RFR: 8288564: C2: LShiftLNode::Ideal produces wrong result after JDK-8278114
Vladimir Kozlov
kvn at openjdk.org
Thu Jun 16 17:09:51 UTC 2022
On Thu, 16 Jun 2022 15:27:42 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!NgSy4cXXFdOP3hWRU50qCPbZdHy6I2V2k2esqYQl6jBwX_TiRxQJbDUu0gKCA-vG6C2PUQrKuk0DDhTxUPeguKWqvQ$
>
> 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
Good. Just one comment.
src/hotspot/share/opto/mulnode.cpp line 816:
> 814: if (add1->in(1) == add1->in(2)) {
> 815: // Convert "(x + x) << c0" into "x << (c0 + 1)"
> 816: assert(con != BitsPerJavaInteger - 1, "sanity check, optimization cannot be applied for con == 31");
The assert is useless because there is check at line 812 `(con < 16).
-------------
Marked as reviewed by kvn (Reviewer).
PR: https://git.openjdk.org/jdk19/pull/29
More information about the hotspot-compiler-dev
mailing list