1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-25 20:23:11 +01:00
llvm-mirror/lib/CodeGen/LiveRangeShrink.cpp
Hongtao Yu 20454adfae [CSSPGO] Unblock optimizations with pseudo probe instrumentation.
The IR/MIR pseudo probe intrinsics don't get materialized into real machine instructions and therefore they don't incur runtime cost directly. However, they come with indirect cost by blocking certain optimizations. Some of the blocking are intentional (such as blocking code merge) for better counts quality while the others are accidental. This change unblocks perf-critical optimizations that do not affect counts quality. They include:

1. IR InstCombine, sinking load operation to shorten lifetimes.
2. MIR LiveRangeShrink, similar to #1
3. MIR TwoAddressInstructionPass, i.e, opeq transform
4. MIR function argument copy elision
5. IR stack protection. (though not perf-critical but nice to have).

Reviewed By: wmi

Differential Revision: https://reviews.llvm.org/D95982
2021-02-10 12:43:17 -08:00

247 lines
8.7 KiB
C++

//===- LiveRangeShrink.cpp - Move instructions to shrink live range -------===//
//
// 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
/// This pass moves instructions close to the definition of its operands to
/// shrink live range of the def instruction. The code motion is limited within
/// the basic block. The moved instruction should have 1 def, and more than one
/// uses, all of which are the only use of the def.
///
///===---------------------------------------------------------------------===//
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/iterator_range.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include <iterator>
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "lrshrink"
STATISTIC(NumInstrsHoistedToShrinkLiveRange,
"Number of insructions hoisted to shrink live range.");
namespace {
class LiveRangeShrink : public MachineFunctionPass {
public:
static char ID;
LiveRangeShrink() : MachineFunctionPass(ID) {
initializeLiveRangeShrinkPass(*PassRegistry::getPassRegistry());
}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
MachineFunctionPass::getAnalysisUsage(AU);
}
StringRef getPassName() const override { return "Live Range Shrink"; }
bool runOnMachineFunction(MachineFunction &MF) override;
};
} // end anonymous namespace
char LiveRangeShrink::ID = 0;
char &llvm::LiveRangeShrinkID = LiveRangeShrink::ID;
INITIALIZE_PASS(LiveRangeShrink, "lrshrink", "Live Range Shrink Pass", false,
false)
using InstOrderMap = DenseMap<MachineInstr *, unsigned>;
/// Returns \p New if it's dominated by \p Old, otherwise return \p Old.
/// \p M maintains a map from instruction to its dominating order that satisfies
/// M[A] > M[B] guarantees that A is dominated by B.
/// If \p New is not in \p M, return \p Old. Otherwise if \p Old is null, return
/// \p New.
static MachineInstr *FindDominatedInstruction(MachineInstr &New,
MachineInstr *Old,
const InstOrderMap &M) {
auto NewIter = M.find(&New);
if (NewIter == M.end())
return Old;
if (Old == nullptr)
return &New;
unsigned OrderOld = M.find(Old)->second;
unsigned OrderNew = NewIter->second;
if (OrderOld != OrderNew)
return OrderOld < OrderNew ? &New : Old;
// OrderOld == OrderNew, we need to iterate down from Old to see if it
// can reach New, if yes, New is dominated by Old.
for (MachineInstr *I = Old->getNextNode(); M.find(I)->second == OrderNew;
I = I->getNextNode())
if (I == &New)
return &New;
return Old;
}
/// Builds Instruction to its dominating order number map \p M by traversing
/// from instruction \p Start.
static void BuildInstOrderMap(MachineBasicBlock::iterator Start,
InstOrderMap &M) {
M.clear();
unsigned i = 0;
for (MachineInstr &I : make_range(Start, Start->getParent()->end()))
M[&I] = i++;
}
bool LiveRangeShrink::runOnMachineFunction(MachineFunction &MF) {
if (skipFunction(MF.getFunction()))
return false;
MachineRegisterInfo &MRI = MF.getRegInfo();
LLVM_DEBUG(dbgs() << "**** Analysing " << MF.getName() << '\n');
InstOrderMap IOM;
// Map from register to instruction order (value of IOM) where the
// register is used last. When moving instructions up, we need to
// make sure all its defs (including dead def) will not cross its
// last use when moving up.
DenseMap<unsigned, std::pair<unsigned, MachineInstr *>> UseMap;
for (MachineBasicBlock &MBB : MF) {
if (MBB.empty())
continue;
bool SawStore = false;
BuildInstOrderMap(MBB.begin(), IOM);
UseMap.clear();
for (MachineBasicBlock::iterator Next = MBB.begin(); Next != MBB.end();) {
MachineInstr &MI = *Next;
++Next;
if (MI.isPHI() || MI.isDebugInstr())
continue;
if (MI.mayStore())
SawStore = true;
unsigned CurrentOrder = IOM[&MI];
unsigned Barrier = 0;
MachineInstr *BarrierMI = nullptr;
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || MO.isDebug())
continue;
if (MO.isUse())
UseMap[MO.getReg()] = std::make_pair(CurrentOrder, &MI);
else if (MO.isDead() && UseMap.count(MO.getReg()))
// Barrier is the last instruction where MO get used. MI should not
// be moved above Barrier.
if (Barrier < UseMap[MO.getReg()].first) {
Barrier = UseMap[MO.getReg()].first;
BarrierMI = UseMap[MO.getReg()].second;
}
}
if (!MI.isSafeToMove(nullptr, SawStore)) {
// If MI has side effects, it should become a barrier for code motion.
// IOM is rebuild from the next instruction to prevent later
// instructions from being moved before this MI.
if (MI.hasUnmodeledSideEffects() && !MI.isPseudoProbe() &&
Next != MBB.end()) {
BuildInstOrderMap(Next, IOM);
SawStore = false;
}
continue;
}
const MachineOperand *DefMO = nullptr;
MachineInstr *Insert = nullptr;
// Number of live-ranges that will be shortened. We do not count
// live-ranges that are defined by a COPY as it could be coalesced later.
unsigned NumEligibleUse = 0;
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || MO.isDead() || MO.isDebug())
continue;
Register Reg = MO.getReg();
// Do not move the instruction if it def/uses a physical register,
// unless it is a constant physical register or a noreg.
if (!Register::isVirtualRegister(Reg)) {
if (!Reg || MRI.isConstantPhysReg(Reg))
continue;
Insert = nullptr;
break;
}
if (MO.isDef()) {
// Do not move if there is more than one def.
if (DefMO) {
Insert = nullptr;
break;
}
DefMO = &MO;
} else if (MRI.hasOneNonDBGUse(Reg) && MRI.hasOneDef(Reg) && DefMO &&
MRI.getRegClass(DefMO->getReg()) ==
MRI.getRegClass(MO.getReg())) {
// The heuristic does not handle different register classes yet
// (registers of different sizes, looser/tighter constraints). This
// is because it needs more accurate model to handle register
// pressure correctly.
MachineInstr &DefInstr = *MRI.def_instr_begin(Reg);
if (!DefInstr.isCopy())
NumEligibleUse++;
Insert = FindDominatedInstruction(DefInstr, Insert, IOM);
} else {
Insert = nullptr;
break;
}
}
// If Barrier equals IOM[I], traverse forward to find if BarrierMI is
// after Insert, if yes, then we should not hoist.
for (MachineInstr *I = Insert; I && IOM[I] == Barrier;
I = I->getNextNode())
if (I == BarrierMI) {
Insert = nullptr;
break;
}
// Move the instruction when # of shrunk live range > 1.
if (DefMO && Insert && NumEligibleUse > 1 && Barrier <= IOM[Insert]) {
MachineBasicBlock::iterator I = std::next(Insert->getIterator());
// Skip all the PHI and debug instructions.
while (I != MBB.end() && (I->isPHI() || I->isDebugInstr()))
I = std::next(I);
if (I == MI.getIterator())
continue;
// Update the dominator order to be the same as the insertion point.
// We do this to maintain a non-decreasing order without need to update
// all instruction orders after the insertion point.
unsigned NewOrder = IOM[&*I];
IOM[&MI] = NewOrder;
NumInstrsHoistedToShrinkLiveRange++;
// Find MI's debug value following MI.
MachineBasicBlock::iterator EndIter = std::next(MI.getIterator());
if (MI.getOperand(0).isReg())
for (; EndIter != MBB.end() && EndIter->isDebugValue() &&
EndIter->getDebugOperandForReg(MI.getOperand(0).getReg());
++EndIter, ++Next)
IOM[&*EndIter] = NewOrder;
MBB.splice(I, &MBB, MI.getIterator(), EndIter);
}
}
}
return false;
}