diff options
Diffstat (limited to 'js/src/jit/shared/IonAssemblerBufferWithConstantPools.h')
| -rw-r--r-- | js/src/jit/shared/IonAssemblerBufferWithConstantPools.h | 1215 |
1 files changed, 1215 insertions, 0 deletions
diff --git a/js/src/jit/shared/IonAssemblerBufferWithConstantPools.h b/js/src/jit/shared/IonAssemblerBufferWithConstantPools.h new file mode 100644 index 0000000000..d21bb7f35b --- /dev/null +++ b/js/src/jit/shared/IonAssemblerBufferWithConstantPools.h @@ -0,0 +1,1215 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- + * vim: set ts=8 sts=2 et sw=2 tw=80: + * This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#ifndef jit_shared_IonAssemblerBufferWithConstantPools_h +#define jit_shared_IonAssemblerBufferWithConstantPools_h + +#include "mozilla/CheckedInt.h" +#include "mozilla/MathAlgorithms.h" + +#include <algorithm> + +#include "jit/JitSpewer.h" +#include "jit/shared/IonAssemblerBuffer.h" + +// [SMDOC] JIT AssemblerBuffer constant pooling (ARM/ARM64/MIPS) +// +// This code extends the AssemblerBuffer to support the pooling of values loaded +// using program-counter relative addressing modes. This is necessary with the +// ARM instruction set because it has a fixed instruction size that can not +// encode all values as immediate arguments in instructions. Pooling the values +// allows the values to be placed in large chunks which minimizes the number of +// forced branches around them in the code. This is used for loading floating +// point constants, for loading 32 bit constants on the ARMv6, for absolute +// branch targets, and in future will be needed for large branches on the ARMv6. +// +// For simplicity of the implementation, the constant pools are always placed +// after the loads referencing them. When a new constant pool load is added to +// the assembler buffer, a corresponding pool entry is added to the current +// pending pool. The finishPool() method copies the current pending pool entries +// into the assembler buffer at the current offset and patches the pending +// constant pool load instructions. +// +// Before inserting instructions or pool entries, it is necessary to determine +// if doing so would place a pending pool entry out of reach of an instruction, +// and if so then the pool must firstly be dumped. With the allocation algorithm +// used below, the recalculation of all the distances between instructions and +// their pool entries can be avoided by noting that there will be a limiting +// instruction and pool entry pair that does not change when inserting more +// instructions. Adding more instructions makes the same increase to the +// distance, between instructions and their pool entries, for all such +// pairs. This pair is recorded as the limiter, and it is updated when new pool +// entries are added, see updateLimiter() +// +// The pools consist of: a guard instruction that branches around the pool, a +// header word that helps identify a pool in the instruction stream, and then +// the pool entries allocated in units of words. The guard instruction could be +// omitted if control does not reach the pool, and this is referred to as a +// natural guard below, but for simplicity the guard branch is always +// emitted. The pool header is an identifiable word that in combination with the +// guard uniquely identifies a pool in the instruction stream. The header also +// encodes the pool size and a flag indicating if the guard is natural. It is +// possible to iterate through the code instructions skipping or examining the +// pools. E.g. it might be necessary to skip pools when search for, or patching, +// an instruction sequence. +// +// It is often required to keep a reference to a pool entry, to patch it after +// the buffer is finished. Each pool entry is assigned a unique index, counting +// up from zero (see the poolEntryCount slot below). These can be mapped back to +// the offset of the pool entry in the finished buffer, see poolEntryOffset(). +// +// The code supports no-pool regions, and for these the size of the region, in +// instructions, must be supplied. This size is used to determine if inserting +// the instructions would place a pool entry out of range, and if so then a pool +// is firstly flushed. The DEBUG code checks that the emitted code is within the +// supplied size to detect programming errors. See enterNoPool() and +// leaveNoPool(). + +// The only planned instruction sets that require inline constant pools are the +// ARM, ARM64, and MIPS, and these all have fixed 32-bit sized instructions so +// for simplicity the code below is specialized for fixed 32-bit sized +// instructions and makes no attempt to support variable length +// instructions. The base assembler buffer which supports variable width +// instruction is used by the x86 and x64 backends. + +// The AssemblerBufferWithConstantPools template class uses static callbacks to +// the provided Asm template argument class: +// +// void Asm::InsertIndexIntoTag(uint8_t* load_, uint32_t index) +// +// When allocEntry() is called to add a constant pool load with an associated +// constant pool entry, this callback is called to encode the index of the +// allocated constant pool entry into the load instruction. +// +// After the constant pool has been placed, PatchConstantPoolLoad() is called +// to update the load instruction with the right load offset. +// +// void Asm::WritePoolGuard(BufferOffset branch, +// Instruction* dest, +// BufferOffset afterPool) +// +// Write out the constant pool guard branch before emitting the pool. +// +// branch +// Offset of the guard branch in the buffer. +// +// dest +// Pointer into the buffer where the guard branch should be emitted. (Same +// as getInst(branch)). Space for guardSize_ instructions has been reserved. +// +// afterPool +// Offset of the first instruction after the constant pool. This includes +// both pool entries and branch veneers added after the pool data. +// +// void Asm::WritePoolHeader(uint8_t* start, Pool* p, bool isNatural) +// +// Write out the pool header which follows the guard branch. +// +// void Asm::PatchConstantPoolLoad(void* loadAddr, void* constPoolAddr) +// +// Re-encode a load of a constant pool entry after the location of the +// constant pool is known. +// +// The load instruction at loadAddr was previously passed to +// InsertIndexIntoTag(). The constPoolAddr is the final address of the +// constant pool in the assembler buffer. +// +// void Asm::PatchShortRangeBranchToVeneer(AssemblerBufferWithConstantPools*, +// unsigned rangeIdx, +// BufferOffset deadline, +// BufferOffset veneer) +// +// Patch a short-range branch to jump through a veneer before it goes out of +// range. +// +// rangeIdx, deadline +// These arguments were previously passed to registerBranchDeadline(). It is +// assumed that PatchShortRangeBranchToVeneer() knows how to compute the +// offset of the short-range branch from this information. +// +// veneer +// Space for a branch veneer, guaranteed to be <= deadline. At this +// position, guardSize_ * InstSize bytes are allocated. They should be +// initialized to the proper unconditional branch instruction. +// +// Unbound branches to the same unbound label are organized as a linked list: +// +// Label::offset -> Branch1 -> Branch2 -> Branch3 -> nil +// +// This callback should insert a new veneer branch into the list: +// +// Label::offset -> Branch1 -> Branch2 -> Veneer -> Branch3 -> nil +// +// When Assembler::bind() rewrites the branches with the real label offset, it +// probably has to bind Branch2 to target the veneer branch instead of jumping +// straight to the label. + +namespace js { +namespace jit { + +// BranchDeadlineSet - Keep track of pending branch deadlines. +// +// Some architectures like arm and arm64 have branch instructions with limited +// range. When assembling a forward branch, it is not always known if the final +// target label will be in range of the branch instruction. +// +// The BranchDeadlineSet data structure is used to keep track of the set of +// pending forward branches. It supports the following fast operations: +// +// 1. Get the earliest deadline in the set. +// 2. Add a new branch deadline. +// 3. Remove a branch deadline. +// +// Architectures may have different branch encodings with different ranges. Each +// supported range is assigned a small integer starting at 0. This data +// structure does not care about the actual range of branch instructions, just +// the latest buffer offset that can be reached - the deadline offset. +// +// Branched are stored as (rangeIdx, deadline) tuples. The target-specific code +// can compute the location of the branch itself from this information. This +// data structure does not need to know. +// +template <unsigned NumRanges> +class BranchDeadlineSet { + // Maintain a list of pending deadlines for each range separately. + // + // The offsets in each vector are always kept in ascending order. + // + // Because we have a separate vector for different ranges, as forward + // branches are added to the assembler buffer, their deadlines will + // always be appended to the vector corresponding to their range. + // + // When binding labels, we expect a more-or-less LIFO order of branch + // resolutions. This would always hold if we had strictly structured control + // flow. + // + // We allow branch deadlines to be added and removed in any order, but + // performance is best in the expected case of near LIFO order. + // + typedef Vector<BufferOffset, 8, LifoAllocPolicy<Fallible>> RangeVector; + + // We really just want "RangeVector deadline_[NumRanges];", but each vector + // needs to be initialized with a LifoAlloc, and C++ doesn't bend that way. + // + // Use raw aligned storage instead and explicitly construct NumRanges + // vectors in our constructor. + mozilla::AlignedStorage2<RangeVector[NumRanges]> deadlineStorage_; + + // Always access the range vectors through this method. + RangeVector& vectorForRange(unsigned rangeIdx) { + MOZ_ASSERT(rangeIdx < NumRanges, "Invalid branch range index"); + return (*deadlineStorage_.addr())[rangeIdx]; + } + + const RangeVector& vectorForRange(unsigned rangeIdx) const { + MOZ_ASSERT(rangeIdx < NumRanges, "Invalid branch range index"); + return (*deadlineStorage_.addr())[rangeIdx]; + } + + // Maintain a precomputed earliest deadline at all times. + // This is unassigned only when all deadline vectors are empty. + BufferOffset earliest_; + + // The range vector owning earliest_. Uninitialized when empty. + unsigned earliestRange_; + + // Recompute the earliest deadline after it's been invalidated. + void recomputeEarliest() { + earliest_ = BufferOffset(); + for (unsigned r = 0; r < NumRanges; r++) { + auto& vec = vectorForRange(r); + if (!vec.empty() && (!earliest_.assigned() || vec[0] < earliest_)) { + earliest_ = vec[0]; + earliestRange_ = r; + } + } + } + + // Update the earliest deadline if needed after inserting (rangeIdx, + // deadline). Always return true for convenience: + // return insert() && updateEarliest(). + bool updateEarliest(unsigned rangeIdx, BufferOffset deadline) { + if (!earliest_.assigned() || deadline < earliest_) { + earliest_ = deadline; + earliestRange_ = rangeIdx; + } + return true; + } + + public: + explicit BranchDeadlineSet(LifoAlloc& alloc) : earliestRange_(0) { + // Manually construct vectors in the uninitialized aligned storage. + // This is because C++ arrays can otherwise only be constructed with + // the default constructor. + for (unsigned r = 0; r < NumRanges; r++) { + new (&vectorForRange(r)) RangeVector(alloc); + } + } + + ~BranchDeadlineSet() { + // Aligned storage doesn't destruct its contents automatically. + for (unsigned r = 0; r < NumRanges; r++) { + vectorForRange(r).~RangeVector(); + } + } + + // Is this set completely empty? + bool empty() const { return !earliest_.assigned(); } + + // Get the total number of deadlines in the set. + size_t size() const { + size_t count = 0; + for (unsigned r = 0; r < NumRanges; r++) { + count += vectorForRange(r).length(); + } + return count; + } + + // Get the number of deadlines for the range with the most elements. + size_t maxRangeSize() const { + size_t count = 0; + for (unsigned r = 0; r < NumRanges; r++) { + count = std::max(count, vectorForRange(r).length()); + } + return count; + } + + // Get the first deadline that is still in the set. + BufferOffset earliestDeadline() const { + MOZ_ASSERT(!empty()); + return earliest_; + } + + // Get the range index corresponding to earliestDeadlineRange(). + unsigned earliestDeadlineRange() const { + MOZ_ASSERT(!empty()); + return earliestRange_; + } + + // Add a (rangeIdx, deadline) tuple to the set. + // + // It is assumed that this tuple is not already in the set. + // This function performs best id the added deadline is later than any + // existing deadline for the same range index. + // + // Return true if the tuple was added, false if the tuple could not be added + // because of an OOM error. + bool addDeadline(unsigned rangeIdx, BufferOffset deadline) { + MOZ_ASSERT(deadline.assigned(), "Can only store assigned buffer offsets"); + // This is the vector where deadline should be saved. + auto& vec = vectorForRange(rangeIdx); + + // Fast case: Simple append to the relevant array. This never affects + // the earliest deadline. + if (!vec.empty() && vec.back() < deadline) { + return vec.append(deadline); + } + + // Fast case: First entry to the vector. We need to update earliest_. + if (vec.empty()) { + return vec.append(deadline) && updateEarliest(rangeIdx, deadline); + } + + return addDeadlineSlow(rangeIdx, deadline); + } + + private: + // General case of addDeadline. This is split into two functions such that + // the common case in addDeadline can be inlined while this part probably + // won't inline. + bool addDeadlineSlow(unsigned rangeIdx, BufferOffset deadline) { + auto& vec = vectorForRange(rangeIdx); + + // Inserting into the middle of the vector. Use a log time binary search + // and a linear time insert(). + // Is it worthwhile special-casing the empty vector? + auto at = std::lower_bound(vec.begin(), vec.end(), deadline); + MOZ_ASSERT(at == vec.end() || *at != deadline, + "Cannot insert duplicate deadlines"); + return vec.insert(at, deadline) && updateEarliest(rangeIdx, deadline); + } + + public: + // Remove a deadline from the set. + // If (rangeIdx, deadline) is not in the set, nothing happens. + void removeDeadline(unsigned rangeIdx, BufferOffset deadline) { + auto& vec = vectorForRange(rangeIdx); + + if (vec.empty()) { + return; + } + + if (deadline == vec.back()) { + // Expected fast case: Structured control flow causes forward + // branches to be bound in reverse order. + vec.popBack(); + } else { + // Slow case: Binary search + linear erase. + auto where = std::lower_bound(vec.begin(), vec.end(), deadline); + if (where == vec.end() || *where != deadline) { + return; + } + vec.erase(where); + } + if (deadline == earliest_) { + recomputeEarliest(); + } + } +}; + +// Specialization for architectures that don't need to track short-range +// branches. +template <> +class BranchDeadlineSet<0u> { + public: + explicit BranchDeadlineSet(LifoAlloc& alloc) {} + bool empty() const { return true; } + size_t size() const { return 0; } + size_t maxRangeSize() const { return 0; } + BufferOffset earliestDeadline() const { MOZ_CRASH(); } + unsigned earliestDeadlineRange() const { MOZ_CRASH(); } + bool addDeadline(unsigned rangeIdx, BufferOffset deadline) { MOZ_CRASH(); } + void removeDeadline(unsigned rangeIdx, BufferOffset deadline) { MOZ_CRASH(); } +}; + +// The allocation unit size for pools. +typedef int32_t PoolAllocUnit; + +// Hysteresis given to short-range branches. +// +// If any short-range branches will go out of range in the next N bytes, +// generate a veneer for them in the current pool. The hysteresis prevents the +// creation of many tiny constant pools for branch veneers. +const size_t ShortRangeBranchHysteresis = 128; + +struct Pool { + private: + // The maximum program-counter relative offset below which the instruction + // set can encode. Different classes of intructions might support different + // ranges but for simplicity the minimum is used here, and for the ARM this + // is constrained to 1024 by the float load instructions. + const size_t maxOffset_; + // An offset to apply to program-counter relative offsets. The ARM has a + // bias of 8. + const unsigned bias_; + + // The content of the pool entries. + Vector<PoolAllocUnit, 8, LifoAllocPolicy<Fallible>> poolData_; + + // Flag that tracks OOM conditions. This is set after any append failed. + bool oom_; + + // The limiting instruction and pool-entry pair. The instruction program + // counter relative offset of this limiting instruction will go out of range + // first as the pool position moves forward. It is more efficient to track + // just this limiting pair than to recheck all offsets when testing if the + // pool needs to be dumped. + // + // 1. The actual offset of the limiting instruction referencing the limiting + // pool entry. + BufferOffset limitingUser; + // 2. The pool entry index of the limiting pool entry. + unsigned limitingUsee; + + public: + // A record of the code offset of instructions that reference pool + // entries. These instructions need to be patched when the actual position + // of the instructions and pools are known, and for the code below this + // occurs when each pool is finished, see finishPool(). + Vector<BufferOffset, 8, LifoAllocPolicy<Fallible>> loadOffsets; + + // Create a Pool. Don't allocate anything from lifoAloc, just capture its + // reference. + explicit Pool(size_t maxOffset, unsigned bias, LifoAlloc& lifoAlloc) + : maxOffset_(maxOffset), + bias_(bias), + poolData_(lifoAlloc), + oom_(false), + limitingUser(), + limitingUsee(INT_MIN), + loadOffsets(lifoAlloc) {} + + // If poolData() returns nullptr then oom_ will also be true. + const PoolAllocUnit* poolData() const { return poolData_.begin(); } + + unsigned numEntries() const { return poolData_.length(); } + + size_t getPoolSize() const { return numEntries() * sizeof(PoolAllocUnit); } + + bool oom() const { return oom_; } + + // Update the instruction/pool-entry pair that limits the position of the + // pool. The nextInst is the actual offset of the new instruction being + // allocated. + // + // This is comparing the offsets, see checkFull() below for the equation, + // but common expressions on both sides have been canceled from the ranges + // being compared. Notably, the poolOffset cancels out, so the limiting pair + // does not depend on where the pool is placed. + void updateLimiter(BufferOffset nextInst) { + ptrdiff_t oldRange = + limitingUsee * sizeof(PoolAllocUnit) - limitingUser.getOffset(); + ptrdiff_t newRange = getPoolSize() - nextInst.getOffset(); + if (!limitingUser.assigned() || newRange > oldRange) { + // We have a new largest range! + limitingUser = nextInst; + limitingUsee = numEntries(); + } + } + + // Check if inserting a pool at the actual offset poolOffset would place + // pool entries out of reach. This is called before inserting instructions + // to check that doing so would not push pool entries out of reach, and if + // so then the pool would need to be firstly dumped. The poolOffset is the + // first word of the pool, after the guard and header and alignment fill. + bool checkFull(size_t poolOffset) const { + // Not full if there are no uses. + if (!limitingUser.assigned()) { + return false; + } + size_t offset = poolOffset + limitingUsee * sizeof(PoolAllocUnit) - + (limitingUser.getOffset() + bias_); + return offset >= maxOffset_; + } + + static const unsigned OOM_FAIL = unsigned(-1); + + unsigned insertEntry(unsigned num, uint8_t* data, BufferOffset off, + LifoAlloc& lifoAlloc) { + if (oom_) { + return OOM_FAIL; + } + unsigned ret = numEntries(); + if (!poolData_.append((PoolAllocUnit*)data, num) || + !loadOffsets.append(off)) { + oom_ = true; + return OOM_FAIL; + } + return ret; + } + + void reset() { + poolData_.clear(); + loadOffsets.clear(); + + limitingUser = BufferOffset(); + limitingUsee = -1; + } +}; + +// Template arguments: +// +// SliceSize +// Number of bytes in each allocated BufferSlice. See +// AssemblerBuffer::SliceSize. +// +// InstSize +// Size in bytes of the fixed-size instructions. This should be equal to +// sizeof(Inst). This is only needed here because the buffer is defined before +// the Instruction. +// +// Inst +// The actual type used to represent instructions. This is only really used as +// the return type of the getInst() method. +// +// Asm +// Class defining the needed static callback functions. See documentation of +// the Asm::* callbacks above. +// +// NumShortBranchRanges +// The number of short branch ranges to support. This can be 0 if no support +// for tracking short range branches is needed. The +// AssemblerBufferWithConstantPools class does not need to know what the range +// of branches is - it deals in branch 'deadlines' which is the last buffer +// position that a short-range forward branch can reach. It is assumed that +// the Asm class is able to find the actual branch instruction given a +// (range-index, deadline) pair. +// +// +template <size_t SliceSize, size_t InstSize, class Inst, class Asm, + unsigned NumShortBranchRanges = 0> +struct AssemblerBufferWithConstantPools + : public AssemblerBuffer<SliceSize, Inst> { + private: + // The PoolEntry index counter. Each PoolEntry is given a unique index, + // counting up from zero, and these can be mapped back to the actual pool + // entry offset after finishing the buffer, see poolEntryOffset(). + size_t poolEntryCount; + + public: + class PoolEntry { + size_t index_; + + public: + explicit PoolEntry(size_t index) : index_(index) {} + + PoolEntry() : index_(-1) {} + + size_t index() const { return index_; } + }; + + private: + typedef AssemblerBuffer<SliceSize, Inst> Parent; + using typename Parent::Slice; + + // The size of a pool guard, in instructions. A branch around the pool. + const unsigned guardSize_; + // The size of the header that is put at the beginning of a full pool, in + // instruction sized units. + const unsigned headerSize_; + + // The maximum pc relative offset encoded in instructions that reference + // pool entries. This is generally set to the maximum offset that can be + // encoded by the instructions, but for testing can be lowered to affect the + // pool placement and frequency of pool placement. + const size_t poolMaxOffset_; + + // The bias on pc relative addressing mode offsets, in units of bytes. The + // ARM has a bias of 8 bytes. + const unsigned pcBias_; + + // The current working pool. Copied out as needed before resetting. + Pool pool_; + + // The buffer should be aligned to this address. + const size_t instBufferAlign_; + + struct PoolInfo { + // The index of the first entry in this pool. + // Pool entries are numbered uniquely across all pools, starting from 0. + unsigned firstEntryIndex; + + // The location of this pool's first entry in the main assembler buffer. + // Note that the pool guard and header come before this offset which + // points directly at the data. + BufferOffset offset; + + explicit PoolInfo(unsigned index, BufferOffset data) + : firstEntryIndex(index), offset(data) {} + }; + + // Info for each pool that has already been dumped. This does not include + // any entries in pool_. + Vector<PoolInfo, 8, LifoAllocPolicy<Fallible>> poolInfo_; + + // Set of short-range forward branches that have not yet been bound. + // We may need to insert veneers if the final label turns out to be out of + // range. + // + // This set stores (rangeIdx, deadline) pairs instead of the actual branch + // locations. + BranchDeadlineSet<NumShortBranchRanges> branchDeadlines_; + + // When true dumping pools is inhibited. + bool canNotPlacePool_; + +#ifdef DEBUG + // State for validating the 'maxInst' argument to enterNoPool(). + // The buffer offset when entering the no-pool region. + size_t canNotPlacePoolStartOffset_; + // The maximum number of word sized instructions declared for the no-pool + // region. + size_t canNotPlacePoolMaxInst_; +#endif + + // Instruction to use for alignment fill. + const uint32_t alignFillInst_; + + // Insert a number of NOP instructions between each requested instruction at + // all locations at which a pool can potentially spill. This is useful for + // checking that instruction locations are correctly referenced and/or + // followed. + const uint32_t nopFillInst_; + const unsigned nopFill_; + + // For inhibiting the insertion of fill NOPs in the dynamic context in which + // they are being inserted. + bool inhibitNops_; + + public: + // A unique id within each JitContext, to identify pools in the debug + // spew. Set by the MacroAssembler, see getNextAssemblerId(). + int id; + + private: + // The buffer slices are in a double linked list. + Slice* getHead() const { return this->head; } + Slice* getTail() const { return this->tail; } + + public: + // Create an assembler buffer. + // Note that this constructor is not allowed to actually allocate memory from + // this->lifoAlloc_ because the MacroAssembler constructor has not yet created + // an AutoJitContextAlloc. + AssemblerBufferWithConstantPools(unsigned guardSize, unsigned headerSize, + size_t instBufferAlign, size_t poolMaxOffset, + unsigned pcBias, uint32_t alignFillInst, + uint32_t nopFillInst, unsigned nopFill = 0) + : poolEntryCount(0), + guardSize_(guardSize), + headerSize_(headerSize), + poolMaxOffset_(poolMaxOffset), + pcBias_(pcBias), + pool_(poolMaxOffset, pcBias, this->lifoAlloc_), + instBufferAlign_(instBufferAlign), + poolInfo_(this->lifoAlloc_), + branchDeadlines_(this->lifoAlloc_), + canNotPlacePool_(false), +#ifdef DEBUG + canNotPlacePoolStartOffset_(0), + canNotPlacePoolMaxInst_(0), +#endif + alignFillInst_(alignFillInst), + nopFillInst_(nopFillInst), + nopFill_(nopFill), + inhibitNops_(false), + id(-1) { + } + + // We need to wait until an AutoJitContextAlloc is created by the + // MacroAssembler before allocating any space. + void initWithAllocator() { + // We hand out references to lifoAlloc_ in the constructor. + // Check that no allocations were made then. + MOZ_ASSERT(this->lifoAlloc_.isEmpty(), + "Illegal LIFO allocations before AutoJitContextAlloc"); + } + + private: + size_t sizeExcludingCurrentPool() const { + // Return the actual size of the buffer, excluding the current pending + // pool. + return this->nextOffset().getOffset(); + } + + public: + size_t size() const { + // Return the current actual size of the buffer. This is only accurate + // if there are no pending pool entries to dump, check. + MOZ_ASSERT_IF(!this->oom(), pool_.numEntries() == 0); + return sizeExcludingCurrentPool(); + } + + private: + void insertNopFill() { + // Insert fill for testing. + if (nopFill_ > 0 && !inhibitNops_ && !canNotPlacePool_) { + inhibitNops_ = true; + + // Fill using a branch-nop rather than a NOP so this can be + // distinguished and skipped. + for (size_t i = 0; i < nopFill_; i++) { + putInt(nopFillInst_); + } + + inhibitNops_ = false; + } + } + + static const unsigned OOM_FAIL = unsigned(-1); + static const unsigned DUMMY_INDEX = unsigned(-2); + + // Check if it is possible to add numInst instructions and numPoolEntries + // constant pool entries without needing to flush the current pool. + bool hasSpaceForInsts(unsigned numInsts, unsigned numPoolEntries) const { + size_t nextOffset = sizeExcludingCurrentPool(); + // Earliest starting offset for the current pool after adding numInsts. + // This is the beginning of the pool entries proper, after inserting a + // guard branch + pool header. + size_t poolOffset = + nextOffset + (numInsts + guardSize_ + headerSize_) * InstSize; + + // Any constant pool loads that would go out of range? + if (pool_.checkFull(poolOffset)) { + return false; + } + + // Any short-range branch that would go out of range? + if (!branchDeadlines_.empty()) { + size_t deadline = branchDeadlines_.earliestDeadline().getOffset(); + size_t poolEnd = poolOffset + pool_.getPoolSize() + + numPoolEntries * sizeof(PoolAllocUnit); + + // When NumShortBranchRanges > 1, is is possible for branch deadlines to + // expire faster than we can insert veneers. Suppose branches are 4 bytes + // each, we could have the following deadline set: + // + // Range 0: 40, 44, 48 + // Range 1: 44, 48 + // + // It is not good enough to start inserting veneers at the 40 deadline; we + // would not be able to create veneers for the second 44 deadline. + // Instead, we need to start at 32: + // + // 32: veneer(40) + // 36: veneer(44) + // 40: veneer(44) + // 44: veneer(48) + // 48: veneer(48) + // + // This is a pretty conservative solution to the problem: If we begin at + // the earliest deadline, we can always emit all veneers for the range + // that currently has the most pending deadlines. That may not leave room + // for veneers for the remaining ranges, so reserve space for those + // secondary range veneers assuming the worst case deadlines. + + // Total pending secondary range veneer size. + size_t secondaryVeneers = guardSize_ * (branchDeadlines_.size() - + branchDeadlines_.maxRangeSize()); + + if (deadline < poolEnd + secondaryVeneers) { + return false; + } + } + + return true; + } + + unsigned insertEntryForwards(unsigned numInst, unsigned numPoolEntries, + uint8_t* inst, uint8_t* data) { + // If inserting pool entries then find a new limiter before we do the + // range check. + if (numPoolEntries) { + pool_.updateLimiter(BufferOffset(sizeExcludingCurrentPool())); + } + + if (!hasSpaceForInsts(numInst, numPoolEntries)) { + if (numPoolEntries) { + JitSpew(JitSpew_Pools, "[%d] Inserting pool entry caused a spill", id); + } else { + JitSpew(JitSpew_Pools, "[%d] Inserting instruction(%zu) caused a spill", + id, sizeExcludingCurrentPool()); + } + + finishPool(numInst * InstSize); + if (this->oom()) { + return OOM_FAIL; + } + return insertEntryForwards(numInst, numPoolEntries, inst, data); + } + if (numPoolEntries) { + unsigned result = pool_.insertEntry(numPoolEntries, data, + this->nextOffset(), this->lifoAlloc_); + if (result == Pool::OOM_FAIL) { + this->fail_oom(); + return OOM_FAIL; + } + return result; + } + + // The pool entry index is returned above when allocating an entry, but + // when not allocating an entry a dummy value is returned - it is not + // expected to be used by the caller. + return DUMMY_INDEX; + } + + public: + // Get the next buffer offset where an instruction would be inserted. + // This may flush the current constant pool before returning nextOffset(). + BufferOffset nextInstrOffset() { + if (!hasSpaceForInsts(/* numInsts= */ 1, /* numPoolEntries= */ 0)) { + JitSpew(JitSpew_Pools, + "[%d] nextInstrOffset @ %d caused a constant pool spill", id, + this->nextOffset().getOffset()); + finishPool(ShortRangeBranchHysteresis); + } + return this->nextOffset(); + } + + MOZ_NEVER_INLINE + BufferOffset allocEntry(size_t numInst, unsigned numPoolEntries, + uint8_t* inst, uint8_t* data, + PoolEntry* pe = nullptr) { + // The allocation of pool entries is not supported in a no-pool region, + // check. + MOZ_ASSERT_IF(numPoolEntries, !canNotPlacePool_); + + if (this->oom()) { + return BufferOffset(); + } + + insertNopFill(); + +#ifdef JS_JITSPEW + if (numPoolEntries && JitSpewEnabled(JitSpew_Pools)) { + JitSpew(JitSpew_Pools, "[%d] Inserting %d entries into pool", id, + numPoolEntries); + JitSpewStart(JitSpew_Pools, "[%d] data is: 0x", id); + size_t length = numPoolEntries * sizeof(PoolAllocUnit); + for (unsigned idx = 0; idx < length; idx++) { + JitSpewCont(JitSpew_Pools, "%02x", data[length - idx - 1]); + if (((idx & 3) == 3) && (idx + 1 != length)) { + JitSpewCont(JitSpew_Pools, "_"); + } + } + JitSpewFin(JitSpew_Pools); + } +#endif + + // Insert the pool value. + unsigned index = insertEntryForwards(numInst, numPoolEntries, inst, data); + if (this->oom()) { + return BufferOffset(); + } + + // Now to get an instruction to write. + PoolEntry retPE; + if (numPoolEntries) { + JitSpew(JitSpew_Pools, "[%d] Entry has index %u, offset %zu", id, index, + sizeExcludingCurrentPool()); + Asm::InsertIndexIntoTag(inst, index); + // Figure out the offset within the pool entries. + retPE = PoolEntry(poolEntryCount); + poolEntryCount += numPoolEntries; + } + // Now inst is a valid thing to insert into the instruction stream. + if (pe != nullptr) { + *pe = retPE; + } + return this->putBytes(numInst * InstSize, inst); + } + + // putInt is the workhorse for the assembler and higher-level buffer + // abstractions: it places one instruction into the instruction stream. + // Under normal circumstances putInt should just check that the constant + // pool does not need to be flushed, that there is space for the single word + // of the instruction, and write that word and update the buffer pointer. + // + // To do better here we need a status variable that handles both nopFill_ + // and capacity, so that we can quickly know whether to go the slow path. + // That could be a variable that has the remaining number of simple + // instructions that can be inserted before a more expensive check, + // which is set to zero when nopFill_ is set. + // + // We assume that we don't have to check this->oom() if there is space to + // insert a plain instruction; there will always come a later time when it + // will be checked anyway. + + MOZ_ALWAYS_INLINE + BufferOffset putInt(uint32_t value) { + if (nopFill_ || + !hasSpaceForInsts(/* numInsts= */ 1, /* numPoolEntries= */ 0)) { + return allocEntry(1, 0, (uint8_t*)&value, nullptr, nullptr); + } + +#if defined(JS_CODEGEN_ARM) || defined(JS_CODEGEN_ARM64) || \ + defined(JS_CODEGEN_MIPS32) || defined(JS_CODEGEN_MIPS64) + return this->putU32Aligned(value); +#else + return this->AssemblerBuffer<SliceSize, Inst>::putInt(value); +#endif + } + + // Register a short-range branch deadline. + // + // After inserting a short-range forward branch, call this method to + // register the branch 'deadline' which is the last buffer offset that the + // branch instruction can reach. + // + // When the branch is bound to a destination label, call + // unregisterBranchDeadline() to stop tracking this branch, + // + // If the assembled code is about to exceed the registered branch deadline, + // and unregisterBranchDeadline() has not yet been called, an + // instruction-sized constant pool entry is allocated before the branch + // deadline. + // + // rangeIdx + // A number < NumShortBranchRanges identifying the range of the branch. + // + // deadline + // The highest buffer offset the the short-range branch can reach + // directly. + // + void registerBranchDeadline(unsigned rangeIdx, BufferOffset deadline) { + if (!this->oom() && !branchDeadlines_.addDeadline(rangeIdx, deadline)) { + this->fail_oom(); + } + } + + // Un-register a short-range branch deadline. + // + // When a short-range branch has been successfully bound to its destination + // label, call this function to stop traching the branch. + // + // The (rangeIdx, deadline) pair must be previously registered. + // + void unregisterBranchDeadline(unsigned rangeIdx, BufferOffset deadline) { + if (!this->oom()) { + branchDeadlines_.removeDeadline(rangeIdx, deadline); + } + } + + private: + // Are any short-range branches about to expire? + bool hasExpirableShortRangeBranches(size_t reservedBytes) const { + if (branchDeadlines_.empty()) { + return false; + } + + // Include branches that would expire in the next N bytes. The reservedBytes + // argument avoids the needless creation of many tiny constant pools. + // + // As the reservedBytes could be of any sizes such as SIZE_MAX, in the case + // of flushPool, we have to check for overflow when comparing the deadline + // with our expected reserved bytes. + size_t deadline = branchDeadlines_.earliestDeadline().getOffset(); + using CheckedSize = mozilla::CheckedInt<size_t>; + CheckedSize current(this->nextOffset().getOffset()); + CheckedSize poolFreeSpace(reservedBytes); + auto future = current + poolFreeSpace; + return !future.isValid() || deadline < future.value(); + } + + bool isPoolEmptyFor(size_t bytes) const { + return pool_.numEntries() == 0 && !hasExpirableShortRangeBranches(bytes); + } + void finishPool(size_t reservedBytes) { + JitSpew(JitSpew_Pools, + "[%d] Attempting to finish pool %zu with %u entries.", id, + poolInfo_.length(), pool_.numEntries()); + + if (reservedBytes < ShortRangeBranchHysteresis) { + reservedBytes = ShortRangeBranchHysteresis; + } + + if (isPoolEmptyFor(reservedBytes)) { + // If there is no data in the pool being dumped, don't dump anything. + JitSpew(JitSpew_Pools, "[%d] Aborting because the pool is empty", id); + return; + } + + // Should not be placing a pool in a no-pool region, check. + MOZ_ASSERT(!canNotPlacePool_); + + // Dump the pool with a guard branch around the pool. + BufferOffset guard = this->putBytes(guardSize_ * InstSize, nullptr); + BufferOffset header = this->putBytes(headerSize_ * InstSize, nullptr); + BufferOffset data = this->putBytesLarge(pool_.getPoolSize(), + (const uint8_t*)pool_.poolData()); + if (this->oom()) { + return; + } + + // Now generate branch veneers for any short-range branches that are + // about to expire. + while (hasExpirableShortRangeBranches(reservedBytes)) { + unsigned rangeIdx = branchDeadlines_.earliestDeadlineRange(); + BufferOffset deadline = branchDeadlines_.earliestDeadline(); + + // Stop tracking this branch. The Asm callback below may register + // new branches to track. + branchDeadlines_.removeDeadline(rangeIdx, deadline); + + // Make room for the veneer. Same as a pool guard branch. + BufferOffset veneer = this->putBytes(guardSize_ * InstSize, nullptr); + if (this->oom()) { + return; + } + + // Fix the branch so it targets the veneer. + // The Asm class knows how to find the original branch given the + // (rangeIdx, deadline) pair. + Asm::PatchShortRangeBranchToVeneer(this, rangeIdx, deadline, veneer); + } + + // We only reserved space for the guard branch and pool header. + // Fill them in. + BufferOffset afterPool = this->nextOffset(); + Asm::WritePoolGuard(guard, this->getInst(guard), afterPool); + Asm::WritePoolHeader((uint8_t*)this->getInst(header), &pool_, false); + + // With the pool's final position determined it is now possible to patch + // the instructions that reference entries in this pool, and this is + // done incrementally as each pool is finished. + size_t poolOffset = data.getOffset(); + + unsigned idx = 0; + for (BufferOffset* iter = pool_.loadOffsets.begin(); + iter != pool_.loadOffsets.end(); ++iter, ++idx) { + // All entries should be before the pool. + MOZ_ASSERT(iter->getOffset() < guard.getOffset()); + + // Everything here is known so we can safely do the necessary + // substitutions. + Inst* inst = this->getInst(*iter); + size_t codeOffset = poolOffset - iter->getOffset(); + + // That is, PatchConstantPoolLoad wants to be handed the address of + // the pool entry that is being loaded. We need to do a non-trivial + // amount of math here, since the pool that we've made does not + // actually reside there in memory. + JitSpew(JitSpew_Pools, "[%d] Fixing entry %d offset to %zu", id, idx, + codeOffset); + Asm::PatchConstantPoolLoad(inst, (uint8_t*)inst + codeOffset); + } + + // Record the pool info. + unsigned firstEntry = poolEntryCount - pool_.numEntries(); + if (!poolInfo_.append(PoolInfo(firstEntry, data))) { + this->fail_oom(); + return; + } + + // Reset everything to the state that it was in when we started. + pool_.reset(); + } + + public: + void flushPool() { + if (this->oom()) { + return; + } + JitSpew(JitSpew_Pools, "[%d] Requesting a pool flush", id); + finishPool(SIZE_MAX); + } + + void enterNoPool(size_t maxInst) { + if (this->oom()) { + return; + } + // Don't allow re-entry. + MOZ_ASSERT(!canNotPlacePool_); + insertNopFill(); + + // Check if the pool will spill by adding maxInst instructions, and if + // so then finish the pool before entering the no-pool region. It is + // assumed that no pool entries are allocated in a no-pool region and + // this is asserted when allocating entries. + if (!hasSpaceForInsts(maxInst, 0)) { + JitSpew(JitSpew_Pools, "[%d] No-Pool instruction(%zu) caused a spill.", + id, sizeExcludingCurrentPool()); + finishPool(maxInst * InstSize); + MOZ_ASSERT(hasSpaceForInsts(maxInst, 0)); + } + +#ifdef DEBUG + // Record the buffer position to allow validating maxInst when leaving + // the region. + canNotPlacePoolStartOffset_ = this->nextOffset().getOffset(); + canNotPlacePoolMaxInst_ = maxInst; +#endif + + canNotPlacePool_ = true; + } + + void leaveNoPool() { + if (this->oom()) { + canNotPlacePool_ = false; + return; + } + MOZ_ASSERT(canNotPlacePool_); + canNotPlacePool_ = false; + + // Validate the maxInst argument supplied to enterNoPool(). + MOZ_ASSERT(this->nextOffset().getOffset() - canNotPlacePoolStartOffset_ <= + canNotPlacePoolMaxInst_ * InstSize); + } + + void enterNoNops() { + MOZ_ASSERT(!inhibitNops_); + inhibitNops_ = true; + } + void leaveNoNops() { + MOZ_ASSERT(inhibitNops_); + inhibitNops_ = false; + } + void assertNoPoolAndNoNops() { + MOZ_ASSERT(inhibitNops_); + MOZ_ASSERT_IF(!this->oom(), isPoolEmptyFor(InstSize) || canNotPlacePool_); + } + + void align(unsigned alignment) { align(alignment, alignFillInst_); } + + void align(unsigned alignment, uint32_t pattern) { + MOZ_ASSERT(mozilla::IsPowerOfTwo(alignment)); + MOZ_ASSERT(alignment >= InstSize); + + // A pool many need to be dumped at this point, so insert NOP fill here. + insertNopFill(); + + // Check if the code position can be aligned without dumping a pool. + unsigned requiredFill = sizeExcludingCurrentPool() & (alignment - 1); + if (requiredFill == 0) { + return; + } + requiredFill = alignment - requiredFill; + + // Add an InstSize because it is probably not useful for a pool to be + // dumped at the aligned code position. + if (!hasSpaceForInsts(requiredFill / InstSize + 1, 0)) { + // Alignment would cause a pool dump, so dump the pool now. + JitSpew(JitSpew_Pools, "[%d] Alignment of %d at %zu caused a spill.", id, + alignment, sizeExcludingCurrentPool()); + finishPool(requiredFill); + } + + bool prevInhibitNops = inhibitNops_; + inhibitNops_ = true; + while ((sizeExcludingCurrentPool() & (alignment - 1)) && !this->oom()) { + putInt(pattern); + } + inhibitNops_ = prevInhibitNops; + } + + public: + void executableCopy(uint8_t* dest) { + if (this->oom()) { + return; + } + // The pools should have all been flushed, check. + MOZ_ASSERT(pool_.numEntries() == 0); + for (Slice* cur = getHead(); cur != nullptr; cur = cur->getNext()) { + memcpy(dest, &cur->instructions[0], cur->length()); + dest += cur->length(); + } + } + + bool appendRawCode(const uint8_t* code, size_t numBytes) { + if (this->oom()) { + return false; + } + // The pools should have all been flushed, check. + MOZ_ASSERT(pool_.numEntries() == 0); + while (numBytes > SliceSize) { + this->putBytes(SliceSize, code); + numBytes -= SliceSize; + code += SliceSize; + } + this->putBytes(numBytes, code); + return !this->oom(); + } + + public: + size_t poolEntryOffset(PoolEntry pe) const { + MOZ_ASSERT(pe.index() < poolEntryCount - pool_.numEntries(), + "Invalid pool entry, or not flushed yet."); + // Find the pool containing pe.index(). + // The array is sorted, so we can use a binary search. + auto b = poolInfo_.begin(), e = poolInfo_.end(); + // A note on asymmetric types in the upper_bound comparator: + // http://permalink.gmane.org/gmane.comp.compilers.clang.devel/10101 + auto i = std::upper_bound(b, e, pe.index(), + [](size_t value, const PoolInfo& entry) { + return value < entry.firstEntryIndex; + }); + // Since upper_bound finds the first pool greater than pe, + // we want the previous one which is the last one less than or equal. + MOZ_ASSERT(i != b, "PoolInfo not sorted or empty?"); + --i; + // The i iterator now points to the pool containing pe.index. + MOZ_ASSERT(i->firstEntryIndex <= pe.index() && + (i + 1 == e || (i + 1)->firstEntryIndex > pe.index())); + // Compute the byte offset into the pool. + unsigned relativeIndex = pe.index() - i->firstEntryIndex; + return i->offset.getOffset() + relativeIndex * sizeof(PoolAllocUnit); + } +}; + +} // namespace jit +} // namespace js + +#endif // jit_shared_IonAssemblerBufferWithConstantPools_h |