133 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck %s -check-prefix=GFX9423 4define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat(ptr %ptr) {5; GFX942-LABEL: flat_atomic_fadd_f32_noret_pat:6; GFX942: ; %bb.0:7; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x248; GFX942-NEXT: v_mov_b32_e32 v2, 4.09; GFX942-NEXT: s_waitcnt lgkmcnt(0)10; GFX942-NEXT: v_mov_b64_e32 v[0:1], s[0:1]11; GFX942-NEXT: buffer_wbl2 sc0 sc112; GFX942-NEXT: flat_atomic_add_f32 v[0:1], v2 sc113; GFX942-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)14; GFX942-NEXT: buffer_inv sc0 sc115; GFX942-NEXT: s_endpgm16 %ret = atomicrmw fadd ptr %ptr, float 4.0 seq_cst, !amdgpu.no.remote.memory !017 ret void18}19 20define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat_ieee(ptr %ptr) #0 {21; GFX942-LABEL: flat_atomic_fadd_f32_noret_pat_ieee:22; GFX942: ; %bb.0:23; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2424; GFX942-NEXT: v_mov_b32_e32 v2, 4.025; GFX942-NEXT: s_waitcnt lgkmcnt(0)26; GFX942-NEXT: v_mov_b64_e32 v[0:1], s[0:1]27; GFX942-NEXT: buffer_wbl2 sc0 sc128; GFX942-NEXT: flat_atomic_add_f32 v[0:1], v2 sc129; GFX942-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)30; GFX942-NEXT: buffer_inv sc0 sc131; GFX942-NEXT: s_endpgm32 %ret = atomicrmw fadd ptr %ptr, float 4.0 seq_cst, !amdgpu.no.remote.memory !033 ret void34}35 36define float @flat_atomic_fadd_f32_rtn_pat(ptr %ptr, float %data) {37; GFX942-LABEL: flat_atomic_fadd_f32_rtn_pat:38; GFX942: ; %bb.0:39; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)40; GFX942-NEXT: v_mov_b32_e32 v2, 4.041; GFX942-NEXT: buffer_wbl2 sc0 sc142; GFX942-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 sc0 sc143; GFX942-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)44; GFX942-NEXT: buffer_inv sc0 sc145; GFX942-NEXT: s_setpc_b64 s[30:31]46 %ret = atomicrmw fadd ptr %ptr, float 4.0 seq_cst, !amdgpu.no.remote.memory !047 ret float %ret48}49 50define <2 x half> @local_atomic_fadd_ret_v2f16_offset(ptr addrspace(3) %ptr, <2 x half> %val) {51; GFX942-LABEL: local_atomic_fadd_ret_v2f16_offset:52; GFX942: ; %bb.0:53; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)54; GFX942-NEXT: ds_pk_add_rtn_f16 v0, v0, v1 offset:6553255; GFX942-NEXT: s_waitcnt lgkmcnt(0)56; GFX942-NEXT: s_setpc_b64 s[30:31]57 %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 1638358 %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst59 ret <2 x half> %result60}61 62define void @local_atomic_fadd_noret_v2f16_offset(ptr addrspace(3) %ptr, <2 x half> %val) {63; GFX942-LABEL: local_atomic_fadd_noret_v2f16_offset:64; GFX942: ; %bb.0:65; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)66; GFX942-NEXT: ds_pk_add_f16 v0, v1 offset:6553267; GFX942-NEXT: s_waitcnt lgkmcnt(0)68; GFX942-NEXT: s_setpc_b64 s[30:31]69 %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 1638370 %unused = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst71 ret void72}73 74define <2 x half> @global_atomic_fadd_ret_v2f16_agent_offset(ptr addrspace(1) %ptr, <2 x half> %val) {75; GFX942-LABEL: global_atomic_fadd_ret_v2f16_agent_offset:76; GFX942: ; %bb.0:77; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)78; GFX942-NEXT: buffer_wbl2 sc179; GFX942-NEXT: global_atomic_pk_add_f16 v0, v[0:1], v2, off offset:1024 sc080; GFX942-NEXT: s_waitcnt vmcnt(0)81; GFX942-NEXT: buffer_inv sc182; GFX942-NEXT: s_setpc_b64 s[30:31]83 %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 25684 %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst85 ret <2 x half> %result86}87 88define void @global_atomic_fadd_noret_v2f16_agent_offset(ptr addrspace(1) %ptr, <2 x half> %val) {89; GFX942-LABEL: global_atomic_fadd_noret_v2f16_agent_offset:90; GFX942: ; %bb.0:91; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)92; GFX942-NEXT: buffer_wbl2 sc193; GFX942-NEXT: global_atomic_pk_add_f16 v[0:1], v2, off offset:102494; GFX942-NEXT: s_waitcnt vmcnt(0)95; GFX942-NEXT: buffer_inv sc196; GFX942-NEXT: s_setpc_b64 s[30:31]97 %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 25698 %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst99 ret void100}101 102define <2 x half> @flat_atomic_fadd_ret_v2f16_agent_offset(ptr %ptr, <2 x half> %val) {103; GFX942-LABEL: flat_atomic_fadd_ret_v2f16_agent_offset:104; GFX942: ; %bb.0:105; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)106; GFX942-NEXT: buffer_wbl2 sc1107; GFX942-NEXT: flat_atomic_pk_add_f16 v0, v[0:1], v2 offset:1024 sc0108; GFX942-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)109; GFX942-NEXT: buffer_inv sc1110; GFX942-NEXT: s_setpc_b64 s[30:31]111 %gep = getelementptr inbounds <2 x half>, ptr %ptr, i32 256112 %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst113 ret <2 x half> %result114}115 116define void @flat_atomic_fadd_noret_v2f16_agent_offset(ptr %ptr, <2 x half> %val) {117; GFX942-LABEL: flat_atomic_fadd_noret_v2f16_agent_offset:118; GFX942: ; %bb.0:119; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)120; GFX942-NEXT: buffer_wbl2 sc1121; GFX942-NEXT: flat_atomic_pk_add_f16 v[0:1], v2 offset:1024122; GFX942-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)123; GFX942-NEXT: buffer_inv sc1124; GFX942-NEXT: s_setpc_b64 s[30:31]125 %gep = getelementptr inbounds <2 x half>, ptr %ptr, i32 256126 %unused = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst127 ret void128}129 130attributes #0 = { "denormal-fp-math-f32"="ieee,ieee" }131 132!0 = !{}133