1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=-unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,GFX7-ALIGNED %s 3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=+unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,GFX7-UNALIGNED %s 4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,GFX9 %s 5 6; Should not merge this to a dword load 7define i32 @global_load_2xi16_align2(i16 addrspace(1)* %p) #0 { 8; GFX7-ALIGNED-LABEL: global_load_2xi16_align2: 9; GFX7-ALIGNED: ; %bb.0: 10; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 11; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v0 12; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 13; GFX7-ALIGNED-NEXT: flat_load_ushort v0, v[0:1] 14; GFX7-ALIGNED-NEXT: flat_load_ushort v1, v[2:3] 15; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 16; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1 17; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 18; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 19; 20; GFX7-UNALIGNED-LABEL: global_load_2xi16_align2: 21; GFX7-UNALIGNED: ; %bb.0: 22; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 23; GFX7-UNALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v0 24; GFX7-UNALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 25; GFX7-UNALIGNED-NEXT: flat_load_ushort v0, v[0:1] 26; GFX7-UNALIGNED-NEXT: flat_load_ushort v1, v[2:3] 27; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 28; GFX7-UNALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1 29; GFX7-UNALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 30; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 31; 32; GFX9-LABEL: global_load_2xi16_align2: 33; GFX9: ; %bb.0: 34; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 35; GFX9-NEXT: global_load_ushort v2, v[0:1], off 36; GFX9-NEXT: global_load_ushort v0, v[0:1], off offset:2 37; GFX9-NEXT: s_waitcnt vmcnt(0) 38; GFX9-NEXT: v_lshl_or_b32 v0, v0, 16, v2 39; GFX9-NEXT: s_setpc_b64 s[30:31] 40 %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1 41 %p.0 = load i16, i16 addrspace(1)* %p, align 2 42 %p.1 = load i16, i16 addrspace(1)* %gep.p, align 2 43 %zext.0 = zext i16 %p.0 to i32 44 %zext.1 = zext i16 %p.1 to i32 45 %shl.1 = shl i32 %zext.1, 16 46 %or = or i32 %zext.0, %shl.1 47 ret i32 %or 48} 49 50; Should not merge this to a dword store 51define amdgpu_kernel void @global_store_2xi16_align2(i16 addrspace(1)* %p, i16 addrspace(1)* %r) #0 { 52; GFX7-ALIGNED-LABEL: global_store_2xi16_align2: 53; GFX7-ALIGNED: ; %bb.0: 54; GFX7-ALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 55; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 1 56; GFX7-ALIGNED-NEXT: s_waitcnt lgkmcnt(0) 57; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 58; GFX7-ALIGNED-NEXT: s_add_u32 s2, s0, 2 59; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 60; GFX7-ALIGNED-NEXT: flat_store_short v[0:1], v2 61; GFX7-ALIGNED-NEXT: s_addc_u32 s3, s1, 0 62; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s2 63; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 2 64; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s3 65; GFX7-ALIGNED-NEXT: flat_store_short v[0:1], v2 66; GFX7-ALIGNED-NEXT: s_endpgm 67; 68; GFX7-UNALIGNED-LABEL: global_store_2xi16_align2: 69; GFX7-UNALIGNED: ; %bb.0: 70; GFX7-UNALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 71; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 1 72; GFX7-UNALIGNED-NEXT: s_waitcnt lgkmcnt(0) 73; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s0 74; GFX7-UNALIGNED-NEXT: s_add_u32 s2, s0, 2 75; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s1 76; GFX7-UNALIGNED-NEXT: flat_store_short v[0:1], v2 77; GFX7-UNALIGNED-NEXT: s_addc_u32 s3, s1, 0 78; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s2 79; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 2 80; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s3 81; GFX7-UNALIGNED-NEXT: flat_store_short v[0:1], v2 82; GFX7-UNALIGNED-NEXT: s_endpgm 83; 84; GFX9-LABEL: global_store_2xi16_align2: 85; GFX9: ; %bb.0: 86; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 87; GFX9-NEXT: v_mov_b32_e32 v0, 0 88; GFX9-NEXT: v_mov_b32_e32 v1, 1 89; GFX9-NEXT: v_mov_b32_e32 v2, 2 90; GFX9-NEXT: s_waitcnt lgkmcnt(0) 91; GFX9-NEXT: global_store_short v0, v1, s[0:1] 92; GFX9-NEXT: global_store_short v0, v2, s[0:1] offset:2 93; GFX9-NEXT: s_endpgm 94 %gep.r = getelementptr i16, i16 addrspace(1)* %r, i64 1 95 store i16 1, i16 addrspace(1)* %r, align 2 96 store i16 2, i16 addrspace(1)* %gep.r, align 2 97 ret void 98} 99 100; Should produce align 1 dword when legal 101define i32 @global_load_2xi16_align1(i16 addrspace(1)* %p) #0 { 102; GFX7-ALIGNED-LABEL: global_load_2xi16_align1: 103; GFX7-ALIGNED: ; %bb.0: 104; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 105; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v0 106; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 107; GFX7-ALIGNED-NEXT: v_add_i32_e32 v4, vcc, 1, v0 108; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc 109; GFX7-ALIGNED-NEXT: v_add_i32_e32 v6, vcc, 3, v0 110; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc 111; GFX7-ALIGNED-NEXT: flat_load_ubyte v0, v[0:1] 112; GFX7-ALIGNED-NEXT: flat_load_ubyte v1, v[6:7] 113; GFX7-ALIGNED-NEXT: flat_load_ubyte v4, v[4:5] 114; GFX7-ALIGNED-NEXT: flat_load_ubyte v2, v[2:3] 115; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(2) 116; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 8, v1 117; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(1) 118; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v3, 8, v4 119; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 120; GFX7-ALIGNED-NEXT: v_or_b32_e32 v1, v1, v2 121; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v3, v0 122; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1 123; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 124; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 125; 126; GFX7-UNALIGNED-LABEL: global_load_2xi16_align1: 127; GFX7-UNALIGNED: ; %bb.0: 128; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 129; GFX7-UNALIGNED-NEXT: flat_load_dword v0, v[0:1] 130; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 131; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 132; 133; GFX9-LABEL: global_load_2xi16_align1: 134; GFX9: ; %bb.0: 135; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 136; GFX9-NEXT: global_load_dword v0, v[0:1], off 137; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff 138; GFX9-NEXT: s_mov_b32 s4, 0xffff 139; GFX9-NEXT: s_waitcnt vmcnt(0) 140; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0 141; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1 142; GFX9-NEXT: s_setpc_b64 s[30:31] 143 %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1 144 %p.0 = load i16, i16 addrspace(1)* %p, align 1 145 %p.1 = load i16, i16 addrspace(1)* %gep.p, align 1 146 %zext.0 = zext i16 %p.0 to i32 147 %zext.1 = zext i16 %p.1 to i32 148 %shl.1 = shl i32 %zext.1, 16 149 %or = or i32 %zext.0, %shl.1 150 ret i32 %or 151} 152 153; Should produce align 1 dword when legal 154define amdgpu_kernel void @global_store_2xi16_align1(i16 addrspace(1)* %p, i16 addrspace(1)* %r) #0 { 155; GFX7-ALIGNED-LABEL: global_store_2xi16_align1: 156; GFX7-ALIGNED: ; %bb.0: 157; GFX7-ALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 158; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 1 159; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v3, 0 160; GFX7-ALIGNED-NEXT: s_waitcnt lgkmcnt(0) 161; GFX7-ALIGNED-NEXT: s_add_u32 s2, s0, 2 162; GFX7-ALIGNED-NEXT: s_addc_u32 s3, s1, 0 163; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 164; GFX7-ALIGNED-NEXT: s_add_u32 s4, s0, 1 165; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 166; GFX7-ALIGNED-NEXT: s_addc_u32 s5, s1, 0 167; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v2 168; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s4 169; GFX7-ALIGNED-NEXT: s_add_u32 s0, s0, 3 170; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s5 171; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v3 172; GFX7-ALIGNED-NEXT: s_addc_u32 s1, s1, 0 173; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 174; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 175; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v3 176; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s2 177; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 2 178; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s3 179; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v2 180; GFX7-ALIGNED-NEXT: s_endpgm 181; 182; GFX7-UNALIGNED-LABEL: global_store_2xi16_align1: 183; GFX7-UNALIGNED: ; %bb.0: 184; GFX7-UNALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 185; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 0x20001 186; GFX7-UNALIGNED-NEXT: s_waitcnt lgkmcnt(0) 187; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s0 188; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s1 189; GFX7-UNALIGNED-NEXT: flat_store_dword v[0:1], v2 190; GFX7-UNALIGNED-NEXT: s_endpgm 191; 192; GFX9-LABEL: global_store_2xi16_align1: 193; GFX9: ; %bb.0: 194; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 195; GFX9-NEXT: v_mov_b32_e32 v0, 0 196; GFX9-NEXT: v_mov_b32_e32 v1, 0x20001 197; GFX9-NEXT: s_waitcnt lgkmcnt(0) 198; GFX9-NEXT: global_store_dword v0, v1, s[0:1] 199; GFX9-NEXT: s_endpgm 200 %gep.r = getelementptr i16, i16 addrspace(1)* %r, i64 1 201 store i16 1, i16 addrspace(1)* %r, align 1 202 store i16 2, i16 addrspace(1)* %gep.r, align 1 203 ret void 204} 205 206; Should merge this to a dword load 207define i32 @global_load_2xi16_align4(i16 addrspace(1)* %p) #0 { 208; GFX7-LABEL: load_2xi16_align4: 209; GFX7: ; %bb.0: 210; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 211; GFX7-NEXT: flat_load_dword v0, v[0:1] 212; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 213; GFX7-NEXT: s_setpc_b64 s[30:31] 214; 215; GFX7-ALIGNED-LABEL: global_load_2xi16_align4: 216; GFX7-ALIGNED: ; %bb.0: 217; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 218; GFX7-ALIGNED-NEXT: flat_load_dword v0, v[0:1] 219; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 220; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 221; 222; GFX7-UNALIGNED-LABEL: global_load_2xi16_align4: 223; GFX7-UNALIGNED: ; %bb.0: 224; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 225; GFX7-UNALIGNED-NEXT: flat_load_dword v0, v[0:1] 226; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 227; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 228; 229; GFX9-LABEL: global_load_2xi16_align4: 230; GFX9: ; %bb.0: 231; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 232; GFX9-NEXT: global_load_dword v0, v[0:1], off 233; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff 234; GFX9-NEXT: s_mov_b32 s4, 0xffff 235; GFX9-NEXT: s_waitcnt vmcnt(0) 236; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0 237; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1 238; GFX9-NEXT: s_setpc_b64 s[30:31] 239 %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1 240 %p.0 = load i16, i16 addrspace(1)* %p, align 4 241 %p.1 = load i16, i16 addrspace(1)* %gep.p, align 2 242 %zext.0 = zext i16 %p.0 to i32 243 %zext.1 = zext i16 %p.1 to i32 244 %shl.1 = shl i32 %zext.1, 16 245 %or = or i32 %zext.0, %shl.1 246 ret i32 %or 247} 248 249; Should merge this to a dword store 250define amdgpu_kernel void @global_store_2xi16_align4(i16 addrspace(1)* %p, i16 addrspace(1)* %r) #0 { 251; GFX7-LABEL: global_store_2xi16_align4: 252; GFX7: ; %bb.0: 253; GFX7-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 254; GFX7-NEXT: v_mov_b32_e32 v2, 0x20001 255; GFX7-NEXT: s_waitcnt lgkmcnt(0) 256; GFX7-NEXT: v_mov_b32_e32 v0, s0 257; GFX7-NEXT: v_mov_b32_e32 v1, s1 258; GFX7-NEXT: flat_store_dword v[0:1], v2 259; GFX7-NEXT: s_endpgm 260; 261; GFX7-ALIGNED-LABEL: global_store_2xi16_align4: 262; GFX7-ALIGNED: ; %bb.0: 263; GFX7-ALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 264; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 0x20001 265; GFX7-ALIGNED-NEXT: s_waitcnt lgkmcnt(0) 266; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 267; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 268; GFX7-ALIGNED-NEXT: flat_store_dword v[0:1], v2 269; GFX7-ALIGNED-NEXT: s_endpgm 270; 271; GFX7-UNALIGNED-LABEL: global_store_2xi16_align4: 272; GFX7-UNALIGNED: ; %bb.0: 273; GFX7-UNALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 274; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 0x20001 275; GFX7-UNALIGNED-NEXT: s_waitcnt lgkmcnt(0) 276; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s0 277; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s1 278; GFX7-UNALIGNED-NEXT: flat_store_dword v[0:1], v2 279; GFX7-UNALIGNED-NEXT: s_endpgm 280; 281; GFX9-LABEL: global_store_2xi16_align4: 282; GFX9: ; %bb.0: 283; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 284; GFX9-NEXT: v_mov_b32_e32 v0, 0 285; GFX9-NEXT: v_mov_b32_e32 v1, 0x20001 286; GFX9-NEXT: s_waitcnt lgkmcnt(0) 287; GFX9-NEXT: global_store_dword v0, v1, s[0:1] 288; GFX9-NEXT: s_endpgm 289 %gep.r = getelementptr i16, i16 addrspace(1)* %r, i64 1 290 store i16 1, i16 addrspace(1)* %r, align 4 291 store i16 2, i16 addrspace(1)* %gep.r, align 2 292 ret void 293} 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327