llvm dev - May 2019 - RFC: On removing magic numbers assuming 8-bit bytes

On Mon, 2019-05-06 at 15:56 -0700, Philip Reames via llvm-dev
wrote:
> > 
> On 5/6/19 2:43 AM, Tim Northover via llvm-dev wrote:
> > On Mon, 6 May 2019 at 10:13, James Courtier-Dutton via llvm-dev
> > <llvm-dev at lists.llvm.org> wrote:
> > > Although the above is mentioning bytes, looking at the "/
> > > 8"   and "& 0x7" makes it look like the author
meant octets and
> > > not bytes.
> > > Bytes can be any size of bits.
> > 
> > I don't think you'll have much luck trying to make that stick
for a
> > general audience, or even a general compiler-writer audience. Byte
> > is
> > far too strongly associated with 8 bits these days.
> 
> +1 Please don't try; insisting on a distinction will confuse many new
> contributors.
Yes, my interpretation is that the community is leaning toward
addressable unit as terminology.
> > > Octets are only ever 8 bits.
> > 
> > You might be able to convert all uses of byte to octet and abandon
> > byte entirely, but at that point why bother? It feels like a change
> > just for the sake of pedantry.
> > 
> > I like the "addressable unit" name, though it's a bit
long
> > (AddrUnit
> > seems OK). It at least signals to a reader that there might be
> > something weird going on. Getting someone writing new code to think
> > in
> > those terms is a different matter, of course, but I don't think
any
> > of
> > the changes under discussion really help there.
> > 
> > BTW, is there an open source backend (in a fork, I assume) that
> > does
> > this? So that we can get some kind of idea of the real scope of the
> > changes needed.
> 
> Strongly agreed.
> 
> My personal take is this is an invasive enough change with enough
> likely
> ongoing maintenance fall out to require substantial justification
> before
> the work was undertaken upstream.
My hope and belief is that having good names instead of these magical
numbers won't be a burden but rather an improvement long-term. 
>   A open source backend proposed for
> inclusion upstream would be one part of that.
That is not on the table right now. However, as the work required to
make such an inclusion happen will be reduced by this cleanup, the
likelihood that it happens in the future should increase.
>   Active contribution from
> the sponsors in other areas would also be a key factor.
I'm not sure how to interpret that, but our team here at Ericsson is
fairly large, we have been working with this out-of-tree backend since
2011 and as a group, we contribute to upstream e.g. by helping out with
the fixedpoint upstreaming, by solving and filing TRs (we're pretty
good at testing I'd say), improving debug information and more.
> 
> > 
> > Cheers.
> > 
> > Tim.
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

On 5/8/19 1:25 AM, Jesper Antonsson wrote:
> On Mon, 2019-05-06 at 15:56 -0700, Philip Reames via llvm-dev wrote:
>> On 5/6/19 2:43 AM, Tim Northover via llvm-dev wrote:
>>> On Mon, 6 May 2019 at 10:13, James Courtier-Dutton via llvm-dev
>>> <llvm-dev at lists.llvm.org> wrote:
>>>> Although the above is mentioning bytes, looking at the "/
>>>> 8"   and "& 0x7" makes it look like the
author meant octets and
>>>> not bytes.
>>>> Bytes can be any size of bits.
>>> I don't think you'll have much luck trying to make that
stick for a
>>> general audience, or even a general compiler-writer audience. Byte
>>> is
>>> far too strongly associated with 8 bits these days.
>> +1 Please don't try; insisting on a distinction will confuse many
new
>> contributors.
> Yes, my interpretation is that the community is leaning toward
> addressable unit as terminology.
>
>>>> Octets are only ever 8 bits.
>>> You might be able to convert all uses of byte to octet and abandon
>>> byte entirely, but at that point why bother? It feels like a change
>>> just for the sake of pedantry.
>>>
>>> I like the "addressable unit" name, though it's a bit
long
>>> (AddrUnit
>>> seems OK). It at least signals to a reader that there might be
>>> something weird going on. Getting someone writing new code to think
>>> in
>>> those terms is a different matter, of course, but I don't think
any
>>> of
>>> the changes under discussion really help there.
>>>
>>> BTW, is there an open source backend (in a fork, I assume) that
>>> does
>>> this? So that we can get some kind of idea of the real scope of the
>>> changes needed.
>> Strongly agreed.
>>
>> My personal take is this is an invasive enough change with enough
>> likely
>> ongoing maintenance fall out to require substantial justification
>> before
>> the work was undertaken upstream.
> My hope and belief is that having good names instead of these magical
> numbers won't be a burden but rather an improvement long-term. 
>
>>   A open source backend proposed for
>> inclusion upstream would be one part of that.
> That is not on the table right now. However, as the work required to
> make such an inclusion happen will be reduced by this cleanup, the
> likelihood that it happens in the future should increase.
Given this, I'm not sure the community as a whole should take on the
burden of supporting non b-byte addressable units.  I see this as a
precondition.  To be clear, I don't care *which* backend there is,
doesn't have to be yours, but the presence of at least one would seem
necessary for testing if nothing else.
>
>>   Active contribution from
>> the sponsors in other areas would also be a key factor.
> I'm not sure how to interpret that, but our team here at Ericsson is
> fairly large, we have been working with this out-of-tree backend since
> 2011 and as a group, we contribute to upstream e.g. by helping out with
> the fixedpoint upstreaming, by solving and filing TRs (we're pretty
> good at testing I'd say), improving debug information and more.
Ok.

p.s. If my wording came across as implying any disrespect, sorry!  I was
making a general point, not thinking about how it might be read in
context. 
>
>>> Cheers.
>>>
>>> Tim.
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

On Wed, 2019-05-08 at 11:12 -0700, Philip Reames wrote:
> On 5/8/19 1:25 AM, Jesper Antonsson wrote:
> > On Mon, 2019-05-06 at 15:56 -0700, Philip Reames via llvm-dev
> > wrote:
> > > On 5/6/19 2:43 AM, Tim Northover via llvm-dev wrote:
> > > > On Mon, 6 May 2019 at 10:13, James Courtier-Dutton via
llvm-dev
> > > > <llvm-dev at lists.llvm.org> wrote:
> > > > > Although the above is mentioning bytes, looking at the
"/
> > > > > 8"   and "& 0x7" makes it look like
the author meant octets
> > > > > and
> > > > > not bytes.
> > > > > Bytes can be any size of bits.
> > > > 
> > > > I don't think you'll have much luck trying to make
that stick
> > > > for a
> > > > general audience, or even a general compiler-writer
audience.
> > > > Byte
> > > > is
> > > > far too strongly associated with 8 bits these days.
> > > 
> > > +1 Please don't try; insisting on a distinction will confuse
many
> > > new
> > > contributors.
> > 
> > Yes, my interpretation is that the community is leaning toward
> > addressable unit as terminology.
> > 
> > > > > Octets are only ever 8 bits.
> > > > 
> > > > You might be able to convert all uses of byte to octet and
> > > > abandon
> > > > byte entirely, but at that point why bother? It feels like a
> > > > change
> > > > just for the sake of pedantry.
> > > > 
> > > > I like the "addressable unit" name, though
it's a bit long
> > > > (AddrUnit
> > > > seems OK). It at least signals to a reader that there might
be
> > > > something weird going on. Getting someone writing new code
to
> > > > think
> > > > in
> > > > those terms is a different matter, of course, but I
don't think
> > > > any
> > > > of
> > > > the changes under discussion really help there.
> > > > 
> > > > BTW, is there an open source backend (in a fork, I assume)
that
> > > > does
> > > > this? So that we can get some kind of idea of the real scope
of
> > > > the
> > > > changes needed.
> > > 
> > > Strongly agreed.
> > > 
> > > My personal take is this is an invasive enough change with enough
> > > likely
> > > ongoing maintenance fall out to require substantial justification
> > > before
> > > the work was undertaken upstream.
> > 
> > My hope and belief is that having good names instead of these
> > magical
> > numbers won't be a burden but rather an improvement long-term. 
> > 
> > >   A open source backend proposed for
> > > inclusion upstream would be one part of that.
> > 
> > That is not on the table right now. However, as the work required
> > to
> > make such an inclusion happen will be reduced by this cleanup, the
> > likelihood that it happens in the future should increase.
> 
> Given this, I'm not sure the community as a whole should take on the
> burden of supporting non b-byte addressable units.  I see this as a
> precondition.  To be clear, I don't care *which* backend there is,
> doesn't have to be yours, but the presence of at least one would seem
> necessary for testing if nothing else.
I agree that an in-tree target is needed for actual support. However,
we're merely suggesting a gradual cleanup of magic numbers in order to
make the code a bit more readable and make life easier for a number of
downstream targets. It will not result in support, but it would make
any effort to create support (or maintain support downstream)
significantly smaller. This would also make it a bit more likely that
LLVM is the compiler of choice for such targets, some of which might
want to upstream eventually.

The onus is on interested parties to maintain any gains, and Ericsson
is offering to do that in a no-drama way with the help of other
companies that have voiced their interest. We continuously merge and
test against top of tree and would act accordingly, if allowed.

As the discussion is subsiding, I'm unsure about how to conclude this
RFC. Several parties have said they support this effort, others have
pitched in with suggestions on terminology and such (which perhaps
indicates that they are not opposed in general). JF Bastien and you ask
for in-tree targets, although JF did indicate that it made sense to
first clean up.

On "byte" vs "addressable unit", we've been thinking a
bit and are
leaning toward staying with the prevalent "byte" terminology for as
long as upstream is 8-bit-only to avoid mixed terminology or larger
patches. However, we're flexible on this, and I've uploaded a twin
patch (in D61725) to my original showcase showing how "addressable
unit" could look. 
> > 
> > >   Active contribution from
> > > the sponsors in other areas would also be a key factor.
> > 
> > I'm not sure how to interpret that, but our team here at Ericsson
> > is
> > fairly large, we have been working with this out-of-tree backend
> > since
> > 2011 and as a group, we contribute to upstream e.g. by helping out
> > with
> > the fixedpoint upstreaming, by solving and filing TRs (we're
pretty
> > good at testing I'd say), improving debug information and more.
> 
> Ok.
> 
> p.s. If my wording came across as implying any disrespect, sorry!  I
> was
> making a general point, not thinking about how it might be read in
> context. 
No problem, thanks!
> 
> > 
> > > > Cheers.
> > > > 
> > > > Tim.
> > > > _______________________________________________
> > > > LLVM Developers mailing list
> > > > llvm-dev at lists.llvm.org
> > > > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> > > 
> > > _______________________________________________
> > > LLVM Developers mailing list
> > > llvm-dev at lists.llvm.org
> > > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev