llvm-project/polly/lib/CodeGen/LoopGenerators.cpp
Karthika Devi C 76672e3349
[Polly] Add vectorize metadata to loops identified as vectorizable by polly (#113994)
This patch introduces the initial implementation for annotating loops
created by Polly. Polly generates RunTimeChecks (RTCs), which result in
loop versioning. Specifically, the loop created by Polly is executed
when the RTCs pass, otherwise, the original loop is executed.

This patch adds the "llvm.loop.vectorize.enable" metadata, setting it to
true for loops created by Polly. Disabling vectorization for the original
fallback loop is already merged in #119188.

This behavior is controlled by the 'polly-annotate-metadata-vectorize'
flag, and the annotations are applied only when this flag is enabled.
This flag is set to false by default.

NOTE: This commit is initial patch in effort to make polly interact with
Loop Vectorizer via metadata.

---------

Co-authored-by: Michael Kruse <github@meinersbur.de>
2025-01-22 14:35:26 +05:30

280 lines
11 KiB
C++

//===------ LoopGenerators.cpp - IR helper to create loops ---------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This file contains functions to create scalar loops and orchestrate the
// creation of parallel loops as LLVM-IR.
//
//===----------------------------------------------------------------------===//
#include "polly/CodeGen/LoopGenerators.h"
#include "polly/Options.h"
#include "polly/ScopDetection.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
using namespace llvm;
using namespace polly;
int polly::PollyNumThreads;
OMPGeneralSchedulingType polly::PollyScheduling;
int polly::PollyChunkSize;
static cl::opt<int, true>
XPollyNumThreads("polly-num-threads",
cl::desc("Number of threads to use (0 = auto)"),
cl::Hidden, cl::location(polly::PollyNumThreads),
cl::init(0), cl::cat(PollyCategory));
cl::opt<bool> PollyVectorizeMetadata(
"polly-annotate-metadata-vectorize",
cl::desc("Append vectorize enable/disable metadata from polly"),
cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory));
static cl::opt<OMPGeneralSchedulingType, true> XPollyScheduling(
"polly-scheduling",
cl::desc("Scheduling type of parallel OpenMP for loops"),
cl::values(clEnumValN(OMPGeneralSchedulingType::StaticChunked, "static",
"Static scheduling"),
clEnumValN(OMPGeneralSchedulingType::Dynamic, "dynamic",
"Dynamic scheduling"),
clEnumValN(OMPGeneralSchedulingType::Guided, "guided",
"Guided scheduling"),
clEnumValN(OMPGeneralSchedulingType::Runtime, "runtime",
"Runtime determined (OMP_SCHEDULE)")),
cl::Hidden, cl::location(polly::PollyScheduling),
cl::init(OMPGeneralSchedulingType::Runtime), cl::Optional,
cl::cat(PollyCategory));
static cl::opt<int, true>
XPollyChunkSize("polly-scheduling-chunksize",
cl::desc("Chunksize to use by the OpenMP runtime calls"),
cl::Hidden, cl::location(polly::PollyChunkSize),
cl::init(0), cl::Optional, cl::cat(PollyCategory));
// We generate a loop of either of the following structures:
//
// BeforeBB BeforeBB
// | |
// v v
// GuardBB PreHeaderBB
// / | | _____
// __ PreHeaderBB | v \/ |
// / \ / | HeaderBB latch
// latch HeaderBB | |\ |
// \ / \ / | \------/
// < \ / |
// \ / v
// ExitBB ExitBB
//
// depending on whether or not we know that it is executed at least once. If
// not, GuardBB checks if the loop is executed at least once. If this is the
// case we branch to PreHeaderBB and subsequently to the HeaderBB, which
// contains the loop iv 'polly.indvar', the incremented loop iv
// 'polly.indvar_next' as well as the condition to check if we execute another
// iteration of the loop. After the loop has finished, we branch to ExitBB.
// We expect the type of UB, LB, UB+Stride to be large enough for values that
// UB may take throughout the execution of the loop, including the computation
// of indvar + Stride before the final abort.
Value *polly::createLoop(Value *LB, Value *UB, Value *Stride,
PollyIRBuilder &Builder, LoopInfo &LI,
DominatorTree &DT, BasicBlock *&ExitBB,
ICmpInst::Predicate Predicate,
ScopAnnotator *Annotator, bool Parallel, bool UseGuard,
bool LoopVectDisabled) {
Function *F = Builder.GetInsertBlock()->getParent();
LLVMContext &Context = F->getContext();
assert(LB->getType() == UB->getType() && "Types of loop bounds do not match");
IntegerType *LoopIVType = dyn_cast<IntegerType>(UB->getType());
assert(LoopIVType && "UB is not integer?");
BasicBlock *BeforeBB = Builder.GetInsertBlock();
BasicBlock *GuardBB =
UseGuard ? BasicBlock::Create(Context, "polly.loop_if", F) : nullptr;
BasicBlock *HeaderBB = BasicBlock::Create(Context, "polly.loop_header", F);
BasicBlock *PreHeaderBB =
BasicBlock::Create(Context, "polly.loop_preheader", F);
// Update LoopInfo
Loop *OuterLoop = LI.getLoopFor(BeforeBB);
Loop *NewLoop = LI.AllocateLoop();
if (OuterLoop)
OuterLoop->addChildLoop(NewLoop);
else
LI.addTopLevelLoop(NewLoop);
if (OuterLoop) {
if (GuardBB)
OuterLoop->addBasicBlockToLoop(GuardBB, LI);
OuterLoop->addBasicBlockToLoop(PreHeaderBB, LI);
}
NewLoop->addBasicBlockToLoop(HeaderBB, LI);
// Notify the annotator (if present) that we have a new loop, but only
// after the header block is set.
if (Annotator)
Annotator->pushLoop(NewLoop, Parallel);
// ExitBB
ExitBB = SplitBlock(BeforeBB, &*Builder.GetInsertPoint(), &DT, &LI);
ExitBB->setName("polly.loop_exit");
// BeforeBB
if (GuardBB) {
BeforeBB->getTerminator()->setSuccessor(0, GuardBB);
DT.addNewBlock(GuardBB, BeforeBB);
// GuardBB
Builder.SetInsertPoint(GuardBB);
Value *LoopGuard;
LoopGuard = Builder.CreateICmp(Predicate, LB, UB);
LoopGuard->setName("polly.loop_guard");
Builder.CreateCondBr(LoopGuard, PreHeaderBB, ExitBB);
DT.addNewBlock(PreHeaderBB, GuardBB);
} else {
BeforeBB->getTerminator()->setSuccessor(0, PreHeaderBB);
DT.addNewBlock(PreHeaderBB, BeforeBB);
}
// PreHeaderBB
Builder.SetInsertPoint(PreHeaderBB);
Builder.CreateBr(HeaderBB);
// HeaderBB
DT.addNewBlock(HeaderBB, PreHeaderBB);
Builder.SetInsertPoint(HeaderBB);
PHINode *IV = Builder.CreatePHI(LoopIVType, 2, "polly.indvar");
IV->addIncoming(LB, PreHeaderBB);
Stride = Builder.CreateZExtOrBitCast(Stride, LoopIVType);
Value *IncrementedIV = Builder.CreateNSWAdd(IV, Stride, "polly.indvar_next");
Value *LoopCondition =
Builder.CreateICmp(Predicate, IncrementedIV, UB, "polly.loop_cond");
// Create the loop latch and annotate it as such.
BranchInst *B = Builder.CreateCondBr(LoopCondition, HeaderBB, ExitBB);
// Don't annotate vectorize metadata when both LoopVectDisabled and
// PollyVectorizeMetadata are disabled. Annotate vectorize metadata to false
// when LoopVectDisabled is true. Otherwise we annotate the vectorize metadata
// to true.
if (Annotator) {
std::optional<bool> EnableVectorizeMetadata;
if (LoopVectDisabled)
EnableVectorizeMetadata = false;
else if (PollyVectorizeMetadata)
EnableVectorizeMetadata = true;
Annotator->annotateLoopLatch(B, Parallel, EnableVectorizeMetadata);
}
IV->addIncoming(IncrementedIV, HeaderBB);
if (GuardBB)
DT.changeImmediateDominator(ExitBB, GuardBB);
else
DT.changeImmediateDominator(ExitBB, HeaderBB);
// The loop body should be added here.
Builder.SetInsertPoint(HeaderBB->getFirstNonPHI());
return IV;
}
Value *ParallelLoopGenerator::createParallelLoop(
Value *LB, Value *UB, Value *Stride, SetVector<Value *> &UsedValues,
ValueMapT &Map, BasicBlock::iterator *LoopBody) {
AllocaInst *Struct = storeValuesIntoStruct(UsedValues);
BasicBlock::iterator BeforeLoop = Builder.GetInsertPoint();
Value *IV;
Function *SubFn;
std::tie(IV, SubFn) = createSubFn(Stride, Struct, UsedValues, Map);
*LoopBody = Builder.GetInsertPoint();
Builder.SetInsertPoint(&*BeforeLoop);
// Add one as the upper bound provided by OpenMP is a < comparison
// whereas the codegenForSequential function creates a <= comparison.
UB = Builder.CreateAdd(UB, ConstantInt::get(LongType, 1));
// Execute the prepared subfunction in parallel.
deployParallelExecution(SubFn, Struct, LB, UB, Stride);
return IV;
}
Function *ParallelLoopGenerator::createSubFnDefinition() {
Function *F = Builder.GetInsertBlock()->getParent();
Function *SubFn = prepareSubFnDefinition(F);
// Certain backends (e.g., NVPTX) do not support '.'s in function names.
// Hence, we ensure that all '.'s are replaced by '_'s.
std::string FunctionName = SubFn->getName().str();
std::replace(FunctionName.begin(), FunctionName.end(), '.', '_');
SubFn->setName(FunctionName);
// Do not run any polly pass on the new function.
SubFn->addFnAttr(PollySkipFnAttr);
return SubFn;
}
AllocaInst *
ParallelLoopGenerator::storeValuesIntoStruct(SetVector<Value *> &Values) {
SmallVector<Type *, 8> Members;
for (Value *V : Values)
Members.push_back(V->getType());
const DataLayout &DL = Builder.GetInsertBlock()->getModule()->getDataLayout();
// We do not want to allocate the alloca inside any loop, thus we allocate it
// in the entry block of the function and use annotations to denote the actual
// live span (similar to clang).
BasicBlock &EntryBB = Builder.GetInsertBlock()->getParent()->getEntryBlock();
BasicBlock::iterator IP = EntryBB.getFirstInsertionPt();
StructType *Ty = StructType::get(Builder.getContext(), Members);
AllocaInst *Struct = new AllocaInst(Ty, DL.getAllocaAddrSpace(), nullptr,
"polly.par.userContext", IP);
for (unsigned i = 0; i < Values.size(); i++) {
Value *Address = Builder.CreateStructGEP(Ty, Struct, i);
Address->setName("polly.subfn.storeaddr." + Values[i]->getName());
Builder.CreateStore(Values[i], Address);
}
return Struct;
}
void ParallelLoopGenerator::extractValuesFromStruct(
SetVector<Value *> OldValues, Type *Ty, Value *Struct, ValueMapT &Map) {
for (unsigned i = 0; i < OldValues.size(); i++) {
Value *Address = Builder.CreateStructGEP(Ty, Struct, i);
Type *ElemTy = cast<GetElementPtrInst>(Address)->getResultElementType();
Value *NewValue = Builder.CreateLoad(ElemTy, Address);
NewValue->setName("polly.subfunc.arg." + OldValues[i]->getName());
Map[OldValues[i]] = NewValue;
}
}
DebugLoc polly::createDebugLocForGeneratedCode(Function *F) {
if (!F)
return DebugLoc();
LLVMContext &Ctx = F->getContext();
DISubprogram *DILScope =
dyn_cast_or_null<DISubprogram>(F->getMetadata(LLVMContext::MD_dbg));
if (!DILScope)
return DebugLoc();
return DILocation::get(Ctx, 0, 0, DILScope);
}