Dump compiler-determined GPU kernel resource usage (#1965)

Adds --kernel-resource-use flag to install.sh to allow dumping per-GPU kernel resource use at compile time (e.g., VGPRs, LDS, SGPRs, scratch, etc.)

CMakeLists.txt

@@ -35,6 +35,7 @@ option(ENABLE_NPKIT                            "Enable NPKit"
 option(ENABLE_IFC                              "Enable indirect function call"                 OFF)
 option(GENERATE_SYM_KERNELS                    "Generate symmetric memory kernels"             OFF)
 option(INSTALL_DEPENDENCIES                    "Force install dependencies"                    OFF)
+option(REPORT_KERNEL_RESOURCE_USE              "Append -Rpass-analysis=kernel to CXX flags"    OFF)
 option(ROCTX                                   "Enable ROCTX"                                  ON)
 option(PROFILE                                 "Enable profiling"                              OFF)
 option(TIMETRACE                               "Enable time-trace during compilation"          OFF)
@@ -1186,6 +1187,10 @@ if (HAVE_KERNARG_PRELOAD)
   target_compile_options(rccl PRIVATE -mllvm --amdgpu-kernarg-preload-count=16)
 endif()
 
+if (REPORT_KERNEL_RESOURCE_USE)
+  target_link_options(rccl PRIVATE -Rpass-analysis=kernel-resource-usage)
+endif()
+
 if (DUMP_ASM) # Save temporary files from kernel compilation
   message(STATUS "Disassembling librccl.so to asm")
   # Maintain symbols but without changing code.  Keep additional data in dwarf section of binary.