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 /* Combine these with & instead of |. */
  47 #define NOOP_WAITCNT    0xcf7f
  48 #define LGKM_CNT        0xc07f
  49 #define EXP_CNT         0xcf0f
  50 #define VM_CNT          0x0f70 /* On GFX9, vmcnt has 6 bits in [0:3] and [14:15] */
  51
  52 struct ac_llvm_flow;
  53
  54 struct ac_llvm_context {
  55         LLVMContextRef context;
  56         LLVMModuleRef module;
  57         LLVMBuilderRef builder;
  58
  59         LLVMTypeRef voidt;
  60         LLVMTypeRef i1;
  61         LLVMTypeRef i8;
  62         LLVMTypeRef i16;
  63         LLVMTypeRef i32;
  64         LLVMTypeRef i64;
  65         LLVMTypeRef intptr;
  66         LLVMTypeRef f16;
  67         LLVMTypeRef f32;
  68         LLVMTypeRef f64;
  69         LLVMTypeRef v2i16;
  70         LLVMTypeRef v2i32;
  71         LLVMTypeRef v3i32;
  72         LLVMTypeRef v4i32;
  73         LLVMTypeRef v2f32;
  74         LLVMTypeRef v3f32;
  75         LLVMTypeRef v4f32;
  76         LLVMTypeRef v8i32;
  77
  78         LLVMValueRef i8_0;
  79         LLVMValueRef i8_1;
  80         LLVMValueRef i16_0;
  81         LLVMValueRef i16_1;
  82         LLVMValueRef i32_0;
  83         LLVMValueRef i32_1;
  84         LLVMValueRef i64_0;
  85         LLVMValueRef i64_1;
  86         LLVMValueRef f16_0;
  87         LLVMValueRef f16_1;
  88         LLVMValueRef f32_0;
  89         LLVMValueRef f32_1;
  90         LLVMValueRef f64_0;
  91         LLVMValueRef f64_1;
  92         LLVMValueRef i1true;
  93         LLVMValueRef i1false;
  94
  95         struct ac_llvm_flow *flow;
  96         unsigned flow_depth;
  97         unsigned flow_depth_max;
  98
  99         unsigned range_md_kind;
 100         unsigned invariant_load_md_kind;
 101         unsigned uniform_md_kind;
 102         unsigned fpmath_md_kind;
 103         LLVMValueRef fpmath_md_2p5_ulp;
 104         LLVMValueRef empty_md;
 105
 106         enum chip_class chip_class;
 107         enum radeon_family family;
 108
 109         LLVMValueRef lds;
 110 };
 111
 112 void
 113 ac_llvm_context_init(struct ac_llvm_context *ctx,
 114                      enum chip_class chip_class, enum radeon_family family);
 115
 116 void
 117 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 118
 119 int
 120 ac_get_llvm_num_components(LLVMValueRef value);
 121
 122 int
 123 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 124
 125 LLVMValueRef
 126 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 127                      LLVMValueRef value,
 128                      int index);
 129
 130 unsigned ac_get_type_size(LLVMTypeRef type);
 131
 132 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 133 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 134 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
 135 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 136 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 137
 138 LLVMValueRef
 139 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 140                    LLVMTypeRef return_type, LLVMValueRef *params,
 141                    unsigned param_count, unsigned attrib_mask);
 142
 143 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 144
 145 LLVMValueRef
 146 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 147              unsigned count_incoming, LLVMValueRef *values,
 148              LLVMBasicBlockRef *blocks);
 149
 150 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 151 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 152                                    LLVMValueRef *pvgpr);
 153
 154 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 155
 156 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 157 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
 158                                  LLVMValueRef value);
 159
 160 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 161
 162 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 163
 164 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 165
 166 LLVMValueRef
 167 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 168                                unsigned value_count, unsigned component);
 169
 170 LLVMValueRef
 171 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 172                                 LLVMValueRef *values,
 173                                 unsigned value_count,
 174                                 unsigned value_stride,
 175                                 bool load,
 176                                 bool always_vector);
 177 LLVMValueRef
 178 ac_build_gather_values(struct ac_llvm_context *ctx,
 179                        LLVMValueRef *values,
 180                        unsigned value_count);
 181 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 182                                      LLVMValueRef value,
 183                                      unsigned num_channels);
 184 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 185
 186 LLVMValueRef
 187 ac_build_fdiv(struct ac_llvm_context *ctx,
 188               LLVMValueRef num,
 189               LLVMValueRef den);
 190
 191 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
 192                                 LLVMValueRef num,
 193                                 LLVMValueRef multiplier,
 194                                 LLVMValueRef pre_shift,
 195                                 LLVMValueRef post_shift,
 196                                 LLVMValueRef increment);
 197 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
 198                                     LLVMValueRef num,
 199                                     LLVMValueRef multiplier,
 200                                     LLVMValueRef pre_shift,
 201                                     LLVMValueRef post_shift,
 202                                     LLVMValueRef increment);
 203 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
 204                                               LLVMValueRef num,
 205                                               LLVMValueRef multiplier,
 206                                               LLVMValueRef post_shift);
 207
 208 void
 209 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 210                        bool is_deriv, bool is_array, bool is_lod,
 211                        LLVMValueRef *coords_arg,
 212                        LLVMValueRef *derivs_arg);
 213
 214
 215 LLVMValueRef
 216 ac_build_fs_interp(struct ac_llvm_context *ctx,
 217                    LLVMValueRef llvm_chan,
 218                    LLVMValueRef attr_number,
 219                    LLVMValueRef params,
 220                    LLVMValueRef i,
 221                    LLVMValueRef j);
 222
 223 LLVMValueRef
 224 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
 225                        LLVMValueRef llvm_chan,
 226                        LLVMValueRef attr_number,
 227                        LLVMValueRef params,
 228                        LLVMValueRef i,
 229                        LLVMValueRef j);
 230
 231 LLVMValueRef
 232 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 233                        LLVMValueRef parameter,
 234                        LLVMValueRef llvm_chan,
 235                        LLVMValueRef attr_number,
 236                        LLVMValueRef params);
 237
 238 LLVMValueRef
 239 ac_build_gep_ptr(struct ac_llvm_context *ctx,
 240                  LLVMValueRef base_ptr,
 241                  LLVMValueRef index);
 242
 243 LLVMValueRef
 244 ac_build_gep0(struct ac_llvm_context *ctx,
 245               LLVMValueRef base_ptr,
 246               LLVMValueRef index);
 247 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 248                                   LLVMValueRef index);
 249
 250 void
 251 ac_build_indexed_store(struct ac_llvm_context *ctx,
 252                        LLVMValueRef base_ptr, LLVMValueRef index,
 253                        LLVMValueRef value);
 254
 255 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 256                            LLVMValueRef index);
 257 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 258                                      LLVMValueRef base_ptr, LLVMValueRef index);
 259 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 260                                    LLVMValueRef base_ptr, LLVMValueRef index);
 261 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 262                                    LLVMValueRef base_ptr, LLVMValueRef index);
 263
 264 void
 265 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 266                             LLVMValueRef rsrc,
 267                             LLVMValueRef vdata,
 268                             unsigned num_channels,
 269                             LLVMValueRef voffset,
 270                             LLVMValueRef soffset,
 271                             unsigned inst_offset,
 272                             bool glc,
 273                             bool slc,
 274                             bool swizzle_enable_hint);
 275
 276 void
 277 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 278                              LLVMValueRef rsrc,
 279                              LLVMValueRef data,
 280                              LLVMValueRef vindex,
 281                              LLVMValueRef voffset,
 282                              unsigned num_channels,
 283                              bool glc,
 284                              bool slc);
 285
 286 LLVMValueRef
 287 ac_build_buffer_load(struct ac_llvm_context *ctx,
 288                      LLVMValueRef rsrc,
 289                      int num_channels,
 290                      LLVMValueRef vindex,
 291                      LLVMValueRef voffset,
 292                      LLVMValueRef soffset,
 293                      unsigned inst_offset,
 294                      unsigned glc,
 295                      unsigned slc,
 296                      bool can_speculate,
 297                      bool allow_smem);
 298
 299 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 300                                          LLVMValueRef rsrc,
 301                                          LLVMValueRef vindex,
 302                                          LLVMValueRef voffset,
 303                                          unsigned num_channels,
 304                                          bool glc,
 305                                          bool can_speculate);
 306
 307 /* load_format that handles the stride & element count better if idxen is
 308  * disabled by LLVM. */
 309 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 310                                                   LLVMValueRef rsrc,
 311                                                   LLVMValueRef vindex,
 312                                                   LLVMValueRef voffset,
 313                                                   unsigned num_channels,
 314                                                   bool glc,
 315                                                   bool can_speculate);
 316
 317 LLVMValueRef
 318 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 319                             LLVMValueRef rsrc,
 320                             LLVMValueRef voffset,
 321                             LLVMValueRef soffset,
 322                             LLVMValueRef immoffset,
 323                             bool glc);
 324
 325 LLVMValueRef
 326 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
 327                            LLVMValueRef rsrc,
 328                            LLVMValueRef voffset,
 329                            LLVMValueRef soffset,
 330                            LLVMValueRef immoffset,
 331                            bool glc);
 332
 333 LLVMValueRef
 334 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
 335                              LLVMValueRef rsrc,
 336                              LLVMValueRef vindex,
 337                              LLVMValueRef voffset,
 338                              LLVMValueRef soffset,
 339                              LLVMValueRef immoffset,
 340                              unsigned num_channels,
 341                              unsigned dfmt,
 342                              unsigned nfmt,
 343                              bool glc,
 344                              bool slc,
 345                              bool can_speculate);
 346
 347 LLVMValueRef
 348 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 349                           LLVMValueRef rsrc,
 350                           LLVMValueRef voffset,
 351                           LLVMValueRef soffset,
 352                           LLVMValueRef immoffset,
 353                           unsigned num_channels,
 354                           unsigned dfmt,
 355                           unsigned nfmt,
 356                           bool glc,
 357                           bool slc,
 358                           bool can_speculate);
 359
 360 /* For ac_build_fetch_format.
 361  *
 362  * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
 363  */
 364 enum {
 365         AC_FETCH_FORMAT_FLOAT = 0,
 366         AC_FETCH_FORMAT_FIXED,
 367         AC_FETCH_FORMAT_UNORM,
 368         AC_FETCH_FORMAT_SNORM,
 369         AC_FETCH_FORMAT_USCALED,
 370         AC_FETCH_FORMAT_SSCALED,
 371         AC_FETCH_FORMAT_UINT,
 372         AC_FETCH_FORMAT_SINT,
 373 };
 374
 375 LLVMValueRef
 376 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
 377                                unsigned log_size,
 378                                unsigned num_channels,
 379                                unsigned format,
 380                                bool reverse,
 381                                bool known_aligned,
 382                                LLVMValueRef rsrc,
 383                                LLVMValueRef vindex,
 384                                LLVMValueRef voffset,
 385                                LLVMValueRef soffset,
 386                                bool glc,
 387                                bool slc,
 388                                bool can_speculate);
 389
 390 void
 391 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
 392                              LLVMValueRef rsrc,
 393                              LLVMValueRef vdata,
 394                              LLVMValueRef voffset,
 395                              LLVMValueRef soffset,
 396                              bool glc);
 397
 398 void
 399 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
 400                             LLVMValueRef rsrc,
 401                             LLVMValueRef vdata,
 402                             LLVMValueRef voffset,
 403                             LLVMValueRef soffset,
 404                             bool glc);
 405
 406 void
 407 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
 408                               LLVMValueRef rsrc,
 409                               LLVMValueRef vdata,
 410                               LLVMValueRef vindex,
 411                               LLVMValueRef voffset,
 412                               LLVMValueRef soffset,
 413                               LLVMValueRef immoffset,
 414                               unsigned num_channels,
 415                               unsigned dfmt,
 416                               unsigned nfmt,
 417                               bool glc,
 418                               bool slc);
 419
 420 void
 421 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
 422                            LLVMValueRef rsrc,
 423                            LLVMValueRef vdata,
 424                            LLVMValueRef voffset,
 425                            LLVMValueRef soffset,
 426                            LLVMValueRef immoffset,
 427                            unsigned num_channels,
 428                            unsigned dfmt,
 429                            unsigned nfmt,
 430                            bool glc,
 431                            bool slc);
 432
 433 LLVMValueRef
 434 ac_get_thread_id(struct ac_llvm_context *ctx);
 435
 436 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 437 #define AC_TID_MASK_TOP      0xfffffffd
 438 #define AC_TID_MASK_LEFT     0xfffffffe
 439
 440 LLVMValueRef
 441 ac_build_ddxy(struct ac_llvm_context *ctx,
 442               uint32_t mask,
 443               int idx,
 444               LLVMValueRef val);
 445
 446 #define AC_SENDMSG_GS 2
 447 #define AC_SENDMSG_GS_DONE 3
 448 #define AC_SENDMSG_GS_ALLOC_REQ 9
 449
 450 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 451 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 452 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 453 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 454
 455 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 456                       uint32_t msg,
 457                       LLVMValueRef wave_id);
 458
 459 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 460                            LLVMValueRef arg,
 461                            LLVMTypeRef dst_type);
 462
 463 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 464                           LLVMValueRef arg,
 465                           LLVMTypeRef dst_type);
 466 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 467                            LLVMValueRef b);
 468 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 469                            LLVMValueRef b);
 470 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 471                            LLVMValueRef b);
 472 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 473                            LLVMValueRef b);
 474 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 475 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 476 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 477
 478 struct ac_export_args {
 479         LLVMValueRef out[4];
 480         unsigned target;
 481         unsigned enabled_channels;
 482         bool compr;
 483         bool done;
 484         bool valid_mask;
 485 };
 486
 487 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 488
 489 void ac_build_export_null(struct ac_llvm_context *ctx);
 490
 491 enum ac_image_opcode {
 492         ac_image_sample,
 493         ac_image_gather4,
 494         ac_image_load,
 495         ac_image_load_mip,
 496         ac_image_store,
 497         ac_image_store_mip,
 498         ac_image_get_lod,
 499         ac_image_get_resinfo,
 500         ac_image_atomic,
 501         ac_image_atomic_cmpswap,
 502 };
 503
 504 enum ac_atomic_op {
 505         ac_atomic_swap,
 506         ac_atomic_add,
 507         ac_atomic_sub,
 508         ac_atomic_smin,
 509         ac_atomic_umin,
 510         ac_atomic_smax,
 511         ac_atomic_umax,
 512         ac_atomic_and,
 513         ac_atomic_or,
 514         ac_atomic_xor,
 515 };
 516
 517 enum ac_image_dim {
 518         ac_image_1d,
 519         ac_image_2d,
 520         ac_image_3d,
 521         ac_image_cube, // includes cube arrays
 522         ac_image_1darray,
 523         ac_image_2darray,
 524         ac_image_2dmsaa,
 525         ac_image_2darraymsaa,
 526 };
 527
 528 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 529 enum ac_image_cache_policy {
 530         ac_glc = 1 << 0,
 531         ac_slc = 1 << 1,
 532 };
 533
 534 struct ac_image_args {
 535         enum ac_image_opcode opcode : 4;
 536         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 537         enum ac_image_dim dim : 3;
 538         unsigned dmask : 4;
 539         unsigned cache_policy : 2;
 540         bool unorm : 1;
 541         bool level_zero : 1;
 542         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 543
 544         LLVMValueRef resource;
 545         LLVMValueRef sampler;
 546         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 547         LLVMValueRef offset;
 548         LLVMValueRef bias;
 549         LLVMValueRef compare;
 550         LLVMValueRef derivs[6];
 551         LLVMValueRef coords[4];
 552         LLVMValueRef lod; // also used by ac_image_get_resinfo
 553 };
 554
 555 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 556                                    struct ac_image_args *a);
 557 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 558                                     LLVMValueRef args[2]);
 559 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 560                                      LLVMValueRef args[2]);
 561 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 562                                      LLVMValueRef args[2]);
 563 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 564                                  LLVMValueRef args[2], unsigned bits, bool hi);
 565 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 566                                  LLVMValueRef args[2], unsigned bits, bool hi);
 567 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 568 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 569 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 570                           LLVMValueRef offset, LLVMValueRef width,
 571                           bool is_signed);
 572 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 573                            LLVMValueRef s1, LLVMValueRef s2);
 574 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 575                            LLVMValueRef s1, LLVMValueRef s2);
 576
 577 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
 578
 579 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 580                            unsigned bitsize);
 581
 582 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 583                             LLVMValueRef src1, LLVMValueRef src2,
 584                             unsigned bitsize);
 585
 586 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 587                             unsigned bitsize);
 588
 589 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 590                             unsigned bitsize);
 591
 592 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 593
 594 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 595                                        LLVMValueRef src0);
 596
 597 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 598                             LLVMValueRef main_fn,
 599                             uint8_t *vs_output_param_offset,
 600                             uint32_t num_outputs,
 601                             uint8_t *num_param_exports);
 602 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 603
 604 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 605 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 606                          LLVMValueRef dw_addr);
 607 void ac_lds_store(struct ac_llvm_context *ctx,
 608                   LLVMValueRef dw_addr, LLVMValueRef value);
 609
 610 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 611                          LLVMTypeRef dst_type,
 612                          LLVMValueRef src0);
 613
 614 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 615 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 616
 617 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 618 void ac_build_break(struct ac_llvm_context *ctx);
 619 void ac_build_continue(struct ac_llvm_context *ctx);
 620 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 621 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 622 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 623 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
 624 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 625                  int lable_id);
 626 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 627                   int lable_id);
 628
 629 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 630                              const char *name);
 631 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 632                                    const char *name);
 633
 634 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 635                          LLVMTypeRef type);
 636
 637 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 638                             unsigned count);
 639
 640 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 641                              unsigned rshift, unsigned bitwidth);
 642
 643 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 644                               LLVMValueRef *addr, bool is_array_tex);
 645
 646 LLVMValueRef
 647 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 648
 649 LLVMValueRef
 650 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 651
 652 LLVMValueRef
 653 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 654
 655 LLVMValueRef
 656 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 657
 658 LLVMValueRef
 659 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 660
 661 LLVMValueRef
 662 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 663
 664 LLVMValueRef
 665 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 666
 667 /**
 668  * Common arguments for a scan/reduce operation that accumulates per-wave
 669  * values across an entire workgroup, while respecting the order of waves.
 670  */
 671 struct ac_wg_scan {
 672         bool enable_reduce;
 673         bool enable_exclusive;
 674         bool enable_inclusive;
 675         nir_op op;
 676         LLVMValueRef src; /* clobbered! */
 677         LLVMValueRef result_reduce;
 678         LLVMValueRef result_exclusive;
 679         LLVMValueRef result_inclusive;
 680         LLVMValueRef extra;
 681         LLVMValueRef waveidx;
 682         LLVMValueRef numwaves; /* only needed for "reduce" operations */
 683
 684         /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
 685         LLVMValueRef scratch;
 686         unsigned maxwaves;
 687 };
 688
 689 void
 690 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 691 void
 692 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 693 void
 694 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 695
 696 void
 697 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 698 void
 699 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 700 void
 701 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
 702
 703 LLVMValueRef
 704 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 705                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 706
 707 LLVMValueRef
 708 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 709
 710 LLVMValueRef
 711 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 712                    unsigned bitsize);
 713
 714 LLVMValueRef
 715 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 716                     unsigned bitsize);
 717
 718 LLVMValueRef
 719 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 720
 721 LLVMValueRef
 722 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 723
 724 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
 725                            LLVMValueRef *args, unsigned num_args);
 726
 727 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 728                                  LLVMValueRef ptr, LLVMValueRef val,
 729                                  const char *sync_scope);
 730
 731 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 732                                       LLVMValueRef cmp, LLVMValueRef val,
 733                                       const char *sync_scope);
 734
 735 #ifdef __cplusplus
 736 }
 737 #endif
 738
 739 #endif