RFR: 8365911: AArch64: Fix encoding error in sve_cpy for negative floats

[Andrew Haley]

On Wed, 27 Aug 2025 01:34:25 GMT, erifan <duke at openjdk.org> wrote:

> The sve_cpy instruction is not correctly implemented for negative floating-point values. The issues include:
> 
> 1. When a negative floating-point number (e.g. `-1.0`) is passed, the `checked_cast<int8_t>(pack(d))` check fails. For example, assume `d = -1.0`:
> - `pack(-1.0)` returns an unsigned int with the 7th bit set, i.e., `0xf0`.
> - `checked_cast<int8_t>(0xf0)` casts `0xf0` to an int8_t value, which is `-16`.
> - Casting this int8_t `-16` back to unsigned int results in `0xfffffff0`.
> - The check compares `0xf0` to `0xfffffff0`, which obviously fails.
> 
> 2. Additionally, the encoding of the negative floating-point number is incorrect:
> - The imm8 field can fall outside the valid range of **[-128, 127]**.
> - Bit **13** should be encoded as **0** for floating-point numbers.
> 
> This PR fixes these issues and renames floating-point `sve_cpy` as `sve_fcpy`.
> 
> Some test cases are added to aarch64-asmtest.py, and all tests passed.

Thanks.

I'm not convinced that the refactoring is necessary. Why not write a replacement for `checked_cast<int8_t>(pack(d))` that does the right thing and fix the first `sve_cpy()` so that it does the right thing for float args?

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

PR Comment: https://git.openjdk.org/jdk/pull/26951#issuecomment-3242601638