[clang][SPIR][SPIRV] Materialize non-generic null pointers via addrspacecast (#161773)

LLVM models ConstantPointerNull as all-zero, but some GPUs (e.g. AMDGPU
and our downstream GPU target) use a non-zero sentinel for null in
private / local address spaces.
SPIR-V is a supported input for our GPU target. This PR preserves a
canonical zero form in the generic AS while allowing later lowering to
substitute the target’s real sentinel.

Author: wenju-he

clang/lib/CodeGen/Targets/SPIR.cpp

@@ -61,6 +61,9 @@ class CommonSPIRTargetCodeGenInfo : public TargetCodeGenInfo {
       QualType SampledType, CodeGenModule &CGM) const;
   void
   setOCLKernelStubCallingConvention(const FunctionType *&FT) const override;
+  llvm::Constant *getNullPointer(const CodeGen::CodeGenModule &CGM,
+                                 llvm::PointerType *T,
+                                 QualType QT) const override;
 };
 class SPIRVTargetCodeGenInfo : public CommonSPIRTargetCodeGenInfo {
 public:
@@ -240,6 +243,29 @@ void CommonSPIRTargetCodeGenInfo::setOCLKernelStubCallingConvention(
       FT, FT->getExtInfo().withCallingConv(CC_SpirFunction));
 }
 
+// LLVM currently assumes a null pointer has the bit pattern 0, but some GPU
+// targets use a non-zero encoding for null in certain address spaces.
+// Because SPIR(-V) is a generic target and the bit pattern of null in
+// non-generic AS is unspecified, materialize null in non-generic AS via an
+// addrspacecast from null in generic AS. This allows later lowering to
+// substitute the target's real sentinel value.
+llvm::Constant *
+CommonSPIRTargetCodeGenInfo::getNullPointer(const CodeGen::CodeGenModule &CGM,
+                                            llvm::PointerType *PT,
+                                            QualType QT) const {
+  LangAS AS = QT->getUnqualifiedDesugaredType()->isNullPtrType()
+                  ? LangAS::Default
+                  : QT->getPointeeType().getAddressSpace();
+  if (AS == LangAS::Default || AS == LangAS::opencl_generic)
+    return llvm::ConstantPointerNull::get(PT);
+
+  auto &Ctx = CGM.getContext();
+  auto NPT = llvm::PointerType::get(
+      PT->getContext(), Ctx.getTargetAddressSpace(LangAS::opencl_generic));
+  return llvm::ConstantExpr::getAddrSpaceCast(
+      llvm::ConstantPointerNull::get(NPT), PT);
+}
+
 LangAS
 SPIRVTargetCodeGenInfo::getGlobalVarAddressSpace(CodeGenModule &CGM,
                                                  const VarDecl *D) const {

clang/test/CodeGenOpenCL/builtins.cl

@@ -62,19 +62,19 @@ void testBranchingOnAddressSpaceCast(generic long* ptr) {
     if (to_global(ptr))
         (void)0;
     // CHECK:       [[P:%[0-9]+]] = call spir_func [[GLOBAL_VOID:ptr addrspace\(1\)]] @__to_global([[GENERIC_VOID:ptr addrspace\(4\)]] {{%[0-9]+}})
-    // CHECK-NEXT:  [[BOOL:%[a-z0-9]+]] = icmp ne ptr addrspace(1) [[P]], null
+    // CHECK-NEXT:  [[BOOL:%[a-z0-9]+]] = icmp ne ptr addrspace(1) [[P]], addrspacecast (ptr addrspace(4) null to ptr addrspace(1))
     // CHECK-NEXT:  br i1 [[BOOL]]
 
     if (to_local(ptr))
         (void)0;
     // CHECK:       [[P:%[0-9]+]] = call spir_func [[LOCAL_VOID:ptr addrspace\(3\)]] @__to_local([[GENERIC_VOID]] {{%[0-9]+}})
-    // CHECK-NEXT:  [[BOOL:%[a-z0-9]+]] = icmp ne ptr addrspace(3) [[P]], null
+    // CHECK-NEXT:  [[BOOL:%[a-z0-9]+]] = icmp ne ptr addrspace(3) [[P]], addrspacecast (ptr addrspace(4) null to ptr addrspace(3))
     // CHECK-NEXT:  br i1 [[BOOL]]
 
     if (to_private(ptr))
         (void)0;
     // CHECK:       [[P:%[0-9]+]] = call spir_func [[PRIVATE_VOID:ptr]] @__to_private([[GENERIC_VOID]] {{%[0-9]+}})
-    // CHECK-NEXT:  [[BOOL:%[a-z0-9]+]] = icmp ne ptr [[P]], null
+    // CHECK-NEXT:  [[BOOL:%[a-z0-9]+]] = icmp ne ptr [[P]], addrspacecast (ptr addrspace(4) null to ptr)
     // CHECK-NEXT:  br i1 [[BOOL]]
 }
 

clang/test/CodeGenSYCL/address-space-conversions.cpp

@@ -35,9 +35,9 @@ void tmpl(T t) {}
   // CHECK-DAG: [[LOC]].ascast = addrspacecast ptr [[LOC]] to ptr addrspace(4)
   // CHECK-DAG: [[PRIV]].ascast = addrspacecast ptr [[PRIV]] to ptr addrspace(4)
   LOC = nullptr;
-  // CHECK-DAG: store ptr addrspace(3) null, ptr addrspace(4) [[LOC]].ascast, align 8
+  // CHECK-DAG: store ptr addrspace(3) addrspacecast (ptr addrspace(4) null to ptr addrspace(3)), ptr addrspace(4) [[LOC]].ascast, align 8
   GLOB = nullptr;
-  // CHECK-DAG: store ptr addrspace(1) null, ptr addrspace(4) [[GLOB]].ascast, align 8
+  // CHECK-DAG: store ptr addrspace(1) addrspacecast (ptr addrspace(4) null to ptr addrspace(1)), ptr addrspace(4) [[GLOB]].ascast, align 8
 
   // Explicit conversions
   // From named address spaces to default address space