[sycl-post-link] Don't remove llvm.compiler.used for NVPTX

llvm/test/CodeGen/AMDGPU/remove-compiler-used.ll

@@ -0,0 +1,19 @@
+;; This test verifies llc on AMDGCN will delete the llvm.compiler.used symbol
+;; while keeping the symbol in the outputted ASM.
+
+; RUN: llc < %s -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 | FileCheck %s
+; RUN: llc < %s -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 | FileCheck %s
+; RUN: llc < %s -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a | FileCheck %s
+
+@keep_this = internal global i32 2, align 4
+@llvm.compiler.used = appending global [1 x ptr] [ptr @keep_this], section "llvm.metadata"
+
+; CHECK-NOT: llvm.metadata
+; CHECK-NOT: llvm{{.*}}used
+; CHECK-NOT: llvm{{.*}}compiler{{.*}}used
+
+; CHECK: .type keep_this,@object ;
+
+; CHECK-NOT: llvm.metadata
+; CHECK-NOT: llvm{{.*}}used
+; CHECK-NOT: llvm{{.*}}compiler{{.*}}used

llvm/test/CodeGen/NVPTX/remove-compiler-used.ll

@@ -0,0 +1,17 @@
+;; This test verifies llc on NVPTX will delete the llvm.compiler.used symbol
+;; while keeping the symbol in the outputted ASM.
+
+; RUN: llc < %s -march=nvptx64 | FileCheck %s
+
+@keep_this = internal global i32 2, align 4
+@llvm.compiler.used = appending global [1 x ptr] [ptr @keep_this], section "llvm.metadata"
+
+; CHECK-NOT: llvm.metadata
+; CHECK-NOT: llvm{{.*}}used
+; CHECK-NOT: llvm{{.*}}compiler{{.*}}used
+
+; CHECK: .global .align 4 .u32 keep_this
+
+; CHECK-NOT: llvm.metadata
+; CHECK-NOT: llvm{{.*}}used
+; CHECK-NOT: llvm{{.*}}compiler{{.*}}used

llvm/tools/sycl-post-link/sycl-post-link.cpp

@@ -784,11 +784,11 @@ processInputModule(std::unique_ptr<Module> M) {
   Modified |= removeSYCLKernelsConstRefArray(*M.get());
 
   // There may be device_global variables kept alive in "llvm.compiler.used"
-  // to keep the optimizer from wrongfully removing them. Since it has served
-  // its purpose, these device_global variables can be removed. If they are not
-  // used inside the device code after they have been removed from
-  // "llvm.compiler.used" they can be erased safely.
-  Modified |= removeDeviceGlobalFromCompilerUsed(*M.get());
+  // to keep the optimizer from wrongfully removing them. llvm.compiler.used
+  // symbols are usually removed at backend lowering, but this is handled here
+  // for SPIR-V since SYCL compilation uses llvm-spirv, not the SPIR-V backend.
+  if (auto Triple = M->getTargetTriple().find("spir") != std::string::npos)
+    Modified |= removeDeviceGlobalFromCompilerUsed(*M.get());
 
   // Instrument each image scope device globals if the module has been
   // instrumented by sanitizer pass.

sycl/test-e2e/DeviceGlobal/device_global_static.cpp

@@ -0,0 +1,31 @@
+// RUN: %{build} -o %t.out
+// RUN: %{run} %t.out
+//
+// The OpenCL GPU backends do not currently support device_global backend
+// calls.
+// UNSUPPORTED: opencl && gpu
+//
+// Tests static device_global access through device kernels.
+
+#include "common.hpp"
+
+static device_global<int[4], TestProperties> DeviceGlobalVar;
+
+int main() {
+  queue Q;
+
+  Q.single_task([=]() { DeviceGlobalVar.get()[0] = 42; });
+  // Make sure that the write happens before subsequent read
+  Q.wait();
+
+  int OutVal = 0;
+  {
+    buffer<int, 1> OutBuf(&OutVal, 1);
+    Q.submit([&](handler &CGH) {
+      auto OutAcc = OutBuf.get_access<access::mode::write>(CGH);
+      CGH.single_task([=]() { OutAcc[0] = DeviceGlobalVar.get()[0]; });
+    });
+  }
+  assert(OutVal == 42 && "Read value does not match.");
+  return 0;
+}

sycl/test/check_device_code/device_global_static.cpp

@@ -0,0 +1,36 @@
+// Tests that the llvm.compiler.used symbol, which is used to implement static
+// device globals, is removed at some point in compilation. For SPIR-V this
+// symbol is removed at sycl-post-link and for NVPTX/AMDGCN it is removed at
+// lowering.
+//
+// It also checks that the symbol can be found in an object file for a given
+// triple, thus validating that `llvm-strings` can successfully be used to
+// check for the presence of the symbol.
+
+// UNSUPPORTED: windows
+
+// RUN: %clangxx -fsycl -fsycl-device-only %s -o %t
+// RUN: llvm-strings %t | grep "llvm.compiler.used"
+// RUN: %clangxx -fsycl %s -o %t
+// RUN: llvm-strings %t | not grep "llvm.compiler.used"
+
+// RUN: %if cuda %{ %clangxx -fsycl -fsycl-device-only -fsycl-targets=nvptx64-nvidia-cuda %s -o %t %}
+// RUN: %if cuda %{ llvm-strings %t | grep "llvm.compiler.used" %}
+// RUN: %if cuda %{ %clangxx -fsycl -fsycl-targets=nvptx64-nvidia-cuda %s -o %t %}
+// RUN: %if cuda %{ llvm-strings %t | not grep "llvm.compiler.used" %}
+
+// RUN: %if hip_amd %{ %clangxx -fsycl -fsycl-device-only -fsycl-targets=amd_gpu_gfx906 %s -o %t %}
+// RUN: %if hip_amd %{ llvm-strings %t | grep "llvm.compiler.used" %}
+// RUN: %if hip_amd %{ %clangxx -fsycl -fsycl-targets=amd_gpu_gfx906 %s -o %t %}
+// RUN: %if hip_amd %{ llvm-strings %t | not grep "llvm.compiler.used" %}
+
+#include <sycl/sycl.hpp>
+
+using namespace sycl;
+using namespace sycl::ext::oneapi::experimental;
+
+static device_global<int> DeviceGlobalVar;
+
+int main() {
+  sycl::queue{}.single_task([=] { volatile int ReadVal = DeviceGlobalVar; });
+}