What if e1 or e2 have a side effect? They should be a save_expr in that case.

Actually, to clarify, they should be saved before extracting the real and imag parts.

So can I just move d_save_expr out of build_float_identity and apply it on the 'top-level' expressions?

OK.

So can I just move d_save_expr out of build_float_identity and apply it on the 'top-level' expressions?

Right. You don't need to save anything for ordinary float comparison (I might be forgetting something though).

You don't need to save anything for ordinary float comparison (I might be forgetting something though).

The original code did save float expressions though, so I've also kept that in the commit I just pushed.

Is this ready for gcc-patches?

I don't see why not. You could always just post to gcc-patches anyway and discuss the patch and make changes there. ;-)

I am yet to get round to fixing up this repo to use as a mirror/ci.

@ibuclaw

• Patch for first commit posted here: https://gcc.gnu.org/ml/gcc-patches/2018-11/msg02019.html
• Backport of [GDC] Do not include padding when hashing floating point types dlang/druntime#2356 in second commit, for CI testing. Only had to change some static foreach statements, everything else is 1:1 upstream. The allSatisfy function was taken from our stable GDC phobos, not upstream druntime (because of static foreach).
• Third commit removes special cases from Makefiles

Note that the XGDC bugs in convert.d seem to be fixed as well, or at least I can't reproduce these on x86_64.