PR #2596 introduced regression when configuring with -O3 optimization #2678
Comments
hppritcha
referenced
this issue
Jan 6, 2017
Newer x86 processors have a core invariant tsc. On these systems it is safe to use the rtdtsc instruction as a monotonic timer. This commit adds a new function to the opal timer code to check if the timer backend is monotonic. On x86 it checks the appropriate bit and on other architectures it parrots back the OPAL_TIMER_MONOTONIC value. Signed-off-by: Nathan Hjelm <[email protected]>
|
Hmm, I tested this only with gcc on CTS-1. Optimized build (defaults to -O2). Will look at the backtrace and see what is going on. |
hjelmn
added a commit
to hjelmn/ompi
that referenced
this issue
Jan 7, 2017
This commit fixes a bug in the timer check. When -fPIC is used we need to save/restore ebx. The code copied from patcher was meant for 32-bit systems and did not work correctly on 64-bit systems. This commit updates the save/restore to use rbx instead of ebx. Fixes open-mpi#2678 Signed-off-by: Nathan Hjelm <[email protected]>
|
Rather embarrassing. The cpuid code I copied from patcher only saved/restored the lower 32 bits of register bx. PR open to fix. |
hjelmn
added a commit
to hjelmn/ompi
that referenced
this issue
Jan 7, 2017
This commit fixes a bug in the timer check. When -fPIC is used we need to save/restore ebx. The code copied from patcher was meant for 32-bit systems and did not work correctly on 64-bit systems. This commit updates the save/restore to use rbx instead of ebx. Fixes open-mpi#2678 Signed-off-by: Nathan Hjelm <[email protected]>
hjelmn
added a commit
to hjelmn/ompi
that referenced
this issue
Jan 7, 2017
This commit fixes a bug in the timer check. When -fPIC is used we need to save/restore ebx. The code copied from patcher was meant for 32-bit systems and did not work correctly on 64-bit systems. This commit updates the save/restore to use rbx instead of ebx. Fixes open-mpi#2678 Signed-off-by: Nathan Hjelm <[email protected]> (cherry picked from commit 5b70ae3) Signed-off-by: Nathan Hjelm <[email protected]>
hjelmn
added a commit
to hjelmn/ompi
that referenced
this issue
Jan 7, 2017
This commit fixes a bug in the timer check. When -fPIC is used we need to save/restore ebx. The code copied from patcher was meant for 32-bit systems and did not work correctly on 64-bit systems. This commit updates the save/restore to use rbx instead of ebx. Fixes open-mpi#2678 Signed-off-by: Nathan Hjelm <[email protected]> (cherry picked from commit 5b70ae3) Signed-off-by: Nathan Hjelm <[email protected]>
hjelmn
added a commit
to hjelmn/ompi
that referenced
this issue
Jan 7, 2017
This commit fixes a bug in the timer check. When -fPIC is used we need to save/restore ebx. The code copied from patcher was meant for 32-bit systems and did not work correctly on 64-bit systems. This commit updates the save/restore to use rbx instead of ebx. Fixes open-mpi#2678 Signed-off-by: Nathan Hjelm <[email protected]> (cherry picked from commit 5b70ae3) Signed-off-by: Nathan Hjelm <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
#2596, which has been merged back in to v2.x and v.2.0.x introduced a major regression. Open MPI can no longer be compiled with gcc (at least) at -O2 or -O3 for x86_64 without immediately segfaulting in MPI_Init. The problem has been observed both for gcc 4.9 and 6.3.
orterun also fails with a similar segfault pattern. If one uses the cpuid instruction directly (avoiding all the exb exchange stuff - which isn't necessary for x86_64), the problem vanishes.
See comments to commit a718743.
Although the problem was originally reported on a system with Opteron processors,
@marksantcroos thanks for catching this.
The text was updated successfully, but these errors were encountered: