RFR: 8351515: C2 incorrectly removes double negation for double and float

[Christian Hagedorn]

On Fri, 21 Mar 2025 10:20:53 GMT, Manuel Hässig <duke at openjdk.org> wrote:

> # Issue Summary
> 
> A fuzzer run discovered that code of the form `0 - (0 - x)` yields the wrong result for `x = -0.0`. According to the [Java Language Spec 15.8.2](https://docs.oracle.com/javase/specs/jls/se23/html/jls-15.html#jls-15.18.2) the sum of floating point zeroes of opposite sign is positive zero. Hence, `(0 - (0 - (-0.0))` must result in `+0.0`, but due to the folding of all double negations in `SubNode::Identity` this was optimized to `-0.0`
> 
> # Changeset overview
> 
> To fix this issue, I excluded floating point numbers from the folding of double negations, which also includes `Float16` values. This might seem excessive at first glance, but we do not track range information for floating point types. Hence, we could still perform the folding of the double negation if a floating point constant is not `-0.0`. However, constant folding already takes care of this.
> 
> Changes:
>  - IR-Framework: fix `IRNode.SUB`not matching `SubHFNode`
>  - Add a regression IR-test
>  - Exclude floating point `SubNodes` from folding double negations
> 
> # Testing
> - [Github Actions](https://github.com/mhaessig/jdk/actions/runs/13989207348)
> - `tier1` through `tier5` plus Oracle internal testing

Looks good to me. Nice to also see a `Float16` IR test :-)

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

Marked as reviewed by chagedorn (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/24150#pullrequestreview-2705470867