Re: Defining a Super Specific uArch

On 13/01/21 11:45 am, squirvel voxbox via crossgcc wrote:
> Hi All,
>
> I asked a fairly similar question over on the irc channel, and didn't get a
> response, so I'm trying here.
https://github.com/crosstool-ng/crosstool-ng is probably the best 
support channel these days. I keep an eye on IRC but it's sporadic at best.
> Let's say I'm attempting to target a super specific CPU, (in this case a
> RK3328, which is a 64bit armv8-a+crypto cpu). What would be the proper way
> to define the cpu?
My initial thought is to make a generic aarch64 toolchain and use 
-march= in CFLAGS to tune for your desired system (the 
aarch64-unknown-linux-gnu sample is probably good enough for this).
> I am currently defining:
>      target architecture: arm
>      suffix to the arch-part: v8-a+crypto (according to gcc, the dash is
> required to define v8-a)
The suffix is just used in naming the tools so it could be anything 
(often people add things "le", "be" or "hf"). Adding a '-' breaks things 
that use the toolchain tuple to figure stuff out. It'd be nice if we 
could stop users from putting a '-' in there but I'm not sure that 
Kconfig has a way of validating string input fields.
>      emit assembly for CPU: cortex-a53
>      other items are set based on the sample aarch64-rpi3 configuration
>
> But ct-ng build fails due to:
>      "Invalid configuration `aarch64v8-a+crypto-RK3328-linux-gnu': more than
> four components" (so more than 4 dashes according to the script source code)
> or if I disable the vendor string (set to RK3328)
>      "Invalid configuration `aarch64v8-a+crypto-linux-gnu': machine
> `aarch64v8-a+crypto' not recognized"
>
> In theory I could just add all the specific uarch flags to CFLAGS/LDFLAGS
> such as -march=, but that feels rather hacky especially since there is an
> arch suffix field.
I think what you want is to set CT_ARCH_ARCH ("Architecture level:") to 
armv8-a+crypto which should have the effect of making this the default 
when no -march= is specified in CFLAGS. The fact that you're already 
setting the vendor string to RK3328 is probably good enough to 
distinguish the resulting toolchain from any others you might have.
> Is the hacky way the way I should be defining this cpu, or am I just
> specifying this incorrectly? If the latter, what is the proper way to
> specify a super specific uarch?
>
> Thanks,
> squirvel
> --
> For unsubscribe information see http://sourceware.org/lists.html#faq
--
For unsubscribe information see http://sourceware.org/lists.html#faq