`test-code-cache` fails on a Linux/riscv64 build with FFI enabled.

[sxa]

Version

main branch

Platform

linux/riscv64

Subsystem

FFI

What steps will reproduce the bug?

Running the parallel/test-code-cache test on a build with FFI enabled. This is spun out from the comment in nodejs/build#4099 (comment)

As background FFI was first able to be enabled after #63794 and superceded by the update in #64040 which included the upstream fix.

How often does it reproduce? Is there a required condition?

Seems to be 100%

What is the expected behavior? Why is that the expected behavior?

Test passes.

What do you see instead?

Here are the first fiew lines of failures - the lines towards the end are replicated with different module names (i.e. it's not just internal/streams for example assert and net are included in these lines int he full log:

Path: parallel/test-code-cache
node:internal/assert/utils:146
  throw error;
  ^

AssertionError [ERR_ASSERTION]: "internal/abort_controller" should've been compiled **with** code cache
"internal/streams/end-of-stream" should've been compiled **with** code cache
"internal/streams/operators" should've been compiled **with** code cache
"internal/streams/destroy" should've been compiled **with** code cache
"internal/streams/legacy" should've been compiled **with** code cache
"internal/streams/add-abort-signal" should've been compiled **with** code cache
[...]

Additional information

No response

[sxa]

@nodejs/ffi @joyeecheung i haven't gone digging yet but would any of you have any idea why enabling FFI in the build might be breaking this test? I'm assuming this hasn't been spotted on any other platform so far and really is specific to Linux/riscv64.

[kxxt]

Which means the parallel/test-code-cache problem seems to be the one remaining that will need addressing. Noting that node -pe 'process.config.variables.node_use_node_code_cache' returns true on the native builds but false on the unofficial (cross-compiled) builds so those pass. It looks like it started failing when the FFI support was added, and I've confirmed (sample size of 1!) that building with --without-ffi seems to allow that test to pass. We need to understand why this might only be affecting RISC-V so I'll split out another issue to cover this failure.

Hmm, I didn't encounter this test failure with a native build on SG2042.

Test logs available at https://github.com/riscv-forks/node-riscv/actions/runs/30152090741/job/89671914200.

./node  -pe 'process.config.variables.node_use_node_code_cache'
true
<]$ out/Release/node test/parallel/test-code-cache.js                        
NOTE: The test started as a child_process using these flags: [ '--expose-internals' ] And these environment variables: {} Use NODE_SKIP_FLAG_CHECK to run the test with the original flags.
(node:1603257) internal/test/binding: These APIs are for internal testing only. Do not use them.
(Use `node --trace-warnings ...` to show where the warning was created)
(node:1603257) internal/test/transfer: These APIs are for internal testing only. Do not use them.
(node:1603257) [DEP0040] DeprecationWarning: The `punycode` module is deprecated. Please use a userland alternative instead.
[kxxt@centiskorch node-riscv]$ echo $?
0
[kxxt@centiskorch node-riscv]$ ./node --experimental-ffi -e "require('node:ffi');console.log('ffi is enabled')"                               
ffi is enabled
(node:1603620) ExperimentalWarning: FFI is an experimental feature and might change at any time
(Use `node --trace-warnings ...` to show where the warning was created)

[kxxt]

would any of you have any idea why enabling FFI in the build might be breaking this test?

The last time I encounter such weird issues is that configuring node.js with --shared broke some tests on riscv64. But it later turned out to be a mis-compilation done by GCC (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116057). Maybe it is worthwhile to check the compiler version and if you could reproduce it with clang.