[Mlir-commits] [mlir] Add non-temporal support for LLVM masked loads (PR #104598)

[Krzysztof Drewniak]

================
@@ -842,22 +842,26 @@ def LLVM_GetActiveLaneMaskOp
 /// Create a call to Masked Load intrinsic.
 def LLVM_MaskedLoadOp : LLVM_OneResultIntrOp<"masked.load"> {
   let arguments = (ins LLVM_AnyPointer:$data, LLVM_VectorOf<I1>:$mask,
-                   Variadic<LLVM_AnyVector>:$pass_thru, I32Attr:$alignment);
+                   Variadic<LLVM_AnyVector>:$pass_thru, I32Attr:$alignment,
+                   DefaultValuedAttr<BoolAttr, "false">:$nontemporal);
----------------
krzysz00 wrote:

... Also, if you're doing this, you might want to take a followup PR where you take `ScalarizeMaskedMemIntr` and have it copy over the `!nontemporal` (and other load-related metadata like alias analysis annotations) over to the loads

https://github.com/llvm/llvm-project/pull/104598