1 ; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs< %s | FileCheck -check-prefix=SI %s 2 3 ; SI-LABEL: @local_i32_load 4 ; SI: DS_READ_B32 [[REG:v[0-9]+]], v{{[0-9]+}}, 0x1c, [M0] 5 ; SI: BUFFER_STORE_DWORD [[REG]], 6 define void @local_i32_load(i32 addrspace(1)* %out, i32 addrspace(3)* %in) nounwind { 7 %gep = getelementptr i32 addrspace(3)* %in, i32 7 8 %val = load i32 addrspace(3)* %gep, align 4 9 store i32 %val, i32 addrspace(1)* %out, align 4 10 ret void 11 } 12 13 ; SI-LABEL: @local_i32_load_0_offset 14 ; SI: DS_READ_B32 [[REG:v[0-9]+]], v{{[0-9]+}}, 0x0, [M0] 15 ; SI: BUFFER_STORE_DWORD [[REG]], 16 define void @local_i32_load_0_offset(i32 addrspace(1)* %out, i32 addrspace(3)* %in) nounwind { 17 %val = load i32 addrspace(3)* %in, align 4 18 store i32 %val, i32 addrspace(1)* %out, align 4 19 ret void 20 } 21 22 ; SI-LABEL: @local_i8_load_i16_max_offset 23 ; SI-NOT: ADD 24 ; SI: DS_READ_U8 [[REG:v[0-9]+]], {{v[0-9]+}}, 0xffff, [M0] 25 ; SI: BUFFER_STORE_BYTE [[REG]], 26 define void @local_i8_load_i16_max_offset(i8 addrspace(1)* %out, i8 addrspace(3)* %in) nounwind { 27 %gep = getelementptr i8 addrspace(3)* %in, i32 65535 28 %val = load i8 addrspace(3)* %gep, align 4 29 store i8 %val, i8 addrspace(1)* %out, align 4 30 ret void 31 } 32 33 ; SI-LABEL: @local_i8_load_over_i16_max_offset 34 ; SI: S_ADD_I32 [[ADDR:s[0-9]+]], s{{[0-9]+}}, 0x10000 35 ; SI: V_MOV_B32_e32 [[VREGADDR:v[0-9]+]], [[ADDR]] 36 ; SI: DS_READ_U8 [[REG:v[0-9]+]], [[VREGADDR]], 0x0, [M0] 37 ; SI: BUFFER_STORE_BYTE [[REG]], 38 define void @local_i8_load_over_i16_max_offset(i8 addrspace(1)* %out, i8 addrspace(3)* %in) nounwind { 39 %gep = getelementptr i8 addrspace(3)* %in, i32 65536 40 %val = load i8 addrspace(3)* %gep, align 4 41 store i8 %val, i8 addrspace(1)* %out, align 4 42 ret void 43 } 44 45 ; SI-LABEL: @local_i64_load 46 ; SI-NOT: ADD 47 ; SI: DS_READ_B64 [[REG:v[[0-9]+:[0-9]+]]], v{{[0-9]+}}, 0x38, [M0] 48 ; SI: BUFFER_STORE_DWORDX2 [[REG]], 49 define void @local_i64_load(i64 addrspace(1)* %out, i64 addrspace(3)* %in) nounwind { 50 %gep = getelementptr i64 addrspace(3)* %in, i32 7 51 %val = load i64 addrspace(3)* %gep, align 8 52 store i64 %val, i64 addrspace(1)* %out, align 8 53 ret void 54 } 55 56 ; SI-LABEL: @local_i64_load_0_offset 57 ; SI: DS_READ_B64 [[REG:v\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, 0x0, [M0] 58 ; SI: BUFFER_STORE_DWORDX2 [[REG]], 59 define void @local_i64_load_0_offset(i64 addrspace(1)* %out, i64 addrspace(3)* %in) nounwind { 60 %val = load i64 addrspace(3)* %in, align 8 61 store i64 %val, i64 addrspace(1)* %out, align 8 62 ret void 63 } 64 65 ; SI-LABEL: @local_f64_load 66 ; SI-NOT: ADD 67 ; SI: DS_READ_B64 [[REG:v[[0-9]+:[0-9]+]]], v{{[0-9]+}}, 0x38, [M0] 68 ; SI: BUFFER_STORE_DWORDX2 [[REG]], 69 define void @local_f64_load(double addrspace(1)* %out, double addrspace(3)* %in) nounwind { 70 %gep = getelementptr double addrspace(3)* %in, i32 7 71 %val = load double addrspace(3)* %gep, align 8 72 store double %val, double addrspace(1)* %out, align 8 73 ret void 74 } 75 76 ; SI-LABEL: @local_f64_load_0_offset 77 ; SI: DS_READ_B64 [[REG:v\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, 0x0, [M0] 78 ; SI: BUFFER_STORE_DWORDX2 [[REG]], 79 define void @local_f64_load_0_offset(double addrspace(1)* %out, double addrspace(3)* %in) nounwind { 80 %val = load double addrspace(3)* %in, align 8 81 store double %val, double addrspace(1)* %out, align 8 82 ret void 83 } 84 85 ; SI-LABEL: @local_i64_store 86 ; SI-NOT: ADD 87 ; SI: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x38 [M0] 88 define void @local_i64_store(i64 addrspace(3)* %out) nounwind { 89 %gep = getelementptr i64 addrspace(3)* %out, i32 7 90 store i64 5678, i64 addrspace(3)* %gep, align 8 91 ret void 92 } 93 94 ; SI-LABEL: @local_i64_store_0_offset 95 ; SI-NOT: ADD 96 ; SI: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x0 [M0] 97 define void @local_i64_store_0_offset(i64 addrspace(3)* %out) nounwind { 98 store i64 1234, i64 addrspace(3)* %out, align 8 99 ret void 100 } 101 102 ; SI-LABEL: @local_f64_store 103 ; SI-NOT: ADD 104 ; SI: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x38 [M0] 105 define void @local_f64_store(double addrspace(3)* %out) nounwind { 106 %gep = getelementptr double addrspace(3)* %out, i32 7 107 store double 16.0, double addrspace(3)* %gep, align 8 108 ret void 109 } 110 111 ; SI-LABEL: @local_f64_store_0_offset 112 ; SI: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x0 [M0] 113 define void @local_f64_store_0_offset(double addrspace(3)* %out) nounwind { 114 store double 20.0, double addrspace(3)* %out, align 8 115 ret void 116 } 117 118 ; SI-LABEL: @local_v2i64_store 119 ; SI-NOT: ADD 120 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x78 [M0] 121 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x70 [M0] 122 define void @local_v2i64_store(<2 x i64> addrspace(3)* %out) nounwind { 123 %gep = getelementptr <2 x i64> addrspace(3)* %out, i32 7 124 store <2 x i64> <i64 5678, i64 5678>, <2 x i64> addrspace(3)* %gep, align 16 125 ret void 126 } 127 128 ; SI-LABEL: @local_v2i64_store_0_offset 129 ; SI-NOT: ADD 130 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x8 [M0] 131 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x0 [M0] 132 define void @local_v2i64_store_0_offset(<2 x i64> addrspace(3)* %out) nounwind { 133 store <2 x i64> <i64 1234, i64 1234>, <2 x i64> addrspace(3)* %out, align 16 134 ret void 135 } 136 137 ; SI-LABEL: @local_v4i64_store 138 ; SI-NOT: ADD 139 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0xf8 [M0] 140 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0xf0 [M0] 141 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0xe8 [M0] 142 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0xe0 [M0] 143 define void @local_v4i64_store(<4 x i64> addrspace(3)* %out) nounwind { 144 %gep = getelementptr <4 x i64> addrspace(3)* %out, i32 7 145 store <4 x i64> <i64 5678, i64 5678, i64 5678, i64 5678>, <4 x i64> addrspace(3)* %gep, align 16 146 ret void 147 } 148 149 ; SI-LABEL: @local_v4i64_store_0_offset 150 ; SI-NOT: ADD 151 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x18 [M0] 152 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x10 [M0] 153 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x8 [M0] 154 ; SI-DAG: DS_WRITE_B64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, 0x0 [M0] 155 define void @local_v4i64_store_0_offset(<4 x i64> addrspace(3)* %out) nounwind { 156 store <4 x i64> <i64 1234, i64 1234, i64 1234, i64 1234>, <4 x i64> addrspace(3)* %out, align 16 157 ret void 158 } 159
