-fdiagnostics-color=always can sometimes be added when it isn't supported

[DavisVaughan]

When running pkgbuild::build(), if your C compiler supports this flag then it will get unconditionally added to an internal Makevars that it used while compiling the package. This can be problematic if your C++ compiler doesn't support this flag.

This situation arose when testing cpp11. We have one GHA workflow that tests on Ubuntu against an old gcc version, in particular, gcc 4.8. Note that support for this flag was added in gcc 4.9. The workflow was setting:

echo $'CXX1X=g++-4.8\nCXX11=g++-4.8' >> ~/.R/Makevars

which meant the C++ compiler was using gcc 4.8, but the C compiler was still using the default, which was a much newer version.

When the package was being built, the FAQ vignette which had {cpp11} knitr chunks in it failed to compile those chunks, because it tried to do so with the -fdiagnostics-color=always flag set, which wasn't recognized on this old version of gcc.

This prompted me to patch this in cpp11 by adding CC=gcc-4.8 to the Makevars as well, so it also looks like we are using an old C compiler r-lib/cpp11#279 (comment), but in theory I wouldn't have to do that.

build() sets the compiler flags here:

pkgbuild/R/build.R

Line 54 in 80762db

withr::local_makevars(compiler_flags(debug = FALSE), .assignment = "+=")

And here is where compiler_flags() adds it

pkgbuild/R/compiler-flags.R

Lines 51 to 59 in 80762db

	if (crayon::has_color() && has_compiler_colored_diagnostics()) {
	flags <- c(
	"CFLAGS", "CXXFLAGS", "CXX11FLAGS",
	"CXX14FLAGS", "CXX17FLAGS", "CXX20FLAGS"
	)
	res[flags] <- paste(res[flags], "-fdiagnostics-color=always")
	}
	res
	}

has_compiler_colored_diagnostics() returns TRUE if the C file that is compiled by has_compiler() compiles successfully while CFLAGS = "-fdiagnostics-color=always" is set

pkgbuild/R/compiler-flags.R

Line 76 in 80762db

res <- withr::with_makevars(c(CFLAGS = "-fdiagnostics-color=always"), has_compiler())

I wonder if it should only return TRUE if both a C and C++ file can successfully be compiled with color turned on? But I guess we also have to worry about exactly which version of C++ we are testing with (CXX11 vs CXX14, etc), so that makes this a little tough to check ahead of time.

[gaborcsardi]

I wonder if it should only return TRUE if both a C and C++ file can successfully be compiled with color turned on? But I guess we also have to worry about exactly which version of C++ we are testing with (CXX11 vs CXX14, etc), so that makes this a little tough to check ahead of time.

Yeah, we should do this. I would not worry too much about CXX11 etc. I think it would be pretty weird if CXX supports this flag and the other CXXxx compilers don't.

[DavisVaughan]

Would that have been the case above though? i.e. where we set

echo $'CXX1X=g++-4.8\nCXX11=g++-4.8' >> ~/.R/Makevars

But we didn't set CXX (which I'm assuming defaulted to a much newer version?)

[gaborcsardi]

Why don't you set CXX? I get it that it is not used by the package in this case, but I would guess that most people use the same compiler for everything, no?

[DavisVaughan]

I assume they mainly targeted CXX11 because cpp11 has SystemRequirements: C++11? I'm not sure why they didn't just use CXX instead if that would have also worked

[DavisVaughan]

Some notes here r-lib/cpp11#78 (comment)

[gaborcsardi]

Anyway, we can check CXX11 as well. Or check what the package uses.

[gaborcsardi]

It is not trivial to check for each CXX* compiler separately and it is also quite slow. So we are not going to do that now.

We can have an env var to opt out, that is probably enough to work around cases where a mixture of compiler versions is used.