ultra_circuit_builder_nonnative.test.cpp

Go to the documentation of this file.

#include "barretenberg/circuit_checker/circuit_checker.hpp"
#include "barretenberg/numeric/uint256/uint256.hpp"
#include "barretenberg/stdlib_circuit_builders/ultra_circuit_builder.hpp"
 
#include <gtest/gtest.h>
 
using namespace bb;
 
class UltraCircuitBuilderNonNative : public ::testing::Test {
  protected:
    static constexpr size_t LIMB_BITS = UltraCircuitBuilder::DEFAULT_NON_NATIVE_FIELD_LIMB_BITS;
    static inline const uint512_t BINARY_BASIS_MODULUS = uint512_t(1) << (LIMB_BITS * 4);
 
    // Number of NNF gates produced by a single partial multiplication
    static constexpr size_t NNF_GATES_PER_PARTIAL_MUL = 4;
    // Number of padding gates added to NNF block by finalize_circuit()
    static constexpr size_t NNF_ENSURE_NONZERO_PADDING = 0;
 
    // Generate 4 random field elements (one per limb)
    static std::array<fr, 4> random_limbs()
    {
        return { fr::random_element(), fr::random_element(), fr::random_element(), fr::random_element() };
    }
 
    // Splits a 256-bit integer into 4 68-bit limbs
    static std::array<fr, 4> split_into_limbs(const uint512_t& input)
    {
        std::array<fr, 4> limbs;
        limbs[0] = input.slice(0, LIMB_BITS).lo;
        limbs[1] = input.slice(LIMB_BITS * 1, LIMB_BITS * 2).lo;
        limbs[2] = input.slice(LIMB_BITS * 2, LIMB_BITS * 3).lo;
        limbs[3] = input.slice(LIMB_BITS * 3, LIMB_BITS * 4).lo;
        return limbs;
    }
 
    // Adds 4 limbs as circuit variables and returns their indices
    static std::array<uint32_t, 4> get_limb_witness_indices(UltraCircuitBuilder& builder,
                                                            const std::array<fr, 4>& limbs)
    {
        std::array<uint32_t, 4> limb_indices;
        limb_indices[0] = builder.add_variable(limbs[0]);
        limb_indices[1] = builder.add_variable(limbs[1]);
        limb_indices[2] = builder.add_variable(limbs[2]);
        limb_indices[3] = builder.add_variable(limbs[3]);
        return limb_indices;
    }
 
    // Helper to set up and evaluate a non-native field multiplication
    static std::array<uint32_t, 2> create_non_native_multiplication(UltraCircuitBuilder& builder,
                                                                    const uint256_t& a,
                                                                    const uint256_t& b,
                                                                    const uint256_t& q,
                                                                    const uint256_t& r,
                                                                    const uint256_t& modulus)
    {
        // Compute negative modulus: (-p) := 2^T - p
        auto modulus_limbs = split_into_limbs(BINARY_BASIS_MODULUS - uint512_t(modulus));
 
        // Add a, b, q, r as circuit variables
        const auto a_indices = get_limb_witness_indices(builder, split_into_limbs(uint256_t(a)));
        const auto b_indices = get_limb_witness_indices(builder, split_into_limbs(uint256_t(b)));
        const auto q_indices = get_limb_witness_indices(builder, split_into_limbs(uint256_t(q)));
        const auto r_indices = get_limb_witness_indices(builder, split_into_limbs(uint256_t(r)));
 
        // Prepare inputs for non-native multiplication gadget
        non_native_multiplication_witnesses<fr> inputs{
            a_indices, b_indices, q_indices, r_indices, modulus_limbs,
        };
        const auto [lo_1_idx, hi_1_idx] = builder.evaluate_non_native_field_multiplication(inputs);
 
        return { lo_1_idx, hi_1_idx };
    }
 
    // Compute quotient and remainder for a * b mod p
    static std::pair<uint256_t, uint256_t> compute_quotient_remainder(const uint256_t& a,
                                                                      const uint256_t& b,
                                                                      const uint256_t& modulus)
    {
        uint1024_t a_big = uint512_t(a);
        uint1024_t b_big = uint512_t(b);
        uint1024_t p_big = uint512_t(modulus);
 
        uint1024_t q_big = (a_big * b_big) / p_big;
        uint1024_t r_big = (a_big * b_big) % p_big;
 
        return { uint256_t(q_big.lo.lo), uint256_t(r_big.lo.lo) };
    }
 
    // Type aliases matching builder API for addition/subtraction
    using scaled_witness = std::pair<uint32_t, fr>;
    using add_simple = std::tuple<scaled_witness, scaled_witness, fr>;
 
    // Data structure for addition/subtraction test inputs
    struct AddSubData {
        std::array<fr, 4> x_limbs;
        std::array<fr, 4> y_limbs;
        fr x_prime;
        fr y_prime;
        std::array<fr, 4> x_scales;
        std::array<fr, 4> y_scales;
        std::array<fr, 4> addconsts;
        fr addconstp;
    };
 
    // Create random limb values for testing
    static AddSubData create_random_add_sub_data()
    {
        return AddSubData{
            .x_limbs = random_limbs(),
            .y_limbs = random_limbs(),
            .x_prime = fr::random_element(),
            .y_prime = fr::random_element(),
            .x_scales = { fr(1), fr(1), fr(1), fr(1) },
            .y_scales = { fr(1), fr(1), fr(1), fr(1) },
            .addconsts = { fr(0), fr(0), fr(0), fr(0) },
            .addconstp = fr(0),
        };
    }
 
    // Create add_simple tuples from test data and witness indices
    static std::tuple<add_simple, add_simple, add_simple, add_simple, std::tuple<uint32_t, uint32_t, fr>>
    create_add_sub_inputs(UltraCircuitBuilder& builder, const AddSubData& data)
    {
        // Add witness variables
        std::array<uint32_t, 4> x_idx, y_idx;
        for (size_t i = 0; i < 4; i++) {
            x_idx[i] = builder.add_variable(data.x_limbs[i]);
            y_idx[i] = builder.add_variable(data.y_limbs[i]);
        }
        uint32_t x_p_idx = builder.add_variable(data.x_prime);
        uint32_t y_p_idx = builder.add_variable(data.y_prime);
 
        // Build add_simple tuples: ((x_idx, x_scale), (y_idx, y_scale), addconst)
        add_simple limb0 = { { x_idx[0], data.x_scales[0] }, { y_idx[0], data.y_scales[0] }, data.addconsts[0] };
        add_simple limb1 = { { x_idx[1], data.x_scales[1] }, { y_idx[1], data.y_scales[1] }, data.addconsts[1] };
        add_simple limb2 = { { x_idx[2], data.x_scales[2] }, { y_idx[2], data.y_scales[2] }, data.addconsts[2] };
        add_simple limb3 = { { x_idx[3], data.x_scales[3] }, { y_idx[3], data.y_scales[3] }, data.addconsts[3] };
        auto limbp = std::make_tuple(x_p_idx, y_p_idx, data.addconstp);
 
        return { limb0, limb1, limb2, limb3, limbp };
    }
 
    // Helper to create partial multiplication inputs and queue the operation
    static std::array<uint32_t, 2> create_and_queue_partial_multiplication(UltraCircuitBuilder& builder,
                                                                           const std::array<fr, 4>& a_limbs,
                                                                           const std::array<fr, 4>& b_limbs)
    {
        const auto a_indices = get_limb_witness_indices(builder, a_limbs);
        const auto b_indices = get_limb_witness_indices(builder, b_limbs);
 
        non_native_partial_multiplication_witnesses<fr> input{ .a = a_indices, .b = b_indices };
        return builder.queue_partial_non_native_field_multiplication(input);
    }
 
    // Compute expected lo_0 and hi_1 values for partial multiplication.
    //
    // Given two non-native field elements represented as 4 limbs each:
    //   a = a[0] + a[1]*L + a[2]*L² + a[3]*L³   (where L = 2^LIMB_BITS)
    //   b = b[0] + b[1]*L + b[2]*L² + b[3]*L³
    //
    // The product a*b expands via schoolbook multiplication to terms at powers of L:
    //   L⁰: a[0]*b[0]
    //   L¹: a[1]*b[0] + a[0]*b[1]
    //   L²: a[2]*b[0] + a[1]*b[1] + a[0]*b[2]
    //   L³: a[3]*b[0] + a[2]*b[1] + a[1]*b[2] + a[0]*b[3]
    //   (higher powers of L can be ignored)
    //
    // The partial products group these terms:
    //   lo_0 = L⁰ and L¹ terms (low 2*LIMB_BITS portion)
    //   hi_1 = L² and L³ terms (high portion, split into hi_0 + remaining for constraint efficiency)
    static std::pair<fr, fr> compute_expected_partial_products(const std::array<fr, 4>& a, const std::array<fr, 4>& b)
    {
        const fr LIMB_SHIFT = fr(uint256_t(1) << LIMB_BITS);
 
        fr lo_0 = a[0] * b[0] + ((a[1] * b[0] + a[0] * b[1]) * LIMB_SHIFT);
        fr hi_0 = a[2] * b[0] + a[0] * b[2] + ((a[0] * b[3] + a[3] * b[0]) * LIMB_SHIFT);
        fr hi_1 = hi_0 + a[1] * b[1] + ((a[1] * b[2] + a[2] * b[1]) * LIMB_SHIFT);
 
        return { lo_0, hi_1 };
    }
};
 
// Verifies that valid non-native field multiplications pass the circuit checker
TEST_F(UltraCircuitBuilderNonNative, Multiplication)
{
    const size_t num_iterations = 50;
    for (size_t i = 0; i < num_iterations; i++) {
        UltraCircuitBuilder builder;
 
        uint256_t a = uint256_t(fq::random_element());
        uint256_t b = uint256_t(fq::random_element());
        uint256_t modulus = fq::modulus;
 
        auto [q, r] = compute_quotient_remainder(a, b, modulus);
        const auto [lo_1_idx, hi_1_idx] = create_non_native_multiplication(builder, a, b, q, r, modulus);
 
        // Range check the carry (output) lo and hi limbs
        const bool is_low_70_bits = uint256_t(builder.get_variable(lo_1_idx)).get_msb() < 70;
        const bool is_high_70_bits = uint256_t(builder.get_variable(hi_1_idx)).get_msb() < 70;
        if (is_low_70_bits && is_high_70_bits) {
            // Uses more efficient NNF range check if both limbs are < 2^70
            builder.range_constrain_two_limbs(lo_1_idx, hi_1_idx, 70, 70);
        } else {
            // Fallback to default range checks
            builder.create_limbed_range_constraint(lo_1_idx, 72);
            builder.create_limbed_range_constraint(hi_1_idx, 72);
        }
 
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
}
 
// Regression test: carry limb > 2^70 requires fallback to default range checks
TEST_F(UltraCircuitBuilderNonNative, MultiplicationLargeCarryRegression)
{
    UltraCircuitBuilder builder;
 
    // Edge case values that produce carry > 2^70
    uint256_t a = uint256_t("0x00ab1504deacff852326adf4a01099e9340f232e2a631042852fce3c4eb8a51b");
    uint256_t b = uint256_t("0x1be457323502cfcd85f8cfa54c8c4fea146b9db2a7d86b29d966d61b714ee249");
    uint256_t q_expected = uint256_t("0x00629b9d576dfc6b5c28a4a254d5e8e3384124f6a898858e95265254a01414d5");
    uint256_t r_expected = uint256_t("0x2c1590eb70a48dce72f7686bbf79b59bf7926c99bc16aba92e474c65a04ea2a0");
    uint256_t modulus = fq::modulus;
 
    // Verify that our expected q, r are correct
    auto [q_computed, r_computed] = compute_quotient_remainder(a, b, modulus);
    EXPECT_EQ(q_computed, q_expected);
    EXPECT_EQ(r_computed, r_expected);
 
    // This edge case leads to the carry limb being > 2^70, so it used to fail when applying a 2^70 range check
    // (with range_constrain_two_limbs). Now it should work since we fallback to default range checks in such a case.
    const auto [lo_1_idx, hi_1_idx] = create_non_native_multiplication(builder, a, b, q_expected, r_expected, modulus);
 
    // Range check the carry (output) lo and hi limbs
    const bool is_high_70_bits = uint256_t(builder.get_variable(hi_1_idx)).get_msb() < 70;
    BB_ASSERT(is_high_70_bits == false); // Regression should hit this case
 
    // Decompose into default range: these should work even if the limbs are > 2^70
    builder.create_limbed_range_constraint(lo_1_idx, 72);
    builder.create_limbed_range_constraint(hi_1_idx, 72);
    EXPECT_TRUE(CircuitChecker::check(builder));
 
    // Using NNF range check should fail here
    builder.range_constrain_two_limbs(lo_1_idx, hi_1_idx, 70, 70);
    EXPECT_FALSE(CircuitChecker::check(builder));
    EXPECT_EQ(builder.err(), "range_constrain_two_limbs: hi limb.");
}
 
// Verifies non-native field addition with various scaling factors
TEST_F(UltraCircuitBuilderNonNative, Addition)
{
    // Test with identity scaling (z = x + y)
    {
        UltraCircuitBuilder builder;
        auto data = create_random_add_sub_data();
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_addition(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Test with different scaling factors per limb and non-zero addconstp
    {
        UltraCircuitBuilder builder;
        auto data = create_random_add_sub_data();
        data.x_scales = { fr(2), fr(3), fr(5), fr(7) };
        data.y_scales = { fr(11), fr(13), fr(17), fr(19) };
        data.addconsts = { fr(100), fr(200), fr(300), fr(400) };
        data.addconstp = fr(500);
 
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_addition(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Test with negative and zero scaling factors
    {
        UltraCircuitBuilder builder;
        auto data = create_random_add_sub_data();
        data.x_scales = { fr(-1), fr(0), fr(1), fr(-2) };
        data.y_scales = { fr(1), fr(-1), fr(0), fr(2) };
        data.addconsts = { fr(0), fr(-50), fr(50), fr(0) };
        data.addconstp = fr(-100);
 
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_addition(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Edge case: all zeros
    {
        UltraCircuitBuilder builder;
        AddSubData data{
            .x_limbs = { fr(0), fr(0), fr(0), fr(0) },
            .y_limbs = { fr(0), fr(0), fr(0), fr(0) },
            .x_prime = fr(0),
            .y_prime = fr(0),
            .x_scales = { fr(1), fr(1), fr(1), fr(1) },
            .y_scales = { fr(1), fr(1), fr(1), fr(1) },
            .addconsts = { fr(0), fr(0), fr(0), fr(0) },
            .addconstp = fr(0),
        };
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_addition(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Edge case: x == y (doubling)
    {
        UltraCircuitBuilder builder;
        fr val0 = fr::random_element();
        fr val1 = fr::random_element();
        fr val2 = fr::random_element();
        fr val3 = fr::random_element();
        fr valp = fr::random_element();
        AddSubData data{
            .x_limbs = { val0, val1, val2, val3 },
            .y_limbs = { val0, val1, val2, val3 },
            .x_prime = valp,
            .y_prime = valp,
            .x_scales = { fr(1), fr(1), fr(1), fr(1) },
            .y_scales = { fr(1), fr(1), fr(1), fr(1) },
            .addconsts = { fr(0), fr(0), fr(0), fr(0) },
            .addconstp = fr(0),
        };
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_addition(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
}
 
// Verifies non-native field subtraction with various scaling factors
TEST_F(UltraCircuitBuilderNonNative, Subtraction)
{
    // Test with identity scaling (z = x - y)
    {
        UltraCircuitBuilder builder;
        auto data = create_random_add_sub_data();
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_subtraction(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Test with different scaling factors per limb and non-zero addconstp
    {
        UltraCircuitBuilder builder;
        auto data = create_random_add_sub_data();
        data.x_scales = { fr(2), fr(3), fr(5), fr(7) };
        data.y_scales = { fr(11), fr(13), fr(17), fr(19) };
        data.addconsts = { fr(100), fr(200), fr(300), fr(400) };
        data.addconstp = fr(500);
 
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_subtraction(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Test with negative and zero scaling factors
    {
        UltraCircuitBuilder builder;
        auto data = create_random_add_sub_data();
        data.x_scales = { fr(-1), fr(0), fr(1), fr(-2) };
        data.y_scales = { fr(1), fr(-1), fr(0), fr(2) };
        data.addconsts = { fr(0), fr(-50), fr(50), fr(0) };
        data.addconstp = fr(-100);
 
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_subtraction(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Edge case: all zeros
    {
        UltraCircuitBuilder builder;
        AddSubData data{
            .x_limbs = { fr(0), fr(0), fr(0), fr(0) },
            .y_limbs = { fr(0), fr(0), fr(0), fr(0) },
            .x_prime = fr(0),
            .y_prime = fr(0),
            .x_scales = { fr(1), fr(1), fr(1), fr(1) },
            .y_scales = { fr(1), fr(1), fr(1), fr(1) },
            .addconsts = { fr(0), fr(0), fr(0), fr(0) },
            .addconstp = fr(0),
        };
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_subtraction(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
 
    // Edge case: x == y (result is zero)
    {
        UltraCircuitBuilder builder;
        fr val0 = fr::random_element();
        fr val1 = fr::random_element();
        fr val2 = fr::random_element();
        fr val3 = fr::random_element();
        fr valp = fr::random_element();
        AddSubData data{
            .x_limbs = { val0, val1, val2, val3 },
            .y_limbs = { val0, val1, val2, val3 },
            .x_prime = valp,
            .y_prime = valp,
            .x_scales = { fr(1), fr(1), fr(1), fr(1) },
            .y_scales = { fr(1), fr(1), fr(1), fr(1) },
            .addconsts = { fr(0), fr(0), fr(0), fr(0) },
            .addconstp = fr(0),
        };
        auto [limb0, limb1, limb2, limb3, limbp] = create_add_sub_inputs(builder, data);
        builder.evaluate_non_native_field_subtraction(limb0, limb1, limb2, limb3, limbp);
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
}
 
// Verifies that providing incorrect witnesses to multiplication causes failure
TEST_F(UltraCircuitBuilderNonNative, MultiplicationInvalidWitnessFailure)
{
    // Helper to test that providing incorrect quotient/remainder causes failure
    auto test_incorrect_qr = [](bool tamper_q, size_t limb_idx) {
        UltraCircuitBuilder builder;
 
        uint256_t a = uint256_t(fq::random_element());
        uint256_t b = uint256_t(fq::random_element());
        uint256_t modulus = fq::modulus;
        auto [q, r] = compute_quotient_remainder(a, b, modulus);
 
        // Tamper with quotient or remainder
        if (tamper_q) {
            // Add 1 to a specific limb of q
            auto q_limbs = split_into_limbs(uint256_t(q));
            q_limbs[limb_idx] += fr(1);
            q = uint256_t(q_limbs[0]) + (uint256_t(q_limbs[1]) << LIMB_BITS) +
                (uint256_t(q_limbs[2]) << (LIMB_BITS * 2)) + (uint256_t(q_limbs[3]) << (LIMB_BITS * 3));
        } else {
            // Add 1 to a specific limb of r
            auto r_limbs = split_into_limbs(uint256_t(r));
            r_limbs[limb_idx] += fr(1);
            r = uint256_t(r_limbs[0]) + (uint256_t(r_limbs[1]) << LIMB_BITS) +
                (uint256_t(r_limbs[2]) << (LIMB_BITS * 2)) + (uint256_t(r_limbs[3]) << (LIMB_BITS * 3));
        }
 
        // Now a*b != q*p + r (the multiplication identity is violated)
        const auto [lo_idx, hi_idx] = create_non_native_multiplication(builder, a, b, q, r, modulus);
 
        // Add range constraints
        builder.create_limbed_range_constraint(lo_idx, 72);
        builder.create_limbed_range_constraint(hi_idx, 72);
 
        EXPECT_FALSE(CircuitChecker::check(builder));
    };
 
    // Test tampering with each limb of q and r
    for (size_t limb = 0; limb < 4; limb++) {
        test_incorrect_qr(true, limb);  // tamper q
        test_incorrect_qr(false, limb); // tamper r
    }
}
 
// Verifies that queue_partial_non_native_field_multiplication computes correct intermediate values
TEST_F(UltraCircuitBuilderNonNative, PartialMultiplicationBasic)
{
    const size_t num_iterations = 10;
    for (size_t i = 0; i < num_iterations; i++) {
        UltraCircuitBuilder builder;
 
        std::array<fr, 4> a_limbs = random_limbs();
        std::array<fr, 4> b_limbs = random_limbs();
 
        const auto [lo_0_idx, hi_1_idx] = create_and_queue_partial_multiplication(builder, a_limbs, b_limbs);
 
        // Verify returned witnesses contain expected values
        auto [expected_lo_0, expected_hi_1] = compute_expected_partial_products(a_limbs, b_limbs);
        EXPECT_EQ(builder.get_variable(lo_0_idx), expected_lo_0);
        EXPECT_EQ(builder.get_variable(hi_1_idx), expected_hi_1);
 
        // Verify circuit passes after finalization (which calls process_non_native_field_multiplications)
        EXPECT_TRUE(CircuitChecker::check(builder));
    }
}
 
// Verifies that duplicate partial multiplications are deduplicated
TEST_F(UltraCircuitBuilderNonNative, PartialMultiplicationDeduplication)
{
    UltraCircuitBuilder builder;
    std::array<fr, 4> a_limbs = random_limbs();
    std::array<fr, 4> b_limbs = random_limbs();
 
    // Add witnesses once (to be reused)
    const auto a_indices = get_limb_witness_indices(builder, a_limbs);
    const auto b_indices = get_limb_witness_indices(builder, b_limbs);
 
    // Queue the same multiplication multiple times using the same witness indices
    const size_t num_duplicates = 5;
    std::vector<std::array<uint32_t, 2>> results;
    for (size_t i = 0; i < num_duplicates; i++) {
        non_native_partial_multiplication_witnesses<fr> input{ .a = a_indices, .b = b_indices };
        results.push_back(builder.queue_partial_non_native_field_multiplication(input));
    }
 
    // Cache should have all entries before finalization
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), num_duplicates);
 
    // Verify circuit passes (finalization happens internally via copy)
    EXPECT_TRUE(CircuitChecker::check(builder));
 
    // Finalize to check deduplication results
    builder.finalize_circuit();
 
    // After finalization, cache should be deduplicated to 1 entry
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), 1U);
 
    // NNF block size should be the same as single multiplication (deduplication worked)
    EXPECT_EQ(builder.blocks.nnf.size(), NNF_GATES_PER_PARTIAL_MUL + NNF_ENSURE_NONZERO_PADDING);
 
    // All results should have the same expected values
    auto [expected_lo_0, expected_hi_1] = compute_expected_partial_products(a_limbs, b_limbs);
    for (const auto& result : results) {
        EXPECT_EQ(builder.get_variable(result[0]), expected_lo_0);
        EXPECT_EQ(builder.get_variable(result[1]), expected_hi_1);
    }
}
 
// Verifies deduplication correctly handles duplicates that result from an assert_equal
TEST_F(UltraCircuitBuilderNonNative, PartialMultiplicationDedupeAssertEqual)
{
    UltraCircuitBuilder builder;
    std::array<fr, 4> a_limbs = random_limbs();
    std::array<fr, 4> b_limbs = random_limbs();
 
    // Create two sets of witness indices (different indices, same underlying values)
    const auto a1_indices = get_limb_witness_indices(builder, a_limbs);
    const auto b1_indices = get_limb_witness_indices(builder, b_limbs);
    const auto a2_indices = get_limb_witness_indices(builder, a_limbs);
    const auto b2_indices = get_limb_witness_indices(builder, b_limbs);
 
    // Without assert_equal, these would be distinct indices and wouldn't deduplicate.
    // assert_equal should make them behave as duplicates.
    for (size_t i = 0; i < 4; i++) {
        builder.assert_equal(a1_indices[i], a2_indices[i]);
        builder.assert_equal(b1_indices[i], b2_indices[i]);
    }
 
    // Queue partial multiplications using both sets of indices
    non_native_partial_multiplication_witnesses<fr> input1{ .a = a1_indices, .b = b1_indices };
    non_native_partial_multiplication_witnesses<fr> input2{ .a = a2_indices, .b = b2_indices };
    builder.queue_partial_non_native_field_multiplication(input1);
    builder.queue_partial_non_native_field_multiplication(input2);
 
    // Before finalization, cache has 2 entries
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), 2U);
 
    // Verify circuit passes
    EXPECT_TRUE(CircuitChecker::check(builder));
 
    // Finalize to check deduplication
    builder.finalize_circuit();
 
    // After finalization, should be deduplicated to 1 entry
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), 1U);
 
    // NNF block size should be the same as single multiplication
    EXPECT_EQ(builder.blocks.nnf.size(), NNF_GATES_PER_PARTIAL_MUL + NNF_ENSURE_NONZERO_PADDING);
}
 
// Verifies multiple distinct partial multiplications all produce correct constraints
TEST_F(UltraCircuitBuilderNonNative, PartialMultiplicationMultipleDistinct)
{
    UltraCircuitBuilder builder;
 
    const size_t num_multiplications = 5;
    std::vector<std::array<fr, 4>> a_values(num_multiplications);
    std::vector<std::array<fr, 4>> b_values(num_multiplications);
    std::vector<std::array<uint32_t, 2>> results(num_multiplications);
 
    // Queue multiple distinct partial multiplications
    for (size_t i = 0; i < num_multiplications; i++) {
        a_values[i] = random_limbs();
        b_values[i] = random_limbs();
        results[i] = create_and_queue_partial_multiplication(builder, a_values[i], b_values[i]);
    }
 
    // All entries should be in the cache
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), num_multiplications);
 
    // Verify circuit passes
    EXPECT_TRUE(CircuitChecker::check(builder));
 
    // Finalize to check gate counts
    builder.finalize_circuit();
 
    // After finalization, all distinct entries should remain (no deduplication)
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), num_multiplications);
 
    // NNF block size should scale with number of multiplications
    EXPECT_EQ(builder.blocks.nnf.size(), num_multiplications * NNF_GATES_PER_PARTIAL_MUL + NNF_ENSURE_NONZERO_PADDING);
 
    // Verify each result has correct values
    for (size_t i = 0; i < num_multiplications; i++) {
        auto [expected_lo_0, expected_hi_1] = compute_expected_partial_products(a_values[i], b_values[i]);
        EXPECT_EQ(builder.get_variable(results[i][0]), expected_lo_0);
        EXPECT_EQ(builder.get_variable(results[i][1]), expected_hi_1);
    }
}
 
// Verifies that finalization with empty cache doesn't cause issues
TEST_F(UltraCircuitBuilderNonNative, ProcessNonNativeFieldMultiplicationsEmptyCache)
{
    UltraCircuitBuilder builder;
 
    // Create some non-NNF constraints to ensure the circuit isn't completely empty
    fr a = fr::random_element();
    fr b = fr::random_element();
    uint32_t a_idx = builder.add_variable(a);
    uint32_t b_idx = builder.add_variable(b);
    uint32_t c_idx = builder.add_variable(a + b);
    builder.create_add_gate({ a_idx, b_idx, c_idx, fr(1), fr(1), fr(-1), fr(0) });
 
    // No NNF operations queued
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), 0U);
 
    // Finalization should succeed without issues
    EXPECT_TRUE(CircuitChecker::check(builder));
 
    // Finalize to inspect NNF block
    builder.finalize_circuit();
 
    // Cache should still be empty
    EXPECT_EQ(builder.cached_partial_non_native_field_multiplications.size(), 0U);
 
    // NNF block should only have ensure_nonzero padding (no actual NNF gates)
    EXPECT_EQ(builder.blocks.nnf.size(), NNF_ENSURE_NONZERO_PADDING);
}