Reason to remove address spaces

KavinduZoysa · July 2, 2021, 12:07pm

Hi all,

Can you please help me to understand the reason to remove address spaces as shown in julia/llvm-remove-addrspaces.cpp at 5650c93f1aa45a17c10c9482e5d3e0042cf4cc6b · JuliaLang/julia · GitHub

Thank you,
Kavindu

maleadt · July 2, 2021, 1:04pm

Only the Julia-specific address space information (used to model the garbage collector) is removed, because certain LLVM back-ends don’t handle that information. It also makes the LLVM IR look ugly (if you pass raw=true to code_llvm the information is retained).

KavinduZoysa · July 2, 2021, 1:13pm

@maleadt Thank you, I was able to see the address space by passing raw=true as you mentioned.

So address spaces are removing using RemoveJuliaAddrspaces pass, because those information cannot be hadled by the LLVM back-end.

Can you please let me know at which phase address spaces are removed?

maleadt · July 2, 2021, 1:18pm

github.com

JuliaLang/julia/blob/4270d3b44e53e9fa68dcdbb6cb416d232615db4c/src/disasm.cpp#L448


      
          LineNumberAnnotatedWriter AAW{debuginfo};
          if (!llvmf->getParent()) {
              // print the function declaration as-is
              llvmf->print(stream, &AAW);
              delete llvmf;
          }
          else {
              Module *m = llvmf->getParent();
              if (strip_ir_metadata) {
                  std::string llvmfn(llvmf->getName());
                  jl_strip_llvm_addrspaces(m);
                  jl_strip_llvm_debug(m, true, &AAW);
                  // rewriting the function type creates a new function, so look it up again
                  llvmf = m->getFunction(llvmfn);
              }
              if (dump_module) {
                  m->print(stream, &AAW);
              }
              else {
                  llvmf->print(stream, &AAW);
              }

We don’t for the main CPU back-ends because those handle Julia’s address spaces (i.e. they ignore it). GPU back-ends do need to put the pass in their pipeline.

KavinduZoysa · July 2, 2021, 1:40pm

Since address spaces are used to identify GC-tracked pointers within the optimizations phases, after the optimization phases, is address space information used for some other reason?

jpsamaroo · July 2, 2021, 1:43pm

Not the Julia addrspaces, but for GPU backends, non-zero addrspaces are used to identify the GPU memory region being accessed. For example, in the AMDGPU LLVM backend, 0 is used for Generic pointers, 1 is for Global pointers, 5 is used for Private (stack) pointers, etc.

KavinduZoysa · July 2, 2021, 4:18pm

@jpsamaroo, thank you for your input.

Let me summarize this. Please correct me if I am wrong.

Julia uses address spaces for two purposes,
1. Julia specific address spaces(AS) to identify GC-tracked pointers for optimization phases
2. GPU backend-specific AS to identify GPU memory regions
Julia specific AS are removed after the optimization phases.
GPU backend-specific AS are not removed like that.

jpsamaroo · July 2, 2021, 6:23pm

That sounds about right to me!

If you don’t mind me asking, why are you asking about these addrspaces? Are you doing any work on Julia’s optimizations passes?

KavinduZoysa · July 2, 2021, 7:22pm

Nope, I am reading on LLVM’s non-integral address space. Julia is the only thing I found that is used the non-integral address spaces.