More discussion at conda/conda#7031

I think we definitely regret that overlap. Unfortunately, it's now used in many recipes. It is not merely a matter of tweaking the compiler activation scripts. I'd like to change it, but we definitely can't undertake the recipe fixes right now.

I definitely understand the cost of fixing something like this. Since I'm not concerned with being able to build your recipes (most of the time), I'll experiment with having an extra activation script that resets HOST back to hostname (or do it in the script all of my teammates use to activate the one conda environment they typically use).

Thanks for the pointer to conda/conda/#7031. Let us know if there's anything we can do to help fix this.

I personally think we could change it, just need to coordinate that with everyone and I think we probably should.

A good start would be to define the new variables, so people can start switching over their recipes as soon as possible as part of a deprecation period for HOST.

Sooo. Trying to build strace, I realized how your recipes are using HOST and BUILD to hold the values that you have to pass through to configure --host=$HOST --build=$BUILD in order for autoconf to understand the pseudo-cross nature of the toolchains.

At one point during the early days of conda-build 3, you guys had variables that were set only when CONDA_BUILD was set in the environment. It looks like you've moved away from that and you're setting rpath all the time now (which is what I wanted, thanks).

I was confused. Looking at the gcc activation script instead of binutils, I see what I remembered and we are setting CFLAGS and LDFLAGS differently for CONDA_BUILD.

Could we:

1. Define the new variables.
2. Guard the setting of HOST with something like [[ "$CONDA_TOOLCHAIN_SET_HOST" = 1 || "$CONDA_BUILD" = 1 ]]

Having the CONDA_TOOLCHAIN_SET_HOST variable would let any users that have come to depend on this behavior outside of conda-build keep getting the behavior and conda-build could still have the behavior until everyone's recipes are cleaned up.

Also, I've had users ask that we have the behavior of CONDA_BUILD in the gcc activation scripts (i.e. that we have the conda include path and rpath by default. I could get that by setting CONDA_BUILD in our environment but I'd rather have a CONDA_TOOLCHAIN_USE_CONDA_INCLUDE_AND_RPATH variable that targets that specific behavior.