Bug ID | 109794 |
---|---|
Summary | LLVM not properly shutdown in `si_pipe.c`? |
Product | Mesa |
Version | git |
Hardware | x86-64 (AMD64) |
OS | Linux (All) |
Status | NEW |
Severity | minor |
Priority | medium |
Component | Drivers/Gallium/radeonsi |
Assignee | dri-devel@lists.freedesktop.org |
Reporter | zegentzy@protonmail.com |
QA Contact | dri-devel@lists.freedesktop.org |
tl;dr; Shouldn't we be calling `LLVMShutdown()` somewhere? I had a program which was calling: ``` eglInitialize(disp, &majoir, &minor); eglTerminate(disp); eglInitialize(disp, &majoir, &minor); ``` Which lead to mesa outputting the following before continuing on normally during the second call to `eglInitalize`: ``` mesa: for the -simplifycfg-sink-common option: may only occur zero or one times! mesa: for the -global-isel-abort option: may only occur zero or one times! ``` This prompted an investigation me to investigate, and with the aid of @imirkin, I think I got the cause. You see, in an `eglInitialize` call `radeonsi_dri.so` gets loaded, which also loads in `libLLVM-9svn.so`, and eventually `ac_init_llvm_once()` gets called more than once. That's OK, for as the name implies, it only happens once. Issue is, when `eglTerminate` is called, `radeonsi_dri.so` gets unloaded, but not `libLLVM-9svn.so`. This means that the next time we all `eglInitialize`, LLVM's `ManagedStatic`s will be preserved, but not our static var `ac_init_llvm_target_once_flag`. We can fix this by calling `LLVMShutdown` before unloading `radeonsi_dri.so`, that way LLVM's `ManagedStatic`s are also reset. Alas, I'm not familiar enough with the mesa codebase, so I got no clue where the appropriate place to plop this function is. Any aid would be well appreciated.