[Mlir-commits] [mlir] [MLIR][LLVMIR] Fix llvm.extractvalue folder (PR #201838)

[Christian Ulmann]

================
@@ -185,6 +185,30 @@ llvm.func @fold_extract_sparse() -> f32 {
 
 // -----
 
+// CHECK-LABEL: no_fold_extract_splat_rank_mismatch
+llvm.func @no_fold_extract_splat_rank_mismatch() -> vector<2xi32> {
+  %0 = llvm.mlir.constant(dense<12> : vector<2xi32>) : vector<2xi32>
+  %1 = llvm.mlir.constant(dense<23> : vector<4x2xi32>) : !llvm.array<4 x vector<2xi32>>
+  // CHECK: extractvalue
+  %2 = llvm.extractvalue %1[0] : !llvm.array<4 x vector<2xi32>>
+  %3 = llvm.shl %0, %2 : vector<2xi32>
----------------
Dinistro wrote:

Nit: I'm not sure if we need the `llvm.shl` in the test. Can we not directly return the extracted value or does this struggle with materialization somehow?
I'm aware that the original issue showed up due to the shift, though. 

https://github.com/llvm/llvm-project/pull/201838