[zkvm] Precompiles refactor

[dhil]

No description provided.

[dhil]

The diff is kind of pointless here as the target has diverted from main. I'll retarget to main instead, though, I am not seeking to get this patch included in main at the moment.

[dhil]

@goodlyrottenapple The actual patch starts on commit fa1e9a5. Each precompile refactor has its own commit, so it is probably best reviewed on a commit-by-commit basis.

(the base of this patch is #2090)

[goodlyrottenapple]

Might it be worth rolling most of the new *_execute functions into a templated generic one?

enum class PrecompileImplOpts
{
    FreeOnNull,
    SuccessOnNull,
};

template <typename>
struct impl_out_size;

template <size_t N>
struct impl_out_size<
    PrecompileImplResult (*)(byte_string_view, std::span<uint8_t, N>)>
{
    static constexpr size_t value = N;
};

template <auto Impl, PrecompileImplOpts... Opts>
static PrecompileResult alloc_and_execute(byte_string_view const input)
{
    constexpr bool free_on_null =
        ((Opts == PrecompileImplOpts::FreeOnNull) || ...);
    constexpr bool success_on_null =
        ((Opts == PrecompileImplOpts::SuccessOnNull) || ...);

    constexpr size_t N = impl_out_size<decltype(Impl)>::value;
    auto *const out = static_cast<uint8_t *>(std::malloc(N));
    MONAD_ASSERT(out != nullptr);
    auto const result = Impl(input, std::span<uint8_t, N>{out, N});

    if (result.data == nullptr) {
        MONAD_DEBUG_ASSERT(result.size == 0);
        if constexpr (free_on_null) {
            std::free(out);
        }
        if constexpr (success_on_null) {
            return {EVMC_SUCCESS, nullptr, 0};
        }
        return PrecompileResult::failure();
    }
    return {EVMC_SUCCESS, result.data, result.size};
}

then in cases, have e.g.:

CASE(0x02, sha256_gas_cost, alloc_and_execute<sha256_impl>);
...
CASE(
    0x09, blake2bf_gas_cost<traits>,
    alloc_and_execute<blake2bf_impl, PrecompileImplOpts::FreeOnNull>);
...
CASE(
    0x0100, p256_verify_gas_cost,
    alloc_and_execute<
        p256_verify_impl, PrecompileImplOpts::FreeOnNull,
        PrecompileImplOpts::SuccessOnNull>
...

[dhil]

Might it be worth rolling most of the new *_execute functions into a templated generic one?

enum class PrecompileImplOpts
{
    FreeOnNull,
    SuccessOnNull,
};

template <typename>
struct impl_out_size;

template <size_t N>
struct impl_out_size<
    PrecompileImplResult (*)(byte_string_view, std::span<uint8_t, N>)>
{
    static constexpr size_t value = N;
};

template <auto Impl, PrecompileImplOpts... Opts>
static PrecompileResult alloc_and_execute(byte_string_view const input)
{
    constexpr bool free_on_null =
        ((Opts == PrecompileImplOpts::FreeOnNull) || ...);
    constexpr bool success_on_null =
        ((Opts == PrecompileImplOpts::SuccessOnNull) || ...);

    constexpr size_t N = impl_out_size<decltype(Impl)>::value;
    auto *const out = static_cast<uint8_t *>(std::malloc(N));
    MONAD_ASSERT(out != nullptr);
    auto const result = Impl(input, std::span<uint8_t, N>{out, N});

    if (result.data == nullptr) {
        MONAD_DEBUG_ASSERT(result.size == 0);
        if constexpr (free_on_null) {
            std::free(out);
        }
        if constexpr (success_on_null) {
            return {EVMC_SUCCESS, nullptr, 0};
        }
        return PrecompileResult::failure();
    }
    return {EVMC_SUCCESS, result.data, result.size};
}

then in cases, have e.g.:

CASE(0x02, sha256_gas_cost, alloc_and_execute<sha256_impl>);
...
CASE(
    0x09, blake2bf_gas_cost<traits>,
    alloc_and_execute<blake2bf_impl, PrecompileImplOpts::FreeOnNull>);
...
CASE(
    0x0100, p256_verify_gas_cost,
    alloc_and_execute<
        p256_verify_impl, PrecompileImplOpts::FreeOnNull,
        PrecompileImplOpts::SuccessOnNull>
...

Yes. I think this might be a good way to cut down the boilerplate. I will look into this.

[mkolosick]

I dug more into the precompiles I hadn't looked at yet, and I think that template generic'ing is actually premature here. I think there are more opportunities to follow the pattern in ecrecover_execute of pulling the common error handling/parsing up into the _execute layer once we've removed silkpre. I started doing this for the other precompiles but too many of them are tied up to internal silkpre implementation at the moment, so I pulled back on that for now.

In light of that I'm just cleaning up the commits and rebasing my other changes and then I would be good with the commit.