Skip to contentSkip to conversation

ARCv2: Add explcit unaligned access support (and ability to disable too)

PATCH Guenter Roeck 2 participants 5 messages Feb 28, 2019 – Mar 4, 2019 Synopsys ARC architecture

lore
Share & tools
lore
#0 lore raw
Quoted 11 lines · Eugeniy Paltsev

As of today we enable unaligned access unconditionally on ARCv2. Do this under a Kconfig option to allow disable it for test, benchmarking etc. Also while at it

- Select HAVE_EFFICIENT_UNALIGNED_ACCESS
- Although gcc defaults to unaligned access (since GNU 2018.03), add the right toggles for enabling or disabling as appropriate
- update bootlog to prints both HW feature status (exists, enabled/disabled) and SW status (used / not used).
- wire up the relaxed memcpy for unaligned access

Quoted 3 lines
Signed-off-by Eugeniy Paltsev <Eugeniy.Paltsev at synopsys.com>
Signed-off-by Vineet Gupta <vgupta at synopsys.com>
[vgupta: squashed patches, handle gcc -mno-unaligned-access quick]

Hi,

with my brand-new gcc-8.3.0, built from upstream gcc, this patch results in

arch/arc/kernel/setup.c: In function 'arc_cpu_mumbojumbo.constprop':
include/linux/compiler.h:348:38: error:
call to '__compiletime_assert_326' declared with attribute error: gcc doesn't support -mno-unaligned-access

when trying to build allnoconfig or tinyconfig. I get the same results with gcc-8.2.0.

Do I need some special compiler configuration flag when building the arcv2 toolchain ? Or am I simply out of luck for using upstream gcc ?

Thanks,
Guenter

#1 lore raw
Quoted 15 lines · Guenter Roeck
Quoted 11 lines · Eugeniy Paltsev

As of today we enable unaligned access unconditionally on ARCv2. Do this under a Kconfig option to allow disable it for test, benchmarking etc. Also while at it

- Select HAVE_EFFICIENT_UNALIGNED_ACCESS
- Although gcc defaults to unaligned access (since GNU 2018.03), add the right toggles for enabling or disabling as appropriate
- update bootlog to prints both HW feature status (exists, enabled/disabled) and SW status (used / not used).
- wire up the relaxed memcpy for unaligned access

Quoted 3 lines
Signed-off-by Eugeniy Paltsev <Eugeniy.Paltsev at synopsys.com>
Signed-off-by Vineet Gupta <vgupta at synopsys.com>
[vgupta: squashed patches, handle gcc -mno-unaligned-access quick]
Quoted 8 lines

Hi,

with my brand-new gcc-8.3.0, built from upstream gcc, this patch results in

arch/arc/kernel/setup.c: In function 'arc_cpu_mumbojumbo.constprop':
include/linux/compiler.h:348:38: error:
call to '__compiletime_assert_326' declared with attribute error: gcc doesn't support -mno-unaligned-access

Sorry it broke ur build - but it did catch a deficiency in compiler as intended - yay !

Quoted 2 lines

when trying to build allnoconfig or tinyconfig. I get the same results with gcc-8.2.0.

Unaligned access code gen was enabled by default in 7.x actually, but us OS folks were late to find that which led to the recent patch to disable it which was broken in gcc.
Claudiu fixed gcc promptly but that patch didn't make it to upstream, which I've posted now.

http://lists.infradead.org/pipermail/linux-snps-arc/2019-February/005522.html

Quoted 2 lines

Do I need some special compiler configuration flag when building the arcv2 toolchain ? Or am I simply out of luck for using upstream gcc ?

Once the patch hits mainline you could use that. In the mean time I'll back out the "detector" as it might trip other people too.

Many Thx for your testing service.
-Vineet

#2 lore raw
Quoted 37 lines · Vineet Gupta
Quoted 15 lines · Guenter Roeck
Quoted 11 lines · Eugeniy Paltsev

As of today we enable unaligned access unconditionally on ARCv2. Do this under a Kconfig option to allow disable it for test, benchmarking etc. Also while at it

- Select HAVE_EFFICIENT_UNALIGNED_ACCESS
- Although gcc defaults to unaligned access (since GNU 2018.03), add the right toggles for enabling or disabling as appropriate
- update bootlog to prints both HW feature status (exists, enabled/disabled) and SW status (used / not used).
- wire up the relaxed memcpy for unaligned access

Quoted 3 lines
Signed-off-by Eugeniy Paltsev <Eugeniy.Paltsev at synopsys.com>
Signed-off-by Vineet Gupta <vgupta at synopsys.com>
[vgupta: squashed patches, handle gcc -mno-unaligned-access quick]
Quoted 8 lines

Hi,

with my brand-new gcc-8.3.0, built from upstream gcc, this patch results in

arch/arc/kernel/setup.c: In function 'arc_cpu_mumbojumbo.constprop':
include/linux/compiler.h:348:38: error:
call to '__compiletime_assert_326' declared with attribute error: gcc doesn't support -mno-unaligned-access

Sorry it broke ur build - but it did catch a deficiency in compiler as intended - yay !

Quoted 2 lines

when trying to build allnoconfig or tinyconfig. I get the same results with gcc-8.2.0.

Unaligned access code gen was enabled by default in 7.x actually, but us OS folks were late to find that which led to the recent patch to disable it which was broken in gcc.
Claudiu fixed gcc promptly but that patch didn't make it to upstream, which I've posted now.

Quoted 8 lines

http://lists.infradead.org/pipermail/linux-snps-arc/2019-February/005522.html

Quoted 2 lines

Do I need some special compiler configuration flag when building the arcv2 toolchain ? Or am I simply out of luck for using upstream gcc ?

Once the patch hits mainline you could use that. In the mean time I'll back out the "detector" as it might trip other people too.

Can you possibly use $(cc-option,-mno-unaligned-access), or would that defeat the purpose ?

Quoted 1 line

Many Thx for your testing service.

My pleasure.

Thanks,
Guenter

#3 lore raw
Quoted 6 lines · Guenter Roeck
Quoted 3 lines

Once the patch hits mainline you could use that. In the mean time I'll back out the "detector" as it might trip other people too.

Can you possibly use $(cc-option,-mno-unaligned-access), or would that defeat the purpose ?

Indeed I was looking into Kconfig documentation and found something similar. However there's an anti-dependency. It is enabled by default and can only be disabled with $(cc-option,-mno-unaligned-access)
So guess we will have to change ARC_USE_UNALIGNED_MEM_ACCESS to ARC_*LACKS*_UNALIGNED_MEM_ACCESS

I'll tinker a bit and post an update soon.

-Vineet

#4 lore raw
Quoted 7 lines · Vineet Gupta
Quoted 6 lines · Guenter Roeck
Quoted 3 lines

Once the patch hits mainline you could use that. In the mean time I'll back out the "detector" as it might trip other people too.

Can you possibly use $(cc-option,-mno-unaligned-access), or would that defeat the purpose ?

Quoted 7 lines

Indeed I was looking into Kconfig documentation and found something similar. However there's an anti-dependency. It is enabled by default and can only be disabled with $(cc-option,-mno-unaligned-access)
So guess we will have to change ARC_USE_UNALIGNED_MEM_ACCESS to ARC_*LACKS*_UNALIGNED_MEM_ACCESS

I'll tinker a bit and post an update soon.

I got around to testing a patch with inverted option, unaligned access is default and ARC_UNALIGNED_ACCESS_NONE disables it. This option can depend on $(cc-option,-mno-unaligned-access). The issue however is "broken" gcc silently ignoring -mno-unaligned-access. So this option can't be auto disable with affected gcc. So this approach won't work. The only way out is remove the stringent check and just wait for upstream gcc to be fixed.