114 lines · plain
1; RUN: llc -O0 -mtriple=spirv32-unknown-unknown %s -o - | FileCheck %s --check-prefix=CHECK-SPIRV2 3; TODO: This test currently fails with LLVM_ENABLE_EXPENSIVE_CHECKS enabled4; XFAIL: expensive_checks5 6; CHECK-SPIRV-DAG: %[[#]] = OpGroupAsyncCopy %[[#]] %[[#Scope:]]7; CHECK-SPIRV-DAG: %[[#Scope]] = OpConstant %[[#]]8 9%opencl.event_t = type opaque10 11define spir_kernel void @test_fn(<2 x i8> addrspace(1)* %src, <2 x i8> addrspace(1)* %dst, <2 x i8> addrspace(3)* %localBuffer, i32 %copiesPerWorkgroup, i32 %copiesPerWorkItem) {12entry:13 %src.addr = alloca <2 x i8> addrspace(1)*, align 414 %dst.addr = alloca <2 x i8> addrspace(1)*, align 415 %localBuffer.addr = alloca <2 x i8> addrspace(3)*, align 416 %copiesPerWorkgroup.addr = alloca i32, align 417 %copiesPerWorkItem.addr = alloca i32, align 418 %i = alloca i32, align 419 %event = alloca %opencl.event_t*, align 420 store <2 x i8> addrspace(1)* %src, <2 x i8> addrspace(1)** %src.addr, align 421 store <2 x i8> addrspace(1)* %dst, <2 x i8> addrspace(1)** %dst.addr, align 422 store <2 x i8> addrspace(3)* %localBuffer, <2 x i8> addrspace(3)** %localBuffer.addr, align 423 store i32 %copiesPerWorkgroup, i32* %copiesPerWorkgroup.addr, align 424 store i32 %copiesPerWorkItem, i32* %copiesPerWorkItem.addr, align 425 store i32 0, i32* %i, align 426 br label %for.cond27 28for.cond: ; preds = %for.inc, %entry29 %0 = load i32, i32* %i, align 430 %1 = load i32, i32* %copiesPerWorkItem.addr, align 431 %cmp = icmp slt i32 %0, %132 br i1 %cmp, label %for.body, label %for.end33 34for.body: ; preds = %for.cond35 %call = call spir_func i32 @_Z12get_local_idj(i32 0)36 %2 = load i32, i32* %copiesPerWorkItem.addr, align 437 %mul = mul i32 %call, %238 %3 = load i32, i32* %i, align 439 %add = add i32 %mul, %340 %4 = load <2 x i8> addrspace(3)*, <2 x i8> addrspace(3)** %localBuffer.addr, align 441 %arrayidx = getelementptr inbounds <2 x i8>, <2 x i8> addrspace(3)* %4, i32 %add42 store <2 x i8> zeroinitializer, <2 x i8> addrspace(3)* %arrayidx, align 243 br label %for.inc44 45for.inc: ; preds = %for.body46 %5 = load i32, i32* %i, align 447 %inc = add nsw i32 %5, 148 store i32 %inc, i32* %i, align 449 br label %for.cond50 51for.end: ; preds = %for.cond52 call spir_func void @_Z7barrierj(i32 1)53 store i32 0, i32* %i, align 454 br label %for.cond155 56for.cond1: ; preds = %for.inc12, %for.end57 %6 = load i32, i32* %i, align 458 %7 = load i32, i32* %copiesPerWorkItem.addr, align 459 %cmp2 = icmp slt i32 %6, %760 br i1 %cmp2, label %for.body3, label %for.end1461 62for.body3: ; preds = %for.cond163 %call4 = call spir_func i32 @_Z13get_global_idj(i32 0)64 %8 = load i32, i32* %copiesPerWorkItem.addr, align 465 %mul5 = mul i32 %call4, %866 %9 = load i32, i32* %i, align 467 %add6 = add i32 %mul5, %968 %10 = load <2 x i8> addrspace(1)*, <2 x i8> addrspace(1)** %src.addr, align 469 %arrayidx7 = getelementptr inbounds <2 x i8>, <2 x i8> addrspace(1)* %10, i32 %add670 %11 = load <2 x i8>, <2 x i8> addrspace(1)* %arrayidx7, align 271 %call8 = call spir_func i32 @_Z12get_local_idj(i32 0)72 %12 = load i32, i32* %copiesPerWorkItem.addr, align 473 %mul9 = mul i32 %call8, %1274 %13 = load i32, i32* %i, align 475 %add10 = add i32 %mul9, %1376 %14 = load <2 x i8> addrspace(3)*, <2 x i8> addrspace(3)** %localBuffer.addr, align 477 %arrayidx11 = getelementptr inbounds <2 x i8>, <2 x i8> addrspace(3)* %14, i32 %add1078 store <2 x i8> %11, <2 x i8> addrspace(3)* %arrayidx11, align 279 br label %for.inc1280 81for.inc12: ; preds = %for.body382 %15 = load i32, i32* %i, align 483 %inc13 = add nsw i32 %15, 184 store i32 %inc13, i32* %i, align 485 br label %for.cond186 87for.end14: ; preds = %for.cond188 call spir_func void @_Z7barrierj(i32 1)89 %16 = load <2 x i8> addrspace(1)*, <2 x i8> addrspace(1)** %dst.addr, align 490 %17 = load i32, i32* %copiesPerWorkgroup.addr, align 491 %call15 = call spir_func i32 @_Z12get_group_idj(i32 0)92 %mul16 = mul i32 %17, %call1593 %add.ptr = getelementptr inbounds <2 x i8>, <2 x i8> addrspace(1)* %16, i32 %mul1694 %18 = load <2 x i8> addrspace(3)*, <2 x i8> addrspace(3)** %localBuffer.addr, align 495 %19 = load i32, i32* %copiesPerWorkgroup.addr, align 496 %call17 = call spir_func %opencl.event_t* @_Z21async_work_group_copyPU3AS1Dv2_cPKU3AS3S_j9ocl_event(<2 x i8> addrspace(1)* %add.ptr, <2 x i8> addrspace(3)* %18, i32 %19, %opencl.event_t* null)97 store %opencl.event_t* %call17, %opencl.event_t** %event, align 498 %20 = addrspacecast %opencl.event_t** %event to %opencl.event_t* addrspace(4)*99 call spir_func void @_Z17wait_group_eventsiPU3AS49ocl_event(i32 1, %opencl.event_t* addrspace(4)* %20)100 ret void101}102 103declare spir_func i32 @_Z12get_local_idj(i32)104 105declare spir_func void @_Z7barrierj(i32)106 107declare spir_func i32 @_Z13get_global_idj(i32)108 109declare spir_func %opencl.event_t* @_Z21async_work_group_copyPU3AS1Dv2_cPKU3AS3S_j9ocl_event(<2 x i8> addrspace(1)*, <2 x i8> addrspace(3)*, i32, %opencl.event_t*)110 111declare spir_func i32 @_Z12get_group_idj(i32)112 113declare spir_func void @_Z17wait_group_eventsiPU3AS49ocl_event(i32, %opencl.event_t* addrspace(4)*)114