llvm · skatrak · Feb 2, 2024 · Jan 16, 2024
diff --git a/flang/include/flang/Tools/CrossToolHelpers.h b/flang/include/flang/Tools/CrossToolHelpers.h
@@ -109,17 +109,6 @@ void setOffloadModuleInterfaceAttributes(
   }
 }
 
-//  Shares assinging of the OpenMP OffloadModuleInterface and its TargetCPU
-//  attribute accross Flang tools (bbc/flang)
-void setOffloadModuleInterfaceTargetAttribute(mlir::ModuleOp &module,
-    llvm::StringRef targetCPU, llvm::StringRef targetFeatures) {
-  // Should be registered by the OpenMPDialect
-  if (auto offloadMod = llvm::dyn_cast<mlir::omp::OffloadModuleInterface>(
-          module.getOperation())) {
-    offloadMod.setTarget(targetCPU, targetFeatures);
-  }
-}
-
 void setOpenMPVersionAttribute(mlir::ModuleOp &module, int64_t version) {
   module.getOperation()->setAttr(
       mlir::StringAttr::get(module.getContext(), llvm::Twine{"omp.version"}),

diff --git a/flang/lib/Frontend/FrontendActions.cpp b/flang/lib/Frontend/FrontendActions.cpp
@@ -301,9 +301,6 @@ bool CodeGenAction::beginSourceFileAction() {
           Fortran::common::LanguageFeature::OpenMP)) {
     setOffloadModuleInterfaceAttributes(*mlirModule,
                                         ci.getInvocation().getLangOpts());
-    setOffloadModuleInterfaceTargetAttribute(
-        *mlirModule, targetMachine.getTargetCPU(),
-        targetMachine.getTargetFeatureString());
     setOpenMPVersionAttribute(*mlirModule,
                               ci.getInvocation().getLangOpts().OpenMPVersion);
   }

diff --git a/flang/test/Lower/OpenMP/FIR/target_cpu_features.f90 b/flang/test/Lower/OpenMP/FIR/target_cpu_features.f90
@@ -1,22 +1,19 @@
 !REQUIRES: amdgpu-registered-target, nvptx-registered-target
-!RUN: %flang_fc1 -emit-fir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck %s
-!RUN: %flang_fc1 -emit-hlfir -triple nvptx64-nvidia-cuda -target-cpu sm_80 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck --check-prefix=NVPTX %s
-
+!RUN: %flang_fc1 -emit-fir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck --check-prefix=AMDGCN %s
+!RUN: %flang_fc1 -emit-fir -triple nvptx64-nvidia-cuda -target-cpu sm_80 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck --check-prefix=NVPTX %s
 
 !===============================================================================
 ! Target_Enter Simple
 !===============================================================================
 
-!CHECK: omp.target = #omp.target<target_cpu = "gfx908",
-!CHECK-SAME: target_features = "+16-bit-insts,+ci-insts,+dl-insts,+dot1-insts,+dot10-insts,
-!CHECK-SAME: +dot2-insts,+dot3-insts,+dot4-insts,+dot5-insts,+dot6-insts,+dot7-insts,+dpp,
-!CHECK-SAME: +gfx8-insts,+gfx9-insts,+gws,+image-insts,+mai-insts,+s-memrealtime,+s-memtime-inst,
-!CHECK-SAME: +wavefrontsize64">
-!NVPTX: omp.target = #omp.target<target_cpu = "sm_80", target_features = "+ptx61,+sm_80">
-!CHECK-LABEL: func.func @_QPomp_target_simple()
-subroutine omp_target_simple
-  ! Directive needed to prevent subroutine from being filtered out when
-  ! compiling for the device.
-  !$omp declare target
-end subroutine omp_target_simple
+!AMDGCN: module attributes {
+!AMDGCN-SAME: fir.target_cpu = "gfx908"
+!AMDGCN-SAME: fir.target_features = #llvm.target_features<["+16-bit-insts", "+ci-insts",
+!AMDGCN-SAME: "+dl-insts", "+dot1-insts", "+dot10-insts", "+dot2-insts", "+dot3-insts",
+!AMDGCN-SAME: "+dot4-insts", "+dot5-insts", "+dot6-insts", "+dot7-insts", "+dpp",
+!AMDGCN-SAME: "+gfx8-insts", "+gfx9-insts", "+gws", "+image-insts", "+mai-insts",
+!AMDGCN-SAME: "+s-memrealtime", "+s-memtime-inst", "+wavefrontsize64"]>
 
+!NVPTX: module attributes {
+!NVPTX-SAME: fir.target_cpu = "sm_80"
+!NVPTX-SAME: fir.target_features = #llvm.target_features<["+ptx61", "+sm_80"]>
diff --git a/flang/test/Lower/OpenMP/target_cpu_features.f90 b/flang/test/Lower/OpenMP/target_cpu_features.f90
@@ -1,21 +1,19 @@
 !REQUIRES: amdgpu-registered-target, nvptx-registered-target
-!RUN: %flang_fc1 -emit-hlfir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck %s
+!RUN: %flang_fc1 -emit-hlfir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck --check-prefix=AMDGCN %s
 !RUN: %flang_fc1 -emit-hlfir -triple nvptx64-nvidia-cuda -target-cpu sm_80 -fopenmp -fopenmp-is-target-device %s -o - | FileCheck --check-prefix=NVPTX %s
 
 !===============================================================================
 ! Target_Enter Simple
 !===============================================================================
 
-!CHECK: omp.target = #omp.target<target_cpu = "gfx908",
-!CHECK-SAME: target_features = "+16-bit-insts,+ci-insts,+dl-insts,+dot1-insts,+dot10-insts,
-!CHECK-SAME: +dot2-insts,+dot3-insts,+dot4-insts,+dot5-insts,+dot6-insts,+dot7-insts,+dpp,
-!CHECK-SAME: +gfx8-insts,+gfx9-insts,+gws,+image-insts,+mai-insts,+s-memrealtime,+s-memtime-inst,
-!CHECK-SAME: +wavefrontsize64">
-!NVPTX: omp.target = #omp.target<target_cpu = "sm_80", target_features = "+ptx61,+sm_80">
-!CHECK-LABEL: func.func @_QPomp_target_simple()
-subroutine omp_target_simple
-  ! Directive needed to prevent subroutine from being filtered out when
-  ! compiling for the device.
-  !$omp declare target
-end subroutine omp_target_simple
+!AMDGCN: module attributes {
+!AMDGCN-SAME: fir.target_cpu = "gfx908"
+!AMDGCN-SAME: fir.target_features = #llvm.target_features<["+16-bit-insts", "+ci-insts",
+!AMDGCN-SAME: "+dl-insts", "+dot1-insts", "+dot10-insts", "+dot2-insts", "+dot3-insts",
+!AMDGCN-SAME: "+dot4-insts", "+dot5-insts", "+dot6-insts", "+dot7-insts", "+dpp",
+!AMDGCN-SAME: "+gfx8-insts", "+gfx9-insts", "+gws", "+image-insts", "+mai-insts",
+!AMDGCN-SAME: "+s-memrealtime", "+s-memtime-inst", "+wavefrontsize64"]>
 
+!NVPTX: module attributes {
+!NVPTX-SAME: fir.target_cpu = "sm_80"
+!NVPTX-SAME: fir.target_features = #llvm.target_features<["+ptx61", "+sm_80"]>
diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
@@ -72,15 +72,6 @@ def FlagsAttr : OpenMP_Attr<"Flags", "flags"> {
   let assemblyFormat = "`<` struct(params) `>`";
 }
 
-def TargetAttr : OpenMP_Attr<"Target", "target"> {
-  let parameters = (ins
-    StringRefParameter<>:$target_cpu,
-    StringRefParameter<>:$target_features
-  );
-
-  let assemblyFormat = "`<` struct(params) `>`";
-}
-
 
 class OpenMP_Op<string mnemonic, list<Trait> traits = []> :
       Op<OpenMP_Dialect, mnemonic, traits>;

diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPOpsInterfaces.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPOpsInterfaces.td
@@ -205,34 +205,6 @@ def OffloadModuleInterface : OpInterface<"OffloadModuleInterface"> {
                       assumeTeamsOversubscription, assumeThreadsOversubscription,
                       assumeNoThreadState, assumeNoNestedParallelism, noGPULib, openmpDeviceVersion));
       }]>,
-    InterfaceMethod<
-      /*description=*/[{
-        Get the Target attribute on the current module if it exists
-        and return the attribute, if it doesn't exist it returns a nullptr.
-      }],
-      /*retTy=*/"mlir::omp::TargetAttr",
-      /*methodName=*/"getTarget",
-      (ins), [{}], [{
-        if (Attribute flags = $_op->getAttr("omp.target"))
-          return ::llvm::dyn_cast_or_null<mlir::omp::TargetAttr>(flags);
-        return nullptr;
-      }]>,
-    InterfaceMethod<
-      /*description=*/[{
-        Set the attribute target on the current module with the
-        specified string arguments - name of cpu and corresponding features.
-      }],
-      /*retTy=*/"void",
-      /*methodName=*/"setTarget",
-      (ins "llvm::StringRef":$targetCPU,
-           "llvm::StringRef":$targetFeatures), [{}], [{
-        if (targetCPU.empty())
-          return;
-        $_op->setAttr(("omp." + mlir::omp::TargetAttr::getMnemonic()).str(),
-                  mlir::omp::TargetAttr::get($_op->getContext(),
-                                             targetCPU.str(),
-                                             targetFeatures.str()));
-      }]>,
     InterfaceMethod<
       /*description=*/[{
         Set a StringAttr on the current module containing the host IR file path. This

diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2336,6 +2336,7 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
   if (!targetOpSupported(opInst))
     return failure();
 
+  auto parentFn = opInst.getParentOfType<LLVM::LLVMFuncOp>();
   auto targetOp = cast<omp::TargetOp>(opInst);
   auto &targetRegion = targetOp.getRegion();
   DataLayout dl = DataLayout(opInst.getParentOfType<ModuleOp>());
@@ -2345,6 +2346,22 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
   using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
   auto bodyCB = [&](InsertPointTy allocaIP,
                     InsertPointTy codeGenIP) -> InsertPointTy {
+    // Forward target-cpu and target-features function attributes from the
+    // original function to the new outlined function.
+    llvm::Function *llvmParentFn =
+        moduleTranslation.lookupFunction(parentFn.getName());
+    llvm::Function *llvmOutlinedFn = codeGenIP.getBlock()->getParent();
+    assert(llvmParentFn && llvmOutlinedFn &&
+           "Both parent and outlined functions must exist at this point");
+
+    if (auto attr = llvmParentFn->getFnAttribute("target-cpu");
+        attr.isStringAttribute())
+      llvmOutlinedFn->addFnAttr(attr);
+
+    if (auto attr = llvmParentFn->getFnAttribute("target-features");
+        attr.isStringAttribute())
+      llvmOutlinedFn->addFnAttr(attr);
+
     builder.restoreIP(codeGenIP);
     unsigned argIndex = 0;
     for (auto &mapOp : mapOperands) {
@@ -2363,7 +2380,7 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
   };
 
   llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
-  StringRef parentName = opInst.getParentOfType<LLVM::LLVMFuncOp>().getName();
+  StringRef parentName = parentFn.getName();
 
   llvm::TargetRegionEntryInfo entryInfo;
 

diff --git a/mlir/test/Target/LLVMIR/omptarget-parallel-llvm.mlir b/mlir/test/Target/LLVMIR/omptarget-parallel-llvm.mlir
@@ -3,7 +3,7 @@
 // The aim of the test is to check the LLVM IR codegen for the device
 // for omp target parallel construct
 
-module attributes {dlti.dl_spec = #dlti.dl_spec<#dlti.dl_entry<"dlti.alloca_memory_space", 5 : ui32>>, 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-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.is_gpu = true, omp.is_target_device = true, omp.target = #omp.target<target_cpu = "gfx90a", target_features = "">} {
+module attributes {dlti.dl_spec = #dlti.dl_spec<#dlti.dl_entry<"dlti.alloca_memory_space", 5 : ui32>>, 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-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.is_gpu = true, omp.is_target_device = true} {
   llvm.func @_QQmain_omp_outline_1(%arg0: !llvm.ptr) attributes {omp.declare_target = #omp.declaretarget<device_type = (host), capture_clause = (to)>} {
     %0 = omp.map_info var_ptr(%arg0 : !llvm.ptr, i32) map_clauses(from) capture(ByRef) -> !llvm.ptr {name = "d"}
     omp.target map_entries(%0 -> %arg2 : !llvm.ptr) {

diff --git a/mlir/test/Target/LLVMIR/omptarget-target-cpu-features.mlir b/mlir/test/Target/LLVMIR/omptarget-target-cpu-features.mlir
@@ -0,0 +1,23 @@
+// Test that the target_features and target_cpu llvm.func attributes are
+// forwarded to outlined target region functions.
+
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+module attributes {omp.is_target_device = false} {
+  llvm.func @omp_target_region() attributes {
+    target_cpu = "x86-64",
+    target_features = #llvm.target_features<["+mmx", "+sse"]>
+  } {
+    omp.target {
+      omp.terminator
+    }
+    llvm.return
+  }
+}
+
+// CHECK: define void @omp_target_region() #[[ATTRS:.*]] {
+// CHECK: define internal void @__omp_offloading_{{.*}}_omp_target_region_{{.*}}() #[[ATTRS]] {
+
+// CHECK: attributes #[[ATTRS]] = {
+// CHECK-SAME: "target-cpu"="x86-64"
+// CHECK-SAME: "target-features"="+mmx,+sse"