From: Justin Lebar <jlebar@google.com>
Date: Thu, 30 Jun 2016 18:41:33 +0000 (+0000)
Subject: [CUDA] Give templated device functions internal linkage, templated kernels external... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=56a9798df3b2217e296a63f3798678eb681d66c7;p=clang

[CUDA] Give templated device functions internal linkage, templated kernels external linkage.

Summary:
This lets LLVM perform IPO over these functions.  In particular, it
allows LLVM to emit ld.global.nc for loads to __restrict pointers in
kernels that are never written to.

Reviewers: rsmith

Subscribers: cfe-commits, tra

Differential Revision: http://reviews.llvm.org/D21337

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@274261 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/CodeGen/CodeGenModule.cpp b/lib/CodeGen/CodeGenModule.cpp
index bbec03842c..bf2e66da69 100644
--- a/lib/CodeGen/CodeGenModule.cpp
+++ b/lib/CodeGen/CodeGenModule.cpp
@@ -2671,9 +2671,18 @@ llvm::GlobalValue::LinkageTypes CodeGenModule::getLLVMLinkageForDeclarator(
   // explicit instantiations can occur in multiple translation units
   // and must all be equivalent. However, we are not allowed to
   // throw away these explicit instantiations.
-  if (Linkage == GVA_StrongODR)
-    return !Context.getLangOpts().AppleKext ? llvm::Function::WeakODRLinkage
-                                            : llvm::Function::ExternalLinkage;
+  //
+  // We don't currently support CUDA device code spread out across multiple TUs,
+  // so say that CUDA templates are either external (for kernels) or internal.
+  // This lets llvm perform aggressive inter-procedural optimizations.
+  if (Linkage == GVA_StrongODR) {
+    if (Context.getLangOpts().AppleKext)
+      return llvm::Function::ExternalLinkage;
+    if (Context.getLangOpts().CUDA && Context.getLangOpts().CUDAIsDevice)
+      return D->hasAttr<CUDAGlobalAttr>() ? llvm::Function::ExternalLinkage
+                                          : llvm::Function::InternalLinkage;
+    return llvm::Function::WeakODRLinkage;
+  }
 
   // C++ doesn't have tentative definitions and thus cannot have common
   // linkage.
diff --git a/test/CodeGenCUDA/ptx-kernels.cu b/test/CodeGenCUDA/ptx-kernels.cu
index 034cef6b60..1d330bdf6a 100644
--- a/test/CodeGenCUDA/ptx-kernels.cu
+++ b/test/CodeGenCUDA/ptx-kernels.cu
@@ -19,11 +19,11 @@ __global__ void global_function() {
 
 // Make sure host-instantiated kernels are preserved on device side.
 template <typename T> __global__ void templated_kernel(T param) {}
-// CHECK-DAG: define weak_odr void @_Z16templated_kernelIiEvT_(
+// CHECK-DAG: define void @_Z16templated_kernelIiEvT_(
 
 namespace {
 __global__ void anonymous_ns_kernel() {}
-// CHECK-DAG: define weak_odr void @_ZN12_GLOBAL__N_119anonymous_ns_kernelEv(
+// CHECK-DAG: define void @_ZN12_GLOBAL__N_119anonymous_ns_kernelEv(
 }
 
 void host_function() {