[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] Questions/comments on TCG
From: |
Blue Swirl |
Subject: |
Re: [Qemu-devel] Questions/comments on TCG |
Date: |
Fri, 7 Mar 2008 20:47:03 +0200 |
On 3/7/08, Stuart Brady <address@hidden> wrote:
> On Fri, Mar 07, 2008 at 06:07:32PM +0200, Blue Swirl wrote:
> > On 3/7/08, Stuart Brady <address@hidden> wrote:
>
> > > I do understand that the current SPARC TCG code is preliminary work.
>
> > > However, in some ways, I feel it still serves as a better reference than
> > > i386 and x86_64
> >
> > Well, I'd still recommend using x86 as a reference until Sparc works
> > or you may copy a faulty design.
>
>
> Don't worry -- I still checked with the x86 targets. I only really
> needed a quick idea of what was required for a new target.
>
>
> > > Which registers should go in tcg_target_reg_alloc_order[]? I notice
> > > that i386 includes ESP, which tcg_target_init() marks as reserved, and
> > > x86_64 includes RBX and RBP, which are again marked as reserved.
> >
> > I put there only the registers that should be safe to use, the G
> > registers may have issues or they are already used as global
> > registers. Also we should not need frame pointer.
>
>
> Sounds reasonable. I think I really meant to ask what _shouldn't_ go in
> tcg_target_reg_alloc_order[]. I was mainly confused by the inclusion of
> registers which are marked as reserved on the x86 targets.
>
>
> > > Furthermore, x86_64's tcg_target_reg_alloc_order[] contains 16 elements
> > > (TCG_TARGET_NB_REGS), but only 15 are specified -- the last element is
> > > left as 0, which is TCG_REG_RAX. SPARC also does this, but with
> > > TARGET_REG_G0 (which is marked as reserved, as it's hardwired to zero).
> >
> > Maybe I missed something, but g0 isn't in the reg_alloc_order?
>
>
> tcg_target_reg_alloc_order[] has 32 elements, but only 14 are used.
> The rest hold 0, specifying TCG_REG_G0.
I see. That could be asking for trouble.
> > > On SPARC, I notice that goto_tb is handled using CALL and JMPL, placing
> > > the return address in o7... but we're returning from a TB, or jumping to
> > > another one, so surely we shouldn't link here? Also, TCG_TYPE_TL is
> > > used for exit_tb's return value, I think this should be the host's long
> > > (using TCG_TYPE_PTR) instead.
> >
> > These are bugs, thanks for spotting. I was using o7 if a register is
> > needed, it will be clobbered anyway.
>
>
> I don't understand -- o7 is required when returning in exit_tb, so if it
> is used, it must be saved and restored.
Not exit_tb, but call.
> > > Also on SPARC, could the indentation of the OP_32_64s be improved?
> > > Yeah, it's not a serious problem, but I feel it would make the code
> > > slightly easier to read.
> >
> > It's not my fault, Emacs wants to do it this way. I'm open to your
> > suggestions.
>
>
> Oh dear, I'm such a vim user, I don't even have Emacs installed. :)
>
> How about something like this?
>
> #if defined(__sparc_v9__) && !defined(__sparc_v8plus__)
> #define OP_32_64(x) \
> glue(glue(case INDEX_op_, x), _i32:) \
> glue(glue(case INDEX_op_, x), _i64)
> #else
> #define OP_32_64(x) \
> glue(glue(case INDEX_op_, x), _i32)
> #endif
> ...
> OP_32_64(ld8u):
> tcg_out_ldst(s, args[0], args[1], args[2], LDUB);
> break;
> ...
>
> The macro might be a bit sick, but hopefully it would make Emacs happy,
> and I feel ':' does make a certain amount of sense, here.
Unfortunately it confuses Emacs even more.
> It probably wouldn't help with indentation, but you could always do
> something like this:
>
> #if defined(__sparc_v9__) && !defined(__sparc_v8plus__)
> #define v9(x) x
> #else
> #define v9(x)
> #endif
> ...
> case INDEX_op_ld8u_i32:
> v9( case INDEX_op_ld8u_i64: )
> tcg_out_ldst(s, args[0], args[1], args[2], LDUB);
> break;
> ...
>
> I'll admit, that looks unusual, but it would avoid breaking searches for
> ld8u_i32 or op_ld8u.
This also does not work.
[Qemu-devel] [PATCH] Remove blank elements in tcg_target_reg_alloc_order[], Stuart Brady, 2008/03/07