[Mlir-commits] [mlir] [MLIR][SROA] Replace pattern based approach with a one-shot one (PR #85437)

[Christian Ulmann]

================
@@ -289,14 +293,18 @@ llvm.func @memcpy_src(%other_array: !llvm.ptr) -> i32 {
 
 // CHECK-LABEL: llvm.func @memcpy_double
 llvm.func @memcpy_double() -> i32 {
-  // CHECK-DAG: %[[ALLOCA_LEN:.*]] = llvm.mlir.constant(1 : i32) : i32
-  // CHECK-DAG: %[[MEMCPY_LEN:.*]] = llvm.mlir.constant(4 : i32) : i32
+  // CHECK: %[[ALLOCA_LEN:.*]] = llvm.mlir.constant(1 : i32) : i32
   %0 = llvm.mlir.constant(1 : i32) : i32
-  // CHECK-COUNT-2: = llvm.alloca %[[ALLOCA_LEN]] x i32
+  // CHECK: = llvm.alloca %[[ALLOCA_LEN]] x i32
+  // TODO: This should also disappear.
+  // CHECK: = llvm.alloca %[[ALLOCA_LEN]] x !llvm.array<1 x i32>
----------------
Dinistro wrote:

Found the issue:
This is due to changing from the rewriter to a builder. The rewriter was able to fold the GEPOp zero indexing GEPOp away, while this is not done be the builder.
Somehow, SROA is not happy with this intermediate GEP, which degrades it's performance in such cases.

I'll address this as part of the improvement efforts.

https://github.com/llvm/llvm-project/pull/85437