[PATCH] D79528: [ConstantFold] Optimize xor undef, undef to undef

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In D79528#2023931 <https://reviews.llvm.org/D79528#2023931>, @efriedma wrote:

> This is correct, strictly speaking, but has the situation with misguided users really improved any?  That same code using _mm_xor_ps() or whatever to zero __m128 variables is probably still out there, and clang still doesn't print any diagnostic by default if you try to do that.  And this still isn't really a useful fold.


Clang does recognize this given the right flags, but I agree there's more risk than reward here - folding to "0" should be almost as effective for optimization. There's code out there that will probably start misbehaving in very difficult-to-diagnose ways if we do this.

  $ cat xor.c 
  #include <xmmintrin.h>
  
  int f() {
    int x;
    return x ^ x;
  }
  
  __m128 g() {
    __m128 x;
    return _mm_xor_si128(x, x);
  }
  $ ./clang xor.c -c
  $ ./clang xor.c -c -Wall
  xor.c:5:10: warning: variable 'x' is uninitialized when used here [-Wuninitialized]
    return x ^ x;
           ^
  xor.c:4:8: note: initialize the variable 'x' to silence this warning
    int x;
         ^
          = 0
  xor.c:10:24: warning: variable 'x' is uninitialized when used here [-Wuninitialized]
    return _mm_xor_si128(x, x);
                         ^
  xor.c:9:3: note: variable 'x' is declared here
    __m128 x;
    ^
  2 warnings generated.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D79528/new/

https://reviews.llvm.org/D79528