Re: [PATCH] tools build: Use -Xpreprocessor instead of -Wp and leave pathnames intact

[Dave Martin]

On Fri, Apr 13, 2018 at 02:53:10PM +0100, Will Deacon wrote:
> Build.include invokes the pre-processor via GCC in order to generate a
> dependency list for the input file. Since these options are passed using
> '-Wp,-M...,$(depfile)' it is important that $(depfile) does not contain
> any commas, so these are substituted with underscores. This substitution
> will break the build if the directory name of the output directory happens
> to include a comma, e.g. when using "aiaiai" for bisection testing:
> 
>   | cc1: fatal error: x86/tools/objtool/fixdep.o: No such file or directory
>   | compilation terminated.
>   | cat: /tmp/aiaiai-test-patchset.qroS/before/obj.defconfig_x86/tools/objtool/.fixdep.o.d: No such file or directory
>   | make[5]: *** [tools/objtool/fixdep.o] Error 1
> 
> We can address this by using -Xpreprocessor instead of -Wp, which allows
> us to pass down an unmodified pathname.
> 
> Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
> Cc: Dave Martin <Dave.Martin@xxxxxxx>
> Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> Signed-off-by: Will Deacon <will.deacon@xxxxxxx>
> ---
> 
> As an aside, the way we currently pass the depfile to -MD appears to be
> in direct contradiction with the preprocessor documentation, although it
> does work with the cc1 implementation.

Hmmm, I try cc1 --help, and it gives

  ...

  -I <dir>
  -M
  -MD
  -MF <file>
  -MG
  -MM

  ...

so it looks like even cc1 shouldn't really be parsing a depfile name
argument after -MD.

The only way to get -MD <file> parsed in the undocumented way seems to
be with gcc -Wp,-MD,... or direct invocation of cc1.  The cpp
frontend, and the gcc frontend itself seem to follow the documentation
and don't parse <file> as the depfile name here:

[...]

> diff --git a/tools/build/Build.include b/tools/build/Build.include

We should probably address this everywhere when we've figured out what
to do.

> index 418871d02ebf..e1914f8e2328 100644
> --- a/tools/build/Build.include
> +++ b/tools/build/Build.include
> @@ -22,9 +22,7 @@ dot-target = $(dir $@).$(notdir $@)
>  basetarget = $(basename $(notdir $@))
>  
>  ###
> -# The temporary file to save gcc -MD generated dependencies must not
> -# contain a comma
> -depfile = $(subst $(comma),_,$(dot-target).d)
> +depfile = $(dot-target).d
>  
>  ###
>  # Check if both arguments has same arguments. Result is empty string if equal.
> @@ -89,12 +87,12 @@ if_changed = $(if $(strip $(any-prereq) $(arg-check)),                   \
>  # - per target C flags
>  # - per object C flags
>  # - BUILD_STR macro to allow '-D"$(variable)"' constructs
> -c_flags_1 = -Wp,-MD,$(depfile) -Wp,-MT,$@ $(CFLAGS) -D"BUILD_STR(s)=\#s" $(CFLAGS_$(basetarget).o) $(CFLAGS_$(obj))
> +c_flags_1 = -Xpreprocessor -MD -Xpreprocessor $(depfile) -Xpreprocessor -MT -Xpreprocessor $@ $(CFLAGS) -D"BUILD_STR(s)=\#s" $(CFLAGS_$(basetarget).o) $(CFLAGS_$(obj))
>  c_flags_2 = $(filter-out $(CFLAGS_REMOVE_$(basetarget).o), $(c_flags_1))
>  c_flags   = $(filter-out $(CFLAGS_REMOVE_$(obj)), $(c_flags_2))
> -cxx_flags = -Wp,-MD,$(depfile) -Wp,-MT,$@ $(CXXFLAGS) -D"BUILD_STR(s)=\#s" $(CXXFLAGS_$(basetarget).o) $(CXXFLAGS_$(obj))
> +cxx_flags = -Xpreprocessor -MD -Xpreprocessor $(depfile) -Xpreprocessor -MT -Xpreprocessor $@ $(CXXFLAGS) -D"BUILD_STR(s)=\#s" $(CXXFLAGS_$(basetarget).o) $(CXXFLAGS_$(obj))
>  
>  ###
>  ## HOSTCC C flags
>  
> -host_c_flags = -Wp,-MD,$(depfile) -Wp,-MT,$@ $(CHOSTFLAGS) -D"BUILD_STR(s)=\#s" $(CHOSTFLAGS_$(basetarget).o) $(CHOSTFLAGS_$(obj))
> +host_c_flags = -Xpreprocessor -MD -Xpreprocessor $(depfile) -Xpreprocessor -MT -Xpreprocessor $@ $(CHOSTFLAGS) -D"BUILD_STR(s)=\#s" $(CHOSTFLAGS_$(basetarget).o) $(CHOSTFLAGS_$(obj))

Any idea why we use -Wp here other than as a bug compatibility hack?
The gcc/clang support the depfile options directly.  It's possible that
gcc didn't support them, or didn't support -MF, sometime in the distant
past.  This use in the kernel makefiles predates git.

I'm wondering whether we should actually switch to using -M -MF, or -MD
-MF (strictly without -Wp or -Xpreprocessor) rather than relying on
a combination of undocumented interactions between -Wp and cc1, and
cc1 violating its own documentation.

Cheers
---Dave