Re: [PATCH bpf-next] bpf: abstract loop unrolling pragmas in BPF selftests

["Jose E. Marchesi" <jose.marchesi@oracle.com>]

> On 2/8/24 10:04 AM, Yonghong Song wrote:
>>
>> On 2/8/24 8:51 AM, Jose E. Marchesi wrote:
>>>> On Thu, 2024-02-08 at 16:35 +0100, Jose E. Marchesi wrote:
>>>> [...]
>>>>
>>>>> If the compiler generates assembly code the same code for
>>>>> profile2.c for
>>>>> before and after, that means that the loop does _not_ get
>>>>> unrolled when
>>>>> profiler.inc.h is built with -O2 but without #pragma unroll.
>>>>>
>>>>> But what if #pragma unroll is used?  If it unrolls then, that
>>>>> would mean
>>>>> that the pragma does something more than -funroll-loops/-O2.
>>>>>
>>>>> Sorry if I am not making sense.  Stuff like this confuses me to no end
>>>>> ;)
>>>> Sorry, I messed up while switching branches :(
>>>> Here are the correct stats:
>>>>
>>>> | File            | insn # | insn # |
>>>> |                 | before |  after |
>>>> |-----------------+--------+--------|
>>>> | profiler1.bpf.o |  16716 |   4813 |
>>> This means:
>>>
>>> - With both `#pragma unroll' and -O2 we get 16716 instructions.
>>> - Without `#pragma unroll' and with -O2 we get 4813 instructions.
>>>
>>> Weird.
>>
>> Thanks for the analysis. I can reproduce with vs. without '#pragma
>> unroll' at -O2
>> level, the number of generated insns is indeed different, quite
>> dramatically
>> as the above numbers. I will do some checking in compiler.
>
> Okay, a quick checking compiler found that
>   - with "#pragma unroll" means no profitability test and do full
>    unroll as instructed


I don't think clang's `#pragma unroll' does full unroll.

On one side, AFAIK `pragma unroll' is supposed to be equivalent to
`pragma clang loop(enable)', which is different to `pragma clang loop
unroll(full)'.

On the other, if you replace `pragma unroll' with `pragma clang loop
unroll(full)' in the BPF selftests you will get branch instruction
overflows.

What criteria `pragma unroll' in clang uses in order to determine how
much it unrolls the loop, compared to -O2|-funroll-loops, I don't know.

>   - without "#pragma unroll" mean compiler will do profitability for full unroll,
>     if compiler thinks full unroll is not profitable, there will be no unrolling.
>
> So for gcc, even users saying '#pragma unroll', gcc still do
> profitability test?

GCC doesn't support `#pragma unroll'.

Hence in my original patch the macro __pragma_unroll expands to nothing
with GCC.  That will lead to the compiler perhaps not unrolling the loop
even with -O2|-funroll-loops.

>
>>
>>>
>>>> | profiler2.bpf.o |   2088 |   2050 |
>>> - Without `#pragma unroll' and with -O2 we get 2088 instructions.
>>> - With `#pragma loop unroll(disable)' and with -O2 we get 2050
>>>    instructions.
>>>
>>> Also surprising.
>>>
>>>> | profiler3.bpf.o |   4465 |   1690 |
>>