Files
swift-mirror/lib/LLVMPasses/LLVMStackPromotion.cpp
Erik Eckstein 0dd045ca04 Add a stack promotion optimization.
It promotes allocations of native swift objects and array buffers to the stack if it is possible.
The SIL StackPromotion pass is the main part of the optimization. For details see comments there.
Unfortunately we need an additional LLVM pass to handle array buffers, which is not very nice.
I hope that we can get rid of it in future (again: for details see the comments in StackPromotion.cpp)

The optimization gives performance improvements in some benchmarks, mostly related to array literals:
ArrayLiteral: +12%
Combos: +16%
DictionaryLiteral: + 37%
RIPEMD: +10%
StringBuilder: +27%
StringInterpolation: +11%
And last but not least the new benchmark which is dedicated to test stack promotion:
ObjectAllocation: +52%
2015-11-05 16:52:59 -08:00

197 lines
7.0 KiB
C++

//===--- LLVMStackPromotion.cpp - Replace allocation calls with alloca ----===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2015 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See http://swift.org/LICENSE.txt for license information
// See http://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
//
// This pass performs the last part of stack promotion for array buffers.
// The SIL StackPromotion pass generates a pair of swift_bufferAllocateOnStack
// and swift_bufferDeallocateFromStack calls. In this pass the final decision
// is made if stack promotion should be done. If yes, the
// swift_bufferAllocateOnStack is replace with an alloca plus a call to
// swift_initStackObject and the swift_bufferDeallocateFromStack is removed.
// TODO: This is a hack and eventually this pass should not be required at all.
// For details see the comments for the SIL StackPromoter.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "swift-stack-promotion"
#include "swift/LLVMPasses/Passes.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/Pass.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
using namespace swift;
STATISTIC(NumBufferAllocsPromoted,
"Number of swift_bufferAllocate promoted");
cl::opt<int> LimitOpt("stack-promotion-limit",
llvm::cl::init(1024), llvm::cl::Hidden);
//===----------------------------------------------------------------------===//
// SwiftStackPromotion Pass
//===----------------------------------------------------------------------===//
char SwiftStackPromotion::ID = 0;
INITIALIZE_PASS_BEGIN(SwiftStackPromotion,
"swift-stack-promotion", "Swift stack promotion pass",
false, false)
INITIALIZE_PASS_END(SwiftStackPromotion,
"swift-stack-promotion", "Swift stack promotion pass",
false, false)
llvm::FunctionPass *swift::createSwiftStackPromotionPass() {
initializeSwiftStackPromotionPass(*llvm::PassRegistry::getPassRegistry());
return new SwiftStackPromotion();
}
void SwiftStackPromotion::getAnalysisUsage(llvm::AnalysisUsage &AU) const {
AU.setPreservesCFG();
}
/// Checks if we can promote a buffer and returns the size of the buffer.
/// The \a align parameter is set to the alignment of the buffer.
int canPromote(CallInst *CI, unsigned &align, int maxSize) {
if (CI->getNumArgOperands() != 3)
return 0;
auto *SizeConst = dyn_cast<ConstantInt>(CI->getArgOperand(1));
if (!SizeConst)
return 0;
auto *AlignMaskConst = dyn_cast<ConstantInt>(CI->getArgOperand(2));
if (!AlignMaskConst)
return 0;
int size = SizeConst->getValue().getSExtValue();
if (size > maxSize)
return 0;
align = AlignMaskConst->getValue().getZExtValue() + 1;
return size;
}
bool SwiftStackPromotion::runOnFunction(Function &F) {
bool Changed = false;
Constant *allocFunc = nullptr;
Constant *initFunc = nullptr;
int maxSize = LimitOpt;
Module *M = F.getParent();
const DataLayout &DL = M->getDataLayout();
IntegerType *AllocType = nullptr;
IntegerType *IntType = nullptr;
SmallVector<CallInst *, 8> BufferAllocs;
SmallPtrSet<CallInst *, 8> PromotedAllocs;
SmallVector<CallInst *, 8> BufferDeallocs;
// Search for allocation- and deallocation-calls in the function.
for (BasicBlock &BB : F) {
for (auto Iter = BB.begin(); Iter != BB.end(); ) {
Instruction *I = Iter;
Iter++;
if (auto *AI = dyn_cast<AllocaInst>(I)) {
int Size = 1;
if (auto *SizeConst = dyn_cast<ConstantInt>(AI->getArraySize()))
Size = SizeConst->getValue().getSExtValue();
// Count the existing alloca sizes against the limit.
maxSize -= DL.getTypeAllocSize(AI->getAllocatedType()) * Size;
}
auto *CI = dyn_cast<CallInst>(I);
if (!CI)
continue;
Function *Callee = CI->getCalledFunction();
if (!Callee)
continue;
if (Callee->getName() == "swift_bufferAllocateOnStack") {
BufferAllocs.push_back(CI);
} else if (Callee->getName() == "swift_bufferDeallocateFromStack") {
BufferDeallocs.push_back(CI);
}
}
}
// First handle allocations.
for (CallInst *CI : BufferAllocs) {
Function *Callee = CI->getCalledFunction();
assert(Callee);
unsigned align = 0;
if (int size = canPromote(CI, align, maxSize)) {
maxSize -= size;
if (!AllocType) {
// Create the swift_initStackObject function and all required types.
AllocType = IntegerType::get(M->getContext(), 8);
IntType = IntegerType::get(M->getContext(), 32);
auto *OrigFT = Callee->getFunctionType();
auto *HeapObjTy = OrigFT->getReturnType();
auto *MetaDataTy = OrigFT->getParamType(0);
auto *NewFTy = FunctionType::get(HeapObjTy,
{MetaDataTy, HeapObjTy},
false);
initFunc = M->getOrInsertFunction("swift_initStackObject", NewFTy);
}
// Replace the allocation call with an alloca.
Value *AllocA = new AllocaInst(AllocType,
ConstantInt::get(IntType, size),
align, "buffer", F.front().begin());
// And initialize it with a call to swift_initStackObject.
IRBuilder<> B(CI);
Value *casted = B.CreateBitCast(AllocA, CI->getType());
CallInst *initCall = B.CreateCall(initFunc,
{CI->getArgOperand(0), casted});
CI->replaceAllUsesWith(initCall);
CI->eraseFromParent();
PromotedAllocs.insert(initCall);
++NumBufferAllocsPromoted;
} else {
// We don't do stack promotion. Replace the call with a call to the
// regular swift_bufferAllocate.
if (!allocFunc) {
allocFunc = M->getOrInsertFunction("swift_bufferAllocate",
Callee->getFunctionType());
}
CI->setCalledFunction(allocFunc);
}
Changed = true;
}
// After we made the decision for all allocations we can handle the
// deallocations.
for (CallInst *CI : BufferDeallocs) {
CallInst *Alloc = dyn_cast<CallInst>(CI->getArgOperand(0));
assert(Alloc && "alloc buffer obfuscated");
if (PromotedAllocs.count(Alloc)) {
IRBuilder<> B(CI);
// This has two purposes:
// 1. Tell LLVM the lifetime of the allocated stack memory.
// 2. Avoid tail-call optimization which may convert the call to the final
// release to a jump, which is done after the stack frame is
// destructed.
B.CreateLifetimeEnd(CI->getArgOperand(0));
}
// Other than inserting the end-of-lifetime, the deallocation is a no-op.
CI->eraseFromParent();
Changed = true;
}
return Changed;
}