[Mlir-commits] [mlir] Atomic Options Support (PR #140932)

[Anchu Rajendran S]

https://github.com/anchuraj created https://github.com/llvm/llvm-project/pull/140932

Adding support for atomic control options for amd gpus. These are to support options `-f[no-]atomic-remote-memory, -f[no-]atomic-fine-grained-memory,
-f[no-]atomic-ignore-denormal-mode`



>From 28b688e72b6cb91777cdfc56413ff66fc02ceb18 Mon Sep 17 00:00:00 2001
From: Anchu Rajendran <asudhaku at amd.com>
Date: Tue, 20 May 2025 16:27:15 -0500
Subject: [PATCH] Atomic Options Support

---
 mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td | 32 +++++++++++++++++----
 mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp  |  7 ++++-
 mlir/test/Target/LLVMIR/llvmir.mlir         |  9 ++++++
 3 files changed, 42 insertions(+), 6 deletions(-)

diff --git a/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td b/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
index f19f9d5a3083c..7da8368c98b7c 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
@@ -2164,12 +2164,18 @@ def LLVM_AtomicRMWOp : LLVM_MemAccessOpBase<"atomicrmw", [
               LLVM_AtomicRMWType:$val, AtomicOrdering:$ordering,
               OptionalAttr<StrAttr>:$syncscope,
               OptionalAttr<I64Attr>:$alignment,
-              UnitAttr:$volatile_);
+              UnitAttr:$volatile_,
+              UnitAttr:$amdgpu_no_fine_grained_memory,
+              UnitAttr:$amdgpu_ignore_denormal_mode,
+              UnitAttr:$amdgpu_no_remote_memory);
   // Append the aliasing related attributes defined in LLVM_MemAccessOpBase.
   let arguments = !con(args, aliasAttrs);
   let results = (outs LLVM_AtomicRMWType:$res);
   let assemblyFormat = [{
-    (`volatile` $volatile_^)? $bin_op $ptr `,` $val
+    (`volatile` $volatile_^)?
+    (`amdgpu_no_fine_grained_memory` $amdgpu_no_fine_grained_memory^)?
+    (`amdgpu_ignore_denormal_mode` $amdgpu_ignore_denormal_mode^)?
+    (`amdgpu_no_remote_memory` $amdgpu_no_remote_memory^)? $bin_op $ptr `,` $val
     (`syncscope` `(` $syncscope^ `)`)? $ordering attr-dict `:`
     qualified(type($ptr)) `,` type($val)
   }];
@@ -2179,6 +2185,18 @@ def LLVM_AtomicRMWOp : LLVM_MemAccessOpBase<"atomicrmw", [
         convertAtomicBinOpToLLVM($bin_op), $ptr, $val, llvm::MaybeAlign(),
         convertAtomicOrderingToLLVM($ordering));
     $res = inst;
+    if($amdgpu_no_fine_grained_memory) {
+      llvm::MDNode *metadata = llvm::MDNode::get(inst->getContext(), std::nullopt);
+      inst->setMetadata((inst->getContext()).getMDKindID("amdgpu.no.fine.grained.memory"), metadata);
+    }
+    if($amdgpu_ignore_denormal_mode) {
+      llvm::MDNode *metadata = llvm::MDNode::get(inst->getContext(), std::nullopt);
+      inst->setMetadata((inst->getContext()).getMDKindID("amdgpu.ignore.denormal.mode"), metadata);
+    }
+    if($amdgpu_no_remote_memory) {
+      llvm::MDNode *metadata = llvm::MDNode::get(inst->getContext(), std::nullopt);
+      inst->setMetadata((inst->getContext()).getMDKindID("amdgpu.no.remote.memory"), metadata);
+    }
   }] # setVolatileCode
      # setSyncScopeCode
      # setAlignmentCode
@@ -2192,12 +2210,16 @@ def LLVM_AtomicRMWOp : LLVM_MemAccessOpBase<"atomicrmw", [
         convertAtomicOrderingFromLLVM(atomicInst->getOrdering()),
         getLLVMSyncScope(atomicInst), alignment, atomicInst->isVolatile());
   }];
-  list<int> llvmArgIndices = [-1, 0, 1, -1, -1, -1, -1, -1, -1, -1, -1];
+  list<int> llvmArgIndices = [-1, 0, 1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
+                              -1];
   let builders = [
-    OpBuilder<(ins "LLVM::AtomicBinOp":$binOp, "Value":$ptr, "Value":$val,
+      OpBuilder<(ins "LLVM::AtomicBinOp":$binOp, "Value":$ptr,"Value":$val,
       "LLVM::AtomicOrdering":$ordering,
       CArg<"StringRef", "StringRef()">:$syncscope,
-      CArg<"unsigned", "0">:$alignment, CArg<"bool", "false">:$isVolatile
+      CArg<"unsigned", "0">:$alignment, CArg<"bool", "false">:$isVolatile,
+      CArg<"bool", "false">:$isAmdgpuNoFineGrainedMemory,
+      CArg<"bool", "false">:$isAmdgpuIgnoreDenormalMode,
+      CArg<"bool", "false">:$isAmdgpuNoRemoteMemory
     )>
   ];
   let hasVerifier = 1;
diff --git a/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp b/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
index e17b9fd6eb8d3..4abff0320b645 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
@@ -3274,10 +3274,15 @@ OpFoldResult LLVM::ConstantOp::fold(FoldAdaptor) { return getValue(); }
 void AtomicRMWOp::build(OpBuilder &builder, OperationState &state,
                         AtomicBinOp binOp, Value ptr, Value val,
                         AtomicOrdering ordering, StringRef syncscope,
-                        unsigned alignment, bool isVolatile) {
+                        unsigned alignment, bool isVolatile,
+                        bool isAmdgpuNoFineGrainedMemory,
+                        bool isAmdgpuIgnoreDenormalMode,
+                        bool isAmdgpuNoRemoteMemory) {
   build(builder, state, val.getType(), binOp, ptr, val, ordering,
         !syncscope.empty() ? builder.getStringAttr(syncscope) : nullptr,
         alignment ? builder.getI64IntegerAttr(alignment) : nullptr, isVolatile,
+        isAmdgpuNoFineGrainedMemory, isAmdgpuIgnoreDenormalMode,
+        isAmdgpuNoRemoteMemory,
         /*access_groups=*/nullptr,
         /*alias_scopes=*/nullptr, /*noalias_scopes=*/nullptr, /*tbaa=*/nullptr);
 }
diff --git a/mlir/test/Target/LLVMIR/llvmir.mlir b/mlir/test/Target/LLVMIR/llvmir.mlir
index 4ef68fa83a70d..5b5a70d8cc76c 100644
--- a/mlir/test/Target/LLVMIR/llvmir.mlir
+++ b/mlir/test/Target/LLVMIR/llvmir.mlir
@@ -1568,6 +1568,15 @@ llvm.func @atomicrmw(
   // CHECK-SAME:  syncscope("singlethread")
   // CHECK-SAME:  align 8
   %27 = llvm.atomicrmw volatile udec_wrap %i32_ptr, %i32 syncscope("singlethread") monotonic {alignment = 8 : i64} : !llvm.ptr, i32
+  // CHECK: atomicrmw
+  // CHECK-SAME: !amdgpu.no.fine.grained.memory
+  %28 = llvm.atomicrmw amdgpu_no_fine_grained_memory udec_wrap %i32_ptr, %i32 monotonic {alignment = 8 : i64} : !llvm.ptr, i32
+  // CHECK: atomicrmw
+  // CHECK-SAME: !amdgpu.no.remote.memory
+  %29 = llvm.atomicrmw amdgpu_no_remote_memory udec_wrap %i32_ptr, %i32 monotonic {alignment = 8 : i64} : !llvm.ptr, i32
+  // CHECK: atomicrmw
+  // CHECK-SAME: !amdgpu.ignore.denormal.mode
+  %30 = llvm.atomicrmw amdgpu_ignore_denormal_mode udec_wrap %i32_ptr, %i32 monotonic {alignment = 8 : i64} : !llvm.ptr, i32
   llvm.return
 }