[PATCH] D149550: [clang][Interp] Fix compound assign operator evaluation order

Thu May 4 12:53:05 PDT 2023

shafik added a subscriber: rsmith.
shafik added inline comments.

================
Comment at: clang/lib/AST/Interp/ByteCodeExprGen.cpp:683-685
+  // C++17 onwards require that we evaluate the RHS first.
+  // Compute RHS and save it in a temporary variable so we can
+  // load it again later.
----------------
aaron.ballman wrote:
> tbaeder wrote:
> > aaron.ballman wrote:
> > > tbaeder wrote:
> > > > tbaeder wrote:
> > > > > aaron.ballman wrote:
> > > > > > In C, the evaluation of the operands are unsequenced. C doesn't currently have constexpr functions, only constexpr objects, but that eliminates mutating operations like compound assignment... for the moment. Perhaps a FIXME comment for figuring out how to handle C?
> > > > > > 
> > > > > > (The situation I'm worried about in C is with UB dealing with unsequenced operations, like rejecting: https://godbolt.org/z/W11jchrKc)
> > > > > Could C make them sequenced when introducing constexpr functions? :)
> > > > Since we already emit a warning for this, we could in the future just check if the statement is in a constexpr function and emit an error instead? We're emitting the warning for c++ pre-17 as well but we don't make it an error, I guess because it's not UB there?
> > > If we're actually leaving the operations unsequenced before C++17, then we should reject that code because it is UB: http://eel.is/c++draft/basic#intro.execution-10
> > > 
> > > The wording in C++14 for assignment operations is:
> > > > In all cases, the assignment is sequenced after the value computation of the right and left operands, and before the value computation of the assignment expression.
> > > 
> > > So the left and right operands are unsequenced relative to one another.
> > > 
> > I was looking at the existing implementation when writing this patch: https://github.com/llvm/llvm-project/blob/1a7a00bdc99fa2b2ca19ecd2d1069991b3c1006b/clang/lib/AST/ExprConstant.cpp#L8545-L8568 which always seems to evaluate RHS first (and actually abort for C++ <= 14, but for unrelated reasons, probably because this statement is just not supported in a constexpr context there).
> I think the existing implementation is incorrect to accept this in C++14 and earlier:
> 
> https://eel.is/c++draft/basic.exec#intro.execution-10.sentence-4
> http://eel.is/c++draft/expr.const#5.8
I filed a bug on this a while ago that we don't catch unsequenced modifications in constant expressions contexts: https://github.com/llvm/llvm-project/issues/37768

I am almost sure I had a conversation with @rsmith about this.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D149550/new/

https://reviews.llvm.org/D149550