blob: 209f9f7255a5749683c3ad0e4216adf5a1e34fc9 [file] [log] [blame]
//===- AArch64FalkorHWPFFix.cpp - Avoid HW prefetcher pitfalls on Falkor --===//
//
// 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 For Falkor, we want to avoid HW prefetcher instruction tag collisions
/// that may inhibit the HW prefetching. This is done in two steps. Before
/// ISel, we mark strided loads (i.e. those that will likely benefit from
/// prefetching) with metadata. Then, after opcodes have been finalized, we
/// insert MOVs and re-write loads to prevent unintentional tag collisions.
// ===---------------------------------------------------------------------===//
#include "AArch64.h"
#include "AArch64InstrInfo.h"
#include "AArch64Subtarget.h"
#include "AArch64TargetMachine.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/None.h"
#include "llvm/ADT/Optional.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/CodeGen/LiveRegUnits.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Metadata.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/DebugCounter.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
#include <iterator>
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "aarch64-falkor-hwpf-fix"
STATISTIC(NumStridedLoadsMarked, "Number of strided loads marked");
STATISTIC(NumCollisionsAvoided,
"Number of HW prefetch tag collisions avoided");
STATISTIC(NumCollisionsNotAvoided,
"Number of HW prefetch tag collisions not avoided due to lack of registers");
DEBUG_COUNTER(FixCounter, "falkor-hwpf",
"Controls which tag collisions are avoided");
namespace {
class FalkorMarkStridedAccesses {
public:
FalkorMarkStridedAccesses(LoopInfo &LI, ScalarEvolution &SE)
: LI(LI), SE(SE) {}
bool run();
private:
bool runOnLoop(Loop &L);
LoopInfo &LI;
ScalarEvolution &SE;
};
class FalkorMarkStridedAccessesLegacy : public FunctionPass {
public:
static char ID; // Pass ID, replacement for typeid
FalkorMarkStridedAccessesLegacy() : FunctionPass(ID) {
initializeFalkorMarkStridedAccessesLegacyPass(
*PassRegistry::getPassRegistry());
}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<TargetPassConfig>();
AU.addPreserved<DominatorTreeWrapperPass>();
AU.addRequired<LoopInfoWrapperPass>();
AU.addPreserved<LoopInfoWrapperPass>();
AU.addRequired<ScalarEvolutionWrapperPass>();
AU.addPreserved<ScalarEvolutionWrapperPass>();
}
bool runOnFunction(Function &F) override;
};
} // end anonymous namespace
char FalkorMarkStridedAccessesLegacy::ID = 0;
INITIALIZE_PASS_BEGIN(FalkorMarkStridedAccessesLegacy, DEBUG_TYPE,
"Falkor HW Prefetch Fix", false, false)
INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
INITIALIZE_PASS_END(FalkorMarkStridedAccessesLegacy, DEBUG_TYPE,
"Falkor HW Prefetch Fix", false, false)
FunctionPass *llvm::createFalkorMarkStridedAccessesPass() {
return new FalkorMarkStridedAccessesLegacy();
}
bool FalkorMarkStridedAccessesLegacy::runOnFunction(Function &F) {
TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
const AArch64Subtarget *ST =
TPC.getTM<AArch64TargetMachine>().getSubtargetImpl(F);
if (ST->getProcFamily() != AArch64Subtarget::Falkor)
return false;
if (skipFunction(F))
return false;
LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
FalkorMarkStridedAccesses LDP(LI, SE);
return LDP.run();
}
bool FalkorMarkStridedAccesses::run() {
bool MadeChange = false;
for (Loop *L : LI)
for (auto LIt = df_begin(L), LE = df_end(L); LIt != LE; ++LIt)
MadeChange |= runOnLoop(**LIt);
return MadeChange;
}
bool FalkorMarkStridedAccesses::runOnLoop(Loop &L) {
// Only mark strided loads in the inner-most loop
if (!L.isInnermost())
return false;
bool MadeChange = false;
for (BasicBlock *BB : L.blocks()) {
for (Instruction &I : *BB) {
LoadInst *LoadI = dyn_cast<LoadInst>(&I);
if (!LoadI)
continue;
Value *PtrValue = LoadI->getPointerOperand();
if (L.isLoopInvariant(PtrValue))
continue;
const SCEV *LSCEV = SE.getSCEV(PtrValue);
const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
if (!LSCEVAddRec || !LSCEVAddRec->isAffine())
continue;
LoadI->setMetadata(FALKOR_STRIDED_ACCESS_MD,
MDNode::get(LoadI->getContext(), {}));
++NumStridedLoadsMarked;
LLVM_DEBUG(dbgs() << "Load: " << I << " marked as strided\n");
MadeChange = true;
}
}
return MadeChange;
}
namespace {
class FalkorHWPFFix : public MachineFunctionPass {
public:
static char ID;
FalkorHWPFFix() : MachineFunctionPass(ID) {
initializeFalkorHWPFFixPass(*PassRegistry::getPassRegistry());
}
bool runOnMachineFunction(MachineFunction &Fn) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
AU.addRequired<MachineLoopInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
MachineFunctionProperties getRequiredProperties() const override {
return MachineFunctionProperties().set(
MachineFunctionProperties::Property::NoVRegs);
}
private:
void runOnLoop(MachineLoop &L, MachineFunction &Fn);
const AArch64InstrInfo *TII;
const TargetRegisterInfo *TRI;
DenseMap<unsigned, SmallVector<MachineInstr *, 4>> TagMap;
bool Modified;
};
/// Bits from load opcodes used to compute HW prefetcher instruction tags.
struct LoadInfo {
LoadInfo() = default;
Register DestReg;
Register BaseReg;
int BaseRegIdx = -1;
const MachineOperand *OffsetOpnd = nullptr;
bool IsPrePost = false;
};
} // end anonymous namespace
char FalkorHWPFFix::ID = 0;
INITIALIZE_PASS_BEGIN(FalkorHWPFFix, "aarch64-falkor-hwpf-fix-late",
"Falkor HW Prefetch Fix Late Phase", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
INITIALIZE_PASS_END(FalkorHWPFFix, "aarch64-falkor-hwpf-fix-late",
"Falkor HW Prefetch Fix Late Phase", false, false)
static unsigned makeTag(unsigned Dest, unsigned Base, unsigned Offset) {
return (Dest & 0xf) | ((Base & 0xf) << 4) | ((Offset & 0x3f) << 8);
}
static Optional<LoadInfo> getLoadInfo(const MachineInstr &MI) {
int DestRegIdx;
int BaseRegIdx;
int OffsetIdx;
bool IsPrePost;
switch (MI.getOpcode()) {
default:
return None;
case AArch64::LD1i64:
case AArch64::LD2i64:
DestRegIdx = 0;
BaseRegIdx = 3;
OffsetIdx = -1;
IsPrePost = false;
break;
case AArch64::LD1i8:
case AArch64::LD1i16:
case AArch64::LD1i32:
case AArch64::LD2i8:
case AArch64::LD2i16:
case AArch64::LD2i32:
case AArch64::LD3i8:
case AArch64::LD3i16:
case AArch64::LD3i32:
case AArch64::LD3i64:
case AArch64::LD4i8:
case AArch64::LD4i16:
case AArch64::LD4i32:
case AArch64::LD4i64:
DestRegIdx = -1;
BaseRegIdx = 3;
OffsetIdx = -1;
IsPrePost = false;
break;
case AArch64::LD1Onev1d:
case AArch64::LD1Onev2s:
case AArch64::LD1Onev4h:
case AArch64::LD1Onev8b:
case AArch64::LD1Onev2d:
case AArch64::LD1Onev4s:
case AArch64::LD1Onev8h:
case AArch64::LD1Onev16b:
case AArch64::LD1Rv1d:
case AArch64::LD1Rv2s:
case AArch64::LD1Rv4h:
case AArch64::LD1Rv8b:
case AArch64::LD1Rv2d:
case AArch64::LD1Rv4s:
case AArch64::LD1Rv8h:
case AArch64::LD1Rv16b:
DestRegIdx = 0;
BaseRegIdx = 1;
OffsetIdx = -1;
IsPrePost = false;
break;
case AArch64::LD1Twov1d:
case AArch64::LD1Twov2s:
case AArch64::LD1Twov4h:
case AArch64::LD1Twov8b:
case AArch64::LD1Twov2d:
case AArch64::LD1Twov4s:
case AArch64::LD1Twov8h:
case AArch64::LD1Twov16b:
case AArch64::LD1Threev1d:
case AArch64::LD1Threev2s:
case AArch64::LD1Threev4h:
case AArch64::LD1Threev8b:
case AArch64::LD1Threev2d:
case AArch64::LD1Threev4s:
case AArch64::LD1Threev8h:
case AArch64::LD1Threev16b:
case AArch64::LD1Fourv1d:
case AArch64::LD1Fourv2s:
case AArch64::LD1Fourv4h:
case AArch64::LD1Fourv8b:
case AArch64::LD1Fourv2d:
case AArch64::LD1Fourv4s:
case AArch64::LD1Fourv8h:
case AArch64::LD1Fourv16b:
case AArch64::LD2Twov2s:
case AArch64::LD2Twov4s:
case AArch64::LD2Twov8b:
case AArch64::LD2Twov2d:
case AArch64::LD2Twov4h:
case AArch64::LD2Twov8h:
case AArch64::LD2Twov16b:
case AArch64::LD2Rv1d:
case AArch64::LD2Rv2s:
case AArch64::LD2Rv4s:
case AArch64::LD2Rv8b:
case AArch64::LD2Rv2d:
case AArch64::LD2Rv4h:
case AArch64::LD2Rv8h:
case AArch64::LD2Rv16b:
case AArch64::LD3Threev2s:
case AArch64::LD3Threev4h:
case AArch64::LD3Threev8b:
case AArch64::LD3Threev2d:
case AArch64::LD3Threev4s:
case AArch64::LD3Threev8h:
case AArch64::LD3Threev16b:
case AArch64::LD3Rv1d:
case AArch64::LD3Rv2s:
case AArch64::LD3Rv4h:
case AArch64::LD3Rv8b:
case AArch64::LD3Rv2d:
case AArch64::LD3Rv4s:
case AArch64::LD3Rv8h:
case AArch64::LD3Rv16b:
case AArch64::LD4Fourv2s:
case AArch64::LD4Fourv4h:
case AArch64::LD4Fourv8b:
case AArch64::LD4Fourv2d:
case AArch64::LD4Fourv4s:
case AArch64::LD4Fourv8h:
case AArch64::LD4Fourv16b:
case AArch64::LD4Rv1d:
case AArch64::LD4Rv2s:
case AArch64::LD4Rv4h:
case AArch64::LD4Rv8b:
case AArch64::LD4Rv2d:
case AArch64::LD4Rv4s:
case AArch64::LD4Rv8h:
case AArch64::LD4Rv16b:
DestRegIdx = -1;
BaseRegIdx = 1;
OffsetIdx = -1;
IsPrePost = false;
break;
case AArch64::LD1i64_POST:
case AArch64::LD2i64_POST:
DestRegIdx = 1;
BaseRegIdx = 4;
OffsetIdx = 5;
IsPrePost = true;
break;
case AArch64::LD1i8_POST:
case AArch64::LD1i16_POST:
case AArch64::LD1i32_POST:
case AArch64::LD2i8_POST:
case AArch64::LD2i16_POST:
case AArch64::LD2i32_POST:
case AArch64::LD3i8_POST:
case AArch64::LD3i16_POST:
case AArch64::LD3i32_POST:
case AArch64::LD3i64_POST:
case AArch64::LD4i8_POST:
case AArch64::LD4i16_POST:
case AArch64::LD4i32_POST:
case AArch64::LD4i64_POST:
DestRegIdx = -1;
BaseRegIdx = 4;
OffsetIdx = 5;
IsPrePost = true;
break;
case AArch64::LD1Onev1d_POST:
case AArch64::LD1Onev2s_POST:
case AArch64::LD1Onev4h_POST:
case AArch64::LD1Onev8b_POST:
case AArch64::LD1Onev2d_POST:
case AArch64::LD1Onev4s_POST:
case AArch64::LD1Onev8h_POST:
case AArch64::LD1Onev16b_POST:
case AArch64::LD1Rv1d_POST:
case AArch64::LD1Rv2s_POST:
case AArch64::LD1Rv4h_POST:
case AArch64::LD1Rv8b_POST:
case AArch64::LD1Rv2d_POST:
case AArch64::LD1Rv4s_POST:
case AArch64::LD1Rv8h_POST:
case AArch64::LD1Rv16b_POST:
DestRegIdx = 1;
BaseRegIdx = 2;
OffsetIdx = 3;
IsPrePost = true;
break;
case AArch64::LD1Twov1d_POST:
case AArch64::LD1Twov2s_POST:
case AArch64::LD1Twov4h_POST:
case AArch64::LD1Twov8b_POST:
case AArch64::LD1Twov2d_POST:
case AArch64::LD1Twov4s_POST:
case AArch64::LD1Twov8h_POST:
case AArch64::LD1Twov16b_POST:
case AArch64::LD1Threev1d_POST:
case AArch64::LD1Threev2s_POST:
case AArch64::LD1Threev4h_POST:
case AArch64::LD1Threev8b_POST:
case AArch64::LD1Threev2d_POST:
case AArch64::LD1Threev4s_POST:
case AArch64::LD1Threev8h_POST:
case AArch64::LD1Threev16b_POST:
case AArch64::LD1Fourv1d_POST:
case AArch64::LD1Fourv2s_POST:
case AArch64::LD1Fourv4h_POST:
case AArch64::LD1Fourv8b_POST:
case AArch64::LD1Fourv2d_POST:
case AArch64::LD1Fourv4s_POST:
case AArch64::LD1Fourv8h_POST:
case AArch64::LD1Fourv16b_POST:
case AArch64::LD2Twov2s_POST:
case AArch64::LD2Twov4s_POST:
case AArch64::LD2Twov8b_POST:
case AArch64::LD2Twov2d_POST:
case AArch64::LD2Twov4h_POST:
case AArch64::LD2Twov8h_POST:
case AArch64::LD2Twov16b_POST:
case AArch64::LD2Rv1d_POST:
case AArch64::LD2Rv2s_POST:
case AArch64::LD2Rv4s_POST:
case AArch64::LD2Rv8b_POST:
case AArch64::LD2Rv2d_POST:
case AArch64::LD2Rv4h_POST:
case AArch64::LD2Rv8h_POST:
case AArch64::LD2Rv16b_POST:
case AArch64::LD3Threev2s_POST:
case AArch64::LD3Threev4h_POST:
case AArch64::LD3Threev8b_POST:
case AArch64::LD3Threev2d_POST:
case AArch64::LD3Threev4s_POST:
case AArch64::LD3Threev8h_POST:
case AArch64::LD3Threev16b_POST:
case AArch64::LD3Rv1d_POST:
case AArch64::LD3Rv2s_POST:
case AArch64::LD3Rv4h_POST:
case AArch64::LD3Rv8b_POST:
case AArch64::LD3Rv2d_POST:
case AArch64::LD3Rv4s_POST:
case AArch64::LD3Rv8h_POST:
case AArch64::LD3Rv16b_POST:
case AArch64::LD4Fourv2s_POST:
case AArch64::LD4Fourv4h_POST:
case AArch64::LD4Fourv8b_POST:
case AArch64::LD4Fourv2d_POST:
case AArch64::LD4Fourv4s_POST:
case AArch64::LD4Fourv8h_POST:
case AArch64::LD4Fourv16b_POST:
case AArch64::LD4Rv1d_POST:
case AArch64::LD4Rv2s_POST:
case AArch64::LD4Rv4h_POST:
case AArch64::LD4Rv8b_POST:
case AArch64::LD4Rv2d_POST:
case AArch64::LD4Rv4s_POST:
case AArch64::LD4Rv8h_POST:
case AArch64::LD4Rv16b_POST:
DestRegIdx = -1;
BaseRegIdx = 2;
OffsetIdx = 3;
IsPrePost = true;
break;
case AArch64::LDRBBroW:
case AArch64::LDRBBroX:
case AArch64::LDRBBui:
case AArch64::LDRBroW:
case AArch64::LDRBroX:
case AArch64::LDRBui:
case AArch64::LDRDl:
case AArch64::LDRDroW:
case AArch64::LDRDroX:
case AArch64::LDRDui:
case AArch64::LDRHHroW:
case AArch64::LDRHHroX:
case AArch64::LDRHHui:
case AArch64::LDRHroW:
case AArch64::LDRHroX:
case AArch64::LDRHui:
case AArch64::LDRQl:
case AArch64::LDRQroW:
case AArch64::LDRQroX:
case AArch64::LDRQui:
case AArch64::LDRSBWroW:
case AArch64::LDRSBWroX:
case AArch64::LDRSBWui:
case AArch64::LDRSBXroW:
case AArch64::LDRSBXroX:
case AArch64::LDRSBXui:
case AArch64::LDRSHWroW:
case AArch64::LDRSHWroX:
case AArch64::LDRSHWui:
case AArch64::LDRSHXroW:
case AArch64::LDRSHXroX:
case AArch64::LDRSHXui:
case AArch64::LDRSWl:
case AArch64::LDRSWroW:
case AArch64::LDRSWroX:
case AArch64::LDRSWui:
case AArch64::LDRSl:
case AArch64::LDRSroW:
case AArch64::LDRSroX:
case AArch64::LDRSui:
case AArch64::LDRWl:
case AArch64::LDRWroW:
case AArch64::LDRWroX:
case AArch64::LDRWui:
case AArch64::LDRXl:
case AArch64::LDRXroW:
case AArch64::LDRXroX:
case AArch64::LDRXui:
case AArch64::LDURBBi:
case AArch64::LDURBi:
case AArch64::LDURDi:
case AArch64::LDURHHi:
case AArch64::LDURHi:
case AArch64::LDURQi:
case AArch64::LDURSBWi:
case AArch64::LDURSBXi:
case AArch64::LDURSHWi:
case AArch64::LDURSHXi:
case AArch64::LDURSWi:
case AArch64::LDURSi:
case AArch64::LDURWi:
case AArch64::LDURXi:
DestRegIdx = 0;
BaseRegIdx = 1;
OffsetIdx = 2;
IsPrePost = false;
break;
case AArch64::LDRBBpost:
case AArch64::LDRBBpre:
case AArch64::LDRBpost:
case AArch64::LDRBpre:
case AArch64::LDRDpost:
case AArch64::LDRDpre:
case AArch64::LDRHHpost:
case AArch64::LDRHHpre:
case AArch64::LDRHpost:
case AArch64::LDRHpre:
case AArch64::LDRQpost:
case AArch64::LDRQpre:
case AArch64::LDRSBWpost:
case AArch64::LDRSBWpre:
case AArch64::LDRSBXpost:
case AArch64::LDRSBXpre:
case AArch64::LDRSHWpost:
case AArch64::LDRSHWpre:
case AArch64::LDRSHXpost:
case AArch64::LDRSHXpre:
case AArch64::LDRSWpost:
case AArch64::LDRSWpre:
case AArch64::LDRSpost:
case AArch64::LDRSpre:
case AArch64::LDRWpost:
case AArch64::LDRWpre:
case AArch64::LDRXpost:
case AArch64::LDRXpre:
DestRegIdx = 1;
BaseRegIdx = 2;
OffsetIdx = 3;
IsPrePost = true;
break;
case AArch64::LDNPDi:
case AArch64::LDNPQi:
case AArch64::LDNPSi:
case AArch64::LDPQi:
case AArch64::LDPDi:
case AArch64::LDPSi:
DestRegIdx = -1;
BaseRegIdx = 2;
OffsetIdx = 3;
IsPrePost = false;
break;
case AArch64::LDPSWi:
case AArch64::LDPWi:
case AArch64::LDPXi:
DestRegIdx = 0;
BaseRegIdx = 2;
OffsetIdx = 3;
IsPrePost = false;
break;
case AArch64::LDPQpost:
case AArch64::LDPQpre:
case AArch64::LDPDpost:
case AArch64::LDPDpre:
case AArch64::LDPSpost:
case AArch64::LDPSpre:
DestRegIdx = -1;
BaseRegIdx = 3;
OffsetIdx = 4;
IsPrePost = true;
break;
case AArch64::LDPSWpost:
case AArch64::LDPSWpre:
case AArch64::LDPWpost:
case AArch64::LDPWpre:
case AArch64::LDPXpost:
case AArch64::LDPXpre:
DestRegIdx = 1;
BaseRegIdx = 3;
OffsetIdx = 4;
IsPrePost = true;
break;
}
// Loads from the stack pointer don't get prefetched.
Register BaseReg = MI.getOperand(BaseRegIdx).getReg();
if (BaseReg == AArch64::SP || BaseReg == AArch64::WSP)
return None;
LoadInfo LI;
LI.DestReg = DestRegIdx == -1 ? Register() : MI.getOperand(DestRegIdx).getReg();
LI.BaseReg = BaseReg;
LI.BaseRegIdx = BaseRegIdx;
LI.OffsetOpnd = OffsetIdx == -1 ? nullptr : &MI.getOperand(OffsetIdx);
LI.IsPrePost = IsPrePost;
return LI;
}
static Optional<unsigned> getTag(const TargetRegisterInfo *TRI,
const MachineInstr &MI, const LoadInfo &LI) {
unsigned Dest = LI.DestReg ? TRI->getEncodingValue(LI.DestReg) : 0;
unsigned Base = TRI->getEncodingValue(LI.BaseReg);
unsigned Off;
if (LI.OffsetOpnd == nullptr)
Off = 0;
else if (LI.OffsetOpnd->isGlobal() || LI.OffsetOpnd->isSymbol() ||
LI.OffsetOpnd->isCPI())
return None;
else if (LI.OffsetOpnd->isReg())
Off = (1 << 5) | TRI->getEncodingValue(LI.OffsetOpnd->getReg());
else
Off = LI.OffsetOpnd->getImm() >> 2;
return makeTag(Dest, Base, Off);
}
void FalkorHWPFFix::runOnLoop(MachineLoop &L, MachineFunction &Fn) {
// Build the initial tag map for the whole loop.
TagMap.clear();
for (MachineBasicBlock *MBB : L.getBlocks())
for (MachineInstr &MI : *MBB) {
Optional<LoadInfo> LInfo = getLoadInfo(MI);
if (!LInfo)
continue;
Optional<unsigned> Tag = getTag(TRI, MI, *LInfo);
if (!Tag)
continue;
TagMap[*Tag].push_back(&MI);
}
bool AnyCollisions = false;
for (auto &P : TagMap) {
auto Size = P.second.size();
if (Size > 1) {
for (auto *MI : P.second) {
if (TII->isStridedAccess(*MI)) {
AnyCollisions = true;
break;
}
}
}
if (AnyCollisions)
break;
}
// Nothing to fix.
if (!AnyCollisions)
return;
MachineRegisterInfo &MRI = Fn.getRegInfo();
// Go through all the basic blocks in the current loop and fix any streaming
// loads to avoid collisions with any other loads.
LiveRegUnits LR(*TRI);
for (MachineBasicBlock *MBB : L.getBlocks()) {
LR.clear();
LR.addLiveOuts(*MBB);
for (auto I = MBB->rbegin(); I != MBB->rend(); LR.stepBackward(*I), ++I) {
MachineInstr &MI = *I;
if (!TII->isStridedAccess(MI))
continue;
Optional<LoadInfo> OptLdI = getLoadInfo(MI);
if (!OptLdI)
continue;
LoadInfo LdI = *OptLdI;
Optional<unsigned> OptOldTag = getTag(TRI, MI, LdI);
if (!OptOldTag)
continue;
auto &OldCollisions = TagMap[*OptOldTag];
if (OldCollisions.size() <= 1)
continue;
bool Fixed = false;
LLVM_DEBUG(dbgs() << "Attempting to fix tag collision: " << MI);
if (!DebugCounter::shouldExecute(FixCounter)) {
LLVM_DEBUG(dbgs() << "Skipping fix due to debug counter:\n " << MI);
continue;
}
// Add the non-base registers of MI as live so we don't use them as
// scratch registers.
for (unsigned OpI = 0, OpE = MI.getNumOperands(); OpI < OpE; ++OpI) {
if (OpI == static_cast<unsigned>(LdI.BaseRegIdx))
continue;
MachineOperand &MO = MI.getOperand(OpI);
if (MO.isReg() && MO.readsReg())
LR.addReg(MO.getReg());
}
for (unsigned ScratchReg : AArch64::GPR64RegClass) {
if (!LR.available(ScratchReg) || MRI.isReserved(ScratchReg))
continue;
LoadInfo NewLdI(LdI);
NewLdI.BaseReg = ScratchReg;
unsigned NewTag = *getTag(TRI, MI, NewLdI);
// Scratch reg tag would collide too, so don't use it.
if (TagMap.count(NewTag))
continue;
LLVM_DEBUG(dbgs() << "Changing base reg to: "
<< printReg(ScratchReg, TRI) << '\n');
// Rewrite:
// Xd = LOAD Xb, off
// to:
// Xc = MOV Xb
// Xd = LOAD Xc, off
DebugLoc DL = MI.getDebugLoc();
BuildMI(*MBB, &MI, DL, TII->get(AArch64::ORRXrs), ScratchReg)
.addReg(AArch64::XZR)
.addReg(LdI.BaseReg)
.addImm(0);
MachineOperand &BaseOpnd = MI.getOperand(LdI.BaseRegIdx);
BaseOpnd.setReg(ScratchReg);
// If the load does a pre/post increment, then insert a MOV after as
// well to update the real base register.
if (LdI.IsPrePost) {
LLVM_DEBUG(dbgs() << "Doing post MOV of incremented reg: "
<< printReg(ScratchReg, TRI) << '\n');
MI.getOperand(0).setReg(
ScratchReg); // Change tied operand pre/post update dest.
BuildMI(*MBB, std::next(MachineBasicBlock::iterator(MI)), DL,
TII->get(AArch64::ORRXrs), LdI.BaseReg)
.addReg(AArch64::XZR)
.addReg(ScratchReg)
.addImm(0);
}
for (int I = 0, E = OldCollisions.size(); I != E; ++I)
if (OldCollisions[I] == &MI) {
std::swap(OldCollisions[I], OldCollisions[E - 1]);
OldCollisions.pop_back();
break;
}
// Update TagMap to reflect instruction changes to reduce the number
// of later MOVs to be inserted. This needs to be done after
// OldCollisions is updated since it may be relocated by this
// insertion.
TagMap[NewTag].push_back(&MI);
++NumCollisionsAvoided;
Fixed = true;
Modified = true;
break;
}
if (!Fixed)
++NumCollisionsNotAvoided;
}
}
}
bool FalkorHWPFFix::runOnMachineFunction(MachineFunction &Fn) {
auto &ST = static_cast<const AArch64Subtarget &>(Fn.getSubtarget());
if (ST.getProcFamily() != AArch64Subtarget::Falkor)
return false;
if (skipFunction(Fn.getFunction()))
return false;
TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo());
TRI = ST.getRegisterInfo();
MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
Modified = false;
for (MachineLoop *I : LI)
for (auto L = df_begin(I), LE = df_end(I); L != LE; ++L)
// Only process inner-loops
if (L->isInnermost())
runOnLoop(**L, Fn);
return Modified;
}
FunctionPass *llvm::createFalkorHWPFFixPass() { return new FalkorHWPFFix(); }