[AMDGPU] Fix lowering enqueue_kernel

[yaxunl]

Summary

Two issues were fixed:

1. runtime has difficulty to allocate memory for an external symbol of a

kernel and set the address of the external symbol, therefore make the runtime
handle of an enqueued kernel an ordinary global variable. Runtime only needs
to store the address of the loaded kernel to the handle and has verified
that this approach works.

2. handle the situation where __enqueue_kernel* gets inlined therefore

the enqueued kernel may be used through a constant expr instead
of an instruction.

[rampitec]

lib/Target/AMDGPU/AMDGPUOpenCLEnqueuedBlockLowering.cpp

96

You can call callectCallers only if F was not already in the set, e.g. if (Funcs.insert(F).second).

[yaxunl]

Revise by Stas' comments and add the change about linkage omitted by accident.

[yaxunl]

ping

[rampitec]

LGTM

lib/Target/AMDGPU/AMDGPUOpenCLEnqueuedBlockLowering.cpp

 }
 
 /// Collect direct or indrect callers of \p F and save them
 /// to \p Callers.
 static void collectCallers(Function *F, DenseSet<Function *> &Callers) {
   for (auto U : F->users()) {
     if (auto *CI = dyn_cast<CallInst>(&*U)) {
       auto *Caller = CI->getParent()->getParent();
-      if (Callers.count(Caller))
-        continue;
-      Callers.insert(Caller);
+      if (Callers.insert(Caller).second)
         collectCallers(Caller, Callers);
     }
   }
 }
 
+/// If \p U is instruction or constant, collect functions which directly or
+/// indirectly use it.
+static void collectFunctionUsers(User *U, DenseSet<Function *> &Funcs) {
+  if (auto *I = dyn_cast<Instruction>(U)) {
+    auto *F = I->getParent()->getParent();
+    if (Funcs.insert(F).second)
+      collectCallers(F, Funcs);
+    return;
+  }
+  if (!isa<Constant>(U))
+    return;
+  for (auto UU : U->users())
+    collectFunctionUsers(&*UU, Funcs);
+}
+
 bool AMDGPUOpenCLEnqueuedBlockLowering::runOnModule(Module &M) {
   DenseSet<Function *> Callers;
   auto &C = M.getContext();
   bool Changed = false;
   for (auto &F : M.functions()) {
     if (F.hasFnAttribute("enqueued-block")) {
       if (!F.hasName()) {
         SmallString<64> Name;
         Mangler::getNameWithPrefix(Name, "__amdgpu_enqueued_kernel",
                                    M.getDataLayout());
         F.setName(Name);
       }
+      DEBUG(dbgs() << "found enqueued kernel: " << F.getName() << '\n');
       auto RuntimeHandle = (F.getName() + ".runtime_handle").str();
+      auto T = Type::getInt8Ty(C)->getPointerTo(AMDGPUAS::GLOBAL_ADDRESS);
       auto *GV = new GlobalVariable(
-          M, Type::getInt8Ty(C)->getPointerTo(AMDGPUAS::GLOBAL_ADDRESS),
-          /*IsConstant=*/true, GlobalValue::ExternalLinkage,
-          /*Initializer=*/nullptr, RuntimeHandle, /*InsertBefore=*/nullptr,
-          GlobalValue::NotThreadLocal, AMDGPUAS::GLOBAL_ADDRESS,
-          /*IsExternallyInitialized=*/true);
+          M, T,
+          /*IsConstant=*/false, GlobalValue::ExternalLinkage,
+          /*Initializer=*/Constant::getNullValue(T), RuntimeHandle,
+          /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal,
+          AMDGPUAS::GLOBAL_ADDRESS,
+          /*IsExternallyInitialized=*/false);
       DEBUG(dbgs() << "runtime handle created: " << *GV << '\n');
 
       for (auto U : F.users()) {
-        if (!isa<ConstantExpr>(&*U))
+        auto *UU = &*U;
+        if (!isa<ConstantExpr>(UU))
           continue;
-        auto *BitCast = cast<ConstantExpr>(&*U);
+        collectFunctionUsers(UU, Callers);
+        auto *BitCast = cast<ConstantExpr>(UU);
         auto *NewPtr = ConstantExpr::getPointerCast(GV, BitCast->getType());
         BitCast->replaceAllUsesWith(NewPtr);
         F.addFnAttr("runtime-handle", RuntimeHandle);
         F.setLinkage(GlobalValue::ExternalLinkage);
-
-        // Collect direct or indirect callers of enqueue_kernel.
-        for (auto U : NewPtr->users()) {
-          if (auto *I = dyn_cast<Instruction>(&*U)) {
-            auto *F = I->getParent()->getParent();
-            Callers.insert(F);
-            collectCallers(F, Callers);
-          }
-        }
         Changed = true;
       }
     }
   }
 
   for (auto F : Callers) {
     if (F->getCallingConv() != CallingConv::AMDGPU_KERNEL)
       continue;
     F->addFnAttr("calls-enqueue-kernel");
+    DEBUG(dbgs() << "mark enqueue_kernel caller:" << F->getName() << '\n');
   }
   return Changed;
 }

test/CodeGen/AMDGPU/enqueue-kernel.ll

 ; RUN: opt -data-layout=A5 -amdgpu-lower-enqueued-block -S < %s | FileCheck %s
 
-; CHECK: @__test_block_invoke_kernel.runtime_handle = external addrspace(1) externally_initialized constant i8 addrspace(1)*
-; CHECK: @__test_block_invoke_2_kernel.runtime_handle = external addrspace(1) externally_initialized constant i8 addrspace(1)*
-; CHECK: @__amdgpu_enqueued_kernel.runtime_handle = external addrspace(1) externally_initialized constant i8 addrspace(1)*
-; CHECK: @__amdgpu_enqueued_kernel.1.runtime_handle = external addrspace(1) externally_initialized constant i8 addrspace(1)*
+; CHECK: @__test_block_invoke_kernel.runtime_handle = addrspace(1) global i8 addrspace(1)* null
+; CHECK: @__test_block_invoke_2_kernel.runtime_handle = addrspace(1) global i8 addrspace(1)* null
+; CHECK: @__amdgpu_enqueued_kernel.runtime_handle = addrspace(1) global i8 addrspace(1)* null
+; CHECK: @__amdgpu_enqueued_kernel.1.runtime_handle = addrspace(1) global i8 addrspace(1)* null
 
 %struct.ndrange_t = type { i32 }
 %opencl.queue_t = type opaque
 
 ; CHECK-LABEL: define amdgpu_kernel void @non_caller
 ; CHECK-NOT: #{{[0-9]+}}
 define amdgpu_kernel void @non_caller(i8 addrspace(1)* %a, i8 %b, i64 addrspace(1)* %c, i64 %d) local_unnamed_addr
   !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !5 !kernel_arg_base_type !5 !kernel_arg_type_qual !6 {
entry:
   store i64 %d, i64 addrspace(5)* %block.captured10, align 8
   %tmp6 = bitcast <{ i32, i32, i8 addrspace(1)*, i64 addrspace(1)*, i64, i8 }> addrspace(5)* %block2 to void () addrspace(5)*
   %tmp8 = addrspacecast void () addrspace(5)* %tmp6 to i8*
   %tmp9 = call i32 @__enqueue_kernel_basic(%opencl.queue_t addrspace(1)* undef, i32 0, %struct.ndrange_t addrspace(5)* byval nonnull %tmp3,
     i8* bitcast (void (<{ i32, i32, i8 addrspace(1)*, i64 addrspace(1)*, i64, i8 }>)* @__test_block_invoke_2_kernel to i8*), i8* nonnull %tmp8) #2
   ret void
 }
 
+; __enqueue_kernel* functions may get inlined
+; CHECK-LABEL: define amdgpu_kernel void @inlined_caller
+; CHECK-SAME: #[[AT_CALLER]]
+; CHECK-NOT: @__test_block_invoke_kernel
+; CHECK: load i64, i64 addrspace(1)* bitcast (i8 addrspace(1)* addrspace(1)* @__test_block_invoke_kernel.runtime_handle to i64 addrspace(1)*)
+define amdgpu_kernel void @inlined_caller(i8 addrspace(1)* %a, i8 %b, i64 addrspace(1)* %c, i64 %d) local_unnamed_addr
+  !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !5 !kernel_arg_base_type !5 !kernel_arg_type_qual !6 {
+entry:
+  %tmp = load i64, i64 addrspace(1)* addrspacecast (i64* bitcast (void (<{ i32, i32, i8 addrspace(1)*, i8 }>)* @__test_block_invoke_kernel to i64*) to i64 addrspace(1)*)
+  store i64 %tmp, i64 addrspace(1)* %c
+  ret void
+}
+
 ; CHECK-LABEL: define dso_local amdgpu_kernel void @__test_block_invoke_kernel
 ; CHECK-SAME: #[[AT1:[0-9]+]]
 define internal amdgpu_kernel void @__test_block_invoke_kernel(<{ i32, i32, i8 addrspace(1)*, i8 }> %arg) #0
   !kernel_arg_addr_space !14 !kernel_arg_access_qual !15 !kernel_arg_type !16 !kernel_arg_base_type !16 !kernel_arg_type_qual !17 {
 entry:
   %.fca.3.extract = extractvalue <{ i32, i32, i8 addrspace(1)*, i8 }> %arg, 2
   %.fca.4.extract = extractvalue <{ i32, i32, i8 addrspace(1)*, i8 }> %arg, 3
   store i8 %.fca.4.extract, i8 addrspace(1)* %.fca.3.extract, align 1