[MLIR] Fix `isValidIndex`

[frgossen]

Summary

Depends On D100634

[herhut]

Thanks. I can see this to be useful as we use -1 for unknown quite often.

[frgossen]

Thanks!
I ran into this where an index was constant-folded to -1 for an unreachable branch and the canonicalization for an extract op crashed subsequently.

mlir/lib/IR/BuiltinAttributes.cpp

bool ElementsAttr::isValidIndex(ArrayRef<uint64_t> index) const {
   if (rank == 0 && index.size() == 1 && index[0] == 0)
     return true;
   if (rank != static_cast<int64_t>(index.size()))
     return false;
 
   // Verify that all of the indices are within the shape dimensions.
   auto shape = type.getShape();
   return llvm::all_of(llvm::seq<int>(0, rank), [&](int i) {
-    return static_cast<int64_t>(index[i]) < shape[i];
+    int64_t dim = static_cast<int64_t>(index[i]);
+    return 0 <= dim && dim < shape[i];
   });
 }
 
 ElementsAttr
 ElementsAttr::mapValues(Type newElementType,
                         function_ref<APInt(const APInt &)> mapping) const {
   if (auto intOrFpAttr = dyn_cast<DenseElementsAttr>())
     return intOrFpAttr.mapValues(newElementType, mapping);