[LLVMdev] Greedy Register Allocation in LLVM 3.0

[Jakob Stoklund Olesen]

On Dec 9, 2011, at 12:10 AM, 陳韋任 <chenwj at iis.sinica.edu.tw> wrote:
>  After reading your blog article, I have some questions. :-) In [1], it says: 
> 
> "It was an important design goal to make the algorithm as flexible as possible,
> and to avoid introducing arbitrary constraints. It is possible to change machine
> code and live ranges at any time. Simply evict the relevant live ranges, make
> the change, and put them back on the queue."
> 
> Q1.
>  The "machine code" mentioned above means LLVM MI (MachineInstr) or? 

Yes.

> Q2.
>  In [2], Evan illustrates current LLVM MI flow below.
> 
> 1. DAG to MI lowering (and pre-RA schedule)
> 2. MI optimizations (LICM, CSE, etc.)
> 3. Register allocation super pass
>   3a. De-ssa (2-address, phi slim)
>   3b. Coalescing
>   3c. Actual register allocation
> 4. Post-RA optimizations
> 5. PEI
> 6. Post-RA scheduling
> 
>  Does "change machine code and live ranges at any time" means we want to change
> the machine code at some points in the above flow? If so, could you point it out?
> If you can give me a small example, that would be great!

The blog post is only about "3c. Actual register allocation". The other passes naturally change the machine code.

In the classical formulations of register allocation like linear scan and graph coloring, it is assumed that the machine instructions stay constant while only the register assignments are changed. The register allocator may insert spill code and copies, but the original instructions are never touched.

I want to allow the register allocator to change the original instructions as well. For example, the scheduler likes to hoist loads:

  xmm3 = load(gpr2)
  xmm2 = addps xmm2, xmm1

The code requires three xmm registers, but if you sink the load, two is enough:

  xmm2 = addps xmm2, xmm1
  xmm1 = load(gpr2)

> Q3.
>  In the section "Lessons learned from linear scan", you say "advanced register
> allocation algorithms often need to build expensive data structures, or they make
> assumptions about live ranges being invariant".
> 
>  Does that imply the graph coloring algorithm? If an algorithm A requires the
> live ranges being invariant, does it mean when we assign a register to a variable
> then it's done and cannot be changed latter on?

No, graph coloring algorithms can change previous assignments.

The invariant live ranges problem is the same as above. If you move instructions around, the live ranges change, and the interference graph is invalidated.

/jakob