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/TargetMachine.h>
  30 #include "compiler/nir/nir.h"
  31 #include "amd_family.h"
  32
  33 #ifdef __cplusplus
  34 extern "C" {
  35 #endif
  36
  37 #define HAVE_32BIT_POINTERS (HAVE_LLVM >= 0x0700)
  38
  39 enum {
  40         AC_ADDR_SPACE_FLAT = HAVE_LLVM >= 0x0700 ? 0 : 4, /* Slower than global. */
  41         AC_ADDR_SPACE_GLOBAL = 1,
  42         AC_ADDR_SPACE_GDS = HAVE_LLVM >= 0x0700 ? 2 : 5,
  43         AC_ADDR_SPACE_LDS = 3,
  44         AC_ADDR_SPACE_CONST = HAVE_LLVM >= 0x0700 ? 4 : 2, /* Global allowing SMEM. */
  45         AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
  46 };
  47
  48 /* Combine these with & instead of |. */
  49 #define NOOP_WAITCNT    0xcf7f
  50 #define LGKM_CNT        0xc07f
  51 #define EXP_CNT         0xcf0f
  52 #define VM_CNT          0x0f70 /* On GFX9, vmcnt has 6 bits in [0:3] and [14:15] */
  53
  54 struct ac_llvm_flow;
  55
  56 struct ac_llvm_context {
  57         LLVMContextRef context;
  58         LLVMModuleRef module;
  59         LLVMBuilderRef builder;
  60
  61         LLVMTypeRef voidt;
  62         LLVMTypeRef i1;
  63         LLVMTypeRef i8;
  64         LLVMTypeRef i16;
  65         LLVMTypeRef i32;
  66         LLVMTypeRef i64;
  67         LLVMTypeRef intptr;
  68         LLVMTypeRef f16;
  69         LLVMTypeRef f32;
  70         LLVMTypeRef f64;
  71         LLVMTypeRef v2i16;
  72         LLVMTypeRef v2i32;
  73         LLVMTypeRef v3i32;
  74         LLVMTypeRef v4i32;
  75         LLVMTypeRef v2f32;
  76         LLVMTypeRef v4f32;
  77         LLVMTypeRef v8i32;
  78
  79         LLVMValueRef i16_0;
  80         LLVMValueRef i16_1;
  81         LLVMValueRef i32_0;
  82         LLVMValueRef i32_1;
  83         LLVMValueRef i64_0;
  84         LLVMValueRef i64_1;
  85         LLVMValueRef f32_0;
  86         LLVMValueRef f32_1;
  87         LLVMValueRef f64_0;
  88         LLVMValueRef f64_1;
  89         LLVMValueRef i1true;
  90         LLVMValueRef i1false;
  91
  92         struct ac_llvm_flow *flow;
  93         unsigned flow_depth;
  94         unsigned flow_depth_max;
  95
  96         unsigned range_md_kind;
  97         unsigned invariant_load_md_kind;
  98         unsigned uniform_md_kind;
  99         unsigned fpmath_md_kind;
 100         LLVMValueRef fpmath_md_2p5_ulp;
 101         LLVMValueRef empty_md;
 102
 103         enum chip_class chip_class;
 104         enum radeon_family family;
 105
 106         LLVMValueRef lds;
 107 };
 108
 109 void
 110 ac_llvm_context_init(struct ac_llvm_context *ctx,
 111                      enum chip_class chip_class, enum radeon_family family);
 112
 113 void
 114 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 115
 116 int
 117 ac_get_llvm_num_components(LLVMValueRef value);
 118
 119 int
 120 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 121
 122 LLVMValueRef
 123 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 124                      LLVMValueRef value,
 125                      int index);
 126
 127 unsigned ac_get_type_size(LLVMTypeRef type);
 128
 129 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 130 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 131 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 132 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 133
 134 LLVMValueRef
 135 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 136                    LLVMTypeRef return_type, LLVMValueRef *params,
 137                    unsigned param_count, unsigned attrib_mask);
 138
 139 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 140
 141 LLVMValueRef
 142 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 143              unsigned count_incoming, LLVMValueRef *values,
 144              LLVMBasicBlockRef *blocks);
 145
 146 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 147 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 148                                    LLVMValueRef *pvgpr);
 149
 150 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 151
 152 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 153
 154 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 155
 156 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 157
 158 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 159
 160 LLVMValueRef
 161 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 162                                unsigned value_count, unsigned component);
 163
 164 LLVMValueRef
 165 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 166                                 LLVMValueRef *values,
 167                                 unsigned value_count,
 168                                 unsigned value_stride,
 169                                 bool load,
 170                                 bool always_vector);
 171 LLVMValueRef
 172 ac_build_gather_values(struct ac_llvm_context *ctx,
 173                        LLVMValueRef *values,
 174                        unsigned value_count);
 175 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 176                                      LLVMValueRef value,
 177                                      unsigned num_channels);
 178 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 179
 180 LLVMValueRef
 181 ac_build_fdiv(struct ac_llvm_context *ctx,
 182               LLVMValueRef num,
 183               LLVMValueRef den);
 184
 185 void
 186 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 187                        bool is_deriv, bool is_array, bool is_lod,
 188                        LLVMValueRef *coords_arg,
 189                        LLVMValueRef *derivs_arg);
 190
 191
 192 LLVMValueRef
 193 ac_build_fs_interp(struct ac_llvm_context *ctx,
 194                    LLVMValueRef llvm_chan,
 195                    LLVMValueRef attr_number,
 196                    LLVMValueRef params,
 197                    LLVMValueRef i,
 198                    LLVMValueRef j);
 199
 200 LLVMValueRef
 201 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 202                        LLVMValueRef parameter,
 203                        LLVMValueRef llvm_chan,
 204                        LLVMValueRef attr_number,
 205                        LLVMValueRef params);
 206
 207 LLVMValueRef
 208 ac_build_gep0(struct ac_llvm_context *ctx,
 209               LLVMValueRef base_ptr,
 210               LLVMValueRef index);
 211 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 212                                   LLVMValueRef index);
 213
 214 void
 215 ac_build_indexed_store(struct ac_llvm_context *ctx,
 216                        LLVMValueRef base_ptr, LLVMValueRef index,
 217                        LLVMValueRef value);
 218
 219 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 220                            LLVMValueRef index);
 221 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 222                                      LLVMValueRef base_ptr, LLVMValueRef index);
 223 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 224                                    LLVMValueRef base_ptr, LLVMValueRef index);
 225 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 226                                    LLVMValueRef base_ptr, LLVMValueRef index);
 227
 228 void
 229 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 230                             LLVMValueRef rsrc,
 231                             LLVMValueRef vdata,
 232                             unsigned num_channels,
 233                             LLVMValueRef voffset,
 234                             LLVMValueRef soffset,
 235                             unsigned inst_offset,
 236                             bool glc,
 237                             bool slc,
 238                             bool writeonly_memory,
 239                             bool swizzle_enable_hint);
 240 LLVMValueRef
 241 ac_build_buffer_load(struct ac_llvm_context *ctx,
 242                      LLVMValueRef rsrc,
 243                      int num_channels,
 244                      LLVMValueRef vindex,
 245                      LLVMValueRef voffset,
 246                      LLVMValueRef soffset,
 247                      unsigned inst_offset,
 248                      unsigned glc,
 249                      unsigned slc,
 250                      bool can_speculate,
 251                      bool allow_smem);
 252
 253 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 254                                          LLVMValueRef rsrc,
 255                                          LLVMValueRef vindex,
 256                                          LLVMValueRef voffset,
 257                                          unsigned num_channels,
 258                                          bool glc,
 259                                          bool can_speculate);
 260
 261 /* load_format that handles the stride & element count better if idxen is
 262  * disabled by LLVM. */
 263 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 264                                                   LLVMValueRef rsrc,
 265                                                   LLVMValueRef vindex,
 266                                                   LLVMValueRef voffset,
 267                                                   unsigned num_channels,
 268                                                   bool glc,
 269                                                   bool can_speculate);
 270
 271 LLVMValueRef
 272 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 273                             LLVMValueRef rsrc,
 274                             LLVMValueRef vindex,
 275                             LLVMValueRef voffset,
 276                                 LLVMValueRef soffset,
 277                                 LLVMValueRef immoffset);
 278
 279 LLVMValueRef
 280 ac_get_thread_id(struct ac_llvm_context *ctx);
 281
 282 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 283 #define AC_TID_MASK_TOP      0xfffffffd
 284 #define AC_TID_MASK_LEFT     0xfffffffe
 285
 286 LLVMValueRef
 287 ac_build_ddxy(struct ac_llvm_context *ctx,
 288               uint32_t mask,
 289               int idx,
 290               LLVMValueRef val);
 291
 292 #define AC_SENDMSG_GS 2
 293 #define AC_SENDMSG_GS_DONE 3
 294
 295 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 296 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 297 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 298 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 299
 300 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 301                       uint32_t msg,
 302                       LLVMValueRef wave_id);
 303
 304 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 305                            LLVMValueRef arg,
 306                            LLVMTypeRef dst_type);
 307
 308 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 309                           LLVMValueRef arg,
 310                           LLVMTypeRef dst_type);
 311 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 312                            LLVMValueRef b);
 313 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 314                            LLVMValueRef b);
 315 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 316                            LLVMValueRef b);
 317 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 318                            LLVMValueRef b);
 319 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 320 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 321
 322 struct ac_export_args {
 323         LLVMValueRef out[4];
 324         unsigned target;
 325         unsigned enabled_channels;
 326         bool compr;
 327         bool done;
 328         bool valid_mask;
 329 };
 330
 331 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 332
 333 void ac_build_export_null(struct ac_llvm_context *ctx);
 334
 335 enum ac_image_opcode {
 336         ac_image_sample,
 337         ac_image_gather4,
 338         ac_image_load,
 339         ac_image_load_mip,
 340         ac_image_store,
 341         ac_image_store_mip,
 342         ac_image_get_lod,
 343         ac_image_get_resinfo,
 344         ac_image_atomic,
 345         ac_image_atomic_cmpswap,
 346 };
 347
 348 enum ac_atomic_op {
 349         ac_atomic_swap,
 350         ac_atomic_add,
 351         ac_atomic_sub,
 352         ac_atomic_smin,
 353         ac_atomic_umin,
 354         ac_atomic_smax,
 355         ac_atomic_umax,
 356         ac_atomic_and,
 357         ac_atomic_or,
 358         ac_atomic_xor,
 359 };
 360
 361 enum ac_image_dim {
 362         ac_image_1d,
 363         ac_image_2d,
 364         ac_image_3d,
 365         ac_image_cube, // includes cube arrays
 366         ac_image_1darray,
 367         ac_image_2darray,
 368         ac_image_2dmsaa,
 369         ac_image_2darraymsaa,
 370 };
 371
 372 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 373 enum ac_image_cache_policy {
 374         ac_glc = 1 << 0,
 375         ac_slc = 1 << 1,
 376 };
 377
 378 struct ac_image_args {
 379         enum ac_image_opcode opcode : 4;
 380         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 381         enum ac_image_dim dim : 3;
 382         unsigned dmask : 4;
 383         unsigned cache_policy : 2;
 384         bool unorm : 1;
 385         bool level_zero : 1;
 386         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 387
 388         LLVMValueRef resource;
 389         LLVMValueRef sampler;
 390         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 391         LLVMValueRef offset;
 392         LLVMValueRef bias;
 393         LLVMValueRef compare;
 394         LLVMValueRef derivs[6];
 395         LLVMValueRef coords[4];
 396         LLVMValueRef lod; // also used by ac_image_get_resinfo
 397 };
 398
 399 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 400                                    struct ac_image_args *a);
 401 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 402                                     LLVMValueRef args[2]);
 403 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 404                                      LLVMValueRef args[2]);
 405 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 406                                      LLVMValueRef args[2]);
 407 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 408                                  LLVMValueRef args[2], unsigned bits, bool hi);
 409 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 410                                  LLVMValueRef args[2], unsigned bits, bool hi);
 411 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 412 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 413 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 414                           LLVMValueRef offset, LLVMValueRef width,
 415                           bool is_signed);
 416 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 417                            LLVMValueRef s1, LLVMValueRef s2);
 418 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 419                            LLVMValueRef s1, LLVMValueRef s2);
 420
 421 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
 422
 423 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 424                            unsigned bitsize);
 425
 426 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 427                             unsigned bitsize);
 428
 429 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 430                             unsigned bitsize);
 431
 432 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 433
 434 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 435                                        LLVMValueRef src0);
 436
 437 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 438                             LLVMValueRef main_fn,
 439                             uint8_t *vs_output_param_offset,
 440                             uint32_t num_outputs,
 441                             uint8_t *num_param_exports);
 442 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 443
 444 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 445 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 446                          LLVMValueRef dw_addr);
 447 void ac_lds_store(struct ac_llvm_context *ctx,
 448                   LLVMValueRef dw_addr, LLVMValueRef value);
 449
 450 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 451                          LLVMTypeRef dst_type,
 452                          LLVMValueRef src0);
 453
 454 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 455 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 456
 457 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 458 void ac_build_break(struct ac_llvm_context *ctx);
 459 void ac_build_continue(struct ac_llvm_context *ctx);
 460 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 461 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 462 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 463 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 464                  int lable_id);
 465 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 466                   int lable_id);
 467
 468 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 469                              const char *name);
 470 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 471                                    const char *name);
 472
 473 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 474                          LLVMTypeRef type);
 475
 476 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 477                             unsigned count);
 478
 479 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 480                              unsigned rshift, unsigned bitwidth);
 481
 482 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 483                               LLVMValueRef *addr, bool is_array_tex);
 484
 485 LLVMValueRef
 486 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 487
 488 LLVMValueRef
 489 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 490
 491 LLVMValueRef
 492 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 493
 494 LLVMValueRef
 495 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 496
 497 LLVMValueRef
 498 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 499
 500 LLVMValueRef
 501 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 502
 503 LLVMValueRef
 504 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 505
 506 LLVMValueRef
 507 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 508                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 509
 510 LLVMValueRef
 511 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 512
 513 #ifdef __cplusplus
 514 }
 515 #endif
 516
 517 #endif