[rocm-libraries] ROCm/rocm-libraries#7665 (commit 4767702)

[rocPRIM] build benchmark_device_adjacent_difference at -O1
 (#7665)

Build `benchmark_device_adjacent_difference` at `-O1` to work around
ROCm/llvm-project#2616: clang/llc hangs indefinitely in AMDGPU
SelectionDAG ISel on this TU at `-O2`/`-O3` for `gfx950`. The hang is a
DAG combine freeze cycle in shared SDAG code.

Confirmed: with `-O1` the TU builds in ~25 s instead of hanging. Other
opt levels and other targets are unaffected.

Revert this commit once the upstream codegen fix lands and is in
TheRock.

Addresses ROCM-24721

Author: lamb-j

CHANGELOG.md

@@ -108,6 +108,10 @@ Full documentation for rocPRIM is available at [https://rocm.docs.amd.com/projec
 * Added `rocprim::apply` which applies a function to a `rocprim::tuple`.
 
 
+### Known issues
+
+* benchmark_device_adjacent_difference build hangs due to https://github.com/ROCm/llvm-project/issues/2616.  Workaround is to build with -O1.
+
 ### Changed
 
 * Changed tests to support `ptr-to-const` output in `/test/rocprim/test_device_batch_memcpy.cpp`.

benchmark/CMakeLists.txt

@@ -172,6 +172,11 @@ add_rocprim_benchmark(benchmark_block_scan.cpp)
 add_rocprim_benchmark(benchmark_block_sort.cpp)
 add_rocprim_benchmark(benchmark_config_dispatch.cpp)
 add_rocprim_benchmark(benchmark_device_adjacent_difference.cpp)
+# Workaround for ROCm/llvm-project#2616: DAG combine cycle hangs codegen at
+# -O2/-O3 on gfx950 for this TU. Build at -O1 until the codegen fix lands.
+if(TARGET benchmark_device_adjacent_difference)
+  target_compile_options(benchmark_device_adjacent_difference PRIVATE -O1)
+endif()
 add_rocprim_benchmark(benchmark_device_adjacent_find.cpp)
 add_rocprim_benchmark(benchmark_device_batch_memcpy.cpp)
 add_rocprim_benchmark(benchmark_device_binary_search.cpp)