1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN %s 2; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GCN %s 3 4declare float @llvm.fma.f32(float, float, float) #1 5declare double @llvm.fma.f64(double, double, double) #1 6declare float @llvm.fmuladd.f32(float, float, float) #1 7declare float @llvm.amdgcn.div.fixup.f32(float, float, float) #1 8 9 10; GCN-LABEL: {{^}}test_sgpr_use_twice_binop: 11; GCN: s_load_dword [[SGPR:s[0-9]+]], 12; GCN: v_add_f32_e64 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]] 13; GCN: buffer_store_dword [[RESULT]] 14define amdgpu_kernel void @test_sgpr_use_twice_binop(float addrspace(1)* %out, float %a) #0 { 15 %dbl = fadd float %a, %a 16 store float %dbl, float addrspace(1)* %out, align 4 17 ret void 18} 19 20; GCN-LABEL: {{^}}test_sgpr_use_three_ternary_op: 21; GCN: s_load_dword [[SGPR:s[0-9]+]], 22; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], [[SGPR]] 23; GCN: buffer_store_dword [[RESULT]] 24define amdgpu_kernel void @test_sgpr_use_three_ternary_op(float addrspace(1)* %out, float %a) #0 { 25 %fma = call float @llvm.fma.f32(float %a, float %a, float %a) #1 26 store float %fma, float addrspace(1)* %out, align 4 27 ret void 28} 29 30; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_b: 31; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 32; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 33; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]] 34; GCN: v_fma_f32 [[RESULT:v[0-9]+]], s[[SGPR0]], s[[SGPR0]], [[VGPR1]] 35; GCN: buffer_store_dword [[RESULT]] 36define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_b(float addrspace(1)* %out, float %a, float %b) #0 { 37 %fma = call float @llvm.fma.f32(float %a, float %a, float %b) #1 38 store float %fma, float addrspace(1)* %out, align 4 39 ret void 40} 41 42; GCN-LABEL: {{^}}test_use_s_v_s: 43; GCN-DAG: s_load_dwordx2 s{{\[}}[[SA:[0-9]+]]:[[SB:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}} 44; SI: buffer_load_dword [[VA0:v[0-9]+]] 45; SI-NEXT: buffer_load_dword [[VA1:v[0-9]+]] 46 47; GCN-NOT: v_mov_b32 48 49; VI: buffer_load_dword [[VA0:v[0-9]+]] 50; VI-NEXT: buffer_load_dword [[VA1:v[0-9]+]] 51 52; GCN-NOT: v_mov_b32 53; GCN: v_mov_b32_e32 [[VB:v[0-9]+]], s[[SB]] 54; GCN-NOT: v_mov_b32 55 56; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[SA]], [[VA0]], [[VB]] 57; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[SA]], [[VA1]], [[VB]] 58; GCN: buffer_store_dword [[RESULT0]] 59; GCN: buffer_store_dword [[RESULT1]] 60define amdgpu_kernel void @test_use_s_v_s(float addrspace(1)* %out, float %a, float %b, float addrspace(1)* %in) #0 { 61 %va0 = load volatile float, float addrspace(1)* %in 62 %va1 = load volatile float, float addrspace(1)* %in 63 %fma0 = call float @llvm.fma.f32(float %a, float %va0, float %b) #1 64 %fma1 = call float @llvm.fma.f32(float %a, float %va1, float %b) #1 65 store volatile float %fma0, float addrspace(1)* %out 66 store volatile float %fma1, float addrspace(1)* %out 67 ret void 68} 69 70; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_b_a: 71; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 72; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 73; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]] 74; GCN: v_fma_f32 [[RESULT:v[0-9]+]], s[[SGPR0]], [[VGPR1]], s[[SGPR0]] 75; GCN: buffer_store_dword [[RESULT]] 76define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_b_a(float addrspace(1)* %out, float %a, float %b) #0 { 77 %fma = call float @llvm.fma.f32(float %a, float %b, float %a) #1 78 store float %fma, float addrspace(1)* %out, align 4 79 ret void 80} 81 82; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_b_a_a: 83; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 84; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 85; GCN: v_mov_b32_e32 [[VGPR1:v[0-9]+]], s[[SGPR1]] 86; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[VGPR1]], s[[SGPR0]], s[[SGPR0]] 87; GCN: buffer_store_dword [[RESULT]] 88define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_b_a_a(float addrspace(1)* %out, float %a, float %b) #0 { 89 %fma = call float @llvm.fma.f32(float %b, float %a, float %a) #1 90 store float %fma, float addrspace(1)* %out, align 4 91 ret void 92} 93 94; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_imm: 95; GCN: s_load_dword [[SGPR:s[0-9]+]] 96; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], 2.0 97; GCN: buffer_store_dword [[RESULT]] 98define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_imm(float addrspace(1)* %out, float %a) #0 { 99 %fma = call float @llvm.fma.f32(float %a, float %a, float 2.0) #1 100 store float %fma, float addrspace(1)* %out, align 4 101 ret void 102} 103 104; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_imm_a: 105; GCN: s_load_dword [[SGPR:s[0-9]+]] 106; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], 2.0, [[SGPR]] 107; GCN: buffer_store_dword [[RESULT]] 108define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_imm_a(float addrspace(1)* %out, float %a) #0 { 109 %fma = call float @llvm.fma.f32(float %a, float 2.0, float %a) #1 110 store float %fma, float addrspace(1)* %out, align 4 111 ret void 112} 113 114; Don't use fma since fma c, x, y is canonicalized to fma x, c, y 115; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_imm_a_a: 116; GCN: s_load_dword [[SGPR:s[0-9]+]] 117; GCN: v_div_fixup_f32 [[RESULT:v[0-9]+]], 2.0, [[SGPR]], [[SGPR]] 118; GCN: buffer_store_dword [[RESULT]] 119define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_imm_a_a(float addrspace(1)* %out, float %a) #0 { 120 %val = call float @llvm.amdgcn.div.fixup.f32(float 2.0, float %a, float %a) #1 121 store float %val, float addrspace(1)* %out, align 4 122 ret void 123} 124 125; GCN-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_a_kimm: 126; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]] 127; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 128; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], [[VK]] 129; GCN: buffer_store_dword [[RESULT]] 130define amdgpu_kernel void @test_sgpr_use_twice_ternary_op_a_a_kimm(float addrspace(1)* %out, float %a) #0 { 131 %fma = call float @llvm.fma.f32(float %a, float %a, float 1024.0) #1 132 store float %fma, float addrspace(1)* %out, align 4 133 ret void 134} 135 136; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_k_s: 137; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]] 138; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 139; GCN: v_fma_f32 [[RESULT0:v[0-9]+]], [[VK]], [[VK]], [[SGPR]] 140; GCN: buffer_store_dword [[RESULT0]] 141define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_k_s(float addrspace(1)* %out, float %a) #0 { 142 %fma = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %a) #1 143 store float %fma, float addrspace(1)* %out 144 ret void 145} 146 147; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_k_s_x2: 148; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 149; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 150; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 151; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], [[VK]], [[VK]], s[[SGPR0]] 152; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], [[VK]], [[VK]], s[[SGPR1]] 153; GCN: buffer_store_dword [[RESULT0]] 154; GCN: buffer_store_dword [[RESULT1]] 155; GCN: s_endpgm 156define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_k_s_x2(float addrspace(1)* %out, float %a, float %b) #0 { 157 %fma0 = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %a) #1 158 %fma1 = call float @llvm.fma.f32(float 1024.0, float 1024.0, float %b) #1 159 store volatile float %fma0, float addrspace(1)* %out 160 store volatile float %fma1, float addrspace(1)* %out 161 ret void 162} 163 164; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_s_k: 165; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]] 166; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 167; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[VK]], [[VK]] 168; GCN: buffer_store_dword [[RESULT]] 169define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_s_k(float addrspace(1)* %out, float %a) #0 { 170 %fma = call float @llvm.fma.f32(float 1024.0, float %a, float 1024.0) #1 171 store float %fma, float addrspace(1)* %out 172 ret void 173} 174 175; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_k_s_k_x2: 176; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 177; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 178; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 179; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[SGPR0]], [[VK]], [[VK]] 180; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[SGPR1]], [[VK]], [[VK]] 181; GCN: buffer_store_dword [[RESULT0]] 182; GCN: buffer_store_dword [[RESULT1]] 183; GCN: s_endpgm 184define amdgpu_kernel void @test_literal_use_twice_ternary_op_k_s_k_x2(float addrspace(1)* %out, float %a, float %b) #0 { 185 %fma0 = call float @llvm.fma.f32(float 1024.0, float %a, float 1024.0) #1 186 %fma1 = call float @llvm.fma.f32(float 1024.0, float %b, float 1024.0) #1 187 store volatile float %fma0, float addrspace(1)* %out 188 store volatile float %fma1, float addrspace(1)* %out 189 ret void 190} 191 192; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_s_k_k: 193; GCN-DAG: s_load_dword [[SGPR:s[0-9]+]] 194; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 195; GCN: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], [[VK]], [[VK]] 196; GCN: buffer_store_dword [[RESULT]] 197define amdgpu_kernel void @test_literal_use_twice_ternary_op_s_k_k(float addrspace(1)* %out, float %a) #0 { 198 %fma = call float @llvm.fma.f32(float %a, float 1024.0, float 1024.0) #1 199 store float %fma, float addrspace(1)* %out 200 ret void 201} 202 203; GCN-LABEL: {{^}}test_literal_use_twice_ternary_op_s_k_k_x2: 204; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 205; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 206; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000 207; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[SGPR0]], [[VK]], [[VK]] 208; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[SGPR1]], [[VK]], [[VK]] 209; GCN: buffer_store_dword [[RESULT0]] 210; GCN: buffer_store_dword [[RESULT1]] 211; GCN: s_endpgm 212define amdgpu_kernel void @test_literal_use_twice_ternary_op_s_k_k_x2(float addrspace(1)* %out, float %a, float %b) #0 { 213 %fma0 = call float @llvm.fma.f32(float %a, float 1024.0, float 1024.0) #1 214 %fma1 = call float @llvm.fma.f32(float %b, float 1024.0, float 1024.0) #1 215 store volatile float %fma0, float addrspace(1)* %out 216 store volatile float %fma1, float addrspace(1)* %out 217 ret void 218} 219 220; GCN-LABEL: {{^}}test_s0_s1_k_f32: 221; SI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb 222; VI-DAG: s_load_dwordx2 s{{\[}}[[SGPR0:[0-9]+]]:[[SGPR1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x2c 223; GCN-DAG: v_mov_b32_e32 [[VK0:v[0-9]+]], 0x44800000 224; GCN-DAG: v_mov_b32_e32 [[VS1:v[0-9]+]], s[[SGPR1]] 225 226; GCN-DAG: v_fma_f32 [[RESULT0:v[0-9]+]], s[[SGPR0]], [[VS1]], [[VK0]] 227; GCN-DAG: v_mov_b32_e32 [[VK1:v[0-9]+]], 0x45800000 228; GCN-DAG: v_fma_f32 [[RESULT1:v[0-9]+]], s[[SGPR0]], [[VS1]], [[VK1]] 229 230; GCN: buffer_store_dword [[RESULT0]] 231; GCN: buffer_store_dword [[RESULT1]] 232define amdgpu_kernel void @test_s0_s1_k_f32(float addrspace(1)* %out, float %a, float %b) #0 { 233 %fma0 = call float @llvm.fma.f32(float %a, float %b, float 1024.0) #1 234 %fma1 = call float @llvm.fma.f32(float %a, float %b, float 4096.0) #1 235 store volatile float %fma0, float addrspace(1)* %out 236 store volatile float %fma1, float addrspace(1)* %out 237 ret void 238} 239 240; FIXME: Immediate in SGPRs just copied to VGPRs 241; GCN-LABEL: {{^}}test_s0_s1_k_f64: 242; GCN-DAG: s_load_dwordx2 [[SGPR0:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, {{0x13|0x4c}} 243; GCN-DAG: s_load_dwordx2 s{{\[}}[[SGPR1_SUB0:[0-9]+]]:[[SGPR1_SUB1:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0x1d|0x74}} 244; GCN-DAG: v_mov_b32_e32 v[[VK0_SUB1:[0-9]+]], 0x40900000 245; GCN-DAG: v_mov_b32_e32 v[[VZERO:[0-9]+]], 0{{$}} 246 247; GCN-DAG: v_mov_b32_e32 v[[VS1_SUB0:[0-9]+]], s[[SGPR1_SUB0]] 248; GCN-DAG: v_mov_b32_e32 v[[VS1_SUB1:[0-9]+]], s[[SGPR1_SUB1]] 249; GCN: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[SGPR0]], v{{\[}}[[VS1_SUB0]]:[[VS1_SUB1]]{{\]}}, v{{\[}}[[VZERO]]:[[VK0_SUB1]]{{\]}} 250 251; Same zero component is re-used for half of each immediate. 252; GCN: v_mov_b32_e32 v[[VK1_SUB1:[0-9]+]], 0x40b00000 253; GCN: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[SGPR0]], v{{\[}}[[VS1_SUB0]]:[[VS1_SUB1]]{{\]}}, v{{\[}}[[VZERO]]:[[VK1_SUB1]]{{\]}} 254 255; GCN: buffer_store_dwordx2 [[RESULT0]] 256; GCN: buffer_store_dwordx2 [[RESULT1]] 257define amdgpu_kernel void @test_s0_s1_k_f64(double addrspace(1)* %out, [8 x i32], double %a, [8 x i32], double %b) #0 { 258 %fma0 = call double @llvm.fma.f64(double %a, double %b, double 1024.0) #1 259 %fma1 = call double @llvm.fma.f64(double %a, double %b, double 4096.0) #1 260 store volatile double %fma0, double addrspace(1)* %out 261 store volatile double %fma1, double addrspace(1)* %out 262 ret void 263} 264 265attributes #0 = { nounwind } 266attributes #1 = { nounwind readnone } 267