[PGO][OpenMP] Instrumentation for GPU devices (Revision of #76587) (#102691)

This pull request is a revised version of #76587. This pull request
fixes some build issues that were present in the previous version of
this change.

> This pull request is the first part of an ongoing effort to extends
PGO instrumentation to GPU device code. This PR makes the following
changes:
>
> - Adds blank registration functions to device RTL
> - Gives PGO globals protected visibility when targeting a supported
GPU
> - Handles any addrspace casts for PGO calls
> - Implements PGO global extraction in GPU plugins (currently only
dumps info)
>
> These changes can be tested by supplying `-fprofile-instrument=clang`
while targeting a GPU.

Author: EthanLuisMcDonough

clang/lib/CodeGen/CodeGenPGO.cpp

@@ -1195,10 +1195,15 @@ void CodeGenPGO::emitCounterSetOrIncrement(CGBuilderTy &Builder, const Stmt *S,
 
   unsigned Counter = (*RegionCounterMap)[S];
 
-  llvm::Value *Args[] = {FuncNameVar,
-                         Builder.getInt64(FunctionHash),
-                         Builder.getInt32(NumRegionCounters),
-                         Builder.getInt32(Counter), StepV};
+  // Make sure that pointer to global is passed in with zero addrspace
+  // This is relevant during GPU profiling
+  auto *NormalizedFuncNameVarPtr =
+      llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+          FuncNameVar, llvm::PointerType::get(CGM.getLLVMContext(), 0));
+
+  llvm::Value *Args[] = {
+      NormalizedFuncNameVarPtr, Builder.getInt64(FunctionHash),
+      Builder.getInt32(NumRegionCounters), Builder.getInt32(Counter), StepV};
 
   if (llvm::EnableSingleByteCoverage)
     Builder.CreateCall(CGM.getIntrinsic(llvm::Intrinsic::instrprof_cover),

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

@@ -506,6 +506,9 @@ __OMP_RTL(__kmpc_barrier_simple_generic, false, Void, IdentPtr, Int32)
 __OMP_RTL(__kmpc_warp_active_thread_mask, false, Int64,)
 __OMP_RTL(__kmpc_syncwarp, false, Void, Int64)
 
+__OMP_RTL(__llvm_profile_register_function, false, Void, VoidPtr)
+__OMP_RTL(__llvm_profile_register_names_function, false, Void, VoidPtr, Int64)
+
 __OMP_RTL(__last, false, Void, )
 
 #undef __OMP_RTL

llvm/include/llvm/ProfileData/InstrProf.h

@@ -181,6 +181,10 @@ inline StringRef getInstrProfBitmapBiasVarName() {
 /// Return the marker used to separate PGO names during serialization.
 inline StringRef getInstrProfNameSeparator() { return "\01"; }
 
+/// Determines whether module targets a GPU eligable for PGO
+/// instrumentation
+bool isGPUProfTarget(const Module &M);
+
 /// Please use getIRPGOFuncName for LLVM IR instrumentation. This function is
 /// for front-end (Clang, etc) instrumentation.
 /// Return the modified name for function \c F suitable to be

llvm/lib/ProfileData/InstrProf.cpp

@@ -437,13 +437,31 @@ std::string getPGOFuncNameVarName(StringRef FuncName,
   return VarName;
 }
 
+bool isGPUProfTarget(const Module &M) {
+  const auto &T = Triple(M.getTargetTriple());
+  return T.isAMDGPU() || T.isNVPTX();
+}
+
+void setPGOFuncVisibility(Module &M, GlobalVariable *FuncNameVar) {
+  // If the target is a GPU, make the symbol protected so it can
+  // be read from the host device
+  if (isGPUProfTarget(M))
+    FuncNameVar->setVisibility(GlobalValue::ProtectedVisibility);
+  // Hide the symbol so that we correctly get a copy for each executable.
+  else if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
+    FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
+}
+
 GlobalVariable *createPGOFuncNameVar(Module &M,
                                      GlobalValue::LinkageTypes Linkage,
                                      StringRef PGOFuncName) {
+  // Ensure profiling variables on GPU are visible to be read from host
+  if (isGPUProfTarget(M))
+    Linkage = GlobalValue::ExternalLinkage;
   // We generally want to match the function's linkage, but available_externally
   // and extern_weak both have the wrong semantics, and anything that doesn't
   // need to link across compilation units doesn't need to be visible at all.
-  if (Linkage == GlobalValue::ExternalWeakLinkage)
+  else if (Linkage == GlobalValue::ExternalWeakLinkage)
     Linkage = GlobalValue::LinkOnceAnyLinkage;
   else if (Linkage == GlobalValue::AvailableExternallyLinkage)
     Linkage = GlobalValue::LinkOnceODRLinkage;
@@ -457,10 +475,7 @@ GlobalVariable *createPGOFuncNameVar(Module &M,
       new GlobalVariable(M, Value->getType(), true, Linkage, Value,
                          getPGOFuncNameVarName(PGOFuncName, Linkage));
 
-  // Hide the symbol so that we correctly get a copy for each executable.
-  if (!GlobalValue::isLocalLinkage(FuncNameVar->getLinkage()))
-    FuncNameVar->setVisibility(GlobalValue::HiddenVisibility);
-
+  setPGOFuncVisibility(M, FuncNameVar);
   return FuncNameVar;
 }
 

llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp

@@ -1059,6 +1059,8 @@ void InstrLowerer::lowerValueProfileInst(InstrProfValueProfileInst *Ind) {
                       llvm::InstrProfValueKind::IPVK_MemOPSize);
   CallInst *Call = nullptr;
   auto *TLI = &GetTLI(*Ind->getFunction());
+  auto *NormalizedDataVarPtr = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+      DataVar, PointerType::get(M.getContext(), 0));
 
   // To support value profiling calls within Windows exception handlers, funclet
   // information contained within operand bundles needs to be copied over to
@@ -1067,11 +1069,13 @@ void InstrLowerer::lowerValueProfileInst(InstrProfValueProfileInst *Ind) {
   SmallVector<OperandBundleDef, 1> OpBundles;
   Ind->getOperandBundlesAsDefs(OpBundles);
   if (!IsMemOpSize) {
-    Value *Args[3] = {Ind->getTargetValue(), DataVar, Builder.getInt32(Index)};
+    Value *Args[3] = {Ind->getTargetValue(), NormalizedDataVarPtr,
+                      Builder.getInt32(Index)};
     Call = Builder.CreateCall(getOrInsertValueProfilingCall(M, *TLI), Args,
                               OpBundles);
   } else {
-    Value *Args[3] = {Ind->getTargetValue(), DataVar, Builder.getInt32(Index)};
+    Value *Args[3] = {Ind->getTargetValue(), NormalizedDataVarPtr,
+                      Builder.getInt32(Index)};
     Call = Builder.CreateCall(
         getOrInsertValueProfilingCall(M, *TLI, ValueProfilingCallType::MemOp),
         Args, OpBundles);
@@ -1814,7 +1818,8 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
         getInstrProfSectionName(IPSK_vals, TT.getObjectFormat()));
     ValuesVar->setAlignment(Align(8));
     maybeSetComdat(ValuesVar, Fn, CntsVarName);
-    ValuesPtrExpr = ValuesVar;
+    ValuesPtrExpr = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+        ValuesVar, PointerType::get(Fn->getContext(), 0));
   }
 
   uint64_t NumCounters = Inc->getNumCounters()->getZExtValue();
@@ -1838,6 +1843,10 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
     Int16ArrayVals[Kind] = ConstantInt::get(Int16Ty, PD.NumValueSites[Kind]);
 
+  if (isGPUProfTarget(M)) {
+    Linkage = GlobalValue::ExternalLinkage;
+    Visibility = GlobalValue::ProtectedVisibility;
+  }
   // If the data variable is not referenced by code (if we don't emit
   // @llvm.instrprof.value.profile, NS will be 0), and the counter keeps the
   // data variable live under linker GC, the data variable can be private. This
@@ -1849,9 +1858,9 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
   // If profd is in a deduplicate comdat, NS==0 with a hash suffix guarantees
   // that other copies must have the same CFG and cannot have value profiling.
   // If no hash suffix, other profd copies may be referenced by code.
-  if (NS == 0 && !(DataReferencedByCode && NeedComdat && !Renamed) &&
-      (TT.isOSBinFormatELF() ||
-       (!DataReferencedByCode && TT.isOSBinFormatCOFF()))) {
+  else if (NS == 0 && !(DataReferencedByCode && NeedComdat && !Renamed) &&
+           (TT.isOSBinFormatELF() ||
+            (!DataReferencedByCode && TT.isOSBinFormatCOFF()))) {
     Linkage = GlobalValue::PrivateLinkage;
     Visibility = GlobalValue::DefaultVisibility;
   }
@@ -1974,6 +1983,13 @@ void InstrLowerer::emitNameData() {
   NamesVar = new GlobalVariable(M, NamesVal->getType(), true,
                                 GlobalValue::PrivateLinkage, NamesVal,
                                 getInstrProfNamesVarName());
+
+  // Make names variable public if current target is a GPU
+  if (isGPUProfTarget(M)) {
+    NamesVar->setLinkage(GlobalValue::ExternalLinkage);
+    NamesVar->setVisibility(GlobalValue::VisibilityTypes::ProtectedVisibility);
+  }
+
   NamesSize = CompressedNameStr.size();
   setGlobalVariableLargeSection(TT, *NamesVar);
   NamesVar->setSection(
@@ -2040,10 +2056,13 @@ void InstrLowerer::emitRegistration() {
   IRBuilder<> IRB(BasicBlock::Create(M.getContext(), "", RegisterF));
   for (Value *Data : CompilerUsedVars)
     if (!isa<Function>(Data))
-      IRB.CreateCall(RuntimeRegisterF, Data);
+      // Check for addrspace cast when profiling GPU
+      IRB.CreateCall(RuntimeRegisterF,
+                     IRB.CreatePointerBitCastOrAddrSpaceCast(Data, VoidPtrTy));
   for (Value *Data : UsedVars)
     if (Data != NamesVar && !isa<Function>(Data))
-      IRB.CreateCall(RuntimeRegisterF, Data);
+      IRB.CreateCall(RuntimeRegisterF,
+                     IRB.CreatePointerBitCastOrAddrSpaceCast(Data, VoidPtrTy));
 
   if (NamesVar) {
     Type *ParamTypes[] = {VoidPtrTy, Int64Ty};
@@ -2052,7 +2071,9 @@ void InstrLowerer::emitRegistration() {
     auto *NamesRegisterF =
         Function::Create(NamesRegisterTy, GlobalVariable::ExternalLinkage,
                          getInstrProfNamesRegFuncName(), M);
-    IRB.CreateCall(NamesRegisterF, {NamesVar, IRB.getInt64(NamesSize)});
+    IRB.CreateCall(NamesRegisterF, {IRB.CreatePointerBitCastOrAddrSpaceCast(
+                                        NamesVar, VoidPtrTy),
+                                    IRB.getInt64(NamesSize)});
   }
 
   IRB.CreateRetVoid();
@@ -2073,7 +2094,10 @@ bool InstrLowerer::emitRuntimeHook() {
   auto *Var =
       new GlobalVariable(M, Int32Ty, false, GlobalValue::ExternalLinkage,
                          nullptr, getInstrProfRuntimeHookVarName());
-  Var->setVisibility(GlobalValue::HiddenVisibility);
+  if (isGPUProfTarget(M))
+    Var->setVisibility(GlobalValue::ProtectedVisibility);
+  else
+    Var->setVisibility(GlobalValue::HiddenVisibility);
 
   if (TT.isOSBinFormatELF() && !TT.isPS()) {
     // Mark the user variable as used so that it isn't stripped out.

llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp

@@ -909,14 +909,18 @@ void FunctionInstrumenter::instrument() {
   auto Name = FuncInfo.FuncNameVar;
   auto CFGHash =
       ConstantInt::get(Type::getInt64Ty(M.getContext()), FuncInfo.FunctionHash);
+  // Make sure that pointer to global is passed in with zero addrspace
+  // This is relevant during GPU profiling
+  auto *NormalizedNamePtr = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+      Name, PointerType::get(M.getContext(), 0));
   if (PGOFunctionEntryCoverage) {
     auto &EntryBB = F.getEntryBlock();
     IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
     // llvm.instrprof.cover(i8* <name>, i64 <hash>, i32 <num-counters>,
     //                      i32 <index>)
     Builder.CreateCall(
         Intrinsic::getDeclaration(&M, Intrinsic::instrprof_cover),
-        {Name, CFGHash, Builder.getInt32(1), Builder.getInt32(0)});
+        {NormalizedNamePtr, CFGHash, Builder.getInt32(1), Builder.getInt32(0)});
     return;
   }
 
@@ -971,7 +975,8 @@ void FunctionInstrumenter::instrument() {
     //                          i32 <index>)
     Builder.CreateCall(
         Intrinsic::getDeclaration(&M, Intrinsic::instrprof_timestamp),
-        {Name, CFGHash, Builder.getInt32(NumCounters), Builder.getInt32(I)});
+        {NormalizedNamePtr, CFGHash, Builder.getInt32(NumCounters),
+         Builder.getInt32(I)});
     I += PGOBlockCoverage ? 8 : 1;
   }
 
@@ -985,7 +990,8 @@ void FunctionInstrumenter::instrument() {
         Intrinsic::getDeclaration(&M, PGOBlockCoverage
                                           ? Intrinsic::instrprof_cover
                                           : Intrinsic::instrprof_increment),
-        {Name, CFGHash, Builder.getInt32(NumCounters), Builder.getInt32(I++)});
+        {NormalizedNamePtr, CFGHash, Builder.getInt32(NumCounters),
+         Builder.getInt32(I++)});
   }
 
   // Now instrument select instructions:
@@ -1028,11 +1034,14 @@ void FunctionInstrumenter::instrument() {
         ToProfile = Builder.CreatePtrToInt(Cand.V, Builder.getInt64Ty());
       assert(ToProfile && "value profiling Value is of unexpected type");
 
+      auto *NormalizedNamePtr = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+          Name, PointerType::get(M.getContext(), 0));
+
       SmallVector<OperandBundleDef, 1> OpBundles;
       populateEHOperandBundle(Cand, BlockColors, OpBundles);
       Builder.CreateCall(
           Intrinsic::getDeclaration(&M, Intrinsic::instrprof_value_profile),
-          {FuncInfo.FuncNameVar, Builder.getInt64(FuncInfo.FunctionHash),
+          {NormalizedNamePtr, Builder.getInt64(FuncInfo.FunctionHash),
            ToProfile, Builder.getInt32(Kind), Builder.getInt32(SiteIndex++)},
           OpBundles);
     }
@@ -1709,10 +1718,13 @@ void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) {
   IRBuilder<> Builder(&SI);
   Type *Int64Ty = Builder.getInt64Ty();
   auto *Step = Builder.CreateZExt(SI.getCondition(), Int64Ty);
+  auto *NormalizedFuncNameVarPtr =
+      ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+          FuncNameVar, PointerType::get(M->getContext(), 0));
   Builder.CreateCall(
       Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment_step),
-      {FuncNameVar, Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs),
-       Builder.getInt32(*CurCtrIdx), Step});
+      {NormalizedFuncNameVarPtr, Builder.getInt64(FuncHash),
+       Builder.getInt32(TotalNumCtrs), Builder.getInt32(*CurCtrIdx), Step});
   ++(*CurCtrIdx);
 }
 

offload/DeviceRTL/CMakeLists.txt

@@ -77,6 +77,7 @@ set(include_files
   ${include_directory}/Interface.h
   ${include_directory}/LibC.h
   ${include_directory}/Mapping.h
+  ${include_directory}/Profiling.h
   ${include_directory}/State.h
   ${include_directory}/Synchronization.h
   ${include_directory}/Types.h
@@ -93,6 +94,7 @@ set(src_files
   ${source_directory}/Mapping.cpp
   ${source_directory}/Misc.cpp
   ${source_directory}/Parallelism.cpp
+  ${source_directory}/Profiling.cpp
   ${source_directory}/Reduction.cpp
   ${source_directory}/State.cpp
   ${source_directory}/Synchronization.cpp

offload/DeviceRTL/include/Profiling.h

@@ -0,0 +1,21 @@
+//===-------- Profiling.h - OpenMP interface ---------------------- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef OMPTARGET_DEVICERTL_PROFILING_H
+#define OMPTARGET_DEVICERTL_PROFILING_H
+
+extern "C" {
+void __llvm_profile_register_function(void *Ptr);
+void __llvm_profile_register_names_function(void *Ptr, long int I);
+void __llvm_profile_instrument_memop(long int I, void *Ptr, int I2);
+}
+
+#endif

offload/DeviceRTL/src/Profiling.cpp

@@ -0,0 +1,22 @@
+//===------- Profiling.cpp ---------------------------------------- C++ ---===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "Profiling.h"
+
+#pragma omp begin declare target device_type(nohost)
+
+extern "C" {
+
+// Provides empty implementations for certain functions in compiler-rt
+// that are emitted by the PGO instrumentation.
+void __llvm_profile_register_function(void *Ptr) {}
+void __llvm_profile_register_names_function(void *Ptr, long int I) {}
+void __llvm_profile_instrument_memop(long int I, void *Ptr, int I2) {}
+}
+
+#pragma omp end declare target

offload/plugins-nextgen/common/CMakeLists.txt

@@ -7,7 +7,7 @@ add_library(PluginCommon OBJECT
   src/RPC.cpp
   src/Utils/ELF.cpp
 )
-add_dependencies(PluginCommon intrinsics_gen)
+add_dependencies(PluginCommon intrinsics_gen LLVMProfileData)
 
 # Only enable JIT for those targets that LLVM can support.
 set(supported_jit_targets AMDGPU NVPTX)
@@ -52,6 +52,7 @@ target_compile_definitions(PluginCommon PRIVATE
 
 target_compile_options(PluginCommon PUBLIC ${offload_compile_flags})
 target_link_options(PluginCommon PUBLIC ${offload_link_flags})
+target_link_libraries(PluginCommon PRIVATE LLVMProfileData)
 
 target_include_directories(PluginCommon PUBLIC
   ${CMAKE_CURRENT_SOURCE_DIR}/include