[SYCL] Mark calls to barrier and work-item functions as convergent

Calls to work-item functions and barrier function must be convergent to
preserve semantics of these functions during llvm optimizations.

Signed-off-by: Artur Gainullin <[email protected]>

Author: againull

llvm/lib/SYCLLowerIR/LowerWGScope.cpp

@@ -874,10 +874,15 @@ Value *spirv::genLinearLocalID(Instruction &Before, const Triple &TT) {
     IRBuilder<> Bld(Ctx);
     Bld.SetInsertPoint(&Before);
 
+    AttributeList Attr;
+    Attr = Attr.addAttribute(Ctx, AttributeList::FunctionIndex,
+                             Attribute::Convergent);
+
 #define CREATE_CALLEE(NAME, FN_NAME)                                           \
-  FunctionCallee FnCallee##NAME = M.getOrInsertFunction(FN_NAME, RetTy);       \
+  FunctionCallee FnCallee##NAME = M.getOrInsertFunction(FN_NAME, Attr, RetTy); \
   assert(FnCallee##NAME && "spirv intrinsic creation failed");                 \
-  auto NAME = Bld.CreateCall(FnCallee##NAME, {});
+  auto NAME = Bld.CreateCall(FnCallee##NAME, {});                              \
+  NAME->addAttribute(AttributeList::FunctionIndex, Attribute::Convergent);
 
     CREATE_CALLEE(LocalInvocationId_X, "_Z27__spirv_LocalInvocationId_xv");
     CREATE_CALLEE(LocalInvocationId_Y, "_Z27__spirv_LocalInvocationId_yv");
@@ -937,8 +942,11 @@ Instruction *spirv::genWGBarrier(Instruction &Before, const Triple &TT) {
   Type *SemanticsTy = Type::getInt32Ty(Ctx);
   Type *RetTy = Type::getVoidTy(Ctx);
 
+  AttributeList Attr;
+  Attr = Attr.addAttribute(Ctx, AttributeList::FunctionIndex,
+                           Attribute::Convergent);
   FunctionCallee FC =
-      M.getOrInsertFunction(Name, RetTy, ScopeTy, ScopeTy, SemanticsTy);
+      M.getOrInsertFunction(Name, Attr, RetTy, ScopeTy, ScopeTy, SemanticsTy);
   assert(FC.getCallee() && "spirv intrinsic creation failed");
 
   IRBuilder<> Bld(Ctx);
@@ -948,5 +956,8 @@ Instruction *spirv::genWGBarrier(Instruction &Before, const Triple &TT) {
   auto ArgSema = ConstantInt::get(
       ScopeTy, asUInt(spirv::MemorySemantics::SequentiallyConsistent) |
                    asUInt(spirv::MemorySemantics::WorkgroupMemory));
-  return Bld.CreateCall(FC, {ArgExec, ArgMem, ArgSema});
+  auto BarrierCall = Bld.CreateCall(FC, {ArgExec, ArgMem, ArgSema});
+  BarrierCall->addAttribute(llvm::AttributeList::FunctionIndex,
+                            llvm::Attribute::Convergent);
+  return BarrierCall;
 }

llvm/test/SYCLLowerIR/convergent.ll

@@ -0,0 +1,43 @@
+; RUN: opt < %s -LowerWGScope -S | FileCheck %s
+; RUN: opt < %s -LowerWGScope --mtriple=nvptx -S | FileCheck %s -check-prefix=CHECK-PTX
+
+
+%struct.baz = type { i64 }
+
+define internal spir_func void @wibble(%struct.baz* byval(%struct.baz) %arg1) !work_group_scope !0 {
+; CHECK-PTX:   %1 = call i64 @_Z27__spirv_LocalInvocationId_xv() #0
+; CHECK-PTX:   %2 = call i64 @_Z27__spirv_LocalInvocationId_yv() #0
+; CHECK-PTX:   %3 = call i64 @_Z27__spirv_LocalInvocationId_zv() #0
+; CHECK-PTX:   %4 = call i64 @_Z23__spirv_WorkgroupSize_yv() #0
+; CHECK-PTX:   %5 = call i64 @_Z23__spirv_WorkgroupSize_zv() #0
+; CHECK-PTX:   call void @_Z22__spirv_ControlBarrierN5__spv5ScopeES0_j(i32 2, i32 2, i32 272) #0
+; CHECK:   call void @__spirv_ControlBarrier(i32 2, i32 2, i32 272) #1
+  ret void
+}
+
+; CHECK-PTX: ; Function Attrs: convergent
+; CHECK-PTX: declare i64 @_Z27__spirv_LocalInvocationId_xv() #0
+
+; CHECK-PTX: ; Function Attrs: convergent
+; CHECK-PTX: declare i64 @_Z27__spirv_LocalInvocationId_yv() #0
+
+; CHECK-PTX: ; Function Attrs: convergent
+; CHECK-PTX: declare i64 @_Z27__spirv_LocalInvocationId_zv() #0
+
+; CHECK-PTX: ; Function Attrs: convergent
+; CHECK-PTX: declare i64 @_Z23__spirv_WorkgroupSize_yv() #0
+
+; CHECK-PTX: ; Function Attrs: convergent
+; CHECK-PTX: declare i64 @_Z23__spirv_WorkgroupSize_zv() #0
+
+; CHECK-PTX: ; Function Attrs: convergent
+; CHECK-PTX: declare void @_Z22__spirv_ControlBarrierN5__spv5ScopeES0_j(i32, i32, i32) #0
+
+; CHECK-PTX: attributes #0 = { convergent }
+
+; CHECK: ; Function Attrs: convergent
+; CHECK: declare void @__spirv_ControlBarrier(i32, i32, i32) #1
+
+; CHECK: attributes #1 = { convergent }
+
+!0 = !{}