[LLVMdev] Spills and values present in both registers & stack

David Blaikie dblaikie at gmail.com
Thu Jul 28 16:05:49 PDT 2011

On Tue, Jul 26, 2011 at 11:35 AM, Taral <taralx at gmail.com> wrote:
> One piece of code I'm writing has a lot of intermediates, and I'm
> trying to optimize down the number of memory accesses. Here's a
> snippet from the start of the function, where I think there is some
> low-hanging fruit:
> # BB#0:
>        pushq   %rbp
>        pushq   %r15
>        pushq   %r14
>        pushq   %r13
>        pushq   %r12
>        pushq   %rbx
>        movq    %rdx, %rcx
>        movq    %rdi, -16(%rsp)         # 8-byte Spill
>        movq    (%rsi), %rdi
>        movq    8(%rsi), %r8
>        movq    8(%rcx), %rax
>        movq    %rax, -24(%rsp)         # 8-byte Spill
>        movq    16(%rcx), %rax
>        movq    %rax, -8(%rsp)          # 8-byte Spill
>        movq    %rdi, %rax
>        mulq    -24(%rsp)               # 8-byte Folded Reload
> You'll note that rbx,r12,r13,r14,r15,rbp are all dead after the
> pushes. But the spill code still insists on using rax to load the
> spilled values, forcing them to be reloaded later.

I'm not the most familiar with this sort of thing - but a small
example (of llvm bitcode) & the optimization flags you used, etc,
might be helpful (& I might be able to have a go at explaining it, if
no one else does).

> Is the register
> allocator (pbqp, I think) capable of having values in registers and on
> the stack at the same time?

I can at least confirm that the default register allocator on x86
isn't PBQP, it's the greedy allocator (unless your'e compiling with
-O0, in which case it's the fast allocator).

- David

More information about the llvm-dev mailing list