RFR: 8346664: C2: Optimize mask check with constant offset [v5]

[Matthias Ernst]

On Fri, 24 Jan 2025 16:05:57 GMT, Quan Anh Mai <qamai at openjdk.org> wrote:

>> src/hotspot/share/opto/mulnode.cpp line 2112:
>> 
>>> 2110: 
>>> 2111:   jint zeros = AndIL_min_trailing_zeros(phase, expr, bt);
>>> 2112:   return zeros > 0 && ((((jlong)1) << zeros) > mask_t->hi_as_long() && mask_t->lo_as_long() >= 0);
>> 
>> FWIW I came up with another way to formulate this. I think it makes even more clear how this is all about "expr is shifted to be completely left of the mask", i.e. we're comparing right-most value bit position to left-most mask bit position, all other bits are irrelevant.
>> 
>> This could also inform how to bias the random generator. Let me know what you think.
>> 
>> Suggestion:
>> 
>>   jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
>>   if (expr_trailing_zeros == 0) return false;
>> 
>>   const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
>>   if (mask_t == nullptr || mask_t->lo_as_long() < 0) return false;
>> 
>>   jint mask_bit_width = mask_t->hi_as_long() == 0 ? 0 : (BitsPerLong - count_leading_zeros(mask_t->hi_as_long()));
>>   return expr_trailing_zeros >= mask_bit_width;
>
> `mask_t->lo_as_long() == 0` does not imply `mask == 0`, though. Other than that I think it is a great suggestion.

You're right. Fixed above. I was trying to get around special casing 0 (count_*_zeros don't like them).

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/22856#discussion_r1928929675