2015-12-19 04:18:05

by Andrew Pinski

[permalink] [raw]
Subject: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

The problem here is that GCC 6 and above emits .arch now
for each function so now the global .arch_extension has
no effect. This fixes the problem by putting
.arch_extension inside ARM64_LSE_ATOMIC_INSN so
it is enabled for each place where LSE is used.

Signed-off-by: Andrew Pinski <[email protected]>
---
arch/arm64/include/asm/lse.h | 4 +---
1 files changed, 1 insertions(+), 3 deletions(-)

diff --git a/arch/arm64/include/asm/lse.h b/arch/arm64/include/asm/lse.h
index 3de42d6..625601f 100644
--- a/arch/arm64/include/asm/lse.h
+++ b/arch/arm64/include/asm/lse.h
@@ -17,8 +17,6 @@

#else /* __ASSEMBLER__ */

-__asm__(".arch_extension lse");
-
/* Move the ll/sc atomics out-of-line */
#define __LL_SC_INLINE
#define __LL_SC_PREFIX(x) __ll_sc_##x
@@ -29,7 +27,7 @@ __asm__(".arch_extension lse");

/* In-line patching at runtime */
#define ARM64_LSE_ATOMIC_INSN(llsc, lse) \
- ALTERNATIVE(llsc, lse, ARM64_HAS_LSE_ATOMICS)
+ ALTERNATIVE(".arch_extension lse\n" llsc, lse, ARM64_HAS_LSE_ATOMICS)

#endif /* __ASSEMBLER__ */
#else /* CONFIG_AS_LSE && CONFIG_ARM64_LSE_ATOMICS */
--
1.7.2.5


2015-12-21 12:38:45

by Will Deacon

[permalink] [raw]
Subject: Re: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
> The problem here is that GCC 6 and above emits .arch now
> for each function so now the global .arch_extension has
> no effect. This fixes the problem by putting
> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
> it is enabled for each place where LSE is used.

Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
.arch_extension is used for virt, mp and sec over there, and it may be
tricky to isolate the actual instruction usage (at least, virt looks
lost in kvm/arm.c).

Why can't gas have an option to accept all instruction encodings that it
knows about, inspite of any .arch directives?

Will

> Signed-off-by: Andrew Pinski <[email protected]>
> ---
> arch/arm64/include/asm/lse.h | 4 +---
> 1 files changed, 1 insertions(+), 3 deletions(-)
>
> diff --git a/arch/arm64/include/asm/lse.h b/arch/arm64/include/asm/lse.h
> index 3de42d6..625601f 100644
> --- a/arch/arm64/include/asm/lse.h
> +++ b/arch/arm64/include/asm/lse.h
> @@ -17,8 +17,6 @@
>
> #else /* __ASSEMBLER__ */
>
> -__asm__(".arch_extension lse");
> -
> /* Move the ll/sc atomics out-of-line */
> #define __LL_SC_INLINE
> #define __LL_SC_PREFIX(x) __ll_sc_##x
> @@ -29,7 +27,7 @@ __asm__(".arch_extension lse");
>
> /* In-line patching at runtime */
> #define ARM64_LSE_ATOMIC_INSN(llsc, lse) \
> - ALTERNATIVE(llsc, lse, ARM64_HAS_LSE_ATOMICS)
> + ALTERNATIVE(".arch_extension lse\n" llsc, lse, ARM64_HAS_LSE_ATOMICS)
>
> #endif /* __ASSEMBLER__ */
> #else /* CONFIG_AS_LSE && CONFIG_ARM64_LSE_ATOMICS */
> --
> 1.7.2.5
>
>
> _______________________________________________
> linux-arm-kernel mailing list
> [email protected]
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>

2015-12-21 12:46:26

by Ard Biesheuvel

[permalink] [raw]
Subject: Re: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

On 21 December 2015 at 13:38, Will Deacon <[email protected]> wrote:
> On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
>> The problem here is that GCC 6 and above emits .arch now
>> for each function so now the global .arch_extension has
>> no effect. This fixes the problem by putting
>> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
>> it is enabled for each place where LSE is used.
>
> Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
> .arch_extension is used for virt, mp and sec over there, and it may be
> tricky to isolate the actual instruction usage (at least, virt looks
> lost in kvm/arm.c).
>
> Why can't gas have an option to accept all instruction encodings that it
> knows about, inspite of any .arch directives?
>

Modern GAS supports things like -march=armv7-a+mp+sec+virt, so it
probably makes sense to pass that on the command line when building
for v7 (or +sec only for v6) if the assembler is found to support it
at build time.


>> Signed-off-by: Andrew Pinski <[email protected]>
>> ---
>> arch/arm64/include/asm/lse.h | 4 +---
>> 1 files changed, 1 insertions(+), 3 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/lse.h b/arch/arm64/include/asm/lse.h
>> index 3de42d6..625601f 100644
>> --- a/arch/arm64/include/asm/lse.h
>> +++ b/arch/arm64/include/asm/lse.h
>> @@ -17,8 +17,6 @@
>>
>> #else /* __ASSEMBLER__ */
>>
>> -__asm__(".arch_extension lse");
>> -
>> /* Move the ll/sc atomics out-of-line */
>> #define __LL_SC_INLINE
>> #define __LL_SC_PREFIX(x) __ll_sc_##x
>> @@ -29,7 +27,7 @@ __asm__(".arch_extension lse");
>>
>> /* In-line patching at runtime */
>> #define ARM64_LSE_ATOMIC_INSN(llsc, lse) \
>> - ALTERNATIVE(llsc, lse, ARM64_HAS_LSE_ATOMICS)
>> + ALTERNATIVE(".arch_extension lse\n" llsc, lse, ARM64_HAS_LSE_ATOMICS)
>>
>> #endif /* __ASSEMBLER__ */
>> #else /* CONFIG_AS_LSE && CONFIG_ARM64_LSE_ATOMICS */
>> --
>> 1.7.2.5
>>
>>
>> _______________________________________________
>> linux-arm-kernel mailing list
>> [email protected]
>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>>
>
> _______________________________________________
> linux-arm-kernel mailing list
> [email protected]
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

2015-12-21 12:51:50

by Will Deacon

[permalink] [raw]
Subject: Re: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

On Mon, Dec 21, 2015 at 01:46:22PM +0100, Ard Biesheuvel wrote:
> On 21 December 2015 at 13:38, Will Deacon <[email protected]> wrote:
> > On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
> >> The problem here is that GCC 6 and above emits .arch now
> >> for each function so now the global .arch_extension has
> >> no effect. This fixes the problem by putting
> >> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
> >> it is enabled for each place where LSE is used.
> >
> > Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
> > .arch_extension is used for virt, mp and sec over there, and it may be
> > tricky to isolate the actual instruction usage (at least, virt looks
> > lost in kvm/arm.c).
> >
> > Why can't gas have an option to accept all instruction encodings that it
> > knows about, inspite of any .arch directives?
> >
>
> Modern GAS supports things like -march=armv7-a+mp+sec+virt, so it
> probably makes sense to pass that on the command line when building
> for v7 (or +sec only for v6) if the assembler is found to support it
> at build time.

Does that override a more restrictive .arch directive emitted by the
compiler?

Will

2015-12-21 12:58:33

by Ard Biesheuvel

[permalink] [raw]
Subject: Re: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

On 21 December 2015 at 13:51, Will Deacon <[email protected]> wrote:
> On Mon, Dec 21, 2015 at 01:46:22PM +0100, Ard Biesheuvel wrote:
>> On 21 December 2015 at 13:38, Will Deacon <[email protected]> wrote:
>> > On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
>> >> The problem here is that GCC 6 and above emits .arch now
>> >> for each function so now the global .arch_extension has
>> >> no effect. This fixes the problem by putting
>> >> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
>> >> it is enabled for each place where LSE is used.
>> >
>> > Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
>> > .arch_extension is used for virt, mp and sec over there, and it may be
>> > tricky to isolate the actual instruction usage (at least, virt looks
>> > lost in kvm/arm.c).
>> >
>> > Why can't gas have an option to accept all instruction encodings that it
>> > knows about, inspite of any .arch directives?
>> >
>>
>> Modern GAS supports things like -march=armv7-a+mp+sec+virt, so it
>> probably makes sense to pass that on the command line when building
>> for v7 (or +sec only for v6) if the assembler is found to support it
>> at build time.
>
> Does that override a more restrictive .arch directive emitted by the
> compiler?
>

It seems to be additive: -march=armv7-a+mp+sec allows a .S file
containing a virt arch_extension + both hvc and smc instructions to be
assembled.

2015-12-21 13:03:50

by Russell King - ARM Linux

[permalink] [raw]
Subject: Re: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

On Mon, Dec 21, 2015 at 01:46:22PM +0100, Ard Biesheuvel wrote:
> On 21 December 2015 at 13:38, Will Deacon <[email protected]> wrote:
> > On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
> >> The problem here is that GCC 6 and above emits .arch now
> >> for each function so now the global .arch_extension has
> >> no effect. This fixes the problem by putting
> >> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
> >> it is enabled for each place where LSE is used.
> >
> > Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
> > .arch_extension is used for virt, mp and sec over there, and it may be
> > tricky to isolate the actual instruction usage (at least, virt looks
> > lost in kvm/arm.c).
> >
> > Why can't gas have an option to accept all instruction encodings that it
> > knows about, inspite of any .arch directives?
> >
>
> Modern GAS supports things like -march=armv7-a+mp+sec+virt, so it
> probably makes sense to pass that on the command line when building
> for v7 (or +sec only for v6) if the assembler is found to support it
> at build time.

Modern GCCs (GCC 4 at least) add .arch / .cpu pseudo-instructions in
the assembly output which means using -Wa,-march= and -Wa,-mcpu= to
GCC are now totally useless.

(From the bug reports I've seen, I don't deem GCC 5 to be anything but
a total failure, and so as far as I'm concerned, GCC 5 is not to be
used for ARM, and effectively doesn't exist. I have no experience or
opinions on GCC 6. Hence, "GCC 4" still counts as "modern" being the
only recent compilers that produce sane results.)

--
RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

2015-12-21 13:37:32

by Will Deacon

[permalink] [raw]
Subject: Re: [PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

On Mon, Dec 21, 2015 at 01:58:30PM +0100, Ard Biesheuvel wrote:
> On 21 December 2015 at 13:51, Will Deacon <[email protected]> wrote:
> > On Mon, Dec 21, 2015 at 01:46:22PM +0100, Ard Biesheuvel wrote:
> >> On 21 December 2015 at 13:38, Will Deacon <[email protected]> wrote:
> >> > On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
> >> >> The problem here is that GCC 6 and above emits .arch now
> >> >> for each function so now the global .arch_extension has
> >> >> no effect. This fixes the problem by putting
> >> >> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
> >> >> it is enabled for each place where LSE is used.
> >> >
> >> > Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
> >> > .arch_extension is used for virt, mp and sec over there, and it may be
> >> > tricky to isolate the actual instruction usage (at least, virt looks
> >> > lost in kvm/arm.c).
> >> >
> >> > Why can't gas have an option to accept all instruction encodings that it
> >> > knows about, inspite of any .arch directives?
> >> >
> >>
> >> Modern GAS supports things like -march=armv7-a+mp+sec+virt, so it
> >> probably makes sense to pass that on the command line when building
> >> for v7 (or +sec only for v6) if the assembler is found to support it
> >> at build time.
> >
> > Does that override a more restrictive .arch directive emitted by the
> > compiler?
> >
>
> It seems to be additive: -march=armv7-a+mp+sec allows a .S file
> containing a virt arch_extension + both hvc and smc instructions to be
> assembled.

The problem I'm seeing is if I have something like:

.arch_extension lse

before something like:

.cpu cortex-a57+fp+simd+crc

-or-

.arch armv8-a+fp+simd+crc

then I can no longer assemble lse instructions. So the .cpu/.arch
directive is undoing the .arch_extension. We can fix this by following
Andrew's suggestion to have .arch_extension before every point of use,
but the whole thing would be much simpler if we could just tell gas to
assemble harder.

Maybe we just need to construct the mother of all -march options based
on build-time checks in the Makefile?

Will