[PATCH v5 00/23] ILP32 for ARM64

[Kapoor, Prasun]

On 10/2/15, 2:37 AM, "Catalin Marinas" <catalin.marinas at arm.com> wrote:

>On Thu, Oct 01, 2015 at 09:49:46PM +0000, Pinski, Andrew wrote:
>> Ok, we will rewrite these patches using 32bit time_t and 32bit off_t
>> and redo the toolchain support for them.  Note this is going back to
>> the abi I had originally done when I submitted my original version
>> when it was asked to change time_t to be 64bit.
>
>One of the key aspects of kernel development is the ability to adapt
>quickly to new requests/insights. This implies releasing early and often
>rather than a new version roughly twice a year (IIRC, v1 was posted
>September 2013). Moreover, the success of the kernel is partly based on
>not getting stuck on old decisions (well, unless it breaks accepted user
>ABI).
>
>So, at the time, following x32 discussions, we thought of using the
>native ABI as much as possible. However, two important things happened
>since:
>
>1. libc community didn't like breaking the POSIX compliance
>2. No-one seems desperate for ILP32 on AArch64
>
>(1) is a fair point and I would rather be careful as we don't know the
>extent of the code affected. In the meantime, we've also had ongoing
>work for addressing the 2038 issue on 32-bit architectures.
>
>The second point is equally important. The benchmarks I've seen didn't
>show a significant improvement and the messages I got on various
>channels pretty much labeled ILP32 as a transitional stage to full LP64.
>In this case, we need to balance the benefits of a close to native ABI
>(future proof, slightly higher performance) vs. the cost of maintaining
>such ABI in the kernel on the long term, especially if it's not widely
>used/tested.


For us ILP32  is not about putting this into our product flier at all, it
is about supporting real applications. We have an existing product line of
MIPS based SoCs where a large number of N32 (an exact equivalent of ILP32)
applications are currently in production. Our customers are looking to
bring those applications (mostly in Networking and Telecom space) over to
ARMv8. 

We think its an extremely risky strategy to say either future processors
should incur the additional cost (power and complexity) of implementing
Aarch32 instruction set or have no way of  supporting 32 bit applications
at all.

Apart from there being an installed base of 32 bit networking and telecom
applications, we have also seen non-trivial performance gains with ILP32
(for example, our SPECINT score goes up by 7% with ILP32 compared to
LP64).  



>
>We've seen the kernel patches and, following discussions on the lists,
>decided to change the original recommendation. IIRC, the main ideas (but
>you need to read various threads as I can't remember the details from
>6-7 months ago):
>
>a) separate syscall table for ILP32
>b) close to compat ABI with 32-bit time_t, off_t
>c) asm-generic/unistd.h rather than asm/unistd32.h (that's where it
>   would differ from compat, together with places where pointers are
>   passed)
>
>As I said previously, I'm not going to pay any attention to the patches
>in this series, it's nothing more than a rebase of a version I already
>reviewed.
>
>-- 
>Catalin