Home | History | Annotate | Download | only in AMDGPU
      1 //===-- SIIntrinsics.td - SI Intrinsic defs ----------------*- tablegen -*-===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // SI Intrinsic Definitions
     11 //
     12 //===----------------------------------------------------------------------===//
     13 
     14 
     15 let TargetPrefix = "SI", isTarget = 1 in {
     16 
     17   def int_SI_tid : Intrinsic <[llvm_i32_ty], [], [IntrNoMem]>;
     18   def int_SI_packf16 : Intrinsic <[llvm_i32_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
     19   def int_SI_export : Intrinsic <[], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_float_ty, llvm_float_ty, llvm_float_ty, llvm_float_ty], []>;
     20   def int_SI_load_const : Intrinsic <[llvm_float_ty], [llvm_anyint_ty, llvm_i32_ty], [IntrNoMem]>;
     21   def int_SI_vs_load_input : Intrinsic <[llvm_v4f32_ty], [llvm_anyint_ty, llvm_i16_ty, llvm_i32_ty], [IntrNoMem]> ;
     22 
     23   // Fully-flexible TBUFFER_STORE_FORMAT_* except for the ADDR64 bit, which is not exposed
     24   def int_SI_tbuffer_store : Intrinsic <
     25     [],
     26     [llvm_anyint_ty, // rsrc(SGPR)
     27      llvm_anyint_ty, // vdata(VGPR), overloaded for types i32, v2i32, v4i32
     28      llvm_i32_ty,    // num_channels(imm), selects opcode suffix: 1=X, 2=XY, 3=XYZ, 4=XYZW
     29      llvm_i32_ty,    // vaddr(VGPR)
     30      llvm_i32_ty,    // soffset(SGPR)
     31      llvm_i32_ty,    // inst_offset(imm)
     32      llvm_i32_ty,    // dfmt(imm)
     33      llvm_i32_ty,    // nfmt(imm)
     34      llvm_i32_ty,    // offen(imm)
     35      llvm_i32_ty,    // idxen(imm)
     36      llvm_i32_ty,    // glc(imm)
     37      llvm_i32_ty,    // slc(imm)
     38      llvm_i32_ty],   // tfe(imm)
     39     []>;
     40 
     41   // Fully-flexible BUFFER_LOAD_DWORD_* except for the ADDR64 bit, which is not exposed
     42   def int_SI_buffer_load_dword : Intrinsic <
     43     [llvm_anyint_ty], // vdata(VGPR), overloaded for types i32, v2i32, v4i32
     44     [llvm_anyint_ty,  // rsrc(SGPR)
     45      llvm_anyint_ty,  // vaddr(VGPR)
     46      llvm_i32_ty,     // soffset(SGPR)
     47      llvm_i32_ty,     // inst_offset(imm)
     48      llvm_i32_ty,     // offen(imm)
     49      llvm_i32_ty,     // idxen(imm)
     50      llvm_i32_ty,     // glc(imm)
     51      llvm_i32_ty,     // slc(imm)
     52      llvm_i32_ty],    // tfe(imm)
     53     [IntrReadArgMem]>;
     54 
     55   def int_SI_sendmsg : Intrinsic <[], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
     56 
     57   // Fully-flexible SAMPLE instruction.
     58   class SampleRaw : Intrinsic <
     59     [llvm_v4f32_ty],    // vdata(VGPR)
     60     [llvm_anyint_ty,    // vaddr(VGPR)
     61      llvm_v8i32_ty,     // rsrc(SGPR)
     62      llvm_v4i32_ty,     // sampler(SGPR)
     63      llvm_i32_ty,       // dmask(imm)
     64      llvm_i32_ty,       // unorm(imm)
     65      llvm_i32_ty,       // r128(imm)
     66      llvm_i32_ty,       // da(imm)
     67      llvm_i32_ty,       // glc(imm)
     68      llvm_i32_ty,       // slc(imm)
     69      llvm_i32_ty,       // tfe(imm)
     70      llvm_i32_ty],      // lwe(imm)
     71     [IntrNoMem]>;
     72 
     73   // Image instruction without a sampler.
     74   class Image : Intrinsic <
     75     [llvm_v4f32_ty],    // vdata(VGPR)
     76     [llvm_anyint_ty,    // vaddr(VGPR)
     77      llvm_v8i32_ty,     // rsrc(SGPR)
     78      llvm_i32_ty,       // dmask(imm)
     79      llvm_i32_ty,       // unorm(imm)
     80      llvm_i32_ty,       // r128(imm)
     81      llvm_i32_ty,       // da(imm)
     82      llvm_i32_ty,       // glc(imm)
     83      llvm_i32_ty,       // slc(imm)
     84      llvm_i32_ty,       // tfe(imm)
     85      llvm_i32_ty],      // lwe(imm)
     86     [IntrNoMem]>;
     87 
     88   // Basic sample
     89   def int_SI_image_sample : SampleRaw;
     90   def int_SI_image_sample_cl : SampleRaw;
     91   def int_SI_image_sample_d : SampleRaw;
     92   def int_SI_image_sample_d_cl : SampleRaw;
     93   def int_SI_image_sample_l : SampleRaw;
     94   def int_SI_image_sample_b : SampleRaw;
     95   def int_SI_image_sample_b_cl : SampleRaw;
     96   def int_SI_image_sample_lz : SampleRaw;
     97   def int_SI_image_sample_cd : SampleRaw;
     98   def int_SI_image_sample_cd_cl : SampleRaw;
     99 
    100   // Sample with comparison
    101   def int_SI_image_sample_c : SampleRaw;
    102   def int_SI_image_sample_c_cl : SampleRaw;
    103   def int_SI_image_sample_c_d : SampleRaw;
    104   def int_SI_image_sample_c_d_cl : SampleRaw;
    105   def int_SI_image_sample_c_l : SampleRaw;
    106   def int_SI_image_sample_c_b : SampleRaw;
    107   def int_SI_image_sample_c_b_cl : SampleRaw;
    108   def int_SI_image_sample_c_lz : SampleRaw;
    109   def int_SI_image_sample_c_cd : SampleRaw;
    110   def int_SI_image_sample_c_cd_cl : SampleRaw;
    111 
    112   // Sample with offsets
    113   def int_SI_image_sample_o : SampleRaw;
    114   def int_SI_image_sample_cl_o : SampleRaw;
    115   def int_SI_image_sample_d_o : SampleRaw;
    116   def int_SI_image_sample_d_cl_o : SampleRaw;
    117   def int_SI_image_sample_l_o : SampleRaw;
    118   def int_SI_image_sample_b_o : SampleRaw;
    119   def int_SI_image_sample_b_cl_o : SampleRaw;
    120   def int_SI_image_sample_lz_o : SampleRaw;
    121   def int_SI_image_sample_cd_o : SampleRaw;
    122   def int_SI_image_sample_cd_cl_o : SampleRaw;
    123 
    124   // Sample with comparison and offsets
    125   def int_SI_image_sample_c_o : SampleRaw;
    126   def int_SI_image_sample_c_cl_o : SampleRaw;
    127   def int_SI_image_sample_c_d_o : SampleRaw;
    128   def int_SI_image_sample_c_d_cl_o : SampleRaw;
    129   def int_SI_image_sample_c_l_o : SampleRaw;
    130   def int_SI_image_sample_c_b_o : SampleRaw;
    131   def int_SI_image_sample_c_b_cl_o : SampleRaw;
    132   def int_SI_image_sample_c_lz_o : SampleRaw;
    133   def int_SI_image_sample_c_cd_o : SampleRaw;
    134   def int_SI_image_sample_c_cd_cl_o : SampleRaw;
    135 
    136   // Basic gather4
    137   def int_SI_gather4 : SampleRaw;
    138   def int_SI_gather4_cl : SampleRaw;
    139   def int_SI_gather4_l : SampleRaw;
    140   def int_SI_gather4_b : SampleRaw;
    141   def int_SI_gather4_b_cl : SampleRaw;
    142   def int_SI_gather4_lz : SampleRaw;
    143 
    144   // Gather4 with comparison
    145   def int_SI_gather4_c : SampleRaw;
    146   def int_SI_gather4_c_cl : SampleRaw;
    147   def int_SI_gather4_c_l : SampleRaw;
    148   def int_SI_gather4_c_b : SampleRaw;
    149   def int_SI_gather4_c_b_cl : SampleRaw;
    150   def int_SI_gather4_c_lz : SampleRaw;
    151 
    152   // Gather4 with offsets
    153   def int_SI_gather4_o : SampleRaw;
    154   def int_SI_gather4_cl_o : SampleRaw;
    155   def int_SI_gather4_l_o : SampleRaw;
    156   def int_SI_gather4_b_o : SampleRaw;
    157   def int_SI_gather4_b_cl_o : SampleRaw;
    158   def int_SI_gather4_lz_o : SampleRaw;
    159 
    160   // Gather4 with comparison and offsets
    161   def int_SI_gather4_c_o : SampleRaw;
    162   def int_SI_gather4_c_cl_o : SampleRaw;
    163   def int_SI_gather4_c_l_o : SampleRaw;
    164   def int_SI_gather4_c_b_o : SampleRaw;
    165   def int_SI_gather4_c_b_cl_o : SampleRaw;
    166   def int_SI_gather4_c_lz_o : SampleRaw;
    167 
    168   def int_SI_getlod : SampleRaw;
    169 
    170   // Image instrinsics.
    171   def int_SI_image_load : Image;
    172   def int_SI_image_load_mip : Image;
    173   def int_SI_getresinfo : Image;
    174 
    175   // Deprecated image and sample intrinsics.
    176   class Sample : Intrinsic <[llvm_v4f32_ty], [llvm_anyvector_ty, llvm_v32i8_ty, llvm_anyint_ty, llvm_i32_ty], [IntrNoMem]>;
    177 
    178   def int_SI_sample : Sample;
    179   def int_SI_sampleb : Sample;
    180   def int_SI_sampled : Sample;
    181   def int_SI_samplel : Sample;
    182   def int_SI_imageload : Intrinsic <[llvm_v4i32_ty], [llvm_anyvector_ty, llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>;
    183   def int_SI_resinfo : Intrinsic <[llvm_v4i32_ty], [llvm_i32_ty, llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>;
    184 
    185   /* Interpolation Intrinsics */
    186 
    187   def int_SI_fs_constant : Intrinsic <[llvm_float_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
    188   def int_SI_fs_interp : Intrinsic <[llvm_float_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_v2i32_ty], [IntrNoMem]>;
    189 
    190   /* Control flow Intrinsics */
    191 
    192   def int_SI_if : Intrinsic<[llvm_i64_ty], [llvm_i1_ty, llvm_empty_ty], []>;
    193   def int_SI_else : Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_empty_ty], []>;
    194   def int_SI_break : Intrinsic<[llvm_i64_ty], [llvm_i64_ty], []>;
    195   def int_SI_if_break : Intrinsic<[llvm_i64_ty], [llvm_i1_ty, llvm_i64_ty], []>;
    196   def int_SI_else_break : Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], []>;
    197   def int_SI_loop : Intrinsic<[], [llvm_i64_ty, llvm_empty_ty], []>;
    198   def int_SI_end_cf : Intrinsic<[], [llvm_i64_ty], []>;
    199 }
    200