On Thu, Aug 15, 2024 at 2:47 PM Eduard Zingerman <eddyz87@xxxxxxxxx> wrote:
>
> On Thu, 2024-08-15 at 14:32 -0700, Yonghong Song wrote:
>
> [...]
>
> > > +__success
> > > +/* program entry for main(), regular function prologue */
> > > +__jit_x86(" endbr64")
> > > +__jit_x86(" nopl (%rax,%rax)")
> > > +__jit_x86(" xorq %rax, %rax")
> > > +__jit_x86(" pushq %rbp")
> > > +__jit_x86(" movq %rsp, %rbp")
> >
> > How do we hanble multi architectures (x86, arm64, riscv64)?
> >
> > Do we support the following?
> >
> > __jit_x86(...)
> > __jit_x86(...)
> > ...
> >
> > __jit_arm64(...)
> > __jit_arm64(...)
> > ...
> >
> > __jit_riscv64(...)
> > __jit_riscv64(...)
> > ...
>
> ^^^^
> I was thinking about this variant (and this is how things are now implemented).
> Whenever there would be a need for that, just add one more arch
> specific macro.
>
> >
> > Or we can use macro like
> >
> > #ifdef __TARGET_ARCH_x86
> > __jit(...)
> > ...
> > #elif defined(__TARGET_ARCH_arm64)
> > __jit(...)
> > ...
> > #elif defined(...)
> >
> > Or we can have
> >
> > __arch_x86_64
> > __jit(...) // code for x86
> > ...
> >
> > __arch_arm64
> > __jit(...) // code for arm64
> > ...
> >
> > __arch_riscv
> > __jit(...) // code for riscv
> > ...
>
> This also looks good, and will work better with "*_next" and "*_not"
> variants if we are going to borrow from llvm-lit/FileCheck.
>
shorter __jit() and then arch-specific __arch_blah seems pretty clean,
so if it's not too hard, let's do this.
BTW, in your implementation you are collecting expected messages for
all specified architectures, but really there will always be just one
valid subset. So maybe just discard all non-host architectures upfront
during "parsing" of decl tags?
> > For xlated, different archs could share the same code.
> > Bot for jited code, different arch has different encoding,
> > so we need to figure out a format suitable for multiple
> > archs.
>
> I'll go with whatever way mailing list likes better.
>
> [...]
>
