Crash asking me to report it here #96135
Labels
clang:codegen
clang:openmp
OpenMP related changes to Clang
confirmed
Verified by a second party
crash
Prefer [crash-on-valid] or [crash-on-invalid]
needs-reduction
Large reproducer that should be reduced into a simpler form
Comments
|
Update: even when I remove everything related to openmp in the function that is indicated in the crash report, it still crashs and apparently in the same function |
EugeneZelenko
added
clang:codegen
crash
|
@llvm/issue-subscribers-clang-codegen Author: Line (Line-fr)
Hi, I have been trying to port my code to openMP recently and while it seemed to work fine for the first part of my code, for the second, I made it quite unhappy. Instead of getting a proper error allowing me to know what to do about it, it crashed giving me this: (the files do not have the exact same name but from a previous compilation without any modification
compilation command: clang++ main.cpp -fopenmp --offload-arch=gfx90a I hope you will be able to solve this crash for the future nice error messages by my side, I am gonna try to solve it blindly |
|
Could you please try 18 or |
|
HI Thank you for your reply, I am gonna try recompiling it but I have never done such thing and I am not a sudo user in the server in question. I will keep you in touch of my progress for this compilation and using of another llvm version |
|
Okay I succeeded and got the error again: |
|
I found why in my code! It was not in the function printed by the error. I don't know if it is supposed to make such a crash but #pragma omp parallel for reduction(+:allresults[0: (2*localqbits)]) breaks it #pragma omp parallel for reduction(+:allresults[0:64]) works This new functionnality of openMP 4.5 allowing to reduce arrays seems to break on non compiling time constant size array I hope this issue was not useless to you, sorry for the time |
It's absolutely not. We shouldn't be crashing on both good inputs and bad inputs, so there's something to fix here. |
|
@llvm/issue-subscribers-openmp Author: Line (Line-fr)
Hi, I have been trying to port my code to openMP recently and while it seemed to work fine for the first part of my code, for the second, I made it quite unhappy. Instead of getting a proper error allowing me to know what to do about it, it crashed giving me this: (the files do not have the exact same name but from a previous compilation without any modification
compilation command: clang++ main.cpp -fopenmp --offload-arch=gfx90a I hope you will be able to solve this crash for the future nice error messages by my side, I am gonna try to solve it blindly |
The compiler should never crash like this. It would be very helpful if you could create a small reproducer and either post it here or through https://godbolt.org/. |
|
Hi, it was not so easy to isolate exactly at first but by copying my code and removing progressively, I ended up reducing it drastically!
#include<omp.h>
#include<iostream>
using namespace std;
int main(){
int localqbits = 7;
double allresults[64]; //allocating on gpu or not do not change anything, allresults just needs to be a pointer
#pragma omp target device(0) is_device_ptr(allresults)
{
#pragma omp parallel for reduction(+:allresults[0:2*localqbits])
for (int m = 0; m < 1; m++){
}
}
return 0;
}Note that if I add "const int localqbits" it stops crashing. the bug is on the default clang of my system. Here is how I compiled main branch to test again:
result of compilation checks It crashes on main branch too I hope that it helped you and that I did a good enough job at reducing it! |
|
@Line-fr Thank you, you did a good job! Clang 18.1.0 crash: |
|
If you compile it with an NVPTX architecture you get a warning bout VLAs. I wonder why you don't need this in the AMDGPU version, considering we don't fully support VLAs there either. |
EugeneZelenko
added
clang:openmp
|
FWIW, this works if you make the |
|
Hi all! I saw this just recently on a recent rocm installation with the provided clang, but obviously it's an upstream bug. #include <malloc.h>
int main(int argc, char* argv[])
{
int* arr = (int*) malloc(1000 * sizeof(int));
#pragma omp target enter data map(alloc: arr[0:1000])
#pragma omp target teams distribute parallel for
for(int i=0; i<1000; i++) { arr[i] = -1.0; }
int len = 1000;
#pragma omp target teams distribute parallel for reduction(+:arr[0:len])
for(int i = 0; i < 1000; i++)
arr[i] += 1;
}See also https://godbolt.org/z/xEGxYYfrv |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi, I have been trying to port my code to openMP recently and while it seemed to work fine for the first part of my code, for the second, I made it quite unhappy. Instead of getting a proper error allowing me to know what to do about it, it crashed giving me this: (the files do not have the exact same name but from a previous compilation without any modification
compilation command: clang++ main.cpp -fopenmp --offload-arch=gfx90a
I use MI250x with rocm which worked well with openmp up to now
errormessage.txt
main-bd92e5.zip
I hope you will be able to solve this crash for the future nice error messages by my side, I am gonna try to solve it blindly
Thank you
The text was updated successfully, but these errors were encountered: