[OpenMP] Prevent OpenMPOpt from internalizing uncalled functions

[jhuber6]

Summary

Currently OpenMPOpt will only check if a function is a kernel before deciding not to internalize it. Any uncalled function that gets internalized will be trivially dead in the module so this is unnnecessary.

Depends on D102423

[jdoerfert]

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

2655

Up to this point it makes sense. (clang-tidy/format!)

2661

This change is what irritates me. We basically won't emit a remark anymore, nor will we try hard to optimize a function, but we should. The user now doesn't know that we kinda ignored "unsued" below even though it can be very much used from outside the translation unit. We still didn't run H2S on it or reported that globalization was left. Instead, we should report there is a external version of a function (regardless of the internalization) and we should report if the external version has leftover globalization, for example.

[jhuber6]

Adding uncalled functions to SCC. This will make the module pass actually run on uncalled functions.

[jdoerfert]

LGTM

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

 PreservedAnalyses OpenMPOptPass::run(Module &M, ModuleAnalysisManager &AM) {
   if (!containsOpenMP(M))
     return PreservedAnalyses::all();
   if (DisableOpenMPOptimizations)
     return PreservedAnalyses::all();
 
   KernelSet Kernels = getDeviceKernels(M);
 
-  // Create internal copies of each function if this is a kernel Module.
+  auto IsCalled = [&](Function &F) {
+    if (Kernels.contains(&F))
+      return true;
+    for (const User *U : F.users())
+      if (!isa<BlockAddress>(U))
+        return true;
+    return false;
+  };
+
+  // Create internal copies of each function if this is a kernel Module. This
+  // allows iterprocedural passes to see every call edge.
   DenseSet<const Function *> InternalizedFuncs;
   if (isOpenMPDevice(M))
     for (Function &F : M)
-      if (!F.isDeclaration() && !Kernels.contains(&F))
+      if (!F.isDeclaration() && !Kernels.contains(&F) && IsCalled(F))
         if (Attributor::internalizeFunction(F, /* Force */ true))
           InternalizedFuncs.insert(&F);
 
-  // Look at every function definition in the Module that wasn't internalized.
+  // Look at every function in the Module unless it was internalized.
   SmallVector<Function *, 16> SCC;
   for (Function &F : M)
     if (!F.isDeclaration() && !InternalizedFuncs.contains(&F))
       SCC.push_back(&F);
 
   if (SCC.empty())
     return PreservedAnalyses::all();
 
   FunctionAnalysisManager &FAM =
       AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
 
   AnalysisGetter AG(FAM);

llvm/test/Transforms/OpenMP/remove_globalization.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature
 ; RUN: opt -S -passes=openmp-opt < %s | FileCheck %s
 ; RUN: opt -passes=openmp-opt -pass-remarks=openmp-opt -pass-remarks-missed=openmp-opt -disable-output < %s 2>&1 | FileCheck %s -check-prefix=CHECK-REMARKS
 target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
 target triple = "nvptx64"
 
 ; CHECK-REMARKS: remark: remove_globalization.c:4:2: Could not move globalized variable to the stack. Variable is potentially captured.
 ; CHECK-REMARKS: remark: remove_globalization.c:2:2: Moving globalized variable to the stack.
+; CHECK-REMARKS: remark: remove_globalization.c:6:2: Moving globalized variable to the stack.
 
 @S = external local_unnamed_addr global i8*
 
 define void @kernel() {
 ; CHECK-LABEL: define {{[^@]+}}@kernel() {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    call void @foo() #[[ATTR0:[0-9]+]]
 ; CHECK-NEXT:    call void @bar() #[[ATTR0]]
 ; CHECK-NEXT:    ret void
 ;
 entry:
   call void @foo()
   call void @bar()
   ret void
 }
 
 define internal void @foo() {
 ; CHECK-LABEL: define {{[^@]+}}@foo
 ; CHECK-SAME: () #[[ATTR0]] {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = alloca i8, i64 4, align 1
 ; CHECK-NEXT:    ret void
 ;
 entry:
-  %0 = call i8* @__kmpc_alloc_shared(i64 4), !dbg !11
+  %0 = call i8* @__kmpc_alloc_shared(i64 4), !dbg !12
   call void @use(i8* %0)
   call void @__kmpc_free_shared(i8* %0)
   ret void
 }
 
 define internal void @bar() {
 ; CHECK-LABEL: define {{[^@]+}}@bar
 ; CHECK-SAME: () #[[ATTR0]] {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = call i8* @__kmpc_alloc_shared(i64 noundef 4) #[[ATTR0]]
 ; CHECK-NEXT:    call void @share(i8* nofree writeonly [[TMP0]]) #[[ATTR2:[0-9]+]]
 ; CHECK-NEXT:    call void @__kmpc_free_shared(i8* [[TMP0]]) #[[ATTR0]]
 ; CHECK-NEXT:    ret void
 ;
 entry:
-  %0 = call i8* @__kmpc_alloc_shared(i64 4), !dbg !12
+  %0 = call i8* @__kmpc_alloc_shared(i64 4), !dbg !13
   call void @share(i8* %0)
   call void @__kmpc_free_shared(i8* %0)
   ret void
 }
 
 define internal void @use(i8* %x) {
 entry:
   ret void
 }
 
 define internal void @share(i8* %x) {
 ; CHECK-LABEL: define {{[^@]+}}@share
 ; CHECK-SAME: (i8* nofree writeonly [[X:%.*]]) #[[ATTR1:[0-9]+]] {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    store i8* [[X]], i8** @S, align 8
 ; CHECK-NEXT:    ret void
 ;
 entry:
   store i8* %x, i8** @S
   ret void
 }
 
+define void @unused() {
+entry:
+  %0 = call i8* @__kmpc_alloc_shared(i64 4), !dbg !14
+  call void @use(i8* %0)
+  call void @__kmpc_free_shared(i8* %0)
+  ret void
+}
+
 ; CHECK: declare i8* @__kmpc_alloc_shared(i64)
 declare i8* @__kmpc_alloc_shared(i64)
 
 ; CHECK: declare void @__kmpc_free_shared(i8* nocapture)
 declare void @__kmpc_free_shared(i8*)
 
 !llvm.dbg.cu = !{!0}
 !llvm.module.flags = !{!3, !4, !6, !7}
 !nvvm.annotations = !{!5}
 
 !0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 13.0.0", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, splitDebugInlining: false, nameTableKind: None)
 !1 = !DIFile(filename: "remove_globalization.c", directory: "/tmp/remove_globalization.c")
 !2 = !{}
 !3 = !{i32 2, !"Debug Info Version", i32 3}
 !4 = !{i32 1, !"wchar_size", i32 4}
 !5 = !{void ()* @kernel, !"kernel", i32 1}
 !6 = !{i32 7, !"openmp", i32 50}
 !7 = !{i32 7, !"openmp-device", i32 50}
-!8 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !10, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
-!9 = distinct !DISubprogram(name: "bar", scope: !1, file: !1, line: 1, type: !10, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
-!10 = !DISubroutineType(types: !2)
-!11 = !DILocation(line: 2, column: 2, scope: !8)
-!12 = !DILocation(line: 4, column: 2, scope: !9)
+!8 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !11, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
+!9 = distinct !DISubprogram(name: "bar", scope: !1, file: !1, line: 1, type: !11, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
+!10 = distinct !DISubprogram(name: "unused", scope: !1, file: !1, line: 1, type: !11, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
+!11 = !DISubroutineType(types: !2)
+!12 = !DILocation(line: 2, column: 2, scope: !8)
+!13 = !DILocation(line: 4, column: 2, scope: !9)
+!14 = !DILocation(line: 6, column: 2, scope: !9)