172 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 < %s | FileCheck --check-prefix=GFX12 %s3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -mattr=+cumode < %s | FileCheck --check-prefix=GFX12 %s4 5 6define amdgpu_kernel void @buffer_last_use_load_0(ptr addrspace(7) %in, ptr addrspace(7) %out) {7; GFX12-LABEL: buffer_last_use_load_0:8; GFX12: ; %bb.0: ; %entry9; GFX12-NEXT: s_clause 0x110; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x011; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x1012; GFX12-NEXT: s_mov_b32 s12, 013; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)14; GFX12-NEXT: s_mov_b32 s7, s1215; GFX12-NEXT: s_mov_b32 s9, s1216; GFX12-NEXT: s_wait_kmcnt 0x017; GFX12-NEXT: s_mov_b32 s6, s318; GFX12-NEXT: s_mov_b32 s8, s119; GFX12-NEXT: s_or_b64 s[10:11], s[6:7], s[12:13]20; GFX12-NEXT: s_mov_b32 s13, s221; GFX12-NEXT: v_mov_b32_e32 v0, s022; GFX12-NEXT: s_or_b64 s[8:9], s[8:9], s[12:13]23; GFX12-NEXT: s_clause 0x124; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x3025; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x2026; GFX12-NEXT: s_mov_b32 s5, s1227; GFX12-NEXT: s_wait_kmcnt 0x028; GFX12-NEXT: v_mov_b32_e32 v1, s029; GFX12-NEXT: buffer_load_b32 v0, v0, s[8:11], null offen th:TH_LOAD_LU30; GFX12-NEXT: s_mov_b32 s4, s331; GFX12-NEXT: s_mov_b32 s3, s1232; GFX12-NEXT: s_or_b64 s[6:7], s[4:5], s[12:13]33; GFX12-NEXT: s_mov_b32 s13, s234; GFX12-NEXT: s_mov_b32 s2, s135; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)36; GFX12-NEXT: s_or_b64 s[4:5], s[2:3], s[12:13]37; GFX12-NEXT: s_wait_loadcnt 0x038; GFX12-NEXT: buffer_store_b32 v0, v1, s[4:7], null offen39; GFX12-NEXT: s_endpgm40entry:41 %val = load i32, ptr addrspace(7) %in, !amdgpu.last.use !{}42 store i32 %val, ptr addrspace(7) %out43 ret void44}45 46define amdgpu_kernel void @buffer_last_use_load_1(ptr addrspace(7) %in, ptr addrspace(7) %out) {47; GFX12-LABEL: buffer_last_use_load_1:48; GFX12: ; %bb.0: ; %entry49; GFX12-NEXT: s_clause 0x150; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x051; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x1052; GFX12-NEXT: v_and_b32_e32 v0, 0x3ff, v053; GFX12-NEXT: s_mov_b32 s12, 054; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)55; GFX12-NEXT: s_mov_b32 s7, s1256; GFX12-NEXT: s_mov_b32 s9, s1257; GFX12-NEXT: s_wait_kmcnt 0x058; GFX12-NEXT: s_mov_b32 s6, s359; GFX12-NEXT: v_lshl_add_u32 v0, v0, 2, s060; GFX12-NEXT: s_mov_b32 s8, s161; GFX12-NEXT: s_or_b64 s[10:11], s[6:7], s[12:13]62; GFX12-NEXT: s_mov_b32 s13, s263; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)64; GFX12-NEXT: s_or_b64 s[8:9], s[8:9], s[12:13]65; GFX12-NEXT: s_clause 0x166; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x3067; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x2068; GFX12-NEXT: buffer_load_b32 v0, v0, s[8:11], null offen th:TH_LOAD_LU69; GFX12-NEXT: s_mov_b32 s5, s1270; GFX12-NEXT: s_wait_kmcnt 0x071; GFX12-NEXT: s_mov_b32 s4, s372; GFX12-NEXT: v_mov_b32_e32 v1, s073; GFX12-NEXT: s_or_b64 s[6:7], s[4:5], s[12:13]74; GFX12-NEXT: s_mov_b32 s13, s275; GFX12-NEXT: s_mov_b32 s2, s176; GFX12-NEXT: s_mov_b32 s3, s1277; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)78; GFX12-NEXT: s_or_b64 s[4:5], s[2:3], s[12:13]79; GFX12-NEXT: s_wait_loadcnt 0x080; GFX12-NEXT: buffer_store_b32 v0, v1, s[4:7], null offen81; GFX12-NEXT: s_endpgm82entry:83 %tid = call i32 @llvm.amdgcn.workitem.id.x()84 %val.gep = getelementptr inbounds i32, ptr addrspace(7) %in, i32 %tid85 %val = load i32, ptr addrspace(7) %val.gep, align 4, !amdgpu.last.use !{}86 store i32 %val, ptr addrspace(7) %out87 ret void88}89 90define amdgpu_kernel void @buffer_last_use_and_volatile_load(ptr addrspace(7) %in, ptr addrspace(7) %out) {91; GFX12-LABEL: buffer_last_use_and_volatile_load:92; GFX12: ; %bb.0: ; %entry93; GFX12-NEXT: s_clause 0x194; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x095; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x1096; GFX12-NEXT: s_mov_b32 s12, 097; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)98; GFX12-NEXT: s_mov_b32 s7, s1299; GFX12-NEXT: s_mov_b32 s9, s12100; GFX12-NEXT: s_wait_kmcnt 0x0101; GFX12-NEXT: s_mov_b32 s6, s3102; GFX12-NEXT: s_mov_b32 s8, s1103; GFX12-NEXT: s_or_b64 s[10:11], s[6:7], s[12:13]104; GFX12-NEXT: s_mov_b32 s13, s2105; GFX12-NEXT: v_mov_b32_e32 v0, s0106; GFX12-NEXT: s_or_b64 s[8:9], s[8:9], s[12:13]107; GFX12-NEXT: s_clause 0x1108; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x30109; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x20110; GFX12-NEXT: s_mov_b32 s5, s12111; GFX12-NEXT: s_wait_kmcnt 0x0112; GFX12-NEXT: v_mov_b32_e32 v1, s0113; GFX12-NEXT: buffer_load_b32 v0, v0, s[8:11], null offen th:TH_LOAD_BYPASS scope:SCOPE_SYS114; GFX12-NEXT: s_wait_loadcnt 0x0115; GFX12-NEXT: s_mov_b32 s4, s3116; GFX12-NEXT: s_mov_b32 s3, s12117; GFX12-NEXT: s_or_b64 s[6:7], s[4:5], s[12:13]118; GFX12-NEXT: s_mov_b32 s13, s2119; GFX12-NEXT: s_mov_b32 s2, s1120; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)121; GFX12-NEXT: s_or_b64 s[4:5], s[2:3], s[12:13]122; GFX12-NEXT: buffer_store_b32 v0, v1, s[4:7], null offen123; GFX12-NEXT: s_endpgm124entry:125 %val = load volatile i32, ptr addrspace(7) %in, !amdgpu.last.use !{}126 store i32 %val, ptr addrspace(7) %out127 ret void128}129 130define amdgpu_kernel void @buffer_last_use_and_nontemporal_load(ptr addrspace(7) %in, ptr addrspace(7) %out) {131; GFX12-LABEL: buffer_last_use_and_nontemporal_load:132; GFX12: ; %bb.0: ; %entry133; GFX12-NEXT: s_clause 0x1134; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x0135; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x10136; GFX12-NEXT: s_mov_b32 s12, 0137; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)138; GFX12-NEXT: s_mov_b32 s7, s12139; GFX12-NEXT: s_mov_b32 s9, s12140; GFX12-NEXT: s_wait_kmcnt 0x0141; GFX12-NEXT: s_mov_b32 s6, s3142; GFX12-NEXT: s_mov_b32 s8, s1143; GFX12-NEXT: s_or_b64 s[10:11], s[6:7], s[12:13]144; GFX12-NEXT: s_mov_b32 s13, s2145; GFX12-NEXT: v_mov_b32_e32 v0, s0146; GFX12-NEXT: s_or_b64 s[8:9], s[8:9], s[12:13]147; GFX12-NEXT: s_clause 0x1148; GFX12-NEXT: s_load_b32 s13, s[4:5], 0x30149; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x20150; GFX12-NEXT: s_mov_b32 s5, s12151; GFX12-NEXT: s_wait_kmcnt 0x0152; GFX12-NEXT: v_mov_b32_e32 v1, s0153; GFX12-NEXT: buffer_load_b32 v0, v0, s[8:11], null offen th:TH_LOAD_LU154; GFX12-NEXT: s_mov_b32 s4, s3155; GFX12-NEXT: s_mov_b32 s3, s12156; GFX12-NEXT: s_or_b64 s[6:7], s[4:5], s[12:13]157; GFX12-NEXT: s_mov_b32 s13, s2158; GFX12-NEXT: s_mov_b32 s2, s1159; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)160; GFX12-NEXT: s_or_b64 s[4:5], s[2:3], s[12:13]161; GFX12-NEXT: s_wait_loadcnt 0x0162; GFX12-NEXT: buffer_store_b32 v0, v1, s[4:7], null offen163; GFX12-NEXT: s_endpgm164entry:165 %val = load i32, ptr addrspace(7) %in, !amdgpu.last.use !{}, !nontemporal !0166 store i32 %val, ptr addrspace(7) %out167 ret void168}169 170!0 = !{i32 1}171declare i32 @llvm.amdgcn.workitem.id.x()172