1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
|
//===-- SIPostRABundler.cpp -----------------------------------------------===//
//
// 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 creates bundles of memory instructions to protect adjacent loads
/// and stores from being rescheduled apart from each other post-RA.
///
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
using namespace llvm;
#define DEBUG_TYPE "si-post-ra-bundler"
namespace {
class SIPostRABundler : public MachineFunctionPass {
public:
static char ID;
public:
SIPostRABundler() : MachineFunctionPass(ID) {
initializeSIPostRABundlerPass(*PassRegistry::getPassRegistry());
}
bool runOnMachineFunction(MachineFunction &MF) override;
StringRef getPassName() const override {
return "SI post-RA bundler";
}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesAll();
MachineFunctionPass::getAnalysisUsage(AU);
}
private:
const SIRegisterInfo *TRI;
SmallSet<Register, 16> Defs;
void collectUsedRegUnits(const MachineInstr &MI,
BitVector &UsedRegUnits) const;
bool isBundleCandidate(const MachineInstr &MI) const;
bool isDependentLoad(const MachineInstr &MI) const;
bool canBundle(const MachineInstr &MI, const MachineInstr &NextMI) const;
};
constexpr uint64_t MemFlags = SIInstrFlags::MTBUF | SIInstrFlags::MUBUF |
SIInstrFlags::SMRD | SIInstrFlags::DS |
SIInstrFlags::FLAT | SIInstrFlags::MIMG;
} // End anonymous namespace.
INITIALIZE_PASS(SIPostRABundler, DEBUG_TYPE, "SI post-RA bundler", false, false)
char SIPostRABundler::ID = 0;
char &llvm::SIPostRABundlerID = SIPostRABundler::ID;
FunctionPass *llvm::createSIPostRABundlerPass() {
return new SIPostRABundler();
}
bool SIPostRABundler::isDependentLoad(const MachineInstr &MI) const {
if (!MI.mayLoad())
return false;
for (const MachineOperand &Op : MI.explicit_operands()) {
if (!Op.isReg())
continue;
Register Reg = Op.getReg();
for (Register Def : Defs)
if (TRI->regsOverlap(Reg, Def))
return true;
}
return false;
}
void SIPostRABundler::collectUsedRegUnits(const MachineInstr &MI,
BitVector &UsedRegUnits) const {
if (MI.isDebugInstr())
return;
for (const MachineOperand &Op : MI.operands()) {
if (!Op.isReg() || !Op.readsReg())
continue;
Register Reg = Op.getReg();
assert(!Op.getSubReg() &&
"subregister indexes should not be present after RA");
for (MCRegUnit Unit : TRI->regunits(Reg))
UsedRegUnits.set(Unit);
}
}
bool SIPostRABundler::isBundleCandidate(const MachineInstr &MI) const {
const uint64_t IMemFlags = MI.getDesc().TSFlags & MemFlags;
return IMemFlags != 0 && MI.mayLoadOrStore() && !MI.isBundled();
}
bool SIPostRABundler::canBundle(const MachineInstr &MI,
const MachineInstr &NextMI) const {
const uint64_t IMemFlags = MI.getDesc().TSFlags & MemFlags;
return (IMemFlags != 0 && MI.mayLoadOrStore() && !NextMI.isBundled() &&
NextMI.mayLoad() == MI.mayLoad() && NextMI.mayStore() == MI.mayStore() &&
((NextMI.getDesc().TSFlags & MemFlags) == IMemFlags) &&
!isDependentLoad(NextMI));
}
bool SIPostRABundler::runOnMachineFunction(MachineFunction &MF) {
if (skipFunction(MF.getFunction()))
return false;
TRI = MF.getSubtarget<GCNSubtarget>().getRegisterInfo();
BitVector BundleUsedRegUnits(TRI->getNumRegUnits());
BitVector KillUsedRegUnits(TRI->getNumRegUnits());
bool Changed = false;
for (MachineBasicBlock &MBB : MF) {
bool HasIGLPInstrs = llvm::any_of(MBB.instrs(), [](MachineInstr &MI) {
unsigned Opc = MI.getOpcode();
return Opc == AMDGPU::SCHED_GROUP_BARRIER || Opc == AMDGPU::IGLP_OPT;
});
// Don't cluster with IGLP instructions.
if (HasIGLPInstrs)
continue;
MachineBasicBlock::instr_iterator Next;
MachineBasicBlock::instr_iterator B = MBB.instr_begin();
MachineBasicBlock::instr_iterator E = MBB.instr_end();
for (auto I = B; I != E; I = Next) {
Next = std::next(I);
if (!isBundleCandidate(*I))
continue;
assert(Defs.empty());
if (I->getNumExplicitDefs() != 0)
Defs.insert(I->defs().begin()->getReg());
MachineBasicBlock::instr_iterator BundleStart = I;
MachineBasicBlock::instr_iterator BundleEnd = I;
unsigned ClauseLength = 1;
for (I = Next; I != E; I = Next) {
Next = std::next(I);
assert(BundleEnd != I);
if (canBundle(*BundleEnd, *I)) {
BundleEnd = I;
if (I->getNumExplicitDefs() != 0)
Defs.insert(I->defs().begin()->getReg());
++ClauseLength;
} else if (!I->isMetaInstruction()) {
// Allow meta instructions in between bundle candidates, but do not
// start or end a bundle on one.
//
// TODO: It may be better to move meta instructions like dbg_value
// after the bundle. We're relying on the memory legalizer to unbundle
// these.
break;
}
}
Next = std::next(BundleEnd);
if (ClauseLength > 1) {
Changed = true;
// Before register allocation, kills are inserted after potential soft
// clauses to hint register allocation. Look for kills that look like
// this, and erase them.
if (Next != E && Next->isKill()) {
// TODO: Should maybe back-propagate kill flags to the bundle.
for (const MachineInstr &BundleMI : make_range(BundleStart, Next))
collectUsedRegUnits(BundleMI, BundleUsedRegUnits);
BundleUsedRegUnits.flip();
while (Next != E && Next->isKill()) {
MachineInstr &Kill = *Next;
collectUsedRegUnits(Kill, KillUsedRegUnits);
KillUsedRegUnits &= BundleUsedRegUnits;
// Erase the kill if it's a subset of the used registers.
//
// TODO: Should we just remove all kills? Is there any real reason to
// keep them after RA?
if (KillUsedRegUnits.none()) {
++Next;
Kill.eraseFromParent();
} else
break;
KillUsedRegUnits.reset();
}
BundleUsedRegUnits.reset();
}
finalizeBundle(MBB, BundleStart, Next);
}
Defs.clear();
}
}
return Changed;
}
|