I think we mostly agree it’s unfortunate that C conflated (for example) “character” and “small integer” and “raw memory” into the same type, “char.” Ideally those would be distinct types, and “integer” would support arithmetic +-*/etc. but not also support text- or bit-specific operations at all. (That’s a major reason we created a std::byte that supports bit-ops, but not text or arithmetic.)
So, in that light…
> the status quo is inconsistent, and satisfies neither the "<< means multiplication" nor the
> "<< means bit shift" perspective. I want one or the other, or a really good reason for the status quo.
… for an integer type we’d want (in an ideal world) all operations to be arithmetic, and no bit-ops. Right? If so, +1 for multiplication.
Of course there may be other issues here, such as historical and/or consistency. I’m just throwing in an ‘ideal world’ perspective as a data point in case it helps inform deciding between feasible options, I’m not steeped in this to know whether it’s the right answer all-up in this case.
From: email@example.com <firstname.lastname@example.org>
On Behalf Of Richard Smith
Sent: Friday, March 16, 2018 3:48 PM
To: ub <email@example.com>
Subject: Re: [ub] An update on signed integers
On Fri, 16 Mar 2018, 14:20 Arthur O'Dwyer, <firstname.lastname@example.org> wrote:
(Dropped some marginally relevant lists I get mod-queued on anyway)
> The notes should help clarify RIchard's rationale for proposing this.
http://wiki.edg.com/bin/view/Wg21jacksonville2018/Sg6P0907R0 has some discussion from Richard (Smith?) about (1<<32) needing to remain IDB-at-best on x86 platforms (Ctrl+F "crypto"). Total agreement on that front.
I don't see any notes on that page which are relevant to (1<<31) — there's just the straw poll results with zero prior discussion notated.
My rationale for wanting *some* change here is that the status quo is inconsistent, and satisfies neither the "<< means multiplication" nor the "<< means bit shift" perspective. I want one or the other, or a really good reason for the status quo.
And, FWIW, the bit shift option was my preference.
Shifting out of range (1 << 31) as currently defined (if shift has defined value when interpreted as unsigned, you get the value as signed, you can dhigt into but not past the sign, shifting past sign bit is UB), 1<<31 was defined and still is (as of Howard's paper), 2<<31 was UB and still is, WG14 is currently considering whether to adopt Howard's paper which made this. Should we take it back to undefined to do 1<<31?
SS F N A SA
2 7 5 1 1
Am I looking at the wrong page, perhaps?
P.S. — Alternatively, since 0x80000000 is allowed to be a trap representation in P0907r1 but 0xC0000000 is not, perhaps the intent of this straw poll was that 1<<31 should be UB but 3<<30 should be well-defined. That would be extra weird, though, so I hope that's not what happened.
On Fri, Mar 16, 2018 at 11:06 AM, JF Bastien <email@example.com> wrote:
On Fri, Mar 16, 2018 at 2:04 PM, Arthur O'Dwyer <firstname.lastname@example.org> wrote:
P0907r1 proposes this addition relative to the WD:
> If overflow caused by an operation which would require representing an integer which cannot be represented by the type, the behavior is undefined.
However, it simultaneously proposes these deletions relative to the WD:
[ Note: Operators can be regrouped according to the usual mathematical rules only where the operators really are associative or commutative...
> an operation that would have undefined behavior as specified in Clause 4 through 19 of this document [Note: including, for example,
signed integer overflow,certain pointer arithmetic, division by zero, or certain shift operations —end note]
The removals are all non-normative, but they seem to be aimed at eliminating references to "signed overflow is UB", even though signed overflow is still UB. Was there a sense in the room that we wanted to downplay the importance of teaching signed UB in C++, but not actually eliminate the UB itself? Or what's the point of re-wording these existing notes?
There was extensive discussion of this note, what to add / remove, etc, and no direction as to where to go. I'll ask EWG today.
Also, separately and less importantly, I'd love to hear someone's rationale for making (1<<31) UB in C++2a when it's IDB in C++17. The straw poll result in P0907r1 sounds unambiguous, but I don't understand what rationale could exist for taking this construct from IDB into UB. Is the assumption that people haven't yet had a chance to write any programs whose correctness depends on the value of (1<<31), so if we change it back to UB fast enough, nobody will notice?
The notes should help clarify RIchard's rationale for proposing this.
On Fri, Mar 16, 2018 at 8:56 AM, JF Bastien <email@example.com> wrote:
SG6 and SG12 discussed wg21.link/P0907r0 Signed Integers are Two’s Complement and provided extensive feedback.
I've attached an updated paper listing polls and addressing most feedback (except some wording fiddle) to the EWG wiki for this afternoon's discussion:
ub mailing list
ub mailing list