[PATCH] D33391: [DAGCombine] fold (add/uaddo (xor a, -1), 1) -> (sub 0, a)

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

Oh, oops, posted the wrong version.  Corrected:

  #include <stdio.h>
  __attribute((noinline)) int f(unsigned a, unsigned *ovf) {
     return __builtin_usub_overflow(0, a, ovf);
  }
  __attribute((noinline)) int f2(unsigned a, unsigned *ovf) {
    return __builtin_uadd_overflow(a ^ -1, 1, ovf);
  }
  int main() {
    unsigned sum, ovf;
    ovf = f(0, &sum); printf("%u %u\n", sum, ovf);
    ovf = f2(0, &sum); printf("%u %u\n", sum, ovf);
    ovf = f(1, &sum); printf("%u %u\n", sum, ovf);
    ovf = f2(1, &sum); printf("%u %u\n", sum, ovf);
  }

(I have no idea why clang generates such weird code for overloaded versions of the intrinsics.)


https://reviews.llvm.org/D33391