@@ -933,9 +933,53 @@ static void addPreloadKernArgHint(Function &F, TargetMachine &TM) {
933933 }
934934}
935935
936- class AMDGPUAttributor : public ModulePass {
936+ static bool runImpl (Module &M, AnalysisGetter &AG, TargetMachine &TM) {
937+ SetVector<Function *> Functions;
938+ for (Function &F : M) {
939+ if (!F.isIntrinsic ())
940+ Functions.insert (&F);
941+ }
942+
943+ CallGraphUpdater CGUpdater;
944+ BumpPtrAllocator Allocator;
945+ AMDGPUInformationCache InfoCache (M, AG, Allocator, nullptr , TM);
946+ DenseSet<const char *> Allowed (
947+ {&AAAMDAttributes::ID, &AAUniformWorkGroupSize::ID,
948+ &AAPotentialValues::ID, &AAAMDFlatWorkGroupSize::ID,
949+ &AAAMDWavesPerEU::ID, &AACallEdges::ID, &AAPointerInfo::ID,
950+ &AAPotentialConstantValues::ID, &AAUnderlyingObjects::ID});
951+
952+ AttributorConfig AC (CGUpdater);
953+ AC.Allowed = &Allowed;
954+ AC.IsModulePass = true ;
955+ AC.DefaultInitializeLiveInternals = false ;
956+ AC.IPOAmendableCB = [](const Function &F) {
957+ return F.getCallingConv () == CallingConv::AMDGPU_KERNEL;
958+ };
959+
960+ Attributor A (Functions, InfoCache, AC);
961+
962+ for (Function &F : M) {
963+ if (!F.isIntrinsic ()) {
964+ A.
getOrCreateAAFor <AAAMDAttributes>(
IRPosition::function (F));
965+ A.
getOrCreateAAFor <AAUniformWorkGroupSize>(
IRPosition::function (F));
966+ CallingConv::ID CC = F.getCallingConv ();
967+ if (!AMDGPU::isEntryFunctionCC (CC)) {
968+ A.
getOrCreateAAFor <AAAMDFlatWorkGroupSize>(
IRPosition::function (F));
969+ A.
getOrCreateAAFor <AAAMDWavesPerEU>(
IRPosition::function (F));
970+ } else if (CC == CallingConv::AMDGPU_KERNEL) {
971+ addPreloadKernArgHint (F, TM);
972+ }
973+ }
974+ }
975+
976+ ChangeStatus Change = A.run ();
977+ return Change == ChangeStatus::CHANGED;
978+ }
979+
980+ class AMDGPUAttributorLegacy : public ModulePass {
937981public:
938- AMDGPUAttributor () : ModulePass(ID) {}
982+ AMDGPUAttributorLegacy () : ModulePass(ID) {}
939983
940984 // / doInitialization - Virtual method overridden by subclasses to do
941985 // / any necessary initialization before any pass is run.
@@ -949,48 +993,8 @@ class AMDGPUAttributor : public ModulePass {
949993 }
950994
951995 bool runOnModule (Module &M) override {
952- SetVector<Function *> Functions;
953996 AnalysisGetter AG (this );
954- for (Function &F : M) {
955- if (!F.isIntrinsic ())
956- Functions.insert (&F);
957- }
958-
959- CallGraphUpdater CGUpdater;
960- BumpPtrAllocator Allocator;
961- AMDGPUInformationCache InfoCache (M, AG, Allocator, nullptr , *TM);
962- DenseSet<const char *> Allowed (
963- {&AAAMDAttributes::ID, &AAUniformWorkGroupSize::ID,
964- &AAPotentialValues::ID, &AAAMDFlatWorkGroupSize::ID,
965- &AAAMDWavesPerEU::ID, &AACallEdges::ID, &AAPointerInfo::ID,
966- &AAPotentialConstantValues::ID, &AAUnderlyingObjects::ID});
967-
968- AttributorConfig AC (CGUpdater);
969- AC.Allowed = &Allowed;
970- AC.IsModulePass = true ;
971- AC.DefaultInitializeLiveInternals = false ;
972- AC.IPOAmendableCB = [](const Function &F) {
973- return F.getCallingConv () == CallingConv::AMDGPU_KERNEL;
974- };
975-
976- Attributor A (Functions, InfoCache, AC);
977-
978- for (Function &F : M) {
979- if (!F.isIntrinsic ()) {
980- A.
getOrCreateAAFor <AAAMDAttributes>(
IRPosition::function (F));
981- A.
getOrCreateAAFor <AAUniformWorkGroupSize>(
IRPosition::function (F));
982- CallingConv::ID CC = F.getCallingConv ();
983- if (!AMDGPU::isEntryFunctionCC (CC)) {
984- A.
getOrCreateAAFor <AAAMDFlatWorkGroupSize>(
IRPosition::function (F));
985- A.
getOrCreateAAFor <AAAMDWavesPerEU>(
IRPosition::function (F));
986- } else if (CC == CallingConv::AMDGPU_KERNEL) {
987- addPreloadKernArgHint (F, *TM);
988- }
989- }
990- }
991-
992- ChangeStatus Change = A.run ();
993- return Change == ChangeStatus::CHANGED;
997+ return runImpl (M, AG, *TM);
994998 }
995999
9961000 void getAnalysisUsage (AnalysisUsage &AU) const override {
@@ -1003,11 +1007,25 @@ class AMDGPUAttributor : public ModulePass {
10031007};
10041008} // namespace
10051009
1006- char AMDGPUAttributor::ID = 0 ;
1010+ PreservedAnalyses llvm::AMDGPUAttributorPass::run (Module &M,
1011+ ModuleAnalysisManager &AM) {
10071012
1008- Pass *llvm::createAMDGPUAttributorPass () { return new AMDGPUAttributor (); }
1009- INITIALIZE_PASS_BEGIN (AMDGPUAttributor, DEBUG_TYPE, " AMDGPU Attributor" false ,
1010- false )
1013+ FunctionAnalysisManager &FAM =
1014+ AM.getResult <FunctionAnalysisManagerModuleProxy>(M).getManager ();
1015+ AnalysisGetter AG (FAM);
1016+
1017+ // TODO: Probably preserves CFG
1018+ return runImpl (M, AG, TM) ? PreservedAnalyses::none ()
1019+ : PreservedAnalyses::all ();
1020+ }
1021+
1022+ char AMDGPUAttributorLegacy::ID = 0 ;
1023+
1024+ Pass *llvm::createAMDGPUAttributorLegacyPass () {
1025+ return new AMDGPUAttributorLegacy ();
1026+ }
1027+ INITIALIZE_PASS_BEGIN (AMDGPUAttributorLegacy, DEBUG_TYPE, " AMDGPU Attributor"
1028+ false , false )
10111029INITIALIZE_PASS_DEPENDENCY(CycleInfoWrapperPass);
1012- INITIALIZE_PASS_END (AMDGPUAttributor , DEBUG_TYPE, " AMDGPU Attributor" , false ,
1013- false )
1030+ INITIALIZE_PASS_END (AMDGPUAttributorLegacy , DEBUG_TYPE, " AMDGPU Attributor"
1031+ false , false )
0 commit comments