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-prefix=GFX7-ALIGNED %s 3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=+unaligned-access-mode < %s | FileCheck --check-prefix=GFX7-UNALIGNED %s 4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+unaligned-access-mode < %s | FileCheck --check-prefix=GFX9 %s 5; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+unaligned-access-mode < %s | FileCheck --check-prefix=GFX10 %s 6; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -mattr=+unaligned-access-mode < %s | FileCheck --check-prefix=GFX11 %s 7 8; Should not merge this to a dword load 9define i32 @global_load_2xi16_align2(i16 addrspace(1)* %p) #0 { 10; GFX7-ALIGNED-LABEL: global_load_2xi16_align2: 11; GFX7-ALIGNED: ; %bb.0: 12; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 13; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v0 14; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 15; GFX7-ALIGNED-NEXT: flat_load_ushort v2, v[2:3] 16; GFX7-ALIGNED-NEXT: flat_load_ushort v0, v[0:1] 17; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(1) 18; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v2 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: global_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 v2, vcc, 2, v0 27; GFX7-UNALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 28; GFX7-UNALIGNED-NEXT: flat_load_ushort v2, v[2:3] 29; GFX7-UNALIGNED-NEXT: flat_load_ushort v0, v[0:1] 30; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(1) 31; GFX7-UNALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v2 32; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 33; GFX7-UNALIGNED-NEXT: v_or_b32_e32 v0, v0, v1 34; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 35; 36; GFX9-LABEL: global_load_2xi16_align2: 37; GFX9: ; %bb.0: 38; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 39; GFX9-NEXT: global_load_ushort v2, v[0:1], off 40; GFX9-NEXT: global_load_ushort v3, v[0:1], off offset:2 41; GFX9-NEXT: s_waitcnt vmcnt(0) 42; GFX9-NEXT: v_lshl_or_b32 v0, v3, 16, v2 43; GFX9-NEXT: s_setpc_b64 s[30:31] 44; 45; GFX10-LABEL: global_load_2xi16_align2: 46; GFX10: ; %bb.0: 47; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 48; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 49; GFX10-NEXT: s_clause 0x1 50; GFX10-NEXT: global_load_ushort v2, v[0:1], off 51; GFX10-NEXT: global_load_ushort v3, v[0:1], off offset:2 52; GFX10-NEXT: s_waitcnt vmcnt(0) 53; GFX10-NEXT: v_lshl_or_b32 v0, v3, 16, v2 54; GFX10-NEXT: s_setpc_b64 s[30:31] 55; 56; GFX11-LABEL: global_load_2xi16_align2: 57; GFX11: ; %bb.0: 58; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 59; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 60; GFX11-NEXT: s_clause 0x1 61; GFX11-NEXT: global_load_u16 v2, v[0:1], off 62; GFX11-NEXT: global_load_u16 v0, v[0:1], off offset:2 63; GFX11-NEXT: s_waitcnt vmcnt(0) 64; GFX11-NEXT: v_lshl_or_b32 v0, v0, 16, v2 65; GFX11-NEXT: s_setpc_b64 s[30:31] 66 %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1 67 %p.0 = load i16, i16 addrspace(1)* %p, align 2 68 %p.1 = load i16, i16 addrspace(1)* %gep.p, align 2 69 %zext.0 = zext i16 %p.0 to i32 70 %zext.1 = zext i16 %p.1 to i32 71 %shl.1 = shl i32 %zext.1, 16 72 %or = or i32 %zext.0, %shl.1 73 ret i32 %or 74} 75 76; Should not merge this to a dword store 77define amdgpu_kernel void @global_store_2xi16_align2(i16 addrspace(1)* %p, i16 addrspace(1)* %r) #0 { 78; GFX7-ALIGNED-LABEL: global_store_2xi16_align2: 79; GFX7-ALIGNED: ; %bb.0: 80; GFX7-ALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 81; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 1 82; GFX7-ALIGNED-NEXT: s_waitcnt lgkmcnt(0) 83; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 84; GFX7-ALIGNED-NEXT: s_add_u32 s2, s0, 2 85; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 86; GFX7-ALIGNED-NEXT: s_addc_u32 s3, s1, 0 87; GFX7-ALIGNED-NEXT: flat_store_short v[0:1], v2 88; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s2 89; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 2 90; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s3 91; GFX7-ALIGNED-NEXT: flat_store_short v[0:1], v2 92; GFX7-ALIGNED-NEXT: s_endpgm 93; 94; GFX7-UNALIGNED-LABEL: global_store_2xi16_align2: 95; GFX7-UNALIGNED: ; %bb.0: 96; GFX7-UNALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 97; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 1 98; GFX7-UNALIGNED-NEXT: s_waitcnt lgkmcnt(0) 99; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s0 100; GFX7-UNALIGNED-NEXT: s_add_u32 s2, s0, 2 101; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s1 102; GFX7-UNALIGNED-NEXT: s_addc_u32 s3, s1, 0 103; GFX7-UNALIGNED-NEXT: flat_store_short v[0:1], v2 104; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s2 105; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 2 106; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s3 107; GFX7-UNALIGNED-NEXT: flat_store_short v[0:1], v2 108; GFX7-UNALIGNED-NEXT: s_endpgm 109; 110; GFX9-LABEL: global_store_2xi16_align2: 111; GFX9: ; %bb.0: 112; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 113; GFX9-NEXT: v_mov_b32_e32 v0, 0 114; GFX9-NEXT: v_mov_b32_e32 v1, 1 115; GFX9-NEXT: v_mov_b32_e32 v2, 2 116; GFX9-NEXT: s_waitcnt lgkmcnt(0) 117; GFX9-NEXT: global_store_short v0, v1, s[0:1] 118; GFX9-NEXT: global_store_short v0, v2, s[0:1] offset:2 119; GFX9-NEXT: s_endpgm 120; 121; GFX10-LABEL: global_store_2xi16_align2: 122; GFX10: ; %bb.0: 123; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 124; GFX10-NEXT: v_mov_b32_e32 v0, 0 125; GFX10-NEXT: v_mov_b32_e32 v1, 1 126; GFX10-NEXT: v_mov_b32_e32 v2, 2 127; GFX10-NEXT: s_waitcnt lgkmcnt(0) 128; GFX10-NEXT: global_store_short v0, v1, s[0:1] 129; GFX10-NEXT: global_store_short v0, v2, s[0:1] offset:2 130; GFX10-NEXT: s_endpgm 131; 132; GFX11-LABEL: global_store_2xi16_align2: 133; GFX11: ; %bb.0: 134; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x8 135; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 1 136; GFX11-NEXT: v_mov_b32_e32 v2, 2 137; GFX11-NEXT: s_waitcnt lgkmcnt(0) 138; GFX11-NEXT: s_clause 0x1 139; GFX11-NEXT: global_store_b16 v0, v1, s[0:1] 140; GFX11-NEXT: global_store_b16 v0, v2, s[0:1] offset:2 141; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) 142; GFX11-NEXT: s_endpgm 143 %gep.r = getelementptr i16, i16 addrspace(1)* %r, i64 1 144 store i16 1, i16 addrspace(1)* %r, align 2 145 store i16 2, i16 addrspace(1)* %gep.r, align 2 146 ret void 147} 148 149; Should produce align 1 dword when legal 150define i32 @global_load_2xi16_align1(i16 addrspace(1)* %p) #0 { 151; GFX7-ALIGNED-LABEL: global_load_2xi16_align1: 152; GFX7-ALIGNED: ; %bb.0: 153; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 154; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 1, v0 155; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 156; GFX7-ALIGNED-NEXT: flat_load_ubyte v4, v[2:3] 157; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 3, v0 158; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc 159; GFX7-ALIGNED-NEXT: flat_load_ubyte v5, v[0:1] 160; GFX7-ALIGNED-NEXT: flat_load_ubyte v2, v[2:3] 161; GFX7-ALIGNED-NEXT: v_add_i32_e32 v0, vcc, 2, v0 162; GFX7-ALIGNED-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc 163; GFX7-ALIGNED-NEXT: flat_load_ubyte v0, v[0:1] 164; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(3) 165; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 8, v4 166; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(2) 167; GFX7-ALIGNED-NEXT: v_or_b32_e32 v1, v1, v5 168; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(1) 169; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v2, 8, v2 170; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 171; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v2, v0 172; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v0, 16, v0 173; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v1, v0 174; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 175; 176; GFX7-UNALIGNED-LABEL: global_load_2xi16_align1: 177; GFX7-UNALIGNED: ; %bb.0: 178; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 179; GFX7-UNALIGNED-NEXT: flat_load_dword v0, v[0:1] 180; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 181; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 182; 183; GFX9-LABEL: global_load_2xi16_align1: 184; GFX9: ; %bb.0: 185; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 186; GFX9-NEXT: global_load_dword v0, v[0:1], off 187; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff 188; GFX9-NEXT: s_mov_b32 s4, 0xffff 189; GFX9-NEXT: s_waitcnt vmcnt(0) 190; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0 191; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1 192; GFX9-NEXT: s_setpc_b64 s[30:31] 193; 194; GFX10-LABEL: global_load_2xi16_align1: 195; GFX10: ; %bb.0: 196; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 197; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 198; GFX10-NEXT: global_load_dword v0, v[0:1], off 199; GFX10-NEXT: s_waitcnt vmcnt(0) 200; GFX10-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 201; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 202; GFX10-NEXT: s_setpc_b64 s[30:31] 203; 204; GFX11-LABEL: global_load_2xi16_align1: 205; GFX11: ; %bb.0: 206; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 207; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 208; GFX11-NEXT: global_load_b32 v0, v[0:1], off 209; GFX11-NEXT: s_waitcnt vmcnt(0) 210; GFX11-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 211; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) 212; GFX11-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 213; GFX11-NEXT: s_setpc_b64 s[30:31] 214 %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1 215 %p.0 = load i16, i16 addrspace(1)* %p, align 1 216 %p.1 = load i16, i16 addrspace(1)* %gep.p, align 1 217 %zext.0 = zext i16 %p.0 to i32 218 %zext.1 = zext i16 %p.1 to i32 219 %shl.1 = shl i32 %zext.1, 16 220 %or = or i32 %zext.0, %shl.1 221 ret i32 %or 222} 223 224; Should produce align 1 dword when legal 225define amdgpu_kernel void @global_store_2xi16_align1(i16 addrspace(1)* %p, i16 addrspace(1)* %r) #0 { 226; GFX7-ALIGNED-LABEL: global_store_2xi16_align1: 227; GFX7-ALIGNED: ; %bb.0: 228; GFX7-ALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 229; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 1 230; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v3, 0 231; GFX7-ALIGNED-NEXT: s_waitcnt lgkmcnt(0) 232; GFX7-ALIGNED-NEXT: s_add_u32 s2, s0, 2 233; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 234; GFX7-ALIGNED-NEXT: s_addc_u32 s3, s1, 0 235; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 236; GFX7-ALIGNED-NEXT: s_add_u32 s4, s0, 1 237; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v2 238; GFX7-ALIGNED-NEXT: s_addc_u32 s5, s1, 0 239; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s4 240; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s5 241; GFX7-ALIGNED-NEXT: s_add_u32 s0, s0, 3 242; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v3 243; GFX7-ALIGNED-NEXT: s_addc_u32 s1, s1, 0 244; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 245; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 246; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v3 247; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s2 248; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 2 249; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s3 250; GFX7-ALIGNED-NEXT: flat_store_byte v[0:1], v2 251; GFX7-ALIGNED-NEXT: s_endpgm 252; 253; GFX7-UNALIGNED-LABEL: global_store_2xi16_align1: 254; GFX7-UNALIGNED: ; %bb.0: 255; GFX7-UNALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 256; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 0x20001 257; GFX7-UNALIGNED-NEXT: s_waitcnt lgkmcnt(0) 258; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s0 259; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s1 260; GFX7-UNALIGNED-NEXT: flat_store_dword v[0:1], v2 261; GFX7-UNALIGNED-NEXT: s_endpgm 262; 263; GFX9-LABEL: global_store_2xi16_align1: 264; GFX9: ; %bb.0: 265; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 266; GFX9-NEXT: v_mov_b32_e32 v0, 0 267; GFX9-NEXT: v_mov_b32_e32 v1, 0x20001 268; GFX9-NEXT: s_waitcnt lgkmcnt(0) 269; GFX9-NEXT: global_store_dword v0, v1, s[0:1] 270; GFX9-NEXT: s_endpgm 271; 272; GFX10-LABEL: global_store_2xi16_align1: 273; GFX10: ; %bb.0: 274; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 275; GFX10-NEXT: v_mov_b32_e32 v0, 0 276; GFX10-NEXT: v_mov_b32_e32 v1, 0x20001 277; GFX10-NEXT: s_waitcnt lgkmcnt(0) 278; GFX10-NEXT: global_store_dword v0, v1, s[0:1] 279; GFX10-NEXT: s_endpgm 280; 281; GFX11-LABEL: global_store_2xi16_align1: 282; GFX11: ; %bb.0: 283; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x8 284; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x20001 285; GFX11-NEXT: s_waitcnt lgkmcnt(0) 286; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] 287; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) 288; GFX11-NEXT: s_endpgm 289 %gep.r = getelementptr i16, i16 addrspace(1)* %r, i64 1 290 store i16 1, i16 addrspace(1)* %r, align 1 291 store i16 2, i16 addrspace(1)* %gep.r, align 1 292 ret void 293} 294 295; Should merge this to a dword load 296define i32 @global_load_2xi16_align4(i16 addrspace(1)* %p) #0 { 297; GFX7-LABEL: load_2xi16_align4: 298; GFX7: ; %bb.0: 299; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 300; GFX7-NEXT: flat_load_dword v0, v[0:1] 301; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 302; GFX7-NEXT: s_setpc_b64 s[30:31] 303; 304; GFX7-ALIGNED-LABEL: global_load_2xi16_align4: 305; GFX7-ALIGNED: ; %bb.0: 306; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 307; GFX7-ALIGNED-NEXT: flat_load_dword v0, v[0:1] 308; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) 309; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31] 310; 311; GFX7-UNALIGNED-LABEL: global_load_2xi16_align4: 312; GFX7-UNALIGNED: ; %bb.0: 313; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 314; GFX7-UNALIGNED-NEXT: flat_load_dword v0, v[0:1] 315; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) 316; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31] 317; 318; GFX9-LABEL: global_load_2xi16_align4: 319; GFX9: ; %bb.0: 320; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 321; GFX9-NEXT: global_load_dword v0, v[0:1], off 322; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff 323; GFX9-NEXT: s_mov_b32 s4, 0xffff 324; GFX9-NEXT: s_waitcnt vmcnt(0) 325; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0 326; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1 327; GFX9-NEXT: s_setpc_b64 s[30:31] 328; 329; GFX10-LABEL: global_load_2xi16_align4: 330; GFX10: ; %bb.0: 331; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 332; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 333; GFX10-NEXT: global_load_dword v0, v[0:1], off 334; GFX10-NEXT: s_waitcnt vmcnt(0) 335; GFX10-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 336; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 337; GFX10-NEXT: s_setpc_b64 s[30:31] 338; 339; GFX11-LABEL: global_load_2xi16_align4: 340; GFX11: ; %bb.0: 341; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 342; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 343; GFX11-NEXT: global_load_b32 v0, v[0:1], off 344; GFX11-NEXT: s_waitcnt vmcnt(0) 345; GFX11-NEXT: v_bfi_b32 v1, 0xffff, 0, v0 346; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) 347; GFX11-NEXT: v_and_or_b32 v0, 0xffff, v0, v1 348; GFX11-NEXT: s_setpc_b64 s[30:31] 349 %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1 350 %p.0 = load i16, i16 addrspace(1)* %p, align 4 351 %p.1 = load i16, i16 addrspace(1)* %gep.p, align 2 352 %zext.0 = zext i16 %p.0 to i32 353 %zext.1 = zext i16 %p.1 to i32 354 %shl.1 = shl i32 %zext.1, 16 355 %or = or i32 %zext.0, %shl.1 356 ret i32 %or 357} 358 359; Should merge this to a dword store 360define amdgpu_kernel void @global_store_2xi16_align4(i16 addrspace(1)* %p, i16 addrspace(1)* %r) #0 { 361; GFX7-LABEL: global_store_2xi16_align4: 362; GFX7: ; %bb.0: 363; GFX7-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 364; GFX7-NEXT: v_mov_b32_e32 v2, 0x20001 365; GFX7-NEXT: s_waitcnt lgkmcnt(0) 366; GFX7-NEXT: v_mov_b32_e32 v0, s0 367; GFX7-NEXT: v_mov_b32_e32 v1, s1 368; GFX7-NEXT: flat_store_dword v[0:1], v2 369; GFX7-NEXT: s_endpgm 370; 371; GFX7-ALIGNED-LABEL: global_store_2xi16_align4: 372; GFX7-ALIGNED: ; %bb.0: 373; GFX7-ALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 374; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v2, 0x20001 375; GFX7-ALIGNED-NEXT: s_waitcnt lgkmcnt(0) 376; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, s0 377; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v1, s1 378; GFX7-ALIGNED-NEXT: flat_store_dword v[0:1], v2 379; GFX7-ALIGNED-NEXT: s_endpgm 380; 381; GFX7-UNALIGNED-LABEL: global_store_2xi16_align4: 382; GFX7-UNALIGNED: ; %bb.0: 383; GFX7-UNALIGNED-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 384; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v2, 0x20001 385; GFX7-UNALIGNED-NEXT: s_waitcnt lgkmcnt(0) 386; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, s0 387; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v1, s1 388; GFX7-UNALIGNED-NEXT: flat_store_dword v[0:1], v2 389; GFX7-UNALIGNED-NEXT: s_endpgm 390; 391; GFX9-LABEL: global_store_2xi16_align4: 392; GFX9: ; %bb.0: 393; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 394; GFX9-NEXT: v_mov_b32_e32 v0, 0 395; GFX9-NEXT: v_mov_b32_e32 v1, 0x20001 396; GFX9-NEXT: s_waitcnt lgkmcnt(0) 397; GFX9-NEXT: global_store_dword v0, v1, s[0:1] 398; GFX9-NEXT: s_endpgm 399; 400; GFX10-LABEL: global_store_2xi16_align4: 401; GFX10: ; %bb.0: 402; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 403; GFX10-NEXT: v_mov_b32_e32 v0, 0 404; GFX10-NEXT: v_mov_b32_e32 v1, 0x20001 405; GFX10-NEXT: s_waitcnt lgkmcnt(0) 406; GFX10-NEXT: global_store_dword v0, v1, s[0:1] 407; GFX10-NEXT: s_endpgm 408; 409; GFX11-LABEL: global_store_2xi16_align4: 410; GFX11: ; %bb.0: 411; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x8 412; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x20001 413; GFX11-NEXT: s_waitcnt lgkmcnt(0) 414; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] 415; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) 416; GFX11-NEXT: s_endpgm 417 %gep.r = getelementptr i16, i16 addrspace(1)* %r, i64 1 418 store i16 1, i16 addrspace(1)* %r, align 4 419 store i16 2, i16 addrspace(1)* %gep.r, align 2 420 ret void 421} 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455