Skip to content

Allow skip inlined fastpath #19

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
qinsoon merged 3 commits into from
Jun 30, 2023
Merged

Allow skip inlined fastpath #19

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ec9a7a0
Allow skip inlined fastpath
qinsoon Jun 29, 2023
4c901a0
Add missing derefAttr
qinsoon Jun 29, 2023
e6cdfea
Update llvm-final-gc-lowering.cpp
qinsoon Jun 29, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
146 changes: 78 additions & 68 deletions src/llvm-final-gc-lowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -281,74 +281,84 @@ Value *FinalLowerGC::lowerGCAllocBytes(CallInst *target, Function &F)
auto pool_osize_i32 = ConstantInt::get(Type::getInt32Ty(F.getContext()), osize);
auto pool_osize = ConstantInt::get(Type::getInt64Ty(F.getContext()), osize);

// Assuming we use the first immix allocator.
// FIXME: We should get the allocator index and type from MMTk.
auto allocator_offset = offsetof(jl_tls_states_t, mmtk_mutator) + offsetof(MMTkMutatorContext, allocators) + offsetof(Allocators, immix);

auto cursor_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), allocator_offset + offsetof(ImmixAllocator, cursor));
auto limit_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), allocator_offset + offsetof(ImmixAllocator, limit));

auto cursor_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, cursor_pos);
auto cursor_ptr = builder.CreateBitCast(cursor_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "cursor_ptr");
auto cursor = builder.CreateLoad(Type::getInt64Ty(target->getContext()), cursor_ptr, "cursor");

// offset = 8
auto delta_offset = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), ConstantInt::get(Type::getInt64Ty(target->getContext()), 8));
auto delta_cursor = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), cursor);
auto delta_op = builder.CreateNSWAdd(delta_offset, delta_cursor);
// alignment 16 (15 = 16 - 1)
auto delta = builder.CreateAnd(delta_op, ConstantInt::get(Type::getInt64Ty(target->getContext()), 15), "delta");
auto result = builder.CreateNSWAdd(cursor, delta, "result");

auto new_cursor = builder.CreateNSWAdd(result, pool_osize);

auto limit_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, limit_pos);
auto limit_ptr = builder.CreateBitCast(limit_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "limit_ptr");
auto limit = builder.CreateLoad(Type::getInt64Ty(target->getContext()), limit_ptr, "limit");

auto gt_limit = builder.CreateICmpSGT(new_cursor, limit);

auto current_block = target->getParent();
builder.SetInsertPoint(target->getNextNode());
auto phiNode = builder.CreatePHI(poolAllocFunc->getReturnType(), 2, "phi_fast_slow");
auto top_cont = current_block->splitBasicBlock(target->getNextNode(), "top_cont");

auto slowpath = BasicBlock::Create(target->getContext(), "slowpath", target->getFunction());
auto fastpath = BasicBlock::Create(target->getContext(), "fastpath", target->getFunction(), top_cont);

auto next_br = current_block->getTerminator();
next_br->eraseFromParent();
builder.SetInsertPoint(current_block);
builder.CreateCondBr(gt_limit, slowpath, fastpath);

// slowpath
builder.SetInsertPoint(slowpath);
auto pool_offs = ConstantInt::get(Type::getInt32Ty(F.getContext()), 1);
auto new_call = builder.CreateCall(poolAllocFunc, { ptls, pool_offs, pool_osize_i32 });
new_call->setAttributes(new_call->getCalledFunction()->getAttributes());
builder.CreateBr(top_cont);

// // fastpath
builder.SetInsertPoint(fastpath);
builder.CreateStore(new_cursor, cursor_ptr);

// ptls->gc_num.allocd += osize;
auto pool_alloc_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), offsetof(jl_tls_states_t, gc_num));
auto pool_alloc_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, pool_alloc_pos);
auto pool_alloc_tls = builder.CreateBitCast(pool_alloc_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "pool_alloc");
auto pool_allocd = builder.CreateLoad(Type::getInt64Ty(target->getContext()), pool_alloc_tls);
auto pool_allocd_total = builder.CreateAdd(pool_allocd, pool_osize);
builder.CreateStore(pool_allocd_total, pool_alloc_tls);

auto v_raw = builder.CreateNSWAdd(result, ConstantInt::get(Type::getInt64Ty(target->getContext()), sizeof(jl_taggedvalue_t)));
auto v_as_ptr = builder.CreateIntToPtr(v_raw, poolAllocFunc->getReturnType());
builder.CreateBr(top_cont);

phiNode->addIncoming(new_call, slowpath);
phiNode->addIncoming(v_as_ptr, fastpath);
phiNode->takeName(target);

return phiNode;
// Should we generate fastpath allocation sequence here? We should always generate fastpath here for MMTk.
// Setting this to false will increase allocation overhead a lot, and should only be used for debugging.
const bool INLINE_FASTPATH_ALLOCATION = true;

if (INLINE_FASTPATH_ALLOCATION) {
// Assuming we use the first immix allocator.
// FIXME: We should get the allocator index and type from MMTk.
auto allocator_offset = offsetof(jl_tls_states_t, mmtk_mutator) + offsetof(MMTkMutatorContext, allocators) + offsetof(Allocators, immix);

auto cursor_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), allocator_offset + offsetof(ImmixAllocator, cursor));
auto limit_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), allocator_offset + offsetof(ImmixAllocator, limit));

auto cursor_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, cursor_pos);
auto cursor_ptr = builder.CreateBitCast(cursor_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "cursor_ptr");
auto cursor = builder.CreateLoad(Type::getInt64Ty(target->getContext()), cursor_ptr, "cursor");

// offset = 8
auto delta_offset = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), ConstantInt::get(Type::getInt64Ty(target->getContext()), 8));
auto delta_cursor = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), cursor);
auto delta_op = builder.CreateNSWAdd(delta_offset, delta_cursor);
// alignment 16 (15 = 16 - 1)
auto delta = builder.CreateAnd(delta_op, ConstantInt::get(Type::getInt64Ty(target->getContext()), 15), "delta");
auto result = builder.CreateNSWAdd(cursor, delta, "result");

auto new_cursor = builder.CreateNSWAdd(result, pool_osize);

auto limit_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, limit_pos);
auto limit_ptr = builder.CreateBitCast(limit_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "limit_ptr");
auto limit = builder.CreateLoad(Type::getInt64Ty(target->getContext()), limit_ptr, "limit");

auto gt_limit = builder.CreateICmpSGT(new_cursor, limit);

auto current_block = target->getParent();
builder.SetInsertPoint(target->getNextNode());
auto phiNode = builder.CreatePHI(poolAllocFunc->getReturnType(), 2, "phi_fast_slow");
auto top_cont = current_block->splitBasicBlock(target->getNextNode(), "top_cont");

auto slowpath = BasicBlock::Create(target->getContext(), "slowpath", target->getFunction());
auto fastpath = BasicBlock::Create(target->getContext(), "fastpath", target->getFunction(), top_cont);

auto next_br = current_block->getTerminator();
next_br->eraseFromParent();
builder.SetInsertPoint(current_block);
builder.CreateCondBr(gt_limit, slowpath, fastpath);

// slowpath
builder.SetInsertPoint(slowpath);
auto pool_offs = ConstantInt::get(Type::getInt32Ty(F.getContext()), 1);
auto new_call = builder.CreateCall(poolAllocFunc, { ptls, pool_offs, pool_osize_i32 });
new_call->setAttributes(new_call->getCalledFunction()->getAttributes());
builder.CreateBr(top_cont);

// // fastpath
builder.SetInsertPoint(fastpath);
builder.CreateStore(new_cursor, cursor_ptr);

// ptls->gc_num.allocd += osize;
auto pool_alloc_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), offsetof(jl_tls_states_t, gc_num));
auto pool_alloc_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, pool_alloc_pos);
auto pool_alloc_tls = builder.CreateBitCast(pool_alloc_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "pool_alloc");
auto pool_allocd = builder.CreateLoad(Type::getInt64Ty(target->getContext()), pool_alloc_tls);
auto pool_allocd_total = builder.CreateAdd(pool_allocd, pool_osize);
builder.CreateStore(pool_allocd_total, pool_alloc_tls);

auto v_raw = builder.CreateNSWAdd(result, ConstantInt::get(Type::getInt64Ty(target->getContext()), sizeof(jl_taggedvalue_t)));
auto v_as_ptr = builder.CreateIntToPtr(v_raw, poolAllocFunc->getReturnType());
builder.CreateBr(top_cont);

phiNode->addIncoming(new_call, slowpath);
phiNode->addIncoming(v_as_ptr, fastpath);
phiNode->takeName(target);

return phiNode;
} else {
auto pool_offs = ConstantInt::get(Type::getInt32Ty(F.getContext()), 1);
newI = builder.CreateCall(poolAllocFunc, { ptls, pool_offs, pool_osize_i32 });
derefAttr = Attribute::getWithDereferenceableBytes(F.getContext(), osize);
}
#endif // MMTK_GC
}
} else {
Expand Down