[Bug libgomp/126370] New: Reuse threads with same affinity

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126370

            Bug ID: 126370
           Summary: Reuse threads with same affinity
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libgomp
          Assignee: unassigned at gcc dot gnu.org
          Reporter: avieira at gcc dot gnu.org
                CC: jakub at gcc dot gnu.org
  Target Milestone: ---

Created attachment 65115
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65115&action=edit
reproducer

First off let me start saying I do not think this is a bug. The behavior
libgomp exhibits seems to be entirely compliant with the standard, it just
happens to differ from the LLVM and Intel implementation in a way that some
users believe to be undesirable.

The attached test.cpp case creates N threads (based on the number of cpus
available), then halves them and then sets them back to N. The N or half N here
is not extremely relevant as long as N > 4, the larger it is, the more likely
it is to show the result, in my case N = 16.

The behaviour we are trying to show here is that with libgomp the threads
between N and N/2, keep their affinity when using OMP_PROC_BIND, but they
change their thread ID and gettid. The thread ID's seem to be re-used between
those, and shuffled around, the gettid's are all higher. Whereas for libgomp
they stay the same.

using gomp:
g++ test.cpp -g -std=c++17 -O3 -fstrict-aliasing -xc++ -fopenmp -ffast-math
-lnuma -lgomp -pthread -o gcc_gomp

using llvm omp:
g++ -g -std=c++20 -O3 -fstrict-aliasing -xc++ -fopenmp -ffast-math -c test.cpp
-I /usr/lib/llvm-14/include/
g++ test.o -lnuma -pthread -lomp -L /usr/lib/llvm-14/lib -o gcc_llvmomp 


and then:
OMP_PROC_BIND=true ./gcc_gomp

...
thread 0: 281474842439712, gettid: 41819--> same thread id  --> same gettid
thread 1: 281474836918368, gettid: 41820--> same thread id  --> same gettid
thread 2: 281474828464224, gettid: 41821--> same thread id  --> same gettid
thread 3: 281474820010080, gettid: 41822--> same thread id  --> same gettid
thread 4: 281474811555936, gettid: 41823--> same thread id  --> same gettid
thread 5: 281474803101792, gettid: 41824--> same thread id  --> same gettid
thread 6: 281474794647648, gettid: 41825--> same thread id  --> same gettid
thread 7: 281474786193504, gettid: 41826--> same thread id  --> same gettid
thread 8: 281474769285216, gettid: 41835--> different thread id to:
281474777739360  --> different gettid to: 41827
thread 9: 281474752376928, gettid: 41836--> different thread id to:
281474769285216  --> different gettid to: 41828
thread 10: 281474735468640, gettid: 41837--> different thread id to:
281474760831072  --> different gettid to: 41829
thread 11: 281474727014496, gettid: 41838--> different thread id to:
281474752376928  --> different gettid to: 41830
thread 12: 281474760831072, gettid: 41839--> different thread id to:
281474743922784  --> different gettid to: 41831
thread 13: 281474777739360, gettid: 41840--> different thread id to:
281474735468640  --> different gettid to: 41832
thread 14: 281474743922784, gettid: 41841--> different thread id to:
281474727014496  --> different gettid to: 41833
thread 15: 281474718560352, gettid: 41842--> same thread id  --> different
gettid to: 41834
...

and:

OMP_PROC_BIND=true ./gcc_llvmomp

...
thread 0: 281474842439712, gettid: 41863--> same thread id  --> same gettid
thread 1: 281474836063584, gettid: 41864--> same thread id  --> same gettid
thread 2: 281474827605472, gettid: 41865--> same thread id  --> same gettid
thread 3: 281474819147360, gettid: 41866--> same thread id  --> same gettid
thread 4: 281474810689248, gettid: 41867--> same thread id  --> same gettid
thread 5: 281474802231136, gettid: 41868--> same thread id  --> same gettid
thread 6: 281474793773024, gettid: 41869--> same thread id  --> same gettid
thread 7: 281474785314912, gettid: 41870--> same thread id  --> same gettid
thread 8: 281474439833824, gettid: 41871--> same thread id  --> same gettid
thread 9: 281474431375712, gettid: 41872--> same thread id  --> same gettid
thread 10: 281474422917600, gettid: 41873--> same thread id  --> same gettid
thread 11: 281474414459488, gettid: 41874--> same thread id  --> same gettid
thread 12: 281474406001376, gettid: 41875--> same thread id  --> same gettid
thread 13: 281474397543264, gettid: 41876--> same thread id  --> same gettid
thread 14: 281474389085152, gettid: 41877--> same thread id  --> same gettid
thread 15: 281473902963808, gettid: 41878--> same thread id  --> same gettid