[llvm-dev] The undef story

[Hal Finkel via llvm-dev]

On 06/29/2017 12:43 PM, Peter Lawrence wrote:
>
>> On Jun 29, 2017, at 9:32 AM, Hal Finkel <hfinkel at anl.gov 
>> <mailto:hfinkel at anl.gov>> wrote:
>>
>>
>> On 06/29/2017 10:41 AM, Peter Lawrence wrote:
>>>
>>>> On Jun 29, 2017, at 4:39 AM, Hal Finkel <hfinkel at anl.gov 
>>>> <mailto:hfinkel at anl.gov>> wrote:
>>>>
>>>> On 06/28/2017 05:33 PM, Peter Lawrence wrote:
>>>>> Chandler,
>>>>>                where we disagree is in whether the current project 
>>>>> is moving the issue
>>>>> forward.  It is not.  It is making the compiler more complex for 
>>>>> no additional value.
>>>>>
>>>>> The current project is not based in evidence, I have asked for any 
>>>>> SPEC benchmark
>>>>> that shows performance gain by the compiler taking advantage of 
>>>>> “undefined behavior”
>>>>> and no one can show that.
>>>>
>>>> I can't comment on SPEC, but this does remind me of code I was 
>>>> working on recently. To abstract the relevant parts, it looked 
>>>> something like this:
>>>>
>>>> template <typename T>
>>>> int do_something(T mask, bool cond) {
>>>> if (mask & 2)
>>>> return 1;
>>>>
>>>> if (cond) {
>>>> T high_mask = mask >> 48;
>>>> if (high_mask > 5)
>>>> do_something_1(high_mask);
>>>> else if (high_mask > 3)
>>>> do_something_2();
>>>> }
>>>>
>>>> return 0;
>>>> }
>>>>
>>>> This function ended up being instantiated on different types T 
>>>> (e.g. unsigned char, unsigned int, unsigned long, etc.) and, 
>>>> dynamically, cond was always false when T was char. The question 
>>>> is: Can the compiler eliminate all of the code predicated on cond 
>>>> for the smaller types? In this case, this code was hot, and 
>>>> moreover, performance depended on the fact that, for T = unsigned 
>>>> char, the function was inlined and the branch on cond was 
>>>> eliminated. In the relevant translation unit, however, the compiler 
>>>> would never see how cond was set.
>>>>
>>>> Luckily, we do the right thing here currently. In the case where T 
>>>> = unsigned char, we end up folding both of the high_mask tests as 
>>>> though they were false. That entire part of the code is eliminated, 
>>>> the function is inlined, and everyone is happy.
>>>>
>>>> Why was I looking at this? As it turns out, if the 'else if' in 
>>>> this example is just 'else', we don't actually eliminate both sides 
>>>> of the branch. The same is true for many other variants of the 
>>>> conditionals (i.e. we don't recognize all of the code as dead).
>>>
>>>
>>> I apologize in advance if I have missed something here and am 
>>> misreading your example...
>>>
>>> This doesn’t make sense to me, a shift amount of 48 is “undefined” 
>>> for unsigned char,
>>> How do we know this isn’t a source code bug,
>>> What makes us think the the user intended the result to be “0”.
>>
>> As I said, this is representation of what the real code did, and 
>> looked like, after other inlining had taken place, etc. In the 
>> original form, the user's intent was clear. That code is never 
>> executed when T is a small integer type.
>
>
> I will still have a hard time believing this until I see a real 
> example, can you fill in the details ?

As I recall, it was roughly like this: There is some collection of 
objects that have some flags (stored in a bit mask). Some of these 
objects optionally supported tracing, and for such objects, the tracing 
flags are stored in the high bits of the bit mask. Some of the low bits 
had universal meanings. For objects that didn't support tracing at all, 
they used a smaller bit mask (to save on space, etc.). External logic 
guaranteed that the tracing bits were only checked for objects that 
supported tracing (in addition to other conditions being true). There 
was a template class that handled the bit masks and interpreted some of 
the tracing bits (on supported objects, but only if tracing was 
enabled). This function was a member of that template class. The 48 was 
not directly specified in the function like this, but was a parameter, 
and so only appeared like this after inlining.

  -Hal

>
>
> Peter Lawrence.
>
>
>
>
>

-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170629/d3693a19/attachment.html>