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         /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
 112          * are two ac_llvm_contexts. Declare a pointer here, so that the control
 113          * flow stack is shared by both ac_llvm_contexts.
 114          */
 115         struct ac_llvm_flow_state *flow;
 116
 117         unsigned range_md_kind;
 118         unsigned invariant_load_md_kind;
 119         unsigned uniform_md_kind;
 120         unsigned fpmath_md_kind;
 121         LLVMValueRef fpmath_md_2p5_ulp;
 122         LLVMValueRef empty_md;
 123
 124         enum chip_class chip_class;
 125         enum radeon_family family;
 126
 127         unsigned wave_size;
 128         unsigned ballot_mask_bits;
 129
 130         unsigned float_mode;
 131
 132         LLVMValueRef lds;
 133 };
 134
 135 void
 136 ac_llvm_context_init(struct ac_llvm_context *ctx,
 137                      struct ac_llvm_compiler *compiler,
 138                      enum chip_class chip_class, enum radeon_family family,
 139                      enum ac_float_mode float_mode, unsigned wave_size,
 140                      unsigned ballot_mask_bits);
 141
 142 void
 143 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 144
 145 int
 146 ac_get_llvm_num_components(LLVMValueRef value);
 147
 148 int
 149 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 150
 151 LLVMValueRef
 152 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 153                      LLVMValueRef value,
 154                      int index);
 155
 156 unsigned ac_get_type_size(LLVMTypeRef type);
 157
 158 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 159 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 160 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 161 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 162 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 163
 164 LLVMValueRef
 165 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 166                    LLVMTypeRef return_type, LLVMValueRef *params,
 167                    unsigned param_count, unsigned attrib_mask);
 168
 169 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 170
 171 LLVMValueRef
 172 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 173              unsigned count_incoming, LLVMValueRef *values,
 174              LLVMBasicBlockRef *blocks);
 175
 176 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 177 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 178                                    LLVMValueRef *pvgpr);
 179
 180 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 181
 182 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 183 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 184                                  LLVMValueRef value);
 185
 186 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 187
 188 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 189
 190 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 191
 192 LLVMValueRef
 193 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 194                                unsigned value_count, unsigned component);
 195
 196 LLVMValueRef
 197 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 198                                 LLVMValueRef *values,
 199                                 unsigned value_count,
 200                                 unsigned value_stride,
 201                                 bool load,
 202                                 bool always_vector);
 203 LLVMValueRef
 204 ac_build_gather_values(struct ac_llvm_context *ctx,
 205                        LLVMValueRef *values,
 206                        unsigned value_count);
 207
 208 LLVMValueRef
 209 ac_extract_components(struct ac_llvm_context *ctx,
 210                       LLVMValueRef value,
 211                       unsigned start,
 212                       unsigned channels);
 213
 214 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 215                                      LLVMValueRef value,
 216                                      unsigned num_channels);
 217 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 218
 219 LLVMValueRef
 220 ac_build_fdiv(struct ac_llvm_context *ctx,
 221               LLVMValueRef num,
 222               LLVMValueRef den);
 223
 224 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 225                                 LLVMValueRef num,
 226                                 LLVMValueRef multiplier,
 227                                 LLVMValueRef pre_shift,
 228                                 LLVMValueRef post_shift,
 229                                 LLVMValueRef increment);
 230 LLVMValueRef ac_build_fast_udiv_nuw(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_u31_d_not_one(struct ac_llvm_context *ctx,
 237                                               LLVMValueRef num,
 238                                               LLVMValueRef multiplier,
 239                                               LLVMValueRef post_shift);
 240
 241 void
 242 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 243                        bool is_deriv, bool is_array, bool is_lod,
 244                        LLVMValueRef *coords_arg,
 245                        LLVMValueRef *derivs_arg);
 246
 247
 248 LLVMValueRef
 249 ac_build_fs_interp(struct ac_llvm_context *ctx,
 250                    LLVMValueRef llvm_chan,
 251                    LLVMValueRef attr_number,
 252                    LLVMValueRef params,
 253                    LLVMValueRef i,
 254                    LLVMValueRef j);
 255
 256 LLVMValueRef
 257 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 258                        LLVMValueRef llvm_chan,
 259                        LLVMValueRef attr_number,
 260                        LLVMValueRef params,
 261                        LLVMValueRef i,
 262                        LLVMValueRef j);
 263
 264 LLVMValueRef
 265 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 266                        LLVMValueRef parameter,
 267                        LLVMValueRef llvm_chan,
 268                        LLVMValueRef attr_number,
 269                        LLVMValueRef params);
 270
 271 LLVMValueRef
 272 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 273                  LLVMValueRef base_ptr,
 274                  LLVMValueRef index);
 275
 276 LLVMValueRef
 277 ac_build_gep0(struct ac_llvm_context *ctx,
 278               LLVMValueRef base_ptr,
 279               LLVMValueRef index);
 280 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 281                                   LLVMValueRef index);
 282
 283 void
 284 ac_build_indexed_store(struct ac_llvm_context *ctx,
 285                        LLVMValueRef base_ptr, LLVMValueRef index,
 286                        LLVMValueRef value);
 287
 288 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 289                            LLVMValueRef index);
 290 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 291                                      LLVMValueRef base_ptr, LLVMValueRef index);
 292 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 293                                    LLVMValueRef base_ptr, LLVMValueRef index);
 294 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 295                                    LLVMValueRef base_ptr, LLVMValueRef index);
 296
 297 void
 298 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 299                             LLVMValueRef rsrc,
 300                             LLVMValueRef vdata,
 301                             unsigned num_channels,
 302                             LLVMValueRef voffset,
 303                             LLVMValueRef soffset,
 304                             unsigned inst_offset,
 305                             unsigned cache_policy);
 306
 307 void
 308 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 309                              LLVMValueRef rsrc,
 310                              LLVMValueRef data,
 311                              LLVMValueRef vindex,
 312                              LLVMValueRef voffset,
 313                              unsigned num_channels,
 314                              unsigned cache_policy);
 315
 316 LLVMValueRef
 317 ac_build_buffer_load(struct ac_llvm_context *ctx,
 318                      LLVMValueRef rsrc,
 319                      int num_channels,
 320                      LLVMValueRef vindex,
 321                      LLVMValueRef voffset,
 322                      LLVMValueRef soffset,
 323                      unsigned inst_offset,
 324                      unsigned cache_policy,
 325                      bool can_speculate,
 326                      bool allow_smem);
 327
 328 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 329                                          LLVMValueRef rsrc,
 330                                          LLVMValueRef vindex,
 331                                          LLVMValueRef voffset,
 332                                          unsigned num_channels,
 333                                          unsigned cache_policy,
 334                                          bool can_speculate);
 335
 336 LLVMValueRef
 337 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 338                             LLVMValueRef rsrc,
 339                             LLVMValueRef voffset,
 340                             LLVMValueRef soffset,
 341                             LLVMValueRef immoffset,
 342                             unsigned cache_policy);
 343
 344 LLVMValueRef
 345 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 346                            LLVMValueRef rsrc,
 347                            LLVMValueRef voffset,
 348                            LLVMValueRef soffset,
 349                            LLVMValueRef immoffset,
 350                            unsigned cache_policy);
 351
 352 LLVMValueRef
 353 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 354                              LLVMValueRef rsrc,
 355                              LLVMValueRef vindex,
 356                              LLVMValueRef voffset,
 357                              LLVMValueRef soffset,
 358                              LLVMValueRef immoffset,
 359                              unsigned num_channels,
 360                              unsigned dfmt,
 361                              unsigned nfmt,
 362                              unsigned cache_policy,
 363                              bool can_speculate);
 364
 365 LLVMValueRef
 366 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 367                           LLVMValueRef rsrc,
 368                           LLVMValueRef voffset,
 369                           LLVMValueRef soffset,
 370                           LLVMValueRef immoffset,
 371                           unsigned num_channels,
 372                           unsigned dfmt,
 373                           unsigned nfmt,
 374                           unsigned cache_policy,
 375                           bool can_speculate);
 376
 377 /* For ac_build_fetch_format.
 378  *
 379  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 380  */
 381 enum {
 382         AC_FETCH_FORMAT_FLOAT = 0,
 383         AC_FETCH_FORMAT_FIXED,
 384         AC_FETCH_FORMAT_UNORM,
 385         AC_FETCH_FORMAT_SNORM,
 386         AC_FETCH_FORMAT_USCALED,
 387         AC_FETCH_FORMAT_SSCALED,
 388         AC_FETCH_FORMAT_UINT,
 389         AC_FETCH_FORMAT_SINT,
 390 };
 391
 392 LLVMValueRef
 393 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 394                                unsigned log_size,
 395                                unsigned num_channels,
 396                                unsigned format,
 397                                bool reverse,
 398                                bool known_aligned,
 399                                LLVMValueRef rsrc,
 400                                LLVMValueRef vindex,
 401                                LLVMValueRef voffset,
 402                                LLVMValueRef soffset,
 403                                unsigned cache_policy,
 404                                bool can_speculate);
 405
 406 void
 407 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 408                              LLVMValueRef rsrc,
 409                              LLVMValueRef vdata,
 410                              LLVMValueRef voffset,
 411                              LLVMValueRef soffset,
 412                              unsigned cache_policy);
 413
 414 void
 415 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 416                             LLVMValueRef rsrc,
 417                             LLVMValueRef vdata,
 418                             LLVMValueRef voffset,
 419                             LLVMValueRef soffset,
 420                             unsigned cache_policy);
 421
 422 void
 423 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 424                               LLVMValueRef rsrc,
 425                               LLVMValueRef vdata,
 426                               LLVMValueRef vindex,
 427                               LLVMValueRef voffset,
 428                               LLVMValueRef soffset,
 429                               LLVMValueRef immoffset,
 430                               unsigned num_channels,
 431                               unsigned dfmt,
 432                               unsigned nfmt,
 433                               unsigned cache_policy);
 434
 435 void
 436 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 437                            LLVMValueRef rsrc,
 438                            LLVMValueRef vdata,
 439                            LLVMValueRef voffset,
 440                            LLVMValueRef soffset,
 441                            LLVMValueRef immoffset,
 442                            unsigned num_channels,
 443                            unsigned dfmt,
 444                            unsigned nfmt,
 445                            unsigned cache_policy);
 446
 447 LLVMValueRef
 448 ac_get_thread_id(struct ac_llvm_context *ctx);
 449
 450 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 451 #define AC_TID_MASK_TOP      0xfffffffd
 452 #define AC_TID_MASK_LEFT     0xfffffffe
 453
 454 LLVMValueRef
 455 ac_build_ddxy(struct ac_llvm_context *ctx,
 456               uint32_t mask,
 457               int idx,
 458               LLVMValueRef val);
 459
 460 #define AC_SENDMSG_GS 2
 461 #define AC_SENDMSG_GS_DONE 3
 462 #define AC_SENDMSG_GS_ALLOC_REQ 9
 463
 464 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 465 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 466 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 467 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 468
 469 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 470                       uint32_t msg,
 471                       LLVMValueRef wave_id);
 472
 473 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 474                            LLVMValueRef arg,
 475                            LLVMTypeRef dst_type);
 476
 477 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 478                           LLVMValueRef arg,
 479                           LLVMTypeRef dst_type);
 480 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 481                            LLVMValueRef b);
 482 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 483                            LLVMValueRef b);
 484 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 485                            LLVMValueRef b);
 486 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 487                            LLVMValueRef b);
 488 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 489 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 490 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 491
 492 struct ac_export_args {
 493         LLVMValueRef out[4];
 494         unsigned target;
 495         unsigned enabled_channels;
 496         bool compr;
 497         bool done;
 498         bool valid_mask;
 499 };
 500
 501 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 502
 503 void ac_build_export_null(struct ac_llvm_context *ctx);
 504
 505 enum ac_image_opcode {
 506         ac_image_sample,
 507         ac_image_gather4,
 508         ac_image_load,
 509         ac_image_load_mip,
 510         ac_image_store,
 511         ac_image_store_mip,
 512         ac_image_get_lod,
 513         ac_image_get_resinfo,
 514         ac_image_atomic,
 515         ac_image_atomic_cmpswap,
 516 };
 517
 518 enum ac_atomic_op {
 519         ac_atomic_swap,
 520         ac_atomic_add,
 521         ac_atomic_sub,
 522         ac_atomic_smin,
 523         ac_atomic_umin,
 524         ac_atomic_smax,
 525         ac_atomic_umax,
 526         ac_atomic_and,
 527         ac_atomic_or,
 528         ac_atomic_xor,
 529         ac_atomic_inc_wrap,
 530         ac_atomic_dec_wrap,
 531 };
 532
 533 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 534 enum ac_image_cache_policy {
 535         ac_glc = 1 << 0, /* per-CU cache control */
 536         ac_slc = 1 << 1, /* global L2 cache control */
 537         ac_dlc = 1 << 2, /* per-shader-array cache control */
 538         ac_swizzled = 1 << 3, /* the access is swizzled, disabling load/store merging */
 539 };
 540
 541 struct ac_image_args {
 542         enum ac_image_opcode opcode : 4;
 543         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 544         enum ac_image_dim dim : 3;
 545         unsigned dmask : 4;
 546         unsigned cache_policy : 3;
 547         bool unorm : 1;
 548         bool level_zero : 1;
 549         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 550
 551         LLVMValueRef resource;
 552         LLVMValueRef sampler;
 553         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 554         LLVMValueRef offset;
 555         LLVMValueRef bias;
 556         LLVMValueRef compare;
 557         LLVMValueRef derivs[6];
 558         LLVMValueRef coords[4];
 559         LLVMValueRef lod; // also used by ac_image_get_resinfo
 560 };
 561
 562 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 563                                    struct ac_image_args *a);
 564 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
 565                                              LLVMValueRef rsrc);
 566 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 567                                     LLVMValueRef args[2]);
 568 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 569                                      LLVMValueRef args[2]);
 570 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 571                                      LLVMValueRef args[2]);
 572 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 573                                  LLVMValueRef args[2], unsigned bits, bool hi);
 574 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 575                                  LLVMValueRef args[2], unsigned bits, bool hi);
 576 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 577 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 578 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 579                           LLVMValueRef offset, LLVMValueRef width,
 580                           bool is_signed);
 581 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 582                            LLVMValueRef s1, LLVMValueRef s2);
 583 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 584                            LLVMValueRef s1, LLVMValueRef s2);
 585
 586 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 587
 588 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 589                            unsigned bitsize);
 590
 591 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 592                             LLVMValueRef src1, LLVMValueRef src2,
 593                             unsigned bitsize);
 594
 595 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 596                             unsigned bitsize);
 597
 598 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 599                             unsigned bitsize);
 600
 601 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 602
 603 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 604                                        LLVMValueRef src0);
 605
 606 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 607                             LLVMValueRef main_fn,
 608                             uint8_t *vs_output_param_offset,
 609                             uint32_t num_outputs,
 610                             uint8_t *num_param_exports);
 611 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 612
 613 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 614 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 615                          LLVMValueRef dw_addr);
 616 void ac_lds_store(struct ac_llvm_context *ctx,
 617                   LLVMValueRef dw_addr, LLVMValueRef value);
 618
 619 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 620                          LLVMTypeRef dst_type,
 621                          LLVMValueRef src0);
 622
 623 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 624 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 625
 626 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 627 void ac_build_break(struct ac_llvm_context *ctx);
 628 void ac_build_continue(struct ac_llvm_context *ctx);
 629 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 630 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 631 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 632 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 633 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 634                  int lable_id);
 635 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 636                   int lable_id);
 637
 638 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 639                              const char *name);
 640 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 641                                    const char *name);
 642
 643 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 644                          LLVMTypeRef type);
 645
 646 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 647                             unsigned count);
 648
 649 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 650                              unsigned rshift, unsigned bitwidth);
 651
 652 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 653                               LLVMValueRef *addr, bool is_array_tex);
 654
 655 LLVMValueRef
 656 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 657
 658 LLVMValueRef
 659 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 660
 661 LLVMValueRef
 662 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 663
 664 LLVMValueRef
 665 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 666
 667 LLVMValueRef
 668 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 669
 670 LLVMValueRef
 671 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 672
 673 LLVMValueRef
 674 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 675
 676 /**
 677  * Common arguments for a scan/reduce operation that accumulates per-wave
 678  * values across an entire workgroup, while respecting the order of waves.
 679  */
 680 struct ac_wg_scan {
 681         bool enable_reduce;
 682         bool enable_exclusive;
 683         bool enable_inclusive;
 684         nir_op op;
 685         LLVMValueRef src; /* clobbered! */
 686         LLVMValueRef result_reduce;
 687         LLVMValueRef result_exclusive;
 688         LLVMValueRef result_inclusive;
 689         LLVMValueRef extra;
 690         LLVMValueRef waveidx;
 691         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 692
 693         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 694         LLVMValueRef scratch;
 695         unsigned maxwaves;
 696 };
 697
 698 void
 699 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 700 void
 701 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 702 void
 703 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 704
 705 void
 706 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 707 void
 708 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 709 void
 710 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 711
 712 LLVMValueRef
 713 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 714                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 715
 716 LLVMValueRef
 717 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 718
 719 LLVMValueRef
 720 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 721                    unsigned bitsize);
 722
 723 LLVMValueRef
 724 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 725                     unsigned bitsize);
 726
 727 LLVMValueRef
 728 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
 729                       unsigned bitsize);
 730
 731 LLVMValueRef
 732 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 733
 734 LLVMValueRef
 735 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 736
 737 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 738                            LLVMValueRef *args, unsigned num_args);
 739
 740 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 741                                  LLVMValueRef ptr, LLVMValueRef val,
 742                                  const char *sync_scope);
 743
 744 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 745                                       LLVMValueRef cmp, LLVMValueRef val,
 746                                       const char *sync_scope);
 747
 748 void
 749 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
 750                 LLVMValueRef stencil, LLVMValueRef samplemask,
 751                 struct ac_export_args *args);
 752
 753 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
 754                                    LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt);
 755
 756 struct ac_ngg_prim {
 757         unsigned num_vertices;
 758         LLVMValueRef isnull;
 759         LLVMValueRef index[3];
 760         LLVMValueRef edgeflag[3];
 761         LLVMValueRef passthrough;
 762 };
 763
 764 LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
 765                                  const struct ac_ngg_prim *prim);
 766 void ac_build_export_prim(struct ac_llvm_context *ctx,
 767                           const struct ac_ngg_prim *prim);
 768
 769 static inline LLVMValueRef
 770 ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
 771 {
 772         assert(arg.used);
 773         return LLVMGetParam(ctx->main_function, arg.arg_index);
 774 }
 775
 776 enum ac_llvm_calling_convention {
 777         AC_LLVM_AMDGPU_VS = 87,
 778         AC_LLVM_AMDGPU_GS = 88,
 779         AC_LLVM_AMDGPU_PS = 89,
 780         AC_LLVM_AMDGPU_CS = 90,
 781         AC_LLVM_AMDGPU_HS = 93,
 782 };
 783
 784 LLVMValueRef ac_build_main(const struct ac_shader_args *args,
 785                            struct ac_llvm_context *ctx,
 786                            enum ac_llvm_calling_convention convention,
 787                            const char *name, LLVMTypeRef ret_type,
 788                            LLVMModuleRef module);
 789
 790 #ifdef __cplusplus
 791 }
 792 #endif
 793
 794 #endif