[libc] Do not find system binaries for offloading tools

[jhuber6]

Summary

We use `find_program` to identify a few programs we use for offloading.
Namely, `clang-offload-packger`, `amdgpu-arch`, and `nvptx-arch`.
Currently the logic allows these to bind to any tool matching this name,
so it will find it on the system. This meant that if the installation
was deleted or it found a broken binary the compilation would fail. We
should only pull these from the current LLVM binary directory.

[arsenm]

Can this try to use the imported target instead of searching

[jhuber6]

In D158203#4596285, @arsenm wrote:

Can this try to use the imported target instead of searching

So this is part of the runtimes which is a separate cmake job so it doesn't share the targets.

[arsenm]

In D158203#4596291, @jhuber6 wrote:

In D158203#4596285, @arsenm wrote:

Can this try to use the imported target instead of searching

So this is part of the runtimes which is a separate cmake job so it doesn't share the targets.

I'd still hope the runtimes build would use exported targets for these sorts of tools

[jhuber6]

In D158203#4596300, @arsenm wrote:

In D158203#4596291, @jhuber6 wrote:

In D158203#4596285, @arsenm wrote:

Can this try to use the imported target instead of searching

So this is part of the runtimes which is a separate cmake job so it doesn't share the targets.

I'd still hope the runtimes build would use exported targets for these sorts of tools

I'm not very familiar with that, do we export tools in this way? I thought it was all just binaries in a directory once you leave the main CMake invocation.

libc/cmake/modules/prepare_libc_gpu_build.cmake

 endif()
 if(NOT LLVM_LIBC_FULL_BUILD)
   message(FATAL_ERROR "LLVM_LIBC_FULL_BUILD must be enabled to build libc for "
                       "GPU.")
 endif()
 
 # Identify any locally installed AMD GPUs on the system using 'amdgpu-arch'.
 find_program(LIBC_AMDGPU_ARCH
-             NAMES amdgpu-arch
+             NAMES amdgpu-arch NO_DEFAULT_PATH
              PATHS ${LLVM_BINARY_DIR}/bin /opt/rocm/llvm/bin/)
 
 # Identify any locally installed NVIDIA GPUs on the system using 'nvptx-arch'.
 find_program(LIBC_NVPTX_ARCH
-             NAMES nvptx-arch
+             NAMES nvptx-arch NO_DEFAULT_PATH
              PATHS ${LLVM_BINARY_DIR}/bin)
 
 # Get the list of all natively supported GPU architectures.
 set(detected_gpu_architectures "")
 foreach(arch_tool ${LIBC_NVPTX_ARCH} ${LIBC_AMDGPU_ARCH})
   if(arch_tool)
     execute_process(COMMAND ${arch_tool}
                     OUTPUT_VARIABLE arch_tool_output
elseif(LIBC_GPU_ARCHITECTURES STREQUAL "native")
   endif()
   set(LIBC_GPU_ARCHITECTURES ${detected_gpu_architectures})
 endif()
 message(STATUS "Building libc for the following GPU architecture(s): "
                "${LIBC_GPU_ARCHITECTURES}")
 
 # Identify the program used to package multiple images into a single binary.
 find_program(LIBC_CLANG_OFFLOAD_PACKAGER
-             NAMES clang-offload-packager
+             NAMES clang-offload-packager NO_DEFAULT_PATH
              PATHS ${LLVM_BINARY_DIR}/bin)
 if(NOT LIBC_CLANG_OFFLOAD_PACKAGER)
   message(FATAL_ERROR "Cannot find the 'clang-offload-packager' for the GPU "
                       "build")
 endif()
 
 # Optionally set up a job pool to limit the number of GPU tests run in parallel.
 # This is sometimes necessary as running too many tests in parallel can cause