87 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc < %s -mtriple=amdgcn -mcpu=tonga -show-mc-encoding | FileCheck -enable-var-scope -check-prefixes=UNPACKED %s3; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 | FileCheck -enable-var-scope -check-prefixes=PACKED %s4; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 | FileCheck -enable-var-scope -check-prefixes=PACKED %s5 6define amdgpu_ps half @buffer_load_format_d16_x(ptr addrspace(8) inreg %rsrc) {7; UNPACKED-LABEL: buffer_load_format_d16_x:8; UNPACKED: ; %bb.0: ; %main_body9; UNPACKED-NEXT: buffer_load_format_d16_x v0, off, s[0:3], 0 ; encoding: [0x00,0x00,0x20,0xe0,0x00,0x00,0x00,0x80]10; UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]11; UNPACKED-NEXT: ; return to shader part epilog12;13; PACKED-LABEL: buffer_load_format_d16_x:14; PACKED: ; %bb.0: ; %main_body15; PACKED-NEXT: buffer_load_format_d16_x v0, off, s[0:3], 016; PACKED-NEXT: s_waitcnt vmcnt(0)17; PACKED-NEXT: ; return to shader part epilog18main_body:19 %data = call half @llvm.amdgcn.raw.ptr.buffer.load.format.f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)20 ret half %data21}22 23define amdgpu_ps half @buffer_load_format_d16_xy(ptr addrspace(8) inreg %rsrc) {24; UNPACKED-LABEL: buffer_load_format_d16_xy:25; UNPACKED: ; %bb.0: ; %main_body26; UNPACKED-NEXT: buffer_load_format_d16_xy v[0:1], off, s[0:3], 0 ; encoding: [0x00,0x00,0x24,0xe0,0x00,0x00,0x00,0x80]27; UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]28; UNPACKED-NEXT: v_mov_b32_e32 v0, v1 ; encoding: [0x01,0x03,0x00,0x7e]29; UNPACKED-NEXT: ; return to shader part epilog30;31; PACKED-LABEL: buffer_load_format_d16_xy:32; PACKED: ; %bb.0: ; %main_body33; PACKED-NEXT: buffer_load_format_d16_xy v0, off, s[0:3], 034; PACKED-NEXT: s_waitcnt vmcnt(0)35; PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v036; PACKED-NEXT: ; return to shader part epilog37main_body:38 %data = call <2 x half> @llvm.amdgcn.raw.ptr.buffer.load.format.v2f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)39 %elt = extractelement <2 x half> %data, i32 140 ret half %elt41}42 43define amdgpu_ps half @buffer_load_format_d16_xyz(ptr addrspace(8) inreg %rsrc) {44; UNPACKED-LABEL: buffer_load_format_d16_xyz:45; UNPACKED: ; %bb.0: ; %main_body46; UNPACKED-NEXT: buffer_load_format_d16_xyz v[0:2], off, s[0:3], 0 ; encoding: [0x00,0x00,0x28,0xe0,0x00,0x00,0x00,0x80]47; UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]48; UNPACKED-NEXT: v_mov_b32_e32 v0, v2 ; encoding: [0x02,0x03,0x00,0x7e]49; UNPACKED-NEXT: ; return to shader part epilog50;51; PACKED-LABEL: buffer_load_format_d16_xyz:52; PACKED: ; %bb.0: ; %main_body53; PACKED-NEXT: buffer_load_format_d16_xyz v[0:1], off, s[0:3], 054; PACKED-NEXT: s_waitcnt vmcnt(0)55; PACKED-NEXT: v_mov_b32_e32 v0, v156; PACKED-NEXT: ; return to shader part epilog57main_body:58 %data = call <3 x half> @llvm.amdgcn.raw.ptr.buffer.load.format.v3f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)59 %elt = extractelement <3 x half> %data, i32 260 ret half %elt61}62 63define amdgpu_ps half @buffer_load_format_d16_xyzw(ptr addrspace(8) inreg %rsrc) {64; UNPACKED-LABEL: buffer_load_format_d16_xyzw:65; UNPACKED: ; %bb.0: ; %main_body66; UNPACKED-NEXT: buffer_load_format_d16_xyzw v[0:3], off, s[0:3], 0 ; encoding: [0x00,0x00,0x2c,0xe0,0x00,0x00,0x00,0x80]67; UNPACKED-NEXT: s_waitcnt vmcnt(0) ; encoding: [0x70,0x0f,0x8c,0xbf]68; UNPACKED-NEXT: v_mov_b32_e32 v0, v3 ; encoding: [0x03,0x03,0x00,0x7e]69; UNPACKED-NEXT: ; return to shader part epilog70;71; PACKED-LABEL: buffer_load_format_d16_xyzw:72; PACKED: ; %bb.0: ; %main_body73; PACKED-NEXT: buffer_load_format_d16_xyzw v[0:1], off, s[0:3], 074; PACKED-NEXT: s_waitcnt vmcnt(0)75; PACKED-NEXT: v_lshrrev_b32_e32 v0, 16, v176; PACKED-NEXT: ; return to shader part epilog77main_body:78 %data = call <4 x half> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)79 %elt = extractelement <4 x half> %data, i32 380 ret half %elt81}82 83declare half @llvm.amdgcn.raw.ptr.buffer.load.format.f16(ptr addrspace(8), i32, i32, i32)84declare <2 x half> @llvm.amdgcn.raw.ptr.buffer.load.format.v2f16(ptr addrspace(8), i32, i32, i32)85declare <3 x half> @llvm.amdgcn.raw.ptr.buffer.load.format.v3f16(ptr addrspace(8), i32, i32, i32)86declare <4 x half> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f16(ptr addrspace(8), i32, i32, i32)87