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: v_mov_b32_e32 v2, 0 10; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0x9 11; GCN-NEXT: s_mov_b32 s7, 0xf000 12; GCN-NEXT: s_mov_b32 s6, -1 13; GCN-NEXT: s_waitcnt lgkmcnt(0) 14; GCN-NEXT: v_cvt_f32_u32_e32 v0, s12 15; GCN-NEXT: v_cvt_f32_u32_e32 v1, s13 16; GCN-NEXT: s_sub_u32 s2, 0, s12 17; GCN-NEXT: s_subb_u32 s3, 0, s13 18; GCN-NEXT: s_mov_b32 s4, s8 19; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 20; GCN-NEXT: v_rcp_f32_e32 v0, v0 21; GCN-NEXT: v_mov_b32_e32 v1, 0 22; GCN-NEXT: s_mov_b32 s5, s9 23; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 24; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v0 25; GCN-NEXT: v_trunc_f32_e32 v3, v3 26; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v3 27; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 28; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 29; GCN-NEXT: v_mul_hi_u32 v5, s2, v0 30; GCN-NEXT: v_mul_lo_u32 v4, s2, v3 31; GCN-NEXT: v_mul_lo_u32 v7, s3, v0 32; GCN-NEXT: v_mul_lo_u32 v6, s2, v0 33; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 34; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v7 35; GCN-NEXT: v_mul_hi_u32 v5, v0, v6 36; GCN-NEXT: v_mul_lo_u32 v7, v0, v4 37; GCN-NEXT: v_mul_hi_u32 v9, v0, v4 38; GCN-NEXT: v_mul_lo_u32 v8, v3, v6 39; GCN-NEXT: v_mul_hi_u32 v6, v3, v6 40; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v7 41; GCN-NEXT: v_addc_u32_e32 v7, vcc, v2, v9, vcc 42; GCN-NEXT: v_mul_hi_u32 v9, v3, v4 43; GCN-NEXT: v_mul_lo_u32 v4, v3, v4 44; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 45; GCN-NEXT: v_addc_u32_e32 v5, vcc, v7, v6, vcc 46; GCN-NEXT: v_addc_u32_e32 v6, vcc, v9, v1, vcc 47; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 48; GCN-NEXT: v_add_i32_e64 v0, s[0:1], v0, v4 49; GCN-NEXT: v_addc_u32_e32 v5, vcc, v2, v6, vcc 50; GCN-NEXT: v_addc_u32_e64 v4, vcc, v3, v5, s[0:1] 51; GCN-NEXT: v_mul_lo_u32 v6, s2, v4 52; GCN-NEXT: v_mul_hi_u32 v7, s2, v0 53; GCN-NEXT: v_mul_lo_u32 v8, s3, v0 54; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 55; GCN-NEXT: v_mul_lo_u32 v7, s2, v0 56; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 57; GCN-NEXT: v_mul_lo_u32 v10, v0, v6 58; GCN-NEXT: v_mul_hi_u32 v12, v0, v6 59; GCN-NEXT: v_mul_hi_u32 v11, v0, v7 60; GCN-NEXT: v_mul_hi_u32 v9, v4, v7 61; GCN-NEXT: v_mul_lo_u32 v7, v4, v7 62; GCN-NEXT: v_mul_hi_u32 v8, v4, v6 63; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 64; GCN-NEXT: v_addc_u32_e32 v11, vcc, v2, v12, vcc 65; GCN-NEXT: v_mul_lo_u32 v4, v4, v6 66; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 67; GCN-NEXT: v_addc_u32_e32 v7, vcc, v11, v9, vcc 68; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v1, vcc 69; GCN-NEXT: v_add_i32_e32 v4, vcc, v7, v4 70; GCN-NEXT: v_addc_u32_e32 v6, vcc, v2, v6, vcc 71; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 72; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v6, s[0:1] 73; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4 74; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 75; GCN-NEXT: v_mul_lo_u32 v4, s10, v3 76; GCN-NEXT: v_mul_hi_u32 v5, s10, v0 77; GCN-NEXT: v_mul_hi_u32 v6, s10, v3 78; GCN-NEXT: v_mul_hi_u32 v7, s11, v3 79; GCN-NEXT: v_mul_lo_u32 v3, s11, v3 80; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 81; GCN-NEXT: v_addc_u32_e32 v5, vcc, v2, v6, vcc 82; GCN-NEXT: v_mul_lo_u32 v6, s11, v0 83; GCN-NEXT: v_mul_hi_u32 v0, s11, v0 84; GCN-NEXT: v_add_i32_e32 v4, vcc, v6, v4 85; GCN-NEXT: v_addc_u32_e32 v0, vcc, v5, v0, vcc 86; GCN-NEXT: v_addc_u32_e32 v1, vcc, v7, v1, vcc 87; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v3 88; GCN-NEXT: v_addc_u32_e32 v1, vcc, v2, v1, vcc 89; GCN-NEXT: v_mul_lo_u32 v1, s12, v1 90; GCN-NEXT: v_mul_hi_u32 v2, s12, v0 91; GCN-NEXT: v_mul_lo_u32 v3, s13, v0 92; GCN-NEXT: v_mul_lo_u32 v0, s12, v0 93; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 94; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 95; GCN-NEXT: v_sub_i32_e32 v2, vcc, s11, v1 96; GCN-NEXT: v_mov_b32_e32 v3, s13 97; GCN-NEXT: v_sub_i32_e32 v0, vcc, s10, v0 98; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 99; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s12, v0 100; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 101; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s13, v5 102; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 103; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 104; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s12, v4 105; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s12, v4 106; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 107; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s13, v5 108; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 109; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 110; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 111; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 112; GCN-NEXT: v_mov_b32_e32 v5, s11 113; GCN-NEXT: v_subb_u32_e32 v1, vcc, v5, v1, vcc 114; GCN-NEXT: v_cmp_le_u32_e32 vcc, s13, v1 115; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 116; GCN-NEXT: v_cmp_le_u32_e32 vcc, s12, v0 117; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 118; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s13, v1 119; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 120; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 121; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 122; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 123; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 124; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 125; GCN-NEXT: s_endpgm 126; 127; GCN-IR-LABEL: s_test_srem: 128; GCN-IR: ; %bb.0: ; %_udiv-special-cases 129; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 130; GCN-IR-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xd 131; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 132; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[8:9], s[6:7], 0 133; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], s[2:3], 0 134; GCN-IR-NEXT: s_flbit_i32_b32 s10, s2 135; GCN-IR-NEXT: s_or_b64 s[8:9], s[0:1], s[8:9] 136; GCN-IR-NEXT: s_flbit_i32_b32 s0, s6 137; GCN-IR-NEXT: s_flbit_i32_b32 s11, s3 138; GCN-IR-NEXT: s_add_i32 s10, s10, 32 139; GCN-IR-NEXT: v_mov_b32_e32 v0, s11 140; GCN-IR-NEXT: v_mov_b32_e32 v1, s10 141; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s3, 0 142; GCN-IR-NEXT: s_add_i32 s0, s0, 32 143; GCN-IR-NEXT: s_flbit_i32_b32 s1, s7 144; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 145; GCN-IR-NEXT: v_mov_b32_e32 v0, s1 146; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 147; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s7, 0 148; GCN-IR-NEXT: v_cndmask_b32_e32 v3, v0, v1, vcc 149; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v2, v3 150; GCN-IR-NEXT: v_subb_u32_e64 v1, s[0:1], 0, 0, vcc 151; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 152; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], 63, v[0:1] 153; GCN-IR-NEXT: s_or_b64 s[8:9], s[8:9], vcc 154; GCN-IR-NEXT: s_or_b64 s[0:1], s[8:9], s[0:1] 155; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 156; GCN-IR-NEXT: s_cbranch_vccz BB0_2 157; GCN-IR-NEXT: ; %bb.1: 158; GCN-IR-NEXT: v_mov_b32_e32 v0, s7 159; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[8:9] 160; GCN-IR-NEXT: v_mov_b32_e32 v0, s6 161; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[8:9] 162; GCN-IR-NEXT: s_branch BB0_7 163; GCN-IR-NEXT: BB0_2: ; %udiv-bb1 164; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 1, v0 165; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc 166; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[4:5], v[0:1] 167; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 168; GCN-IR-NEXT: v_lshl_b64 v[0:1], s[6:7], v0 169; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 170; GCN-IR-NEXT: s_cbranch_vccz BB0_4 171; GCN-IR-NEXT: ; %bb.3: 172; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 173; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 174; GCN-IR-NEXT: s_branch BB0_6 175; GCN-IR-NEXT: BB0_4: ; %udiv-preheader 176; GCN-IR-NEXT: v_not_b32_e32 v2, v2 177; GCN-IR-NEXT: s_add_u32 s8, s2, -1 178; GCN-IR-NEXT: v_lshr_b64 v[6:7], s[6:7], v4 179; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, v2, v3 180; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 181; GCN-IR-NEXT: s_addc_u32 s9, s3, -1 182; GCN-IR-NEXT: v_addc_u32_e64 v5, s[0:1], -1, 0, vcc 183; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 184; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 185; GCN-IR-NEXT: BB0_5: ; %udiv-do-while 186; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 187; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 188; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 189; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 190; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 191; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 192; GCN-IR-NEXT: v_mov_b32_e32 v2, s9 193; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, s8, v6 194; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, v2, v7, vcc 195; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 196; GCN-IR-NEXT: v_and_b32_e32 v10, s2, v8 197; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 198; GCN-IR-NEXT: v_and_b32_e32 v11, s3, v8 199; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 200; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 201; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 202; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 203; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 204; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 205; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 206; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 207; GCN-IR-NEXT: v_subb_u32_e64 v7, s[0:1], v7, v11, s[0:1] 208; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 209; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 210; GCN-IR-NEXT: s_cbranch_vccz BB0_5 211; GCN-IR-NEXT: BB0_6: ; %udiv-loop-exit 212; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 213; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 214; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 215; GCN-IR-NEXT: BB0_7: ; %udiv-end 216; GCN-IR-NEXT: v_mul_lo_u32 v1, s2, v1 217; GCN-IR-NEXT: v_mul_hi_u32 v2, s2, v0 218; GCN-IR-NEXT: v_mul_lo_u32 v3, s3, v0 219; GCN-IR-NEXT: v_mul_lo_u32 v0, s2, v0 220; GCN-IR-NEXT: s_mov_b32 s11, 0xf000 221; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 222; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 223; GCN-IR-NEXT: v_mov_b32_e32 v2, s7 224; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 225; GCN-IR-NEXT: s_mov_b32 s10, -1 226; GCN-IR-NEXT: s_mov_b32 s8, s4 227; GCN-IR-NEXT: s_mov_b32 s9, s5 228; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 229; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 230; GCN-IR-NEXT: s_endpgm 231 %result = urem i64 %x, %y 232 store i64 %result, i64 addrspace(1)* %out 233 ret void 234} 235 236define i64 @v_test_srem(i64 %x, i64 %y) { 237; GCN-LABEL: v_test_srem: 238; GCN: ; %bb.0: 239; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 240; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v3 241; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 242; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v4, vcc 243; GCN-NEXT: v_xor_b32_e32 v3, v3, v4 244; GCN-NEXT: v_xor_b32_e32 v2, v2, v4 245; GCN-NEXT: v_cvt_f32_u32_e32 v4, v2 246; GCN-NEXT: v_cvt_f32_u32_e32 v5, v3 247; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v2 248; GCN-NEXT: v_subb_u32_e32 v7, vcc, 0, v3, vcc 249; GCN-NEXT: v_mov_b32_e32 v14, 0 250; GCN-NEXT: v_mac_f32_e32 v4, 0x4f800000, v5 251; GCN-NEXT: v_rcp_f32_e32 v4, v4 252; GCN-NEXT: v_mov_b32_e32 v13, 0 253; GCN-NEXT: v_mul_f32_e32 v4, 0x5f7ffffc, v4 254; GCN-NEXT: v_mul_f32_e32 v5, 0x2f800000, v4 255; GCN-NEXT: v_trunc_f32_e32 v5, v5 256; GCN-NEXT: v_mac_f32_e32 v4, 0xcf800000, v5 257; GCN-NEXT: v_cvt_u32_f32_e32 v4, v4 258; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5 259; GCN-NEXT: v_mul_hi_u32 v8, v6, v4 260; GCN-NEXT: v_mul_lo_u32 v9, v6, v5 261; GCN-NEXT: v_mul_lo_u32 v10, v7, v4 262; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v9 263; GCN-NEXT: v_mul_lo_u32 v9, v6, v4 264; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v10 265; GCN-NEXT: v_mul_lo_u32 v11, v4, v8 266; GCN-NEXT: v_mul_hi_u32 v10, v4, v8 267; GCN-NEXT: v_mul_hi_u32 v12, v4, v9 268; GCN-NEXT: v_mul_hi_u32 v15, v5, v8 269; GCN-NEXT: v_mul_lo_u32 v8, v5, v8 270; GCN-NEXT: v_add_i32_e32 v11, vcc, v12, v11 271; GCN-NEXT: v_mul_lo_u32 v12, v5, v9 272; GCN-NEXT: v_mul_hi_u32 v9, v5, v9 273; GCN-NEXT: v_addc_u32_e32 v10, vcc, v14, v10, vcc 274; GCN-NEXT: v_add_i32_e32 v11, vcc, v11, v12 275; GCN-NEXT: v_addc_u32_e32 v9, vcc, v10, v9, vcc 276; GCN-NEXT: v_addc_u32_e32 v10, vcc, v15, v13, vcc 277; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 278; GCN-NEXT: v_add_i32_e64 v4, s[4:5], v4, v8 279; GCN-NEXT: v_addc_u32_e32 v9, vcc, v14, v10, vcc 280; GCN-NEXT: v_addc_u32_e64 v8, vcc, v5, v9, s[4:5] 281; GCN-NEXT: v_mul_lo_u32 v10, v6, v8 282; GCN-NEXT: v_mul_hi_u32 v11, v6, v4 283; GCN-NEXT: v_mul_lo_u32 v7, v7, v4 284; GCN-NEXT: v_mul_lo_u32 v6, v6, v4 285; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 286; GCN-NEXT: v_add_i32_e32 v7, vcc, v10, v7 287; GCN-NEXT: v_mul_lo_u32 v12, v4, v7 288; GCN-NEXT: v_mul_hi_u32 v15, v4, v6 289; GCN-NEXT: v_mul_hi_u32 v16, v4, v7 290; GCN-NEXT: v_mul_hi_u32 v11, v8, v6 291; GCN-NEXT: v_mul_lo_u32 v6, v8, v6 292; GCN-NEXT: v_add_i32_e32 v12, vcc, v15, v12 293; GCN-NEXT: v_mul_hi_u32 v10, v8, v7 294; GCN-NEXT: v_addc_u32_e32 v15, vcc, v14, v16, vcc 295; GCN-NEXT: v_mul_lo_u32 v7, v8, v7 296; GCN-NEXT: v_add_i32_e32 v6, vcc, v12, v6 297; GCN-NEXT: v_addc_u32_e32 v6, vcc, v15, v11, vcc 298; GCN-NEXT: v_addc_u32_e32 v8, vcc, v10, v13, vcc 299; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 300; GCN-NEXT: v_addc_u32_e32 v7, vcc, v14, v8, vcc 301; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v9 302; GCN-NEXT: v_addc_u32_e64 v5, vcc, v5, v7, s[4:5] 303; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v6 304; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc 305; GCN-NEXT: v_ashrrev_i32_e32 v6, 31, v1 306; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v6 307; GCN-NEXT: v_xor_b32_e32 v0, v0, v6 308; GCN-NEXT: v_mul_lo_u32 v7, v0, v5 309; GCN-NEXT: v_mul_hi_u32 v8, v0, v4 310; GCN-NEXT: v_mul_hi_u32 v9, v0, v5 311; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v6, vcc 312; GCN-NEXT: v_xor_b32_e32 v1, v1, v6 313; GCN-NEXT: v_add_i32_e32 v7, vcc, v8, v7 314; GCN-NEXT: v_addc_u32_e32 v8, vcc, v14, v9, vcc 315; GCN-NEXT: v_mul_lo_u32 v9, v1, v4 316; GCN-NEXT: v_mul_hi_u32 v4, v1, v4 317; GCN-NEXT: v_mul_hi_u32 v10, v1, v5 318; GCN-NEXT: v_mul_lo_u32 v5, v1, v5 319; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v9 320; GCN-NEXT: v_addc_u32_e32 v4, vcc, v8, v4, vcc 321; GCN-NEXT: v_addc_u32_e32 v7, vcc, v10, v13, vcc 322; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 323; GCN-NEXT: v_addc_u32_e32 v5, vcc, v14, v7, vcc 324; GCN-NEXT: v_mul_lo_u32 v5, v2, v5 325; GCN-NEXT: v_mul_hi_u32 v7, v2, v4 326; GCN-NEXT: v_mul_lo_u32 v8, v3, v4 327; GCN-NEXT: v_mul_lo_u32 v4, v2, v4 328; GCN-NEXT: v_add_i32_e32 v5, vcc, v7, v5 329; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v8 330; GCN-NEXT: v_sub_i32_e32 v7, vcc, v1, v5 331; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 332; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v7, v3, vcc 333; GCN-NEXT: v_sub_i32_e64 v7, s[4:5], v0, v2 334; GCN-NEXT: v_subbrev_u32_e64 v8, s[6:7], 0, v4, s[4:5] 335; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v8, v3 336; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 337; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[6:7] 338; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v7, v2 339; GCN-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[6:7] 340; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v8, v3 341; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3 342; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v3, s[4:5] 343; GCN-NEXT: v_cndmask_b32_e64 v9, v9, v10, s[6:7] 344; GCN-NEXT: v_sub_i32_e64 v10, s[4:5], v7, v2 345; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 346; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2 347; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 348; GCN-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc 349; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 350; GCN-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc 351; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v9 352; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 353; GCN-NEXT: v_cndmask_b32_e64 v2, v7, v10, s[4:5] 354; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 355; GCN-NEXT: v_cndmask_b32_e64 v4, v8, v4, s[4:5] 356; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc 357; GCN-NEXT: v_xor_b32_e32 v0, v0, v6 358; GCN-NEXT: v_xor_b32_e32 v1, v1, v6 359; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v6 360; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v6, vcc 361; GCN-NEXT: s_setpc_b64 s[30:31] 362; 363; GCN-IR-LABEL: v_test_srem: 364; GCN-IR: ; %bb.0: ; %_udiv-special-cases 365; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 366; GCN-IR-NEXT: v_ashrrev_i32_e32 v4, 31, v1 367; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v4 368; GCN-IR-NEXT: v_ashrrev_i32_e32 v6, 31, v3 369; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v4 370; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 371; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v4, vcc 372; GCN-IR-NEXT: v_xor_b32_e32 v2, v2, v6 373; GCN-IR-NEXT: v_sub_i32_e32 v5, vcc, v2, v6 374; GCN-IR-NEXT: v_xor_b32_e32 v3, v3, v6 375; GCN-IR-NEXT: v_subb_u32_e32 v6, vcc, v3, v6, vcc 376; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[5:6] 377; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 378; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v5 379; GCN-IR-NEXT: s_or_b64 s[6:7], vcc, s[4:5] 380; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, 32, v3 381; GCN-IR-NEXT: v_ffbh_u32_e32 v7, v6 382; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v6 383; GCN-IR-NEXT: v_cndmask_b32_e32 v12, v7, v3, vcc 384; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v0 385; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, 32, v3 386; GCN-IR-NEXT: v_ffbh_u32_e32 v7, v1 387; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1 388; GCN-IR-NEXT: v_cndmask_b32_e32 v14, v7, v3, vcc 389; GCN-IR-NEXT: v_sub_i32_e32 v7, vcc, v12, v14 390; GCN-IR-NEXT: v_subb_u32_e64 v8, s[4:5], 0, 0, vcc 391; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[7:8] 392; GCN-IR-NEXT: v_cmp_ne_u64_e64 s[4:5], 63, v[7:8] 393; GCN-IR-NEXT: s_or_b64 s[6:7], s[6:7], vcc 394; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 395; GCN-IR-NEXT: s_xor_b64 s[8:9], s[6:7], -1 396; GCN-IR-NEXT: v_mov_b32_e32 v2, v4 397; GCN-IR-NEXT: v_mov_b32_e32 v15, v13 398; GCN-IR-NEXT: v_cndmask_b32_e64 v3, v1, 0, s[6:7] 399; GCN-IR-NEXT: s_and_b64 s[4:5], s[8:9], s[4:5] 400; GCN-IR-NEXT: v_cndmask_b32_e64 v9, v0, 0, s[6:7] 401; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 402; GCN-IR-NEXT: s_cbranch_execz BB1_6 403; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 404; GCN-IR-NEXT: v_add_i32_e32 v9, vcc, 1, v7 405; GCN-IR-NEXT: v_addc_u32_e32 v10, vcc, 0, v8, vcc 406; GCN-IR-NEXT: v_sub_i32_e64 v3, s[4:5], 63, v7 407; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[9:10], v[7:8] 408; GCN-IR-NEXT: v_mov_b32_e32 v10, 0 409; GCN-IR-NEXT: v_lshl_b64 v[7:8], v[0:1], v3 410; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 411; GCN-IR-NEXT: v_mov_b32_e32 v11, 0 412; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 413; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 414; GCN-IR-NEXT: s_cbranch_execz BB1_5 415; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 416; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, -1, v5 417; GCN-IR-NEXT: v_lshr_b64 v[16:17], v[0:1], v9 418; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, -1, v6, vcc 419; GCN-IR-NEXT: v_not_b32_e32 v10, v12 420; GCN-IR-NEXT: v_not_b32_e32 v11, v13 421; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, v10, v14 422; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, v11, v15, vcc 423; GCN-IR-NEXT: v_mov_b32_e32 v14, 0 424; GCN-IR-NEXT: v_mov_b32_e32 v15, 0 425; GCN-IR-NEXT: BB1_3: ; %udiv-do-while 426; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 427; GCN-IR-NEXT: v_lshl_b64 v[16:17], v[16:17], 1 428; GCN-IR-NEXT: v_lshrrev_b32_e32 v10, 31, v8 429; GCN-IR-NEXT: v_or_b32_e32 v16, v16, v10 430; GCN-IR-NEXT: v_lshl_b64 v[7:8], v[7:8], 1 431; GCN-IR-NEXT: v_sub_i32_e32 v10, vcc, v3, v16 432; GCN-IR-NEXT: v_subb_u32_e32 v10, vcc, v9, v17, vcc 433; GCN-IR-NEXT: v_or_b32_e32 v7, v14, v7 434; GCN-IR-NEXT: v_ashrrev_i32_e32 v14, 31, v10 435; GCN-IR-NEXT: v_and_b32_e32 v19, v14, v5 436; GCN-IR-NEXT: v_and_b32_e32 v10, 1, v14 437; GCN-IR-NEXT: v_and_b32_e32 v18, v14, v6 438; GCN-IR-NEXT: v_add_i32_e32 v14, vcc, 1, v12 439; GCN-IR-NEXT: v_or_b32_e32 v8, v15, v8 440; GCN-IR-NEXT: v_addc_u32_e32 v15, vcc, 0, v13, vcc 441; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[14:15], v[12:13] 442; GCN-IR-NEXT: v_mov_b32_e32 v12, v14 443; GCN-IR-NEXT: v_mov_b32_e32 v11, 0 444; GCN-IR-NEXT: v_sub_i32_e64 v16, s[4:5], v16, v19 445; GCN-IR-NEXT: v_mov_b32_e32 v13, v15 446; GCN-IR-NEXT: v_mov_b32_e32 v15, v11 447; GCN-IR-NEXT: v_subb_u32_e64 v17, s[4:5], v17, v18, s[4:5] 448; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 449; GCN-IR-NEXT: v_mov_b32_e32 v14, v10 450; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 451; GCN-IR-NEXT: s_cbranch_execnz BB1_3 452; GCN-IR-NEXT: ; %bb.4: ; %Flow 453; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 454; GCN-IR-NEXT: BB1_5: ; %Flow3 455; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 456; GCN-IR-NEXT: v_lshl_b64 v[7:8], v[7:8], 1 457; GCN-IR-NEXT: v_or_b32_e32 v3, v11, v8 458; GCN-IR-NEXT: v_or_b32_e32 v9, v10, v7 459; GCN-IR-NEXT: BB1_6: ; %Flow4 460; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 461; GCN-IR-NEXT: v_mul_lo_u32 v3, v5, v3 462; GCN-IR-NEXT: v_mul_hi_u32 v7, v5, v9 463; GCN-IR-NEXT: v_mul_lo_u32 v6, v6, v9 464; GCN-IR-NEXT: v_mul_lo_u32 v5, v5, v9 465; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, v7, v3 466; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, v3, v6 467; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v5 468; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 469; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v4 470; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 471; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 472; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 473; GCN-IR-NEXT: s_setpc_b64 s[30:31] 474 %result = srem i64 %x, %y 475 ret i64 %result 476} 477 478define amdgpu_kernel void @s_test_srem23_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 479; GCN-LABEL: s_test_srem23_64: 480; GCN: ; %bb.0: 481; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 482; GCN-NEXT: s_load_dword s1, s[0:1], 0xe 483; GCN-NEXT: s_mov_b32 s3, 0xf000 484; GCN-NEXT: s_mov_b32 s2, -1 485; GCN-NEXT: s_waitcnt lgkmcnt(0) 486; GCN-NEXT: s_ashr_i64 s[6:7], s[6:7], 41 487; GCN-NEXT: s_ashr_i64 s[0:1], s[0:1], 41 488; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 489; GCN-NEXT: v_cvt_f32_i32_e32 v1, s6 490; GCN-NEXT: s_xor_b32 s1, s6, s0 491; GCN-NEXT: s_ashr_i32 s1, s1, 30 492; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 493; GCN-NEXT: s_or_b32 s1, s1, 1 494; GCN-NEXT: v_mov_b32_e32 v3, s1 495; GCN-NEXT: s_mov_b32 s1, s5 496; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 497; GCN-NEXT: v_trunc_f32_e32 v2, v2 498; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 499; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 500; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 501; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 502; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 503; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 504; GCN-NEXT: s_mov_b32 s0, s4 505; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 506; GCN-NEXT: v_bfe_i32 v0, v0, 0, 23 507; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 508; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 509; GCN-NEXT: s_endpgm 510; 511; GCN-IR-LABEL: s_test_srem23_64: 512; GCN-IR: ; %bb.0: 513; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 514; GCN-IR-NEXT: s_load_dword s1, s[0:1], 0xe 515; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 516; GCN-IR-NEXT: s_mov_b32 s2, -1 517; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 518; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[6:7], 41 519; GCN-IR-NEXT: s_ashr_i64 s[0:1], s[0:1], 41 520; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s0 521; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s6 522; GCN-IR-NEXT: s_xor_b32 s1, s6, s0 523; GCN-IR-NEXT: s_ashr_i32 s1, s1, 30 524; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 525; GCN-IR-NEXT: s_or_b32 s1, s1, 1 526; GCN-IR-NEXT: v_mov_b32_e32 v3, s1 527; GCN-IR-NEXT: s_mov_b32 s1, s5 528; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 529; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 530; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 531; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 532; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 533; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 534; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 535; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 536; GCN-IR-NEXT: s_mov_b32 s0, s4 537; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 538; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 23 539; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 540; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 541; GCN-IR-NEXT: s_endpgm 542 %1 = ashr i64 %x, 41 543 %2 = ashr i64 %y, 41 544 %result = srem i64 %1, %2 545 store i64 %result, i64 addrspace(1)* %out 546 ret void 547} 548 549define amdgpu_kernel void @s_test_srem24_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 550; GCN-LABEL: s_test_srem24_64: 551; GCN: ; %bb.0: 552; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 553; GCN-NEXT: s_load_dword s1, s[0:1], 0xe 554; GCN-NEXT: s_mov_b32 s3, 0xf000 555; GCN-NEXT: s_mov_b32 s2, -1 556; GCN-NEXT: s_waitcnt lgkmcnt(0) 557; GCN-NEXT: s_ashr_i64 s[6:7], s[6:7], 40 558; GCN-NEXT: s_ashr_i64 s[0:1], s[0:1], 40 559; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 560; GCN-NEXT: v_cvt_f32_i32_e32 v1, s6 561; GCN-NEXT: s_xor_b32 s1, s6, s0 562; GCN-NEXT: s_ashr_i32 s1, s1, 30 563; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 564; GCN-NEXT: s_or_b32 s1, s1, 1 565; GCN-NEXT: v_mov_b32_e32 v3, s1 566; GCN-NEXT: s_mov_b32 s1, s5 567; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 568; GCN-NEXT: v_trunc_f32_e32 v2, v2 569; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 570; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 571; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 572; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 573; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 574; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 575; GCN-NEXT: s_mov_b32 s0, s4 576; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 577; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 578; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 579; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 580; GCN-NEXT: s_endpgm 581; 582; GCN-IR-LABEL: s_test_srem24_64: 583; GCN-IR: ; %bb.0: 584; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 585; GCN-IR-NEXT: s_load_dword s1, s[0:1], 0xe 586; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 587; GCN-IR-NEXT: s_mov_b32 s2, -1 588; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 589; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[6:7], 40 590; GCN-IR-NEXT: s_ashr_i64 s[0:1], s[0:1], 40 591; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s0 592; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s6 593; GCN-IR-NEXT: s_xor_b32 s1, s6, s0 594; GCN-IR-NEXT: s_ashr_i32 s1, s1, 30 595; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 596; GCN-IR-NEXT: s_or_b32 s1, s1, 1 597; GCN-IR-NEXT: v_mov_b32_e32 v3, s1 598; GCN-IR-NEXT: s_mov_b32 s1, s5 599; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 600; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 601; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 602; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 603; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 604; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 605; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 606; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 607; GCN-IR-NEXT: s_mov_b32 s0, s4 608; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 609; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 610; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 611; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 612; GCN-IR-NEXT: s_endpgm 613 %1 = ashr i64 %x, 40 614 %2 = ashr i64 %y, 40 615 %result = srem i64 %1, %2 616 store i64 %result, i64 addrspace(1)* %out 617 ret void 618} 619 620define i64 @v_test_srem24_64(i64 %x, i64 %y) { 621; GCN-LABEL: v_test_srem24_64: 622; GCN: ; %bb.0: 623; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 624; GCN-NEXT: v_ashr_i64 v[2:3], v[2:3], 40 625; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 626; GCN-NEXT: v_cvt_f32_i32_e32 v3, v2 627; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 628; GCN-NEXT: v_xor_b32_e32 v5, v0, v2 629; GCN-NEXT: v_ashrrev_i32_e32 v5, 30, v5 630; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v3 631; GCN-NEXT: v_or_b32_e32 v5, 1, v5 632; GCN-NEXT: v_mul_f32_e32 v4, v1, v4 633; GCN-NEXT: v_trunc_f32_e32 v4, v4 634; GCN-NEXT: v_mad_f32 v1, -v4, v3, v1 635; GCN-NEXT: v_cvt_i32_f32_e32 v4, v4 636; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v3| 637; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc 638; GCN-NEXT: v_add_i32_e32 v1, vcc, v4, v1 639; GCN-NEXT: v_mul_lo_u32 v1, v1, v2 640; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 641; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 642; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 643; GCN-NEXT: s_setpc_b64 s[30:31] 644; 645; GCN-IR-LABEL: v_test_srem24_64: 646; GCN-IR: ; %bb.0: 647; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 648; GCN-IR-NEXT: v_ashr_i64 v[2:3], v[2:3], 40 649; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 650; GCN-IR-NEXT: v_cvt_f32_i32_e32 v3, v2 651; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 652; GCN-IR-NEXT: v_xor_b32_e32 v5, v0, v2 653; GCN-IR-NEXT: v_ashrrev_i32_e32 v5, 30, v5 654; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v4, v3 655; GCN-IR-NEXT: v_or_b32_e32 v5, 1, v5 656; GCN-IR-NEXT: v_mul_f32_e32 v4, v1, v4 657; GCN-IR-NEXT: v_trunc_f32_e32 v4, v4 658; GCN-IR-NEXT: v_mad_f32 v1, -v4, v3, v1 659; GCN-IR-NEXT: v_cvt_i32_f32_e32 v4, v4 660; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v3| 661; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc 662; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v4, v1 663; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v2 664; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 665; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 666; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 667; GCN-IR-NEXT: s_setpc_b64 s[30:31] 668 %1 = ashr i64 %x, 40 669 %2 = ashr i64 %y, 40 670 %result = srem i64 %1, %2 671 ret i64 %result 672} 673 674define amdgpu_kernel void @s_test_srem25_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 675; GCN-LABEL: s_test_srem25_64: 676; GCN: ; %bb.0: 677; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 678; GCN-NEXT: s_load_dword s1, s[0:1], 0xe 679; GCN-NEXT: s_mov_b32 s3, 0xf000 680; GCN-NEXT: s_mov_b32 s2, -1 681; GCN-NEXT: s_waitcnt lgkmcnt(0) 682; GCN-NEXT: s_ashr_i64 s[6:7], s[6:7], 39 683; GCN-NEXT: s_ashr_i64 s[0:1], s[0:1], 39 684; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 685; GCN-NEXT: v_cvt_f32_i32_e32 v1, s6 686; GCN-NEXT: s_xor_b32 s1, s6, s0 687; GCN-NEXT: s_ashr_i32 s1, s1, 30 688; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 689; GCN-NEXT: s_or_b32 s1, s1, 1 690; GCN-NEXT: v_mov_b32_e32 v3, s1 691; GCN-NEXT: s_mov_b32 s1, s5 692; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 693; GCN-NEXT: v_trunc_f32_e32 v2, v2 694; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 695; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 696; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 697; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 698; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 699; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 700; GCN-NEXT: s_mov_b32 s0, s4 701; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 702; GCN-NEXT: v_bfe_i32 v0, v0, 0, 25 703; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 704; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 705; GCN-NEXT: s_endpgm 706; 707; GCN-IR-LABEL: s_test_srem25_64: 708; GCN-IR: ; %bb.0: 709; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 710; GCN-IR-NEXT: s_load_dword s1, s[0:1], 0xe 711; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 712; GCN-IR-NEXT: s_mov_b32 s2, -1 713; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 714; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[6:7], 39 715; GCN-IR-NEXT: s_ashr_i64 s[0:1], s[0:1], 39 716; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s0 717; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s6 718; GCN-IR-NEXT: s_xor_b32 s1, s6, s0 719; GCN-IR-NEXT: s_ashr_i32 s1, s1, 30 720; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 721; GCN-IR-NEXT: s_or_b32 s1, s1, 1 722; GCN-IR-NEXT: v_mov_b32_e32 v3, s1 723; GCN-IR-NEXT: s_mov_b32 s1, s5 724; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 725; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 726; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 727; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 728; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 729; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 730; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 731; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 732; GCN-IR-NEXT: s_mov_b32 s0, s4 733; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 734; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 25 735; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 736; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 737; GCN-IR-NEXT: s_endpgm 738 %1 = ashr i64 %x, 39 739 %2 = ashr i64 %y, 39 740 %result = srem i64 %1, %2 741 store i64 %result, i64 addrspace(1)* %out 742 ret void 743} 744 745define amdgpu_kernel void @s_test_srem31_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 746; GCN-LABEL: s_test_srem31_64: 747; GCN: ; %bb.0: 748; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 749; GCN-NEXT: s_load_dword s1, s[0:1], 0xe 750; GCN-NEXT: s_mov_b32 s3, 0xf000 751; GCN-NEXT: s_mov_b32 s2, -1 752; GCN-NEXT: s_waitcnt lgkmcnt(0) 753; GCN-NEXT: s_ashr_i64 s[6:7], s[6:7], 33 754; GCN-NEXT: s_ashr_i64 s[0:1], s[0:1], 33 755; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 756; GCN-NEXT: v_cvt_f32_i32_e32 v1, s6 757; GCN-NEXT: s_xor_b32 s1, s6, s0 758; GCN-NEXT: s_ashr_i32 s1, s1, 30 759; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 760; GCN-NEXT: s_or_b32 s1, s1, 1 761; GCN-NEXT: v_mov_b32_e32 v3, s1 762; GCN-NEXT: s_mov_b32 s1, s5 763; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 764; GCN-NEXT: v_trunc_f32_e32 v2, v2 765; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 766; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 767; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 768; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 769; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 770; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 771; GCN-NEXT: s_mov_b32 s0, s4 772; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 773; GCN-NEXT: v_bfe_i32 v0, v0, 0, 31 774; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 775; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 776; GCN-NEXT: s_endpgm 777; 778; GCN-IR-LABEL: s_test_srem31_64: 779; GCN-IR: ; %bb.0: 780; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 781; GCN-IR-NEXT: s_load_dword s1, s[0:1], 0xe 782; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 783; GCN-IR-NEXT: s_mov_b32 s2, -1 784; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 785; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[6:7], 33 786; GCN-IR-NEXT: s_ashr_i64 s[0:1], s[0:1], 33 787; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s0 788; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s6 789; GCN-IR-NEXT: s_xor_b32 s1, s6, s0 790; GCN-IR-NEXT: s_ashr_i32 s1, s1, 30 791; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 792; GCN-IR-NEXT: s_or_b32 s1, s1, 1 793; GCN-IR-NEXT: v_mov_b32_e32 v3, s1 794; GCN-IR-NEXT: s_mov_b32 s1, s5 795; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 796; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 797; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 798; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 799; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 800; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 801; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 802; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 803; GCN-IR-NEXT: s_mov_b32 s0, s4 804; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 805; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 31 806; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 807; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 808; GCN-IR-NEXT: s_endpgm 809 %1 = ashr i64 %x, 33 810 %2 = ashr i64 %y, 33 811 %result = srem i64 %1, %2 812 store i64 %result, i64 addrspace(1)* %out 813 ret void 814} 815 816; 32 known sign bits 817define amdgpu_kernel void @s_test_srem32_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 818; GCN-LABEL: s_test_srem32_64: 819; GCN: ; %bb.0: 820; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 821; GCN-NEXT: s_load_dword s0, s[0:1], 0xe 822; GCN-NEXT: s_mov_b32 s3, 0xf000 823; GCN-NEXT: s_mov_b32 s2, -1 824; GCN-NEXT: s_waitcnt lgkmcnt(0) 825; GCN-NEXT: v_cvt_f32_i32_e32 v1, s7 826; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 827; GCN-NEXT: s_xor_b32 s1, s7, s0 828; GCN-NEXT: s_ashr_i32 s1, s1, 30 829; GCN-NEXT: s_or_b32 s1, s1, 1 830; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 831; GCN-NEXT: v_mov_b32_e32 v3, s1 832; GCN-NEXT: s_mov_b32 s1, s5 833; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 834; GCN-NEXT: v_trunc_f32_e32 v2, v2 835; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 836; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 837; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 838; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 839; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 840; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 841; GCN-NEXT: s_mov_b32 s0, s4 842; GCN-NEXT: v_sub_i32_e32 v0, vcc, s7, v0 843; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 844; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 845; GCN-NEXT: s_endpgm 846; 847; GCN-IR-LABEL: s_test_srem32_64: 848; GCN-IR: ; %bb.0: 849; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 850; GCN-IR-NEXT: s_load_dword s0, s[0:1], 0xe 851; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 852; GCN-IR-NEXT: s_mov_b32 s2, -1 853; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 854; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, s7 855; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s0 856; GCN-IR-NEXT: s_xor_b32 s1, s7, s0 857; GCN-IR-NEXT: s_ashr_i32 s1, s1, 30 858; GCN-IR-NEXT: s_or_b32 s1, s1, 1 859; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v0 860; GCN-IR-NEXT: v_mov_b32_e32 v3, s1 861; GCN-IR-NEXT: s_mov_b32 s1, s5 862; GCN-IR-NEXT: v_mul_f32_e32 v2, v1, v2 863; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 864; GCN-IR-NEXT: v_mad_f32 v1, -v2, v0, v1 865; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 866; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 867; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 868; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 869; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 870; GCN-IR-NEXT: s_mov_b32 s0, s4 871; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s7, v0 872; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 873; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 874; GCN-IR-NEXT: s_endpgm 875 %1 = ashr i64 %x, 32 876 %2 = ashr i64 %y, 32 877 %result = srem i64 %1, %2 878 store i64 %result, i64 addrspace(1)* %out 879 ret void 880} 881 882; 33 known sign bits 883define amdgpu_kernel void @s_test_srem33_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 884; GCN-LABEL: s_test_srem33_64: 885; GCN: ; %bb.0: 886; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0x9 887; GCN-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd 888; GCN-NEXT: v_mov_b32_e32 v7, 0 889; GCN-NEXT: s_mov_b32 s7, 0xf000 890; GCN-NEXT: s_mov_b32 s6, -1 891; GCN-NEXT: s_waitcnt lgkmcnt(0) 892; GCN-NEXT: s_ashr_i64 s[2:3], s[10:11], 31 893; GCN-NEXT: s_ashr_i64 s[4:5], s[0:1], 31 894; GCN-NEXT: s_ashr_i32 s0, s1, 31 895; GCN-NEXT: s_add_u32 s4, s4, s0 896; GCN-NEXT: s_mov_b32 s1, s0 897; GCN-NEXT: s_addc_u32 s5, s5, s0 898; GCN-NEXT: s_xor_b64 s[12:13], s[4:5], s[0:1] 899; GCN-NEXT: v_cvt_f32_u32_e32 v0, s12 900; GCN-NEXT: v_cvt_f32_u32_e32 v1, s13 901; GCN-NEXT: s_sub_u32 s4, 0, s12 902; GCN-NEXT: s_subb_u32 s5, 0, s13 903; GCN-NEXT: s_ashr_i32 s10, s11, 31 904; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 905; GCN-NEXT: v_rcp_f32_e32 v0, v0 906; GCN-NEXT: v_mov_b32_e32 v1, 0 907; GCN-NEXT: s_mov_b32 s11, s10 908; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 909; GCN-NEXT: v_mul_f32_e32 v2, 0x2f800000, v0 910; GCN-NEXT: v_trunc_f32_e32 v2, v2 911; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v2 912; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 913; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 914; GCN-NEXT: v_mul_hi_u32 v4, s4, v0 915; GCN-NEXT: v_mul_lo_u32 v3, s4, v2 916; GCN-NEXT: v_mul_lo_u32 v6, s5, v0 917; GCN-NEXT: v_mul_lo_u32 v5, s4, v0 918; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 919; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v6 920; GCN-NEXT: v_mul_hi_u32 v4, v0, v5 921; GCN-NEXT: v_mul_lo_u32 v6, v0, v3 922; GCN-NEXT: v_mul_hi_u32 v8, v0, v3 923; GCN-NEXT: v_mul_hi_u32 v9, v2, v3 924; GCN-NEXT: v_mul_lo_u32 v3, v2, v3 925; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v6 926; GCN-NEXT: v_addc_u32_e32 v6, vcc, v7, v8, vcc 927; GCN-NEXT: v_mul_lo_u32 v8, v2, v5 928; GCN-NEXT: v_mul_hi_u32 v5, v2, v5 929; GCN-NEXT: v_add_i32_e32 v4, vcc, v8, v4 930; GCN-NEXT: v_addc_u32_e32 v4, vcc, v6, v5, vcc 931; GCN-NEXT: v_addc_u32_e32 v5, vcc, v9, v1, vcc 932; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 933; GCN-NEXT: v_add_i32_e64 v0, s[0:1], v0, v3 934; GCN-NEXT: v_addc_u32_e32 v4, vcc, v7, v5, vcc 935; GCN-NEXT: v_addc_u32_e64 v3, vcc, v2, v4, s[0:1] 936; GCN-NEXT: v_mul_lo_u32 v5, s4, v3 937; GCN-NEXT: v_mul_hi_u32 v6, s4, v0 938; GCN-NEXT: v_mul_lo_u32 v8, s5, v0 939; GCN-NEXT: s_mov_b32 s5, s9 940; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 941; GCN-NEXT: v_mul_lo_u32 v6, s4, v0 942; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 943; GCN-NEXT: v_mul_lo_u32 v10, v0, v5 944; GCN-NEXT: v_mul_hi_u32 v12, v0, v5 945; GCN-NEXT: v_mul_hi_u32 v11, v0, v6 946; GCN-NEXT: v_mul_hi_u32 v9, v3, v6 947; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 948; GCN-NEXT: v_mul_hi_u32 v8, v3, v5 949; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 950; GCN-NEXT: v_addc_u32_e32 v11, vcc, v7, v12, vcc 951; GCN-NEXT: v_mul_lo_u32 v3, v3, v5 952; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v10 953; GCN-NEXT: v_addc_u32_e32 v6, vcc, v11, v9, vcc 954; GCN-NEXT: v_addc_u32_e32 v5, vcc, v8, v1, vcc 955; GCN-NEXT: v_add_i32_e32 v3, vcc, v6, v3 956; GCN-NEXT: v_addc_u32_e32 v5, vcc, v7, v5, vcc 957; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 958; GCN-NEXT: v_addc_u32_e64 v2, vcc, v2, v5, s[0:1] 959; GCN-NEXT: s_add_u32 s0, s2, s10 960; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v3 961; GCN-NEXT: s_addc_u32 s1, s3, s10 962; GCN-NEXT: s_xor_b64 s[14:15], s[0:1], s[10:11] 963; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc 964; GCN-NEXT: v_mul_lo_u32 v3, s14, v2 965; GCN-NEXT: v_mul_hi_u32 v4, s14, v0 966; GCN-NEXT: v_mul_hi_u32 v5, s14, v2 967; GCN-NEXT: v_mul_hi_u32 v6, s15, v2 968; GCN-NEXT: v_mul_lo_u32 v2, s15, v2 969; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 970; GCN-NEXT: v_addc_u32_e32 v4, vcc, v7, v5, vcc 971; GCN-NEXT: v_mul_lo_u32 v5, s15, v0 972; GCN-NEXT: v_mul_hi_u32 v0, s15, v0 973; GCN-NEXT: s_mov_b32 s4, s8 974; GCN-NEXT: v_add_i32_e32 v3, vcc, v5, v3 975; GCN-NEXT: v_addc_u32_e32 v0, vcc, v4, v0, vcc 976; GCN-NEXT: v_addc_u32_e32 v1, vcc, v6, v1, vcc 977; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 978; GCN-NEXT: v_addc_u32_e32 v1, vcc, v7, v1, vcc 979; GCN-NEXT: v_mul_lo_u32 v1, s12, v1 980; GCN-NEXT: v_mul_hi_u32 v2, s12, v0 981; GCN-NEXT: v_mul_lo_u32 v3, s13, v0 982; GCN-NEXT: v_mul_lo_u32 v0, s12, v0 983; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 984; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 985; GCN-NEXT: v_sub_i32_e32 v2, vcc, s15, v1 986; GCN-NEXT: v_mov_b32_e32 v3, s13 987; GCN-NEXT: v_sub_i32_e32 v0, vcc, s14, v0 988; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 989; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s12, v0 990; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 991; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s13, v5 992; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 993; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 994; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s12, v4 995; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s12, v4 996; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 997; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s13, v5 998; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 999; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 1000; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 1001; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 1002; GCN-NEXT: v_mov_b32_e32 v5, s15 1003; GCN-NEXT: v_subb_u32_e32 v1, vcc, v5, v1, vcc 1004; GCN-NEXT: v_cmp_le_u32_e32 vcc, s13, v1 1005; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 1006; GCN-NEXT: v_cmp_le_u32_e32 vcc, s12, v0 1007; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 1008; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s13, v1 1009; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 1010; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 1011; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 1012; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 1013; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 1014; GCN-NEXT: v_xor_b32_e32 v0, s10, v0 1015; GCN-NEXT: v_xor_b32_e32 v1, s10, v1 1016; GCN-NEXT: v_mov_b32_e32 v2, s10 1017; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s10, v0 1018; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1019; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1020; GCN-NEXT: s_endpgm 1021; 1022; GCN-IR-LABEL: s_test_srem33_64: 1023; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1024; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 1025; GCN-IR-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd 1026; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1027; GCN-IR-NEXT: s_ashr_i32 s2, s7, 31 1028; GCN-IR-NEXT: s_ashr_i64 s[10:11], s[0:1], 31 1029; GCN-IR-NEXT: s_ashr_i32 s0, s1, 31 1030; GCN-IR-NEXT: s_mov_b32 s1, s0 1031; GCN-IR-NEXT: s_ashr_i64 s[8:9], s[6:7], 31 1032; GCN-IR-NEXT: s_mov_b32 s3, s2 1033; GCN-IR-NEXT: s_xor_b64 s[6:7], s[8:9], s[2:3] 1034; GCN-IR-NEXT: s_xor_b64 s[10:11], s[10:11], s[0:1] 1035; GCN-IR-NEXT: s_sub_u32 s8, s6, s2 1036; GCN-IR-NEXT: s_subb_u32 s9, s7, s2 1037; GCN-IR-NEXT: s_sub_u32 s10, s10, s0 1038; GCN-IR-NEXT: s_subb_u32 s11, s11, s0 1039; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], s[10:11], 0 1040; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[6:7], s[8:9], 0 1041; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s11, 0 1042; GCN-IR-NEXT: s_or_b64 s[6:7], s[0:1], s[6:7] 1043; GCN-IR-NEXT: s_flbit_i32_b32 s0, s10 1044; GCN-IR-NEXT: s_add_i32 s0, s0, 32 1045; GCN-IR-NEXT: s_flbit_i32_b32 s1, s11 1046; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 1047; GCN-IR-NEXT: s_flbit_i32_b32 s0, s8 1048; GCN-IR-NEXT: v_mov_b32_e32 v0, s1 1049; GCN-IR-NEXT: s_add_i32 s0, s0, 32 1050; GCN-IR-NEXT: s_flbit_i32_b32 s1, s9 1051; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 1052; GCN-IR-NEXT: v_mov_b32_e32 v0, s1 1053; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 1054; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s9, 0 1055; GCN-IR-NEXT: v_cndmask_b32_e32 v3, v0, v1, vcc 1056; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v2, v3 1057; GCN-IR-NEXT: v_subb_u32_e64 v1, s[0:1], 0, 0, vcc 1058; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 1059; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], 63, v[0:1] 1060; GCN-IR-NEXT: s_or_b64 s[6:7], s[6:7], vcc 1061; GCN-IR-NEXT: s_or_b64 s[0:1], s[6:7], s[0:1] 1062; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 1063; GCN-IR-NEXT: s_cbranch_vccz BB8_2 1064; GCN-IR-NEXT: ; %bb.1: 1065; GCN-IR-NEXT: v_mov_b32_e32 v0, s9 1066; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[6:7] 1067; GCN-IR-NEXT: v_mov_b32_e32 v0, s8 1068; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[6:7] 1069; GCN-IR-NEXT: s_branch BB8_7 1070; GCN-IR-NEXT: BB8_2: ; %udiv-bb1 1071; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 1, v0 1072; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc 1073; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[4:5], v[0:1] 1074; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 1075; GCN-IR-NEXT: v_lshl_b64 v[0:1], s[8:9], v0 1076; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 1077; GCN-IR-NEXT: s_cbranch_vccz BB8_4 1078; GCN-IR-NEXT: ; %bb.3: 1079; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 1080; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1081; GCN-IR-NEXT: s_branch BB8_6 1082; GCN-IR-NEXT: BB8_4: ; %udiv-preheader 1083; GCN-IR-NEXT: v_not_b32_e32 v2, v2 1084; GCN-IR-NEXT: s_add_u32 s6, s10, -1 1085; GCN-IR-NEXT: v_lshr_b64 v[6:7], s[8:9], v4 1086; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, v2, v3 1087; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 1088; GCN-IR-NEXT: s_addc_u32 s7, s11, -1 1089; GCN-IR-NEXT: v_addc_u32_e64 v5, s[0:1], -1, 0, vcc 1090; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1091; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1092; GCN-IR-NEXT: BB8_5: ; %udiv-do-while 1093; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1094; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 1095; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 1096; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1097; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 1098; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 1099; GCN-IR-NEXT: v_mov_b32_e32 v2, s7 1100; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, s6, v6 1101; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, v2, v7, vcc 1102; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 1103; GCN-IR-NEXT: v_and_b32_e32 v10, s10, v8 1104; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 1105; GCN-IR-NEXT: v_and_b32_e32 v11, s11, v8 1106; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 1107; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 1108; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 1109; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 1110; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 1111; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 1112; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1113; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 1114; GCN-IR-NEXT: v_subb_u32_e64 v7, s[0:1], v7, v11, s[0:1] 1115; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 1116; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 1117; GCN-IR-NEXT: s_cbranch_vccz BB8_5 1118; GCN-IR-NEXT: BB8_6: ; %udiv-loop-exit 1119; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1120; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 1121; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 1122; GCN-IR-NEXT: BB8_7: ; %udiv-end 1123; GCN-IR-NEXT: v_mul_lo_u32 v1, s10, v1 1124; GCN-IR-NEXT: v_mul_hi_u32 v2, s10, v0 1125; GCN-IR-NEXT: v_mul_lo_u32 v3, s11, v0 1126; GCN-IR-NEXT: v_mul_lo_u32 v0, s10, v0 1127; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 1128; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1129; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1130; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s8, v0 1131; GCN-IR-NEXT: v_mov_b32_e32 v2, s9 1132; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 1133; GCN-IR-NEXT: v_xor_b32_e32 v0, s2, v0 1134; GCN-IR-NEXT: v_xor_b32_e32 v1, s3, v1 1135; GCN-IR-NEXT: v_mov_b32_e32 v2, s3 1136; GCN-IR-NEXT: v_subrev_i32_e32 v0, vcc, s2, v0 1137; GCN-IR-NEXT: s_mov_b32 s6, -1 1138; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1139; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1140; GCN-IR-NEXT: s_endpgm 1141 %1 = ashr i64 %x, 31 1142 %2 = ashr i64 %y, 31 1143 %result = srem i64 %1, %2 1144 store i64 %result, i64 addrspace(1)* %out 1145 ret void 1146} 1147 1148define amdgpu_kernel void @s_test_srem24_48(i48 addrspace(1)* %out, i48 %x, i48 %y) { 1149; GCN-LABEL: s_test_srem24_48: 1150; GCN: ; %bb.0: 1151; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 1152; GCN-NEXT: s_load_dword s2, s[0:1], 0xb 1153; GCN-NEXT: s_load_dword s3, s[0:1], 0xc 1154; GCN-NEXT: s_load_dword s6, s[0:1], 0xd 1155; GCN-NEXT: s_load_dword s0, s[0:1], 0xe 1156; GCN-NEXT: s_mov_b32 s7, 0xf000 1157; GCN-NEXT: s_waitcnt lgkmcnt(0) 1158; GCN-NEXT: v_mov_b32_e32 v2, s2 1159; GCN-NEXT: s_sext_i32_i16 s1, s3 1160; GCN-NEXT: v_mov_b32_e32 v0, s6 1161; GCN-NEXT: s_sext_i32_i16 s0, s0 1162; GCN-NEXT: v_alignbit_b32 v0, s0, v0, 24 1163; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 1164; GCN-NEXT: v_alignbit_b32 v2, s1, v2, 24 1165; GCN-NEXT: v_cvt_f32_i32_e32 v3, v2 1166; GCN-NEXT: v_xor_b32_e32 v5, v2, v0 1167; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v1 1168; GCN-NEXT: v_ashrrev_i32_e32 v5, 30, v5 1169; GCN-NEXT: v_or_b32_e32 v5, 1, v5 1170; GCN-NEXT: s_mov_b32 s6, -1 1171; GCN-NEXT: v_mul_f32_e32 v4, v3, v4 1172; GCN-NEXT: v_trunc_f32_e32 v4, v4 1173; GCN-NEXT: v_mad_f32 v3, -v4, v1, v3 1174; GCN-NEXT: v_cvt_i32_f32_e32 v4, v4 1175; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v1| 1176; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc 1177; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v4 1178; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 1179; GCN-NEXT: v_sub_i32_e32 v0, vcc, v2, v0 1180; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1181; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1182; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0 1183; GCN-NEXT: buffer_store_short v1, off, s[4:7], 0 offset:4 1184; GCN-NEXT: s_endpgm 1185; 1186; GCN-IR-LABEL: s_test_srem24_48: 1187; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1188; GCN-IR-NEXT: s_load_dword s2, s[0:1], 0xb 1189; GCN-IR-NEXT: s_load_dword s3, s[0:1], 0xc 1190; GCN-IR-NEXT: s_load_dword s4, s[0:1], 0xd 1191; GCN-IR-NEXT: s_load_dword s5, s[0:1], 0xe 1192; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1193; GCN-IR-NEXT: s_sext_i32_i16 s3, s3 1194; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[2:3], 24 1195; GCN-IR-NEXT: s_sext_i32_i16 s5, s5 1196; GCN-IR-NEXT: s_ashr_i32 s2, s3, 31 1197; GCN-IR-NEXT: s_ashr_i64 s[8:9], s[4:5], 24 1198; GCN-IR-NEXT: s_ashr_i32 s4, s5, 31 1199; GCN-IR-NEXT: s_mov_b32 s3, s2 1200; GCN-IR-NEXT: s_mov_b32 s5, s4 1201; GCN-IR-NEXT: s_xor_b64 s[6:7], s[6:7], s[2:3] 1202; GCN-IR-NEXT: s_xor_b64 s[8:9], s[8:9], s[4:5] 1203; GCN-IR-NEXT: s_sub_u32 s6, s6, s2 1204; GCN-IR-NEXT: s_subb_u32 s7, s7, s2 1205; GCN-IR-NEXT: s_sub_u32 s8, s8, s4 1206; GCN-IR-NEXT: s_subb_u32 s9, s9, s4 1207; GCN-IR-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 1208; GCN-IR-NEXT: s_flbit_i32_b32 s0, s8 1209; GCN-IR-NEXT: s_add_i32 s0, s0, 32 1210; GCN-IR-NEXT: s_flbit_i32_b32 s1, s9 1211; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 1212; GCN-IR-NEXT: s_flbit_i32_b32 s0, s6 1213; GCN-IR-NEXT: v_mov_b32_e32 v0, s1 1214; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s9, 0 1215; GCN-IR-NEXT: s_add_i32 s0, s0, 32 1216; GCN-IR-NEXT: s_flbit_i32_b32 s1, s7 1217; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 1218; GCN-IR-NEXT: v_mov_b32_e32 v0, s1 1219; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 1220; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s7, 0 1221; GCN-IR-NEXT: v_cndmask_b32_e32 v3, v0, v1, vcc 1222; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v2, v3 1223; GCN-IR-NEXT: v_subb_u32_e64 v1, s[0:1], 0, 0, vcc 1224; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[10:11], s[8:9], 0 1225; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[12:13], s[6:7], 0 1226; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 1227; GCN-IR-NEXT: s_or_b64 s[10:11], s[10:11], s[12:13] 1228; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], 63, v[0:1] 1229; GCN-IR-NEXT: s_or_b64 s[10:11], s[10:11], vcc 1230; GCN-IR-NEXT: s_or_b64 s[0:1], s[10:11], s[0:1] 1231; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 1232; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1233; GCN-IR-NEXT: s_mov_b64 vcc, vcc 1234; GCN-IR-NEXT: s_cbranch_vccz BB9_2 1235; GCN-IR-NEXT: ; %bb.1: 1236; GCN-IR-NEXT: v_mov_b32_e32 v0, s7 1237; GCN-IR-NEXT: v_cndmask_b32_e64 v1, v0, 0, s[10:11] 1238; GCN-IR-NEXT: v_mov_b32_e32 v0, s6 1239; GCN-IR-NEXT: v_cndmask_b32_e64 v0, v0, 0, s[10:11] 1240; GCN-IR-NEXT: s_branch BB9_7 1241; GCN-IR-NEXT: BB9_2: ; %udiv-bb1 1242; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 1, v0 1243; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc 1244; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[4:5], v[0:1] 1245; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 1246; GCN-IR-NEXT: v_lshl_b64 v[0:1], s[6:7], v0 1247; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 1248; GCN-IR-NEXT: s_cbranch_vccz BB9_4 1249; GCN-IR-NEXT: ; %bb.3: 1250; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 1251; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1252; GCN-IR-NEXT: s_branch BB9_6 1253; GCN-IR-NEXT: BB9_4: ; %udiv-preheader 1254; GCN-IR-NEXT: v_not_b32_e32 v2, v2 1255; GCN-IR-NEXT: s_add_u32 s10, s8, -1 1256; GCN-IR-NEXT: v_lshr_b64 v[6:7], s[6:7], v4 1257; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, v2, v3 1258; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 1259; GCN-IR-NEXT: s_addc_u32 s11, s9, -1 1260; GCN-IR-NEXT: v_addc_u32_e64 v5, s[0:1], -1, 0, vcc 1261; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1262; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1263; GCN-IR-NEXT: BB9_5: ; %udiv-do-while 1264; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1265; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 1266; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 1267; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1268; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 1269; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 1270; GCN-IR-NEXT: v_mov_b32_e32 v2, s11 1271; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, s10, v6 1272; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, v2, v7, vcc 1273; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 1274; GCN-IR-NEXT: v_and_b32_e32 v10, s8, v8 1275; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 1276; GCN-IR-NEXT: v_and_b32_e32 v11, s9, v8 1277; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 1278; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 1279; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 1280; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 1281; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 1282; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 1283; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1284; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 1285; GCN-IR-NEXT: v_subb_u32_e64 v7, s[0:1], v7, v11, s[0:1] 1286; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 1287; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 1288; GCN-IR-NEXT: s_cbranch_vccz BB9_5 1289; GCN-IR-NEXT: BB9_6: ; %udiv-loop-exit 1290; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1291; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 1292; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 1293; GCN-IR-NEXT: BB9_7: ; %udiv-end 1294; GCN-IR-NEXT: v_mul_lo_u32 v1, s8, v1 1295; GCN-IR-NEXT: v_mul_hi_u32 v2, s8, v0 1296; GCN-IR-NEXT: v_mul_lo_u32 v3, s9, v0 1297; GCN-IR-NEXT: v_mul_lo_u32 v0, s8, v0 1298; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1299; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1300; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 1301; GCN-IR-NEXT: v_mov_b32_e32 v2, s7 1302; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc 1303; GCN-IR-NEXT: v_xor_b32_e32 v0, s2, v0 1304; GCN-IR-NEXT: v_xor_b32_e32 v1, s3, v1 1305; GCN-IR-NEXT: v_mov_b32_e32 v2, s3 1306; GCN-IR-NEXT: v_subrev_i32_e32 v0, vcc, s2, v0 1307; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1308; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 1309; GCN-IR-NEXT: s_mov_b32 s6, -1 1310; GCN-IR-NEXT: buffer_store_short v1, off, s[4:7], 0 offset:4 1311; GCN-IR-NEXT: buffer_store_dword v0, off, s[4:7], 0 1312; GCN-IR-NEXT: s_endpgm 1313 %1 = ashr i48 %x, 24 1314 %2 = ashr i48 %y, 24 1315 %result = srem i48 %1, %2 1316 store i48 %result, i48 addrspace(1)* %out 1317 ret void 1318} 1319 1320define amdgpu_kernel void @s_test_srem_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 1321; GCN-LABEL: s_test_srem_k_num_i64: 1322; GCN: ; %bb.0: 1323; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 1324; GCN-NEXT: v_mov_b32_e32 v2, 0 1325; GCN-NEXT: s_waitcnt lgkmcnt(0) 1326; GCN-NEXT: s_ashr_i32 s0, s7, 31 1327; GCN-NEXT: s_add_u32 s2, s6, s0 1328; GCN-NEXT: s_addc_u32 s3, s7, s0 1329; GCN-NEXT: s_mov_b32 s1, s0 1330; GCN-NEXT: s_xor_b64 s[8:9], s[2:3], s[0:1] 1331; GCN-NEXT: v_cvt_f32_u32_e32 v0, s8 1332; GCN-NEXT: v_cvt_f32_u32_e32 v1, s9 1333; GCN-NEXT: s_sub_u32 s2, 0, s8 1334; GCN-NEXT: s_subb_u32 s3, 0, s9 1335; GCN-NEXT: s_mov_b32 s7, 0xf000 1336; GCN-NEXT: v_mac_f32_e32 v0, 0x4f800000, v1 1337; GCN-NEXT: v_rcp_f32_e32 v0, v0 1338; GCN-NEXT: v_mov_b32_e32 v1, 0 1339; GCN-NEXT: s_mov_b32 s6, -1 1340; GCN-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 1341; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v0 1342; GCN-NEXT: v_trunc_f32_e32 v3, v3 1343; GCN-NEXT: v_mac_f32_e32 v0, 0xcf800000, v3 1344; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 1345; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 1346; GCN-NEXT: v_mul_hi_u32 v5, s2, v0 1347; GCN-NEXT: v_mul_lo_u32 v4, s2, v3 1348; GCN-NEXT: v_mul_lo_u32 v7, s3, v0 1349; GCN-NEXT: v_mul_lo_u32 v6, s2, v0 1350; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 1351; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v7 1352; GCN-NEXT: v_mul_hi_u32 v5, v0, v6 1353; GCN-NEXT: v_mul_lo_u32 v7, v0, v4 1354; GCN-NEXT: v_mul_hi_u32 v9, v0, v4 1355; GCN-NEXT: v_mul_lo_u32 v8, v3, v6 1356; GCN-NEXT: v_mul_hi_u32 v6, v3, v6 1357; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v7 1358; GCN-NEXT: v_addc_u32_e32 v7, vcc, v2, v9, vcc 1359; GCN-NEXT: v_mul_hi_u32 v9, v3, v4 1360; GCN-NEXT: v_mul_lo_u32 v4, v3, v4 1361; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 1362; GCN-NEXT: v_addc_u32_e32 v5, vcc, v7, v6, vcc 1363; GCN-NEXT: v_addc_u32_e32 v6, vcc, v9, v1, vcc 1364; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 1365; GCN-NEXT: v_add_i32_e64 v0, s[0:1], v0, v4 1366; GCN-NEXT: v_addc_u32_e32 v5, vcc, v2, v6, vcc 1367; GCN-NEXT: v_addc_u32_e64 v4, vcc, v3, v5, s[0:1] 1368; GCN-NEXT: v_mul_lo_u32 v6, s2, v4 1369; GCN-NEXT: v_mul_hi_u32 v7, s2, v0 1370; GCN-NEXT: v_mul_lo_u32 v8, s3, v0 1371; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1372; GCN-NEXT: v_mul_lo_u32 v7, s2, v0 1373; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 1374; GCN-NEXT: v_mul_lo_u32 v10, v0, v6 1375; GCN-NEXT: v_mul_hi_u32 v12, v0, v6 1376; GCN-NEXT: v_mul_hi_u32 v11, v0, v7 1377; GCN-NEXT: v_mul_hi_u32 v9, v4, v7 1378; GCN-NEXT: v_mul_lo_u32 v7, v4, v7 1379; GCN-NEXT: v_mul_hi_u32 v8, v4, v6 1380; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 1381; GCN-NEXT: v_addc_u32_e32 v11, vcc, v2, v12, vcc 1382; GCN-NEXT: v_mul_lo_u32 v4, v4, v6 1383; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 1384; GCN-NEXT: v_addc_u32_e32 v7, vcc, v11, v9, vcc 1385; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v1, vcc 1386; GCN-NEXT: v_add_i32_e32 v4, vcc, v7, v4 1387; GCN-NEXT: v_addc_u32_e32 v6, vcc, v2, v6, vcc 1388; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 1389; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v6, s[0:1] 1390; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4 1391; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 1392; GCN-NEXT: v_mul_hi_u32 v5, 24, v0 1393; GCN-NEXT: v_mul_lo_u32 v4, v3, 24 1394; GCN-NEXT: v_mul_hi_u32 v6, 24, v3 1395; GCN-NEXT: v_mul_hi_u32 v0, 0, v0 1396; GCN-NEXT: v_mul_hi_u32 v3, 0, v3 1397; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 1398; GCN-NEXT: v_addc_u32_e32 v2, vcc, v2, v6, vcc 1399; GCN-NEXT: v_add_i32_e32 v4, vcc, 0, v4 1400; GCN-NEXT: v_addc_u32_e32 v0, vcc, v2, v0, vcc 1401; GCN-NEXT: v_addc_u32_e32 v1, vcc, v3, v1, vcc 1402; GCN-NEXT: v_mul_lo_u32 v1, s8, v1 1403; GCN-NEXT: v_mul_hi_u32 v2, s8, v0 1404; GCN-NEXT: v_mul_lo_u32 v3, s9, v0 1405; GCN-NEXT: v_mul_lo_u32 v0, s8, v0 1406; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1407; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1408; GCN-NEXT: v_sub_i32_e32 v2, vcc, 0, v1 1409; GCN-NEXT: v_mov_b32_e32 v3, s9 1410; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1411; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, vcc 1412; GCN-NEXT: v_subrev_i32_e64 v4, s[0:1], s8, v0 1413; GCN-NEXT: v_subbrev_u32_e64 v5, s[2:3], 0, v2, s[0:1] 1414; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s9, v5 1415; GCN-NEXT: v_subb_u32_e64 v2, s[0:1], v2, v3, s[0:1] 1416; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[2:3] 1417; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s8, v4 1418; GCN-NEXT: v_subrev_i32_e64 v3, s[0:1], s8, v4 1419; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[2:3] 1420; GCN-NEXT: v_cmp_eq_u32_e64 s[2:3], s9, v5 1421; GCN-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1422; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[2:3] 1423; GCN-NEXT: v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1] 1424; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6 1425; GCN-NEXT: v_cmp_le_u32_e32 vcc, s9, v1 1426; GCN-NEXT: v_cndmask_b32_e64 v2, v5, v2, s[0:1] 1427; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 1428; GCN-NEXT: v_cmp_le_u32_e32 vcc, s8, v0 1429; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 1430; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s9, v1 1431; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc 1432; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 1433; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc 1434; GCN-NEXT: v_cndmask_b32_e64 v2, v4, v3, s[0:1] 1435; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc 1436; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1437; GCN-NEXT: s_endpgm 1438; 1439; GCN-IR-LABEL: s_test_srem_k_num_i64: 1440; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1441; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 1442; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 1443; GCN-IR-NEXT: s_ashr_i32 s0, s7, 31 1444; GCN-IR-NEXT: s_mov_b32 s1, s0 1445; GCN-IR-NEXT: s_xor_b64 s[2:3], s[6:7], s[0:1] 1446; GCN-IR-NEXT: s_sub_u32 s2, s2, s0 1447; GCN-IR-NEXT: s_subb_u32 s3, s3, s0 1448; GCN-IR-NEXT: s_flbit_i32_b32 s0, s2 1449; GCN-IR-NEXT: s_add_i32 s0, s0, 32 1450; GCN-IR-NEXT: s_flbit_i32_b32 s1, s3 1451; GCN-IR-NEXT: v_mov_b32_e32 v0, s1 1452; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 1453; GCN-IR-NEXT: v_cmp_eq_u32_e64 vcc, s3, 0 1454; GCN-IR-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc 1455; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, 0xffffffc5, v2 1456; GCN-IR-NEXT: v_addc_u32_e64 v1, s[0:1], 0, -1, vcc 1457; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[8:9], s[2:3], 0 1458; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[0:1] 1459; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[0:1], 63, v[0:1] 1460; GCN-IR-NEXT: s_or_b64 s[8:9], s[8:9], vcc 1461; GCN-IR-NEXT: s_or_b64 s[0:1], s[8:9], s[0:1] 1462; GCN-IR-NEXT: s_mov_b32 s6, -1 1463; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 1464; GCN-IR-NEXT: s_cbranch_vccz BB10_2 1465; GCN-IR-NEXT: ; %bb.1: 1466; GCN-IR-NEXT: v_mov_b32_e32 v1, 0 1467; GCN-IR-NEXT: v_cndmask_b32_e64 v0, 24, 0, s[8:9] 1468; GCN-IR-NEXT: s_branch BB10_7 1469; GCN-IR-NEXT: BB10_2: ; %udiv-bb1 1470; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, 1, v0 1471; GCN-IR-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc 1472; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[0:1], v[3:4], v[0:1] 1473; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 63, v0 1474; GCN-IR-NEXT: v_lshl_b64 v[0:1], 24, v0 1475; GCN-IR-NEXT: s_and_b64 vcc, exec, s[0:1] 1476; GCN-IR-NEXT: s_cbranch_vccz BB10_4 1477; GCN-IR-NEXT: ; %bb.3: 1478; GCN-IR-NEXT: v_mov_b32_e32 v2, 0 1479; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1480; GCN-IR-NEXT: s_branch BB10_6 1481; GCN-IR-NEXT: BB10_4: ; %udiv-preheader 1482; GCN-IR-NEXT: s_add_u32 s7, s2, -1 1483; GCN-IR-NEXT: v_lshr_b64 v[6:7], 24, v3 1484; GCN-IR-NEXT: v_sub_i32_e32 v4, vcc, 58, v2 1485; GCN-IR-NEXT: v_mov_b32_e32 v8, 0 1486; GCN-IR-NEXT: s_addc_u32 s8, s3, -1 1487; GCN-IR-NEXT: v_subb_u32_e64 v5, s[0:1], 0, 0, vcc 1488; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1489; GCN-IR-NEXT: v_mov_b32_e32 v3, 0 1490; GCN-IR-NEXT: BB10_5: ; %udiv-do-while 1491; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1492; GCN-IR-NEXT: v_lshl_b64 v[6:7], v[6:7], 1 1493; GCN-IR-NEXT: v_lshrrev_b32_e32 v2, 31, v1 1494; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1495; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v2 1496; GCN-IR-NEXT: v_or_b32_e32 v0, v8, v0 1497; GCN-IR-NEXT: v_mov_b32_e32 v2, s8 1498; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, s7, v6 1499; GCN-IR-NEXT: v_subb_u32_e32 v2, vcc, v2, v7, vcc 1500; GCN-IR-NEXT: v_ashrrev_i32_e32 v8, 31, v2 1501; GCN-IR-NEXT: v_and_b32_e32 v10, s2, v8 1502; GCN-IR-NEXT: v_and_b32_e32 v2, 1, v8 1503; GCN-IR-NEXT: v_and_b32_e32 v11, s3, v8 1504; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 1, v4 1505; GCN-IR-NEXT: v_or_b32_e32 v1, v9, v1 1506; GCN-IR-NEXT: v_addc_u32_e32 v9, vcc, 0, v5, vcc 1507; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[8:9], v[4:5] 1508; GCN-IR-NEXT: v_mov_b32_e32 v4, v8 1509; GCN-IR-NEXT: v_sub_i32_e64 v6, s[0:1], v6, v10 1510; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1511; GCN-IR-NEXT: v_mov_b32_e32 v9, v3 1512; GCN-IR-NEXT: v_subb_u32_e64 v7, s[0:1], v7, v11, s[0:1] 1513; GCN-IR-NEXT: s_and_b64 vcc, exec, vcc 1514; GCN-IR-NEXT: v_mov_b32_e32 v8, v2 1515; GCN-IR-NEXT: s_cbranch_vccz BB10_5 1516; GCN-IR-NEXT: BB10_6: ; %udiv-loop-exit 1517; GCN-IR-NEXT: v_lshl_b64 v[0:1], v[0:1], 1 1518; GCN-IR-NEXT: v_or_b32_e32 v0, v2, v0 1519; GCN-IR-NEXT: v_or_b32_e32 v1, v3, v1 1520; GCN-IR-NEXT: BB10_7: ; %udiv-end 1521; GCN-IR-NEXT: v_mul_lo_u32 v1, s2, v1 1522; GCN-IR-NEXT: v_mul_hi_u32 v2, s2, v0 1523; GCN-IR-NEXT: v_mul_lo_u32 v3, s3, v0 1524; GCN-IR-NEXT: v_mul_lo_u32 v0, s2, v0 1525; GCN-IR-NEXT: s_mov_b32 s7, 0xf000 1526; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1527; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v1, v3 1528; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1529; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1530; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1531; GCN-IR-NEXT: s_endpgm 1532 %result = srem i64 24, %x 1533 store i64 %result, i64 addrspace(1)* %out 1534 ret void 1535} 1536 1537define i64 @v_test_srem_k_num_i64(i64 %x) { 1538; GCN-LABEL: v_test_srem_k_num_i64: 1539; GCN: ; %bb.0: 1540; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1541; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1542; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1543; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 1544; GCN-NEXT: v_xor_b32_e32 v1, v1, v2 1545; GCN-NEXT: v_xor_b32_e32 v0, v0, v2 1546; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 1547; GCN-NEXT: v_cvt_f32_u32_e32 v3, v1 1548; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v0 1549; GCN-NEXT: v_subb_u32_e32 v5, vcc, 0, v1, vcc 1550; GCN-NEXT: v_mov_b32_e32 v12, 0 1551; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 1552; GCN-NEXT: v_rcp_f32_e32 v2, v2 1553; GCN-NEXT: v_mov_b32_e32 v11, 0 1554; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 1555; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 1556; GCN-NEXT: v_trunc_f32_e32 v3, v3 1557; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 1558; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 1559; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 1560; GCN-NEXT: v_mul_hi_u32 v6, v4, v2 1561; GCN-NEXT: v_mul_lo_u32 v7, v4, v3 1562; GCN-NEXT: v_mul_lo_u32 v8, v5, v2 1563; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 1564; GCN-NEXT: v_mul_lo_u32 v7, v4, v2 1565; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 1566; GCN-NEXT: v_mul_lo_u32 v9, v2, v6 1567; GCN-NEXT: v_mul_hi_u32 v8, v2, v6 1568; GCN-NEXT: v_mul_hi_u32 v10, v2, v7 1569; GCN-NEXT: v_mul_hi_u32 v13, v3, v6 1570; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 1571; GCN-NEXT: v_add_i32_e32 v9, vcc, v10, v9 1572; GCN-NEXT: v_mul_lo_u32 v10, v3, v7 1573; GCN-NEXT: v_mul_hi_u32 v7, v3, v7 1574; GCN-NEXT: v_addc_u32_e32 v8, vcc, v12, v8, vcc 1575; GCN-NEXT: v_add_i32_e32 v9, vcc, v9, v10 1576; GCN-NEXT: v_addc_u32_e32 v7, vcc, v8, v7, vcc 1577; GCN-NEXT: v_addc_u32_e32 v8, vcc, v13, v11, vcc 1578; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1579; GCN-NEXT: v_add_i32_e64 v2, s[4:5], v2, v6 1580; GCN-NEXT: v_addc_u32_e32 v7, vcc, v12, v8, vcc 1581; GCN-NEXT: v_addc_u32_e64 v6, vcc, v3, v7, s[4:5] 1582; GCN-NEXT: v_mul_lo_u32 v8, v4, v6 1583; GCN-NEXT: v_mul_hi_u32 v9, v4, v2 1584; GCN-NEXT: v_mul_lo_u32 v5, v5, v2 1585; GCN-NEXT: v_mul_lo_u32 v4, v4, v2 1586; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1587; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 1588; GCN-NEXT: v_mul_lo_u32 v10, v2, v5 1589; GCN-NEXT: v_mul_hi_u32 v13, v2, v4 1590; GCN-NEXT: v_mul_hi_u32 v14, v2, v5 1591; GCN-NEXT: v_mul_hi_u32 v9, v6, v4 1592; GCN-NEXT: v_mul_lo_u32 v4, v6, v4 1593; GCN-NEXT: v_add_i32_e32 v10, vcc, v13, v10 1594; GCN-NEXT: v_mul_hi_u32 v8, v6, v5 1595; GCN-NEXT: v_addc_u32_e32 v13, vcc, v12, v14, vcc 1596; GCN-NEXT: v_mul_lo_u32 v5, v6, v5 1597; GCN-NEXT: v_add_i32_e32 v4, vcc, v10, v4 1598; GCN-NEXT: v_addc_u32_e32 v4, vcc, v13, v9, vcc 1599; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v11, vcc 1600; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 1601; GCN-NEXT: v_addc_u32_e32 v5, vcc, v12, v6, vcc 1602; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v7 1603; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v5, s[4:5] 1604; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 1605; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 1606; GCN-NEXT: v_mul_hi_u32 v5, 24, v2 1607; GCN-NEXT: v_mul_lo_u32 v4, v3, 24 1608; GCN-NEXT: v_mul_hi_u32 v6, 24, v3 1609; GCN-NEXT: v_mul_hi_u32 v2, 0, v2 1610; GCN-NEXT: v_mul_hi_u32 v3, 0, v3 1611; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4 1612; GCN-NEXT: v_addc_u32_e32 v5, vcc, v12, v6, vcc 1613; GCN-NEXT: v_add_i32_e32 v4, vcc, 0, v4 1614; GCN-NEXT: v_addc_u32_e32 v2, vcc, v5, v2, vcc 1615; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v11, vcc 1616; GCN-NEXT: v_mul_lo_u32 v3, v0, v3 1617; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 1618; GCN-NEXT: v_mul_lo_u32 v5, v1, v2 1619; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 1620; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1621; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 1622; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 1623; GCN-NEXT: v_sub_i32_e32 v2, vcc, 24, v2 1624; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 1625; GCN-NEXT: v_sub_i32_e64 v5, s[4:5], v2, v0 1626; GCN-NEXT: v_subbrev_u32_e64 v6, s[6:7], 0, v4, s[4:5] 1627; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v6, v1 1628; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[6:7] 1629; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v5, v0 1630; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[6:7] 1631; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v6, v1 1632; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, s[4:5] 1633; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 1634; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[6:7] 1635; GCN-NEXT: v_sub_i32_e64 v8, s[4:5], v5, v0 1636; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 1637; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v1 1638; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v7 1639; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, vcc 1640; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 1641; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc 1642; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1 1643; GCN-NEXT: v_cndmask_b32_e32 v0, v7, v0, vcc 1644; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v8, s[4:5] 1645; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 1646; GCN-NEXT: v_cndmask_b32_e64 v1, v6, v4, s[4:5] 1647; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc 1648; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc 1649; GCN-NEXT: s_setpc_b64 s[30:31] 1650; 1651; GCN-IR-LABEL: v_test_srem_k_num_i64: 1652; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1653; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1654; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1655; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1656; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 1657; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1658; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1659; GCN-IR-NEXT: v_ffbh_u32_e32 v2, v0 1660; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, 32, v2 1661; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v1 1662; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1 1663; GCN-IR-NEXT: v_cndmask_b32_e32 v8, v3, v2, vcc 1664; GCN-IR-NEXT: s_movk_i32 s6, 0xffc5 1665; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, s6, v8 1666; GCN-IR-NEXT: v_addc_u32_e64 v4, s[6:7], 0, -1, vcc 1667; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 1668; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[3:4] 1669; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1670; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], vcc 1671; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[3:4] 1672; GCN-IR-NEXT: v_cndmask_b32_e64 v2, 24, 0, s[4:5] 1673; GCN-IR-NEXT: s_xor_b64 s[4:5], s[4:5], -1 1674; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1675; GCN-IR-NEXT: s_and_b64 s[4:5], s[4:5], vcc 1676; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1677; GCN-IR-NEXT: s_cbranch_execz BB11_6 1678; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1679; GCN-IR-NEXT: v_add_i32_e32 v5, vcc, 1, v3 1680; GCN-IR-NEXT: v_addc_u32_e32 v6, vcc, 0, v4, vcc 1681; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 63, v3 1682; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[5:6], v[3:4] 1683; GCN-IR-NEXT: v_mov_b32_e32 v6, 0 1684; GCN-IR-NEXT: v_lshl_b64 v[2:3], 24, v2 1685; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 1686; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1687; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1688; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 1689; GCN-IR-NEXT: s_cbranch_execz BB11_5 1690; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1691; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, -1, v0 1692; GCN-IR-NEXT: v_lshr_b64 v[10:11], 24, v5 1693; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc 1694; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, 58, v8 1695; GCN-IR-NEXT: v_mov_b32_e32 v12, 0 1696; GCN-IR-NEXT: v_subb_u32_e32 v9, vcc, 0, v9, vcc 1697; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 1698; GCN-IR-NEXT: BB11_3: ; %udiv-do-while 1699; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1700; GCN-IR-NEXT: v_lshl_b64 v[10:11], v[10:11], 1 1701; GCN-IR-NEXT: v_lshrrev_b32_e32 v6, 31, v3 1702; GCN-IR-NEXT: v_or_b32_e32 v10, v10, v6 1703; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1704; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, v4, v10 1705; GCN-IR-NEXT: v_subb_u32_e32 v6, vcc, v5, v11, vcc 1706; GCN-IR-NEXT: v_or_b32_e32 v2, v12, v2 1707; GCN-IR-NEXT: v_ashrrev_i32_e32 v12, 31, v6 1708; GCN-IR-NEXT: v_and_b32_e32 v15, v12, v0 1709; GCN-IR-NEXT: v_and_b32_e32 v6, 1, v12 1710; GCN-IR-NEXT: v_and_b32_e32 v14, v12, v1 1711; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, 1, v8 1712; GCN-IR-NEXT: v_or_b32_e32 v3, v13, v3 1713; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, 0, v9, vcc 1714; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[12:13], v[8:9] 1715; GCN-IR-NEXT: v_mov_b32_e32 v8, v12 1716; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1717; GCN-IR-NEXT: v_sub_i32_e64 v10, s[4:5], v10, v15 1718; GCN-IR-NEXT: v_mov_b32_e32 v9, v13 1719; GCN-IR-NEXT: v_mov_b32_e32 v13, v7 1720; GCN-IR-NEXT: v_subb_u32_e64 v11, s[4:5], v11, v14, s[4:5] 1721; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 1722; GCN-IR-NEXT: v_mov_b32_e32 v12, v6 1723; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 1724; GCN-IR-NEXT: s_cbranch_execnz BB11_3 1725; GCN-IR-NEXT: ; %bb.4: ; %Flow 1726; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1727; GCN-IR-NEXT: BB11_5: ; %Flow3 1728; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1729; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1730; GCN-IR-NEXT: v_or_b32_e32 v5, v7, v3 1731; GCN-IR-NEXT: v_or_b32_e32 v2, v6, v2 1732; GCN-IR-NEXT: BB11_6: ; %Flow4 1733; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1734; GCN-IR-NEXT: v_mul_lo_u32 v3, v0, v5 1735; GCN-IR-NEXT: v_mul_hi_u32 v4, v0, v2 1736; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v2 1737; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, v2 1738; GCN-IR-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1739; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v3, v1 1740; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1741; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1742; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1743 %result = srem i64 24, %x 1744 ret i64 %result 1745} 1746 1747define i64 @v_test_srem_pow2_k_num_i64(i64 %x) { 1748; GCN-LABEL: v_test_srem_pow2_k_num_i64: 1749; GCN: ; %bb.0: 1750; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1751; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1752; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 1753; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 1754; GCN-NEXT: v_xor_b32_e32 v1, v1, v2 1755; GCN-NEXT: v_xor_b32_e32 v0, v0, v2 1756; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 1757; GCN-NEXT: v_cvt_f32_u32_e32 v3, v1 1758; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v0 1759; GCN-NEXT: v_subb_u32_e32 v5, vcc, 0, v1, vcc 1760; GCN-NEXT: v_mov_b32_e32 v12, 0 1761; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 1762; GCN-NEXT: v_rcp_f32_e32 v2, v2 1763; GCN-NEXT: v_mov_b32_e32 v11, 0 1764; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 1765; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 1766; GCN-NEXT: v_trunc_f32_e32 v3, v3 1767; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 1768; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 1769; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 1770; GCN-NEXT: v_mul_hi_u32 v6, v4, v2 1771; GCN-NEXT: v_mul_lo_u32 v7, v4, v3 1772; GCN-NEXT: v_mul_lo_u32 v8, v5, v2 1773; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 1774; GCN-NEXT: v_mul_lo_u32 v7, v4, v2 1775; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 1776; GCN-NEXT: v_mul_lo_u32 v9, v2, v6 1777; GCN-NEXT: v_mul_hi_u32 v8, v2, v6 1778; GCN-NEXT: v_mul_hi_u32 v10, v2, v7 1779; GCN-NEXT: v_mul_hi_u32 v13, v3, v6 1780; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 1781; GCN-NEXT: v_add_i32_e32 v9, vcc, v10, v9 1782; GCN-NEXT: v_mul_lo_u32 v10, v3, v7 1783; GCN-NEXT: v_mul_hi_u32 v7, v3, v7 1784; GCN-NEXT: v_addc_u32_e32 v8, vcc, v12, v8, vcc 1785; GCN-NEXT: v_add_i32_e32 v9, vcc, v9, v10 1786; GCN-NEXT: v_addc_u32_e32 v7, vcc, v8, v7, vcc 1787; GCN-NEXT: v_addc_u32_e32 v8, vcc, v13, v11, vcc 1788; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 1789; GCN-NEXT: v_add_i32_e64 v2, s[4:5], v2, v6 1790; GCN-NEXT: v_addc_u32_e32 v7, vcc, v12, v8, vcc 1791; GCN-NEXT: v_addc_u32_e64 v6, vcc, v3, v7, s[4:5] 1792; GCN-NEXT: v_mul_lo_u32 v8, v4, v6 1793; GCN-NEXT: v_mul_hi_u32 v9, v4, v2 1794; GCN-NEXT: v_mul_lo_u32 v5, v5, v2 1795; GCN-NEXT: v_mul_lo_u32 v4, v4, v2 1796; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1797; GCN-NEXT: v_add_i32_e32 v5, vcc, v8, v5 1798; GCN-NEXT: v_mul_lo_u32 v10, v2, v5 1799; GCN-NEXT: v_mul_hi_u32 v13, v2, v4 1800; GCN-NEXT: v_mul_hi_u32 v14, v2, v5 1801; GCN-NEXT: v_mul_hi_u32 v9, v6, v4 1802; GCN-NEXT: v_mul_lo_u32 v4, v6, v4 1803; GCN-NEXT: v_add_i32_e32 v10, vcc, v13, v10 1804; GCN-NEXT: v_mul_hi_u32 v8, v6, v5 1805; GCN-NEXT: v_addc_u32_e32 v13, vcc, v12, v14, vcc 1806; GCN-NEXT: v_mul_lo_u32 v5, v6, v5 1807; GCN-NEXT: v_add_i32_e32 v4, vcc, v10, v4 1808; GCN-NEXT: v_addc_u32_e32 v4, vcc, v13, v9, vcc 1809; GCN-NEXT: v_addc_u32_e32 v6, vcc, v8, v11, vcc 1810; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v5 1811; GCN-NEXT: v_addc_u32_e32 v5, vcc, v12, v6, vcc 1812; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v7 1813; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v5, s[4:5] 1814; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 1815; GCN-NEXT: s_mov_b32 s4, 0x8000 1816; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 1817; GCN-NEXT: v_mul_hi_u32 v4, s4, v2 1818; GCN-NEXT: v_mul_hi_u32 v5, s4, v3 1819; GCN-NEXT: v_lshlrev_b32_e32 v6, 15, v3 1820; GCN-NEXT: v_mul_hi_u32 v2, 0, v2 1821; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v6 1822; GCN-NEXT: v_mul_hi_u32 v3, 0, v3 1823; GCN-NEXT: v_addc_u32_e32 v5, vcc, v12, v5, vcc 1824; GCN-NEXT: v_add_i32_e32 v4, vcc, 0, v4 1825; GCN-NEXT: v_addc_u32_e32 v2, vcc, v5, v2, vcc 1826; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v11, vcc 1827; GCN-NEXT: v_mul_lo_u32 v3, v0, v3 1828; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 1829; GCN-NEXT: v_mul_lo_u32 v5, v1, v2 1830; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 1831; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1832; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 1833; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 1834; GCN-NEXT: v_sub_i32_e32 v2, vcc, s4, v2 1835; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 1836; GCN-NEXT: v_sub_i32_e64 v5, s[4:5], v2, v0 1837; GCN-NEXT: v_subbrev_u32_e64 v6, s[6:7], 0, v4, s[4:5] 1838; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v6, v1 1839; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[6:7] 1840; GCN-NEXT: v_cmp_ge_u32_e64 s[6:7], v5, v0 1841; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[6:7] 1842; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], v6, v1 1843; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, s[4:5] 1844; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 1845; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[6:7] 1846; GCN-NEXT: v_sub_i32_e64 v8, s[4:5], v5, v0 1847; GCN-NEXT: v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5] 1848; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v1 1849; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v7 1850; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, vcc 1851; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 1852; GCN-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc 1853; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1 1854; GCN-NEXT: v_cndmask_b32_e32 v0, v7, v0, vcc 1855; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v8, s[4:5] 1856; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 1857; GCN-NEXT: v_cndmask_b32_e64 v1, v6, v4, s[4:5] 1858; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc 1859; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc 1860; GCN-NEXT: s_setpc_b64 s[30:31] 1861; 1862; GCN-IR-LABEL: v_test_srem_pow2_k_num_i64: 1863; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1864; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1865; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1866; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1867; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 1868; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1869; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1870; GCN-IR-NEXT: v_ffbh_u32_e32 v2, v0 1871; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, 32, v2 1872; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v1 1873; GCN-IR-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1 1874; GCN-IR-NEXT: v_cndmask_b32_e32 v8, v3, v2, vcc 1875; GCN-IR-NEXT: s_movk_i32 s6, 0xffd0 1876; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, s6, v8 1877; GCN-IR-NEXT: v_addc_u32_e64 v3, s[6:7], 0, -1, vcc 1878; GCN-IR-NEXT: v_cmp_eq_u64_e64 s[4:5], 0, v[0:1] 1879; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, 63, v[2:3] 1880; GCN-IR-NEXT: s_mov_b32 s8, 0x8000 1881; GCN-IR-NEXT: s_or_b64 s[4:5], s[4:5], vcc 1882; GCN-IR-NEXT: v_mov_b32_e32 v4, s8 1883; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[2:3] 1884; GCN-IR-NEXT: v_mov_b32_e32 v9, 0 1885; GCN-IR-NEXT: v_cndmask_b32_e64 v4, v4, 0, s[4:5] 1886; GCN-IR-NEXT: s_xor_b64 s[4:5], s[4:5], -1 1887; GCN-IR-NEXT: s_mov_b32 s9, 0 1888; GCN-IR-NEXT: v_mov_b32_e32 v5, v9 1889; GCN-IR-NEXT: s_and_b64 s[4:5], s[4:5], vcc 1890; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 1891; GCN-IR-NEXT: s_cbranch_execz BB12_6 1892; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 1893; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, 1, v2 1894; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, 0, v3, vcc 1895; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[4:5], v[2:3] 1896; GCN-IR-NEXT: v_sub_i32_e64 v2, s[4:5], 63, v2 1897; GCN-IR-NEXT: v_lshl_b64 v[2:3], s[8:9], v2 1898; GCN-IR-NEXT: v_mov_b32_e32 v6, 0 1899; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 1900; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1901; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 1902; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 1903; GCN-IR-NEXT: s_cbranch_execz BB12_5 1904; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 1905; GCN-IR-NEXT: s_mov_b32 s5, 0 1906; GCN-IR-NEXT: s_mov_b32 s4, 0x8000 1907; GCN-IR-NEXT: v_lshr_b64 v[10:11], s[4:5], v4 1908; GCN-IR-NEXT: v_add_i32_e32 v4, vcc, -1, v0 1909; GCN-IR-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc 1910; GCN-IR-NEXT: v_sub_i32_e32 v8, vcc, 47, v8 1911; GCN-IR-NEXT: v_mov_b32_e32 v12, 0 1912; GCN-IR-NEXT: v_subb_u32_e32 v9, vcc, 0, v9, vcc 1913; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 1914; GCN-IR-NEXT: BB12_3: ; %udiv-do-while 1915; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 1916; GCN-IR-NEXT: v_lshl_b64 v[10:11], v[10:11], 1 1917; GCN-IR-NEXT: v_lshrrev_b32_e32 v6, 31, v3 1918; GCN-IR-NEXT: v_or_b32_e32 v10, v10, v6 1919; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1920; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, v4, v10 1921; GCN-IR-NEXT: v_subb_u32_e32 v6, vcc, v5, v11, vcc 1922; GCN-IR-NEXT: v_or_b32_e32 v2, v12, v2 1923; GCN-IR-NEXT: v_ashrrev_i32_e32 v12, 31, v6 1924; GCN-IR-NEXT: v_and_b32_e32 v15, v12, v0 1925; GCN-IR-NEXT: v_and_b32_e32 v6, 1, v12 1926; GCN-IR-NEXT: v_and_b32_e32 v14, v12, v1 1927; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, 1, v8 1928; GCN-IR-NEXT: v_or_b32_e32 v3, v13, v3 1929; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, 0, v9, vcc 1930; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[12:13], v[8:9] 1931; GCN-IR-NEXT: v_mov_b32_e32 v8, v12 1932; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 1933; GCN-IR-NEXT: v_sub_i32_e64 v10, s[4:5], v10, v15 1934; GCN-IR-NEXT: v_mov_b32_e32 v9, v13 1935; GCN-IR-NEXT: v_mov_b32_e32 v13, v7 1936; GCN-IR-NEXT: v_subb_u32_e64 v11, s[4:5], v11, v14, s[4:5] 1937; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 1938; GCN-IR-NEXT: v_mov_b32_e32 v12, v6 1939; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 1940; GCN-IR-NEXT: s_cbranch_execnz BB12_3 1941; GCN-IR-NEXT: ; %bb.4: ; %Flow 1942; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 1943; GCN-IR-NEXT: BB12_5: ; %Flow3 1944; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 1945; GCN-IR-NEXT: v_lshl_b64 v[2:3], v[2:3], 1 1946; GCN-IR-NEXT: v_or_b32_e32 v5, v7, v3 1947; GCN-IR-NEXT: v_or_b32_e32 v4, v6, v2 1948; GCN-IR-NEXT: BB12_6: ; %Flow4 1949; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 1950; GCN-IR-NEXT: v_mul_lo_u32 v2, v0, v5 1951; GCN-IR-NEXT: v_mul_hi_u32 v3, v0, v4 1952; GCN-IR-NEXT: v_mul_lo_u32 v1, v1, v4 1953; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, v4 1954; GCN-IR-NEXT: v_add_i32_e32 v2, vcc, v3, v2 1955; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1956; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 1957; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, 0, v1, vcc 1958; GCN-IR-NEXT: s_setpc_b64 s[30:31] 1959 %result = srem i64 32768, %x 1960 ret i64 %result 1961} 1962 1963define i64 @v_test_srem_pow2_k_den_i64(i64 %x) { 1964; GCN-LABEL: v_test_srem_pow2_k_den_i64: 1965; GCN: ; %bb.0: 1966; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1967; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1968; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v2 1969; GCN-NEXT: v_add_i32_e32 v2, vcc, v0, v2 1970; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 1971; GCN-NEXT: v_and_b32_e32 v2, 0xffff8000, v2 1972; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1973; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 1974; GCN-NEXT: s_setpc_b64 s[30:31] 1975; 1976; GCN-IR-LABEL: v_test_srem_pow2_k_den_i64: 1977; GCN-IR: ; %bb.0: ; %_udiv-special-cases 1978; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1979; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1980; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 1981; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1982; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v2 1983; GCN-IR-NEXT: v_ffbh_u32_e32 v3, v0 1984; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc 1985; GCN-IR-NEXT: v_add_i32_e64 v3, s[4:5], 32, v3 1986; GCN-IR-NEXT: v_ffbh_u32_e32 v4, v1 1987; GCN-IR-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v1 1988; GCN-IR-NEXT: v_cndmask_b32_e64 v8, v4, v3, s[4:5] 1989; GCN-IR-NEXT: v_sub_i32_e64 v4, s[4:5], 48, v8 1990; GCN-IR-NEXT: v_subb_u32_e64 v5, s[4:5], 0, 0, s[4:5] 1991; GCN-IR-NEXT: v_cmp_eq_u64_e32 vcc, 0, v[0:1] 1992; GCN-IR-NEXT: v_cmp_lt_u64_e64 s[4:5], 63, v[4:5] 1993; GCN-IR-NEXT: v_mov_b32_e32 v3, v2 1994; GCN-IR-NEXT: s_or_b64 s[4:5], vcc, s[4:5] 1995; GCN-IR-NEXT: v_cmp_ne_u64_e32 vcc, 63, v[4:5] 1996; GCN-IR-NEXT: s_xor_b64 s[6:7], s[4:5], -1 1997; GCN-IR-NEXT: v_cndmask_b32_e64 v7, v1, 0, s[4:5] 1998; GCN-IR-NEXT: v_cndmask_b32_e64 v6, v0, 0, s[4:5] 1999; GCN-IR-NEXT: s_and_b64 s[4:5], s[6:7], vcc 2000; GCN-IR-NEXT: s_and_saveexec_b64 s[6:7], s[4:5] 2001; GCN-IR-NEXT: s_cbranch_execz BB13_6 2002; GCN-IR-NEXT: ; %bb.1: ; %udiv-bb1 2003; GCN-IR-NEXT: v_add_i32_e32 v9, vcc, 1, v4 2004; GCN-IR-NEXT: v_addc_u32_e32 v10, vcc, 0, v5, vcc 2005; GCN-IR-NEXT: v_cmp_ge_u64_e32 vcc, v[9:10], v[4:5] 2006; GCN-IR-NEXT: v_sub_i32_e64 v4, s[4:5], 63, v4 2007; GCN-IR-NEXT: v_mov_b32_e32 v6, 0 2008; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[0:1], v4 2009; GCN-IR-NEXT: s_mov_b64 s[8:9], 0 2010; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 2011; GCN-IR-NEXT: s_and_saveexec_b64 s[4:5], vcc 2012; GCN-IR-NEXT: s_xor_b64 s[10:11], exec, s[4:5] 2013; GCN-IR-NEXT: s_cbranch_execz BB13_5 2014; GCN-IR-NEXT: ; %bb.2: ; %udiv-preheader 2015; GCN-IR-NEXT: v_mov_b32_e32 v12, 0 2016; GCN-IR-NEXT: v_lshr_b64 v[10:11], v[0:1], v9 2017; GCN-IR-NEXT: v_add_i32_e32 v8, vcc, 0xffffffcf, v8 2018; GCN-IR-NEXT: v_addc_u32_e64 v9, s[4:5], 0, -1, vcc 2019; GCN-IR-NEXT: v_mov_b32_e32 v13, 0 2020; GCN-IR-NEXT: s_movk_i32 s12, 0x7fff 2021; GCN-IR-NEXT: BB13_3: ; %udiv-do-while 2022; GCN-IR-NEXT: ; =>This Inner Loop Header: Depth=1 2023; GCN-IR-NEXT: v_lshl_b64 v[10:11], v[10:11], 1 2024; GCN-IR-NEXT: v_lshrrev_b32_e32 v6, 31, v5 2025; GCN-IR-NEXT: v_or_b32_e32 v10, v10, v6 2026; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[4:5], 1 2027; GCN-IR-NEXT: v_sub_i32_e32 v6, vcc, s12, v10 2028; GCN-IR-NEXT: v_subb_u32_e32 v6, vcc, 0, v11, vcc 2029; GCN-IR-NEXT: v_or_b32_e32 v4, v12, v4 2030; GCN-IR-NEXT: v_ashrrev_i32_e32 v12, 31, v6 2031; GCN-IR-NEXT: v_and_b32_e32 v14, 0x8000, v12 2032; GCN-IR-NEXT: v_and_b32_e32 v6, 1, v12 2033; GCN-IR-NEXT: v_add_i32_e32 v12, vcc, 1, v8 2034; GCN-IR-NEXT: v_or_b32_e32 v5, v13, v5 2035; GCN-IR-NEXT: v_addc_u32_e32 v13, vcc, 0, v9, vcc 2036; GCN-IR-NEXT: v_cmp_lt_u64_e32 vcc, v[12:13], v[8:9] 2037; GCN-IR-NEXT: v_mov_b32_e32 v8, v12 2038; GCN-IR-NEXT: v_mov_b32_e32 v7, 0 2039; GCN-IR-NEXT: v_mov_b32_e32 v9, v13 2040; GCN-IR-NEXT: v_mov_b32_e32 v13, v7 2041; GCN-IR-NEXT: v_mov_b32_e32 v15, 0 2042; GCN-IR-NEXT: v_sub_i32_e64 v10, s[4:5], v10, v14 2043; GCN-IR-NEXT: v_subb_u32_e64 v11, s[4:5], v11, v15, s[4:5] 2044; GCN-IR-NEXT: s_or_b64 s[8:9], vcc, s[8:9] 2045; GCN-IR-NEXT: v_mov_b32_e32 v12, v6 2046; GCN-IR-NEXT: s_andn2_b64 exec, exec, s[8:9] 2047; GCN-IR-NEXT: s_cbranch_execnz BB13_3 2048; GCN-IR-NEXT: ; %bb.4: ; %Flow 2049; GCN-IR-NEXT: s_or_b64 exec, exec, s[8:9] 2050; GCN-IR-NEXT: BB13_5: ; %Flow3 2051; GCN-IR-NEXT: s_or_b64 exec, exec, s[10:11] 2052; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[4:5], 1 2053; GCN-IR-NEXT: v_or_b32_e32 v7, v7, v5 2054; GCN-IR-NEXT: v_or_b32_e32 v6, v6, v4 2055; GCN-IR-NEXT: BB13_6: ; %Flow4 2056; GCN-IR-NEXT: s_or_b64 exec, exec, s[6:7] 2057; GCN-IR-NEXT: v_lshl_b64 v[4:5], v[6:7], 15 2058; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 2059; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 2060; GCN-IR-NEXT: v_xor_b32_e32 v0, v0, v2 2061; GCN-IR-NEXT: v_xor_b32_e32 v1, v1, v3 2062; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 2063; GCN-IR-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 2064; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2065 %result = srem i64 %x, 32768 2066 ret i64 %result 2067} 2068 2069define amdgpu_kernel void @s_test_srem24_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 2070; GCN-LABEL: s_test_srem24_k_num_i64: 2071; GCN: ; %bb.0: 2072; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 2073; GCN-NEXT: s_mov_b32 s6, 0x41c00000 2074; GCN-NEXT: s_waitcnt lgkmcnt(0) 2075; GCN-NEXT: s_ashr_i64 s[4:5], s[2:3], 40 2076; GCN-NEXT: v_cvt_f32_i32_e32 v0, s4 2077; GCN-NEXT: s_ashr_i32 s5, s4, 30 2078; GCN-NEXT: s_or_b32 s5, s5, 1 2079; GCN-NEXT: v_mov_b32_e32 v3, s5 2080; GCN-NEXT: v_rcp_iflag_f32_e32 v1, v0 2081; GCN-NEXT: s_mov_b32 s3, 0xf000 2082; GCN-NEXT: s_mov_b32 s2, -1 2083; GCN-NEXT: v_mul_f32_e32 v1, s6, v1 2084; GCN-NEXT: v_trunc_f32_e32 v1, v1 2085; GCN-NEXT: v_mad_f32 v2, -v1, v0, s6 2086; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1 2087; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, |v0| 2088; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 2089; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 2090; GCN-NEXT: v_mul_lo_u32 v0, v0, s4 2091; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2092; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2093; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2094; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 2095; GCN-NEXT: s_endpgm 2096; 2097; GCN-IR-LABEL: s_test_srem24_k_num_i64: 2098; GCN-IR: ; %bb.0: 2099; GCN-IR-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 2100; GCN-IR-NEXT: s_mov_b32 s6, 0x41c00000 2101; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 2102; GCN-IR-NEXT: s_ashr_i64 s[4:5], s[2:3], 40 2103; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s4 2104; GCN-IR-NEXT: s_ashr_i32 s5, s4, 30 2105; GCN-IR-NEXT: s_or_b32 s5, s5, 1 2106; GCN-IR-NEXT: v_mov_b32_e32 v3, s5 2107; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v1, v0 2108; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 2109; GCN-IR-NEXT: s_mov_b32 s2, -1 2110; GCN-IR-NEXT: v_mul_f32_e32 v1, s6, v1 2111; GCN-IR-NEXT: v_trunc_f32_e32 v1, v1 2112; GCN-IR-NEXT: v_mad_f32 v2, -v1, v0, s6 2113; GCN-IR-NEXT: v_cvt_i32_f32_e32 v1, v1 2114; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, |v0| 2115; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 2116; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v1 2117; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s4 2118; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2119; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2120; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2121; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 2122; GCN-IR-NEXT: s_endpgm 2123 %x.shr = ashr i64 %x, 40 2124 %result = srem i64 24, %x.shr 2125 store i64 %result, i64 addrspace(1)* %out 2126 ret void 2127} 2128 2129define amdgpu_kernel void @s_test_srem24_k_den_i64(i64 addrspace(1)* %out, i64 %x) { 2130; GCN-LABEL: s_test_srem24_k_den_i64: 2131; GCN: ; %bb.0: 2132; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 2133; GCN-NEXT: s_mov_b32 s8, 0x46b6fe00 2134; GCN-NEXT: s_mov_b32 s3, 0xf000 2135; GCN-NEXT: s_mov_b32 s2, -1 2136; GCN-NEXT: s_waitcnt lgkmcnt(0) 2137; GCN-NEXT: s_ashr_i64 s[6:7], s[6:7], 40 2138; GCN-NEXT: v_cvt_f32_i32_e32 v0, s6 2139; GCN-NEXT: s_ashr_i32 s0, s6, 30 2140; GCN-NEXT: s_or_b32 s0, s0, 1 2141; GCN-NEXT: v_mov_b32_e32 v1, s0 2142; GCN-NEXT: v_mul_f32_e32 v2, 0x38331158, v0 2143; GCN-NEXT: v_trunc_f32_e32 v2, v2 2144; GCN-NEXT: v_mad_f32 v0, -v2, s8, v0 2145; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 2146; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s8 2147; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc 2148; GCN-NEXT: s_movk_i32 s0, 0x5b7f 2149; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 2150; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 2151; GCN-NEXT: s_mov_b32 s0, s4 2152; GCN-NEXT: s_mov_b32 s1, s5 2153; GCN-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 2154; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2155; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2156; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 2157; GCN-NEXT: s_endpgm 2158; 2159; GCN-IR-LABEL: s_test_srem24_k_den_i64: 2160; GCN-IR: ; %bb.0: 2161; GCN-IR-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 2162; GCN-IR-NEXT: s_mov_b32 s8, 0x46b6fe00 2163; GCN-IR-NEXT: s_mov_b32 s3, 0xf000 2164; GCN-IR-NEXT: s_mov_b32 s2, -1 2165; GCN-IR-NEXT: s_waitcnt lgkmcnt(0) 2166; GCN-IR-NEXT: s_ashr_i64 s[6:7], s[6:7], 40 2167; GCN-IR-NEXT: v_cvt_f32_i32_e32 v0, s6 2168; GCN-IR-NEXT: s_ashr_i32 s0, s6, 30 2169; GCN-IR-NEXT: s_or_b32 s0, s0, 1 2170; GCN-IR-NEXT: v_mov_b32_e32 v1, s0 2171; GCN-IR-NEXT: v_mul_f32_e32 v2, 0x38331158, v0 2172; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 2173; GCN-IR-NEXT: v_mad_f32 v0, -v2, s8, v0 2174; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 2175; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s8 2176; GCN-IR-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc 2177; GCN-IR-NEXT: s_movk_i32 s0, 0x5b7f 2178; GCN-IR-NEXT: v_add_i32_e32 v0, vcc, v0, v2 2179; GCN-IR-NEXT: v_mul_lo_u32 v0, v0, s0 2180; GCN-IR-NEXT: s_mov_b32 s0, s4 2181; GCN-IR-NEXT: s_mov_b32 s1, s5 2182; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, s6, v0 2183; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2184; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2185; GCN-IR-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 2186; GCN-IR-NEXT: s_endpgm 2187 %x.shr = ashr i64 %x, 40 2188 %result = srem i64 %x.shr, 23423 2189 store i64 %result, i64 addrspace(1)* %out 2190 ret void 2191} 2192 2193define i64 @v_test_srem24_k_num_i64(i64 %x) { 2194; GCN-LABEL: v_test_srem24_k_num_i64: 2195; GCN: ; %bb.0: 2196; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2197; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2198; GCN-NEXT: s_mov_b32 s4, 0x41c00000 2199; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 2200; GCN-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2201; GCN-NEXT: v_or_b32_e32 v3, 1, v3 2202; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v1 2203; GCN-NEXT: v_mul_f32_e32 v2, s4, v2 2204; GCN-NEXT: v_trunc_f32_e32 v2, v2 2205; GCN-NEXT: v_mad_f32 v4, -v2, v1, s4 2206; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 2207; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2208; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2209; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2210; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 2211; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2212; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2213; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2214; GCN-NEXT: s_setpc_b64 s[30:31] 2215; 2216; GCN-IR-LABEL: v_test_srem24_k_num_i64: 2217; GCN-IR: ; %bb.0: 2218; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2219; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2220; GCN-IR-NEXT: s_mov_b32 s4, 0x41c00000 2221; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 2222; GCN-IR-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2223; GCN-IR-NEXT: v_or_b32_e32 v3, 1, v3 2224; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v1 2225; GCN-IR-NEXT: v_mul_f32_e32 v2, s4, v2 2226; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 2227; GCN-IR-NEXT: v_mad_f32 v4, -v2, v1, s4 2228; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 2229; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2230; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2231; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2232; GCN-IR-NEXT: v_mul_lo_u32 v0, v1, v0 2233; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 2234; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2235; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2236; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2237 %x.shr = ashr i64 %x, 40 2238 %result = srem i64 24, %x.shr 2239 ret i64 %result 2240} 2241 2242define i64 @v_test_srem24_pow2_k_num_i64(i64 %x) { 2243; GCN-LABEL: v_test_srem24_pow2_k_num_i64: 2244; GCN: ; %bb.0: 2245; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2246; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2247; GCN-NEXT: s_mov_b32 s4, 0x47000000 2248; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 2249; GCN-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2250; GCN-NEXT: v_or_b32_e32 v3, 1, v3 2251; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v1 2252; GCN-NEXT: v_mul_f32_e32 v2, s4, v2 2253; GCN-NEXT: v_trunc_f32_e32 v2, v2 2254; GCN-NEXT: v_mad_f32 v4, -v2, v1, s4 2255; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2 2256; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2257; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2258; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2259; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 2260; GCN-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 2261; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 2262; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2263; GCN-NEXT: s_setpc_b64 s[30:31] 2264; 2265; GCN-IR-LABEL: v_test_srem24_pow2_k_num_i64: 2266; GCN-IR: ; %bb.0: 2267; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2268; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2269; GCN-IR-NEXT: s_mov_b32 s4, 0x47000000 2270; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 2271; GCN-IR-NEXT: v_ashrrev_i32_e32 v3, 30, v0 2272; GCN-IR-NEXT: v_or_b32_e32 v3, 1, v3 2273; GCN-IR-NEXT: v_rcp_iflag_f32_e32 v2, v1 2274; GCN-IR-NEXT: v_mul_f32_e32 v2, s4, v2 2275; GCN-IR-NEXT: v_trunc_f32_e32 v2, v2 2276; GCN-IR-NEXT: v_mad_f32 v4, -v2, v1, s4 2277; GCN-IR-NEXT: v_cvt_i32_f32_e32 v2, v2 2278; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 2279; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc 2280; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v2, v1 2281; GCN-IR-NEXT: v_mul_lo_u32 v0, v1, v0 2282; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 2283; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2284; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2285; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2286 %x.shr = ashr i64 %x, 40 2287 %result = srem i64 32768, %x.shr 2288 ret i64 %result 2289} 2290 2291define i64 @v_test_srem24_pow2_k_den_i64(i64 %x) { 2292; GCN-LABEL: v_test_srem24_pow2_k_den_i64: 2293; GCN: ; %bb.0: 2294; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2295; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2296; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v1 2297; GCN-NEXT: v_add_i32_e32 v2, vcc, v0, v2 2298; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 2299; GCN-NEXT: v_and_b32_e32 v2, 0xffff8000, v2 2300; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 2301; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 2302; GCN-NEXT: s_setpc_b64 s[30:31] 2303; 2304; GCN-IR-LABEL: v_test_srem24_pow2_k_den_i64: 2305; GCN-IR: ; %bb.0: 2306; GCN-IR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 2307; GCN-IR-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 2308; GCN-IR-NEXT: s_mov_b32 s4, 0x47000000 2309; GCN-IR-NEXT: v_cvt_f32_i32_e32 v1, v0 2310; GCN-IR-NEXT: v_ashrrev_i32_e32 v2, 30, v0 2311; GCN-IR-NEXT: v_or_b32_e32 v2, 1, v2 2312; GCN-IR-NEXT: v_mul_f32_e32 v3, 0x38000000, v1 2313; GCN-IR-NEXT: v_trunc_f32_e32 v3, v3 2314; GCN-IR-NEXT: v_mad_f32 v1, -v3, s4, v1 2315; GCN-IR-NEXT: v_cvt_i32_f32_e32 v3, v3 2316; GCN-IR-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, s4 2317; GCN-IR-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc 2318; GCN-IR-NEXT: v_add_i32_e32 v1, vcc, v3, v1 2319; GCN-IR-NEXT: v_lshlrev_b32_e32 v1, 15, v1 2320; GCN-IR-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 2321; GCN-IR-NEXT: v_bfe_i32 v0, v0, 0, 24 2322; GCN-IR-NEXT: v_ashrrev_i32_e32 v1, 31, v0 2323; GCN-IR-NEXT: s_setpc_b64 s[30:31] 2324 %x.shr = ashr i64 %x, 40 2325 %result = srem i64 %x.shr, 32768 2326 ret i64 %result 2327} 2328