llvm-project/llvm/lib/Transforms/Utils/PredicateInfo.cpp

641 lines
25 KiB
C++
Raw Normal View History

Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
//===-- PredicateInfo.cpp - PredicateInfo Builder--------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------===//
//
// This file implements the PredicateInfo class.
//
//===----------------------------------------------------------------===//
#include "llvm/Transforms/Utils/PredicateInfo.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/CFG.h"
#include "llvm/Analysis/OrderedBasicBlock.h"
#include "llvm/IR/AssemblyAnnotationWriter.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/FormattedStream.h"
#include "llvm/Transforms/Scalar.h"
#include <algorithm>
#define DEBUG_TYPE "predicateinfo"
using namespace llvm;
using namespace PatternMatch;
using namespace llvm::PredicateInfoClasses;
INITIALIZE_PASS_BEGIN(PredicateInfoPrinterLegacyPass, "print-predicateinfo",
"PredicateInfo Printer", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
INITIALIZE_PASS_END(PredicateInfoPrinterLegacyPass, "print-predicateinfo",
"PredicateInfo Printer", false, false)
static cl::opt<bool> VerifyPredicateInfo(
"verify-predicateinfo", cl::init(false), cl::Hidden,
cl::desc("Verify PredicateInfo in legacy printer pass."));
namespace llvm {
namespace PredicateInfoClasses {
enum LocalNum {
// Operations that must appear first in the block.
LN_First,
// Operations that are somewhere in the middle of the block, and are sorted on
// demand.
LN_Middle,
// Operations that must appear last in a block, like successor phi node uses.
LN_Last
};
// Associate global and local DFS info with defs and uses, so we can sort them
// into a global domination ordering.
struct ValueDFS {
int DFSIn = 0;
int DFSOut = 0;
unsigned int LocalNum = LN_Middle;
PredicateBase *PInfo = nullptr;
// Only one of Def or Use will be set.
Value *Def = nullptr;
Use *U = nullptr;
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
};
// This compares ValueDFS structures, creating OrderedBasicBlocks where
// necessary to compare uses/defs in the same block. Doing so allows us to walk
// the minimum number of instructions necessary to compute our def/use ordering.
struct ValueDFS_Compare {
DenseMap<const BasicBlock *, std::unique_ptr<OrderedBasicBlock>> &OBBMap;
ValueDFS_Compare(
DenseMap<const BasicBlock *, std::unique_ptr<OrderedBasicBlock>> &OBBMap)
: OBBMap(OBBMap) {}
bool operator()(const ValueDFS &A, const ValueDFS &B) const {
if (&A == &B)
return false;
// The only case we can't directly compare them is when they in the same
// block, and both have localnum == middle. In that case, we have to use
// comesbefore to see what the real ordering is, because they are in the
// same basic block.
bool SameBlock = std::tie(A.DFSIn, A.DFSOut) == std::tie(B.DFSIn, B.DFSOut);
if (!SameBlock || A.LocalNum != LN_Middle || B.LocalNum != LN_Middle)
return std::tie(A.DFSIn, A.DFSOut, A.LocalNum, A.Def, A.U) <
std::tie(B.DFSIn, B.DFSOut, B.LocalNum, B.Def, B.U);
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
return localComesBefore(A, B);
}
// Get the definition of an instruction that occurs in the middle of a block.
Value *getMiddleDef(const ValueDFS &VD) const {
if (VD.Def)
return VD.Def;
// It's possible for the defs and uses to be null. For branches, the local
// numbering will say the placed predicaeinfos should go first (IE
// LN_beginning), so we won't be in this function. For assumes, we will end
// up here, beause we need to order the def we will place relative to the
// assume. So for the purpose of ordering, we pretend the def is the assume
// because that is where we will insert the info.
if (!VD.U) {
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
assert(VD.PInfo &&
"No def, no use, and no predicateinfo should not occur");
assert(isa<PredicateAssume>(VD.PInfo) &&
"Middle of block should only occur for assumes");
return cast<PredicateAssume>(VD.PInfo)->AssumeInst;
}
return nullptr;
}
// Return either the Def, if it's not null, or the user of the Use, if the def
// is null.
const Instruction *getDefOrUser(const Value *Def, const Use *U) const {
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
if (Def)
return cast<Instruction>(Def);
return cast<Instruction>(U->getUser());
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
}
// This performs the necessary local basic block ordering checks to tell
// whether A comes before B, where both are in the same basic block.
bool localComesBefore(const ValueDFS &A, const ValueDFS &B) const {
auto *ADef = getMiddleDef(A);
auto *BDef = getMiddleDef(B);
// See if we have real values or uses. If we have real values, we are
// guaranteed they are instructions or arguments. No matter what, we are
// guaranteed they are in the same block if they are instructions.
auto *ArgA = dyn_cast_or_null<Argument>(ADef);
auto *ArgB = dyn_cast_or_null<Argument>(BDef);
if (ArgA && !ArgB)
return true;
if (ArgB && !ArgA)
return false;
if (ArgA && ArgB)
return ArgA->getArgNo() < ArgB->getArgNo();
auto *AInst = getDefOrUser(ADef, A.U);
auto *BInst = getDefOrUser(BDef, B.U);
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
auto *BB = AInst->getParent();
auto LookupResult = OBBMap.find(BB);
if (LookupResult != OBBMap.end())
return LookupResult->second->dominates(AInst, BInst);
else {
auto Result = OBBMap.insert({BB, make_unique<OrderedBasicBlock>(BB)});
return Result.first->second->dominates(AInst, BInst);
}
return std::tie(ADef, A.U) < std::tie(BDef, B.U);
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
}
};
} // namespace PredicateInfoClasses
bool PredicateInfo::stackIsInScope(const ValueDFSStack &Stack, int DFSIn,
int DFSOut) const {
if (Stack.empty())
return false;
return DFSIn >= Stack.back().DFSIn && DFSOut <= Stack.back().DFSOut;
}
void PredicateInfo::popStackUntilDFSScope(ValueDFSStack &Stack, int DFSIn,
int DFSOut) {
while (!Stack.empty() && !stackIsInScope(Stack, DFSIn, DFSOut))
Stack.pop_back();
}
// Convert the uses of Op into a vector of uses, associating global and local
// DFS info with each one.
void PredicateInfo::convertUsesToDFSOrdered(
Value *Op, SmallVectorImpl<ValueDFS> &DFSOrderedSet) {
for (auto &U : Op->uses()) {
if (auto *I = dyn_cast<Instruction>(U.getUser())) {
ValueDFS VD;
// Put the phi node uses in the incoming block.
BasicBlock *IBlock;
if (auto *PN = dyn_cast<PHINode>(I)) {
IBlock = PN->getIncomingBlock(U);
// Make phi node users appear last in the incoming block
// they are from.
VD.LocalNum = LN_Last;
} else {
// If it's not a phi node use, it is somewhere in the middle of the
// block.
IBlock = I->getParent();
VD.LocalNum = LN_Middle;
}
DomTreeNode *DomNode = DT.getNode(IBlock);
// It's possible our use is in an unreachable block. Skip it if so.
if (!DomNode)
continue;
VD.DFSIn = DomNode->getDFSNumIn();
VD.DFSOut = DomNode->getDFSNumOut();
VD.U = &U;
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
DFSOrderedSet.push_back(VD);
}
}
}
// Collect relevant operations from Comparison that we may want to insert copies
// for.
void collectCmpOps(CmpInst *Comparison, SmallVectorImpl<Value *> &CmpOperands) {
auto *Op0 = Comparison->getOperand(0);
auto *Op1 = Comparison->getOperand(1);
if (Op0 == Op1)
return;
CmpOperands.push_back(Comparison);
// Only want real values, not constants. Additionally, operands with one use
// are only being used in the comparison, which means they will not be useful
// for us to consider for predicateinfo.
//
// FIXME: LLVM crashes trying to create an intrinsic declaration of some
// pointer to function types that return structs, so we avoid them.
if ((isa<Instruction>(Op0) || isa<Argument>(Op0)) && !Op0->hasOneUse() &&
!(Op0->getType()->isPointerTy() &&
Op0->getType()->getPointerElementType()->isFunctionTy()))
CmpOperands.push_back(Op0);
if ((isa<Instruction>(Op1) || isa<Argument>(Op1)) && !Op1->hasOneUse() &&
!(Op1->getType()->isPointerTy() &&
Op1->getType()->getPointerElementType()->isFunctionTy()))
CmpOperands.push_back(Op1);
}
// Process an assume instruction and place relevant operations we want to rename
// into OpsToRename.
void PredicateInfo::processAssume(IntrinsicInst *II, BasicBlock *AssumeBB,
SmallPtrSetImpl<Value *> &OpsToRename) {
SmallVector<Value *, 8> CmpOperands;
// Second, see if we have a comparison we support
SmallVector<Value *, 2> ComparisonsToProcess;
CmpInst::Predicate Pred;
Value *Operand = II->getOperand(0);
if (m_c_And(m_Cmp(Pred, m_Value(), m_Value()),
m_Cmp(Pred, m_Value(), m_Value()))
.match(II->getOperand(0))) {
ComparisonsToProcess.push_back(
cast<BinaryOperator>(Operand)->getOperand(0));
ComparisonsToProcess.push_back(
cast<BinaryOperator>(Operand)->getOperand(1));
} else {
ComparisonsToProcess.push_back(Operand);
}
for (auto Comparison : ComparisonsToProcess) {
if (auto *Cmp = dyn_cast<CmpInst>(Comparison)) {
collectCmpOps(Cmp, CmpOperands);
// Now add our copy infos for our operands
for (auto *Op : CmpOperands) {
OpsToRename.insert(Op);
auto &OperandInfo = getOrCreateValueInfo(Op);
PredicateBase *PB = new PredicateAssume(Op, II, Cmp);
AllInfos.push_back(PB);
OperandInfo.Infos.push_back(PB);
}
CmpOperands.clear();
}
}
}
// Process a block terminating branch, and place relevant operations to be
// renamed into OpsToRename.
void PredicateInfo::processBranch(BranchInst *BI, BasicBlock *BranchBB,
SmallPtrSetImpl<Value *> &OpsToRename) {
SmallVector<Value *, 8> CmpOperands;
BasicBlock *FirstBB = BI->getSuccessor(0);
BasicBlock *SecondBB = BI->getSuccessor(1);
bool FirstSinglePred = FirstBB->getSinglePredecessor();
bool SecondSinglePred = SecondBB->getSinglePredecessor();
SmallVector<BasicBlock *, 2> SuccsToProcess;
bool isAnd = false;
bool isOr = false;
// First make sure we have single preds for these successors, as we can't
// usefully propagate true/false info to them if there are multiple paths to
// them.
if (FirstSinglePred)
SuccsToProcess.push_back(FirstBB);
if (SecondSinglePred)
SuccsToProcess.push_back(SecondBB);
if (SuccsToProcess.empty())
return;
// Second, see if we have a comparison we support
SmallVector<Value *, 2> ComparisonsToProcess;
CmpInst::Predicate Pred;
// Match combinations of conditions.
if (match(BI->getCondition(), m_And(m_Cmp(Pred, m_Value(), m_Value()),
m_Cmp(Pred, m_Value(), m_Value()))) ||
match(BI->getCondition(), m_Or(m_Cmp(Pred, m_Value(), m_Value()),
m_Cmp(Pred, m_Value(), m_Value())))) {
auto *BinOp = cast<BinaryOperator>(BI->getCondition());
if (BinOp->getOpcode() == Instruction::And)
isAnd = true;
else if (BinOp->getOpcode() == Instruction::Or)
isOr = true;
ComparisonsToProcess.push_back(BinOp->getOperand(0));
ComparisonsToProcess.push_back(BinOp->getOperand(1));
} else {
ComparisonsToProcess.push_back(BI->getCondition());
}
for (auto Comparison : ComparisonsToProcess) {
if (auto *Cmp = dyn_cast<CmpInst>(Comparison)) {
collectCmpOps(Cmp, CmpOperands);
// Now add our copy infos for our operands
for (auto *Op : CmpOperands) {
OpsToRename.insert(Op);
auto &OperandInfo = getOrCreateValueInfo(Op);
for (auto *Succ : SuccsToProcess) {
bool TakenEdge = (Succ == FirstBB);
// For and, only insert on the true edge
// For or, only insert on the false edge
if ((isAnd && !TakenEdge) || (isOr && TakenEdge))
continue;
PredicateBase *PB =
new PredicateBranch(Op, BranchBB, Succ, Cmp, TakenEdge);
AllInfos.push_back(PB);
OperandInfo.Infos.push_back(PB);
}
}
CmpOperands.clear();
}
}
}
// Build predicate info for our function
void PredicateInfo::buildPredicateInfo() {
DT.updateDFSNumbers();
// Collect operands to rename from all conditional branch terminators, as well
// as assume statements.
SmallPtrSet<Value *, 8> OpsToRename;
for (auto DTN : depth_first(DT.getRootNode())) {
BasicBlock *BranchBB = DTN->getBlock();
if (auto *BI = dyn_cast<BranchInst>(BranchBB->getTerminator())) {
if (!BI->isConditional())
continue;
processBranch(BI, BranchBB, OpsToRename);
}
}
for (auto &Assume : AC.assumptions()) {
if (auto *II = dyn_cast_or_null<IntrinsicInst>(Assume))
processAssume(II, II->getParent(), OpsToRename);
}
// Now rename all our operations.
renameUses(OpsToRename);
}
Value *PredicateInfo::materializeStack(unsigned int &Counter,
ValueDFSStack &RenameStack,
Value *OrigOp) {
// Find the first thing we have to materialize
auto RevIter = RenameStack.rbegin();
for (; RevIter != RenameStack.rend(); ++RevIter)
if (RevIter->Def)
break;
size_t Start = RevIter - RenameStack.rbegin();
// The maximum number of things we should be trying to materialize at once
// right now is 4, depending on if we had an assume, a branch, and both used
// and of conditions.
for (auto RenameIter = RenameStack.end() - Start;
RenameIter != RenameStack.end(); ++RenameIter) {
auto *Op =
RenameIter == RenameStack.begin() ? OrigOp : (RenameIter - 1)->Def;
ValueDFS &Result = *RenameIter;
auto *ValInfo = Result.PInfo;
// For branches, we can just place the operand in the split block. For
// assume, we have to place it right before the assume to ensure we dominate
// all of our uses.
if (isa<PredicateBranch>(ValInfo)) {
auto *PBranch = cast<PredicateBranch>(ValInfo);
// It's possible we are trying to insert multiple predicateinfos in the
// same block at the beginning of the block. When we do this, we need to
// insert them one after the other, not one before the other. To see if we
// have already inserted predicateinfo into this block, we see if Op !=
// OrigOp && Op->getParent() == PBranch->SplitBB. Op must be an
// instruction we inserted if it's not the original op.
BasicBlock::iterator InsertPt;
if (Op == OrigOp ||
cast<Instruction>(Op)->getParent() != PBranch->SplitBB) {
InsertPt = PBranch->SplitBB->begin();
// Insert after last phi node.
while (isa<PHINode>(InsertPt))
++InsertPt;
} else {
// Insert after op.
InsertPt = ++(cast<Instruction>(Op)->getIterator());
}
IRBuilder<> B(PBranch->SplitBB, InsertPt);
Function *IF = Intrinsic::getDeclaration(
F.getParent(), Intrinsic::ssa_copy, Op->getType());
Value *PIC = B.CreateCall(IF, Op, Op->getName() + "." + Twine(Counter++));
PredicateMap.insert({PIC, ValInfo});
Result.Def = PIC;
} else {
auto *PAssume = dyn_cast<PredicateAssume>(ValInfo);
assert(PAssume &&
"Should not have gotten here without it being an assume");
// Unlike above, this should already insert in the right order when we
// insert multiple predicateinfos in the same block. Because we are
// always inserting right before the assume (instead of the beginning of a
// block), newer insertions will end up after older ones.
IRBuilder<> B(PAssume->AssumeInst->getParent(),
PAssume->AssumeInst->getIterator());
Function *IF = Intrinsic::getDeclaration(
F.getParent(), Intrinsic::ssa_copy, Op->getType());
Value *PIC = B.CreateCall(IF, Op);
PredicateMap.insert({PIC, ValInfo});
Result.Def = PIC;
}
}
return RenameStack.back().Def;
}
// Instead of the standard SSA renaming algorithm, which is O(Number of
// instructions), and walks the entire dominator tree, we walk only the defs +
// uses. The standard SSA renaming algorithm does not really rely on the
// dominator tree except to order the stack push/pops of the renaming stacks, so
// that defs end up getting pushed before hitting the correct uses. This does
// not require the dominator tree, only the *order* of the dominator tree. The
// complete and correct ordering of the defs and uses, in dominator tree is
// contained in the DFS numbering of the dominator tree. So we sort the defs and
// uses into the DFS ordering, and then just use the renaming stack as per
// normal, pushing when we hit a def (which is a predicateinfo instruction),
// popping when we are out of the dfs scope for that def, and replacing any uses
// with top of stack if it exists. In order to handle liveness without
// propagating liveness info, we don't actually insert the predicateinfo
// instruction def until we see a use that it would dominate. Once we see such
// a use, we materialize the predicateinfo instruction in the right place and
// use it.
//
// TODO: Use this algorithm to perform fast single-variable renaming in
// promotememtoreg and memoryssa.
void PredicateInfo::renameUses(SmallPtrSetImpl<Value *> &OpsToRename) {
ValueDFS_Compare Compare(OBBMap);
// Compute liveness, and rename in O(uses) per Op.
for (auto *Op : OpsToRename) {
unsigned Counter = 0;
SmallVector<ValueDFS, 16> OrderedUses;
const auto &ValueInfo = getValueInfo(Op);
// Insert the possible copies into the def/use list.
// They will become real copies if we find a real use for them, and never
// created otherwise.
for (auto &PossibleCopy : ValueInfo.Infos) {
ValueDFS VD;
BasicBlock *CopyBB = nullptr;
// Determine where we are going to place the copy by the copy type.
// The predicate info for branches always come first, they will get
// materialized in the split block at the top of the block.
// The predicate info for assumes will be somewhere in the middle,
// it will get materialized in front of the assume.
if (const auto *PBranch = dyn_cast<PredicateBranch>(PossibleCopy)) {
CopyBB = PBranch->SplitBB;
VD.LocalNum = LN_First;
} else if (const auto *PAssume =
dyn_cast<PredicateAssume>(PossibleCopy)) {
CopyBB = PAssume->AssumeInst->getParent();
VD.LocalNum = LN_Middle;
} else
llvm_unreachable("Unhandled predicate info type");
DomTreeNode *DomNode = DT.getNode(CopyBB);
if (!DomNode)
continue;
VD.DFSIn = DomNode->getDFSNumIn();
VD.DFSOut = DomNode->getDFSNumOut();
VD.PInfo = PossibleCopy;
OrderedUses.push_back(VD);
}
convertUsesToDFSOrdered(Op, OrderedUses);
std::sort(OrderedUses.begin(), OrderedUses.end(), Compare);
SmallVector<ValueDFS, 8> RenameStack;
// For each use, sorted into dfs order, push values and replaces uses with
// top of stack, which will represent the reaching def.
for (auto &VD : OrderedUses) {
// We currently do not materialize copy over copy, but we should decide if
// we want to.
bool PossibleCopy = VD.PInfo != nullptr;
if (RenameStack.empty()) {
DEBUG(dbgs() << "Rename Stack is empty\n");
} else {
DEBUG(dbgs() << "Rename Stack Top DFS numbers are ("
<< RenameStack.back().DFSIn << ","
<< RenameStack.back().DFSOut << ")\n");
}
DEBUG(dbgs() << "Current DFS numbers are (" << VD.DFSIn << ","
<< VD.DFSOut << ")\n");
bool ShouldPush = (VD.Def || PossibleCopy);
bool OutOfScope = !stackIsInScope(RenameStack, VD.DFSIn, VD.DFSOut);
if (OutOfScope || ShouldPush) {
// Sync to our current scope.
popStackUntilDFSScope(RenameStack, VD.DFSIn, VD.DFSOut);
ShouldPush |= (VD.Def || PossibleCopy);
if (ShouldPush) {
RenameStack.push_back(VD);
}
}
// If we get to this point, and the stack is empty we must have a use
// with no renaming needed, just skip it.
if (RenameStack.empty())
continue;
// Skip values, only want to rename the uses
if (VD.Def || PossibleCopy)
continue;
ValueDFS &Result = RenameStack.back();
// If the possible copy dominates something, materialize our stack up to
// this point. This ensures every comparison that affects our operation
// ends up with predicateinfo.
if (!Result.Def)
Result.Def = materializeStack(Counter, RenameStack, Op);
DEBUG(dbgs() << "Found replacement " << *Result.Def << " for "
<< *VD.U->get() << " in " << *(VD.U->getUser()) << "\n");
assert(DT.dominates(cast<Instruction>(Result.Def), *VD.U) &&
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
"Predicateinfo def should have dominated this use");
VD.U->set(Result.Def);
Add PredicateInfo utility and printing pass Summary: This patch adds a utility to build extended SSA (see "ABCD: eliminating array bounds checks on demand"), and an intrinsic to support it. This is then used to get functionality equivalent to propagateEquality in GVN, in NewGVN (without having to replace instructions as we go). It would work similarly in SCCP or other passes. This has been talked about a few times, so i built a real implementation and tried to productionize it. Copies are inserted for operands used in assumes and conditional branches that are based on comparisons (see below for more) Every use affected by the predicate is renamed to the appropriate intrinsic result. E.g. %cmp = icmp eq i32 %x, 50 br i1 %cmp, label %true, label %false true: ret i32 %x false: ret i32 1 will become %cmp = icmp eq i32, %x, 50 br i1 %cmp, label %true, label %false true: ; Has predicate info ; branch predicate info { TrueEdge: 1 Comparison: %cmp = icmp eq i32 %x, 50 } %x.0 = call @llvm.ssa_copy.i32(i32 %x) ret i32 %x.0 false: ret i23 1 (you can use -print-predicateinfo to get an annotated-with-predicateinfo dump) This enables us to easily determine what operations are affected by a given predicate, and how operations affected by a chain of predicates. Reviewers: davide, sanjoy Subscribers: mgorny, llvm-commits, Prazek Differential Revision: https://reviews.llvm.org/D29519 Update for review comments Fix a bug Nuno noticed where we are giving information about and/or on edges where the info is not useful and easy to use wrong Update for review comments llvm-svn: 294351
2017-02-08 05:10:46 +08:00
}
}
}
PredicateInfo::ValueInfo &PredicateInfo::getOrCreateValueInfo(Value *Operand) {
auto OIN = ValueInfoNums.find(Operand);
if (OIN == ValueInfoNums.end()) {
// This will grow it
ValueInfos.resize(ValueInfos.size() + 1);
// This will use the new size and give us a 0 based number of the info
auto InsertResult = ValueInfoNums.insert({Operand, ValueInfos.size() - 1});
assert(InsertResult.second && "Value info number already existed?");
return ValueInfos[InsertResult.first->second];
}
return ValueInfos[OIN->second];
}
const PredicateInfo::ValueInfo &
PredicateInfo::getValueInfo(Value *Operand) const {
auto OINI = ValueInfoNums.lookup(Operand);
assert(OINI != 0 && "Operand was not really in the Value Info Numbers");
assert(OINI < ValueInfos.size() &&
"Value Info Number greater than size of Value Info Table");
return ValueInfos[OINI];
}
PredicateInfo::PredicateInfo(Function &F, DominatorTree &DT,
AssumptionCache &AC)
: F(F), DT(DT), AC(AC) {
// Push an empty operand info so that we can detect 0 as not finding one
ValueInfos.resize(1);
buildPredicateInfo();
}
PredicateInfo::~PredicateInfo() {}
void PredicateInfo::verifyPredicateInfo() const {}
char PredicateInfoPrinterLegacyPass::ID = 0;
PredicateInfoPrinterLegacyPass::PredicateInfoPrinterLegacyPass()
: FunctionPass(ID) {
initializePredicateInfoPrinterLegacyPassPass(
*PassRegistry::getPassRegistry());
}
void PredicateInfoPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<DominatorTreeWrapperPass>();
AU.addRequired<AssumptionCacheTracker>();
}
bool PredicateInfoPrinterLegacyPass::runOnFunction(Function &F) {
auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
auto PredInfo = make_unique<PredicateInfo>(F, DT, AC);
PredInfo->print(dbgs());
if (VerifyPredicateInfo)
PredInfo->verifyPredicateInfo();
return false;
}
PreservedAnalyses PredicateInfoPrinterPass::run(Function &F,
FunctionAnalysisManager &AM) {
auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
auto &AC = AM.getResult<AssumptionAnalysis>(F);
OS << "PredicateInfo for function: " << F.getName() << "\n";
make_unique<PredicateInfo>(F, DT, AC)->print(OS);
return PreservedAnalyses::all();
}
/// \brief An assembly annotator class to print PredicateInfo information in
/// comments.
class PredicateInfoAnnotatedWriter : public AssemblyAnnotationWriter {
friend class PredicateInfo;
const PredicateInfo *PredInfo;
public:
PredicateInfoAnnotatedWriter(const PredicateInfo *M) : PredInfo(M) {}
virtual void emitBasicBlockStartAnnot(const BasicBlock *BB,
formatted_raw_ostream &OS) {}
virtual void emitInstructionAnnot(const Instruction *I,
formatted_raw_ostream &OS) {
if (const auto *PI = PredInfo->getPredicateInfoFor(I)) {
OS << "; Has predicate info\n";
if (const auto *PB = dyn_cast<PredicateBranch>(PI))
OS << "; branch predicate info { TrueEdge: " << PB->TrueEdge
<< " Comparison:" << *PB->Comparison << " }\n";
else if (const auto *PA = dyn_cast<PredicateAssume>(PI))
OS << "; assume predicate info {"
<< " Comparison:" << *PA->Comparison << " }\n";
}
}
};
void PredicateInfo::print(raw_ostream &OS) const {
PredicateInfoAnnotatedWriter Writer(this);
F.print(OS, &Writer);
}
void PredicateInfo::dump() const {
PredicateInfoAnnotatedWriter Writer(this);
F.print(dbgs(), &Writer);
}
PreservedAnalyses PredicateInfoVerifierPass::run(Function &F,
FunctionAnalysisManager &AM) {
auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
auto &AC = AM.getResult<AssumptionAnalysis>(F);
make_unique<PredicateInfo>(F, DT, AC)->verifyPredicateInfo();
return PreservedAnalyses::all();
}
}