[flang][OMPIRBuilder][MLIR][llvm] Backend support for atomic control options

Author: anchuraj

flang/test/Driver/atomic-control-options.f90

@@ -0,0 +1,20 @@
+! REQUIRES: amdgpu-registered-target
+! RUN: %flang_fc1 -emit-llvm -triple amdgcn-amd-amdhsa -fopenmp -fopenmp-is-device -munsafe-fp-atomics %s -o -|FileCheck -check-prefix=UNSAFE-FP-ATOMICS %s
+! RUN: %flang_fc1 -emit-llvm -triple amdgcn-amd-amdhsa -fopenmp -fopenmp-is-device -fatomic-ignore-denormal-mode %s -o -|FileCheck -check-prefix=IGNORE-DENORMAL-MODE %s
+! RUN: %flang_fc1 -emit-llvm -triple amdgcn-amd-amdhsa -fopenmp -fopenmp-is-device -fatomic-fine-grained-memory %s -o -|FileCheck -check-prefix=FINE-GRAINED-MEMORY %s
+! RUN: %flang_fc1 -emit-llvm -triple amdgcn-amd-amdhsa -fopenmp -fopenmp-is-device -fatomic-remote-memory %s -o -|FileCheck -check-prefix=REMOTE-MEMORY %s
+program test
+    implicit none
+    integer :: A, threads
+    threads = 128
+    A = 0
+    !$omp target parallel num_threads(threads)
+    !$omp atomic
+    A =  A + 1
+    !$omp end target parallel
+end program test
+
+!UNSAFE-FP-ATOMICS: %{{.*}} = atomicrmw add ptr {{.*}}, i32 1 monotonic, align 4, !amdgpu.ignore.denormal.mode !{{.*}}, !amdgpu.no.fine.grained.memory !{{.*}}, !amdgpu.no.remote.memory !{{.*}}
+!IGNORE-DENORMAL-MODE: %{{.*}} = atomicrmw add ptr {{.*}}, i32 1 monotonic, align 4, !amdgpu.ignore.denormal.mode !{{.*}}, !amdgpu.no.fine.grained.memory !{{.*}}, !amdgpu.no.remote.memory !{{.*}}
+!FINE-GRAINED-MEMORY: %{{.*}} = atomicrmw add ptr {{.*}}, i32 1 monotonic, align 4, !amdgpu.no.remote.memory !{{.*}}
+!REMOTE-MEMORY: %{{.*}} = atomicrmw add ptr {{.*}}, i32 1 monotonic, align 4, !amdgpu.no.fine.grained.memory !{{.*}}

llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h

@@ -3286,7 +3286,8 @@ class OpenMPIRBuilder {
   emitAtomicUpdate(InsertPointTy AllocaIP, Value *X, Type *XElemTy, Value *Expr,
                    AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
                    AtomicUpdateCallbackTy &UpdateOp, bool VolatileX,
-                   bool IsXBinopExpr);
+                   bool IsXBinopExpr, bool IsIgnoreDenormalMode,
+                   bool IsFineGrainedMemory, bool IsRemoteMemory);
 
   /// Emit the binary op. described by \p RMWOp, using \p Src1 and \p Src2 .
   ///
@@ -3359,7 +3360,9 @@ class OpenMPIRBuilder {
   LLVM_ABI InsertPointOrErrorTy createAtomicUpdate(
       const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X,
       Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
-      AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr);
+      AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr,
+      bool IsIgnoreDenormalMode = false, bool IsFineGrainedMemory = false,
+      bool IsRemoteMemory = false);
 
   /// Emit atomic update for constructs: --- Only Scalar data types
   /// V = X; X = X BinOp Expr ,
@@ -3394,7 +3397,9 @@ class OpenMPIRBuilder {
       const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X,
       AtomicOpValue &V, Value *Expr, AtomicOrdering AO,
       AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp,
-      bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr);
+      bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr,
+      bool IsIgnoreDenormalMode = false, bool IsFineGrainedMemory = false,
+      bool IsRemoteMemory = false);
 
   /// Emit atomic compare for constructs: --- Only scalar data types
   /// cond-expr-stmt:

llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp

@@ -8957,7 +8957,8 @@ OpenMPIRBuilder::createAtomicWrite(const LocationDescription &Loc,
 OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createAtomicUpdate(
     const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X,
     Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
-    AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr) {
+    AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr,
+    bool IsIgnoreDenormalMode, bool IsFineGrainedMemory, bool IsRemoteMemory) {
   assert(!isConflictIP(Loc.IP, AllocaIP) && "IPs must not be ambiguous");
   if (!updateToLocation(Loc))
     return Loc.IP;
@@ -8975,9 +8976,9 @@ OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createAtomicUpdate(
            "OpenMP atomic does not support LT or GT operations");
   });
 
-  Expected<std::pair<Value *, Value *>> AtomicResult =
-      emitAtomicUpdate(AllocaIP, X.Var, X.ElemTy, Expr, AO, RMWOp, UpdateOp,
-                       X.IsVolatile, IsXBinopExpr);
+  Expected<std::pair<Value *, Value *>> AtomicResult = emitAtomicUpdate(
+      AllocaIP, X.Var, X.ElemTy, Expr, AO, RMWOp, UpdateOp, X.IsVolatile,
+      IsXBinopExpr, IsIgnoreDenormalMode, IsFineGrainedMemory, IsRemoteMemory);
   if (!AtomicResult)
     return AtomicResult.takeError();
   checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Update);
@@ -9024,7 +9025,8 @@ Value *OpenMPIRBuilder::emitRMWOpAsInstruction(Value *Src1, Value *Src2,
 Expected<std::pair<Value *, Value *>> OpenMPIRBuilder::emitAtomicUpdate(
     InsertPointTy AllocaIP, Value *X, Type *XElemTy, Value *Expr,
     AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
-    AtomicUpdateCallbackTy &UpdateOp, bool VolatileX, bool IsXBinopExpr) {
+    AtomicUpdateCallbackTy &UpdateOp, bool VolatileX, bool IsXBinopExpr,
+    bool IsIgnoreDenormalMode, bool IsFineGrainedMemory, bool IsRemoteMemory) {
   // TODO: handle the case where XElemTy is not byte-sized or not a power of 2
   // or a complex datatype.
   bool emitRMWOp = false;
@@ -9047,7 +9049,20 @@ Expected<std::pair<Value *, Value *>> OpenMPIRBuilder::emitAtomicUpdate(
 
   std::pair<Value *, Value *> Res;
   if (emitRMWOp) {
-    Res.first = Builder.CreateAtomicRMW(RMWOp, X, Expr, llvm::MaybeAlign(), AO);
+    AtomicRMWInst *RMWInst =
+        Builder.CreateAtomicRMW(RMWOp, X, Expr, llvm::MaybeAlign(), AO);
+    if (T.isAMDGPU()) {
+      if (IsIgnoreDenormalMode)
+        RMWInst->setMetadata("amdgpu.ignore.denormal.mode",
+                             llvm::MDNode::get(Builder.getContext(), {}));
+      if (!IsFineGrainedMemory)
+        RMWInst->setMetadata("amdgpu.no.fine.grained.memory",
+                             llvm::MDNode::get(Builder.getContext(), {}));
+      if (!IsRemoteMemory)
+        RMWInst->setMetadata("amdgpu.no.remote.memory",
+                             llvm::MDNode::get(Builder.getContext(), {}));
+    }
+    Res.first = RMWInst;
     // not needed except in case of postfix captures. Generate anyway for
     // consistency with the else part. Will be removed with any DCE pass.
     // AtomicRMWInst::Xchg does not have a coressponding instruction.
@@ -9179,7 +9194,8 @@ OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createAtomicCapture(
     const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X,
     AtomicOpValue &V, Value *Expr, AtomicOrdering AO,
     AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp,
-    bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr) {
+    bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr,
+    bool IsIgnoreDenormalMode, bool IsFineGrainedMemory, bool IsRemoteMemory) {
   if (!updateToLocation(Loc))
     return Loc.IP;
 
@@ -9198,9 +9214,9 @@ OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createAtomicCapture(
   // If UpdateExpr is 'x' updated with some `expr` not based on 'x',
   // 'x' is simply atomically rewritten with 'expr'.
   AtomicRMWInst::BinOp AtomicOp = (UpdateExpr ? RMWOp : AtomicRMWInst::Xchg);
-  Expected<std::pair<Value *, Value *>> AtomicResult =
-      emitAtomicUpdate(AllocaIP, X.Var, X.ElemTy, Expr, AO, AtomicOp, UpdateOp,
-                       X.IsVolatile, IsXBinopExpr);
+  Expected<std::pair<Value *, Value *>> AtomicResult = emitAtomicUpdate(
+      AllocaIP, X.Var, X.ElemTy, Expr, AO, AtomicOp, UpdateOp, X.IsVolatile,
+      IsXBinopExpr, IsIgnoreDenormalMode, IsFineGrainedMemory, IsRemoteMemory);
   if (!AtomicResult)
     return AtomicResult.takeError();
   Value *CapturedVal =

mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp

@@ -3269,13 +3269,25 @@ convertOmpAtomicUpdate(omp::AtomicUpdateOp &opInst,
     return moduleTranslation.lookupValue(yieldop.getResults()[0]);
   };
 
+  bool isIgnoreDenormalMode = false;
+  bool isFineGrainedMemory = false;
+  bool isRemoteMemory = false;
+  if (opInst->hasAttr(opInst.getAtomicControlAttrName())) {
+    mlir::omp::AtomicControlAttr atomicControlAttr =
+        opInst.getAtomicControlAttr();
+    isIgnoreDenormalMode = atomicControlAttr.getIgnoreDenormalMode();
+    isFineGrainedMemory = atomicControlAttr.getFineGrainedMemory();
+    isRemoteMemory = atomicControlAttr.getRemoteMemory();
+  }
+
   // Handle ambiguous alloca, if any.
   auto allocaIP = findAllocaInsertPoint(builder, moduleTranslation);
   llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
   llvm::OpenMPIRBuilder::InsertPointOrErrorTy afterIP =
       ompBuilder->createAtomicUpdate(ompLoc, allocaIP, llvmAtomicX, llvmExpr,
                                      atomicOrdering, binop, updateFn,
-                                     isXBinopExpr);
+                                     isXBinopExpr, isIgnoreDenormalMode,
+                                     isFineGrainedMemory, isRemoteMemory);
 
   if (failed(handleError(afterIP, *opInst)))
     return failure();
@@ -3364,13 +3376,26 @@ convertOmpAtomicCapture(omp::AtomicCaptureOp atomicCaptureOp,
     return moduleTranslation.lookupValue(yieldop.getResults()[0]);
   };
 
+  bool isIgnoreDenormalMode = false;
+  bool isFineGrainedMemory = false;
+  bool isRemoteMemory = false;
+  if (atomicUpdateOp &&
+      atomicUpdateOp->hasAttr(atomicUpdateOp.getAtomicControlAttrName())) {
+    mlir::omp::AtomicControlAttr atomicControlAttr =
+        atomicUpdateOp.getAtomicControlAttr();
+    isIgnoreDenormalMode = atomicControlAttr.getIgnoreDenormalMode();
+    isFineGrainedMemory = atomicControlAttr.getFineGrainedMemory();
+    isRemoteMemory = atomicControlAttr.getRemoteMemory();
+  }
+
   // Handle ambiguous alloca, if any.
   auto allocaIP = findAllocaInsertPoint(builder, moduleTranslation);
   llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
   llvm::OpenMPIRBuilder::InsertPointOrErrorTy afterIP =
       ompBuilder->createAtomicCapture(
           ompLoc, allocaIP, llvmAtomicX, llvmAtomicV, llvmExpr, atomicOrdering,
-          binop, updateFn, atomicUpdateOp, isPostfixUpdate, isXBinopExpr);
+          binop, updateFn, atomicUpdateOp, isPostfixUpdate, isXBinopExpr,
+          isIgnoreDenormalMode, isFineGrainedMemory, isRemoteMemory);
 
   if (failed(handleError(afterIP, *atomicCaptureOp)))
     return failure();

mlir/test/Target/LLVMIR/omptarget-atomic-control-options.mlir

@@ -0,0 +1,36 @@
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+// CHECK: atomicrmw add ptr %loadgep_, i32 1 monotonic, align 4, !amdgpu.ignore.denormal.mode !{{.*}}, !amdgpu.no.fine.grained.memory !{{.*}}, !amdgpu.no.remote.memory !{{.*}}
+
+module attributes {dlti.dl_spec = #dlti.dl_spec<!llvm.ptr = dense<64> : vector<4xi64>, !llvm.ptr<1> = dense<64> : vector<4xi64>, !llvm.ptr<2> = dense<32> : vector<4xi64>, !llvm.ptr<3> = dense<32> : vector<4xi64>, !llvm.ptr<4> = dense<64> : vector<4xi64>, !llvm.ptr<5> = dense<32> : vector<4xi64>, !llvm.ptr<6> = dense<32> : vector<4xi64>, !llvm.ptr<7> = dense<[160, 256, 256, 32]> : vector<4xi64>, !llvm.ptr<8> = dense<[128, 128, 128, 48]> : vector<4xi64>, !llvm.ptr<9> = dense<[192, 256, 256, 32]> : vector<4xi64>, i64 = dense<64> : vector<2xi64>, i1 = dense<8> : vector<2xi64>, i8 = dense<8> : vector<2xi64>, i16 = dense<16> : vector<2xi64>, i32 = dense<32> : vector<2xi64>, f16 = dense<16> : vector<2xi64>, f64 = dense<64> : vector<2xi64>, f128 = dense<128> : vector<2xi64>, "dlti.endianness" = "little", "dlti.legal_int_widths" = array<i32: 32, 64>, "dlti.stack_alignment" = 32 : i64, "dlti.alloca_memory_space" = 5 : ui64, "dlti.global_memory_space" = 1 : ui64>, fir.atomic_ignore_denormal_mode, fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", fir.target_cpu = "generic-hsa", llvm.data_layout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128:128:48-p9:192:256:256:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8:9", llvm.target_triple = "amdgcn-amd-amdhsa", omp.flags = #omp.flags<openmp_device_version = 31>, omp.is_gpu = true, omp.is_target_device = true, omp.requires = #omp<clause_requires none>, omp.target_triples = [], omp.version = #omp.version<version = 31>} {
+  llvm.func @_QQmain() attributes {fir.bindc_name = "TEST", omp.declare_target = #omp.declaretarget<device_type = (host), capture_clause = (to)>, target_cpu = "generic-hsa"} {
+    %0 = llvm.mlir.constant(1 : i64) : i64
+    %1 = llvm.alloca %0 x i32 {bindc_name = "threads"} : (i64) -> !llvm.ptr<5>
+    %2 = llvm.addrspacecast %1 : !llvm.ptr<5> to !llvm.ptr
+    %3 = llvm.mlir.constant(1 : i64) : i64
+    %4 = llvm.alloca %3 x i32 {bindc_name = "a"} : (i64) -> !llvm.ptr<5>
+    %5 = llvm.addrspacecast %4 : !llvm.ptr<5> to !llvm.ptr
+    %6 = llvm.mlir.constant(0 : i32) : i32
+    %7 = llvm.mlir.constant(128 : i32) : i32
+    %8 = llvm.mlir.constant(1 : i64) : i64
+    %9 = llvm.mlir.constant(1 : i64) : i64
+    llvm.store %7, %2 : i32, !llvm.ptr
+    llvm.store %6, %5 : i32, !llvm.ptr
+    %10 = omp.map.info var_ptr(%2 : !llvm.ptr, i32) map_clauses(implicit, exit_release_or_enter_alloc) capture(ByCopy) -> !llvm.ptr {name = "threads"}
+    %11 = omp.map.info var_ptr(%5 : !llvm.ptr, i32) map_clauses(implicit, exit_release_or_enter_alloc) capture(ByCopy) -> !llvm.ptr {name = "a"}
+    omp.target map_entries(%10 -> %arg0, %11 -> %arg1 : !llvm.ptr, !llvm.ptr) {
+      %12 = llvm.mlir.constant(1 : i32) : i32
+      %13 = llvm.load %arg0 : !llvm.ptr -> i32
+      omp.parallel num_threads(%13 : i32) {
+        omp.atomic.update %arg1 : !llvm.ptr {
+        ^bb0(%arg2: i32):
+          %14 = llvm.add %arg2, %12 : i32
+          omp.yield(%14 : i32)
+        } {atomic_control = #omp.atomic_control<ignore_denormal_mode = true>}
+        omp.terminator
+      }
+      omp.terminator
+    }
+    llvm.return
+  }
+}