[Mlir-commits] [mlir] [mlir][LLVM] Prevent vector bit-width overflow (PR #220898)

[Mehdi Amini]

================
@@ -1589,9 +1589,13 @@ def LLVM_vector_insert
     "type($srcvec) `into` type($res)";
   let extraClassDeclaration = [{
     uint64_t getVectorBitWidth(Type vector) {
-      return getVectorNumElements(vector).getKnownMinValue() *
-             ::llvm::cast<VectorType>(vector).getElementType()
-                .getIntOrFloatBitWidth();
+      auto vectorType = ::llvm::cast<VectorType>(vector);
+      uint64_t numElements = vectorType.getDimSize(0);
----------------
joker-eph wrote:

Yes, it should be fine: for a scalable vector, `getDimSize(0)` is the known minimum element count; scalability is tracked separately. So this implementation preserves the previous `getKnownMinValue()` semantics, while avoiding the narrowing through `ElementCount`. I clarified that in the code (and added a test with scalable vec)
 

https://github.com/llvm/llvm-project/pull/220898