Drop non-profiling inputs to profiling build

[aherrmann]

The profiling builds took the non-profiling object file outputs as an additional input with a comment stating that these are required for template Haskell. However, it seems these inputs are unnecessary as removing them does not seem to negatively affect bazel test //tests/... -c dbg. As I understand, this is because we are passing -fexternal-interpreter on profiling builds. Removing that flag would indeed cause errors due to missing object files.

[guibou]

I fully agree with theses changes, however have a look at #438

There is two options for profiling:

• building twice (with and without profiling)
• building with -fexternal-interpreter

The current situation is that, starting from the first solution, the flag for the second solution was introduced in a totally unrelated commit, see #438 for all the details, and progressively the code evolved such that the first solution was not working anymore and that's what your are removing now (which is a good thing).

But considering that -fexternal-interpreter is ~ 15x times slower than the other approach when it involves template haskell, see https://gitlab.haskell.org/ghc/ghc/issues/16256 I have concerns.

That's a tradeoff between simplifying rules_haskell by fully switching to -fexternal-interpreter or removing the flag and fixing the previous approach. Pro and cons of approach:

• Fixing the first will be more work than the second
• The first generates more code in rules_haskell
• The second is slower
• The second will be faster if the bug if fixed upstream.

[guibou]

This can be extended by removing the objets generation code at all. But I think that bazel won't generate them because they are not part of the dependency graph anymore.

See my other comment before merging.

[aherrmann]

@guibou Thanks for the pointer. It's a tricky tradeoff. As you point out this PR doesn't change the current state of affairs.

This can be extended by removing the objets generation code at all.

I'm working on this for a separate PR. I'll continue the discussion there.