111 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; xUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -filetype=obj < %t.gfx7.bc | llvm-readelf --notes - | FileCheck --check-prefixes=CHECK,PRE-GFX9 %s5; xUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -filetype=obj < %t.gfx8.bc | llvm-readelf --notes - | FileCheck --check-prefixes=CHECK,PRE-GFX9 %s6; xUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -filetype=obj < %t.gfx9.bc | llvm-readelf --notes - | FileCheck --check-prefixes=CHECK,GFX9 %s7; xUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 < %t.gfx7.bc | FileCheck --check-prefixes=CHECK,PRE-GFX9 %s8; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 < %t.gfx8.bc | FileCheck --check-prefixes=CHECK,PRE-GFX9 %s9; xUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %t.gfx9.bc | FileCheck --check-prefixes=CHECK,GFX9 %s10 11 12; On gfx8, the queue ptr is required for this addrspacecast on cov413; CHECK: - .args:14 15; PRE-GFX9: .value_kind: hidden_private_base16; PRE-GFX9: .value_kind: hidden_shared_base17 18; GFX9-NOT: hidden_queue_ptr19; GFX9-NOT: hidden_private_base20; GFX9-NOT: hidden_shared_base21 22; CHECK-LABEL: .name: addrspacecast_requires_queue_ptr23define amdgpu_kernel void @addrspacecast_requires_queue_ptr(ptr addrspace(5) %ptr.private, ptr addrspace(3) %ptr.local) {24 %flat.private = addrspacecast ptr addrspace(5) %ptr.private to ptr25 %flat.local = addrspacecast ptr addrspace(3) %ptr.local to ptr26 store volatile i32 1, ptr %flat.private27 store volatile i32 2, ptr %flat.local28 ret void29}30 31; CHECK: - .args:32; PRE-GFX9: hidden_shared_base33; GFX9-NOT: hidden_shared_base34; CHECK-LABEL: .name: is_shared_requires_queue_ptr35define amdgpu_kernel void @is_shared_requires_queue_ptr(ptr %ptr) {36 %is.shared = call i1 @llvm.amdgcn.is.shared(ptr %ptr)37 %zext = zext i1 %is.shared to i3238 store volatile i32 %zext, ptr addrspace(1) poison39 ret void40}41 42; CHECK: - .args:43; PRE-GFX9: hidden_private_base44; GFX9-NOT: hidden_private_base45; CHECK-LABEL: .name: is_private_requires_queue_ptr46define amdgpu_kernel void @is_private_requires_queue_ptr(ptr %ptr) {47 %is.private = call i1 @llvm.amdgcn.is.private(ptr %ptr)48 %zext = zext i1 %is.private to i3249 store volatile i32 %zext, ptr addrspace(1) poison50 ret void51}52 53; CHECK: - .args:54 55; PRE-GFX9: hidden_queue_ptr56; GFX9-NOT: hidden_queue_ptr57; CHECK-LABEL: .name: trap_requires_queue_ptr58define amdgpu_kernel void @trap_requires_queue_ptr() {59 call void @llvm.trap()60 unreachable61}62 63; CHECK: - .args:64 65; PRE-GFX9: hidden_queue_ptr66; GFX9-NOT: hidden_queue_ptr67; CHECK-LABEL: .name: debugtrap_requires_queue_ptr68define amdgpu_kernel void @debugtrap_requires_queue_ptr() {69 call void @llvm.debugtrap()70 unreachable71}72 73; CHECK: - .args:74 75; PRE-GFX9: hidden_queue_ptr76; GFX9-NOT: hidden_queue_ptr77; CHECK-LABEL: .name: ubsantrap_requires_queue_ptr78define amdgpu_kernel void @ubsantrap_requires_queue_ptr() {79 call void @llvm.ubsantrap(i8 0)80 unreachable81}82 83; CHECK: - .args:84; CHECK: hidden_queue_ptr85; CHECK-LABEL: .name: amdgcn_queue_ptr_requires_queue_ptr86define amdgpu_kernel void @amdgcn_queue_ptr_requires_queue_ptr(ptr addrspace(1) %ptr) {87 %queue.ptr = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()88 %implicitarg.ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()89 %dispatch.ptr = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()90 %dispatch.id = call i64 @llvm.amdgcn.dispatch.id()91 %queue.load = load volatile i8, ptr addrspace(4) %queue.ptr92 %implicitarg.load = load volatile i8, ptr addrspace(4) %implicitarg.ptr93 %dispatch.load = load volatile i8, ptr addrspace(4) %dispatch.ptr94 store volatile i64 %dispatch.id, ptr addrspace(1) %ptr95 ret void96}97 98 99declare noalias ptr addrspace(4) @llvm.amdgcn.queue.ptr()100declare noalias ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()101declare i64 @llvm.amdgcn.dispatch.id()102declare noalias ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()103declare i1 @llvm.amdgcn.is.shared(ptr)104declare i1 @llvm.amdgcn.is.private(ptr)105declare void @llvm.trap()106declare void @llvm.debugtrap()107declare void @llvm.ubsantrap(i8 immarg)108 109!llvm.module.flags = !{!0}110!0 = !{i32 1, !"amdhsa_code_object_version", i32 500}111