In the following specification the term XLEN refers to the width of an x register in bits, as described in the RISC-V ISA specification. The term NRET refers to the maximum number of instructions that the core under test can retire in one cycle. If more than one of the retired instruction writes the same register, the channel with the highest index contains the instruction that wins the conflict. The term ILEN refers to the maximum instruction width for the processor under test.
The Interface consists only of output signals. Each signal is a concatenation of NRET values of constant width, effectively creating NRET channels. For simplicity, the following descriptions refer to one such channel. For example, we refer to rvfi_valid as a 1-bit signal, not a NRET-bits signal.
output [NRET - 1 : 0] rvfi_valid
output [NRET * 64 - 1 : 0] rvfi_order
output [NRET * ILEN - 1 : 0] rvfi_insn
output [NRET - 1 : 0] rvfi_trap
output [NRET - 1 : 0] rvfi_halt
output [NRET - 1 : 0] rvfi_intr
output [NRET * 2 - 1 : 0] rvfi_mode
output [NRET * 2 - 1 : 0] rvfi_ixl
When the core retires an instruction, it asserts the rvfi_valid signal and uses the signals described below to output the details of the retired instruction. The signals below are only valid during such a cycle and can be driven to arbitrary values in a cycle in which rvfi_valid is not asserted.
The rvfi_order field must be set to the instruction index. No indices must be used twice and there must be no gaps. Instructions may be retired in a reordered fashion, as long as causality is preserved (register and memory write operations must be retired before the read operations that depend on them).
rvfi_insn is the instruction word for the retired instruction. In case of an instruction with fewer than ILEN bits, the upper bits of this output must be all zero. For compressed instructions the compressed instruction word must be output on this port. For fused instructions the complete fused instruction sequence must be output.
rvfi_trap must be set for an instruction that cannot be decoded as a legal instruction, such as 0x00000000.
In addition, rvfi_trap must be set for a misaligned memory read or write in PMAs that don't allow misaligned access, or other memory access violations. rvfi_trap must also be set for a jump instruction that jumps to a misaligned instruction.
The signal rvfi_halt must be set when the instruction is the last instruction that the core retires before halting execution. It should not be set for an instruction that triggers a trap condition if the CPU reacts to the trap by executing a trap handler. This signal enables verification of liveness properties.
rvfi_intr must be set for the first instruction that is part of a trap handler, i.e. an instruction that has a rvfi_pc_rdata that does not match the rvfi_pc_wdata of the previous instruction.
rvfi_mode must be set to the current privilege level, using the following encoding: 0=U-Mode, 1=S-Mode, 2=Reserved, 3=M-Mode
Finally rvfi_ixl must be set to the value of MXL/SXL/UXL in the current privilege level, using the following encoding: 1=32, 2=64
output [NRET * 5 - 1 : 0] rvfi_rs1_addr
output [NRET * 5 - 1 : 0] rvfi_rs2_addr
output [NRET * XLEN - 1 : 0] rvfi_rs1_rdata
output [NRET * XLEN - 1 : 0] rvfi_rs2_rdata
rvfi_rs1_addr and rvfi_rs2_addr are the decoded rs1 and rs1 register addresses for the retired instruction. For an instruction that reads no rs1/rs2 register, this output can have an arbitrary value. However, if this output is nonzero then rvfi_rs1_rdata must carry the value stored in that register in the pre-state.
rvfi_rs1_rdata/rvfi_rs2_rdata is the value of the x register addressed by rs1/rs2 before execution of this instruction. This output must be zero when rs1/rs2 is zero.
output [NRET * 5 - 1 : 0] rvfi_rd_addr
output [NRET * XLEN - 1 : 0] rvfi_rd_wdata
rvfi_rd_addr is the decoded rd register address for the retired instruction. For an instruction that writes no rd register, this output must always be zero.
rvfi_rd_wdata is the value of the x register addressed by rd after execution of this instruction. This output must be zero when rd is zero.
output [NRET * XLEN - 1 : 0] rvfi_pc_rdata
output [NRET * XLEN - 1 : 0] rvfi_pc_wdata
This is the program counter (pc) before (rvfi_pc_rdata) and after (rvfi_pc_wdata) execution of this instruction. I.e. this is the address of the retired instruction and the address of the next instruction.
output [NRET * XLEN - 1 : 0] rvfi_mem_addr
output [NRET * XLEN/8 - 1 : 0] rvfi_mem_rmask
output [NRET * XLEN/8 - 1 : 0] rvfi_mem_wmask
output [NRET * XLEN - 1 : 0] rvfi_mem_rdata
output [NRET * XLEN - 1 : 0] rvfi_mem_wdata
For memory operations (rvfi_mem_rmask and/or rvfi_mem_wmask are non-zero), rvfi_mem_addr holds the accessed memory location.
When the define RISCV_FORMAL_ALIGNED_MEM is set, the address must have a 4-byte alignment for XLEN=32 and an 8-byte alignment for XLEN=64. When the define is not set, then the address must point directly to the LSB or the word / half word / byte that is accessed.
rvfi_mem_rmask is a bitmask that specifies which bytes in rvfi_mem_rdata contain valid read data from rvfi_mem_addr.
rvfi_mem_wmask is a bitmask that specifies which bytes in rvfi_mem_wdata contain valid data that is written to rvfi_mem_addr.
rvfi_mem_rdata is the pre-state data read from rvfi_mem_addr. rvfi_mem_rmask specifies which bytes are valid.
rvfi_mem_wdata is the post-state data written to rvfi_mem_addr. rvfi_mem_wmask specifies which bytes are valid.
When RISCV_FORMAL_ALIGNED_MEM is set then riscv-formal assumes that unaligned memory access causes a trap.
Some arithmetic operations (such as multiplication and division) are beyond to practical capabilities of even modern hardware model checkers. In order to still be able to verify things like bypassing for the arithmetic units performing those operations we define a set of alternative arithmetic operations. When the define RISCV_FORMAL_ALTOPS is set riscv-formal will expect the processor under test to implement those alternative operations instead.
Commutative operations (like multiplication) are replaced with addition followed by applying XOR with a bitmask that indicates the type of the operation. Noncommutative operations (like division) are replaced with subtraction followed by applying XOR with a bitmask that indicates the type of the operation.
The bitmasks are 64 bits wide. RV32 implementations only use the lower 32 bits of the bitmasks. The *W instructions in RV64 (suchg as MULW) are implemented by adding or subtracting the lower 32 bits of the operands, then XORing with the lower 32 bits of the bitmask, then sign extending the result to 64 bits.
| Operation | Add/Sub | Bitmask |
|---|---|---|
| MUL | Add | 0x2cdf52a55876063e |
| MULH | Add | 0x15d01651f6583fb7 |
| MULHSU | Sub | 0xea3969edecfbe137 |
| MULHU | Add | 0xd13db50d949ce5e8 |
| DIV | Sub | 0x29bbf66f7f8529ec |
| DIVU | Sub | 0x8c629acb10e8fd70 |
| REM | Sub | 0xf5b7d8538da68fa5 |
| REMU | Sub | 0xbc4402413138d0e1 |
For each supported CSR there are four additional output ports:
output [NRET * XLEN - 1 : 0] rvfi_csr_<csrname>_rmask
output [NRET * XLEN - 1 : 0] rvfi_csr_<csrname>_wmask
output [NRET * XLEN - 1 : 0] rvfi_csr_<csrname>_rdata
output [NRET * XLEN - 1 : 0] rvfi_csr_<csrname>_wdata
The rmask and wmask ports specify which bits of rdata and wdata are valid.
It is always valid for an instruction to activate more rmask/rdata bits
than required by the instruction, as long as the reported bits correctly reflect
the machine state.
If reading a CSR has side effects, those side effects are not triggered by raised
rmask bits but by the type of the instruction.
The Verilog define RISCV_FORMAL_CSR_<CSRNAME> must be set for each CSR traced
via RVFI by the core under test.
See RISC-V Formal CSR Sematics for the exact semantics of CSR values output via RVFI.
Out-of-order cores that execute speculatively can commit speculative instructions on RVFI.
Rollbacks must be output via the rollback interface, that is enabled when RISCV_FORMAL_ROLLBACK is defined:
output [ 0 : 0] rvfi_rollback_valid
output [63 : 0] rvfi_rollback_order
All RVFI packets output prior to the cycle with asserted rvfi_rollback_valid with a rvfi_order field of greater or equal to rvfi_rollback_order are invalidated by a rollback event.
RVFI packets output in the same cycle as rvfi_rollback_valid are already part of the new instruction stream re-starting at the instruction number indicated in rvfi_rollback_order.
The following section contains notes on future extensions to RVFI. They will come part of the spec as soon as there is at least one core that implements the feature, and a matching formal check that utilises the feature. In many cases the additional ports will only be used (and expected from the core) when additional to-be-defined RISCV_FORMAL_* Verilog defines are set.
Fused instructions are simply handled as larger instructions in RVFI. Additional rvfi_rs* ports (or even rvfi_rd* ports) may be added to accommodate the fused instructions.
No instruction models for fused instructions have been created yet.
Alternatively fused instructions may be output as individual instructions in separate RVFI channels.
The following is the proposed RVFI extension for floating point ISAs:
output [NRET * 5 - 1 : 0] rvfi_frs1_addr
output [NRET * 5 - 1 : 0] rvfi_frs2_addr
output [NRET * 5 - 1 : 0] rvfi_frs3_addr
output [NRET * 5 - 1 : 0] rvfi_frd_addr
output [NRET - 1 : 0] rvfi_frs1_rvalid
output [NRET - 1 : 0] rvfi_frs2_rvalid
output [NRET - 1 : 0] rvfi_frs3_rvalid
output [NRET - 1 : 0] rvfi_frd_wvalid
output [NRET * FLEN - 1 : 0] rvfi_frs1_rdata
output [NRET * FLEN - 1 : 0] rvfi_frs2_rdata
output [NRET * FLEN - 1 : 0] rvfi_frs3_rdata
output [NRET * FLEN - 1 : 0] rvfi_frd_wdata
output [NRET * XLEN - 1 : 0] rvfi_csr_fcsr_rmask
output [NRET * XLEN - 1 : 0] rvfi_csr_fcsr_wmask
output [NRET * XLEN - 1 : 0] rvfi_csr_fcsr_rdata
output [NRET * XLEN - 1 : 0] rvfi_csr_fcsr_wdata
Since f0 is not a zero register, additional *_[rw]valid signals are required to indicate if frs1, frs2, frs3, and frd and their corresponding pre- or post-values are valid.
Alternative arithmetic operations (RISCV_FORMAL_ALTOPS) will be defined for all non-trivial floating point operations.
For processors with support for S-mode and virtual memory we define the following additional RVFI signals for data load/stores:
output [NRET * 64 - 1 : 0] rvfi_mem_paddr
output [NRET * XLEN - 1 : 0] rvfi_mem_pte0
output [NRET * XLEN - 1 : 0] rvfi_mem_pte1
output [NRET * XLEN - 1 : 0] rvfi_mem_pte2
output [NRET * XLEN - 1 : 0] rvfi_mem_pte3
And the following additional RVFI signals for instruction fetches:
output [NRET * 64 - 1 : 0] rvfi_pc_paddr
output [NRET * XLEN - 1 : 0] rvfi_pc_pte0
output [NRET * XLEN - 1 : 0] rvfi_pc_pte1
output [NRET * XLEN - 1 : 0] rvfi_pc_pte2
output [NRET * XLEN - 1 : 0] rvfi_pc_pte3
And we require that the satp CSR is observable through RVFI:
output [NRET * XLEN - 1 : 0] rvfi_csr_satp_rmask
output [NRET * XLEN - 1 : 0] rvfi_csr_satp_wmask
output [NRET * XLEN - 1 : 0] rvfi_csr_satp_rdata
output [NRET * XLEN - 1 : 0] rvfi_csr_satp_wdata
The rvfi_mem_paddr field carries the physical address of the memory access. The rvfi_mem_pte[0123] fields carry the values of the page table entries used to convert rvfi_mem_addr to rvfi_mem_paddr. Unused rvfi_mem_pte[0123] fields must always be set to zero.
For memory accesses in M-mode, or with satp.MODE=0, rvfi_mem_paddr must have the same value as rvfi_mem_addr and all four rvfi_mem_pte[0123] fields must be set to zero.
For example in Sv32 mode, modulo missing fences, rvfi_mem_pte1 must carry the value of the 32-bit word at the following memory location:
pt1 = rvfi_csr_satp_rdata & 0x003fffff
vpn1 = (rvfi_mem_addr >> 22) & 0x3ff
pte1_addr = (pt1 << 12) | (vpn1 << 2)
And rvfi_mem_pte0 must carry the value of the 32-bit word at the following memory location (or zero if pte1.X or pte1.R or !pte1.V):
pt0 = rvfi_mem_pte1 >> 10
vpn0 = (rvfi_mem_addr >> 12) & 0x3ff
pte0_addr = (pt0 << 12) | (vpn0 << 2)
Finally, rvfi_mem_paddr must be set to the following address:
ppn = rvfi_mem_pte0 >> 10
offset = rvfi_mem_addr & 0xfff
rvfi_mem_paddr = (ppn << 12) | offset
AMO instructions (AMOSWAP.W, etc.) can be modelled using the existing rvfi_mem_* interface by asserting bits in both rvfi_mem_rmask and rvfi_mem_wmask.
There is also no extension to the RVFI port necessary to accommodate the LR, SC, FENCE and FENCE.I instructions.
Verification of this instructions for a single-core systems can be done using the RVFI port only. A strategy must be defined to verify their correct behavior in multicore systems.
For atomic instructions with rd = x0 a core might have no way of knowing the old or new value of the memory location. For those situations we add an additional RVFI output port:
output [NRET - 1 : 0] rvfi_mem_extamo
When rvfi_mem_extamo is set, rvfi_mem_wdata carries the rs2 value used with the atomic instruction instead of the new value in the memory location. rvfi_mem_rmask is all-zeros in this case.
Consider the following sequence of instructions:
....
add t0,t1,t2
beqz t3,label
sub t0,t1,t3
label:
....
When t3 has a non-zero value the processor could decide not to schedule the add instruction because its value is never going to be used. In this case the processor would be unable to produce a valid RVFI trace for the instruction sequence.
An additional signal can be added to RVFI that can be used to mark such instructions:
output [NRET - 1 : 0] rvfi_skip
When rvfi_skip is high the core may output arbitrary data on the *_rdata and *_wdata ports (excluding rvfi_pc_rdata and rvfi_pc_wdata). The register values written by such intrustions may only be observed by other skipped instructions. An additional formal proof must be added to check this property.
Memory operations (rvfi_mem_rmask and/or rvfi_mem_wmask are non-zero) can not be skipped.