[Flang][OpenMP][MLIR] Add support for -nogpulib option (#71045)

If -nogpulib option is passed by the user, then the OpenMP device
runtime is not used and we should not emit globals to configure
debugging at compile-time for the device runtime.

Link to -nogpulib flag implementation for Clang:
https://reviews.llvm.org/D125314

Author: DominikAdamski

clang/include/clang/Driver/Options.td

@@ -5197,7 +5197,7 @@ def nohipwrapperinc : Flag<["-"], "nohipwrapperinc">, Group<IncludePath_Group>,
   HelpText<"Do not include the default HIP wrapper headers and include paths">;
 def : Flag<["-"], "nocudainc">, Alias<nogpuinc>;
 def nogpulib : Flag<["-"], "nogpulib">, MarshallingInfoFlag<LangOpts<"NoGPULib">>,
-  Visibility<[ClangOption, CC1Option]>,
+  Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>,
   HelpText<"Do not link device library for CUDA/HIP device compilation">;
 def : Flag<["-"], "nocudalib">, Alias<nogpulib>;
 def gpulibc : Flag<["-"], "gpulibc">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>,

clang/lib/Driver/ToolChains/Flang.cpp

@@ -428,6 +428,8 @@ void Flang::addOffloadOptions(Compilation &C, const InputInfoList &Inputs,
       CmdArgs.push_back("-fopenmp-assume-no-thread-state");
     if (Args.hasArg(options::OPT_fopenmp_assume_no_nested_parallelism))
       CmdArgs.push_back("-fopenmp-assume-no-nested-parallelism");
+    if (Args.hasArg(options::OPT_nogpulib))
+      CmdArgs.push_back("-nogpulib");
   }
 }
 

flang/include/flang/Frontend/LangOptions.def

@@ -21,6 +21,8 @@ LANGOPT(Name, Bits, Default)
 
 ENUM_LANGOPT(FPContractMode, FPModeKind, 2, FPM_Fast) ///< FP Contract Mode (off/fast)
 
+/// Indicate a build without the standard GPU libraries.
+LANGOPT(NoGPULib  , 1, false)
 /// Permit floating point optimization without regard to infinities
 LANGOPT(NoHonorInfs, 1, false)
 /// Permit floating point optimization without regard to NaN

flang/include/flang/Tools/CrossToolHelpers.h

@@ -56,14 +56,16 @@ struct OffloadModuleOpts {
   OffloadModuleOpts(uint32_t OpenMPTargetDebug, bool OpenMPTeamSubscription,
       bool OpenMPThreadSubscription, bool OpenMPNoThreadState,
       bool OpenMPNoNestedParallelism, bool OpenMPIsTargetDevice,
-      bool OpenMPIsGPU, uint32_t OpenMPVersion, std::string OMPHostIRFile = {})
+      bool OpenMPIsGPU, uint32_t OpenMPVersion, std::string OMPHostIRFile = {},
+      bool NoGPULib = false)
       : OpenMPTargetDebug(OpenMPTargetDebug),
         OpenMPTeamSubscription(OpenMPTeamSubscription),
         OpenMPThreadSubscription(OpenMPThreadSubscription),
         OpenMPNoThreadState(OpenMPNoThreadState),
         OpenMPNoNestedParallelism(OpenMPNoNestedParallelism),
         OpenMPIsTargetDevice(OpenMPIsTargetDevice), OpenMPIsGPU(OpenMPIsGPU),
-        OpenMPVersion(OpenMPVersion), OMPHostIRFile(OMPHostIRFile) {}
+        OpenMPVersion(OpenMPVersion), OMPHostIRFile(OMPHostIRFile),
+        NoGPULib(NoGPULib) {}
 
   OffloadModuleOpts(Fortran::frontend::LangOptions &Opts)
       : OpenMPTargetDebug(Opts.OpenMPTargetDebug),
@@ -73,7 +75,7 @@ struct OffloadModuleOpts {
         OpenMPNoNestedParallelism(Opts.OpenMPNoNestedParallelism),
         OpenMPIsTargetDevice(Opts.OpenMPIsTargetDevice),
         OpenMPIsGPU(Opts.OpenMPIsGPU), OpenMPVersion(Opts.OpenMPVersion),
-        OMPHostIRFile(Opts.OMPHostIRFile) {}
+        OMPHostIRFile(Opts.OMPHostIRFile), NoGPULib(Opts.NoGPULib) {}
 
   uint32_t OpenMPTargetDebug = 0;
   bool OpenMPTeamSubscription = false;
@@ -84,6 +86,7 @@ struct OffloadModuleOpts {
   bool OpenMPIsGPU = false;
   uint32_t OpenMPVersion = 11;
   std::string OMPHostIRFile = {};
+  bool NoGPULib = false;
 };
 
 //  Shares assinging of the OpenMP OffloadModuleInterface and its assorted
@@ -98,7 +101,7 @@ void setOffloadModuleInterfaceAttributes(
     if (Opts.OpenMPIsTargetDevice) {
       offloadMod.setFlags(Opts.OpenMPTargetDebug, Opts.OpenMPTeamSubscription,
           Opts.OpenMPThreadSubscription, Opts.OpenMPNoThreadState,
-          Opts.OpenMPNoNestedParallelism, Opts.OpenMPVersion);
+          Opts.OpenMPNoNestedParallelism, Opts.OpenMPVersion, Opts.NoGPULib);
 
       if (!Opts.OMPHostIRFile.empty())
         offloadMod.setHostIRFilePath(Opts.OMPHostIRFile);

flang/lib/Frontend/CompilerInvocation.cpp

@@ -935,6 +935,8 @@ static bool parseDialectArgs(CompilerInvocation &res, llvm::opt::ArgList &args,
             args.hasArg(clang::driver::options::OPT_fopenmp_target_debug))
           res.getLangOpts().OpenMPTargetDebug = 1;
       }
+      if (args.hasArg(clang::driver::options::OPT_nogpulib))
+        res.getLangOpts().NoGPULib = 1;
     }
 
     switch (llvm::Triple(res.getTargetOpts().triple).getArch()) {

flang/test/Driver/driver-help-hidden.f90

@@ -127,6 +127,7 @@
 ! CHECK-NEXT: --no-offload-arch=<value>
 ! CHECK-NEXT:                         Remove CUDA/HIP offloading device architecture (e.g. sm_35, gfx906) from the list of devices to compile for. 'all' resets the list to its default value.
 ! CHECK-NEXT: -nocpp                  Disable predefined and command line preprocessor macros
+! CHECK-NEXT: -nogpulib               Do not link device library for CUDA/HIP device compilation
 ! CHECK-NEXT: --offload-arch=<value>  Specify an offloading device architecture for CUDA, HIP, or OpenMP. (e.g. sm_35). If 'native' is used the compiler will detect locally installed architectures. For HIP offloading, the device architecture can be followed by target ID features delimited by a colon (e.g. gfx908:xnack+:sramecc-). May be specified more than once.
 ! CHECK-NEXT: --offload-device-only   Only compile for the offloading device.
 ! CHECK-NEXT: --offload-host-device   Compile for both the offloading host and device (default).

flang/test/Driver/driver-help.f90

@@ -113,6 +113,7 @@
 ! HELP-NEXT: --no-offload-arch=<value>
 ! HELP-NEXT:                         Remove CUDA/HIP offloading device architecture (e.g. sm_35, gfx906) from the list of devices to compile for. 'all' resets the list to its default value.
 ! HELP-NEXT: -nocpp                  Disable predefined and command line preprocessor macros
+! HELP-NEXT: -nogpulib               Do not link device library for CUDA/HIP device compilation
 ! HELP-NEXT: --offload-arch=<value>  Specify an offloading device architecture for CUDA, HIP, or OpenMP. (e.g. sm_35). If 'native' is used the compiler will detect locally installed architectures. For HIP offloading, the device architecture can be followed by target ID features delimited by a colon (e.g. gfx908:xnack+:sramecc-). May be specified more than once.
 ! HELP-NEXT: --offload-device-only   Only compile for the offloading device.
 ! HELP-NEXT: --offload-host-device   Compile for both the offloading host and device (default).
@@ -249,6 +250,7 @@
 ! HELP-FC1-NEXT: -mvscale-max=<value>    Specify the vscale maximum. Defaults to the vector length agnostic value of "0". (AArch64/RISC-V only)
 ! HELP-FC1-NEXT: -mvscale-min=<value>    Specify the vscale minimum. Defaults to "1". (AArch64/RISC-V only)
 ! HELP-FC1-NEXT: -nocpp                  Disable predefined and command line preprocessor macros
+! HELP-FC1-NEXT: -nogpulib               Do not link device library for CUDA/HIP device compilation
 ! HELP-FC1-NEXT: -opt-record-file <value>
 ! HELP-FC1-NEXT:                         File name to use for YAML optimization record output
 ! HELP-FC1-NEXT: -opt-record-format <value>

flang/test/Lower/OpenMP/nogpulib.f90

@@ -0,0 +1,12 @@
+!REQUIRES: amdgpu-registered-target
+
+!RUN: %flang_fc1 -triple amdgcn-amd-amdhsa -emit-hlfir -fopenmp -fopenmp-is-target-device %s -o - | FileCheck %s
+!RUN: bbc -fopenmp -fopenmp-is-target-device -fopenmp-is-gpu -emit-hlfir -o - %s | FileCheck %s
+!RUN: %flang_fc1 -triple amdgcn-amd-amdhsa -emit-hlfir -fopenmp -fopenmp-is-target-device -nogpulib %s -o - | FileCheck %s -check-prefix=FLAG_SET
+!RUN: bbc -fopenmp -fopenmp-is-target-device -fopenmp-is-gpu -emit-hlfir -nogpulib -o - %s | FileCheck %s -check-prefix=FLAG_SET
+
+!CHECK-NOT: module attributes {{{.*}}no_gpu_lib
+!FLAG_SET: module attributes {{{.*}}no_gpu_lib = true
+subroutine omp_subroutine()
+end subroutine omp_subroutine
+

flang/tools/bbc/bbc.cpp

@@ -181,6 +181,12 @@ static llvm::cl::opt<bool> setOpenMPNoNestedParallelism(
                    "a parallel region."),
     llvm::cl::init(false));
 
+static llvm::cl::opt<bool>
+    setNoGPULib("nogpulib",
+                llvm::cl::desc("Do not link device library for CUDA/HIP device "
+                               "compilation"),
+                llvm::cl::init(false));
+
 static llvm::cl::opt<bool> enableOpenACC("fopenacc",
                                          llvm::cl::desc("enable openacc"),
                                          llvm::cl::init(false));
@@ -349,7 +355,7 @@ static mlir::LogicalResult convertFortranSourceToMLIR(
         OffloadModuleOpts(setOpenMPTargetDebug, setOpenMPTeamSubscription,
                           setOpenMPThreadSubscription, setOpenMPNoThreadState,
                           setOpenMPNoNestedParallelism, enableOpenMPDevice,
-                          enableOpenMPGPU, setOpenMPVersion);
+                          enableOpenMPGPU, setOpenMPVersion, "", setNoGPULib);
     setOffloadModuleInterfaceAttributes(mlirModule, offloadModuleOpts);
     setOpenMPVersionAttribute(mlirModule, setOpenMPVersion);
   }

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

@@ -65,6 +65,7 @@ def FlagsAttr : OpenMP_Attr<"Flags", "flags"> {
     DefaultValuedParameter<"bool", "false">:$assume_threads_oversubscription,
     DefaultValuedParameter<"bool", "false">:$assume_no_thread_state,
     DefaultValuedParameter<"bool", "false">:$assume_no_nested_parallelism,
+    DefaultValuedParameter<"bool", "false">:$no_gpu_lib,
     DefaultValuedParameter<"uint32_t", "50">:$openmp_device_version
   );
 

mlir/include/mlir/Dialect/OpenMP/OpenMPOpsInterfaces.td

@@ -198,11 +198,12 @@ def OffloadModuleInterface : OpInterface<"OffloadModuleInterface"> {
             "bool":$assumeThreadsOversubscription,
             "bool":$assumeNoThreadState,
             "bool":$assumeNoNestedParallelism,
-            "uint32_t":$openmpDeviceVersion), [{}], [{
+            "uint32_t":$openmpDeviceVersion,
+            "bool":$noGPULib), [{}], [{
         $_op->setAttr(("omp." + mlir::omp::FlagsAttr::getMnemonic()).str(),
                   mlir::omp::FlagsAttr::get($_op->getContext(), debugKind,
                       assumeTeamsOversubscription, assumeThreadsOversubscription,
-                      assumeNoThreadState, assumeNoNestedParallelism, openmpDeviceVersion));
+                      assumeNoThreadState, assumeNoNestedParallelism, noGPULib, openmpDeviceVersion));
       }]>,
     InterfaceMethod<
       /*description=*/[{

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

@@ -2035,6 +2035,12 @@ LogicalResult convertFlagsAttr(Operation *op, mlir::omp::FlagsAttr attribute,
 
   llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
 
+  ompBuilder->M.addModuleFlag(llvm::Module::Max, "openmp-device",
+                              attribute.getOpenmpDeviceVersion());
+
+  if (attribute.getNoGpuLib())
+    return success();
+
   ompBuilder->createGlobalFlag(
       attribute.getDebugKind() /*LangOpts().OpenMPTargetDebug*/,
       "__omp_rtl_debug_kind");
@@ -2056,8 +2062,6 @@ LogicalResult convertFlagsAttr(Operation *op, mlir::omp::FlagsAttr attribute,
           .getAssumeNoNestedParallelism() /*LangOpts().OpenMPNoNestedParallelism*/
       ,
       "__omp_rtl_assume_no_nested_parallelism");
-  ompBuilder->M.addModuleFlag(llvm::Module::Max, "openmp-device",
-                              attribute.getOpenmpDeviceVersion());
   return success();
 }
 

mlir/test/Dialect/OpenMP/attr.mlir

@@ -54,6 +54,12 @@ module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true,
 // CHECK: module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true, assume_no_thread_state = true, openmp_device_version = 51>} {
 module attributes {omp.flags = #omp.flags<assume_no_thread_state = true, assume_teams_oversubscription = true, openmp_device_version = 51>} {}
 
+// CHECK: module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true, assume_no_thread_state = true, no_gpu_lib = true, openmp_device_version = 51>} {
+module attributes {omp.flags = #omp.flags<assume_no_thread_state = true, assume_teams_oversubscription = true, no_gpu_lib = true, openmp_device_version = 51>} {}
+
+// CHECK: module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true, openmp_device_version = 51>} {
+module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true, no_gpu_lib = false, openmp_device_version = 51>} {}
+
 // CHECK: module attributes {omp.version = #omp.version<version = 51>} {
 module attributes {omp.version = #omp.version<version = 51>} {}
 

mlir/test/Target/LLVMIR/openmp-llvm.mlir

@@ -2530,6 +2530,16 @@ module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true,
 
 // -----
 
+// CHECK-NOT: @__omp_rtl_debug_kind = weak_odr hidden constant i32 0
+// CHECK-NOT: @__omp_rtl_assume_teams_oversubscription = weak_odr hidden constant i32 1
+// CHECK-NOT: @__omp_rtl_assume_threads_oversubscription = weak_odr hidden constant i32 0
+// CHECK-NOT: @__omp_rtl_assume_no_thread_state = weak_odr hidden constant i32 1
+// CHECK-NOT: @__omp_rtl_assume_no_nested_parallelism = weak_odr hidden constant i32 0
+module attributes {omp.flags = #omp.flags<assume_teams_oversubscription = true, assume_no_thread_state = true,
+                                          no_gpu_lib=true>} {}
+
+// -----
+
 module attributes {omp.is_target_device = false} {
   // CHECK: define void @filter_nohost
   llvm.func @filter_nohost() -> ()