67 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s2; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GCN,GFX10PLUS,GFX10 %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 < %s | FileCheck -check-prefixes=GCN,GFX10PLUS,GFX11 %s4 5; GCN-LABEL: {{^}}test_add_lit:6; GFX10PLUS: v_add_co_u32{{(_e64)?}} v{{[0-9]+}}, vcc_lo, 0x80992bff, v{{[0-9]+}}7; GFX11: v_add_co_ci_u32_e64 v{{[0-9]+}}, null, 0xe7, v{{[0-9]+}}, vcc_lo8; GFX10: v_add_co_ci_u32_e32 v{{[0-9]+}}, vcc_lo, 0xe7, v{{[0-9]+}}, vcc_lo9; GFX9: v_mov_b32_e32 [[C2:v[0-9]+]], 0xe710; GFX9: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0x80992bff, v{{[0-9]+}}11; GFX9: v_addc_co_u32_e32 v{{[0-9]+}}, vcc, v{{[0-9]+}}, [[C2]], vcc12define amdgpu_kernel void @test_add_lit(ptr addrspace(1) %p) {13 %id = tail call i32 @llvm.amdgcn.workitem.id.x()14 %ptr = getelementptr inbounds i64, ptr addrspace(1) %p, i32 %id15 %load = load i64, ptr addrspace(1) %ptr, align 816 %add = add nsw i64 %load, 99429496729517 store i64 %add, ptr addrspace(1) %ptr, align 818 ret void19}20 21; GCN-LABEL: {{^}}test_cndmask_lit:22; GFX10PLUS: v_cndmask_b32_e32 v{{[0-9]+}}, 0x3039, v{{[0-9]+}}, vcc_lo23; GFX9: v_mov_b32_e32 [[C:v[0-9]+]], 0x303924; GFX9: v_cndmask_b32_e32 v{{[0-9]+}}, [[C]], v{{[0-9]+}}, vcc25define amdgpu_kernel void @test_cndmask_lit(ptr addrspace(1) %p) {26 %id = tail call i32 @llvm.amdgcn.workitem.id.x()27 %n = add nuw nsw i32 %id, 128 %p1 = getelementptr inbounds i32, ptr addrspace(1) %p, i32 %id29 %v1 = load i32, ptr addrspace(1) %p1, align 430 %p2 = getelementptr inbounds i32, ptr addrspace(1) %p, i32 %n31 %v2 = load i32, ptr addrspace(1) %p2, align 432 %cmp = icmp sgt i32 %v1, 033 %sel = select i1 %cmp, i32 12345, i32 %v234 store i32 %sel, ptr addrspace(1) %p1, align 435 ret void36}37 38; GCN-LABEL: {{^}}test_bfe_2lit_s:39; GFX10PLUS: v_mov_b32_e32 [[C1:v[0-9]+]], 0xddd540; GFX10PLUS: v_bfe_u32 v{{[0-9]+}}, 0x3039, s{{[0-9]+}}, [[C1]]41; GFX9-DAG: v_mov_b32_e32 [[C2:v[0-9]+]], 0xddd542; GFX9-DAG: s_movk_i32 [[C1:s[0-9]+]], 0x303943; GFX9: v_bfe_u32 v{{[0-9]+}}, [[C1]], v{{[0-9]+}}, [[C2]]44define amdgpu_kernel void @test_bfe_2lit_s(ptr addrspace(1) %p, i32 %src) {45 %bfe = call i32 @llvm.amdgcn.ubfe.i32(i32 12345, i32 %src, i32 56789)46 store i32 %bfe, ptr addrspace(1) %p, align 447 ret void48}49 50; GCN-LABEL: {{^}}test_bfe_2lit_v:51; GFX10PLUS: s_movk_i32 [[C1:s[0-9]+]], 0x303952; GFX10PLUS: v_bfe_u32 v{{[0-9]+}}, [[C1]], v{{[0-9]+}}, 0xddd553; GFX9-DAG: v_mov_b32_e32 [[C2:v[0-9]+]], 0xddd554; GFX9-DAG: s_movk_i32 [[C1:s[0-9]+]], 0x303955; GFX9: v_bfe_u32 v{{[0-9]+}}, [[C1]], v{{[0-9]+}}, [[C2]]56define amdgpu_kernel void @test_bfe_2lit_v(ptr addrspace(1) %p) {57 %id = tail call i32 @llvm.amdgcn.workitem.id.x()58 %ptr = getelementptr inbounds i32, ptr addrspace(1) %p, i32 %id59 %load = load i32, ptr addrspace(1) %ptr, align 460 %bfe = call i32 @llvm.amdgcn.ubfe.i32(i32 12345, i32 %load, i32 56789)61 store i32 %bfe, ptr addrspace(1) %ptr, align 462 ret void63}64 65declare i32 @llvm.amdgcn.workitem.id.x()66declare i32 @llvm.amdgcn.ubfe.i32(i32, i32, i32)67