RFR: 8288564: C2: LShiftLNode::Ideal produces wrong result after JDK-8278114 [v2]

[Tobias Hartmann]

On Fri, 17 Jun 2022 08:46:39 GMT, Christian Hagedorn <chagedorn at openjdk.org> wrote:

>> [JDK-8278114](https://bugs.openjdk.org/browse/JDK-8278114) added the following transformation for integer and long left shifts:
>> 
>>  "(x + x) << c0" into "x << (c0 + 1)"
>> 
>> However, in the long shift case, this transformation is not correct if `c0` is 63:
>> 
>> 
>> (x + x) << 63 = 2x << 63
>> 
>> while
>> 
>>  (x + x) << 63 --transform--> x << 64 = x << 0 = x
>> 
>> which is not the same. For example, if `x = 1`:
>> 
>> 2x << 63 = 2 << 63 = 0 != 1
>> 
>> This optimization does not account for the fact that `x << 64` is the same as `x << 0 = x`. According to the [Java spec, chapter 15.19](https://docs.oracle.com/javase/specs/jls/se18/html/jls-15.html#jls-15.19), we only consider the six lowest-order bits of the right-hand operand (i.e. `"right-hand operand" & 0b111111`). Therefore, `x << 64` is the same as `x << 0` (`64 = 0b10000000 & 0b0111111 = 0`).
>> 
>> Integer shifts are not affected because we do not apply this transformation if `c0 >= 16`:
>> 
>> https://urldefense.com/v3/__https://github.com/openjdk/jdk19/blob/729164f53499f146579a48ba1b466c687802f330/src/hotspot/share/opto/mulnode.cpp*L810-L817__;Iw!!ACWV5N9M2RV99hQ!JWPzfSx6w6SrUyFy8xTwioPUJKolldgkYbjgaHAQigbBdFDqXDDLnS7FXS7pcF7i_wc-unEfP3EP_gRoSGk8CKqRj9ywZ4GZAA$ 
>> 
>> The fix I propose is to not apply this optimization for long left shifts if `c0 == 63`. I've added an additional sanity assertion for integer left shifts just in case this optimization is moved at some point and ending up outside the check for `con < 16`.
>> 
>> Thanks,
>> Christian
>
> Christian Hagedorn has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - typo
>  - change assert to comment

Nice analysis. Looks good.

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

Marked as reviewed by thartmann (Reviewer).

PR: https://git.openjdk.org/jdk19/pull/29