[clang] [C++26][clang] Implement P2795R5 'Erroneous behaviour for uninitialized reads' (PR #177614)

[Corentin Jabot via cfe-commits]

cor3ntin wrote:

Thanks. 

A few comments:
  - I feel like the constant evaluation warnings are a regression. As a user, i want to know that the variable was not initialized. whether it is indeterminate or not does not matter during constant evaluation.
I do not mind tracking it, but i don't think it should be user-observable. The only significant change in expr.const is
https://eel.is/c++draft/expr.const#9.8
So we should keep saying "variable 'x' is uninitialized when used here"

- The wording tries to be very specific about how bit_cast preserves indeterminate-ness, we probably should have tests for that. 

Then there is the load bearing aspect of this paper.
This sentence https://eel.is/c++draft/intro.abstract#7.sentence-1
> An implementation should issue a diagnostic when such an operation is executed[.](https://eel.is/c++draft/intro.abstract#7.sentence-1)

There are a couple valid interpretations
 - We should do static analysis (which you are doing)
 - We should diagnose at runtime somehow.

I we wanted to diagnose at runtime somehow, which could be work for a future PR, there are a few ways to go about it
 - Sanitizers should know about indeterminate and erroneous behavior (but we already diagnose reading indeterminate value in memsan)
 - We should check that this implementation does not make sanitizers worse
 - Debuggers might want to know about the indeterminate attribute 
 - Then there is the question of whether we want to force more instrumentation in clang somehow - although that is probably premature.








https://github.com/llvm/llvm-project/pull/177614