Bug 235: Fix ICE with array slice assign.

[ibuclaw]

I've been thinking of every angle to tackle this. But actually, the best way not to fall into a myriad of edge cases is just to not return every static array or aggregate in a TARGET_EXPR.

There may be an opportunity to move this to d_save_expr, but for now I'm just going with the conservative choice.

[ibuclaw]

Still need to mark calls returning via NRVO with RETURN_SLOT_OPT.

Should only be applicable to INIT_EXPR assignments.

[ibuclaw]

@jpf91 - should be ready for review.

[jpf91]

LGTM. I assume you tested i686 manually? (the multilib configuration does not really run all 32 bit tests, IIRC the 32 bit unittests don't run in multilib configurations)

I'll test ARM as well, though that will probably take some time ;-)

[ibuclaw]

Yes, I checked and it doesn't create a self referencing target expr.

[jpf91]

OK, ARM tests passed. BTW: Is va_args C++ mangling architecture specific? I get this error with GCC 6.1.0 on ARM:

/tmp/ccKMJDyv.o: In function `myprintf(char const*, ...)':
cppa.d:(.text+0xcdc): undefined reference to `myvprintf(char const*, __va_list)'
/tmp/ccWjjwHd.o: In function `myvprintf(char const*, std::__va_list)':
cppb.cpp:(.text+0x52c): undefined reference to `myvprintfx(char const*, std::__va_list)'
collect2: error: ld returned 1 exit status
compiler exited with status 1
PASS: runnable/cppa.d (test for excess errors)

[ibuclaw]

Yes, it is indeed given an std namespace on arm. This is something to handle in a hook like Target::mangle in the frontend.