1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=-unaligned-scratch-access < %s | FileCheck --check-prefix=GFX7-ALIGNED %s 3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=+unaligned-scratch-access < %s | FileCheck --check-prefix=GFX7-UNALIGNED %s 4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+unaligned-scratch-access < %s | FileCheck --check-prefix=GFX9 %s 5; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+unaligned-scratch-access -amdgpu-enable-flat-scratch < %s | FileCheck --check-prefix=GFX9-FLASTSCR %s 6; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+unaligned-scratch-access < %s | FileCheck --check-prefix=GFX10 %s 7; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+unaligned-scratch-access -amdgpu-enable-flat-scratch < %s | FileCheck --check-prefix=GFX10-FLASTSCR %s 8 9; Should not merge this to a dword load 10define i32 @private_load_2xi16_align2(i16 addrspace(5)* %p) #0 { 11; GFX7-ALIGNED-LABEL: private_load_2xi16_align2: 12; GFX7-ALIGNED: ; %bb.0: 13; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 14; GFX7-ALIGNED-NEXT: v_add_i32_e32 v1, vcc, 2, v0 15; GFX7-ALIGNED-NEXT: buffer_load_ushort v1, v1, s[0:3], 0 offen 16; GFX7-ALIGNED-NEXT: buffer_load_ushort v0, v0, s[0:3], 0 offen 17; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(1) 18; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1 19; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 20; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 21; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 22; 23; GFX7-UNALIGNED-LABEL: private_load_2xi16_align2: 24; GFX7-UNALIGNED: ; %bb.0: 25; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 26; GFX7-UNALIGNED-NEXT: v_add_i32_e32 v1, vcc, 2, v0 27; GFX7-UNALIGNED-NEXT: buffer_load_ushort v1, v1, s[0:3], 0 offen 28; GFX7-UNALIGNED-NEXT: buffer_load_ushort v0, v0, s[0:3], 0 offen 29; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(1) 30; GFX7-UNALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1 31; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 32; GFX7-UNALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 33; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 34; 35; GFX9-LABEL: private_load_2xi16_align2: 36; GFX9: ; %bb.0: 37; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 38; GFX9-NEXT: buffer_load_ushort v1, v0, s[0:3], 0 offen 39; GFX9-NEXT: buffer_load_ushort v2, v0, s[0:3], 0 offen offset:2 40; GFX9-NEXT: s_waitcnt vmcnt(0) 41; GFX9-NEXT: v_lshl_or_b32 v0, v2, 16, v1 42; GFX9-NEXT: s_setpc_b64 s[30:31] 43; 44; GFX9-FLASTSCR-LABEL: private_load_2xi16_align2: 45; GFX9-FLASTSCR: ; %bb.0: 46; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 47; GFX9-FLASTSCR-NEXT: scratch_load_ushort v1, v0, off 48; GFX9-FLASTSCR-NEXT: scratch_load_ushort v2, v0, off offset:2 49; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 50; GFX9-FLASTSCR-NEXT: v_lshl_or_b32 v0, v2, 16, v1 51; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 52; 53; GFX10-LABEL: private_load_2xi16_align2: 54; GFX10: ; %bb.0: 55; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 56; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 57; GFX10-NEXT: s_clause 0x1 58; GFX10-NEXT: buffer_load_ushort v1, v0, s[0:3], 0 offen 59; GFX10-NEXT: buffer_load_ushort v2, v0, s[0:3], 0 offen offset:2 60; GFX10-NEXT: s_waitcnt vmcnt(0) 61; GFX10-NEXT: v_lshl_or_b32 v0, v2, 16, v1 62; GFX10-NEXT: s_setpc_b64 s[30:31] 63; 64; GFX10-FLASTSCR-LABEL: private_load_2xi16_align2: 65; GFX10-FLASTSCR: ; %bb.0: 66; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 67; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 68; GFX10-FLASTSCR-NEXT: s_clause 0x1 69; GFX10-FLASTSCR-NEXT: scratch_load_ushort v1, v0, off 70; GFX10-FLASTSCR-NEXT: scratch_load_ushort v2, v0, off offset:2 71; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 72; GFX10-FLASTSCR-NEXT: v_lshl_or_b32 v0, v2, 16, v1 73; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 74 %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1 75 %p.0 = load i16, i16 addrspace(5)* %p, align 2 76 %p.1 = load i16, i16 addrspace(5)* %gep.p, align 2 77 %zext.0 = zext i16 %p.0 to i32 78 %zext.1 = zext i16 %p.1 to i32 79 %shl.1 = shl i32 %zext.1, 16 80 %or = or i32 %zext.0, %shl.1 81 ret i32 %or 82} 83 84; Should not merge this to a dword store 85define void @private_store_2xi16_align2(i16 addrspace(5)* %p, i16 addrspace(5)* %r) #0 { 86; GFX7-ALIGNED-LABEL: private_store_2xi16_align2: 87; GFX7-ALIGNED: ; %bb.0: 88; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 89; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v3, 1 90; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, 2 91; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v1 92; GFX7-ALIGNED-NEXT: buffer_store_short v3, v1, s[0:3], 0 offen 93; GFX7-ALIGNED-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen 94; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 95; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 96; 97; GFX7-UNALIGNED-LABEL: private_store_2xi16_align2: 98; GFX7-UNALIGNED: ; %bb.0: 99; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 100; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v3, 1 101; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, 2 102; GFX7-UNALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v1 103; GFX7-UNALIGNED-NEXT: buffer_store_short v3, v1, s[0:3], 0 offen 104; GFX7-UNALIGNED-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen 105; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 106; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 107; 108; GFX9-LABEL: private_store_2xi16_align2: 109; GFX9: ; %bb.0: 110; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 111; GFX9-NEXT: v_mov_b32_e32 v0, 1 112; GFX9-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen 113; GFX9-NEXT: v_mov_b32_e32 v0, 2 114; GFX9-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen offset:2 115; GFX9-NEXT: s_waitcnt vmcnt(0) 116; GFX9-NEXT: s_setpc_b64 s[30:31] 117; 118; GFX9-FLASTSCR-LABEL: private_store_2xi16_align2: 119; GFX9-FLASTSCR: ; %bb.0: 120; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 121; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 1 122; GFX9-FLASTSCR-NEXT: scratch_store_short v1, v0, off 123; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 2 124; GFX9-FLASTSCR-NEXT: scratch_store_short v1, v0, off offset:2 125; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 126; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 127; 128; GFX10-LABEL: private_store_2xi16_align2: 129; GFX10: ; %bb.0: 130; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 131; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 132; GFX10-NEXT: v_mov_b32_e32 v0, 1 133; GFX10-NEXT: v_mov_b32_e32 v2, 2 134; GFX10-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen 135; GFX10-NEXT: buffer_store_short v2, v1, s[0:3], 0 offen offset:2 136; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 137; GFX10-NEXT: s_setpc_b64 s[30:31] 138; 139; GFX10-FLASTSCR-LABEL: private_store_2xi16_align2: 140; GFX10-FLASTSCR: ; %bb.0: 141; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 142; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 143; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v0, 1 144; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v2, 2 145; GFX10-FLASTSCR-NEXT: scratch_store_short v1, v0, off 146; GFX10-FLASTSCR-NEXT: scratch_store_short v1, v2, off offset:2 147; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 148; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 149 %gep.r = getelementptr i16, i16 addrspace(5)* %r, i64 1 150 store i16 1, i16 addrspace(5)* %r, align 2 151 store i16 2, i16 addrspace(5)* %gep.r, align 2 152 ret void 153} 154 155; Should produce align 1 dword when legal 156define i32 @private_load_2xi16_align1(i16 addrspace(5)* %p) #0 { 157; GFX7-ALIGNED-LABEL: private_load_2xi16_align1: 158; GFX7-ALIGNED: ; %bb.0: 159; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 160; GFX7-ALIGNED-NEXT: v_add_i32_e32 v1, vcc, 2, v0 161; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 1, v0 162; GFX7-ALIGNED-NEXT: v_add_i32_e32 v3, vcc, 3, v0 163; GFX7-ALIGNED-NEXT: buffer_load_ubyte v2, v2, s[0:3], 0 offen 164; GFX7-ALIGNED-NEXT: buffer_load_ubyte v3, v3, s[0:3], 0 offen 165; GFX7-ALIGNED-NEXT: buffer_load_ubyte v0, v0, s[0:3], 0 offen 166; GFX7-ALIGNED-NEXT: buffer_load_ubyte v1, v1, s[0:3], 0 offen 167; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(3) 168; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v2, 8, v2 169; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(2) 170; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v3, 8, v3 171; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(1) 172; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v2, v0 173; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 174; GFX7-ALIGNED-NEXT: v_or_b32_e32 v1, v3, v1 175; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1 176; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 177; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 178; 179; GFX7-UNALIGNED-LABEL: private_load_2xi16_align1: 180; GFX7-UNALIGNED: ; %bb.0: 181; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 182; GFX7-UNALIGNED-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 183; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 184; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 185; 186; GFX9-LABEL: private_load_2xi16_align1: 187; GFX9: ; %bb.0: 188; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 189; GFX9-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 190; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff 191; GFX9-NEXT: s_mov_b32 s4, 0xffff 192; GFX9-NEXT: s_waitcnt vmcnt(0) 193; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0 194; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1 195; GFX9-NEXT: s_setpc_b64 s[30:31] 196; 197; GFX9-FLASTSCR-LABEL: private_load_2xi16_align1: 198; GFX9-FLASTSCR: ; %bb.0: 199; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 200; GFX9-FLASTSCR-NEXT: scratch_load_dword v0, v0, off 201; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v1, 0xffff 202; GFX9-FLASTSCR-NEXT: s_mov_b32 s0, 0xffff 203; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 204; GFX9-FLASTSCR-NEXT: v_bfi_b32 v1, v1, 0, v0 205; GFX9-FLASTSCR-NEXT: v_and_or_b32 v0, v0, s0, v1 206; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 207; 208; GFX10-LABEL: private_load_2xi16_align1: 209; GFX10: ; %bb.0: 210; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 211; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 212; GFX10-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 213; GFX10-NEXT: s_waitcnt vmcnt(0) 214; GFX10-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 215; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 216; GFX10-NEXT: s_setpc_b64 s[30:31] 217; 218; GFX10-FLASTSCR-LABEL: private_load_2xi16_align1: 219; GFX10-FLASTSCR: ; %bb.0: 220; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 221; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 222; GFX10-FLASTSCR-NEXT: scratch_load_dword v0, v0, off 223; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 224; GFX10-FLASTSCR-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 225; GFX10-FLASTSCR-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 226; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 227 %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1 228 %p.0 = load i16, i16 addrspace(5)* %p, align 1 229 %p.1 = load i16, i16 addrspace(5)* %gep.p, align 1 230 %zext.0 = zext i16 %p.0 to i32 231 %zext.1 = zext i16 %p.1 to i32 232 %shl.1 = shl i32 %zext.1, 16 233 %or = or i32 %zext.0, %shl.1 234 ret i32 %or 235} 236 237; Should produce align 1 dword when legal 238define void @private_store_2xi16_align1(i16 addrspace(5)* %p, i16 addrspace(5)* %r) #0 { 239; GFX7-ALIGNED-LABEL: private_store_2xi16_align1: 240; GFX7-ALIGNED: ; %bb.0: 241; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 242; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v3, 1 243; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v1 244; GFX7-ALIGNED-NEXT: buffer_store_byte v3, v1, s[0:3], 0 offen 245; GFX7-ALIGNED-NEXT: v_add_i32_e32 v3, vcc, 1, v1 246; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v4, 0 247; GFX7-ALIGNED-NEXT: v_add_i32_e32 v1, vcc, 3, v1 248; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, 2 249; GFX7-ALIGNED-NEXT: buffer_store_byte v4, v3, s[0:3], 0 offen 250; GFX7-ALIGNED-NEXT: buffer_store_byte v4, v1, s[0:3], 0 offen 251; GFX7-ALIGNED-NEXT: buffer_store_byte v0, v2, s[0:3], 0 offen 252; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 253; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 254; 255; GFX7-UNALIGNED-LABEL: private_store_2xi16_align1: 256; GFX7-UNALIGNED: ; %bb.0: 257; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 258; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, 0x20001 259; GFX7-UNALIGNED-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 260; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 261; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 262; 263; GFX9-LABEL: private_store_2xi16_align1: 264; GFX9: ; %bb.0: 265; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 266; GFX9-NEXT: v_mov_b32_e32 v0, 0x20001 267; GFX9-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 268; GFX9-NEXT: s_waitcnt vmcnt(0) 269; GFX9-NEXT: s_setpc_b64 s[30:31] 270; 271; GFX9-FLASTSCR-LABEL: private_store_2xi16_align1: 272; GFX9-FLASTSCR: ; %bb.0: 273; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 274; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001 275; GFX9-FLASTSCR-NEXT: scratch_store_dword v1, v0, off 276; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 277; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 278; 279; GFX10-LABEL: private_store_2xi16_align1: 280; GFX10: ; %bb.0: 281; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 282; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 283; GFX10-NEXT: v_mov_b32_e32 v0, 0x20001 284; GFX10-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 285; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 286; GFX10-NEXT: s_setpc_b64 s[30:31] 287; 288; GFX10-FLASTSCR-LABEL: private_store_2xi16_align1: 289; GFX10-FLASTSCR: ; %bb.0: 290; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 291; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 292; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001 293; GFX10-FLASTSCR-NEXT: scratch_store_dword v1, v0, off 294; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 295; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 296 %gep.r = getelementptr i16, i16 addrspace(5)* %r, i64 1 297 store i16 1, i16 addrspace(5)* %r, align 1 298 store i16 2, i16 addrspace(5)* %gep.r, align 1 299 ret void 300} 301 302; Should merge this to a dword load 303define i32 @private_load_2xi16_align4(i16 addrspace(5)* %p) #0 { 304; GFX7-LABEL: load_2xi16_align4: 305; GFX7: ; %bb.0: 306; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 307; GFX7-NEXT: flat_load_dword v0, v[0:1] 308; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 309; GFX7-NEXT: s_setpc_b64 s[30:31] 310; 311; GFX7-ALIGNED-LABEL: private_load_2xi16_align4: 312; GFX7-ALIGNED: ; %bb.0: 313; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 314; GFX7-ALIGNED-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 315; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 316; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 317; 318; GFX7-UNALIGNED-LABEL: private_load_2xi16_align4: 319; GFX7-UNALIGNED: ; %bb.0: 320; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 321; GFX7-UNALIGNED-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 322; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 323; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 324; 325; GFX9-LABEL: private_load_2xi16_align4: 326; GFX9: ; %bb.0: 327; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 328; GFX9-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 329; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff 330; GFX9-NEXT: s_mov_b32 s4, 0xffff 331; GFX9-NEXT: s_waitcnt vmcnt(0) 332; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0 333; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1 334; GFX9-NEXT: s_setpc_b64 s[30:31] 335; 336; GFX9-FLASTSCR-LABEL: private_load_2xi16_align4: 337; GFX9-FLASTSCR: ; %bb.0: 338; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 339; GFX9-FLASTSCR-NEXT: scratch_load_dword v0, v0, off 340; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v1, 0xffff 341; GFX9-FLASTSCR-NEXT: s_mov_b32 s0, 0xffff 342; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 343; GFX9-FLASTSCR-NEXT: v_bfi_b32 v1, v1, 0, v0 344; GFX9-FLASTSCR-NEXT: v_and_or_b32 v0, v0, s0, v1 345; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 346; 347; GFX10-LABEL: private_load_2xi16_align4: 348; GFX10: ; %bb.0: 349; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 350; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 351; GFX10-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen 352; GFX10-NEXT: s_waitcnt vmcnt(0) 353; GFX10-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 354; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 355; GFX10-NEXT: s_setpc_b64 s[30:31] 356; 357; GFX10-FLASTSCR-LABEL: private_load_2xi16_align4: 358; GFX10-FLASTSCR: ; %bb.0: 359; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 360; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 361; GFX10-FLASTSCR-NEXT: scratch_load_dword v0, v0, off 362; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 363; GFX10-FLASTSCR-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 364; GFX10-FLASTSCR-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 365; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 366 %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1 367 %p.0 = load i16, i16 addrspace(5)* %p, align 4 368 %p.1 = load i16, i16 addrspace(5)* %gep.p, align 2 369 %zext.0 = zext i16 %p.0 to i32 370 %zext.1 = zext i16 %p.1 to i32 371 %shl.1 = shl i32 %zext.1, 16 372 %or = or i32 %zext.0, %shl.1 373 ret i32 %or 374} 375 376; Should merge this to a dword store 377define void @private_store_2xi16_align4(i16 addrspace(5)* %p, i16 addrspace(5)* %r) #0 { 378; GFX7-LABEL: private_store_2xi16_align4: 379; GFX7: ; %bb.0: 380; GFX7-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 381; GFX7-NEXT: v_mov_b32_e32 v2, 0x20001 382; GFX7-NEXT: s_waitcnt lgkmcnt(0) 383; GFX7-NEXT: v_mov_b32_e32 v0, s0 384; GFX7-NEXT: v_mov_b32_e32 v1, s1 385; GFX7-NEXT: flat_store_dword v[0:1], v2 386; GFX7-NEXT: s_endpgm 387; 388; GFX7-ALIGNED-LABEL: private_store_2xi16_align4: 389; GFX7-ALIGNED: ; %bb.0: 390; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 391; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, 0x20001 392; GFX7-ALIGNED-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 393; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 394; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 395; 396; GFX7-UNALIGNED-LABEL: private_store_2xi16_align4: 397; GFX7-UNALIGNED: ; %bb.0: 398; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 399; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, 0x20001 400; GFX7-UNALIGNED-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 401; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 402; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 403; 404; GFX9-LABEL: private_store_2xi16_align4: 405; GFX9: ; %bb.0: 406; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 407; GFX9-NEXT: v_mov_b32_e32 v0, 0x20001 408; GFX9-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 409; GFX9-NEXT: s_waitcnt vmcnt(0) 410; GFX9-NEXT: s_setpc_b64 s[30:31] 411; 412; GFX9-FLASTSCR-LABEL: private_store_2xi16_align4: 413; GFX9-FLASTSCR: ; %bb.0: 414; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 415; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001 416; GFX9-FLASTSCR-NEXT: scratch_store_dword v1, v0, off 417; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) 418; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 419; 420; GFX10-LABEL: private_store_2xi16_align4: 421; GFX10: ; %bb.0: 422; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 423; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 424; GFX10-NEXT: v_mov_b32_e32 v0, 0x20001 425; GFX10-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen 426; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 427; GFX10-NEXT: s_setpc_b64 s[30:31] 428; 429; GFX10-FLASTSCR-LABEL: private_store_2xi16_align4: 430; GFX10-FLASTSCR: ; %bb.0: 431; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 432; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 433; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001 434; GFX10-FLASTSCR-NEXT: scratch_store_dword v1, v0, off 435; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0 436; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31] 437 %gep.r = getelementptr i16, i16 addrspace(5)* %r, i64 1 438 store i16 1, i16 addrspace(5)* %r, align 4 439 store i16 2, i16 addrspace(5)* %gep.r, align 2 440 ret void 441} 442