[PATCH] D55057: [Headers] Make max_align_t match GCC's implementation.

Thu Nov 29 13:52:56 PST 2018

uweigand added inline comments.

================
Comment at: lib/Headers/__stddef_max_align_t.h:40
       __attribute__((__aligned__(__alignof__(long double))));
+#ifdef __i386__
+  __float128 __clang_max_align_nonce3
----------------
jyknight wrote:
> Can you fix clang to consistently define `__SIZEOF_FLOAT128__` in InitPreprocessor alongside the rest of the SIZEOF macros?
> 
> And then use that to determine whether to add float128 to the union? This change, as is, will break on any target which is i386 but doesn't define __float128, e.g. i386-unknown-unknown.
> 
> The only additional target which will be modified with that (that is: the only other target which has a float128 type, but for which max_align isn't already 16) is systemz-*-linux.
> 
> But I think that's actually indicating a pre-existing bug in the SystemZ config -- it's configured for a 16-byte long double, with 8-byte alignment, but the ABI doc seems to call for 16-byte alignment. +Ulrich for comment on that.
That's a bug in the ABI doc which we'll fix once we get around to releasing an updated version.

long double on SystemZ must be 8-byte aligned, which is the maximum alignment of all standard types on Z, and this was chosen because historically the ABI only guarantees an 8-byte stack alignment, so 16-byte aligned standard types would be awkward.

Repository:
  rC Clang

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D55057/new/

https://reviews.llvm.org/D55057