Skip to content

Bug with PNCLIPUP.W instruction #2350

Description

@mattiamirigaldi

PNCLIPUP.W: Packed Narrowing Unsigned Clip (64-bit to 32-bit).
Spike casts the 64-bit source registers to a signed 64-bit integer (sreg_t) before evaluating saturation. Because of this, any 64-bit input with the 63rd bit set (MSB = 1) is incorrectly evaluated as a negative number rather than as an unsigned number. This leads Spike to incorrectly saturate the result to 0x00000000 instead of the expected unsigned maximum 0xFFFFFFFF.

Example:

  • Input: rs1 = 0x8ed30e0630d20a98, rs2 = 0x83b9d1d6ae9ffccb

  • Operation:

  • Lane 0 (rs1): 0x8ed30e0630d20a98 >_u 0x00000000FFFFFFFF (True, must saturate to 0xFFFFFFFF)

  • Lane 1 (rs2): 0x83b9d1d6ae9ffccb >_u 0x00000000FFFFFFFF (True, must saturate to 0xFFFFFFFF)

  • Expected result: 0xFFFFFFFFFFFFFFFF

  • Spike result: 0x0000000000000000

Possible Cause: In the header file, s1 and s2 are explicitly cast to (sreg_t). When an input has the highest bit set to 1, the C++ cast interprets it as a negative value (< 0). The P_USAT_FULL macro detects a value less than zero and incorrectly clamps the final output to 0.

Possible Fix: Remove the (sreg_t) cast, or updated the clamping logic to explicitly handle 64-bit unsigned comparisons.


Spike hash : b2d6fbf

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions