1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -march=amdgcn -mcpu=gfx600 -amdgpu-bypass-slow-div=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s 3; RUN: llc -march=amdgcn -mcpu=gfx600 -amdgpu-bypass-slow-div=0 -amdgpu-codegenprepare-expand-div64 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN-IR %s 4 5define amdgpu_kernel void @s_test_urem_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 6; GCN-LABEL: s_test_urem_i64: 7; GCN: ; %bb.0: 8; GCN-NEXT: s_load_dwordx2 s[12:13], s[0:1], 0xd 9; GCN-NEXT: v_mov_b32_e32 v2, 0 10; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0x9 11; GCN-NEXT: s_mov_b32 s7, 0xf000 12; GCN-NEXT: s_mov_b32 s6, -1 13; GCN-NEXT: s_waitcnt lgkmcnt(0) 14; GCN-NEXT: v_cvt_f32_u32_e32 v0, s12 15; GCN-NEXT: v_cvt_f32_u32_e32 v1, s13 16; GCN-NEXT: s_sub_u32 s2, 0, s12 17; GCN-NEXT: s_subb_u32 s3, 0, s13 18; GCN-NEXT: s_mov_b32 s4, s8 19; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 20; GCN-NEXT: v_rcp_f32_e32 v0, v0 21; GCN-NEXT: v_mov_b32_e32 v1, 0 22; GCN-NEXT: s_mov_b32 s5, s9 23; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 24; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v0 25; GCN-NEXT: v_trunc_f32_e32 v3, v3 26; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v3 27; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 28; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 29; GCN-NEXT: v_mul_hi_u32 v5, s2, v0 30; GCN-NEXT: v_mul_lo_u32 v4, s2, v3 31; GCN-NEXT: v_mul_lo_u32 v7, s3, v0 32; GCN-NEXT: v_mul_lo_u32 v6, s2, v0 33; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 34; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v7 35; GCN-NEXT: v_mul_hi_u32 v5, v0, v6 36; GCN-NEXT: v_mul_lo_u32 v7, v0, v4 37; GCN-NEXT: v_mul_hi_u32 v9, v0, v4 38; GCN-NEXT: v_mul_lo_u32 v8, v3, v6 39; GCN-NEXT: v_mul_hi_u32 v6, v3, v6 40; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v7 41; GCN-NEXT: v_addc_u32_e32 v7, vcc, v2, v9, vcc 42; GCN-NEXT: v_mul_hi_u32 v9, v3, v4 43; GCN-NEXT: v_mul_lo_u32 v4, v3, v4 44; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 45; GCN-NEXT: v_addc_u32_e32 v5, vcc, v7, v6, vcc 46; GCN-NEXT: v_addc_u32_e32 v6, vcc, v9, v1, vcc 47; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 48; GCN-NEXT: v_add_i32_e64 v0, s[0:1], v0, v4 49; GCN-NEXT: v_addc_u32_e32 v5, vcc, v2, v6, vcc 50; GCN-NEXT: v_addc_u32_e64 v4, vcc, v3, v5, s[0:1] 51; GCN-NEXT: v_mul_lo_u32 v6, s2, v4 52; GCN-NEXT: v_mul_hi_u32 v7, s2, v0 53; GCN-NEXT: v_mul_lo_u32 v8, s3, v0 54; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 55; GCN-NEXT: v_mul_lo_u32 v7, s2, v0 56; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 57; GCN-NEXT: v_mul_lo_u32 v10, v0, v6 58; GCN-NEXT: v_mul_hi_u32 v12, v0, v6 59; GCN-NEXT: v_mul_hi_u32 v11, v0, v7 60; GCN-NEXT: v_mul_hi_u32 v9, v4, v7 61; GCN-NEXT: v_mul_lo_u32 v7, v4, v7 62; GCN-NEXT: v_mul_hi_u32 v8, v4, v6 63; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 64; GCN-NEXT: v_addc_u32_e32 v11, vcc, v2, v12, vcc 65; GCN-NEXT: v_mul_lo_u32 v4, v4, v6 66; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 67; GCN-NEXT: v_addc_u32_e32 v7, vcc, v11, v9, vcc 68; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v1, vcc 69; GCN-NEXT: v_add_i32_e32 v4, vcc, v7, v4 70; GCN-NEXT: v_addc_u32_e32 v6, vcc, v2, v6, vcc 71; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 72; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v6, s[0:1] 73; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4 74; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 75; GCN-NEXT: v_mul_lo_u32 v4, s10, v3 76; GCN-NEXT: v_mul_hi_u32 v5, s10, v0 77; GCN-NEXT: v_mul_hi_u32 v6, s10, v3 78; GCN-NEXT: v_mul_hi_u32 v7, s11, v3 79; GCN-NEXT: v_mul_lo_u32 v3, s11, v3 80; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 81; GCN-NEXT: v_addc_u32_e32 v5, vcc, v2, v6, vcc 82; GCN-NEXT: v_mul_lo_u32 v6, s11, v0 83; GCN-NEXT: v_mul_hi_u32 v0, s11, v0 84; GCN-NEXT: v_add_i32_e32 v4, vcc, v6, v4 85; GCN-NEXT: v_addc_u32_e32 v0, vcc, v5, v0, vcc 86; GCN-NEXT: v_addc_u32_e32 v1, vcc, v7, v1, vcc 87; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v3 88; GCN-NEXT: v_addc_u32_e32 v1, vcc, v2, v1, vcc 89; GCN-NEXT: v_mul_lo_u32 v1, s12, v1 90; GCN-NEXT: v_mul_hi_u32 v2, s12, v0 91; GCN-NEXT: v_mul_lo_u32 v3, s13, v0 92; GCN-NEXT: v_mul_lo_u32 v0, s12, v0 93; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 94; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 95; GCN-NEXT: v_sub_i32_e32 v2, vcc, s11, v1 96; GCN-NEXT: v_mov_b32_e32 v3, s13 97; GCN-NEXT: v_sub_i32_e32 v0, vcc, s10, v0 98; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 99; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s12, v0 100; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 101; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s13, v5 102; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 103; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 104; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s12, v4 105; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s12, v4 106; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 107; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s13, v5 108; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 109; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 110; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 111; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 112; GCN-NEXT: v_mov_b32_e32 v5, s11 113; GCN-NEXT: v_subb_u32_e32 v1, vcc, v5, v1, vcc 114; GCN-NEXT: v_cmp_le_u32_e32 vcc, s13, v1 115; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 116; GCN-NEXT: v_cmp_le_u32_e32 vcc, s12, v0 117; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 118; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s13, v1 119; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 120; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 121; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 122; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 123; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 124; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 125; GCN-NEXT: s_endpgm 126; 127; GCN-IR-LABEL: s_test_urem_i64: 128; GCN-IR: ; %bb.0: ; %_udiv-special-cases 129; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 130; GCN-IR-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xd 131; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 132; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[8:9], s[6:7], 0 133; GCN-IR-NEXT: s_flbit_i32_b32 s10, s2 134; GCN-IR-NEXT: s_add_i32 s10, s10, 32 135; GCN-IR-NEXT: s_flbit_i32_b32 s11, s3 136; GCN-IR-NEXT: v_mov_b32_e32 v1, s10 137; GCN-IR-NEXT: s_flbit_i32_b32 s10, s6 138; GCN-IR-NEXT: v_mov_b32_e32 v0, s11 139; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s3, 0 140; GCN-IR-NEXT: s_add_i32 s10, s10, 32 141; GCN-IR-NEXT: s_flbit_i32_b32 s11, s7 142; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 143; GCN-IR-NEXT: v_mov_b32_e32 v0, s11 144; GCN-IR-NEXT: v_mov_b32_e32 v1, s10 145; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s7, 0 146; GCN-IR-NEXT: v_cndmask_b32_e32 v3, v0, v1, vcc 147; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v2, v3 148; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], s[2:3], 0 149; GCN-IR-NEXT: v_subb_u32_e64 v1, s[10:11], 0, 0, vcc 150; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 151; GCN-IR-NEXT: s_or_b64 s[0:1], s[0:1], s[8:9] 152; GCN-IR-NEXT: s_or_b64 s[0:1], s[0:1], vcc 153; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[0:1] 154; GCN-IR-NEXT: s_xor_b64 s[8:9], s[0:1], -1 155; GCN-IR-NEXT: s_and_b64 s[8:9], s[8:9], vcc 156; GCN-IR-NEXT: s_and_b64 vcc, exec, s[8:9] 157; GCN-IR-NEXT: s_cbranch_vccz BB0_4 158; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 159; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 1, v0 160; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc 161; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[4:5], v[0:1] 162; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 163; GCN-IR-NEXT: v_lshl_b64 v[0:1], s[6:7], v0 164; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[0:1] 165; GCN-IR-NEXT: s_cbranch_vccz BB0_5 166; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 167; GCN-IR-NEXT: v_not_b32_e32 v2, v2 168; GCN-IR-NEXT: s_add_u32 s8, s2, -1 169; GCN-IR-NEXT: v_lshr_b64 v[6:7], s[6:7], v4 170; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, v2, v3 171; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 172; GCN-IR-NEXT: s_addc_u32 s9, s3, -1 173; GCN-IR-NEXT: v_addc_u32_e64 v5, s[0:1], -1, 0, vcc 174; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 175; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 176; GCN-IR-NEXT: BB0_3: ; %udiv-do-while 177; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 178; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 179; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 180; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 181; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 182; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 183; GCN-IR-NEXT: v_mov_b32_e32 v2, s9 184; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, s8, v6 185; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, v2, v7, vcc 186; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 187; GCN-IR-NEXT: v_and_b32_e32 v10, s2, v8 188; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 189; GCN-IR-NEXT: v_and_b32_e32 v11, s3, v8 190; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 191; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 192; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 193; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 194; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 195; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 196; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 197; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 198; GCN-IR-NEXT: v_subb_u32_e64 v7, s[0:1], v7, v11, s[0:1] 199; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 200; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 201; GCN-IR-NEXT: s_cbranch_vccz BB0_3 202; GCN-IR-NEXT: s_branch BB0_6 203; GCN-IR-NEXT: BB0_4: 204; GCN-IR-NEXT: v_mov_b32_e32 v0, s7 205; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[0:1] 206; GCN-IR-NEXT: v_mov_b32_e32 v0, s6 207; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1] 208; GCN-IR-NEXT: s_branch BB0_7 209; GCN-IR-NEXT: BB0_5: 210; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 211; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 212; GCN-IR-NEXT: BB0_6: ; %Flow6 213; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 214; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 215; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 216; GCN-IR-NEXT: BB0_7: ; %udiv-end 217; GCN-IR-NEXT: v_mul_lo_u32 v1, s2, v1 218; GCN-IR-NEXT: v_mul_hi_u32 v2, s2, v0 219; GCN-IR-NEXT: v_mul_lo_u32 v3, s3, v0 220; GCN-IR-NEXT: v_mul_lo_u32 v0, s2, v0 221; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 222; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 223; GCN-IR-NEXT: v_mov_b32_e32 v2, s7 224; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 225; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 226; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 227; GCN-IR-NEXT: s_mov_b32 s6, -1 228; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 229; GCN-IR-NEXT: s_endpgm 230 %result = urem i64 %x, %y 231 store i64 %result, i64 addrspace(1)* %out 232 ret void 233} 234 235define i64 @v_test_urem_i64(i64 %x, i64 %y) { 236; GCN-LABEL: v_test_urem_i64: 237; GCN: ; %bb.0: 238; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 239; GCN-NEXT: v_cvt_f32_u32_e32 v4, v2 240; GCN-NEXT: v_cvt_f32_u32_e32 v5, v3 241; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v2 242; GCN-NEXT: v_subb_u32_e32 v7, vcc, 0, v3, vcc 243; GCN-NEXT: v_mov_b32_e32 v14, 0 244; GCN-NEXT: v_mac_f32_e32 v4, 0x4f800000, v5 245; GCN-NEXT: v_rcp_f32_e32 v4, v4 246; GCN-NEXT: v_mov_b32_e32 v13, 0 247; GCN-NEXT: v_mul_f32_e32 v4, 0x5f7ffffc, v4 248; GCN-NEXT: v_mul_f32_e32 v5, 0x2f800000, v4 249; GCN-NEXT: v_trunc_f32_e32 v5, v5 250; GCN-NEXT: v_mac_f32_e32 v4, 0xcf800000, v5 251; GCN-NEXT: v_cvt_u32_f32_e32 v4, v4 252; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5 253; GCN-NEXT: v_mul_hi_u32 v8, v6, v4 254; GCN-NEXT: v_mul_lo_u32 v9, v6, v5 255; GCN-NEXT: v_mul_lo_u32 v10, v7, v4 256; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v9 257; GCN-NEXT: v_mul_lo_u32 v9, v6, v4 258; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v10 259; GCN-NEXT: v_mul_lo_u32 v11, v4, v8 260; GCN-NEXT: v_mul_hi_u32 v10, v4, v8 261; GCN-NEXT: v_mul_hi_u32 v12, v4, v9 262; GCN-NEXT: v_mul_hi_u32 v15, v5, v8 263; GCN-NEXT: v_mul_lo_u32 v8, v5, v8 264; GCN-NEXT: v_add_i32_e32 v11, vcc, v12, v11 265; GCN-NEXT: v_mul_lo_u32 v12, v5, v9 266; GCN-NEXT: v_mul_hi_u32 v9, v5, v9 267; GCN-NEXT: v_addc_u32_e32 v10, vcc, v14, v10, vcc 268; GCN-NEXT: v_add_i32_e32 v11, vcc, v11, v12 269; GCN-NEXT: v_addc_u32_e32 v9, vcc, v10, v9, vcc 270; GCN-NEXT: v_addc_u32_e32 v10, vcc, v15, v13, vcc 271; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 272; GCN-NEXT: v_add_i32_e64 v4, s[4:5], v4, v8 273; GCN-NEXT: v_addc_u32_e32 v9, vcc, v14, v10, vcc 274; GCN-NEXT: v_addc_u32_e64 v8, vcc, v5, v9, s[4:5] 275; GCN-NEXT: v_mul_lo_u32 v10, v6, v8 276; GCN-NEXT: v_mul_hi_u32 v11, v6, v4 277; GCN-NEXT: v_mul_lo_u32 v7, v7, v4 278; GCN-NEXT: v_mul_lo_u32 v6, v6, v4 279; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 280; GCN-NEXT: v_add_i32_e32 v7, vcc, v10, v7 281; GCN-NEXT: v_mul_lo_u32 v12, v4, v7 282; GCN-NEXT: v_mul_hi_u32 v15, v4, v6 283; GCN-NEXT: v_mul_hi_u32 v16, v4, v7 284; GCN-NEXT: v_mul_hi_u32 v11, v8, v6 285; GCN-NEXT: v_mul_lo_u32 v6, v8, v6 286; GCN-NEXT: v_add_i32_e32 v12, vcc, v15, v12 287; GCN-NEXT: v_mul_hi_u32 v10, v8, v7 288; GCN-NEXT: v_addc_u32_e32 v15, vcc, v14, v16, vcc 289; GCN-NEXT: v_mul_lo_u32 v7, v8, v7 290; GCN-NEXT: v_add_i32_e32 v6, vcc, v12, v6 291; GCN-NEXT: v_addc_u32_e32 v6, vcc, v15, v11, vcc 292; GCN-NEXT: v_addc_u32_e32 v8, vcc, v10, v13, vcc 293; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 294; GCN-NEXT: v_addc_u32_e32 v7, vcc, v14, v8, vcc 295; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v9 296; GCN-NEXT: v_addc_u32_e64 v5, vcc, v5, v7, s[4:5] 297; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v6 298; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc 299; GCN-NEXT: v_mul_lo_u32 v6, v0, v5 300; GCN-NEXT: v_mul_hi_u32 v7, v0, v4 301; GCN-NEXT: v_mul_hi_u32 v8, v0, v5 302; GCN-NEXT: v_mul_hi_u32 v9, v1, v5 303; GCN-NEXT: v_mul_lo_u32 v5, v1, v5 304; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 305; GCN-NEXT: v_addc_u32_e32 v7, vcc, v14, v8, vcc 306; GCN-NEXT: v_mul_lo_u32 v8, v1, v4 307; GCN-NEXT: v_mul_hi_u32 v4, v1, v4 308; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 309; GCN-NEXT: v_addc_u32_e32 v4, vcc, v7, v4, vcc 310; GCN-NEXT: v_addc_u32_e32 v6, vcc, v9, v13, vcc 311; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 312; GCN-NEXT: v_addc_u32_e32 v5, vcc, v14, v6, vcc 313; GCN-NEXT: v_mul_lo_u32 v5, v2, v5 314; GCN-NEXT: v_mul_hi_u32 v6, v2, v4 315; GCN-NEXT: v_mul_lo_u32 v7, v3, v4 316; GCN-NEXT: v_mul_lo_u32 v4, v2, v4 317; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 318; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v7 319; GCN-NEXT: v_sub_i32_e32 v6, vcc, v1, v5 320; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 321; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v6, v3, vcc 322; GCN-NEXT: v_sub_i32_e64 v6, s[4:5], v0, v2 323; GCN-NEXT: v_subbrev_u32_e64 v7, s[6:7], 0, v4, s[4:5] 324; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v7, v3 325; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 326; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[6:7] 327; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v6, v2 328; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[6:7] 329; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v7, v3 330; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3 331; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v3, s[4:5] 332; GCN-NEXT: v_cndmask_b32_e64 v8, v8, v9, s[6:7] 333; GCN-NEXT: v_sub_i32_e64 v9, s[4:5], v6, v2 334; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 335; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2 336; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 337; GCN-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc 338; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 339; GCN-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc 340; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v8 341; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 342; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v9, s[4:5] 343; GCN-NEXT: v_cndmask_b32_e64 v2, v7, v4, s[4:5] 344; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v6, vcc 345; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 346; GCN-NEXT: s_setpc_b64 s[30:31] 347; 348; GCN-IR-LABEL: v_test_urem_i64: 349; GCN-IR: ; %bb.0: ; %_udiv-special-cases 350; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 351; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[2:3] 352; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 353; GCN-IR-NEXT: v_ffbh_u32_e32 v4, v2 354; GCN-IR-NEXT: s_or_b64 s[4:5], vcc, s[4:5] 355; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 32, v4 356; GCN-IR-NEXT: v_ffbh_u32_e32 v5, v3 357; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3 358; GCN-IR-NEXT: v_cndmask_b32_e32 v10, v5, v4, vcc 359; GCN-IR-NEXT: v_ffbh_u32_e32 v4, v0 360; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 32, v4 361; GCN-IR-NEXT: v_ffbh_u32_e32 v5, v1 362; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1 363; GCN-IR-NEXT: v_cndmask_b32_e32 v12, v5, v4, vcc 364; GCN-IR-NEXT: v_sub_i32_e32 v5, vcc, v10, v12 365; GCN-IR-NEXT: v_subb_u32_e64 v6, s[6:7], 0, 0, vcc 366; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[5:6] 367; GCN-IR-NEXT: v_mov_b32_e32 v11, 0 368; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], vcc 369; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[5:6] 370; GCN-IR-NEXT: s_xor_b64 s[6:7], s[4:5], -1 371; GCN-IR-NEXT: v_mov_b32_e32 v13, v11 372; GCN-IR-NEXT: v_cndmask_b32_e64 v7, v1, 0, s[4:5] 373; GCN-IR-NEXT: v_cndmask_b32_e64 v4, v0, 0, s[4:5] 374; GCN-IR-NEXT: s_and_b64 s[4:5], s[6:7], vcc 375; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 376; GCN-IR-NEXT: s_cbranch_execz BB1_6 377; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 378; GCN-IR-NEXT: v_add_i32_e32 v7, vcc, 1, v5 379; GCN-IR-NEXT: v_addc_u32_e32 v8, vcc, 0, v6, vcc 380; GCN-IR-NEXT: v_sub_i32_e64 v4, s[4:5], 63, v5 381; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[7:8], v[5:6] 382; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 383; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[0:1], v4 384; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 385; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 386; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 387; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 388; GCN-IR-NEXT: s_cbranch_execz BB1_5 389; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 390; GCN-IR-NEXT: v_add_i32_e32 v6, vcc, -1, v2 391; GCN-IR-NEXT: v_lshr_b64 v[14:15], v[0:1], v7 392; GCN-IR-NEXT: v_addc_u32_e32 v7, vcc, -1, v3, vcc 393; GCN-IR-NEXT: v_not_b32_e32 v8, v10 394; GCN-IR-NEXT: v_not_b32_e32 v9, v11 395; GCN-IR-NEXT: v_add_i32_e32 v10, vcc, v8, v12 396; GCN-IR-NEXT: v_addc_u32_e32 v11, vcc, v9, v13, vcc 397; GCN-IR-NEXT: v_mov_b32_e32 v12, 0 398; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 399; GCN-IR-NEXT: BB1_3: ; %udiv-do-while 400; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 401; GCN-IR-NEXT: v_lshl_b64 v[14:15], v[14:15], 1 402; GCN-IR-NEXT: v_lshrrev_b32_e32 v8, 31, v5 403; GCN-IR-NEXT: v_or_b32_e32 v14, v14, v8 404; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[4:5], 1 405; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, v6, v14 406; GCN-IR-NEXT: v_subb_u32_e32 v8, vcc, v7, v15, vcc 407; GCN-IR-NEXT: v_or_b32_e32 v4, v12, v4 408; GCN-IR-NEXT: v_ashrrev_i32_e32 v12, 31, v8 409; GCN-IR-NEXT: v_and_b32_e32 v17, v12, v2 410; GCN-IR-NEXT: v_and_b32_e32 v8, 1, v12 411; GCN-IR-NEXT: v_and_b32_e32 v16, v12, v3 412; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, 1, v10 413; GCN-IR-NEXT: v_or_b32_e32 v5, v13, v5 414; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, 0, v11, vcc 415; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[12:13], v[10:11] 416; GCN-IR-NEXT: v_mov_b32_e32 v10, v12 417; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 418; GCN-IR-NEXT: v_sub_i32_e64 v14, s[4:5], v14, v17 419; GCN-IR-NEXT: v_mov_b32_e32 v11, v13 420; GCN-IR-NEXT: v_mov_b32_e32 v13, v9 421; GCN-IR-NEXT: v_subb_u32_e64 v15, s[4:5], v15, v16, s[4:5] 422; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 423; GCN-IR-NEXT: v_mov_b32_e32 v12, v8 424; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 425; GCN-IR-NEXT: s_cbranch_execnz BB1_3 426; GCN-IR-NEXT: ; %bb.4: ; %Flow 427; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 428; GCN-IR-NEXT: BB1_5: ; %Flow3 429; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 430; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[4:5], 1 431; GCN-IR-NEXT: v_or_b32_e32 v7, v9, v5 432; GCN-IR-NEXT: v_or_b32_e32 v4, v8, v4 433; GCN-IR-NEXT: BB1_6: ; %Flow4 434; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 435; GCN-IR-NEXT: v_mul_lo_u32 v5, v2, v7 436; GCN-IR-NEXT: v_mul_hi_u32 v6, v2, v4 437; GCN-IR-NEXT: v_mul_lo_u32 v3, v3, v4 438; GCN-IR-NEXT: v_mul_lo_u32 v2, v2, v4 439; GCN-IR-NEXT: v_add_i32_e32 v5, vcc, v6, v5 440; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, v5, v3 441; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 442; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 443; GCN-IR-NEXT: s_setpc_b64 s[30:31] 444 %result = urem i64 %x, %y 445 ret i64 %result 446} 447 448define amdgpu_kernel void @s_test_urem31_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 449; GCN-LABEL: s_test_urem31_i64: 450; GCN: ; %bb.0: 451; GCN-NEXT: s_load_dword s2, s[0:1], 0xe 452; GCN-NEXT: s_waitcnt lgkmcnt(0) 453; GCN-NEXT: s_lshr_b32 s4, s2, 1 454; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 455; GCN-NEXT: v_cvt_f32_u32_e32 v0, s4 456; GCN-NEXT: s_waitcnt lgkmcnt(0) 457; GCN-NEXT: s_mov_b32 s2, -1 458; GCN-NEXT: s_lshr_b32 s5, s3, 1 459; GCN-NEXT: v_cvt_f32_u32_e32 v1, s5 460; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 461; GCN-NEXT: s_mov_b32 s3, 0xf000 462; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 463; GCN-NEXT: v_trunc_f32_e32 v2, v2 464; GCN-NEXT: v_cvt_u32_f32_e32 v3, v2 465; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 466; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 467; GCN-NEXT: v_mov_b32_e32 v1, 0 468; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v3, vcc 469; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 470; GCN-NEXT: v_sub_i32_e32 v0, vcc, s5, v0 471; GCN-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0 472; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 473; GCN-NEXT: s_endpgm 474; 475; GCN-IR-LABEL: s_test_urem31_i64: 476; GCN-IR: ; %bb.0: 477; GCN-IR-NEXT: s_load_dword s2, s[0:1], 0xe 478; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 479; GCN-IR-NEXT: s_lshr_b32 s4, s2, 1 480; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 481; GCN-IR-NEXT: v_cvt_f32_u32_e32 v0, s4 482; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 483; GCN-IR-NEXT: s_mov_b32 s2, -1 484; GCN-IR-NEXT: s_lshr_b32 s5, s3, 1 485; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, s5 486; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 487; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 488; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 489; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 490; GCN-IR-NEXT: v_cvt_u32_f32_e32 v3, v2 491; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 492; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 493; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 494; GCN-IR-NEXT: v_addc_u32_e32 v0, vcc, 0, v3, vcc 495; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 496; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s5, v0 497; GCN-IR-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0 498; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 499; GCN-IR-NEXT: s_endpgm 500 %1 = lshr i64 %x, 33 501 %2 = lshr i64 %y, 33 502 %result = urem i64 %1, %2 503 store i64 %result, i64 addrspace(1)* %out 504 ret void 505} 506 507define amdgpu_kernel void @s_test_urem31_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> %x, <2 x i64> %y) { 508; GCN-LABEL: s_test_urem31_v2i64: 509; GCN: ; %bb.0: 510; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 511; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0xd 512; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x11 513; GCN-NEXT: s_mov_b32 s7, 0xf000 514; GCN-NEXT: s_mov_b32 s6, -1 515; GCN-NEXT: s_waitcnt lgkmcnt(0) 516; GCN-NEXT: s_lshr_b32 s2, s9, 1 517; GCN-NEXT: s_lshr_b32 s0, s1, 1 518; GCN-NEXT: v_cvt_f32_u32_e32 v0, s0 519; GCN-NEXT: v_cvt_f32_u32_e32 v1, s2 520; GCN-NEXT: s_lshr_b32 s3, s3, 1 521; GCN-NEXT: v_cvt_f32_u32_e32 v4, s3 522; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 523; GCN-NEXT: s_lshr_b32 s1, s11, 1 524; GCN-NEXT: v_cvt_f32_u32_e32 v3, s1 525; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 526; GCN-NEXT: v_trunc_f32_e32 v2, v2 527; GCN-NEXT: v_cvt_u32_f32_e32 v5, v2 528; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 529; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v4 530; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 531; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v5, vcc 532; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 533; GCN-NEXT: v_mul_f32_e32 v2, v3, v2 534; GCN-NEXT: v_trunc_f32_e32 v2, v2 535; GCN-NEXT: v_cvt_u32_f32_e32 v5, v2 536; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 537; GCN-NEXT: v_mad_f32 v2, -v2, v4, v3 538; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v4 539; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v5, vcc 540; GCN-NEXT: v_mul_lo_u32 v2, v2, s3 541; GCN-NEXT: v_mov_b32_e32 v1, 0 542; GCN-NEXT: s_brev_b32 s0, -2 543; GCN-NEXT: v_and_b32_e32 v0, s0, v0 544; GCN-NEXT: v_sub_i32_e32 v2, vcc, s1, v2 545; GCN-NEXT: v_and_b32_e32 v2, s0, v2 546; GCN-NEXT: v_mov_b32_e32 v3, v1 547; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 548; GCN-NEXT: s_endpgm 549; 550; GCN-IR-LABEL: s_test_urem31_v2i64: 551; GCN-IR: ; %bb.0: 552; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 553; GCN-IR-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0xd 554; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x11 555; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 556; GCN-IR-NEXT: s_mov_b32 s6, -1 557; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 558; GCN-IR-NEXT: s_lshr_b32 s2, s9, 1 559; GCN-IR-NEXT: s_lshr_b32 s0, s1, 1 560; GCN-IR-NEXT: v_cvt_f32_u32_e32 v0, s0 561; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, s2 562; GCN-IR-NEXT: s_lshr_b32 s3, s3, 1 563; GCN-IR-NEXT: v_cvt_f32_u32_e32 v4, s3 564; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 565; GCN-IR-NEXT: s_lshr_b32 s1, s11, 1 566; GCN-IR-NEXT: v_cvt_f32_u32_e32 v3, s1 567; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 568; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 569; GCN-IR-NEXT: v_cvt_u32_f32_e32 v5, v2 570; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 571; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v4 572; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 573; GCN-IR-NEXT: v_addc_u32_e32 v0, vcc, 0, v5, vcc 574; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 575; GCN-IR-NEXT: v_mul_f32_e32 v2, v3, v2 576; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 577; GCN-IR-NEXT: v_cvt_u32_f32_e32 v5, v2 578; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 579; GCN-IR-NEXT: v_mad_f32 v2, -v2, v4, v3 580; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v4 581; GCN-IR-NEXT: v_addc_u32_e32 v2, vcc, 0, v5, vcc 582; GCN-IR-NEXT: v_mul_lo_u32 v2, v2, s3 583; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 584; GCN-IR-NEXT: s_brev_b32 s0, -2 585; GCN-IR-NEXT: v_and_b32_e32 v0, s0, v0 586; GCN-IR-NEXT: v_sub_i32_e32 v2, vcc, s1, v2 587; GCN-IR-NEXT: v_and_b32_e32 v2, s0, v2 588; GCN-IR-NEXT: v_mov_b32_e32 v3, v1 589; GCN-IR-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 590; GCN-IR-NEXT: s_endpgm 591 %1 = lshr <2 x i64> %x, <i64 33, i64 33> 592 %2 = lshr <2 x i64> %y, <i64 33, i64 33> 593 %result = urem <2 x i64> %1, %2 594 store <2 x i64> %result, <2 x i64> addrspace(1)* %out 595 ret void 596} 597 598define amdgpu_kernel void @s_test_urem24_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 599; GCN-LABEL: s_test_urem24_i64: 600; GCN: ; %bb.0: 601; GCN-NEXT: s_load_dword s2, s[0:1], 0xe 602; GCN-NEXT: s_waitcnt lgkmcnt(0) 603; GCN-NEXT: s_lshr_b32 s4, s2, 8 604; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 605; GCN-NEXT: v_cvt_f32_u32_e32 v0, s4 606; GCN-NEXT: s_waitcnt lgkmcnt(0) 607; GCN-NEXT: s_mov_b32 s2, -1 608; GCN-NEXT: s_lshr_b32 s5, s3, 8 609; GCN-NEXT: v_cvt_f32_u32_e32 v1, s5 610; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 611; GCN-NEXT: s_mov_b32 s3, 0xf000 612; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 613; GCN-NEXT: v_trunc_f32_e32 v2, v2 614; GCN-NEXT: v_cvt_u32_f32_e32 v3, v2 615; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 616; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 617; GCN-NEXT: v_mov_b32_e32 v1, 0 618; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v3, vcc 619; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 620; GCN-NEXT: v_sub_i32_e32 v0, vcc, s5, v0 621; GCN-NEXT: v_and_b32_e32 v0, 0xffffff, v0 622; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 623; GCN-NEXT: s_endpgm 624; 625; GCN-IR-LABEL: s_test_urem24_i64: 626; GCN-IR: ; %bb.0: 627; GCN-IR-NEXT: s_load_dword s2, s[0:1], 0xe 628; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 629; GCN-IR-NEXT: s_lshr_b32 s4, s2, 8 630; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 631; GCN-IR-NEXT: v_cvt_f32_u32_e32 v0, s4 632; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 633; GCN-IR-NEXT: s_mov_b32 s2, -1 634; GCN-IR-NEXT: s_lshr_b32 s5, s3, 8 635; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, s5 636; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 637; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 638; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 639; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 640; GCN-IR-NEXT: v_cvt_u32_f32_e32 v3, v2 641; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 642; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 643; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 644; GCN-IR-NEXT: v_addc_u32_e32 v0, vcc, 0, v3, vcc 645; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 646; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s5, v0 647; GCN-IR-NEXT: v_and_b32_e32 v0, 0xffffff, v0 648; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 649; GCN-IR-NEXT: s_endpgm 650 %1 = lshr i64 %x, 40 651 %2 = lshr i64 %y, 40 652 %result = urem i64 %1, %2 653 store i64 %result, i64 addrspace(1)* %out 654 ret void 655} 656 657define amdgpu_kernel void @s_test_urem23_64_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> %x, <2 x i64> %y) { 658; GCN-LABEL: s_test_urem23_64_v2i64: 659; GCN: ; %bb.0: 660; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 661; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0xd 662; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x11 663; GCN-NEXT: s_mov_b32 s7, 0xf000 664; GCN-NEXT: s_mov_b32 s6, -1 665; GCN-NEXT: s_waitcnt lgkmcnt(0) 666; GCN-NEXT: s_lshr_b32 s2, s9, 1 667; GCN-NEXT: s_lshr_b32 s0, s1, 1 668; GCN-NEXT: v_cvt_f32_u32_e32 v0, s0 669; GCN-NEXT: v_cvt_f32_u32_e32 v1, s2 670; GCN-NEXT: s_lshr_b32 s3, s3, 9 671; GCN-NEXT: v_cvt_f32_u32_e32 v4, s3 672; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 673; GCN-NEXT: s_lshr_b32 s1, s11, 9 674; GCN-NEXT: v_cvt_f32_u32_e32 v3, s1 675; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 676; GCN-NEXT: v_trunc_f32_e32 v2, v2 677; GCN-NEXT: v_cvt_u32_f32_e32 v5, v2 678; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 679; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v4 680; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 681; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v5, vcc 682; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 683; GCN-NEXT: v_mul_f32_e32 v2, v3, v2 684; GCN-NEXT: v_trunc_f32_e32 v2, v2 685; GCN-NEXT: v_cvt_u32_f32_e32 v5, v2 686; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 687; GCN-NEXT: v_mad_f32 v2, -v2, v4, v3 688; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v4 689; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v5, vcc 690; GCN-NEXT: v_mul_lo_u32 v2, v2, s3 691; GCN-NEXT: v_mov_b32_e32 v1, 0 692; GCN-NEXT: s_brev_b32 s0, -2 693; GCN-NEXT: v_and_b32_e32 v0, s0, v0 694; GCN-NEXT: v_sub_i32_e32 v2, vcc, s1, v2 695; GCN-NEXT: v_and_b32_e32 v2, s0, v2 696; GCN-NEXT: v_mov_b32_e32 v3, v1 697; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 698; GCN-NEXT: s_endpgm 699; 700; GCN-IR-LABEL: s_test_urem23_64_v2i64: 701; GCN-IR: ; %bb.0: 702; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 703; GCN-IR-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0xd 704; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x11 705; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 706; GCN-IR-NEXT: s_mov_b32 s6, -1 707; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 708; GCN-IR-NEXT: s_lshr_b32 s2, s9, 1 709; GCN-IR-NEXT: s_lshr_b32 s0, s1, 1 710; GCN-IR-NEXT: v_cvt_f32_u32_e32 v0, s0 711; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, s2 712; GCN-IR-NEXT: s_lshr_b32 s3, s3, 9 713; GCN-IR-NEXT: v_cvt_f32_u32_e32 v4, s3 714; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 715; GCN-IR-NEXT: s_lshr_b32 s1, s11, 9 716; GCN-IR-NEXT: v_cvt_f32_u32_e32 v3, s1 717; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 718; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 719; GCN-IR-NEXT: v_cvt_u32_f32_e32 v5, v2 720; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 721; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v4 722; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 723; GCN-IR-NEXT: v_addc_u32_e32 v0, vcc, 0, v5, vcc 724; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 725; GCN-IR-NEXT: v_mul_f32_e32 v2, v3, v2 726; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 727; GCN-IR-NEXT: v_cvt_u32_f32_e32 v5, v2 728; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 729; GCN-IR-NEXT: v_mad_f32 v2, -v2, v4, v3 730; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v4 731; GCN-IR-NEXT: v_addc_u32_e32 v2, vcc, 0, v5, vcc 732; GCN-IR-NEXT: v_mul_lo_u32 v2, v2, s3 733; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 734; GCN-IR-NEXT: s_brev_b32 s0, -2 735; GCN-IR-NEXT: v_and_b32_e32 v0, s0, v0 736; GCN-IR-NEXT: v_sub_i32_e32 v2, vcc, s1, v2 737; GCN-IR-NEXT: v_and_b32_e32 v2, s0, v2 738; GCN-IR-NEXT: v_mov_b32_e32 v3, v1 739; GCN-IR-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 740; GCN-IR-NEXT: s_endpgm 741 %1 = lshr <2 x i64> %x, <i64 33, i64 41> 742 %2 = lshr <2 x i64> %y, <i64 33, i64 41> 743 %result = urem <2 x i64> %1, %2 744 store <2 x i64> %result, <2 x i64> addrspace(1)* %out 745 ret void 746} 747 748define amdgpu_kernel void @s_test_urem_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 749; GCN-LABEL: s_test_urem_k_num_i64: 750; GCN: ; %bb.0: 751; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 752; GCN-NEXT: v_mov_b32_e32 v2, 0 753; GCN-NEXT: s_mov_b32 s11, 0xf000 754; GCN-NEXT: s_mov_b32 s10, -1 755; GCN-NEXT: s_waitcnt lgkmcnt(0) 756; GCN-NEXT: v_cvt_f32_u32_e32 v0, s6 757; GCN-NEXT: v_cvt_f32_u32_e32 v1, s7 758; GCN-NEXT: s_sub_u32 s2, 0, s6 759; GCN-NEXT: s_subb_u32 s3, 0, s7 760; GCN-NEXT: s_mov_b32 s8, s4 761; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 762; GCN-NEXT: v_rcp_f32_e32 v0, v0 763; GCN-NEXT: v_mov_b32_e32 v1, 0 764; GCN-NEXT: s_mov_b32 s9, s5 765; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 766; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v0 767; GCN-NEXT: v_trunc_f32_e32 v3, v3 768; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v3 769; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 770; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 771; GCN-NEXT: v_mul_hi_u32 v5, s2, v0 772; GCN-NEXT: v_mul_lo_u32 v4, s2, v3 773; GCN-NEXT: v_mul_lo_u32 v7, s3, v0 774; GCN-NEXT: v_mul_lo_u32 v6, s2, v0 775; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 776; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v7 777; GCN-NEXT: v_mul_hi_u32 v5, v0, v6 778; GCN-NEXT: v_mul_lo_u32 v7, v0, v4 779; GCN-NEXT: v_mul_hi_u32 v9, v0, v4 780; GCN-NEXT: v_mul_lo_u32 v8, v3, v6 781; GCN-NEXT: v_mul_hi_u32 v6, v3, v6 782; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v7 783; GCN-NEXT: v_mul_hi_u32 v10, v3, v4 784; GCN-NEXT: v_addc_u32_e32 v7, vcc, v2, v9, vcc 785; GCN-NEXT: v_mul_lo_u32 v4, v3, v4 786; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 787; GCN-NEXT: v_addc_u32_e32 v5, vcc, v7, v6, vcc 788; GCN-NEXT: v_addc_u32_e32 v6, vcc, v10, v1, vcc 789; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 790; GCN-NEXT: v_add_i32_e64 v0, s[0:1], v0, v4 791; GCN-NEXT: v_addc_u32_e32 v5, vcc, v2, v6, vcc 792; GCN-NEXT: v_addc_u32_e64 v4, vcc, v3, v5, s[0:1] 793; GCN-NEXT: v_mul_lo_u32 v6, s2, v4 794; GCN-NEXT: v_mul_hi_u32 v7, s2, v0 795; GCN-NEXT: v_mul_lo_u32 v8, s3, v0 796; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 797; GCN-NEXT: v_mul_lo_u32 v7, s2, v0 798; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 799; GCN-NEXT: v_mul_lo_u32 v10, v0, v6 800; GCN-NEXT: v_mul_hi_u32 v12, v0, v6 801; GCN-NEXT: v_mul_hi_u32 v11, v0, v7 802; GCN-NEXT: v_mul_hi_u32 v9, v4, v7 803; GCN-NEXT: v_mul_lo_u32 v7, v4, v7 804; GCN-NEXT: v_mul_hi_u32 v8, v4, v6 805; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 806; GCN-NEXT: v_addc_u32_e32 v11, vcc, v2, v12, vcc 807; GCN-NEXT: v_mul_lo_u32 v4, v4, v6 808; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 809; GCN-NEXT: v_addc_u32_e32 v7, vcc, v11, v9, vcc 810; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v1, vcc 811; GCN-NEXT: v_add_i32_e32 v4, vcc, v7, v4 812; GCN-NEXT: v_addc_u32_e32 v6, vcc, v2, v6, vcc 813; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 814; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v6, s[0:1] 815; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4 816; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 817; GCN-NEXT: v_mul_lo_u32 v4, v3, 24 818; GCN-NEXT: v_mul_hi_u32 v0, v0, 24 819; GCN-NEXT: v_mul_hi_u32 v3, v3, 24 820; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4 821; GCN-NEXT: v_addc_u32_e32 v0, vcc, v2, v3, vcc 822; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc 823; GCN-NEXT: v_mul_lo_u32 v1, s6, v1 824; GCN-NEXT: v_mul_hi_u32 v2, s6, v0 825; GCN-NEXT: v_mul_lo_u32 v3, s7, v0 826; GCN-NEXT: v_mul_lo_u32 v0, s6, v0 827; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 828; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 829; GCN-NEXT: v_sub_i32_e32 v2, vcc, 0, v1 830; GCN-NEXT: v_mov_b32_e32 v3, s7 831; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 832; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 833; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s6, v0 834; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 835; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s7, v5 836; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 837; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 838; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s6, v4 839; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s6, v4 840; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 841; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s7, v5 842; GCN-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 843; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 844; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 845; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 846; GCN-NEXT: v_cmp_le_u32_e32 vcc, s7, v1 847; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 848; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 849; GCN-NEXT: v_cmp_le_u32_e32 vcc, s6, v0 850; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 851; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s7, v1 852; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 853; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 854; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 855; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 856; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 857; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 858; GCN-NEXT: s_endpgm 859; 860; GCN-IR-LABEL: s_test_urem_k_num_i64: 861; GCN-IR: ; %bb.0: ; %_udiv-special-cases 862; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 863; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 864; GCN-IR-NEXT: s_flbit_i32_b32 s2, s6 865; GCN-IR-NEXT: s_flbit_i32_b32 s3, s7 866; GCN-IR-NEXT: s_add_i32 s2, s2, 32 867; GCN-IR-NEXT: v_mov_b32_e32 v0, s3 868; GCN-IR-NEXT: v_mov_b32_e32 v1, s2 869; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s7, 0 870; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 871; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, 0xffffffc5, v2 872; GCN-IR-NEXT: v_addc_u32_e64 v1, s[2:3], 0, -1, vcc 873; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], 0 874; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 875; GCN-IR-NEXT: s_or_b64 s[0:1], s[0:1], vcc 876; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[0:1] 877; GCN-IR-NEXT: s_xor_b64 s[2:3], s[0:1], -1 878; GCN-IR-NEXT: s_and_b64 s[2:3], s[2:3], vcc 879; GCN-IR-NEXT: s_and_b64 vcc, exec, s[2:3] 880; GCN-IR-NEXT: s_cbranch_vccz BB6_4 881; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 882; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, 1, v0 883; GCN-IR-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc 884; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[3:4], v[0:1] 885; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 886; GCN-IR-NEXT: v_lshl_b64 v[0:1], 24, v0 887; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[0:1] 888; GCN-IR-NEXT: s_cbranch_vccz BB6_5 889; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 890; GCN-IR-NEXT: s_add_u32 s2, s6, -1 891; GCN-IR-NEXT: v_lshr_b64 v[6:7], 24, v3 892; GCN-IR-NEXT: v_sub_i32_e32 v4, vcc, 58, v2 893; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 894; GCN-IR-NEXT: s_addc_u32 s3, s7, -1 895; GCN-IR-NEXT: v_subb_u32_e64 v5, s[0:1], 0, 0, vcc 896; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 897; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 898; GCN-IR-NEXT: BB6_3: ; %udiv-do-while 899; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 900; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 901; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 902; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 903; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 904; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 905; GCN-IR-NEXT: v_mov_b32_e32 v2, s3 906; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, s2, v6 907; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, v2, v7, vcc 908; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 909; GCN-IR-NEXT: v_and_b32_e32 v10, s6, v8 910; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 911; GCN-IR-NEXT: v_and_b32_e32 v11, s7, v8 912; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 913; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 914; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 915; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 916; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 917; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 918; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 919; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 920; GCN-IR-NEXT: v_subb_u32_e64 v7, s[0:1], v7, v11, s[0:1] 921; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 922; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 923; GCN-IR-NEXT: s_cbranch_vccz BB6_3 924; GCN-IR-NEXT: s_branch BB6_6 925; GCN-IR-NEXT: BB6_4: 926; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 927; GCN-IR-NEXT: v_cndmask_b32_e64 v0, 24, 0, s[0:1] 928; GCN-IR-NEXT: s_branch BB6_7 929; GCN-IR-NEXT: BB6_5: 930; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 931; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 932; GCN-IR-NEXT: BB6_6: ; %Flow5 933; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 934; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 935; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 936; GCN-IR-NEXT: BB6_7: ; %udiv-end 937; GCN-IR-NEXT: v_mul_lo_u32 v1, s6, v1 938; GCN-IR-NEXT: v_mul_hi_u32 v2, s6, v0 939; GCN-IR-NEXT: v_mul_lo_u32 v3, s7, v0 940; GCN-IR-NEXT: v_mul_lo_u32 v0, s6, v0 941; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 942; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 943; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 944; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 945; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 946; GCN-IR-NEXT: s_mov_b32 s6, -1 947; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 948; GCN-IR-NEXT: s_endpgm 949 %result = urem i64 24, %x 950 store i64 %result, i64 addrspace(1)* %out 951 ret void 952} 953 954define amdgpu_kernel void @s_test_urem_k_den_i64(i64 addrspace(1)* %out, i64 %x) { 955; GCN-LABEL: s_test_urem_k_den_i64: 956; GCN: ; %bb.0: 957; GCN-NEXT: v_mov_b32_e32 v0, 0x4f800000 958; GCN-NEXT: v_madak_f32 v0, 0, v0, 0x41c00000 959; GCN-NEXT: v_rcp_f32_e32 v0, v0 960; GCN-NEXT: s_movk_i32 s2, 0xffe8 961; GCN-NEXT: v_mov_b32_e32 v8, 0 962; GCN-NEXT: v_mov_b32_e32 v7, 0 963; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 964; GCN-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0 965; GCN-NEXT: v_trunc_f32_e32 v1, v1 966; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v1 967; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 968; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1 969; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 970; GCN-NEXT: s_mov_b32 s11, 0xf000 971; GCN-NEXT: v_mul_hi_u32 v2, v0, s2 972; GCN-NEXT: v_mul_lo_u32 v3, v1, s2 973; GCN-NEXT: v_mul_lo_u32 v4, v0, s2 974; GCN-NEXT: s_mov_b32 s10, -1 975; GCN-NEXT: v_subrev_i32_e32 v2, vcc, v0, v2 976; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 977; GCN-NEXT: v_mul_lo_u32 v5, v0, v2 978; GCN-NEXT: v_mul_hi_u32 v6, v0, v4 979; GCN-NEXT: v_mul_hi_u32 v3, v0, v2 980; GCN-NEXT: v_mul_hi_u32 v9, v1, v2 981; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 982; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 983; GCN-NEXT: v_mul_lo_u32 v6, v1, v4 984; GCN-NEXT: v_mul_hi_u32 v4, v1, v4 985; GCN-NEXT: v_addc_u32_e32 v3, vcc, v8, v3, vcc 986; GCN-NEXT: s_waitcnt lgkmcnt(0) 987; GCN-NEXT: s_mov_b32 s8, s4 988; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 989; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v4, vcc 990; GCN-NEXT: v_addc_u32_e32 v4, vcc, v9, v7, vcc 991; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 992; GCN-NEXT: v_add_i32_e64 v0, s[0:1], v0, v2 993; GCN-NEXT: v_addc_u32_e32 v3, vcc, v8, v4, vcc 994; GCN-NEXT: v_mul_hi_u32 v4, v0, s2 995; GCN-NEXT: v_addc_u32_e64 v2, vcc, v1, v3, s[0:1] 996; GCN-NEXT: v_mul_lo_u32 v5, v2, s2 997; GCN-NEXT: v_mul_lo_u32 v6, v0, s2 998; GCN-NEXT: v_subrev_i32_e32 v4, vcc, v0, v4 999; GCN-NEXT: s_mov_b32 s9, s5 1000; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 1001; GCN-NEXT: v_mul_lo_u32 v5, v0, v4 1002; GCN-NEXT: v_mul_hi_u32 v9, v0, v6 1003; GCN-NEXT: v_mul_hi_u32 v10, v0, v4 1004; GCN-NEXT: v_mul_hi_u32 v11, v2, v4 1005; GCN-NEXT: v_add_i32_e32 v5, vcc, v9, v5 1006; GCN-NEXT: v_addc_u32_e32 v9, vcc, v8, v10, vcc 1007; GCN-NEXT: v_mul_lo_u32 v10, v2, v6 1008; GCN-NEXT: v_mul_hi_u32 v6, v2, v6 1009; GCN-NEXT: v_mul_lo_u32 v2, v2, v4 1010; GCN-NEXT: v_add_i32_e32 v5, vcc, v10, v5 1011; GCN-NEXT: v_addc_u32_e32 v5, vcc, v9, v6, vcc 1012; GCN-NEXT: v_addc_u32_e32 v4, vcc, v11, v7, vcc 1013; GCN-NEXT: v_add_i32_e32 v2, vcc, v5, v2 1014; GCN-NEXT: v_addc_u32_e32 v4, vcc, v8, v4, vcc 1015; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1016; GCN-NEXT: v_addc_u32_e64 v1, vcc, v1, v4, s[0:1] 1017; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1018; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc 1019; GCN-NEXT: v_mul_lo_u32 v2, s6, v1 1020; GCN-NEXT: v_mul_hi_u32 v3, s6, v0 1021; GCN-NEXT: v_mul_hi_u32 v4, s6, v1 1022; GCN-NEXT: v_mul_hi_u32 v5, s7, v1 1023; GCN-NEXT: v_mul_lo_u32 v1, s7, v1 1024; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1025; GCN-NEXT: v_addc_u32_e32 v3, vcc, v8, v4, vcc 1026; GCN-NEXT: v_mul_lo_u32 v4, s7, v0 1027; GCN-NEXT: v_mul_hi_u32 v0, s7, v0 1028; GCN-NEXT: v_add_i32_e32 v2, vcc, v4, v2 1029; GCN-NEXT: v_addc_u32_e32 v0, vcc, v3, v0, vcc 1030; GCN-NEXT: v_addc_u32_e32 v2, vcc, v5, v7, vcc 1031; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 1032; GCN-NEXT: v_addc_u32_e32 v1, vcc, v8, v2, vcc 1033; GCN-NEXT: v_mul_hi_u32 v2, v0, 24 1034; GCN-NEXT: v_mul_lo_u32 v1, v1, 24 1035; GCN-NEXT: v_mul_lo_u32 v0, v0, 24 1036; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1037; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 1038; GCN-NEXT: v_mov_b32_e32 v2, s7 1039; GCN-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 1040; GCN-NEXT: v_subrev_i32_e32 v2, vcc, 24, v0 1041; GCN-NEXT: v_subbrev_u32_e32 v3, vcc, 0, v1, vcc 1042; GCN-NEXT: v_subrev_i32_e32 v4, vcc, 24, v2 1043; GCN-NEXT: v_subbrev_u32_e32 v5, vcc, 0, v3, vcc 1044; GCN-NEXT: v_cmp_lt_u32_e32 vcc, 23, v2 1045; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 1046; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3 1047; GCN-NEXT: v_cndmask_b32_e32 v6, -1, v6, vcc 1048; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6 1049; GCN-NEXT: v_cmp_lt_u32_e64 s[0:1], 23, v0 1050; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc 1051; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, s[0:1] 1052; GCN-NEXT: v_cmp_eq_u32_e64 s[0:1], 0, v1 1053; GCN-NEXT: v_cndmask_b32_e64 v5, -1, v5, s[0:1] 1054; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v5 1055; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc 1056; GCN-NEXT: v_cndmask_b32_e64 v1, v1, v3, s[0:1] 1057; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1] 1058; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 1059; GCN-NEXT: s_endpgm 1060; 1061; GCN-IR-LABEL: s_test_urem_k_den_i64: 1062; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1063; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 1064; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1065; GCN-IR-NEXT: s_flbit_i32_b32 s2, s6 1066; GCN-IR-NEXT: s_flbit_i32_b32 s3, s7 1067; GCN-IR-NEXT: s_add_i32 s2, s2, 32 1068; GCN-IR-NEXT: v_mov_b32_e32 v0, s3 1069; GCN-IR-NEXT: v_mov_b32_e32 v1, s2 1070; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s7, 0 1071; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 1072; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 59, v2 1073; GCN-IR-NEXT: v_subb_u32_e64 v1, s[2:3], 0, 0, vcc 1074; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], 0 1075; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 1076; GCN-IR-NEXT: s_or_b64 s[0:1], s[0:1], vcc 1077; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[0:1] 1078; GCN-IR-NEXT: s_xor_b64 s[2:3], s[0:1], -1 1079; GCN-IR-NEXT: s_and_b64 s[2:3], s[2:3], vcc 1080; GCN-IR-NEXT: s_and_b64 vcc, exec, s[2:3] 1081; GCN-IR-NEXT: s_cbranch_vccz BB7_4 1082; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1083; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, 1, v0 1084; GCN-IR-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc 1085; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[3:4], v[0:1] 1086; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 1087; GCN-IR-NEXT: v_lshl_b64 v[0:1], s[6:7], v0 1088; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[0:1] 1089; GCN-IR-NEXT: s_cbranch_vccz BB7_5 1090; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1091; GCN-IR-NEXT: v_lshr_b64 v[6:7], s[6:7], v3 1092; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 0xffffffc4, v2 1093; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 1094; GCN-IR-NEXT: v_addc_u32_e64 v5, s[0:1], 0, -1, vcc 1095; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1096; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1097; GCN-IR-NEXT: BB7_3: ; %udiv-do-while 1098; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1099; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 1100; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 1101; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 1102; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1103; GCN-IR-NEXT: v_sub_i32_e32 v2, vcc, 23, v6 1104; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, 0, v7, vcc 1105; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 1106; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 1107; GCN-IR-NEXT: v_and_b32_e32 v10, 24, v8 1108; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 1109; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 1110; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 1111; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 1112; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 1113; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 1114; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 1115; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1116; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 1117; GCN-IR-NEXT: v_subbrev_u32_e64 v7, s[0:1], 0, v7, s[0:1] 1118; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 1119; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 1120; GCN-IR-NEXT: s_cbranch_vccz BB7_3 1121; GCN-IR-NEXT: s_branch BB7_6 1122; GCN-IR-NEXT: BB7_4: 1123; GCN-IR-NEXT: v_mov_b32_e32 v0, s7 1124; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[0:1] 1125; GCN-IR-NEXT: v_mov_b32_e32 v0, s6 1126; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[0:1] 1127; GCN-IR-NEXT: s_branch BB7_7 1128; GCN-IR-NEXT: BB7_5: 1129; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 1130; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1131; GCN-IR-NEXT: BB7_6: ; %Flow5 1132; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1133; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 1134; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 1135; GCN-IR-NEXT: BB7_7: ; %udiv-end 1136; GCN-IR-NEXT: v_mul_hi_u32 v2, v0, 24 1137; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, 24 1138; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, 24 1139; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1140; GCN-IR-NEXT: v_mov_b32_e32 v2, s7 1141; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 1142; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 1143; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 1144; GCN-IR-NEXT: s_mov_b32 s6, -1 1145; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1146; GCN-IR-NEXT: s_endpgm 1147 %result = urem i64 %x, 24 1148 store i64 %result, i64 addrspace(1)* %out 1149 ret void 1150} 1151 1152; FIXME: Constant bus violation 1153; define i64 @v_test_urem_k_num_i64(i64 %x) { 1154; %result = urem i64 24, %x 1155; ret i64 %result 1156; } 1157 1158define i64 @v_test_urem_pow2_k_num_i64(i64 %x) { 1159; GCN-LABEL: v_test_urem_pow2_k_num_i64: 1160; GCN: ; %bb.0: 1161; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1162; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 1163; GCN-NEXT: v_cvt_f32_u32_e32 v3, v1 1164; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v0 1165; GCN-NEXT: v_subb_u32_e32 v5, vcc, 0, v1, vcc 1166; GCN-NEXT: v_mov_b32_e32 v12, 0 1167; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 1168; GCN-NEXT: v_rcp_f32_e32 v2, v2 1169; GCN-NEXT: v_mov_b32_e32 v11, 0 1170; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 1171; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 1172; GCN-NEXT: v_trunc_f32_e32 v3, v3 1173; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 1174; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 1175; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 1176; GCN-NEXT: v_mul_hi_u32 v7, v4, v2 1177; GCN-NEXT: v_mul_lo_u32 v6, v4, v3 1178; GCN-NEXT: v_mul_lo_u32 v8, v5, v2 1179; GCN-NEXT: v_mul_lo_u32 v9, v4, v2 1180; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1181; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 1182; GCN-NEXT: v_mul_lo_u32 v8, v2, v6 1183; GCN-NEXT: v_mul_hi_u32 v10, v2, v9 1184; GCN-NEXT: v_mul_hi_u32 v7, v2, v6 1185; GCN-NEXT: v_mul_hi_u32 v13, v3, v6 1186; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 1187; GCN-NEXT: v_add_i32_e32 v8, vcc, v10, v8 1188; GCN-NEXT: v_mul_lo_u32 v10, v3, v9 1189; GCN-NEXT: v_mul_hi_u32 v9, v3, v9 1190; GCN-NEXT: v_addc_u32_e32 v7, vcc, v12, v7, vcc 1191; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v10 1192; GCN-NEXT: v_addc_u32_e32 v7, vcc, v7, v9, vcc 1193; GCN-NEXT: v_addc_u32_e32 v8, vcc, v13, v11, vcc 1194; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1195; GCN-NEXT: v_add_i32_e64 v2, s[4:5], v2, v6 1196; GCN-NEXT: v_addc_u32_e32 v7, vcc, v12, v8, vcc 1197; GCN-NEXT: v_addc_u32_e64 v6, vcc, v3, v7, s[4:5] 1198; GCN-NEXT: v_mul_lo_u32 v8, v4, v6 1199; GCN-NEXT: v_mul_hi_u32 v9, v4, v2 1200; GCN-NEXT: v_mul_lo_u32 v5, v5, v2 1201; GCN-NEXT: v_mul_lo_u32 v4, v4, v2 1202; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1203; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 1204; GCN-NEXT: v_mul_lo_u32 v10, v2, v5 1205; GCN-NEXT: v_mul_hi_u32 v13, v2, v4 1206; GCN-NEXT: v_mul_hi_u32 v14, v2, v5 1207; GCN-NEXT: v_mul_hi_u32 v9, v6, v4 1208; GCN-NEXT: v_mul_lo_u32 v4, v6, v4 1209; GCN-NEXT: v_add_i32_e32 v10, vcc, v13, v10 1210; GCN-NEXT: v_mul_hi_u32 v8, v6, v5 1211; GCN-NEXT: v_addc_u32_e32 v13, vcc, v12, v14, vcc 1212; GCN-NEXT: v_mul_lo_u32 v5, v6, v5 1213; GCN-NEXT: v_add_i32_e32 v4, vcc, v10, v4 1214; GCN-NEXT: v_addc_u32_e32 v4, vcc, v13, v9, vcc 1215; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v11, vcc 1216; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 1217; GCN-NEXT: v_addc_u32_e32 v5, vcc, v12, v6, vcc 1218; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v7 1219; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v5, s[4:5] 1220; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 1221; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v3, vcc 1222; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v2 1223; GCN-NEXT: v_mul_lo_u32 v3, v1, v2 1224; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 1225; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 1226; GCN-NEXT: s_mov_b32 s4, 0x8000 1227; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1228; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 1229; GCN-NEXT: v_sub_i32_e32 v2, vcc, s4, v2 1230; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 1231; GCN-NEXT: v_sub_i32_e64 v5, s[4:5], v2, v0 1232; GCN-NEXT: v_subbrev_u32_e64 v6, s[6:7], 0, v4, s[4:5] 1233; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v6, v1 1234; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[6:7] 1235; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v5, v0 1236; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[6:7] 1237; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v6, v1 1238; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, s[4:5] 1239; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 1240; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[6:7] 1241; GCN-NEXT: v_sub_i32_e64 v8, s[4:5], v5, v0 1242; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 1243; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v1 1244; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v7 1245; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, vcc 1246; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 1247; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc 1248; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1 1249; GCN-NEXT: v_cndmask_b32_e32 v0, v7, v0, vcc 1250; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v8, s[4:5] 1251; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 1252; GCN-NEXT: v_cndmask_b32_e64 v1, v6, v4, s[4:5] 1253; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc 1254; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc 1255; GCN-NEXT: s_setpc_b64 s[30:31] 1256; 1257; GCN-IR-LABEL: v_test_urem_pow2_k_num_i64: 1258; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1259; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1260; GCN-IR-NEXT: v_ffbh_u32_e32 v2, v0 1261; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, 32, v2 1262; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v1 1263; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1 1264; GCN-IR-NEXT: v_cndmask_b32_e32 v8, v3, v2, vcc 1265; GCN-IR-NEXT: s_movk_i32 s6, 0xffd0 1266; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, s6, v8 1267; GCN-IR-NEXT: v_addc_u32_e64 v3, s[6:7], 0, -1, vcc 1268; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 1269; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[2:3] 1270; GCN-IR-NEXT: s_mov_b32 s8, 0x8000 1271; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], vcc 1272; GCN-IR-NEXT: v_mov_b32_e32 v4, s8 1273; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[2:3] 1274; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1275; GCN-IR-NEXT: v_cndmask_b32_e64 v4, v4, 0, s[4:5] 1276; GCN-IR-NEXT: s_xor_b64 s[4:5], s[4:5], -1 1277; GCN-IR-NEXT: s_mov_b32 s9, 0 1278; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1279; GCN-IR-NEXT: s_and_b64 s[4:5], s[4:5], vcc 1280; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1281; GCN-IR-NEXT: s_cbranch_execz BB8_6 1282; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1283; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 1, v2 1284; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, 0, v3, vcc 1285; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[4:5], v[2:3] 1286; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 63, v2 1287; GCN-IR-NEXT: v_lshl_b64 v[2:3], s[8:9], v2 1288; GCN-IR-NEXT: v_mov_b32_e32 v6, 0 1289; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 1290; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1291; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1292; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 1293; GCN-IR-NEXT: s_cbranch_execz BB8_5 1294; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1295; GCN-IR-NEXT: s_mov_b32 s5, 0 1296; GCN-IR-NEXT: s_mov_b32 s4, 0x8000 1297; GCN-IR-NEXT: v_lshr_b64 v[10:11], s[4:5], v4 1298; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, -1, v0 1299; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc 1300; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, 47, v8 1301; GCN-IR-NEXT: v_mov_b32_e32 v12, 0 1302; GCN-IR-NEXT: v_subb_u32_e32 v9, vcc, 0, v9, vcc 1303; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 1304; GCN-IR-NEXT: BB8_3: ; %udiv-do-while 1305; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1306; GCN-IR-NEXT: v_lshl_b64 v[10:11], v[10:11], 1 1307; GCN-IR-NEXT: v_lshrrev_b32_e32 v6, 31, v3 1308; GCN-IR-NEXT: v_or_b32_e32 v10, v10, v6 1309; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1310; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, v4, v10 1311; GCN-IR-NEXT: v_subb_u32_e32 v6, vcc, v5, v11, vcc 1312; GCN-IR-NEXT: v_or_b32_e32 v2, v12, v2 1313; GCN-IR-NEXT: v_ashrrev_i32_e32 v12, 31, v6 1314; GCN-IR-NEXT: v_and_b32_e32 v15, v12, v0 1315; GCN-IR-NEXT: v_and_b32_e32 v6, 1, v12 1316; GCN-IR-NEXT: v_and_b32_e32 v14, v12, v1 1317; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, 1, v8 1318; GCN-IR-NEXT: v_or_b32_e32 v3, v13, v3 1319; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, 0, v9, vcc 1320; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[12:13], v[8:9] 1321; GCN-IR-NEXT: v_mov_b32_e32 v8, v12 1322; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1323; GCN-IR-NEXT: v_sub_i32_e64 v10, s[4:5], v10, v15 1324; GCN-IR-NEXT: v_mov_b32_e32 v9, v13 1325; GCN-IR-NEXT: v_mov_b32_e32 v13, v7 1326; GCN-IR-NEXT: v_subb_u32_e64 v11, s[4:5], v11, v14, s[4:5] 1327; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 1328; GCN-IR-NEXT: v_mov_b32_e32 v12, v6 1329; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 1330; GCN-IR-NEXT: s_cbranch_execnz BB8_3 1331; GCN-IR-NEXT: ; %bb.4: ; %Flow 1332; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1333; GCN-IR-NEXT: BB8_5: ; %Flow3 1334; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1335; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1336; GCN-IR-NEXT: v_or_b32_e32 v5, v7, v3 1337; GCN-IR-NEXT: v_or_b32_e32 v4, v6, v2 1338; GCN-IR-NEXT: BB8_6: ; %Flow4 1339; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1340; GCN-IR-NEXT: v_mul_lo_u32 v2, v0, v5 1341; GCN-IR-NEXT: v_mul_hi_u32 v3, v0, v4 1342; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v4 1343; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, v4 1344; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1345; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1346; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 1347; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1348; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1349 %result = urem i64 32768, %x 1350 ret i64 %result 1351} 1352 1353define i64 @v_test_urem_pow2_k_den_i64(i64 %x) { 1354; GCN-LABEL: v_test_urem_pow2_k_den_i64: 1355; GCN: ; %bb.0: 1356; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1357; GCN-NEXT: v_and_b32_e32 v0, 0x7fff, v0 1358; GCN-NEXT: v_mov_b32_e32 v1, 0 1359; GCN-NEXT: s_setpc_b64 s[30:31] 1360; 1361; GCN-IR-LABEL: v_test_urem_pow2_k_den_i64: 1362; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1363; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1364; GCN-IR-NEXT: v_ffbh_u32_e32 v2, v0 1365; GCN-IR-NEXT: v_add_i32_e64 v2, s[4:5], 32, v2 1366; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v1 1367; GCN-IR-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v1 1368; GCN-IR-NEXT: v_cndmask_b32_e64 v6, v3, v2, s[4:5] 1369; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 48, v6 1370; GCN-IR-NEXT: v_subb_u32_e64 v3, s[4:5], 0, 0, s[4:5] 1371; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[0:1] 1372; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[4:5], 63, v[2:3] 1373; GCN-IR-NEXT: s_or_b64 s[4:5], vcc, s[4:5] 1374; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[2:3] 1375; GCN-IR-NEXT: s_xor_b64 s[6:7], s[4:5], -1 1376; GCN-IR-NEXT: v_cndmask_b32_e64 v5, v1, 0, s[4:5] 1377; GCN-IR-NEXT: v_cndmask_b32_e64 v4, v0, 0, s[4:5] 1378; GCN-IR-NEXT: s_and_b64 s[4:5], s[6:7], vcc 1379; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1380; GCN-IR-NEXT: s_cbranch_execz BB9_6 1381; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1382; GCN-IR-NEXT: v_add_i32_e32 v7, vcc, 1, v2 1383; GCN-IR-NEXT: v_addc_u32_e32 v8, vcc, 0, v3, vcc 1384; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[7:8], v[2:3] 1385; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 63, v2 1386; GCN-IR-NEXT: v_mov_b32_e32 v4, 0 1387; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[0:1], v2 1388; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 1389; GCN-IR-NEXT: v_mov_b32_e32 v5, 0 1390; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1391; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 1392; GCN-IR-NEXT: s_cbranch_execz BB9_5 1393; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1394; GCN-IR-NEXT: v_mov_b32_e32 v10, 0 1395; GCN-IR-NEXT: v_lshr_b64 v[8:9], v[0:1], v7 1396; GCN-IR-NEXT: v_add_i32_e32 v6, vcc, 0xffffffcf, v6 1397; GCN-IR-NEXT: v_addc_u32_e64 v7, s[4:5], 0, -1, vcc 1398; GCN-IR-NEXT: v_mov_b32_e32 v11, 0 1399; GCN-IR-NEXT: s_movk_i32 s12, 0x7fff 1400; GCN-IR-NEXT: BB9_3: ; %udiv-do-while 1401; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1402; GCN-IR-NEXT: v_lshl_b64 v[8:9], v[8:9], 1 1403; GCN-IR-NEXT: v_lshrrev_b32_e32 v4, 31, v3 1404; GCN-IR-NEXT: v_or_b32_e32 v8, v8, v4 1405; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1406; GCN-IR-NEXT: v_sub_i32_e32 v4, vcc, s12, v8 1407; GCN-IR-NEXT: v_subb_u32_e32 v4, vcc, 0, v9, vcc 1408; GCN-IR-NEXT: v_or_b32_e32 v2, v10, v2 1409; GCN-IR-NEXT: v_ashrrev_i32_e32 v10, 31, v4 1410; GCN-IR-NEXT: v_and_b32_e32 v12, 0x8000, v10 1411; GCN-IR-NEXT: v_and_b32_e32 v4, 1, v10 1412; GCN-IR-NEXT: v_add_i32_e32 v10, vcc, 1, v6 1413; GCN-IR-NEXT: v_or_b32_e32 v3, v11, v3 1414; GCN-IR-NEXT: v_addc_u32_e32 v11, vcc, 0, v7, vcc 1415; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[10:11], v[6:7] 1416; GCN-IR-NEXT: v_mov_b32_e32 v6, v10 1417; GCN-IR-NEXT: v_mov_b32_e32 v5, 0 1418; GCN-IR-NEXT: v_mov_b32_e32 v7, v11 1419; GCN-IR-NEXT: v_mov_b32_e32 v11, v5 1420; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 1421; GCN-IR-NEXT: v_sub_i32_e64 v8, s[4:5], v8, v12 1422; GCN-IR-NEXT: v_subb_u32_e64 v9, s[4:5], v9, v13, s[4:5] 1423; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 1424; GCN-IR-NEXT: v_mov_b32_e32 v10, v4 1425; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 1426; GCN-IR-NEXT: s_cbranch_execnz BB9_3 1427; GCN-IR-NEXT: ; %bb.4: ; %Flow 1428; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1429; GCN-IR-NEXT: BB9_5: ; %Flow3 1430; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1431; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1432; GCN-IR-NEXT: v_or_b32_e32 v5, v5, v3 1433; GCN-IR-NEXT: v_or_b32_e32 v4, v4, v2 1434; GCN-IR-NEXT: BB9_6: ; %Flow4 1435; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1436; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[4:5], 15 1437; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1438; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 1439; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1440 %result = urem i64 %x, 32768 1441 ret i64 %result 1442} 1443 1444define amdgpu_kernel void @s_test_urem24_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 1445; GCN-LABEL: s_test_urem24_k_num_i64: 1446; GCN: ; %bb.0: 1447; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1448; GCN-NEXT: s_mov_b32 s5, 0x41c00000 1449; GCN-NEXT: s_waitcnt lgkmcnt(0) 1450; GCN-NEXT: s_mov_b32 s2, -1 1451; GCN-NEXT: s_lshr_b32 s4, s3, 8 1452; GCN-NEXT: v_cvt_f32_u32_e32 v0, s4 1453; GCN-NEXT: s_mov_b32 s3, 0xf000 1454; GCN-NEXT: v_rcp_iflag_f32_e32 v1, v0 1455; GCN-NEXT: v_mul_f32_e32 v1, s5, v1 1456; GCN-NEXT: v_trunc_f32_e32 v1, v1 1457; GCN-NEXT: v_cvt_u32_f32_e32 v2, v1 1458; GCN-NEXT: v_mad_f32 v1, -v1, v0, s5 1459; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 1460; GCN-NEXT: v_mov_b32_e32 v1, 0 1461; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v2, vcc 1462; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 1463; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1464; GCN-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1465; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 1466; GCN-NEXT: s_endpgm 1467; 1468; GCN-IR-LABEL: s_test_urem24_k_num_i64: 1469; GCN-IR: ; %bb.0: 1470; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1471; GCN-IR-NEXT: s_mov_b32 s5, 0x41c00000 1472; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1473; GCN-IR-NEXT: s_mov_b32 s2, -1 1474; GCN-IR-NEXT: s_lshr_b32 s4, s3, 8 1475; GCN-IR-NEXT: v_cvt_f32_u32_e32 v0, s4 1476; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 1477; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v1, v0 1478; GCN-IR-NEXT: v_mul_f32_e32 v1, s5, v1 1479; GCN-IR-NEXT: v_trunc_f32_e32 v1, v1 1480; GCN-IR-NEXT: v_cvt_u32_f32_e32 v2, v1 1481; GCN-IR-NEXT: v_mad_f32 v1, -v1, v0, s5 1482; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, v0 1483; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1484; GCN-IR-NEXT: v_addc_u32_e32 v0, vcc, 0, v2, vcc 1485; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 1486; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1487; GCN-IR-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1488; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 1489; GCN-IR-NEXT: s_endpgm 1490 %x.shr = lshr i64 %x, 40 1491 %result = urem i64 24, %x.shr 1492 store i64 %result, i64 addrspace(1)* %out 1493 ret void 1494} 1495 1496define amdgpu_kernel void @s_test_urem24_k_den_i64(i64 addrspace(1)* %out, i64 %x) { 1497; GCN-LABEL: s_test_urem24_k_den_i64: 1498; GCN: ; %bb.0: 1499; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 1500; GCN-NEXT: s_mov_b32 s1, 0x46b6fe00 1501; GCN-NEXT: s_movk_i32 s0, 0x5b7f 1502; GCN-NEXT: s_mov_b32 s3, 0xf000 1503; GCN-NEXT: s_mov_b32 s2, -1 1504; GCN-NEXT: s_waitcnt lgkmcnt(0) 1505; GCN-NEXT: s_lshr_b32 s6, s7, 8 1506; GCN-NEXT: v_cvt_f32_u32_e32 v0, s6 1507; GCN-NEXT: v_mul_f32_e32 v1, 0x38331158, v0 1508; GCN-NEXT: v_trunc_f32_e32 v1, v1 1509; GCN-NEXT: v_cvt_u32_f32_e32 v2, v1 1510; GCN-NEXT: v_mad_f32 v0, -v1, s1, v0 1511; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s1 1512; GCN-NEXT: s_mov_b32 s1, s5 1513; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v2, vcc 1514; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 1515; GCN-NEXT: s_mov_b32 s0, s4 1516; GCN-NEXT: v_mov_b32_e32 v1, 0 1517; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 1518; GCN-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1519; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 1520; GCN-NEXT: s_endpgm 1521; 1522; GCN-IR-LABEL: s_test_urem24_k_den_i64: 1523; GCN-IR: ; %bb.0: 1524; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 1525; GCN-IR-NEXT: s_mov_b32 s1, 0x46b6fe00 1526; GCN-IR-NEXT: s_movk_i32 s0, 0x5b7f 1527; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 1528; GCN-IR-NEXT: s_mov_b32 s2, -1 1529; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1530; GCN-IR-NEXT: s_lshr_b32 s6, s7, 8 1531; GCN-IR-NEXT: v_cvt_f32_u32_e32 v0, s6 1532; GCN-IR-NEXT: v_mul_f32_e32 v1, 0x38331158, v0 1533; GCN-IR-NEXT: v_trunc_f32_e32 v1, v1 1534; GCN-IR-NEXT: v_cvt_u32_f32_e32 v2, v1 1535; GCN-IR-NEXT: v_mad_f32 v0, -v1, s1, v0 1536; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s1 1537; GCN-IR-NEXT: s_mov_b32 s1, s5 1538; GCN-IR-NEXT: v_addc_u32_e32 v0, vcc, 0, v2, vcc 1539; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 1540; GCN-IR-NEXT: s_mov_b32 s0, s4 1541; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1542; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 1543; GCN-IR-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1544; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 1545; GCN-IR-NEXT: s_endpgm 1546 %x.shr = lshr i64 %x, 40 1547 %result = urem i64 %x.shr, 23423 1548 store i64 %result, i64 addrspace(1)* %out 1549 ret void 1550} 1551 1552define i64 @v_test_urem24_k_num_i64(i64 %x) { 1553; GCN-LABEL: v_test_urem24_k_num_i64: 1554; GCN: ; %bb.0: 1555; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1556; GCN-NEXT: v_lshrrev_b32_e32 v0, 8, v1 1557; GCN-NEXT: v_cvt_f32_u32_e32 v1, v0 1558; GCN-NEXT: s_mov_b32 s4, 0x41c00000 1559; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v1 1560; GCN-NEXT: v_mul_f32_e32 v2, s4, v2 1561; GCN-NEXT: v_trunc_f32_e32 v2, v2 1562; GCN-NEXT: v_cvt_u32_f32_e32 v3, v2 1563; GCN-NEXT: v_mad_f32 v2, -v2, v1, s4 1564; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v1 1565; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc 1566; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 1567; GCN-NEXT: v_mov_b32_e32 v1, 0 1568; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1569; GCN-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1570; GCN-NEXT: s_setpc_b64 s[30:31] 1571; 1572; GCN-IR-LABEL: v_test_urem24_k_num_i64: 1573; GCN-IR: ; %bb.0: 1574; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1575; GCN-IR-NEXT: v_lshrrev_b32_e32 v0, 8, v1 1576; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, v0 1577; GCN-IR-NEXT: s_mov_b32 s4, 0x41c00000 1578; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v1 1579; GCN-IR-NEXT: v_mul_f32_e32 v2, s4, v2 1580; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 1581; GCN-IR-NEXT: v_cvt_u32_f32_e32 v3, v2 1582; GCN-IR-NEXT: v_mad_f32 v2, -v2, v1, s4 1583; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v1 1584; GCN-IR-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc 1585; GCN-IR-NEXT: v_mul_lo_u32 v0, v1, v0 1586; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1587; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1588; GCN-IR-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1589; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1590 %x.shr = lshr i64 %x, 40 1591 %result = urem i64 24, %x.shr 1592 ret i64 %result 1593} 1594 1595define i64 @v_test_urem24_pow2_k_num_i64(i64 %x) { 1596; GCN-LABEL: v_test_urem24_pow2_k_num_i64: 1597; GCN: ; %bb.0: 1598; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1599; GCN-NEXT: v_lshrrev_b32_e32 v0, 8, v1 1600; GCN-NEXT: v_cvt_f32_u32_e32 v1, v0 1601; GCN-NEXT: s_mov_b32 s4, 0x47000000 1602; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v1 1603; GCN-NEXT: v_mul_f32_e32 v2, s4, v2 1604; GCN-NEXT: v_trunc_f32_e32 v2, v2 1605; GCN-NEXT: v_cvt_u32_f32_e32 v3, v2 1606; GCN-NEXT: v_mad_f32 v2, -v2, v1, s4 1607; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v1 1608; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc 1609; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 1610; GCN-NEXT: v_mov_b32_e32 v1, 0 1611; GCN-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 1612; GCN-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1613; GCN-NEXT: s_setpc_b64 s[30:31] 1614; 1615; GCN-IR-LABEL: v_test_urem24_pow2_k_num_i64: 1616; GCN-IR: ; %bb.0: 1617; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1618; GCN-IR-NEXT: v_lshrrev_b32_e32 v0, 8, v1 1619; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, v0 1620; GCN-IR-NEXT: s_mov_b32 s4, 0x47000000 1621; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v1 1622; GCN-IR-NEXT: v_mul_f32_e32 v2, s4, v2 1623; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 1624; GCN-IR-NEXT: v_cvt_u32_f32_e32 v3, v2 1625; GCN-IR-NEXT: v_mad_f32 v2, -v2, v1, s4 1626; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, v1 1627; GCN-IR-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc 1628; GCN-IR-NEXT: v_mul_lo_u32 v0, v1, v0 1629; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1630; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 1631; GCN-IR-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1632; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1633 %x.shr = lshr i64 %x, 40 1634 %result = urem i64 32768, %x.shr 1635 ret i64 %result 1636} 1637 1638define i64 @v_test_urem24_pow2_k_den_i64(i64 %x) { 1639; GCN-LABEL: v_test_urem24_pow2_k_den_i64: 1640; GCN: ; %bb.0: 1641; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1642; GCN-NEXT: v_bfe_u32 v0, v1, 8, 15 1643; GCN-NEXT: v_mov_b32_e32 v1, 0 1644; GCN-NEXT: s_setpc_b64 s[30:31] 1645; 1646; GCN-IR-LABEL: v_test_urem24_pow2_k_den_i64: 1647; GCN-IR: ; %bb.0: 1648; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1649; GCN-IR-NEXT: v_lshrrev_b32_e32 v0, 8, v1 1650; GCN-IR-NEXT: v_cvt_f32_u32_e32 v1, v0 1651; GCN-IR-NEXT: s_mov_b32 s4, 0x47000000 1652; GCN-IR-NEXT: v_mul_f32_e32 v2, 0x38000000, v1 1653; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 1654; GCN-IR-NEXT: v_cvt_u32_f32_e32 v3, v2 1655; GCN-IR-NEXT: v_mad_f32 v1, -v2, s4, v1 1656; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, s4 1657; GCN-IR-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc 1658; GCN-IR-NEXT: v_lshlrev_b32_e32 v1, 15, v1 1659; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 1660; GCN-IR-NEXT: v_and_b32_e32 v0, 0xffffff, v0 1661; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1662; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1663 %x.shr = lshr i64 %x, 40 1664 %result = urem i64 %x.shr, 32768 1665 ret i64 %result 1666} 1667