[llvm] [UniformityAnalysis] Track uniform values for conservative divergence queries (PR #180509)

[Pankaj Dwivedi via llvm-commits]

================
@@ -392,7 +402,15 @@ template <typename ContextT> class GenericUniformityAnalysisImpl {
   };
 
   /// \brief Whether \p Val is divergent at its definition.
-  bool isDivergent(ConstValueRefT V) const { return DivergentValues.count(V); }
+  /// When UniformValues is populated (after finalization), values not in
+  /// either set (e.g. newly created) are conservatively treated as divergent.
+  bool isDivergent(ConstValueRefT V) const {
+    if (ContextT::isNeverDivergent(V))
+      return false;
+    if (UniformValues.empty())
----------------
PankajDwivedi-25 wrote:

> If I follow everything, we're using `UniformValues.empty()` to differentiate the moment before/after `finalizeUniformValues` was called. Right?

Yes. 


> I think we could then have two versions of this method:
> 
> ```c++
> public:
> bool isDivergent(ConstValueRefT V) const {
>   if (ContextT::isNeverDivergent(V))
>     return false;
>   return !UniformValues.contains(V);
> }
> 
> protected:
> // only used during the construction of the analysis (names are difficult)
> bool isDivergentUnderConstruction(ConstValueRefT V) const {
>   if (ContextT::isNeverDivergent(V))
>     return false;
>   return DivergentValues.contains(V);
> }
> ```
> 
> During the construction of the analysis we only call (2). The users of the analysis only call (1).
> 
> After `finalizeUniformValues`, `DivergentValues` is not used anymore, so you can clear it

Yes, that would be a nice separation of these two uses of a single API.

It will require massive cleanup. 

https://github.com/llvm/llvm-project/pull/180509