llvm dev - Jul 2015 - [LLVMdev] C as used/implemented in practice: analysis of responses

On 1 July 2015 at 17:15, Russell Wallace <russell.wallace at gmail.com>
wrote:
> I'm proposing that LLVM unilaterally replace most undefined behaviour
with
> implementation-defined behaviour.
That's precisely the problem. Which behaviour?

Let's have an example:

struct Foo {
  long a[95];
  char b[4];
  double c[2];
};

void fuzz(Foo &F) {
  for (int i=0; i<100; i++)
    F.a[i] = 123;
}

There are many ways I can do this "right":

1. Only go up to 95, since you're using an integer to set the value.
2. Go up to 96, since char is an integer type.
2. Go all the way to 100, but casting "123" to double from 97 onwards,
in pairs
3. Go all the way to 100, and set integer 123 bitwise (for whatever fp
representation that is) from 97
4. Do any of above, and emit a warning
5. Bail on error

Compilers prefer not to bail on error, since the standard permits it.
A warning would be a good thing, though.

Now, since it's a warning, I *have* to output something. What? Even
considering one compiler, you'll have to convince *most* <compilerX>
engineers to agree on something, and that's not trivial.

Moreover, this loop is very easy to vectorise, and that would give me
4x speed improvements for 4-way vectorization. That's too much for
compilers to pass.

If I create a vectorised loop that goes all the way to 92, I'll have
to create a tail loop. If I don't want to create a tail loop, I have
to override 'b' (and probably 'c') on a vector write. If I
implement
the variations where I can do that, the vectoriser will be very happy.
People generally like when the vectoriser is happy.

Now, you have a "safe mode" where these things don't happen.
Let's say
you and me agree that it should only go to 95, since this is "probably
what the user wants". But some programmers *use* that as a feature,
and the standard allow it, so we *have* to implement it *both*.

Best case scenario, you have now implemented two completely different
behaviours for every undefined behaviour in each standard. Worse
still, you have divided the programmers in two classes: those that
play it safe, and those that don't, essentially creating two different
programming languages. Code that compiles and work with
compilerA+safe_mode will not necessarily compile/work with
compilerB+safe_mode or compilerA+full_mode either.

C and C++ are already complicated enough, with so many standard levels
to implement (C90, C99, C11, C++03, C++11, C++14, etc) that
duplicating each and everyone of them, *per compiler*, is not
something you want to do.

That will, ultimately, move compilers away from each other, which is
not what most users really want.

cheers,
--renato

Your option 3 is the preferred way to handle your example, because it
allows one to reason about the behaviour of the program. As a bonus, it
does so without interfering in any way with optimisation.

What most users really want - and what many of them /expect/ even though
it's currently not actually the case - is for C to behave like high-level
assembly language, for code to translate straightforwardly into machine
operations. Yes, it would be nice to get consistent behaviour between
compilers. But that's not going to happen. Failing that, given that there
is no way to improve all compilers, improving one compiler would be much
better than nothing.

On Wed, Jul 1, 2015 at 5:58 PM, Renato Golin <renato.golin at linaro.org>
wrote:
> On 1 July 2015 at 17:15, Russell Wallace <russell.wallace at
gmail.com>
> wrote:
> > I'm proposing that LLVM unilaterally replace most undefined
behaviour
> with
> > implementation-defined behaviour.
>
> That's precisely the problem. Which behaviour?
>
> Let's have an example:
>
> struct Foo {
>   long a[95];
>   char b[4];
>   double c[2];
> };
>
> void fuzz(Foo &F) {
>   for (int i=0; i<100; i++)
>     F.a[i] = 123;
> }
>
> There are many ways I can do this "right":
>
> 1. Only go up to 95, since you're using an integer to set the value.
> 2. Go up to 96, since char is an integer type.
> 2. Go all the way to 100, but casting "123" to double from 97
onwards, in
> pairs
> 3. Go all the way to 100, and set integer 123 bitwise (for whatever fp
> representation that is) from 97
> 4. Do any of above, and emit a warning
> 5. Bail on error
>
> Compilers prefer not to bail on error, since the standard permits it.
> A warning would be a good thing, though.
>
> Now, since it's a warning, I *have* to output something. What? Even
> considering one compiler, you'll have to convince *most*
<compilerX>
> engineers to agree on something, and that's not trivial.
>
> Moreover, this loop is very easy to vectorise, and that would give me
> 4x speed improvements for 4-way vectorization. That's too much for
> compilers to pass.
>
> If I create a vectorised loop that goes all the way to 92, I'll have
> to create a tail loop. If I don't want to create a tail loop, I have
> to override 'b' (and probably 'c') on a vector write. If I
implement
> the variations where I can do that, the vectoriser will be very happy.
> People generally like when the vectoriser is happy.
>
> Now, you have a "safe mode" where these things don't happen.
Let's say
> you and me agree that it should only go to 95, since this is "probably
> what the user wants". But some programmers *use* that as a feature,
> and the standard allow it, so we *have* to implement it *both*.
>
> Best case scenario, you have now implemented two completely different
> behaviours for every undefined behaviour in each standard. Worse
> still, you have divided the programmers in two classes: those that
> play it safe, and those that don't, essentially creating two different
> programming languages. Code that compiles and work with
> compilerA+safe_mode will not necessarily compile/work with
> compilerB+safe_mode or compilerA+full_mode either.
>
> C and C++ are already complicated enough, with so many standard levels
> to implement (C90, C99, C11, C++03, C++11, C++14, etc) that
> duplicating each and everyone of them, *per compiler*, is not
> something you want to do.
>
> That will, ultimately, move compilers away from each other, which is
> not what most users really want.
>
> cheers,
> --renato
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150701/d5e9d899/attachment.html>

On Wed, Jul 1, 2015 at 10:19 AM, Russell Wallace <russell.wallace at
gmail.com>
wrote:
> Your option 3 is the preferred way to handle your example, because it
> allows one to reason about the behaviour of the program. As a bonus, it
> does so without interfering in any way with optimisation.
>
The problem is that setting it bit-wise is platform dependent. For example,
that concrete example (with 95 replaced by 94) would have padding after the
char array depending on sizeof(long), so some of the stores might not even
affect any of the fields declared in the struct. So you're basically asking
the compiler to do something pretty arbitrary and platform dependent, which
is what it is already basically doing (minus a lot of the platform
dependence). This is the sort of thing that I was talking about upthread.

It could be argued that deleting the loop (which would lead to an obvious
error immediately, with any basic testing) is better than a "works on my
machine" situation where the latent bug is one due to bitwise layout
differences -- I wouldn't want to debug that. Obviously an intelligible
static diagnosis of the situation is the ideal case, but doing so is a hard
problem (and may be too noisy anyway).

-- Sean Silva

>
> What most users really want - and what many of them /expect/ even though
> it's currently not actually the case - is for C to behave like
high-level
> assembly language, for code to translate straightforwardly into machine
> operations. Yes, it would be nice to get consistent behaviour between
> compilers. But that's not going to happen. Failing that, given that
there
> is no way to improve all compilers, improving one compiler would be much
> better than nothing.
>
> On Wed, Jul 1, 2015 at 5:58 PM, Renato Golin <renato.golin at
linaro.org>
> wrote:
>
>> On 1 July 2015 at 17:15, Russell Wallace <russell.wallace at
gmail.com>
>> wrote:
>> > I'm proposing that LLVM unilaterally replace most undefined
behaviour
>> with
>> > implementation-defined behaviour.
>>
>> That's precisely the problem. Which behaviour?
>>
>> Let's have an example:
>>
>> struct Foo {
>>   long a[95];
>>   char b[4];
>>   double c[2];
>> };
>>
>> void fuzz(Foo &F) {
>>   for (int i=0; i<100; i++)
>>     F.a[i] = 123;
>> }
>>
>> There are many ways I can do this "right":
>>
>> 1. Only go up to 95, since you're using an integer to set the
value.
>> 2. Go up to 96, since char is an integer type.
>> 2. Go all the way to 100, but casting "123" to double from 97
onwards, in
>> pairs
>> 3. Go all the way to 100, and set integer 123 bitwise (for whatever fp
>> representation that is) from 97
>> 4. Do any of above, and emit a warning
>> 5. Bail on error
>>
>> Compilers prefer not to bail on error, since the standard permits it.
>> A warning would be a good thing, though.
>>
>> Now, since it's a warning, I *have* to output something. What? Even
>> considering one compiler, you'll have to convince *most*
<compilerX>
>> engineers to agree on something, and that's not trivial.
>>
>> Moreover, this loop is very easy to vectorise, and that would give me
>> 4x speed improvements for 4-way vectorization. That's too much for
>> compilers to pass.
>>
>> If I create a vectorised loop that goes all the way to 92, I'll
have
>> to create a tail loop. If I don't want to create a tail loop, I
have
>> to override 'b' (and probably 'c') on a vector write.
If I implement
>> the variations where I can do that, the vectoriser will be very happy.
>> People generally like when the vectoriser is happy.
>>
>> Now, you have a "safe mode" where these things don't
happen. Let's say
>> you and me agree that it should only go to 95, since this is
"probably
>> what the user wants". But some programmers *use* that as a
feature,
>> and the standard allow it, so we *have* to implement it *both*.
>>
>> Best case scenario, you have now implemented two completely different
>> behaviours for every undefined behaviour in each standard. Worse
>> still, you have divided the programmers in two classes: those that
>> play it safe, and those that don't, essentially creating two
different
>> programming languages. Code that compiles and work with
>> compilerA+safe_mode will not necessarily compile/work with
>> compilerB+safe_mode or compilerA+full_mode either.
>>
>> C and C++ are already complicated enough, with so many standard levels
>> to implement (C90, C99, C11, C++03, C++11, C++14, etc) that
>> duplicating each and everyone of them, *per compiler*, is not
>> something you want to do.
>>
>> That will, ultimately, move compilers away from each other, which is
>> not what most users really want.
>>
>> cheers,
>> --renato
>>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150701/dd747e97/attachment.html>