llvm-project/llvm/lib/CodeGen/LiveDebugValues/InstrRefBasedImpl.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

3697 lines
141 KiB
C++
Raw Normal View History

//===- InstrRefBasedImpl.cpp - Tracking Debug Value MIs -------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
/// \file InstrRefBasedImpl.cpp
///
/// This is a separate implementation of LiveDebugValues, see
/// LiveDebugValues.cpp and VarLocBasedImpl.cpp for more information.
///
/// This pass propagates variable locations between basic blocks, resolving
/// control flow conflicts between them. The problem is SSA construction, where
/// each debug instruction assigns the *value* that a variable has, and every
/// instruction where the variable is in scope uses that variable. The resulting
/// map of instruction-to-value is then translated into a register (or spill)
/// location for each variable over each instruction.
///
/// The primary difference from normal SSA construction is that we cannot
/// _create_ PHI values that contain variable values. CodeGen has already
/// completed, and we can't alter it just to make debug-info complete. Thus:
/// we can identify function positions where we would like a PHI value for a
/// variable, but must search the MachineFunction to see whether such a PHI is
/// available. If no such PHI exists, the variable location must be dropped.
///
/// To achieve this, we perform two kinds of analysis. First, we identify
/// every value defined by every instruction (ignoring those that only move
/// another value), then re-compute an SSA-form representation of the
/// MachineFunction, using value propagation to eliminate any un-necessary
/// PHI values. This gives us a map of every value computed in the function,
/// and its location within the register file / stack.
///
/// Secondly, for each variable we perform the same analysis, where each debug
/// instruction is considered a def, and every instruction where the variable
/// is in lexical scope as a use. Value propagation is used again to eliminate
/// any un-necessary PHIs. This gives us a map of each variable to the value
/// it should have in a block.
///
/// Once both are complete, we have two maps for each block:
/// * Variables to the values they should have,
/// * Values to the register / spill slot they are located in.
/// After which we can marry-up variable values with a location, and emit
/// DBG_VALUE instructions specifying those locations. Variable locations may
/// be dropped in this process due to the desired variable value not being
/// resident in any machine location, or because there is no PHI value in any
/// location that accurately represents the desired value. The building of
/// location lists for each block is left to DbgEntityHistoryCalculator.
///
/// This pass is kept efficient because the size of the first SSA problem
/// is proportional to the working-set size of the function, which the compiler
/// tries to keep small. (It's also proportional to the number of blocks).
/// Additionally, we repeatedly perform the second SSA problem analysis with
/// only the variables and blocks in a single lexical scope, exploiting their
/// locality.
///
/// ### Terminology
///
/// A machine location is a register or spill slot, a value is something that's
/// defined by an instruction or PHI node, while a variable value is the value
/// assigned to a variable. A variable location is a machine location, that must
/// contain the appropriate variable value. A value that is a PHI node is
/// occasionally called an mphi.
///
/// The first SSA problem is the "machine value location" problem,
/// because we're determining which machine locations contain which values.
/// The "locations" are constant: what's unknown is what value they contain.
///
/// The second SSA problem (the one for variables) is the "variable value
/// problem", because it's determining what values a variable has, rather than
/// what location those values are placed in.
///
/// TODO:
/// Overlapping fragments
/// Entry values
/// Add back DEBUG statements for debugging this
/// Collect statistics
///
//===----------------------------------------------------------------------===//
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/IteratedDominanceFrontier.h"
#include "llvm/CodeGen/LexicalScopes.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineInstrBundle.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/CodeGen/TargetFrameLowering.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/IR/DIBuilder.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Module.h"
#include "llvm/InitializePasses.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/TypeSize.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/Utils/SSAUpdaterImpl.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <functional>
#include <limits.h>
#include <limits>
#include <queue>
#include <tuple>
#include <utility>
#include <vector>
#include "InstrRefBasedImpl.h"
#include "LiveDebugValues.h"
using namespace llvm;
using namespace LiveDebugValues;
// SSAUpdaterImple sets DEBUG_TYPE, change it.
#undef DEBUG_TYPE
#define DEBUG_TYPE "livedebugvalues"
// Act more like the VarLoc implementation, by propagating some locations too
// far and ignoring some transfers.
static cl::opt<bool> EmulateOldLDV("emulate-old-livedebugvalues", cl::Hidden,
cl::desc("Act like old LiveDebugValues did"),
cl::init(false));
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
// Limit for the maximum number of stack slots we should track, past which we
// will ignore any spills. InstrRefBasedLDV gathers detailed information on all
// stack slots which leads to high memory consumption, and in some scenarios
// (such as asan with very many locals) the working set of the function can be
// very large, causing many spills. In these scenarios, it is very unlikely that
// the developer has hundreds of variables live at the same time that they're
// carefully thinking about -- instead, they probably autogenerated the code.
// When this happens, gracefully stop tracking excess spill slots, rather than
// consuming all the developer's memory.
static cl::opt<unsigned>
StackWorkingSetLimit("livedebugvalues-max-stack-slots", cl::Hidden,
cl::desc("livedebugvalues-stack-ws-limit"),
cl::init(250));
/// Tracker for converting machine value locations and variable values into
/// variable locations (the output of LiveDebugValues), recorded as DBG_VALUEs
/// specifying block live-in locations and transfers within blocks.
///
/// Operating on a per-block basis, this class takes a (pre-loaded) MLocTracker
/// and must be initialized with the set of variable values that are live-in to
/// the block. The caller then repeatedly calls process(). TransferTracker picks
/// out variable locations for the live-in variable values (if there _is_ a
/// location) and creates the corresponding DBG_VALUEs. Then, as the block is
/// stepped through, transfers of values between machine locations are
/// identified and if profitable, a DBG_VALUE created.
///
/// This is where debug use-before-defs would be resolved: a variable with an
/// unavailable value could materialize in the middle of a block, when the
/// value becomes available. Or, we could detect clobbers and re-specify the
/// variable in a backup location. (XXX these are unimplemented).
class TransferTracker {
public:
const TargetInstrInfo *TII;
const TargetLowering *TLI;
/// This machine location tracker is assumed to always contain the up-to-date
/// value mapping for all machine locations. TransferTracker only reads
/// information from it. (XXX make it const?)
MLocTracker *MTracker;
MachineFunction &MF;
bool ShouldEmitDebugEntryValues;
/// Record of all changes in variable locations at a block position. Awkwardly
/// we allow inserting either before or after the point: MBB != nullptr
/// indicates it's before, otherwise after.
struct Transfer {
MachineBasicBlock::instr_iterator Pos; /// Position to insert DBG_VALUes
MachineBasicBlock *MBB; /// non-null if we should insert after.
SmallVector<MachineInstr *, 4> Insts; /// Vector of DBG_VALUEs to insert.
};
2021-06-19 11:36:44 -07:00
struct LocAndProperties {
LocIdx Loc;
DbgValueProperties Properties;
2021-06-19 11:36:44 -07:00
};
/// Collection of transfers (DBG_VALUEs) to be inserted.
SmallVector<Transfer, 32> Transfers;
/// Local cache of what-value-is-in-what-LocIdx. Used to identify differences
/// between TransferTrackers view of variable locations and MLocTrackers. For
/// example, MLocTracker observes all clobbers, but TransferTracker lazily
/// does not.
SmallVector<ValueIDNum, 32> VarLocs;
/// Map from LocIdxes to which DebugVariables are based that location.
/// Mantained while stepping through the block. Not accurate if
/// VarLocs[Idx] != MTracker->LocIdxToIDNum[Idx].
DenseMap<LocIdx, SmallSet<DebugVariable, 4>> ActiveMLocs;
/// Map from DebugVariable to it's current location and qualifying meta
/// information. To be used in conjunction with ActiveMLocs to construct
/// enough information for the DBG_VALUEs for a particular LocIdx.
DenseMap<DebugVariable, LocAndProperties> ActiveVLocs;
/// Temporary cache of DBG_VALUEs to be entered into the Transfers collection.
SmallVector<MachineInstr *, 4> PendingDbgValues;
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
/// Record of a use-before-def: created when a value that's live-in to the
/// current block isn't available in any machine location, but it will be
/// defined in this block.
struct UseBeforeDef {
/// Value of this variable, def'd in block.
ValueIDNum ID;
/// Identity of this variable.
DebugVariable Var;
/// Additional variable properties.
DbgValueProperties Properties;
};
/// Map from instruction index (within the block) to the set of UseBeforeDefs
/// that become defined at that instruction.
DenseMap<unsigned, SmallVector<UseBeforeDef, 1>> UseBeforeDefs;
/// The set of variables that are in UseBeforeDefs and can become a location
/// once the relevant value is defined. An element being erased from this
/// collection prevents the use-before-def materializing.
DenseSet<DebugVariable> UseBeforeDefVariables;
const TargetRegisterInfo &TRI;
const BitVector &CalleeSavedRegs;
TransferTracker(const TargetInstrInfo *TII, MLocTracker *MTracker,
MachineFunction &MF, const TargetRegisterInfo &TRI,
const BitVector &CalleeSavedRegs, const TargetPassConfig &TPC)
: TII(TII), MTracker(MTracker), MF(MF), TRI(TRI),
CalleeSavedRegs(CalleeSavedRegs) {
TLI = MF.getSubtarget().getTargetLowering();
auto &TM = TPC.getTM<TargetMachine>();
ShouldEmitDebugEntryValues = TM.Options.ShouldEmitDebugEntryValues();
}
/// Load object with live-in variable values. \p mlocs contains the live-in
/// values in each machine location, while \p vlocs the live-in variable
/// values. This method picks variable locations for the live-in variables,
/// creates DBG_VALUEs and puts them in #Transfers, then prepares the other
/// object fields to track variable locations as we step through the block.
/// FIXME: could just examine mloctracker instead of passing in \p mlocs?
void
loadInlocs(MachineBasicBlock &MBB, ValueTable &MLocs,
const SmallVectorImpl<std::pair<DebugVariable, DbgValue>> &VLocs,
unsigned NumLocs) {
ActiveMLocs.clear();
ActiveVLocs.clear();
VarLocs.clear();
VarLocs.reserve(NumLocs);
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
UseBeforeDefs.clear();
UseBeforeDefVariables.clear();
auto isCalleeSaved = [&](LocIdx L) {
unsigned Reg = MTracker->LocIdxToLocID[L];
if (Reg >= MTracker->NumRegs)
return false;
for (MCRegAliasIterator RAI(Reg, &TRI, true); RAI.isValid(); ++RAI)
if (CalleeSavedRegs.test(*RAI))
return true;
return false;
};
// Map of the preferred location for each value.
DenseMap<ValueIDNum, LocIdx> ValueToLoc;
// Initialized the preferred-location map with illegal locations, to be
// filled in later.
for (auto &VLoc : VLocs)
if (VLoc.second.Kind == DbgValue::Def)
ValueToLoc.insert({VLoc.second.ID, LocIdx::MakeIllegalLoc()});
ActiveMLocs.reserve(VLocs.size());
ActiveVLocs.reserve(VLocs.size());
// Produce a map of value numbers to the current machine locs they live
// in. When emulating VarLocBasedImpl, there should only be one
// location; when not, we get to pick.
for (auto Location : MTracker->locations()) {
LocIdx Idx = Location.Idx;
ValueIDNum &VNum = MLocs[Idx.asU64()];
VarLocs.push_back(VNum);
// Is there a variable that wants a location for this value? If not, skip.
auto VIt = ValueToLoc.find(VNum);
if (VIt == ValueToLoc.end())
continue;
LocIdx CurLoc = VIt->second;
// In order of preference, pick:
// * Callee saved registers,
// * Other registers,
// * Spill slots.
if (CurLoc.isIllegal() || MTracker->isSpill(CurLoc) ||
(!isCalleeSaved(CurLoc) && isCalleeSaved(Idx.asU64()))) {
// Insert, or overwrite if insertion failed.
VIt->second = Idx;
}
}
// Now map variables to their picked LocIdxes.
for (const auto &Var : VLocs) {
if (Var.second.Kind == DbgValue::Const) {
PendingDbgValues.push_back(
emitMOLoc(*Var.second.MO, Var.first, Var.second.Properties));
continue;
}
// If the value has no location, we can't make a variable location.
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
const ValueIDNum &Num = Var.second.ID;
auto ValuesPreferredLoc = ValueToLoc.find(Num);
if (ValuesPreferredLoc->second.isIllegal()) {
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
// If it's a def that occurs in this block, register it as a
// use-before-def to be resolved as we step through the block.
if (Num.getBlock() == (unsigned)MBB.getNumber() && !Num.isPHI())
addUseBeforeDef(Var.first, Var.second.Properties, Num);
else
recoverAsEntryValue(Var.first, Var.second.Properties, Num);
continue;
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
}
LocIdx M = ValuesPreferredLoc->second;
auto NewValue = LocAndProperties{M, Var.second.Properties};
auto Result = ActiveVLocs.insert(std::make_pair(Var.first, NewValue));
if (!Result.second)
Result.first->second = NewValue;
ActiveMLocs[M].insert(Var.first);
PendingDbgValues.push_back(
MTracker->emitLoc(M, Var.first, Var.second.Properties));
}
flushDbgValues(MBB.begin(), &MBB);
}
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
/// Record that \p Var has value \p ID, a value that becomes available
/// later in the function.
void addUseBeforeDef(const DebugVariable &Var,
const DbgValueProperties &Properties, ValueIDNum ID) {
UseBeforeDef UBD = {ID, Var, Properties};
UseBeforeDefs[ID.getInst()].push_back(UBD);
UseBeforeDefVariables.insert(Var);
}
/// After the instruction at index \p Inst and position \p pos has been
/// processed, check whether it defines a variable value in a use-before-def.
/// If so, and the variable value hasn't changed since the start of the
/// block, create a DBG_VALUE.
void checkInstForNewValues(unsigned Inst, MachineBasicBlock::iterator pos) {
auto MIt = UseBeforeDefs.find(Inst);
if (MIt == UseBeforeDefs.end())
return;
for (auto &Use : MIt->second) {
LocIdx L = Use.ID.getLoc();
// If something goes very wrong, we might end up labelling a COPY
// instruction or similar with an instruction number, where it doesn't
// actually define a new value, instead it moves a value. In case this
// happens, discard.
if (MTracker->readMLoc(L) != Use.ID)
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
continue;
// If a different debug instruction defined the variable value / location
// since the start of the block, don't materialize this use-before-def.
if (!UseBeforeDefVariables.count(Use.Var))
continue;
PendingDbgValues.push_back(MTracker->emitLoc(L, Use.Var, Use.Properties));
}
flushDbgValues(pos, nullptr);
}
/// Helper to move created DBG_VALUEs into Transfers collection.
void flushDbgValues(MachineBasicBlock::iterator Pos, MachineBasicBlock *MBB) {
if (PendingDbgValues.size() == 0)
return;
// Pick out the instruction start position.
MachineBasicBlock::instr_iterator BundleStart;
if (MBB && Pos == MBB->begin())
BundleStart = MBB->instr_begin();
else
BundleStart = getBundleStart(Pos->getIterator());
Transfers.push_back({BundleStart, MBB, PendingDbgValues});
PendingDbgValues.clear();
}
bool isEntryValueVariable(const DebugVariable &Var,
const DIExpression *Expr) const {
if (!Var.getVariable()->isParameter())
return false;
if (Var.getInlinedAt())
return false;
if (Expr->getNumElements() > 0)
return false;
return true;
}
bool isEntryValueValue(const ValueIDNum &Val) const {
// Must be in entry block (block number zero), and be a PHI / live-in value.
if (Val.getBlock() || !Val.isPHI())
return false;
// Entry values must enter in a register.
if (MTracker->isSpill(Val.getLoc()))
return false;
Register SP = TLI->getStackPointerRegisterToSaveRestore();
Register FP = TRI.getFrameRegister(MF);
Register Reg = MTracker->LocIdxToLocID[Val.getLoc()];
return Reg != SP && Reg != FP;
}
bool recoverAsEntryValue(const DebugVariable &Var,
const DbgValueProperties &Prop,
const ValueIDNum &Num) {
// Is this variable location a candidate to be an entry value. First,
// should we be trying this at all?
if (!ShouldEmitDebugEntryValues)
return false;
// Is the variable appropriate for entry values (i.e., is a parameter).
if (!isEntryValueVariable(Var, Prop.DIExpr))
return false;
// Is the value assigned to this variable still the entry value?
if (!isEntryValueValue(Num))
return false;
// Emit a variable location using an entry value expression.
DIExpression *NewExpr =
DIExpression::prepend(Prop.DIExpr, DIExpression::EntryValue);
Register Reg = MTracker->LocIdxToLocID[Num.getLoc()];
MachineOperand MO = MachineOperand::CreateReg(Reg, false);
PendingDbgValues.push_back(emitMOLoc(MO, Var, {NewExpr, Prop.Indirect}));
return true;
}
/// Change a variable value after encountering a DBG_VALUE inside a block.
void redefVar(const MachineInstr &MI) {
DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
MI.getDebugLoc()->getInlinedAt());
DbgValueProperties Properties(MI);
const MachineOperand &MO = MI.getOperand(0);
// Ignore non-register locations, we don't transfer those.
if (!MO.isReg() || MO.getReg() == 0) {
auto It = ActiveVLocs.find(Var);
if (It != ActiveVLocs.end()) {
ActiveMLocs[It->second.Loc].erase(Var);
ActiveVLocs.erase(It);
}
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
// Any use-before-defs no longer apply.
UseBeforeDefVariables.erase(Var);
return;
}
Register Reg = MO.getReg();
LocIdx NewLoc = MTracker->getRegMLoc(Reg);
redefVar(MI, Properties, NewLoc);
}
/// Handle a change in variable location within a block. Terminate the
/// variables current location, and record the value it now refers to, so
/// that we can detect location transfers later on.
void redefVar(const MachineInstr &MI, const DbgValueProperties &Properties,
Optional<LocIdx> OptNewLoc) {
DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
MI.getDebugLoc()->getInlinedAt());
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
// Any use-before-defs no longer apply.
UseBeforeDefVariables.erase(Var);
// Erase any previous location,
auto It = ActiveVLocs.find(Var);
if (It != ActiveVLocs.end())
ActiveMLocs[It->second.Loc].erase(Var);
// If there _is_ no new location, all we had to do was erase.
if (!OptNewLoc)
return;
LocIdx NewLoc = *OptNewLoc;
// Check whether our local copy of values-by-location in #VarLocs is out of
// date. Wipe old tracking data for the location if it's been clobbered in
// the meantime.
if (MTracker->readMLoc(NewLoc) != VarLocs[NewLoc.asU64()]) {
for (auto &P : ActiveMLocs[NewLoc]) {
ActiveVLocs.erase(P);
}
ActiveMLocs[NewLoc.asU64()].clear();
VarLocs[NewLoc.asU64()] = MTracker->readMLoc(NewLoc);
}
ActiveMLocs[NewLoc].insert(Var);
if (It == ActiveVLocs.end()) {
ActiveVLocs.insert(
std::make_pair(Var, LocAndProperties{NewLoc, Properties}));
} else {
It->second.Loc = NewLoc;
It->second.Properties = Properties;
}
}
/// Account for a location \p mloc being clobbered. Examine the variable
/// locations that will be terminated: and try to recover them by using
/// another location. Optionally, given \p MakeUndef, emit a DBG_VALUE to
/// explicitly terminate a location if it can't be recovered.
void clobberMloc(LocIdx MLoc, MachineBasicBlock::iterator Pos,
bool MakeUndef = true) {
auto ActiveMLocIt = ActiveMLocs.find(MLoc);
if (ActiveMLocIt == ActiveMLocs.end())
return;
// What was the old variable value?
ValueIDNum OldValue = VarLocs[MLoc.asU64()];
VarLocs[MLoc.asU64()] = ValueIDNum::EmptyValue;
// Examine the remaining variable locations: if we can find the same value
// again, we can recover the location.
Optional<LocIdx> NewLoc = None;
for (auto Loc : MTracker->locations())
if (Loc.Value == OldValue)
NewLoc = Loc.Idx;
// If there is no location, and we weren't asked to make the variable
// explicitly undef, then stop here.
if (!NewLoc && !MakeUndef) {
// Try and recover a few more locations with entry values.
for (auto &Var : ActiveMLocIt->second) {
auto &Prop = ActiveVLocs.find(Var)->second.Properties;
recoverAsEntryValue(Var, Prop, OldValue);
}
flushDbgValues(Pos, nullptr);
return;
}
// Examine all the variables based on this location.
DenseSet<DebugVariable> NewMLocs;
for (auto &Var : ActiveMLocIt->second) {
auto ActiveVLocIt = ActiveVLocs.find(Var);
// Re-state the variable location: if there's no replacement then NewLoc
// is None and a $noreg DBG_VALUE will be created. Otherwise, a DBG_VALUE
// identifying the alternative location will be emitted.
const DbgValueProperties &Properties = ActiveVLocIt->second.Properties;
PendingDbgValues.push_back(MTracker->emitLoc(NewLoc, Var, Properties));
// Update machine locations <=> variable locations maps. Defer updating
// ActiveMLocs to avoid invalidaing the ActiveMLocIt iterator.
if (!NewLoc) {
ActiveVLocs.erase(ActiveVLocIt);
} else {
ActiveVLocIt->second.Loc = *NewLoc;
NewMLocs.insert(Var);
}
}
// Commit any deferred ActiveMLoc changes.
if (!NewMLocs.empty())
for (auto &Var : NewMLocs)
ActiveMLocs[*NewLoc].insert(Var);
// We lazily track what locations have which values; if we've found a new
// location for the clobbered value, remember it.
if (NewLoc)
VarLocs[NewLoc->asU64()] = OldValue;
flushDbgValues(Pos, nullptr);
// Re-find ActiveMLocIt, iterator could have been invalidated.
ActiveMLocIt = ActiveMLocs.find(MLoc);
ActiveMLocIt->second.clear();
}
/// Transfer variables based on \p Src to be based on \p Dst. This handles
/// both register copies as well as spills and restores. Creates DBG_VALUEs
/// describing the movement.
void transferMlocs(LocIdx Src, LocIdx Dst, MachineBasicBlock::iterator Pos) {
// Does Src still contain the value num we expect? If not, it's been
// clobbered in the meantime, and our variable locations are stale.
if (VarLocs[Src.asU64()] != MTracker->readMLoc(Src))
return;
// assert(ActiveMLocs[Dst].size() == 0);
//^^^ Legitimate scenario on account of un-clobbered slot being assigned to?
// Move set of active variables from one location to another.
auto MovingVars = ActiveMLocs[Src];
ActiveMLocs[Dst] = MovingVars;
VarLocs[Dst.asU64()] = VarLocs[Src.asU64()];
// For each variable based on Src; create a location at Dst.
for (auto &Var : MovingVars) {
auto ActiveVLocIt = ActiveVLocs.find(Var);
assert(ActiveVLocIt != ActiveVLocs.end());
ActiveVLocIt->second.Loc = Dst;
MachineInstr *MI =
MTracker->emitLoc(Dst, Var, ActiveVLocIt->second.Properties);
PendingDbgValues.push_back(MI);
}
ActiveMLocs[Src].clear();
flushDbgValues(Pos, nullptr);
// XXX XXX XXX "pretend to be old LDV" means dropping all tracking data
// about the old location.
if (EmulateOldLDV)
VarLocs[Src.asU64()] = ValueIDNum::EmptyValue;
}
MachineInstrBuilder emitMOLoc(const MachineOperand &MO,
const DebugVariable &Var,
const DbgValueProperties &Properties) {
DebugLoc DL = DILocation::get(Var.getVariable()->getContext(), 0, 0,
Var.getVariable()->getScope(),
const_cast<DILocation *>(Var.getInlinedAt()));
auto MIB = BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE));
MIB.add(MO);
if (Properties.Indirect)
MIB.addImm(0);
else
MIB.addReg(0);
MIB.addMetadata(Var.getVariable());
MIB.addMetadata(Properties.DIExpr);
return MIB;
}
};
//===----------------------------------------------------------------------===//
// Implementation
//===----------------------------------------------------------------------===//
ValueIDNum ValueIDNum::EmptyValue = {UINT_MAX, UINT_MAX, UINT_MAX};
ValueIDNum ValueIDNum::TombstoneValue = {UINT_MAX, UINT_MAX, UINT_MAX - 1};
#ifndef NDEBUG
void DbgValue::dump(const MLocTracker *MTrack) const {
if (Kind == Const) {
MO->dump();
} else if (Kind == NoVal) {
dbgs() << "NoVal(" << BlockNo << ")";
} else if (Kind == VPHI) {
dbgs() << "VPHI(" << BlockNo << "," << MTrack->IDAsString(ID) << ")";
} else {
assert(Kind == Def);
dbgs() << MTrack->IDAsString(ID);
}
if (Properties.Indirect)
dbgs() << " indir";
if (Properties.DIExpr)
dbgs() << " " << *Properties.DIExpr;
}
#endif
MLocTracker::MLocTracker(MachineFunction &MF, const TargetInstrInfo &TII,
const TargetRegisterInfo &TRI,
const TargetLowering &TLI)
: MF(MF), TII(TII), TRI(TRI), TLI(TLI),
LocIdxToIDNum(ValueIDNum::EmptyValue), LocIdxToLocID(0) {
NumRegs = TRI.getNumRegs();
reset();
LocIDToLocIdx.resize(NumRegs, LocIdx::MakeIllegalLoc());
assert(NumRegs < (1u << NUM_LOC_BITS)); // Detect bit packing failure
// Always track SP. This avoids the implicit clobbering caused by regmasks
// from affectings its values. (LiveDebugValues disbelieves calls and
// regmasks that claim to clobber SP).
Register SP = TLI.getStackPointerRegisterToSaveRestore();
if (SP) {
unsigned ID = getLocID(SP);
(void)lookupOrTrackRegister(ID);
for (MCRegAliasIterator RAI(SP, &TRI, true); RAI.isValid(); ++RAI)
SPAliases.insert(*RAI);
}
// Build some common stack positions -- full registers being spilt to the
// stack.
StackSlotIdxes.insert({{8, 0}, 0});
StackSlotIdxes.insert({{16, 0}, 1});
StackSlotIdxes.insert({{32, 0}, 2});
StackSlotIdxes.insert({{64, 0}, 3});
StackSlotIdxes.insert({{128, 0}, 4});
StackSlotIdxes.insert({{256, 0}, 5});
StackSlotIdxes.insert({{512, 0}, 6});
// Traverse all the subregister idxes, and ensure there's an index for them.
// Duplicates are no problem: we're interested in their position in the
// stack slot, we don't want to type the slot.
for (unsigned int I = 1; I < TRI.getNumSubRegIndices(); ++I) {
unsigned Size = TRI.getSubRegIdxSize(I);
unsigned Offs = TRI.getSubRegIdxOffset(I);
unsigned Idx = StackSlotIdxes.size();
// Some subregs have -1, -2 and so forth fed into their fields, to mean
// special backend things. Ignore those.
if (Size > 60000 || Offs > 60000)
continue;
StackSlotIdxes.insert({{Size, Offs}, Idx});
}
for (auto &Idx : StackSlotIdxes)
StackIdxesToPos[Idx.second] = Idx.first;
NumSlotIdxes = StackSlotIdxes.size();
}
LocIdx MLocTracker::trackRegister(unsigned ID) {
assert(ID != 0);
LocIdx NewIdx = LocIdx(LocIdxToIDNum.size());
LocIdxToIDNum.grow(NewIdx);
LocIdxToLocID.grow(NewIdx);
// Default: it's an mphi.
ValueIDNum ValNum = {CurBB, 0, NewIdx};
// Was this reg ever touched by a regmask?
for (const auto &MaskPair : reverse(Masks)) {
if (MaskPair.first->clobbersPhysReg(ID)) {
// There was an earlier def we skipped.
ValNum = {CurBB, MaskPair.second, NewIdx};
break;
}
}
LocIdxToIDNum[NewIdx] = ValNum;
LocIdxToLocID[NewIdx] = ID;
return NewIdx;
}
void MLocTracker::writeRegMask(const MachineOperand *MO, unsigned CurBB,
unsigned InstID) {
// Def any register we track have that isn't preserved. The regmask
// terminates the liveness of a register, meaning its value can't be
// relied upon -- we represent this by giving it a new value.
for (auto Location : locations()) {
unsigned ID = LocIdxToLocID[Location.Idx];
// Don't clobber SP, even if the mask says it's clobbered.
if (ID < NumRegs && !SPAliases.count(ID) && MO->clobbersPhysReg(ID))
defReg(ID, CurBB, InstID);
}
Masks.push_back(std::make_pair(MO, InstID));
}
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
Optional<SpillLocationNo> MLocTracker::getOrTrackSpillLoc(SpillLoc L) {
SpillLocationNo SpillID(SpillLocs.idFor(L));
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
if (SpillID.id() == 0) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
// If there is no location, and we have reached the limit of how many stack
// slots to track, then don't track this one.
if (SpillLocs.size() >= StackWorkingSetLimit)
return None;
// Spill location is untracked: create record for this one, and all
// subregister slots too.
SpillID = SpillLocationNo(SpillLocs.insert(L));
for (unsigned StackIdx = 0; StackIdx < NumSlotIdxes; ++StackIdx) {
unsigned L = getSpillIDWithIdx(SpillID, StackIdx);
LocIdx Idx = LocIdx(LocIdxToIDNum.size()); // New idx
LocIdxToIDNum.grow(Idx);
LocIdxToLocID.grow(Idx);
LocIDToLocIdx.push_back(Idx);
LocIdxToLocID[Idx] = L;
// Initialize to PHI value; corresponds to the location's live-in value
// during transfer function construction.
LocIdxToIDNum[Idx] = ValueIDNum(CurBB, 0, Idx);
}
}
return SpillID;
}
std::string MLocTracker::LocIdxToName(LocIdx Idx) const {
unsigned ID = LocIdxToLocID[Idx];
if (ID >= NumRegs) {
StackSlotPos Pos = locIDToSpillIdx(ID);
ID -= NumRegs;
unsigned Slot = ID / NumSlotIdxes;
return Twine("slot ")
.concat(Twine(Slot).concat(Twine(" sz ").concat(Twine(Pos.first)
.concat(Twine(" offs ").concat(Twine(Pos.second))))))
.str();
} else {
return TRI.getRegAsmName(ID).str();
}
}
std::string MLocTracker::IDAsString(const ValueIDNum &Num) const {
std::string DefName = LocIdxToName(Num.getLoc());
return Num.asString(DefName);
}
#ifndef NDEBUG
LLVM_DUMP_METHOD void MLocTracker::dump() {
for (auto Location : locations()) {
std::string MLocName = LocIdxToName(Location.Value.getLoc());
std::string DefName = Location.Value.asString(MLocName);
dbgs() << LocIdxToName(Location.Idx) << " --> " << DefName << "\n";
}
}
LLVM_DUMP_METHOD void MLocTracker::dump_mloc_map() {
for (auto Location : locations()) {
std::string foo = LocIdxToName(Location.Idx);
dbgs() << "Idx " << Location.Idx.asU64() << " " << foo << "\n";
}
}
#endif
MachineInstrBuilder MLocTracker::emitLoc(Optional<LocIdx> MLoc,
const DebugVariable &Var,
const DbgValueProperties &Properties) {
DebugLoc DL = DILocation::get(Var.getVariable()->getContext(), 0, 0,
Var.getVariable()->getScope(),
const_cast<DILocation *>(Var.getInlinedAt()));
auto MIB = BuildMI(MF, DL, TII.get(TargetOpcode::DBG_VALUE));
const DIExpression *Expr = Properties.DIExpr;
if (!MLoc) {
// No location -> DBG_VALUE $noreg
MIB.addReg(0);
MIB.addReg(0);
} else if (LocIdxToLocID[*MLoc] >= NumRegs) {
unsigned LocID = LocIdxToLocID[*MLoc];
SpillLocationNo SpillID = locIDToSpill(LocID);
StackSlotPos StackIdx = locIDToSpillIdx(LocID);
unsigned short Offset = StackIdx.second;
// TODO: support variables that are located in spill slots, with non-zero
// offsets from the start of the spill slot. It would require some more
// complex DIExpression calculations. This doesn't seem to be produced by
// LLVM right now, so don't try and support it.
// Accept no-subregister slots and subregisters where the offset is zero.
// The consumer should already have type information to work out how large
// the variable is.
if (Offset == 0) {
const SpillLoc &Spill = SpillLocs[SpillID.id()];
Expr = TRI.prependOffsetExpression(Expr, DIExpression::ApplyOffset,
Spill.SpillOffset);
unsigned Base = Spill.SpillBase;
MIB.addReg(Base);
MIB.addImm(0);
// Being on the stack makes this location indirect; if it was _already_
// indirect though, we need to add extra indirection. See this test for
// a scenario where this happens:
// llvm/test/DebugInfo/X86/spill-nontrivial-param.ll
if (Properties.Indirect) {
std::vector<uint64_t> Elts = {dwarf::DW_OP_deref};
Expr = DIExpression::append(Expr, Elts);
}
} else {
// This is a stack location with a weird subregister offset: emit an undef
// DBG_VALUE instead.
MIB.addReg(0);
MIB.addReg(0);
}
} else {
// Non-empty, non-stack slot, must be a plain register.
unsigned LocID = LocIdxToLocID[*MLoc];
MIB.addReg(LocID);
if (Properties.Indirect)
MIB.addImm(0);
else
MIB.addReg(0);
}
MIB.addMetadata(Var.getVariable());
MIB.addMetadata(Expr);
return MIB;
}
/// Default construct and initialize the pass.
InstrRefBasedLDV::InstrRefBasedLDV() = default;
bool InstrRefBasedLDV::isCalleeSaved(LocIdx L) const {
unsigned Reg = MTracker->LocIdxToLocID[L];
for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
if (CalleeSavedRegs.test(*RAI))
return true;
return false;
}
//===----------------------------------------------------------------------===//
// Debug Range Extension Implementation
//===----------------------------------------------------------------------===//
#ifndef NDEBUG
// Something to restore in the future.
// void InstrRefBasedLDV::printVarLocInMBB(..)
#endif
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
Optional<SpillLocationNo>
InstrRefBasedLDV::extractSpillBaseRegAndOffset(const MachineInstr &MI) {
assert(MI.hasOneMemOperand() &&
"Spill instruction does not have exactly one memory operand?");
auto MMOI = MI.memoperands_begin();
const PseudoSourceValue *PVal = (*MMOI)->getPseudoValue();
assert(PVal->kind() == PseudoSourceValue::FixedStack &&
"Inconsistent memory operand in spill instruction");
int FI = cast<FixedStackPseudoSourceValue>(PVal)->getFrameIndex();
const MachineBasicBlock *MBB = MI.getParent();
Register Reg;
StackOffset Offset = TFI->getFrameIndexReference(*MBB->getParent(), FI, Reg);
return MTracker->getOrTrackSpillLoc({Reg, Offset});
}
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
Optional<LocIdx>
InstrRefBasedLDV::findLocationForMemOperand(const MachineInstr &MI) {
Optional<SpillLocationNo> SpillLoc = extractSpillBaseRegAndOffset(MI);
if (!SpillLoc)
return None;
// Where in the stack slot is this value defined -- i.e., what size of value
// is this? An important question, because it could be loaded into a register
// from the stack at some point. Happily the memory operand will tell us
// the size written to the stack.
auto *MemOperand = *MI.memoperands_begin();
unsigned SizeInBits = MemOperand->getSizeInBits();
// Find that position in the stack indexes we're tracking.
auto IdxIt = MTracker->StackSlotIdxes.find({SizeInBits, 0});
if (IdxIt == MTracker->StackSlotIdxes.end())
// That index is not tracked. This is suprising, and unlikely to ever
// occur, but the safe action is to indicate the variable is optimised out.
return None;
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
unsigned SpillID = MTracker->getSpillIDWithIdx(*SpillLoc, IdxIt->second);
return MTracker->getSpillMLoc(SpillID);
}
/// End all previous ranges related to @MI and start a new range from @MI
/// if it is a DBG_VALUE instr.
bool InstrRefBasedLDV::transferDebugValue(const MachineInstr &MI) {
if (!MI.isDebugValue())
return false;
const DILocalVariable *Var = MI.getDebugVariable();
const DIExpression *Expr = MI.getDebugExpression();
const DILocation *DebugLoc = MI.getDebugLoc();
const DILocation *InlinedAt = DebugLoc->getInlinedAt();
assert(Var->isValidLocationForIntrinsic(DebugLoc) &&
"Expected inlined-at fields to agree");
DebugVariable V(Var, Expr, InlinedAt);
DbgValueProperties Properties(MI);
// If there are no instructions in this lexical scope, do no location tracking
// at all, this variable shouldn't get a legitimate location range.
auto *Scope = LS.findLexicalScope(MI.getDebugLoc().get());
if (Scope == nullptr)
return true; // handled it; by doing nothing
// For now, ignore DBG_VALUE_LISTs when extending ranges. Allow it to
// contribute to locations in this block, but don't propagate further.
// Interpret it like a DBG_VALUE $noreg.
if (MI.isDebugValueList()) {
if (VTracker)
VTracker->defVar(MI, Properties, None);
if (TTracker)
TTracker->redefVar(MI, Properties, None);
return true;
}
const MachineOperand &MO = MI.getOperand(0);
// MLocTracker needs to know that this register is read, even if it's only
// read by a debug inst.
if (MO.isReg() && MO.getReg() != 0)
(void)MTracker->readReg(MO.getReg());
// If we're preparing for the second analysis (variables), the machine value
// locations are already solved, and we report this DBG_VALUE and the value
// it refers to to VLocTracker.
if (VTracker) {
if (MO.isReg()) {
// Feed defVar the new variable location, or if this is a
// DBG_VALUE $noreg, feed defVar None.
if (MO.getReg())
VTracker->defVar(MI, Properties, MTracker->readReg(MO.getReg()));
else
VTracker->defVar(MI, Properties, None);
} else if (MI.getOperand(0).isImm() || MI.getOperand(0).isFPImm() ||
MI.getOperand(0).isCImm()) {
VTracker->defVar(MI, MI.getOperand(0));
}
}
// If performing final tracking of transfers, report this variable definition
// to the TransferTracker too.
if (TTracker)
TTracker->redefVar(MI);
return true;
}
bool InstrRefBasedLDV::transferDebugInstrRef(MachineInstr &MI,
const ValueTable *MLiveOuts,
const ValueTable *MLiveIns) {
if (!MI.isDebugRef())
return false;
// Only handle this instruction when we are building the variable value
// transfer function.
if (!VTracker && !TTracker)
return false;
unsigned InstNo = MI.getOperand(0).getImm();
unsigned OpNo = MI.getOperand(1).getImm();
const DILocalVariable *Var = MI.getDebugVariable();
const DIExpression *Expr = MI.getDebugExpression();
const DILocation *DebugLoc = MI.getDebugLoc();
const DILocation *InlinedAt = DebugLoc->getInlinedAt();
assert(Var->isValidLocationForIntrinsic(DebugLoc) &&
"Expected inlined-at fields to agree");
DebugVariable V(Var, Expr, InlinedAt);
auto *Scope = LS.findLexicalScope(MI.getDebugLoc().get());
if (Scope == nullptr)
return true; // Handled by doing nothing. This variable is never in scope.
const MachineFunction &MF = *MI.getParent()->getParent();
// Various optimizations may have happened to the value during codegen,
// recorded in the value substitution table. Apply any substitutions to
// the instruction / operand number in this DBG_INSTR_REF, and collect
// any subregister extractions performed during optimization.
// Create dummy substitution with Src set, for lookup.
auto SoughtSub =
MachineFunction::DebugSubstitution({InstNo, OpNo}, {0, 0}, 0);
SmallVector<unsigned, 4> SeenSubregs;
auto LowerBoundIt = llvm::lower_bound(MF.DebugValueSubstitutions, SoughtSub);
while (LowerBoundIt != MF.DebugValueSubstitutions.end() &&
LowerBoundIt->Src == SoughtSub.Src) {
std::tie(InstNo, OpNo) = LowerBoundIt->Dest;
SoughtSub.Src = LowerBoundIt->Dest;
if (unsigned Subreg = LowerBoundIt->Subreg)
SeenSubregs.push_back(Subreg);
LowerBoundIt = llvm::lower_bound(MF.DebugValueSubstitutions, SoughtSub);
}
// Default machine value number is <None> -- if no instruction defines
// the corresponding value, it must have been optimized out.
Optional<ValueIDNum> NewID = None;
// Try to lookup the instruction number, and find the machine value number
// that it defines. It could be an instruction, or a PHI.
auto InstrIt = DebugInstrNumToInstr.find(InstNo);
auto PHIIt = std::lower_bound(DebugPHINumToValue.begin(),
DebugPHINumToValue.end(), InstNo);
if (InstrIt != DebugInstrNumToInstr.end()) {
const MachineInstr &TargetInstr = *InstrIt->second.first;
uint64_t BlockNo = TargetInstr.getParent()->getNumber();
// Pick out the designated operand. It might be a memory reference, if
// a register def was folded into a stack store.
if (OpNo == MachineFunction::DebugOperandMemNumber &&
TargetInstr.hasOneMemOperand()) {
Optional<LocIdx> L = findLocationForMemOperand(TargetInstr);
if (L)
NewID = ValueIDNum(BlockNo, InstrIt->second.second, *L);
} else if (OpNo != MachineFunction::DebugOperandMemNumber) {
// Permit the debug-info to be completely wrong: identifying a nonexistant
// operand, or one that is not a register definition, means something
// unexpected happened during optimisation. Broken debug-info, however,
// shouldn't crash the compiler -- instead leave the variable value as
// None, which will make it appear "optimised out".
if (OpNo < TargetInstr.getNumOperands()) {
const MachineOperand &MO = TargetInstr.getOperand(OpNo);
if (MO.isReg() && MO.isDef() && MO.getReg()) {
unsigned LocID = MTracker->getLocID(MO.getReg());
LocIdx L = MTracker->LocIDToLocIdx[LocID];
NewID = ValueIDNum(BlockNo, InstrIt->second.second, L);
}
}
if (!NewID) {
LLVM_DEBUG(
{ dbgs() << "Seen instruction reference to illegal operand\n"; });
}
}
// else: NewID is left as None.
} else if (PHIIt != DebugPHINumToValue.end() && PHIIt->InstrNum == InstNo) {
// It's actually a PHI value. Which value it is might not be obvious, use
// the resolver helper to find out.
NewID = resolveDbgPHIs(*MI.getParent()->getParent(), MLiveOuts, MLiveIns,
MI, InstNo);
}
// Apply any subregister extractions, in reverse. We might have seen code
// like this:
// CALL64 @foo, implicit-def $rax
// %0:gr64 = COPY $rax
// %1:gr32 = COPY %0.sub_32bit
// %2:gr16 = COPY %1.sub_16bit
// %3:gr8 = COPY %2.sub_8bit
// In which case each copy would have been recorded as a substitution with
// a subregister qualifier. Apply those qualifiers now.
if (NewID && !SeenSubregs.empty()) {
unsigned Offset = 0;
unsigned Size = 0;
// Look at each subregister that we passed through, and progressively
// narrow in, accumulating any offsets that occur. Substitutions should
// only ever be the same or narrower width than what they read from;
// iterate in reverse order so that we go from wide to small.
for (unsigned Subreg : reverse(SeenSubregs)) {
unsigned ThisSize = TRI->getSubRegIdxSize(Subreg);
unsigned ThisOffset = TRI->getSubRegIdxOffset(Subreg);
Offset += ThisOffset;
Size = (Size == 0) ? ThisSize : std::min(Size, ThisSize);
}
// If that worked, look for an appropriate subregister with the register
// where the define happens. Don't look at values that were defined during
// a stack write: we can't currently express register locations within
// spills.
LocIdx L = NewID->getLoc();
if (NewID && !MTracker->isSpill(L)) {
// Find the register class for the register where this def happened.
// FIXME: no index for this?
Register Reg = MTracker->LocIdxToLocID[L];
const TargetRegisterClass *TRC = nullptr;
for (auto *TRCI : TRI->regclasses())
if (TRCI->contains(Reg))
TRC = TRCI;
assert(TRC && "Couldn't find target register class?");
// If the register we have isn't the right size or in the right place,
// Try to find a subregister inside it.
unsigned MainRegSize = TRI->getRegSizeInBits(*TRC);
if (Size != MainRegSize || Offset) {
// Enumerate all subregisters, searching.
Register NewReg = 0;
for (MCSubRegIterator SRI(Reg, TRI, false); SRI.isValid(); ++SRI) {
unsigned Subreg = TRI->getSubRegIndex(Reg, *SRI);
unsigned SubregSize = TRI->getSubRegIdxSize(Subreg);
unsigned SubregOffset = TRI->getSubRegIdxOffset(Subreg);
if (SubregSize == Size && SubregOffset == Offset) {
NewReg = *SRI;
break;
}
}
// If we didn't find anything: there's no way to express our value.
if (!NewReg) {
NewID = None;
} else {
// Re-state the value as being defined within the subregister
// that we found.
LocIdx NewLoc = MTracker->lookupOrTrackRegister(NewReg);
NewID = ValueIDNum(NewID->getBlock(), NewID->getInst(), NewLoc);
}
}
} else {
// If we can't handle subregisters, unset the new value.
NewID = None;
}
}
// We, we have a value number or None. Tell the variable value tracker about
// it. The rest of this LiveDebugValues implementation acts exactly the same
// for DBG_INSTR_REFs as DBG_VALUEs (just, the former can refer to values that
// aren't immediately available).
DbgValueProperties Properties(Expr, false);
if (VTracker)
VTracker->defVar(MI, Properties, NewID);
// If we're on the final pass through the function, decompose this INSTR_REF
// into a plain DBG_VALUE.
if (!TTracker)
return true;
// Pick a location for the machine value number, if such a location exists.
// (This information could be stored in TransferTracker to make it faster).
Optional<LocIdx> FoundLoc = None;
for (auto Location : MTracker->locations()) {
LocIdx CurL = Location.Idx;
ValueIDNum ID = MTracker->readMLoc(CurL);
if (NewID && ID == NewID) {
// If this is the first location with that value, pick it. Otherwise,
// consider whether it's a "longer term" location.
if (!FoundLoc) {
FoundLoc = CurL;
continue;
}
if (MTracker->isSpill(CurL))
FoundLoc = CurL; // Spills are a longer term location.
else if (!MTracker->isSpill(*FoundLoc) &&
!MTracker->isSpill(CurL) &&
!isCalleeSaved(*FoundLoc) &&
isCalleeSaved(CurL))
FoundLoc = CurL; // Callee saved regs are longer term than normal.
}
}
// Tell transfer tracker that the variable value has changed.
TTracker->redefVar(MI, Properties, FoundLoc);
[DebugInstrRef] Handle DBG_INSTR_REFs use-before-defs in LiveDebugValues Deciding where to place debugging instructions when normal instructions sink between blocks is difficult -- see PR44117. Dealing with this with instruction-referencing variable locations is simple: we just tolerate DBG_INSTR_REFs referring to values that haven't been computed yet. This patch adds support into InstrRefBasedLDV to record when a variable value appears in the middle of a block, and should have a DBG_VALUE added when it appears (a debug use before def). While described simply, this relies heavily on the value-propagation algorithm in InstrRefBasedLDV. The implementation doesn't attempt to verify the location of a value unless something non-trivial occurs to merge variable values in vlocJoin. This means that a variable with a value that has no location can retain it across all control flow (including loops). It's only when another debug instruction specifies a different variable value that we have to check, and find there's no location. This property means that if a machine value is defined in a block dominated by a DBG_INSTR_REF that refers to it, all the successor blocks can automatically find a location for that value (if it's not clobbered). Thus in a sense, InstrRefBasedLDV is already supporting and implementing use-before-defs. This patch allows us to specify a variable location in the block where it's defined. When loading live-in variable locations, TransferTracker currently discards those where it can't find a location for the variable value. However, we can tell from the machine value number whether the value is defined in this block. If it is, add it to a set of use-before-def records. Then, once the relevant instruction has been processed, emit a DBG_VALUE immediately after it. Differential Revision: https://reviews.llvm.org/D85775
2020-10-23 14:53:15 +01:00
// If there was a value with no location; but the value is defined in a
// later instruction in this block, this is a block-local use-before-def.
if (!FoundLoc && NewID && NewID->getBlock() == CurBB &&
NewID->getInst() > CurInst)
TTracker->addUseBeforeDef(V, {MI.getDebugExpression(), false}, *NewID);
// Produce a DBG_VALUE representing what this DBG_INSTR_REF meant.
// This DBG_VALUE is potentially a $noreg / undefined location, if
// FoundLoc is None.
// (XXX -- could morph the DBG_INSTR_REF in the future).
MachineInstr *DbgMI = MTracker->emitLoc(FoundLoc, V, Properties);
TTracker->PendingDbgValues.push_back(DbgMI);
TTracker->flushDbgValues(MI.getIterator(), nullptr);
return true;
}
bool InstrRefBasedLDV::transferDebugPHI(MachineInstr &MI) {
if (!MI.isDebugPHI())
return false;
// Analyse these only when solving the machine value location problem.
if (VTracker || TTracker)
return true;
// First operand is the value location, either a stack slot or register.
// Second is the debug instruction number of the original PHI.
const MachineOperand &MO = MI.getOperand(0);
unsigned InstrNum = MI.getOperand(1).getImm();
auto EmitBadPHI = [this, &MI, InstrNum](void) -> bool {
// Helper lambda to do any accounting when we fail to find a location for
// a DBG_PHI. This can happen if DBG_PHIs are malformed, or refer to a
// dead stack slot, for example.
// Record a DebugPHIRecord with an empty value + location.
DebugPHINumToValue.push_back({InstrNum, MI.getParent(), None, None});
return true;
};
if (MO.isReg() && MO.getReg()) {
// The value is whatever's currently in the register. Read and record it,
// to be analysed later.
Register Reg = MO.getReg();
ValueIDNum Num = MTracker->readReg(Reg);
auto PHIRec = DebugPHIRecord(
{InstrNum, MI.getParent(), Num, MTracker->lookupOrTrackRegister(Reg)});
DebugPHINumToValue.push_back(PHIRec);
// Ensure this register is tracked.
for (MCRegAliasIterator RAI(MO.getReg(), TRI, true); RAI.isValid(); ++RAI)
MTracker->lookupOrTrackRegister(*RAI);
} else if (MO.isFI()) {
// The value is whatever's in this stack slot.
unsigned FI = MO.getIndex();
// If the stack slot is dead, then this was optimized away.
// FIXME: stack slot colouring should account for slots that get merged.
if (MFI->isDeadObjectIndex(FI))
return EmitBadPHI();
// Identify this spill slot, ensure it's tracked.
Register Base;
StackOffset Offs = TFI->getFrameIndexReference(*MI.getMF(), FI, Base);
SpillLoc SL = {Base, Offs};
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
Optional<SpillLocationNo> SpillNo = MTracker->getOrTrackSpillLoc(SL);
// We might be able to find a value, but have chosen not to, to avoid
// tracking too much stack information.
if (!SpillNo)
return EmitBadPHI();
// Problem: what value should we extract from the stack? LLVM does not
// record what size the last store to the slot was, and it would become
// sketchy after stack slot colouring anyway. Take a look at what values
// are stored on the stack, and pick the largest one that wasn't def'd
// by a spill (i.e., the value most likely to have been def'd in a register
// and then spilt.
std::array<unsigned, 4> CandidateSizes = {64, 32, 16, 8};
Optional<ValueIDNum> Result = None;
Optional<LocIdx> SpillLoc = None;
for (unsigned CS : CandidateSizes) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
unsigned SpillID = MTracker->getLocID(*SpillNo, {CS, 0});
SpillLoc = MTracker->getSpillMLoc(SpillID);
ValueIDNum Val = MTracker->readMLoc(*SpillLoc);
// If this value was defined in it's own position, then it was probably
// an aliasing index of a small value that was spilt.
if (Val.getLoc() != SpillLoc->asU64()) {
Result = Val;
break;
}
}
// If we didn't find anything, we're probably looking at a PHI, or a memory
// store folded into an instruction. FIXME: Take a guess that's it's 64
// bits. This isn't ideal, but tracking the size that the spill is
// "supposed" to be is more complex, and benefits a small number of
// locations.
if (!Result) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
unsigned SpillID = MTracker->getLocID(*SpillNo, {64, 0});
SpillLoc = MTracker->getSpillMLoc(SpillID);
Result = MTracker->readMLoc(*SpillLoc);
}
// Record this DBG_PHI for later analysis.
auto DbgPHI =
DebugPHIRecord({InstrNum, MI.getParent(), *Result, *SpillLoc});
DebugPHINumToValue.push_back(DbgPHI);
} else {
// Else: if the operand is neither a legal register or a stack slot, then
// we're being fed illegal debug-info. Record an empty PHI, so that any
// debug users trying to read this number will be put off trying to
// interpret the value.
LLVM_DEBUG(
{ dbgs() << "Seen DBG_PHI with unrecognised operand format\n"; });
return EmitBadPHI();
}
return true;
}
void InstrRefBasedLDV::transferRegisterDef(MachineInstr &MI) {
// Meta Instructions do not affect the debug liveness of any register they
// define.
if (MI.isImplicitDef()) {
// Except when there's an implicit def, and the location it's defining has
// no value number. The whole point of an implicit def is to announce that
// the register is live, without be specific about it's value. So define
// a value if there isn't one already.
ValueIDNum Num = MTracker->readReg(MI.getOperand(0).getReg());
// Has a legitimate value -> ignore the implicit def.
if (Num.getLoc() != 0)
return;
// Otherwise, def it here.
} else if (MI.isMetaInstruction())
return;
// We always ignore SP defines on call instructions, they don't actually
// change the value of the stack pointer... except for win32's _chkstk. This
// is rare: filter quickly for the common case (no stack adjustments, not a
// call, etc). If it is a call that modifies SP, recognise the SP register
// defs.
bool CallChangesSP = false;
if (AdjustsStackInCalls && MI.isCall() && MI.getOperand(0).isSymbol() &&
!strcmp(MI.getOperand(0).getSymbolName(), StackProbeSymbolName.data()))
CallChangesSP = true;
// Test whether we should ignore a def of this register due to it being part
// of the stack pointer.
auto IgnoreSPAlias = [this, &MI, CallChangesSP](Register R) -> bool {
if (CallChangesSP)
return false;
return MI.isCall() && MTracker->SPAliases.count(R);
};
// Find the regs killed by MI, and find regmasks of preserved regs.
// Max out the number of statically allocated elements in `DeadRegs`, as this
// prevents fallback to std::set::count() operations.
SmallSet<uint32_t, 32> DeadRegs;
SmallVector<const uint32_t *, 4> RegMasks;
SmallVector<const MachineOperand *, 4> RegMaskPtrs;
for (const MachineOperand &MO : MI.operands()) {
// Determine whether the operand is a register def.
if (MO.isReg() && MO.isDef() && MO.getReg() &&
Register::isPhysicalRegister(MO.getReg()) &&
!IgnoreSPAlias(MO.getReg())) {
// Remove ranges of all aliased registers.
for (MCRegAliasIterator RAI(MO.getReg(), TRI, true); RAI.isValid(); ++RAI)
// FIXME: Can we break out of this loop early if no insertion occurs?
DeadRegs.insert(*RAI);
} else if (MO.isRegMask()) {
RegMasks.push_back(MO.getRegMask());
RegMaskPtrs.push_back(&MO);
}
}
// Tell MLocTracker about all definitions, of regmasks and otherwise.
for (uint32_t DeadReg : DeadRegs)
MTracker->defReg(DeadReg, CurBB, CurInst);
for (auto *MO : RegMaskPtrs)
MTracker->writeRegMask(MO, CurBB, CurInst);
// If this instruction writes to a spill slot, def that slot.
if (hasFoldedStackStore(MI)) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
if (Optional<SpillLocationNo> SpillNo = extractSpillBaseRegAndOffset(MI)) {
for (unsigned int I = 0; I < MTracker->NumSlotIdxes; ++I) {
unsigned SpillID = MTracker->getSpillIDWithIdx(*SpillNo, I);
LocIdx L = MTracker->getSpillMLoc(SpillID);
MTracker->setMLoc(L, ValueIDNum(CurBB, CurInst, L));
}
}
}
if (!TTracker)
return;
// When committing variable values to locations: tell transfer tracker that
// we've clobbered things. It may be able to recover the variable from a
// different location.
// Inform TTracker about any direct clobbers.
for (uint32_t DeadReg : DeadRegs) {
LocIdx Loc = MTracker->lookupOrTrackRegister(DeadReg);
TTracker->clobberMloc(Loc, MI.getIterator(), false);
}
// Look for any clobbers performed by a register mask. Only test locations
// that are actually being tracked.
if (!RegMaskPtrs.empty()) {
for (auto L : MTracker->locations()) {
// Stack locations can't be clobbered by regmasks.
if (MTracker->isSpill(L.Idx))
continue;
Register Reg = MTracker->LocIdxToLocID[L.Idx];
if (IgnoreSPAlias(Reg))
continue;
for (auto *MO : RegMaskPtrs)
if (MO->clobbersPhysReg(Reg))
TTracker->clobberMloc(L.Idx, MI.getIterator(), false);
}
}
// Tell TTracker about any folded stack store.
if (hasFoldedStackStore(MI)) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
if (Optional<SpillLocationNo> SpillNo = extractSpillBaseRegAndOffset(MI)) {
for (unsigned int I = 0; I < MTracker->NumSlotIdxes; ++I) {
unsigned SpillID = MTracker->getSpillIDWithIdx(*SpillNo, I);
LocIdx L = MTracker->getSpillMLoc(SpillID);
TTracker->clobberMloc(L, MI.getIterator(), true);
}
}
}
}
void InstrRefBasedLDV::performCopy(Register SrcRegNum, Register DstRegNum) {
// In all circumstances, re-def all aliases. It's definitely a new value now.
for (MCRegAliasIterator RAI(DstRegNum, TRI, true); RAI.isValid(); ++RAI)
MTracker->defReg(*RAI, CurBB, CurInst);
ValueIDNum SrcValue = MTracker->readReg(SrcRegNum);
MTracker->setReg(DstRegNum, SrcValue);
// Copy subregisters from one location to another.
for (MCSubRegIndexIterator SRI(SrcRegNum, TRI); SRI.isValid(); ++SRI) {
unsigned SrcSubReg = SRI.getSubReg();
unsigned SubRegIdx = SRI.getSubRegIndex();
unsigned DstSubReg = TRI->getSubReg(DstRegNum, SubRegIdx);
if (!DstSubReg)
continue;
// Do copy. There are two matching subregisters, the source value should
// have been def'd when the super-reg was, the latter might not be tracked
// yet.
// This will force SrcSubReg to be tracked, if it isn't yet. Will read
// mphi values if it wasn't tracked.
LocIdx SrcL = MTracker->lookupOrTrackRegister(SrcSubReg);
LocIdx DstL = MTracker->lookupOrTrackRegister(DstSubReg);
(void)SrcL;
(void)DstL;
ValueIDNum CpyValue = MTracker->readReg(SrcSubReg);
MTracker->setReg(DstSubReg, CpyValue);
}
}
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
Optional<SpillLocationNo>
InstrRefBasedLDV::isSpillInstruction(const MachineInstr &MI,
MachineFunction *MF) {
// TODO: Handle multiple stores folded into one.
if (!MI.hasOneMemOperand())
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
return None;
// Reject any memory operand that's aliased -- we can't guarantee its value.
auto MMOI = MI.memoperands_begin();
const PseudoSourceValue *PVal = (*MMOI)->getPseudoValue();
if (PVal->isAliased(MFI))
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
return None;
if (!MI.getSpillSize(TII) && !MI.getFoldedSpillSize(TII))
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
return None; // This is not a spill instruction, since no valid size was
// returned from either function.
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
return extractSpillBaseRegAndOffset(MI);
}
bool InstrRefBasedLDV::isLocationSpill(const MachineInstr &MI,
MachineFunction *MF, unsigned &Reg) {
if (!isSpillInstruction(MI, MF))
return false;
int FI;
Reg = TII->isStoreToStackSlotPostFE(MI, FI);
return Reg != 0;
}
Optional<SpillLocationNo>
InstrRefBasedLDV::isRestoreInstruction(const MachineInstr &MI,
MachineFunction *MF, unsigned &Reg) {
if (!MI.hasOneMemOperand())
return None;
// FIXME: Handle folded restore instructions with more than one memory
// operand.
if (MI.getRestoreSize(TII)) {
Reg = MI.getOperand(0).getReg();
return extractSpillBaseRegAndOffset(MI);
}
return None;
}
bool InstrRefBasedLDV::transferSpillOrRestoreInst(MachineInstr &MI) {
// XXX -- it's too difficult to implement VarLocBasedImpl's stack location
// limitations under the new model. Therefore, when comparing them, compare
// versions that don't attempt spills or restores at all.
if (EmulateOldLDV)
return false;
// Strictly limit ourselves to plain loads and stores, not all instructions
// that can access the stack.
int DummyFI = -1;
if (!TII->isStoreToStackSlotPostFE(MI, DummyFI) &&
!TII->isLoadFromStackSlotPostFE(MI, DummyFI))
return false;
MachineFunction *MF = MI.getMF();
unsigned Reg;
LLVM_DEBUG(dbgs() << "Examining instruction: "; MI.dump(););
// Strictly limit ourselves to plain loads and stores, not all instructions
// that can access the stack.
int FIDummy;
if (!TII->isStoreToStackSlotPostFE(MI, FIDummy) &&
!TII->isLoadFromStackSlotPostFE(MI, FIDummy))
return false;
// First, if there are any DBG_VALUEs pointing at a spill slot that is
// written to, terminate that variable location. The value in memory
// will have changed. DbgEntityHistoryCalculator doesn't try to detect this.
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
if (Optional<SpillLocationNo> Loc = isSpillInstruction(MI, MF)) {
// Un-set this location and clobber, so that earlier locations don't
// continue past this store.
for (unsigned SlotIdx = 0; SlotIdx < MTracker->NumSlotIdxes; ++SlotIdx) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
unsigned SpillID = MTracker->getSpillIDWithIdx(*Loc, SlotIdx);
Optional<LocIdx> MLoc = MTracker->getSpillMLoc(SpillID);
if (!MLoc)
continue;
// We need to over-write the stack slot with something (here, a def at
// this instruction) to ensure no values are preserved in this stack slot
// after the spill. It also prevents TTracker from trying to recover the
// location and re-installing it in the same place.
ValueIDNum Def(CurBB, CurInst, *MLoc);
MTracker->setMLoc(*MLoc, Def);
if (TTracker)
TTracker->clobberMloc(*MLoc, MI.getIterator());
}
}
// Try to recognise spill and restore instructions that may transfer a value.
if (isLocationSpill(MI, MF, Reg)) {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
// isLocationSpill returning true should guarantee we can extract a
// location.
SpillLocationNo Loc = *extractSpillBaseRegAndOffset(MI);
auto DoTransfer = [&](Register SrcReg, unsigned SpillID) {
auto ReadValue = MTracker->readReg(SrcReg);
LocIdx DstLoc = MTracker->getSpillMLoc(SpillID);
MTracker->setMLoc(DstLoc, ReadValue);
if (TTracker) {
LocIdx SrcLoc = MTracker->getRegMLoc(SrcReg);
TTracker->transferMlocs(SrcLoc, DstLoc, MI.getIterator());
}
};
// Then, transfer subreg bits.
for (MCSubRegIterator SRI(Reg, TRI, false); SRI.isValid(); ++SRI) {
// Ensure this reg is tracked,
(void)MTracker->lookupOrTrackRegister(*SRI);
unsigned SubregIdx = TRI->getSubRegIndex(Reg, *SRI);
unsigned SpillID = MTracker->getLocID(Loc, SubregIdx);
DoTransfer(*SRI, SpillID);
}
// Directly lookup size of main source reg, and transfer.
unsigned Size = TRI->getRegSizeInBits(Reg, *MRI);
unsigned SpillID = MTracker->getLocID(Loc, {Size, 0});
DoTransfer(Reg, SpillID);
} else {
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
Optional<SpillLocationNo> Loc = isRestoreInstruction(MI, MF, Reg);
if (!Loc)
return false;
// Assumption: we're reading from the base of the stack slot, not some
// offset into it. It seems very unlikely LLVM would ever generate
// restores where this wasn't true. This then becomes a question of what
// subregisters in the destination register line up with positions in the
// stack slot.
// Def all registers that alias the destination.
for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
MTracker->defReg(*RAI, CurBB, CurInst);
// Now find subregisters within the destination register, and load values
// from stack slot positions.
auto DoTransfer = [&](Register DestReg, unsigned SpillID) {
LocIdx SrcIdx = MTracker->getSpillMLoc(SpillID);
auto ReadValue = MTracker->readMLoc(SrcIdx);
MTracker->setReg(DestReg, ReadValue);
if (TTracker) {
LocIdx DstLoc = MTracker->getRegMLoc(DestReg);
TTracker->transferMlocs(SrcIdx, DstLoc, MI.getIterator());
}
};
for (MCSubRegIterator SRI(Reg, TRI, false); SRI.isValid(); ++SRI) {
unsigned Subreg = TRI->getSubRegIndex(Reg, *SRI);
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
unsigned SpillID = MTracker->getLocID(*Loc, Subreg);
DoTransfer(*SRI, SpillID);
}
// Directly look up this registers slot idx by size, and transfer.
unsigned Size = TRI->getRegSizeInBits(Reg, *MRI);
Re-apply 3fab2d138e30, now with a triple added Was reverted in 1c1b670a73a9 as it broke all non-x86 bots. Original commit message: [DebugInfo][InstrRef] Add a max-stack-slots-to-track cut-out In certain circumstances with things like autogenerated code and asan, you can end up with thousands of Values live at the same time, causing a large working set and a lot of information spilled to the stack. Unfortunately InstrRefBasedLDV doesn't cope well with this and consumes a lot of memory when there are many many stack slots. See the reproducer in D116821. It seems very unlikely that a developer would be able to reason about hundreds of live named local variables at the same time, so a huge working set and many stack slots is an indicator that we're likely analysing autogenerated or instrumented code. In those cases: gracefully degrade by setting an upper bound on the amount of stack slots to track. This limits peak memory consumption, at the cost of dropping some variable locations, but in a rare scenario where it's unlikely someone is actually going to use them. In terms of the patch, this adds a cl::opt for max number of stack slots to track, and has the stack-slot-numbering code optionally return None. That then filters through a number of code paths, which can then chose to not track a spill / restore if it touches an untracked spill slot. The added test checks that we drop variable locations that are on the stack, if we set the limit to zero. Differential Revision: https://reviews.llvm.org/D118601
2022-02-01 19:19:20 +00:00
unsigned SpillID = MTracker->getLocID(*Loc, {Size, 0});
DoTransfer(Reg, SpillID);
}
return true;
}
bool InstrRefBasedLDV::transferRegisterCopy(MachineInstr &MI) {
auto DestSrc = TII->isCopyInstr(MI);
if (!DestSrc)
return false;
const MachineOperand *DestRegOp = DestSrc->Destination;
const MachineOperand *SrcRegOp = DestSrc->Source;
auto isCalleeSavedReg = [&](unsigned Reg) {
for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
if (CalleeSavedRegs.test(*RAI))
return true;
return false;
};
Register SrcReg = SrcRegOp->getReg();
Register DestReg = DestRegOp->getReg();
// Ignore identity copies. Yep, these make it as far as LiveDebugValues.
if (SrcReg == DestReg)
return true;
// For emulating VarLocBasedImpl:
// We want to recognize instructions where destination register is callee
// saved register. If register that could be clobbered by the call is
// included, there would be a great chance that it is going to be clobbered
// soon. It is more likely that previous register, which is callee saved, is
// going to stay unclobbered longer, even if it is killed.
//
// For InstrRefBasedImpl, we can track multiple locations per value, so
// ignore this condition.
if (EmulateOldLDV && !isCalleeSavedReg(DestReg))
return false;
// InstrRefBasedImpl only followed killing copies.
if (EmulateOldLDV && !SrcRegOp->isKill())
return false;
// Copy MTracker info, including subregs if available.
InstrRefBasedLDV::performCopy(SrcReg, DestReg);
// Only produce a transfer of DBG_VALUE within a block where old LDV
// would have. We might make use of the additional value tracking in some
// other way, later.
if (TTracker && isCalleeSavedReg(DestReg) && SrcRegOp->isKill())
TTracker->transferMlocs(MTracker->getRegMLoc(SrcReg),
MTracker->getRegMLoc(DestReg), MI.getIterator());
// VarLocBasedImpl would quit tracking the old location after copying.
if (EmulateOldLDV && SrcReg != DestReg)
MTracker->defReg(SrcReg, CurBB, CurInst);
// Finally, the copy might have clobbered variables based on the destination
// register. Tell TTracker about it, in case a backup location exists.
if (TTracker) {
for (MCRegAliasIterator RAI(DestReg, TRI, true); RAI.isValid(); ++RAI) {
LocIdx ClobberedLoc = MTracker->getRegMLoc(*RAI);
TTracker->clobberMloc(ClobberedLoc, MI.getIterator(), false);
}
}
return true;
}
/// Accumulate a mapping between each DILocalVariable fragment and other
/// fragments of that DILocalVariable which overlap. This reduces work during
/// the data-flow stage from "Find any overlapping fragments" to "Check if the
/// known-to-overlap fragments are present".
/// \param MI A previously unprocessed debug instruction to analyze for
/// fragment usage.
void InstrRefBasedLDV::accumulateFragmentMap(MachineInstr &MI) {
assert(MI.isDebugValue() || MI.isDebugRef());
DebugVariable MIVar(MI.getDebugVariable(), MI.getDebugExpression(),
MI.getDebugLoc()->getInlinedAt());
FragmentInfo ThisFragment = MIVar.getFragmentOrDefault();
// If this is the first sighting of this variable, then we are guaranteed
// there are currently no overlapping fragments either. Initialize the set
// of seen fragments, record no overlaps for the current one, and return.
auto SeenIt = SeenFragments.find(MIVar.getVariable());
if (SeenIt == SeenFragments.end()) {
SmallSet<FragmentInfo, 4> OneFragment;
OneFragment.insert(ThisFragment);
SeenFragments.insert({MIVar.getVariable(), OneFragment});
OverlapFragments.insert({{MIVar.getVariable(), ThisFragment}, {}});
return;
}
// If this particular Variable/Fragment pair already exists in the overlap
// map, it has already been accounted for.
auto IsInOLapMap =
OverlapFragments.insert({{MIVar.getVariable(), ThisFragment}, {}});
if (!IsInOLapMap.second)
return;
auto &ThisFragmentsOverlaps = IsInOLapMap.first->second;
auto &AllSeenFragments = SeenIt->second;
// Otherwise, examine all other seen fragments for this variable, with "this"
// fragment being a previously unseen fragment. Record any pair of
// overlapping fragments.
for (auto &ASeenFragment : AllSeenFragments) {
// Does this previously seen fragment overlap?
if (DIExpression::fragmentsOverlap(ThisFragment, ASeenFragment)) {
// Yes: Mark the current fragment as being overlapped.
ThisFragmentsOverlaps.push_back(ASeenFragment);
// Mark the previously seen fragment as being overlapped by the current
// one.
auto ASeenFragmentsOverlaps =
OverlapFragments.find({MIVar.getVariable(), ASeenFragment});
assert(ASeenFragmentsOverlaps != OverlapFragments.end() &&
"Previously seen var fragment has no vector of overlaps");
ASeenFragmentsOverlaps->second.push_back(ThisFragment);
}
}
AllSeenFragments.insert(ThisFragment);
}
void InstrRefBasedLDV::process(MachineInstr &MI, const ValueTable *MLiveOuts,
const ValueTable *MLiveIns) {
// Try to interpret an MI as a debug or transfer instruction. Only if it's
// none of these should we interpret it's register defs as new value
// definitions.
if (transferDebugValue(MI))
return;
if (transferDebugInstrRef(MI, MLiveOuts, MLiveIns))
return;
if (transferDebugPHI(MI))
return;
if (transferRegisterCopy(MI))
return;
if (transferSpillOrRestoreInst(MI))
return;
transferRegisterDef(MI);
}
void InstrRefBasedLDV::produceMLocTransferFunction(
MachineFunction &MF, SmallVectorImpl<MLocTransferMap> &MLocTransfer,
unsigned MaxNumBlocks) {
// Because we try to optimize around register mask operands by ignoring regs
// that aren't currently tracked, we set up something ugly for later: RegMask
// operands that are seen earlier than the first use of a register, still need
// to clobber that register in the transfer function. But this information
// isn't actively recorded. Instead, we track each RegMask used in each block,
// and accumulated the clobbered but untracked registers in each block into
// the following bitvector. Later, if new values are tracked, we can add
// appropriate clobbers.
SmallVector<BitVector, 32> BlockMasks;
BlockMasks.resize(MaxNumBlocks);
// Reserve one bit per register for the masks described above.
unsigned BVWords = MachineOperand::getRegMaskSize(TRI->getNumRegs());
for (auto &BV : BlockMasks)
BV.resize(TRI->getNumRegs(), true);
// Step through all instructions and inhale the transfer function.
for (auto &MBB : MF) {
// Object fields that are read by trackers to know where we are in the
// function.
CurBB = MBB.getNumber();
CurInst = 1;
// Set all machine locations to a PHI value. For transfer function
// production only, this signifies the live-in value to the block.
MTracker->reset();
MTracker->setMPhis(CurBB);
// Step through each instruction in this block.
for (auto &MI : MBB) {
// Pass in an empty unique_ptr for the value tables when accumulating the
// machine transfer function.
process(MI, nullptr, nullptr);
// Also accumulate fragment map.
if (MI.isDebugValue() || MI.isDebugRef())
accumulateFragmentMap(MI);
// Create a map from the instruction number (if present) to the
// MachineInstr and its position.
if (uint64_t InstrNo = MI.peekDebugInstrNum()) {
auto InstrAndPos = std::make_pair(&MI, CurInst);
auto InsertResult =
DebugInstrNumToInstr.insert(std::make_pair(InstrNo, InstrAndPos));
// There should never be duplicate instruction numbers.
assert(InsertResult.second);
(void)InsertResult;
}
++CurInst;
}
// Produce the transfer function, a map of machine location to new value. If
// any machine location has the live-in phi value from the start of the
// block, it's live-through and doesn't need recording in the transfer
// function.
for (auto Location : MTracker->locations()) {
LocIdx Idx = Location.Idx;
ValueIDNum &P = Location.Value;
if (P.isPHI() && P.getLoc() == Idx.asU64())
continue;
// Insert-or-update.
auto &TransferMap = MLocTransfer[CurBB];
auto Result = TransferMap.insert(std::make_pair(Idx.asU64(), P));
if (!Result.second)
Result.first->second = P;
}
// Accumulate any bitmask operands into the clobberred reg mask for this
// block.
for (auto &P : MTracker->Masks) {
BlockMasks[CurBB].clearBitsNotInMask(P.first->getRegMask(), BVWords);
}
}
// Compute a bitvector of all the registers that are tracked in this block.
BitVector UsedRegs(TRI->getNumRegs());
for (auto Location : MTracker->locations()) {
unsigned ID = MTracker->LocIdxToLocID[Location.Idx];
// Ignore stack slots, and aliases of the stack pointer.
if (ID >= TRI->getNumRegs() || MTracker->SPAliases.count(ID))
continue;
UsedRegs.set(ID);
}
// Check that any regmask-clobber of a register that gets tracked, is not
// live-through in the transfer function. It needs to be clobbered at the
// very least.
for (unsigned int I = 0; I < MaxNumBlocks; ++I) {
BitVector &BV = BlockMasks[I];
BV.flip();
BV &= UsedRegs;
// This produces all the bits that we clobber, but also use. Check that
// they're all clobbered or at least set in the designated transfer
// elem.
for (unsigned Bit : BV.set_bits()) {
unsigned ID = MTracker->getLocID(Bit);
LocIdx Idx = MTracker->LocIDToLocIdx[ID];
auto &TransferMap = MLocTransfer[I];
// Install a value representing the fact that this location is effectively
// written to in this block. As there's no reserved value, instead use
// a value number that is never generated. Pick the value number for the
// first instruction in the block, def'ing this location, which we know
// this block never used anyway.
ValueIDNum NotGeneratedNum = ValueIDNum(I, 1, Idx);
auto Result =
TransferMap.insert(std::make_pair(Idx.asU64(), NotGeneratedNum));
if (!Result.second) {
ValueIDNum &ValueID = Result.first->second;
if (ValueID.getBlock() == I && ValueID.isPHI())
// It was left as live-through. Set it to clobbered.
ValueID = NotGeneratedNum;
}
}
}
}
bool InstrRefBasedLDV::mlocJoin(
MachineBasicBlock &MBB, SmallPtrSet<const MachineBasicBlock *, 16> &Visited,
FuncValueTable &OutLocs, ValueTable &InLocs) {
LLVM_DEBUG(dbgs() << "join MBB: " << MBB.getNumber() << "\n");
bool Changed = false;
// Handle value-propagation when control flow merges on entry to a block. For
// any location without a PHI already placed, the location has the same value
// as its predecessors. If a PHI is placed, test to see whether it's now a
// redundant PHI that we can eliminate.
SmallVector<const MachineBasicBlock *, 8> BlockOrders;
for (auto Pred : MBB.predecessors())
BlockOrders.push_back(Pred);
// Visit predecessors in RPOT order.
auto Cmp = [&](const MachineBasicBlock *A, const MachineBasicBlock *B) {
return BBToOrder.find(A)->second < BBToOrder.find(B)->second;
};
llvm::sort(BlockOrders, Cmp);
// Skip entry block.
if (BlockOrders.size() == 0)
return false;
// Step through all machine locations, look at each predecessor and test
// whether we can eliminate redundant PHIs.
for (auto Location : MTracker->locations()) {
LocIdx Idx = Location.Idx;
// Pick out the first predecessors live-out value for this location. It's
// guaranteed to not be a backedge, as we order by RPO.
ValueIDNum FirstVal = OutLocs[BlockOrders[0]->getNumber()][Idx.asU64()];
// If we've already eliminated a PHI here, do no further checking, just
// propagate the first live-in value into this block.
if (InLocs[Idx.asU64()] != ValueIDNum(MBB.getNumber(), 0, Idx)) {
if (InLocs[Idx.asU64()] != FirstVal) {
InLocs[Idx.asU64()] = FirstVal;
Changed |= true;
}
continue;
}
// We're now examining a PHI to see whether it's un-necessary. Loop around
// the other live-in values and test whether they're all the same.
bool Disagree = false;
for (unsigned int I = 1; I < BlockOrders.size(); ++I) {
const MachineBasicBlock *PredMBB = BlockOrders[I];
const ValueIDNum &PredLiveOut =
OutLocs[PredMBB->getNumber()][Idx.asU64()];
// Incoming values agree, continue trying to eliminate this PHI.
if (FirstVal == PredLiveOut)
continue;
// We can also accept a PHI value that feeds back into itself.
if (PredLiveOut == ValueIDNum(MBB.getNumber(), 0, Idx))
continue;
// Live-out of a predecessor disagrees with the first predecessor.
Disagree = true;
}
// No disagreement? No PHI. Otherwise, leave the PHI in live-ins.
if (!Disagree) {
InLocs[Idx.asU64()] = FirstVal;
Changed |= true;
}
}
// TODO: Reimplement NumInserted and NumRemoved.
return Changed;
}
void InstrRefBasedLDV::findStackIndexInterference(
SmallVectorImpl<unsigned> &Slots) {
// We could spend a bit of time finding the exact, minimal, set of stack
// indexes that interfere with each other, much like reg units. Or, we can
// rely on the fact that:
// * The smallest / lowest index will interfere with everything at zero
// offset, which will be the largest set of registers,
// * Most indexes with non-zero offset will end up being interference units
// anyway.
// So just pick those out and return them.
// We can rely on a single-byte stack index existing already, because we
// initialize them in MLocTracker.
auto It = MTracker->StackSlotIdxes.find({8, 0});
assert(It != MTracker->StackSlotIdxes.end());
Slots.push_back(It->second);
// Find anything that has a non-zero offset and add that too.
for (auto &Pair : MTracker->StackSlotIdxes) {
// Is offset zero? If so, ignore.
if (!Pair.first.second)
continue;
Slots.push_back(Pair.second);
}
}
void InstrRefBasedLDV::placeMLocPHIs(
MachineFunction &MF, SmallPtrSetImpl<MachineBasicBlock *> &AllBlocks,
FuncValueTable &MInLocs, SmallVectorImpl<MLocTransferMap> &MLocTransfer) {
SmallVector<unsigned, 4> StackUnits;
findStackIndexInterference(StackUnits);
// To avoid repeatedly running the PHI placement algorithm, leverage the
// fact that a def of register MUST also def its register units. Find the
// units for registers, place PHIs for them, and then replicate them for
// aliasing registers. Some inputs that are never def'd (DBG_PHIs of
// arguments) don't lead to register units being tracked, just place PHIs for
// those registers directly. Stack slots have their own form of "unit",
// store them to one side.
SmallSet<Register, 32> RegUnitsToPHIUp;
SmallSet<LocIdx, 32> NormalLocsToPHI;
SmallSet<SpillLocationNo, 32> StackSlots;
for (auto Location : MTracker->locations()) {
LocIdx L = Location.Idx;
if (MTracker->isSpill(L)) {
StackSlots.insert(MTracker->locIDToSpill(MTracker->LocIdxToLocID[L]));
continue;
}
Register R = MTracker->LocIdxToLocID[L];
SmallSet<Register, 8> FoundRegUnits;
bool AnyIllegal = false;
for (MCRegUnitIterator RUI(R.asMCReg(), TRI); RUI.isValid(); ++RUI) {
for (MCRegUnitRootIterator URoot(*RUI, TRI); URoot.isValid(); ++URoot){
if (!MTracker->isRegisterTracked(*URoot)) {
// Not all roots were loaded into the tracking map: this register
// isn't actually def'd anywhere, we only read from it. Generate PHIs
// for this reg, but don't iterate units.
AnyIllegal = true;
} else {
FoundRegUnits.insert(*URoot);
}
}
}
if (AnyIllegal) {
NormalLocsToPHI.insert(L);
continue;
}
RegUnitsToPHIUp.insert(FoundRegUnits.begin(), FoundRegUnits.end());
}
// Lambda to fetch PHIs for a given location, and write into the PHIBlocks
// collection.
SmallVector<MachineBasicBlock *, 32> PHIBlocks;
auto CollectPHIsForLoc = [&](LocIdx L) {
// Collect the set of defs.
SmallPtrSet<MachineBasicBlock *, 32> DefBlocks;
for (unsigned int I = 0; I < OrderToBB.size(); ++I) {
MachineBasicBlock *MBB = OrderToBB[I];
const auto &TransferFunc = MLocTransfer[MBB->getNumber()];
if (TransferFunc.find(L) != TransferFunc.end())
DefBlocks.insert(MBB);
}
// The entry block defs the location too: it's the live-in / argument value.
// Only insert if there are other defs though; everything is trivially live
// through otherwise.
if (!DefBlocks.empty())
DefBlocks.insert(&*MF.begin());
// Ask the SSA construction algorithm where we should put PHIs. Clear
// anything that might have been hanging around from earlier.
PHIBlocks.clear();
BlockPHIPlacement(AllBlocks, DefBlocks, PHIBlocks);
};
auto InstallPHIsAtLoc = [&PHIBlocks, &MInLocs](LocIdx L) {
for (const MachineBasicBlock *MBB : PHIBlocks)
MInLocs[MBB->getNumber()][L.asU64()] = ValueIDNum(MBB->getNumber(), 0, L);
};
// For locations with no reg units, just place PHIs.
for (LocIdx L : NormalLocsToPHI) {
CollectPHIsForLoc(L);
// Install those PHI values into the live-in value array.
InstallPHIsAtLoc(L);
}
// For stack slots, calculate PHIs for the equivalent of the units, then
// install for each index.
for (SpillLocationNo Slot : StackSlots) {
for (unsigned Idx : StackUnits) {
unsigned SpillID = MTracker->getSpillIDWithIdx(Slot, Idx);
LocIdx L = MTracker->getSpillMLoc(SpillID);
CollectPHIsForLoc(L);
InstallPHIsAtLoc(L);
// Find anything that aliases this stack index, install PHIs for it too.
unsigned Size, Offset;
std::tie(Size, Offset) = MTracker->StackIdxesToPos[Idx];
for (auto &Pair : MTracker->StackSlotIdxes) {
unsigned ThisSize, ThisOffset;
std::tie(ThisSize, ThisOffset) = Pair.first;
if (ThisSize + ThisOffset <= Offset || Size + Offset <= ThisOffset)
continue;
unsigned ThisID = MTracker->getSpillIDWithIdx(Slot, Pair.second);
LocIdx ThisL = MTracker->getSpillMLoc(ThisID);
InstallPHIsAtLoc(ThisL);
}
}
}
// For reg units, place PHIs, and then place them for any aliasing registers.
for (Register R : RegUnitsToPHIUp) {
LocIdx L = MTracker->lookupOrTrackRegister(R);
CollectPHIsForLoc(L);
// Install those PHI values into the live-in value array.
InstallPHIsAtLoc(L);
// Now find aliases and install PHIs for those.
for (MCRegAliasIterator RAI(R, TRI, true); RAI.isValid(); ++RAI) {
// Super-registers that are "above" the largest register read/written by
// the function will alias, but will not be tracked.
if (!MTracker->isRegisterTracked(*RAI))
continue;
LocIdx AliasLoc = MTracker->lookupOrTrackRegister(*RAI);
InstallPHIsAtLoc(AliasLoc);
}
}
}
void InstrRefBasedLDV::buildMLocValueMap(
MachineFunction &MF, FuncValueTable &MInLocs, FuncValueTable &MOutLocs,
SmallVectorImpl<MLocTransferMap> &MLocTransfer) {
std::priority_queue<unsigned int, std::vector<unsigned int>,
std::greater<unsigned int>>
Worklist, Pending;
// We track what is on the current and pending worklist to avoid inserting
// the same thing twice. We could avoid this with a custom priority queue,
// but this is probably not worth it.
SmallPtrSet<MachineBasicBlock *, 16> OnPending, OnWorklist;
// Initialize worklist with every block to be visited. Also produce list of
// all blocks.
SmallPtrSet<MachineBasicBlock *, 32> AllBlocks;
for (unsigned int I = 0; I < BBToOrder.size(); ++I) {
Worklist.push(I);
OnWorklist.insert(OrderToBB[I]);
AllBlocks.insert(OrderToBB[I]);
}
// Initialize entry block to PHIs. These represent arguments.
for (auto Location : MTracker->locations())
MInLocs[0][Location.Idx.asU64()] = ValueIDNum(0, 0, Location.Idx);
MTracker->reset();
// Start by placing PHIs, using the usual SSA constructor algorithm. Consider
// any machine-location that isn't live-through a block to be def'd in that
// block.
placeMLocPHIs(MF, AllBlocks, MInLocs, MLocTransfer);
// Propagate values to eliminate redundant PHIs. At the same time, this
// produces the table of Block x Location => Value for the entry to each
// block.
// The kind of PHIs we can eliminate are, for example, where one path in a
// conditional spills and restores a register, and the register still has
// the same value once control flow joins, unbeknowns to the PHI placement
// code. Propagating values allows us to identify such un-necessary PHIs and
// remove them.
SmallPtrSet<const MachineBasicBlock *, 16> Visited;
while (!Worklist.empty() || !Pending.empty()) {
// Vector for storing the evaluated block transfer function.
SmallVector<std::pair<LocIdx, ValueIDNum>, 32> ToRemap;
while (!Worklist.empty()) {
MachineBasicBlock *MBB = OrderToBB[Worklist.top()];
CurBB = MBB->getNumber();
Worklist.pop();
// Join the values in all predecessor blocks.
bool InLocsChanged;
InLocsChanged = mlocJoin(*MBB, Visited, MOutLocs, MInLocs[CurBB]);
InLocsChanged |= Visited.insert(MBB).second;
// Don't examine transfer function if we've visited this loc at least
// once, and inlocs haven't changed.
if (!InLocsChanged)
continue;
// Load the current set of live-ins into MLocTracker.
MTracker->loadFromArray(MInLocs[CurBB], CurBB);
// Each element of the transfer function can be a new def, or a read of
// a live-in value. Evaluate each element, and store to "ToRemap".
ToRemap.clear();
for (auto &P : MLocTransfer[CurBB]) {
if (P.second.getBlock() == CurBB && P.second.isPHI()) {
// This is a movement of whatever was live in. Read it.
ValueIDNum NewID = MTracker->readMLoc(P.second.getLoc());
ToRemap.push_back(std::make_pair(P.first, NewID));
} else {
// It's a def. Just set it.
assert(P.second.getBlock() == CurBB);
ToRemap.push_back(std::make_pair(P.first, P.second));
}
}
// Commit the transfer function changes into mloc tracker, which
// transforms the contents of the MLocTracker into the live-outs.
for (auto &P : ToRemap)
MTracker->setMLoc(P.first, P.second);
// Now copy out-locs from mloc tracker into out-loc vector, checking
// whether changes have occurred. These changes can have come from both
// the transfer function, and mlocJoin.
bool OLChanged = false;
for (auto Location : MTracker->locations()) {
OLChanged |= MOutLocs[CurBB][Location.Idx.asU64()] != Location.Value;
MOutLocs[CurBB][Location.Idx.asU64()] = Location.Value;
}
MTracker->reset();
// No need to examine successors again if out-locs didn't change.
if (!OLChanged)
continue;
// All successors should be visited: put any back-edges on the pending
// list for the next pass-through, and any other successors to be
// visited this pass, if they're not going to be already.
for (auto s : MBB->successors()) {
// Does branching to this successor represent a back-edge?
if (BBToOrder[s] > BBToOrder[MBB]) {
// No: visit it during this dataflow iteration.
if (OnWorklist.insert(s).second)
Worklist.push(BBToOrder[s]);
} else {
// Yes: visit it on the next iteration.
if (OnPending.insert(s).second)
Pending.push(BBToOrder[s]);
}
}
}
Worklist.swap(Pending);
std::swap(OnPending, OnWorklist);
OnPending.clear();
// At this point, pending must be empty, since it was just the empty
// worklist
assert(Pending.empty() && "Pending should be empty");
}
// Once all the live-ins don't change on mlocJoin(), we've eliminated all
// redundant PHIs.
}
void InstrRefBasedLDV::BlockPHIPlacement(
const SmallPtrSetImpl<MachineBasicBlock *> &AllBlocks,
const SmallPtrSetImpl<MachineBasicBlock *> &DefBlocks,
SmallVectorImpl<MachineBasicBlock *> &PHIBlocks) {
// Apply IDF calculator to the designated set of location defs, storing
// required PHIs into PHIBlocks. Uses the dominator tree stored in the
// InstrRefBasedLDV object.
IDFCalculatorBase<MachineBasicBlock, false> IDF(DomTree->getBase());
IDF.setLiveInBlocks(AllBlocks);
IDF.setDefiningBlocks(DefBlocks);
IDF.calculate(PHIBlocks);
}
Optional<ValueIDNum> InstrRefBasedLDV::pickVPHILoc(
const MachineBasicBlock &MBB, const DebugVariable &Var,
const LiveIdxT &LiveOuts, FuncValueTable &MOutLocs,
const SmallVectorImpl<const MachineBasicBlock *> &BlockOrders) {
// Collect a set of locations from predecessor where its live-out value can
// be found.
SmallVector<SmallVector<LocIdx, 4>, 8> Locs;
SmallVector<const DbgValueProperties *, 4> Properties;
unsigned NumLocs = MTracker->getNumLocs();
// No predecessors means no PHIs.
if (BlockOrders.empty())
return None;
for (auto p : BlockOrders) {
unsigned ThisBBNum = p->getNumber();
auto OutValIt = LiveOuts.find(p);
if (OutValIt == LiveOuts.end())
// If we have a predecessor not in scope, we'll never find a PHI position.
return None;
const DbgValue &OutVal = *OutValIt->second;
if (OutVal.Kind == DbgValue::Const || OutVal.Kind == DbgValue::NoVal)
// Consts and no-values cannot have locations we can join on.
return None;
Properties.push_back(&OutVal.Properties);
// Create new empty vector of locations.
Locs.resize(Locs.size() + 1);
// If the live-in value is a def, find the locations where that value is
// present. Do the same for VPHIs where we know the VPHI value.
if (OutVal.Kind == DbgValue::Def ||
(OutVal.Kind == DbgValue::VPHI && OutVal.BlockNo != MBB.getNumber() &&
OutVal.ID != ValueIDNum::EmptyValue)) {
ValueIDNum ValToLookFor = OutVal.ID;
// Search the live-outs of the predecessor for the specified value.
for (unsigned int I = 0; I < NumLocs; ++I) {
if (MOutLocs[ThisBBNum][I] == ValToLookFor)
Locs.back().push_back(LocIdx(I));
}
} else {
assert(OutVal.Kind == DbgValue::VPHI);
// For VPHIs where we don't know the location, we definitely can't find
// a join loc.
if (OutVal.BlockNo != MBB.getNumber())
return None;
// Otherwise: this is a VPHI on a backedge feeding back into itself, i.e.
// a value that's live-through the whole loop. (It has to be a backedge,
// because a block can't dominate itself). We can accept as a PHI location
// any location where the other predecessors agree, _and_ the machine
// locations feed back into themselves. Therefore, add all self-looping
// machine-value PHI locations.
for (unsigned int I = 0; I < NumLocs; ++I) {
ValueIDNum MPHI(MBB.getNumber(), 0, LocIdx(I));
if (MOutLocs[ThisBBNum][I] == MPHI)
Locs.back().push_back(LocIdx(I));
}
}
}
// We should have found locations for all predecessors, or returned.
assert(Locs.size() == BlockOrders.size());
// Check that all properties are the same. We can't pick a location if they're
// not.
const DbgValueProperties *Properties0 = Properties[0];
for (auto *Prop : Properties)
if (*Prop != *Properties0)
return None;
// Starting with the first set of locations, take the intersection with
// subsequent sets.
SmallVector<LocIdx, 4> CandidateLocs = Locs[0];
for (unsigned int I = 1; I < Locs.size(); ++I) {
auto &LocVec = Locs[I];
SmallVector<LocIdx, 4> NewCandidates;
std::set_intersection(CandidateLocs.begin(), CandidateLocs.end(),
LocVec.begin(), LocVec.end(), std::inserter(NewCandidates, NewCandidates.begin()));
CandidateLocs = NewCandidates;
}
if (CandidateLocs.empty())
return None;
// We now have a set of LocIdxes that contain the right output value in
// each of the predecessors. Pick the lowest; if there's a register loc,
// that'll be it.
LocIdx L = *CandidateLocs.begin();
// Return a PHI-value-number for the found location.
ValueIDNum PHIVal = {(unsigned)MBB.getNumber(), 0, L};
return PHIVal;
}
bool InstrRefBasedLDV::vlocJoin(
MachineBasicBlock &MBB, LiveIdxT &VLOCOutLocs,
SmallPtrSet<const MachineBasicBlock *, 8> &BlocksToExplore,
DbgValue &LiveIn) {
LLVM_DEBUG(dbgs() << "join MBB: " << MBB.getNumber() << "\n");
bool Changed = false;
// Order predecessors by RPOT order, for exploring them in that order.
SmallVector<MachineBasicBlock *, 8> BlockOrders(MBB.predecessors());
auto Cmp = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
return BBToOrder[A] < BBToOrder[B];
};
llvm::sort(BlockOrders, Cmp);
unsigned CurBlockRPONum = BBToOrder[&MBB];
// Collect all the incoming DbgValues for this variable, from predecessor
// live-out values.
SmallVector<InValueT, 8> Values;
bool Bail = false;
int BackEdgesStart = 0;
for (auto p : BlockOrders) {
// If the predecessor isn't in scope / to be explored, we'll never be
// able to join any locations.
if (!BlocksToExplore.contains(p)) {
Bail = true;
break;
}
// All Live-outs will have been initialized.
DbgValue &OutLoc = *VLOCOutLocs.find(p)->second;
// Keep track of where back-edges begin in the Values vector. Relies on
// BlockOrders being sorted by RPO.
unsigned ThisBBRPONum = BBToOrder[p];
if (ThisBBRPONum < CurBlockRPONum)
++BackEdgesStart;
Values.push_back(std::make_pair(p, &OutLoc));
}
// If there were no values, or one of the predecessors couldn't have a
// value, then give up immediately. It's not safe to produce a live-in
// value. Leave as whatever it was before.
if (Bail || Values.size() == 0)
return false;
// All (non-entry) blocks have at least one non-backedge predecessor.
// Pick the variable value from the first of these, to compare against
// all others.
const DbgValue &FirstVal = *Values[0].second;
// If the old live-in value is not a PHI then either a) no PHI is needed
// here, or b) we eliminated the PHI that was here. If so, we can just
// propagate in the first parent's incoming value.
if (LiveIn.Kind != DbgValue::VPHI || LiveIn.BlockNo != MBB.getNumber()) {
Changed = LiveIn != FirstVal;
if (Changed)
LiveIn = FirstVal;
return Changed;
}
// Scan for variable values that can never be resolved: if they have
// different DIExpressions, different indirectness, or are mixed constants /
// non-constants.
for (auto &V : Values) {
if (V.second->Properties != FirstVal.Properties)
return false;
if (V.second->Kind == DbgValue::NoVal)
return false;
if (V.second->Kind == DbgValue::Const && FirstVal.Kind != DbgValue::Const)
return false;
}
// Try to eliminate this PHI. Do the incoming values all agree?
bool Disagree = false;
for (auto &V : Values) {
if (*V.second == FirstVal)
continue; // No disagreement.
// Eliminate if a backedge feeds a VPHI back into itself.
if (V.second->Kind == DbgValue::VPHI &&
V.second->BlockNo == MBB.getNumber() &&
// Is this a backedge?
std::distance(Values.begin(), &V) >= BackEdgesStart)
continue;
Disagree = true;
}
// No disagreement -> live-through value.
if (!Disagree) {
Changed = LiveIn != FirstVal;
if (Changed)
LiveIn = FirstVal;
return Changed;
} else {
// Otherwise use a VPHI.
DbgValue VPHI(MBB.getNumber(), FirstVal.Properties, DbgValue::VPHI);
Changed = LiveIn != VPHI;
if (Changed)
LiveIn = VPHI;
return Changed;
}
}
void InstrRefBasedLDV::getBlocksForScope(
const DILocation *DILoc,
SmallPtrSetImpl<const MachineBasicBlock *> &BlocksToExplore,
const SmallPtrSetImpl<MachineBasicBlock *> &AssignBlocks) {
// Get the set of "normal" in-lexical-scope blocks.
LS.getMachineBasicBlocks(DILoc, BlocksToExplore);
// VarLoc LiveDebugValues tracks variable locations that are defined in
// blocks not in scope. This is something we could legitimately ignore, but
// lets allow it for now for the sake of coverage.
BlocksToExplore.insert(AssignBlocks.begin(), AssignBlocks.end());
// Storage for artificial blocks we intend to add to BlocksToExplore.
DenseSet<const MachineBasicBlock *> ToAdd;
// To avoid needlessly dropping large volumes of variable locations, propagate
// variables through aritifical blocks, i.e. those that don't have any
// instructions in scope at all. To accurately replicate VarLoc
// LiveDebugValues, this means exploring all artificial successors too.
// Perform a depth-first-search to enumerate those blocks.
for (auto *MBB : BlocksToExplore) {
// Depth-first-search state: each node is a block and which successor
// we're currently exploring.
SmallVector<std::pair<const MachineBasicBlock *,
MachineBasicBlock::const_succ_iterator>,
8>
DFS;
// Find any artificial successors not already tracked.
for (auto *succ : MBB->successors()) {
if (BlocksToExplore.count(succ))
continue;
if (!ArtificialBlocks.count(succ))
continue;
ToAdd.insert(succ);
DFS.push_back({succ, succ->succ_begin()});
}
// Search all those blocks, depth first.
while (!DFS.empty()) {
const MachineBasicBlock *CurBB = DFS.back().first;
MachineBasicBlock::const_succ_iterator &CurSucc = DFS.back().second;
// Walk back if we've explored this blocks successors to the end.
if (CurSucc == CurBB->succ_end()) {
DFS.pop_back();
continue;
}
// If the current successor is artificial and unexplored, descend into
// it.
if (!ToAdd.count(*CurSucc) && ArtificialBlocks.count(*CurSucc)) {
ToAdd.insert(*CurSucc);
DFS.push_back({*CurSucc, (*CurSucc)->succ_begin()});
continue;
}
++CurSucc;
}
};
BlocksToExplore.insert(ToAdd.begin(), ToAdd.end());
}
void InstrRefBasedLDV::buildVLocValueMap(
const DILocation *DILoc, const SmallSet<DebugVariable, 4> &VarsWeCareAbout,
SmallPtrSetImpl<MachineBasicBlock *> &AssignBlocks, LiveInsT &Output,
FuncValueTable &MOutLocs, FuncValueTable &MInLocs,
SmallVectorImpl<VLocTracker> &AllTheVLocs) {
// This method is much like buildMLocValueMap: but focuses on a single
// LexicalScope at a time. Pick out a set of blocks and variables that are
// to have their value assignments solved, then run our dataflow algorithm
// until a fixedpoint is reached.
std::priority_queue<unsigned int, std::vector<unsigned int>,
std::greater<unsigned int>>
Worklist, Pending;
SmallPtrSet<MachineBasicBlock *, 16> OnWorklist, OnPending;
// The set of blocks we'll be examining.
SmallPtrSet<const MachineBasicBlock *, 8> BlocksToExplore;
// The order in which to examine them (RPO).
SmallVector<MachineBasicBlock *, 8> BlockOrders;
// RPO ordering function.
auto Cmp = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
return BBToOrder[A] < BBToOrder[B];
};
getBlocksForScope(DILoc, BlocksToExplore, AssignBlocks);
// Single block scope: not interesting! No propagation at all. Note that
// this could probably go above ArtificialBlocks without damage, but
// that then produces output differences from original-live-debug-values,
// which propagates from a single block into many artificial ones.
if (BlocksToExplore.size() == 1)
return;
// Convert a const set to a non-const set. LexicalScopes
// getMachineBasicBlocks returns const MBB pointers, IDF wants mutable ones.
// (Neither of them mutate anything).
SmallPtrSet<MachineBasicBlock *, 8> MutBlocksToExplore;
for (const auto *MBB : BlocksToExplore)
MutBlocksToExplore.insert(const_cast<MachineBasicBlock *>(MBB));
// Picks out relevants blocks RPO order and sort them.
for (auto *MBB : BlocksToExplore)
BlockOrders.push_back(const_cast<MachineBasicBlock *>(MBB));
llvm::sort(BlockOrders, Cmp);
unsigned NumBlocks = BlockOrders.size();
// Allocate some vectors for storing the live ins and live outs. Large.
SmallVector<DbgValue, 32> LiveIns, LiveOuts;
LiveIns.reserve(NumBlocks);
LiveOuts.reserve(NumBlocks);
// Initialize all values to start as NoVals. This signifies "it's live
// through, but we don't know what it is".
DbgValueProperties EmptyProperties(EmptyExpr, false);
for (unsigned int I = 0; I < NumBlocks; ++I) {
DbgValue EmptyDbgValue(I, EmptyProperties, DbgValue::NoVal);
LiveIns.push_back(EmptyDbgValue);
LiveOuts.push_back(EmptyDbgValue);
}
// Produce by-MBB indexes of live-in/live-outs, to ease lookup within
// vlocJoin.
LiveIdxT LiveOutIdx, LiveInIdx;
LiveOutIdx.reserve(NumBlocks);
LiveInIdx.reserve(NumBlocks);
for (unsigned I = 0; I < NumBlocks; ++I) {
LiveOutIdx[BlockOrders[I]] = &LiveOuts[I];
LiveInIdx[BlockOrders[I]] = &LiveIns[I];
}
// Loop over each variable and place PHIs for it, then propagate values
// between blocks. This keeps the locality of working on one lexical scope at
// at time, but avoids re-processing variable values because some other
// variable has been assigned.
for (auto &Var : VarsWeCareAbout) {
// Re-initialize live-ins and live-outs, to clear the remains of previous
// variables live-ins / live-outs.
for (unsigned int I = 0; I < NumBlocks; ++I) {
DbgValue EmptyDbgValue(I, EmptyProperties, DbgValue::NoVal);
LiveIns[I] = EmptyDbgValue;
LiveOuts[I] = EmptyDbgValue;
}
// Place PHIs for variable values, using the LLVM IDF calculator.
// Collect the set of blocks where variables are def'd.
SmallPtrSet<MachineBasicBlock *, 32> DefBlocks;
for (const MachineBasicBlock *ExpMBB : BlocksToExplore) {
auto &TransferFunc = AllTheVLocs[ExpMBB->getNumber()].Vars;
if (TransferFunc.find(Var) != TransferFunc.end())
DefBlocks.insert(const_cast<MachineBasicBlock *>(ExpMBB));
}
SmallVector<MachineBasicBlock *, 32> PHIBlocks;
// Request the set of PHIs we should insert for this variable. If there's
// only one value definition, things are very simple.
if (DefBlocks.size() == 1) {
placePHIsForSingleVarDefinition(MutBlocksToExplore, *DefBlocks.begin(),
AllTheVLocs, Var, Output);
continue;
}
// Otherwise: we need to place PHIs through SSA and propagate values.
BlockPHIPlacement(MutBlocksToExplore, DefBlocks, PHIBlocks);
// Insert PHIs into the per-block live-in tables for this variable.
for (MachineBasicBlock *PHIMBB : PHIBlocks) {
unsigned BlockNo = PHIMBB->getNumber();
DbgValue *LiveIn = LiveInIdx[PHIMBB];
*LiveIn = DbgValue(BlockNo, EmptyProperties, DbgValue::VPHI);
}
for (auto *MBB : BlockOrders) {
Worklist.push(BBToOrder[MBB]);
OnWorklist.insert(MBB);
}
// Iterate over all the blocks we selected, propagating the variables value.
// This loop does two things:
// * Eliminates un-necessary VPHIs in vlocJoin,
// * Evaluates the blocks transfer function (i.e. variable assignments) and
// stores the result to the blocks live-outs.
// Always evaluate the transfer function on the first iteration, and when
// the live-ins change thereafter.
bool FirstTrip = true;
while (!Worklist.empty() || !Pending.empty()) {
while (!Worklist.empty()) {
auto *MBB = OrderToBB[Worklist.top()];
CurBB = MBB->getNumber();
Worklist.pop();
auto LiveInsIt = LiveInIdx.find(MBB);
assert(LiveInsIt != LiveInIdx.end());
DbgValue *LiveIn = LiveInsIt->second;
// Join values from predecessors. Updates LiveInIdx, and writes output
// into JoinedInLocs.
bool InLocsChanged =
vlocJoin(*MBB, LiveOutIdx, BlocksToExplore, *LiveIn);
SmallVector<const MachineBasicBlock *, 8> Preds;
for (const auto *Pred : MBB->predecessors())
Preds.push_back(Pred);
// If this block's live-in value is a VPHI, try to pick a machine-value
// for it. This makes the machine-value available and propagated
// through all blocks by the time value propagation finishes. We can't
// do this any earlier as it needs to read the block live-outs.
if (LiveIn->Kind == DbgValue::VPHI && LiveIn->BlockNo == (int)CurBB) {
// There's a small possibility that on a preceeding path, a VPHI is
// eliminated and transitions from VPHI-with-location to
// live-through-value. As a result, the selected location of any VPHI
// might change, so we need to re-compute it on each iteration.
Optional<ValueIDNum> ValueNum =
pickVPHILoc(*MBB, Var, LiveOutIdx, MOutLocs, Preds);
if (ValueNum) {
InLocsChanged |= LiveIn->ID != *ValueNum;
LiveIn->ID = *ValueNum;
}
}
if (!InLocsChanged && !FirstTrip)
continue;
DbgValue *LiveOut = LiveOutIdx[MBB];
bool OLChanged = false;
// Do transfer function.
auto &VTracker = AllTheVLocs[MBB->getNumber()];
auto TransferIt = VTracker.Vars.find(Var);
if (TransferIt != VTracker.Vars.end()) {
// Erase on empty transfer (DBG_VALUE $noreg).
if (TransferIt->second.Kind == DbgValue::Undef) {
DbgValue NewVal(MBB->getNumber(), EmptyProperties, DbgValue::NoVal);
if (*LiveOut != NewVal) {
*LiveOut = NewVal;
OLChanged = true;
}
} else {
// Insert new variable value; or overwrite.
if (*LiveOut != TransferIt->second) {
*LiveOut = TransferIt->second;
OLChanged = true;
}
}
} else {
// Just copy live-ins to live-outs, for anything not transferred.
if (*LiveOut != *LiveIn) {
*LiveOut = *LiveIn;
OLChanged = true;
}
}
// If no live-out value changed, there's no need to explore further.
if (!OLChanged)
continue;
// We should visit all successors. Ensure we'll visit any non-backedge
// successors during this dataflow iteration; book backedge successors
// to be visited next time around.
for (auto s : MBB->successors()) {
// Ignore out of scope / not-to-be-explored successors.
if (LiveInIdx.find(s) == LiveInIdx.end())
continue;
if (BBToOrder[s] > BBToOrder[MBB]) {
if (OnWorklist.insert(s).second)
Worklist.push(BBToOrder[s]);
} else if (OnPending.insert(s).second && (FirstTrip || OLChanged)) {
Pending.push(BBToOrder[s]);
}
}
}
Worklist.swap(Pending);
std::swap(OnWorklist, OnPending);
OnPending.clear();
assert(Pending.empty());
FirstTrip = false;
}
// Save live-ins to output vector. Ignore any that are still marked as being
// VPHIs with no location -- those are variables that we know the value of,
// but are not actually available in the register file.
for (auto *MBB : BlockOrders) {
DbgValue *BlockLiveIn = LiveInIdx[MBB];
if (BlockLiveIn->Kind == DbgValue::NoVal)
continue;
if (BlockLiveIn->Kind == DbgValue::VPHI &&
BlockLiveIn->ID == ValueIDNum::EmptyValue)
continue;
if (BlockLiveIn->Kind == DbgValue::VPHI)
BlockLiveIn->Kind = DbgValue::Def;
assert(BlockLiveIn->Properties.DIExpr->getFragmentInfo() ==
Var.getFragment() && "Fragment info missing during value prop");
Output[MBB->getNumber()].push_back(std::make_pair(Var, *BlockLiveIn));
}
} // Per-variable loop.
BlockOrders.clear();
BlocksToExplore.clear();
}
void InstrRefBasedLDV::placePHIsForSingleVarDefinition(
const SmallPtrSetImpl<MachineBasicBlock *> &InScopeBlocks,
MachineBasicBlock *AssignMBB, SmallVectorImpl<VLocTracker> &AllTheVLocs,
const DebugVariable &Var, LiveInsT &Output) {
// If there is a single definition of the variable, then working out it's
// value everywhere is very simple: it's every block dominated by the
// definition. At the dominance frontier, the usual algorithm would:
// * Place PHIs,
// * Propagate values into them,
// * Find there's no incoming variable value from the other incoming branches
// of the dominance frontier,
// * Specify there's no variable value in blocks past the frontier.
// This is a common case, hence it's worth special-casing it.
// Pick out the variables value from the block transfer function.
VLocTracker &VLocs = AllTheVLocs[AssignMBB->getNumber()];
auto ValueIt = VLocs.Vars.find(Var);
const DbgValue &Value = ValueIt->second;
// If it's an explicit assignment of "undef", that means there is no location
// anyway, anywhere.
if (Value.Kind == DbgValue::Undef)
return;
// Assign the variable value to entry to each dominated block that's in scope.
// Skip the definition block -- it's assigned the variable value in the middle
// of the block somewhere.
for (auto *ScopeBlock : InScopeBlocks) {
if (!DomTree->properlyDominates(AssignMBB, ScopeBlock))
continue;
Output[ScopeBlock->getNumber()].push_back({Var, Value});
}
// All blocks that aren't dominated have no live-in value, thus no variable
// value will be given to them.
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void InstrRefBasedLDV::dump_mloc_transfer(
const MLocTransferMap &mloc_transfer) const {
for (auto &P : mloc_transfer) {
std::string foo = MTracker->LocIdxToName(P.first);
std::string bar = MTracker->IDAsString(P.second);
dbgs() << "Loc " << foo << " --> " << bar << "\n";
}
}
#endif
void InstrRefBasedLDV::initialSetup(MachineFunction &MF) {
// Build some useful data structures.
LLVMContext &Context = MF.getFunction().getContext();
EmptyExpr = DIExpression::get(Context, {});
auto hasNonArtificialLocation = [](const MachineInstr &MI) -> bool {
if (const DebugLoc &DL = MI.getDebugLoc())
return DL.getLine() != 0;
return false;
};
// Collect a set of all the artificial blocks.
for (auto &MBB : MF)
if (none_of(MBB.instrs(), hasNonArtificialLocation))
ArtificialBlocks.insert(&MBB);
// Compute mappings of block <=> RPO order.
ReversePostOrderTraversal<MachineFunction *> RPOT(&MF);
unsigned int RPONumber = 0;
for (MachineBasicBlock *MBB : RPOT) {
OrderToBB[RPONumber] = MBB;
BBToOrder[MBB] = RPONumber;
BBNumToRPO[MBB->getNumber()] = RPONumber;
++RPONumber;
}
// Order value substitutions by their "source" operand pair, for quick lookup.
llvm::sort(MF.DebugValueSubstitutions);
#ifdef EXPENSIVE_CHECKS
// As an expensive check, test whether there are any duplicate substitution
// sources in the collection.
if (MF.DebugValueSubstitutions.size() > 2) {
for (auto It = MF.DebugValueSubstitutions.begin();
It != std::prev(MF.DebugValueSubstitutions.end()); ++It) {
assert(It->Src != std::next(It)->Src && "Duplicate variable location "
"substitution seen");
}
}
#endif
}
// Produce an "ejection map" for blocks, i.e., what's the highest-numbered
// lexical scope it's used in. When exploring in DFS order and we pass that
// scope, the block can be processed and any tracking information freed.
void InstrRefBasedLDV::makeDepthFirstEjectionMap(
SmallVectorImpl<unsigned> &EjectionMap,
const ScopeToDILocT &ScopeToDILocation,
ScopeToAssignBlocksT &ScopeToAssignBlocks) {
SmallPtrSet<const MachineBasicBlock *, 8> BlocksToExplore;
SmallVector<std::pair<LexicalScope *, ssize_t>, 4> WorkStack;
auto *TopScope = LS.getCurrentFunctionScope();
// Unlike lexical scope explorers, we explore in reverse order, to find the
// "last" lexical scope used for each block early.
WorkStack.push_back({TopScope, TopScope->getChildren().size() - 1});
while (!WorkStack.empty()) {
auto &ScopePosition = WorkStack.back();
LexicalScope *WS = ScopePosition.first;
ssize_t ChildNum = ScopePosition.second--;
const SmallVectorImpl<LexicalScope *> &Children = WS->getChildren();
if (ChildNum >= 0) {
// If ChildNum is positive, there are remaining children to explore.
// Push the child and its children-count onto the stack.
auto &ChildScope = Children[ChildNum];
WorkStack.push_back(
std::make_pair(ChildScope, ChildScope->getChildren().size() - 1));
} else {
WorkStack.pop_back();
// We've explored all children and any later blocks: examine all blocks
// in our scope. If they haven't yet had an ejection number set, then
// this scope will be the last to use that block.
auto DILocationIt = ScopeToDILocation.find(WS);
if (DILocationIt != ScopeToDILocation.end()) {
getBlocksForScope(DILocationIt->second, BlocksToExplore,
ScopeToAssignBlocks.find(WS)->second);
for (auto *MBB : BlocksToExplore) {
unsigned BBNum = MBB->getNumber();
if (EjectionMap[BBNum] == 0)
EjectionMap[BBNum] = WS->getDFSOut();
}
BlocksToExplore.clear();
}
}
}
}
bool InstrRefBasedLDV::depthFirstVLocAndEmit(
unsigned MaxNumBlocks, const ScopeToDILocT &ScopeToDILocation,
const ScopeToVarsT &ScopeToVars, ScopeToAssignBlocksT &ScopeToAssignBlocks,
LiveInsT &Output, FuncValueTable &MOutLocs, FuncValueTable &MInLocs,
SmallVectorImpl<VLocTracker> &AllTheVLocs, MachineFunction &MF,
DenseMap<DebugVariable, unsigned> &AllVarsNumbering,
const TargetPassConfig &TPC) {
TTracker = new TransferTracker(TII, MTracker, MF, *TRI, CalleeSavedRegs, TPC);
unsigned NumLocs = MTracker->getNumLocs();
VTracker = nullptr;
// No scopes? No variable locations.
if (!LS.getCurrentFunctionScope())
return false;
// Build map from block number to the last scope that uses the block.
SmallVector<unsigned, 16> EjectionMap;
EjectionMap.resize(MaxNumBlocks, 0);
makeDepthFirstEjectionMap(EjectionMap, ScopeToDILocation,
ScopeToAssignBlocks);
// Helper lambda for ejecting a block -- if nothing is going to use the block,
// we can translate the variable location information into DBG_VALUEs and then
// free all of InstrRefBasedLDV's data structures.
auto EjectBlock = [&](MachineBasicBlock &MBB) -> void {
unsigned BBNum = MBB.getNumber();
AllTheVLocs[BBNum].clear();
// Prime the transfer-tracker, and then step through all the block
// instructions, installing transfers.
MTracker->reset();
MTracker->loadFromArray(MInLocs[BBNum], BBNum);
TTracker->loadInlocs(MBB, MInLocs[BBNum], Output[BBNum], NumLocs);
CurBB = BBNum;
CurInst = 1;
for (auto &MI : MBB) {
process(MI, MOutLocs.get(), MInLocs.get());
TTracker->checkInstForNewValues(CurInst, MI.getIterator());
++CurInst;
}
// Free machine-location tables for this block.
MInLocs[BBNum].reset();
MOutLocs[BBNum].reset();
// We don't need live-in variable values for this block either.
Output[BBNum].clear();
AllTheVLocs[BBNum].clear();
};
SmallPtrSet<const MachineBasicBlock *, 8> BlocksToExplore;
SmallVector<std::pair<LexicalScope *, ssize_t>, 4> WorkStack;
WorkStack.push_back({LS.getCurrentFunctionScope(), 0});
unsigned HighestDFSIn = 0;
// Proceed to explore in depth first order.
while (!WorkStack.empty()) {
auto &ScopePosition = WorkStack.back();
LexicalScope *WS = ScopePosition.first;
ssize_t ChildNum = ScopePosition.second++;
// We obesrve scopes with children twice here, once descending in, once
// ascending out of the scope nest. Use HighestDFSIn as a ratchet to ensure
// we don't process a scope twice. Additionally, ignore scopes that don't
// have a DILocation -- by proxy, this means we never tracked any variable
// assignments in that scope.
auto DILocIt = ScopeToDILocation.find(WS);
if (HighestDFSIn <= WS->getDFSIn() && DILocIt != ScopeToDILocation.end()) {
const DILocation *DILoc = DILocIt->second;
auto &VarsWeCareAbout = ScopeToVars.find(WS)->second;
auto &BlocksInScope = ScopeToAssignBlocks.find(WS)->second;
buildVLocValueMap(DILoc, VarsWeCareAbout, BlocksInScope, Output, MOutLocs,
MInLocs, AllTheVLocs);
}
HighestDFSIn = std::max(HighestDFSIn, WS->getDFSIn());
// Descend into any scope nests.
const SmallVectorImpl<LexicalScope *> &Children = WS->getChildren();
if (ChildNum < (ssize_t)Children.size()) {
// There are children to explore -- push onto stack and continue.
auto &ChildScope = Children[ChildNum];
WorkStack.push_back(std::make_pair(ChildScope, 0));
} else {
WorkStack.pop_back();
// We've explored a leaf, or have explored all the children of a scope.
// Try to eject any blocks where this is the last scope it's relevant to.
auto DILocationIt = ScopeToDILocation.find(WS);
if (DILocationIt == ScopeToDILocation.end())
continue;
getBlocksForScope(DILocationIt->second, BlocksToExplore,
ScopeToAssignBlocks.find(WS)->second);
for (auto *MBB : BlocksToExplore)
if (WS->getDFSOut() == EjectionMap[MBB->getNumber()])
EjectBlock(const_cast<MachineBasicBlock &>(*MBB));
BlocksToExplore.clear();
}
}
// Some artificial blocks may not have been ejected, meaning they're not
// connected to an actual legitimate scope. This can technically happen
// with things like the entry block. In theory, we shouldn't need to do
// anything for such out-of-scope blocks, but for the sake of being similar
// to VarLocBasedLDV, eject these too.
for (auto *MBB : ArtificialBlocks)
if (MOutLocs[MBB->getNumber()])
EjectBlock(*MBB);
return emitTransfers(AllVarsNumbering);
}
bool InstrRefBasedLDV::emitTransfers(
DenseMap<DebugVariable, unsigned> &AllVarsNumbering) {
// Go through all the transfers recorded in the TransferTracker -- this is
// both the live-ins to a block, and any movements of values that happen
// in the middle.
for (const auto &P : TTracker->Transfers) {
// We have to insert DBG_VALUEs in a consistent order, otherwise they
// appear in DWARF in different orders. Use the order that they appear
// when walking through each block / each instruction, stored in
// AllVarsNumbering.
SmallVector<std::pair<unsigned, MachineInstr *>> Insts;
for (MachineInstr *MI : P.Insts) {
DebugVariable Var(MI->getDebugVariable(), MI->getDebugExpression(),
MI->getDebugLoc()->getInlinedAt());
Insts.emplace_back(AllVarsNumbering.find(Var)->second, MI);
}
llvm::sort(Insts,
[](const auto &A, const auto &B) { return A.first < B.first; });
// Insert either before or after the designated point...
if (P.MBB) {
MachineBasicBlock &MBB = *P.MBB;
for (const auto &Pair : Insts)
MBB.insert(P.Pos, Pair.second);
} else {
// Terminators, like tail calls, can clobber things. Don't try and place
// transfers after them.
if (P.Pos->isTerminator())
continue;
MachineBasicBlock &MBB = *P.Pos->getParent();
for (const auto &Pair : Insts)
MBB.insertAfterBundle(P.Pos, Pair.second);
}
}
return TTracker->Transfers.size() != 0;
}
/// Calculate the liveness information for the given machine function and
/// extend ranges across basic blocks.
bool InstrRefBasedLDV::ExtendRanges(MachineFunction &MF,
MachineDominatorTree *DomTree,
TargetPassConfig *TPC,
unsigned InputBBLimit,
unsigned InputDbgValLimit) {
// No subprogram means this function contains no debuginfo.
if (!MF.getFunction().getSubprogram())
return false;
LLVM_DEBUG(dbgs() << "\nDebug Range Extension\n");
this->TPC = TPC;
this->DomTree = DomTree;
TRI = MF.getSubtarget().getRegisterInfo();
MRI = &MF.getRegInfo();
TII = MF.getSubtarget().getInstrInfo();
TFI = MF.getSubtarget().getFrameLowering();
TFI->getCalleeSaves(MF, CalleeSavedRegs);
MFI = &MF.getFrameInfo();
LS.initialize(MF);
const auto &STI = MF.getSubtarget();
AdjustsStackInCalls = MFI->adjustsStack() &&
STI.getFrameLowering()->stackProbeFunctionModifiesSP();
if (AdjustsStackInCalls)
StackProbeSymbolName = STI.getTargetLowering()->getStackProbeSymbolName(MF);
MTracker =
new MLocTracker(MF, *TII, *TRI, *MF.getSubtarget().getTargetLowering());
VTracker = nullptr;
TTracker = nullptr;
SmallVector<MLocTransferMap, 32> MLocTransfer;
SmallVector<VLocTracker, 8> vlocs;
LiveInsT SavedLiveIns;
int MaxNumBlocks = -1;
for (auto &MBB : MF)
MaxNumBlocks = std::max(MBB.getNumber(), MaxNumBlocks);
assert(MaxNumBlocks >= 0);
++MaxNumBlocks;
MLocTransfer.resize(MaxNumBlocks);
vlocs.resize(MaxNumBlocks, VLocTracker(OverlapFragments, EmptyExpr));
SavedLiveIns.resize(MaxNumBlocks);
initialSetup(MF);
produceMLocTransferFunction(MF, MLocTransfer, MaxNumBlocks);
// Allocate and initialize two array-of-arrays for the live-in and live-out
// machine values. The outer dimension is the block number; while the inner
// dimension is a LocIdx from MLocTracker.
FuncValueTable MOutLocs = std::make_unique<ValueTable[]>(MaxNumBlocks);
FuncValueTable MInLocs = std::make_unique<ValueTable[]>(MaxNumBlocks);
unsigned NumLocs = MTracker->getNumLocs();
for (int i = 0; i < MaxNumBlocks; ++i) {
// These all auto-initialize to ValueIDNum::EmptyValue
MOutLocs[i] = std::make_unique<ValueIDNum[]>(NumLocs);
MInLocs[i] = std::make_unique<ValueIDNum[]>(NumLocs);
}
// Solve the machine value dataflow problem using the MLocTransfer function,
// storing the computed live-ins / live-outs into the array-of-arrays. We use
// both live-ins and live-outs for decision making in the variable value
// dataflow problem.
buildMLocValueMap(MF, MInLocs, MOutLocs, MLocTransfer);
// Patch up debug phi numbers, turning unknown block-live-in values into
// either live-through machine values, or PHIs.
for (auto &DBG_PHI : DebugPHINumToValue) {
// Identify unresolved block-live-ins.
if (!DBG_PHI.ValueRead)
continue;
ValueIDNum &Num = *DBG_PHI.ValueRead;
if (!Num.isPHI())
continue;
unsigned BlockNo = Num.getBlock();
LocIdx LocNo = Num.getLoc();
Num = MInLocs[BlockNo][LocNo.asU64()];
}
// Later, we'll be looking up ranges of instruction numbers.
llvm::sort(DebugPHINumToValue);
// Walk back through each block / instruction, collecting DBG_VALUE
// instructions and recording what machine value their operands refer to.
for (auto &OrderPair : OrderToBB) {
MachineBasicBlock &MBB = *OrderPair.second;
CurBB = MBB.getNumber();
VTracker = &vlocs[CurBB];
VTracker->MBB = &MBB;
MTracker->loadFromArray(MInLocs[CurBB], CurBB);
CurInst = 1;
for (auto &MI : MBB) {
process(MI, MOutLocs.get(), MInLocs.get());
++CurInst;
}
MTracker->reset();
}
// Number all variables in the order that they appear, to be used as a stable
// insertion order later.
DenseMap<DebugVariable, unsigned> AllVarsNumbering;
// Map from one LexicalScope to all the variables in that scope.
ScopeToVarsT ScopeToVars;
// Map from One lexical scope to all blocks where assignments happen for
// that scope.
ScopeToAssignBlocksT ScopeToAssignBlocks;
// Store map of DILocations that describes scopes.
ScopeToDILocT ScopeToDILocation;
// To mirror old LiveDebugValues, enumerate variables in RPOT order. Otherwise
// the order is unimportant, it just has to be stable.
unsigned VarAssignCount = 0;
for (unsigned int I = 0; I < OrderToBB.size(); ++I) {
auto *MBB = OrderToBB[I];
auto *VTracker = &vlocs[MBB->getNumber()];
// Collect each variable with a DBG_VALUE in this block.
for (auto &idx : VTracker->Vars) {
const auto &Var = idx.first;
const DILocation *ScopeLoc = VTracker->Scopes[Var];
assert(ScopeLoc != nullptr);
auto *Scope = LS.findLexicalScope(ScopeLoc);
// No insts in scope -> shouldn't have been recorded.
assert(Scope != nullptr);
AllVarsNumbering.insert(std::make_pair(Var, AllVarsNumbering.size()));
ScopeToVars[Scope].insert(Var);
ScopeToAssignBlocks[Scope].insert(VTracker->MBB);
ScopeToDILocation[Scope] = ScopeLoc;
++VarAssignCount;
}
}
bool Changed = false;
// If we have an extremely large number of variable assignments and blocks,
// bail out at this point. We've burnt some time doing analysis already,
// however we should cut our losses.
if ((unsigned)MaxNumBlocks > InputBBLimit &&
VarAssignCount > InputDbgValLimit) {
LLVM_DEBUG(dbgs() << "Disabling InstrRefBasedLDV: " << MF.getName()
<< " has " << MaxNumBlocks << " basic blocks and "
<< VarAssignCount
<< " variable assignments, exceeding limits.\n");
} else {
// Optionally, solve the variable value problem and emit to blocks by using
// a lexical-scope-depth search. It should be functionally identical to
// the "else" block of this condition.
Changed = depthFirstVLocAndEmit(
MaxNumBlocks, ScopeToDILocation, ScopeToVars, ScopeToAssignBlocks,
SavedLiveIns, MOutLocs, MInLocs, vlocs, MF, AllVarsNumbering, *TPC);
}
delete MTracker;
delete TTracker;
MTracker = nullptr;
VTracker = nullptr;
TTracker = nullptr;
ArtificialBlocks.clear();
OrderToBB.clear();
BBToOrder.clear();
BBNumToRPO.clear();
DebugInstrNumToInstr.clear();
DebugPHINumToValue.clear();
OverlapFragments.clear();
SeenFragments.clear();
SeenDbgPHIs.clear();
return Changed;
}
LDVImpl *llvm::makeInstrRefBasedLiveDebugValues() {
return new InstrRefBasedLDV();
}
namespace {
class LDVSSABlock;
class LDVSSAUpdater;
// Pick a type to identify incoming block values as we construct SSA. We
// can't use anything more robust than an integer unfortunately, as SSAUpdater
// expects to zero-initialize the type.
typedef uint64_t BlockValueNum;
/// Represents an SSA PHI node for the SSA updater class. Contains the block
/// this PHI is in, the value number it would have, and the expected incoming
/// values from parent blocks.
class LDVSSAPhi {
public:
SmallVector<std::pair<LDVSSABlock *, BlockValueNum>, 4> IncomingValues;
LDVSSABlock *ParentBlock;
BlockValueNum PHIValNum;
LDVSSAPhi(BlockValueNum PHIValNum, LDVSSABlock *ParentBlock)
: ParentBlock(ParentBlock), PHIValNum(PHIValNum) {}
LDVSSABlock *getParent() { return ParentBlock; }
};
/// Thin wrapper around a block predecessor iterator. Only difference from a
/// normal block iterator is that it dereferences to an LDVSSABlock.
class LDVSSABlockIterator {
public:
MachineBasicBlock::pred_iterator PredIt;
LDVSSAUpdater &Updater;
LDVSSABlockIterator(MachineBasicBlock::pred_iterator PredIt,
LDVSSAUpdater &Updater)
: PredIt(PredIt), Updater(Updater) {}
bool operator!=(const LDVSSABlockIterator &OtherIt) const {
return OtherIt.PredIt != PredIt;
}
LDVSSABlockIterator &operator++() {
++PredIt;
return *this;
}
LDVSSABlock *operator*();
};
/// Thin wrapper around a block for SSA Updater interface. Necessary because
/// we need to track the PHI value(s) that we may have observed as necessary
/// in this block.
class LDVSSABlock {
public:
MachineBasicBlock &BB;
LDVSSAUpdater &Updater;
using PHIListT = SmallVector<LDVSSAPhi, 1>;
/// List of PHIs in this block. There should only ever be one.
PHIListT PHIList;
LDVSSABlock(MachineBasicBlock &BB, LDVSSAUpdater &Updater)
: BB(BB), Updater(Updater) {}
LDVSSABlockIterator succ_begin() {
return LDVSSABlockIterator(BB.succ_begin(), Updater);
}
LDVSSABlockIterator succ_end() {
return LDVSSABlockIterator(BB.succ_end(), Updater);
}
/// SSAUpdater has requested a PHI: create that within this block record.
LDVSSAPhi *newPHI(BlockValueNum Value) {
PHIList.emplace_back(Value, this);
return &PHIList.back();
}
/// SSAUpdater wishes to know what PHIs already exist in this block.
PHIListT &phis() { return PHIList; }
};
/// Utility class for the SSAUpdater interface: tracks blocks, PHIs and values
/// while SSAUpdater is exploring the CFG. It's passed as a handle / baton to
// SSAUpdaterTraits<LDVSSAUpdater>.
class LDVSSAUpdater {
public:
/// Map of value numbers to PHI records.
DenseMap<BlockValueNum, LDVSSAPhi *> PHIs;
/// Map of which blocks generate Undef values -- blocks that are not
/// dominated by any Def.
DenseMap<MachineBasicBlock *, BlockValueNum> UndefMap;
/// Map of machine blocks to our own records of them.
DenseMap<MachineBasicBlock *, LDVSSABlock *> BlockMap;
/// Machine location where any PHI must occur.
LocIdx Loc;
/// Table of live-in machine value numbers for blocks / locations.
const ValueTable *MLiveIns;
LDVSSAUpdater(LocIdx L, const ValueTable *MLiveIns)
: Loc(L), MLiveIns(MLiveIns) {}
void reset() {
for (auto &Block : BlockMap)
delete Block.second;
PHIs.clear();
UndefMap.clear();
BlockMap.clear();
}
~LDVSSAUpdater() { reset(); }
/// For a given MBB, create a wrapper block for it. Stores it in the
/// LDVSSAUpdater block map.
LDVSSABlock *getSSALDVBlock(MachineBasicBlock *BB) {
auto it = BlockMap.find(BB);
if (it == BlockMap.end()) {
BlockMap[BB] = new LDVSSABlock(*BB, *this);
it = BlockMap.find(BB);
}
return it->second;
}
/// Find the live-in value number for the given block. Looks up the value at
/// the PHI location on entry.
BlockValueNum getValue(LDVSSABlock *LDVBB) {
return MLiveIns[LDVBB->BB.getNumber()][Loc.asU64()].asU64();
}
};
LDVSSABlock *LDVSSABlockIterator::operator*() {
return Updater.getSSALDVBlock(*PredIt);
}
#ifndef NDEBUG
raw_ostream &operator<<(raw_ostream &out, const LDVSSAPhi &PHI) {
out << "SSALDVPHI " << PHI.PHIValNum;
return out;
}
#endif
} // namespace
namespace llvm {
/// Template specialization to give SSAUpdater access to CFG and value
/// information. SSAUpdater calls methods in these traits, passing in the
/// LDVSSAUpdater object, to learn about blocks and the values they define.
/// It also provides methods to create PHI nodes and track them.
template <> class SSAUpdaterTraits<LDVSSAUpdater> {
public:
using BlkT = LDVSSABlock;
using ValT = BlockValueNum;
using PhiT = LDVSSAPhi;
using BlkSucc_iterator = LDVSSABlockIterator;
// Methods to access block successors -- dereferencing to our wrapper class.
static BlkSucc_iterator BlkSucc_begin(BlkT *BB) { return BB->succ_begin(); }
static BlkSucc_iterator BlkSucc_end(BlkT *BB) { return BB->succ_end(); }
/// Iterator for PHI operands.
class PHI_iterator {
private:
LDVSSAPhi *PHI;
unsigned Idx;
public:
explicit PHI_iterator(LDVSSAPhi *P) // begin iterator
: PHI(P), Idx(0) {}
PHI_iterator(LDVSSAPhi *P, bool) // end iterator
: PHI(P), Idx(PHI->IncomingValues.size()) {}
PHI_iterator &operator++() {
Idx++;
return *this;
}
bool operator==(const PHI_iterator &X) const { return Idx == X.Idx; }
bool operator!=(const PHI_iterator &X) const { return !operator==(X); }
BlockValueNum getIncomingValue() { return PHI->IncomingValues[Idx].second; }
LDVSSABlock *getIncomingBlock() { return PHI->IncomingValues[Idx].first; }
};
static inline PHI_iterator PHI_begin(PhiT *PHI) { return PHI_iterator(PHI); }
static inline PHI_iterator PHI_end(PhiT *PHI) {
return PHI_iterator(PHI, true);
}
/// FindPredecessorBlocks - Put the predecessors of BB into the Preds
/// vector.
static void FindPredecessorBlocks(LDVSSABlock *BB,
SmallVectorImpl<LDVSSABlock *> *Preds) {
for (MachineBasicBlock *Pred : BB->BB.predecessors())
Preds->push_back(BB->Updater.getSSALDVBlock(Pred));
}
/// GetUndefVal - Normally creates an IMPLICIT_DEF instruction with a new
/// register. For LiveDebugValues, represents a block identified as not having
/// any DBG_PHI predecessors.
static BlockValueNum GetUndefVal(LDVSSABlock *BB, LDVSSAUpdater *Updater) {
// Create a value number for this block -- it needs to be unique and in the
// "undef" collection, so that we know it's not real. Use a number
// representing a PHI into this block.
BlockValueNum Num = ValueIDNum(BB->BB.getNumber(), 0, Updater->Loc).asU64();
Updater->UndefMap[&BB->BB] = Num;
return Num;
}
/// CreateEmptyPHI - Create a (representation of a) PHI in the given block.
/// SSAUpdater will populate it with information about incoming values. The
/// value number of this PHI is whatever the machine value number problem
/// solution determined it to be. This includes non-phi values if SSAUpdater
/// tries to create a PHI where the incoming values are identical.
static BlockValueNum CreateEmptyPHI(LDVSSABlock *BB, unsigned NumPreds,
LDVSSAUpdater *Updater) {
BlockValueNum PHIValNum = Updater->getValue(BB);
LDVSSAPhi *PHI = BB->newPHI(PHIValNum);
Updater->PHIs[PHIValNum] = PHI;
return PHIValNum;
}
/// AddPHIOperand - Add the specified value as an operand of the PHI for
/// the specified predecessor block.
static void AddPHIOperand(LDVSSAPhi *PHI, BlockValueNum Val, LDVSSABlock *Pred) {
PHI->IncomingValues.push_back(std::make_pair(Pred, Val));
}
/// ValueIsPHI - Check if the instruction that defines the specified value
/// is a PHI instruction.
static LDVSSAPhi *ValueIsPHI(BlockValueNum Val, LDVSSAUpdater *Updater) {
auto PHIIt = Updater->PHIs.find(Val);
if (PHIIt == Updater->PHIs.end())
return nullptr;
return PHIIt->second;
}
/// ValueIsNewPHI - Like ValueIsPHI but also check if the PHI has no source
/// operands, i.e., it was just added.
static LDVSSAPhi *ValueIsNewPHI(BlockValueNum Val, LDVSSAUpdater *Updater) {
LDVSSAPhi *PHI = ValueIsPHI(Val, Updater);
if (PHI && PHI->IncomingValues.size() == 0)
return PHI;
return nullptr;
}
/// GetPHIValue - For the specified PHI instruction, return the value
/// that it defines.
static BlockValueNum GetPHIValue(LDVSSAPhi *PHI) { return PHI->PHIValNum; }
};
} // end namespace llvm
Optional<ValueIDNum> InstrRefBasedLDV::resolveDbgPHIs(
MachineFunction &MF, const ValueTable *MLiveOuts,
const ValueTable *MLiveIns, MachineInstr &Here, uint64_t InstrNum) {
assert(MLiveOuts && MLiveIns &&
"Tried to resolve DBG_PHI before location "
"tables allocated?");
// This function will be called twice per DBG_INSTR_REF, and might end up
// computing lots of SSA information: memoize it.
auto SeenDbgPHIIt = SeenDbgPHIs.find(&Here);
if (SeenDbgPHIIt != SeenDbgPHIs.end())
return SeenDbgPHIIt->second;
Optional<ValueIDNum> Result =
resolveDbgPHIsImpl(MF, MLiveOuts, MLiveIns, Here, InstrNum);
SeenDbgPHIs.insert({&Here, Result});
return Result;
}
Optional<ValueIDNum> InstrRefBasedLDV::resolveDbgPHIsImpl(
MachineFunction &MF, const ValueTable *MLiveOuts,
const ValueTable *MLiveIns, MachineInstr &Here, uint64_t InstrNum) {
// Pick out records of DBG_PHI instructions that have been observed. If there
// are none, then we cannot compute a value number.
auto RangePair = std::equal_range(DebugPHINumToValue.begin(),
DebugPHINumToValue.end(), InstrNum);
auto LowerIt = RangePair.first;
auto UpperIt = RangePair.second;
// No DBG_PHI means there can be no location.
if (LowerIt == UpperIt)
return None;
// If any DBG_PHIs referred to a location we didn't understand, don't try to
// compute a value. There might be scenarios where we could recover a value
// for some range of DBG_INSTR_REFs, but at this point we can have high
// confidence that we've seen a bug.
auto DBGPHIRange = make_range(LowerIt, UpperIt);
for (const DebugPHIRecord &DBG_PHI : DBGPHIRange)
if (!DBG_PHI.ValueRead)
return None;
// If there's only one DBG_PHI, then that is our value number.
if (std::distance(LowerIt, UpperIt) == 1)
return *LowerIt->ValueRead;
// Pick out the location (physreg, slot) where any PHIs must occur. It's
// technically possible for us to merge values in different registers in each
// block, but highly unlikely that LLVM will generate such code after register
// allocation.
LocIdx Loc = *LowerIt->ReadLoc;
// We have several DBG_PHIs, and a use position (the Here inst). All each
// DBG_PHI does is identify a value at a program position. We can treat each
// DBG_PHI like it's a Def of a value, and the use position is a Use of a
// value, just like SSA. We use the bulk-standard LLVM SSA updater class to
// determine which Def is used at the Use, and any PHIs that happen along
// the way.
// Adapted LLVM SSA Updater:
LDVSSAUpdater Updater(Loc, MLiveIns);
// Map of which Def or PHI is the current value in each block.
DenseMap<LDVSSABlock *, BlockValueNum> AvailableValues;
// Set of PHIs that we have created along the way.
SmallVector<LDVSSAPhi *, 8> CreatedPHIs;
// Each existing DBG_PHI is a Def'd value under this model. Record these Defs
// for the SSAUpdater.
for (const auto &DBG_PHI : DBGPHIRange) {
LDVSSABlock *Block = Updater.getSSALDVBlock(DBG_PHI.MBB);
const ValueIDNum &Num = *DBG_PHI.ValueRead;
AvailableValues.insert(std::make_pair(Block, Num.asU64()));
}
LDVSSABlock *HereBlock = Updater.getSSALDVBlock(Here.getParent());
const auto &AvailIt = AvailableValues.find(HereBlock);
if (AvailIt != AvailableValues.end()) {
// Actually, we already know what the value is -- the Use is in the same
// block as the Def.
return ValueIDNum::fromU64(AvailIt->second);
}
// Otherwise, we must use the SSA Updater. It will identify the value number
// that we are to use, and the PHIs that must happen along the way.
SSAUpdaterImpl<LDVSSAUpdater> Impl(&Updater, &AvailableValues, &CreatedPHIs);
BlockValueNum ResultInt = Impl.GetValue(Updater.getSSALDVBlock(Here.getParent()));
ValueIDNum Result = ValueIDNum::fromU64(ResultInt);
// We have the number for a PHI, or possibly live-through value, to be used
// at this Use. There are a number of things we have to check about it though:
// * Does any PHI use an 'Undef' (like an IMPLICIT_DEF) value? If so, this
// Use was not completely dominated by DBG_PHIs and we should abort.
// * Are the Defs or PHIs clobbered in a block? SSAUpdater isn't aware that
// we've left SSA form. Validate that the inputs to each PHI are the
// expected values.
// * Is a PHI we've created actually a merging of values, or are all the
// predecessor values the same, leading to a non-PHI machine value number?
// (SSAUpdater doesn't know that either). Remap validated PHIs into the
// the ValidatedValues collection below to sort this out.
DenseMap<LDVSSABlock *, ValueIDNum> ValidatedValues;
// Define all the input DBG_PHI values in ValidatedValues.
for (const auto &DBG_PHI : DBGPHIRange) {
LDVSSABlock *Block = Updater.getSSALDVBlock(DBG_PHI.MBB);
const ValueIDNum &Num = *DBG_PHI.ValueRead;
ValidatedValues.insert(std::make_pair(Block, Num));
}
// Sort PHIs to validate into RPO-order.
SmallVector<LDVSSAPhi *, 8> SortedPHIs;
for (auto &PHI : CreatedPHIs)
SortedPHIs.push_back(PHI);
std::sort(
SortedPHIs.begin(), SortedPHIs.end(), [&](LDVSSAPhi *A, LDVSSAPhi *B) {
return BBToOrder[&A->getParent()->BB] < BBToOrder[&B->getParent()->BB];
});
for (auto &PHI : SortedPHIs) {
ValueIDNum ThisBlockValueNum =
MLiveIns[PHI->ParentBlock->BB.getNumber()][Loc.asU64()];
// Are all these things actually defined?
for (auto &PHIIt : PHI->IncomingValues) {
// Any undef input means DBG_PHIs didn't dominate the use point.
if (Updater.UndefMap.find(&PHIIt.first->BB) != Updater.UndefMap.end())
return None;
ValueIDNum ValueToCheck;
const ValueTable &BlockLiveOuts = MLiveOuts[PHIIt.first->BB.getNumber()];
auto VVal = ValidatedValues.find(PHIIt.first);
if (VVal == ValidatedValues.end()) {
// We cross a loop, and this is a backedge. LLVMs tail duplication
// happens so late that DBG_PHI instructions should not be able to
// migrate into loops -- meaning we can only be live-through this
// loop.
ValueToCheck = ThisBlockValueNum;
} else {
// Does the block have as a live-out, in the location we're examining,
// the value that we expect? If not, it's been moved or clobbered.
ValueToCheck = VVal->second;
}
if (BlockLiveOuts[Loc.asU64()] != ValueToCheck)
return None;
}
// Record this value as validated.
ValidatedValues.insert({PHI->ParentBlock, ThisBlockValueNum});
}
// All the PHIs are valid: we can return what the SSAUpdater said our value
// number was.
return Result;
}