1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s 3 4define amdgpu_kernel void @s_test_srem(i64 addrspace(1)* %out, i64 %x, i64 %y) { 5; GCN-LABEL: s_test_srem: 6; GCN: ; %bb.0: 7; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0x9 8; GCN-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xd 9; GCN-NEXT: s_mov_b32 s7, 0xf000 10; GCN-NEXT: s_mov_b32 s6, -1 11; GCN-NEXT: v_mov_b32_e32 v0, 0 12; GCN-NEXT: v_mov_b32_e32 v1, 0 13; GCN-NEXT: s_waitcnt lgkmcnt(0) 14; GCN-NEXT: s_mov_b32 s4, s8 15; GCN-NEXT: s_mov_b32 s5, s9 16; GCN-NEXT: v_cvt_f32_u32_e32 v2, s2 17; GCN-NEXT: v_cvt_f32_u32_e32 v3, s3 18; GCN-NEXT: s_sub_u32 s8, 0, s2 19; GCN-NEXT: v_mov_b32_e32 v4, s3 20; GCN-NEXT: v_mov_b32_e32 v5, s11 21; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 22; GCN-NEXT: s_subb_u32 s9, 0, s3 23; GCN-NEXT: v_rcp_f32_e32 v2, v2 24; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 25; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 26; GCN-NEXT: v_trunc_f32_e32 v3, v3 27; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 28; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 29; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 30; GCN-NEXT: v_mul_lo_u32 v6, s8, v3 31; GCN-NEXT: v_mul_lo_u32 v7, s9, v2 32; GCN-NEXT: v_mul_hi_u32 v8, s8, v2 33; GCN-NEXT: v_mul_lo_u32 v9, s8, v2 34; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 35; GCN-NEXT: v_mul_hi_u32 v8, v2, v9 36; GCN-NEXT: v_mul_hi_u32 v10, v3, v9 37; GCN-NEXT: v_mul_lo_u32 v9, v3, v9 38; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 39; GCN-NEXT: v_mul_hi_u32 v7, v2, v6 40; GCN-NEXT: v_mul_lo_u32 v11, v2, v6 41; GCN-NEXT: v_mul_hi_u32 v12, v3, v6 42; GCN-NEXT: v_mul_lo_u32 v6, v3, v6 43; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v11 44; GCN-NEXT: v_addc_u32_e32 v7, vcc, v1, v7, vcc 45; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 46; GCN-NEXT: v_addc_u32_e32 v7, vcc, v7, v10, vcc 47; GCN-NEXT: v_addc_u32_e32 v8, vcc, v12, v0, vcc 48; GCN-NEXT: v_add_i32_e32 v6, vcc, v7, v6 49; GCN-NEXT: v_addc_u32_e32 v7, vcc, v1, v8, vcc 50; GCN-NEXT: v_add_i32_e64 v2, s[0:1], v2, v6 51; GCN-NEXT: v_addc_u32_e64 v6, vcc, v3, v7, s[0:1] 52; GCN-NEXT: v_mul_hi_u32 v8, s8, v2 53; GCN-NEXT: v_mul_lo_u32 v9, s9, v2 54; GCN-NEXT: v_mul_lo_u32 v10, s8, v2 55; GCN-NEXT: v_mul_lo_u32 v11, s8, v6 56; GCN-NEXT: v_mul_hi_u32 v12, v6, v10 57; GCN-NEXT: v_mul_lo_u32 v13, v6, v10 58; GCN-NEXT: v_mul_hi_u32 v10, v2, v10 59; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v11 60; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 61; GCN-NEXT: v_mul_hi_u32 v9, v6, v8 62; GCN-NEXT: v_mul_hi_u32 v11, v2, v8 63; GCN-NEXT: v_mul_lo_u32 v14, v2, v8 64; GCN-NEXT: v_mul_lo_u32 v6, v6, v8 65; GCN-NEXT: v_add_i32_e32 v8, vcc, v10, v14 66; GCN-NEXT: v_addc_u32_e32 v10, vcc, v1, v11, vcc 67; GCN-NEXT: v_add_i32_e32 v8, vcc, v13, v8 68; GCN-NEXT: v_addc_u32_e32 v8, vcc, v10, v12, vcc 69; GCN-NEXT: v_addc_u32_e32 v9, vcc, v9, v0, vcc 70; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v7 71; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 72; GCN-NEXT: v_addc_u32_e32 v7, vcc, v1, v9, vcc 73; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v7, s[0:1] 74; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6 75; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 76; GCN-NEXT: v_mul_hi_u32 v6, s10, v2 77; GCN-NEXT: v_mul_hi_u32 v7, s11, v2 78; GCN-NEXT: v_mul_lo_u32 v2, s11, v2 79; GCN-NEXT: v_mul_hi_u32 v8, s10, v3 80; GCN-NEXT: v_mul_lo_u32 v9, s10, v3 81; GCN-NEXT: v_mul_hi_u32 v10, s11, v3 82; GCN-NEXT: v_mul_lo_u32 v3, s11, v3 83; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v9 84; GCN-NEXT: v_addc_u32_e32 v8, vcc, v1, v8, vcc 85; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6 86; GCN-NEXT: v_addc_u32_e32 v2, vcc, v8, v7, vcc 87; GCN-NEXT: v_addc_u32_e32 v0, vcc, v10, v0, vcc 88; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v3 89; GCN-NEXT: v_addc_u32_e32 v0, vcc, v1, v0, vcc 90; GCN-NEXT: v_mul_hi_u32 v1, s2, v2 91; GCN-NEXT: v_mul_lo_u32 v3, s3, v2 92; GCN-NEXT: v_mul_lo_u32 v2, s2, v2 93; GCN-NEXT: v_mul_lo_u32 v0, s2, v0 94; GCN-NEXT: v_add_i32_e32 v0, vcc, v1, v0 95; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v3 96; GCN-NEXT: v_sub_i32_e32 v1, vcc, s11, v0 97; GCN-NEXT: v_sub_i32_e32 v2, vcc, s10, v2 98; GCN-NEXT: v_subb_u32_e64 v1, s[0:1], v1, v4, vcc 99; GCN-NEXT: v_subb_u32_e32 v0, vcc, v5, v0, vcc 100; GCN-NEXT: v_cmp_le_u32_e32 vcc, s2, v2 101; GCN-NEXT: v_cndmask_b32_e64 v3, 0, -1, vcc 102; GCN-NEXT: v_subrev_i32_e32 v5, vcc, s2, v2 103; GCN-NEXT: v_subb_u32_e64 v4, s[0:1], v1, v4, vcc 104; GCN-NEXT: v_subbrev_u32_e32 v1, vcc, 0, v1, vcc 105; GCN-NEXT: v_cmp_le_u32_e32 vcc, s2, v5 106; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 107; GCN-NEXT: v_subrev_i32_e32 v7, vcc, s2, v5 108; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v0 109; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[0:1] 110; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v1 111; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[0:1] 112; GCN-NEXT: v_subbrev_u32_e32 v4, vcc, 0, v4, vcc 113; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s3, v0 114; GCN-NEXT: v_cndmask_b32_e32 v3, v8, v3, vcc 115; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s3, v1 116; GCN-NEXT: v_cndmask_b32_e32 v6, v9, v6, vcc 117; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6 118; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc 119; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v3 120; GCN-NEXT: v_cndmask_b32_e64 v1, v0, v1, s[0:1] 121; GCN-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc 122; GCN-NEXT: v_cndmask_b32_e64 v0, v2, v0, s[0:1] 123; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 124; GCN-NEXT: s_endpgm 125 %result = urem i64 %x, %y 126 store i64 %result, i64 addrspace(1)* %out 127 ret void 128} 129 130define i64 @v_test_srem(i64 %x, i64 %y) { 131; GCN-LABEL: v_test_srem: 132; GCN: ; %bb.0: 133; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 134; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v3 135; GCN-NEXT: v_mov_b32_e32 v5, 0 136; GCN-NEXT: v_mov_b32_e32 v6, 0 137; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v1 138; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v4 139; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v4, vcc 140; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v7 141; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v7, vcc 142; GCN-NEXT: v_xor_b32_e32 v3, v3, v4 143; GCN-NEXT: v_xor_b32_e32 v2, v2, v4 144; GCN-NEXT: v_xor_b32_e32 v1, v1, v7 145; GCN-NEXT: v_xor_b32_e32 v0, v0, v7 146; GCN-NEXT: v_cvt_f32_u32_e32 v4, v2 147; GCN-NEXT: v_cvt_f32_u32_e32 v8, v3 148; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v2 149; GCN-NEXT: v_subb_u32_e32 v10, vcc, 0, v3, vcc 150; GCN-NEXT: v_mac_f32_e32 v4, 0x4f800000, v8 151; GCN-NEXT: v_rcp_f32_e32 v4, v4 152; GCN-NEXT: v_mul_f32_e32 v4, 0x5f7ffffc, v4 153; GCN-NEXT: v_mul_f32_e32 v8, 0x2f800000, v4 154; GCN-NEXT: v_trunc_f32_e32 v8, v8 155; GCN-NEXT: v_mac_f32_e32 v4, 0xcf800000, v8 156; GCN-NEXT: v_cvt_u32_f32_e32 v8, v8 157; GCN-NEXT: v_cvt_u32_f32_e32 v4, v4 158; GCN-NEXT: v_mul_lo_u32 v11, v9, v8 159; GCN-NEXT: v_mul_lo_u32 v12, v10, v4 160; GCN-NEXT: v_mul_hi_u32 v13, v9, v4 161; GCN-NEXT: v_mul_lo_u32 v14, v9, v4 162; GCN-NEXT: v_add_i32_e32 v11, vcc, v13, v11 163; GCN-NEXT: v_mul_hi_u32 v13, v4, v14 164; GCN-NEXT: v_mul_hi_u32 v15, v8, v14 165; GCN-NEXT: v_mul_lo_u32 v14, v8, v14 166; GCN-NEXT: v_add_i32_e32 v11, vcc, v11, v12 167; GCN-NEXT: v_mul_hi_u32 v12, v4, v11 168; GCN-NEXT: v_mul_lo_u32 v16, v4, v11 169; GCN-NEXT: v_mul_hi_u32 v17, v8, v11 170; GCN-NEXT: v_mul_lo_u32 v11, v8, v11 171; GCN-NEXT: v_add_i32_e32 v13, vcc, v13, v16 172; GCN-NEXT: v_addc_u32_e32 v12, vcc, v6, v12, vcc 173; GCN-NEXT: v_add_i32_e32 v13, vcc, v14, v13 174; GCN-NEXT: v_addc_u32_e32 v12, vcc, v12, v15, vcc 175; GCN-NEXT: v_addc_u32_e32 v13, vcc, v17, v5, vcc 176; GCN-NEXT: v_add_i32_e32 v11, vcc, v12, v11 177; GCN-NEXT: v_addc_u32_e32 v12, vcc, v6, v13, vcc 178; GCN-NEXT: v_add_i32_e64 v4, s[4:5], v4, v11 179; GCN-NEXT: v_addc_u32_e64 v11, vcc, v8, v12, s[4:5] 180; GCN-NEXT: v_mul_hi_u32 v13, v9, v4 181; GCN-NEXT: v_mul_lo_u32 v10, v10, v4 182; GCN-NEXT: v_mul_lo_u32 v14, v9, v4 183; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v12 184; GCN-NEXT: v_mul_lo_u32 v9, v9, v11 185; GCN-NEXT: v_mul_hi_u32 v12, v11, v14 186; GCN-NEXT: v_mul_lo_u32 v15, v11, v14 187; GCN-NEXT: v_mul_hi_u32 v14, v4, v14 188; GCN-NEXT: v_add_i32_e32 v9, vcc, v13, v9 189; GCN-NEXT: v_add_i32_e32 v9, vcc, v9, v10 190; GCN-NEXT: v_mul_hi_u32 v10, v11, v9 191; GCN-NEXT: v_mul_hi_u32 v13, v4, v9 192; GCN-NEXT: v_mul_lo_u32 v16, v4, v9 193; GCN-NEXT: v_mul_lo_u32 v9, v11, v9 194; GCN-NEXT: v_add_i32_e32 v11, vcc, v14, v16 195; GCN-NEXT: v_addc_u32_e32 v13, vcc, v6, v13, vcc 196; GCN-NEXT: v_add_i32_e32 v11, vcc, v15, v11 197; GCN-NEXT: v_addc_u32_e32 v11, vcc, v13, v12, vcc 198; GCN-NEXT: v_addc_u32_e32 v10, vcc, v10, v5, vcc 199; GCN-NEXT: v_add_i32_e32 v9, vcc, v11, v9 200; GCN-NEXT: v_addc_u32_e32 v10, vcc, v6, v10, vcc 201; GCN-NEXT: v_addc_u32_e64 v8, vcc, v8, v10, s[4:5] 202; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v9 203; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc 204; GCN-NEXT: v_mul_hi_u32 v9, v0, v4 205; GCN-NEXT: v_mul_hi_u32 v10, v1, v4 206; GCN-NEXT: v_mul_lo_u32 v4, v1, v4 207; GCN-NEXT: v_mul_hi_u32 v11, v0, v8 208; GCN-NEXT: v_mul_lo_u32 v12, v0, v8 209; GCN-NEXT: v_mul_hi_u32 v13, v1, v8 210; GCN-NEXT: v_mul_lo_u32 v8, v1, v8 211; GCN-NEXT: v_add_i32_e32 v9, vcc, v9, v12 212; GCN-NEXT: v_addc_u32_e32 v11, vcc, v6, v11, vcc 213; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v9 214; GCN-NEXT: v_addc_u32_e32 v4, vcc, v11, v10, vcc 215; GCN-NEXT: v_addc_u32_e32 v5, vcc, v13, v5, vcc 216; GCN-NEXT: v_add_i32_e32 v4, vcc, v4, v8 217; GCN-NEXT: v_addc_u32_e32 v5, vcc, v6, v5, vcc 218; GCN-NEXT: v_mul_hi_u32 v6, v2, v4 219; GCN-NEXT: v_mul_lo_u32 v8, v3, v4 220; GCN-NEXT: v_mul_lo_u32 v4, v2, v4 221; GCN-NEXT: v_mul_lo_u32 v5, v2, v5 222; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 223; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v8 224; GCN-NEXT: v_sub_i32_e32 v6, vcc, v1, v5 225; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4 226; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v6, v3, vcc 227; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v5, vcc 228; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2 229; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 230; GCN-NEXT: v_sub_i32_e32 v6, vcc, v0, v2 231; GCN-NEXT: v_subb_u32_e64 v8, s[4:5], v4, v3, vcc 232; GCN-NEXT: v_subbrev_u32_e32 v4, vcc, 0, v4, vcc 233; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v6, v2 234; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, vcc 235; GCN-NEXT: v_sub_i32_e32 v2, vcc, v6, v2 236; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v1, v3 237; GCN-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[4:5] 238; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v4, v3 239; GCN-NEXT: v_cndmask_b32_e64 v11, 0, -1, s[4:5] 240; GCN-NEXT: v_subbrev_u32_e32 v8, vcc, 0, v8, vcc 241; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 242; GCN-NEXT: v_cndmask_b32_e32 v5, v10, v5, vcc 243; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 244; GCN-NEXT: v_cndmask_b32_e32 v3, v11, v9, vcc 245; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3 246; GCN-NEXT: v_cndmask_b32_e32 v3, v4, v8, vcc 247; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v5 248; GCN-NEXT: v_cndmask_b32_e64 v1, v1, v3, s[4:5] 249; GCN-NEXT: v_cndmask_b32_e32 v2, v6, v2, vcc 250; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5] 251; GCN-NEXT: v_xor_b32_e32 v1, v1, v7 252; GCN-NEXT: v_xor_b32_e32 v0, v0, v7 253; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v7 254; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v7, vcc 255; GCN-NEXT: s_setpc_b64 s[30:31] 256 %result = srem i64 %x, %y 257 ret i64 %result 258} 259 260define amdgpu_kernel void @s_test_srem23_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 261; GCN-LABEL: s_test_srem23_64: 262; GCN: ; %bb.0: 263; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 264; GCN-NEXT: s_load_dword s9, s[0:1], 0xe 265; GCN-NEXT: s_mov_b32 s3, 0xf000 266; GCN-NEXT: s_mov_b32 s2, -1 267; GCN-NEXT: s_waitcnt lgkmcnt(0) 268; GCN-NEXT: s_mov_b32 s0, s4 269; GCN-NEXT: s_mov_b32 s1, s5 270; GCN-NEXT: s_ashr_i64 s[4:5], s[6:7], 41 271; GCN-NEXT: s_ashr_i64 s[6:7], s[8:9], 41 272; GCN-NEXT: s_xor_b32 s5, s4, s6 273; GCN-NEXT: v_cvt_f32_i32_e32 v0, s6 274; GCN-NEXT: v_cvt_f32_i32_e32 v1, s4 275; GCN-NEXT: s_ashr_i32 s5, s5, 30 276; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 277; GCN-NEXT: s_or_b32 s5, s5, 1 278; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 279; GCN-NEXT: v_trunc_f32_e32 v2, v2 280; GCN-NEXT: v_mov_b32_e32 v3, s5 281; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 282; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 283; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 284; GCN-NEXT: v_cvt_i32_f32_e32 v1, v2 285; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 286; GCN-NEXT: v_mul_lo_u32 v0, v0, s6 287; GCN-NEXT: v_sub_i32_e32 v0, vcc, s4, v0 288; GCN-NEXT: v_bfe_i32 v0, v0, 0, 23 289; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 290; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 291; GCN-NEXT: s_endpgm 292 %1 = ashr i64 %x, 41 293 %2 = ashr i64 %y, 41 294 %result = srem i64 %1, %2 295 store i64 %result, i64 addrspace(1)* %out 296 ret void 297} 298 299define amdgpu_kernel void @s_test_srem24_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 300; GCN-LABEL: s_test_srem24_64: 301; GCN: ; %bb.0: 302; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 303; GCN-NEXT: s_load_dword s9, s[0:1], 0xe 304; GCN-NEXT: s_mov_b32 s3, 0xf000 305; GCN-NEXT: s_mov_b32 s2, -1 306; GCN-NEXT: s_waitcnt lgkmcnt(0) 307; GCN-NEXT: s_mov_b32 s0, s4 308; GCN-NEXT: s_mov_b32 s1, s5 309; GCN-NEXT: s_ashr_i64 s[4:5], s[6:7], 40 310; GCN-NEXT: s_ashr_i64 s[6:7], s[8:9], 40 311; GCN-NEXT: s_xor_b32 s5, s4, s6 312; GCN-NEXT: v_cvt_f32_i32_e32 v0, s6 313; GCN-NEXT: v_cvt_f32_i32_e32 v1, s4 314; GCN-NEXT: s_ashr_i32 s5, s5, 30 315; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v0 316; GCN-NEXT: s_or_b32 s5, s5, 1 317; GCN-NEXT: v_mul_f32_e32 v2, v1, v2 318; GCN-NEXT: v_trunc_f32_e32 v2, v2 319; GCN-NEXT: v_mov_b32_e32 v3, s5 320; GCN-NEXT: v_mad_f32 v1, -v2, v0, v1 321; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v0| 322; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v3, vcc 323; GCN-NEXT: v_cvt_i32_f32_e32 v1, v2 324; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 325; GCN-NEXT: v_mul_lo_u32 v0, v0, s6 326; GCN-NEXT: v_sub_i32_e32 v0, vcc, s4, v0 327; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 328; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 329; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 330; GCN-NEXT: s_endpgm 331 %1 = ashr i64 %x, 40 332 %2 = ashr i64 %y, 40 333 %result = srem i64 %1, %2 334 store i64 %result, i64 addrspace(1)* %out 335 ret void 336} 337 338define i64 @v_test_srem24_64(i64 %x, i64 %y) { 339; GCN-LABEL: v_test_srem24_64: 340; GCN: ; %bb.0: 341; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 342; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 343; GCN-NEXT: v_ashr_i64 v[1:2], v[2:3], 40 344; GCN-NEXT: v_xor_b32_e32 v2, v0, v1 345; GCN-NEXT: v_cvt_f32_i32_e32 v3, v0 346; GCN-NEXT: v_cvt_f32_i32_e32 v4, v1 347; GCN-NEXT: v_ashrrev_i32_e32 v2, 30, v2 348; GCN-NEXT: v_rcp_iflag_f32_e32 v5, v4 349; GCN-NEXT: v_or_b32_e32 v2, 1, v2 350; GCN-NEXT: v_mul_f32_e32 v5, v3, v5 351; GCN-NEXT: v_trunc_f32_e32 v5, v5 352; GCN-NEXT: v_mad_f32 v3, -v5, v4, v3 353; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v4| 354; GCN-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc 355; GCN-NEXT: v_cvt_i32_f32_e32 v3, v5 356; GCN-NEXT: v_add_i32_e32 v2, vcc, v3, v2 357; GCN-NEXT: v_mul_lo_u32 v1, v2, v1 358; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 359; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 360; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 361; GCN-NEXT: s_setpc_b64 s[30:31] 362 %1 = ashr i64 %x, 40 363 %2 = ashr i64 %y, 40 364 %result = srem i64 %1, %2 365 ret i64 %result 366} 367 368define amdgpu_kernel void @s_test_srem25_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 369; GCN-LABEL: s_test_srem25_64: 370; GCN: ; %bb.0: 371; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 372; GCN-NEXT: s_load_dword s3, s[0:1], 0xe 373; GCN-NEXT: s_mov_b32 s11, 0xf000 374; GCN-NEXT: s_mov_b32 s10, -1 375; GCN-NEXT: s_waitcnt lgkmcnt(0) 376; GCN-NEXT: s_mov_b32 s8, s4 377; GCN-NEXT: s_mov_b32 s9, s5 378; GCN-NEXT: s_ashr_i64 s[0:1], s[6:7], 39 379; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 39 380; GCN-NEXT: s_ashr_i32 s1, s2, 31 381; GCN-NEXT: s_ashr_i32 s4, s0, 31 382; GCN-NEXT: s_add_i32 s2, s2, s1 383; GCN-NEXT: s_add_i32 s0, s0, s4 384; GCN-NEXT: s_xor_b32 s5, s2, s1 385; GCN-NEXT: s_xor_b32 s2, s0, s4 386; GCN-NEXT: v_cvt_f32_u32_e32 v0, s5 387; GCN-NEXT: v_rcp_iflag_f32_e32 v0, v0 388; GCN-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0 389; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 390; GCN-NEXT: v_mul_hi_u32 v1, v0, s5 391; GCN-NEXT: v_mul_lo_u32 v2, v0, s5 392; GCN-NEXT: v_sub_i32_e32 v3, vcc, 0, v2 393; GCN-NEXT: v_cmp_eq_u32_e64 s[0:1], 0, v1 394; GCN-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[0:1] 395; GCN-NEXT: v_mul_hi_u32 v1, v1, v0 396; GCN-NEXT: v_add_i32_e32 v2, vcc, v1, v0 397; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v1, v0 398; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1] 399; GCN-NEXT: v_mul_hi_u32 v0, v0, s2 400; GCN-NEXT: v_mul_lo_u32 v0, v0, s5 401; GCN-NEXT: v_sub_i32_e32 v1, vcc, s2, v0 402; GCN-NEXT: v_add_i32_e32 v2, vcc, s5, v1 403; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], s2, v0 404; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s5, v1 405; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s5, v1 406; GCN-NEXT: s_and_b64 vcc, s[2:3], s[0:1] 407; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc 408; GCN-NEXT: v_cndmask_b32_e64 v0, v2, v0, s[0:1] 409; GCN-NEXT: v_xor_b32_e32 v0, s4, v0 410; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s4, v0 411; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 412; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 413; GCN-NEXT: s_endpgm 414 %1 = ashr i64 %x, 39 415 %2 = ashr i64 %y, 39 416 %result = srem i64 %1, %2 417 store i64 %result, i64 addrspace(1)* %out 418 ret void 419} 420 421define amdgpu_kernel void @s_test_srem31_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 422; GCN-LABEL: s_test_srem31_64: 423; GCN: ; %bb.0: 424; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 425; GCN-NEXT: s_load_dword s3, s[0:1], 0xe 426; GCN-NEXT: s_mov_b32 s11, 0xf000 427; GCN-NEXT: s_mov_b32 s10, -1 428; GCN-NEXT: s_waitcnt lgkmcnt(0) 429; GCN-NEXT: s_mov_b32 s8, s4 430; GCN-NEXT: s_mov_b32 s9, s5 431; GCN-NEXT: s_ashr_i64 s[0:1], s[6:7], 33 432; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 33 433; GCN-NEXT: s_ashr_i32 s1, s2, 31 434; GCN-NEXT: s_ashr_i32 s4, s0, 31 435; GCN-NEXT: s_add_i32 s2, s2, s1 436; GCN-NEXT: s_add_i32 s0, s0, s4 437; GCN-NEXT: s_xor_b32 s5, s2, s1 438; GCN-NEXT: s_xor_b32 s2, s0, s4 439; GCN-NEXT: v_cvt_f32_u32_e32 v0, s5 440; GCN-NEXT: v_rcp_iflag_f32_e32 v0, v0 441; GCN-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0 442; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 443; GCN-NEXT: v_mul_hi_u32 v1, v0, s5 444; GCN-NEXT: v_mul_lo_u32 v2, v0, s5 445; GCN-NEXT: v_sub_i32_e32 v3, vcc, 0, v2 446; GCN-NEXT: v_cmp_eq_u32_e64 s[0:1], 0, v1 447; GCN-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[0:1] 448; GCN-NEXT: v_mul_hi_u32 v1, v1, v0 449; GCN-NEXT: v_add_i32_e32 v2, vcc, v1, v0 450; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v1, v0 451; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1] 452; GCN-NEXT: v_mul_hi_u32 v0, v0, s2 453; GCN-NEXT: v_mul_lo_u32 v0, v0, s5 454; GCN-NEXT: v_sub_i32_e32 v1, vcc, s2, v0 455; GCN-NEXT: v_add_i32_e32 v2, vcc, s5, v1 456; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], s2, v0 457; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s5, v1 458; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s5, v1 459; GCN-NEXT: s_and_b64 vcc, s[2:3], s[0:1] 460; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc 461; GCN-NEXT: v_cndmask_b32_e64 v0, v2, v0, s[0:1] 462; GCN-NEXT: v_xor_b32_e32 v0, s4, v0 463; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s4, v0 464; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 465; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 466; GCN-NEXT: s_endpgm 467 %1 = ashr i64 %x, 33 468 %2 = ashr i64 %y, 33 469 %result = srem i64 %1, %2 470 store i64 %result, i64 addrspace(1)* %out 471 ret void 472} 473 474; 32 known sign bits 475define amdgpu_kernel void @s_test_srem32_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 476; GCN-LABEL: s_test_srem32_64: 477; GCN: ; %bb.0: 478; GCN-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9 479; GCN-NEXT: s_load_dword s2, s[0:1], 0xe 480; GCN-NEXT: s_mov_b32 s11, 0xf000 481; GCN-NEXT: s_mov_b32 s10, -1 482; GCN-NEXT: s_waitcnt lgkmcnt(0) 483; GCN-NEXT: s_mov_b32 s8, s4 484; GCN-NEXT: s_mov_b32 s9, s5 485; GCN-NEXT: s_ashr_i32 s0, s2, 31 486; GCN-NEXT: s_ashr_i32 s4, s7, 31 487; GCN-NEXT: s_add_i32 s2, s2, s0 488; GCN-NEXT: s_add_i32 s1, s7, s4 489; GCN-NEXT: s_xor_b32 s5, s2, s0 490; GCN-NEXT: s_xor_b32 s2, s1, s4 491; GCN-NEXT: v_cvt_f32_u32_e32 v0, s5 492; GCN-NEXT: v_rcp_iflag_f32_e32 v0, v0 493; GCN-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0 494; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0 495; GCN-NEXT: v_mul_hi_u32 v1, v0, s5 496; GCN-NEXT: v_mul_lo_u32 v2, v0, s5 497; GCN-NEXT: v_sub_i32_e32 v3, vcc, 0, v2 498; GCN-NEXT: v_cmp_eq_u32_e64 s[0:1], 0, v1 499; GCN-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[0:1] 500; GCN-NEXT: v_mul_hi_u32 v1, v1, v0 501; GCN-NEXT: v_add_i32_e32 v2, vcc, v1, v0 502; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v1, v0 503; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[0:1] 504; GCN-NEXT: v_mul_hi_u32 v0, v0, s2 505; GCN-NEXT: v_mul_lo_u32 v0, v0, s5 506; GCN-NEXT: v_sub_i32_e32 v1, vcc, s2, v0 507; GCN-NEXT: v_add_i32_e32 v2, vcc, s5, v1 508; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], s2, v0 509; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s5, v1 510; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s5, v1 511; GCN-NEXT: s_and_b64 vcc, s[2:3], s[0:1] 512; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc 513; GCN-NEXT: v_cndmask_b32_e64 v0, v2, v0, s[0:1] 514; GCN-NEXT: v_xor_b32_e32 v0, s4, v0 515; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s4, v0 516; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 517; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 518; GCN-NEXT: s_endpgm 519 %1 = ashr i64 %x, 32 520 %2 = ashr i64 %y, 32 521 %result = srem i64 %1, %2 522 store i64 %result, i64 addrspace(1)* %out 523 ret void 524} 525 526; 33 known sign bits 527define amdgpu_kernel void @s_test_srem33_64(i64 addrspace(1)* %out, i64 %x, i64 %y) { 528; GCN-LABEL: s_test_srem33_64: 529; GCN: ; %bb.0: 530; GCN-NEXT: s_load_dwordx4 s[8:11], s[0:1], 0x9 531; GCN-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xd 532; GCN-NEXT: s_mov_b32 s7, 0xf000 533; GCN-NEXT: s_mov_b32 s6, -1 534; GCN-NEXT: v_mov_b32_e32 v0, 0 535; GCN-NEXT: v_mov_b32_e32 v1, 0 536; GCN-NEXT: s_waitcnt lgkmcnt(0) 537; GCN-NEXT: s_mov_b32 s4, s8 538; GCN-NEXT: s_mov_b32 s5, s9 539; GCN-NEXT: s_ashr_i64 s[0:1], s[10:11], 31 540; GCN-NEXT: s_ashr_i32 s8, s3, 31 541; GCN-NEXT: s_ashr_i64 s[2:3], s[2:3], 31 542; GCN-NEXT: s_mov_b32 s9, s8 543; GCN-NEXT: s_add_u32 s2, s2, s8 544; GCN-NEXT: s_addc_u32 s3, s3, s8 545; GCN-NEXT: s_xor_b64 s[2:3], s[2:3], s[8:9] 546; GCN-NEXT: v_cvt_f32_u32_e32 v2, s2 547; GCN-NEXT: v_cvt_f32_u32_e32 v3, s3 548; GCN-NEXT: s_sub_u32 s12, 0, s2 549; GCN-NEXT: v_mov_b32_e32 v4, s3 550; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 551; GCN-NEXT: s_subb_u32 s13, 0, s3 552; GCN-NEXT: s_ashr_i32 s8, s11, 31 553; GCN-NEXT: v_rcp_f32_e32 v2, v2 554; GCN-NEXT: s_mov_b32 s9, s8 555; GCN-NEXT: s_add_u32 s0, s0, s8 556; GCN-NEXT: v_mov_b32_e32 v3, s8 557; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 558; GCN-NEXT: s_addc_u32 s1, s1, s8 559; GCN-NEXT: v_mul_f32_e32 v5, 0x2f800000, v2 560; GCN-NEXT: s_xor_b64 s[10:11], s[0:1], s[8:9] 561; GCN-NEXT: v_trunc_f32_e32 v5, v5 562; GCN-NEXT: v_mov_b32_e32 v6, s11 563; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v5 564; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5 565; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 566; GCN-NEXT: v_mul_lo_u32 v7, s12, v5 567; GCN-NEXT: v_mul_lo_u32 v8, s13, v2 568; GCN-NEXT: v_mul_hi_u32 v9, s12, v2 569; GCN-NEXT: v_mul_lo_u32 v10, s12, v2 570; GCN-NEXT: v_add_i32_e32 v7, vcc, v9, v7 571; GCN-NEXT: v_mul_hi_u32 v9, v2, v10 572; GCN-NEXT: v_mul_hi_u32 v11, v5, v10 573; GCN-NEXT: v_mul_lo_u32 v10, v5, v10 574; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v8 575; GCN-NEXT: v_mul_hi_u32 v8, v2, v7 576; GCN-NEXT: v_mul_lo_u32 v12, v2, v7 577; GCN-NEXT: v_mul_hi_u32 v13, v5, v7 578; GCN-NEXT: v_mul_lo_u32 v7, v5, v7 579; GCN-NEXT: v_add_i32_e32 v9, vcc, v9, v12 580; GCN-NEXT: v_addc_u32_e32 v8, vcc, v1, v8, vcc 581; GCN-NEXT: v_add_i32_e32 v9, vcc, v10, v9 582; GCN-NEXT: v_addc_u32_e32 v8, vcc, v8, v11, vcc 583; GCN-NEXT: v_addc_u32_e32 v9, vcc, v13, v0, vcc 584; GCN-NEXT: v_add_i32_e32 v7, vcc, v8, v7 585; GCN-NEXT: v_addc_u32_e32 v8, vcc, v1, v9, vcc 586; GCN-NEXT: v_add_i32_e64 v2, s[0:1], v2, v7 587; GCN-NEXT: v_addc_u32_e64 v7, vcc, v5, v8, s[0:1] 588; GCN-NEXT: v_mul_hi_u32 v9, s12, v2 589; GCN-NEXT: v_mul_lo_u32 v10, s13, v2 590; GCN-NEXT: v_mul_lo_u32 v11, s12, v2 591; GCN-NEXT: v_mul_lo_u32 v12, s12, v7 592; GCN-NEXT: v_mul_hi_u32 v13, v7, v11 593; GCN-NEXT: v_mul_lo_u32 v14, v7, v11 594; GCN-NEXT: v_mul_hi_u32 v11, v2, v11 595; GCN-NEXT: v_add_i32_e32 v9, vcc, v9, v12 596; GCN-NEXT: v_add_i32_e32 v9, vcc, v10, v9 597; GCN-NEXT: v_mul_hi_u32 v10, v7, v9 598; GCN-NEXT: v_mul_hi_u32 v12, v2, v9 599; GCN-NEXT: v_mul_lo_u32 v15, v2, v9 600; GCN-NEXT: v_mul_lo_u32 v7, v7, v9 601; GCN-NEXT: v_add_i32_e32 v9, vcc, v11, v15 602; GCN-NEXT: v_addc_u32_e32 v11, vcc, v1, v12, vcc 603; GCN-NEXT: v_add_i32_e32 v9, vcc, v14, v9 604; GCN-NEXT: v_addc_u32_e32 v9, vcc, v11, v13, vcc 605; GCN-NEXT: v_addc_u32_e32 v10, vcc, v10, v0, vcc 606; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v8 607; GCN-NEXT: v_add_i32_e32 v7, vcc, v9, v7 608; GCN-NEXT: v_addc_u32_e32 v8, vcc, v1, v10, vcc 609; GCN-NEXT: v_addc_u32_e64 v5, vcc, v5, v8, s[0:1] 610; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v7 611; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc 612; GCN-NEXT: v_mul_hi_u32 v7, s10, v2 613; GCN-NEXT: v_mul_hi_u32 v8, s11, v2 614; GCN-NEXT: v_mul_lo_u32 v2, s11, v2 615; GCN-NEXT: v_mul_hi_u32 v9, s10, v5 616; GCN-NEXT: v_mul_lo_u32 v10, s10, v5 617; GCN-NEXT: v_mul_hi_u32 v11, s11, v5 618; GCN-NEXT: v_mul_lo_u32 v5, s11, v5 619; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 620; GCN-NEXT: v_addc_u32_e32 v9, vcc, v1, v9, vcc 621; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v7 622; GCN-NEXT: v_addc_u32_e32 v2, vcc, v9, v8, vcc 623; GCN-NEXT: v_addc_u32_e32 v0, vcc, v11, v0, vcc 624; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v5 625; GCN-NEXT: v_addc_u32_e32 v0, vcc, v1, v0, vcc 626; GCN-NEXT: v_mul_hi_u32 v1, s2, v2 627; GCN-NEXT: v_mul_lo_u32 v5, s3, v2 628; GCN-NEXT: v_mul_lo_u32 v2, s2, v2 629; GCN-NEXT: v_mul_lo_u32 v0, s2, v0 630; GCN-NEXT: v_add_i32_e32 v0, vcc, v1, v0 631; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v5 632; GCN-NEXT: v_sub_i32_e32 v1, vcc, s11, v0 633; GCN-NEXT: v_sub_i32_e32 v2, vcc, s10, v2 634; GCN-NEXT: v_subb_u32_e64 v1, s[0:1], v1, v4, vcc 635; GCN-NEXT: v_subb_u32_e32 v0, vcc, v6, v0, vcc 636; GCN-NEXT: v_cmp_le_u32_e32 vcc, s2, v2 637; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 638; GCN-NEXT: v_subrev_i32_e32 v6, vcc, s2, v2 639; GCN-NEXT: v_subb_u32_e64 v4, s[0:1], v1, v4, vcc 640; GCN-NEXT: v_subbrev_u32_e32 v1, vcc, 0, v1, vcc 641; GCN-NEXT: v_cmp_le_u32_e32 vcc, s2, v6 642; GCN-NEXT: v_cndmask_b32_e64 v7, 0, -1, vcc 643; GCN-NEXT: v_subrev_i32_e32 v8, vcc, s2, v6 644; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v0 645; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[0:1] 646; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v1 647; GCN-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[0:1] 648; GCN-NEXT: v_subbrev_u32_e32 v4, vcc, 0, v4, vcc 649; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s3, v0 650; GCN-NEXT: v_cndmask_b32_e32 v5, v9, v5, vcc 651; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s3, v1 652; GCN-NEXT: v_cndmask_b32_e32 v7, v10, v7, vcc 653; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7 654; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc 655; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v5 656; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[0:1] 657; GCN-NEXT: v_cndmask_b32_e32 v1, v6, v8, vcc 658; GCN-NEXT: v_cndmask_b32_e64 v1, v2, v1, s[0:1] 659; GCN-NEXT: v_xor_b32_e32 v2, s8, v0 660; GCN-NEXT: v_xor_b32_e32 v0, s8, v1 661; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s8, v0 662; GCN-NEXT: v_subb_u32_e32 v1, vcc, v2, v3, vcc 663; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 664; GCN-NEXT: s_endpgm 665 %1 = ashr i64 %x, 31 666 %2 = ashr i64 %y, 31 667 %result = srem i64 %1, %2 668 store i64 %result, i64 addrspace(1)* %out 669 ret void 670} 671 672define amdgpu_kernel void @s_test_srem24_48(i48 addrspace(1)* %out, i48 %x, i48 %y) { 673; GCN-LABEL: s_test_srem24_48: 674; GCN: ; %bb.0: 675; GCN-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9 676; GCN-NEXT: s_load_dword s2, s[0:1], 0xb 677; GCN-NEXT: s_load_dword s3, s[0:1], 0xc 678; GCN-NEXT: s_load_dword s8, s[0:1], 0xd 679; GCN-NEXT: s_load_dword s9, s[0:1], 0xe 680; GCN-NEXT: s_mov_b32 s7, 0xf000 681; GCN-NEXT: s_mov_b32 s6, -1 682; GCN-NEXT: s_waitcnt lgkmcnt(0) 683; GCN-NEXT: s_sext_i32_i16 s0, s3 684; GCN-NEXT: s_sext_i32_i16 s1, s9 685; GCN-NEXT: v_mov_b32_e32 v0, s8 686; GCN-NEXT: v_mov_b32_e32 v1, s2 687; GCN-NEXT: v_alignbit_b32 v0, s1, v0, 24 688; GCN-NEXT: v_alignbit_b32 v1, s0, v1, 24 689; GCN-NEXT: v_xor_b32_e32 v2, v1, v0 690; GCN-NEXT: v_cvt_f32_i32_e32 v3, v1 691; GCN-NEXT: v_cvt_f32_i32_e32 v4, v0 692; GCN-NEXT: v_ashrrev_i32_e32 v2, 30, v2 693; GCN-NEXT: v_rcp_iflag_f32_e32 v5, v4 694; GCN-NEXT: v_or_b32_e32 v2, 1, v2 695; GCN-NEXT: v_mul_f32_e32 v5, v3, v5 696; GCN-NEXT: v_trunc_f32_e32 v5, v5 697; GCN-NEXT: v_mad_f32 v3, -v5, v4, v3 698; GCN-NEXT: v_cvt_i32_f32_e32 v5, v5 699; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v4| 700; GCN-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc 701; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v5 702; GCN-NEXT: v_mul_lo_u32 v0, v2, v0 703; GCN-NEXT: v_sub_i32_e32 v0, vcc, v1, v0 704; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 705; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 706; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0 707; GCN-NEXT: buffer_store_short v1, off, s[4:7], 0 offset:4 708; GCN-NEXT: s_endpgm 709 %1 = ashr i48 %x, 24 710 %2 = ashr i48 %y, 24 711 %result = srem i48 %1, %2 712 store i48 %result, i48 addrspace(1)* %out 713 ret void 714} 715 716define amdgpu_kernel void @s_test_srem_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 717; GCN-LABEL: s_test_srem_k_num_i64: 718; GCN: ; %bb.0: 719; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 720; GCN-NEXT: s_mov_b32 s7, 0xf000 721; GCN-NEXT: s_mov_b32 s6, -1 722; GCN-NEXT: v_mov_b32_e32 v0, 0 723; GCN-NEXT: s_waitcnt lgkmcnt(0) 724; GCN-NEXT: s_ashr_i32 s8, s3, 31 725; GCN-NEXT: v_mov_b32_e32 v1, 0 726; GCN-NEXT: s_mov_b32 s4, s0 727; GCN-NEXT: s_mov_b32 s5, s1 728; GCN-NEXT: s_mov_b32 s9, s8 729; GCN-NEXT: s_add_u32 s0, s2, s8 730; GCN-NEXT: s_addc_u32 s1, s3, s8 731; GCN-NEXT: s_xor_b64 s[2:3], s[0:1], s[8:9] 732; GCN-NEXT: v_cvt_f32_u32_e32 v2, s2 733; GCN-NEXT: v_cvt_f32_u32_e32 v3, s3 734; GCN-NEXT: s_sub_u32 s8, 0, s2 735; GCN-NEXT: v_mov_b32_e32 v4, s3 736; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v3 737; GCN-NEXT: s_subb_u32 s9, 0, s3 738; GCN-NEXT: v_rcp_f32_e32 v2, v2 739; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 740; GCN-NEXT: v_mul_f32_e32 v3, 0x2f800000, v2 741; GCN-NEXT: v_trunc_f32_e32 v3, v3 742; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v3 743; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3 744; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 745; GCN-NEXT: v_mul_lo_u32 v5, s8, v3 746; GCN-NEXT: v_mul_lo_u32 v6, s9, v2 747; GCN-NEXT: v_mul_hi_u32 v7, s8, v2 748; GCN-NEXT: v_mul_lo_u32 v8, s8, v2 749; GCN-NEXT: v_add_i32_e32 v5, vcc, v7, v5 750; GCN-NEXT: v_mul_hi_u32 v7, v2, v8 751; GCN-NEXT: v_mul_hi_u32 v9, v3, v8 752; GCN-NEXT: v_mul_lo_u32 v8, v3, v8 753; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v6 754; GCN-NEXT: v_mul_hi_u32 v6, v2, v5 755; GCN-NEXT: v_mul_lo_u32 v10, v2, v5 756; GCN-NEXT: v_mul_hi_u32 v11, v3, v5 757; GCN-NEXT: v_mul_lo_u32 v5, v3, v5 758; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 759; GCN-NEXT: v_addc_u32_e32 v6, vcc, v1, v6, vcc 760; GCN-NEXT: v_add_i32_e32 v7, vcc, v8, v7 761; GCN-NEXT: v_addc_u32_e32 v6, vcc, v6, v9, vcc 762; GCN-NEXT: v_addc_u32_e32 v7, vcc, v11, v0, vcc 763; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v5 764; GCN-NEXT: v_addc_u32_e32 v6, vcc, v1, v7, vcc 765; GCN-NEXT: v_add_i32_e64 v2, s[0:1], v2, v5 766; GCN-NEXT: v_addc_u32_e64 v5, vcc, v3, v6, s[0:1] 767; GCN-NEXT: v_mul_hi_u32 v7, s8, v2 768; GCN-NEXT: v_mul_lo_u32 v8, s9, v2 769; GCN-NEXT: v_mul_lo_u32 v9, s8, v2 770; GCN-NEXT: v_mul_lo_u32 v10, s8, v5 771; GCN-NEXT: v_mul_hi_u32 v11, v5, v9 772; GCN-NEXT: v_mul_lo_u32 v12, v5, v9 773; GCN-NEXT: v_mul_hi_u32 v9, v2, v9 774; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v10 775; GCN-NEXT: v_add_i32_e32 v7, vcc, v8, v7 776; GCN-NEXT: v_mul_hi_u32 v8, v5, v7 777; GCN-NEXT: v_mul_hi_u32 v10, v2, v7 778; GCN-NEXT: v_mul_lo_u32 v13, v2, v7 779; GCN-NEXT: v_mul_lo_u32 v5, v5, v7 780; GCN-NEXT: v_add_i32_e32 v7, vcc, v9, v13 781; GCN-NEXT: v_addc_u32_e32 v9, vcc, v1, v10, vcc 782; GCN-NEXT: v_add_i32_e32 v7, vcc, v12, v7 783; GCN-NEXT: v_addc_u32_e32 v7, vcc, v9, v11, vcc 784; GCN-NEXT: v_addc_u32_e32 v8, vcc, v8, v0, vcc 785; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v6 786; GCN-NEXT: v_add_i32_e32 v5, vcc, v7, v5 787; GCN-NEXT: v_addc_u32_e32 v6, vcc, v1, v8, vcc 788; GCN-NEXT: v_addc_u32_e64 v3, vcc, v3, v6, s[0:1] 789; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v5 790; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc 791; GCN-NEXT: v_mul_hi_u32 v5, 24, v2 792; GCN-NEXT: v_mul_hi_u32 v2, 0, v2 793; GCN-NEXT: v_mul_hi_u32 v6, 24, v3 794; GCN-NEXT: v_mul_lo_u32 v7, v3, 24 795; GCN-NEXT: v_mul_hi_u32 v3, 0, v3 796; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v7 797; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v6, vcc 798; GCN-NEXT: v_add_i32_e32 v5, vcc, 0, v5 799; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 800; GCN-NEXT: v_addc_u32_e32 v0, vcc, v3, v0, vcc 801; GCN-NEXT: v_mul_hi_u32 v2, s2, v1 802; GCN-NEXT: v_mul_lo_u32 v3, s3, v1 803; GCN-NEXT: v_mul_lo_u32 v1, s2, v1 804; GCN-NEXT: v_mul_lo_u32 v0, s2, v0 805; GCN-NEXT: v_add_i32_e32 v0, vcc, v2, v0 806; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v3 807; GCN-NEXT: v_sub_i32_e32 v2, vcc, 0, v0 808; GCN-NEXT: v_sub_i32_e32 v3, vcc, 24, v1 809; GCN-NEXT: v_subb_u32_e64 v1, s[0:1], v2, v4, vcc 810; GCN-NEXT: v_subb_u32_e32 v0, vcc, 0, v0, vcc 811; GCN-NEXT: v_cmp_le_u32_e32 vcc, s2, v3 812; GCN-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc 813; GCN-NEXT: v_subrev_i32_e32 v5, vcc, s2, v3 814; GCN-NEXT: v_subb_u32_e64 v4, s[0:1], v1, v4, vcc 815; GCN-NEXT: v_subbrev_u32_e32 v1, vcc, 0, v1, vcc 816; GCN-NEXT: v_cmp_le_u32_e32 vcc, s2, v5 817; GCN-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc 818; GCN-NEXT: v_subrev_i32_e32 v7, vcc, s2, v5 819; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v0 820; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[0:1] 821; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v1 822; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[0:1] 823; GCN-NEXT: v_subbrev_u32_e32 v4, vcc, 0, v4, vcc 824; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s3, v0 825; GCN-NEXT: v_cndmask_b32_e32 v2, v8, v2, vcc 826; GCN-NEXT: v_cmp_eq_u32_e32 vcc, s3, v1 827; GCN-NEXT: v_cndmask_b32_e32 v6, v9, v6, vcc 828; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6 829; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc 830; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v2 831; GCN-NEXT: v_cndmask_b32_e64 v1, v0, v1, s[0:1] 832; GCN-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc 833; GCN-NEXT: v_cndmask_b32_e64 v0, v3, v0, s[0:1] 834; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 835; GCN-NEXT: s_endpgm 836 %result = srem i64 24, %x 837 store i64 %result, i64 addrspace(1)* %out 838 ret void 839} 840 841define i64 @v_test_srem_k_num_i64(i64 %x) { 842; GCN-LABEL: v_test_srem_k_num_i64: 843; GCN: ; %bb.0: 844; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 845; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 846; GCN-NEXT: v_mov_b32_e32 v3, 0 847; GCN-NEXT: v_mov_b32_e32 v4, 0 848; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 849; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 850; GCN-NEXT: v_xor_b32_e32 v1, v1, v2 851; GCN-NEXT: v_xor_b32_e32 v0, v0, v2 852; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 853; GCN-NEXT: v_cvt_f32_u32_e32 v5, v1 854; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v0 855; GCN-NEXT: v_subb_u32_e32 v7, vcc, 0, v1, vcc 856; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v5 857; GCN-NEXT: v_rcp_f32_e32 v2, v2 858; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 859; GCN-NEXT: v_mul_f32_e32 v5, 0x2f800000, v2 860; GCN-NEXT: v_trunc_f32_e32 v5, v5 861; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v5 862; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5 863; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 864; GCN-NEXT: v_mul_lo_u32 v8, v6, v5 865; GCN-NEXT: v_mul_lo_u32 v9, v7, v2 866; GCN-NEXT: v_mul_hi_u32 v10, v6, v2 867; GCN-NEXT: v_mul_lo_u32 v11, v6, v2 868; GCN-NEXT: v_add_i32_e32 v8, vcc, v10, v8 869; GCN-NEXT: v_mul_hi_u32 v10, v2, v11 870; GCN-NEXT: v_mul_hi_u32 v12, v5, v11 871; GCN-NEXT: v_mul_lo_u32 v11, v5, v11 872; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v9 873; GCN-NEXT: v_mul_hi_u32 v9, v2, v8 874; GCN-NEXT: v_mul_lo_u32 v13, v2, v8 875; GCN-NEXT: v_mul_hi_u32 v14, v5, v8 876; GCN-NEXT: v_mul_lo_u32 v8, v5, v8 877; GCN-NEXT: v_add_i32_e32 v10, vcc, v10, v13 878; GCN-NEXT: v_addc_u32_e32 v9, vcc, v4, v9, vcc 879; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 880; GCN-NEXT: v_addc_u32_e32 v9, vcc, v9, v12, vcc 881; GCN-NEXT: v_addc_u32_e32 v10, vcc, v14, v3, vcc 882; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 883; GCN-NEXT: v_addc_u32_e32 v9, vcc, v4, v10, vcc 884; GCN-NEXT: v_add_i32_e64 v2, s[4:5], v2, v8 885; GCN-NEXT: v_addc_u32_e64 v8, vcc, v5, v9, s[4:5] 886; GCN-NEXT: v_mul_hi_u32 v10, v6, v2 887; GCN-NEXT: v_mul_lo_u32 v7, v7, v2 888; GCN-NEXT: v_mul_lo_u32 v11, v6, v2 889; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v9 890; GCN-NEXT: v_mul_lo_u32 v6, v6, v8 891; GCN-NEXT: v_mul_hi_u32 v9, v8, v11 892; GCN-NEXT: v_mul_lo_u32 v12, v8, v11 893; GCN-NEXT: v_mul_hi_u32 v11, v2, v11 894; GCN-NEXT: v_add_i32_e32 v6, vcc, v10, v6 895; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 896; GCN-NEXT: v_mul_hi_u32 v7, v8, v6 897; GCN-NEXT: v_mul_hi_u32 v10, v2, v6 898; GCN-NEXT: v_mul_lo_u32 v13, v2, v6 899; GCN-NEXT: v_mul_lo_u32 v6, v8, v6 900; GCN-NEXT: v_add_i32_e32 v8, vcc, v11, v13 901; GCN-NEXT: v_addc_u32_e32 v10, vcc, v4, v10, vcc 902; GCN-NEXT: v_add_i32_e32 v8, vcc, v12, v8 903; GCN-NEXT: v_addc_u32_e32 v8, vcc, v10, v9, vcc 904; GCN-NEXT: v_addc_u32_e32 v7, vcc, v7, v3, vcc 905; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 906; GCN-NEXT: v_addc_u32_e32 v7, vcc, v4, v7, vcc 907; GCN-NEXT: v_addc_u32_e64 v5, vcc, v5, v7, s[4:5] 908; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6 909; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc 910; GCN-NEXT: v_mul_hi_u32 v6, 24, v2 911; GCN-NEXT: v_mul_hi_u32 v2, 0, v2 912; GCN-NEXT: v_mul_hi_u32 v7, 24, v5 913; GCN-NEXT: v_mul_lo_u32 v8, v5, 24 914; GCN-NEXT: v_mul_hi_u32 v5, 0, v5 915; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 916; GCN-NEXT: v_addc_u32_e32 v4, vcc, v4, v7, vcc 917; GCN-NEXT: v_add_i32_e32 v6, vcc, 0, v6 918; GCN-NEXT: v_addc_u32_e32 v2, vcc, v4, v2, vcc 919; GCN-NEXT: v_addc_u32_e32 v3, vcc, v5, v3, vcc 920; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 921; GCN-NEXT: v_mul_lo_u32 v5, v1, v2 922; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 923; GCN-NEXT: v_mul_lo_u32 v3, v0, v3 924; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 925; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 926; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 927; GCN-NEXT: v_sub_i32_e32 v2, vcc, 24, v2 928; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 929; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 930; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 931; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 932; GCN-NEXT: v_sub_i32_e32 v6, vcc, v2, v0 933; GCN-NEXT: v_subb_u32_e64 v7, s[4:5], v4, v1, vcc 934; GCN-NEXT: v_subbrev_u32_e32 v4, vcc, 0, v4, vcc 935; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v6, v0 936; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, vcc 937; GCN-NEXT: v_sub_i32_e32 v0, vcc, v6, v0 938; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v3, v1 939; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[4:5] 940; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v4, v1 941; GCN-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[4:5] 942; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], v3, v1 943; GCN-NEXT: v_cndmask_b32_e64 v5, v9, v5, s[4:5] 944; GCN-NEXT: v_subbrev_u32_e32 v7, vcc, 0, v7, vcc 945; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1 946; GCN-NEXT: v_cndmask_b32_e32 v1, v10, v8, vcc 947; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1 948; GCN-NEXT: v_cndmask_b32_e32 v0, v6, v0, vcc 949; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v5 950; GCN-NEXT: v_cndmask_b32_e64 v0, v2, v0, s[4:5] 951; GCN-NEXT: v_cndmask_b32_e32 v1, v4, v7, vcc 952; GCN-NEXT: v_cndmask_b32_e64 v1, v3, v1, s[4:5] 953; GCN-NEXT: s_setpc_b64 s[30:31] 954 %result = srem i64 24, %x 955 ret i64 %result 956} 957 958define i64 @v_test_srem_pow2_k_num_i64(i64 %x) { 959; GCN-LABEL: v_test_srem_pow2_k_num_i64: 960; GCN: ; %bb.0: 961; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 962; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 963; GCN-NEXT: v_mov_b32_e32 v3, 0 964; GCN-NEXT: v_mov_b32_e32 v4, 0 965; GCN-NEXT: s_mov_b32 s6, 0x8000 966; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2 967; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc 968; GCN-NEXT: v_xor_b32_e32 v1, v1, v2 969; GCN-NEXT: v_xor_b32_e32 v0, v0, v2 970; GCN-NEXT: v_cvt_f32_u32_e32 v2, v0 971; GCN-NEXT: v_cvt_f32_u32_e32 v5, v1 972; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v0 973; GCN-NEXT: v_subb_u32_e32 v7, vcc, 0, v1, vcc 974; GCN-NEXT: v_mac_f32_e32 v2, 0x4f800000, v5 975; GCN-NEXT: v_rcp_f32_e32 v2, v2 976; GCN-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 977; GCN-NEXT: v_mul_f32_e32 v5, 0x2f800000, v2 978; GCN-NEXT: v_trunc_f32_e32 v5, v5 979; GCN-NEXT: v_mac_f32_e32 v2, 0xcf800000, v5 980; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5 981; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2 982; GCN-NEXT: v_mul_lo_u32 v8, v6, v5 983; GCN-NEXT: v_mul_lo_u32 v9, v7, v2 984; GCN-NEXT: v_mul_hi_u32 v10, v6, v2 985; GCN-NEXT: v_mul_lo_u32 v11, v6, v2 986; GCN-NEXT: v_add_i32_e32 v8, vcc, v10, v8 987; GCN-NEXT: v_mul_hi_u32 v10, v2, v11 988; GCN-NEXT: v_mul_hi_u32 v12, v5, v11 989; GCN-NEXT: v_mul_lo_u32 v11, v5, v11 990; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v9 991; GCN-NEXT: v_mul_hi_u32 v9, v2, v8 992; GCN-NEXT: v_mul_lo_u32 v13, v2, v8 993; GCN-NEXT: v_mul_hi_u32 v14, v5, v8 994; GCN-NEXT: v_mul_lo_u32 v8, v5, v8 995; GCN-NEXT: v_add_i32_e32 v10, vcc, v10, v13 996; GCN-NEXT: v_addc_u32_e32 v9, vcc, v4, v9, vcc 997; GCN-NEXT: v_add_i32_e32 v10, vcc, v11, v10 998; GCN-NEXT: v_addc_u32_e32 v9, vcc, v9, v12, vcc 999; GCN-NEXT: v_addc_u32_e32 v10, vcc, v14, v3, vcc 1000; GCN-NEXT: v_add_i32_e32 v8, vcc, v9, v8 1001; GCN-NEXT: v_addc_u32_e32 v9, vcc, v4, v10, vcc 1002; GCN-NEXT: v_add_i32_e64 v2, s[4:5], v2, v8 1003; GCN-NEXT: v_addc_u32_e64 v8, vcc, v5, v9, s[4:5] 1004; GCN-NEXT: v_mul_hi_u32 v10, v6, v2 1005; GCN-NEXT: v_mul_lo_u32 v7, v7, v2 1006; GCN-NEXT: v_mul_lo_u32 v11, v6, v2 1007; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v9 1008; GCN-NEXT: v_mul_lo_u32 v6, v6, v8 1009; GCN-NEXT: v_mul_hi_u32 v9, v8, v11 1010; GCN-NEXT: v_mul_lo_u32 v12, v8, v11 1011; GCN-NEXT: v_mul_hi_u32 v11, v2, v11 1012; GCN-NEXT: v_add_i32_e32 v6, vcc, v10, v6 1013; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v7 1014; GCN-NEXT: v_mul_hi_u32 v7, v8, v6 1015; GCN-NEXT: v_mul_hi_u32 v10, v2, v6 1016; GCN-NEXT: v_mul_lo_u32 v13, v2, v6 1017; GCN-NEXT: v_mul_lo_u32 v6, v8, v6 1018; GCN-NEXT: v_add_i32_e32 v8, vcc, v11, v13 1019; GCN-NEXT: v_addc_u32_e32 v10, vcc, v4, v10, vcc 1020; GCN-NEXT: v_add_i32_e32 v8, vcc, v12, v8 1021; GCN-NEXT: v_addc_u32_e32 v8, vcc, v10, v9, vcc 1022; GCN-NEXT: v_addc_u32_e32 v7, vcc, v7, v3, vcc 1023; GCN-NEXT: v_add_i32_e32 v6, vcc, v8, v6 1024; GCN-NEXT: v_addc_u32_e32 v7, vcc, v4, v7, vcc 1025; GCN-NEXT: v_addc_u32_e64 v5, vcc, v5, v7, s[4:5] 1026; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6 1027; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc 1028; GCN-NEXT: v_mul_hi_u32 v6, s6, v2 1029; GCN-NEXT: v_mul_hi_u32 v2, 0, v2 1030; GCN-NEXT: v_mul_hi_u32 v7, s6, v5 1031; GCN-NEXT: v_lshlrev_b32_e32 v8, 15, v5 1032; GCN-NEXT: v_mul_hi_u32 v5, 0, v5 1033; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v8 1034; GCN-NEXT: v_addc_u32_e32 v4, vcc, v4, v7, vcc 1035; GCN-NEXT: v_add_i32_e32 v6, vcc, 0, v6 1036; GCN-NEXT: v_addc_u32_e32 v2, vcc, v4, v2, vcc 1037; GCN-NEXT: v_addc_u32_e32 v3, vcc, v5, v3, vcc 1038; GCN-NEXT: v_mul_hi_u32 v4, v0, v2 1039; GCN-NEXT: v_mul_lo_u32 v5, v1, v2 1040; GCN-NEXT: v_mul_lo_u32 v2, v0, v2 1041; GCN-NEXT: v_mul_lo_u32 v3, v0, v3 1042; GCN-NEXT: v_add_i32_e32 v3, vcc, v4, v3 1043; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v5 1044; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v3 1045; GCN-NEXT: v_sub_i32_e32 v2, vcc, s6, v2 1046; GCN-NEXT: v_subb_u32_e64 v4, s[4:5], v4, v1, vcc 1047; GCN-NEXT: v_subb_u32_e32 v3, vcc, 0, v3, vcc 1048; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v0 1049; GCN-NEXT: v_cndmask_b32_e64 v5, 0, -1, vcc 1050; GCN-NEXT: v_sub_i32_e32 v6, vcc, v2, v0 1051; GCN-NEXT: v_subb_u32_e64 v7, s[4:5], v4, v1, vcc 1052; GCN-NEXT: v_subbrev_u32_e32 v4, vcc, 0, v4, vcc 1053; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v6, v0 1054; GCN-NEXT: v_cndmask_b32_e64 v8, 0, -1, vcc 1055; GCN-NEXT: v_sub_i32_e32 v0, vcc, v6, v0 1056; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v3, v1 1057; GCN-NEXT: v_cndmask_b32_e64 v9, 0, -1, s[4:5] 1058; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v4, v1 1059; GCN-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[4:5] 1060; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], v3, v1 1061; GCN-NEXT: v_cndmask_b32_e64 v5, v9, v5, s[4:5] 1062; GCN-NEXT: v_subbrev_u32_e32 v7, vcc, 0, v7, vcc 1063; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1 1064; GCN-NEXT: v_cndmask_b32_e32 v1, v10, v8, vcc 1065; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1 1066; GCN-NEXT: v_cndmask_b32_e32 v0, v6, v0, vcc 1067; GCN-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v5 1068; GCN-NEXT: v_cndmask_b32_e64 v0, v2, v0, s[4:5] 1069; GCN-NEXT: v_cndmask_b32_e32 v1, v4, v7, vcc 1070; GCN-NEXT: v_cndmask_b32_e64 v1, v3, v1, s[4:5] 1071; GCN-NEXT: s_setpc_b64 s[30:31] 1072 %result = srem i64 32768, %x 1073 ret i64 %result 1074} 1075 1076define i64 @v_test_srem_pow2_k_den_i64(i64 %x) { 1077; GCN-LABEL: v_test_srem_pow2_k_den_i64: 1078; GCN: ; %bb.0: 1079; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1080; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1 1081; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v2 1082; GCN-NEXT: v_add_i32_e32 v2, vcc, v0, v2 1083; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 1084; GCN-NEXT: v_and_b32_e32 v2, 0xffff8000, v2 1085; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1086; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 1087; GCN-NEXT: s_setpc_b64 s[30:31] 1088 %result = srem i64 %x, 32768 1089 ret i64 %result 1090} 1091 1092define amdgpu_kernel void @s_test_srem24_k_num_i64(i64 addrspace(1)* %out, i64 %x) { 1093; GCN-LABEL: s_test_srem24_k_num_i64: 1094; GCN: ; %bb.0: 1095; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1096; GCN-NEXT: s_mov_b32 s7, 0xf000 1097; GCN-NEXT: s_mov_b32 s6, -1 1098; GCN-NEXT: s_waitcnt lgkmcnt(0) 1099; GCN-NEXT: s_mov_b32 s2, 0x41c00000 1100; GCN-NEXT: s_mov_b32 s4, s0 1101; GCN-NEXT: s_mov_b32 s5, s1 1102; GCN-NEXT: s_ashr_i64 s[0:1], s[2:3], 40 1103; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 1104; GCN-NEXT: s_ashr_i32 s1, s0, 30 1105; GCN-NEXT: v_rcp_iflag_f32_e32 v1, v0 1106; GCN-NEXT: s_or_b32 s1, s1, 1 1107; GCN-NEXT: v_mul_f32_e32 v1, s2, v1 1108; GCN-NEXT: v_mov_b32_e32 v2, s1 1109; GCN-NEXT: v_trunc_f32_e32 v1, v1 1110; GCN-NEXT: v_mad_f32 v3, -v1, v0, s2 1111; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v0| 1112; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc 1113; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1 1114; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 1115; GCN-NEXT: v_mul_lo_u32 v0, v0, s0 1116; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1117; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1118; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1119; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1120; GCN-NEXT: s_endpgm 1121 %x.shr = ashr i64 %x, 40 1122 %result = srem i64 24, %x.shr 1123 store i64 %result, i64 addrspace(1)* %out 1124 ret void 1125} 1126 1127define amdgpu_kernel void @s_test_srem24_k_den_i64(i64 addrspace(1)* %out, i64 %x) { 1128; GCN-LABEL: s_test_srem24_k_den_i64: 1129; GCN: ; %bb.0: 1130; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 1131; GCN-NEXT: s_mov_b32 s7, 0xf000 1132; GCN-NEXT: s_mov_b32 s6, -1 1133; GCN-NEXT: s_waitcnt lgkmcnt(0) 1134; GCN-NEXT: s_mov_b32 s2, 0x46b6fe00 1135; GCN-NEXT: s_mov_b32 s4, s0 1136; GCN-NEXT: s_mov_b32 s5, s1 1137; GCN-NEXT: s_ashr_i64 s[0:1], s[2:3], 40 1138; GCN-NEXT: v_cvt_f32_i32_e32 v0, s0 1139; GCN-NEXT: s_ashr_i32 s1, s0, 30 1140; GCN-NEXT: v_mul_f32_e32 v1, 0x38331158, v0 1141; GCN-NEXT: s_or_b32 s1, s1, 1 1142; GCN-NEXT: v_trunc_f32_e32 v1, v1 1143; GCN-NEXT: v_mov_b32_e32 v2, s1 1144; GCN-NEXT: v_mad_f32 v0, -v1, s2, v0 1145; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1 1146; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v0|, s2 1147; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc 1148; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1 1149; GCN-NEXT: s_movk_i32 s1, 0x5b7f 1150; GCN-NEXT: v_mul_lo_u32 v0, v0, s1 1151; GCN-NEXT: v_sub_i32_e32 v0, vcc, s0, v0 1152; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1153; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1154; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 1155; GCN-NEXT: s_endpgm 1156 %x.shr = ashr i64 %x, 40 1157 %result = srem i64 %x.shr, 23423 1158 store i64 %result, i64 addrspace(1)* %out 1159 ret void 1160} 1161 1162define i64 @v_test_srem24_k_num_i64(i64 %x) { 1163; GCN-LABEL: v_test_srem24_k_num_i64: 1164; GCN: ; %bb.0: 1165; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1166; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 1167; GCN-NEXT: s_mov_b32 s4, 0x41c00000 1168; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 1169; GCN-NEXT: v_ashrrev_i32_e32 v2, 30, v0 1170; GCN-NEXT: v_rcp_iflag_f32_e32 v3, v1 1171; GCN-NEXT: v_or_b32_e32 v2, 1, v2 1172; GCN-NEXT: v_mul_f32_e32 v3, s4, v3 1173; GCN-NEXT: v_trunc_f32_e32 v3, v3 1174; GCN-NEXT: v_mad_f32 v4, -v3, v1, s4 1175; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 1176; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc 1177; GCN-NEXT: v_cvt_i32_f32_e32 v2, v3 1178; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1179; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 1180; GCN-NEXT: v_sub_i32_e32 v0, vcc, 24, v0 1181; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1182; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1183; GCN-NEXT: s_setpc_b64 s[30:31] 1184 %x.shr = ashr i64 %x, 40 1185 %result = srem i64 24, %x.shr 1186 ret i64 %result 1187} 1188 1189define i64 @v_test_srem24_pow2_k_num_i64(i64 %x) { 1190; GCN-LABEL: v_test_srem24_pow2_k_num_i64: 1191; GCN: ; %bb.0: 1192; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1193; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 1194; GCN-NEXT: s_mov_b32 s4, 0x47000000 1195; GCN-NEXT: v_cvt_f32_i32_e32 v1, v0 1196; GCN-NEXT: v_ashrrev_i32_e32 v2, 30, v0 1197; GCN-NEXT: v_rcp_iflag_f32_e32 v3, v1 1198; GCN-NEXT: v_or_b32_e32 v2, 1, v2 1199; GCN-NEXT: v_mul_f32_e32 v3, s4, v3 1200; GCN-NEXT: v_trunc_f32_e32 v3, v3 1201; GCN-NEXT: v_mad_f32 v4, -v3, v1, s4 1202; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v4|, |v1| 1203; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc 1204; GCN-NEXT: v_cvt_i32_f32_e32 v2, v3 1205; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1 1206; GCN-NEXT: v_mul_lo_u32 v0, v1, v0 1207; GCN-NEXT: v_sub_i32_e32 v0, vcc, 0x8000, v0 1208; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24 1209; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 1210; GCN-NEXT: s_setpc_b64 s[30:31] 1211 %x.shr = ashr i64 %x, 40 1212 %result = srem i64 32768, %x.shr 1213 ret i64 %result 1214} 1215 1216define i64 @v_test_srem24_pow2_k_den_i64(i64 %x) { 1217; GCN-LABEL: v_test_srem24_pow2_k_den_i64: 1218; GCN: ; %bb.0: 1219; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 1220; GCN-NEXT: v_ashr_i64 v[0:1], v[0:1], 40 1221; GCN-NEXT: v_lshrrev_b32_e32 v2, 17, v1 1222; GCN-NEXT: v_add_i32_e32 v2, vcc, v0, v2 1223; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 1224; GCN-NEXT: v_and_b32_e32 v2, 0xffff8000, v2 1225; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2 1226; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc 1227; GCN-NEXT: s_setpc_b64 s[30:31] 1228 %x.shr = ashr i64 %x, 40 1229 %result = srem i64 %x.shr, 32768 1230 ret i64 %result 1231} 1232