[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH v2 6/7] tcg: Introduce zero and sign-extended ve
From: |
Richard Henderson |
Subject: |
Re: [Qemu-devel] [PATCH v2 6/7] tcg: Introduce zero and sign-extended versions of load helpers |
Date: |
Fri, 30 Aug 2013 13:53:37 -0700 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:17.0) Gecko/20130805 Thunderbird/17.0.8 |
On 08/30/2013 12:12 PM, Aurelien Jarno wrote:
> On Fri, Aug 30, 2013 at 10:20:23AM -0700, Richard Henderson wrote:
>> On 08/30/2013 09:55 AM, Aurelien Jarno wrote:
>>> While it works for x86 and some other architectures, it makes the
>>> assumption that only part of the register can be used later by the TCG
>>> code. It won't be the case if we later (and I hope we will) implement a
>>> MIPS64 TCG target. In that case, a 32-bit value has to be returned
>>> signed extended, which won't be the case for example for a 32-bit guest
>>> loading a 16-bit unsigned value.
>>
>> This doesn't break the mips64 abi, since we'll be returning a 64-bit value,
>> not
>> a 32-bit value that needs sign-extension.
>>
>> Given a mips64 host with 32-bit guest, the sign-extension of the 32-bit load
>> can either happen by using helper_ret_ldsl_mmu in the table of helper
>> functions, or by using an sll insn instead of a move to put the value into
>> place at the end of the slow path.
>
> That's indeed a possibility. That said while the MIPS64 ABI is then
> still followed, it would have break a MIPS backend as the ABI between
> the helper and the TCG code is broken.
How's that? We're passing a value extended to tcg_target_ulong.
For the 32-bit mips backend running o32 (or even a theoretical n32 backend),
that type is uint32_t. That gets returned from C exactly how C returns that
type. For o32 it's the full width of the register, full stop. For n32, it
would be returned sign-extended in the 64-bit register.
Please explain exactly the failure mode you imagine, because I don't think
there is one.
> I am therefore concerned that we might break some of our 64-bit
> backends. x86-64 and ia64 should be fine, I don't know about aarch64,
> ppc64, sparc64 or s390x.
Nope, all 4 of those will be fine. Not least of which because the later 3
are still using the original helper functions, not the new helper_ret_* ones.
But certainly all 4 of those are, in the gcc sense, TRULY_NOOP_TRUNCATION
machines, meaning we can truncate to 32 bits merely by ignoring the garbage
in the high bits. In practice it means that they have different 32-bit and
64-bit comparison instructions.
r~
- [Qemu-devel] [PATCH v2 3/7] exec: Rename USUFFIX to LSUFFIX, (continued)
[Qemu-devel] [PATCH v2 7/7] tcg-i386: Make use of zero-extended memory helper routines, Richard Henderson, 2013/08/29