[PowerPC] Diagnose musttail instead of crash inside backend (llvm#93267)

musttail is not often possible to be generated on PPC targets as when
calling to a function defined in another module, PPC needs to restore
the TOC pointer. To restore the TOC pointer, compiler needs to emit a
nop after the call to let linker generate codes to restore TOC pointer.
Tail call cannot generate expected call sequence for this case.

To avoid the crash inside the compiler backend, a diagnosis is added in
the frontend.

Fixes llvm#63214

Author: Chen Zheng

clang/include/clang/Basic/DiagnosticCommonKinds.td

@@ -367,6 +367,15 @@ def warn_target_unrecognized_env : Warning<
 def err_target_unsupported_abi_with_fpu : Error<
   "'%0' ABI is not supported with FPU">;
 
+def err_ppc_impossible_musttail: Error<
+  "'musttail' attribute for this call is impossible because %select{"
+  "long calls can not be tail called on PPC|"
+  "indirect calls can not be tail called on PPC|"
+  "external calls can not be tail called on PPC}0"
+  >;
+def err_aix_musttail_unsupported: Error<
+  "'musttail' attribute is not supported on AIX">;
+
 // Source manager
 def err_cannot_open_file : Error<"cannot open file '%0': %1">, DefaultFatal;
 def err_file_modified : Error<

clang/lib/Basic/Targets/PPC.cpp

@@ -93,6 +93,8 @@ bool PPCTargetInfo::handleTargetFeatures(std::vector<std::string> &Features,
       HasQuadwordAtomics = true;
     } else if (Feature == "+aix-shared-lib-tls-model-opt") {
       HasAIXShLibTLSModelOpt = true;
+    } else if (Feature == "+longcall") {
+      UseLongCalls = true;
     }
     // TODO: Finish this list and add an assert that we've handled them
     // all.
@@ -728,6 +730,7 @@ bool PPCTargetInfo::hasFeature(StringRef Feature) const {
       .Case("isa-v31-instructions", IsISA3_1)
       .Case("quadword-atomics", HasQuadwordAtomics)
       .Case("aix-shared-lib-tls-model-opt", HasAIXShLibTLSModelOpt)
+      .Case("longcall", UseLongCalls)
       .Default(false);
 }
 

clang/lib/Basic/Targets/PPC.h

@@ -82,6 +82,7 @@ class LLVM_LIBRARY_VISIBILITY PPCTargetInfo : public TargetInfo {
   bool IsISA3_1 = false;
   bool HasQuadwordAtomics = false;
   bool HasAIXShLibTLSModelOpt = false;
+  bool UseLongCalls = false;
 
 protected:
   std::string ABI;

clang/lib/CodeGen/CGCall.cpp

@@ -5751,8 +5751,35 @@ RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
   if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
     if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
       Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
-    else if (IsMustTail)
+    else if (IsMustTail) {
+      if (getTarget().getTriple().isPPC()) {
+        if (getTarget().getTriple().isOSAIX())
+          CGM.getDiags().Report(Loc, diag::err_aix_musttail_unsupported);
+        else if (!getTarget().hasFeature("pcrelative-memops")) {
+          if (getTarget().hasFeature("longcall"))
+            CGM.getDiags().Report(Loc, diag::err_ppc_impossible_musttail) << 0;
+          else if (Call->isIndirectCall())
+            CGM.getDiags().Report(Loc, diag::err_ppc_impossible_musttail) << 1;
+          else if (isa_and_nonnull<FunctionDecl>(TargetDecl)) {
+            if (!cast<FunctionDecl>(TargetDecl)->isDefined())
+              // The undefined callee may be a forward declaration. Without
+              // knowning all symbols in the module, we won't know the symbol is
+              // defined or not. Collect all these symbols for later diagnosing.
+              CGM.addUndefinedGlobalForTailCall(
+                  {cast<FunctionDecl>(TargetDecl), Loc});
+            else {
+              llvm::GlobalValue::LinkageTypes Linkage = CGM.getFunctionLinkage(
+                  GlobalDecl(cast<FunctionDecl>(TargetDecl)));
+              if (llvm::GlobalValue::isWeakForLinker(Linkage) ||
+                  llvm::GlobalValue::isDiscardableIfUnused(Linkage))
+                CGM.getDiags().Report(Loc, diag::err_ppc_impossible_musttail)
+                    << 2;
+            }
+          }
+        }
+      }
       Call->setTailCallKind(llvm::CallInst::TCK_MustTail);
+    }
   }
 
   // Add metadata for calls to MSAllocator functions

clang/lib/CodeGen/CodeGenModule.cpp

@@ -1394,6 +1394,21 @@ void CodeGenModule::Release() {
   // that might affect the DLL storage class or the visibility, and
   // before anything that might act on these.
   setVisibilityFromDLLStorageClass(LangOpts, getModule());
+
+  // Check the tail call symbols are truly undefined.
+  if (getTriple().isPPC() && !MustTailCallUndefinedGlobals.empty()) {
+    for (auto &I : MustTailCallUndefinedGlobals) {
+      if (!I.first->isDefined())
+        getDiags().Report(I.second, diag::err_ppc_impossible_musttail) << 2;
+      else {
+        StringRef MangledName = getMangledName(GlobalDecl(I.first));
+        llvm::GlobalValue *Entry = GetGlobalValue(MangledName);
+        if (!Entry || Entry->isWeakForLinker() ||
+            Entry->isDeclarationForLinker())
+          getDiags().Report(I.second, diag::err_ppc_impossible_musttail) << 2;
+      }
+    }
+  }
 }
 
 void CodeGenModule::EmitOpenCLMetadata() {

clang/lib/CodeGen/CodeGenModule.h

@@ -485,6 +485,14 @@ class CodeGenModule : public CodeGenTypeCache {
   typedef std::pair<OrderGlobalInitsOrStermFinalizers, llvm::Function *>
       GlobalInitData;
 
+  // When a tail call is performed on an "undefined" symbol, on PPC without pc
+  // relative feature, the tail call is not allowed. In "EmitCall" for such
+  // tail calls, the "undefined" symbols may be forward declarations, their
+  // definitions are provided in the module after the callsites. For such tail
+  // calls, diagnose message should not be emitted.
+  llvm::SmallSetVector<std::pair<const FunctionDecl *, SourceLocation>, 4>
+      MustTailCallUndefinedGlobals;
+
   struct GlobalInitPriorityCmp {
     bool operator()(const GlobalInitData &LHS,
                     const GlobalInitData &RHS) const {
@@ -1647,6 +1655,11 @@ class CodeGenModule : public CodeGenTypeCache {
     return getTriple().isSPIRVLogical();
   }
 
+  void addUndefinedGlobalForTailCall(
+      std::pair<const FunctionDecl *, SourceLocation> Global) {
+    MustTailCallUndefinedGlobals.insert(Global);
+  }
+
 private:
   bool shouldDropDLLAttribute(const Decl *D, const llvm::GlobalValue *GV) const;
 

clang/test/CodeGen/PowerPC/musttail-forward-declaration-inline.c

@@ -0,0 +1,12 @@
+// RUN: %clang_cc1 %s -triple powerpc64le-unknown-linux-gnu -o /dev/null -emit-llvm -verify
+// RUN: %clang_cc1 %s -triple powerpc64-unknown-linux-gnu -o /dev/null -emit-llvm -verify
+
+inline int func2(int i);
+int external_call2(int i) {
+  // expected-error@+1 {{'musttail' attribute for this call is impossible because external calls can not be tail called on PPC}}
+  [[clang::musttail]] return func2(i);
+}
+
+inline int func2(int i) {
+  return 0;
+}

clang/test/CodeGen/PowerPC/musttail-forward-declaration-weak.c

@@ -0,0 +1,12 @@
+// RUN: %clang_cc1 %s -triple powerpc64le-unknown-linux-gnu -o /dev/null -emit-llvm -verify
+// RUN: %clang_cc1 %s -triple powerpc64-unknown-linux-gnu -o /dev/null -emit-llvm -verify
+
+int func2(int i);
+int external_call2(int i) {
+  // expected-error@+1 {{'musttail' attribute for this call is impossible because external calls can not be tail called on PPC}}
+  [[clang::musttail]] return func2(i);
+}
+
+__attribute__((weak)) int func2(int i) {
+  return 0;
+}

clang/test/CodeGen/PowerPC/musttail-forward-declaration.c

@@ -0,0 +1,11 @@
+// RUN: %clang_cc1 %s -triple powerpc64le-unknown-linux-gnu -o /dev/null -emit-llvm -verify=good
+// RUN: %clang_cc1 %s -triple powerpc64-unknown-linux-gnu -o /dev/null -emit-llvm -verify=good
+
+int func2(int i);
+int external_call2(int i) {
+  // good-no-diagnostics
+  [[clang::musttail]] return func2(i);
+}
+int func2(int i) {
+  return 0;
+}

clang/test/CodeGen/PowerPC/musttail-indirect.cpp

@@ -0,0 +1,8 @@
+// RUN: %clang_cc1 %s -triple powerpc64-unknown-linux-gnu -o /dev/null -emit-llvm -verify
+// RUN: %clang_cc1 %s -triple powerpc-unknown-linux-gnu -o /dev/null -emit-llvm -verify
+
+void name(int *params) {
+  auto fn = (void (*)(int *))1;
+  // expected-error@+1 {{'musttail' attribute for this call is impossible because indirect calls can not be tail called on PPC}}
+  [[clang::musttail]] return fn(params);
+}