shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity

[AArch64] Remove AArch64AddressTypePromotion pass

Browse Source 
Summary:
Remove the AArch64AddressTypePromotion pass as we migrated all transformations
done in this pass into CGP in r299379.

Reviewers: qcolombet, jmolloy, javed.absar, mcrosier

Reviewed By: qcolombet

Subscribers: aemerson, rengolin, mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D31623

llvm-svn: 302245
This commit is contained in:
Jun Bum Lim 2017-05-05 16:05:41 +00:00
parent c1ddedc657
commit 94d42533eb
4 changed files with 0 additions and 505 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
llvm/lib/Target/AArch64/AArch64.h
View File

@ -41,7 +41,6 @@ FunctionPass *createAArch64LoadStoreOptimizationPass();
FunctionPass *createAArch64VectorByElementOptPass();
ModulePass *createAArch64PromoteConstantPass();
FunctionPass *createAArch64ConditionOptimizerPass();
FunctionPass *createAArch64AddressTypePromotionPass();
FunctionPass *createAArch64A57FPLoadBalancing();
FunctionPass *createAArch64A53Fix835769();
@ -54,7 +53,6 @@ createAArch64InstructionSelector(const AArch64TargetMachine &,
void initializeAArch64A53Fix835769Pass(PassRegistry&);
void initializeAArch64A57FPLoadBalancingPass(PassRegistry&);
void initializeAArch64AddressTypePromotionPass(PassRegistry&);
void initializeAArch64AdvSIMDScalarPass(PassRegistry&);
void initializeAArch64CollectLOHPass(PassRegistry&);
void initializeAArch64ConditionalComparesPass(PassRegistry&);

493
llvm/lib/Target/AArch64/AArch64AddressTypePromotion.cpp
View File

@ -1,493 +0,0 @@
//===-- AArch64AddressTypePromotion.cpp --- Promote type for addr accesses -==//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass tries to promote the computations use to obtained a sign extended
// value used into memory accesses.
// E.g.
// a = add nsw i32 b, 3
// d = sext i32 a to i64
// e = getelementptr ..., i64 d
//
// =>
// f = sext i32 b to i64
// a = add nsw i64 f, 3
// e = getelementptr ..., i64 a
//
// This is legal to do if the computations are marked with either nsw or nuw
// markers. Moreover, the current heuristic is simple: it does not create new
// sext operations, i.e., it gives up when a sext would have forked (e.g., if a
// = add i32 b, c, two sexts are required to promote the computation).
//
// FIXME: This pass may be useful for other targets too.
// ===---------------------------------------------------------------------===//
#include "AArch64.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Use.h"
#include "llvm/IR/User.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
using namespace llvm;
#define DEBUG_TYPE "aarch64-type-promotion"
static cl::opt<bool>
EnableMerge("aarch64-type-promotion-merge", cl::Hidden,
cl::desc("Enable merging of redundant sexts when one is dominating"
" the other."),
cl::init(true));
#define AARCH64_TYPE_PROMO_NAME "AArch64 Address Type Promotion"
//===----------------------------------------------------------------------===//
// AArch64AddressTypePromotion
//===----------------------------------------------------------------------===//
namespace {
class AArch64AddressTypePromotion : public FunctionPass {
public:
static char ID;
AArch64AddressTypePromotion() : FunctionPass(ID) {
initializeAArch64AddressTypePromotionPass(*PassRegistry::getPassRegistry());
}
StringRef getPassName() const override { return AARCH64_TYPE_PROMO_NAME; }
/// Iterate over the functions and promote the computation of interesting
// sext instructions.
bool runOnFunction(Function &F) override;
private:
/// The current function.
Function *Func = nullptr;
/// Filter out all sexts that does not have this type.
/// Currently initialized with Int64Ty.
Type *ConsideredSExtType = nullptr;
// This transformation requires dominator info.
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
AU.addRequired<DominatorTreeWrapperPass>();
AU.addPreserved<DominatorTreeWrapperPass>();
FunctionPass::getAnalysisUsage(AU);
}
typedef SmallPtrSet<Instruction *, 32> SetOfInstructions;
typedef SmallVector<Instruction *, 16> Instructions;
typedef DenseMap<Value *, Instructions> ValueToInsts;
/// Check if it is profitable to move a sext through this instruction.
/// Currently, we consider it is profitable if:
/// - Inst is used only once (no need to insert truncate).
/// - Inst has only one operand that will require a sext operation (we do
/// do not create new sext operation).
bool shouldGetThrough(const Instruction *Inst);
/// Check if it is possible and legal to move a sext through this
/// instruction.
/// Current heuristic considers that we can get through:
/// - Arithmetic operation marked with the nsw or nuw flag.
/// - Other sext operation.
/// - Truncate operation if it was just dropping sign extended bits.
bool canGetThrough(const Instruction *Inst);
/// Move sext operations through safe to sext instructions.
bool propagateSignExtension(Instructions &SExtInsts);
/// Is this sext should be considered for code motion.
/// We look for sext with ConsideredSExtType and uses in at least one
// GetElementPtrInst.
bool shouldConsiderSExt(const Instruction *SExt) const;
/// Collect all interesting sext operations, i.e., the ones with the right
/// type and used in memory accesses.
/// More precisely, a sext instruction is considered as interesting if it
/// is used in a "complex" getelementptr or it exits at least another
/// sext instruction that sign extended the same initial value.
/// A getelementptr is considered as "complex" if it has more than 2
// operands.
void analyzeSExtension(Instructions &SExtInsts);
/// Merge redundant sign extension operations in common dominator.
void mergeSExts(ValueToInsts &ValToSExtendedUses,
SetOfInstructions &ToRemove);
};
} // end anonymous namespace
char AArch64AddressTypePromotion::ID = 0;
INITIALIZE_PASS_BEGIN(AArch64AddressTypePromotion, "aarch64-type-promotion",
AARCH64_TYPE_PROMO_NAME, false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(AArch64AddressTypePromotion, "aarch64-type-promotion",
AARCH64_TYPE_PROMO_NAME, false, false)
FunctionPass *llvm::createAArch64AddressTypePromotionPass() {
return new AArch64AddressTypePromotion();
}
bool AArch64AddressTypePromotion::canGetThrough(const Instruction *Inst) {
if (isa<SExtInst>(Inst))
return true;
const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
(BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
return true;
// sext(trunc(sext)) --> sext
if (isa<TruncInst>(Inst) && isa<SExtInst>(Inst->getOperand(0))) {
const Instruction *Opnd = cast<Instruction>(Inst->getOperand(0));
// Check that the truncate just drop sign extended bits.
if (Inst->getType()->getIntegerBitWidth() >=
Opnd->getOperand(0)->getType()->getIntegerBitWidth() &&
Inst->getOperand(0)->getType()->getIntegerBitWidth() <=
ConsideredSExtType->getIntegerBitWidth())
return true;
}
return false;
}
bool AArch64AddressTypePromotion::shouldGetThrough(const Instruction *Inst) {
// If the type of the sext is the same as the considered one, this sext
// will become useless.
// Otherwise, we will have to do something to preserve the original value,
// unless it is used once.
if (isa<SExtInst>(Inst) &&
(Inst->getType() == ConsideredSExtType || Inst->hasOneUse()))
return true;
// If the Inst is used more that once, we may need to insert truncate
// operations and we don't do that at the moment.
if (!Inst->hasOneUse())
return false;
// This truncate is used only once, thus if we can get thourgh, it will become
// useless.
if (isa<TruncInst>(Inst))
return true;
// If both operands are not constant, a new sext will be created here.
// Current heuristic is: each step should be profitable.
// Therefore we don't allow to increase the number of sext even if it may
// be profitable later on.
if (isa<BinaryOperator>(Inst) && isa<ConstantInt>(Inst->getOperand(1)))
return true;
return false;
}
static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
return !(isa<SelectInst>(Inst) && OpIdx == 0);
}
bool
AArch64AddressTypePromotion::shouldConsiderSExt(const Instruction *SExt) const {
if (SExt->getType() != ConsideredSExtType)
return false;
for (const User *U : SExt->users()) {
if (isa<GetElementPtrInst>(U))
return true;
}
return false;
}
// Input:
// - SExtInsts contains all the sext instructions that are used directly in
// GetElementPtrInst, i.e., access to memory.
// Algorithm:
// - For each sext operation in SExtInsts:
// Let var be the operand of sext.
// while it is profitable (see shouldGetThrough), legal, and safe
// (see canGetThrough) to move sext through var's definition:
// * promote the type of var's definition.
// * fold var into sext uses.
// * move sext above var's definition.
// * update sext operand to use the operand of var that should be sign
// extended (by construction there is only one).
//
// E.g.,
// a = ... i32 c, 3
// b = sext i32 a to i64 <- is it legal/safe/profitable to get through 'a'
// ...
// = b
// => Yes, update the code
// b = sext i32 c to i64
// a = ... i64 b, 3
// ...
// = a
// Iterate on 'c'.
bool
AArch64AddressTypePromotion::propagateSignExtension(Instructions &SExtInsts) {
DEBUG(dbgs() << "*** Propagate Sign Extension ***\n");
bool LocalChange = false;
SetOfInstructions ToRemove;
ValueToInsts ValToSExtendedUses;
while (!SExtInsts.empty()) {
// Get through simple chain.
Instruction *SExt = SExtInsts.pop_back_val();
DEBUG(dbgs() << "Consider:\n" << *SExt << '\n');
// If this SExt has already been merged continue.
if (SExt->use_empty() && ToRemove.count(SExt)) {
DEBUG(dbgs() << "No uses => marked as delete\n");
continue;
}
// Now try to get through the chain of definitions.
while (auto *Inst = dyn_cast<Instruction>(SExt->getOperand(0))) {
DEBUG(dbgs() << "Try to get through:\n" << *Inst << '\n');
if (!canGetThrough(Inst) || !shouldGetThrough(Inst)) {
// We cannot get through something that is not an Instruction
// or not safe to SExt.
DEBUG(dbgs() << "Cannot get through\n");
break;
}
LocalChange = true;
// If this is a sign extend, it becomes useless.
if (isa<SExtInst>(Inst) || isa<TruncInst>(Inst)) {
DEBUG(dbgs() << "SExt or trunc, mark it as to remove\n");
// We cannot use replaceAllUsesWith here because we may trigger some
// assertion on the type as all involved sext operation may have not
// been moved yet.
while (!Inst->use_empty()) {
Use &U = *Inst->use_begin();
Instruction *User = dyn_cast<Instruction>(U.getUser());
assert(User && "User of sext is not an Instruction!");
User->setOperand(U.getOperandNo(), SExt);
}
ToRemove.insert(Inst);
SExt->setOperand(0, Inst->getOperand(0));
SExt->moveBefore(Inst);
continue;
}
// Get through the Instruction:
// 1. Update its type.
// 2. Replace the uses of SExt by Inst.
// 3. Sign extend each operand that needs to be sign extended.
// Step #1.
Inst->mutateType(SExt->getType());
// Step #2.
SExt->replaceAllUsesWith(Inst);
// Step #3.
Instruction *SExtForOpnd = SExt;
DEBUG(dbgs() << "Propagate SExt to operands\n");
for (int OpIdx = 0, EndOpIdx = Inst->getNumOperands(); OpIdx != EndOpIdx;
++OpIdx) {
DEBUG(dbgs() << "Operand:\n" << *(Inst->getOperand(OpIdx)) << '\n');
if (Inst->getOperand(OpIdx)->getType() == SExt->getType() ||
!shouldSExtOperand(Inst, OpIdx)) {
DEBUG(dbgs() << "No need to propagate\n");
continue;
}
// Check if we can statically sign extend the operand.
Value *Opnd = Inst->getOperand(OpIdx);
if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
DEBUG(dbgs() << "Statically sign extend\n");
Inst->setOperand(OpIdx, ConstantInt::getSigned(SExt->getType(),
Cst->getSExtValue()));
continue;
}
// UndefValue are typed, so we have to statically sign extend them.
if (isa<UndefValue>(Opnd)) {
DEBUG(dbgs() << "Statically sign extend\n");
Inst->setOperand(OpIdx, UndefValue::get(SExt->getType()));
continue;
}
// Otherwise we have to explicity sign extend it.
assert(SExtForOpnd &&
"Only one operand should have been sign extended");
SExtForOpnd->setOperand(0, Opnd);
DEBUG(dbgs() << "Move before:\n" << *Inst << "\nSign extend\n");
// Move the sign extension before the insertion point.
SExtForOpnd->moveBefore(Inst);
Inst->setOperand(OpIdx, SExtForOpnd);
// If more sext are required, new instructions will have to be created.
SExtForOpnd = nullptr;
}
if (SExtForOpnd == SExt) {
DEBUG(dbgs() << "Sign extension is useless now\n");
ToRemove.insert(SExt);
break;
}
}
// If the use is already of the right type, connect its uses to its argument
// and delete it.
// This can happen for an Instruction all uses of which are sign extended.
if (!ToRemove.count(SExt) &&
SExt->getType() == SExt->getOperand(0)->getType()) {
DEBUG(dbgs() << "Sign extension is useless, attach its use to "
"its argument\n");
SExt->replaceAllUsesWith(SExt->getOperand(0));
ToRemove.insert(SExt);
} else
ValToSExtendedUses[SExt->getOperand(0)].push_back(SExt);
}
if (EnableMerge)
mergeSExts(ValToSExtendedUses, ToRemove);
// Remove all instructions marked as ToRemove.
for (Instruction *I: ToRemove)
I->eraseFromParent();
return LocalChange;
}
void AArch64AddressTypePromotion::mergeSExts(ValueToInsts &ValToSExtendedUses,
SetOfInstructions &ToRemove) {
DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
for (auto &Entry : ValToSExtendedUses) {
Instructions &Insts = Entry.second;
Instructions CurPts;
for (Instruction *Inst : Insts) {
if (ToRemove.count(Inst))
continue;
bool inserted = false;
for (auto &Pt : CurPts) {
if (DT.dominates(Inst, Pt)) {
DEBUG(dbgs() << "Replace all uses of:\n" << *Pt << "\nwith:\n"
<< *Inst << '\n');
Pt->replaceAllUsesWith(Inst);
ToRemove.insert(Pt);
Pt = Inst;
inserted = true;
break;
}
if (!DT.dominates(Pt, Inst))
// Give up if we need to merge in a common dominator as the
// expermients show it is not profitable.
continue;
DEBUG(dbgs() << "Replace all uses of:\n" << *Inst << "\nwith:\n"
<< *Pt << '\n');
Inst->replaceAllUsesWith(Pt);
ToRemove.insert(Inst);
inserted = true;
break;
}
if (!inserted)
CurPts.push_back(Inst);
}
}
}
void AArch64AddressTypePromotion::analyzeSExtension(Instructions &SExtInsts) {
DEBUG(dbgs() << "*** Analyze Sign Extensions ***\n");
DenseMap<Value *, Instruction *> SeenChains;
for (auto &BB : *Func) {
for (auto &II : BB) {
Instruction *SExt = &II;
// Collect all sext operation per type.
if (!isa<SExtInst>(SExt) || !shouldConsiderSExt(SExt))
continue;
DEBUG(dbgs() << "Found:\n" << (*SExt) << '\n');
// Cases where we actually perform the optimization:
// 1. SExt is used in a getelementptr with more than 2 operand =>
// likely we can merge some computation if they are done on 64 bits.
// 2. The beginning of the SExt chain is SExt several time. =>
// code sharing is possible.
bool insert = false;
// #1.
for (const User *U : SExt->users()) {
const Instruction *Inst = dyn_cast<GetElementPtrInst>(U);
if (Inst && Inst->getNumOperands() > 2) {
DEBUG(dbgs() << "Interesting use in GetElementPtrInst\n" << *Inst
<< '\n');
insert = true;
break;
}
}
// #2.
// Check the head of the chain.
Instruction *Inst = SExt;
Value *Last;
do {
int OpdIdx = 0;
const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
if (BinOp && isa<ConstantInt>(BinOp->getOperand(0)))
OpdIdx = 1;
Last = Inst->getOperand(OpdIdx);
Inst = dyn_cast<Instruction>(Last);
} while (Inst && canGetThrough(Inst) && shouldGetThrough(Inst));
DEBUG(dbgs() << "Head of the chain:\n" << *Last << '\n');
DenseMap<Value *, Instruction *>::iterator AlreadySeen =
SeenChains.find(Last);
if (insert || AlreadySeen != SeenChains.end()) {
DEBUG(dbgs() << "Insert\n");
SExtInsts.push_back(SExt);
if (AlreadySeen != SeenChains.end() && AlreadySeen->second != nullptr) {
DEBUG(dbgs() << "Insert chain member\n");
SExtInsts.push_back(AlreadySeen->second);
SeenChains[Last] = nullptr;
}
} else {
DEBUG(dbgs() << "Record its chain membership\n");
SeenChains[Last] = SExt;
}
}
}
}
bool AArch64AddressTypePromotion::runOnFunction(Function &F) {
if (skipFunction(F))
return false;
if (F.isDeclaration())
return false;
Func = &F;
ConsideredSExtType = Type::getInt64Ty(Func->getContext());
DEBUG(dbgs() << "*** " << getPassName() << ": " << Func->getName() << '\n');
Instructions SExtInsts;
analyzeSExtension(SExtInsts);
return propagateSignExtension(SExtInsts);
}

9
llvm/lib/Target/AArch64/AArch64TargetMachine.cpp
View File

@ -108,11 +108,6 @@ EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden,
cl::desc("Work around Cortex-A53 erratum 835769"),
cl::init(false));
static cl::opt<bool>
EnableAddressTypePromotion("aarch64-enable-type-promotion", cl::Hidden,
cl::desc("Enable the type promotion pass"),
cl::init(false));
static cl::opt<bool>
EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden,
cl::desc("Enable optimizations on complex GEPs"),
@ -146,7 +141,6 @@ extern "C" void LLVMInitializeAArch64Target() {
initializeGlobalISel(*PR);
initializeAArch64A53Fix835769Pass(*PR);
initializeAArch64A57FPLoadBalancingPass(*PR);
initializeAArch64AddressTypePromotionPass(*PR);
initializeAArch64AdvSIMDScalarPass(*PR);
initializeAArch64CollectLOHPass(*PR);
initializeAArch64ConditionalComparesPass(*PR);
@ -382,9 +376,6 @@ bool AArch64PassConfig::addPreISel() {
addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize));
}
if (TM->getOptLevel() != CodeGenOpt::None && EnableAddressTypePromotion)
addPass(createAArch64AddressTypePromotionPass());
return false;
}

1
llvm/lib/Target/AArch64/CMakeLists.txt
View File

@ -39,7 +39,6 @@ endif()
add_llvm_target(AArch64CodeGen
AArch64A57FPLoadBalancing.cpp
AArch64AddressTypePromotion.cpp
AArch64AdvSIMDScalarPass.cpp
AArch64AsmPrinter.cpp
AArch64CleanupLocalDynamicTLSPass.cpp