198 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; gfx8 required knowing no overflow happened to fold the addressing mode3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 < %s | FileCheck -check-prefix=GFX8 %s4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s5 6define void @gep_noflags_alloca(i32 %idx, i32 %val) #0 {7; GFX8-LABEL: gep_noflags_alloca:8; GFX8: ; %bb.0:9; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)10; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v011; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s3212; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v213; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v014; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen15; GFX8-NEXT: s_waitcnt vmcnt(0)16; GFX8-NEXT: s_setpc_b64 s[30:31]17;18; GFX9-LABEL: gep_noflags_alloca:19; GFX9: ; %bb.0:20; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)21; GFX9-NEXT: s_lshr_b32 s4, s32, 622; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, s423; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:1624; GFX9-NEXT: s_waitcnt vmcnt(0)25; GFX9-NEXT: s_setpc_b64 s[30:31]26 %alloca = alloca [32 x i32], addrspace(5)27 %gep0 = getelementptr [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx28 %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 429 store volatile i32 %val, ptr addrspace(5) %gep130 ret void31}32 33define void @gep_inbounds_alloca(i32 %idx, i32 %val) #0 {34; GFX8-LABEL: gep_inbounds_alloca:35; GFX8: ; %bb.0:36; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)37; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v038; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s3239; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v240; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v041; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen42; GFX8-NEXT: s_waitcnt vmcnt(0)43; GFX8-NEXT: s_setpc_b64 s[30:31]44;45; GFX9-LABEL: gep_inbounds_alloca:46; GFX9: ; %bb.0:47; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)48; GFX9-NEXT: s_lshr_b32 s4, s32, 649; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, s450; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:1651; GFX9-NEXT: s_waitcnt vmcnt(0)52; GFX9-NEXT: s_setpc_b64 s[30:31]53 %alloca = alloca [32 x i32], addrspace(5)54 %gep0 = getelementptr inbounds [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx55 %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 456 store volatile i32 %val, ptr addrspace(5) %gep157 ret void58}59 60define void @gep_nuw_alloca(i32 %idx, i32 %val) #0 {61; GFX8-LABEL: gep_nuw_alloca:62; GFX8: ; %bb.0:63; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)64; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v065; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s3266; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v267; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v068; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen69; GFX8-NEXT: s_waitcnt vmcnt(0)70; GFX8-NEXT: s_setpc_b64 s[30:31]71;72; GFX9-LABEL: gep_nuw_alloca:73; GFX9: ; %bb.0:74; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)75; GFX9-NEXT: s_lshr_b32 s4, s32, 676; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, s477; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:1678; GFX9-NEXT: s_waitcnt vmcnt(0)79; GFX9-NEXT: s_setpc_b64 s[30:31]80 %alloca = alloca [32 x i32], addrspace(5)81 %gep0 = getelementptr nuw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx82 %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 483 store volatile i32 %val, ptr addrspace(5) %gep184 ret void85}86 87define void @gep_nusw_alloca(i32 %idx, i32 %val) #0 {88; GFX8-LABEL: gep_nusw_alloca:89; GFX8: ; %bb.0:90; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)91; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v092; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s3293; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v294; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v095; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen96; GFX8-NEXT: s_waitcnt vmcnt(0)97; GFX8-NEXT: s_setpc_b64 s[30:31]98;99; GFX9-LABEL: gep_nusw_alloca:100; GFX9: ; %bb.0:101; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)102; GFX9-NEXT: s_lshr_b32 s4, s32, 6103; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, s4104; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16105; GFX9-NEXT: s_waitcnt vmcnt(0)106; GFX9-NEXT: s_setpc_b64 s[30:31]107 %alloca = alloca [32 x i32], addrspace(5)108 %gep0 = getelementptr nusw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx109 %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4110 store volatile i32 %val, ptr addrspace(5) %gep1111 ret void112}113 114define void @gep_inbounds_nuw_alloca(i32 %idx, i32 %val) #0 {115; GFX8-LABEL: gep_inbounds_nuw_alloca:116; GFX8: ; %bb.0:117; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)118; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0119; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32120; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2121; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0122; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen123; GFX8-NEXT: s_waitcnt vmcnt(0)124; GFX8-NEXT: s_setpc_b64 s[30:31]125;126; GFX9-LABEL: gep_inbounds_nuw_alloca:127; GFX9: ; %bb.0:128; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)129; GFX9-NEXT: s_lshr_b32 s4, s32, 6130; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, s4131; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16132; GFX9-NEXT: s_waitcnt vmcnt(0)133; GFX9-NEXT: s_setpc_b64 s[30:31]134 %alloca = alloca [32 x i32], addrspace(5)135 %gep0 = getelementptr inbounds nuw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx136 %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4137 store volatile i32 %val, ptr addrspace(5) %gep1138 ret void139}140 141define void @gep_nusw_nuw_alloca(i32 %idx, i32 %val) #0 {142; GFX8-LABEL: gep_nusw_nuw_alloca:143; GFX8: ; %bb.0:144; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)145; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0146; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32147; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2148; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0149; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen150; GFX8-NEXT: s_waitcnt vmcnt(0)151; GFX8-NEXT: s_setpc_b64 s[30:31]152;153; GFX9-LABEL: gep_nusw_nuw_alloca:154; GFX9: ; %bb.0:155; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)156; GFX9-NEXT: s_lshr_b32 s4, s32, 6157; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, s4158; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16159; GFX9-NEXT: s_waitcnt vmcnt(0)160; GFX9-NEXT: s_setpc_b64 s[30:31]161 %alloca = alloca [32 x i32], addrspace(5)162 %gep0 = getelementptr nusw nuw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx163 %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4164 store volatile i32 %val, ptr addrspace(5) %gep1165 ret void166}167 168define void @gep_inbounds_nuw_alloca_nonpow2_scale(i32 %idx, i32 %val) #0 {169; GFX8-LABEL: gep_inbounds_nuw_alloca_nonpow2_scale:170; GFX8: ; %bb.0:171; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)172; GFX8-NEXT: s_movk_i32 s4, 0x84173; GFX8-NEXT: v_mul_lo_u32 v0, v0, s4174; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32175; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2176; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0177; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen178; GFX8-NEXT: s_waitcnt vmcnt(0)179; GFX8-NEXT: s_setpc_b64 s[30:31]180;181; GFX9-LABEL: gep_inbounds_nuw_alloca_nonpow2_scale:182; GFX9: ; %bb.0:183; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)184; GFX9-NEXT: s_movk_i32 s4, 0x84185; GFX9-NEXT: v_mul_lo_u32 v0, v0, s4186; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32187; GFX9-NEXT: v_add_u32_e32 v0, v0, v2188; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16189; GFX9-NEXT: s_waitcnt vmcnt(0)190; GFX9-NEXT: s_setpc_b64 s[30:31]191 %alloca = alloca [5 x [33 x i32]], align 4, addrspace(5)192 %gep1 = getelementptr inbounds nuw [5 x [33 x i32]], ptr addrspace(5) %alloca, i32 0, i32 %idx, i32 4193 store volatile i32 %val, ptr addrspace(5) %gep1, align 4194 ret void195}196 197attributes #0 = { nounwind }198