I have also found this. Here is the code I am using:

#include <stdint.h>

struct UInt96
  {
  uint32_t d2, d1, d0;
  };

void ADD96_v1(const UInt96& s, UInt96& d)
  {
  uint64_t sum = uint64_t(d.d0) + uint64_t(s.d0);
  d.d0 = sum; sum >>= 32;
  sum += uint64_t(d.d1) + uint64_t(s.d1);
  d.d1 = sum; sum >>= 32;
  sum += uint64_t(d.d2) + uint64_t(s.d2);
  d.d2 = sum;
  }

void ADD96_v2(const UInt96& s, UInt96& d)
  {
  uint32_t carry;

  d.d0 = __builtin_addc(d.d0, s.d0, 0, &carry);
  d.d1 = __builtin_addc(d.d1, s.d1, carry, &carry);
  d.d2 = __builtin_addc(d.d2, s.d2, carry, &carry);
  }

ADD96_v1 was my original function, and ADD96_v2 is my attempt to rewrite it using __builtin_addc. While the second version works, I am surprised that it is less optimal than the first version: using the Godbolt Compiler Explorer, the generated ARM code contains an additional three seemingly unnecessary instructions; worse the PPC code utilises branch operations!

My compiler options are:

-target arm-unknown-linux-gnueabihf -mcpu=cortex-a9 -O2

and:

-target powerpc-unknown-linux-gnu -mcpu=603e -O2

It seems that the __builtin_addc* functions don't really have an advantage; is that true?

FYI – https://reviews.llvm.org/D70079

Current Codegen: https://godbolt.org/z/cVSccY

rG257acbf6aee9 solves much of the poor codegen generically in DAGCombine

x86/arm code works as expected

ppc struggles as the target still needs to be updated to use ADDCARRY/SUBCARRY

Andy - maybe open a PPC specific bug?

Adding some PPC guys who might know whats best to do to improve PPC support and reduce the branching.

Do we need to extend the ISD::ADD/SUBCARRY combines (where possible) to better support ISD::ADD/SUBE or can we move PPC away from ISD::ADD/SUBC + ISD::ADD/SUBE?

We aren't currently working on moving away from the old carry setting/using nodes to the new ones, but we do not have any fundamental issue with doing so. I am not sure this statement is any help though.