144 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=amdgcn -mcpu=tonga -show-mc-encoding | FileCheck -enable-var-scope -check-prefixes=PREGFX10-UNPACKED %s3; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s4; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s5; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 | FileCheck -enable-var-scope -check-prefixes=GFX10-PACKED %s6; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s7; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s8 9define amdgpu_ps half @tbuffer_load_d16_x(ptr addrspace(8) inreg %rsrc) {10; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_x:11; PREGFX10-UNPACKED: ; %bb.0: ; %main_body12; PREGFX10-UNPACKED-NEXT: tbuffer_load_format_d16_x v0, off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] ; encoding: [0x00,0x00,0xb4,0xe8,0x00,0x00,0x00,0x80]13; PREGFX10-UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]14; PREGFX10-UNPACKED-NEXT: ; return to shader part epilog15;16; PREGFX10-PACKED-LABEL: tbuffer_load_d16_x:17; PREGFX10-PACKED: ; %bb.0: ; %main_body18; PREGFX10-PACKED-NEXT: tbuffer_load_format_d16_x v0, off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM]19; PREGFX10-PACKED-NEXT: s_waitcnt vmcnt(0)20; PREGFX10-PACKED-NEXT: ; return to shader part epilog21;22; GFX10-PACKED-LABEL: tbuffer_load_d16_x:23; GFX10-PACKED: ; %bb.0: ; %main_body24; GFX10-PACKED-NEXT: tbuffer_load_format_d16_x v0, off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]25; GFX10-PACKED-NEXT: s_waitcnt vmcnt(0)26; GFX10-PACKED-NEXT: ; return to shader part epilog27;28; GFX11-PACKED-LABEL: tbuffer_load_d16_x:29; GFX11-PACKED: ; %bb.0: ; %main_body30; GFX11-PACKED-NEXT: tbuffer_load_d16_format_x v0, off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]31; GFX11-PACKED-NEXT: s_waitcnt vmcnt(0)32; GFX11-PACKED-NEXT: ; return to shader part epilog33main_body:34 %data = call half @llvm.amdgcn.raw.ptr.tbuffer.load.f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 22, i32 0)35 ret half %data36}37 38define amdgpu_ps half @tbuffer_load_d16_xy(ptr addrspace(8) inreg %rsrc) {39; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_xy:40; PREGFX10-UNPACKED: ; %bb.0: ; %main_body41; PREGFX10-UNPACKED-NEXT: tbuffer_load_format_d16_xy v[0:1], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] ; encoding: [0x00,0x80,0xb4,0xe8,0x00,0x00,0x00,0x80]42; PREGFX10-UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]43; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, v1 ; encoding: [0x01,0x03,0x00,0x7e]44; PREGFX10-UNPACKED-NEXT: ; return to shader part epilog45;46; PREGFX10-PACKED-LABEL: tbuffer_load_d16_xy:47; PREGFX10-PACKED: ; %bb.0: ; %main_body48; PREGFX10-PACKED-NEXT: tbuffer_load_format_d16_xy v0, off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM]49; PREGFX10-PACKED-NEXT: s_waitcnt vmcnt(0)50; PREGFX10-PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v051; PREGFX10-PACKED-NEXT: ; return to shader part epilog52;53; GFX10-PACKED-LABEL: tbuffer_load_d16_xy:54; GFX10-PACKED: ; %bb.0: ; %main_body55; GFX10-PACKED-NEXT: tbuffer_load_format_d16_xy v0, off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]56; GFX10-PACKED-NEXT: s_waitcnt vmcnt(0)57; GFX10-PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v058; GFX10-PACKED-NEXT: ; return to shader part epilog59;60; GFX11-PACKED-LABEL: tbuffer_load_d16_xy:61; GFX11-PACKED: ; %bb.0: ; %main_body62; GFX11-PACKED-NEXT: tbuffer_load_d16_format_xy v0, off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]63; GFX11-PACKED-NEXT: s_waitcnt vmcnt(0)64; GFX11-PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v065; GFX11-PACKED-NEXT: ; return to shader part epilog66main_body:67 %data = call <2 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v2f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 22, i32 0)68 %elt = extractelement <2 x half> %data, i32 169 ret half %elt70}71 72define amdgpu_ps half @tbuffer_load_d16_xyz(ptr addrspace(8) inreg %rsrc) {73; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_xyz:74; PREGFX10-UNPACKED: ; %bb.0: ; %main_body75; PREGFX10-UNPACKED-NEXT: tbuffer_load_format_d16_xyz v[0:2], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] ; encoding: [0x00,0x00,0xb5,0xe8,0x00,0x00,0x00,0x80]76; PREGFX10-UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]77; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, v2 ; encoding: [0x02,0x03,0x00,0x7e]78; PREGFX10-UNPACKED-NEXT: ; return to shader part epilog79;80; PREGFX10-PACKED-LABEL: tbuffer_load_d16_xyz:81; PREGFX10-PACKED: ; %bb.0: ; %main_body82; PREGFX10-PACKED-NEXT: tbuffer_load_format_d16_xyz v[0:1], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM]83; PREGFX10-PACKED-NEXT: s_waitcnt vmcnt(0)84; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v0, v185; PREGFX10-PACKED-NEXT: ; return to shader part epilog86;87; GFX10-PACKED-LABEL: tbuffer_load_d16_xyz:88; GFX10-PACKED: ; %bb.0: ; %main_body89; GFX10-PACKED-NEXT: tbuffer_load_format_d16_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]90; GFX10-PACKED-NEXT: s_waitcnt vmcnt(0)91; GFX10-PACKED-NEXT: v_mov_b32_e32 v0, v192; GFX10-PACKED-NEXT: ; return to shader part epilog93;94; GFX11-PACKED-LABEL: tbuffer_load_d16_xyz:95; GFX11-PACKED: ; %bb.0: ; %main_body96; GFX11-PACKED-NEXT: tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]97; GFX11-PACKED-NEXT: s_waitcnt vmcnt(0)98; GFX11-PACKED-NEXT: v_mov_b32_e32 v0, v199; GFX11-PACKED-NEXT: ; return to shader part epilog100main_body:101 %data = call <3 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v3f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 22, i32 0)102 %elt = extractelement <3 x half> %data, i32 2103 ret half %elt104}105 106define amdgpu_ps half @tbuffer_load_d16_xyzw(ptr addrspace(8) inreg %rsrc) {107; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_xyzw:108; PREGFX10-UNPACKED: ; %bb.0: ; %main_body109; PREGFX10-UNPACKED-NEXT: tbuffer_load_format_d16_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] ; encoding: [0x00,0x80,0xb5,0xe8,0x00,0x00,0x00,0x80]110; PREGFX10-UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]111; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, v3 ; encoding: [0x03,0x03,0x00,0x7e]112; PREGFX10-UNPACKED-NEXT: ; return to shader part epilog113;114; PREGFX10-PACKED-LABEL: tbuffer_load_d16_xyzw:115; PREGFX10-PACKED: ; %bb.0: ; %main_body116; PREGFX10-PACKED-NEXT: tbuffer_load_format_d16_xyzw v[0:1], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM]117; PREGFX10-PACKED-NEXT: s_waitcnt vmcnt(0)118; PREGFX10-PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v1119; PREGFX10-PACKED-NEXT: ; return to shader part epilog120;121; GFX10-PACKED-LABEL: tbuffer_load_d16_xyzw:122; GFX10-PACKED: ; %bb.0: ; %main_body123; GFX10-PACKED-NEXT: tbuffer_load_format_d16_xyzw v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]124; GFX10-PACKED-NEXT: s_waitcnt vmcnt(0)125; GFX10-PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v1126; GFX10-PACKED-NEXT: ; return to shader part epilog127;128; GFX11-PACKED-LABEL: tbuffer_load_d16_xyzw:129; GFX11-PACKED: ; %bb.0: ; %main_body130; GFX11-PACKED-NEXT: tbuffer_load_d16_format_xyzw v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]131; GFX11-PACKED-NEXT: s_waitcnt vmcnt(0)132; GFX11-PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v1133; GFX11-PACKED-NEXT: ; return to shader part epilog134main_body:135 %data = call <4 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 22, i32 0)136 %elt = extractelement <4 x half> %data, i32 3137 ret half %elt138}139 140declare half @llvm.amdgcn.raw.ptr.tbuffer.load.f16(ptr addrspace(8), i32, i32, i32, i32)141declare <2 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v2f16(ptr addrspace(8), i32, i32, i32, i32)142declare <3 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v3f16(ptr addrspace(8), i32, i32, i32, i32)143declare <4 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f16(ptr addrspace(8), i32, i32, i32, i32)144