[AMDGPU] double free or corruption (!prev) in runTreeUpDown while linking RCCL

When building RCCL using LLVM 21 on Debian, I am seeing a crash in the linking stage. Admittedly, since RCCL is built using `-fgpu-rdc`, that is most of the compilation process. The crashing function is `void (anonymous namespace)::runTreeUpDown, ProtoSimple<1, 1, 2, 0, 0>, 2>(int, int, ncclDevWorkColl*)` as per the backtrace:

```
double free or corruption (!prev)
PLEASE submit a bug report to
https://github.com/llvm/llvm-project/issues/ and include the crash
backtrace.
Stack dump:
0. Running pass 'CallGraph Pass Manager' on module 'ld-temp.o'.
1. Running pass 'Machine Instruction Scheduler' on function
'@_ZN12_GLOBAL__N_113runTreeUpDownIj13FuncPreMulSumIjE11ProtoSimpleILi1ELi1ELi2ELi0ELi0EELi2EEEviiP15ncclDevWorkColl'
double free or corruption (!prev)
clang++-21: error: unable to execute command: Aborted
clang++-21: error: amdgcn-link command failed due to signal (use -v to
see invocation)
Debian clang version 21.1.4 (1)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/lib/llvm-21/bin
clang++-21: warning: treating 'c' input as 'c++' when in C++ mode, this
behavior is deprecated [-Wdeprecated]
clang++-21: warning: treating 'c' input as 'c++' when in C++ mode, this
behavior is deprecated [-Wdeprecated]
clang++-21: warning: treating 'c' input as 'c++' when in C++ mode, this
behavior is deprecated [-Wdeprecated]
clang++-21: note: diagnostic msg: Error generating preprocessed source(s).
failed to execute:/usr/bin/clang++-21 --driver-mode=g++ --hip-link
--offload-arch=gfx803 --offload-arch=gfx900 --offload-arch=gfx906
--offload-arch=gfx908 --offload-arch=gfx90a --offload-arch=gfx1010
--offload-arch=gfx1030 --offload-arch=gfx1100 --offload-arch=gfx1101
--offload-arch=gfx1102 -fPIC -gz -g -O2
-ffile-prefix-map=/root/rccl/rccl=. -Xarch_host -fstack-protector-strong
-fstack-clash-protection -Wformat -Werror=format-security -Xarch_host
-fcf-protection -Wdate-time -D_FORTIFY_SOURCE=2 -O3 -DNDEBUG -Xlinker
--dependency-file=CMakeFiles/rccl.dir/link.d -v -Wl,-z,relro -Wl,-z,now
-shared -Wl,-soname,librccl.so.1 -o "librccl.so.1.0"
CMakeFiles/rccl.dir/hipify/src/bootstrap.cc.o
CMakeFiles/rccl.dir/hipify/src/channel.cc.o
CMakeFiles/rccl.dir/hipify/src/collectives.cc.o
CMakeFiles/rccl.dir/hipify/src/debug.cc.o
CMakeFiles/rccl.dir/hipify/src/enqueue.cc.o
CMakeFiles/rccl.dir/hipify/src/group.cc.o
CMakeFiles/rccl.dir/hipify/src/init.cc.o
CMakeFiles/rccl.dir/hipify/src/init_nvtx.cc.o
CMakeFiles/rccl.dir/hipify/src/net.cc.o
CMakeFiles/rccl.dir/hipify/src/msccl.cc.o
CMakeFiles/rccl.dir/hipify/src/proxy.cc.o
CMakeFiles/rccl.dir/hipify/src/rccl_wrap.cc.o
CMakeFiles/rccl.dir/hipify/src/register.cc.o
CMakeFiles/rccl.dir/hipify/src/transport.cc.o
CMakeFiles/rccl.dir/hipify/src/device/common.cu.cpp.o
CMakeFiles/rccl.dir/hipify/src/device/onerank.cu.cpp.o
CMakeFiles/rccl.dir/hipify/src/graph/connect.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/paths.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/rings.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/rome_models.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/search.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/topo.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/trees.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/tuning.cc.o
CMakeFiles/rccl.dir/hipify/src/graph/xml.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/alt_rsmi.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/archinfo.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/argcheck.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/api_trace.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/ibvsymbols.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/ibvwrap.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/ipcsocket.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/npkit.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/nvmlwrap_stub.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/param.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/profiler.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/rocm_smi_wrap.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/rocmwrap.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/roctx.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/shmutils.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/signals.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/socket.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/strongstream.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/tuner.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/utils.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/msccl/msccl_lifecycle.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/msccl/msccl_parser.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/msccl/msccl_setup.cc.o
CMakeFiles/rccl.dir/hipify/src/misc/msccl/msccl_status.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/coll_net.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/generic.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/net_tmp.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/net_ib.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/net_socket.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/nvls.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/p2p.cc.o
CMakeFiles/rccl.dir/hipify/src/transport/shm.cc.o
CMakeFiles/rccl.dir/hipify/gensrc/all_gather_sum_i8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_minmax_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_premulsum_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_prod_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sum_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sumpostdiv_i32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sumpostdiv_i64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sumpostdiv_i8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sumpostdiv_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sumpostdiv_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/all_reduce_sumpostdiv_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/alltoall_pivot_sum_i8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/broadcast_sum_i8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/device_table.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/host_table.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_minmax_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_premulsum_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_prod_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_minmax_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_premulsum_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_prod_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sum_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sumpostdiv_i32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sumpostdiv_i64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sumpostdiv_i8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sumpostdiv_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sumpostdiv_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_scatter_sumpostdiv_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_bf16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_bf8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_f16.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_f32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_f64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_f8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sum_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sumpostdiv_i32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sumpostdiv_i64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sumpostdiv_i8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sumpostdiv_u32.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sumpostdiv_u64.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/reduce_sumpostdiv_u8.cpp.o
CMakeFiles/rccl.dir/hipify/gensrc/sendrecv_sum_i8.cpp.o
CMakeFiles/rccl.dir/git_version.cpp.o -fgpu-rdc -ldl
/usr/lib/x86_64-linux-gnu/librocm_smi64.so.1.0
/usr/lib/x86_64-linux-gnu/libamdhip64.so.5.7.31921 --hip-link
/usr/lib/llvm-21/lib/clang/21/lib/linux/libclang_rt.builtins-x86_64.a
-lpthread -lrt -ldl
```

This output was produced using clang-21 1:21.1.4-1 on Debian Unstable. I'll try to follow up on this issue after going thorugh with [the instructions on for identifying if this is a front-end, middle-end, or back-end issue](https://llvm.org/docs/HowToSubmitABug.html#crashes-the-compiler) and attempting to reproduce the issue with the official LLVM 21.1.4 release.

Originally reported by @tflink in [Debian Bug #1118478](https://bugs.debian.org/1118478).

Reproduce the failure with the reported clang++-21 --hip-link command and the RCCL build using LLVM 21 on Debian. Start with the amdgcn-link crash in the Machine Instruction Scheduler and the named runTreeUpDown function, then inspect the relevant LLVM compiler and AMDGPU linker paths. Done means the RCCL link completes without the double-free abort and the regression is covered by an appropriate test.

Written by the indexing model from the issue text.