[Mlir-commits] [mlir] 3ae3fa4 - [mlir][llvm] Add nontemporal field to llvm.intr.masked.store (#205566)

[llvmlistbot at llvm.org]

Author: lonely eagle
Date: 2026-06-25T00:10:39+08:00
New Revision: 3ae3fa4642b7e9acd7167014ce6ed395186f99f7

URL: https://github.com/llvm/llvm-project/commit/3ae3fa4642b7e9acd7167014ce6ed395186f99f7
DIFF: https://github.com/llvm/llvm-project/commit/3ae3fa4642b7e9acd7167014ce6ed395186f99f7.diff

LOG: [mlir][llvm] Add nontemporal field to llvm.intr.masked.store (#205566)

Add `nontemporal` field to `llvm.intr.masked.store`. Since `nontemporal`
is a missing field for `llvm.intr.masked.store`, you can refer to
https://github.com/llvm/llvm-project/blob/e68e8d35c91b4fd3ba0ae3ef12d79b41d92580b2/llvm/test/CodeGen/AArch64/sve-nontemporal-masked-ldst.ll#L28.

---------

Co-authored-by: Tobias Gysi <tobias.gysi at nextsilicon.com>

Added: 
    

Modified: 
    mlir/include/mlir/Dialect/LLVMIR/LLVMIntrinsicOps.td
    mlir/test/Target/LLVMIR/Import/intrinsic.ll
    mlir/test/Target/LLVMIR/llvmir-intrinsics.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Dialect/LLVMIR/LLVMIntrinsicOps.td b/mlir/include/mlir/Dialect/LLVMIR/LLVMIntrinsicOps.td
index 92f0193cdadad..73629b95a06e5 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/LLVMIntrinsicOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/LLVMIntrinsicOps.td
@@ -1124,22 +1124,25 @@ def LLVM_MaskedLoadOp : LLVM_OneResultIntrOp<"masked.load"> {
 /// Create a call to Masked Store intrinsic.
 def LLVM_MaskedStoreOp : LLVM_ZeroResultIntrOp<"masked.store"> {
   let arguments = (ins LLVM_AnyVector:$value, LLVM_AnyPointer:$data,
-                   LLVM_VectorOf<I1>:$mask, I32Attr:$alignment);
+                   LLVM_VectorOf<I1>:$mask, I32Attr:$alignment,
+                   UnitAttr:$nontemporal);
   let builders = [LLVM_VoidResultTypeOpBuilder, LLVM_ZeroResultOpBuilder];
   let assemblyFormat = "$value `,` $data `,` $mask attr-dict `:` "
     "type($value) `,` type($mask) `into` qualified(type($data))";
 
   string llvmBuilder = [{
-    builder.CreateMaskedStore(
+    auto *inst = builder.CreateMaskedStore(
       $value, $data, llvm::Align($alignment), $mask);
-  }];
+  }] #setNonTemporalMetadataCode;
   string mlirBuilder = [{
     auto *intrinInst = dyn_cast<llvm::IntrinsicInst>(inst);
     llvm::Align alignment = intrinInst->getParamAlign(1).valueOrOne();
+    bool nontemporal = intrinInst->hasMetadata(llvm::LLVMContext::MD_nontemporal);
     $_op = LLVM::MaskedStoreOp::create($_builder, $_location,
-      $value, $data, $mask, $_builder.getI32IntegerAttr(alignment.value()));
+      $value, $data, $mask, $_builder.getI32IntegerAttr(alignment.value()),
+      nontemporal ? $_builder.getUnitAttr() : nullptr);
   }];
-  list<int> llvmArgIndices = [0, 1, 2, -1];
+  list<int> llvmArgIndices = [0, 1, 2, -1, -1];
 }
 
 /// Create a call to Masked Gather intrinsic.

diff  --git a/mlir/test/Target/LLVMIR/Import/intrinsic.ll b/mlir/test/Target/LLVMIR/Import/intrinsic.ll
index 959a04fff6dca..d3427b4ab9771 100644
--- a/mlir/test/Target/LLVMIR/Import/intrinsic.ll
+++ b/mlir/test/Target/LLVMIR/Import/intrinsic.ll
@@ -517,6 +517,9 @@ define void @masked_load_store_intrinsics(ptr %vec, <7 x i1> %mask) {
   ; CHECK:  llvm.intr.masked.store %[[VAL2]], %[[VEC]], %[[MASK]] {alignment = 8 : i32}
   ; CHECK-SAME:  vector<7xf32>, vector<7xi1> into !llvm.ptr
   call void @llvm.masked.store.v7f32.p0(<7 x float> %2, ptr %vec, i32 8, <7 x i1> %mask)
+  ; CHECK:  llvm.intr.masked.store %[[VAL2]], %[[VEC]], %[[MASK]] {alignment = 8 : i32, nontemporal}
+  ; CHECK-SAME:  vector<7xf32>, vector<7xi1> into !llvm.ptr 
+  call void @llvm.masked.store.v7f32.p0(<7 x float> %2, ptr %vec, i32 8, <7 x i1> %mask), !nontemporal !{i32 1}
   ret void
 }
 

diff  --git a/mlir/test/Target/LLVMIR/llvmir-intrinsics.mlir b/mlir/test/Target/LLVMIR/llvmir-intrinsics.mlir
index 293a56a82b23c..ea393dd445eda 100644
--- a/mlir/test/Target/LLVMIR/llvmir-intrinsics.mlir
+++ b/mlir/test/Target/LLVMIR/llvmir-intrinsics.mlir
@@ -566,6 +566,9 @@ llvm.func @masked_load_store_intrinsics(%A: !llvm.ptr, %mask: vector<7xi1>) {
   // CHECK: call void @llvm.masked.store.v7f32.p0(<7 x float> %{{.*}}, ptr align 1 %0, <7 x i1> %{{.*}})
   llvm.intr.masked.store %b, %A, %mask { alignment = 1: i32} :
     vector<7xf32>, vector<7xi1> into !llvm.ptr
+  // CHECK: call void @llvm.masked.store.v7f32.p0(<7 x float> %{{.*}}, ptr align 1 %0, <7 x i1> %{{.*}}), !nontemporal !{{.*}}
+  llvm.intr.masked.store %b, %A, %mask { alignment = 1: i32, nontemporal} :
+    vector<7xf32>, vector<7xi1> into !llvm.ptr
   llvm.return
 }