VMFPU mask flag desynchronized from FPU output

[ianfield]

FP comparison instructions (vmfge, vmfeq, vmflt, etc.) produce results that should be routed to the mask unit. Under pipeline pressure from filler instructions, the routing flag reads the wrong instruction and the comparison result is written to the VRF instead of the mask unit.

Reproducer

vsetivli zero, 4, e32, m1, ta, mu
vle32.v  v4, (a1)
vle32.v  v6, (a2)
vfadd.vv v24, v24, v26
vmfge.vf v0, v6, fa5
vfadd.vv v4, v6, v6, v0.t
vse32.v  v4, (a3)

Steps to Reproduce

git clone https://github.com/ianfield/ara-test.git
cd ara-test/
export LLVM_RISCV=/opt/homebrew/Cellar/llvm/21.1.7/bin/
./sh/checkout.sh
./sh/build.sh
./sh/run.sh --sim spike -k test_vfadd_hazard
./sh/run.sh --sim fork -k test_vfadd_hazard
./sh/run.sh --sim base -k test_vfadd_hazard

Proposed change

vmfpu.sv:211

case (op) inside
  VFDIV, VFRDIV, VFSQRT:  fpu_latency = LatFDivSqrt;
  [VFREDMIN:VFREDMAX]:    fpu_latency = LatFNonComp;
  [VFCVTXUF:VFCVTFF]:     fpu_latency = LatFConv;
  [VFMIN:VFSGNJX]:        fpu_latency = LatFNonComp;
  [VMFEQ:VMFGE]:          fpu_latency = LatFNonComp;
  default: begin