133 lines · plain
1; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -passes=amdgpu-attributor -o %t.gfx7.bc %s2; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -passes=amdgpu-attributor -o %t.gfx8.bc %s3; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -passes=amdgpu-attributor -o %t.gfx9.bc %s4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 < %t.gfx7.bc | FileCheck --check-prefixes=CHECK,PRE-GFX9 %s5; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 < %t.gfx8.bc | FileCheck --check-prefixes=CHECK,PRE-GFX9 %s6; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %t.gfx9.bc | FileCheck --check-prefixes=CHECK,GFX9 %s7 8; CHECK: addrspacecast_requires_queue_ptr9; PRE-GFX9: .amdhsa_user_sgpr_queue_ptr 110; GFX9: .amdhsa_user_sgpr_queue_ptr 011 12; CHECK: is_shared_requires_queue_ptr13; PRE-GFX9: .amdhsa_user_sgpr_queue_ptr 114; GFX9: .amdhsa_user_sgpr_queue_ptr 015 16; CHECK: is_private_requires_queue_ptr17; PRE-GFX9: .amdhsa_user_sgpr_queue_ptr 118; GFX9: .amdhsa_user_sgpr_queue_ptr 019 20; CHECK: trap_requires_queue_ptr21; PRE-GFX9: .amdhsa_user_sgpr_queue_ptr 122; GFX9: .amdhsa_user_sgpr_queue_ptr 023 24; CHECK: debugtrap_requires_queue_ptr25; PRE-GFX9: .amdhsa_user_sgpr_queue_ptr 126; GFX9: .amdhsa_user_sgpr_queue_ptr 027 28; CHECK: ubsantrap_requires_queue_ptr29; PRE-GFX9: .amdhsa_user_sgpr_queue_ptr 130; GFX9: .amdhsa_user_sgpr_queue_ptr 031 32; CHECK: amdgcn_queue_ptr_requires_queue_ptr33; CHECK: .amdhsa_user_sgpr_queue_ptr 134 35 36; On gfx8, the queue ptr is required for this addrspacecast.37; CHECK: - .args:38; CHECK-NOT: hidden_queue_ptr39; CHECK-LABEL: .name: addrspacecast_requires_queue_ptr40define amdgpu_kernel void @addrspacecast_requires_queue_ptr(ptr addrspace(5) %ptr.private, ptr addrspace(3) %ptr.local) {41 %flat.private = addrspacecast ptr addrspace(5) %ptr.private to ptr42 %flat.local = addrspacecast ptr addrspace(3) %ptr.local to ptr43 store volatile i32 1, ptr %flat.private44 store volatile i32 2, ptr %flat.local45 ret void46}47 48 49; CHECK: - .args:50; CHECK-NOT: hidden_shared_base51; CHECK-LABEL: .name: is_shared_requires_queue_ptr52define amdgpu_kernel void @is_shared_requires_queue_ptr(ptr %ptr) {53 %is.shared = call i1 @llvm.amdgcn.is.shared(ptr %ptr)54 %zext = zext i1 %is.shared to i3255 store volatile i32 %zext, ptr addrspace(1) poison56 ret void57}58 59; CHECK: - .args:60; CHECK-NOT: hidden_shared_base61; CHECK-NOT: hidden_private_base62; CHECK-NOT: hidden_queue_ptr63; CHECK-LABEL: .name: is_private_requires_queue_ptr64define amdgpu_kernel void @is_private_requires_queue_ptr(ptr %ptr) {65 %is.private = call i1 @llvm.amdgcn.is.private(ptr %ptr)66 %zext = zext i1 %is.private to i3267 store volatile i32 %zext, ptr addrspace(1) poison68 ret void69}70 71; CHECK: - .args:72 73; CHECK-NOT: hidden_shared_base74; CHECK-NOT: hidden_private_base75; CHECK-NOT: hidden_queue_ptr76; CHECK-LABEL: .name: trap_requires_queue_ptr77define amdgpu_kernel void @trap_requires_queue_ptr() {78 call void @llvm.trap()79 unreachable80}81 82; CHECK: - .args:83 84; CHECK-NOT: hidden_shared_base85; CHECK-NOT: hidden_private_base86; CHECK-NOT: hidden_queue_ptr87; CHECK-LABEL: .name: debugtrap_requires_queue_ptr88define amdgpu_kernel void @debugtrap_requires_queue_ptr() {89 call void @llvm.debugtrap()90 unreachable91}92 93; CHECK: - .args:94 95; CHECK-NOT: hidden_shared_base96; CHECK-NOT: hidden_private_base97; CHECK-NOT: hidden_queue_ptr98; CHECK-LABEL: .name: ubsantrap_requires_queue_ptr99define amdgpu_kernel void @ubsantrap_requires_queue_ptr() {100 call void @llvm.ubsantrap(i8 0)101 unreachable102}103 104; CHECK: - .args:105; CHECK-NOT: hidden_queue_ptr106; CHECK-NOT: hidden_shared_base107; CHECK-NOT: hidden_private_base108; CHECK-LABEL: .name: amdgcn_queue_ptr_requires_queue_ptr109define amdgpu_kernel void @amdgcn_queue_ptr_requires_queue_ptr(ptr addrspace(1) %ptr) {110 %queue.ptr = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()111 %implicitarg.ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()112 %dispatch.ptr = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()113 %dispatch.id = call i64 @llvm.amdgcn.dispatch.id()114 %queue.load = load volatile i8, ptr addrspace(4) %queue.ptr115 %implicitarg.load = load volatile i8, ptr addrspace(4) %implicitarg.ptr116 %dispatch.load = load volatile i8, ptr addrspace(4) %dispatch.ptr117 store volatile i64 %dispatch.id, ptr addrspace(1) %ptr118 ret void119}120 121declare noalias ptr addrspace(4) @llvm.amdgcn.queue.ptr()122declare noalias ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()123declare i64 @llvm.amdgcn.dispatch.id()124declare noalias ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()125declare i1 @llvm.amdgcn.is.shared(ptr)126declare i1 @llvm.amdgcn.is.private(ptr)127declare void @llvm.trap()128declare void @llvm.debugtrap()129declare void @llvm.ubsantrap(i8 immarg)130 131!llvm.module.flags = !{!0}132!0 = !{i32 1, !"amdhsa_code_object_version", i32 400}133