[libc] Make NVPTX benchmarks use LTO for linking

Summary:
Now that we can do LTO, we can make the benchmarks more accurate by
allowing optimization + inlining of the implementation.

Author: jhuber6

libc/benchmarks/gpu/CMakeLists.txt

@@ -24,6 +24,8 @@ function(add_benchmark benchmark_name)
       libc.src.stdio.printf
       ${BENCHMARK_DEPENDS}
     ${BENCHMARK_UNPARSED_ARGUMENTS}
+    COMPILE_OPTIONS
+      -flto
   )
   get_fq_target_name(${benchmark_name} fq_target_name)
   set(fq_build_target_name ${fq_target_name}.__build__)

libc/cmake/modules/LLVMLibCTestRules.cmake

@@ -474,6 +474,8 @@ function(add_integration_test test_name)
     target_link_options(${fq_build_target_name} PRIVATE
       ${LIBC_COMPILE_OPTIONS_DEFAULT} -Wno-multi-gpu
       "-Wl,--suppress-stack-size-warning"
+      "-Wl,-mllvm,-nvptx-lower-global-ctor-dtor=1"
+      "-Wl,-mllvm,-nvptx-emit-init-fini-kernel"
       -march=${LIBC_GPU_TARGET_ARCHITECTURE} -nostdlib -static
       "--cuda-path=${LIBC_CUDA_ROOT}")
   elseif(LIBC_CC_SUPPORTS_NOSTDLIBPP)
@@ -657,6 +659,8 @@ function(add_libc_hermetic test_name)
     target_link_options(${fq_build_target_name} PRIVATE
       ${LIBC_COMPILE_OPTIONS_DEFAULT} -Wno-multi-gpu
       "-Wl,--suppress-stack-size-warning"
+      "-Wl,-mllvm,-nvptx-lower-global-ctor-dtor=1"
+      "-Wl,-mllvm,-nvptx-emit-init-fini-kernel"
       -march=${LIBC_GPU_TARGET_ARCHITECTURE} -nostdlib -static
       "--cuda-path=${LIBC_CUDA_ROOT}")
   elseif(LIBC_CC_SUPPORTS_NOSTDLIBPP)