Analyze rv32emu

contributed by <weishiuan>

Goal

Building on sammer1077's contributions, the objective is to ensure rv32emu-next successfully passes the riscv-arch-test. This involves:

1. Identifying and solving any technical problems encountered during the process and submitting pull requests accordingly.
2. Providing runtime information for rv32emu, including CPU cycle, branch count, and instruction count.
3. Identifying the top 10 most frequently used instructions during runtime.
4. Optionally integrating RV32C into rv32emu-next based on either PR#3 or PR#4.

Arch-Test

The RISC-V architecture test ensures a CPU model meets the RISC-V specification:

1. Includes the Reference signature.
2. Runs the test-specific program, overwriting data in specific parts called test-signature.
3. Compares test-signature with the Reference signature. If they are exactly equal, the compliance test is passed.

After compiling and running tests, find the ELF files in riscv-arch-test/work/C directories. Use riscv-none-embed-objdump -D to get more information.

Test Points

Compliance Test checks specific instructions for functionality and decode correctness. Ensure correctness of results, x0 cannot be modified, and decode correctness. Any difference in the signature leads to test failure.

Standard Instruction for Compressed Instruction

Compressed instructions are shorter in special cases:

• Small immediate/address offset.
• One of the registers is x0, x1, x2.
• Destination register equals source register.
• Using 8 most popular registers.

The C-extension is compatible with other standard instructions and is interleaved with the 16-bit boundary.

Progress on Pull Request (PR#3)

Failed tests:

• c.lui
• c.srli
• c.srai
• c.andi
• c.jalr
• c.jal
• c.ebreak
• c.addi16sp

These tests are not supported by Compute Goto implementation.

RV32C Compatibility with Other Instruction Sets

Any instruction fetched by rv32emu-next can be checked for the last 2 bits:

• If the last 2 bits are 3 (11), it's a standard uncompressed instruction, and the next PC should be PC+4.
• Otherwise, it's a compressed instruction, and the next PC should be PC+2.

Implementation Example

while (rv->csr_cycle < cycles_target && !rv->halt) {
    // Fetch the next instruction
    inst = rv->io.mem_ifetch(rv, rv->PC);

    if ((inst & 3) == 3) {
        // Standard uncompressed instruction
        uint32_t index = (inst & INST_6_2) >> 2;
        TABLE_TYPE op = jump_table[index];
        assert(op);

        rv->inst_len = INST_32;

        // Dispatch and execute the opcode
        if (!op(rv, inst))
            break;

        // Increment the cycles csr
        rv->csr_cycle++;
    } else {
        // Compressed instruction
        const uint16_t c_index = ((inst & FR_C_15_13) >> 11) | (inst & FR_C_1_0);
        const c_opcode_t op = c_opcodes[c_index];
        assert(op);

        rv->inst_len = INST_16;

        // Dispatch and execute the compressed opcode
        if (!op(rv, inst))
            break;

        // Increment the cycles csr
        rv->csr_cycle++;
    }
}

Debug with Compliance Test

After running the compliance test, check and record each failed instruction. Go to the work directories in riscv-arch-tests and find the corresponding .elf and .diff files.

1. .diff illustrates the differences in test signature and reference signature.
   .elf can be opened with riscv-none-embed-objdump -D to get the instruction of the test code.

2. Run ./rv32emu –trace FailInstructionTest.elf to trace the runtime information. Check each line, especially the test instruction. Manually calculate and compare the result to your printed message. If something is wrong, find the position of the error code based on the conclusion.

Note that some instruction tests will use another instruction; ensure correctness in other instruction usage.

Example - Fixing Arithmetic Error in c.lui

c.lui

1. c.lui loads the non-zero 6-bit immediate field into bits 17-12 of the destination register, clears the bottom 12 bits, and sign-extends bit 17 into all higher bits of the destination.
2. The code points with nzimm=0 are reserved.
3. The remaining code points with rd=x0 are HINTS.

if (rd == 2) {
        // C.ADDI16SP
        uint32_t tmp = (inst & 0x1000) >> 3 | (inst & 0x40) >> 2 | (inst & 0x20) << 1 |
                       (inst & 0x18) << 4 | (inst & 0x4) << 3;

        const uint32_t imm = (tmp & 0x200) ? (0xfffffc00 | tmp) : tmp;

        if (imm != 0) {
            rv->X[rd] += imm;
        } else {
            // Handle reserved parts (add-on)
        }
    } else if (rd != 0) {
        // C.LUI
        uint32_t tmp = (inst & 0x1000) << 5 | (inst & 0x7c) << 10;
        const int32_t imm = (tmp & 0x20000) ? (0xfffc0000 | tmp) : tmp;

        if (imm != 0) {
            rv->X[rd] = imm;
        } else {
            // Handle reserved parts
        }
    } else {
        // HINTS
    }

    rv->PC += rv->inst_len;
    return true;
}

c.jalr

c.jalr expands to jalr x1, 0(rs1). Special part in the test code:

inst_8:
    auipc ra, 0x0
    addi ra, ra, 16        # 80000240 <inst_8+0x10>
    jalr ra
    xori ra, ra, 2
    j 80000244             # Jump to inst_8+0x14
    xori ra, ra, 3
    auipc a1, 0x0
    addi a1, a1, -20        # 80000230 <inst_8>
    andi a1, a1, -4
    sub ra, ra, a1
    sw ra, 32(a0)

Error Parts

Directly save Next PC to ra and jump to the instruction line to x's value:

// C.JALR
const uint32_t next_pc = rv->PC + 2;
rv->X[1] = next_pc;
rv->PC = rv->X[rs1];

if (rv->PC & 1) {
    rv_except_inst_misaligned(rv, rv->PC);
    return false;
}

// Branch is possible
return false;

Corrected Parts

Use a variable (rs_value) to store the value in register x, save the Next PC to ra, and jump to rs_value line:

// C.JALR
const int32_t rs_value = rv->X[rs1];
const uint32_t next_pc = rv->PC + rv->inst_len;
rv->X[rv_reg_ra] = next_pc;
rv->PC = rs_value;

if (rv->PC & 0x1) {
    rv_except_inst_misaligned(rv, rv->PC);
    return false;
}

// Branch is possible
return false;

Optimizing Control Flow with Compute Goto in RV32C Integration

Compute Goto in rv32emu-next

In the rv32emu-next project, Compute Goto is utilized after all instruction calls, employing a unified DISPATH process. Despite the absence of reused code, the use of defined macros streamlines program maintenance. The implementation involves defining the macro op_xxxx to ensure all instruction handlers have unique jumps, which are not shared.

#define TARGET(instr)         \
    op_##instr : EXEC(instr); \
    DISPATCH();

This approach significantly improves the speed of rv32emu-next by enhancing the branch prediction of the program itself, rather than the RISC-V model's branch prediction.

Integration of Compute Goto with RV32C

To integrate Compute Goto with RV32C, several components are prepared:

Jump table

A table storing pointers to functions is created to facilitate efficient dispatching.

typedef struct {
    TABLE_TYPE_RVC entries[4][4];
} JumpTableRVC;

JumpTableRVC jump_table_rvc = {
    {
        {OP(caddi4spn), OP(caddi), OP(cslli), OP(unimp)},   // 00
        {OP(cfld), OP(cjal), OP(cfldsp), OP(unimp)},         // 01
        {OP(clw), OP(cli), OP(clwsp), OP(unimp)},            // 10
        {OP(cflw), OP(clui), OP(cflwsp), OP(unimp)}          // 11
    },
    {
        {OP(unimp), OP(cmisc_alu), OP(ccr), OP(unimp)},      // 00
        {OP(cfsd), OP(cj), OP(cfsdsp), OP(unimp)},           // 01
        {OP(csw), OP(cbeqz), OP(cswsp), OP(unimp)},          // 10
        {OP(cfsw), OP(cbnez), OP(cfswsp), OP(unimp)}         // 11
    }
};

DISPATCH

This component checks whether the CPU has halted or if the trace cycle meets the target. It then fetches the instruction and determines whether it is compressed or uncompressed, adjusting the instruction length accordingly.

#define DISPATCH()                                                          \
    {                                                                       \
        if (rv->csr_cycle >= cycles_target || rv->halt)                     \
            goto quit;                                                      \
        /* fetch the next instruction */                                    \
        inst = rv->io.mem_ifetch(rv, rv->PC);                               \
        /* standard uncompressed instruction */                             \
        if ((inst & 3) == 3) {                                              \
            uint32_t index = (inst & INST_6_2) >> 2;                        \
            rv->inst_len = INST_32;                                         \
            handle_uncompressed_instruction(rv, inst, jump_table);         \
        } else {                                                            \
            /* Compressed Extension Instruction */                          \
            inst &= 0x0000FFFF;                                             \
            int16_t c_index = (inst & FC_FUNC3) >> 11 | (inst & FC_OPCODE); \
            rv->inst_len = INST_16;                                         \
            handle_compressed_instruction(rv, inst, jump_table_rvc);            \
        }                                                                   \
    }

EXEC

The EXEC component remains unchanged. It dispatches the opcode and increments the cycle counter.

#define EXEC(instr)                  \
    do {                             \
        if (!execute_opcode(rv, instr, inst))  \
            goto quit;               \
        rv->csr_cycle++;             \
    } while (0)

TARGET

This part ensures instruction fetching after handling the instruction.

#ifdef ENABLE_RV32C
    EXEC(caddi4spn);
    EXEC(caddi);
    EXEC(cslli);
    EXEC(cjal);
    EXEC(clw);
    EXEC(cli);
    EXEC(clwsp);
    EXEC(clui);
    EXEC(cmisc_alu);
    EXEC(ccr);
    EXEC(cj);
    EXEC(csw);
    EXEC(cbeqz);
    EXEC(cswsp);
    EXEC(cbnez);
#endif

Pull Request

The project is currently in a pull request stage. Extensive modifications are planned in the commit messages based on the reviewer's advice. The team is committed to delivering the best results by the end of the project.

Conclusion

This project focuses on integrating RV32C, contributed by ccs100203 and Uduru0522, with Compute Goto in rv32emu-next. RV32C, a part of the Standard extension for compressed instructions, optimizes instruction length based on specific conditions. Compliance test considerations include ensuring correct instruction functionality and checking for any "reserved" or "HINT" decode results. The Compute Goto technique enhances branch prediction accuracy by maintaining independently predicted jumps.