Re: [ptxdist] Toolchain for STM32MP15x

[Guillermo Rodriguez Garcia <guille.rodriguez@gmail.com>]

Hi Michael,

El mar., 26 nov. 2019 a las 8:24, Michael Olbrich
(<m.olbrich@pengutronix.de>) escribió:
>
> On Mon, Nov 25, 2019 at 05:43:01PM +0100, Guillermo Rodriguez Garcia wrote:
> > Hi all,
> >
> > I have setup the toolchain with the settings from my previous email
> > (-mcpu=cortex-a7 -mfpu=neon-vfpv4)
> >
> > When I try to build the Linux kernel I get lots of warnings about
> > conflicting toolchain options:
> >
> > arch/arm/kernel/elf.c:1:0: warning: switch -mcpu=cortex-a7 conflicts
> > with -march=armv5t switch
>
> The problem is, that the kernel tries to add it's own -march=armv7-a and I
> guess that causes warnings in combination with -mcpu=cortex-a7 (at least
> with some gcc versions). This is tested with -Werror, so it fails. The

You are right, this causes a warning:

$ /opt/OSELAS.Toolchain-2016.06.1/arm-v7a-linux-gnueabihf/gcc-5.4.0-glibc-2.23-binutils-2.26-kernel-4.6-sanitized/bin/arm-v7a-linux-gnueabihf-gcc
-march=armv7-a -mcpu=cortex-a7 -o hello hello.c
hello.c:1:0: warning: switch -mcpu=cortex-a7 conflicts with
-march=armv7-a switch

I didn't understand why the conflict happens with -mcpu=cortex-a7 but
not with -mcpu=cortex-a9 (as both of which are ARMv7-A). But the I
found that Cortex A7 actually implements the ARMv7ve variant of ARMv7.
Indeed the following does not result in warnings:

$ /opt/OSELAS.Toolchain-2016.06.1/arm-v7a-linux-gnueabihf/gcc-5.4.0-glibc-2.23-binutils-2.26-kernel-4.6-sanitized/bin/arm-v7a-linux-gnueabihf-gcc
-march=armv7ve -mcpu=cortex-a7 -o hello hello.c

> fallback is -march=armv5t ... (see arch/arm/Makefile in the kernel).
>
> You cannot use -mcpu= in the extra cflags like this.
>
> > This is with OSELAS.Toolchain-2016.06.1, using arm-v7a-linux-gnueabihf.
> >
> > I am not sure where the -march=armv5t is coming from, since this is an
> > arm-v7a toolchain.
> >
> > Also if I use -mcpu=cortex-a9 instead of cortex-a7, there are no warnings :-?
>
> This will have the same problem with gcc 9.x.
>
> > Can anyone sehd some light ?
>
> See above. Combining -march= with -mcpu= is not really supported and the
> result is undefined.

According to the gcc docs this should be supported:

-march=name[+extension…]
    [...] This option can be used in conjunction with or instead of
the -mcpu= option

-mcpu=name[+extension…]
    [...] Where this option is used in conjunction with -march or
-mtune, those options take precedence over the appropriate part of
this option.

The problem seems to be that although this is supported, there is a
warning because the explicit -march (armv7-a) does not match the
architecture derived from -mcpu (armv7ve). And as you said,
specifically for the kernel build, this is tested with -Werror, so the
test fails and the armv5t fallback is used instead (which also fails).


> I'd say, just use -mfpu=neon-vfpv4. Have you tested if
> -mcpu=cortex-a7 actually makes a difference?

No, didn't get that far (the build failed..)

But I guess I can replace -mcpu=cortex-a7 with -mtune=cortex-a7 (that
does not result in any errors)

Thank you! This was helpful.

BR,

Guillermo Rodriguez Garcia
guille.rodriguez@gmail.com

_______________________________________________
ptxdist mailing list
ptxdist@pengutronix.de