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_srem(i64 addrspace(1)* %out, i64 %x, i64 %y) { 6; GCN-LABEL: s_test_srem: 7; GCN: ; %bb.0: 8; GCN-NEXT: s_load_dwordx2 s[12:13], s[0:1], 0xd 9; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0x9 10; GCN-NEXT: s_mov_b32 s7, 0xf000 11; GCN-NEXT: s_mov_b32 s6, -1 12; GCN-NEXT: s_waitcnt lgkmcnt(0) 13; GCN-NEXT: v_cvt_f32_u32_e32 v0, s12 14; GCN-NEXT: v_cvt_f32_u32_e32 v1, s13 15; GCN-NEXT: s_sub_u32 s0, 0, s12 16; GCN-NEXT: s_subb_u32 s1, 0, s13 17; GCN-NEXT: s_mov_b32 s4, s8 18; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 19; GCN-NEXT: v_rcp_f32_e32 v0, v0 20; GCN-NEXT: s_mov_b32 s5, s9 21; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 22; GCN-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0 23; GCN-NEXT: v_trunc_f32_e32 v1, v1 24; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v1 25; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1 26; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 27; GCN-NEXT: v_mul_lo_u32 v2, s0, v1 28; GCN-NEXT: v_mul_hi_u32 v3, s0, v0 29; GCN-NEXT: v_mul_lo_u32 v5, s1, v0 30; GCN-NEXT: v_mul_lo_u32 v4, s0, v0 31; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 32; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v5 33; GCN-NEXT: v_mul_hi_u32 v3, v0, v4 34; GCN-NEXT: v_mul_lo_u32 v5, v0, v2 35; GCN-NEXT: v_mul_hi_u32 v7, v0, v2 36; GCN-NEXT: v_mul_hi_u32 v6, v1, v4 37; GCN-NEXT: v_mul_lo_u32 v4, v1, v4 38; GCN-NEXT: v_mul_hi_u32 v8, v1, v2 39; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 40; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v7, vcc 41; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 42; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v4 43; GCN-NEXT: v_addc_u32_e32 v3, vcc, v5, v6, vcc 44; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v8, vcc 45; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 46; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 47; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 48; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v3, vcc 49; GCN-NEXT: v_mul_lo_u32 v2, s0, v1 50; GCN-NEXT: v_mul_hi_u32 v3, s0, v0 51; GCN-NEXT: v_mul_lo_u32 v4, s1, v0 52; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 53; GCN-NEXT: v_mul_lo_u32 v3, s0, v0 54; GCN-NEXT: v_add_i32_e32 v2, vcc, v4, v2 55; GCN-NEXT: v_mul_lo_u32 v6, v0, v2 56; GCN-NEXT: v_mul_hi_u32 v7, v0, v3 57; GCN-NEXT: v_mul_hi_u32 v8, v0, v2 58; GCN-NEXT: v_mul_hi_u32 v5, v1, v3 59; GCN-NEXT: v_mul_lo_u32 v3, v1, v3 60; GCN-NEXT: v_mul_hi_u32 v4, v1, v2 61; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 62; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v8, vcc 63; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 64; GCN-NEXT: v_add_i32_e32 v3, vcc, v6, v3 65; GCN-NEXT: v_addc_u32_e32 v3, vcc, v7, v5, vcc 66; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc 67; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 68; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 69; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 70; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v3, vcc 71; GCN-NEXT: v_mul_lo_u32 v2, s10, v1 72; GCN-NEXT: v_mul_hi_u32 v3, s10, v0 73; GCN-NEXT: v_mul_hi_u32 v4, s10, v1 74; GCN-NEXT: v_mul_hi_u32 v5, s11, v1 75; GCN-NEXT: v_mul_lo_u32 v1, s11, v1 76; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 77; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 78; GCN-NEXT: v_mul_lo_u32 v4, s11, v0 79; GCN-NEXT: v_mul_hi_u32 v0, s11, v0 80; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 81; GCN-NEXT: v_addc_u32_e32 v0, vcc, v3, v0, vcc 82; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v5, vcc 83; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 84; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v2, vcc 85; GCN-NEXT: v_mul_lo_u32 v1, s12, v1 86; GCN-NEXT: v_mul_hi_u32 v2, s12, v0 87; GCN-NEXT: v_mul_lo_u32 v3, s13, v0 88; GCN-NEXT: v_mul_lo_u32 v0, s12, v0 89; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 90; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 91; GCN-NEXT: v_sub_i32_e32 v2, vcc, s11, v1 92; GCN-NEXT: v_mov_b32_e32 v3, s13 93; GCN-NEXT: v_sub_i32_e32 v0, vcc, s10, v0 94; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 95; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s12, v0 96; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 97; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s13, v5 98; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 99; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s12, v4 100; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 101; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 102; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s13, v5 103; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s12, v4 104; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 105; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 106; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 107; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 108; GCN-NEXT: v_mov_b32_e32 v5, s11 109; GCN-NEXT: v_subb_u32_e32 v1, vcc, v5, v1, vcc 110; GCN-NEXT: v_cmp_le_u32_e32 vcc, s13, v1 111; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 112; GCN-NEXT: v_cmp_le_u32_e32 vcc, s12, v0 113; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 114; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s13, v1 115; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 116; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 117; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 118; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 119; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 120; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 121; GCN-NEXT: s_endpgm 122; 123; GCN-IR-LABEL: s_test_srem: 124; GCN-IR: ; %bb.0: ; %_udiv-special-cases 125; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 126; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 127; GCN-IR-NEXT: s_mov_b64 s[6:7], 0 128; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 129; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[8:9], s[4:5], 0 130; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[10:11], s[2:3], 0 131; GCN-IR-NEXT: s_flbit_i32_b32 s12, s4 132; GCN-IR-NEXT: s_add_i32 s14, s12, 32 133; GCN-IR-NEXT: s_or_b64 s[12:13], s[8:9], s[10:11] 134; GCN-IR-NEXT: s_flbit_i32_b32 s8, s5 135; GCN-IR-NEXT: s_min_u32 s10, s14, s8 136; GCN-IR-NEXT: s_flbit_i32_b32 s8, s2 137; GCN-IR-NEXT: s_add_i32 s8, s8, 32 138; GCN-IR-NEXT: s_flbit_i32_b32 s9, s3 139; GCN-IR-NEXT: s_min_u32 s14, s8, s9 140; GCN-IR-NEXT: s_sub_u32 s8, s10, s14 141; GCN-IR-NEXT: s_subb_u32 s9, 0, 0 142; GCN-IR-NEXT: v_cmp_gt_u64_e64 s[16:17], s[8:9], 63 143; GCN-IR-NEXT: s_mov_b32 s11, 0 144; GCN-IR-NEXT: s_or_b64 s[12:13], s[12:13], s[16:17] 145; GCN-IR-NEXT: v_cmp_ne_u64_e64 s[16:17], s[8:9], 63 146; GCN-IR-NEXT: s_xor_b64 s[18:19], s[12:13], -1 147; GCN-IR-NEXT: s_and_b64 s[16:17], s[18:19], s[16:17] 148; GCN-IR-NEXT: s_and_b64 vcc, exec, s[16:17] 149; GCN-IR-NEXT: s_cbranch_vccz .LBB0_5 150; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 151; GCN-IR-NEXT: s_add_u32 s12, s8, 1 152; GCN-IR-NEXT: s_addc_u32 s13, s9, 0 153; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[16:17], s[12:13], 0 154; GCN-IR-NEXT: s_sub_i32 s8, 63, s8 155; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[16:17] 156; GCN-IR-NEXT: s_lshl_b64 s[8:9], s[2:3], s8 157; GCN-IR-NEXT: s_cbranch_vccz .LBB0_4 158; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 159; GCN-IR-NEXT: s_lshr_b64 s[12:13], s[2:3], s12 160; GCN-IR-NEXT: s_add_u32 s16, s4, -1 161; GCN-IR-NEXT: s_addc_u32 s17, s5, -1 162; GCN-IR-NEXT: s_not_b64 s[6:7], s[10:11] 163; GCN-IR-NEXT: s_add_u32 s10, s6, s14 164; GCN-IR-NEXT: s_addc_u32 s11, s7, 0 165; GCN-IR-NEXT: s_mov_b64 s[14:15], 0 166; GCN-IR-NEXT: s_mov_b32 s7, 0 167; GCN-IR-NEXT: .LBB0_3: ; %udiv-do-while 168; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 169; GCN-IR-NEXT: s_lshl_b64 s[12:13], s[12:13], 1 170; GCN-IR-NEXT: s_lshr_b32 s6, s9, 31 171; GCN-IR-NEXT: s_lshl_b64 s[8:9], s[8:9], 1 172; GCN-IR-NEXT: s_or_b64 s[12:13], s[12:13], s[6:7] 173; GCN-IR-NEXT: s_or_b64 s[8:9], s[14:15], s[8:9] 174; GCN-IR-NEXT: s_sub_u32 s6, s16, s12 175; GCN-IR-NEXT: s_subb_u32 s6, s17, s13 176; GCN-IR-NEXT: s_ashr_i32 s14, s6, 31 177; GCN-IR-NEXT: s_mov_b32 s15, s14 178; GCN-IR-NEXT: s_and_b32 s6, s14, 1 179; GCN-IR-NEXT: s_and_b64 s[14:15], s[14:15], s[4:5] 180; GCN-IR-NEXT: s_sub_u32 s12, s12, s14 181; GCN-IR-NEXT: s_subb_u32 s13, s13, s15 182; GCN-IR-NEXT: s_add_u32 s10, s10, 1 183; GCN-IR-NEXT: s_addc_u32 s11, s11, 0 184; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[18:19], s[10:11], 0 185; GCN-IR-NEXT: s_mov_b64 s[14:15], s[6:7] 186; GCN-IR-NEXT: s_and_b64 vcc, exec, s[18:19] 187; GCN-IR-NEXT: s_cbranch_vccz .LBB0_3 188; GCN-IR-NEXT: .LBB0_4: ; %Flow6 189; GCN-IR-NEXT: s_lshl_b64 s[8:9], s[8:9], 1 190; GCN-IR-NEXT: s_or_b64 s[6:7], s[6:7], s[8:9] 191; GCN-IR-NEXT: v_mov_b32_e32 v0, s6 192; GCN-IR-NEXT: v_mov_b32_e32 v1, s7 193; GCN-IR-NEXT: s_branch .LBB0_6 194; GCN-IR-NEXT: .LBB0_5: 195; GCN-IR-NEXT: v_mov_b32_e32 v0, s3 196; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[12:13] 197; GCN-IR-NEXT: v_mov_b32_e32 v0, s2 198; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[12:13] 199; GCN-IR-NEXT: .LBB0_6: ; %udiv-end 200; GCN-IR-NEXT: v_mul_lo_u32 v1, s4, v1 201; GCN-IR-NEXT: v_mul_hi_u32 v2, s4, v0 202; GCN-IR-NEXT: v_mul_lo_u32 v3, s5, v0 203; GCN-IR-NEXT: v_mul_lo_u32 v0, s4, v0 204; GCN-IR-NEXT: s_mov_b32 s11, 0xf000 205; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 206; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 207; GCN-IR-NEXT: v_mov_b32_e32 v2, s3 208; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 209; GCN-IR-NEXT: s_mov_b32 s10, -1 210; GCN-IR-NEXT: s_mov_b32 s8, s0 211; GCN-IR-NEXT: s_mov_b32 s9, s1 212; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 213; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 214; GCN-IR-NEXT: s_endpgm 215 %result = urem i64 %x, %y 216 store i64 %result, i64 addrspace(1)* %out 217 ret void 218} 219 220define i64 @v_test_srem(i64 %x, i64 %y) { 221; GCN-LABEL: v_test_srem: 222; GCN: ; %bb.0: 223; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 224; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v3 225; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 226; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v4, vcc 227; GCN-NEXT: v_xor_b32_e32 v3, v3, v4 228; GCN-NEXT: v_xor_b32_e32 v2, v2, v4 229; GCN-NEXT: v_cvt_f32_u32_e32 v4, v2 230; GCN-NEXT: v_cvt_f32_u32_e32 v5, v3 231; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v2 232; GCN-NEXT: v_subb_u32_e32 v7, vcc, 0, v3, vcc 233; GCN-NEXT: v_mac_f32_e32 v4, 0x4f800000, v5 234; GCN-NEXT: v_rcp_f32_e32 v4, v4 235; GCN-NEXT: v_mul_f32_e32 v4, 0x5f7ffffc, v4 236; GCN-NEXT: v_mul_f32_e32 v5, 0x2f800000, v4 237; GCN-NEXT: v_trunc_f32_e32 v5, v5 238; GCN-NEXT: v_mac_f32_e32 v4, 0xcf800000, v5 239; GCN-NEXT: v_cvt_u32_f32_e32 v4, v4 240; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5 241; GCN-NEXT: v_mul_hi_u32 v8, v6, v4 242; GCN-NEXT: v_mul_lo_u32 v9, v6, v5 243; GCN-NEXT: v_mul_lo_u32 v10, v7, v4 244; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v9 245; GCN-NEXT: v_mul_lo_u32 v9, v6, v4 246; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v10 247; GCN-NEXT: v_mul_lo_u32 v10, v4, v8 248; GCN-NEXT: v_mul_hi_u32 v11, v4, v9 249; GCN-NEXT: v_mul_hi_u32 v12, v4, v8 250; GCN-NEXT: v_mul_hi_u32 v13, v5, v8 251; GCN-NEXT: v_mul_lo_u32 v8, v5, v8 252; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 253; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v12, vcc 254; GCN-NEXT: v_mul_lo_u32 v12, v5, v9 255; GCN-NEXT: v_mul_hi_u32 v9, v5, v9 256; GCN-NEXT: v_add_i32_e32 v10, vcc, v10, v12 257; GCN-NEXT: v_addc_u32_e32 v9, vcc, v11, v9, vcc 258; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v13, vcc 259; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 260; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v10, vcc 261; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v8 262; GCN-NEXT: v_addc_u32_e32 v5, vcc, v5, v9, vcc 263; GCN-NEXT: v_mul_lo_u32 v8, v6, v5 264; GCN-NEXT: v_mul_hi_u32 v9, v6, v4 265; GCN-NEXT: v_mul_lo_u32 v7, v7, v4 266; GCN-NEXT: v_mul_lo_u32 v6, v6, v4 267; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 268; GCN-NEXT: v_add_i32_e32 v7, vcc, v8, v7 269; GCN-NEXT: v_mul_lo_u32 v10, v4, v7 270; GCN-NEXT: v_mul_hi_u32 v11, v4, v6 271; GCN-NEXT: v_mul_hi_u32 v12, v4, v7 272; GCN-NEXT: v_mul_hi_u32 v9, v5, v6 273; GCN-NEXT: v_mul_lo_u32 v6, v5, v6 274; GCN-NEXT: v_mul_hi_u32 v8, v5, v7 275; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 276; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v12, vcc 277; GCN-NEXT: v_mul_lo_u32 v7, v5, v7 278; GCN-NEXT: v_add_i32_e32 v6, vcc, v10, v6 279; GCN-NEXT: v_addc_u32_e32 v6, vcc, v11, v9, vcc 280; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc 281; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 282; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v8, vcc 283; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v6 284; GCN-NEXT: v_addc_u32_e32 v5, vcc, v5, v7, vcc 285; GCN-NEXT: v_ashrrev_i32_e32 v6, 31, v1 286; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v6 287; GCN-NEXT: v_xor_b32_e32 v0, v0, v6 288; GCN-NEXT: v_mul_lo_u32 v7, v0, v5 289; GCN-NEXT: v_mul_hi_u32 v8, v0, v4 290; GCN-NEXT: v_mul_hi_u32 v9, v0, v5 291; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v6, vcc 292; GCN-NEXT: v_xor_b32_e32 v1, v1, v6 293; GCN-NEXT: v_add_i32_e32 v7, vcc, v8, v7 294; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v9, vcc 295; GCN-NEXT: v_mul_lo_u32 v9, v1, v4 296; GCN-NEXT: v_mul_hi_u32 v4, v1, v4 297; GCN-NEXT: v_mul_hi_u32 v10, v1, v5 298; GCN-NEXT: v_mul_lo_u32 v5, v1, v5 299; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v9 300; GCN-NEXT: v_addc_u32_e32 v4, vcc, v8, v4, vcc 301; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v10, vcc 302; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 303; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v7, vcc 304; GCN-NEXT: v_mul_lo_u32 v5, v2, v5 305; GCN-NEXT: v_mul_hi_u32 v7, v2, v4 306; GCN-NEXT: v_mul_lo_u32 v8, v3, v4 307; GCN-NEXT: v_mul_lo_u32 v4, v2, v4 308; GCN-NEXT: v_add_i32_e32 v5, vcc, v7, v5 309; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v8 310; GCN-NEXT: v_sub_i32_e32 v7, vcc, v1, v5 311; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 312; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v7, v3, vcc 313; GCN-NEXT: v_sub_i32_e64 v7, s[4:5], v0, v2 314; GCN-NEXT: v_subbrev_u32_e64 v8, s[6:7], 0, v4, s[4:5] 315; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v8, v3 316; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[6:7] 317; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v7, v2 318; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 319; GCN-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[6:7] 320; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v8, v3 321; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v3, s[4:5] 322; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3 323; GCN-NEXT: v_cndmask_b32_e64 v9, v9, v10, s[6:7] 324; GCN-NEXT: v_sub_i32_e64 v10, s[4:5], v7, v2 325; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 326; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2 327; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 328; GCN-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc 329; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 330; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v9 331; GCN-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc 332; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 333; GCN-NEXT: v_cndmask_b32_e64 v2, v7, v10, s[4:5] 334; GCN-NEXT: v_cndmask_b32_e64 v4, v8, v4, s[4:5] 335; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 336; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc 337; GCN-NEXT: v_xor_b32_e32 v0, v0, v6 338; GCN-NEXT: v_xor_b32_e32 v1, v1, v6 339; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v6 340; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v6, vcc 341; GCN-NEXT: s_setpc_b64 s[30:31] 342; 343; GCN-IR-LABEL: v_test_srem: 344; GCN-IR: ; %bb.0: ; %_udiv-special-cases 345; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 346; GCN-IR-NEXT: v_ashrrev_i32_e32 v4, 31, v1 347; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v4 348; GCN-IR-NEXT: v_ashrrev_i32_e32 v6, 31, v3 349; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v4 350; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 351; GCN-IR-NEXT: v_xor_b32_e32 v2, v2, v6 352; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v4, vcc 353; GCN-IR-NEXT: v_xor_b32_e32 v3, v3, v6 354; GCN-IR-NEXT: v_sub_i32_e32 v2, vcc, v2, v6 355; GCN-IR-NEXT: v_subb_u32_e32 v3, vcc, v3, v6, vcc 356; GCN-IR-NEXT: v_ffbh_u32_e32 v6, v2 357; GCN-IR-NEXT: v_add_i32_e64 v6, s[6:7], 32, v6 358; GCN-IR-NEXT: v_ffbh_u32_e32 v7, v3 359; GCN-IR-NEXT: v_min_u32_e32 v10, v6, v7 360; GCN-IR-NEXT: v_ffbh_u32_e32 v6, v0 361; GCN-IR-NEXT: v_add_i32_e64 v6, s[6:7], 32, v6 362; GCN-IR-NEXT: v_ffbh_u32_e32 v7, v1 363; GCN-IR-NEXT: v_min_u32_e32 v11, v6, v7 364; GCN-IR-NEXT: v_sub_i32_e64 v7, s[6:7], v10, v11 365; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[2:3] 366; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 367; GCN-IR-NEXT: v_subb_u32_e64 v8, s[6:7], 0, 0, s[6:7] 368; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[6:7], 63, v[7:8] 369; GCN-IR-NEXT: s_or_b64 s[4:5], vcc, s[4:5] 370; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], s[6:7] 371; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[7:8] 372; GCN-IR-NEXT: s_xor_b64 s[6:7], s[4:5], -1 373; GCN-IR-NEXT: v_mov_b32_e32 v5, v4 374; GCN-IR-NEXT: v_cndmask_b32_e64 v9, v1, 0, s[4:5] 375; GCN-IR-NEXT: v_cndmask_b32_e64 v6, v0, 0, s[4:5] 376; GCN-IR-NEXT: s_and_b64 s[4:5], s[6:7], vcc 377; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 378; GCN-IR-NEXT: s_cbranch_execz .LBB1_6 379; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 380; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, 1, v7 381; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, 0, v8, vcc 382; GCN-IR-NEXT: v_sub_i32_e64 v6, s[4:5], 63, v7 383; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[12:13] 384; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[0:1], v6 385; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 386; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 387; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 388; GCN-IR-NEXT: s_xor_b64 s[8:9], exec, s[4:5] 389; GCN-IR-NEXT: s_cbranch_execz .LBB1_5 390; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 391; GCN-IR-NEXT: v_add_i32_e32 v16, vcc, -1, v2 392; GCN-IR-NEXT: v_addc_u32_e32 v17, vcc, -1, v3, vcc 393; GCN-IR-NEXT: v_not_b32_e32 v9, v10 394; GCN-IR-NEXT: v_lshr_b64 v[12:13], v[0:1], v12 395; GCN-IR-NEXT: v_not_b32_e32 v8, 0 396; GCN-IR-NEXT: v_add_i32_e32 v10, vcc, v9, v11 397; GCN-IR-NEXT: v_mov_b32_e32 v14, 0 398; GCN-IR-NEXT: v_addc_u32_e32 v11, vcc, 0, v8, vcc 399; GCN-IR-NEXT: s_mov_b64 s[10:11], 0 400; GCN-IR-NEXT: v_mov_b32_e32 v15, 0 401; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 402; GCN-IR-NEXT: .LBB1_3: ; %udiv-do-while 403; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 404; GCN-IR-NEXT: v_lshl_b64 v[12:13], v[12:13], 1 405; GCN-IR-NEXT: v_lshrrev_b32_e32 v8, 31, v7 406; GCN-IR-NEXT: v_or_b32_e32 v12, v12, v8 407; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 408; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, v16, v12 409; GCN-IR-NEXT: v_subb_u32_e32 v8, vcc, v17, v13, vcc 410; GCN-IR-NEXT: v_or_b32_e32 v6, v14, v6 411; GCN-IR-NEXT: v_ashrrev_i32_e32 v14, 31, v8 412; GCN-IR-NEXT: v_add_i32_e32 v10, vcc, 1, v10 413; GCN-IR-NEXT: v_or_b32_e32 v7, v15, v7 414; GCN-IR-NEXT: v_and_b32_e32 v8, 1, v14 415; GCN-IR-NEXT: v_and_b32_e32 v15, v14, v3 416; GCN-IR-NEXT: v_and_b32_e32 v14, v14, v2 417; GCN-IR-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc 418; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[10:11] 419; GCN-IR-NEXT: v_sub_i32_e64 v12, s[4:5], v12, v14 420; GCN-IR-NEXT: v_subb_u32_e64 v13, s[4:5], v13, v15, s[4:5] 421; GCN-IR-NEXT: v_mov_b32_e32 v15, v9 422; GCN-IR-NEXT: s_or_b64 s[10:11], vcc, s[10:11] 423; GCN-IR-NEXT: v_mov_b32_e32 v14, v8 424; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[10:11] 425; GCN-IR-NEXT: s_cbranch_execnz .LBB1_3 426; GCN-IR-NEXT: ; %bb.4: ; %Flow 427; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 428; GCN-IR-NEXT: .LBB1_5: ; %Flow3 429; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 430; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 431; GCN-IR-NEXT: v_or_b32_e32 v9, v9, v7 432; GCN-IR-NEXT: v_or_b32_e32 v6, v8, v6 433; GCN-IR-NEXT: .LBB1_6: ; %Flow4 434; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 435; GCN-IR-NEXT: v_mul_lo_u32 v7, v2, v9 436; GCN-IR-NEXT: v_mul_hi_u32 v8, v2, v6 437; GCN-IR-NEXT: v_mul_lo_u32 v3, v3, v6 438; GCN-IR-NEXT: v_mul_lo_u32 v2, v2, v6 439; GCN-IR-NEXT: v_add_i32_e32 v7, vcc, v8, v7 440; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, v7, 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: v_xor_b32_e32 v0, v0, v4 444; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v5 445; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 446; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 447; GCN-IR-NEXT: s_setpc_b64 s[30:31] 448 %result = srem i64 %x, %y 449 ret i64 %result 450} 451 452define amdgpu_kernel void @s_test_srem23_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 453; GCN-LABEL: s_test_srem23_64: 454; GCN: ; %bb.0: 455; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 456; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 457; GCN-NEXT: s_mov_b32 s7, 0xf000 458; GCN-NEXT: s_mov_b32 s6, -1 459; GCN-NEXT: s_waitcnt lgkmcnt(0) 460; GCN-NEXT: s_ashr_i64 s[4:5], s[4:5], 41 461; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 462; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 41 463; GCN-NEXT: v_cvt_f32_i32_e32 v1, s2 464; GCN-NEXT: s_xor_b32 s3, s2, s4 465; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 466; GCN-NEXT: s_ashr_i32 s3, s3, 30 467; GCN-NEXT: s_or_b32 s3, s3, 1 468; GCN-NEXT: v_mov_b32_e32 v3, s3 469; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 470; GCN-NEXT: v_trunc_f32_e32 v2, v2 471; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 472; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 473; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 474; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 475; GCN-NEXT: s_mov_b32 s5, s1 476; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 477; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 478; GCN-NEXT: s_mov_b32 s4, s0 479; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 480; GCN-NEXT: v_bfe_i32 v0, v0, 0, 23 481; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 482; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 483; GCN-NEXT: s_endpgm 484; 485; GCN-IR-LABEL: s_test_srem23_64: 486; GCN-IR: ; %bb.0: 487; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 488; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 489; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 490; GCN-IR-NEXT: s_mov_b32 s6, -1 491; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 492; GCN-IR-NEXT: s_ashr_i64 s[4:5], s[4:5], 41 493; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 494; GCN-IR-NEXT: s_ashr_i64 s[2:3], s[2:3], 41 495; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s2 496; GCN-IR-NEXT: s_xor_b32 s3, s2, s4 497; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 498; GCN-IR-NEXT: s_ashr_i32 s3, s3, 30 499; GCN-IR-NEXT: s_or_b32 s3, s3, 1 500; GCN-IR-NEXT: v_mov_b32_e32 v3, s3 501; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 502; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 503; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 504; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 505; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 506; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 507; GCN-IR-NEXT: s_mov_b32 s5, s1 508; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 509; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 510; GCN-IR-NEXT: s_mov_b32 s4, s0 511; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 512; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 23 513; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 514; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 515; GCN-IR-NEXT: s_endpgm 516 %1 = ashr i64 %x, 41 517 %2 = ashr i64 %y, 41 518 %result = srem i64 %1, %2 519 store i64 %result, i64 addrspace(1)* %out 520 ret void 521} 522 523define amdgpu_kernel void @s_test_srem24_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 524; GCN-LABEL: s_test_srem24_64: 525; GCN: ; %bb.0: 526; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 527; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 528; GCN-NEXT: s_mov_b32 s7, 0xf000 529; GCN-NEXT: s_mov_b32 s6, -1 530; GCN-NEXT: s_waitcnt lgkmcnt(0) 531; GCN-NEXT: s_ashr_i64 s[4:5], s[4:5], 40 532; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 533; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 40 534; GCN-NEXT: v_cvt_f32_i32_e32 v1, s2 535; GCN-NEXT: s_xor_b32 s3, s2, s4 536; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 537; GCN-NEXT: s_ashr_i32 s3, s3, 30 538; GCN-NEXT: s_or_b32 s3, s3, 1 539; GCN-NEXT: v_mov_b32_e32 v3, s3 540; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 541; GCN-NEXT: v_trunc_f32_e32 v2, v2 542; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 543; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 544; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 545; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 546; GCN-NEXT: s_mov_b32 s5, s1 547; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 548; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 549; GCN-NEXT: s_mov_b32 s4, s0 550; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 551; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 552; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 553; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 554; GCN-NEXT: s_endpgm 555; 556; GCN-IR-LABEL: s_test_srem24_64: 557; GCN-IR: ; %bb.0: 558; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 559; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 560; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 561; GCN-IR-NEXT: s_mov_b32 s6, -1 562; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 563; GCN-IR-NEXT: s_ashr_i64 s[4:5], s[4:5], 40 564; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 565; GCN-IR-NEXT: s_ashr_i64 s[2:3], s[2:3], 40 566; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s2 567; GCN-IR-NEXT: s_xor_b32 s3, s2, s4 568; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 569; GCN-IR-NEXT: s_ashr_i32 s3, s3, 30 570; GCN-IR-NEXT: s_or_b32 s3, s3, 1 571; GCN-IR-NEXT: v_mov_b32_e32 v3, s3 572; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 573; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 574; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 575; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 576; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 577; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 578; GCN-IR-NEXT: s_mov_b32 s5, s1 579; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 580; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 581; GCN-IR-NEXT: s_mov_b32 s4, s0 582; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 583; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 584; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 585; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 586; GCN-IR-NEXT: s_endpgm 587 %1 = ashr i64 %x, 40 588 %2 = ashr i64 %y, 40 589 %result = srem i64 %1, %2 590 store i64 %result, i64 addrspace(1)* %out 591 ret void 592} 593 594define i64 @v_test_srem24_64(i64 %x, i64 %y) { 595; GCN-LABEL: v_test_srem24_64: 596; GCN: ; %bb.0: 597; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 598; GCN-NEXT: v_ashr_i64 v[2:3], v[2:3], 40 599; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 600; GCN-NEXT: v_cvt_f32_i32_e32 v3, v2 601; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 602; GCN-NEXT: v_xor_b32_e32 v5, v0, v2 603; GCN-NEXT: v_ashrrev_i32_e32 v5, 30, v5 604; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v3 605; GCN-NEXT: v_or_b32_e32 v5, 1, v5 606; GCN-NEXT: v_mul_f32_e32 v4, v1, v4 607; GCN-NEXT: v_trunc_f32_e32 v4, v4 608; GCN-NEXT: v_mad_f32 v1, -v4, v3, v1 609; GCN-NEXT: v_cvt_i32_f32_e32 v4, v4 610; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v3| 611; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc 612; GCN-NEXT: v_add_i32_e32 v1, vcc, v4, v1 613; GCN-NEXT: v_mul_lo_u32 v1, v1, v2 614; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 615; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 616; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 617; GCN-NEXT: s_setpc_b64 s[30:31] 618; 619; GCN-IR-LABEL: v_test_srem24_64: 620; GCN-IR: ; %bb.0: 621; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 622; GCN-IR-NEXT: v_ashr_i64 v[2:3], v[2:3], 40 623; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 624; GCN-IR-NEXT: v_cvt_f32_i32_e32 v3, v2 625; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 626; GCN-IR-NEXT: v_xor_b32_e32 v5, v0, v2 627; GCN-IR-NEXT: v_ashrrev_i32_e32 v5, 30, v5 628; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v4, v3 629; GCN-IR-NEXT: v_or_b32_e32 v5, 1, v5 630; GCN-IR-NEXT: v_mul_f32_e32 v4, v1, v4 631; GCN-IR-NEXT: v_trunc_f32_e32 v4, v4 632; GCN-IR-NEXT: v_mad_f32 v1, -v4, v3, v1 633; GCN-IR-NEXT: v_cvt_i32_f32_e32 v4, v4 634; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v3| 635; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc 636; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v4, v1 637; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v2 638; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 639; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 640; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 641; GCN-IR-NEXT: s_setpc_b64 s[30:31] 642 %1 = ashr i64 %x, 40 643 %2 = ashr i64 %y, 40 644 %result = srem i64 %1, %2 645 ret i64 %result 646} 647 648define amdgpu_kernel void @s_test_srem25_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 649; GCN-LABEL: s_test_srem25_64: 650; GCN: ; %bb.0: 651; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 652; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 653; GCN-NEXT: s_mov_b32 s7, 0xf000 654; GCN-NEXT: s_mov_b32 s6, -1 655; GCN-NEXT: s_waitcnt lgkmcnt(0) 656; GCN-NEXT: s_ashr_i64 s[4:5], s[4:5], 39 657; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 658; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 39 659; GCN-NEXT: v_cvt_f32_i32_e32 v1, s2 660; GCN-NEXT: s_xor_b32 s3, s2, s4 661; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 662; GCN-NEXT: s_ashr_i32 s3, s3, 30 663; GCN-NEXT: s_or_b32 s3, s3, 1 664; GCN-NEXT: v_mov_b32_e32 v3, s3 665; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 666; GCN-NEXT: v_trunc_f32_e32 v2, v2 667; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 668; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 669; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 670; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 671; GCN-NEXT: s_mov_b32 s5, s1 672; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 673; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 674; GCN-NEXT: s_mov_b32 s4, s0 675; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 676; GCN-NEXT: v_bfe_i32 v0, v0, 0, 25 677; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 678; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 679; GCN-NEXT: s_endpgm 680; 681; GCN-IR-LABEL: s_test_srem25_64: 682; GCN-IR: ; %bb.0: 683; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 684; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 685; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 686; GCN-IR-NEXT: s_mov_b32 s6, -1 687; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 688; GCN-IR-NEXT: s_ashr_i64 s[4:5], s[4:5], 39 689; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 690; GCN-IR-NEXT: s_ashr_i64 s[2:3], s[2:3], 39 691; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s2 692; GCN-IR-NEXT: s_xor_b32 s3, s2, s4 693; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 694; GCN-IR-NEXT: s_ashr_i32 s3, s3, 30 695; GCN-IR-NEXT: s_or_b32 s3, s3, 1 696; GCN-IR-NEXT: v_mov_b32_e32 v3, s3 697; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 698; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 699; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 700; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 701; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 702; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 703; GCN-IR-NEXT: s_mov_b32 s5, s1 704; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 705; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 706; GCN-IR-NEXT: s_mov_b32 s4, s0 707; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 708; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 25 709; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 710; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 711; GCN-IR-NEXT: s_endpgm 712 %1 = ashr i64 %x, 39 713 %2 = ashr i64 %y, 39 714 %result = srem i64 %1, %2 715 store i64 %result, i64 addrspace(1)* %out 716 ret void 717} 718 719define amdgpu_kernel void @s_test_srem31_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 720; GCN-LABEL: s_test_srem31_64: 721; GCN: ; %bb.0: 722; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 723; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 724; GCN-NEXT: s_mov_b32 s7, 0xf000 725; GCN-NEXT: s_mov_b32 s6, -1 726; GCN-NEXT: s_waitcnt lgkmcnt(0) 727; GCN-NEXT: s_ashr_i64 s[4:5], s[4:5], 33 728; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 729; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 33 730; GCN-NEXT: v_cvt_f32_i32_e32 v1, s2 731; GCN-NEXT: s_xor_b32 s3, s2, s4 732; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 733; GCN-NEXT: s_ashr_i32 s3, s3, 30 734; GCN-NEXT: s_or_b32 s3, s3, 1 735; GCN-NEXT: v_mov_b32_e32 v3, s3 736; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 737; GCN-NEXT: v_trunc_f32_e32 v2, v2 738; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 739; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 740; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 741; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 742; GCN-NEXT: s_mov_b32 s5, s1 743; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 744; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 745; GCN-NEXT: s_mov_b32 s4, s0 746; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 747; GCN-NEXT: v_bfe_i32 v0, v0, 0, 31 748; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 749; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 750; GCN-NEXT: s_endpgm 751; 752; GCN-IR-LABEL: s_test_srem31_64: 753; GCN-IR: ; %bb.0: 754; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd 755; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 756; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 757; GCN-IR-NEXT: s_mov_b32 s6, -1 758; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 759; GCN-IR-NEXT: s_ashr_i64 s[4:5], s[4:5], 33 760; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 761; GCN-IR-NEXT: s_ashr_i64 s[2:3], s[2:3], 33 762; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s2 763; GCN-IR-NEXT: s_xor_b32 s3, s2, s4 764; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 765; GCN-IR-NEXT: s_ashr_i32 s3, s3, 30 766; GCN-IR-NEXT: s_or_b32 s3, s3, 1 767; GCN-IR-NEXT: v_mov_b32_e32 v3, s3 768; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 769; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 770; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 771; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 772; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 773; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 774; GCN-IR-NEXT: s_mov_b32 s5, s1 775; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 776; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 777; GCN-IR-NEXT: s_mov_b32 s4, s0 778; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 779; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 31 780; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 781; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 782; GCN-IR-NEXT: s_endpgm 783 %1 = ashr i64 %x, 33 784 %2 = ashr i64 %y, 33 785 %result = srem i64 %1, %2 786 store i64 %result, i64 addrspace(1)* %out 787 ret void 788} 789 790; 32 known sign bits 791define amdgpu_kernel void @s_test_srem32_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 792; GCN-LABEL: s_test_srem32_64: 793; GCN: ; %bb.0: 794; GCN-NEXT: s_load_dword s4, s[0:1], 0xe 795; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 796; GCN-NEXT: s_mov_b32 s7, 0xf000 797; GCN-NEXT: s_mov_b32 s6, -1 798; GCN-NEXT: s_waitcnt lgkmcnt(0) 799; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 800; GCN-NEXT: v_cvt_f32_i32_e32 v1, s3 801; GCN-NEXT: s_xor_b32 s2, s3, s4 802; GCN-NEXT: s_ashr_i32 s2, s2, 30 803; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 804; GCN-NEXT: s_or_b32 s2, s2, 1 805; GCN-NEXT: v_mov_b32_e32 v3, s2 806; GCN-NEXT: s_mov_b32 s5, s1 807; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 808; GCN-NEXT: v_trunc_f32_e32 v2, v2 809; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 810; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 811; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 812; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 813; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 814; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 815; GCN-NEXT: s_mov_b32 s4, s0 816; GCN-NEXT: v_sub_i32_e32 v0, vcc, s3, v0 817; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 818; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 819; GCN-NEXT: s_endpgm 820; 821; GCN-IR-LABEL: s_test_srem32_64: 822; GCN-IR: ; %bb.0: 823; GCN-IR-NEXT: s_load_dword s4, s[0:1], 0xe 824; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 825; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 826; GCN-IR-NEXT: s_mov_b32 s6, -1 827; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 828; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 829; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s3 830; GCN-IR-NEXT: s_xor_b32 s2, s3, s4 831; GCN-IR-NEXT: s_ashr_i32 s2, s2, 30 832; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 833; GCN-IR-NEXT: s_or_b32 s2, s2, 1 834; GCN-IR-NEXT: v_mov_b32_e32 v3, s2 835; GCN-IR-NEXT: s_mov_b32 s5, s1 836; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 837; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 838; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 839; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 840; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 841; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 842; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 843; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 844; GCN-IR-NEXT: s_mov_b32 s4, s0 845; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s3, v0 846; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 847; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 848; GCN-IR-NEXT: s_endpgm 849 %1 = ashr i64 %x, 32 850 %2 = ashr i64 %y, 32 851 %result = srem i64 %1, %2 852 store i64 %result, i64 addrspace(1)* %out 853 ret void 854} 855 856; 33 known sign bits 857define amdgpu_kernel void @s_test_srem33_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 858; GCN-LABEL: s_test_srem33_64: 859; GCN: ; %bb.0: 860; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 861; GCN-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd 862; GCN-NEXT: s_mov_b32 s11, 0xf000 863; GCN-NEXT: s_mov_b32 s10, -1 864; GCN-NEXT: s_waitcnt lgkmcnt(0) 865; GCN-NEXT: s_ashr_i64 s[2:3], s[6:7], 31 866; GCN-NEXT: s_ashr_i64 s[8:9], s[0:1], 31 867; GCN-NEXT: s_ashr_i32 s0, s1, 31 868; GCN-NEXT: s_add_u32 s8, s8, s0 869; GCN-NEXT: s_mov_b32 s1, s0 870; GCN-NEXT: s_addc_u32 s9, s9, s0 871; GCN-NEXT: s_xor_b64 s[12:13], s[8:9], s[0:1] 872; GCN-NEXT: v_cvt_f32_u32_e32 v0, s12 873; GCN-NEXT: v_cvt_f32_u32_e32 v1, s13 874; GCN-NEXT: s_sub_u32 s0, 0, s12 875; GCN-NEXT: s_subb_u32 s1, 0, s13 876; GCN-NEXT: s_ashr_i32 s6, s7, 31 877; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 878; GCN-NEXT: v_rcp_f32_e32 v0, v0 879; GCN-NEXT: s_mov_b32 s7, s6 880; GCN-NEXT: s_mov_b32 s8, s4 881; GCN-NEXT: s_mov_b32 s9, s5 882; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 883; GCN-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0 884; GCN-NEXT: v_trunc_f32_e32 v1, v1 885; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v1 886; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1 887; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 888; GCN-NEXT: v_mul_lo_u32 v2, s0, v1 889; GCN-NEXT: v_mul_hi_u32 v3, s0, v0 890; GCN-NEXT: v_mul_lo_u32 v5, s1, v0 891; GCN-NEXT: v_mul_lo_u32 v4, s0, v0 892; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 893; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v5 894; GCN-NEXT: v_mul_hi_u32 v3, v0, v4 895; GCN-NEXT: v_mul_lo_u32 v5, v0, v2 896; GCN-NEXT: v_mul_hi_u32 v6, v0, v2 897; GCN-NEXT: v_mul_hi_u32 v7, v1, v2 898; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 899; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 900; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v6, vcc 901; GCN-NEXT: v_mul_lo_u32 v6, v1, v4 902; GCN-NEXT: v_mul_hi_u32 v4, v1, v4 903; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v6 904; GCN-NEXT: v_addc_u32_e32 v3, vcc, v5, v4, vcc 905; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v7, vcc 906; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 907; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 908; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 909; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v3, vcc 910; GCN-NEXT: v_mul_lo_u32 v2, s0, v1 911; GCN-NEXT: v_mul_hi_u32 v3, s0, v0 912; GCN-NEXT: v_mul_lo_u32 v4, s1, v0 913; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 914; GCN-NEXT: v_mul_lo_u32 v3, s0, v0 915; GCN-NEXT: v_add_i32_e32 v2, vcc, v4, v2 916; GCN-NEXT: v_mul_lo_u32 v6, v0, v2 917; GCN-NEXT: v_mul_hi_u32 v7, v0, v3 918; GCN-NEXT: v_mul_hi_u32 v8, v0, v2 919; GCN-NEXT: v_mul_hi_u32 v5, v1, v3 920; GCN-NEXT: v_mul_lo_u32 v3, v1, v3 921; GCN-NEXT: v_mul_hi_u32 v4, v1, v2 922; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 923; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v8, vcc 924; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 925; GCN-NEXT: v_add_i32_e32 v3, vcc, v6, v3 926; GCN-NEXT: v_addc_u32_e32 v3, vcc, v7, v5, vcc 927; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc 928; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 929; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 930; GCN-NEXT: s_add_u32 s0, s2, s6 931; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 932; GCN-NEXT: s_addc_u32 s1, s3, s6 933; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v3, vcc 934; GCN-NEXT: s_xor_b64 s[14:15], s[0:1], s[6:7] 935; GCN-NEXT: v_mul_lo_u32 v2, s14, v1 936; GCN-NEXT: v_mul_hi_u32 v3, s14, v0 937; GCN-NEXT: v_mul_hi_u32 v4, s14, v1 938; GCN-NEXT: v_mul_hi_u32 v5, s15, v1 939; GCN-NEXT: v_mul_lo_u32 v1, s15, v1 940; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 941; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 942; GCN-NEXT: v_mul_lo_u32 v4, s15, v0 943; GCN-NEXT: v_mul_hi_u32 v0, s15, v0 944; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 945; GCN-NEXT: v_addc_u32_e32 v0, vcc, v3, v0, vcc 946; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v5, vcc 947; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 948; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v2, vcc 949; GCN-NEXT: v_mul_lo_u32 v1, s12, v1 950; GCN-NEXT: v_mul_hi_u32 v2, s12, v0 951; GCN-NEXT: v_mul_lo_u32 v3, s13, v0 952; GCN-NEXT: v_mul_lo_u32 v0, s12, v0 953; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 954; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 955; GCN-NEXT: v_sub_i32_e32 v2, vcc, s15, v1 956; GCN-NEXT: v_mov_b32_e32 v3, s13 957; GCN-NEXT: v_sub_i32_e32 v0, vcc, s14, v0 958; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 959; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s12, v0 960; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 961; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s13, v5 962; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 963; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s12, v4 964; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 965; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 966; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s13, v5 967; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s12, v4 968; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 969; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 970; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 971; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 972; GCN-NEXT: v_mov_b32_e32 v5, s15 973; GCN-NEXT: v_subb_u32_e32 v1, vcc, v5, v1, vcc 974; GCN-NEXT: v_cmp_le_u32_e32 vcc, s13, v1 975; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 976; GCN-NEXT: v_cmp_le_u32_e32 vcc, s12, v0 977; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 978; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s13, v1 979; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 980; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 981; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 982; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 983; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 984; GCN-NEXT: v_xor_b32_e32 v0, s6, v0 985; GCN-NEXT: v_xor_b32_e32 v1, s6, v1 986; GCN-NEXT: v_mov_b32_e32 v2, s6 987; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s6, v0 988; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 989; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 990; GCN-NEXT: s_endpgm 991; 992; GCN-IR-LABEL: s_test_srem33_64: 993; GCN-IR: ; %bb.0: ; %_udiv-special-cases 994; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 995; GCN-IR-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd 996; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 997; GCN-IR-NEXT: s_ashr_i64 s[2:3], s[6:7], 31 998; GCN-IR-NEXT: s_ashr_i64 s[8:9], s[0:1], 31 999; GCN-IR-NEXT: s_ashr_i32 s0, s7, 31 1000; GCN-IR-NEXT: s_ashr_i32 s6, s1, 31 1001; GCN-IR-NEXT: s_mov_b32 s1, s0 1002; GCN-IR-NEXT: s_mov_b32 s7, s6 1003; GCN-IR-NEXT: s_xor_b64 s[2:3], s[2:3], s[0:1] 1004; GCN-IR-NEXT: s_xor_b64 s[8:9], s[8:9], s[6:7] 1005; GCN-IR-NEXT: s_sub_u32 s2, s2, s0 1006; GCN-IR-NEXT: s_subb_u32 s3, s3, s0 1007; GCN-IR-NEXT: s_sub_u32 s8, s8, s6 1008; GCN-IR-NEXT: s_subb_u32 s9, s9, s6 1009; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[10:11], s[8:9], 0 1010; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[12:13], s[2:3], 0 1011; GCN-IR-NEXT: s_mov_b64 s[6:7], 0 1012; GCN-IR-NEXT: s_or_b64 s[14:15], s[10:11], s[12:13] 1013; GCN-IR-NEXT: s_flbit_i32_b32 s10, s8 1014; GCN-IR-NEXT: s_add_i32 s10, s10, 32 1015; GCN-IR-NEXT: s_flbit_i32_b32 s11, s9 1016; GCN-IR-NEXT: s_min_u32 s12, s10, s11 1017; GCN-IR-NEXT: s_flbit_i32_b32 s10, s2 1018; GCN-IR-NEXT: s_add_i32 s10, s10, 32 1019; GCN-IR-NEXT: s_flbit_i32_b32 s11, s3 1020; GCN-IR-NEXT: s_min_u32 s16, s10, s11 1021; GCN-IR-NEXT: s_sub_u32 s10, s12, s16 1022; GCN-IR-NEXT: s_subb_u32 s11, 0, 0 1023; GCN-IR-NEXT: v_cmp_gt_u64_e64 s[18:19], s[10:11], 63 1024; GCN-IR-NEXT: s_mov_b32 s13, 0 1025; GCN-IR-NEXT: s_or_b64 s[14:15], s[14:15], s[18:19] 1026; GCN-IR-NEXT: v_cmp_ne_u64_e64 s[18:19], s[10:11], 63 1027; GCN-IR-NEXT: s_xor_b64 s[20:21], s[14:15], -1 1028; GCN-IR-NEXT: s_and_b64 s[18:19], s[20:21], s[18:19] 1029; GCN-IR-NEXT: s_and_b64 vcc, exec, s[18:19] 1030; GCN-IR-NEXT: s_cbranch_vccz .LBB8_5 1031; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1032; GCN-IR-NEXT: s_add_u32 s14, s10, 1 1033; GCN-IR-NEXT: s_addc_u32 s15, s11, 0 1034; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[18:19], s[14:15], 0 1035; GCN-IR-NEXT: s_sub_i32 s10, 63, s10 1036; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[18:19] 1037; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[2:3], s10 1038; GCN-IR-NEXT: s_cbranch_vccz .LBB8_4 1039; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1040; GCN-IR-NEXT: s_lshr_b64 s[14:15], s[2:3], s14 1041; GCN-IR-NEXT: s_add_u32 s18, s8, -1 1042; GCN-IR-NEXT: s_addc_u32 s19, s9, -1 1043; GCN-IR-NEXT: s_not_b64 s[6:7], s[12:13] 1044; GCN-IR-NEXT: s_add_u32 s12, s6, s16 1045; GCN-IR-NEXT: s_addc_u32 s13, s7, 0 1046; GCN-IR-NEXT: s_mov_b64 s[16:17], 0 1047; GCN-IR-NEXT: s_mov_b32 s7, 0 1048; GCN-IR-NEXT: .LBB8_3: ; %udiv-do-while 1049; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1050; GCN-IR-NEXT: s_lshl_b64 s[14:15], s[14:15], 1 1051; GCN-IR-NEXT: s_lshr_b32 s6, s11, 31 1052; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[10:11], 1 1053; GCN-IR-NEXT: s_or_b64 s[14:15], s[14:15], s[6:7] 1054; GCN-IR-NEXT: s_or_b64 s[10:11], s[16:17], s[10:11] 1055; GCN-IR-NEXT: s_sub_u32 s6, s18, s14 1056; GCN-IR-NEXT: s_subb_u32 s6, s19, s15 1057; GCN-IR-NEXT: s_ashr_i32 s16, s6, 31 1058; GCN-IR-NEXT: s_mov_b32 s17, s16 1059; GCN-IR-NEXT: s_and_b32 s6, s16, 1 1060; GCN-IR-NEXT: s_and_b64 s[16:17], s[16:17], s[8:9] 1061; GCN-IR-NEXT: s_sub_u32 s14, s14, s16 1062; GCN-IR-NEXT: s_subb_u32 s15, s15, s17 1063; GCN-IR-NEXT: s_add_u32 s12, s12, 1 1064; GCN-IR-NEXT: s_addc_u32 s13, s13, 0 1065; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[20:21], s[12:13], 0 1066; GCN-IR-NEXT: s_mov_b64 s[16:17], s[6:7] 1067; GCN-IR-NEXT: s_and_b64 vcc, exec, s[20:21] 1068; GCN-IR-NEXT: s_cbranch_vccz .LBB8_3 1069; GCN-IR-NEXT: .LBB8_4: ; %Flow6 1070; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[10:11], 1 1071; GCN-IR-NEXT: s_or_b64 s[6:7], s[6:7], s[10:11] 1072; GCN-IR-NEXT: v_mov_b32_e32 v0, s6 1073; GCN-IR-NEXT: v_mov_b32_e32 v1, s7 1074; GCN-IR-NEXT: s_branch .LBB8_6 1075; GCN-IR-NEXT: .LBB8_5: 1076; GCN-IR-NEXT: v_mov_b32_e32 v0, s3 1077; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[14:15] 1078; GCN-IR-NEXT: v_mov_b32_e32 v0, s2 1079; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[14:15] 1080; GCN-IR-NEXT: .LBB8_6: ; %udiv-end 1081; GCN-IR-NEXT: v_mul_lo_u32 v1, s8, v1 1082; GCN-IR-NEXT: v_mul_hi_u32 v2, s8, v0 1083; GCN-IR-NEXT: v_mul_lo_u32 v3, s9, v0 1084; GCN-IR-NEXT: v_mul_lo_u32 v0, s8, v0 1085; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 1086; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1087; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1088; GCN-IR-NEXT: v_mov_b32_e32 v2, s3 1089; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 1090; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 1091; GCN-IR-NEXT: v_xor_b32_e32 v0, s0, v0 1092; GCN-IR-NEXT: v_xor_b32_e32 v1, s1, v1 1093; GCN-IR-NEXT: v_mov_b32_e32 v2, s1 1094; GCN-IR-NEXT: v_subrev_i32_e32 v0, vcc, s0, v0 1095; GCN-IR-NEXT: s_mov_b32 s6, -1 1096; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1097; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1098; GCN-IR-NEXT: s_endpgm 1099 %1 = ashr i64 %x, 31 1100 %2 = ashr i64 %y, 31 1101 %result = srem i64 %1, %2 1102 store i64 %result, i64 addrspace(1)* %out 1103 ret void 1104} 1105 1106define amdgpu_kernel void @s_test_srem24_48(i48 addrspace(1)* %out, i48 %x, i48 %y) { 1107; GCN-LABEL: s_test_srem24_48: 1108; GCN: ; %bb.0: 1109; GCN-NEXT: s_load_dword s2, s[0:1], 0xc 1110; GCN-NEXT: s_load_dword s3, s[0:1], 0xe 1111; GCN-NEXT: s_load_dword s6, s[0:1], 0xd 1112; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 1113; GCN-NEXT: s_load_dword s0, s[0:1], 0xb 1114; GCN-NEXT: s_waitcnt lgkmcnt(0) 1115; GCN-NEXT: s_sext_i32_i16 s1, s2 1116; GCN-NEXT: s_sext_i32_i16 s2, s3 1117; GCN-NEXT: v_mov_b32_e32 v0, s6 1118; GCN-NEXT: v_alignbit_b32 v0, s2, v0, 24 1119; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 1120; GCN-NEXT: v_mov_b32_e32 v2, s0 1121; GCN-NEXT: v_alignbit_b32 v2, s1, v2, 24 1122; GCN-NEXT: v_cvt_f32_i32_e32 v3, v2 1123; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v1 1124; GCN-NEXT: v_xor_b32_e32 v5, v2, v0 1125; GCN-NEXT: v_ashrrev_i32_e32 v5, 30, v5 1126; GCN-NEXT: v_or_b32_e32 v5, 1, v5 1127; GCN-NEXT: v_mul_f32_e32 v4, v3, v4 1128; GCN-NEXT: v_trunc_f32_e32 v4, v4 1129; GCN-NEXT: v_mad_f32 v3, -v4, v1, v3 1130; GCN-NEXT: v_cvt_i32_f32_e32 v4, v4 1131; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v1| 1132; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc 1133; GCN-NEXT: s_mov_b32 s7, 0xf000 1134; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v4 1135; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 1136; GCN-NEXT: s_mov_b32 s6, -1 1137; GCN-NEXT: v_sub_i32_e32 v0, vcc, v2, v0 1138; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1139; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1140; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0 1141; GCN-NEXT: buffer_store_short v1, off, s[4:7], 0 offset:4 1142; GCN-NEXT: s_endpgm 1143; 1144; GCN-IR-LABEL: s_test_srem24_48: 1145; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1146; GCN-IR-NEXT: s_load_dword s3, s[0:1], 0xc 1147; GCN-IR-NEXT: s_load_dword s5, s[0:1], 0xe 1148; GCN-IR-NEXT: s_load_dword s2, s[0:1], 0xb 1149; GCN-IR-NEXT: s_load_dword s4, s[0:1], 0xd 1150; GCN-IR-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9 1151; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1152; GCN-IR-NEXT: s_sext_i32_i16 s3, s3 1153; GCN-IR-NEXT: s_sext_i32_i16 s5, s5 1154; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[2:3], 24 1155; GCN-IR-NEXT: s_ashr_i32 s2, s3, 31 1156; GCN-IR-NEXT: s_ashr_i32 s10, s5, 31 1157; GCN-IR-NEXT: s_ashr_i64 s[8:9], s[4:5], 24 1158; GCN-IR-NEXT: s_mov_b32 s3, s2 1159; GCN-IR-NEXT: s_mov_b32 s11, s10 1160; GCN-IR-NEXT: s_xor_b64 s[4:5], s[6:7], s[2:3] 1161; GCN-IR-NEXT: s_xor_b64 s[6:7], s[8:9], s[10:11] 1162; GCN-IR-NEXT: s_sub_u32 s4, s4, s2 1163; GCN-IR-NEXT: s_subb_u32 s5, s5, s2 1164; GCN-IR-NEXT: s_sub_u32 s6, s6, s10 1165; GCN-IR-NEXT: s_subb_u32 s7, s7, s10 1166; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[10:11], s[6:7], 0 1167; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[12:13], s[4:5], 0 1168; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 1169; GCN-IR-NEXT: s_or_b64 s[14:15], s[10:11], s[12:13] 1170; GCN-IR-NEXT: s_flbit_i32_b32 s10, s6 1171; GCN-IR-NEXT: s_add_i32 s10, s10, 32 1172; GCN-IR-NEXT: s_flbit_i32_b32 s11, s7 1173; GCN-IR-NEXT: s_min_u32 s12, s10, s11 1174; GCN-IR-NEXT: s_flbit_i32_b32 s10, s4 1175; GCN-IR-NEXT: s_add_i32 s10, s10, 32 1176; GCN-IR-NEXT: s_flbit_i32_b32 s11, s5 1177; GCN-IR-NEXT: s_min_u32 s16, s10, s11 1178; GCN-IR-NEXT: s_sub_u32 s10, s12, s16 1179; GCN-IR-NEXT: s_subb_u32 s11, 0, 0 1180; GCN-IR-NEXT: v_cmp_gt_u64_e64 s[18:19], s[10:11], 63 1181; GCN-IR-NEXT: s_mov_b32 s13, 0 1182; GCN-IR-NEXT: s_or_b64 s[14:15], s[14:15], s[18:19] 1183; GCN-IR-NEXT: v_cmp_ne_u64_e64 s[18:19], s[10:11], 63 1184; GCN-IR-NEXT: s_xor_b64 s[20:21], s[14:15], -1 1185; GCN-IR-NEXT: s_and_b64 s[18:19], s[20:21], s[18:19] 1186; GCN-IR-NEXT: s_and_b64 vcc, exec, s[18:19] 1187; GCN-IR-NEXT: s_cbranch_vccz .LBB9_5 1188; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1189; GCN-IR-NEXT: s_add_u32 s14, s10, 1 1190; GCN-IR-NEXT: s_addc_u32 s15, s11, 0 1191; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[18:19], s[14:15], 0 1192; GCN-IR-NEXT: s_sub_i32 s10, 63, s10 1193; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[18:19] 1194; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[4:5], s10 1195; GCN-IR-NEXT: s_cbranch_vccz .LBB9_4 1196; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1197; GCN-IR-NEXT: s_lshr_b64 s[14:15], s[4:5], s14 1198; GCN-IR-NEXT: s_add_u32 s18, s6, -1 1199; GCN-IR-NEXT: s_addc_u32 s19, s7, -1 1200; GCN-IR-NEXT: s_not_b64 s[8:9], s[12:13] 1201; GCN-IR-NEXT: s_add_u32 s12, s8, s16 1202; GCN-IR-NEXT: s_addc_u32 s13, s9, 0 1203; GCN-IR-NEXT: s_mov_b64 s[16:17], 0 1204; GCN-IR-NEXT: s_mov_b32 s9, 0 1205; GCN-IR-NEXT: .LBB9_3: ; %udiv-do-while 1206; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1207; GCN-IR-NEXT: s_lshl_b64 s[14:15], s[14:15], 1 1208; GCN-IR-NEXT: s_lshr_b32 s8, s11, 31 1209; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[10:11], 1 1210; GCN-IR-NEXT: s_or_b64 s[14:15], s[14:15], s[8:9] 1211; GCN-IR-NEXT: s_or_b64 s[10:11], s[16:17], s[10:11] 1212; GCN-IR-NEXT: s_sub_u32 s8, s18, s14 1213; GCN-IR-NEXT: s_subb_u32 s8, s19, s15 1214; GCN-IR-NEXT: s_ashr_i32 s16, s8, 31 1215; GCN-IR-NEXT: s_mov_b32 s17, s16 1216; GCN-IR-NEXT: s_and_b32 s8, s16, 1 1217; GCN-IR-NEXT: s_and_b64 s[16:17], s[16:17], s[6:7] 1218; GCN-IR-NEXT: s_sub_u32 s14, s14, s16 1219; GCN-IR-NEXT: s_subb_u32 s15, s15, s17 1220; GCN-IR-NEXT: s_add_u32 s12, s12, 1 1221; GCN-IR-NEXT: s_addc_u32 s13, s13, 0 1222; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[20:21], s[12:13], 0 1223; GCN-IR-NEXT: s_mov_b64 s[16:17], s[8:9] 1224; GCN-IR-NEXT: s_and_b64 vcc, exec, s[20:21] 1225; GCN-IR-NEXT: s_cbranch_vccz .LBB9_3 1226; GCN-IR-NEXT: .LBB9_4: ; %Flow3 1227; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[10:11], 1 1228; GCN-IR-NEXT: s_or_b64 s[8:9], s[8:9], s[10:11] 1229; GCN-IR-NEXT: v_mov_b32_e32 v0, s8 1230; GCN-IR-NEXT: v_mov_b32_e32 v1, s9 1231; GCN-IR-NEXT: s_branch .LBB9_6 1232; GCN-IR-NEXT: .LBB9_5: 1233; GCN-IR-NEXT: v_mov_b32_e32 v0, s5 1234; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[14:15] 1235; GCN-IR-NEXT: v_mov_b32_e32 v0, s4 1236; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[14:15] 1237; GCN-IR-NEXT: .LBB9_6: ; %udiv-end 1238; GCN-IR-NEXT: v_mul_lo_u32 v1, s6, v1 1239; GCN-IR-NEXT: v_mul_hi_u32 v2, s6, v0 1240; GCN-IR-NEXT: v_mul_lo_u32 v3, s7, v0 1241; GCN-IR-NEXT: v_mul_lo_u32 v0, s6, v0 1242; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1243; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1244; GCN-IR-NEXT: v_mov_b32_e32 v2, s5 1245; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s4, v0 1246; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 1247; GCN-IR-NEXT: v_xor_b32_e32 v0, s2, v0 1248; GCN-IR-NEXT: v_xor_b32_e32 v1, s3, v1 1249; GCN-IR-NEXT: v_mov_b32_e32 v2, s3 1250; GCN-IR-NEXT: v_subrev_i32_e32 v0, vcc, s2, v0 1251; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1252; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 1253; GCN-IR-NEXT: s_mov_b32 s2, -1 1254; GCN-IR-NEXT: buffer_store_short v1, off, s[0:3], 0 offset:4 1255; GCN-IR-NEXT: buffer_store_dword v0, off, s[0:3], 0 1256; GCN-IR-NEXT: s_endpgm 1257 %1 = ashr i48 %x, 24 1258 %2 = ashr i48 %y, 24 1259 %result = srem i48 %1, %2 1260 store i48 %result, i48 addrspace(1)* %out 1261 ret void 1262} 1263 1264define amdgpu_kernel void @s_test_srem_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 1265; GCN-LABEL: s_test_srem_k_num_i64: 1266; GCN: ; %bb.0: 1267; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1268; GCN-NEXT: s_mov_b32 s7, 0xf000 1269; GCN-NEXT: s_mov_b32 s6, -1 1270; GCN-NEXT: s_waitcnt lgkmcnt(0) 1271; GCN-NEXT: s_ashr_i32 s4, s3, 31 1272; GCN-NEXT: s_add_u32 s2, s2, s4 1273; GCN-NEXT: s_mov_b32 s5, s4 1274; GCN-NEXT: s_addc_u32 s3, s3, s4 1275; GCN-NEXT: s_xor_b64 s[8:9], s[2:3], s[4:5] 1276; GCN-NEXT: v_cvt_f32_u32_e32 v0, s8 1277; GCN-NEXT: v_cvt_f32_u32_e32 v1, s9 1278; GCN-NEXT: s_sub_u32 s2, 0, s8 1279; GCN-NEXT: s_subb_u32 s3, 0, s9 1280; GCN-NEXT: s_mov_b32 s4, s0 1281; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 1282; GCN-NEXT: v_rcp_f32_e32 v0, v0 1283; GCN-NEXT: s_mov_b32 s5, s1 1284; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 1285; GCN-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0 1286; GCN-NEXT: v_trunc_f32_e32 v1, v1 1287; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v1 1288; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1 1289; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 1290; GCN-NEXT: v_mul_lo_u32 v2, s2, v1 1291; GCN-NEXT: v_mul_hi_u32 v3, s2, v0 1292; GCN-NEXT: v_mul_lo_u32 v5, s3, v0 1293; GCN-NEXT: v_mul_lo_u32 v4, s2, v0 1294; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1295; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v5 1296; GCN-NEXT: v_mul_hi_u32 v3, v0, v4 1297; GCN-NEXT: v_mul_lo_u32 v5, v0, v2 1298; GCN-NEXT: v_mul_hi_u32 v7, v0, v2 1299; GCN-NEXT: v_mul_hi_u32 v6, v1, v4 1300; GCN-NEXT: v_mul_lo_u32 v4, v1, v4 1301; GCN-NEXT: v_mul_hi_u32 v8, v1, v2 1302; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 1303; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v7, vcc 1304; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 1305; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v4 1306; GCN-NEXT: v_addc_u32_e32 v3, vcc, v5, v6, vcc 1307; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v8, vcc 1308; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1309; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 1310; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1311; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v3, vcc 1312; GCN-NEXT: v_mul_lo_u32 v2, s2, v1 1313; GCN-NEXT: v_mul_hi_u32 v3, s2, v0 1314; GCN-NEXT: v_mul_lo_u32 v4, s3, v0 1315; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1316; GCN-NEXT: v_mul_lo_u32 v3, s2, v0 1317; GCN-NEXT: v_add_i32_e32 v2, vcc, v4, v2 1318; GCN-NEXT: v_mul_lo_u32 v6, v0, v2 1319; GCN-NEXT: v_mul_hi_u32 v7, v0, v3 1320; GCN-NEXT: v_mul_hi_u32 v8, v0, v2 1321; GCN-NEXT: v_mul_hi_u32 v5, v1, v3 1322; GCN-NEXT: v_mul_lo_u32 v3, v1, v3 1323; GCN-NEXT: v_mul_hi_u32 v4, v1, v2 1324; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1325; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v8, vcc 1326; GCN-NEXT: v_mul_lo_u32 v2, v1, v2 1327; GCN-NEXT: v_add_i32_e32 v3, vcc, v6, v3 1328; GCN-NEXT: v_addc_u32_e32 v3, vcc, v7, v5, vcc 1329; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc 1330; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1331; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v4, vcc 1332; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1333; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v3, vcc 1334; GCN-NEXT: v_mul_lo_u32 v2, v1, 24 1335; GCN-NEXT: v_mul_hi_u32 v0, v0, 24 1336; GCN-NEXT: v_mul_hi_u32 v1, v1, 24 1337; GCN-NEXT: v_mov_b32_e32 v3, s9 1338; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1339; GCN-NEXT: v_addc_u32_e32 v0, vcc, 0, v1, vcc 1340; GCN-NEXT: v_mul_lo_u32 v1, s9, v0 1341; GCN-NEXT: v_mul_hi_u32 v2, s8, v0 1342; GCN-NEXT: v_mul_lo_u32 v0, s8, v0 1343; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1344; GCN-NEXT: v_sub_i32_e32 v2, vcc, 0, v1 1345; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1346; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 1347; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s8, v0 1348; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 1349; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s9, v5 1350; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 1351; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s8, v4 1352; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 1353; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 1354; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s9, v5 1355; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s8, v4 1356; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 1357; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 1358; GCN-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1359; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 1360; GCN-NEXT: v_cmp_le_u32_e32 vcc, s9, v1 1361; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 1362; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 1363; GCN-NEXT: v_cmp_le_u32_e32 vcc, s8, v0 1364; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 1365; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s9, v1 1366; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 1367; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 1368; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 1369; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 1370; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 1371; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1372; GCN-NEXT: s_endpgm 1373; 1374; GCN-IR-LABEL: s_test_srem_k_num_i64: 1375; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1376; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1377; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1378; GCN-IR-NEXT: s_ashr_i32 s6, s3, 31 1379; GCN-IR-NEXT: s_mov_b32 s7, s6 1380; GCN-IR-NEXT: s_xor_b64 s[2:3], s[2:3], s[6:7] 1381; GCN-IR-NEXT: s_sub_u32 s4, s2, s6 1382; GCN-IR-NEXT: s_subb_u32 s5, s3, s6 1383; GCN-IR-NEXT: s_flbit_i32_b32 s2, s4 1384; GCN-IR-NEXT: s_add_i32 s2, s2, 32 1385; GCN-IR-NEXT: s_flbit_i32_b32 s3, s5 1386; GCN-IR-NEXT: s_min_u32 s8, s2, s3 1387; GCN-IR-NEXT: s_add_u32 s6, s8, 0xffffffc5 1388; GCN-IR-NEXT: s_addc_u32 s7, 0, -1 1389; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[10:11], s[4:5], 0 1390; GCN-IR-NEXT: v_cmp_gt_u64_e64 s[12:13], s[6:7], 63 1391; GCN-IR-NEXT: s_mov_b64 s[2:3], 0 1392; GCN-IR-NEXT: s_or_b64 s[10:11], s[10:11], s[12:13] 1393; GCN-IR-NEXT: v_cmp_ne_u64_e64 s[12:13], s[6:7], 63 1394; GCN-IR-NEXT: s_xor_b64 s[14:15], s[10:11], -1 1395; GCN-IR-NEXT: s_and_b64 s[12:13], s[14:15], s[12:13] 1396; GCN-IR-NEXT: s_and_b64 vcc, exec, s[12:13] 1397; GCN-IR-NEXT: s_cbranch_vccz .LBB10_5 1398; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1399; GCN-IR-NEXT: s_add_u32 s10, s6, 1 1400; GCN-IR-NEXT: s_addc_u32 s11, s7, 0 1401; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[12:13], s[10:11], 0 1402; GCN-IR-NEXT: s_sub_i32 s6, 63, s6 1403; GCN-IR-NEXT: s_andn2_b64 vcc, exec, s[12:13] 1404; GCN-IR-NEXT: s_lshl_b64 s[6:7], 24, s6 1405; GCN-IR-NEXT: s_cbranch_vccz .LBB10_4 1406; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1407; GCN-IR-NEXT: s_lshr_b64 s[10:11], 24, s10 1408; GCN-IR-NEXT: s_add_u32 s14, s4, -1 1409; GCN-IR-NEXT: s_addc_u32 s15, s5, -1 1410; GCN-IR-NEXT: s_sub_u32 s8, 58, s8 1411; GCN-IR-NEXT: s_subb_u32 s9, 0, 0 1412; GCN-IR-NEXT: s_mov_b64 s[12:13], 0 1413; GCN-IR-NEXT: s_mov_b32 s3, 0 1414; GCN-IR-NEXT: .LBB10_3: ; %udiv-do-while 1415; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1416; GCN-IR-NEXT: s_lshl_b64 s[10:11], s[10:11], 1 1417; GCN-IR-NEXT: s_lshr_b32 s2, s7, 31 1418; GCN-IR-NEXT: s_lshl_b64 s[6:7], s[6:7], 1 1419; GCN-IR-NEXT: s_or_b64 s[10:11], s[10:11], s[2:3] 1420; GCN-IR-NEXT: s_or_b64 s[6:7], s[12:13], s[6:7] 1421; GCN-IR-NEXT: s_sub_u32 s2, s14, s10 1422; GCN-IR-NEXT: s_subb_u32 s2, s15, s11 1423; GCN-IR-NEXT: s_ashr_i32 s12, s2, 31 1424; GCN-IR-NEXT: s_mov_b32 s13, s12 1425; GCN-IR-NEXT: s_and_b32 s2, s12, 1 1426; GCN-IR-NEXT: s_and_b64 s[12:13], s[12:13], s[4:5] 1427; GCN-IR-NEXT: s_sub_u32 s10, s10, s12 1428; GCN-IR-NEXT: s_subb_u32 s11, s11, s13 1429; GCN-IR-NEXT: s_add_u32 s8, s8, 1 1430; GCN-IR-NEXT: s_addc_u32 s9, s9, 0 1431; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[16:17], s[8:9], 0 1432; GCN-IR-NEXT: s_mov_b64 s[12:13], s[2:3] 1433; GCN-IR-NEXT: s_and_b64 vcc, exec, s[16:17] 1434; GCN-IR-NEXT: s_cbranch_vccz .LBB10_3 1435; GCN-IR-NEXT: .LBB10_4: ; %Flow5 1436; GCN-IR-NEXT: s_lshl_b64 s[6:7], s[6:7], 1 1437; GCN-IR-NEXT: s_or_b64 s[2:3], s[2:3], s[6:7] 1438; GCN-IR-NEXT: v_mov_b32_e32 v0, s2 1439; GCN-IR-NEXT: v_mov_b32_e32 v1, s3 1440; GCN-IR-NEXT: s_branch .LBB10_6 1441; GCN-IR-NEXT: .LBB10_5: 1442; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1443; GCN-IR-NEXT: v_cndmask_b32_e64 v0, 24, 0, s[10:11] 1444; GCN-IR-NEXT: .LBB10_6: ; %udiv-end 1445; GCN-IR-NEXT: v_mul_lo_u32 v1, s4, v1 1446; GCN-IR-NEXT: v_mul_hi_u32 v2, s4, v0 1447; GCN-IR-NEXT: v_mul_lo_u32 v3, s5, v0 1448; GCN-IR-NEXT: v_mul_lo_u32 v0, s4, v0 1449; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 1450; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1451; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1452; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1453; GCN-IR-NEXT: s_mov_b32 s2, -1 1454; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1455; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 1456; GCN-IR-NEXT: s_endpgm 1457 %result = srem i64 24, %x 1458 store i64 %result, i64 addrspace(1)* %out 1459 ret void 1460} 1461 1462define i64 @v_test_srem_k_num_i64(i64 %x) { 1463; GCN-LABEL: v_test_srem_k_num_i64: 1464; GCN: ; %bb.0: 1465; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1466; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1467; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1468; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 1469; GCN-NEXT: v_xor_b32_e32 v1, v1, v2 1470; GCN-NEXT: v_xor_b32_e32 v0, v0, v2 1471; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 1472; GCN-NEXT: v_cvt_f32_u32_e32 v3, v1 1473; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v0 1474; GCN-NEXT: v_subb_u32_e32 v5, vcc, 0, v1, vcc 1475; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 1476; GCN-NEXT: v_rcp_f32_e32 v2, v2 1477; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 1478; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 1479; GCN-NEXT: v_trunc_f32_e32 v3, v3 1480; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 1481; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 1482; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 1483; GCN-NEXT: v_mul_hi_u32 v6, v4, v2 1484; GCN-NEXT: v_mul_lo_u32 v7, v4, v3 1485; GCN-NEXT: v_mul_lo_u32 v8, v5, v2 1486; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 1487; GCN-NEXT: v_mul_lo_u32 v7, v4, v2 1488; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 1489; GCN-NEXT: v_mul_lo_u32 v8, v2, v6 1490; GCN-NEXT: v_mul_hi_u32 v9, v2, v7 1491; GCN-NEXT: v_mul_hi_u32 v10, v2, v6 1492; GCN-NEXT: v_mul_hi_u32 v11, v3, v6 1493; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 1494; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1495; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v10, vcc 1496; GCN-NEXT: v_mul_lo_u32 v10, v3, v7 1497; GCN-NEXT: v_mul_hi_u32 v7, v3, v7 1498; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v10 1499; GCN-NEXT: v_addc_u32_e32 v7, vcc, v9, v7, vcc 1500; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v11, vcc 1501; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1502; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v8, vcc 1503; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6 1504; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v7, vcc 1505; GCN-NEXT: v_mul_lo_u32 v6, v4, v3 1506; GCN-NEXT: v_mul_hi_u32 v7, v4, v2 1507; GCN-NEXT: v_mul_lo_u32 v5, v5, v2 1508; GCN-NEXT: v_mul_lo_u32 v4, v4, v2 1509; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1510; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 1511; GCN-NEXT: v_mul_lo_u32 v8, v2, v5 1512; GCN-NEXT: v_mul_hi_u32 v9, v2, v4 1513; GCN-NEXT: v_mul_hi_u32 v10, v2, v5 1514; GCN-NEXT: v_mul_hi_u32 v7, v3, v4 1515; GCN-NEXT: v_mul_lo_u32 v4, v3, v4 1516; GCN-NEXT: v_mul_hi_u32 v6, v3, v5 1517; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1518; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v10, vcc 1519; GCN-NEXT: v_mul_lo_u32 v5, v3, v5 1520; GCN-NEXT: v_add_i32_e32 v4, vcc, v8, v4 1521; GCN-NEXT: v_addc_u32_e32 v4, vcc, v9, v7, vcc 1522; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc 1523; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 1524; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v6, vcc 1525; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 1526; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v5, vcc 1527; GCN-NEXT: v_mul_lo_u32 v4, v3, 24 1528; GCN-NEXT: v_mul_hi_u32 v2, v2, 24 1529; GCN-NEXT: v_mul_hi_u32 v3, v3, 24 1530; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 1531; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v3, vcc 1532; GCN-NEXT: v_mul_lo_u32 v3, v1, v2 1533; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 1534; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 1535; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1536; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 1537; GCN-NEXT: v_sub_i32_e32 v2, vcc, 24, v2 1538; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 1539; GCN-NEXT: v_sub_i32_e64 v5, s[4:5], v2, v0 1540; GCN-NEXT: v_subbrev_u32_e64 v6, s[6:7], 0, v4, s[4:5] 1541; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v6, v1 1542; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[6:7] 1543; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v5, v0 1544; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[6:7] 1545; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v6, v1 1546; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, s[4:5] 1547; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[6:7] 1548; GCN-NEXT: v_sub_i32_e64 v8, s[4:5], v5, v0 1549; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 1550; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 1551; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v1 1552; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v7 1553; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, vcc 1554; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 1555; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc 1556; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1 1557; GCN-NEXT: v_cndmask_b32_e32 v0, v7, v0, vcc 1558; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v8, s[4:5] 1559; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 1560; GCN-NEXT: v_cndmask_b32_e64 v1, v6, v4, s[4:5] 1561; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc 1562; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc 1563; GCN-NEXT: s_setpc_b64 s[30:31] 1564; 1565; GCN-IR-LABEL: v_test_srem_k_num_i64: 1566; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1567; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1568; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1569; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1570; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 1571; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1572; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1573; GCN-IR-NEXT: v_ffbh_u32_e32 v2, v0 1574; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, 32, v2 1575; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v1 1576; GCN-IR-NEXT: v_min_u32_e32 v6, v2, v3 1577; GCN-IR-NEXT: s_movk_i32 s6, 0xffc5 1578; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, s6, v6 1579; GCN-IR-NEXT: v_addc_u32_e64 v5, s[6:7], 0, -1, vcc 1580; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 1581; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[4:5] 1582; GCN-IR-NEXT: v_cmp_ne_u64_e64 s[6:7], 63, v[4:5] 1583; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], vcc 1584; GCN-IR-NEXT: v_cndmask_b32_e64 v3, 24, 0, s[4:5] 1585; GCN-IR-NEXT: s_xor_b64 s[4:5], s[4:5], -1 1586; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 1587; GCN-IR-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7] 1588; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1589; GCN-IR-NEXT: s_cbranch_execz .LBB11_6 1590; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1591; GCN-IR-NEXT: v_add_i32_e32 v7, vcc, 1, v4 1592; GCN-IR-NEXT: v_addc_u32_e32 v8, vcc, 0, v5, vcc 1593; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 63, v4 1594; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[7:8] 1595; GCN-IR-NEXT: v_lshl_b64 v[2:3], 24, v2 1596; GCN-IR-NEXT: v_mov_b32_e32 v4, 0 1597; GCN-IR-NEXT: v_mov_b32_e32 v5, 0 1598; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1599; GCN-IR-NEXT: s_xor_b64 s[8:9], exec, s[4:5] 1600; GCN-IR-NEXT: s_cbranch_execz .LBB11_5 1601; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1602; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, -1, v0 1603; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, -1, v1, vcc 1604; GCN-IR-NEXT: v_lshr_b64 v[8:9], 24, v7 1605; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, 58, v6 1606; GCN-IR-NEXT: v_mov_b32_e32 v10, 0 1607; GCN-IR-NEXT: v_subb_u32_e64 v7, s[4:5], 0, 0, vcc 1608; GCN-IR-NEXT: s_mov_b64 s[10:11], 0 1609; GCN-IR-NEXT: v_mov_b32_e32 v11, 0 1610; GCN-IR-NEXT: v_mov_b32_e32 v5, 0 1611; GCN-IR-NEXT: .LBB11_3: ; %udiv-do-while 1612; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1613; GCN-IR-NEXT: v_lshl_b64 v[8:9], v[8:9], 1 1614; GCN-IR-NEXT: v_lshrrev_b32_e32 v4, 31, v3 1615; GCN-IR-NEXT: v_or_b32_e32 v8, v8, v4 1616; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1617; GCN-IR-NEXT: v_sub_i32_e32 v4, vcc, v12, v8 1618; GCN-IR-NEXT: v_subb_u32_e32 v4, vcc, v13, v9, vcc 1619; GCN-IR-NEXT: v_or_b32_e32 v2, v10, v2 1620; GCN-IR-NEXT: v_ashrrev_i32_e32 v10, 31, v4 1621; GCN-IR-NEXT: v_add_i32_e32 v6, vcc, 1, v6 1622; GCN-IR-NEXT: v_or_b32_e32 v3, v11, v3 1623; GCN-IR-NEXT: v_and_b32_e32 v4, 1, v10 1624; GCN-IR-NEXT: v_and_b32_e32 v11, v10, v1 1625; GCN-IR-NEXT: v_and_b32_e32 v10, v10, v0 1626; GCN-IR-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc 1627; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[6:7] 1628; GCN-IR-NEXT: v_sub_i32_e64 v8, s[4:5], v8, v10 1629; GCN-IR-NEXT: v_subb_u32_e64 v9, s[4:5], v9, v11, s[4:5] 1630; GCN-IR-NEXT: v_mov_b32_e32 v11, v5 1631; GCN-IR-NEXT: s_or_b64 s[10:11], vcc, s[10:11] 1632; GCN-IR-NEXT: v_mov_b32_e32 v10, v4 1633; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[10:11] 1634; GCN-IR-NEXT: s_cbranch_execnz .LBB11_3 1635; GCN-IR-NEXT: ; %bb.4: ; %Flow 1636; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1637; GCN-IR-NEXT: .LBB11_5: ; %Flow3 1638; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1639; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[2:3], 1 1640; GCN-IR-NEXT: v_or_b32_e32 v2, v5, v7 1641; GCN-IR-NEXT: v_or_b32_e32 v3, v4, v6 1642; GCN-IR-NEXT: .LBB11_6: ; %Flow4 1643; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1644; GCN-IR-NEXT: v_mul_lo_u32 v2, v0, v2 1645; GCN-IR-NEXT: v_mul_hi_u32 v4, v0, v3 1646; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v3 1647; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, v3 1648; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, v4, v2 1649; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1650; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1651; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1652; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1653 %result = srem i64 24, %x 1654 ret i64 %result 1655} 1656 1657define i64 @v_test_srem_pow2_k_num_i64(i64 %x) { 1658; GCN-LABEL: v_test_srem_pow2_k_num_i64: 1659; GCN: ; %bb.0: 1660; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1661; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1662; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1663; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 1664; GCN-NEXT: v_xor_b32_e32 v1, v1, v2 1665; GCN-NEXT: v_xor_b32_e32 v0, v0, v2 1666; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 1667; GCN-NEXT: v_cvt_f32_u32_e32 v3, v1 1668; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v0 1669; GCN-NEXT: v_subb_u32_e32 v5, vcc, 0, v1, vcc 1670; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 1671; GCN-NEXT: v_rcp_f32_e32 v2, v2 1672; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 1673; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 1674; GCN-NEXT: v_trunc_f32_e32 v3, v3 1675; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 1676; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 1677; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 1678; GCN-NEXT: v_mul_hi_u32 v6, v4, v2 1679; GCN-NEXT: v_mul_lo_u32 v7, v4, v3 1680; GCN-NEXT: v_mul_lo_u32 v8, v5, v2 1681; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 1682; GCN-NEXT: v_mul_lo_u32 v7, v4, v2 1683; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 1684; GCN-NEXT: v_mul_lo_u32 v8, v2, v6 1685; GCN-NEXT: v_mul_hi_u32 v9, v2, v7 1686; GCN-NEXT: v_mul_hi_u32 v10, v2, v6 1687; GCN-NEXT: v_mul_hi_u32 v11, v3, v6 1688; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 1689; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1690; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v10, vcc 1691; GCN-NEXT: v_mul_lo_u32 v10, v3, v7 1692; GCN-NEXT: v_mul_hi_u32 v7, v3, v7 1693; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v10 1694; GCN-NEXT: v_addc_u32_e32 v7, vcc, v9, v7, vcc 1695; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v11, vcc 1696; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1697; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v8, vcc 1698; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6 1699; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v7, vcc 1700; GCN-NEXT: v_mul_lo_u32 v6, v4, v3 1701; GCN-NEXT: v_mul_hi_u32 v7, v4, v2 1702; GCN-NEXT: v_mul_lo_u32 v5, v5, v2 1703; GCN-NEXT: v_mul_lo_u32 v4, v4, v2 1704; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1705; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 1706; GCN-NEXT: v_mul_lo_u32 v8, v2, v5 1707; GCN-NEXT: v_mul_hi_u32 v9, v2, v4 1708; GCN-NEXT: v_mul_hi_u32 v10, v2, v5 1709; GCN-NEXT: v_mul_hi_u32 v7, v3, v4 1710; GCN-NEXT: v_mul_lo_u32 v4, v3, v4 1711; GCN-NEXT: v_mul_hi_u32 v6, v3, v5 1712; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1713; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v10, vcc 1714; GCN-NEXT: v_mul_lo_u32 v5, v3, v5 1715; GCN-NEXT: v_add_i32_e32 v4, vcc, v8, v4 1716; GCN-NEXT: v_addc_u32_e32 v4, vcc, v9, v7, vcc 1717; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc 1718; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 1719; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v6, vcc 1720; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 1721; GCN-NEXT: v_addc_u32_e32 v2, vcc, v3, v5, vcc 1722; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v2 1723; GCN-NEXT: v_mul_lo_u32 v3, v1, v2 1724; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 1725; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 1726; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1727; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 1728; GCN-NEXT: v_sub_i32_e32 v2, vcc, 0x8000, v2 1729; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 1730; GCN-NEXT: v_sub_i32_e64 v5, s[4:5], v2, v0 1731; GCN-NEXT: v_subbrev_u32_e64 v6, s[6:7], 0, v4, s[4:5] 1732; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v6, v1 1733; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[6:7] 1734; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v5, v0 1735; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[6:7] 1736; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v6, v1 1737; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, s[4:5] 1738; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[6:7] 1739; GCN-NEXT: v_sub_i32_e64 v8, s[4:5], v5, v0 1740; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 1741; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 1742; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v1 1743; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v7 1744; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, vcc 1745; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 1746; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc 1747; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1 1748; GCN-NEXT: v_cndmask_b32_e32 v0, v7, v0, vcc 1749; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v8, s[4:5] 1750; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 1751; GCN-NEXT: v_cndmask_b32_e64 v1, v6, v4, s[4:5] 1752; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc 1753; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc 1754; GCN-NEXT: s_setpc_b64 s[30:31] 1755; 1756; GCN-IR-LABEL: v_test_srem_pow2_k_num_i64: 1757; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1758; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1759; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1760; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1761; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 1762; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1763; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1764; GCN-IR-NEXT: v_ffbh_u32_e32 v2, v0 1765; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, 32, v2 1766; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v1 1767; GCN-IR-NEXT: v_min_u32_e32 v6, v2, v3 1768; GCN-IR-NEXT: s_movk_i32 s6, 0xffd0 1769; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, s6, v6 1770; GCN-IR-NEXT: v_addc_u32_e64 v4, s[6:7], 0, -1, vcc 1771; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 1772; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[3:4] 1773; GCN-IR-NEXT: v_mov_b32_e32 v5, 0x8000 1774; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], vcc 1775; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[3:4] 1776; GCN-IR-NEXT: v_cndmask_b32_e64 v5, v5, 0, s[4:5] 1777; GCN-IR-NEXT: s_xor_b64 s[4:5], s[4:5], -1 1778; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 1779; GCN-IR-NEXT: s_mov_b64 s[8:9], 0x8000 1780; GCN-IR-NEXT: s_and_b64 s[4:5], s[4:5], vcc 1781; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1782; GCN-IR-NEXT: s_cbranch_execz .LBB12_6 1783; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1784; GCN-IR-NEXT: v_add_i32_e32 v7, vcc, 1, v3 1785; GCN-IR-NEXT: v_addc_u32_e32 v8, vcc, 0, v4, vcc 1786; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 63, v3 1787; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[7:8] 1788; GCN-IR-NEXT: v_lshl_b64 v[2:3], s[8:9], v2 1789; GCN-IR-NEXT: v_mov_b32_e32 v4, 0 1790; GCN-IR-NEXT: v_mov_b32_e32 v5, 0 1791; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1792; GCN-IR-NEXT: s_xor_b64 s[8:9], exec, s[4:5] 1793; GCN-IR-NEXT: s_cbranch_execz .LBB12_5 1794; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1795; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, -1, v0 1796; GCN-IR-NEXT: s_mov_b64 s[4:5], 0x8000 1797; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, -1, v1, vcc 1798; GCN-IR-NEXT: v_lshr_b64 v[8:9], s[4:5], v7 1799; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, 47, v6 1800; GCN-IR-NEXT: v_mov_b32_e32 v10, 0 1801; GCN-IR-NEXT: v_subb_u32_e64 v7, s[4:5], 0, 0, vcc 1802; GCN-IR-NEXT: s_mov_b64 s[10:11], 0 1803; GCN-IR-NEXT: v_mov_b32_e32 v11, 0 1804; GCN-IR-NEXT: v_mov_b32_e32 v5, 0 1805; GCN-IR-NEXT: .LBB12_3: ; %udiv-do-while 1806; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1807; GCN-IR-NEXT: v_lshl_b64 v[8:9], v[8:9], 1 1808; GCN-IR-NEXT: v_lshrrev_b32_e32 v4, 31, v3 1809; GCN-IR-NEXT: v_or_b32_e32 v8, v8, v4 1810; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1811; GCN-IR-NEXT: v_sub_i32_e32 v4, vcc, v12, v8 1812; GCN-IR-NEXT: v_subb_u32_e32 v4, vcc, v13, v9, vcc 1813; GCN-IR-NEXT: v_or_b32_e32 v2, v10, v2 1814; GCN-IR-NEXT: v_ashrrev_i32_e32 v10, 31, v4 1815; GCN-IR-NEXT: v_add_i32_e32 v6, vcc, 1, v6 1816; GCN-IR-NEXT: v_or_b32_e32 v3, v11, v3 1817; GCN-IR-NEXT: v_and_b32_e32 v4, 1, v10 1818; GCN-IR-NEXT: v_and_b32_e32 v11, v10, v1 1819; GCN-IR-NEXT: v_and_b32_e32 v10, v10, v0 1820; GCN-IR-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc 1821; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[6:7] 1822; GCN-IR-NEXT: v_sub_i32_e64 v8, s[4:5], v8, v10 1823; GCN-IR-NEXT: v_subb_u32_e64 v9, s[4:5], v9, v11, s[4:5] 1824; GCN-IR-NEXT: v_mov_b32_e32 v11, v5 1825; GCN-IR-NEXT: s_or_b64 s[10:11], vcc, s[10:11] 1826; GCN-IR-NEXT: v_mov_b32_e32 v10, v4 1827; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[10:11] 1828; GCN-IR-NEXT: s_cbranch_execnz .LBB12_3 1829; GCN-IR-NEXT: ; %bb.4: ; %Flow 1830; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1831; GCN-IR-NEXT: .LBB12_5: ; %Flow3 1832; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1833; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[2:3], 1 1834; GCN-IR-NEXT: v_or_b32_e32 v2, v5, v7 1835; GCN-IR-NEXT: v_or_b32_e32 v5, v4, v6 1836; GCN-IR-NEXT: .LBB12_6: ; %Flow4 1837; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1838; GCN-IR-NEXT: v_mul_lo_u32 v2, v0, v2 1839; GCN-IR-NEXT: v_mul_hi_u32 v3, v0, v5 1840; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v5 1841; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, v5 1842; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1843; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1844; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 1845; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1846; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1847 %result = srem i64 32768, %x 1848 ret i64 %result 1849} 1850 1851define i64 @v_test_srem_pow2_k_den_i64(i64 %x) { 1852; GCN-LABEL: v_test_srem_pow2_k_den_i64: 1853; GCN: ; %bb.0: 1854; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1855; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1856; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v2 1857; GCN-NEXT: v_add_i32_e32 v2, vcc, v0, v2 1858; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 1859; GCN-NEXT: v_and_b32_e32 v2, 0xffff8000, v2 1860; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1861; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 1862; GCN-NEXT: s_setpc_b64 s[30:31] 1863; 1864; GCN-IR-LABEL: v_test_srem_pow2_k_den_i64: 1865; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1866; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1867; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1868; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1869; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 1870; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1871; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1872; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v0 1873; GCN-IR-NEXT: v_add_i32_e64 v3, s[4:5], 32, v3 1874; GCN-IR-NEXT: v_ffbh_u32_e32 v4, v1 1875; GCN-IR-NEXT: v_min_u32_e32 v8, v3, v4 1876; GCN-IR-NEXT: v_sub_i32_e64 v4, s[4:5], 48, v8 1877; GCN-IR-NEXT: v_subb_u32_e64 v5, s[4:5], 0, 0, s[4:5] 1878; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[0:1] 1879; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[4:5], 63, v[4:5] 1880; GCN-IR-NEXT: v_mov_b32_e32 v3, v2 1881; GCN-IR-NEXT: s_or_b64 s[4:5], vcc, s[4:5] 1882; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[4:5] 1883; GCN-IR-NEXT: s_xor_b64 s[6:7], s[4:5], -1 1884; GCN-IR-NEXT: v_cndmask_b32_e64 v7, v1, 0, s[4:5] 1885; GCN-IR-NEXT: v_cndmask_b32_e64 v6, v0, 0, s[4:5] 1886; GCN-IR-NEXT: s_and_b64 s[4:5], s[6:7], vcc 1887; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1888; GCN-IR-NEXT: s_cbranch_execz .LBB13_6 1889; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1890; GCN-IR-NEXT: v_add_i32_e32 v9, vcc, 1, v4 1891; GCN-IR-NEXT: v_addc_u32_e32 v10, vcc, 0, v5, vcc 1892; GCN-IR-NEXT: v_sub_i32_e64 v4, s[4:5], 63, v4 1893; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[9:10] 1894; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[0:1], v4 1895; GCN-IR-NEXT: v_mov_b32_e32 v6, 0 1896; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1897; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1898; GCN-IR-NEXT: s_xor_b64 s[8:9], exec, s[4:5] 1899; GCN-IR-NEXT: s_cbranch_execz .LBB13_5 1900; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1901; GCN-IR-NEXT: v_lshr_b64 v[10:11], v[0:1], v9 1902; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 0xffffffcf, v8 1903; GCN-IR-NEXT: v_mov_b32_e32 v12, 0 1904; GCN-IR-NEXT: v_addc_u32_e64 v9, s[4:5], 0, -1, vcc 1905; GCN-IR-NEXT: s_mov_b64 s[10:11], 0 1906; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 1907; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1908; GCN-IR-NEXT: s_movk_i32 s12, 0x7fff 1909; GCN-IR-NEXT: .LBB13_3: ; %udiv-do-while 1910; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1911; GCN-IR-NEXT: v_lshl_b64 v[10:11], v[10:11], 1 1912; GCN-IR-NEXT: v_lshrrev_b32_e32 v6, 31, v5 1913; GCN-IR-NEXT: v_or_b32_e32 v10, v10, v6 1914; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, s12, v10 1915; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[4:5], 1 1916; GCN-IR-NEXT: v_subb_u32_e32 v6, vcc, 0, v11, vcc 1917; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v8 1918; GCN-IR-NEXT: v_or_b32_e32 v4, v12, v4 1919; GCN-IR-NEXT: v_ashrrev_i32_e32 v12, 31, v6 1920; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc 1921; GCN-IR-NEXT: v_and_b32_e32 v6, 1, v12 1922; GCN-IR-NEXT: v_and_b32_e32 v12, 0x8000, v12 1923; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[8:9] 1924; GCN-IR-NEXT: v_or_b32_e32 v5, v13, v5 1925; GCN-IR-NEXT: v_sub_i32_e64 v10, s[4:5], v10, v12 1926; GCN-IR-NEXT: v_mov_b32_e32 v13, v7 1927; GCN-IR-NEXT: v_subbrev_u32_e64 v11, s[4:5], 0, v11, s[4:5] 1928; GCN-IR-NEXT: s_or_b64 s[10:11], vcc, s[10:11] 1929; GCN-IR-NEXT: v_mov_b32_e32 v12, v6 1930; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[10:11] 1931; GCN-IR-NEXT: s_cbranch_execnz .LBB13_3 1932; GCN-IR-NEXT: ; %bb.4: ; %Flow 1933; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1934; GCN-IR-NEXT: .LBB13_5: ; %Flow3 1935; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1936; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[4:5], 1 1937; GCN-IR-NEXT: v_or_b32_e32 v7, v7, v5 1938; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v4 1939; GCN-IR-NEXT: .LBB13_6: ; %Flow4 1940; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1941; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[6:7], 15 1942; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 1943; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 1944; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1945; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v3 1946; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1947; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 1948; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1949 %result = srem i64 %x, 32768 1950 ret i64 %result 1951} 1952 1953define amdgpu_kernel void @s_test_srem24_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 1954; GCN-LABEL: s_test_srem24_k_num_i64: 1955; GCN: ; %bb.0: 1956; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1957; GCN-NEXT: s_mov_b32 s6, 0x41c00000 1958; GCN-NEXT: s_waitcnt lgkmcnt(0) 1959; GCN-NEXT: s_ashr_i64 s[4:5], s[2:3], 40 1960; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 1961; GCN-NEXT: s_ashr_i32 s5, s4, 30 1962; GCN-NEXT: s_or_b32 s5, s5, 1 1963; GCN-NEXT: v_mov_b32_e32 v3, s5 1964; GCN-NEXT: v_rcp_iflag_f32_e32 v1, v0 1965; GCN-NEXT: s_mov_b32 s3, 0xf000 1966; GCN-NEXT: s_mov_b32 s2, -1 1967; GCN-NEXT: v_mul_f32_e32 v1, 0x41c00000, v1 1968; GCN-NEXT: v_trunc_f32_e32 v1, v1 1969; GCN-NEXT: v_mad_f32 v2, -v1, v0, s6 1970; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1 1971; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, |v0| 1972; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 1973; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 1974; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 1975; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1976; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1977; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1978; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 1979; GCN-NEXT: s_endpgm 1980; 1981; GCN-IR-LABEL: s_test_srem24_k_num_i64: 1982; GCN-IR: ; %bb.0: 1983; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1984; GCN-IR-NEXT: s_mov_b32 s6, 0x41c00000 1985; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1986; GCN-IR-NEXT: s_ashr_i64 s[4:5], s[2:3], 40 1987; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 1988; GCN-IR-NEXT: s_ashr_i32 s5, s4, 30 1989; GCN-IR-NEXT: s_or_b32 s5, s5, 1 1990; GCN-IR-NEXT: v_mov_b32_e32 v3, s5 1991; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v1, v0 1992; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 1993; GCN-IR-NEXT: s_mov_b32 s2, -1 1994; GCN-IR-NEXT: v_mul_f32_e32 v1, 0x41c00000, v1 1995; GCN-IR-NEXT: v_trunc_f32_e32 v1, v1 1996; GCN-IR-NEXT: v_mad_f32 v2, -v1, v0, s6 1997; GCN-IR-NEXT: v_cvt_i32_f32_e32 v1, v1 1998; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, |v0| 1999; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 2000; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v1 2001; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 2002; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2003; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2004; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2005; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 2006; GCN-IR-NEXT: s_endpgm 2007 %x.shr = ashr i64 %x, 40 2008 %result = srem i64 24, %x.shr 2009 store i64 %result, i64 addrspace(1)* %out 2010 ret void 2011} 2012 2013define amdgpu_kernel void @s_test_srem24_k_den_i64(i64 addrspace(1)* %out, i64 %x) { 2014; GCN-LABEL: s_test_srem24_k_den_i64: 2015; GCN: ; %bb.0: 2016; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 2017; GCN-NEXT: s_mov_b32 s4, 0x46b6fe00 2018; GCN-NEXT: s_mov_b32 s7, 0xf000 2019; GCN-NEXT: s_mov_b32 s6, -1 2020; GCN-NEXT: s_waitcnt lgkmcnt(0) 2021; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 40 2022; GCN-NEXT: v_cvt_f32_i32_e32 v0, s2 2023; GCN-NEXT: s_ashr_i32 s3, s2, 30 2024; GCN-NEXT: s_or_b32 s3, s3, 1 2025; GCN-NEXT: v_mov_b32_e32 v1, s3 2026; GCN-NEXT: v_mul_f32_e32 v2, 0x38331158, v0 2027; GCN-NEXT: v_trunc_f32_e32 v2, v2 2028; GCN-NEXT: v_mad_f32 v0, -v2, s4, v0 2029; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 2030; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s4 2031; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc 2032; GCN-NEXT: s_movk_i32 s3, 0x5b7f 2033; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 2034; GCN-NEXT: v_mul_lo_u32 v0, v0, s3 2035; GCN-NEXT: s_mov_b32 s4, s0 2036; GCN-NEXT: s_mov_b32 s5, s1 2037; GCN-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 2038; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2039; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2040; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 2041; GCN-NEXT: s_endpgm 2042; 2043; GCN-IR-LABEL: s_test_srem24_k_den_i64: 2044; GCN-IR: ; %bb.0: 2045; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 2046; GCN-IR-NEXT: s_mov_b32 s4, 0x46b6fe00 2047; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 2048; GCN-IR-NEXT: s_mov_b32 s6, -1 2049; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 2050; GCN-IR-NEXT: s_ashr_i64 s[2:3], s[2:3], 40 2051; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s2 2052; GCN-IR-NEXT: s_ashr_i32 s3, s2, 30 2053; GCN-IR-NEXT: s_or_b32 s3, s3, 1 2054; GCN-IR-NEXT: v_mov_b32_e32 v1, s3 2055; GCN-IR-NEXT: v_mul_f32_e32 v2, 0x38331158, v0 2056; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 2057; GCN-IR-NEXT: v_mad_f32 v0, -v2, s4, v0 2058; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 2059; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s4 2060; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc 2061; GCN-IR-NEXT: s_movk_i32 s3, 0x5b7f 2062; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 2063; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s3 2064; GCN-IR-NEXT: s_mov_b32 s4, s0 2065; GCN-IR-NEXT: s_mov_b32 s5, s1 2066; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s2, v0 2067; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2068; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2069; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 2070; GCN-IR-NEXT: s_endpgm 2071 %x.shr = ashr i64 %x, 40 2072 %result = srem i64 %x.shr, 23423 2073 store i64 %result, i64 addrspace(1)* %out 2074 ret void 2075} 2076 2077define i64 @v_test_srem24_k_num_i64(i64 %x) { 2078; GCN-LABEL: v_test_srem24_k_num_i64: 2079; GCN: ; %bb.0: 2080; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2081; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2082; GCN-NEXT: s_mov_b32 s4, 0x41c00000 2083; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 2084; GCN-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2085; GCN-NEXT: v_or_b32_e32 v3, 1, v3 2086; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v1 2087; GCN-NEXT: v_mul_f32_e32 v2, 0x41c00000, v2 2088; GCN-NEXT: v_trunc_f32_e32 v2, v2 2089; GCN-NEXT: v_mad_f32 v4, -v2, v1, s4 2090; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 2091; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2092; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2093; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2094; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 2095; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2096; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2097; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2098; GCN-NEXT: s_setpc_b64 s[30:31] 2099; 2100; GCN-IR-LABEL: v_test_srem24_k_num_i64: 2101; GCN-IR: ; %bb.0: 2102; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2103; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2104; GCN-IR-NEXT: s_mov_b32 s4, 0x41c00000 2105; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 2106; GCN-IR-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2107; GCN-IR-NEXT: v_or_b32_e32 v3, 1, v3 2108; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v1 2109; GCN-IR-NEXT: v_mul_f32_e32 v2, 0x41c00000, v2 2110; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 2111; GCN-IR-NEXT: v_mad_f32 v4, -v2, v1, s4 2112; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 2113; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2114; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2115; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2116; GCN-IR-NEXT: v_mul_lo_u32 v0, v1, v0 2117; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2118; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2119; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2120; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2121 %x.shr = ashr i64 %x, 40 2122 %result = srem i64 24, %x.shr 2123 ret i64 %result 2124} 2125 2126define i64 @v_test_srem24_pow2_k_num_i64(i64 %x) { 2127; GCN-LABEL: v_test_srem24_pow2_k_num_i64: 2128; GCN: ; %bb.0: 2129; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2130; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2131; GCN-NEXT: s_mov_b32 s4, 0x47000000 2132; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 2133; GCN-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2134; GCN-NEXT: v_or_b32_e32 v3, 1, v3 2135; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v1 2136; GCN-NEXT: v_mul_f32_e32 v2, 0x47000000, v2 2137; GCN-NEXT: v_trunc_f32_e32 v2, v2 2138; GCN-NEXT: v_mad_f32 v4, -v2, v1, s4 2139; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 2140; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2141; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2142; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2143; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 2144; GCN-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 2145; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2146; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2147; GCN-NEXT: s_setpc_b64 s[30:31] 2148; 2149; GCN-IR-LABEL: v_test_srem24_pow2_k_num_i64: 2150; GCN-IR: ; %bb.0: 2151; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2152; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2153; GCN-IR-NEXT: s_mov_b32 s4, 0x47000000 2154; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 2155; GCN-IR-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2156; GCN-IR-NEXT: v_or_b32_e32 v3, 1, v3 2157; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v1 2158; GCN-IR-NEXT: v_mul_f32_e32 v2, 0x47000000, v2 2159; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 2160; GCN-IR-NEXT: v_mad_f32 v4, -v2, v1, s4 2161; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 2162; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2163; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2164; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2165; GCN-IR-NEXT: v_mul_lo_u32 v0, v1, v0 2166; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 2167; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2168; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2169; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2170 %x.shr = ashr i64 %x, 40 2171 %result = srem i64 32768, %x.shr 2172 ret i64 %result 2173} 2174 2175define i64 @v_test_srem24_pow2_k_den_i64(i64 %x) { 2176; GCN-LABEL: v_test_srem24_pow2_k_den_i64: 2177; GCN: ; %bb.0: 2178; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2179; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2180; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v1 2181; GCN-NEXT: v_add_i32_e32 v2, vcc, v0, v2 2182; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 2183; GCN-NEXT: v_and_b32_e32 v2, 0xffff8000, v2 2184; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 2185; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 2186; GCN-NEXT: s_setpc_b64 s[30:31] 2187; 2188; GCN-IR-LABEL: v_test_srem24_pow2_k_den_i64: 2189; GCN-IR: ; %bb.0: 2190; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2191; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2192; GCN-IR-NEXT: s_mov_b32 s4, 0x47000000 2193; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 2194; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 30, v0 2195; GCN-IR-NEXT: v_or_b32_e32 v2, 1, v2 2196; GCN-IR-NEXT: v_mul_f32_e32 v3, 0x38000000, v1 2197; GCN-IR-NEXT: v_trunc_f32_e32 v3, v3 2198; GCN-IR-NEXT: v_mad_f32 v1, -v3, s4, v1 2199; GCN-IR-NEXT: v_cvt_i32_f32_e32 v3, v3 2200; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, s4 2201; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc 2202; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v3, v1 2203; GCN-IR-NEXT: v_lshlrev_b32_e32 v1, 15, v1 2204; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 2205; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2206; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2207; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2208 %x.shr = ashr i64 %x, 40 2209 %result = srem i64 %x.shr, 32768 2210 ret i64 %result 2211} 2212