[OpenMP] Emit calls to int64_t functions for amdgcn

[JonChesterfield]

Summary

[OpenMP] Emit calls to int64_t functions for amdgcn

Two functions, syncwarp and active_thread_mask, return lanemask_t. Currently
this is assumed to be int32, which is true for nvptx. Patch makes the type
target architecture dependent.

[JonChesterfield]

llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp

221

This seems ugly - is there a better path from module to triple?

[jdoerfert]

Generally yes. Some comments. Can you add the functions to the test that adds attributes. It should "pick up" the signature and add attributes only for the right one. We might need to add the functions to some attribute group in the OMPKinds.def but that would be good anyway, even if we just do nounwind.

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

433

I would suggest to put it in OMPKinds.def similar to size_t:

`OMP_TYPE(SizeTy, M.getDataLayout().getIntPtrType(Ctx))`

I guess you can keep getLanemaskType and use it though.

llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp

66

This shadows the other declaration, doesn't it? See also other comments.

1189

Not needed once in OMPKinds.def, probably the shadowing stuff neither.

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

334 ↗

(On Diff #299192)

Looks again like shadowing, necessary?

[JonChesterfield]

• address some review comments

[JonChesterfield]

Good call on shadowing, those are all gone. Attributes are more difficult as the test is IR, so can't declare the function twice. Haven't managed to tag nounwind onto either yet but will try a few more things.

[JonChesterfield]

• tests for i32 version of active_thread_mask, syncwarp

[JonChesterfield]

• test attributes on syncwarp, thread_mask

[jdoerfert]

LGTM

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

public:
   Constant *getOrCreateSrcLocStr(const LocationDescription &Loc);
 
   /// Return an ident_t* encoding the source location \p SrcLocStr and \p Flags.
   /// TODO: Create a enum class for the Reserve2Flags
   Value *getOrCreateIdent(Constant *SrcLocStr,
                           omp::IdentFlag Flags = omp::IdentFlag(0),
                           unsigned Reserve2Flags = 0);
 
+  // Get the type corresponding to __kmpc_impl_lanemask_t from the deviceRTL
+  Type *getLanemaskType();
+
   /// Generate control flow and cleanup for cancellation.
   ///
   /// \param CancelFlag Flag indicating if the cancellation is performed.
   /// \param CanceledDirective The kind of directive that is cancled.
   void emitCancelationCheckImpl(Value *CancelFlag,
                                 omp::Directive CanceledDirective);
 
   /// Generate a barrier runtime call.
CallInst *CreateCachedThreadPrivate(const LocationDescription &Loc,
                                       const llvm::Twine &Name = Twine(""));
 
   /// Declarations for LLVM-IR types (simple, array, function and structure) are
   /// generated below. Their names are defined and used in OpenMPKinds.def. Here
   /// we provide the declarations, the initializeTypes function will provide the
   /// values.
   ///
   ///{
 #define OMP_TYPE(VarName, InitValue) Type *VarName = nullptr;
 #define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize)                             \
   ArrayType *VarName##Ty = nullptr;                                            \
   PointerType *VarName##PtrTy = nullptr;
 #define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...)                  \
   FunctionType *VarName = nullptr;                                             \
   PointerType *VarName##Ptr = nullptr;
 #define OMP_STRUCT_TYPE(VarName, StrName, ...)                                 \
   StructType *VarName = nullptr;                                               \

llvm/include/llvm/Frontend/OpenMP/OMPKinds.def

 __OMP_TYPE(Int32)
 __OMP_TYPE(Int64)
 __OMP_TYPE(Int8Ptr)
 __OMP_TYPE(Int16Ptr)
 __OMP_TYPE(Int32Ptr)
 __OMP_TYPE(Int64Ptr)
 
 OMP_TYPE(SizeTy, M.getDataLayout().getIntPtrType(Ctx))
+OMP_TYPE(LanemaskTy, getLanemaskType())
 
 #define __OMP_PTR_TYPE(NAME, BASE) OMP_TYPE(NAME, BASE->getPointerTo())
 
 __OMP_PTR_TYPE(VoidPtr, Int8)
 __OMP_PTR_TYPE(VoidPtrPtr, VoidPtr)
 __OMP_PTR_TYPE(VoidPtrPtrPtr, VoidPtrPtr)
 
 __OMP_PTR_TYPE(Int8PtrPtr, Int8Ptr)

 __OMP_RTL(__kmpc_end_sharing_variables, false, Void, )
 __OMP_RTL(__kmpc_get_shared_variables, false, Void, VoidPtrPtrPtr)
 __OMP_RTL(__kmpc_parallel_level, false, Int16, IdentPtr, Int32)
 __OMP_RTL(__kmpc_is_spmd_exec_mode, false, Int8, )
 __OMP_RTL(__kmpc_get_team_static_memory, false, Void, Int16, VoidPtr, SizeTy,
           Int16, VoidPtrPtr)
 __OMP_RTL(__kmpc_restore_team_static_memory, false, Void, Int16, Int16)
 __OMP_RTL(__kmpc_barrier_simple_spmd, false, Void, IdentPtr, Int32)
-__OMP_RTL(__kmpc_warp_active_thread_mask, false, Int32, )
-__OMP_RTL(__kmpc_syncwarp, false, Void, Int32)
+
+__OMP_RTL(__kmpc_warp_active_thread_mask, false, LanemaskTy,)
+__OMP_RTL(__kmpc_syncwarp, false, Void, LanemaskTy)
 
 __OMP_RTL(__last, false, Void, )
 
 #undef __OMP_RTL
 #undef OMP_RTL
 
 #define ParamAttrs(...) ArrayRef<AttributeSet>({__VA_ARGS__})
 #define EnumAttr(Kind) Attribute::get(Ctx, Attribute::AttrKind::Kind)

llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp

 /// convenient way to create LLVM instructions for OpenMP directives.
 ///
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
 
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/StringSwitch.h"
+#include "llvm/ADT/Triple.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/DebugInfo.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/MDBuilder.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Error.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/CodeExtractor.h"
default:
     break;
   }
 }
 
 FunctionCallee
 OpenMPIRBuilder::getOrCreateRuntimeFunction(Module &M, RuntimeFunction FnID) {
   FunctionType *FnTy = nullptr;
   Function *Fn = nullptr;
 
   // Try to find the declation in the module first.
   switch (FnID) {
 #define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...)                          \
   case Enum:                                                                   \
     FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__},        \
                              IsVarArg);                                        \
     Fn = M.getFunction(Str);                                                   \
     break;
auto *GV = new GlobalVariable(M, IdentPtr->getPointerElementType(),
                                   GlobalValue::PrivateLinkage, Initializer);
     GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
     GV->setAlignment(Align(8));
     Ident = GV;
   }
   return Ident;
 }
 
+Type *OpenMPIRBuilder::getLanemaskType() {
+  LLVMContext &Ctx = M.getContext();
+  Triple triple(M.getTargetTriple());
+
+  // This test is adequate until deviceRTL has finer grained lane widths
+  return triple.isAMDGCN() ? Type::getInt64Ty(Ctx) : Type::getInt32Ty(Ctx);
+}
+
 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef LocStr) {
   Constant *&SrcLocStr = SrcLocStrMap[LocStr];
   if (!SrcLocStr) {
     Constant *Initializer =
         ConstantDataArray::getString(M.getContext(), LocStr);
 
     // Look for existing encoding of the location, not needed but minimizes the
     // difference to the existing solution while we transition.
Value *OpenMPIRBuilder::getOMPCriticalRegionLock(StringRef CriticalName) {
   return getOrCreateOMPInternalVariable(KmpCriticalNameTy, Name);
 }
 
 // Create all simple and struct types exposed by the runtime and remember
 // the llvm::PointerTypes of them for easy access later.
 void OpenMPIRBuilder::initializeTypes(Module &M) {
   LLVMContext &Ctx = M.getContext();
   StructType *T;
 #define OMP_TYPE(VarName, InitValue) VarName = InitValue;
 #define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize)                             \
   VarName##Ty = ArrayType::get(ElemTy, ArraySize);                             \
   VarName##PtrTy = PointerType::getUnqual(VarName##Ty);
 #define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...)                  \
   VarName = FunctionType::get(ReturnType, {__VA_ARGS__}, IsVarArg);            \
   VarName##Ptr = PointerType::getUnqual(VarName);
 #define OMP_STRUCT_TYPE(VarName, StructName, ...)                              \
   T = M.getTypeByName(StructName);                                             \

llvm/test/Transforms/OpenMP/add_attributes.ll

 declare void @__kmpc_free(i32, i8*, i8*)
 
 declare i8* @__kmpc_init_allocator(i32, i8*, i32, i8*)
 
 declare void @__kmpc_destroy_allocator(i32, i8*)
 
 declare void @__kmpc_push_target_tripcount(i64, i64)
 
+declare i32 @__kmpc_warp_active_thread_mask()
+
+declare void @__kmpc_syncwarp(i32)
+
 declare i32 @__tgt_target_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**)
 
 declare i32 @__tgt_target_nowait_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**)
 
 declare i32 @__tgt_target_teams_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**, i32, i32)
 
 declare i32 @__tgt_target_teams_nowait_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**, i32, i32)
 

 ; CHECK-NEXT: declare i8* @__kmpc_init_allocator(i32, i8*, i32, i8*)
 
 ; CHECK: ; Function Attrs: nounwind
 ; CHECK-NEXT: declare void @__kmpc_destroy_allocator(i32, i8*)
 
 ; CHECK: ; Function Attrs: nounwind
 ; CHECK-NEXT: declare void @__kmpc_push_target_tripcount(i64, i64)
 
+; CHECK: ; Function Attrs: convergent nounwind
+; CHECK-NEXT: declare i32 @__kmpc_warp_active_thread_mask()
+
+; CHECK: ; Function Attrs: convergent nounwind
+; CHECK-NEXT: declare void @__kmpc_syncwarp(i32)
+
 ; CHECK: ; Function Attrs: nounwind
 ; CHECK-NEXT: declare i32 @__tgt_target_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**)
 
 ; CHECK: ; Function Attrs: nounwind
 ; CHECK-NEXT: declare i32 @__tgt_target_nowait_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**)
 
 ; CHECK: ; Function Attrs: nounwind
 ; CHECK-NEXT: declare i32 @__tgt_target_teams_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**, i32, i32)

 ; OPTIMISTIC-NEXT: declare noalias i8* @__kmpc_init_allocator(i32, i8*, i32, i8*)
 
 ; OPTIMISTIC: ; Function Attrs: nosync nounwind willreturn
 ; OPTIMISTIC-NEXT: declare void @__kmpc_destroy_allocator(i32, i8*)
 
 ; OPTIMISTIC: ; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn writeonly
 ; OPTIMISTIC-NEXT: declare void @__kmpc_push_target_tripcount(i64, i64)
 
+; OPTIMISTIC: ; Function Attrs: convergent nounwind
+; OPTIMISTIC-NEXT: declare i32 @__kmpc_warp_active_thread_mask()
+
+; OPTIMISTIC: ; Function Attrs: convergent nounwind
+; OPTIMISTIC-NEXT: declare void @__kmpc_syncwarp(i32)
+
 ; OPTIMISTIC: ; Function Attrs: nounwind
 ; OPTIMISTIC-NEXT: declare i32 @__tgt_target_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**)
 
 ; OPTIMISTIC: ; Function Attrs: nounwind
 ; OPTIMISTIC-NEXT: declare i32 @__tgt_target_nowait_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**)
 
 ; OPTIMISTIC: ; Function Attrs: nounwind
 ; OPTIMISTIC-NEXT: declare i32 @__tgt_target_teams_mapper(i64, i8*, i32, i8**, i8**, i64*, i64*, i8**, i32, i32)

llvm/test/Transforms/OpenMP/add_attributes_amdgcn.ll

+; RUN: opt < %s -S -openmpopt        | FileCheck %s
+; RUN: opt < %s -S -passes=openmpopt | FileCheck %s
+; RUN: opt < %s -S -openmpopt        -openmp-ir-builder-optimistic-attributes | FileCheck %s --check-prefix=OPTIMISTIC
+; RUN: opt < %s -S -passes=openmpopt -openmp-ir-builder-optimistic-attributes | FileCheck %s --check-prefix=OPTIMISTIC
+
+target triple = "amdgcn-amd-amdhsa"
+
+define void @call_all(i64 %arg) {
+  call void @__kmpc_syncwarp(i64 %arg)
+  call i64 @__kmpc_warp_active_thread_mask()
+  ret void
+}
+
+declare i64 @__kmpc_warp_active_thread_mask()
+
+declare void @__kmpc_syncwarp(i64)
+
+; CHECK: ; Function Attrs: convergent nounwind
+; CHECK-NEXT: declare i64 @__kmpc_warp_active_thread_mask()
+
+; CHECK: ; Function Attrs: convergent nounwind
+; CHECK-NEXT: declare void @__kmpc_syncwarp(i64)
+
+; OPTIMISTIC: ; Function Attrs: convergent nounwind
+; OPTIMISTIC-NEXT: declare i64 @__kmpc_warp_active_thread_mask()
+
+; OPTIMISTIC: ; Function Attrs: convergent nounwind
+; OPTIMISTIC-NEXT: declare void @__kmpc_syncwarp(i64)