[OpenMP] Emit calls to int64_t functions for amdgcn

[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.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D89746

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

@@ -226,6 +226,9 @@ public:
                           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.

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

@@ -153,6 +153,7 @@ __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())
 
@@ -553,8 +554,9 @@ __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, )
 

llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp

@@ -16,6 +16,7 @@
 
 #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"
@@ -217,6 +218,14 @@ Value *OpenMPIRBuilder::getOrCreateIdent(Constant *SrcLocStr,
   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) {

llvm/test/Transforms/OpenMP/add_attributes.ll

@@ -629,6 +629,10 @@ 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**)
@@ -1142,6 +1146,12 @@ declare void @__kmpc_proxy_task_completed_ooo(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**)
 
@@ -1661,6 +1671,12 @@ declare void @__kmpc_proxy_task_completed_ooo(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**)
 

llvm/test/Transforms/OpenMP/add_attributes_amdgcn.ll

@@ -0,0 +1,28 @@
+; 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)