[PATCH] D54425: [AArch64] Add aarch64_vector_pcs function attribute to Clang

[John McCall via Phabricator via cfe-commits]

rjmccall added inline comments.


================
Comment at: lib/CodeGen/CGDebugInfo.cpp:1101
   case CC_AAPCS:
+  case CC_AArch64VectorCall:
     return llvm::dwarf::DW_CC_LLVM_AAPCS;
----------------
rnk wrote:
> sdesmalen wrote:
> > I wasn't really sure whether this requires a corresponding DW_CC_LLVM_AAVPCS record in LLVM, as I couldn't find much about the DW_CC_LLVM_  encodings, specifically whether they align with some agreed encoding that is implemented by GDB/LLDB. Is this defined anywhere, or is it ignored by debuggers at the moment?
> DWARF only allows encoding 256 conventions, and we grabbed 0xC[0-F], I guess for "clang", so we probably want to be careful about adding another. Do you anticipate making debuggers able to call such functions? If not, it's probably not worth it.
They probably should be callable.

It looks like DWARF reserves the first 64 conventions for general/language purposes and treats the rest of the range as "user" conventions.   If those conventions are assumed to be universally unique, that's a really limiting schema once you started dividing it up by vendor.  If I might make a suggestion, while there are certainly many calling conventions that are meant to have universal meaning (e.g. most language-specific conventions), there are also a large number that are inherently target-specific.  DWARF already uses a lot of numbers that only make sense in the context of a target (like register numbers); it would make sense for DWARF to carve out a range of the encoding space (maybe 16 or 32 numbers) for target-specific CCs.  This is hardly the first example; consider also all the variant ARM32 CCs or the i386 fastcall CC.


https://reviews.llvm.org/D54425