1; RUN: opt -S -load-store-vectorizer -mattr=-unaligned-buffer-access,+max-private-element-size-16 < %s | FileCheck -check-prefix=ALIGNED -check-prefix=ALL %s 2; RUN: opt -S -load-store-vectorizer -mattr=+unaligned-buffer-access,+unaligned-scratch-access,+max-private-element-size-16 < %s | FileCheck -check-prefix=UNALIGNED -check-prefix=ALL %s 3; RUN: opt -S -passes='function(load-store-vectorizer)' -mattr=-unaligned-buffer-access,+max-private-element-size-16 < %s | FileCheck -check-prefix=ALIGNED -check-prefix=ALL %s 4; RUN: opt -S -passes='function(load-store-vectorizer)' -mattr=+unaligned-buffer-access,+unaligned-scratch-access,+max-private-element-size-16 < %s | FileCheck -check-prefix=UNALIGNED -check-prefix=ALL %s 5 6target triple = "amdgcn--" 7target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5" 8 9; ALL-LABEL: @load_unknown_offset_align1_i8( 10; ALL: alloca [128 x i8], align 1 11; UNALIGNED: load <2 x i8>, <2 x i8> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 12 13; ALIGNED: load i8, i8 addrspace(5)* %ptr0, align 1{{$}} 14; ALIGNED: load i8, i8 addrspace(5)* %ptr1, align 1{{$}} 15define amdgpu_kernel void @load_unknown_offset_align1_i8(i8 addrspace(1)* noalias %out, i32 %offset) #0 { 16 %alloca = alloca [128 x i8], align 1, addrspace(5) 17 %ptr0 = getelementptr inbounds [128 x i8], [128 x i8] addrspace(5)* %alloca, i32 0, i32 %offset 18 %val0 = load i8, i8 addrspace(5)* %ptr0, align 1 19 %ptr1 = getelementptr inbounds i8, i8 addrspace(5)* %ptr0, i32 1 20 %val1 = load i8, i8 addrspace(5)* %ptr1, align 1 21 %add = add i8 %val0, %val1 22 store i8 %add, i8 addrspace(1)* %out 23 ret void 24} 25 26; ALL-LABEL: @load_unknown_offset_align1_i16( 27; ALL: alloca [128 x i16], align 1, addrspace(5){{$}} 28; UNALIGNED: load <2 x i16>, <2 x i16> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 29 30; ALIGNED: load i16, i16 addrspace(5)* %ptr0, align 1{{$}} 31; ALIGNED: load i16, i16 addrspace(5)* %ptr1, align 1{{$}} 32define amdgpu_kernel void @load_unknown_offset_align1_i16(i16 addrspace(1)* noalias %out, i32 %offset) #0 { 33 %alloca = alloca [128 x i16], align 1, addrspace(5) 34 %ptr0 = getelementptr inbounds [128 x i16], [128 x i16] addrspace(5)* %alloca, i32 0, i32 %offset 35 %val0 = load i16, i16 addrspace(5)* %ptr0, align 1 36 %ptr1 = getelementptr inbounds i16, i16 addrspace(5)* %ptr0, i32 1 37 %val1 = load i16, i16 addrspace(5)* %ptr1, align 1 38 %add = add i16 %val0, %val1 39 store i16 %add, i16 addrspace(1)* %out 40 ret void 41} 42 43; FIXME: Although the offset is unknown here, we know it is a multiple 44; of the element size, so should still be align 4 45 46; ALL-LABEL: @load_unknown_offset_align1_i32( 47; ALL: alloca [128 x i32], align 1 48; UNALIGNED: load <2 x i32>, <2 x i32> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 49 50; ALIGNED: load i32, i32 addrspace(5)* %ptr0, align 1 51; ALIGNED: load i32, i32 addrspace(5)* %ptr1, align 1 52define amdgpu_kernel void @load_unknown_offset_align1_i32(i32 addrspace(1)* noalias %out, i32 %offset) #0 { 53 %alloca = alloca [128 x i32], align 1, addrspace(5) 54 %ptr0 = getelementptr inbounds [128 x i32], [128 x i32] addrspace(5)* %alloca, i32 0, i32 %offset 55 %val0 = load i32, i32 addrspace(5)* %ptr0, align 1 56 %ptr1 = getelementptr inbounds i32, i32 addrspace(5)* %ptr0, i32 1 57 %val1 = load i32, i32 addrspace(5)* %ptr1, align 1 58 %add = add i32 %val0, %val1 59 store i32 %add, i32 addrspace(1)* %out 60 ret void 61} 62 63; FIXME: Should always increase alignment of the load 64; Make sure alloca alignment isn't decreased 65; ALL-LABEL: @load_alloca16_unknown_offset_align1_i32( 66; ALL: alloca [128 x i32], align 16 67 68; UNALIGNED: load <2 x i32>, <2 x i32> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 69; ALIGNED: load <2 x i32>, <2 x i32> addrspace(5)* %{{[0-9]+}}, align 4{{$}} 70define amdgpu_kernel void @load_alloca16_unknown_offset_align1_i32(i32 addrspace(1)* noalias %out, i32 %offset) #0 { 71 %alloca = alloca [128 x i32], align 16, addrspace(5) 72 %ptr0 = getelementptr inbounds [128 x i32], [128 x i32] addrspace(5)* %alloca, i32 0, i32 %offset 73 %val0 = load i32, i32 addrspace(5)* %ptr0, align 1 74 %ptr1 = getelementptr inbounds i32, i32 addrspace(5)* %ptr0, i32 1 75 %val1 = load i32, i32 addrspace(5)* %ptr1, align 1 76 %add = add i32 %val0, %val1 77 store i32 %add, i32 addrspace(1)* %out 78 ret void 79} 80 81; ALL-LABEL: @store_unknown_offset_align1_i8( 82; ALL: alloca [128 x i8], align 1 83; UNALIGNED: store <2 x i8> <i8 9, i8 10>, <2 x i8> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 84 85; ALIGNED: store i8 9, i8 addrspace(5)* %ptr0, align 1{{$}} 86; ALIGNED: store i8 10, i8 addrspace(5)* %ptr1, align 1{{$}} 87define amdgpu_kernel void @store_unknown_offset_align1_i8(i8 addrspace(1)* noalias %out, i32 %offset) #0 { 88 %alloca = alloca [128 x i8], align 1, addrspace(5) 89 %ptr0 = getelementptr inbounds [128 x i8], [128 x i8] addrspace(5)* %alloca, i32 0, i32 %offset 90 store i8 9, i8 addrspace(5)* %ptr0, align 1 91 %ptr1 = getelementptr inbounds i8, i8 addrspace(5)* %ptr0, i32 1 92 store i8 10, i8 addrspace(5)* %ptr1, align 1 93 ret void 94} 95 96; ALL-LABEL: @store_unknown_offset_align1_i16( 97; ALL: alloca [128 x i16], align 1 98; UNALIGNED: store <2 x i16> <i16 9, i16 10>, <2 x i16> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 99 100; ALIGNED: store i16 9, i16 addrspace(5)* %ptr0, align 1{{$}} 101; ALIGNED: store i16 10, i16 addrspace(5)* %ptr1, align 1{{$}} 102define amdgpu_kernel void @store_unknown_offset_align1_i16(i16 addrspace(1)* noalias %out, i32 %offset) #0 { 103 %alloca = alloca [128 x i16], align 1, addrspace(5) 104 %ptr0 = getelementptr inbounds [128 x i16], [128 x i16] addrspace(5)* %alloca, i32 0, i32 %offset 105 store i16 9, i16 addrspace(5)* %ptr0, align 1 106 %ptr1 = getelementptr inbounds i16, i16 addrspace(5)* %ptr0, i32 1 107 store i16 10, i16 addrspace(5)* %ptr1, align 1 108 ret void 109} 110 111; FIXME: Although the offset is unknown here, we know it is a multiple 112; of the element size, so it still should be align 4. 113 114; ALL-LABEL: @store_unknown_offset_align1_i32( 115; ALL: alloca [128 x i32], align 1 116 117; UNALIGNED: store <2 x i32> <i32 9, i32 10>, <2 x i32> addrspace(5)* %{{[0-9]+}}, align 1{{$}} 118 119; ALIGNED: store i32 9, i32 addrspace(5)* %ptr0, align 1 120; ALIGNED: store i32 10, i32 addrspace(5)* %ptr1, align 1 121define amdgpu_kernel void @store_unknown_offset_align1_i32(i32 addrspace(1)* noalias %out, i32 %offset) #0 { 122 %alloca = alloca [128 x i32], align 1, addrspace(5) 123 %ptr0 = getelementptr inbounds [128 x i32], [128 x i32] addrspace(5)* %alloca, i32 0, i32 %offset 124 store i32 9, i32 addrspace(5)* %ptr0, align 1 125 %ptr1 = getelementptr inbounds i32, i32 addrspace(5)* %ptr0, i32 1 126 store i32 10, i32 addrspace(5)* %ptr1, align 1 127 ret void 128} 129 130; ALL-LABEL: @merge_private_store_4_vector_elts_loads_v4i32( 131; ALIGNED: %alloca = alloca [8 x i32], align 4, addrspace(5) 132; ALIGNED: store <4 x i32> <i32 9, i32 1, i32 23, i32 19>, <4 x i32> addrspace(5)* %1, align 4 133 134; UNALIGNED: %alloca = alloca [8 x i32], align 1, addrspace(5) 135; UNALIGNED: store <4 x i32> <i32 9, i32 1, i32 23, i32 19>, <4 x i32> addrspace(5)* %1, align 1 136define amdgpu_kernel void @merge_private_store_4_vector_elts_loads_v4i32() { 137 %alloca = alloca [8 x i32], align 1, addrspace(5) 138 %out = bitcast [8 x i32] addrspace(5)* %alloca to i32 addrspace(5)* 139 %out.gep.1 = getelementptr i32, i32 addrspace(5)* %out, i32 1 140 %out.gep.2 = getelementptr i32, i32 addrspace(5)* %out, i32 2 141 %out.gep.3 = getelementptr i32, i32 addrspace(5)* %out, i32 3 142 143 store i32 9, i32 addrspace(5)* %out, align 1 144 store i32 1, i32 addrspace(5)* %out.gep.1, align 1 145 store i32 23, i32 addrspace(5)* %out.gep.2, align 1 146 store i32 19, i32 addrspace(5)* %out.gep.3, align 1 147 ret void 148} 149 150; ALL-LABEL: @merge_private_store_4_vector_elts_loads_v4i8( 151; ALIGNED: %alloca = alloca [8 x i8], align 4, addrspace(5) 152; ALIGNED: store <4 x i8> <i8 9, i8 1, i8 23, i8 19>, <4 x i8> addrspace(5)* %1, align 4 153 154; UNALIGNED: %alloca = alloca [8 x i8], align 1, addrspace(5) 155; UNALIGNED: store <4 x i8> <i8 9, i8 1, i8 23, i8 19>, <4 x i8> addrspace(5)* %1, align 1 156define amdgpu_kernel void @merge_private_store_4_vector_elts_loads_v4i8() { 157 %alloca = alloca [8 x i8], align 1, addrspace(5) 158 %out = bitcast [8 x i8] addrspace(5)* %alloca to i8 addrspace(5)* 159 %out.gep.1 = getelementptr i8, i8 addrspace(5)* %out, i8 1 160 %out.gep.2 = getelementptr i8, i8 addrspace(5)* %out, i8 2 161 %out.gep.3 = getelementptr i8, i8 addrspace(5)* %out, i8 3 162 163 store i8 9, i8 addrspace(5)* %out, align 1 164 store i8 1, i8 addrspace(5)* %out.gep.1, align 1 165 store i8 23, i8 addrspace(5)* %out.gep.2, align 1 166 store i8 19, i8 addrspace(5)* %out.gep.3, align 1 167 ret void 168} 169 170; ALL-LABEL: @merge_private_load_4_vector_elts_loads_v4i32( 171; ALIGNED: %alloca = alloca [8 x i32], align 4, addrspace(5) 172; ALIGNED: load <4 x i32>, <4 x i32> addrspace(5)* %1, align 4 173 174; UNALIGNED: %alloca = alloca [8 x i32], align 1, addrspace(5) 175; UNALIGNED: load <4 x i32>, <4 x i32> addrspace(5)* %1, align 1 176define amdgpu_kernel void @merge_private_load_4_vector_elts_loads_v4i32() { 177 %alloca = alloca [8 x i32], align 1, addrspace(5) 178 %out = bitcast [8 x i32] addrspace(5)* %alloca to i32 addrspace(5)* 179 %out.gep.1 = getelementptr i32, i32 addrspace(5)* %out, i32 1 180 %out.gep.2 = getelementptr i32, i32 addrspace(5)* %out, i32 2 181 %out.gep.3 = getelementptr i32, i32 addrspace(5)* %out, i32 3 182 183 %load0 = load i32, i32 addrspace(5)* %out, align 1 184 %load1 = load i32, i32 addrspace(5)* %out.gep.1, align 1 185 %load2 = load i32, i32 addrspace(5)* %out.gep.2, align 1 186 %load3 = load i32, i32 addrspace(5)* %out.gep.3, align 1 187 ret void 188} 189 190; ALL-LABEL: @merge_private_load_4_vector_elts_loads_v4i8( 191; ALIGNED: %alloca = alloca [8 x i8], align 4, addrspace(5) 192; ALIGNED: load <4 x i8>, <4 x i8> addrspace(5)* %1, align 4 193 194; UNALIGNED: %alloca = alloca [8 x i8], align 1, addrspace(5) 195; UNALIGNED: load <4 x i8>, <4 x i8> addrspace(5)* %1, align 1 196define amdgpu_kernel void @merge_private_load_4_vector_elts_loads_v4i8() { 197 %alloca = alloca [8 x i8], align 1, addrspace(5) 198 %out = bitcast [8 x i8] addrspace(5)* %alloca to i8 addrspace(5)* 199 %out.gep.1 = getelementptr i8, i8 addrspace(5)* %out, i8 1 200 %out.gep.2 = getelementptr i8, i8 addrspace(5)* %out, i8 2 201 %out.gep.3 = getelementptr i8, i8 addrspace(5)* %out, i8 3 202 203 %load0 = load i8, i8 addrspace(5)* %out, align 1 204 %load1 = load i8, i8 addrspace(5)* %out.gep.1, align 1 205 %load2 = load i8, i8 addrspace(5)* %out.gep.2, align 1 206 %load3 = load i8, i8 addrspace(5)* %out.gep.3, align 1 207 ret void 208} 209 210attributes #0 = { nounwind } 211