build: compile each backend archive with one section per procedure

CHANGELOG.md

@@ -20,6 +20,13 @@
 
 ### Changed
 
+* Each backend archive (`libhipfort-amdgcn.a`, `libhipfort-nvptx.a`) is
+  compiled with one section per procedure (`-ffunction-sections
+  -fdata-sections`). A program linked with `-Wl,--gc-sections` keeps only the
+  wrappers it calls, instead of all of a library's wrappers, and still links
+  when another wrapper names a C function that the installed library does not
+  export. The tests link that way. A shared library that links an archive also
+  needs `-Wl,--exclude-libs,ALL`.
 * **Breaking.** The array overloads of `hipMalloc`, `hipMallocManaged`,
   `hipHostMalloc`, `hipFree`, `hipHostFree`, the `hipMemcpy` families and the
   `hipHostRegister` family are now part of the Fortran 2008 array overloads, like

lib/CMakeLists.txt

@@ -115,6 +115,14 @@ if(HIPFORT_ASSUMED_RANK AND NOT (HIPFORT_USE_FPOINTER_INTERFACES AND CMAKE_Fortr
     "Use the F2018 assumed-rank array interfaces" FORCE)
 endif()
 
+# Each module compiles to one object, and a linker extracts whole objects from an
+# archive, so a program that calls one wrapper links all of that module's. One
+# section per procedure lets a program linked with -Wl,--gc-sections keep only
+# the wrappers it calls; without that flag it changes nothing. Only these
+# compilers split the sections (the Intel ones accept the flags but do not).
+set(HIPFORT_SECTION_FLAGS
+  "$<$<COMPILE_LANG_AND_ID:Fortran,GNU,LLVMFlang,Flang,NVHPC>:-ffunction-sections;-fdata-sections>")
+
 if(HIPFORT_BUILD_AMDGCN)
 set(HIPFORT_ARCH "amdgcn")
     # amdgcn
@@ -138,6 +146,7 @@ set(HIPFORT_ARCH "amdgcn")
     	target_compile_definitions(${HIPFORT_LIB} PRIVATE USE_ASSUMED_SHAPE)
     ENDIF()
     target_compile_definitions(${HIPFORT_LIB} PRIVATE _HIPFORT_ARCH='${HIPFORT_ARCH}')
+    target_compile_options(${HIPFORT_LIB} PRIVATE "${HIPFORT_SECTION_FLAGS}")
     # Install Target hipfort-amdgcn
     # CMAKE_INSTALL_LIBDIR/INCLUDEDIR already carry the toolchain-specific
     # subdirectory (see HIPFORT_MULTITOOLCHAIN_LAYOUT in the top-level
@@ -183,6 +192,7 @@ set(HIPFORT_ARCH "nvptx")
     ENDIF()
     target_compile_definitions(${HIPFORT_LIB} PRIVATE USE_CUDA_NAMES)
     target_compile_definitions(${HIPFORT_LIB} PRIVATE _HIPFORT_ARCH='${HIPFORT_ARCH}')
+    target_compile_options(${HIPFORT_LIB} PRIVATE "${HIPFORT_SECTION_FLAGS}")
     # Install Target hipfort-nvptx
     rocm_install_targets(
         TARGETS ${HIPFORT_LIB}

test/CMakeLists.txt

@@ -52,6 +52,14 @@ if(BUILD_TESTING)
            CMAKE_Fortran_PREPROCESS_SOURCE "${CMAKE_Fortran_PREPROCESS_SOURCE}")
   endif()
 
+  # Keep only the binding wrappers a test calls (the archives are built with one
+  # section per procedure), so a test links even when another wrapper names a C
+  # symbol the installed library does not export. Spelled -Wl, rather than
+  # LINKER:, which CMake 3.27 passes to LLVMFlang as a bare --gc-sections.
+  if(NOT APPLE AND NOT WIN32)
+    add_link_options(-Wl,--gc-sections)
+  endif()
+
   # Which backend the tests link: hipfort::hip carries the archive of the one
   # lib/ built the components for (nvptx when HIPFORT_BUILD_AMDGCN is OFF, even on
   # an AMD host). On the CUDA backend the symbol tests are compiled with