llvm-project/llvm/lib/CodeGen/LiveIntervalAnalysis.cpp

1593 lines
59 KiB
C++
Raw Normal View History

//===- LiveIntervalAnalysis.cpp - Live Interval Analysis ------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file This file implements the LiveInterval analysis pass which is used
/// by the Linear Scan Register allocator. This pass linearizes the
/// basic blocks of the function in DFS order and computes live intervals for
/// each virtual and physical register.
//
//===----------------------------------------------------------------------===//
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "LiveRangeCalc.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/iterator_range.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/LiveInterval.h"
#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBundle.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/SlotIndexes.h"
#include "llvm/CodeGen/VirtRegMap.h"
#include "llvm/MC/LaneBitmask.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/Pass.h"
#include "llvm/Support/BlockFrequency.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetSubtargetInfo.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <iterator>
#include <tuple>
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "regalloc"
2007-05-03 09:11:54 +08:00
char LiveIntervals::ID = 0;
char &llvm::LiveIntervalsID = LiveIntervals::ID;
INITIALIZE_PASS_BEGIN(LiveIntervals, "liveintervals",
"Live Interval Analysis", false, false)
[PM/AA] Rebuild LLVM's alias analysis infrastructure in a way compatible with the new pass manager, and no longer relying on analysis groups. This builds essentially a ground-up new AA infrastructure stack for LLVM. The core ideas are the same that are used throughout the new pass manager: type erased polymorphism and direct composition. The design is as follows: - FunctionAAResults is a type-erasing alias analysis results aggregation interface to walk a single query across a range of results from different alias analyses. Currently this is function-specific as we always assume that aliasing queries are *within* a function. - AAResultBase is a CRTP utility providing stub implementations of various parts of the alias analysis result concept, notably in several cases in terms of other more general parts of the interface. This can be used to implement only a narrow part of the interface rather than the entire interface. This isn't really ideal, this logic should be hoisted into FunctionAAResults as currently it will cause a significant amount of redundant work, but it faithfully models the behavior of the prior infrastructure. - All the alias analysis passes are ported to be wrapper passes for the legacy PM and new-style analysis passes for the new PM with a shared result object. In some cases (most notably CFL), this is an extremely naive approach that we should revisit when we can specialize for the new pass manager. - BasicAA has been restructured to reflect that it is much more fundamentally a function analysis because it uses dominator trees and loop info that need to be constructed for each function. All of the references to getting alias analysis results have been updated to use the new aggregation interface. All the preservation and other pass management code has been updated accordingly. The way the FunctionAAResultsWrapperPass works is to detect the available alias analyses when run, and add them to the results object. This means that we should be able to continue to respect when various passes are added to the pipeline, for example adding CFL or adding TBAA passes should just cause their results to be available and to get folded into this. The exception to this rule is BasicAA which really needs to be a function pass due to using dominator trees and loop info. As a consequence, the FunctionAAResultsWrapperPass directly depends on BasicAA and always includes it in the aggregation. This has significant implications for preserving analyses. Generally, most passes shouldn't bother preserving FunctionAAResultsWrapperPass because rebuilding the results just updates the set of known AA passes. The exception to this rule are LoopPass instances which need to preserve all the function analyses that the loop pass manager will end up needing. This means preserving both BasicAAWrapperPass and the aggregating FunctionAAResultsWrapperPass. Now, when preserving an alias analysis, you do so by directly preserving that analysis. This is only necessary for non-immutable-pass-provided alias analyses though, and there are only three of interest: BasicAA, GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is preserved when needed because it (like DominatorTree and LoopInfo) is marked as a CFG-only pass. I've expanded GlobalsAA into the preserved set everywhere we previously were preserving all of AliasAnalysis, and I've added SCEVAA in the intersection of that with where we preserve SCEV itself. One significant challenge to all of this is that the CGSCC passes were actually using the alias analysis implementations by taking advantage of a pretty amazing set of loop holes in the old pass manager's analysis management code which allowed analysis groups to slide through in many cases. Moving away from analysis groups makes this problem much more obvious. To fix it, I've leveraged the flexibility the design of the new PM components provides to just directly construct the relevant alias analyses for the relevant functions in the IPO passes that need them. This is a bit hacky, but should go away with the new pass manager, and is already in many ways cleaner than the prior state. Another significant challenge is that various facilities of the old alias analysis infrastructure just don't fit any more. The most significant of these is the alias analysis 'counter' pass. That pass relied on the ability to snoop on AA queries at different points in the analysis group chain. Instead, I'm planning to build printing functionality directly into the aggregation layer. I've not included that in this patch merely to keep it smaller. Note that all of this needs a nearly complete rewrite of the AA documentation. I'm planning to do that, but I'd like to make sure the new design settles, and to flesh out a bit more of what it looks like in the new pass manager first. Differential Revision: http://reviews.llvm.org/D12080 llvm-svn: 247167
2015-09-10 01:55:00 +08:00
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
INITIALIZE_PASS_END(LiveIntervals, "liveintervals",
"Live Interval Analysis", false, false)
#ifndef NDEBUG
static cl::opt<bool> EnablePrecomputePhysRegs(
"precompute-phys-liveness", cl::Hidden,
cl::desc("Eagerly compute live intervals for all physreg units."));
#else
static bool EnablePrecomputePhysRegs = false;
#endif // NDEBUG
[LiveIntervalAnalysis] Speed up creation of live ranges for physical registers by using a segment set. The patch addresses a compile-time performance regression in the LiveIntervals analysis pass (see http://llvm.org/bugs/show_bug.cgi?id=18580). This regression is especially critical when compiling long functions. Our analysis had shown that the most of time is taken for generation of live intervals for physical registers. Insertions in the middle of the array of live ranges cause quadratic algorithmic complexity, which is apparently the main reason for the slow-down. Overview of changes: - The patch introduces an additional std::set<Segment>* member in LiveRange for storing segments in the phase of initial creation. The set is used if this member is not NULL, otherwise everything works the old way. - The set of operations on LiveRange used during initial creation (i.e. used by createDeadDefs and extendToUses) have been reimplemented to use the segment set if it is available. - After a live range is created the contents of the set are flushed to the segment vector, because the set is not as efficient as the vector for the later uses of the live range. After the flushing, the set is deleted and cannot be used again. - The set is only for live ranges computed in LiveIntervalAnalysis::computeLiveInRegUnits() and getRegUnit() but not in computeVirtRegs(), because I did not bring any performance benefits to computeVirtRegs() and for some examples even brought a slow down. Patch by Vaidas Gasiunas <vaidas.gasiunas@sap.com> Differential Revision: http://reviews.llvm.org/D6013 llvm-svn: 228421
2015-02-07 02:42:41 +08:00
namespace llvm {
[LiveIntervalAnalysis] Speed up creation of live ranges for physical registers by using a segment set. The patch addresses a compile-time performance regression in the LiveIntervals analysis pass (see http://llvm.org/bugs/show_bug.cgi?id=18580). This regression is especially critical when compiling long functions. Our analysis had shown that the most of time is taken for generation of live intervals for physical registers. Insertions in the middle of the array of live ranges cause quadratic algorithmic complexity, which is apparently the main reason for the slow-down. Overview of changes: - The patch introduces an additional std::set<Segment>* member in LiveRange for storing segments in the phase of initial creation. The set is used if this member is not NULL, otherwise everything works the old way. - The set of operations on LiveRange used during initial creation (i.e. used by createDeadDefs and extendToUses) have been reimplemented to use the segment set if it is available. - After a live range is created the contents of the set are flushed to the segment vector, because the set is not as efficient as the vector for the later uses of the live range. After the flushing, the set is deleted and cannot be used again. - The set is only for live ranges computed in LiveIntervalAnalysis::computeLiveInRegUnits() and getRegUnit() but not in computeVirtRegs(), because I did not bring any performance benefits to computeVirtRegs() and for some examples even brought a slow down. Patch by Vaidas Gasiunas <vaidas.gasiunas@sap.com> Differential Revision: http://reviews.llvm.org/D6013 llvm-svn: 228421
2015-02-07 02:42:41 +08:00
cl::opt<bool> UseSegmentSetForPhysRegs(
"use-segment-set-for-physregs", cl::Hidden, cl::init(true),
cl::desc(
"Use segment set for the computation of the live ranges of physregs."));
} // end namespace llvm
[LiveIntervalAnalysis] Speed up creation of live ranges for physical registers by using a segment set. The patch addresses a compile-time performance regression in the LiveIntervals analysis pass (see http://llvm.org/bugs/show_bug.cgi?id=18580). This regression is especially critical when compiling long functions. Our analysis had shown that the most of time is taken for generation of live intervals for physical registers. Insertions in the middle of the array of live ranges cause quadratic algorithmic complexity, which is apparently the main reason for the slow-down. Overview of changes: - The patch introduces an additional std::set<Segment>* member in LiveRange for storing segments in the phase of initial creation. The set is used if this member is not NULL, otherwise everything works the old way. - The set of operations on LiveRange used during initial creation (i.e. used by createDeadDefs and extendToUses) have been reimplemented to use the segment set if it is available. - After a live range is created the contents of the set are flushed to the segment vector, because the set is not as efficient as the vector for the later uses of the live range. After the flushing, the set is deleted and cannot be used again. - The set is only for live ranges computed in LiveIntervalAnalysis::computeLiveInRegUnits() and getRegUnit() but not in computeVirtRegs(), because I did not bring any performance benefits to computeVirtRegs() and for some examples even brought a slow down. Patch by Vaidas Gasiunas <vaidas.gasiunas@sap.com> Differential Revision: http://reviews.llvm.org/D6013 llvm-svn: 228421
2015-02-07 02:42:41 +08:00
void LiveIntervals::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
[PM/AA] Rebuild LLVM's alias analysis infrastructure in a way compatible with the new pass manager, and no longer relying on analysis groups. This builds essentially a ground-up new AA infrastructure stack for LLVM. The core ideas are the same that are used throughout the new pass manager: type erased polymorphism and direct composition. The design is as follows: - FunctionAAResults is a type-erasing alias analysis results aggregation interface to walk a single query across a range of results from different alias analyses. Currently this is function-specific as we always assume that aliasing queries are *within* a function. - AAResultBase is a CRTP utility providing stub implementations of various parts of the alias analysis result concept, notably in several cases in terms of other more general parts of the interface. This can be used to implement only a narrow part of the interface rather than the entire interface. This isn't really ideal, this logic should be hoisted into FunctionAAResults as currently it will cause a significant amount of redundant work, but it faithfully models the behavior of the prior infrastructure. - All the alias analysis passes are ported to be wrapper passes for the legacy PM and new-style analysis passes for the new PM with a shared result object. In some cases (most notably CFL), this is an extremely naive approach that we should revisit when we can specialize for the new pass manager. - BasicAA has been restructured to reflect that it is much more fundamentally a function analysis because it uses dominator trees and loop info that need to be constructed for each function. All of the references to getting alias analysis results have been updated to use the new aggregation interface. All the preservation and other pass management code has been updated accordingly. The way the FunctionAAResultsWrapperPass works is to detect the available alias analyses when run, and add them to the results object. This means that we should be able to continue to respect when various passes are added to the pipeline, for example adding CFL or adding TBAA passes should just cause their results to be available and to get folded into this. The exception to this rule is BasicAA which really needs to be a function pass due to using dominator trees and loop info. As a consequence, the FunctionAAResultsWrapperPass directly depends on BasicAA and always includes it in the aggregation. This has significant implications for preserving analyses. Generally, most passes shouldn't bother preserving FunctionAAResultsWrapperPass because rebuilding the results just updates the set of known AA passes. The exception to this rule are LoopPass instances which need to preserve all the function analyses that the loop pass manager will end up needing. This means preserving both BasicAAWrapperPass and the aggregating FunctionAAResultsWrapperPass. Now, when preserving an alias analysis, you do so by directly preserving that analysis. This is only necessary for non-immutable-pass-provided alias analyses though, and there are only three of interest: BasicAA, GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is preserved when needed because it (like DominatorTree and LoopInfo) is marked as a CFG-only pass. I've expanded GlobalsAA into the preserved set everywhere we previously were preserving all of AliasAnalysis, and I've added SCEVAA in the intersection of that with where we preserve SCEV itself. One significant challenge to all of this is that the CGSCC passes were actually using the alias analysis implementations by taking advantage of a pretty amazing set of loop holes in the old pass manager's analysis management code which allowed analysis groups to slide through in many cases. Moving away from analysis groups makes this problem much more obvious. To fix it, I've leveraged the flexibility the design of the new PM components provides to just directly construct the relevant alias analyses for the relevant functions in the IPO passes that need them. This is a bit hacky, but should go away with the new pass manager, and is already in many ways cleaner than the prior state. Another significant challenge is that various facilities of the old alias analysis infrastructure just don't fit any more. The most significant of these is the alias analysis 'counter' pass. That pass relied on the ability to snoop on AA queries at different points in the analysis group chain. Instead, I'm planning to build printing functionality directly into the aggregation layer. I've not included that in this patch merely to keep it smaller. Note that all of this needs a nearly complete rewrite of the AA documentation. I'm planning to do that, but I'd like to make sure the new design settles, and to flesh out a bit more of what it looks like in the new pass manager first. Differential Revision: http://reviews.llvm.org/D12080 llvm-svn: 247167
2015-09-10 01:55:00 +08:00
AU.addRequired<AAResultsWrapperPass>();
AU.addPreserved<AAResultsWrapperPass>();
AU.addPreserved<LiveVariables>();
AU.addPreservedID(MachineLoopInfoID);
AU.addRequiredTransitiveID(MachineDominatorsID);
AU.addPreservedID(MachineDominatorsID);
AU.addPreserved<SlotIndexes>();
AU.addRequiredTransitive<SlotIndexes>();
MachineFunctionPass::getAnalysisUsage(AU);
}
LiveIntervals::LiveIntervals() : MachineFunctionPass(ID) {
initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
}
LiveIntervals::~LiveIntervals() {
delete LRCalc;
}
void LiveIntervals::releaseMemory() {
// Free the live intervals themselves.
for (unsigned i = 0, e = VirtRegIntervals.size(); i != e; ++i)
delete VirtRegIntervals[TargetRegisterInfo::index2VirtReg(i)];
VirtRegIntervals.clear();
RegMaskSlots.clear();
RegMaskBits.clear();
RegMaskBlocks.clear();
for (LiveRange *LR : RegUnitRanges)
delete LR;
RegUnitRanges.clear();
// Release VNInfo memory regions, VNInfo objects don't need to be dtor'd.
VNInfoAllocator.Reset();
}
bool LiveIntervals::runOnMachineFunction(MachineFunction &fn) {
MF = &fn;
MRI = &MF->getRegInfo();
TRI = MF->getSubtarget().getRegisterInfo();
TII = MF->getSubtarget().getInstrInfo();
[PM/AA] Rebuild LLVM's alias analysis infrastructure in a way compatible with the new pass manager, and no longer relying on analysis groups. This builds essentially a ground-up new AA infrastructure stack for LLVM. The core ideas are the same that are used throughout the new pass manager: type erased polymorphism and direct composition. The design is as follows: - FunctionAAResults is a type-erasing alias analysis results aggregation interface to walk a single query across a range of results from different alias analyses. Currently this is function-specific as we always assume that aliasing queries are *within* a function. - AAResultBase is a CRTP utility providing stub implementations of various parts of the alias analysis result concept, notably in several cases in terms of other more general parts of the interface. This can be used to implement only a narrow part of the interface rather than the entire interface. This isn't really ideal, this logic should be hoisted into FunctionAAResults as currently it will cause a significant amount of redundant work, but it faithfully models the behavior of the prior infrastructure. - All the alias analysis passes are ported to be wrapper passes for the legacy PM and new-style analysis passes for the new PM with a shared result object. In some cases (most notably CFL), this is an extremely naive approach that we should revisit when we can specialize for the new pass manager. - BasicAA has been restructured to reflect that it is much more fundamentally a function analysis because it uses dominator trees and loop info that need to be constructed for each function. All of the references to getting alias analysis results have been updated to use the new aggregation interface. All the preservation and other pass management code has been updated accordingly. The way the FunctionAAResultsWrapperPass works is to detect the available alias analyses when run, and add them to the results object. This means that we should be able to continue to respect when various passes are added to the pipeline, for example adding CFL or adding TBAA passes should just cause their results to be available and to get folded into this. The exception to this rule is BasicAA which really needs to be a function pass due to using dominator trees and loop info. As a consequence, the FunctionAAResultsWrapperPass directly depends on BasicAA and always includes it in the aggregation. This has significant implications for preserving analyses. Generally, most passes shouldn't bother preserving FunctionAAResultsWrapperPass because rebuilding the results just updates the set of known AA passes. The exception to this rule are LoopPass instances which need to preserve all the function analyses that the loop pass manager will end up needing. This means preserving both BasicAAWrapperPass and the aggregating FunctionAAResultsWrapperPass. Now, when preserving an alias analysis, you do so by directly preserving that analysis. This is only necessary for non-immutable-pass-provided alias analyses though, and there are only three of interest: BasicAA, GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is preserved when needed because it (like DominatorTree and LoopInfo) is marked as a CFG-only pass. I've expanded GlobalsAA into the preserved set everywhere we previously were preserving all of AliasAnalysis, and I've added SCEVAA in the intersection of that with where we preserve SCEV itself. One significant challenge to all of this is that the CGSCC passes were actually using the alias analysis implementations by taking advantage of a pretty amazing set of loop holes in the old pass manager's analysis management code which allowed analysis groups to slide through in many cases. Moving away from analysis groups makes this problem much more obvious. To fix it, I've leveraged the flexibility the design of the new PM components provides to just directly construct the relevant alias analyses for the relevant functions in the IPO passes that need them. This is a bit hacky, but should go away with the new pass manager, and is already in many ways cleaner than the prior state. Another significant challenge is that various facilities of the old alias analysis infrastructure just don't fit any more. The most significant of these is the alias analysis 'counter' pass. That pass relied on the ability to snoop on AA queries at different points in the analysis group chain. Instead, I'm planning to build printing functionality directly into the aggregation layer. I've not included that in this patch merely to keep it smaller. Note that all of this needs a nearly complete rewrite of the AA documentation. I'm planning to do that, but I'd like to make sure the new design settles, and to flesh out a bit more of what it looks like in the new pass manager first. Differential Revision: http://reviews.llvm.org/D12080 llvm-svn: 247167
2015-09-10 01:55:00 +08:00
AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
Indexes = &getAnalysis<SlotIndexes>();
DomTree = &getAnalysis<MachineDominatorTree>();
if (!LRCalc)
LRCalc = new LiveRangeCalc();
// Allocate space for all virtual registers.
VirtRegIntervals.resize(MRI->getNumVirtRegs());
computeVirtRegs();
computeRegMasks();
computeLiveInRegUnits();
if (EnablePrecomputePhysRegs) {
// For stress testing, precompute live ranges of all physical register
// units, including reserved registers.
for (unsigned i = 0, e = TRI->getNumRegUnits(); i != e; ++i)
getRegUnit(i);
}
DEBUG(dump());
return true;
}
void LiveIntervals::print(raw_ostream &OS, const Module* ) const {
OS << "********** INTERVALS **********\n";
// Dump the regunits.
for (unsigned Unit = 0, UnitE = RegUnitRanges.size(); Unit != UnitE; ++Unit)
if (LiveRange *LR = RegUnitRanges[Unit])
OS << PrintRegUnit(Unit, TRI) << ' ' << *LR << '\n';
// Dump the virtregs.
for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
if (hasInterval(Reg))
OS << getInterval(Reg) << '\n';
}
OS << "RegMasks:";
for (SlotIndex Idx : RegMaskSlots)
OS << ' ' << Idx;
OS << '\n';
printInstrs(OS);
}
void LiveIntervals::printInstrs(raw_ostream &OS) const {
OS << "********** MACHINEINSTRS **********\n";
MF->print(OS, Indexes);
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void LiveIntervals::dumpInstrs() const {
printInstrs(dbgs());
}
#endif
LiveInterval* LiveIntervals::createInterval(unsigned reg) {
float Weight = TargetRegisterInfo::isPhysicalRegister(reg) ? huge_valf : 0.0F;
return new LiveInterval(reg, Weight);
}
2007-11-12 14:35:08 +08:00
/// Compute the live interval of a virtual register, based on defs and uses.
void LiveIntervals::computeVirtRegInterval(LiveInterval &LI) {
assert(LRCalc && "LRCalc not initialized.");
assert(LI.empty() && "Should only compute empty intervals.");
LRCalc->reset(MF, getSlotIndexes(), DomTree, &getVNInfoAllocator());
LRCalc->calculate(LI, MRI->shouldTrackSubRegLiveness(LI.reg));
computeDeadValues(LI, nullptr);
}
void LiveIntervals::computeVirtRegs() {
for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
if (MRI->reg_nodbg_empty(Reg))
continue;
createAndComputeVirtRegInterval(Reg);
}
}
void LiveIntervals::computeRegMasks() {
RegMaskBlocks.resize(MF->getNumBlockIDs());
// Find all instructions with regmask operands.
for (const MachineBasicBlock &MBB : *MF) {
std::pair<unsigned, unsigned> &RMB = RegMaskBlocks[MBB.getNumber()];
RMB.first = RegMaskSlots.size();
// Some block starts, such as EH funclets, create masks.
if (const uint32_t *Mask = MBB.getBeginClobberMask(TRI)) {
RegMaskSlots.push_back(Indexes->getMBBStartIdx(&MBB));
RegMaskBits.push_back(Mask);
}
for (const MachineInstr &MI : MBB) {
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isRegMask())
continue;
RegMaskSlots.push_back(Indexes->getInstructionIndex(MI).getRegSlot());
RegMaskBits.push_back(MO.getRegMask());
}
}
// Some block ends, such as funclet returns, create masks. Put the mask on
// the last instruction of the block, because MBB slot index intervals are
// half-open.
if (const uint32_t *Mask = MBB.getEndClobberMask(TRI)) {
assert(!MBB.empty() && "empty return block?");
RegMaskSlots.push_back(
Indexes->getInstructionIndex(MBB.back()).getRegSlot());
RegMaskBits.push_back(Mask);
}
// Compute the number of register mask instructions in this block.
RMB.second = RegMaskSlots.size() - RMB.first;
}
}
//===----------------------------------------------------------------------===//
// Register Unit Liveness
//===----------------------------------------------------------------------===//
//
// Fixed interference typically comes from ABI boundaries: Function arguments
// and return values are passed in fixed registers, and so are exception
// pointers entering landing pads. Certain instructions require values to be
// present in specific registers. That is also represented through fixed
// interference.
//
/// Compute the live range of a register unit, based on the uses and defs of
/// aliasing registers. The range should be empty, or contain only dead
/// phi-defs from ABI blocks.
void LiveIntervals::computeRegUnitRange(LiveRange &LR, unsigned Unit) {
assert(LRCalc && "LRCalc not initialized.");
LRCalc->reset(MF, getSlotIndexes(), DomTree, &getVNInfoAllocator());
// The physregs aliasing Unit are the roots and their super-registers.
// Create all values as dead defs before extending to uses. Note that roots
// may share super-registers. That's OK because createDeadDefs() is
// idempotent. It is very rare for a register unit to have multiple roots, so
// uniquing super-registers is probably not worthwhile.
LiveIntervalAnalysis: Fix alias regunit reserved definition A register in CodeGen can be marked as reserved: In that case we consider the register always live and do not use (or rather ignore) kill/dead/undef operand flags. LiveIntervalAnalysis however tracks liveness per register unit (not per register). We already needed adjustments for this in r292871 to deal with super/sub registers. However I did not look at aliased register there. Looking at ARM: FPSCR (regunits FPSCR, FPSCR~FPSCR_NZCV) aliases with FPSCR_NZCV (regunits FPSCR_NZCV, FPSCR~FPSCR_NZCV) hence they share a register unit (FPSCR~FPSCR_NZCV) that represents the aliased parts of the registers. This shared register unit was previously considered non-reserved, however given that we uses of the reserved FPSCR potentially violate some rules (like uses without defs) we should make FPSCR~FPSCR_NZCV reserved too and stop tracking liveness for it. This patch: - Defines a register unit as reserved when: At least for one root register, the root register and all its super registers are reserved. - Adjust LiveIntervals::computeRegUnitRange() for new reserved definition. - Add MachineRegisterInfo::isReservedRegUnit() to have a canonical way of testing. - Stop computing LiveRanges for reserved register units in HMEditor even with UpdateFlags enabled. - Skip verification of uses of reserved reg units in the machine verifier (this usually didn't happen because there would be no cached liverange but there is no guarantee for that and I would run into this case before the HMEditor tweak, so may as well fix the verifier too). Note that this should only affect ARMs FPSCR/FPSCR_NZCV registers today; aliased registers are rarely used, the only other cases are hexagons P0-P3/P3_0 and C8/USR pairs which are not mixing reserved/non-reserved registers in an alias. Differential Revision: https://reviews.llvm.org/D37356 llvm-svn: 312348
2017-09-02 02:36:26 +08:00
bool IsReserved = false;
for (MCRegUnitRootIterator Root(Unit, TRI); Root.isValid(); ++Root) {
LiveIntervalAnalysis: Fix alias regunit reserved definition A register in CodeGen can be marked as reserved: In that case we consider the register always live and do not use (or rather ignore) kill/dead/undef operand flags. LiveIntervalAnalysis however tracks liveness per register unit (not per register). We already needed adjustments for this in r292871 to deal with super/sub registers. However I did not look at aliased register there. Looking at ARM: FPSCR (regunits FPSCR, FPSCR~FPSCR_NZCV) aliases with FPSCR_NZCV (regunits FPSCR_NZCV, FPSCR~FPSCR_NZCV) hence they share a register unit (FPSCR~FPSCR_NZCV) that represents the aliased parts of the registers. This shared register unit was previously considered non-reserved, however given that we uses of the reserved FPSCR potentially violate some rules (like uses without defs) we should make FPSCR~FPSCR_NZCV reserved too and stop tracking liveness for it. This patch: - Defines a register unit as reserved when: At least for one root register, the root register and all its super registers are reserved. - Adjust LiveIntervals::computeRegUnitRange() for new reserved definition. - Add MachineRegisterInfo::isReservedRegUnit() to have a canonical way of testing. - Stop computing LiveRanges for reserved register units in HMEditor even with UpdateFlags enabled. - Skip verification of uses of reserved reg units in the machine verifier (this usually didn't happen because there would be no cached liverange but there is no guarantee for that and I would run into this case before the HMEditor tweak, so may as well fix the verifier too). Note that this should only affect ARMs FPSCR/FPSCR_NZCV registers today; aliased registers are rarely used, the only other cases are hexagons P0-P3/P3_0 and C8/USR pairs which are not mixing reserved/non-reserved registers in an alias. Differential Revision: https://reviews.llvm.org/D37356 llvm-svn: 312348
2017-09-02 02:36:26 +08:00
bool IsRootReserved = true;
for (MCSuperRegIterator Super(*Root, TRI, /*IncludeSelf=*/true);
Super.isValid(); ++Super) {
unsigned Reg = *Super;
if (!MRI->reg_empty(Reg))
LRCalc->createDeadDefs(LR, Reg);
// A register unit is considered reserved if all its roots and all their
// super registers are reserved.
if (!MRI->isReserved(Reg))
LiveIntervalAnalysis: Fix alias regunit reserved definition A register in CodeGen can be marked as reserved: In that case we consider the register always live and do not use (or rather ignore) kill/dead/undef operand flags. LiveIntervalAnalysis however tracks liveness per register unit (not per register). We already needed adjustments for this in r292871 to deal with super/sub registers. However I did not look at aliased register there. Looking at ARM: FPSCR (regunits FPSCR, FPSCR~FPSCR_NZCV) aliases with FPSCR_NZCV (regunits FPSCR_NZCV, FPSCR~FPSCR_NZCV) hence they share a register unit (FPSCR~FPSCR_NZCV) that represents the aliased parts of the registers. This shared register unit was previously considered non-reserved, however given that we uses of the reserved FPSCR potentially violate some rules (like uses without defs) we should make FPSCR~FPSCR_NZCV reserved too and stop tracking liveness for it. This patch: - Defines a register unit as reserved when: At least for one root register, the root register and all its super registers are reserved. - Adjust LiveIntervals::computeRegUnitRange() for new reserved definition. - Add MachineRegisterInfo::isReservedRegUnit() to have a canonical way of testing. - Stop computing LiveRanges for reserved register units in HMEditor even with UpdateFlags enabled. - Skip verification of uses of reserved reg units in the machine verifier (this usually didn't happen because there would be no cached liverange but there is no guarantee for that and I would run into this case before the HMEditor tweak, so may as well fix the verifier too). Note that this should only affect ARMs FPSCR/FPSCR_NZCV registers today; aliased registers are rarely used, the only other cases are hexagons P0-P3/P3_0 and C8/USR pairs which are not mixing reserved/non-reserved registers in an alias. Differential Revision: https://reviews.llvm.org/D37356 llvm-svn: 312348
2017-09-02 02:36:26 +08:00
IsRootReserved = false;
}
LiveIntervalAnalysis: Fix alias regunit reserved definition A register in CodeGen can be marked as reserved: In that case we consider the register always live and do not use (or rather ignore) kill/dead/undef operand flags. LiveIntervalAnalysis however tracks liveness per register unit (not per register). We already needed adjustments for this in r292871 to deal with super/sub registers. However I did not look at aliased register there. Looking at ARM: FPSCR (regunits FPSCR, FPSCR~FPSCR_NZCV) aliases with FPSCR_NZCV (regunits FPSCR_NZCV, FPSCR~FPSCR_NZCV) hence they share a register unit (FPSCR~FPSCR_NZCV) that represents the aliased parts of the registers. This shared register unit was previously considered non-reserved, however given that we uses of the reserved FPSCR potentially violate some rules (like uses without defs) we should make FPSCR~FPSCR_NZCV reserved too and stop tracking liveness for it. This patch: - Defines a register unit as reserved when: At least for one root register, the root register and all its super registers are reserved. - Adjust LiveIntervals::computeRegUnitRange() for new reserved definition. - Add MachineRegisterInfo::isReservedRegUnit() to have a canonical way of testing. - Stop computing LiveRanges for reserved register units in HMEditor even with UpdateFlags enabled. - Skip verification of uses of reserved reg units in the machine verifier (this usually didn't happen because there would be no cached liverange but there is no guarantee for that and I would run into this case before the HMEditor tweak, so may as well fix the verifier too). Note that this should only affect ARMs FPSCR/FPSCR_NZCV registers today; aliased registers are rarely used, the only other cases are hexagons P0-P3/P3_0 and C8/USR pairs which are not mixing reserved/non-reserved registers in an alias. Differential Revision: https://reviews.llvm.org/D37356 llvm-svn: 312348
2017-09-02 02:36:26 +08:00
IsReserved |= IsRootReserved;
}
LiveIntervalAnalysis: Fix alias regunit reserved definition A register in CodeGen can be marked as reserved: In that case we consider the register always live and do not use (or rather ignore) kill/dead/undef operand flags. LiveIntervalAnalysis however tracks liveness per register unit (not per register). We already needed adjustments for this in r292871 to deal with super/sub registers. However I did not look at aliased register there. Looking at ARM: FPSCR (regunits FPSCR, FPSCR~FPSCR_NZCV) aliases with FPSCR_NZCV (regunits FPSCR_NZCV, FPSCR~FPSCR_NZCV) hence they share a register unit (FPSCR~FPSCR_NZCV) that represents the aliased parts of the registers. This shared register unit was previously considered non-reserved, however given that we uses of the reserved FPSCR potentially violate some rules (like uses without defs) we should make FPSCR~FPSCR_NZCV reserved too and stop tracking liveness for it. This patch: - Defines a register unit as reserved when: At least for one root register, the root register and all its super registers are reserved. - Adjust LiveIntervals::computeRegUnitRange() for new reserved definition. - Add MachineRegisterInfo::isReservedRegUnit() to have a canonical way of testing. - Stop computing LiveRanges for reserved register units in HMEditor even with UpdateFlags enabled. - Skip verification of uses of reserved reg units in the machine verifier (this usually didn't happen because there would be no cached liverange but there is no guarantee for that and I would run into this case before the HMEditor tweak, so may as well fix the verifier too). Note that this should only affect ARMs FPSCR/FPSCR_NZCV registers today; aliased registers are rarely used, the only other cases are hexagons P0-P3/P3_0 and C8/USR pairs which are not mixing reserved/non-reserved registers in an alias. Differential Revision: https://reviews.llvm.org/D37356 llvm-svn: 312348
2017-09-02 02:36:26 +08:00
assert(IsReserved == MRI->isReservedRegUnit(Unit) &&
"reserved computation mismatch");
// Now extend LR to reach all uses.
// Ignore uses of reserved registers. We only track defs of those.
if (!IsReserved) {
for (MCRegUnitRootIterator Root(Unit, TRI); Root.isValid(); ++Root) {
for (MCSuperRegIterator Super(*Root, TRI, /*IncludeSelf=*/true);
Super.isValid(); ++Super) {
unsigned Reg = *Super;
if (!MRI->reg_empty(Reg))
LRCalc->extendToUses(LR, Reg);
}
}
}
[LiveIntervalAnalysis] Speed up creation of live ranges for physical registers by using a segment set. The patch addresses a compile-time performance regression in the LiveIntervals analysis pass (see http://llvm.org/bugs/show_bug.cgi?id=18580). This regression is especially critical when compiling long functions. Our analysis had shown that the most of time is taken for generation of live intervals for physical registers. Insertions in the middle of the array of live ranges cause quadratic algorithmic complexity, which is apparently the main reason for the slow-down. Overview of changes: - The patch introduces an additional std::set<Segment>* member in LiveRange for storing segments in the phase of initial creation. The set is used if this member is not NULL, otherwise everything works the old way. - The set of operations on LiveRange used during initial creation (i.e. used by createDeadDefs and extendToUses) have been reimplemented to use the segment set if it is available. - After a live range is created the contents of the set are flushed to the segment vector, because the set is not as efficient as the vector for the later uses of the live range. After the flushing, the set is deleted and cannot be used again. - The set is only for live ranges computed in LiveIntervalAnalysis::computeLiveInRegUnits() and getRegUnit() but not in computeVirtRegs(), because I did not bring any performance benefits to computeVirtRegs() and for some examples even brought a slow down. Patch by Vaidas Gasiunas <vaidas.gasiunas@sap.com> Differential Revision: http://reviews.llvm.org/D6013 llvm-svn: 228421
2015-02-07 02:42:41 +08:00
// Flush the segment set to the segment vector.
if (UseSegmentSetForPhysRegs)
LR.flushSegmentSet();
}
/// Precompute the live ranges of any register units that are live-in to an ABI
/// block somewhere. Register values can appear without a corresponding def when
/// entering the entry block or a landing pad.
void LiveIntervals::computeLiveInRegUnits() {
RegUnitRanges.resize(TRI->getNumRegUnits());
DEBUG(dbgs() << "Computing live-in reg-units in ABI blocks.\n");
// Keep track of the live range sets allocated.
SmallVector<unsigned, 8> NewRanges;
// Check all basic blocks for live-ins.
for (const MachineBasicBlock &MBB : *MF) {
// We only care about ABI blocks: Entry + landing pads.
if ((&MBB != &MF->front() && !MBB.isEHPad()) || MBB.livein_empty())
continue;
// Create phi-defs at Begin for all live-in registers.
SlotIndex Begin = Indexes->getMBBStartIdx(&MBB);
DEBUG(dbgs() << Begin << "\tBB#" << MBB.getNumber());
for (const auto &LI : MBB.liveins()) {
for (MCRegUnitIterator Units(LI.PhysReg, TRI); Units.isValid(); ++Units) {
unsigned Unit = *Units;
LiveRange *LR = RegUnitRanges[Unit];
if (!LR) {
[LiveIntervalAnalysis] Speed up creation of live ranges for physical registers by using a segment set. The patch addresses a compile-time performance regression in the LiveIntervals analysis pass (see http://llvm.org/bugs/show_bug.cgi?id=18580). This regression is especially critical when compiling long functions. Our analysis had shown that the most of time is taken for generation of live intervals for physical registers. Insertions in the middle of the array of live ranges cause quadratic algorithmic complexity, which is apparently the main reason for the slow-down. Overview of changes: - The patch introduces an additional std::set<Segment>* member in LiveRange for storing segments in the phase of initial creation. The set is used if this member is not NULL, otherwise everything works the old way. - The set of operations on LiveRange used during initial creation (i.e. used by createDeadDefs and extendToUses) have been reimplemented to use the segment set if it is available. - After a live range is created the contents of the set are flushed to the segment vector, because the set is not as efficient as the vector for the later uses of the live range. After the flushing, the set is deleted and cannot be used again. - The set is only for live ranges computed in LiveIntervalAnalysis::computeLiveInRegUnits() and getRegUnit() but not in computeVirtRegs(), because I did not bring any performance benefits to computeVirtRegs() and for some examples even brought a slow down. Patch by Vaidas Gasiunas <vaidas.gasiunas@sap.com> Differential Revision: http://reviews.llvm.org/D6013 llvm-svn: 228421
2015-02-07 02:42:41 +08:00
// Use segment set to speed-up initial computation of the live range.
LR = RegUnitRanges[Unit] = new LiveRange(UseSegmentSetForPhysRegs);
NewRanges.push_back(Unit);
}
VNInfo *VNI = LR->createDeadDef(Begin, getVNInfoAllocator());
(void)VNI;
DEBUG(dbgs() << ' ' << PrintRegUnit(Unit, TRI) << '#' << VNI->id);
}
}
DEBUG(dbgs() << '\n');
}
DEBUG(dbgs() << "Created " << NewRanges.size() << " new intervals.\n");
// Compute the 'normal' part of the ranges.
for (unsigned Unit : NewRanges)
computeRegUnitRange(*RegUnitRanges[Unit], Unit);
}
static void createSegmentsForValues(LiveRange &LR,
iterator_range<LiveInterval::vni_iterator> VNIs) {
for (VNInfo *VNI : VNIs) {
if (VNI->isUnused())
continue;
SlotIndex Def = VNI->def;
LR.addSegment(LiveRange::Segment(Def, Def.getDeadSlot(), VNI));
}
}
using ShrinkToUsesWorkList = SmallVector<std::pair<SlotIndex, VNInfo*>, 16>;
static void extendSegmentsToUses(LiveRange &LR, const SlotIndexes &Indexes,
ShrinkToUsesWorkList &WorkList,
const LiveRange &OldRange) {
// Keep track of the PHIs that are in use.
SmallPtrSet<VNInfo*, 8> UsedPHIs;
// Blocks that have already been added to WorkList as live-out.
SmallPtrSet<const MachineBasicBlock*, 16> LiveOut;
// Extend intervals to reach all uses in WorkList.
while (!WorkList.empty()) {
SlotIndex Idx = WorkList.back().first;
VNInfo *VNI = WorkList.back().second;
WorkList.pop_back();
const MachineBasicBlock *MBB = Indexes.getMBBFromIndex(Idx.getPrevSlot());
SlotIndex BlockStart = Indexes.getMBBStartIdx(MBB);
// Extend the live range for VNI to be live at Idx.
if (VNInfo *ExtVNI = LR.extendInBlock(BlockStart, Idx)) {
assert(ExtVNI == VNI && "Unexpected existing value number");
(void)ExtVNI;
// Is this a PHIDef we haven't seen before?
if (!VNI->isPHIDef() || VNI->def != BlockStart ||
!UsedPHIs.insert(VNI).second)
continue;
// The PHI is live, make sure the predecessors are live-out.
for (const MachineBasicBlock *Pred : MBB->predecessors()) {
if (!LiveOut.insert(Pred).second)
continue;
SlotIndex Stop = Indexes.getMBBEndIdx(Pred);
// A predecessor is not required to have a live-out value for a PHI.
if (VNInfo *PVNI = OldRange.getVNInfoBefore(Stop))
WorkList.push_back(std::make_pair(Stop, PVNI));
}
continue;
}
// VNI is live-in to MBB.
DEBUG(dbgs() << " live-in at " << BlockStart << '\n');
LR.addSegment(LiveRange::Segment(BlockStart, Idx, VNI));
// Make sure VNI is live-out from the predecessors.
for (const MachineBasicBlock *Pred : MBB->predecessors()) {
if (!LiveOut.insert(Pred).second)
continue;
SlotIndex Stop = Indexes.getMBBEndIdx(Pred);
assert(OldRange.getVNInfoBefore(Stop) == VNI &&
"Wrong value out of predecessor");
WorkList.push_back(std::make_pair(Stop, VNI));
}
}
}
bool LiveIntervals::shrinkToUses(LiveInterval *li,
SmallVectorImpl<MachineInstr*> *dead) {
DEBUG(dbgs() << "Shrink: " << *li << '\n');
assert(TargetRegisterInfo::isVirtualRegister(li->reg)
&& "Can only shrink virtual registers");
// Shrink subregister live ranges.
bool NeedsCleanup = false;
for (LiveInterval::SubRange &S : li->subranges()) {
shrinkToUses(S, li->reg);
if (S.empty())
NeedsCleanup = true;
}
if (NeedsCleanup)
li->removeEmptySubRanges();
// Find all the values used, including PHI kills.
ShrinkToUsesWorkList WorkList;
// Visit all instructions reading li->reg.
unsigned Reg = li->reg;
for (MachineInstr &UseMI : MRI->reg_instructions(Reg)) {
if (UseMI.isDebugValue() || !UseMI.readsVirtualRegister(Reg))
continue;
SlotIndex Idx = getInstructionIndex(UseMI).getRegSlot();
LiveQueryResult LRQ = li->Query(Idx);
VNInfo *VNI = LRQ.valueIn();
if (!VNI) {
// This shouldn't happen: readsVirtualRegister returns true, but there is
// no live value. It is likely caused by a target getting <undef> flags
// wrong.
DEBUG(dbgs() << Idx << '\t' << UseMI
<< "Warning: Instr claims to read non-existent value in "
<< *li << '\n');
continue;
}
// Special case: An early-clobber tied operand reads and writes the
// register one slot early.
if (VNInfo *DefVNI = LRQ.valueDefined())
Idx = DefVNI->def;
WorkList.push_back(std::make_pair(Idx, VNI));
}
// Create new live ranges with only minimal live segments per def.
LiveRange NewLR;
createSegmentsForValues(NewLR, make_range(li->vni_begin(), li->vni_end()));
extendSegmentsToUses(NewLR, *Indexes, WorkList, *li);
// Move the trimmed segments back.
li->segments.swap(NewLR.segments);
// Handle dead values.
bool CanSeparate = computeDeadValues(*li, dead);
DEBUG(dbgs() << "Shrunk: " << *li << '\n');
return CanSeparate;
}
bool LiveIntervals::computeDeadValues(LiveInterval &LI,
SmallVectorImpl<MachineInstr*> *dead) {
bool MayHaveSplitComponents = false;
for (VNInfo *VNI : LI.valnos) {
if (VNI->isUnused())
continue;
SlotIndex Def = VNI->def;
LiveRange::iterator I = LI.FindSegmentContaining(Def);
assert(I != LI.end() && "Missing segment for VNI");
// Is the register live before? Otherwise we may have to add a read-undef
// flag for subregister defs.
unsigned VReg = LI.reg;
if (MRI->shouldTrackSubRegLiveness(VReg)) {
if ((I == LI.begin() || std::prev(I)->end < Def) && !VNI->isPHIDef()) {
MachineInstr *MI = getInstructionFromIndex(Def);
MI->setRegisterDefReadUndef(VReg);
}
}
if (I->end != Def.getDeadSlot())
continue;
2011-03-02 08:33:01 +08:00
if (VNI->isPHIDef()) {
// This is a dead PHI. Remove it.
VNI->markUnused();
LI.removeSegment(I);
DEBUG(dbgs() << "Dead PHI at " << Def << " may separate interval\n");
MayHaveSplitComponents = true;
} else {
// This is a dead def. Make sure the instruction knows.
MachineInstr *MI = getInstructionFromIndex(Def);
assert(MI && "No instruction defining live value");
MI->addRegisterDead(LI.reg, TRI);
if (dead && MI->allDefsAreDead()) {
DEBUG(dbgs() << "All defs dead: " << Def << '\t' << *MI);
dead->push_back(MI);
}
}
}
return MayHaveSplitComponents;
}
void LiveIntervals::shrinkToUses(LiveInterval::SubRange &SR, unsigned Reg) {
DEBUG(dbgs() << "Shrink: " << SR << '\n');
assert(TargetRegisterInfo::isVirtualRegister(Reg)
&& "Can only shrink virtual registers");
// Find all the values used, including PHI kills.
ShrinkToUsesWorkList WorkList;
// Visit all instructions reading Reg.
SlotIndex LastIdx;
for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) {
// Skip "undef" uses.
if (!MO.readsReg())
continue;
// Maybe the operand is for a subregister we don't care about.
unsigned SubReg = MO.getSubReg();
if (SubReg != 0) {
LaneBitmask LaneMask = TRI->getSubRegIndexLaneMask(SubReg);
if ((LaneMask & SR.LaneMask).none())
continue;
}
// We only need to visit each instruction once.
MachineInstr *UseMI = MO.getParent();
SlotIndex Idx = getInstructionIndex(*UseMI).getRegSlot();
if (Idx == LastIdx)
continue;
LastIdx = Idx;
LiveQueryResult LRQ = SR.Query(Idx);
VNInfo *VNI = LRQ.valueIn();
// For Subranges it is possible that only undef values are left in that
// part of the subregister, so there is no real liverange at the use
if (!VNI)
continue;
// Special case: An early-clobber tied operand reads and writes the
// register one slot early.
if (VNInfo *DefVNI = LRQ.valueDefined())
Idx = DefVNI->def;
WorkList.push_back(std::make_pair(Idx, VNI));
}
// Create a new live ranges with only minimal live segments per def.
LiveRange NewLR;
createSegmentsForValues(NewLR, make_range(SR.vni_begin(), SR.vni_end()));
extendSegmentsToUses(NewLR, *Indexes, WorkList, SR);
// Move the trimmed ranges back.
SR.segments.swap(NewLR.segments);
// Remove dead PHI value numbers
for (VNInfo *VNI : SR.valnos) {
if (VNI->isUnused())
continue;
const LiveRange::Segment *Segment = SR.getSegmentContaining(VNI->def);
assert(Segment != nullptr && "Missing segment for VNI");
if (Segment->end != VNI->def.getDeadSlot())
continue;
if (VNI->isPHIDef()) {
// This is a dead PHI. Remove it.
DEBUG(dbgs() << "Dead PHI at " << VNI->def << " may separate interval\n");
VNI->markUnused();
SR.removeSegment(*Segment);
}
}
DEBUG(dbgs() << "Shrunk: " << SR << '\n');
}
void LiveIntervals::extendToIndices(LiveRange &LR,
ArrayRef<SlotIndex> Indices,
ArrayRef<SlotIndex> Undefs) {
assert(LRCalc && "LRCalc not initialized.");
LRCalc->reset(MF, getSlotIndexes(), DomTree, &getVNInfoAllocator());
for (SlotIndex Idx : Indices)
LRCalc->extend(LR, Idx, /*PhysReg=*/0, Undefs);
}
void LiveIntervals::pruneValue(LiveRange &LR, SlotIndex Kill,
SmallVectorImpl<SlotIndex> *EndPoints) {
LiveQueryResult LRQ = LR.Query(Kill);
VNInfo *VNI = LRQ.valueOutOrDead();
if (!VNI)
return;
MachineBasicBlock *KillMBB = Indexes->getMBBFromIndex(Kill);
SlotIndex MBBEnd = Indexes->getMBBEndIdx(KillMBB);
// If VNI isn't live out from KillMBB, the value is trivially pruned.
if (LRQ.endPoint() < MBBEnd) {
LR.removeSegment(Kill, LRQ.endPoint());
if (EndPoints) EndPoints->push_back(LRQ.endPoint());
return;
}
// VNI is live out of KillMBB.
LR.removeSegment(Kill, MBBEnd);
if (EndPoints) EndPoints->push_back(MBBEnd);
// Find all blocks that are reachable from KillMBB without leaving VNI's live
// range. It is possible that KillMBB itself is reachable, so start a DFS
// from each successor.
using VisitedTy = df_iterator_default_set<MachineBasicBlock*,9>;
VisitedTy Visited;
for (MachineBasicBlock *Succ : KillMBB->successors()) {
for (df_ext_iterator<MachineBasicBlock*, VisitedTy>
I = df_ext_begin(Succ, Visited), E = df_ext_end(Succ, Visited);
I != E;) {
MachineBasicBlock *MBB = *I;
// Check if VNI is live in to MBB.
SlotIndex MBBStart, MBBEnd;
std::tie(MBBStart, MBBEnd) = Indexes->getMBBRange(MBB);
LiveQueryResult LRQ = LR.Query(MBBStart);
if (LRQ.valueIn() != VNI) {
// This block isn't part of the VNI segment. Prune the search.
I.skipChildren();
continue;
}
// Prune the search if VNI is killed in MBB.
if (LRQ.endPoint() < MBBEnd) {
LR.removeSegment(MBBStart, LRQ.endPoint());
if (EndPoints) EndPoints->push_back(LRQ.endPoint());
I.skipChildren();
continue;
}
// VNI is live through MBB.
LR.removeSegment(MBBStart, MBBEnd);
if (EndPoints) EndPoints->push_back(MBBEnd);
++I;
}
}
}
2007-11-12 14:35:08 +08:00
//===----------------------------------------------------------------------===//
// Register allocator hooks.
//
void LiveIntervals::addKillFlags(const VirtRegMap *VRM) {
// Keep track of regunit ranges.
SmallVector<std::pair<const LiveRange*, LiveRange::const_iterator>, 8> RU;
// Keep track of subregister ranges.
SmallVector<std::pair<const LiveInterval::SubRange*,
LiveRange::const_iterator>, 4> SRs;
for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
if (MRI->reg_nodbg_empty(Reg))
continue;
const LiveInterval &LI = getInterval(Reg);
if (LI.empty())
continue;
// Find the regunit intervals for the assigned register. They may overlap
// the virtual register live range, cancelling any kills.
RU.clear();
for (MCRegUnitIterator Unit(VRM->getPhys(Reg), TRI); Unit.isValid();
++Unit) {
const LiveRange &RURange = getRegUnit(*Unit);
if (RURange.empty())
continue;
RU.push_back(std::make_pair(&RURange, RURange.find(LI.begin()->end)));
}
if (MRI->subRegLivenessEnabled()) {
SRs.clear();
for (const LiveInterval::SubRange &SR : LI.subranges()) {
SRs.push_back(std::make_pair(&SR, SR.find(LI.begin()->end)));
}
}
// Every instruction that kills Reg corresponds to a segment range end
// point.
for (LiveInterval::const_iterator RI = LI.begin(), RE = LI.end(); RI != RE;
++RI) {
// A block index indicates an MBB edge.
if (RI->end.isBlock())
continue;
MachineInstr *MI = getInstructionFromIndex(RI->end);
if (!MI)
continue;
2013-10-05 00:52:58 +08:00
// Check if any of the regunits are live beyond the end of RI. That could
// happen when a physreg is defined as a copy of a virtreg:
//
// %EAX = COPY %vreg5
// FOO %vreg5 <--- MI, cancel kill because %EAX is live.
// BAR %EAX<kill>
//
// There should be no kill flag on FOO when %vreg5 is rewritten as %EAX.
for (auto &RUP : RU) {
const LiveRange &RURange = *RUP.first;
LiveRange::const_iterator &I = RUP.second;
if (I == RURange.end())
continue;
I = RURange.advanceTo(I, RI->end);
if (I == RURange.end() || I->start >= RI->end)
continue;
// I is overlapping RI.
goto CancelKill;
}
if (MRI->subRegLivenessEnabled()) {
// When reading a partial undefined value we must not add a kill flag.
// The regalloc might have used the undef lane for something else.
// Example:
// %vreg1 = ... ; R32: %vreg1
// %vreg2:high16 = ... ; R64: %vreg2
// = read %vreg2<kill> ; R64: %vreg2
// = read %vreg1 ; R32: %vreg1
// The <kill> flag is correct for %vreg2, but the register allocator may
// assign R0L to %vreg1, and R0 to %vreg2 because the low 32bits of R0
// are actually never written by %vreg2. After assignment the <kill>
// flag at the read instruction is invalid.
LaneBitmask DefinedLanesMask;
if (!SRs.empty()) {
// Compute a mask of lanes that are defined.
DefinedLanesMask = LaneBitmask::getNone();
for (auto &SRP : SRs) {
const LiveInterval::SubRange &SR = *SRP.first;
LiveRange::const_iterator &I = SRP.second;
if (I == SR.end())
continue;
I = SR.advanceTo(I, RI->end);
if (I == SR.end() || I->start >= RI->end)
continue;
// I is overlapping RI
DefinedLanesMask |= SR.LaneMask;
}
} else
DefinedLanesMask = LaneBitmask::getAll();
bool IsFullWrite = false;
for (const MachineOperand &MO : MI->operands()) {
if (!MO.isReg() || MO.getReg() != Reg)
continue;
if (MO.isUse()) {
// Reading any undefined lanes?
LaneBitmask UseMask = TRI->getSubRegIndexLaneMask(MO.getSubReg());
if ((UseMask & ~DefinedLanesMask).any())
goto CancelKill;
} else if (MO.getSubReg() == 0) {
// Writing to the full register?
assert(MO.isDef());
IsFullWrite = true;
}
}
// If an instruction writes to a subregister, a new segment starts in
// the LiveInterval. But as this is only overriding part of the register
// adding kill-flags is not correct here after registers have been
// assigned.
if (!IsFullWrite) {
// Next segment has to be adjacent in the subregister write case.
LiveRange::const_iterator N = std::next(RI);
if (N != LI.end() && N->start == RI->end)
goto CancelKill;
}
}
MI->addRegisterKilled(Reg, nullptr);
continue;
CancelKill:
MI->clearRegisterKills(Reg, nullptr);
}
}
}
MachineBasicBlock*
LiveIntervals::intervalIsInOneMBB(const LiveInterval &LI) const {
// A local live range must be fully contained inside the block, meaning it is
// defined and killed at instructions, not at block boundaries. It is not
// live in or or out of any block.
//
// It is technically possible to have a PHI-defined live range identical to a
// single block, but we are going to return false in that case.
SlotIndex Start = LI.beginIndex();
if (Start.isBlock())
return nullptr;
SlotIndex Stop = LI.endIndex();
if (Stop.isBlock())
return nullptr;
// getMBBFromIndex doesn't need to search the MBB table when both indexes
// belong to proper instructions.
MachineBasicBlock *MBB1 = Indexes->getMBBFromIndex(Start);
MachineBasicBlock *MBB2 = Indexes->getMBBFromIndex(Stop);
return MBB1 == MBB2 ? MBB1 : nullptr;
}
bool
LiveIntervals::hasPHIKill(const LiveInterval &LI, const VNInfo *VNI) const {
for (const VNInfo *PHI : LI.valnos) {
if (PHI->isUnused() || !PHI->isPHIDef())
continue;
const MachineBasicBlock *PHIMBB = getMBBFromIndex(PHI->def);
// Conservatively return true instead of scanning huge predecessor lists.
if (PHIMBB->pred_size() > 100)
return true;
for (const MachineBasicBlock *Pred : PHIMBB->predecessors())
if (VNI == LI.getVNInfoBefore(Indexes->getMBBEndIdx(Pred)))
return true;
}
return false;
}
float LiveIntervals::getSpillWeight(bool isDef, bool isUse,
const MachineBlockFrequencyInfo *MBFI,
const MachineInstr &MI) {
BlockFrequency Freq = MBFI->getBlockFreq(MI.getParent());
const float Scale = 1.0f / MBFI->getEntryFreq();
return (isDef + isUse) * (Freq.getFrequency() * Scale);
}
LiveRange::Segment
LiveIntervals::addSegmentToEndOfBlock(unsigned reg, MachineInstr &startInst) {
LiveInterval& Interval = createEmptyInterval(reg);
VNInfo *VN = Interval.getNextValue(
SlotIndex(getInstructionIndex(startInst).getRegSlot()),
getVNInfoAllocator());
LiveRange::Segment S(SlotIndex(getInstructionIndex(startInst).getRegSlot()),
getMBBEndIdx(startInst.getParent()), VN);
Interval.addSegment(S);
return S;
}
//===----------------------------------------------------------------------===//
// Register mask functions
//===----------------------------------------------------------------------===//
bool LiveIntervals::checkRegMaskInterference(LiveInterval &LI,
BitVector &UsableRegs) {
if (LI.empty())
return false;
LiveInterval::iterator LiveI = LI.begin(), LiveE = LI.end();
// Use a smaller arrays for local live ranges.
ArrayRef<SlotIndex> Slots;
ArrayRef<const uint32_t*> Bits;
if (MachineBasicBlock *MBB = intervalIsInOneMBB(LI)) {
Slots = getRegMaskSlotsInBlock(MBB->getNumber());
Bits = getRegMaskBitsInBlock(MBB->getNumber());
} else {
Slots = getRegMaskSlots();
Bits = getRegMaskBits();
}
// We are going to enumerate all the register mask slots contained in LI.
// Start with a binary search of RegMaskSlots to find a starting point.
ArrayRef<SlotIndex>::iterator SlotI =
std::lower_bound(Slots.begin(), Slots.end(), LiveI->start);
ArrayRef<SlotIndex>::iterator SlotE = Slots.end();
// No slots in range, LI begins after the last call.
if (SlotI == SlotE)
return false;
bool Found = false;
while (true) {
assert(*SlotI >= LiveI->start);
// Loop over all slots overlapping this segment.
while (*SlotI < LiveI->end) {
// *SlotI overlaps LI. Collect mask bits.
if (!Found) {
// This is the first overlap. Initialize UsableRegs to all ones.
UsableRegs.clear();
UsableRegs.resize(TRI->getNumRegs(), true);
Found = true;
}
// Remove usable registers clobbered by this mask.
UsableRegs.clearBitsNotInMask(Bits[SlotI-Slots.begin()]);
if (++SlotI == SlotE)
return Found;
}
// *SlotI is beyond the current LI segment.
LiveI = LI.advanceTo(LiveI, *SlotI);
if (LiveI == LiveE)
return Found;
// Advance SlotI until it overlaps.
while (*SlotI < LiveI->start)
if (++SlotI == SlotE)
return Found;
}
}
//===----------------------------------------------------------------------===//
// IntervalUpdate class.
//===----------------------------------------------------------------------===//
/// Toolkit used by handleMove to trim or extend live intervals.
class LiveIntervals::HMEditor {
private:
LiveIntervals& LIS;
const MachineRegisterInfo& MRI;
const TargetRegisterInfo& TRI;
SlotIndex OldIdx;
SlotIndex NewIdx;
SmallPtrSet<LiveRange*, 8> Updated;
bool UpdateFlags;
public:
HMEditor(LiveIntervals& LIS, const MachineRegisterInfo& MRI,
const TargetRegisterInfo& TRI,
SlotIndex OldIdx, SlotIndex NewIdx, bool UpdateFlags)
: LIS(LIS), MRI(MRI), TRI(TRI), OldIdx(OldIdx), NewIdx(NewIdx),
UpdateFlags(UpdateFlags) {}
// FIXME: UpdateFlags is a workaround that creates live intervals for all
// physregs, even those that aren't needed for regalloc, in order to update
// kill flags. This is wasteful. Eventually, LiveVariables will strip all kill
// flags, and postRA passes will use a live register utility instead.
LiveRange *getRegUnitLI(unsigned Unit) {
LiveIntervalAnalysis: Fix alias regunit reserved definition A register in CodeGen can be marked as reserved: In that case we consider the register always live and do not use (or rather ignore) kill/dead/undef operand flags. LiveIntervalAnalysis however tracks liveness per register unit (not per register). We already needed adjustments for this in r292871 to deal with super/sub registers. However I did not look at aliased register there. Looking at ARM: FPSCR (regunits FPSCR, FPSCR~FPSCR_NZCV) aliases with FPSCR_NZCV (regunits FPSCR_NZCV, FPSCR~FPSCR_NZCV) hence they share a register unit (FPSCR~FPSCR_NZCV) that represents the aliased parts of the registers. This shared register unit was previously considered non-reserved, however given that we uses of the reserved FPSCR potentially violate some rules (like uses without defs) we should make FPSCR~FPSCR_NZCV reserved too and stop tracking liveness for it. This patch: - Defines a register unit as reserved when: At least for one root register, the root register and all its super registers are reserved. - Adjust LiveIntervals::computeRegUnitRange() for new reserved definition. - Add MachineRegisterInfo::isReservedRegUnit() to have a canonical way of testing. - Stop computing LiveRanges for reserved register units in HMEditor even with UpdateFlags enabled. - Skip verification of uses of reserved reg units in the machine verifier (this usually didn't happen because there would be no cached liverange but there is no guarantee for that and I would run into this case before the HMEditor tweak, so may as well fix the verifier too). Note that this should only affect ARMs FPSCR/FPSCR_NZCV registers today; aliased registers are rarely used, the only other cases are hexagons P0-P3/P3_0 and C8/USR pairs which are not mixing reserved/non-reserved registers in an alias. Differential Revision: https://reviews.llvm.org/D37356 llvm-svn: 312348
2017-09-02 02:36:26 +08:00
if (UpdateFlags && !MRI.isReservedRegUnit(Unit))
return &LIS.getRegUnit(Unit);
return LIS.getCachedRegUnit(Unit);
}
/// Update all live ranges touched by MI, assuming a move from OldIdx to
/// NewIdx.
void updateAllRanges(MachineInstr *MI) {
DEBUG(dbgs() << "handleMove " << OldIdx << " -> " << NewIdx << ": " << *MI);
bool hasRegMask = false;
for (MachineOperand &MO : MI->operands()) {
if (MO.isRegMask())
hasRegMask = true;
if (!MO.isReg())
continue;
if (MO.isUse()) {
if (!MO.readsReg())
continue;
// Aggressively clear all kill flags.
// They are reinserted by VirtRegRewriter.
MO.setIsKill(false);
}
unsigned Reg = MO.getReg();
if (!Reg)
continue;
if (TargetRegisterInfo::isVirtualRegister(Reg)) {
LiveInterval &LI = LIS.getInterval(Reg);
if (LI.hasSubRanges()) {
unsigned SubReg = MO.getSubReg();
LaneBitmask LaneMask = SubReg ? TRI.getSubRegIndexLaneMask(SubReg)
: MRI.getMaxLaneMaskForVReg(Reg);
for (LiveInterval::SubRange &S : LI.subranges()) {
if ((S.LaneMask & LaneMask).none())
continue;
updateRange(S, Reg, S.LaneMask);
}
}
updateRange(LI, Reg, LaneBitmask::getNone());
continue;
}
// For physregs, only update the regunits that actually have a
// precomputed live range.
for (MCRegUnitIterator Units(Reg, &TRI); Units.isValid(); ++Units)
if (LiveRange *LR = getRegUnitLI(*Units))
updateRange(*LR, *Units, LaneBitmask::getNone());
}
if (hasRegMask)
updateRegMaskSlots();
}
private:
/// Update a single live range, assuming an instruction has been moved from
/// OldIdx to NewIdx.
void updateRange(LiveRange &LR, unsigned Reg, LaneBitmask LaneMask) {
if (!Updated.insert(&LR).second)
return;
DEBUG({
dbgs() << " ";
if (TargetRegisterInfo::isVirtualRegister(Reg)) {
dbgs() << PrintReg(Reg);
if (LaneMask.any())
dbgs() << " L" << PrintLaneMask(LaneMask);
} else {
dbgs() << PrintRegUnit(Reg, &TRI);
}
dbgs() << ":\t" << LR << '\n';
});
if (SlotIndex::isEarlierInstr(OldIdx, NewIdx))
handleMoveDown(LR);
else
handleMoveUp(LR, Reg, LaneMask);
DEBUG(dbgs() << " -->\t" << LR << '\n');
LR.verify();
}
/// Update LR to reflect an instruction has been moved downwards from OldIdx
/// to NewIdx (OldIdx < NewIdx).
void handleMoveDown(LiveRange &LR) {
LiveRange::iterator E = LR.end();
// Segment going into OldIdx.
LiveRange::iterator OldIdxIn = LR.find(OldIdx.getBaseIndex());
// No value live before or after OldIdx? Nothing to do.
if (OldIdxIn == E || SlotIndex::isEarlierInstr(OldIdx, OldIdxIn->start))
return;
LiveRange::iterator OldIdxOut;
// Do we have a value live-in to OldIdx?
if (SlotIndex::isEarlierInstr(OldIdxIn->start, OldIdx)) {
// If the live-in value already extends to NewIdx, there is nothing to do.
if (SlotIndex::isEarlierEqualInstr(NewIdx, OldIdxIn->end))
return;
// Aggressively remove all kill flags from the old kill point.
// Kill flags shouldn't be used while live intervals exist, they will be
// reinserted by VirtRegRewriter.
if (MachineInstr *KillMI = LIS.getInstructionFromIndex(OldIdxIn->end))
for (MIBundleOperands MO(*KillMI); MO.isValid(); ++MO)
if (MO->isReg() && MO->isUse())
MO->setIsKill(false);
// Is there a def before NewIdx which is not OldIdx?
LiveRange::iterator Next = std::next(OldIdxIn);
if (Next != E && !SlotIndex::isSameInstr(OldIdx, Next->start) &&
SlotIndex::isEarlierInstr(Next->start, NewIdx)) {
// If we are here then OldIdx was just a use but not a def. We only have
// to ensure liveness extends to NewIdx.
LiveRange::iterator NewIdxIn =
LR.advanceTo(Next, NewIdx.getBaseIndex());
// Extend the segment before NewIdx if necessary.
if (NewIdxIn == E ||
!SlotIndex::isEarlierInstr(NewIdxIn->start, NewIdx)) {
LiveRange::iterator Prev = std::prev(NewIdxIn);
Prev->end = NewIdx.getRegSlot();
}
// Extend OldIdxIn.
OldIdxIn->end = Next->start;
return;
}
// Adjust OldIdxIn->end to reach NewIdx. This may temporarily make LR
// invalid by overlapping ranges.
bool isKill = SlotIndex::isSameInstr(OldIdx, OldIdxIn->end);
OldIdxIn->end = NewIdx.getRegSlot(OldIdxIn->end.isEarlyClobber());
// If this was not a kill, then there was no def and we're done.
if (!isKill)
return;
// Did we have a Def at OldIdx?
OldIdxOut = Next;
if (OldIdxOut == E || !SlotIndex::isSameInstr(OldIdx, OldIdxOut->start))
return;
} else {
OldIdxOut = OldIdxIn;
}
// If we are here then there is a Definition at OldIdx. OldIdxOut points
// to the segment starting there.
assert(OldIdxOut != E && SlotIndex::isSameInstr(OldIdx, OldIdxOut->start) &&
"No def?");
VNInfo *OldIdxVNI = OldIdxOut->valno;
assert(OldIdxVNI->def == OldIdxOut->start && "Inconsistent def");
// If the defined value extends beyond NewIdx, just move the beginning
// of the segment to NewIdx.
SlotIndex NewIdxDef = NewIdx.getRegSlot(OldIdxOut->start.isEarlyClobber());
if (SlotIndex::isEarlierInstr(NewIdxDef, OldIdxOut->end)) {
OldIdxVNI->def = NewIdxDef;
OldIdxOut->start = OldIdxVNI->def;
return;
}
// If we are here then we have a Definition at OldIdx which ends before
// NewIdx.
// Is there an existing Def at NewIdx?
LiveRange::iterator AfterNewIdx
= LR.advanceTo(OldIdxOut, NewIdx.getRegSlot());
bool OldIdxDefIsDead = OldIdxOut->end.isDead();
if (!OldIdxDefIsDead &&
SlotIndex::isEarlierInstr(OldIdxOut->end, NewIdxDef)) {
// OldIdx is not a dead def, and NewIdxDef is inside a new interval.
VNInfo *DefVNI;
if (OldIdxOut != LR.begin() &&
!SlotIndex::isEarlierInstr(std::prev(OldIdxOut)->end,
OldIdxOut->start)) {
// There is no gap between OldIdxOut and its predecessor anymore,
// merge them.
LiveRange::iterator IPrev = std::prev(OldIdxOut);
DefVNI = OldIdxVNI;
IPrev->end = OldIdxOut->end;
} else {
// The value is live in to OldIdx
LiveRange::iterator INext = std::next(OldIdxOut);
assert(INext != E && "Must have following segment");
// We merge OldIdxOut and its successor. As we're dealing with subreg
// reordering, there is always a successor to OldIdxOut in the same BB
// We don't need INext->valno anymore and will reuse for the new segment
// we create later.
DefVNI = OldIdxVNI;
INext->start = OldIdxOut->end;
INext->valno->def = INext->start;
}
// If NewIdx is behind the last segment, extend that and append a new one.
if (AfterNewIdx == E) {
// OldIdxOut is undef at this point, Slide (OldIdxOut;AfterNewIdx] up
// one position.
// |- ?/OldIdxOut -| |- X0 -| ... |- Xn -| end
// => |- X0/OldIdxOut -| ... |- Xn -| |- undef/NewS -| end
std::copy(std::next(OldIdxOut), E, OldIdxOut);
// The last segment is undefined now, reuse it for a dead def.
LiveRange::iterator NewSegment = std::prev(E);
*NewSegment = LiveRange::Segment(NewIdxDef, NewIdxDef.getDeadSlot(),
DefVNI);
DefVNI->def = NewIdxDef;
LiveRange::iterator Prev = std::prev(NewSegment);
Prev->end = NewIdxDef;
} else {
// OldIdxOut is undef at this point, Slide (OldIdxOut;AfterNewIdx] up
// one position.
// |- ?/OldIdxOut -| |- X0 -| ... |- Xn/AfterNewIdx -| |- Next -|
// => |- X0/OldIdxOut -| ... |- Xn -| |- Xn/AfterNewIdx -| |- Next -|
std::copy(std::next(OldIdxOut), std::next(AfterNewIdx), OldIdxOut);
LiveRange::iterator Prev = std::prev(AfterNewIdx);
// We have two cases:
if (SlotIndex::isEarlierInstr(Prev->start, NewIdxDef)) {
// Case 1: NewIdx is inside a liverange. Split this liverange at
// NewIdxDef into the segment "Prev" followed by "NewSegment".
LiveRange::iterator NewSegment = AfterNewIdx;
*NewSegment = LiveRange::Segment(NewIdxDef, Prev->end, Prev->valno);
Prev->valno->def = NewIdxDef;
*Prev = LiveRange::Segment(Prev->start, NewIdxDef, DefVNI);
DefVNI->def = Prev->start;
} else {
// Case 2: NewIdx is in a lifetime hole. Keep AfterNewIdx as is and
// turn Prev into a segment from NewIdx to AfterNewIdx->start.
*Prev = LiveRange::Segment(NewIdxDef, AfterNewIdx->start, DefVNI);
DefVNI->def = NewIdxDef;
assert(DefVNI != AfterNewIdx->valno);
}
}
return;
}
if (AfterNewIdx != E &&
SlotIndex::isSameInstr(AfterNewIdx->start, NewIdxDef)) {
// There is an existing def at NewIdx. The def at OldIdx is coalesced into
// that value.
assert(AfterNewIdx->valno != OldIdxVNI && "Multiple defs of value?");
LR.removeValNo(OldIdxVNI);
} else {
// There was no existing def at NewIdx. We need to create a dead def
// at NewIdx. Shift segments over the old OldIdxOut segment, this frees
// a new segment at the place where we want to construct the dead def.
// |- OldIdxOut -| |- X0 -| ... |- Xn -| |- AfterNewIdx -|
// => |- X0/OldIdxOut -| ... |- Xn -| |- undef/NewS. -| |- AfterNewIdx -|
assert(AfterNewIdx != OldIdxOut && "Inconsistent iterators");
std::copy(std::next(OldIdxOut), AfterNewIdx, OldIdxOut);
// We can reuse OldIdxVNI now.
LiveRange::iterator NewSegment = std::prev(AfterNewIdx);
VNInfo *NewSegmentVNI = OldIdxVNI;
NewSegmentVNI->def = NewIdxDef;
*NewSegment = LiveRange::Segment(NewIdxDef, NewIdxDef.getDeadSlot(),
NewSegmentVNI);
}
}
/// Update LR to reflect an instruction has been moved upwards from OldIdx
/// to NewIdx (NewIdx < OldIdx).
void handleMoveUp(LiveRange &LR, unsigned Reg, LaneBitmask LaneMask) {
LiveRange::iterator E = LR.end();
// Segment going into OldIdx.
LiveRange::iterator OldIdxIn = LR.find(OldIdx.getBaseIndex());
// No value live before or after OldIdx? Nothing to do.
if (OldIdxIn == E || SlotIndex::isEarlierInstr(OldIdx, OldIdxIn->start))
return;
LiveRange::iterator OldIdxOut;
// Do we have a value live-in to OldIdx?
if (SlotIndex::isEarlierInstr(OldIdxIn->start, OldIdx)) {
// If the live-in value isn't killed here, then we have no Def at
// OldIdx, moreover the value must be live at NewIdx so there is nothing
// to do.
bool isKill = SlotIndex::isSameInstr(OldIdx, OldIdxIn->end);
if (!isKill)
return;
// At this point we have to move OldIdxIn->end back to the nearest
// previous use or (dead-)def but no further than NewIdx.
SlotIndex DefBeforeOldIdx
= std::max(OldIdxIn->start.getDeadSlot(),
NewIdx.getRegSlot(OldIdxIn->end.isEarlyClobber()));
OldIdxIn->end = findLastUseBefore(DefBeforeOldIdx, Reg, LaneMask);
// Did we have a Def at OldIdx? If not we are done now.
OldIdxOut = std::next(OldIdxIn);
if (OldIdxOut == E || !SlotIndex::isSameInstr(OldIdx, OldIdxOut->start))
return;
} else {
OldIdxOut = OldIdxIn;
OldIdxIn = OldIdxOut != LR.begin() ? std::prev(OldIdxOut) : E;
}
// If we are here then there is a Definition at OldIdx. OldIdxOut points
// to the segment starting there.
assert(OldIdxOut != E && SlotIndex::isSameInstr(OldIdx, OldIdxOut->start) &&
"No def?");
VNInfo *OldIdxVNI = OldIdxOut->valno;
assert(OldIdxVNI->def == OldIdxOut->start && "Inconsistent def");
bool OldIdxDefIsDead = OldIdxOut->end.isDead();
// Is there an existing def at NewIdx?
SlotIndex NewIdxDef = NewIdx.getRegSlot(OldIdxOut->start.isEarlyClobber());
LiveRange::iterator NewIdxOut = LR.find(NewIdx.getRegSlot());
if (SlotIndex::isSameInstr(NewIdxOut->start, NewIdx)) {
assert(NewIdxOut->valno != OldIdxVNI &&
"Same value defined more than once?");
// If OldIdx was a dead def remove it.
if (!OldIdxDefIsDead) {
// Remove segment starting at NewIdx and move begin of OldIdxOut to
// NewIdx so it can take its place.
OldIdxVNI->def = NewIdxDef;
OldIdxOut->start = NewIdxDef;
LR.removeValNo(NewIdxOut->valno);
} else {
// Simply remove the dead def at OldIdx.
LR.removeValNo(OldIdxVNI);
}
} else {
// Previously nothing was live after NewIdx, so all we have to do now is
// move the begin of OldIdxOut to NewIdx.
if (!OldIdxDefIsDead) {
// Do we have any intermediate Defs between OldIdx and NewIdx?
if (OldIdxIn != E &&
SlotIndex::isEarlierInstr(NewIdxDef, OldIdxIn->start)) {
// OldIdx is not a dead def and NewIdx is before predecessor start.
LiveRange::iterator NewIdxIn = NewIdxOut;
assert(NewIdxIn == LR.find(NewIdx.getBaseIndex()));
const SlotIndex SplitPos = NewIdxDef;
OldIdxVNI = OldIdxIn->valno;
// Merge the OldIdxIn and OldIdxOut segments into OldIdxOut.
OldIdxOut->valno->def = OldIdxIn->start;
*OldIdxOut = LiveRange::Segment(OldIdxIn->start, OldIdxOut->end,
OldIdxOut->valno);
// OldIdxIn and OldIdxVNI are now undef and can be overridden.
// We Slide [NewIdxIn, OldIdxIn) down one position.
// |- X0/NewIdxIn -| ... |- Xn-1 -||- Xn/OldIdxIn -||- OldIdxOut -|
// => |- undef/NexIdxIn -| |- X0 -| ... |- Xn-1 -| |- Xn/OldIdxOut -|
std::copy_backward(NewIdxIn, OldIdxIn, OldIdxOut);
// NewIdxIn is now considered undef so we can reuse it for the moved
// value.
LiveRange::iterator NewSegment = NewIdxIn;
LiveRange::iterator Next = std::next(NewSegment);
if (SlotIndex::isEarlierInstr(Next->start, NewIdx)) {
// There is no gap between NewSegment and its predecessor.
*NewSegment = LiveRange::Segment(Next->start, SplitPos,
Next->valno);
*Next = LiveRange::Segment(SplitPos, Next->end, OldIdxVNI);
Next->valno->def = SplitPos;
} else {
// There is a gap between NewSegment and its predecessor
// Value becomes live in.
*NewSegment = LiveRange::Segment(SplitPos, Next->start, OldIdxVNI);
NewSegment->valno->def = SplitPos;
}
} else {
// Leave the end point of a live def.
OldIdxOut->start = NewIdxDef;
OldIdxVNI->def = NewIdxDef;
if (OldIdxIn != E && SlotIndex::isEarlierInstr(NewIdx, OldIdxIn->end))
OldIdxIn->end = NewIdx.getRegSlot();
}
} else {
// OldIdxVNI is a dead def. It may have been moved across other values
// in LR, so move OldIdxOut up to NewIdxOut. Slide [NewIdxOut;OldIdxOut)
// down one position.
// |- X0/NewIdxOut -| ... |- Xn-1 -| |- Xn/OldIdxOut -| |- next - |
// => |- undef/NewIdxOut -| |- X0 -| ... |- Xn-1 -| |- next -|
std::copy_backward(NewIdxOut, OldIdxOut, std::next(OldIdxOut));
// OldIdxVNI can be reused now to build a new dead def segment.
LiveRange::iterator NewSegment = NewIdxOut;
VNInfo *NewSegmentVNI = OldIdxVNI;
*NewSegment = LiveRange::Segment(NewIdxDef, NewIdxDef.getDeadSlot(),
NewSegmentVNI);
NewSegmentVNI->def = NewIdxDef;
}
}
}
void updateRegMaskSlots() {
SmallVectorImpl<SlotIndex>::iterator RI =
std::lower_bound(LIS.RegMaskSlots.begin(), LIS.RegMaskSlots.end(),
OldIdx);
assert(RI != LIS.RegMaskSlots.end() && *RI == OldIdx.getRegSlot() &&
"No RegMask at OldIdx.");
*RI = NewIdx.getRegSlot();
assert((RI == LIS.RegMaskSlots.begin() ||
SlotIndex::isEarlierInstr(*std::prev(RI), *RI)) &&
"Cannot move regmask instruction above another call");
assert((std::next(RI) == LIS.RegMaskSlots.end() ||
SlotIndex::isEarlierInstr(*RI, *std::next(RI))) &&
"Cannot move regmask instruction below another call");
}
// Return the last use of reg between NewIdx and OldIdx.
SlotIndex findLastUseBefore(SlotIndex Before, unsigned Reg,
LaneBitmask LaneMask) {
if (TargetRegisterInfo::isVirtualRegister(Reg)) {
SlotIndex LastUse = Before;
for (MachineOperand &MO : MRI.use_nodbg_operands(Reg)) {
if (MO.isUndef())
continue;
unsigned SubReg = MO.getSubReg();
if (SubReg != 0 && LaneMask.any()
&& (TRI.getSubRegIndexLaneMask(SubReg) & LaneMask).none())
continue;
const MachineInstr &MI = *MO.getParent();
SlotIndex InstSlot = LIS.getSlotIndexes()->getInstructionIndex(MI);
if (InstSlot > LastUse && InstSlot < OldIdx)
LastUse = InstSlot.getRegSlot();
}
return LastUse;
}
// This is a regunit interval, so scanning the use list could be very
// expensive. Scan upwards from OldIdx instead.
assert(Before < OldIdx && "Expected upwards move");
SlotIndexes *Indexes = LIS.getSlotIndexes();
MachineBasicBlock *MBB = Indexes->getMBBFromIndex(Before);
// OldIdx may not correspond to an instruction any longer, so set MII to
// point to the next instruction after OldIdx, or MBB->end().
MachineBasicBlock::iterator MII = MBB->end();
if (MachineInstr *MI = Indexes->getInstructionFromIndex(
Indexes->getNextNonNullIndex(OldIdx)))
if (MI->getParent() == MBB)
MII = MI;
MachineBasicBlock::iterator Begin = MBB->begin();
while (MII != Begin) {
if ((--MII)->isDebugValue())
continue;
SlotIndex Idx = Indexes->getInstructionIndex(*MII);
// Stop searching when Before is reached.
if (!SlotIndex::isEarlierInstr(Before, Idx))
return Before;
// Check if MII uses Reg.
for (MIBundleOperands MO(*MII); MO.isValid(); ++MO)
if (MO->isReg() && !MO->isUndef() &&
TargetRegisterInfo::isPhysicalRegister(MO->getReg()) &&
TRI.hasRegUnit(MO->getReg(), Reg))
return Idx.getRegSlot();
}
// Didn't reach Before. It must be the first instruction in the block.
return Before;
}
};
void LiveIntervals::handleMove(MachineInstr &MI, bool UpdateFlags) {
assert(!MI.isBundled() && "Can't handle bundled instructions yet.");
SlotIndex OldIndex = Indexes->getInstructionIndex(MI);
Indexes->removeMachineInstrFromMaps(MI);
SlotIndex NewIndex = Indexes->insertMachineInstrInMaps(MI);
assert(getMBBStartIdx(MI.getParent()) <= OldIndex &&
OldIndex < getMBBEndIdx(MI.getParent()) &&
"Cannot handle moves across basic block boundaries.");
HMEditor HME(*this, *MRI, *TRI, OldIndex, NewIndex, UpdateFlags);
HME.updateAllRanges(&MI);
}
void LiveIntervals::handleMoveIntoBundle(MachineInstr &MI,
MachineInstr &BundleStart,
bool UpdateFlags) {
SlotIndex OldIndex = Indexes->getInstructionIndex(MI);
SlotIndex NewIndex = Indexes->getInstructionIndex(BundleStart);
HMEditor HME(*this, *MRI, *TRI, OldIndex, NewIndex, UpdateFlags);
HME.updateAllRanges(&MI);
}
void LiveIntervals::repairOldRegInRange(const MachineBasicBlock::iterator Begin,
const MachineBasicBlock::iterator End,
const SlotIndex endIdx,
LiveRange &LR, const unsigned Reg,
LaneBitmask LaneMask) {
LiveInterval::iterator LII = LR.find(endIdx);
SlotIndex lastUseIdx;
if (LII == LR.begin()) {
// This happens when the function is called for a subregister that only
// occurs _after_ the range that is to be repaired.
return;
}
if (LII != LR.end() && LII->start < endIdx)
lastUseIdx = LII->end;
else
--LII;
for (MachineBasicBlock::iterator I = End; I != Begin;) {
--I;
MachineInstr &MI = *I;
if (MI.isDebugValue())
continue;
SlotIndex instrIdx = getInstructionIndex(MI);
bool isStartValid = getInstructionFromIndex(LII->start);
bool isEndValid = getInstructionFromIndex(LII->end);
// FIXME: This doesn't currently handle early-clobber or multiple removed
// defs inside of the region to repair.
for (MachineInstr::mop_iterator OI = MI.operands_begin(),
OE = MI.operands_end();
OI != OE; ++OI) {
const MachineOperand &MO = *OI;
if (!MO.isReg() || MO.getReg() != Reg)
continue;
unsigned SubReg = MO.getSubReg();
LaneBitmask Mask = TRI->getSubRegIndexLaneMask(SubReg);
if ((Mask & LaneMask).none())
continue;
if (MO.isDef()) {
if (!isStartValid) {
if (LII->end.isDead()) {
SlotIndex prevStart;
if (LII != LR.begin())
prevStart = std::prev(LII)->start;
// FIXME: This could be more efficient if there was a
// removeSegment method that returned an iterator.
LR.removeSegment(*LII, true);
if (prevStart.isValid())
LII = LR.find(prevStart);
else
LII = LR.begin();
} else {
LII->start = instrIdx.getRegSlot();
LII->valno->def = instrIdx.getRegSlot();
if (MO.getSubReg() && !MO.isUndef())
lastUseIdx = instrIdx.getRegSlot();
else
lastUseIdx = SlotIndex();
continue;
}
}
if (!lastUseIdx.isValid()) {
VNInfo *VNI = LR.getNextValue(instrIdx.getRegSlot(), VNInfoAllocator);
LiveRange::Segment S(instrIdx.getRegSlot(),
instrIdx.getDeadSlot(), VNI);
LII = LR.addSegment(S);
} else if (LII->start != instrIdx.getRegSlot()) {
VNInfo *VNI = LR.getNextValue(instrIdx.getRegSlot(), VNInfoAllocator);
LiveRange::Segment S(instrIdx.getRegSlot(), lastUseIdx, VNI);
LII = LR.addSegment(S);
}
if (MO.getSubReg() && !MO.isUndef())
lastUseIdx = instrIdx.getRegSlot();
else
lastUseIdx = SlotIndex();
} else if (MO.isUse()) {
// FIXME: This should probably be handled outside of this branch,
// either as part of the def case (for defs inside of the region) or
// after the loop over the region.
if (!isEndValid && !LII->end.isBlock())
LII->end = instrIdx.getRegSlot();
if (!lastUseIdx.isValid())
lastUseIdx = instrIdx.getRegSlot();
}
}
}
}
void
LiveIntervals::repairIntervalsInRange(MachineBasicBlock *MBB,
MachineBasicBlock::iterator Begin,
MachineBasicBlock::iterator End,
ArrayRef<unsigned> OrigRegs) {
// Find anchor points, which are at the beginning/end of blocks or at
// instructions that already have indexes.
while (Begin != MBB->begin() && !Indexes->hasIndex(*Begin))
--Begin;
while (End != MBB->end() && !Indexes->hasIndex(*End))
++End;
SlotIndex endIdx;
if (End == MBB->end())
endIdx = getMBBEndIdx(MBB).getPrevSlot();
else
endIdx = getInstructionIndex(*End);
Indexes->repairIndexesInRange(MBB, Begin, End);
for (MachineBasicBlock::iterator I = End; I != Begin;) {
--I;
MachineInstr &MI = *I;
if (MI.isDebugValue())
continue;
for (MachineInstr::const_mop_iterator MOI = MI.operands_begin(),
MOE = MI.operands_end();
MOI != MOE; ++MOI) {
if (MOI->isReg() &&
TargetRegisterInfo::isVirtualRegister(MOI->getReg()) &&
!hasInterval(MOI->getReg())) {
createAndComputeVirtRegInterval(MOI->getReg());
}
}
}
for (unsigned Reg : OrigRegs) {
if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
LiveInterval &LI = getInterval(Reg);
// FIXME: Should we support undefs that gain defs?
if (!LI.hasAtLeastOneValue())
continue;
for (LiveInterval::SubRange &S : LI.subranges())
repairOldRegInRange(Begin, End, endIdx, S, Reg, S.LaneMask);
repairOldRegInRange(Begin, End, endIdx, LI, Reg);
}
}
void LiveIntervals::removePhysRegDefAt(unsigned Reg, SlotIndex Pos) {
for (MCRegUnitIterator Unit(Reg, TRI); Unit.isValid(); ++Unit) {
if (LiveRange *LR = getCachedRegUnit(*Unit))
if (VNInfo *VNI = LR->getVNInfoAt(Pos))
LR->removeValNo(VNI);
}
}
void LiveIntervals::removeVRegDefAt(LiveInterval &LI, SlotIndex Pos) {
// LI may not have the main range computed yet, but its subranges may
// be present.
VNInfo *VNI = LI.getVNInfoAt(Pos);
if (VNI != nullptr) {
assert(VNI->def.getBaseIndex() == Pos.getBaseIndex());
LI.removeValNo(VNI);
}
// Also remove the value defined in subranges.
for (LiveInterval::SubRange &S : LI.subranges()) {
if (VNInfo *SVNI = S.getVNInfoAt(Pos))
if (SVNI->def.getBaseIndex() == Pos.getBaseIndex())
S.removeValNo(SVNI);
}
LI.removeEmptySubRanges();
}
void LiveIntervals::splitSeparateComponents(LiveInterval &LI,
SmallVectorImpl<LiveInterval*> &SplitLIs) {
ConnectedVNInfoEqClasses ConEQ(*this);
unsigned NumComp = ConEQ.Classify(LI);
if (NumComp <= 1)
return;
DEBUG(dbgs() << " Split " << NumComp << " components: " << LI << '\n');
unsigned Reg = LI.reg;
const TargetRegisterClass *RegClass = MRI->getRegClass(Reg);
for (unsigned I = 1; I < NumComp; ++I) {
unsigned NewVReg = MRI->createVirtualRegister(RegClass);
LiveInterval &NewLI = createEmptyInterval(NewVReg);
SplitLIs.push_back(&NewLI);
}
ConEQ.Distribute(LI, SplitLIs.data(), *MRI);
}
void LiveIntervals::constructMainRangeFromSubranges(LiveInterval &LI) {
assert(LRCalc && "LRCalc not initialized.");
LRCalc->reset(MF, getSlotIndexes(), DomTree, &getVNInfoAllocator());
LRCalc->constructMainRangeFromSubranges(LI);
}