FPU: prevent a coprocessor divide/sqrt from corrupting scalar FP registers

[emiliengnr]

Hello, here is a pull request for a bug I found.

In src/main/scala/tile/FPU.scala the shared divide/sqrt writeback selects its address and
data with divSqrt_wen and uses them for both the scalar register file and the coprocessor
response, without checking whether the completion belongs to the coprocessor. A coprocessor
divide/sqrt result can therefore land in a scalar FP register, either when no scalar write
is pending or when a scalar pipelined result writes back on the same cycle and loses the
shared address and data. The fix gives the scalar write separate sources selected only for a
scalar divide/sqrt (divSqrt_wen && !divSqrt_cp) and leaves io.cp_resp unchanged, matching
the !wbInfo(0).cp guard the pipelined path already uses. It is reachable on any core that
attaches a coprocessor on the shared divide/sqrt path, and was found with the Saturn vector
unit as Rocket's coprocessor.

Related issue: none

Type of change: bug report

Impact: no functional change

Development Phase: implementation

Release Notes
Fixed the shared FPU divide/sqrt writeback so a coprocessor divide/sqrt result no longer corrupts a scalar floating-point register.

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: emiliengnr / name: Emilien Garnier (edc8cfa)