[ltsmaster] Emit constructors as .init_array instead of .ctors when appropriate

[valpackett]

AArch64 only supports .init_array, for example. The reason this wasn't discovered earlier is that BFD and gold linkers automatically convert .ctors to .init_array as a performance optimization.

Based on clang code: https://github.com/llvm-mirror/clang/blob/5897428cd24e2deefbcc6f6744c0d7d233aa6747/lib/Driver/ToolChains/Gnu.cpp#L2598-L2620

ldc-developers/druntime#146

[kinke]

Thx; too bad that ugly piece of code isn't in LLVM itself; it apparently defaults to .init_array on Linux.

[valpackett]

it apparently defaults to .init_array on Linux

LLVM itself? hm, why does clang then have a check for the GCC installation on Linux:

(getTriple().getOS() == llvm::Triple::Linux &&
       ((!GCCInstallation.isValid() || !V.isOlderThan(4, 7, 0)) ||
getTriple().isAndroid()))

[kinke]

LLVM itself?

Yep, as can be seen by compiling with -output-s -mtriple={x86_64,aarch64}-linux-gnu; the object file sections are .{init,fini}_array, i.e., not converted by the linker.

why does clang then

Don't know, don't care. ;)

[JohanEngelen]

This looks OK to me, besides a missing test case. No need to test all the different possible triples, but a test with one triple and checking that setting that targetOption flag actually does the right thing would be good.
Any other concerns @kinke ?

[kinke]

Nope, no concerns from my side. We must not forget to cherry-pick it to master afterwards.

[JohanEngelen]

@myfreeweb Can you fix the build errors in the CI? (just use #ifdef to remove the checking of triples unknown to older LLVMs)

[valpackett]

Now passes on travis. Semaphore seems dead