RISC-V Vector Instruction vssub.vv not saturating

[Tony Cole]

I noticed doing a negate ( 0 – 0x80000000 ) using vssub.vv produces an incorrect result of 0x80000000 (should saturate to 0x7FFFFFFF).

Here is the bit of the code:

  vmv.v.i v16, 0
  …
8f040457 vssub.vv v8,v16,v8

I believe the instruction encoding is correct (vssub.vv with vd = v8, vs2 = v16, rs1 = v8), but the result does not saturate in QEMU.

I’ve just tested with what I think is the latest branch ( https://github.com/sifive/qemu/tree/rvv-1.0-upstream-v7 commit 26 Feb 2021: 1151361fa7d45cc90d69086ccf1a4d8397931811 ) and the problem still exists.

[Alistair Francis (alistair2323)]

Thanks for raising this bug case. A fix should be available soon.

[Frank Chang (frankchang0125)]

This should be a quick fix, we will run couple tests again to ensure the fix doesn't break anything. Thanks~

[Alistair Francis (alistair2323)]

This was fixed by commit:

commit 65606f21243a796537bfe4708720a9bf4bb50169
Author: LIU Zhiwei <email address hidden>
Date: Fri Feb 12 23:02:21 2021 +0800

    target/riscv: Fixup saturate subtract function

    The overflow predication ((a - b) ^ a) & (a ^ b) & INT64_MIN is right.
    However, when the predication is ture and a is 0, it should return maximum.

    Signed-off-by: LIU Zhiwei <email address hidden>
    Reviewed-by: Richard Henderson <email address hidden>
    Reviewed-by: Alistair Francis <email address hidden>
    Message-id: <email address hidden>
    Signed-off-by: Alistair Francis <email address hidden>

The fix will be in the QEMU 6.1 release.