1 ; RUN: llc -march=r600 -mcpu=cypress < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s 2 ; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs< %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 3 4 ; FUNC-LABEL: @anyext_load_i8: 5 ; EG: AND_INT 6 ; EG: 255 7 define void @anyext_load_i8(i8 addrspace(1)* nocapture noalias %out, i8 addrspace(1)* nocapture noalias %src) nounwind { 8 %cast = bitcast i8 addrspace(1)* %src to i32 addrspace(1)* 9 %load = load i32 addrspace(1)* %cast, align 1 10 %x = bitcast i32 %load to <4 x i8> 11 %castOut = bitcast i8 addrspace(1)* %out to <4 x i8> addrspace(1)* 12 store <4 x i8> %x, <4 x i8> addrspace(1)* %castOut, align 1 13 ret void 14 } 15 16 ; FUNC-LABEL: @anyext_load_i16: 17 ; EG: AND_INT 18 ; EG: AND_INT 19 ; EG-DAG: 65535 20 ; EG-DAG: -65536 21 define void @anyext_load_i16(i16 addrspace(1)* nocapture noalias %out, i16 addrspace(1)* nocapture noalias %src) nounwind { 22 %cast = bitcast i16 addrspace(1)* %src to i32 addrspace(1)* 23 %load = load i32 addrspace(1)* %cast, align 1 24 %x = bitcast i32 %load to <2 x i16> 25 %castOut = bitcast i16 addrspace(1)* %out to <2 x i16> addrspace(1)* 26 store <2 x i16> %x, <2 x i16> addrspace(1)* %castOut, align 1 27 ret void 28 } 29 30 ; FUNC-LABEL: @anyext_load_lds_i8: 31 ; EG: AND_INT 32 ; EG: 255 33 define void @anyext_load_lds_i8(i8 addrspace(3)* nocapture noalias %out, i8 addrspace(3)* nocapture noalias %src) nounwind { 34 %cast = bitcast i8 addrspace(3)* %src to i32 addrspace(3)* 35 %load = load i32 addrspace(3)* %cast, align 1 36 %x = bitcast i32 %load to <4 x i8> 37 %castOut = bitcast i8 addrspace(3)* %out to <4 x i8> addrspace(3)* 38 store <4 x i8> %x, <4 x i8> addrspace(3)* %castOut, align 1 39 ret void 40 } 41 42 ; FUNC-LABEL: @anyext_load_lds_i16: 43 ; EG: AND_INT 44 ; EG: AND_INT 45 ; EG-DAG: 65535 46 ; EG-DAG: -65536 47 define void @anyext_load_lds_i16(i16 addrspace(3)* nocapture noalias %out, i16 addrspace(3)* nocapture noalias %src) nounwind { 48 %cast = bitcast i16 addrspace(3)* %src to i32 addrspace(3)* 49 %load = load i32 addrspace(3)* %cast, align 1 50 %x = bitcast i32 %load to <2 x i16> 51 %castOut = bitcast i16 addrspace(3)* %out to <2 x i16> addrspace(3)* 52 store <2 x i16> %x, <2 x i16> addrspace(3)* %castOut, align 1 53 ret void 54 } 55 56 ; FUNC-LABEL: @sextload_global_i8_to_i64 57 ; SI: BUFFER_LOAD_SBYTE [[LOAD:v[0-9]+]], 58 ; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 31, [[LOAD]] 59 ; SI: BUFFER_STORE_DWORDX2 60 define void @sextload_global_i8_to_i64(i64 addrspace(1)* %out, i8 addrspace(1)* %in) nounwind { 61 %a = load i8 addrspace(1)* %in, align 8 62 %ext = sext i8 %a to i64 63 store i64 %ext, i64 addrspace(1)* %out, align 8 64 ret void 65 } 66 67 ; FUNC-LABEL: @sextload_global_i16_to_i64 68 ; SI: BUFFER_LOAD_SSHORT [[LOAD:v[0-9]+]], 69 ; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 31, [[LOAD]] 70 ; SI: BUFFER_STORE_DWORDX2 71 define void @sextload_global_i16_to_i64(i64 addrspace(1)* %out, i16 addrspace(1)* %in) nounwind { 72 %a = load i16 addrspace(1)* %in, align 8 73 %ext = sext i16 %a to i64 74 store i64 %ext, i64 addrspace(1)* %out, align 8 75 ret void 76 } 77 78 ; FUNC-LABEL: @sextload_global_i32_to_i64 79 ; SI: BUFFER_LOAD_DWORD [[LOAD:v[0-9]+]], 80 ; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 31, [[LOAD]] 81 ; SI: BUFFER_STORE_DWORDX2 82 define void @sextload_global_i32_to_i64(i64 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind { 83 %a = load i32 addrspace(1)* %in, align 8 84 %ext = sext i32 %a to i64 85 store i64 %ext, i64 addrspace(1)* %out, align 8 86 ret void 87 } 88 89 ; FUNC-LABEL: @zextload_global_i8_to_i64 90 ; SI: S_MOV_B32 [[ZERO:s[0-9]+]], 0 91 ; SI: BUFFER_LOAD_UBYTE [[LOAD:v[0-9]+]], 92 ; SI: V_MOV_B32_e32 {{v[0-9]+}}, [[ZERO]] 93 ; SI: BUFFER_STORE_DWORDX2 94 define void @zextload_global_i8_to_i64(i64 addrspace(1)* %out, i8 addrspace(1)* %in) nounwind { 95 %a = load i8 addrspace(1)* %in, align 8 96 %ext = zext i8 %a to i64 97 store i64 %ext, i64 addrspace(1)* %out, align 8 98 ret void 99 } 100 101 ; FUNC-LABEL: @zextload_global_i16_to_i64 102 ; SI: S_MOV_B32 [[ZERO:s[0-9]+]], 0 103 ; SI: BUFFER_LOAD_USHORT [[LOAD:v[0-9]+]], 104 ; SI: V_MOV_B32_e32 {{v[0-9]+}}, [[ZERO]] 105 ; SI: BUFFER_STORE_DWORDX2 106 define void @zextload_global_i16_to_i64(i64 addrspace(1)* %out, i16 addrspace(1)* %in) nounwind { 107 %a = load i16 addrspace(1)* %in, align 8 108 %ext = zext i16 %a to i64 109 store i64 %ext, i64 addrspace(1)* %out, align 8 110 ret void 111 } 112 113 ; FUNC-LABEL: @zextload_global_i32_to_i64 114 ; SI: S_MOV_B32 [[ZERO:s[0-9]+]], 0 115 ; SI: BUFFER_LOAD_DWORD [[LOAD:v[0-9]+]], 116 ; SI: V_MOV_B32_e32 {{v[0-9]+}}, [[ZERO]] 117 ; SI: BUFFER_STORE_DWORDX2 118 define void @zextload_global_i32_to_i64(i64 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind { 119 %a = load i32 addrspace(1)* %in, align 8 120 %ext = zext i32 %a to i64 121 store i64 %ext, i64 addrspace(1)* %out, align 8 122 ret void 123 } 124
