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