[DeviceSanitizer] Support "-fsanitize-ignorelist=" to disable sanitizing on some of kernels #2055
Conversation
github-actions
bot
added
loader
|
please rebase |
AllanZyne
changed the title
| if (m_KernelMap.find(Kernel) != m_KernelMap.end()) { | ||
| return m_KernelMap[Kernel]; | ||
| } | ||
| return nullptr; |
There was a problem hiding this comment.
Please add assertions for all caller of getKernelInfo to avoid segment fault. I see there are some guard with assertion, but some are not.
There was a problem hiding this comment.
Actually, I think this is better than old implement(return nullptr when not found), then there will be a segv that we will not miss. But adding assertion after a call that may return nullptr is better.
There was a problem hiding this comment.
It needn't assert, I did this on purpose.
It possible that some of kernel are not in m_KernelMap.
I have added "if" checkers when calling getKernelInfo.
You're right, I will add.
There was a problem hiding this comment.
KernelInfo doesn't like other XXXinfo.
It's possible that some of kernel are not in m_KernelMap.
| @@ -111,9 +111,9 @@ struct ProgramInfo { | |||
| ur_program_handle_t Handle; | |||
| std::atomic<int32_t> RefCount = 1; | |||
|
|
|||
| // lock this mutex if following fields are accessed | |||
There was a problem hiding this comment.
If the program is only built once, then this would not be a problem to use mutex. I don't like the "likely" in the comment unless it has proves. I prefer keeping the mutex.
There was a problem hiding this comment.
Synced with @zhaomaosu, mutex can be removed here.
I removed "likely" in comment.
|
Hi @oneapi-src/unified-runtime-maintain, please review, thank you. |
LLVM: intel/llvm#15294
Add a device global "__AsanKernelMetadata" which records the sanitized kernel name, so that we can check if sanitizer layer needs specially handle it.