Adapt to dynamic side metadata address

[qinsoon]

No description provided.

[qinsoon]

julia-version
JULIA_REPO=qinsoon/julia
JULIA_REF=no-const-side-metadata-address

[qinsoon]

Tried with 3 benchmarks from Julia's GCBenchmarks (tree_mutable, tree_immutable and a ported gcbench). This PR slows down the total time by roughly 1%.

I tried a few optimizations for the write barrier fastpath. However, I didn't observe any speed up.

1. Generate a load for the side metadata base address only when we are generating images; otherwise for JIT, we simply emit the address as a constant.
2. Use LLVM's invariant load (assuming the loaded address is always the same)

Value *metadata_base_ptr;
if (jl_generating_output()) {
    F.getParent()->getOrInsertGlobal("MMTK_SIDE_LOG_BIT_BASE_ADDRESS", i8_ptr_ty);
    auto metadata_base_global = F.getParent()->getNamedGlobal("MMTK_SIDE_LOG_BIT_BASE_ADDRESS");
    assert(metadata_base_global != nullptr);
    auto metadata_base_load = builder.CreateAlignedLoad(
        i8_ptr_ty, metadata_base_global, Align(sizeof(void *)), "mmtk_side_log_bit_base");
    metadata_base_load->setMetadata(llvm::LLVMContext::MD_tbaa, get_tbaa_const(F.getContext()));
    metadata_base_load->setMetadata(llvm::LLVMContext::MD_invariant_load,
                                    llvm::MDNode::get(F.getContext(), {}));
    metadata_base_ptr = builder.CreateAlignedLoad(
        i8_ptr_ty,
        metadata_base_global,
        Align(sizeof(void *)),
        "mmtk_side_log_bit_base");
} else {
    intptr_t metadata_base_address = reinterpret_cast<intptr_t>(MMTK_SIDE_LOG_BIT_BASE_ADDRESS);
    auto metadata_base_val = ConstantInt::get(intptr_ty, metadata_base_address);
    metadata_base_ptr = ConstantExpr::getIntToPtr(metadata_base_val, PointerType::get(i8_ty, 0));
}