72 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52 3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 %s -o - | FileCheck %s4 5; Testing codegen for memcpy with scalar reads.6 7 8define void @memcpy_p1_p4_sz16_align_4_4(ptr addrspace(1) align 4 %dst, ptr addrspace(4) align 4 readonly inreg %src) {9; CHECK-LABEL: memcpy_p1_p4_sz16_align_4_4:10; CHECK: ; %bb.0: ; %entry11; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)12; CHECK-NEXT: s_load_dwordx4 s[4:7], s[16:17], 0x013; CHECK-NEXT: s_waitcnt lgkmcnt(0)14; CHECK-NEXT: v_mov_b32_e32 v2, s415; CHECK-NEXT: v_mov_b32_e32 v3, s516; CHECK-NEXT: v_mov_b32_e32 v4, s617; CHECK-NEXT: v_mov_b32_e32 v5, s718; CHECK-NEXT: global_store_dwordx4 v[0:1], v[2:5], off19; CHECK-NEXT: s_setpc_b64 s[30:31]20entry:21 tail call void @llvm.memcpy.p1.p4.i64(ptr addrspace(1) noundef nonnull align 4 %dst, ptr addrspace(4) noundef nonnull align 4 %src, i64 16, i1 false)22 ret void23}24 25define void @memcpy_p1_p4_sz31_align_4_4(ptr addrspace(1) align 4 %dst, ptr addrspace(4) align 4 readonly inreg %src) {26; CHECK-LABEL: memcpy_p1_p4_sz31_align_4_4:27; CHECK: ; %bb.0: ; %entry28; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)29; CHECK-NEXT: s_load_dwordx4 s[4:7], s[16:17], 0x030; CHECK-NEXT: v_mov_b32_e32 v6, 031; CHECK-NEXT: s_waitcnt lgkmcnt(0)32; CHECK-NEXT: v_mov_b32_e32 v2, s433; CHECK-NEXT: v_mov_b32_e32 v3, s534; CHECK-NEXT: v_mov_b32_e32 v4, s635; CHECK-NEXT: v_mov_b32_e32 v5, s736; CHECK-NEXT: global_store_dwordx4 v[0:1], v[2:5], off37; CHECK-NEXT: global_load_dwordx4 v[2:5], v6, s[16:17] offset:1538; CHECK-NEXT: s_waitcnt vmcnt(0)39; CHECK-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:1540; CHECK-NEXT: s_setpc_b64 s[30:31]41entry:42 tail call void @llvm.memcpy.p1.p4.i64(ptr addrspace(1) noundef nonnull align 4 %dst, ptr addrspace(4) noundef nonnull align 4 %src, i64 31, i1 false)43 ret void44}45 46define void @memcpy_p1_p4_sz32_align_4_4(ptr addrspace(1) align 4 %dst, ptr addrspace(4) align 4 readonly inreg %src) {47; CHECK-LABEL: memcpy_p1_p4_sz32_align_4_4:48; CHECK: ; %bb.0: ; %entry49; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)50; CHECK-NEXT: s_load_dwordx8 s[4:11], s[16:17], 0x051; CHECK-NEXT: s_waitcnt lgkmcnt(0)52; CHECK-NEXT: v_mov_b32_e32 v2, s453; CHECK-NEXT: v_mov_b32_e32 v3, s554; CHECK-NEXT: v_mov_b32_e32 v4, s655; CHECK-NEXT: v_mov_b32_e32 v5, s756; CHECK-NEXT: v_mov_b32_e32 v6, s857; CHECK-NEXT: v_mov_b32_e32 v7, s958; CHECK-NEXT: v_mov_b32_e32 v8, s1059; CHECK-NEXT: v_mov_b32_e32 v9, s1160; CHECK-NEXT: global_store_dwordx4 v[0:1], v[2:5], off61; CHECK-NEXT: global_store_dwordx4 v[0:1], v[6:9], off offset:1662; CHECK-NEXT: s_setpc_b64 s[30:31]63entry:64 tail call void @llvm.memcpy.p1.p4.i64(ptr addrspace(1) noundef nonnull align 4 %dst, ptr addrspace(4) noundef nonnull align 4 %src, i64 32, i1 false)65 ret void66}67 68declare void @llvm.memcpy.p1.p4.i64(ptr addrspace(1) noalias nocapture writeonly, ptr addrspace(4) noalias nocapture readonly, i64, i1 immarg) #269 70attributes #0 = { nocallback nofree nounwind willreturn memory(argmem: readwrite) }71 72