[HLSL] Call global constructors inside entry

[beanz]

Summary

HLSL doesn't have a runtime loader model that supports global
construction by a loader or runtime initializer. To allow us to leverage
global constructors with minimal code generation impact we put calls to
the global constructors inside the generated entry function.

[python3kgae]

clang/lib/CodeGen/CGHLSLRuntime.cpp

151

Don't need to generate CtorCalls for lib profile in clang codeGen.
Have to do this when linking anyway.

[beanz]

clang/lib/CodeGen/CGHLSLRuntime.cpp

151

Are you sure? The global constructors contain the createHandle calls for resources, and DXC does currently generate those inside shader entries for lib shaders.

I'm sure we're not generating those _correctly_ with this change, but I think we still need the constructor calls inside any function annotated with the [shader(...)] attribute.

[python3kgae]

clang/lib/CodeGen/CGHLSLRuntime.cpp

151

I saw dxc did call ctor for all entries. Maybe to match the behavior, we need to do the same thing.
But I don't understand why we must do that except to match what dxc does.
Maybe we can also have a test for lib profile to make understanding things easier?

[beanz]

clang/lib/CodeGen/CGHLSLRuntime.cpp

151

I can (and should) add a library shader test too.

In addition to the cases where DXC generates global constructors, we need them in clang in more cases, because I'm using global constructors to reduce the scope of changes HLSL needs in CodeGen.

In the case of global resource variables (as shown in the test in this PR), DXC generates the handle creation calls with a bunch of custom code in clang CodeGen, and inserts that code into the entry functions.

To avoid needing special code for handle creation in Clang, I've put handle creation calls inside the constructors for the resource object. For global definitions, the constructors then get called by the global initialization function automatically, which we can inline into the entry.

[aaron.ballman]

clang/test/CodeGenHLSL/GlobalConstructors.hlsl

3

Can you also add a test using __attribute__((constructor))? And probably one using __attribute__((destructor)) at some point as well?

https://godbolt.org/z/jqedf1qGa

I see the docs said that global destructors are not supported, which is fine, but then what's the behavior when a user tries one anyway? (I would expect we'd get reasonable errors about it not being supported). Feel free to handle the dtor bits in a follow-up though.

[beanz]

Adding test coverage for __attribute__((constructor))

[beanz]

clang/test/CodeGenHLSL/GlobalConstructors.hlsl

3

I need to play with this. We don't currently support global destructors, but I think we're going to in Clang.

I'll try to get a second patch up today or tomorrow to add global destructors.

[aaron.ballman]

clang/lib/CodeGen/CGHLSLRuntime.cpp

167–170

Do we have sema checks for this so users get decent diagnostics instead of a crash (or miscompile)?

[beanz]

clang/lib/CodeGen/CGHLSLRuntime.cpp

167–170

It looks like COMDats are only attached if object format supports them (and DXContainer isn't properly handled there). I don't think there is any sema-level validation because the COMDat just binds to the global value address. I'll get a patch up to llvm::Triple to mark DXContainer as not supporting COMDat.

[aaron.ballman]

clang/lib/CodeGen/CGHLSLRuntime.cpp

167–170

Thanks! Also be sure to add some logic to SemaDeclAttr.cpp to diagnose specifying a priority.

[beanz]

Adding Sema errors for specifying initializer priority, and a test case for library shaders.

[aaron.ballman]

LGTM!

clang/docs/HLSL/EntryFunctions.rst

 profile you are compiling.
 
 In code generation, two functions are generated. One is the user defined
 function, which is code generated as a mangled C++ function with internal
 linkage following normal function code generation.
 
 The actual exported entry function which can be called by the GPU driver is a
 ``void(void)`` function that isn't name mangled. In code generation we generate
-the unmangled entry function, instantiations of the parameters with their
-semantic values populated, and a call to the user-defined function. After the
-call instruction the return value (if any) is saved using a target-appropriate
-intrinsic for storing outputs (for DirectX, the ``llvm.dx.store.output``).
+the unmangled entry function to serve as the actual shader entry. The shader
+entry function is annotated with the ``hlsl.shader`` function attribute
+identifying the entry's pipeline stage.
+
+The body of the unmangled entry function contains first a call to execute global
+constructors, then instantiations of the user-defined entry parameters with
+their semantic values populated, and a call to the user-defined function.
+After the call instruction the return value (if any) is saved using a
+target-appropriate intrinsic for storing outputs (for DirectX, the
+``llvm.dx.store.output``). Global destructors are not supported in HLSL.
 
 .. note::
 
    HLSL support in Clang is currently focused on compute shaders, which do not
    support output semantics. Support for output semantics will not be
    implemented until other shader profiles are supported.
 
 Below is example IR that represents the planned implementation, subject to

clang/lib/CodeGen/CGHLSLRuntime.h

protected:
 
   llvm::Value *emitInputSemantic(llvm::IRBuilder<> &B, const ParmVarDecl &D);
 
 public:
   CGHLSLRuntime(CodeGenModule &CGM) : CGM(CGM) {}
   virtual ~CGHLSLRuntime() {}
 
   void annotateHLSLResource(const VarDecl *D, llvm::GlobalVariable *GV);
+  void generateGlobalCtorCalls();
 
   void finishCodeGen();
 
   void setHLSLEntryAttributes(const FunctionDecl *FD, llvm::Function *Fn);
 
   void emitEntryFunction(const FunctionDecl *FD, llvm::Function *Fn);
 };
 
 } // namespace CodeGen
 } // namespace clang
 
 #endif

clang/lib/CodeGen/CGHLSLRuntime.cpp

 } // namespace
 
 void CGHLSLRuntime::finishCodeGen() {
   auto &TargetOpts = CGM.getTarget().getTargetOpts();
   llvm::Module &M = CGM.getModule();
   Triple T(M.getTargetTriple());
   if (T.getArch() == Triple::ArchType::dxil)
     addDxilValVersion(TargetOpts.DxilValidatorVersion, M);
+
+  generateGlobalCtorCalls();
 }
 
 void CGHLSLRuntime::annotateHLSLResource(const VarDecl *D, GlobalVariable *GV) {
   const Type *Ty = D->getType()->getPointeeOrArrayElementType();
   if (!Ty)
     return;
   const auto *RD = Ty->getAsCXXRecordDecl();
   if (!RD)
for (const auto *Param : FD->parameters()) {
     Args.push_back(emitInputSemantic(B, *Param));
   }
 
   CallInst *CI = B.CreateCall(FunctionCallee(Fn), Args);
   (void)CI;
   // FIXME: Handle codegen for return type semantics.
   B.CreateRetVoid();
 }
+
+void CGHLSLRuntime::generateGlobalCtorCalls() {
+  llvm::Module &M = CGM.getModule();
+  const auto *GlobalCtors = M.getNamedGlobal("llvm.global_ctors");
+  if (!GlobalCtors)
+    return;
+  const auto *CA = dyn_cast<ConstantArray>(GlobalCtors->getInitializer());
+  if (!CA)
+    return;
+  // The global_ctor array elements are a struct [Priority, Fn *, COMDat].
+  // HLSL neither supports priorities or COMDat values, so we will check those
+  // in an assert but not handle them.
+
+  llvm::SmallVector<Function *> CtorFns;
+  for (const auto &Ctor : CA->operands()) {
+    if (isa<ConstantAggregateZero>(Ctor))
+      continue;
+    ConstantStruct *CS = cast<ConstantStruct>(Ctor);
+
+    assert(cast<ConstantInt>(CS->getOperand(0))->getValue() == 65535 &&
+           "HLSL doesn't support setting priority for global ctors.");
+    assert(isa<ConstantPointerNull>(CS->getOperand(2)) &&
+           "HLSL doesn't support COMDat for global ctors.");
+    CtorFns.push_back(cast<Function>(CS->getOperand(1)));
+  }
+
+  // Insert a call to the global constructor at the beginning of the entry block
+  // to externally exported functions. This is a bit of a hack, but HLSL allows
+  // global constructors, but doesn't support driver initialization of globals.
+  for (auto &F : M.functions()) {
+    if (!F.hasFnAttribute("hlsl.shader"))
+      continue;
+    IRBuilder<> B(&F.getEntryBlock(), F.getEntryBlock().begin());
+    for (auto *Fn : CtorFns)
+      B.CreateCall(FunctionCallee(Fn));
+  }
+}

clang/test/CodeGenHLSL/GlobalConstructors.hlsl

+// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.0-compute -x hlsl -S -emit-llvm -disable-llvm-passes %s -o - | FileCheck %s
+
+RWBuffer<float> Buffer;
+
+[numthreads(1,1,1)]
+void main(unsigned GI : SV_GroupIndex) {}
+
+//CHECK:      define void @main()
+//CHECK-NEXT: entry:
+//CHECK-NEXT:   call void @_GLOBAL__sub_I_GlobalConstructors.hlsl()
+//CHECK-NEXT:   %0 = call i32 @llvm.dx.flattened.thread.id.in.group()
+//CHECK-NEXT:   call void @"?main@@YAXI@Z"(i32 %0)
+//CHECK-NEXT:   ret void