src/amd/common/ac_llvm_build.h

   1 /*
   2  * Copyright 2016 Bas Nieuwenhuizen
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  13  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  14  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  15  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  16  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  17  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  18  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  19  *
  20  * The above copyright notice and this permission notice (including the
  21  * next paragraph) shall be included in all copies or substantial portions
  22  * of the Software.
  23  *
  24  */
  25 #ifndef AC_LLVM_BUILD_H
  26 #define AC_LLVM_BUILD_H
  27
  28 #include <stdbool.h>
  29 #include <llvm-c/TargetMachine.h>
  30 #include "compiler/nir/nir.h"
  31 #include "amd_family.h"
  32
  33 #ifdef __cplusplus
  34 extern "C" {
  35 #endif
  36
  37 enum {
  38         AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
  39         AC_ADDR_SPACE_GLOBAL = 1,
  40         AC_ADDR_SPACE_GDS = 2,
  41         AC_ADDR_SPACE_LDS = 3,
  42         AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
  43         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  44 };
  45
  46 /* Combine these with & instead of |. */
  47 #define NOOP_WAITCNT    0xcf7f
  48 #define LGKM_CNT        0xc07f
  49 #define EXP_CNT         0xcf0f
  50 #define VM_CNT          0x0f70 /* On GFX9, vmcnt has 6 bits in [0:3] and [14:15] */
  51
  52 struct ac_llvm_flow;
  53
  54 struct ac_llvm_context {
  55         LLVMContextRef context;
  56         LLVMModuleRef module;
  57         LLVMBuilderRef builder;
  58
  59         LLVMTypeRef voidt;
  60         LLVMTypeRef i1;
  61         LLVMTypeRef i8;
  62         LLVMTypeRef i16;
  63         LLVMTypeRef i32;
  64         LLVMTypeRef i64;
  65         LLVMTypeRef intptr;
  66         LLVMTypeRef f16;
  67         LLVMTypeRef f32;
  68         LLVMTypeRef f64;
  69         LLVMTypeRef v2i16;
  70         LLVMTypeRef v2i32;
  71         LLVMTypeRef v3i32;
  72         LLVMTypeRef v4i32;
  73         LLVMTypeRef v2f32;
  74         LLVMTypeRef v4f32;
  75         LLVMTypeRef v8i32;
  76
  77         LLVMValueRef i8_0;
  78         LLVMValueRef i8_1;
  79         LLVMValueRef i16_0;
  80         LLVMValueRef i16_1;
  81         LLVMValueRef i32_0;
  82         LLVMValueRef i32_1;
  83         LLVMValueRef i64_0;
  84         LLVMValueRef i64_1;
  85         LLVMValueRef f16_0;
  86         LLVMValueRef f16_1;
  87         LLVMValueRef f32_0;
  88         LLVMValueRef f32_1;
  89         LLVMValueRef f64_0;
  90         LLVMValueRef f64_1;
  91         LLVMValueRef i1true;
  92         LLVMValueRef i1false;
  93
  94         struct ac_llvm_flow *flow;
  95         unsigned flow_depth;
  96         unsigned flow_depth_max;
  97
  98         unsigned range_md_kind;
  99         unsigned invariant_load_md_kind;
 100         unsigned uniform_md_kind;
 101         unsigned fpmath_md_kind;
 102         LLVMValueRef fpmath_md_2p5_ulp;
 103         LLVMValueRef empty_md;
 104
 105         enum chip_class chip_class;
 106         enum radeon_family family;
 107
 108         LLVMValueRef lds;
 109 };
 110
 111 void
 112 ac_llvm_context_init(struct ac_llvm_context *ctx,
 113                      enum chip_class chip_class, enum radeon_family family);
 114
 115 void
 116 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 117
 118 int
 119 ac_get_llvm_num_components(LLVMValueRef value);
 120
 121 int
 122 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 123
 124 LLVMValueRef
 125 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 126                      LLVMValueRef value,
 127                      int index);
 128
 129 unsigned ac_get_type_size(LLVMTypeRef type);
 130
 131 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 132 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 133 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 134 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 135 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 136
 137 LLVMValueRef
 138 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 139                    LLVMTypeRef return_type, LLVMValueRef *params,
 140                    unsigned param_count, unsigned attrib_mask);
 141
 142 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 143
 144 LLVMValueRef
 145 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 146              unsigned count_incoming, LLVMValueRef *values,
 147              LLVMBasicBlockRef *blocks);
 148
 149 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 150 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 151                                    LLVMValueRef *pvgpr);
 152
 153 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 154
 155 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 156
 157 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 158
 159 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 160
 161 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 162
 163 LLVMValueRef
 164 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 165                                unsigned value_count, unsigned component);
 166
 167 LLVMValueRef
 168 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 169                                 LLVMValueRef *values,
 170                                 unsigned value_count,
 171                                 unsigned value_stride,
 172                                 bool load,
 173                                 bool always_vector);
 174 LLVMValueRef
 175 ac_build_gather_values(struct ac_llvm_context *ctx,
 176                        LLVMValueRef *values,
 177                        unsigned value_count);
 178 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 179                                      LLVMValueRef value,
 180                                      unsigned num_channels);
 181 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 182
 183 LLVMValueRef
 184 ac_build_fdiv(struct ac_llvm_context *ctx,
 185               LLVMValueRef num,
 186               LLVMValueRef den);
 187
 188 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 189                                 LLVMValueRef num,
 190                                 LLVMValueRef multiplier,
 191                                 LLVMValueRef pre_shift,
 192                                 LLVMValueRef post_shift,
 193                                 LLVMValueRef increment);
 194 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 195                                     LLVMValueRef num,
 196                                     LLVMValueRef multiplier,
 197                                     LLVMValueRef pre_shift,
 198                                     LLVMValueRef post_shift,
 199                                     LLVMValueRef increment);
 200 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 201                                               LLVMValueRef num,
 202                                               LLVMValueRef multiplier,
 203                                               LLVMValueRef post_shift);
 204
 205 void
 206 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 207                        bool is_deriv, bool is_array, bool is_lod,
 208                        LLVMValueRef *coords_arg,
 209                        LLVMValueRef *derivs_arg);
 210
 211
 212 LLVMValueRef
 213 ac_build_fs_interp(struct ac_llvm_context *ctx,
 214                    LLVMValueRef llvm_chan,
 215                    LLVMValueRef attr_number,
 216                    LLVMValueRef params,
 217                    LLVMValueRef i,
 218                    LLVMValueRef j);
 219
 220 LLVMValueRef
 221 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 222                        LLVMValueRef llvm_chan,
 223                        LLVMValueRef attr_number,
 224                        LLVMValueRef params,
 225                        LLVMValueRef i,
 226                        LLVMValueRef j);
 227
 228 LLVMValueRef
 229 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 230                        LLVMValueRef parameter,
 231                        LLVMValueRef llvm_chan,
 232                        LLVMValueRef attr_number,
 233                        LLVMValueRef params);
 234
 235 LLVMValueRef
 236 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 237                  LLVMValueRef base_ptr,
 238                  LLVMValueRef index);
 239
 240 LLVMValueRef
 241 ac_build_gep0(struct ac_llvm_context *ctx,
 242               LLVMValueRef base_ptr,
 243               LLVMValueRef index);
 244 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 245                                   LLVMValueRef index);
 246
 247 void
 248 ac_build_indexed_store(struct ac_llvm_context *ctx,
 249                        LLVMValueRef base_ptr, LLVMValueRef index,
 250                        LLVMValueRef value);
 251
 252 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 253                            LLVMValueRef index);
 254 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 255                                      LLVMValueRef base_ptr, LLVMValueRef index);
 256 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 257                                    LLVMValueRef base_ptr, LLVMValueRef index);
 258 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 259                                    LLVMValueRef base_ptr, LLVMValueRef index);
 260
 261 void
 262 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 263                             LLVMValueRef rsrc,
 264                             LLVMValueRef vdata,
 265                             unsigned num_channels,
 266                             LLVMValueRef voffset,
 267                             LLVMValueRef soffset,
 268                             unsigned inst_offset,
 269                             bool glc,
 270                             bool slc,
 271                             bool writeonly_memory,
 272                             bool swizzle_enable_hint);
 273
 274 void
 275 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 276                              LLVMValueRef rsrc,
 277                              LLVMValueRef data,
 278                              LLVMValueRef vindex,
 279                              LLVMValueRef voffset,
 280                              unsigned num_channels,
 281                              bool glc,
 282                              bool writeonly_memory);
 283
 284 LLVMValueRef
 285 ac_build_buffer_load(struct ac_llvm_context *ctx,
 286                      LLVMValueRef rsrc,
 287                      int num_channels,
 288                      LLVMValueRef vindex,
 289                      LLVMValueRef voffset,
 290                      LLVMValueRef soffset,
 291                      unsigned inst_offset,
 292                      unsigned glc,
 293                      unsigned slc,
 294                      bool can_speculate,
 295                      bool allow_smem);
 296
 297 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 298                                          LLVMValueRef rsrc,
 299                                          LLVMValueRef vindex,
 300                                          LLVMValueRef voffset,
 301                                          unsigned num_channels,
 302                                          bool glc,
 303                                          bool can_speculate);
 304
 305 /* load_format that handles the stride & element count better if idxen is
 306  * disabled by LLVM. */
 307 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 308                                                   LLVMValueRef rsrc,
 309                                                   LLVMValueRef vindex,
 310                                                   LLVMValueRef voffset,
 311                                                   unsigned num_channels,
 312                                                   bool glc,
 313                                                   bool can_speculate);
 314
 315 LLVMValueRef
 316 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 317                             LLVMValueRef rsrc,
 318                             LLVMValueRef voffset,
 319                             LLVMValueRef soffset,
 320                             LLVMValueRef immoffset,
 321                             bool glc);
 322
 323 LLVMValueRef
 324 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 325                            LLVMValueRef rsrc,
 326                            LLVMValueRef voffset,
 327                            LLVMValueRef soffset,
 328                            LLVMValueRef immoffset,
 329                            bool glc);
 330
 331 LLVMValueRef
 332 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 333                              LLVMValueRef rsrc,
 334                              LLVMValueRef vindex,
 335                              LLVMValueRef voffset,
 336                              LLVMValueRef soffset,
 337                              LLVMValueRef immoffset,
 338                              unsigned num_channels,
 339                              unsigned dfmt,
 340                              unsigned nfmt,
 341                              bool glc,
 342                              bool slc,
 343                              bool can_speculate);
 344
 345 LLVMValueRef
 346 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 347                           LLVMValueRef rsrc,
 348                           LLVMValueRef voffset,
 349                           LLVMValueRef soffset,
 350                           LLVMValueRef immoffset,
 351                           unsigned num_channels,
 352                           unsigned dfmt,
 353                           unsigned nfmt,
 354                           bool glc,
 355                           bool slc,
 356                           bool can_speculate);
 357
 358 void
 359 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 360                              LLVMValueRef rsrc,
 361                              LLVMValueRef vdata,
 362                              LLVMValueRef voffset,
 363                              LLVMValueRef soffset,
 364                              bool glc,
 365                              bool writeonly_memory);
 366
 367 void
 368 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 369                             LLVMValueRef rsrc,
 370                             LLVMValueRef vdata,
 371                             LLVMValueRef voffset,
 372                             LLVMValueRef soffset,
 373                             bool glc,
 374                             bool writeonly_memory);
 375
 376 void
 377 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 378                               LLVMValueRef rsrc,
 379                               LLVMValueRef vdata,
 380                               LLVMValueRef vindex,
 381                               LLVMValueRef voffset,
 382                               LLVMValueRef soffset,
 383                               LLVMValueRef immoffset,
 384                               unsigned num_channels,
 385                               unsigned dfmt,
 386                               unsigned nfmt,
 387                               bool glc,
 388                               bool slc,
 389                               bool writeonly_memory);
 390
 391 void
 392 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 393                            LLVMValueRef rsrc,
 394                            LLVMValueRef vdata,
 395                            LLVMValueRef voffset,
 396                            LLVMValueRef soffset,
 397                            LLVMValueRef immoffset,
 398                            unsigned num_channels,
 399                            unsigned dfmt,
 400                            unsigned nfmt,
 401                            bool glc,
 402                            bool slc,
 403                            bool writeonly_memory);
 404
 405 LLVMValueRef
 406 ac_get_thread_id(struct ac_llvm_context *ctx);
 407
 408 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 409 #define AC_TID_MASK_TOP      0xfffffffd
 410 #define AC_TID_MASK_LEFT     0xfffffffe
 411
 412 LLVMValueRef
 413 ac_build_ddxy(struct ac_llvm_context *ctx,
 414               uint32_t mask,
 415               int idx,
 416               LLVMValueRef val);
 417
 418 #define AC_SENDMSG_GS 2
 419 #define AC_SENDMSG_GS_DONE 3
 420
 421 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 422 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 423 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 424 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 425
 426 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 427                       uint32_t msg,
 428                       LLVMValueRef wave_id);
 429
 430 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 431                            LLVMValueRef arg,
 432                            LLVMTypeRef dst_type);
 433
 434 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 435                           LLVMValueRef arg,
 436                           LLVMTypeRef dst_type);
 437 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 438                            LLVMValueRef b);
 439 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 440                            LLVMValueRef b);
 441 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 442                            LLVMValueRef b);
 443 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 444                            LLVMValueRef b);
 445 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 446 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 447 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 448
 449 struct ac_export_args {
 450         LLVMValueRef out[4];
 451         unsigned target;
 452         unsigned enabled_channels;
 453         bool compr;
 454         bool done;
 455         bool valid_mask;
 456 };
 457
 458 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 459
 460 void ac_build_export_null(struct ac_llvm_context *ctx);
 461
 462 enum ac_image_opcode {
 463         ac_image_sample,
 464         ac_image_gather4,
 465         ac_image_load,
 466         ac_image_load_mip,
 467         ac_image_store,
 468         ac_image_store_mip,
 469         ac_image_get_lod,
 470         ac_image_get_resinfo,
 471         ac_image_atomic,
 472         ac_image_atomic_cmpswap,
 473 };
 474
 475 enum ac_atomic_op {
 476         ac_atomic_swap,
 477         ac_atomic_add,
 478         ac_atomic_sub,
 479         ac_atomic_smin,
 480         ac_atomic_umin,
 481         ac_atomic_smax,
 482         ac_atomic_umax,
 483         ac_atomic_and,
 484         ac_atomic_or,
 485         ac_atomic_xor,
 486 };
 487
 488 enum ac_image_dim {
 489         ac_image_1d,
 490         ac_image_2d,
 491         ac_image_3d,
 492         ac_image_cube, // includes cube arrays
 493         ac_image_1darray,
 494         ac_image_2darray,
 495         ac_image_2dmsaa,
 496         ac_image_2darraymsaa,
 497 };
 498
 499 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 500 enum ac_image_cache_policy {
 501         ac_glc = 1 << 0,
 502         ac_slc = 1 << 1,
 503 };
 504
 505 struct ac_image_args {
 506         enum ac_image_opcode opcode : 4;
 507         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 508         enum ac_image_dim dim : 3;
 509         unsigned dmask : 4;
 510         unsigned cache_policy : 2;
 511         bool unorm : 1;
 512         bool level_zero : 1;
 513         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 514
 515         LLVMValueRef resource;
 516         LLVMValueRef sampler;
 517         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 518         LLVMValueRef offset;
 519         LLVMValueRef bias;
 520         LLVMValueRef compare;
 521         LLVMValueRef derivs[6];
 522         LLVMValueRef coords[4];
 523         LLVMValueRef lod; // also used by ac_image_get_resinfo
 524 };
 525
 526 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 527                                    struct ac_image_args *a);
 528 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 529                                     LLVMValueRef args[2]);
 530 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 531                                      LLVMValueRef args[2]);
 532 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 533                                      LLVMValueRef args[2]);
 534 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 535                                  LLVMValueRef args[2], unsigned bits, bool hi);
 536 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 537                                  LLVMValueRef args[2], unsigned bits, bool hi);
 538 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 539 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 540 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 541                           LLVMValueRef offset, LLVMValueRef width,
 542                           bool is_signed);
 543 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 544                            LLVMValueRef s1, LLVMValueRef s2);
 545 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 546                            LLVMValueRef s1, LLVMValueRef s2);
 547
 548 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
 549
 550 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 551                            unsigned bitsize);
 552
 553 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 554                             LLVMValueRef src1, LLVMValueRef src2,
 555                             unsigned bitsize);
 556
 557 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 558                             unsigned bitsize);
 559
 560 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 561                             unsigned bitsize);
 562
 563 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 564
 565 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 566                                        LLVMValueRef src0);
 567
 568 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 569                             LLVMValueRef main_fn,
 570                             uint8_t *vs_output_param_offset,
 571                             uint32_t num_outputs,
 572                             uint8_t *num_param_exports);
 573 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 574
 575 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 576 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 577                          LLVMValueRef dw_addr);
 578 void ac_lds_store(struct ac_llvm_context *ctx,
 579                   LLVMValueRef dw_addr, LLVMValueRef value);
 580
 581 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 582                          LLVMTypeRef dst_type,
 583                          LLVMValueRef src0);
 584
 585 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 586 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 587
 588 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 589 void ac_build_break(struct ac_llvm_context *ctx);
 590 void ac_build_continue(struct ac_llvm_context *ctx);
 591 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 592 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 593 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 594 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 595 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 596                  int lable_id);
 597 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 598                   int lable_id);
 599
 600 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 601                              const char *name);
 602 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 603                                    const char *name);
 604
 605 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 606                          LLVMTypeRef type);
 607
 608 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 609                             unsigned count);
 610
 611 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 612                              unsigned rshift, unsigned bitwidth);
 613
 614 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 615                               LLVMValueRef *addr, bool is_array_tex);
 616
 617 LLVMValueRef
 618 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 619
 620 LLVMValueRef
 621 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 622
 623 LLVMValueRef
 624 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 625
 626 LLVMValueRef
 627 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 628
 629 LLVMValueRef
 630 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 631
 632 LLVMValueRef
 633 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 634
 635 LLVMValueRef
 636 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 637
 638 /**
 639  * Common arguments for a scan/reduce operation that accumulates per-wave
 640  * values across an entire workgroup, while respecting the order of waves.
 641  */
 642 struct ac_wg_scan {
 643         bool enable_reduce;
 644         bool enable_exclusive;
 645         bool enable_inclusive;
 646         nir_op op;
 647         LLVMValueRef src; /* clobbered! */
 648         LLVMValueRef result_reduce;
 649         LLVMValueRef result_exclusive;
 650         LLVMValueRef result_inclusive;
 651         LLVMValueRef extra;
 652         LLVMValueRef waveidx;
 653         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 654
 655         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 656         LLVMValueRef scratch;
 657         unsigned maxwaves;
 658 };
 659
 660 void
 661 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 662 void
 663 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 664 void
 665 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 666
 667 void
 668 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 669 void
 670 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 671 void
 672 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 673
 674 LLVMValueRef
 675 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 676                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 677
 678 LLVMValueRef
 679 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 680
 681 LLVMValueRef
 682 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 683                    unsigned bitsize);
 684
 685 LLVMValueRef
 686 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 687                     unsigned bitsize);
 688
 689 LLVMValueRef
 690 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 691
 692 LLVMValueRef
 693 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 694
 695 #ifdef __cplusplus
 696 }
 697 #endif
 698
 699 #endif