : MFI.getStackSize();
355 auto CurrentIt = FnInfos.find(AP.CurrentFnSym);
356 if (CurrentIt != FnInfos.end())
357 CurrentIt->second.RecordCount++;
359 FnInfos.insert(std::make_pair(AP.CurrentFnSym, FunctionInfo(FrameSize)));
362 void StackMaps::recordStackMap(const MCSymbol &L, const MachineInstr &MI) {
363 assert(MI.getOpcode() == TargetOpcode::STACKMAP && "expected stackmap");
365 StackMapOpers opers(&MI);
366 const int64_t ID = MI.getOperand(PatchPointOpers::IDPos).getImm();
367 recordStackMapOpers(L, MI, ID, std::next(MI.operands_begin(),
372 void StackMaps::recordPatchPoint(const MCSymbol &L, const MachineInstr &MI) {
373 assert(MI.getOpcode() == TargetOpcode::PATCHPOINT && "expected patchpoint");
375 PatchPointOpers opers(&MI);
376 const int64_t ID = opers.getID();
377 auto MOI = std::next(MI.operands_begin(), opers.getStackMapStartIdx());
378 recordStackMapOpers(L, MI, ID, MOI, MI.operands_end(),
379 opers.isAnyReg() && opers.hasDef());
383 auto &Locations = CSInfos.back().Locations;
384 if (opers.isAnyReg()) {
385 unsigned NArgs = opers.getNumCallArgs();
386 for (unsigned i = 0, e = (opers.hasDef() ? NArgs + 1 : NArgs); i != e; ++i)
387 assert(Locations[i].Type == Location::Register &&
388 "anyreg arg must be in reg.");
393 void StackMaps::recordStatepoint(const MCSymbol &L, const MachineInstr &MI) {
394 assert(MI.getOpcode() == TargetOpcode::STATEPOINT && "expected statepoint");
396 StatepointOpers opers(&MI);
397 // Record all the deopt and gc operands (they're contiguous and run from the
398 // initial index to the end of the operand list)
399 const unsigned StartIdx = opers.getVarIdx();
400 recordStackMapOpers(L, MI, opers.getID(), MI.operands_begin() + StartIdx,
401 MI.operands_end(), false);
404 /// Emit the stackmap header.
407 /// uint8 : Stack Map Version (currently 2)
408 /// uint8 : Reserved (expected to be 0)
409 /// uint16 : Reserved (expected to be 0)
411 /// uint32 : NumFunctions
412 /// uint32 : NumConstants
413 /// uint32 : NumRecords
414 void StackMaps::emitStackmapHeader(MCStreamer &OS) {
416 OS.emitIntValue(StackMapVersion, 1); // Version.
417 OS.emitIntValue(0, 1); // Reserved.
418 OS.emitInt16(0); // Reserved.
421 LLVM_DEBUG(dbgs() << WSMP << "#functions = " << FnInfos.size() << '\n');
422 OS.emitInt32(FnInfos.size());
424 LLVM_DEBUG(dbgs() << WSMP << "#constants = " << ConstPool.size() << '\n');
425 OS.emitInt32(ConstPool.size());
427 LLVM_DEBUG(dbgs() << WSMP << "#callsites = " << CSInfos.size() << '\n');
428 OS.emitInt32(CSInfos.size());
431 /// Emit the function frame record for each function.
433 /// StkSizeRecord[NumFunctions] {
434 /// uint64 : Function Address
435 /// uint64 : Stack Size
436 /// uint64 : Record Count
438 void StackMaps::emitFunctionFrameRecords(MCStreamer &OS) {
439 // Function Frame records.
440 LLVM_DEBUG(dbgs() << WSMP << "functions:\n");
441 for (auto const &FR : FnInfos) {
442 LLVM_DEBUG(dbgs() << WSMP << "function addr: " << FR.first
443 << " frame size: " << FR.second.StackSize
444 << " callsite count: " << FR.second.RecordCount << '\n');
445 OS.emitSymbolValue(FR.first, 8);
446 OS.emitIntValue(FR.second.StackSize, 8);
447 OS.emitIntValue(FR.second.RecordCount, 8);
451 /// Emit the constant pool.
453 /// int64 : Constants[NumConstants]
454 void StackMaps::emitConstantPoolEntries(MCStreamer &OS) {
455 // Constant pool entries.
456 LLVM_DEBUG(dbgs() << WSMP << "constants:\n");
457 for (const auto &ConstEntry : ConstPool) {
458 LLVM_DEBUG(dbgs() << WSMP << ConstEntry.second << '\n');
459 OS.emitIntValue(ConstEntry.second, 8);
463 /// Emit the callsite info for each callsite.
465 /// StkMapRecord[NumRecords] {
466 /// uint64 : PatchPoint ID
467 /// uint32 : Instruction Offset
468 /// uint16 : Reserved (record flags)
469 /// uint16 : NumLocations
470 /// Location[NumLocations] {
471 /// uint8 : Register | Direct | Indirect | Constant | ConstantIndex
472 /// uint8 : Size in Bytes
473 /// uint16 : Dwarf RegNum
477 /// uint16 : NumLiveOuts
478 /// LiveOuts[NumLiveOuts] {
479 /// uint16 : Dwarf RegNum
481 /// uint8 : Size in Bytes
483 /// uint32 : Padding (only if required to align to 8 byte)
486 /// Location Encoding, Type, Value:
487 /// 0x1, Register, Reg (value in register)
488 /// 0x2, Direct, Reg + Offset (frame index)
489 /// 0x3, Indirect, [Reg + Offset] (spilled value)
490 /// 0x4, Constant, Offset (small constant)
491 /// 0x5, ConstIndex, Constants[Offset] (large constant)
492 void StackMaps::emitCallsiteEntries(MCStreamer &OS) {
493 LLVM_DEBUG(print(dbgs()));
495 for (const auto &CSI : CSInfos) {
496 const LocationVec &CSLocs = CSI.Locations;
497 const LiveOutVec &LiveOuts = CSI.LiveOuts;
499 // Verify stack map entry. It's better to communicate a problem to the
500 // runtime than crash in case of in-process compilation. Currently, we do
501 // simple overflow checks, but we may eventually communicate other
502 // compilation errors this way.
503 if (CSLocs.size() > UINT16_MAX || LiveOuts.size() > UINT16_MAX) {
504 OS.emitIntValue(UINT64_MAX, 8); // Invalid ID.
505 OS.emitValue(CSI.CSOffsetExpr, 4);
506 OS.emitInt16(0); // Reserved.
507 OS.emitInt16(0); // 0 locations.
508 OS.emitInt16(0); // padding.
509 OS.emitInt16(0); // 0 live-out registers.
510 OS.emitInt32(0); // padding.
514 OS.emitIntValue(CSI.ID, 8);
515 OS.emitValue(CSI.CSOffsetExpr, 4);
517 // Reserved for flags.
519 OS.emitInt16(CSLocs.size());
521 for (const auto &Loc : CSLocs) {
522 OS.emitIntValue(Loc.Type, 1);
523 OS.emitIntValue(0, 1); // Reserved
524 OS.emitInt16(Loc.Size);
525 OS.emitInt16(Loc.Reg);
526 OS.emitInt16(0); // Reserved
527 OS.emitInt32(Loc.Offset);
530 // Emit alignment to 8 byte.
531 OS.emitValueToAlignment(8);
533 // Num live-out registers and padding to align to 4 byte.
535 OS.emitInt16(LiveOuts.size());
537 for (const auto &LO : LiveOuts) {
538 OS.emitInt16(LO.DwarfRegNum);
539 OS.emitIntValue(0, 1);
540 OS.emitIntValue(LO.Size, 1);
542 // Emit alignment to 8 byte.
543 OS.emitValueToAlignment(8);
547 /// Serialize the stackmap data.
548 void StackMaps::serializeToStackMapSection() {
550 // Bail out if there's no stack map data.
551 assert((!CSInfos.empty() || ConstPool.empty()) &&
552 "Expected empty constant pool too!");
553 assert((!CSInfos.empty() || FnInfos.empty()) &&
554 "Expected empty function record too!");
558 MCContext &OutContext = AP.OutStreamer->getContext();
559 MCStreamer &OS = *AP.OutStreamer;
561 // Create the section.
562 MCSection *StackMapSection =
563 OutContext.getObjectFileInfo()->getStackMapSection();
564 OS.SwitchSection(StackMapSection);
566 // Emit a dummy symbol to force section inclusion.
567 OS.emitLabel(OutContext.getOrCreateSymbol(Twine("__LLVM_StackMaps")));
570 LLVM_DEBUG(dbgs() << "********** Stack Map Output **********\n");
571 emitStackmapHeader(OS);
572 emitFunctionFrameRecords(OS);
573 emitConstantPoolEntries(OS);
574 emitCallsiteEntries(OS);