Add hook to limit intrinsic use

Author: MacDue

clang/lib/CodeGen/CGBuiltin.cpp

@@ -2937,12 +2937,21 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
       &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad())
     BuiltinID = mutateLongDoubleBuiltin(BuiltinID);
 
-  // If the builtin has been declared explicitly with an assembler label,
-  // disable the specialized emitting below. Ideally we should communicate the
-  // rename in IR, or at least avoid generating the intrinsic calls that are
-  // likely to get lowered to the renamed library functions.
-  const unsigned BuiltinIDIfNoAsmLabel =
-      FD->hasAttr<AsmLabelAttr>() ? 0 : BuiltinID;
+  const unsigned BuiltinIDIfEmitIntrinsics = [&] {
+    // If the builtin has been declared explicitly with an assembler label,
+    // disable the specialized emitting below. Ideally we should communicate the
+    // rename in IR, or at least avoid generating the intrinsic calls that are
+    // likely to get lowered to the renamed library functions.
+    if (FD->hasAttr<AsmLabelAttr>())
+      return 0U;
+
+    // If the target requests not to use intrinsics for a builtin, disable the
+    // specialized emitting below.
+    if (!getTargetHooks().shouldUseIntrinsicsForBuiltin(BuiltinID))
+      return 0U;
+
+    return BuiltinID;
+  }();
 
   std::optional<bool> ErrnoOverriden;
   // ErrnoOverriden is true if math-errno is overriden via the
@@ -3035,7 +3044,7 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
           ConstWithoutErrnoOrExceptions && ErrnoOverridenToFalseWithOpt;
   }
   if (GenerateIntrinsics) {
-    switch (BuiltinIDIfNoAsmLabel) {
+    switch (BuiltinIDIfEmitIntrinsics) {
     case Builtin::BIacos:
     case Builtin::BIacosf:
     case Builtin::BIacosl:
@@ -3515,7 +3524,7 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
     }
   };
 
-  switch (BuiltinIDIfNoAsmLabel) {
+  switch (BuiltinIDIfEmitIntrinsics) {
   default: break;
   case Builtin::BI__builtin___CFStringMakeConstantString:
   case Builtin::BI__builtin___NSStringMakeConstantString:
@@ -3645,7 +3654,7 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
   case Builtin::BI__builtin_ctzl:
   case Builtin::BI__builtin_ctzll:
   case Builtin::BI__builtin_ctzg: {
-    bool HasFallback = BuiltinIDIfNoAsmLabel == Builtin::BI__builtin_ctzg &&
+    bool HasFallback = BuiltinIDIfEmitIntrinsics == Builtin::BI__builtin_ctzg &&
                        E->getNumArgs() > 1;
 
     Value *ArgValue =
@@ -3677,7 +3686,7 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
   case Builtin::BI__builtin_clzl:
   case Builtin::BI__builtin_clzll:
   case Builtin::BI__builtin_clzg: {
-    bool HasFallback = BuiltinIDIfNoAsmLabel == Builtin::BI__builtin_clzg &&
+    bool HasFallback = BuiltinIDIfEmitIntrinsics == Builtin::BI__builtin_clzg &&
                        E->getNumArgs() > 1;
 
     Value *ArgValue =
@@ -4347,7 +4356,7 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
       Ty = VecTy->getElementType();
     bool IsSigned = Ty->isSignedIntegerType();
     unsigned Opc;
-    if (BuiltinIDIfNoAsmLabel == Builtin::BI__builtin_elementwise_add_sat)
+    if (BuiltinIDIfEmitIntrinsics == Builtin::BI__builtin_elementwise_add_sat)
       Opc = IsSigned ? llvm::Intrinsic::sadd_sat : llvm::Intrinsic::uadd_sat;
     else
       Opc = IsSigned ? llvm::Intrinsic::ssub_sat : llvm::Intrinsic::usub_sat;

clang/lib/CodeGen/TargetInfo.h

@@ -454,6 +454,23 @@ class TargetCodeGenInfo {
   initBranchProtectionFnAttributes(const TargetInfo::BranchProtectionInfo &BPI,
                                    llvm::AttrBuilder &FuncAttrs);
 
+  /// Returns true if an intrinsic should be emitted for a specific builtin. By
+  /// default, this disables emitting intrinsics for (non-prefixed) sincos
+  /// builtins, as on targets that do not set llvm::TargetSubtargetInfo::useAA()
+  /// the intrinsics can worsen codegen.
+  /// TODO: Remove once the intrinsic lowering works well for all targets.
+  virtual bool shouldUseIntrinsicsForBuiltin(unsigned BuiltinID) const {
+    switch (BuiltinID) {
+    case Builtin::BIsincos:
+    case Builtin::BIsincosf:
+    case Builtin::BIsincosl:
+      return false;
+    default:
+      break;
+    }
+    return true;
+  }
+
 protected:
   static std::string qualifyWindowsLibrary(StringRef Lib);
 

clang/lib/CodeGen/Targets/AArch64.cpp

@@ -181,6 +181,10 @@ class AArch64TargetCodeGenInfo : public TargetCodeGenInfo {
   bool wouldInliningViolateFunctionCallABI(
       const FunctionDecl *Caller, const FunctionDecl *Callee) const override;
 
+  bool shouldUseIntrinsicsForBuiltin(unsigned BuiltinID) const override {
+    return true;
+  }
+
 private:
   // Diagnose calls between functions with incompatible Streaming SVE
   // attributes.

clang/test/CodeGen/math-libcalls.c

@@ -660,6 +660,17 @@ void foo(double *d, float f, float *fp, long double *l, int *i, const char *c) {
 // HAS_MAYTRAP: declare float @llvm.experimental.constrained.sinh.f32(
 // HAS_MAYTRAP: declare x86_fp80 @llvm.experimental.constrained.sinh.f80(
 
+sincos(f, d, d);       sincosf(f, fp, fp);        sincosl(f, l, l);
+
+// NO__ERRNO: declare void @sincos(double noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// NO__ERRNO: declare void @sincosf(float noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// NO__ERRNO: declare void @sincosl(x86_fp80 noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// HAS__ERRNO: declare void @sincos(double noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// HAS__ERRNO: declare void @sincosf(float noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// HAS__ERRNO: declare void @sincosl(x86_fp80 noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// HAS_MAYTRAP: declare void @sincos(double noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// HAS_MAYTRAP: declare void @sincosf(float noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
+// HAS_MAYTRAP: declare void @sincosl(x86_fp80 noundef, ptr noundef, ptr noundef) [[NOT_READNONE]]
 
   sqrt(f);       sqrtf(f);      sqrtl(f);