From: Jonas Hahnfeld <hahnjo@hahnjo.de>
Date: Tue, 2 Oct 2018 19:12:54 +0000 (+0000)
Subject: [OpenMP][NVPTX] Simplify codegen for orphaned parallel, NFCI.
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=d08b7d2eb6476853396974d98a75b2985cfc09b8;p=clang

[OpenMP][NVPTX] Simplify codegen for orphaned parallel, NFCI.

Worker threads fork off to the compiler generated worker function
directly after entering the kernel function. Hence, there is no
need to check whether the current thread is the master if we are
outside of a parallel region (neither SPMD nor parallel_level > 0).

Differential Revision: https://reviews.llvm.org/D52732

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@343618 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index 96cf6c1173..a923232e81 100644
--- a/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -2233,30 +2233,24 @@ void CGOpenMPRuntimeNVPTX::emitNonSPMDParallelCall(
     Work.emplace_back(WFn);
   };
 
-  auto &&LNParallelGen = [this, Loc, &SeqGen, &L0ParallelGen, &CodeGen,
-                          &ThreadIDAddr](CodeGenFunction &CGF,
-                                         PrePostActionTy &Action) {
-    RegionCodeGenTy RCG(CodeGen);
+  auto &&LNParallelGen = [this, Loc, &SeqGen, &L0ParallelGen](
+                             CodeGenFunction &CGF, PrePostActionTy &Action) {
     if (IsInParallelRegion) {
       SeqGen(CGF, Action);
     } else if (IsInTargetMasterThreadRegion) {
       L0ParallelGen(CGF, Action);
-    } else if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_NonSPMD) {
-      RCG(CGF);
     } else {
       // Check for master and then parallelism:
       // if (__kmpc_is_spmd_exec_mode() || __kmpc_parallel_level(loc, gtid)) {
-      //  Serialized execution.
-      // } else if (master) {
-      //   Worker call.
+      //   Serialized execution.
       // } else {
-      //   Outlined function call.
+      //   Worker call.
       // }
       CGBuilderTy &Bld = CGF.Builder;
       llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
       llvm::BasicBlock *SeqBB = CGF.createBasicBlock(".sequential");
       llvm::BasicBlock *ParallelCheckBB = CGF.createBasicBlock(".parcheck");
-      llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck");
+      llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
       llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall(
           createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode)));
       Bld.CreateCondBr(IsSPMD, SeqBB, ParallelCheckBB);
@@ -2269,29 +2263,17 @@ void CGOpenMPRuntimeNVPTX::emitNonSPMDParallelCall(
           createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_parallel_level),
           {RTLoc, ThreadID});
       llvm::Value *Res = Bld.CreateIsNotNull(PL);
-      Bld.CreateCondBr(Res, SeqBB, MasterCheckBB);
+      Bld.CreateCondBr(Res, SeqBB, MasterBB);
       CGF.EmitBlock(SeqBB);
       SeqGen(CGF, Action);
       CGF.EmitBranch(ExitBB);
       // There is no need to emit line number for unconditional branch.
       (void)ApplyDebugLocation::CreateEmpty(CGF);
-      CGF.EmitBlock(MasterCheckBB);
-      llvm::BasicBlock *MasterThenBB = CGF.createBasicBlock("master.then");
-      llvm::BasicBlock *ElseBlock = CGF.createBasicBlock("omp_if.else");
-      llvm::Value *IsMaster =
-          Bld.CreateICmpEQ(getNVPTXThreadID(CGF), getMasterThreadID(CGF));
-      Bld.CreateCondBr(IsMaster, MasterThenBB, ElseBlock);
-      CGF.EmitBlock(MasterThenBB);
+      CGF.EmitBlock(MasterBB);
       L0ParallelGen(CGF, Action);
       CGF.EmitBranch(ExitBB);
       // There is no need to emit line number for unconditional branch.
       (void)ApplyDebugLocation::CreateEmpty(CGF);
-      CGF.EmitBlock(ElseBlock);
-      // In the worker need to use the real thread id.
-      ThreadIDAddr = emitThreadIDAddress(CGF, Loc);
-      RCG(CGF);
-      // There is no need to emit line number for unconditional branch.
-      (void)ApplyDebugLocation::CreateEmpty(CGF);
       // Emit the continuation block for code after the if.
       CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
     }
diff --git a/test/OpenMP/nvptx_target_codegen.cpp b/test/OpenMP/nvptx_target_codegen.cpp
index e404978290..01f346121a 100644
--- a/test/OpenMP/nvptx_target_codegen.cpp
+++ b/test/OpenMP/nvptx_target_codegen.cpp
@@ -557,7 +557,6 @@ int baz(int f, double &a) {
   // CHECK: [[STACK:%.+]] = alloca [[GLOBAL_ST:%.+]],
   // CHECK: [[ZERO_ADDR:%.+]] = alloca i32,
   // CHECK: [[GTID:%.+]] = call i32 @__kmpc_global_thread_num(%struct.ident_t*
-  // CHECK: [[GTID_ADDR:%.+]] = alloca i32,
   // CHECK: store i32 0, i32* [[ZERO_ADDR]]
   // CHECK: [[RES:%.+]] = call i8 @__kmpc_is_spmd_exec_mode()
   // CHECK: [[IS_SPMD:%.+]] = icmp ne i8 [[RES]], 0
@@ -583,9 +582,6 @@ int baz(int f, double &a) {
   // CHECK: call void @__kmpc_end_serialized_parallel(%struct.ident_t* @{{.+}}, i32 [[GTID]])
   // CHECK: br label
 
-  // CHECK: icmp eq i32
-  // CHECK: br i1
-
   // CHECK: call void @__kmpc_kernel_prepare_parallel(i8* bitcast (void (i16, i32)* @{{.+}} to i8*), i16 1)
   // CHECK: call void @__kmpc_begin_sharing_variables(i8*** [[SHARED_PTR:%.+]], i{{64|32}} 2)
   // CHECK: [[SHARED:%.+]] = load i8**, i8*** [[SHARED_PTR]],
@@ -597,10 +593,6 @@ int baz(int f, double &a) {
   // CHECK: call void @__kmpc_end_sharing_variables()
   // CHECK: br label
 
-  // CHECK: store i32 [[GTID]], i32* [[GTID_ADDR]],
-  // CHECK: call void [[OUTLINED]](i32* [[GTID_ADDR]], i32* [[ZERO_ADDR]], i32* [[F_PTR]], double* %{{.+}})
-  // CHECK: br label
-
   // CHECK: [[RES:%.+]] = load i32, i32* [[F_PTR]],
   // CHECK: store i32 [[RES]], i32* [[RET:%.+]],
   // CHECK: br i1 [[IS_SPMD]], label