From: Alexei Starovoitov <alexei.starovoi...@gmail.com>
Date: Tue, 18 Apr 2017 15:57:09 -0700
> On Tue, Apr 18, 2017 at 02:37:32PM -0400, David Miller wrote:
>> From: Alexei Starovoitov <alexei.starovoi...@gmail.com>
>> Date: Mon, 17 Apr 2017 22:44:47 -0700
>>
>> > The way llvm generates stack access is:
>> > rX = r10
>> > rX += imm
>> > and that's the only thing verifier recognizes as valid ptr_to_stack.
>> > Like rX -= imm will not be recognized as proper stack offset,
>> > since llvm never does it.
>>
>> That simplifies things significantly for me.
>>
>> I only allow moves from the frame pointer to another register,
>> and when I see that I rewrite it to "add FP, STACK_BIAS, DST_REG"
>
> Sounds good to me. Alternative idea: can the above
> 'add FP, STACK_BIAS, one_of_local_regs' be done once in prologue
> and that register used as substitue for R10 ?
> (assuming non-leaf function)
> I completely forgot by now how 2047 magic works.
Well, I'm trying to understand what that would buy us.
If I'm always going to see that kind of sequence:
rX = r10
rx += offset
then I always have that "r10" move available to deal with the bias.
It comes for free.
Right now the current version of the sparc64 JIT will emit:
add %fp, 2047, bpf2sparc[rX]
add bpf2sparc[rX], offset, bpf2sparc[rX]
for that BPF sequence.
Anyways, thanks to your and Daniel's feedback I'm very confident
with the code now. All that's really left is tail calls, and I'll
set up a BPF compilation and test environment for that.
