[clang] Enable clang build with LLVM_BUILD_INSTRUMENTED without setting LLVM_PROFTDATA

[alexander-shaposhnikov]

Summary

At the moment if LLVM_BUILD_INSTRUMENTED is set to True
one has to set LLVM_PROFTDATA even if it's not necessary (because of message(FATAL_ERROR ...)).
Building instrumented Clang is useful even if one doesn't plan to use the target generate-profdata
(currently that target would use only llvm/tools/clang/utils/perf-training/cxx/hello_world.cpp as a source).
For example, one can run the instrumented version of Clang via a separate build system against a different codebase,
collect the profiles and merge them by lllvm-profdata later.

[vsk]

llvm-profdata is tightly coupled with the host compiler: while this setup may work if you get lucky, I don't think it's resilient to changes in libProfData. Also, using the instrumented llvm-profdata will be slow and create extra profiles.

As an alternative, have you considered building an llvm-profdata compatible with your host compiler in a separate step?

[alexander-shaposhnikov]

yeah, i agree, this is not a good idea. My thoughts were different - right now it's not particularly convenient that one has to specify LLVM_PROFDATA when it's not actually used by the build.
Maybe we can create the target "generate-profdata" only if LLVM_PROFDATA is set (but don't fail otherwise) ?

[vsk]

In D38859#896517, @alexshap wrote:

yeah, i agree, this is not a good idea. My thoughts were different - right now it's not particularly convenient that one has to specify LLVM_PROFDATA when it's not actually used by the build.
Maybe we can create the target "generate-profdata" only if LLVM_PROFDATA is set (but don't fail otherwise) ?

+ 1, sgtm

[alexander-shaposhnikov]

Update the wording

[vsk]

Thanks!

cfe/trunk/utils/perf-training/CMakeLists.txt

add_custom_target(clear-profraw
     COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py clean ${CMAKE_CURRENT_BINARY_DIR} profraw
     COMMENT "Clearing old profraw data")
 
   if(NOT LLVM_PROFDATA)
     find_program(LLVM_PROFDATA llvm-profdata)
   endif()
 
   if(NOT LLVM_PROFDATA)
-    message(FATAL_ERROR "Must set LLVM_PROFDATA to point to llvm-profdata to use for merging PGO data")
-  endif()
-
+    message(STATUS "To enable merging PGO data LLVM_PROFDATA has to point to llvm-profdata")
+  else()
     add_custom_target(generate-profdata
       COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py merge ${LLVM_PROFDATA} ${CMAKE_CURRENT_BINARY_DIR}/clang.profdata ${CMAKE_CURRENT_BINARY_DIR}
       COMMENT "Merging profdata"
       DEPENDS generate-profraw)
   endif()
+endif()
 
 find_program(DTRACE dtrace)
 if(APPLE AND DTRACE)
   configure_lit_site_cfg(
     ${CMAKE_CURRENT_SOURCE_DIR}/order-files.lit.site.cfg.in
     ${CMAKE_CURRENT_BINARY_DIR}/order-files/lit.site.cfg
     )