AMDGPU: Use constant and externally_initialized for block handle

[arsenm]

Summary

Also stop using a null initializer. The runtime initializes this, but
I don't see why we aren't using relocations for this. We should be
able to easily fill in the function address. The private/group size
should also be available by relocations but aren't.

[yaxunl]

I doubt this will work. First, this may result in missing symbol in lld. Second, the runtime currently expects *.runtime_handle global symbols having a certain size and will initialize them by copying to them. If this compiler change requires a corresponding runtime change, it should be coordinated with runtime.

[arsenm]

In D141323#4037714, @yaxunl wrote:

I doubt this will work. First, this may result in missing symbol in lld. Second, the runtime currently expects *.runtime_handle global symbols having a certain size and will initialize them by copying to them. If this compiler change requires a corresponding runtime change, it should be coordinated with runtime.

It does work, this just changes optimizer behavior

[arsenm]

In D141323#4037723, @arsenm wrote:

In D141323#4037714, @yaxunl wrote:

I doubt this will work. First, this may result in missing symbol in lld. Second, the runtime currently expects *.runtime_handle global symbols having a certain size and will initialize them by copying to them. If this compiler change requires a corresponding runtime change, it should be coordinated with runtime.

It does work, this just changes optimizer behavior

Neither the size, name nor linkage changes

[yaxunl]

In D141323#4037724, @arsenm wrote:

In D141323#4037723, @arsenm wrote:

In D141323#4037714, @yaxunl wrote:

I doubt this will work. First, this may result in missing symbol in lld. Second, the runtime currently expects *.runtime_handle global symbols having a certain size and will initialize them by copying to them. If this compiler change requires a corresponding runtime change, it should be coordinated with runtime.

It does work, this just changes optimizer behavior

Neither the size, name nor linkage changes

Then it should be OK.

[arsenm]

I posted the g version of this, still needs the undef initializer. available_externally and constant initializer are the useful parts, and can use undef initializer. the zeroinitializer was pointless since it's overwritten

[arsenm]

Drop initializer change, needs to be undef, zero or the actual content and there's no practical difference

[arsenm]

4fc07e1849488718edbe4a0ec68c2143cd40d7c0

llvm/lib/Target/AMDGPU/AMDGPUOpenCLEnqueuedBlockLowering.cpp

if (F.hasFnAttribute("enqueued-block")) {
         Type *Int32 = Type::getInt32Ty(C);
         HandleTy = StructType::create(
             C, {Type::getInt8Ty(C)->getPointerTo(0), Int32, Int32},
             "block.runtime.handle.t");
       }
 
       auto *GV = new GlobalVariable(
           M, HandleTy,
-          /*isConstant=*/false, GlobalValue::ExternalLinkage,
+          /*isConstant=*/true, GlobalValue::ExternalLinkage,
           /*Initializer=*/Constant::getNullValue(HandleTy), RuntimeHandle,
           /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal,
           AMDGPUAS::GLOBAL_ADDRESS,
-          /*isExternallyInitialized=*/false);
+          /*isExternallyInitialized=*/true);
       LLVM_DEBUG(dbgs() << "runtime handle created: " << *GV << '\n');
 
       F.replaceAllUsesWith(ConstantExpr::getAddrSpaceCast(GV, F.getType()));
       F.addFnAttr("runtime-handle", RuntimeHandle);
       F.setLinkage(GlobalValue::ExternalLinkage);
       Changed = true;
     }
   }
 
   return Changed;
 }

llvm/test/CodeGen/AMDGPU/enqueue-kernel.ll

 
 define internal amdgpu_kernel void @1(<{ i32, i32, ptr addrspace(1), i8 }> %arg) #0 {
   ret void
 }
 
 attributes #0 = { "enqueued-block" }
 ;.
 ; CHECK: @[[KERNEL_ADDRESS_USER:[a-zA-Z0-9_$"\\.-]+]] = global [1 x ptr] [ptr addrspacecast (ptr addrspace(1) @block_has_used_kernel_address.runtime_handle to ptr)]
-; CHECK: @[[__TEST_BLOCK_INVOKE_KERNEL_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) global [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
-; CHECK: @[[__TEST_BLOCK_INVOKE_2_KERNEL_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) global [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
-; CHECK: @[[BLOCK_HAS_USED_KERNEL_ADDRESS_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) global [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
-; CHECK: @[[__AMDGPU_ENQUEUED_KERNEL_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) global [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
-; CHECK: @[[__AMDGPU_ENQUEUED_KERNEL_1_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) global [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
+; CHECK: @[[__TEST_BLOCK_INVOKE_KERNEL_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) externally_initialized constant [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
+; CHECK: @[[__TEST_BLOCK_INVOKE_2_KERNEL_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) externally_initialized constant [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
+; CHECK: @[[BLOCK_HAS_USED_KERNEL_ADDRESS_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) externally_initialized constant [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
+; CHECK: @[[__AMDGPU_ENQUEUED_KERNEL_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) externally_initialized constant [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
+; CHECK: @[[__AMDGPU_ENQUEUED_KERNEL_1_RUNTIME_HANDLE:[a-zA-Z0-9_$"\\.-]+]] = addrspace(1) externally_initialized constant [[BLOCK_RUNTIME_HANDLE_T:%.*]] zeroinitializer
 ;.
 ; CHECK-LABEL: define {{[^@]+}}@non_caller
 ; CHECK-SAME: (ptr addrspace(1) [[A:%.*]], i8 [[B:%.*]], ptr addrspace(1) [[C:%.*]], i64 [[D:%.*]]) {
 ; CHECK-NEXT:    ret void
 ;
 ;
 ; CHECK-LABEL: define {{[^@]+}}@caller
 ; CHECK-SAME: (ptr addrspace(1) [[A:%.*]], i8 [[B:%.*]], ptr addrspace(1) [[C:%.*]], i64 [[D:%.*]]) {