Some unit tests failed on i386

[Harry-Chen]

Describe the bug

As described in #645 (comment), some unit tests are failing on i386.

The detailed logs can be retrieved from https://ci.debian.net/packages/j/jsoncons/testing/i386/66453183/. Some lines that might be useful:

3136s 1/1 Test #1: unit_tests .......................***Failed    2.62 sec
3136s d8 55 48 ff ff 7f ff ff ff 7f 7f 
3136s 82 fa ff 7f ff ff fa 7f 7f ff ff 
3136s 
3136s 
3136s ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
3136s unit_tests is a Catch v2.13.10 host application.
3136s Run with -? for options
3136s 
3136s -------------------------------------------------------------------------------
3136s cbor typed array tests
3136s   Tag 85, float32, little endian
3136s -------------------------------------------------------------------------------
3136s ./test/cbor/src/cbor_typed_array_tests.cpp:742
3136s ...............................................................................
3136s 
3136s ./test/cbor/src/cbor_typed_array_tests.cpp:38: FAILED:
3136s   CHECK( (u == v) )
3136s with expansion:
3136s   false
3136s 
3136s d8 56 50 ff ff ff ff ff ff ef ff ff ff ff ff ff ff ef 7f 
3136s 82 fb ff ef ff ff ff ff ff ff fb 7f ef ff ff ff ff ff ff 
3136s 
3136s -------------------------------------------------------------------------------
3136s cbor typed array tests
3136s   Tag 86, float64, little endian
3136s -------------------------------------------------------------------------------
3136s ./test/cbor/src/cbor_typed_array_tests.cpp:773
3136s ...............................................................................
3136s 
3136s ./test/cbor/src/cbor_typed_array_tests.cpp:38: FAILED:
3136s   CHECK( (u == v) )
3136s with expansion:
3136s   false

...


3136s ===============================================================================
3136s test cases:   838 |   837 passed | 1 failed
3136s assertions: 20292 | 20290 passed | 2 failed
3136s 
3136s 
3136s 
3136s 0% tests passed, 1 tests failed out of 1
3136s 
3136s Total Test time (real) =   2.62 sec
3136s 
3136s The following tests FAILED:
3136s 	  1 - unit_tests (Failed)

Enumerate the steps to reproduce the bug

Run the tests with -g -O2 on i386 machine (or chroot).

What compiler, architecture, and operating system?

• Compiler: GCC 15.2.0
• Architecture (e.g. x86, x64): i386 (x86)
• Operating system: Debian sid

What jsoncons library version?

• Latest release 1.5.0
• Other release ______
• master

[danielaparker]

Thanks.

From your logs, I understand what's happening, but I don't understand why it's happening.

What's happening is that in your environment jsoncons is correctly decoding a typed array of float64 in little endian format,

d8 56 50 ff ff ff ff ff ff ef ff ff ff ff ff ff ff ef 7f 

But when encoding, jsoncons produces, not the expected typed array of float64 in little endian format, but rather a regular CBOR array in big endian format,

82 fb ff ef ff ff ff ff ff ff fb 7f ef ff ff ff ff ff ff 

It's the same output you would get replacing

    auto options = cbor::cbor_options{}
        .use_typed_arrays(true);
    cbor::encode_cbor(u, buf, options);

with

    cbor::encode_cbor(u, buf);

It's valid CBOR, just not the expected typed array little endian representation.

This isn't happening in any of my test environments, which all pass. I'll likely need to give you a small test application that produces more diagnostics.

[danielaparker]

Could you try running the tests again with the latest code on master? Thanks.

[Harry-Chen]

@danielaparker The errors remain exactly the same on the latest master commit (dfb01ef).

[danielaparker]

It can't remain exactly the same, the reported line numbers would have to be different.

[Harry-Chen]

It can't remain exactly the same, the reported line numbers would have to be different.

By "exactly the same", I mean the errors are still from those two tests, with the same tag, and the same expected & actual values.

And since you mentioned line numbers, they are the same. Because the related file, i.e., test/cbor/src/cbor_typed_array_tests.cpp, has not changed since August.

[danielaparker]

Anyway, your results are information, I need to look elsewhere.

[danielaparker]

Can you run the program below with code on branch typed_array_issue and provide me with the output? Thanks.

#include <jsoncons/json.hpp>
#include <jsoncons_ext/cbor/cbor.hpp>
#include <iostream>

namespace cbor = jsoncons::cbor;

//d8 56 50 ff ff ff ff ff ff ef ff ff ff ff ff ff ff ef 7f (typed array)
//82 fb ff ef ff ff ff ff ff ff fb 7f ef ff ff ff ff ff ff (regular array)

int main()
{
    const std::vector<uint8_t> input = {
        0xd8, // Tag
        0x56, // Tag 86, float64, little endian, Typed Array
        0x50, // Byte string value of length 16
        0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xef, 0xff,
        0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xef, 0x7f
    };

    auto u = cbor::decode_cbor<std::vector<double>>(input);

    auto options = cbor::cbor_options{}
        .use_typed_arrays(true);
    std::vector<uint8_t> buf;
    std::cout << "main cbor_options::use_typed_arrays: " << options.use_typed_arrays() << "\n";
    cbor::encode_cbor(u, buf, options);

    for (auto c : input)
    {
        std::cout << std::hex << std::setw(2) << std::setfill('0') << (int)c << " ";
    }
    std::cout << "\n";
    for (auto c : buf)
    {
        std::cout << std::hex << std::setw(2) << std::setfill('0') << (int)c << " ";
    }
    std::cout << "\n\n";
}

[Harry-Chen]

@danielaparker Sorry for being a little bit late. The result is:

main cbor_options::use_typed_arrays: 1
try_encode_cbor cbor_encode_options::use_typed_arrays: 1, is_array_like: 1, is_typed_array: 1
basic_cbor_encoder::use_typed_arrays: 1
encode_traits::try_encode is_array_like: 1, is_typed_array: 1
visit_typed_array use_typed_arrays: 1
write_typed_array_tag 0x56
d8 56 50 ff ff ff ff ff ff ef ff ff ff ff ff ff ff ef 7f
d8 56 50 ff ff ff ff ff ff ef ff ff ff ff ff ff ff ef 7f

[Harry-Chen]

And on this branch (45498a6) and the latest master (6aa53e3), the unit test errors on i386 are both gone.

[danielaparker]

Thanks for confirming,

[danielaparker]

The issue was with the detection of CBOR typed arrays that only manifested itself on your i386 environment. It's fixed on master.