src/amd/llvm/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/Core.h>
  30 #include "compiler/nir/nir.h"
  31 #include "amd_family.h"
  32 #include "ac_shader_util.h"
  33 #include "ac_shader_args.h"
  34 #include "ac_shader_abi.h"
  35
  36 #ifdef __cplusplus
  37 extern "C" {
  38 #endif
  39
  40 enum {
  41         AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
  42         AC_ADDR_SPACE_GLOBAL = 1,
  43         AC_ADDR_SPACE_GDS = 2,
  44         AC_ADDR_SPACE_LDS = 3,
  45         AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
  46         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  47 };
  48
  49 #define AC_WAIT_LGKM    (1 << 0) /* LDS, GDS, constant, message */
  50 #define AC_WAIT_VLOAD   (1 << 1) /* VMEM load/sample instructions */
  51 #define AC_WAIT_VSTORE  (1 << 2) /* VMEM store instructions */
  52
  53 struct ac_llvm_flow;
  54 struct ac_llvm_compiler;
  55 enum ac_float_mode;
  56
  57 struct ac_llvm_flow_state {
  58         struct ac_llvm_flow *stack;
  59         unsigned depth_max;
  60         unsigned depth;
  61 };
  62
  63 struct ac_llvm_context {
  64         LLVMContextRef context;
  65         LLVMModuleRef module;
  66         LLVMBuilderRef builder;
  67
  68         LLVMValueRef main_function;
  69
  70         LLVMTypeRef voidt;
  71         LLVMTypeRef i1;
  72         LLVMTypeRef i8;
  73         LLVMTypeRef i16;
  74         LLVMTypeRef i32;
  75         LLVMTypeRef i64;
  76         LLVMTypeRef i128;
  77         LLVMTypeRef intptr;
  78         LLVMTypeRef f16;
  79         LLVMTypeRef f32;
  80         LLVMTypeRef f64;
  81         LLVMTypeRef v2i16;
  82         LLVMTypeRef v2i32;
  83         LLVMTypeRef v3i32;
  84         LLVMTypeRef v4i32;
  85         LLVMTypeRef v2f32;
  86         LLVMTypeRef v3f32;
  87         LLVMTypeRef v4f32;
  88         LLVMTypeRef v8i32;
  89         LLVMTypeRef iN_wavemask;
  90         LLVMTypeRef iN_ballotmask;
  91
  92         LLVMValueRef i8_0;
  93         LLVMValueRef i8_1;
  94         LLVMValueRef i16_0;
  95         LLVMValueRef i16_1;
  96         LLVMValueRef i32_0;
  97         LLVMValueRef i32_1;
  98         LLVMValueRef i64_0;
  99         LLVMValueRef i64_1;
 100         LLVMValueRef i128_0;
 101         LLVMValueRef i128_1;
 102         LLVMValueRef f16_0;
 103         LLVMValueRef f16_1;
 104         LLVMValueRef f32_0;
 105         LLVMValueRef f32_1;
 106         LLVMValueRef f64_0;
 107         LLVMValueRef f64_1;
 108         LLVMValueRef i1true;
 109         LLVMValueRef i1false;
 110
 111         /* Temporary helper to implement demote_to_helper:
 112          * True = live lanes
 113          * False = demoted lanes
 114          */
 115         LLVMValueRef postponed_kill;
 116
 117         /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
 118          * are two ac_llvm_contexts. Declare a pointer here, so that the control
 119          * flow stack is shared by both ac_llvm_contexts.
 120          */
 121         struct ac_llvm_flow_state *flow;
 122
 123         unsigned range_md_kind;
 124         unsigned invariant_load_md_kind;
 125         unsigned uniform_md_kind;
 126         unsigned fpmath_md_kind;
 127         LLVMValueRef fpmath_md_2p5_ulp;
 128         LLVMValueRef empty_md;
 129
 130         enum chip_class chip_class;
 131         enum radeon_family family;
 132
 133         unsigned wave_size;
 134         unsigned ballot_mask_bits;
 135
 136         unsigned float_mode;
 137
 138         LLVMValueRef lds;
 139 };
 140
 141 void
 142 ac_llvm_context_init(struct ac_llvm_context *ctx,
 143                      struct ac_llvm_compiler *compiler,
 144                      enum chip_class chip_class, enum radeon_family family,
 145                      enum ac_float_mode float_mode, unsigned wave_size,
 146                      unsigned ballot_mask_bits);
 147
 148 void
 149 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 150
 151 int
 152 ac_get_llvm_num_components(LLVMValueRef value);
 153
 154 int
 155 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 156
 157 LLVMValueRef
 158 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 159                      LLVMValueRef value,
 160                      int index);
 161
 162 unsigned ac_get_type_size(LLVMTypeRef type);
 163
 164 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 165 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 166 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 167 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 168 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 169
 170 LLVMValueRef
 171 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 172                    LLVMTypeRef return_type, LLVMValueRef *params,
 173                    unsigned param_count, unsigned attrib_mask);
 174
 175 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 176
 177 LLVMValueRef
 178 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 179              unsigned count_incoming, LLVMValueRef *values,
 180              LLVMBasicBlockRef *blocks);
 181
 182 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 183 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 184                                    LLVMValueRef *pvgpr);
 185
 186 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 187
 188 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 189 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 190                                  LLVMValueRef value);
 191
 192 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 193
 194 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 195
 196 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 197
 198 LLVMValueRef
 199 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 200                                unsigned value_count, unsigned component);
 201
 202 LLVMValueRef
 203 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 204                                 LLVMValueRef *values,
 205                                 unsigned value_count,
 206                                 unsigned value_stride,
 207                                 bool load,
 208                                 bool always_vector);
 209 LLVMValueRef
 210 ac_build_gather_values(struct ac_llvm_context *ctx,
 211                        LLVMValueRef *values,
 212                        unsigned value_count);
 213
 214 LLVMValueRef
 215 ac_extract_components(struct ac_llvm_context *ctx,
 216                       LLVMValueRef value,
 217                       unsigned start,
 218                       unsigned channels);
 219
 220 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 221                                      LLVMValueRef value,
 222                                      unsigned num_channels);
 223 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 224
 225 LLVMValueRef
 226 ac_build_fdiv(struct ac_llvm_context *ctx,
 227               LLVMValueRef num,
 228               LLVMValueRef den);
 229
 230 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 231                                 LLVMValueRef num,
 232                                 LLVMValueRef multiplier,
 233                                 LLVMValueRef pre_shift,
 234                                 LLVMValueRef post_shift,
 235                                 LLVMValueRef increment);
 236 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 237                                     LLVMValueRef num,
 238                                     LLVMValueRef multiplier,
 239                                     LLVMValueRef pre_shift,
 240                                     LLVMValueRef post_shift,
 241                                     LLVMValueRef increment);
 242 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 243                                               LLVMValueRef num,
 244                                               LLVMValueRef multiplier,
 245                                               LLVMValueRef post_shift);
 246
 247 void
 248 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 249                        bool is_deriv, bool is_array, bool is_lod,
 250                        LLVMValueRef *coords_arg,
 251                        LLVMValueRef *derivs_arg);
 252
 253
 254 LLVMValueRef
 255 ac_build_fs_interp(struct ac_llvm_context *ctx,
 256                    LLVMValueRef llvm_chan,
 257                    LLVMValueRef attr_number,
 258                    LLVMValueRef params,
 259                    LLVMValueRef i,
 260                    LLVMValueRef j);
 261
 262 LLVMValueRef
 263 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 264                        LLVMValueRef llvm_chan,
 265                        LLVMValueRef attr_number,
 266                        LLVMValueRef params,
 267                        LLVMValueRef i,
 268                        LLVMValueRef j);
 269
 270 LLVMValueRef
 271 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 272                        LLVMValueRef parameter,
 273                        LLVMValueRef llvm_chan,
 274                        LLVMValueRef attr_number,
 275                        LLVMValueRef params);
 276
 277 LLVMValueRef
 278 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 279                  LLVMValueRef base_ptr,
 280                  LLVMValueRef index);
 281
 282 LLVMValueRef
 283 ac_build_gep0(struct ac_llvm_context *ctx,
 284               LLVMValueRef base_ptr,
 285               LLVMValueRef index);
 286 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 287                                   LLVMValueRef index);
 288
 289 void
 290 ac_build_indexed_store(struct ac_llvm_context *ctx,
 291                        LLVMValueRef base_ptr, LLVMValueRef index,
 292                        LLVMValueRef value);
 293
 294 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 295                            LLVMValueRef index);
 296 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 297                                      LLVMValueRef base_ptr, LLVMValueRef index);
 298 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 299                                    LLVMValueRef base_ptr, LLVMValueRef index);
 300 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 301                                    LLVMValueRef base_ptr, LLVMValueRef index);
 302
 303 void
 304 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 305                             LLVMValueRef rsrc,
 306                             LLVMValueRef vdata,
 307                             unsigned num_channels,
 308                             LLVMValueRef voffset,
 309                             LLVMValueRef soffset,
 310                             unsigned inst_offset,
 311                             unsigned cache_policy);
 312
 313 void
 314 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 315                              LLVMValueRef rsrc,
 316                              LLVMValueRef data,
 317                              LLVMValueRef vindex,
 318                              LLVMValueRef voffset,
 319                              unsigned num_channels,
 320                              unsigned cache_policy);
 321
 322 LLVMValueRef
 323 ac_build_buffer_load(struct ac_llvm_context *ctx,
 324                      LLVMValueRef rsrc,
 325                      int num_channels,
 326                      LLVMValueRef vindex,
 327                      LLVMValueRef voffset,
 328                      LLVMValueRef soffset,
 329                      unsigned inst_offset,
 330                      unsigned cache_policy,
 331                      bool can_speculate,
 332                      bool allow_smem);
 333
 334 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 335                                          LLVMValueRef rsrc,
 336                                          LLVMValueRef vindex,
 337                                          LLVMValueRef voffset,
 338                                          unsigned num_channels,
 339                                          unsigned cache_policy,
 340                                          bool can_speculate);
 341
 342 LLVMValueRef
 343 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 344                             LLVMValueRef rsrc,
 345                             LLVMValueRef voffset,
 346                             LLVMValueRef soffset,
 347                             LLVMValueRef immoffset,
 348                             unsigned cache_policy);
 349
 350 LLVMValueRef
 351 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 352                            LLVMValueRef rsrc,
 353                            LLVMValueRef voffset,
 354                            LLVMValueRef soffset,
 355                            LLVMValueRef immoffset,
 356                            unsigned cache_policy);
 357
 358 LLVMValueRef
 359 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 360                              LLVMValueRef rsrc,
 361                              LLVMValueRef vindex,
 362                              LLVMValueRef voffset,
 363                              LLVMValueRef soffset,
 364                              LLVMValueRef immoffset,
 365                              unsigned num_channels,
 366                              unsigned dfmt,
 367                              unsigned nfmt,
 368                              unsigned cache_policy,
 369                              bool can_speculate);
 370
 371 LLVMValueRef
 372 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 373                           LLVMValueRef rsrc,
 374                           LLVMValueRef voffset,
 375                           LLVMValueRef soffset,
 376                           LLVMValueRef immoffset,
 377                           unsigned num_channels,
 378                           unsigned dfmt,
 379                           unsigned nfmt,
 380                           unsigned cache_policy,
 381                           bool can_speculate);
 382
 383 /* For ac_build_fetch_format.
 384  *
 385  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 386  */
 387 enum {
 388         AC_FETCH_FORMAT_FLOAT = 0,
 389         AC_FETCH_FORMAT_FIXED,
 390         AC_FETCH_FORMAT_UNORM,
 391         AC_FETCH_FORMAT_SNORM,
 392         AC_FETCH_FORMAT_USCALED,
 393         AC_FETCH_FORMAT_SSCALED,
 394         AC_FETCH_FORMAT_UINT,
 395         AC_FETCH_FORMAT_SINT,
 396 };
 397
 398 LLVMValueRef
 399 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 400                                unsigned log_size,
 401                                unsigned num_channels,
 402                                unsigned format,
 403                                bool reverse,
 404                                bool known_aligned,
 405                                LLVMValueRef rsrc,
 406                                LLVMValueRef vindex,
 407                                LLVMValueRef voffset,
 408                                LLVMValueRef soffset,
 409                                unsigned cache_policy,
 410                                bool can_speculate);
 411
 412 void
 413 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 414                              LLVMValueRef rsrc,
 415                              LLVMValueRef vdata,
 416                              LLVMValueRef voffset,
 417                              LLVMValueRef soffset,
 418                              unsigned cache_policy);
 419
 420 void
 421 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 422                             LLVMValueRef rsrc,
 423                             LLVMValueRef vdata,
 424                             LLVMValueRef voffset,
 425                             LLVMValueRef soffset,
 426                             unsigned cache_policy);
 427
 428 void
 429 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 430                               LLVMValueRef rsrc,
 431                               LLVMValueRef vdata,
 432                               LLVMValueRef vindex,
 433                               LLVMValueRef voffset,
 434                               LLVMValueRef soffset,
 435                               LLVMValueRef immoffset,
 436                               unsigned num_channels,
 437                               unsigned dfmt,
 438                               unsigned nfmt,
 439                               unsigned cache_policy);
 440
 441 void
 442 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 443                            LLVMValueRef rsrc,
 444                            LLVMValueRef vdata,
 445                            LLVMValueRef voffset,
 446                            LLVMValueRef soffset,
 447                            LLVMValueRef immoffset,
 448                            unsigned num_channels,
 449                            unsigned dfmt,
 450                            unsigned nfmt,
 451                            unsigned cache_policy);
 452
 453 LLVMValueRef
 454 ac_get_thread_id(struct ac_llvm_context *ctx);
 455
 456 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 457 #define AC_TID_MASK_TOP      0xfffffffd
 458 #define AC_TID_MASK_LEFT     0xfffffffe
 459
 460 LLVMValueRef
 461 ac_build_ddxy(struct ac_llvm_context *ctx,
 462               uint32_t mask,
 463               int idx,
 464               LLVMValueRef val);
 465
 466 #define AC_SENDMSG_GS 2
 467 #define AC_SENDMSG_GS_DONE 3
 468 #define AC_SENDMSG_GS_ALLOC_REQ 9
 469
 470 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 471 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 472 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 473 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 474
 475 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 476                       uint32_t msg,
 477                       LLVMValueRef wave_id);
 478
 479 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 480                            LLVMValueRef arg,
 481                            LLVMTypeRef dst_type);
 482
 483 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 484                           LLVMValueRef arg,
 485                           LLVMTypeRef dst_type);
 486 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 487                            LLVMValueRef b);
 488 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 489                            LLVMValueRef b);
 490 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 491                            LLVMValueRef b);
 492 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 493                            LLVMValueRef b);
 494 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 495 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 496 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 497
 498 struct ac_export_args {
 499         LLVMValueRef out[4];
 500         unsigned target;
 501         unsigned enabled_channels;
 502         bool compr;
 503         bool done;
 504         bool valid_mask;
 505 };
 506
 507 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 508
 509 void ac_build_export_null(struct ac_llvm_context *ctx);
 510
 511 enum ac_image_opcode {
 512         ac_image_sample,
 513         ac_image_gather4,
 514         ac_image_load,
 515         ac_image_load_mip,
 516         ac_image_store,
 517         ac_image_store_mip,
 518         ac_image_get_lod,
 519         ac_image_get_resinfo,
 520         ac_image_atomic,
 521         ac_image_atomic_cmpswap,
 522 };
 523
 524 enum ac_atomic_op {
 525         ac_atomic_swap,
 526         ac_atomic_add,
 527         ac_atomic_sub,
 528         ac_atomic_smin,
 529         ac_atomic_umin,
 530         ac_atomic_smax,
 531         ac_atomic_umax,
 532         ac_atomic_and,
 533         ac_atomic_or,
 534         ac_atomic_xor,
 535         ac_atomic_inc_wrap,
 536         ac_atomic_dec_wrap,
 537 };
 538
 539 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 540 enum ac_image_cache_policy {
 541         ac_glc = 1 << 0, /* per-CU cache control */
 542         ac_slc = 1 << 1, /* global L2 cache control */
 543         ac_dlc = 1 << 2, /* per-shader-array cache control */
 544         ac_swizzled = 1 << 3, /* the access is swizzled, disabling load/store merging */
 545 };
 546
 547 struct ac_image_args {
 548         enum ac_image_opcode opcode : 4;
 549         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 550         enum ac_image_dim dim : 3;
 551         unsigned dmask : 4;
 552         unsigned cache_policy : 3;
 553         bool unorm : 1;
 554         bool level_zero : 1;
 555         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 556
 557         LLVMValueRef resource;
 558         LLVMValueRef sampler;
 559         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 560         LLVMValueRef offset;
 561         LLVMValueRef bias;
 562         LLVMValueRef compare;
 563         LLVMValueRef derivs[6];
 564         LLVMValueRef coords[4];
 565         LLVMValueRef lod; // also used by ac_image_get_resinfo
 566 };
 567
 568 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 569                                    struct ac_image_args *a);
 570 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
 571                                              LLVMValueRef rsrc);
 572 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 573                                     LLVMValueRef args[2]);
 574 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 575                                      LLVMValueRef args[2]);
 576 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 577                                      LLVMValueRef args[2]);
 578 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 579                                  LLVMValueRef args[2], unsigned bits, bool hi);
 580 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 581                                  LLVMValueRef args[2], unsigned bits, bool hi);
 582 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 583 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 584 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 585                           LLVMValueRef offset, LLVMValueRef width,
 586                           bool is_signed);
 587 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 588                            LLVMValueRef s1, LLVMValueRef s2);
 589 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 590                            LLVMValueRef s1, LLVMValueRef s2);
 591
 592 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 593
 594 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 595                            unsigned bitsize);
 596
 597 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 598                             LLVMValueRef src1, LLVMValueRef src2,
 599                             unsigned bitsize);
 600
 601 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 602                             unsigned bitsize);
 603
 604 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 605                             unsigned bitsize);
 606
 607 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 608
 609 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 610                                        LLVMValueRef src0);
 611
 612 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 613                             LLVMValueRef main_fn,
 614                             uint8_t *vs_output_param_offset,
 615                             uint32_t num_outputs,
 616                             uint8_t *num_param_exports);
 617 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 618
 619 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 620 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 621                          LLVMValueRef dw_addr);
 622 void ac_lds_store(struct ac_llvm_context *ctx,
 623                   LLVMValueRef dw_addr, LLVMValueRef value);
 624
 625 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 626                          LLVMTypeRef dst_type,
 627                          LLVMValueRef src0);
 628
 629 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 630 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 631
 632 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 633 void ac_build_break(struct ac_llvm_context *ctx);
 634 void ac_build_continue(struct ac_llvm_context *ctx);
 635 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 636 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 637 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 638 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 639 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 640                  int lable_id);
 641 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 642                   int lable_id);
 643
 644 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 645                              const char *name);
 646 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 647                                    const char *name);
 648
 649 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 650                          LLVMTypeRef type);
 651
 652 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 653                             unsigned count);
 654
 655 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 656                              unsigned rshift, unsigned bitwidth);
 657
 658 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 659                               LLVMValueRef *addr, bool is_array_tex);
 660
 661 LLVMValueRef
 662 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 663
 664 LLVMValueRef ac_build_readlane_no_opt_barrier(struct ac_llvm_context *ctx,
 665                                               LLVMValueRef src, LLVMValueRef lane);
 666
 667 LLVMValueRef
 668 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 669
 670 LLVMValueRef
 671 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 672
 673 LLVMValueRef
 674 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 675
 676 LLVMValueRef
 677 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 678
 679 LLVMValueRef
 680 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 681
 682 LLVMValueRef
 683 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 684
 685 /**
 686  * Common arguments for a scan/reduce operation that accumulates per-wave
 687  * values across an entire workgroup, while respecting the order of waves.
 688  */
 689 struct ac_wg_scan {
 690         bool enable_reduce;
 691         bool enable_exclusive;
 692         bool enable_inclusive;
 693         nir_op op;
 694         LLVMValueRef src; /* clobbered! */
 695         LLVMValueRef result_reduce;
 696         LLVMValueRef result_exclusive;
 697         LLVMValueRef result_inclusive;
 698         LLVMValueRef extra;
 699         LLVMValueRef waveidx;
 700         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 701
 702         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 703         LLVMValueRef scratch;
 704         unsigned maxwaves;
 705 };
 706
 707 void
 708 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 709 void
 710 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 711 void
 712 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 713
 714 void
 715 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 716 void
 717 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 718 void
 719 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 720
 721 LLVMValueRef
 722 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 723                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 724
 725 LLVMValueRef
 726 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 727
 728 LLVMValueRef
 729 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 730                    unsigned bitsize);
 731
 732 LLVMValueRef
 733 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 734                     unsigned bitsize);
 735
 736 LLVMValueRef
 737 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
 738                       unsigned bitsize);
 739
 740 LLVMValueRef
 741 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 742
 743 LLVMValueRef
 744 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 745
 746 LLVMValueRef
 747 ac_build_is_helper_invocation(struct ac_llvm_context *ctx);
 748
 749 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 750                            LLVMValueRef *args, unsigned num_args);
 751
 752 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 753                                  LLVMValueRef ptr, LLVMValueRef val,
 754                                  const char *sync_scope);
 755
 756 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 757                                       LLVMValueRef cmp, LLVMValueRef val,
 758                                       const char *sync_scope);
 759
 760 void
 761 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
 762                 LLVMValueRef stencil, LLVMValueRef samplemask,
 763                 struct ac_export_args *args);
 764
 765 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
 766                                    LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt);
 767
 768 struct ac_ngg_prim {
 769         unsigned num_vertices;
 770         LLVMValueRef isnull;
 771         LLVMValueRef index[3];
 772         LLVMValueRef edgeflag[3];
 773         LLVMValueRef passthrough;
 774 };
 775
 776 LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
 777                                  const struct ac_ngg_prim *prim);
 778 void ac_build_export_prim(struct ac_llvm_context *ctx,
 779                           const struct ac_ngg_prim *prim);
 780
 781 static inline LLVMValueRef
 782 ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
 783 {
 784         assert(arg.used);
 785         return LLVMGetParam(ctx->main_function, arg.arg_index);
 786 }
 787
 788 enum ac_llvm_calling_convention {
 789         AC_LLVM_AMDGPU_VS = 87,
 790         AC_LLVM_AMDGPU_GS = 88,
 791         AC_LLVM_AMDGPU_PS = 89,
 792         AC_LLVM_AMDGPU_CS = 90,
 793         AC_LLVM_AMDGPU_HS = 93,
 794 };
 795
 796 LLVMValueRef ac_build_main(const struct ac_shader_args *args,
 797                            struct ac_llvm_context *ctx,
 798                            enum ac_llvm_calling_convention convention,
 799                            const char *name, LLVMTypeRef ret_type,
 800                            LLVMModuleRef module);
 801 void ac_build_s_endpgm(struct ac_llvm_context *ctx);
 802
 803 LLVMValueRef ac_prefix_bitcount(struct ac_llvm_context *ctx,
 804                                 LLVMValueRef mask, LLVMValueRef index);
 805 LLVMValueRef ac_prefix_bitcount_2x64(struct ac_llvm_context *ctx,
 806                                      LLVMValueRef mask[2], LLVMValueRef index);
 807 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context *ctx,
 808                                                  LLVMValueRef is_odd,
 809                                                  LLVMValueRef flatshade_first,
 810                                                  LLVMValueRef index[3]);
 811
 812 #ifdef __cplusplus
 813 }
 814 #endif
 815
 816 #endif