[PATCH] D88243: [OpenMP] OpenMPOpt Support for Globalization Remarks

[Joseph Huber via Phabricator via llvm-commits]

jhuber6 created this revision.
jhuber6 added a reviewer: jdoerfert.
jhuber6 added projects: OpenMP, LLVM.
Herald added subscribers: llvm-commits, guansong, hiraditya, yaxunl.
jhuber6 requested review of this revision.
Herald added a subscriber: sstefan1.

This patch add support for printing analysis messages relating to data globalization on the GPU. This occurs when data is shared between the threads in a GPU context and must be pushed to global or shared memory.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D88243

Files:
  llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
  llvm/lib/Transforms/IPO/OpenMPOpt.cpp


Index: llvm/lib/Transforms/IPO/OpenMPOpt.cpp
===================================================================
--- llvm/lib/Transforms/IPO/OpenMPOpt.cpp
+++ llvm/lib/Transforms/IPO/OpenMPOpt.cpp
@@ -476,6 +476,12 @@
       : M(*(*SCC.begin())->getParent()), SCC(SCC), CGUpdater(CGUpdater),
         OREGetter(OREGetter), OMPInfoCache(OMPInfoCache), A(A) {}
 
+  /// Check if any remarks are enabled for openmp-opt
+  bool remarksEnabled() {
+    auto &Ctx = M.getContext();
+    return Ctx.getDiagHandlerPtr()->isAnyRemarkEnabled(DEBUG_TYPE);
+  }
+
   /// Run all OpenMP optimizations on the underlying SCC/ModuleSlice.
   bool run() {
     if (SCC.empty())
@@ -503,6 +509,8 @@
     Changed |= deleteParallelRegions();
     if (HideMemoryTransferLatency)
       Changed |= hideMemTransfersLatency();
+    if (remarksEnabled())
+      analysisGlobalization();
 
     return Changed;
   }
@@ -695,6 +703,27 @@
     return Changed;
   }
 
+  void analysisGlobalization() {
+    auto &RFI =
+      OMPInfoCache.RFIs[OMPRTL___kmpc_data_sharing_coalesced_push_stack];
+    for (Function *F : SCC) {
+      auto *UV = RFI.getUseVector(*F);
+      if (!UV)
+        return;
+      for (Use *U : *UV) {
+        if (CallInst *CI = getCallIfRegularCall(*U, &RFI)) {
+          auto Remark = [&](OptimizationRemarkAnalysis ORA) {
+            return ORA << "Found thread data sharing on the GPU. Expect "
+              << "degraded performance due to data globalization.";
+          };
+          emitRemark<OptimizationRemarkAnalysis>(CI, "OpenMPGlobalization",
+              Remark);
+        }
+      }
+    }
+    return;
+  }
+
   /// Maps the values stored in the offload arrays passed as arguments to
   /// \p RuntimeCall into the offload arrays in \p OAs.
   bool getValuesInOffloadArrays(CallInst &RuntimeCall,
Index: llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
===================================================================
--- llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
+++ llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
@@ -144,9 +144,11 @@
 __OMP_TYPE(Void)
 __OMP_TYPE(Int1)
 __OMP_TYPE(Int8)
+__OMP_TYPE(Int16)
 __OMP_TYPE(Int32)
 __OMP_TYPE(Int64)
 __OMP_TYPE(Int8Ptr)
+__OMP_TYPE(Int16Ptr)
 __OMP_TYPE(Int32Ptr)
 __OMP_TYPE(Int64Ptr)
 
@@ -500,6 +502,13 @@
 __OMP_RTL(__kmpc_task_allow_completion_event, false, VoidPtr, IdentPtr,
           /* Int */ Int32, /* kmp_task_t */ VoidPtr)
 
+__OMP_RTL(__kmpc_data_sharing_init_stack, false, Void, )
+__OMP_RTL(__kmpc_data_sharing_init_stack_spmd, false, Void, ) 
+__OMP_RTL(__kmpc_data_sharing_coalesced_push_stack, false, VoidPtr, SizeTy,
+          Int16)
+__OMP_RTL(__kmpc_data_sharing_push_stack, false, VoidPtr, SizeTy, Int16)
+__OMP_RTL(__kmpc_data_sharing_pop_stack, false, Void, VoidPtr)
+
 /// Note that device runtime functions (in the following) do not necessarily
 /// need attributes as we expect to see the definitions.
 __OMP_RTL(__kmpc_kernel_parallel, false, Int1, VoidPtrPtr)


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D88243.294102.patch
Type: text/x-patch
Size: 2964 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200924/da2369bc/attachment.bin>