1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx940 < %s | FileCheck %s -check-prefix=GFX940 3 4define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat(ptr %ptr) { 5; GFX940-LABEL: flat_atomic_fadd_f32_noret_pat: 6; GFX940: ; %bb.0: 7; GFX940-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 8; GFX940-NEXT: v_mov_b32_e32 v2, 4.0 9; GFX940-NEXT: s_waitcnt lgkmcnt(0) 10; GFX940-NEXT: v_mov_b64_e32 v[0:1], s[0:1] 11; GFX940-NEXT: buffer_wbl2 sc0 sc1 12; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 sc1 13; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 14; GFX940-NEXT: buffer_inv sc0 sc1 15; GFX940-NEXT: s_endpgm 16 %ret = atomicrmw fadd ptr %ptr, float 4.0 seq_cst, !amdgpu.no.remote.memory !0 17 ret void 18} 19 20define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat_ieee(ptr %ptr) #0 { 21; GFX940-LABEL: flat_atomic_fadd_f32_noret_pat_ieee: 22; GFX940: ; %bb.0: 23; GFX940-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 24; GFX940-NEXT: v_mov_b32_e32 v2, 4.0 25; GFX940-NEXT: s_waitcnt lgkmcnt(0) 26; GFX940-NEXT: v_mov_b64_e32 v[0:1], s[0:1] 27; GFX940-NEXT: buffer_wbl2 sc0 sc1 28; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 sc1 29; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 30; GFX940-NEXT: buffer_inv sc0 sc1 31; GFX940-NEXT: s_endpgm 32 %ret = atomicrmw fadd ptr %ptr, float 4.0 seq_cst, !amdgpu.no.remote.memory !0 33 ret void 34} 35 36define float @flat_atomic_fadd_f32_rtn_pat(ptr %ptr, float %data) { 37; GFX940-LABEL: flat_atomic_fadd_f32_rtn_pat: 38; GFX940: ; %bb.0: 39; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 40; GFX940-NEXT: v_mov_b32_e32 v2, 4.0 41; GFX940-NEXT: buffer_wbl2 sc0 sc1 42; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 sc0 sc1 43; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 44; GFX940-NEXT: buffer_inv sc0 sc1 45; GFX940-NEXT: s_setpc_b64 s[30:31] 46 %ret = atomicrmw fadd ptr %ptr, float 4.0 seq_cst, !amdgpu.no.remote.memory !0 47 ret float %ret 48} 49 50define <2 x half> @local_atomic_fadd_ret_v2f16_offset(ptr addrspace(3) %ptr, <2 x half> %val) { 51; GFX940-LABEL: local_atomic_fadd_ret_v2f16_offset: 52; GFX940: ; %bb.0: 53; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 54; GFX940-NEXT: ds_pk_add_rtn_f16 v0, v0, v1 offset:65532 55; GFX940-NEXT: s_waitcnt lgkmcnt(0) 56; GFX940-NEXT: s_setpc_b64 s[30:31] 57 %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383 58 %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst 59 ret <2 x half> %result 60} 61 62define void @local_atomic_fadd_noret_v2f16_offset(ptr addrspace(3) %ptr, <2 x half> %val) { 63; GFX940-LABEL: local_atomic_fadd_noret_v2f16_offset: 64; GFX940: ; %bb.0: 65; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 66; GFX940-NEXT: ds_pk_add_f16 v0, v1 offset:65532 67; GFX940-NEXT: s_waitcnt lgkmcnt(0) 68; GFX940-NEXT: s_setpc_b64 s[30:31] 69 %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383 70 %unused = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst 71 ret void 72} 73 74define <2 x half> @global_atomic_fadd_ret_v2f16_agent_offset(ptr addrspace(1) %ptr, <2 x half> %val) { 75; GFX940-LABEL: global_atomic_fadd_ret_v2f16_agent_offset: 76; GFX940: ; %bb.0: 77; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 78; GFX940-NEXT: buffer_wbl2 sc1 79; GFX940-NEXT: global_atomic_pk_add_f16 v0, v[0:1], v2, off offset:1024 sc0 80; GFX940-NEXT: s_waitcnt vmcnt(0) 81; GFX940-NEXT: buffer_inv sc1 82; GFX940-NEXT: s_setpc_b64 s[30:31] 83 %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 256 84 %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst 85 ret <2 x half> %result 86} 87 88define void @global_atomic_fadd_noret_v2f16_agent_offset(ptr addrspace(1) %ptr, <2 x half> %val) { 89; GFX940-LABEL: global_atomic_fadd_noret_v2f16_agent_offset: 90; GFX940: ; %bb.0: 91; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 92; GFX940-NEXT: buffer_wbl2 sc1 93; GFX940-NEXT: global_atomic_pk_add_f16 v[0:1], v2, off offset:1024 94; GFX940-NEXT: s_waitcnt vmcnt(0) 95; GFX940-NEXT: buffer_inv sc1 96; GFX940-NEXT: s_setpc_b64 s[30:31] 97 %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 256 98 %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst 99 ret void 100} 101 102define <2 x half> @flat_atomic_fadd_ret_v2f16_agent_offset(ptr %ptr, <2 x half> %val) { 103; GFX940-LABEL: flat_atomic_fadd_ret_v2f16_agent_offset: 104; GFX940: ; %bb.0: 105; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 106; GFX940-NEXT: buffer_wbl2 sc1 107; GFX940-NEXT: flat_atomic_pk_add_f16 v0, v[0:1], v2 offset:1024 sc0 108; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 109; GFX940-NEXT: buffer_inv sc1 110; GFX940-NEXT: s_setpc_b64 s[30:31] 111 %gep = getelementptr <2 x half>, ptr %ptr, i32 256 112 %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst 113 ret <2 x half> %result 114} 115 116define void @flat_atomic_fadd_noret_v2f16_agent_offset(ptr %ptr, <2 x half> %val) { 117; GFX940-LABEL: flat_atomic_fadd_noret_v2f16_agent_offset: 118; GFX940: ; %bb.0: 119; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 120; GFX940-NEXT: buffer_wbl2 sc1 121; GFX940-NEXT: flat_atomic_pk_add_f16 v[0:1], v2 offset:1024 122; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 123; GFX940-NEXT: buffer_inv sc1 124; GFX940-NEXT: s_setpc_b64 s[30:31] 125 %gep = getelementptr <2 x half>, ptr %ptr, i32 256 126 %unused = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst 127 ret void 128} 129 130attributes #0 = { "denormal-fp-math-f32"="ieee,ieee" } 131 132!0 = !{} 133