208 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-REAL16 %s3; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s4 5define amdgpu_kernel void @zextload_global_i8_to_i16(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {6; GFX11-REAL16-LABEL: zextload_global_i8_to_i16:7; GFX11-REAL16: ; %bb.0:8; GFX11-REAL16-NEXT: s_load_b128 s[0:3], s[4:5], 0x249; GFX11-REAL16-NEXT: v_mov_b32_e32 v1, 010; GFX11-REAL16-NEXT: s_waitcnt lgkmcnt(0)11; GFX11-REAL16-NEXT: global_load_d16_u8 v0, v1, s[2:3]12; GFX11-REAL16-NEXT: s_waitcnt vmcnt(0)13; GFX11-REAL16-NEXT: global_store_b16 v1, v0, s[0:1]14; GFX11-REAL16-NEXT: s_endpgm15;16; GFX11-FAKE16-LABEL: zextload_global_i8_to_i16:17; GFX11-FAKE16: ; %bb.0:18; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x2419; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, 020; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)21; GFX11-FAKE16-NEXT: global_load_u8 v1, v0, s[2:3]22; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)23; GFX11-FAKE16-NEXT: global_store_b16 v0, v1, s[0:1]24; GFX11-FAKE16-NEXT: s_endpgm25 %a = load i8, ptr addrspace(1) %in26 %ext = zext i8 %a to i1627 store i16 %ext, ptr addrspace(1) %out28 ret void29}30 31define amdgpu_kernel void @sextload_global_i8_to_i16(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {32; GFX11-REAL16-LABEL: sextload_global_i8_to_i16:33; GFX11-REAL16: ; %bb.0:34; GFX11-REAL16-NEXT: s_load_b128 s[0:3], s[4:5], 0x2435; GFX11-REAL16-NEXT: v_mov_b32_e32 v1, 036; GFX11-REAL16-NEXT: s_waitcnt lgkmcnt(0)37; GFX11-REAL16-NEXT: global_load_d16_i8 v0, v1, s[2:3]38; GFX11-REAL16-NEXT: s_waitcnt vmcnt(0)39; GFX11-REAL16-NEXT: global_store_b16 v1, v0, s[0:1]40; GFX11-REAL16-NEXT: s_endpgm41;42; GFX11-FAKE16-LABEL: sextload_global_i8_to_i16:43; GFX11-FAKE16: ; %bb.0:44; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x2445; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, 046; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)47; GFX11-FAKE16-NEXT: global_load_i8 v1, v0, s[2:3]48; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)49; GFX11-FAKE16-NEXT: global_store_b16 v0, v1, s[0:1]50; GFX11-FAKE16-NEXT: s_endpgm51 %a = load i8, ptr addrspace(1) %in52 %ext = sext i8 %a to i1653 store i16 %ext, ptr addrspace(1) %out54 ret void55}56 57define amdgpu_kernel void @zextload_global_i8_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {58; GFX11-REAL16-LABEL: zextload_global_i8_to_i64:59; GFX11-REAL16: ; %bb.0:60; GFX11-REAL16-NEXT: s_load_b128 s[0:3], s[4:5], 0x2461; GFX11-REAL16-NEXT: v_mov_b32_e32 v1, 062; GFX11-REAL16-NEXT: s_waitcnt lgkmcnt(0)63; GFX11-REAL16-NEXT: global_load_d16_u8 v0, v1, s[2:3]64; GFX11-REAL16-NEXT: s_waitcnt vmcnt(0)65; GFX11-REAL16-NEXT: v_and_b32_e32 v0, 0xffff, v066; GFX11-REAL16-NEXT: global_store_b64 v1, v[0:1], s[0:1]67; GFX11-REAL16-NEXT: s_endpgm68;69; GFX11-FAKE16-LABEL: zextload_global_i8_to_i64:70; GFX11-FAKE16: ; %bb.0:71; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x2472; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, 073; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)74; GFX11-FAKE16-NEXT: global_load_u8 v0, v1, s[2:3]75; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)76; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v077; GFX11-FAKE16-NEXT: global_store_b64 v1, v[0:1], s[0:1]78; GFX11-FAKE16-NEXT: s_endpgm79 %a = load i8, ptr addrspace(1) %in80 %ext = zext i8 %a to i6481 store i64 %ext, ptr addrspace(1) %out82 ret void83}84 85define amdgpu_kernel void @sextload_global_i8_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {86; GFX11-REAL16-LABEL: sextload_global_i8_to_i64:87; GFX11-REAL16: ; %bb.0:88; GFX11-REAL16-NEXT: s_load_b128 s[0:3], s[4:5], 0x2489; GFX11-REAL16-NEXT: v_mov_b32_e32 v2, 090; GFX11-REAL16-NEXT: s_waitcnt lgkmcnt(0)91; GFX11-REAL16-NEXT: global_load_d16_i8 v0, v2, s[2:3]92; GFX11-REAL16-NEXT: s_waitcnt vmcnt(0)93; GFX11-REAL16-NEXT: v_bfe_i32 v0, v0, 0, 1694; GFX11-REAL16-NEXT: s_delay_alu instid0(VALU_DEP_1)95; GFX11-REAL16-NEXT: v_ashrrev_i32_e32 v1, 31, v096; GFX11-REAL16-NEXT: global_store_b64 v2, v[0:1], s[0:1]97; GFX11-REAL16-NEXT: s_endpgm98;99; GFX11-FAKE16-LABEL: sextload_global_i8_to_i64:100; GFX11-FAKE16: ; %bb.0:101; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24102; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, 0103; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)104; GFX11-FAKE16-NEXT: global_load_i8 v0, v2, s[2:3]105; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)106; GFX11-FAKE16-NEXT: v_bfe_i32 v0, v0, 0, 16107; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)108; GFX11-FAKE16-NEXT: v_ashrrev_i32_e32 v1, 31, v0109; GFX11-FAKE16-NEXT: global_store_b64 v2, v[0:1], s[0:1]110; GFX11-FAKE16-NEXT: s_endpgm111 %a = load i8, ptr addrspace(1) %in112 %ext = sext i8 %a to i64113 store i64 %ext, ptr addrspace(1) %out114 ret void115}116 117define amdgpu_kernel void @zextload_global_i16_to_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {118; GFX11-LABEL: zextload_global_i16_to_i32:119; GFX11: ; %bb.0:120; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24121; GFX11-NEXT: v_mov_b32_e32 v0, 0122; GFX11-NEXT: s_waitcnt lgkmcnt(0)123; GFX11-NEXT: global_load_u16 v1, v0, s[2:3]124; GFX11-NEXT: s_waitcnt vmcnt(0)125; GFX11-NEXT: global_store_b32 v0, v1, s[0:1]126; GFX11-NEXT: s_endpgm127 %a = load i16, ptr addrspace(1) %in128 %ext = zext i16 %a to i32129 store i32 %ext, ptr addrspace(1) %out130 ret void131}132 133define amdgpu_kernel void @sextload_global_i16_to_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {134; GFX11-LABEL: sextload_global_i16_to_i32:135; GFX11: ; %bb.0:136; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24137; GFX11-NEXT: v_mov_b32_e32 v0, 0138; GFX11-NEXT: s_waitcnt lgkmcnt(0)139; GFX11-NEXT: global_load_i16 v1, v0, s[2:3]140; GFX11-NEXT: s_waitcnt vmcnt(0)141; GFX11-NEXT: global_store_b32 v0, v1, s[0:1]142; GFX11-NEXT: s_endpgm143 %a = load i16, ptr addrspace(1) %in144 %ext = sext i16 %a to i32145 store i32 %ext, ptr addrspace(1) %out146 ret void147}148 149define amdgpu_kernel void @zextload_global_i16_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {150; GFX11-REAL16-LABEL: zextload_global_i16_to_i64:151; GFX11-REAL16: ; %bb.0:152; GFX11-REAL16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24153; GFX11-REAL16-NEXT: v_mov_b32_e32 v1, 0154; GFX11-REAL16-NEXT: s_waitcnt lgkmcnt(0)155; GFX11-REAL16-NEXT: global_load_d16_b16 v0, v1, s[2:3]156; GFX11-REAL16-NEXT: s_waitcnt vmcnt(0)157; GFX11-REAL16-NEXT: v_and_b32_e32 v0, 0xffff, v0158; GFX11-REAL16-NEXT: global_store_b64 v1, v[0:1], s[0:1]159; GFX11-REAL16-NEXT: s_endpgm160;161; GFX11-FAKE16-LABEL: zextload_global_i16_to_i64:162; GFX11-FAKE16: ; %bb.0:163; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24164; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, 0165; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)166; GFX11-FAKE16-NEXT: global_load_u16 v0, v1, s[2:3]167; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)168; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0169; GFX11-FAKE16-NEXT: global_store_b64 v1, v[0:1], s[0:1]170; GFX11-FAKE16-NEXT: s_endpgm171 %a = load i16, ptr addrspace(1) %in172 %ext = zext i16 %a to i64173 store i64 %ext, ptr addrspace(1) %out174 ret void175}176 177define amdgpu_kernel void @sextload_global_i16_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) nounwind {178; GFX11-REAL16-LABEL: sextload_global_i16_to_i64:179; GFX11-REAL16: ; %bb.0:180; GFX11-REAL16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24181; GFX11-REAL16-NEXT: v_mov_b32_e32 v2, 0182; GFX11-REAL16-NEXT: s_waitcnt lgkmcnt(0)183; GFX11-REAL16-NEXT: global_load_d16_b16 v0, v2, s[2:3]184; GFX11-REAL16-NEXT: s_waitcnt vmcnt(0)185; GFX11-REAL16-NEXT: v_bfe_i32 v0, v0, 0, 16186; GFX11-REAL16-NEXT: s_delay_alu instid0(VALU_DEP_1)187; GFX11-REAL16-NEXT: v_ashrrev_i32_e32 v1, 31, v0188; GFX11-REAL16-NEXT: global_store_b64 v2, v[0:1], s[0:1]189; GFX11-REAL16-NEXT: s_endpgm190;191; GFX11-FAKE16-LABEL: sextload_global_i16_to_i64:192; GFX11-FAKE16: ; %bb.0:193; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24194; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, 0195; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)196; GFX11-FAKE16-NEXT: global_load_u16 v0, v2, s[2:3]197; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)198; GFX11-FAKE16-NEXT: v_bfe_i32 v0, v0, 0, 16199; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)200; GFX11-FAKE16-NEXT: v_ashrrev_i32_e32 v1, 31, v0201; GFX11-FAKE16-NEXT: global_store_b64 v2, v[0:1], s[0:1]202; GFX11-FAKE16-NEXT: s_endpgm203 %a = load i16, ptr addrspace(1) %in204 %ext = sext i16 %a to i64205 store i64 %ext, ptr addrspace(1) %out206 ret void207}208