Alignment requirements for 64-bit atomics on 386

[Kevin Chowski]

I am debugging an issue in which we are seeing a crash on an atomic access; on go 1.14.7 and 1.15.2 this reproduces with the following error: 

> panic: runtime error: invalid memory address or nil pointer dereference

> [signal SIGSEGV: segmentation violation code=0x1 addr=0x0 pc=0x8049f7c]

On gotip (go version devel +3fd4917 Fri Nov 20 20:02:47 2020 +0000 linux/amd64) I see the following instead:

> panic: unaligned 64-bit atomic operation

(I'm glad the error message got better! Kudos to whomever did that.)

My confusion is that I don't quite see this restriction being documented in sync/atomic's docs: ctrl-f for "alignment" only reveals information related to MIPS and ARM, whereas this issue is for Intel 386. Is this working as intended (e.g. documentation bug) or is this something deeper?

-Kevin

[Ian Lance Taylor]

It's down at the very bottom of https://golang.org/pkg/sync/atomic, in
the "Bugs" section, which for some reason refers to i386 using the
name x86-32.

Ian

[Kevin Chowski]

Thanks Ian for the very quick reply!

I see you're referencing this:

> On ARM, x86-32, and 32-bit MIPS, it is the caller's responsibility to arrange for 64-bit alignment of 64-bit words accessed atomically. The first word in a variable or in an allocated struct, array, or slice can be relied upon to be 64-bit aligned.

Yeah it's a little unfortunate that it doesn't say i386, I ctrl-f'd for "i386" on the page too and I probably would have correctly read that sentence if it were in there. Unfortunately I read that sentence multiple times but I guess I missed the obvious "x86" in there; my brain saw "ARM 32  32-bit MIPS" and misread it as "32-bit ARM and 32-bit MIPS". My mistake for sure.

Thanks again,

Kevin

[Ian Lance Taylor]

On Fri, Nov 20, 2020 at 1:03 PM 'Kevin Chowski' via golang-nuts
<golan...@googlegroups.com> wrote:
>
> Thanks Ian for the very quick reply!
>
> I see you're referencing this:
>
> > On ARM, x86-32, and 32-bit MIPS, it is the caller's responsibility to arrange for 64-bit alignment of 64-bit words accessed atomically. The first word in a variable or in an allocated struct, array, or slice can be relied upon to be 64-bit aligned.
>
> Yeah it's a little unfortunate that it doesn't say i386, I ctrl-f'd for "i386" on the page too and I probably would have correctly read that sentence if it were in there. Unfortunately I read that sentence multiple times but I guess I missed the obvious "x86" in there; my brain saw "ARM 32 32-bit MIPS" and misread it as "32-bit ARM and 32-bit MIPS". My mistake for sure.

Honestly I don't think it was your mistake. I think the docs are
confusing. I think we should change the docs to say "i386", since
that is what we use for a GOARCH value. I will do that if nobody
beats me to it.

Ian

> On Friday, November 20, 2020 at 1:54:28 PM UTC-7 Ian Lance Taylor wrote:
>>
>> On Fri, Nov 20, 2020 at 12:48 PM 'Kevin Chowski' via golang-nuts
>> <golan...@googlegroups.com> wrote:
>> >
>> > I am debugging an issue in which we are seeing a crash on an atomic access; on go 1.14.7 and 1.15.2 this reproduces with the following error:
>> >
>> > > panic: runtime error: invalid memory address or nil pointer dereference
>> > > [signal SIGSEGV: segmentation violation code=0x1 addr=0x0 pc=0x8049f7c]
>> >
>> > On gotip (go version devel +3fd4917 Fri Nov 20 20:02:47 2020 +0000 linux/amd64) I see the following instead:
>> >
>> > > panic: unaligned 64-bit atomic operation
>> >
>> > (I'm glad the error message got better! Kudos to whomever did that.)
>> >
>> > My confusion is that I don't quite see this restriction being documented in sync/atomic's docs: ctrl-f for "alignment" only reveals information related to MIPS and ARM, whereas this issue is for Intel 386. Is this working as intended (e.g. documentation bug) or is this something deeper?
>>
>> It's down at the very bottom of https://golang.org/pkg/sync/atomic, in
>> the "Bugs" section, which for some reason refers to i386 using the
>> name x86-32.
>>
>> Ian
>

> --
> You received this message because you are subscribed to the Google Groups "golang-nuts" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/b92e1561-441a-4ff3-8d7c-6929801a0b9bn%40googlegroups.com.