[CUDA] Add appropriate host/device attribute to builtins.

The changes are part of attribute-based CUDA function overloading (D12453)
and as such are only enabled when it's in effect (-fcuda-target-overloads).

Differential Revision: http://reviews.llvm.org/D12122

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@248296 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/Basic/Builtins.h b/include/clang/Basic/Builtins.h
index 554143d6be7d7d0f728e759d34ec935a1920967f..87e2ac70790bd6e6efd8d564dd91656f7394ed8c 100644 (file)
--- a/include/clang/Basic/Builtins.h
+++ b/include/clang/Basic/Builtins.h
@@ -81,6 +81,11 @@ public:
     return getRecord(ID).Type;
   }
 
+  /// \brief Return true if this function is a target-specific builtin
+  bool isTSBuiltin(unsigned ID) const {
+    return ID >= Builtin::FirstTSBuiltin;
+  }
+
   /// \brief Return true if this function has no side effects and doesn't
   /// read memory.
   bool isConst(unsigned ID) const {

diff --git a/lib/Sema/SemaChecking.cpp b/lib/Sema/SemaChecking.cpp
index f8e6c9d5ee4a943199a3bc4b42be719ef48f269f..ae7f6d6a803bb5ecafcf45862f8b086596256e00 100644 (file)
--- a/lib/Sema/SemaChecking.cpp
+++ b/lib/Sema/SemaChecking.cpp
@@ -529,7 +529,7 @@ Sema::CheckBuiltinFunctionCall(FunctionDecl *FDecl, unsigned BuiltinID,
 
   // Since the target specific builtins for each arch overlap, only check those
   // of the arch we are compiling for.
-  if (BuiltinID >= Builtin::FirstTSBuiltin) {
+  if (Context.BuiltinInfo.isTSBuiltin(BuiltinID)) {
     switch (Context.getTargetInfo().getTriple().getArch()) {
       case llvm::Triple::arm:
       case llvm::Triple::armeb:

diff --git a/lib/Sema/SemaDecl.cpp b/lib/Sema/SemaDecl.cpp
index cdf7d7981bd9294d9df9147d575cd03911cc3084..22d2d3469f65d7221d190741a5903bec1fa60ffb 100644 (file)
--- a/lib/Sema/SemaDecl.cpp
+++ b/lib/Sema/SemaDecl.cpp
@@ -11290,6 +11290,18 @@ void Sema::AddKnownFunctionAttributes(FunctionDecl *FD) {
       FD->addAttr(NoThrowAttr::CreateImplicit(Context, FD->getLocation()));
     if (Context.BuiltinInfo.isConst(BuiltinID) && !FD->hasAttr<ConstAttr>())
       FD->addAttr(ConstAttr::CreateImplicit(Context, FD->getLocation()));
+    if (getLangOpts().CUDA && getLangOpts().CUDATargetOverloads &&
+        Context.BuiltinInfo.isTSBuiltin(BuiltinID) &&
+        !FD->hasAttr<CUDADeviceAttr>() && !FD->hasAttr<CUDAHostAttr>()) {
+      // Target-specific builtins are assumed to be intended for use
+      // in this particular CUDA compilation mode and should have
+      // appropriate attribute set so we can enforce CUDA function
+      // call restrictions.
+      if (getLangOpts().CUDAIsDevice)
+        FD->addAttr(CUDADeviceAttr::CreateImplicit(Context, FD->getLocation()));
+      else
+        FD->addAttr(CUDAHostAttr::CreateImplicit(Context, FD->getLocation()));
+    }
   }
 
   IdentifierInfo *Name = FD->getIdentifier();

diff --git a/test/SemaCUDA/builtins.cu b/test/SemaCUDA/builtins.cu
new file mode 100644 (file)
index 0000000..2c619b5
--- /dev/null
+++ b/test/SemaCUDA/builtins.cu
@@ -0,0 +1,36 @@
+// Tests that target-specific builtins have appropriate host/device
+// attributes and that CUDA call restrictions are enforced. Also
+// verify that non-target builtins can be used from both host and
+// device functions.
+//
+// REQUIRES: x86-registered-target
+// REQUIRES: nvptx-registered-target
+// RUN: %clang_cc1 -triple x86_64-unknown-unknown \
+// RUN:     -fcuda-target-overloads -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-unknown-cuda -fcuda-is-device \
+// RUN:     -fcuda-target-overloads -fsyntax-only -verify %s
+
+
+#ifdef __CUDA_ARCH__
+// Device-side builtins are not allowed to be called from host functions.
+void hf() {
+  int x = __builtin_ptx_read_tid_x(); // expected-note  {{'__builtin_ptx_read_tid_x' declared here}}
+  // expected-error@-1 {{reference to __device__ function '__builtin_ptx_read_tid_x' in __host__ function}}
+  x = __builtin_abs(1);
+}
+__attribute__((device)) void df() {
+  int x = __builtin_ptx_read_tid_x();
+  x = __builtin_abs(1);
+}
+#else
+// Host-side builtins are not allowed to be called from device functions.
+__attribute__((device)) void df() {
+  int x = __builtin_ia32_rdtsc();   // expected-note {{'__builtin_ia32_rdtsc' declared here}}
+  // expected-error@-1 {{reference to __host__ function '__builtin_ia32_rdtsc' in __device__ function}}
+  x = __builtin_abs(1);
+}
+void hf() {
+  int x = __builtin_ia32_rdtsc();
+  x = __builtin_abs(1);
+}
+#endif

diff --git a/test/SemaCUDA/implicit-intrinsic.cu b/test/SemaCUDA/implicit-intrinsic.cu
index 3d24aa719e57d902624925291e807e7ed686cd6e..0793d64b1017a3cc8f5da4cd1a029aad7db227de 100644 (file)
--- a/test/SemaCUDA/implicit-intrinsic.cu
+++ b/test/SemaCUDA/implicit-intrinsic.cu
@@ -1,10 +1,13 @@
-// RUN: %clang_cc1 -std=gnu++11 -triple nvptx64-unknown-unknown -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-unknown-unknown -fcuda-is-device \
+// RUN:     -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-unknown-unknown -fcuda-is-device \
+// RUN:     -fcuda-target-overloads -fsyntax-only -verify %s
 
 #include "Inputs/cuda.h"
 
 // expected-no-diagnostics
 __device__ void __threadfence_system() {
-  // This shouldn't produce an error, since __nvvm_membar_sys is inferred to
-  // be __host__ __device__ and thus callable from device code.
+  // This shouldn't produce an error, since __nvvm_membar_sys should be
+  // __device__ and thus callable from device code.
   __nvvm_membar_sys();
 }