[SYCL] Fix group local memory sharing issue (#3489)

Group local memory functions depend on inlining pass. When all
optimizations are disabled then those functions are not inlined. If
a user adds several group local memory calls with the same template
argument then only one group local memory function is specialized and
then one memory buffer is allocated. Thus all variables that save
a result of the group local memory call point to the same buffer.
Enable AlwaysInliner pass even when all llvm passes are disabled as
a workaround.

Author: mlychkov

clang/lib/CodeGen/BackendUtil.cpp

@@ -998,8 +998,14 @@ void EmitAssemblyHelper::EmitAssembly(BackendAction Action,
   }
 
   // Allocate static local memory in SYCL kernel scope for each allocation call.
-  if (LangOpts.SYCLIsDevice)
+  if (LangOpts.SYCLIsDevice) {
+    // Group local memory pass depends on inlining. Turn it on even in case if
+    // all llvm passes or SYCL early optimizations are disabled.
+    // FIXME: Remove this workaround when dependency on inlining is eliminated.
+    if (CodeGenOpts.DisableLLVMPasses)
+      PerModulePasses.add(createAlwaysInlinerLegacyPass(false));
     PerModulePasses.add(createSYCLLowerWGLocalMemoryLegacyPass());
+  }
 
   switch (Action) {
   case Backend_EmitNothing:

clang/test/CodeGenSYCL/group-local-memory.cpp

@@ -1,8 +1,26 @@
-// RUN: %clang_cc1 -fsycl-is-device -triple spir64-unknown-unknown-sycldevice \
-// RUN:     -S -emit-llvm -mllvm -debug-pass=Structure -disable-llvm-passes \
-// RUN:     -o - %s 2>&1 | FileCheck %s
-// RUN: %clang_cc1 -fsycl-is-device -triple spir64-unknown-unknown-sycldevice \
-// RUN:     -S -emit-llvm -mllvm -debug-pass=Structure -o - %s 2>&1 \
-// RUN:     | FileCheck %s
+// Check that SYCLLowerWGLocalMemory pass is added to the SYCL device
+// compilation pipeline with the inliner pass.
 
+// RUN: %clang_cc1 -fsycl-is-device -triple spir64-unknown-unknown-sycldevice -emit-llvm \
+// RUN:   -mllvm -debug-pass=Structure %s -o - 2>&1 \
+// RUN:   | FileCheck %s
+// CHECK: Function Integration/Inlining
 // CHECK: Replace __sycl_allocateLocalMemory with allocation of memory in local address space
+
+// Check that AlwaysInliner pass is always run for compilation of SYCL device
+// target code, even if all optimizations are disabled.
+
+// RUN: %clang_cc1 -fsycl-is-device -triple spir64-unknown-unknown-sycldevice -emit-llvm \
+// RUN:   -mllvm -debug-pass=Structure %s -o - -disable-llvm-passes 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=CHECK-NOPASSES
+// RUN: %clang_cc1 -fsycl-is-device -triple spir64-unknown-unknown-sycldevice -emit-llvm \
+// RUN:   -mllvm -debug-pass=Structure %s -o - -fno-sycl-early-optimizations 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=CHECK-NOPASSES
+// CHECK-NOPASSES: Inliner for always_inline functions
+// CHECK-NOPASSES: Replace __sycl_allocateLocalMemory with allocation of memory in local address space
+
+// RUN: %clang_cc1 -fsycl-is-device -triple spir64-unknown-unknown-sycldevice -emit-llvm \
+// RUN:   -mllvm -debug-pass=Structure %s -o - -O0 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=CHECK-O0opt
+// CHECK-O0opt: Inliner for always_inline functions
+// CHECK-O0opt: Replace __sycl_allocateLocalMemory with allocation of memory in local address space