[LLVMdev] C as used/implemented in practice: analysis of responses

[Sean Silva]

On Tue, Jul 7, 2015 at 10:26 AM, Chris Lattner <clattner at apple.com> wrote:

>
> On Jul 1, 2015, at 3:20 PM, Sean Silva <chisophugis at gmail.com> wrote:
>
>
>
> On Wed, Jul 1, 2015 at 12:22 PM, Russell Wallace <
> russell.wallace at gmail.com> wrote:
>
>> I am arguing in favor of a point, and I understand you disagree with it,
>> but I don't think I'm dismissing any use cases except a very small
>> performance increment.
>>
>
> I'm sure Google has numbers about how much electricity/server cost they
> save for X% performance improvement.
> I'm sure Apple has numbers about how much money they make with X% improved
> battery life.
> I'm not convinced that the cost of some of these bugs is actually larger
> than the benefit of faster programs. Nor am I convinced about the inverse.
> I'm just pointing out that pointing to a "bad bug" caused by a certain
> optimization without comparing the cost of the bug to the benefit of the
> optimization is basically meaningless. You'll need to quantify "very small
> performance improvement" and put it in context of the bugs you're talking
> about.
>
>
> As with many things, it is more complicated than that.  The performance
> effects of optimizations are often non-linear, and you can take a look at
> many of the worst forms of UB in C and easily show cases where they allow
> 2x speedups, not just 2%.
>
> For example, consider undefined behavior for integer overflow:
>
>   for (int i = 0; i <= N; ++i) {
>      …
>   }
>
> When compiling for a 64-bit machine, you really want to promote the
> induction variable to 64-bits.  Further, knowing the trip count of a loop
> is extremely important for many loop optimizations.  Unfortunately, without
> being able to assume undefined integer wraparound, you get neither of these
> from C.
>
> -fstrict-aliasing is another great example.  In many cases, it makes no
> difference whatsoever.  OTOH, on code like:
>
> void doLoopThing(float *array, int *N) {
>     for (int i = 0; i < *N; ++i) {
>        array[i] = array[i] + 1;
>   }
>
> You can easily get a 2x or more speedup due to auto-vectorization if you
> can assume -fstrict-aliasing.  Of course usually you wouldn’t write this
> code, you’d get this because doLoopThing is a template, and N is passed in
> as a reference.
>


You're absolutely right that it's complicated, but I don't think this is
the best example.

The 2x speedup optimizations you're talking about can be done even without
strict aliasing or signed overflow. You just have to emit runtime checks.
It's analogous to emitting the "remainder" loop when doing
autovectorization. Imagine if the standard made it undefined for a loop
over an array of more than 4 floats to not be suitable for vectorization:
sure, that would be nice and make the vectorizer's life easier, but we can
mostly get the "big bang" speedup without it, in the sense that not having
the standard say it is UB probably results in closer to 2% "slowdown" in
the aggregate across all these loops (factoring in icache, etc.) than 2x.

-- Sean Silva


>
>
> Anyway, I could go on and on here, and I’ve spent a lot of time over the
> years thinking about how to improve the situation: can we make clang detect
> more of these, can we make the optimizer more conservative in certain cases
> etc?  This is why (for example) our TBAA uses simple structural points-to
> analysis before using TBAA.  With GCC’s implementation (circa GCC 4.0, I
> have no idea what they are doing now), GCC would “miscompile” code like:
>
> float bitcast(int x) {
>   return *(float*)&x;
> }
>
> This code is a TBAA violation, but is also “obvious” what the programmer
> meant.  LLVM being “nicer” in this case is a feature.  It is irritating
> that the union version of this is also technically UB or implementation
> defined behavior, so that isn’t portable either (a C programmer needs to
> magically know that memcpy is the safe way to do this).
>
> However, as I’ve continued to dig into this, my feeling is that there
> really is no satisfactory solution to these issues.  The problem here are
> pervasive structural problems in the C language: In the first example
> above, it is that “int” is the default type people generally reach for, not
> “long”, and that array indexing is expressed with integers instead of
> iterators.  This isn’t something that we’re going to “fix" in the C
> language, the C community, or the body of existing C code.  Likewise, while
> C++ has made definite progress here by replacing some of these idioms (e.g.
> with iterators), it adds its own layers of UB on, and doesn’t actually
> *subtract* the worst things in C.
>
> My conclusion is that C, and derivatives like C++, is a very dangerous
> language the write safety/correctness critical software in, and my personal
> opinion is that it is almost impossible to write *security* critical
> software in it.  This isn’t really C’s fault if you consider where it was
> born and evolved from (some joke that it started as a *very* nice high
> level assembler for the PDP11
> https://en.wikipedia.org/wiki/C_(programming_language)#Early_developments
> :-).
>
> There are many more modern and much safer languages that either eliminate
> the UB entirely through language design (e.g. using a garbage collector to
> eliminate an entire class of memory safety issues, completely disallowing
> pointer casts to enable TBAA safely, etc), or by intentionally spending a
> bit of performance to provide a safe and correct programming model (e.g. by
> guaranteeing that integers will trap if they overflow).  My hope is that
> the industry will eventually move to better systems programming languages,
> but that will take a very very long time...
>
> -Chris
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150707/197703e7/attachment.html>