[OpenCL] Add global_device and global_host address spaces

This patch introduces 2 new address spaces in OpenCL: global_device and global_host which are a subset of a global address space, so the address space scheme will be looking like: ``` generic->global->host ->device ->private ->local constant ``` Justification: USM allocations may be associated with both host and device memory. We want to give users a way to tell the compiler the allocation type of a USM pointer for optimization purposes. (Link to the Unified Shared Memory extension: https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/USM/cl_intel_unified_shared_memory.asciidoc) Before this patch USM pointer could be only in opencl_global address space, hence a device backend can't tell if a particular pointer points to host or device memory. On FPGAs at least we can generate more efficient hardware code if the user tells us where the pointer can point - being able to distinguish between these types of pointers at compile time allows us to instantiate simpler load-store units to perform memory transactions. Patch by Dmitry Sidorov. Reviewed By: Anastasia Differential Revision: https://reviews.llvm.org/D82174
llvm · Jul 29, 2020 · 8d27be8 · 8d27be8
1 parent 2c662f3
commit 8d27be8
Show file tree

Hide file tree

Showing 22 changed files with 262 additions and 29 deletions.
diff --git a/clang/include/clang/AST/Type.h b/clang/include/clang/AST/Type.h
@@ -480,6 +480,11 @@ class Qualifiers {
            // Otherwise in OpenCLC v2.0 s6.5.5: every address space except
            // for __constant can be used as __generic.
            (A == LangAS::opencl_generic && B != LangAS::opencl_constant) ||
+           // We also define global_device and global_host address spaces,
+           // to distinguish global pointers allocated on host from pointers
+           // allocated on device, which are a subset of __global.
+           (A == LangAS::opencl_global && (B == LangAS::opencl_global_device ||
+                                           B == LangAS::opencl_global_host)) ||
            // Consider pointer size address spaces to be equivalent to default.
            ((isPtrSizeAddressSpace(A) || A == LangAS::Default) &&
             (isPtrSizeAddressSpace(B) || B == LangAS::Default));

diff --git a/clang/include/clang/Basic/AddressSpaces.h b/clang/include/clang/Basic/AddressSpaces.h
@@ -36,6 +36,8 @@ enum class LangAS : unsigned {
   opencl_constant,
   opencl_private,
   opencl_generic,
+  opencl_global_device,
+  opencl_global_host,
 
   // CUDA specific address spaces.
   cuda_device,

diff --git a/clang/include/clang/Basic/Attr.td b/clang/include/clang/Basic/Attr.td
@@ -1178,6 +1178,16 @@ def OpenCLGlobalAddressSpace : TypeAttr {
   let Documentation = [OpenCLAddressSpaceGlobalDocs];
 }
 
+def OpenCLGlobalDeviceAddressSpace : TypeAttr {
+  let Spellings = [Clang<"opencl_global_device">];
+  let Documentation = [OpenCLAddressSpaceGlobalExtDocs];
+}
+
+def OpenCLGlobalHostAddressSpace : TypeAttr {
+  let Spellings = [Clang<"opencl_global_host">];
+  let Documentation = [OpenCLAddressSpaceGlobalExtDocs];
+}
+
 def OpenCLLocalAddressSpace : TypeAttr {
   let Spellings = [Keyword<"__local">, Keyword<"local">, Clang<"opencl_local">];
   let Documentation = [OpenCLAddressSpaceLocalDocs];

diff --git a/clang/include/clang/Basic/AttrDocs.td b/clang/include/clang/Basic/AttrDocs.td
@@ -3123,6 +3123,30 @@ scope) variables and static local variable as well.
   }];
 }
 
+def OpenCLAddressSpaceGlobalExtDocs : Documentation {
+  let Category = DocOpenCLAddressSpaces;
+  let Heading = "[[clang::opencl_global_device]], [[clang::opencl_global_host]]";
+  let Content = [{
+The ``global_device`` and ``global_host`` address space attributes specify that
+an object is allocated in global memory on the device/host. It helps to
+distinguish USM (Unified Shared Memory) pointers that access global device
+memory from those that access global host memory. These new address spaces are
+a subset of the ``__global/opencl_global`` address space, the full address space
+set model for OpenCL 2.0 with the extension looks as follows:
+  generic->global->host
+                 ->device
+         ->private
+         ->local
+  constant
+
+As ``global_device`` and ``global_host`` are a subset of
+``__global/opencl_global`` address spaces it is allowed to convert
+``global_device`` and ``global_host`` address spaces to
+``__global/opencl_global`` address spaces (following ISO/IEC TR 18037 5.1.3
+"Address space nesting and rules for pointers).
+  }];
+}
+
 def OpenCLAddressSpaceLocalDocs : Documentation {
   let Category = DocOpenCLAddressSpaces;
   let Heading = "__local, local, [[clang::opencl_local]]";

diff --git a/clang/include/clang/Sema/ParsedAttr.h b/clang/include/clang/Sema/ParsedAttr.h
@@ -606,6 +606,10 @@ class ParsedAttr final
       return LangAS::opencl_constant;
     case ParsedAttr::AT_OpenCLGlobalAddressSpace:
       return LangAS::opencl_global;
+    case ParsedAttr::AT_OpenCLGlobalDeviceAddressSpace:
+      return LangAS::opencl_global_device;
+    case ParsedAttr::AT_OpenCLGlobalHostAddressSpace:
+      return LangAS::opencl_global_host;
     case ParsedAttr::AT_OpenCLLocalAddressSpace:
       return LangAS::opencl_local;
     case ParsedAttr::AT_OpenCLPrivateAddressSpace:

diff --git a/clang/lib/AST/ASTContext.cpp b/clang/lib/AST/ASTContext.cpp
@@ -919,18 +919,20 @@ static const LangASMap *getAddressSpaceMap(const TargetInfo &T,
     // The fake address space map must have a distinct entry for each
     // language-specific address space.
     static const unsigned FakeAddrSpaceMap[] = {
-        0, // Default
-        1, // opencl_global
-        3, // opencl_local
-        2, // opencl_constant
-        0, // opencl_private
-        4, // opencl_generic
-        5, // cuda_device
-        6, // cuda_constant
-        7, // cuda_shared
-        8, // ptr32_sptr
-        9, // ptr32_uptr
-        10 // ptr64
+        0,  // Default
+        1,  // opencl_global
+        3,  // opencl_local
+        2,  // opencl_constant
+        0,  // opencl_private
+        4,  // opencl_generic
+        5,  // opencl_global_device
+        6,  // opencl_global_host
+        7,  // cuda_device
+        8,  // cuda_constant
+        9,  // cuda_shared
+        10, // ptr32_sptr
+        11, // ptr32_uptr
+        12  // ptr64
     };
     return &FakeAddrSpaceMap;
   } else {

diff --git a/clang/lib/AST/ItaniumMangle.cpp b/clang/lib/AST/ItaniumMangle.cpp
@@ -2388,16 +2388,39 @@ void CXXNameMangler::mangleQualifiers(Qualifiers Quals, const DependentAddressSp
       switch (AS) {
       default: llvm_unreachable("Not a language specific address space");
       //  <OpenCL-addrspace> ::= "CL" [ "global" | "local" | "constant" |
-      //                                "private"| "generic" ]
-      case LangAS::opencl_global:   ASString = "CLglobal";   break;
-      case LangAS::opencl_local:    ASString = "CLlocal";    break;
-      case LangAS::opencl_constant: ASString = "CLconstant"; break;
-      case LangAS::opencl_private:  ASString = "CLprivate";  break;
-      case LangAS::opencl_generic:  ASString = "CLgeneric";  break;
+      //                                "private"| "generic" | "device" |
+      //                                "host" ]
+      case LangAS::opencl_global:
+        ASString = "CLglobal";
+        break;
+      case LangAS::opencl_global_device:
+        ASString = "CLdevice";
+        break;
+      case LangAS::opencl_global_host:
+        ASString = "CLhost";
+        break;
+      case LangAS::opencl_local:
+        ASString = "CLlocal";
+        break;
+      case LangAS::opencl_constant:
+        ASString = "CLconstant";
+        break;
+      case LangAS::opencl_private:
+        ASString = "CLprivate";
+        break;
+      case LangAS::opencl_generic:
+        ASString = "CLgeneric";
+        break;
       //  <CUDA-addrspace> ::= "CU" [ "device" | "constant" | "shared" ]
-      case LangAS::cuda_device:     ASString = "CUdevice";   break;
-      case LangAS::cuda_constant:   ASString = "CUconstant"; break;
-      case LangAS::cuda_shared:     ASString = "CUshared";   break;
+      case LangAS::cuda_device:
+        ASString = "CUdevice";
+        break;
+      case LangAS::cuda_constant:
+        ASString = "CUconstant";
+        break;
+      case LangAS::cuda_shared:
+        ASString = "CUshared";
+        break;
       //  <ptrsize-addrspace> ::= [ "ptr32_sptr" | "ptr32_uptr" | "ptr64" ]
       case LangAS::ptr32_sptr:
         ASString = "ptr32_sptr";

diff --git a/clang/lib/AST/MicrosoftMangle.cpp b/clang/lib/AST/MicrosoftMangle.cpp
@@ -1798,7 +1798,7 @@ void MicrosoftCXXNameMangler::mangleAddressSpaceType(QualType T,
   // where:
   //  <language_addr_space> ::= <OpenCL-addrspace> | <CUDA-addrspace>
   //    <OpenCL-addrspace> ::= "CL" [ "global" | "local" | "constant" |
-  //                                "private"| "generic" ]
+  //                                "private"| "generic" | "device" | "host" ]
   //    <CUDA-addrspace> ::= "CU" [ "device" | "constant" | "shared" ]
   //    Note that the above were chosen to match the Itanium mangling for this.
   //
@@ -1823,6 +1823,12 @@ void MicrosoftCXXNameMangler::mangleAddressSpaceType(QualType T,
     case LangAS::opencl_global:
       Extra.mangleSourceName("_ASCLglobal");
       break;
+    case LangAS::opencl_global_device:
+      Extra.mangleSourceName("_ASCLdevice");
+      break;
+    case LangAS::opencl_global_host:
+      Extra.mangleSourceName("_ASCLhost");
+      break;
     case LangAS::opencl_local:
       Extra.mangleSourceName("_ASCLlocal");
       break;

diff --git a/clang/lib/AST/TypePrinter.cpp b/clang/lib/AST/TypePrinter.cpp
@@ -1564,6 +1564,8 @@ void TypePrinter::printAttributedAfter(const AttributedType *T,
 
   case attr::OpenCLPrivateAddressSpace:
   case attr::OpenCLGlobalAddressSpace:
+  case attr::OpenCLGlobalDeviceAddressSpace:
+  case attr::OpenCLGlobalHostAddressSpace:
   case attr::OpenCLLocalAddressSpace:
   case attr::OpenCLConstantAddressSpace:
   case attr::OpenCLGenericAddressSpace:
@@ -1866,6 +1868,10 @@ std::string Qualifiers::getAddrSpaceAsString(LangAS AS) {
     return "__constant";
   case LangAS::opencl_generic:
     return "__generic";
+  case LangAS::opencl_global_device:
+    return "__global_device";
+  case LangAS::opencl_global_host:
+    return "__global_host";
   case LangAS::cuda_device:
     return "__device__";
   case LangAS::cuda_constant:

diff --git a/clang/lib/Basic/Targets/AMDGPU.cpp b/clang/lib/Basic/Targets/AMDGPU.cpp
@@ -46,6 +46,8 @@ const LangASMap AMDGPUTargetInfo::AMDGPUDefIsGenMap = {
     Constant, // opencl_constant
     Private,  // opencl_private
     Generic,  // opencl_generic
+    Global,   // opencl_global_device
+    Global,   // opencl_global_host
     Global,   // cuda_device
     Constant, // cuda_constant
     Local,    // cuda_shared
@@ -61,6 +63,8 @@ const LangASMap AMDGPUTargetInfo::AMDGPUDefIsPrivMap = {
     Constant, // opencl_constant
     Private,  // opencl_private
     Generic,  // opencl_generic
+    Global,   // opencl_global_device
+    Global,   // opencl_global_host
     Global,   // cuda_device
     Constant, // cuda_constant
     Local,    // cuda_shared

diff --git a/clang/lib/Basic/Targets/NVPTX.h b/clang/lib/Basic/Targets/NVPTX.h
@@ -30,6 +30,8 @@ static const unsigned NVPTXAddrSpaceMap[] = {
     0, // opencl_private
     // FIXME: generic has to be added to the target
     0, // opencl_generic
+    1, // opencl_global_device
+    1, // opencl_global_host
     1, // cuda_device
     4, // cuda_constant
     3, // cuda_shared

diff --git a/clang/lib/Basic/Targets/SPIR.h b/clang/lib/Basic/Targets/SPIR.h
@@ -28,6 +28,8 @@ static const unsigned SPIRAddrSpaceMap[] = {
     2, // opencl_constant
     0, // opencl_private
     4, // opencl_generic
+    5, // opencl_global_device
+    6, // opencl_global_host
     0, // cuda_device
     0, // cuda_constant
     0, // cuda_shared

diff --git a/clang/lib/Basic/Targets/TCE.h b/clang/lib/Basic/Targets/TCE.h
@@ -35,6 +35,8 @@ static const unsigned TCEOpenCLAddrSpaceMap[] = {
     4, // opencl_local
     5, // opencl_constant
     0, // opencl_private
+    1, // opencl_global_device
+    1, // opencl_global_host
     // FIXME: generic has to be added to the target
     0, // opencl_generic
     0, // cuda_device

diff --git a/clang/lib/Basic/Targets/X86.h b/clang/lib/Basic/Targets/X86.h
@@ -30,6 +30,8 @@ static const unsigned X86AddrSpaceMap[] = {
     0,   // opencl_constant
     0,   // opencl_private
     0,   // opencl_generic
+    0,   // opencl_global_device
+    0,   // opencl_global_host
     0,   // cuda_device
     0,   // cuda_constant
     0,   // cuda_shared

diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -1324,10 +1324,18 @@ static void removeImageAccessQualifier(std::string& TyName) {
 // (basically all single AS CPUs).
 static unsigned ArgInfoAddressSpace(LangAS AS) {
   switch (AS) {
-  case LangAS::opencl_global:   return 1;
-  case LangAS::opencl_constant: return 2;
-  case LangAS::opencl_local:    return 3;
-  case LangAS::opencl_generic:  return 4; // Not in SPIR 2.0 specs.
+  case LangAS::opencl_global:
+    return 1;
+  case LangAS::opencl_constant:
+    return 2;
+  case LangAS::opencl_local:
+    return 3;
+  case LangAS::opencl_generic:
+    return 4; // Not in SPIR 2.0 specs.
+  case LangAS::opencl_global_device:
+    return 5;
+  case LangAS::opencl_global_host:
+    return 6;
   default:
     return 0; // Assume private.
   }
@@ -3792,6 +3800,8 @@ LangAS CodeGenModule::GetGlobalVarAddressSpace(const VarDecl *D) {
   if (LangOpts.OpenCL) {
     AddrSpace = D ? D->getType().getAddressSpace() : LangAS::opencl_global;
     assert(AddrSpace == LangAS::opencl_global ||
+           AddrSpace == LangAS::opencl_global_device ||
+           AddrSpace == LangAS::opencl_global_host ||
            AddrSpace == LangAS::opencl_constant ||
            AddrSpace == LangAS::opencl_local ||
            AddrSpace >= LangAS::FirstTargetAddressSpace);

diff --git a/clang/lib/Sema/SemaType.cpp b/clang/lib/Sema/SemaType.cpp
@@ -7968,6 +7968,8 @@ static bool isAddressSpaceKind(const ParsedAttr &attr) {
   return attrKind == ParsedAttr::AT_AddressSpace ||
          attrKind == ParsedAttr::AT_OpenCLPrivateAddressSpace ||
          attrKind == ParsedAttr::AT_OpenCLGlobalAddressSpace ||
+         attrKind == ParsedAttr::AT_OpenCLGlobalDeviceAddressSpace ||
+         attrKind == ParsedAttr::AT_OpenCLGlobalHostAddressSpace ||
          attrKind == ParsedAttr::AT_OpenCLLocalAddressSpace ||
          attrKind == ParsedAttr::AT_OpenCLConstantAddressSpace ||
          attrKind == ParsedAttr::AT_OpenCLGenericAddressSpace;
@@ -8048,6 +8050,8 @@ static void processTypeAttrs(TypeProcessingState &state, QualType &type,
       break;
     case ParsedAttr::AT_OpenCLPrivateAddressSpace:
     case ParsedAttr::AT_OpenCLGlobalAddressSpace:
+    case ParsedAttr::AT_OpenCLGlobalDeviceAddressSpace:
+    case ParsedAttr::AT_OpenCLGlobalHostAddressSpace:
     case ParsedAttr::AT_OpenCLLocalAddressSpace:
     case ParsedAttr::AT_OpenCLConstantAddressSpace:
     case ParsedAttr::AT_OpenCLGenericAddressSpace:

diff --git a/clang/test/AST/language_address_space_attribute.cpp b/clang/test/AST/language_address_space_attribute.cpp
@@ -17,6 +17,18 @@ void langas() {
   // CHECK: VarDecl {{.*}} z_global '__global int *'
   [[clang::opencl_global]] int *z_global;
 
+  // CHECK: VarDecl {{.*}} x_global_device '__global_device int *'
+  __attribute__((opencl_global_device)) int *x_global_device;
+
+  // CHECK: VarDecl {{.*}} z_global_device '__global_device int *'
+  [[clang::opencl_global_device]] int *z_global_device;
+
+  // CHECK: VarDecl {{.*}} x_global_host '__global_host int *'
+  __attribute__((opencl_global_host)) int *x_global_host;
+
+  // CHECK: VarDecl {{.*}} z_global_host '__global_host int *'
+  [[clang::opencl_global_host]] int *z_global_host;
+
   // CHECK: VarDecl {{.*}} x_local '__local int *'
   __attribute__((opencl_local)) int *x_local;
 

diff --git a/clang/test/CodeGenCXX/mangle-address-space.cpp b/clang/test/CodeGenCXX/mangle-address-space.cpp
@@ -43,6 +43,10 @@ void ocl_f0(char __private *p) { }
 
 struct ocl_OpaqueType;
 typedef ocl_OpaqueType __global * ocl_OpaqueTypePtr;
+typedef ocl_OpaqueType __attribute__((opencl_global_host)) * ocl_OpaqueTypePtrH;
+typedef ocl_OpaqueType
+    __attribute__((opencl_global_device)) *
+    ocl_OpaqueTypePtrD;
 
 // CHECKOCL-LABEL: define {{.*}}void @_Z6ocl_f0PU8CLglobal14ocl_OpaqueType
 // WINOCL-LABEL: define {{.*}}void @"?ocl_f0@@YAXPEAU?$_ASCLglobal@$$CAUocl_OpaqueType@@@__clang@@@Z"
@@ -61,4 +65,12 @@ __constant float *ocl_f1(char __generic const *p) { return 0;}
 // CHECKOCL-LABEL: define {{.*}}float* @_Z6ocl_f2PU9CLgenericKc
 // WINOCL-LABEL: define {{.*}}float* @"?ocl_f2@@YAPEAU?$_ASCLgeneric@$$CAM@__clang@@QEAU?$_ASCLgeneric@$$CBD@2@@Z"
 __generic float *ocl_f2(__generic char const * const p) { return 0;}
+
+// CHECKOCL-LABEL: define {{.*}}void @_Z6ocl_f3PU6CLhost14ocl_OpaqueType
+// WINOCL-LABEL: define {{.*}}void @"?ocl_f3@@YAXPEAU?$_ASCLhost@$$CAUocl_OpaqueType@@@__clang@@@Z"
+void ocl_f3(ocl_OpaqueTypePtrH) {}
+
+// CHECKOCL-LABEL: define {{.*}}void @_Z6ocl_f4PU8CLdevice14ocl_OpaqueType
+// WINOCL-LABEL: define {{.*}}void @"?ocl_f4@@YAXPEAU?$_ASCLdevice@$$CAUocl_OpaqueType@@@__clang@@@Z"
+void ocl_f4(ocl_OpaqueTypePtrD) {}
 #endif
diff --git a/clang/test/CodeGenOpenCL/address-spaces-conversions.cl b/clang/test/CodeGenOpenCL/address-spaces-conversions.cl
@@ -6,7 +6,9 @@
 // pointers to different address spaces
 
 // CHECK: define void @test
-void test(global int *arg_glob, generic int *arg_gen) {
+void test(global int *arg_glob, generic int *arg_gen,
+          __attribute__((opencl_global_device)) int *arg_device,
+          __attribute__((opencl_global_host)) int *arg_host) {
   int var_priv;
   arg_gen = arg_glob; // implicit cast global -> generic
   // CHECK: %{{[0-9]+}} = addrspacecast i32 addrspace(1)* %{{[0-9]+}} to i32 addrspace(4)*
@@ -39,6 +41,30 @@ void test(global int *arg_glob, generic int *arg_gen) {
   // CHECK-NOT: bitcast
   // CHECK-NOFAKE: bitcast
   // CHECK-NOFAKE-NOT: addrspacecast
+
+  arg_glob = arg_device; // implicit cast
+  // CHECK: addrspacecast
+  // CHECK-NOFAKE-NOT: addrspacecast
+
+  arg_glob = arg_host; // implicit cast
+  // CHECK: addrspacecast
+  // CHECK-NOFAKE-NOT: addrspacecast
+
+  arg_glob = (global int *)arg_device; // explicit cast
+  // CHECK: addrspacecast
+  // CHECK-NOFAKE-NOT: addrspacecast
+
+  arg_glob = (global int *)arg_host; // explicit cast
+  // CHECK: addrspacecast
+  // CHECK-NOFAKE-NOT: addrspacecast
+
+  arg_device = (__attribute((opencl_global_device)) int *)arg_glob; // explicit cast
+  // CHECK: addrspacecast
+  // CHECK-NOFAKE-NOT: addrspacecast
+
+  arg_host = (__attribute((opencl_global_host)) int *)arg_glob; // explicit cast
+  // CHECK: addrspacecast
+  // CHECK-NOFAKE-NOT: addrspacecast
 }
 
 // Test ternary operator.