1; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=SI %s 2; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=SI %s 3 4; SI-LABEL: {{^}}gs_const: 5; SI-NOT: v_cmpx 6; SI: s_mov_b64 exec, 0 7define amdgpu_gs void @gs_const() { 8 %tmp = icmp ule i32 0, 3 9 %tmp1 = select i1 %tmp, float 1.000000e+00, float -1.000000e+00 10 %c1 = fcmp oge float %tmp1, 0.0 11 call void @llvm.amdgcn.kill(i1 %c1) 12 %tmp2 = icmp ule i32 3, 0 13 %tmp3 = select i1 %tmp2, float 1.000000e+00, float -1.000000e+00 14 %c2 = fcmp oge float %tmp3, 0.0 15 call void @llvm.amdgcn.kill(i1 %c2) 16 ret void 17} 18 19; SI-LABEL: {{^}}vcc_implicit_def: 20; SI-NOT: v_cmp_gt_f32_e32 vcc, 21; SI: v_cmp_gt_f32_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], 0, v{{[0-9]+}} 22; SI: v_cmpx_le_f32_e32 vcc, 0, v{{[0-9]+}} 23; SI: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1.0, [[CMP]] 24define amdgpu_ps void @vcc_implicit_def(float %arg13, float %arg14) { 25 %tmp0 = fcmp olt float %arg13, 0.000000e+00 26 %c1 = fcmp oge float %arg14, 0.0 27 call void @llvm.amdgcn.kill(i1 %c1) 28 %tmp1 = select i1 %tmp0, float 1.000000e+00, float 0.000000e+00 29 call void @llvm.amdgcn.exp.f32(i32 1, i32 15, float %tmp1, float %tmp1, float %tmp1, float %tmp1, i1 true, i1 true) #0 30 ret void 31} 32 33; SI-LABEL: {{^}}true: 34; SI-NEXT: %bb. 35; SI-NEXT: %bb. 36; SI-NEXT: s_endpgm 37define amdgpu_gs void @true() { 38 call void @llvm.amdgcn.kill(i1 true) 39 ret void 40} 41 42; SI-LABEL: {{^}}false: 43; SI-NOT: v_cmpx 44; SI: s_mov_b64 exec, 0 45define amdgpu_gs void @false() { 46 call void @llvm.amdgcn.kill(i1 false) 47 ret void 48} 49 50; SI-LABEL: {{^}}and: 51; SI: v_cmp_lt_i32 52; SI: v_cmp_lt_i32 53; SI: s_or_b64 s[0:1] 54; SI: s_and_b64 exec, exec, s[0:1] 55define amdgpu_gs void @and(i32 %a, i32 %b, i32 %c, i32 %d) { 56 %c1 = icmp slt i32 %a, %b 57 %c2 = icmp slt i32 %c, %d 58 %x = or i1 %c1, %c2 59 call void @llvm.amdgcn.kill(i1 %x) 60 ret void 61} 62 63; SI-LABEL: {{^}}andn2: 64; SI: v_cmp_lt_i32 65; SI: v_cmp_lt_i32 66; SI: s_xor_b64 s[0:1] 67; SI: s_andn2_b64 exec, exec, s[0:1] 68define amdgpu_gs void @andn2(i32 %a, i32 %b, i32 %c, i32 %d) { 69 %c1 = icmp slt i32 %a, %b 70 %c2 = icmp slt i32 %c, %d 71 %x = xor i1 %c1, %c2 72 %y = xor i1 %x, 1 73 call void @llvm.amdgcn.kill(i1 %y) 74 ret void 75} 76 77; SI-LABEL: {{^}}oeq: 78; SI: v_cmpx_eq_f32 79; SI-NOT: s_and 80define amdgpu_gs void @oeq(float %a) { 81 %c1 = fcmp oeq float %a, 0.0 82 call void @llvm.amdgcn.kill(i1 %c1) 83 ret void 84} 85 86; SI-LABEL: {{^}}ogt: 87; SI: v_cmpx_lt_f32 88; SI-NOT: s_and 89define amdgpu_gs void @ogt(float %a) { 90 %c1 = fcmp ogt float %a, 0.0 91 call void @llvm.amdgcn.kill(i1 %c1) 92 ret void 93} 94 95; SI-LABEL: {{^}}oge: 96; SI: v_cmpx_le_f32 97; SI-NOT: s_and 98define amdgpu_gs void @oge(float %a) { 99 %c1 = fcmp oge float %a, 0.0 100 call void @llvm.amdgcn.kill(i1 %c1) 101 ret void 102} 103 104; SI-LABEL: {{^}}olt: 105; SI: v_cmpx_gt_f32 106; SI-NOT: s_and 107define amdgpu_gs void @olt(float %a) { 108 %c1 = fcmp olt float %a, 0.0 109 call void @llvm.amdgcn.kill(i1 %c1) 110 ret void 111} 112 113; SI-LABEL: {{^}}ole: 114; SI: v_cmpx_ge_f32 115; SI-NOT: s_and 116define amdgpu_gs void @ole(float %a) { 117 %c1 = fcmp ole float %a, 0.0 118 call void @llvm.amdgcn.kill(i1 %c1) 119 ret void 120} 121 122; SI-LABEL: {{^}}one: 123; SI: v_cmpx_lg_f32 124; SI-NOT: s_and 125define amdgpu_gs void @one(float %a) { 126 %c1 = fcmp one float %a, 0.0 127 call void @llvm.amdgcn.kill(i1 %c1) 128 ret void 129} 130 131; SI-LABEL: {{^}}ord: 132; FIXME: This is absolutely unimportant, but we could use the cmpx variant here. 133; SI: v_cmp_o_f32 134define amdgpu_gs void @ord(float %a) { 135 %c1 = fcmp ord float %a, 0.0 136 call void @llvm.amdgcn.kill(i1 %c1) 137 ret void 138} 139 140; SI-LABEL: {{^}}uno: 141; FIXME: This is absolutely unimportant, but we could use the cmpx variant here. 142; SI: v_cmp_u_f32 143define amdgpu_gs void @uno(float %a) { 144 %c1 = fcmp uno float %a, 0.0 145 call void @llvm.amdgcn.kill(i1 %c1) 146 ret void 147} 148 149; SI-LABEL: {{^}}ueq: 150; SI: v_cmpx_nlg_f32 151; SI-NOT: s_and 152define amdgpu_gs void @ueq(float %a) { 153 %c1 = fcmp ueq float %a, 0.0 154 call void @llvm.amdgcn.kill(i1 %c1) 155 ret void 156} 157 158; SI-LABEL: {{^}}ugt: 159; SI: v_cmpx_nge_f32 160; SI-NOT: s_and 161define amdgpu_gs void @ugt(float %a) { 162 %c1 = fcmp ugt float %a, 0.0 163 call void @llvm.amdgcn.kill(i1 %c1) 164 ret void 165} 166 167; SI-LABEL: {{^}}uge: 168; SI: v_cmpx_ngt_f32_e32 vcc, -1.0 169; SI-NOT: s_and 170define amdgpu_gs void @uge(float %a) { 171 %c1 = fcmp uge float %a, -1.0 172 call void @llvm.amdgcn.kill(i1 %c1) 173 ret void 174} 175 176; SI-LABEL: {{^}}ult: 177; SI: v_cmpx_nle_f32_e32 vcc, -2.0 178; SI-NOT: s_and 179define amdgpu_gs void @ult(float %a) { 180 %c1 = fcmp ult float %a, -2.0 181 call void @llvm.amdgcn.kill(i1 %c1) 182 ret void 183} 184 185; SI-LABEL: {{^}}ule: 186; SI: v_cmpx_nlt_f32_e32 vcc, 2.0 187; SI-NOT: s_and 188define amdgpu_gs void @ule(float %a) { 189 %c1 = fcmp ule float %a, 2.0 190 call void @llvm.amdgcn.kill(i1 %c1) 191 ret void 192} 193 194; SI-LABEL: {{^}}une: 195; SI: v_cmpx_neq_f32_e32 vcc, 0 196; SI-NOT: s_and 197define amdgpu_gs void @une(float %a) { 198 %c1 = fcmp une float %a, 0.0 199 call void @llvm.amdgcn.kill(i1 %c1) 200 ret void 201} 202 203; SI-LABEL: {{^}}neg_olt: 204; SI: v_cmpx_ngt_f32_e32 vcc, 1.0 205; SI-NOT: s_and 206define amdgpu_gs void @neg_olt(float %a) { 207 %c1 = fcmp olt float %a, 1.0 208 %c2 = xor i1 %c1, 1 209 call void @llvm.amdgcn.kill(i1 %c2) 210 ret void 211} 212 213; SI-LABEL: {{^}}fcmp_x2: 214; FIXME: LLVM should be able to combine these fcmp opcodes. 215; SI: v_cmp_gt_f32 216; SI: v_cndmask_b32 217; SI: v_cmpx_le_f32 218define amdgpu_ps void @fcmp_x2(float %a) #0 { 219 %ogt = fcmp nsz ogt float %a, 2.500000e-01 220 %k = select i1 %ogt, float -1.000000e+00, float 0.000000e+00 221 %c = fcmp nsz oge float %k, 0.000000e+00 222 call void @llvm.amdgcn.kill(i1 %c) #1 223 ret void 224} 225 226; SI-LABEL: {{^}}wqm: 227; SI: v_cmp_neq_f32_e32 vcc, 0 228; SI: s_wqm_b64 s[0:1], vcc 229; SI: s_and_b64 exec, exec, s[0:1] 230define amdgpu_ps void @wqm(float %a) { 231 %c1 = fcmp une float %a, 0.0 232 %c2 = call i1 @llvm.amdgcn.wqm.vote(i1 %c1) 233 call void @llvm.amdgcn.kill(i1 %c2) 234 ret void 235} 236 237; This checks that we use the 64-bit encoding when the operand is a SGPR. 238; SI-LABEL: {{^}}test_sgpr: 239; SI: v_cmpx_ge_f32_e64 240define amdgpu_ps void @test_sgpr(float inreg %a) #0 { 241 %c = fcmp ole float %a, 1.000000e+00 242 call void @llvm.amdgcn.kill(i1 %c) #1 243 ret void 244} 245 246; SI-LABEL: {{^}}test_non_inline_imm_sgpr: 247; SI-NOT: v_cmpx_ge_f32_e64 248define amdgpu_ps void @test_non_inline_imm_sgpr(float inreg %a) #0 { 249 %c = fcmp ole float %a, 1.500000e+00 250 call void @llvm.amdgcn.kill(i1 %c) #1 251 ret void 252} 253 254; SI-LABEL: {{^}}test_scc_liveness: 255; SI: v_cmp 256; SI: s_and_b64 exec 257; SI: s_cmp 258; SI: s_cbranch_scc 259define amdgpu_ps void @test_scc_liveness() #0 { 260main_body: 261 br label %loop3 262 263loop3: ; preds = %loop3, %main_body 264 %tmp = phi i32 [ 0, %main_body ], [ %tmp5, %loop3 ] 265 %tmp1 = icmp sgt i32 %tmp, 0 266 call void @llvm.amdgcn.kill(i1 %tmp1) #1 267 %tmp5 = add i32 %tmp, 1 268 br i1 %tmp1, label %endloop15, label %loop3 269 270endloop15: ; preds = %loop3 271 ret void 272} 273 274declare void @llvm.amdgcn.kill(i1) #0 275declare void @llvm.amdgcn.exp.f32(i32, i32, float, float, float, float, i1, i1) #0 276declare i1 @llvm.amdgcn.wqm.vote(i1) 277 278attributes #0 = { nounwind } 279