src/amd/common/ac_llvm_build.h

   1 /*
   2  * Copyright 2016 Bas Nieuwenhuizen
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  13  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  14  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  15  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  16  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  17  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  18  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  19  *
  20  * The above copyright notice and this permission notice (including the
  21  * next paragraph) shall be included in all copies or substantial portions
  22  * of the Software.
  23  *
  24  */
  25 #ifndef AC_LLVM_BUILD_H
  26 #define AC_LLVM_BUILD_H
  27
  28 #include <stdbool.h>
  29 #include <llvm-c/Core.h>
  30 #include "compiler/nir/nir.h"
  31 #include "amd_family.h"
  32
  33 #ifdef __cplusplus
  34 extern "C" {
  35 #endif
  36
  37 enum {
  38         AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
  39         AC_ADDR_SPACE_GLOBAL = 1,
  40         AC_ADDR_SPACE_GDS = 2,
  41         AC_ADDR_SPACE_LDS = 3,
  42         AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
  43         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  44 };
  45
  46 #define AC_WAIT_LGKM    (1 << 0) /* LDS, GDS, constant, message */
  47 #define AC_WAIT_VLOAD   (1 << 1) /* VMEM load/sample instructions */
  48 #define AC_WAIT_VSTORE  (1 << 2) /* VMEM store instructions */
  49
  50 struct ac_llvm_flow;
  51
  52 struct ac_llvm_context {
  53         LLVMContextRef context;
  54         LLVMModuleRef module;
  55         LLVMBuilderRef builder;
  56
  57         LLVMTypeRef voidt;
  58         LLVMTypeRef i1;
  59         LLVMTypeRef i8;
  60         LLVMTypeRef i16;
  61         LLVMTypeRef i32;
  62         LLVMTypeRef i64;
  63         LLVMTypeRef intptr;
  64         LLVMTypeRef f16;
  65         LLVMTypeRef f32;
  66         LLVMTypeRef f64;
  67         LLVMTypeRef v2i16;
  68         LLVMTypeRef v2i32;
  69         LLVMTypeRef v3i32;
  70         LLVMTypeRef v4i32;
  71         LLVMTypeRef v2f32;
  72         LLVMTypeRef v3f32;
  73         LLVMTypeRef v4f32;
  74         LLVMTypeRef v8i32;
  75
  76         LLVMValueRef i8_0;
  77         LLVMValueRef i8_1;
  78         LLVMValueRef i16_0;
  79         LLVMValueRef i16_1;
  80         LLVMValueRef i32_0;
  81         LLVMValueRef i32_1;
  82         LLVMValueRef i64_0;
  83         LLVMValueRef i64_1;
  84         LLVMValueRef f16_0;
  85         LLVMValueRef f16_1;
  86         LLVMValueRef f32_0;
  87         LLVMValueRef f32_1;
  88         LLVMValueRef f64_0;
  89         LLVMValueRef f64_1;
  90         LLVMValueRef i1true;
  91         LLVMValueRef i1false;
  92
  93         struct ac_llvm_flow *flow;
  94         unsigned flow_depth;
  95         unsigned flow_depth_max;
  96
  97         unsigned range_md_kind;
  98         unsigned invariant_load_md_kind;
  99         unsigned uniform_md_kind;
 100         unsigned fpmath_md_kind;
 101         LLVMValueRef fpmath_md_2p5_ulp;
 102         LLVMValueRef empty_md;
 103
 104         enum chip_class chip_class;
 105         enum radeon_family family;
 106
 107         LLVMValueRef lds;
 108 };
 109
 110 void
 111 ac_llvm_context_init(struct ac_llvm_context *ctx,
 112                      enum chip_class chip_class, enum radeon_family family);
 113
 114 void
 115 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 116
 117 int
 118 ac_get_llvm_num_components(LLVMValueRef value);
 119
 120 int
 121 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 122
 123 LLVMValueRef
 124 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 125                      LLVMValueRef value,
 126                      int index);
 127
 128 unsigned ac_get_type_size(LLVMTypeRef type);
 129
 130 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 131 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 132 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 133 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 134 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 135
 136 LLVMValueRef
 137 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 138                    LLVMTypeRef return_type, LLVMValueRef *params,
 139                    unsigned param_count, unsigned attrib_mask);
 140
 141 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 142
 143 LLVMValueRef
 144 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 145              unsigned count_incoming, LLVMValueRef *values,
 146              LLVMBasicBlockRef *blocks);
 147
 148 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 149 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 150                                    LLVMValueRef *pvgpr);
 151
 152 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 153
 154 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 155 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 156                                  LLVMValueRef value);
 157
 158 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 159
 160 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 161
 162 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 163
 164 LLVMValueRef
 165 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 166                                unsigned value_count, unsigned component);
 167
 168 LLVMValueRef
 169 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 170                                 LLVMValueRef *values,
 171                                 unsigned value_count,
 172                                 unsigned value_stride,
 173                                 bool load,
 174                                 bool always_vector);
 175 LLVMValueRef
 176 ac_build_gather_values(struct ac_llvm_context *ctx,
 177                        LLVMValueRef *values,
 178                        unsigned value_count);
 179 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 180                                      LLVMValueRef value,
 181                                      unsigned num_channels);
 182 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 183
 184 LLVMValueRef
 185 ac_build_fdiv(struct ac_llvm_context *ctx,
 186               LLVMValueRef num,
 187               LLVMValueRef den);
 188
 189 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 190                                 LLVMValueRef num,
 191                                 LLVMValueRef multiplier,
 192                                 LLVMValueRef pre_shift,
 193                                 LLVMValueRef post_shift,
 194                                 LLVMValueRef increment);
 195 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 196                                     LLVMValueRef num,
 197                                     LLVMValueRef multiplier,
 198                                     LLVMValueRef pre_shift,
 199                                     LLVMValueRef post_shift,
 200                                     LLVMValueRef increment);
 201 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 202                                               LLVMValueRef num,
 203                                               LLVMValueRef multiplier,
 204                                               LLVMValueRef post_shift);
 205
 206 void
 207 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 208                        bool is_deriv, bool is_array, bool is_lod,
 209                        LLVMValueRef *coords_arg,
 210                        LLVMValueRef *derivs_arg);
 211
 212
 213 LLVMValueRef
 214 ac_build_fs_interp(struct ac_llvm_context *ctx,
 215                    LLVMValueRef llvm_chan,
 216                    LLVMValueRef attr_number,
 217                    LLVMValueRef params,
 218                    LLVMValueRef i,
 219                    LLVMValueRef j);
 220
 221 LLVMValueRef
 222 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 223                        LLVMValueRef llvm_chan,
 224                        LLVMValueRef attr_number,
 225                        LLVMValueRef params,
 226                        LLVMValueRef i,
 227                        LLVMValueRef j);
 228
 229 LLVMValueRef
 230 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 231                        LLVMValueRef parameter,
 232                        LLVMValueRef llvm_chan,
 233                        LLVMValueRef attr_number,
 234                        LLVMValueRef params);
 235
 236 LLVMValueRef
 237 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 238                  LLVMValueRef base_ptr,
 239                  LLVMValueRef index);
 240
 241 LLVMValueRef
 242 ac_build_gep0(struct ac_llvm_context *ctx,
 243               LLVMValueRef base_ptr,
 244               LLVMValueRef index);
 245 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 246                                   LLVMValueRef index);
 247
 248 void
 249 ac_build_indexed_store(struct ac_llvm_context *ctx,
 250                        LLVMValueRef base_ptr, LLVMValueRef index,
 251                        LLVMValueRef value);
 252
 253 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 254                            LLVMValueRef index);
 255 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 256                                      LLVMValueRef base_ptr, LLVMValueRef index);
 257 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 258                                    LLVMValueRef base_ptr, LLVMValueRef index);
 259 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 260                                    LLVMValueRef base_ptr, LLVMValueRef index);
 261
 262 void
 263 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 264                             LLVMValueRef rsrc,
 265                             LLVMValueRef vdata,
 266                             unsigned num_channels,
 267                             LLVMValueRef voffset,
 268                             LLVMValueRef soffset,
 269                             unsigned inst_offset,
 270                             unsigned cache_policy,
 271                             bool swizzle_enable_hint);
 272
 273 void
 274 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 275                              LLVMValueRef rsrc,
 276                              LLVMValueRef data,
 277                              LLVMValueRef vindex,
 278                              LLVMValueRef voffset,
 279                              unsigned num_channels,
 280                              unsigned cache_policy);
 281
 282 LLVMValueRef
 283 ac_build_buffer_load(struct ac_llvm_context *ctx,
 284                      LLVMValueRef rsrc,
 285                      int num_channels,
 286                      LLVMValueRef vindex,
 287                      LLVMValueRef voffset,
 288                      LLVMValueRef soffset,
 289                      unsigned inst_offset,
 290                      unsigned cache_policy,
 291                      bool can_speculate,
 292                      bool allow_smem);
 293
 294 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 295                                          LLVMValueRef rsrc,
 296                                          LLVMValueRef vindex,
 297                                          LLVMValueRef voffset,
 298                                          unsigned num_channels,
 299                                          unsigned cache_policy,
 300                                          bool can_speculate);
 301
 302 /* load_format that handles the stride & element count better if idxen is
 303  * disabled by LLVM. */
 304 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 305                                                   LLVMValueRef rsrc,
 306                                                   LLVMValueRef vindex,
 307                                                   LLVMValueRef voffset,
 308                                                   unsigned num_channels,
 309                                                   unsigned cache_policy,
 310                                                   bool can_speculate);
 311
 312 LLVMValueRef
 313 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 314                             LLVMValueRef rsrc,
 315                             LLVMValueRef voffset,
 316                             LLVMValueRef soffset,
 317                             LLVMValueRef immoffset,
 318                             unsigned cache_policy);
 319
 320 LLVMValueRef
 321 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 322                            LLVMValueRef rsrc,
 323                            LLVMValueRef voffset,
 324                            LLVMValueRef soffset,
 325                            LLVMValueRef immoffset,
 326                            unsigned cache_policy);
 327
 328 LLVMValueRef
 329 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 330                              LLVMValueRef rsrc,
 331                              LLVMValueRef vindex,
 332                              LLVMValueRef voffset,
 333                              LLVMValueRef soffset,
 334                              LLVMValueRef immoffset,
 335                              unsigned num_channels,
 336                              unsigned dfmt,
 337                              unsigned nfmt,
 338                              unsigned cache_policy,
 339                              bool can_speculate);
 340
 341 LLVMValueRef
 342 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 343                           LLVMValueRef rsrc,
 344                           LLVMValueRef voffset,
 345                           LLVMValueRef soffset,
 346                           LLVMValueRef immoffset,
 347                           unsigned num_channels,
 348                           unsigned dfmt,
 349                           unsigned nfmt,
 350                           unsigned cache_policy,
 351                           bool can_speculate);
 352
 353 /* For ac_build_fetch_format.
 354  *
 355  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 356  */
 357 enum {
 358         AC_FETCH_FORMAT_FLOAT = 0,
 359         AC_FETCH_FORMAT_FIXED,
 360         AC_FETCH_FORMAT_UNORM,
 361         AC_FETCH_FORMAT_SNORM,
 362         AC_FETCH_FORMAT_USCALED,
 363         AC_FETCH_FORMAT_SSCALED,
 364         AC_FETCH_FORMAT_UINT,
 365         AC_FETCH_FORMAT_SINT,
 366 };
 367
 368 LLVMValueRef
 369 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 370                                unsigned log_size,
 371                                unsigned num_channels,
 372                                unsigned format,
 373                                bool reverse,
 374                                bool known_aligned,
 375                                LLVMValueRef rsrc,
 376                                LLVMValueRef vindex,
 377                                LLVMValueRef voffset,
 378                                LLVMValueRef soffset,
 379                                unsigned cache_policy,
 380                                bool can_speculate);
 381
 382 void
 383 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 384                              LLVMValueRef rsrc,
 385                              LLVMValueRef vdata,
 386                              LLVMValueRef voffset,
 387                              LLVMValueRef soffset,
 388                              unsigned cache_policy);
 389
 390 void
 391 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 392                             LLVMValueRef rsrc,
 393                             LLVMValueRef vdata,
 394                             LLVMValueRef voffset,
 395                             LLVMValueRef soffset,
 396                             unsigned cache_policy);
 397
 398 void
 399 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 400                               LLVMValueRef rsrc,
 401                               LLVMValueRef vdata,
 402                               LLVMValueRef vindex,
 403                               LLVMValueRef voffset,
 404                               LLVMValueRef soffset,
 405                               LLVMValueRef immoffset,
 406                               unsigned num_channels,
 407                               unsigned dfmt,
 408                               unsigned nfmt,
 409                               unsigned cache_policy);
 410
 411 void
 412 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 413                            LLVMValueRef rsrc,
 414                            LLVMValueRef vdata,
 415                            LLVMValueRef voffset,
 416                            LLVMValueRef soffset,
 417                            LLVMValueRef immoffset,
 418                            unsigned num_channels,
 419                            unsigned dfmt,
 420                            unsigned nfmt,
 421                            unsigned cache_policy);
 422
 423 LLVMValueRef
 424 ac_get_thread_id(struct ac_llvm_context *ctx);
 425
 426 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 427 #define AC_TID_MASK_TOP      0xfffffffd
 428 #define AC_TID_MASK_LEFT     0xfffffffe
 429
 430 LLVMValueRef
 431 ac_build_ddxy(struct ac_llvm_context *ctx,
 432               uint32_t mask,
 433               int idx,
 434               LLVMValueRef val);
 435
 436 #define AC_SENDMSG_GS 2
 437 #define AC_SENDMSG_GS_DONE 3
 438 #define AC_SENDMSG_GS_ALLOC_REQ 9
 439
 440 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 441 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 442 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 443 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 444
 445 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 446                       uint32_t msg,
 447                       LLVMValueRef wave_id);
 448
 449 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 450                            LLVMValueRef arg,
 451                            LLVMTypeRef dst_type);
 452
 453 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 454                           LLVMValueRef arg,
 455                           LLVMTypeRef dst_type);
 456 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 457                            LLVMValueRef b);
 458 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 459                            LLVMValueRef b);
 460 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 461                            LLVMValueRef b);
 462 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 463                            LLVMValueRef b);
 464 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 465 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 466 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 467
 468 struct ac_export_args {
 469         LLVMValueRef out[4];
 470         unsigned target;
 471         unsigned enabled_channels;
 472         bool compr;
 473         bool done;
 474         bool valid_mask;
 475 };
 476
 477 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 478
 479 void ac_build_export_null(struct ac_llvm_context *ctx);
 480
 481 enum ac_image_opcode {
 482         ac_image_sample,
 483         ac_image_gather4,
 484         ac_image_load,
 485         ac_image_load_mip,
 486         ac_image_store,
 487         ac_image_store_mip,
 488         ac_image_get_lod,
 489         ac_image_get_resinfo,
 490         ac_image_atomic,
 491         ac_image_atomic_cmpswap,
 492 };
 493
 494 enum ac_atomic_op {
 495         ac_atomic_swap,
 496         ac_atomic_add,
 497         ac_atomic_sub,
 498         ac_atomic_smin,
 499         ac_atomic_umin,
 500         ac_atomic_smax,
 501         ac_atomic_umax,
 502         ac_atomic_and,
 503         ac_atomic_or,
 504         ac_atomic_xor,
 505 };
 506
 507 enum ac_image_dim {
 508         ac_image_1d,
 509         ac_image_2d,
 510         ac_image_3d,
 511         ac_image_cube, // includes cube arrays
 512         ac_image_1darray,
 513         ac_image_2darray,
 514         ac_image_2dmsaa,
 515         ac_image_2darraymsaa,
 516 };
 517
 518 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 519 enum ac_image_cache_policy {
 520         ac_glc = 1 << 0, /* per-CU cache control */
 521         ac_slc = 1 << 1, /* global L2 cache control */
 522         ac_dlc = 1 << 2, /* per-shader-array cache control */
 523 };
 524
 525 struct ac_image_args {
 526         enum ac_image_opcode opcode : 4;
 527         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 528         enum ac_image_dim dim : 3;
 529         unsigned dmask : 4;
 530         unsigned cache_policy : 3;
 531         bool unorm : 1;
 532         bool level_zero : 1;
 533         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 534
 535         LLVMValueRef resource;
 536         LLVMValueRef sampler;
 537         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 538         LLVMValueRef offset;
 539         LLVMValueRef bias;
 540         LLVMValueRef compare;
 541         LLVMValueRef derivs[6];
 542         LLVMValueRef coords[4];
 543         LLVMValueRef lod; // also used by ac_image_get_resinfo
 544 };
 545
 546 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 547                                    struct ac_image_args *a);
 548 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 549                                     LLVMValueRef args[2]);
 550 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 551                                      LLVMValueRef args[2]);
 552 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 553                                      LLVMValueRef args[2]);
 554 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 555                                  LLVMValueRef args[2], unsigned bits, bool hi);
 556 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 557                                  LLVMValueRef args[2], unsigned bits, bool hi);
 558 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 559 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 560 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 561                           LLVMValueRef offset, LLVMValueRef width,
 562                           bool is_signed);
 563 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 564                            LLVMValueRef s1, LLVMValueRef s2);
 565 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 566                            LLVMValueRef s1, LLVMValueRef s2);
 567
 568 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 569
 570 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 571                            unsigned bitsize);
 572
 573 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 574                             LLVMValueRef src1, LLVMValueRef src2,
 575                             unsigned bitsize);
 576
 577 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 578                             unsigned bitsize);
 579
 580 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 581                             unsigned bitsize);
 582
 583 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 584
 585 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 586                                        LLVMValueRef src0);
 587
 588 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 589                             LLVMValueRef main_fn,
 590                             uint8_t *vs_output_param_offset,
 591                             uint32_t num_outputs,
 592                             uint8_t *num_param_exports);
 593 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 594
 595 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 596 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 597                          LLVMValueRef dw_addr);
 598 void ac_lds_store(struct ac_llvm_context *ctx,
 599                   LLVMValueRef dw_addr, LLVMValueRef value);
 600
 601 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 602                          LLVMTypeRef dst_type,
 603                          LLVMValueRef src0);
 604
 605 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 606 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 607
 608 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 609 void ac_build_break(struct ac_llvm_context *ctx);
 610 void ac_build_continue(struct ac_llvm_context *ctx);
 611 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 612 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 613 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 614 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 615 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 616                  int lable_id);
 617 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 618                   int lable_id);
 619
 620 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 621                              const char *name);
 622 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 623                                    const char *name);
 624
 625 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 626                          LLVMTypeRef type);
 627
 628 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 629                             unsigned count);
 630
 631 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 632                              unsigned rshift, unsigned bitwidth);
 633
 634 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 635                               LLVMValueRef *addr, bool is_array_tex);
 636
 637 LLVMValueRef
 638 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 639
 640 LLVMValueRef
 641 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 642
 643 LLVMValueRef
 644 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 645
 646 LLVMValueRef
 647 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 648
 649 LLVMValueRef
 650 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 651
 652 LLVMValueRef
 653 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 654
 655 LLVMValueRef
 656 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 657
 658 /**
 659  * Common arguments for a scan/reduce operation that accumulates per-wave
 660  * values across an entire workgroup, while respecting the order of waves.
 661  */
 662 struct ac_wg_scan {
 663         bool enable_reduce;
 664         bool enable_exclusive;
 665         bool enable_inclusive;
 666         nir_op op;
 667         LLVMValueRef src; /* clobbered! */
 668         LLVMValueRef result_reduce;
 669         LLVMValueRef result_exclusive;
 670         LLVMValueRef result_inclusive;
 671         LLVMValueRef extra;
 672         LLVMValueRef waveidx;
 673         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 674
 675         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 676         LLVMValueRef scratch;
 677         unsigned maxwaves;
 678 };
 679
 680 void
 681 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 682 void
 683 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 684 void
 685 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 686
 687 void
 688 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 689 void
 690 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 691 void
 692 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 693
 694 LLVMValueRef
 695 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 696                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 697
 698 LLVMValueRef
 699 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 700
 701 LLVMValueRef
 702 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 703                    unsigned bitsize);
 704
 705 LLVMValueRef
 706 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 707                     unsigned bitsize);
 708
 709 LLVMValueRef
 710 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 711
 712 LLVMValueRef
 713 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 714
 715 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 716                            LLVMValueRef *args, unsigned num_args);
 717
 718 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 719                                  LLVMValueRef ptr, LLVMValueRef val,
 720                                  const char *sync_scope);
 721
 722 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 723                                       LLVMValueRef cmp, LLVMValueRef val,
 724                                       const char *sync_scope);
 725
 726 #ifdef __cplusplus
 727 }
 728 #endif
 729
 730 #endif