1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 03:02:36 +01:00
llvm-mirror/lib/CodeGen/HardwareLoops.cpp

531 lines
19 KiB
C++
Raw Normal View History

//===-- HardwareLoops.cpp - Target Independent Hardware Loops --*- C++ -*-===//
//
// 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
/// Insert hardware loop intrinsics into loops which are deemed profitable by
/// the target, by querying TargetTransformInfo. A hardware loop comprises of
/// two intrinsics: one, outside the loop, to set the loop iteration count and
/// another, in the exit block, to decrement the counter. The decremented value
/// can either be carried through the loop via a phi or handled in some opaque
/// way by the target.
///
//===----------------------------------------------------------------------===//
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Value.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/PassRegistry.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/LoopUtils.h"
#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
#define DEBUG_TYPE "hardware-loops"
#define HW_LOOPS_NAME "Hardware Loop Insertion"
using namespace llvm;
static cl::opt<bool>
ForceHardwareLoops("force-hardware-loops", cl::Hidden, cl::init(false),
cl::desc("Force hardware loops intrinsics to be inserted"));
static cl::opt<bool>
ForceHardwareLoopPHI(
"force-hardware-loop-phi", cl::Hidden, cl::init(false),
cl::desc("Force hardware loop counter to be updated through a phi"));
static cl::opt<bool>
ForceNestedLoop("force-nested-hardware-loop", cl::Hidden, cl::init(false),
cl::desc("Force allowance of nested hardware loops"));
static cl::opt<unsigned>
LoopDecrement("hardware-loop-decrement", cl::Hidden, cl::init(1),
cl::desc("Set the loop decrement value"));
static cl::opt<unsigned>
CounterBitWidth("hardware-loop-counter-bitwidth", cl::Hidden, cl::init(32),
cl::desc("Set the loop counter bitwidth"));
static cl::opt<bool>
ForceGuardLoopEntry(
"force-hardware-loop-guard", cl::Hidden, cl::init(false),
cl::desc("Force generation of loop guard intrinsic"));
STATISTIC(NumHWLoops, "Number of loops converted to hardware loops");
#ifndef NDEBUG
static void debugHWLoopFailure(const StringRef DebugMsg,
Instruction *I) {
dbgs() << "HWLoops: " << DebugMsg;
if (I)
dbgs() << ' ' << *I;
else
dbgs() << '.';
dbgs() << '\n';
}
#endif
static OptimizationRemarkAnalysis
createHWLoopAnalysis(StringRef RemarkName, Loop *L, Instruction *I) {
Value *CodeRegion = L->getHeader();
DebugLoc DL = L->getStartLoc();
if (I) {
CodeRegion = I->getParent();
// If there is no debug location attached to the instruction, revert back to
// using the loop's.
if (I->getDebugLoc())
DL = I->getDebugLoc();
}
OptimizationRemarkAnalysis R(DEBUG_TYPE, RemarkName, DL, CodeRegion);
R << "hardware-loop not created: ";
return R;
}
namespace {
void reportHWLoopFailure(const StringRef Msg, const StringRef ORETag,
OptimizationRemarkEmitter *ORE, Loop *TheLoop, Instruction *I = nullptr) {
LLVM_DEBUG(debugHWLoopFailure(Msg, I));
ORE->emit(createHWLoopAnalysis(ORETag, TheLoop, I) << Msg);
}
using TTI = TargetTransformInfo;
class HardwareLoops : public FunctionPass {
public:
static char ID;
HardwareLoops() : FunctionPass(ID) {
initializeHardwareLoopsPass(*PassRegistry::getPassRegistry());
}
bool runOnFunction(Function &F) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<LoopInfoWrapperPass>();
AU.addPreserved<LoopInfoWrapperPass>();
AU.addRequired<DominatorTreeWrapperPass>();
AU.addPreserved<DominatorTreeWrapperPass>();
AU.addRequired<ScalarEvolutionWrapperPass>();
AU.addRequired<AssumptionCacheTracker>();
AU.addRequired<TargetTransformInfoWrapperPass>();
AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
}
// Try to convert the given Loop into a hardware loop.
bool TryConvertLoop(Loop *L);
// Given that the target believes the loop to be profitable, try to
// convert it.
bool TryConvertLoop(HardwareLoopInfo &HWLoopInfo);
private:
ScalarEvolution *SE = nullptr;
LoopInfo *LI = nullptr;
const DataLayout *DL = nullptr;
OptimizationRemarkEmitter *ORE = nullptr;
const TargetTransformInfo *TTI = nullptr;
DominatorTree *DT = nullptr;
bool PreserveLCSSA = false;
AssumptionCache *AC = nullptr;
TargetLibraryInfo *LibInfo = nullptr;
Module *M = nullptr;
bool MadeChange = false;
};
class HardwareLoop {
// Expand the trip count scev into a value that we can use.
Value *InitLoopCount();
// Insert the set_loop_iteration intrinsic.
[ARM] Alter t2DoLoopStart to define lr This changes the definition of t2DoLoopStart from t2DoLoopStart rGPR to GPRlr = t2DoLoopStart rGPR This will hopefully mean that low overhead loops are more tied together, and we can more reliably generate loops without reverting or being at the whims of the register allocator. This is a fairly simple change in itself, but leads to a number of other required alterations. - The hardware loop pass, if UsePhi is set, now generates loops of the form: %start = llvm.start.loop.iterations(%N) loop: %p = phi [%start], [%dec] %dec = llvm.loop.decrement.reg(%p, 1) %c = icmp ne %dec, 0 br %c, loop, exit - For this a new llvm.start.loop.iterations intrinsic was added, identical to llvm.set.loop.iterations but produces a value as seen above, gluing the loop together more through def-use chains. - This new instrinsic conceptually produces the same output as input, which is taught to SCEV so that the checks in MVETailPredication are not affected. - Some minor changes are needed to the ARMLowOverheadLoop pass, but it has been left mostly as before. We should now more reliably be able to tell that the t2DoLoopStart is correct without having to prove it, but t2WhileLoopStart and tail-predicated loops will remain the same. - And all the tests have been updated. There are a lot of them! This patch on it's own might cause more trouble that it helps, with more tail-predicated loops being reverted, but some additional patches can hopefully improve upon that to get to something that is better overall. Differential Revision: https://reviews.llvm.org/D89881
2020-11-10 16:57:58 +01:00
Value *InsertIterationSetup(Value *LoopCountInit);
// Insert the loop_decrement intrinsic.
void InsertLoopDec();
// Insert the loop_decrement_reg intrinsic.
Instruction *InsertLoopRegDec(Value *EltsRem);
// If the target requires the counter value to be updated in the loop,
// insert a phi to hold the value. The intended purpose is for use by
// loop_decrement_reg.
PHINode *InsertPHICounter(Value *NumElts, Value *EltsRem);
// Create a new cmp, that checks the returned value of loop_decrement*,
// and update the exit branch to use it.
void UpdateBranch(Value *EltsRem);
public:
HardwareLoop(HardwareLoopInfo &Info, ScalarEvolution &SE,
const DataLayout &DL,
OptimizationRemarkEmitter *ORE) :
SE(SE), DL(DL), ORE(ORE), L(Info.L), M(L->getHeader()->getModule()),
TripCount(Info.TripCount),
CountType(Info.CountType),
ExitBranch(Info.ExitBranch),
LoopDecrement(Info.LoopDecrement),
UsePHICounter(Info.CounterInReg),
UseLoopGuard(Info.PerformEntryTest) { }
void Create();
private:
ScalarEvolution &SE;
const DataLayout &DL;
OptimizationRemarkEmitter *ORE = nullptr;
Loop *L = nullptr;
Module *M = nullptr;
const SCEV *TripCount = nullptr;
Type *CountType = nullptr;
BranchInst *ExitBranch = nullptr;
Value *LoopDecrement = nullptr;
bool UsePHICounter = false;
bool UseLoopGuard = false;
BasicBlock *BeginBB = nullptr;
};
}
char HardwareLoops::ID = 0;
bool HardwareLoops::runOnFunction(Function &F) {
if (skipFunction(F))
return false;
LLVM_DEBUG(dbgs() << "HWLoops: Running on " << F.getName() << "\n");
LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
DL = &F.getParent()->getDataLayout();
ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
LibInfo = TLIP ? &TLIP->getTLI(F) : nullptr;
PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
M = F.getParent();
for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
Loop *L = *I;
if (L->isOutermost())
TryConvertLoop(L);
}
return MadeChange;
}
// Return true if the search should stop, which will be when an inner loop is
// converted and the parent loop doesn't support containing a hardware loop.
bool HardwareLoops::TryConvertLoop(Loop *L) {
// Process nested loops first.
bool AnyChanged = false;
for (Loop *SL : *L)
AnyChanged |= TryConvertLoop(SL);
if (AnyChanged) {
reportHWLoopFailure("nested hardware-loops not supported", "HWLoopNested",
ORE, L);
return true; // Stop search.
}
LLVM_DEBUG(dbgs() << "HWLoops: Loop " << L->getHeader()->getName() << "\n");
HardwareLoopInfo HWLoopInfo(L);
if (!HWLoopInfo.canAnalyze(*LI)) {
reportHWLoopFailure("cannot analyze loop, irreducible control flow",
"HWLoopCannotAnalyze", ORE, L);
return false;
}
if (!ForceHardwareLoops &&
!TTI->isHardwareLoopProfitable(L, *SE, *AC, LibInfo, HWLoopInfo)) {
reportHWLoopFailure("it's not profitable to create a hardware-loop",
"HWLoopNotProfitable", ORE, L);
return false;
}
// Allow overriding of the counter width and loop decrement value.
if (CounterBitWidth.getNumOccurrences())
HWLoopInfo.CountType =
IntegerType::get(M->getContext(), CounterBitWidth);
if (LoopDecrement.getNumOccurrences())
HWLoopInfo.LoopDecrement =
ConstantInt::get(HWLoopInfo.CountType, LoopDecrement);
MadeChange |= TryConvertLoop(HWLoopInfo);
return MadeChange && (!HWLoopInfo.IsNestingLegal && !ForceNestedLoop);
}
bool HardwareLoops::TryConvertLoop(HardwareLoopInfo &HWLoopInfo) {
Loop *L = HWLoopInfo.L;
LLVM_DEBUG(dbgs() << "HWLoops: Try to convert profitable loop: " << *L);
if (!HWLoopInfo.isHardwareLoopCandidate(*SE, *LI, *DT, ForceNestedLoop,
ForceHardwareLoopPHI)) {
// TODO: there can be many reasons a loop is not considered a
// candidate, so we should let isHardwareLoopCandidate fill in the
// reason and then report a better message here.
reportHWLoopFailure("loop is not a candidate", "HWLoopNoCandidate", ORE, L);
return false;
}
assert(
(HWLoopInfo.ExitBlock && HWLoopInfo.ExitBranch && HWLoopInfo.TripCount) &&
"Hardware Loop must have set exit info.");
BasicBlock *Preheader = L->getLoopPreheader();
// If we don't have a preheader, then insert one.
if (!Preheader)
Preheader = InsertPreheaderForLoop(L, DT, LI, nullptr, PreserveLCSSA);
if (!Preheader)
return false;
HardwareLoop HWLoop(HWLoopInfo, *SE, *DL, ORE);
HWLoop.Create();
++NumHWLoops;
return true;
}
void HardwareLoop::Create() {
LLVM_DEBUG(dbgs() << "HWLoops: Converting loop..\n");
Value *LoopCountInit = InitLoopCount();
if (!LoopCountInit) {
reportHWLoopFailure("could not safely create a loop count expression",
"HWLoopNotSafe", ORE, L);
return;
}
[ARM] Alter t2DoLoopStart to define lr This changes the definition of t2DoLoopStart from t2DoLoopStart rGPR to GPRlr = t2DoLoopStart rGPR This will hopefully mean that low overhead loops are more tied together, and we can more reliably generate loops without reverting or being at the whims of the register allocator. This is a fairly simple change in itself, but leads to a number of other required alterations. - The hardware loop pass, if UsePhi is set, now generates loops of the form: %start = llvm.start.loop.iterations(%N) loop: %p = phi [%start], [%dec] %dec = llvm.loop.decrement.reg(%p, 1) %c = icmp ne %dec, 0 br %c, loop, exit - For this a new llvm.start.loop.iterations intrinsic was added, identical to llvm.set.loop.iterations but produces a value as seen above, gluing the loop together more through def-use chains. - This new instrinsic conceptually produces the same output as input, which is taught to SCEV so that the checks in MVETailPredication are not affected. - Some minor changes are needed to the ARMLowOverheadLoop pass, but it has been left mostly as before. We should now more reliably be able to tell that the t2DoLoopStart is correct without having to prove it, but t2WhileLoopStart and tail-predicated loops will remain the same. - And all the tests have been updated. There are a lot of them! This patch on it's own might cause more trouble that it helps, with more tail-predicated loops being reverted, but some additional patches can hopefully improve upon that to get to something that is better overall. Differential Revision: https://reviews.llvm.org/D89881
2020-11-10 16:57:58 +01:00
Value *Setup = InsertIterationSetup(LoopCountInit);
if (UsePHICounter || ForceHardwareLoopPHI) {
Instruction *LoopDec = InsertLoopRegDec(LoopCountInit);
[ARM] Alter t2DoLoopStart to define lr This changes the definition of t2DoLoopStart from t2DoLoopStart rGPR to GPRlr = t2DoLoopStart rGPR This will hopefully mean that low overhead loops are more tied together, and we can more reliably generate loops without reverting or being at the whims of the register allocator. This is a fairly simple change in itself, but leads to a number of other required alterations. - The hardware loop pass, if UsePhi is set, now generates loops of the form: %start = llvm.start.loop.iterations(%N) loop: %p = phi [%start], [%dec] %dec = llvm.loop.decrement.reg(%p, 1) %c = icmp ne %dec, 0 br %c, loop, exit - For this a new llvm.start.loop.iterations intrinsic was added, identical to llvm.set.loop.iterations but produces a value as seen above, gluing the loop together more through def-use chains. - This new instrinsic conceptually produces the same output as input, which is taught to SCEV so that the checks in MVETailPredication are not affected. - Some minor changes are needed to the ARMLowOverheadLoop pass, but it has been left mostly as before. We should now more reliably be able to tell that the t2DoLoopStart is correct without having to prove it, but t2WhileLoopStart and tail-predicated loops will remain the same. - And all the tests have been updated. There are a lot of them! This patch on it's own might cause more trouble that it helps, with more tail-predicated loops being reverted, but some additional patches can hopefully improve upon that to get to something that is better overall. Differential Revision: https://reviews.llvm.org/D89881
2020-11-10 16:57:58 +01:00
Value *EltsRem = InsertPHICounter(Setup, LoopDec);
LoopDec->setOperand(0, EltsRem);
UpdateBranch(LoopDec);
} else
InsertLoopDec();
// Run through the basic blocks of the loop and see if any of them have dead
// PHIs that can be removed.
for (auto I : L->blocks())
DeleteDeadPHIs(I);
}
static bool CanGenerateTest(Loop *L, Value *Count) {
BasicBlock *Preheader = L->getLoopPreheader();
if (!Preheader->getSinglePredecessor())
return false;
BasicBlock *Pred = Preheader->getSinglePredecessor();
if (!isa<BranchInst>(Pred->getTerminator()))
return false;
auto *BI = cast<BranchInst>(Pred->getTerminator());
if (BI->isUnconditional() || !isa<ICmpInst>(BI->getCondition()))
return false;
// Check that the icmp is checking for equality of Count and zero and that
// a non-zero value results in entering the loop.
auto ICmp = cast<ICmpInst>(BI->getCondition());
LLVM_DEBUG(dbgs() << " - Found condition: " << *ICmp << "\n");
if (!ICmp->isEquality())
return false;
auto IsCompareZero = [](ICmpInst *ICmp, Value *Count, unsigned OpIdx) {
if (auto *Const = dyn_cast<ConstantInt>(ICmp->getOperand(OpIdx)))
return Const->isZero() && ICmp->getOperand(OpIdx ^ 1) == Count;
return false;
};
if (!IsCompareZero(ICmp, Count, 0) && !IsCompareZero(ICmp, Count, 1))
return false;
unsigned SuccIdx = ICmp->getPredicate() == ICmpInst::ICMP_NE ? 0 : 1;
if (BI->getSuccessor(SuccIdx) != Preheader)
return false;
return true;
}
Value *HardwareLoop::InitLoopCount() {
LLVM_DEBUG(dbgs() << "HWLoops: Initialising loop counter value:\n");
// Can we replace a conditional branch with an intrinsic that sets the
// loop counter and tests that is not zero?
SCEVExpander SCEVE(SE, DL, "loopcnt");
// If we're trying to use the 'test and set' form of the intrinsic, we need
// to replace a conditional branch that is controlling entry to the loop. It
// is likely (guaranteed?) that the preheader has an unconditional branch to
// the loop header, so also check if it has a single predecessor.
if (SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, TripCount,
SE.getZero(TripCount->getType()))) {
LLVM_DEBUG(dbgs() << " - Attempting to use test.set counter.\n");
UseLoopGuard |= ForceGuardLoopEntry;
} else
UseLoopGuard = false;
BasicBlock *BB = L->getLoopPreheader();
if (UseLoopGuard && BB->getSinglePredecessor() &&
cast<BranchInst>(BB->getTerminator())->isUnconditional()) {
BasicBlock *Predecessor = BB->getSinglePredecessor();
// If it's not safe to create a while loop then don't force it and create a
// do-while loop instead
if (!isSafeToExpandAt(TripCount, Predecessor->getTerminator(), SE))
UseLoopGuard = false;
else
BB = Predecessor;
}
if (!isSafeToExpandAt(TripCount, BB->getTerminator(), SE)) {
LLVM_DEBUG(dbgs() << "- Bailing, unsafe to expand TripCount "
<< *TripCount << "\n");
return nullptr;
}
Value *Count = SCEVE.expandCodeFor(TripCount, CountType,
BB->getTerminator());
// FIXME: We've expanded Count where we hope to insert the counter setting
// intrinsic. But, in the case of the 'test and set' form, we may fallback to
// the just 'set' form and in which case the insertion block is most likely
// different. It means there will be instruction(s) in a block that possibly
// aren't needed. The isLoopEntryGuardedByCond is trying to avoid this issue,
// but it's doesn't appear to work in all cases.
UseLoopGuard = UseLoopGuard && CanGenerateTest(L, Count);
BeginBB = UseLoopGuard ? BB : L->getLoopPreheader();
LLVM_DEBUG(dbgs() << " - Loop Count: " << *Count << "\n"
<< " - Expanded Count in " << BB->getName() << "\n"
<< " - Will insert set counter intrinsic into: "
<< BeginBB->getName() << "\n");
return Count;
}
[ARM] Alter t2DoLoopStart to define lr This changes the definition of t2DoLoopStart from t2DoLoopStart rGPR to GPRlr = t2DoLoopStart rGPR This will hopefully mean that low overhead loops are more tied together, and we can more reliably generate loops without reverting or being at the whims of the register allocator. This is a fairly simple change in itself, but leads to a number of other required alterations. - The hardware loop pass, if UsePhi is set, now generates loops of the form: %start = llvm.start.loop.iterations(%N) loop: %p = phi [%start], [%dec] %dec = llvm.loop.decrement.reg(%p, 1) %c = icmp ne %dec, 0 br %c, loop, exit - For this a new llvm.start.loop.iterations intrinsic was added, identical to llvm.set.loop.iterations but produces a value as seen above, gluing the loop together more through def-use chains. - This new instrinsic conceptually produces the same output as input, which is taught to SCEV so that the checks in MVETailPredication are not affected. - Some minor changes are needed to the ARMLowOverheadLoop pass, but it has been left mostly as before. We should now more reliably be able to tell that the t2DoLoopStart is correct without having to prove it, but t2WhileLoopStart and tail-predicated loops will remain the same. - And all the tests have been updated. There are a lot of them! This patch on it's own might cause more trouble that it helps, with more tail-predicated loops being reverted, but some additional patches can hopefully improve upon that to get to something that is better overall. Differential Revision: https://reviews.llvm.org/D89881
2020-11-10 16:57:58 +01:00
Value* HardwareLoop::InsertIterationSetup(Value *LoopCountInit) {
IRBuilder<> Builder(BeginBB->getTerminator());
Type *Ty = LoopCountInit->getType();
[ARM] Alter t2DoLoopStart to define lr This changes the definition of t2DoLoopStart from t2DoLoopStart rGPR to GPRlr = t2DoLoopStart rGPR This will hopefully mean that low overhead loops are more tied together, and we can more reliably generate loops without reverting or being at the whims of the register allocator. This is a fairly simple change in itself, but leads to a number of other required alterations. - The hardware loop pass, if UsePhi is set, now generates loops of the form: %start = llvm.start.loop.iterations(%N) loop: %p = phi [%start], [%dec] %dec = llvm.loop.decrement.reg(%p, 1) %c = icmp ne %dec, 0 br %c, loop, exit - For this a new llvm.start.loop.iterations intrinsic was added, identical to llvm.set.loop.iterations but produces a value as seen above, gluing the loop together more through def-use chains. - This new instrinsic conceptually produces the same output as input, which is taught to SCEV so that the checks in MVETailPredication are not affected. - Some minor changes are needed to the ARMLowOverheadLoop pass, but it has been left mostly as before. We should now more reliably be able to tell that the t2DoLoopStart is correct without having to prove it, but t2WhileLoopStart and tail-predicated loops will remain the same. - And all the tests have been updated. There are a lot of them! This patch on it's own might cause more trouble that it helps, with more tail-predicated loops being reverted, but some additional patches can hopefully improve upon that to get to something that is better overall. Differential Revision: https://reviews.llvm.org/D89881
2020-11-10 16:57:58 +01:00
bool UsePhi = UsePHICounter || ForceHardwareLoopPHI;
Intrinsic::ID ID = UseLoopGuard ? Intrinsic::test_set_loop_iterations
: (UsePhi ? Intrinsic::start_loop_iterations
: Intrinsic::set_loop_iterations);
Function *LoopIter = Intrinsic::getDeclaration(M, ID, Ty);
Value *SetCount = Builder.CreateCall(LoopIter, LoopCountInit);
// Use the return value of the intrinsic to control the entry of the loop.
if (UseLoopGuard) {
assert((isa<BranchInst>(BeginBB->getTerminator()) &&
cast<BranchInst>(BeginBB->getTerminator())->isConditional()) &&
"Expected conditional branch");
auto *LoopGuard = cast<BranchInst>(BeginBB->getTerminator());
LoopGuard->setCondition(SetCount);
if (LoopGuard->getSuccessor(0) != L->getLoopPreheader())
LoopGuard->swapSuccessors();
}
LLVM_DEBUG(dbgs() << "HWLoops: Inserted loop counter: "
<< *SetCount << "\n");
[ARM] Alter t2DoLoopStart to define lr This changes the definition of t2DoLoopStart from t2DoLoopStart rGPR to GPRlr = t2DoLoopStart rGPR This will hopefully mean that low overhead loops are more tied together, and we can more reliably generate loops without reverting or being at the whims of the register allocator. This is a fairly simple change in itself, but leads to a number of other required alterations. - The hardware loop pass, if UsePhi is set, now generates loops of the form: %start = llvm.start.loop.iterations(%N) loop: %p = phi [%start], [%dec] %dec = llvm.loop.decrement.reg(%p, 1) %c = icmp ne %dec, 0 br %c, loop, exit - For this a new llvm.start.loop.iterations intrinsic was added, identical to llvm.set.loop.iterations but produces a value as seen above, gluing the loop together more through def-use chains. - This new instrinsic conceptually produces the same output as input, which is taught to SCEV so that the checks in MVETailPredication are not affected. - Some minor changes are needed to the ARMLowOverheadLoop pass, but it has been left mostly as before. We should now more reliably be able to tell that the t2DoLoopStart is correct without having to prove it, but t2WhileLoopStart and tail-predicated loops will remain the same. - And all the tests have been updated. There are a lot of them! This patch on it's own might cause more trouble that it helps, with more tail-predicated loops being reverted, but some additional patches can hopefully improve upon that to get to something that is better overall. Differential Revision: https://reviews.llvm.org/D89881
2020-11-10 16:57:58 +01:00
return UseLoopGuard ? LoopCountInit : SetCount;
}
void HardwareLoop::InsertLoopDec() {
IRBuilder<> CondBuilder(ExitBranch);
Function *DecFunc =
Intrinsic::getDeclaration(M, Intrinsic::loop_decrement,
LoopDecrement->getType());
Value *Ops[] = { LoopDecrement };
Value *NewCond = CondBuilder.CreateCall(DecFunc, Ops);
Value *OldCond = ExitBranch->getCondition();
ExitBranch->setCondition(NewCond);
// The false branch must exit the loop.
if (!L->contains(ExitBranch->getSuccessor(0)))
ExitBranch->swapSuccessors();
// The old condition may be dead now, and may have even created a dead PHI
// (the original induction variable).
RecursivelyDeleteTriviallyDeadInstructions(OldCond);
LLVM_DEBUG(dbgs() << "HWLoops: Inserted loop dec: " << *NewCond << "\n");
}
Instruction* HardwareLoop::InsertLoopRegDec(Value *EltsRem) {
IRBuilder<> CondBuilder(ExitBranch);
Function *DecFunc =
Intrinsic::getDeclaration(M, Intrinsic::loop_decrement_reg,
{ EltsRem->getType() });
Value *Ops[] = { EltsRem, LoopDecrement };
Value *Call = CondBuilder.CreateCall(DecFunc, Ops);
LLVM_DEBUG(dbgs() << "HWLoops: Inserted loop dec: " << *Call << "\n");
return cast<Instruction>(Call);
}
PHINode* HardwareLoop::InsertPHICounter(Value *NumElts, Value *EltsRem) {
BasicBlock *Preheader = L->getLoopPreheader();
BasicBlock *Header = L->getHeader();
BasicBlock *Latch = ExitBranch->getParent();
IRBuilder<> Builder(Header->getFirstNonPHI());
PHINode *Index = Builder.CreatePHI(NumElts->getType(), 2);
Index->addIncoming(NumElts, Preheader);
Index->addIncoming(EltsRem, Latch);
LLVM_DEBUG(dbgs() << "HWLoops: PHI Counter: " << *Index << "\n");
return Index;
}
void HardwareLoop::UpdateBranch(Value *EltsRem) {
IRBuilder<> CondBuilder(ExitBranch);
Value *NewCond =
CondBuilder.CreateICmpNE(EltsRem, ConstantInt::get(EltsRem->getType(), 0));
Value *OldCond = ExitBranch->getCondition();
ExitBranch->setCondition(NewCond);
// The false branch must exit the loop.
if (!L->contains(ExitBranch->getSuccessor(0)))
ExitBranch->swapSuccessors();
// The old condition may be dead now, and may have even created a dead PHI
// (the original induction variable).
RecursivelyDeleteTriviallyDeadInstructions(OldCond);
}
INITIALIZE_PASS_BEGIN(HardwareLoops, DEBUG_TYPE, HW_LOOPS_NAME, false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
INITIALIZE_PASS_END(HardwareLoops, DEBUG_TYPE, HW_LOOPS_NAME, false, false)
FunctionPass *llvm::createHardwareLoopsPass() { return new HardwareLoops(); }