[pmp]: MML incorrectly suppresses `pmpcfg` writes for OFF and invalid TOR entries

[KnightGOKU]

Observed Behavior

With mseccfg.MML=1 and mseccfg.RLB=0, Ibex suppresses pmpcfg writes that set L=1 and X=1 even when the resulting entry does not create a valid executable PMP rule.

The default test writes pmpcfg0=0x84, which requests L=1, A=OFF, and X=1. The write is ignored and the test reads back:

observed: 0x00008c00
expected: 0x00008c84

The G=1 variant writes L=1, X=1, and A=NA4. Ibex first converts NA4 to OFF, then suppresses the write in the same way.

The invalid-TOR variant writes an entry with A=TOR, L=1, and X=1, but with pmpaddr[i-1] >= pmpaddr[i]. The test reads back:

observed: 0x000e009c
expected: 0x000e8c9c

Expected Behavior

Ibex should accept these writes because A=OFF does not create a PMP rule, and a TOR entry with pmpaddr[i-1] >= pmpaddr[i] matches no addresses.

The MML restriction should suppress only a write that adds a valid M-mode-only or locked Shared-Region rule with executable privileges. The L bit should remain writable for an OFF entry.

Steps to Reproduce

1. Run ./evidence/run_ibex_core_pmp_mml_off_entry.sh.
2. The assembly enables mseccfg.MML, writes pmpcfg0, reads it back, and stores the result.
3. Repeat with ./evidence/run_ibex_core_pmp_mml_off_entry.sh program_g1_invalid_na4.S.
4. Repeat with ./evidence/run_ibex_core_pmp_mml_off_entry.sh program_invalid_tor.S.
5. Observe that all three writes are suppressed and the readback differs from the requested configuration.

The test intentionally exits through $fatal when the observed readback does not match the expected value.

program.zip

Testbench

evidence/ibex_core_pmp_mml_off_entry_tb.sv

Environment

Ibex source revision: 57d37bf9008640415532d23ad59018b795c642aa.

The test uses Verilator and the real ibex_core with PMPEnable=1. The default case uses PMPGranularity=0; the NA4 variant uses PMPGranularity=1.

The expected rule definition follows the RISC-V Smepmp specification.