
This metadata was intended to mark all accesses within an iteration to be pairwise non-aliasing, in this case because every memory of a base pointer is touched (read or write) at most once. This is typical for 'sweeps' over all data. The stated motivation from D30606 is to ensure that unrolled iterations are considered non-aliasing. Rhe implemention had multiple issues: * The structure of the noalias metadata was malformed. D110026 added check in the verifier for this metadata, and the tests were failing since then. * This is not true for the outer loops of the BLIS matrix multiplication, where it was being inserted. Each element of A, B, C is accessed multiple times, as often as the loop not used as an index is iterating. * Scopes were added to SecondLevelOtherAliasScopeList (used for the !noalias scop list) on-the-fly when another SCEV was seen. This meant that previously visited instructions would not be updated with alias scopes that are only seen later, missing out those SCEVs they should not be aliasing with. * Since the !noalias scope list would ideally consists of all other SCEV for this base pointer, we might run quickly into scalability issues. Especially after unrolling there would probably at least once SCEV per instruction and unroll instance. * The inter-iteration noalias base pointer was not removed after leaving the loop marked with it, effectively marking everything after it to noalias as well. A solution I considered was to mark each instruction as non-aliasing with its own scope. The instruction itself would obviously alias itself, but such construction might also be considered invalid. Duplicating the instruction (e.g. due to speculation) would mark the instruction non-aliasing with its clone. I don't want to go into this territory, especially since the original motivation of determining unrolled instances as noalias based on SCEV is the what scev-aa does as well. This effectively reverts D30606 and D35761.
271 lines
8.3 KiB
C++
271 lines
8.3 KiB
C++
//===------ PollyIRBuilder.cpp --------------------------------------------===//
|
|
//
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
//
|
|
// The Polly IRBuilder file contains Polly specific extensions for the IRBuilder
|
|
// that are used e.g. to emit the llvm.loop.parallel metadata.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include "polly/CodeGen/IRBuilder.h"
|
|
#include "polly/ScopInfo.h"
|
|
#include "polly/Support/ScopHelper.h"
|
|
#include "llvm/ADT/SmallVector.h"
|
|
#include "llvm/IR/Metadata.h"
|
|
|
|
using namespace llvm;
|
|
using namespace polly;
|
|
|
|
static const int MaxArraysInAliasScops = 10;
|
|
|
|
/// Get a self referencing id metadata node.
|
|
///
|
|
/// The MDNode looks like this (if arg0/arg1 are not null):
|
|
///
|
|
/// '!n = distinct !{!n, arg0, arg1}'
|
|
///
|
|
/// @return The self referencing id metadata node.
|
|
static MDNode *getID(LLVMContext &Ctx, Metadata *arg0 = nullptr,
|
|
Metadata *arg1 = nullptr) {
|
|
MDNode *ID;
|
|
SmallVector<Metadata *, 3> Args;
|
|
// Reserve operand 0 for loop id self reference.
|
|
Args.push_back(nullptr);
|
|
|
|
if (arg0)
|
|
Args.push_back(arg0);
|
|
if (arg1)
|
|
Args.push_back(arg1);
|
|
|
|
ID = MDNode::getDistinct(Ctx, Args);
|
|
ID->replaceOperandWith(0, ID);
|
|
return ID;
|
|
}
|
|
|
|
ScopAnnotator::ScopAnnotator() : SE(nullptr), AliasScopeDomain(nullptr) {
|
|
// Push an empty staging BandAttr.
|
|
LoopAttrEnv.emplace_back();
|
|
}
|
|
|
|
ScopAnnotator::~ScopAnnotator() {
|
|
assert(LoopAttrEnv.size() == 1 && "Loop stack imbalance");
|
|
assert(!getStagingAttrEnv() && "Forgot to clear staging attr env");
|
|
}
|
|
|
|
void ScopAnnotator::buildAliasScopes(Scop &S) {
|
|
SE = S.getSE();
|
|
|
|
LLVMContext &Ctx = SE->getContext();
|
|
AliasScopeDomain = getID(Ctx, MDString::get(Ctx, "polly.alias.scope.domain"));
|
|
|
|
AliasScopeMap.clear();
|
|
OtherAliasScopeListMap.clear();
|
|
|
|
// We are only interested in arrays, but no scalar references. Scalars should
|
|
// be handled easily by basicaa.
|
|
SmallVector<ScopArrayInfo *, 10> Arrays;
|
|
for (ScopArrayInfo *Array : S.arrays())
|
|
if (Array->isArrayKind())
|
|
Arrays.push_back(Array);
|
|
|
|
// The construction of alias scopes is quadratic in the number of arrays
|
|
// involved. In case of too many arrays, skip the construction of alias
|
|
// information to avoid quadratic increases in compile time and code size.
|
|
if (Arrays.size() > MaxArraysInAliasScops)
|
|
return;
|
|
|
|
std::string AliasScopeStr = "polly.alias.scope.";
|
|
for (const ScopArrayInfo *Array : Arrays) {
|
|
assert(Array->getBasePtr() && "Base pointer must be present");
|
|
AliasScopeMap[Array->getBasePtr()] =
|
|
getID(Ctx, AliasScopeDomain,
|
|
MDString::get(Ctx, (AliasScopeStr + Array->getName()).c_str()));
|
|
}
|
|
|
|
for (const ScopArrayInfo *Array : Arrays) {
|
|
MDNode *AliasScopeList = MDNode::get(Ctx, {});
|
|
for (const auto &AliasScopePair : AliasScopeMap) {
|
|
if (Array->getBasePtr() == AliasScopePair.first)
|
|
continue;
|
|
|
|
Metadata *Args = {AliasScopePair.second};
|
|
AliasScopeList =
|
|
MDNode::concatenate(AliasScopeList, MDNode::get(Ctx, Args));
|
|
}
|
|
|
|
OtherAliasScopeListMap[Array->getBasePtr()] = AliasScopeList;
|
|
}
|
|
}
|
|
|
|
void ScopAnnotator::pushLoop(Loop *L, bool IsParallel) {
|
|
ActiveLoops.push_back(L);
|
|
|
|
if (IsParallel) {
|
|
LLVMContext &Ctx = SE->getContext();
|
|
MDNode *AccessGroup = MDNode::getDistinct(Ctx, {});
|
|
ParallelLoops.push_back(AccessGroup);
|
|
}
|
|
|
|
// Open an empty BandAttr context for loops nested in this one.
|
|
LoopAttrEnv.emplace_back();
|
|
}
|
|
|
|
void ScopAnnotator::popLoop(bool IsParallel) {
|
|
ActiveLoops.pop_back();
|
|
|
|
if (IsParallel) {
|
|
assert(!ParallelLoops.empty() && "Expected a parallel loop to pop");
|
|
ParallelLoops.pop_back();
|
|
}
|
|
|
|
// Exit the subloop context.
|
|
assert(!getStagingAttrEnv() && "Forgot to clear staging attr env");
|
|
assert(LoopAttrEnv.size() >= 2 && "Popped too many");
|
|
LoopAttrEnv.pop_back();
|
|
}
|
|
|
|
void ScopAnnotator::annotateLoopLatch(BranchInst *B, Loop *L, bool IsParallel,
|
|
bool IsLoopVectorizerDisabled) const {
|
|
LLVMContext &Ctx = SE->getContext();
|
|
SmallVector<Metadata *, 3> Args;
|
|
|
|
// For the LoopID self-reference.
|
|
Args.push_back(nullptr);
|
|
|
|
// Add the user-defined loop properties to the annotation, if any. Any
|
|
// additional properties are appended.
|
|
// FIXME: What to do if these conflict?
|
|
MDNode *MData = nullptr;
|
|
if (BandAttr *AttrEnv = getActiveAttrEnv()) {
|
|
MData = AttrEnv->Metadata;
|
|
if (MData)
|
|
llvm::append_range(Args, drop_begin(MData->operands(), 1));
|
|
}
|
|
|
|
if (IsLoopVectorizerDisabled) {
|
|
MDString *PropName = MDString::get(Ctx, "llvm.loop.vectorize.enable");
|
|
ConstantInt *FalseValue = ConstantInt::get(Type::getInt1Ty(Ctx), 0);
|
|
ValueAsMetadata *PropValue = ValueAsMetadata::get(FalseValue);
|
|
Args.push_back(MDNode::get(Ctx, {PropName, PropValue}));
|
|
}
|
|
|
|
if (IsParallel) {
|
|
MDString *PropName = MDString::get(Ctx, "llvm.loop.parallel_accesses");
|
|
MDNode *AccGroup = ParallelLoops.back();
|
|
Args.push_back(MDNode::get(Ctx, {PropName, AccGroup}));
|
|
}
|
|
|
|
// No metadata to annotate.
|
|
if (!MData && Args.size() <= 1)
|
|
return;
|
|
|
|
// Reuse the MData node if possible, this will avoid having to create another
|
|
// one that cannot be merged because LoopIDs are 'distinct'. However, we have
|
|
// to create a new one if we add properties.
|
|
if (!MData || Args.size() > MData->getNumOperands()) {
|
|
MData = MDNode::getDistinct(Ctx, Args);
|
|
MData->replaceOperandWith(0, MData);
|
|
}
|
|
B->setMetadata(LLVMContext::MD_loop, MData);
|
|
}
|
|
|
|
/// Get the pointer operand
|
|
///
|
|
/// @param Inst The instruction to be analyzed.
|
|
/// @return the pointer operand in case @p Inst is a memory access
|
|
/// instruction and nullptr otherwise.
|
|
static llvm::Value *getMemAccInstPointerOperand(Instruction *Inst) {
|
|
auto MemInst = MemAccInst::dyn_cast(Inst);
|
|
if (!MemInst)
|
|
return nullptr;
|
|
|
|
return MemInst.getPointerOperand();
|
|
}
|
|
|
|
/// Find the base pointer of an array access.
|
|
///
|
|
/// This should be equivalent to ScalarEvolution::getPointerBase, which we
|
|
/// cannot use here the IR is still under construction which ScalarEvolution
|
|
/// assumes to not be modified.
|
|
static Value *findBasePtr(Value *Val) {
|
|
while (true) {
|
|
if (auto *Gep = dyn_cast<GEPOperator>(Val)) {
|
|
Val = Gep->getPointerOperand();
|
|
continue;
|
|
}
|
|
if (auto *Cast = dyn_cast<BitCastOperator>(Val)) {
|
|
Val = Cast->getOperand(0);
|
|
continue;
|
|
}
|
|
|
|
break;
|
|
}
|
|
|
|
return Val;
|
|
}
|
|
|
|
void ScopAnnotator::annotate(Instruction *Inst) {
|
|
if (!Inst->mayReadOrWriteMemory())
|
|
return;
|
|
|
|
switch (ParallelLoops.size()) {
|
|
case 0:
|
|
// Not parallel to anything: no access group needed.
|
|
break;
|
|
case 1:
|
|
// Single parallel loop: use directly.
|
|
Inst->setMetadata(LLVMContext::MD_access_group,
|
|
cast<MDNode>(ParallelLoops.front()));
|
|
break;
|
|
default:
|
|
// Parallel to multiple loops: refer to list of access groups.
|
|
Inst->setMetadata(LLVMContext::MD_access_group,
|
|
MDNode::get(SE->getContext(),
|
|
ArrayRef<Metadata *>(
|
|
(Metadata *const *)ParallelLoops.data(),
|
|
ParallelLoops.size())));
|
|
break;
|
|
}
|
|
|
|
// TODO: Use the ScopArrayInfo once available here.
|
|
if (!AliasScopeDomain)
|
|
return;
|
|
|
|
// Do not apply annotations on memory operations that take more than one
|
|
// pointer. It would be ambiguous to which pointer the annotation applies.
|
|
// FIXME: How can we specify annotations for all pointer arguments?
|
|
if (isa<CallInst>(Inst) && !isa<MemSetInst>(Inst))
|
|
return;
|
|
|
|
auto *Ptr = getMemAccInstPointerOperand(Inst);
|
|
if (!Ptr)
|
|
return;
|
|
|
|
Value *BasePtr = findBasePtr(Ptr);
|
|
if (!BasePtr)
|
|
return;
|
|
|
|
auto AliasScope = AliasScopeMap.lookup(BasePtr);
|
|
|
|
if (!AliasScope) {
|
|
BasePtr = AlternativeAliasBases.lookup(BasePtr);
|
|
if (!BasePtr)
|
|
return;
|
|
|
|
AliasScope = AliasScopeMap.lookup(BasePtr);
|
|
if (!AliasScope)
|
|
return;
|
|
}
|
|
|
|
assert(OtherAliasScopeListMap.count(BasePtr) &&
|
|
"BasePtr either expected in AliasScopeMap and OtherAlias...Map");
|
|
auto *OtherAliasScopeList = OtherAliasScopeListMap[BasePtr];
|
|
|
|
Inst->setMetadata("alias.scope", MDNode::get(SE->getContext(), AliasScope));
|
|
Inst->setMetadata("noalias", OtherAliasScopeList);
|
|
}
|