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
 179 LLVMValueRef
 180 ac_build_fdiv(struct ac_llvm_context *ctx,
 181               LLVMValueRef num,
 182               LLVMValueRef den);
 183
 184 void
 185 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 186                        bool is_deriv, bool is_array, bool is_lod,
 187                        LLVMValueRef *coords_arg,
 188                        LLVMValueRef *derivs_arg);
 189
 190
 191 LLVMValueRef
 192 ac_build_fs_interp(struct ac_llvm_context *ctx,
 193                    LLVMValueRef llvm_chan,
 194                    LLVMValueRef attr_number,
 195                    LLVMValueRef params,
 196                    LLVMValueRef i,
 197                    LLVMValueRef j);
 198
 199 LLVMValueRef
 200 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 201                        LLVMValueRef parameter,
 202                        LLVMValueRef llvm_chan,
 203                        LLVMValueRef attr_number,
 204                        LLVMValueRef params);
 205
 206 LLVMValueRef
 207 ac_build_gep0(struct ac_llvm_context *ctx,
 208               LLVMValueRef base_ptr,
 209               LLVMValueRef index);
 210 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 211                                   LLVMValueRef index);
 212
 213 void
 214 ac_build_indexed_store(struct ac_llvm_context *ctx,
 215                        LLVMValueRef base_ptr, LLVMValueRef index,
 216                        LLVMValueRef value);
 217
 218 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 219                            LLVMValueRef index);
 220 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 221                                      LLVMValueRef base_ptr, LLVMValueRef index);
 222 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 223                                    LLVMValueRef base_ptr, LLVMValueRef index);
 224 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 225                                    LLVMValueRef base_ptr, LLVMValueRef index);
 226
 227 void
 228 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 229                             LLVMValueRef rsrc,
 230                             LLVMValueRef vdata,
 231                             unsigned num_channels,
 232                             LLVMValueRef voffset,
 233                             LLVMValueRef soffset,
 234                             unsigned inst_offset,
 235                             bool glc,
 236                             bool slc,
 237                             bool writeonly_memory,
 238                             bool swizzle_enable_hint);
 239 LLVMValueRef
 240 ac_build_buffer_load(struct ac_llvm_context *ctx,
 241                      LLVMValueRef rsrc,
 242                      int num_channels,
 243                      LLVMValueRef vindex,
 244                      LLVMValueRef voffset,
 245                      LLVMValueRef soffset,
 246                      unsigned inst_offset,
 247                      unsigned glc,
 248                      unsigned slc,
 249                      bool can_speculate,
 250                      bool allow_smem);
 251
 252 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 253                                          LLVMValueRef rsrc,
 254                                          LLVMValueRef vindex,
 255                                          LLVMValueRef voffset,
 256                                          unsigned num_channels,
 257                                          bool glc,
 258                                          bool can_speculate);
 259
 260 /* load_format that handles the stride & element count better if idxen is
 261  * disabled by LLVM. */
 262 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 263                                                   LLVMValueRef rsrc,
 264                                                   LLVMValueRef vindex,
 265                                                   LLVMValueRef voffset,
 266                                                   unsigned num_channels,
 267                                                   bool glc,
 268                                                   bool can_speculate);
 269
 270 LLVMValueRef
 271 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 272                             LLVMValueRef rsrc,
 273                             LLVMValueRef vindex,
 274                             LLVMValueRef voffset,
 275                                 LLVMValueRef soffset,
 276                                 LLVMValueRef immoffset);
 277
 278 LLVMValueRef
 279 ac_get_thread_id(struct ac_llvm_context *ctx);
 280
 281 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 282 #define AC_TID_MASK_TOP      0xfffffffd
 283 #define AC_TID_MASK_LEFT     0xfffffffe
 284
 285 LLVMValueRef
 286 ac_build_ddxy(struct ac_llvm_context *ctx,
 287               uint32_t mask,
 288               int idx,
 289               LLVMValueRef val);
 290
 291 #define AC_SENDMSG_GS 2
 292 #define AC_SENDMSG_GS_DONE 3
 293
 294 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 295 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 296 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 297 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 298
 299 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 300                       uint32_t msg,
 301                       LLVMValueRef wave_id);
 302
 303 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 304                            LLVMValueRef arg,
 305                            LLVMTypeRef dst_type);
 306
 307 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 308                           LLVMValueRef arg,
 309                           LLVMTypeRef dst_type);
 310 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 311                            LLVMValueRef b);
 312 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 313                            LLVMValueRef b);
 314 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 315                            LLVMValueRef b);
 316 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 317                            LLVMValueRef b);
 318 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 319 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 320
 321 struct ac_export_args {
 322         LLVMValueRef out[4];
 323         unsigned target;
 324         unsigned enabled_channels;
 325         bool compr;
 326         bool done;
 327         bool valid_mask;
 328 };
 329
 330 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 331
 332 void ac_build_export_null(struct ac_llvm_context *ctx);
 333
 334 enum ac_image_opcode {
 335         ac_image_sample,
 336         ac_image_gather4,
 337         ac_image_load,
 338         ac_image_load_mip,
 339         ac_image_store,
 340         ac_image_store_mip,
 341         ac_image_get_lod,
 342         ac_image_get_resinfo,
 343         ac_image_atomic,
 344         ac_image_atomic_cmpswap,
 345 };
 346
 347 enum ac_atomic_op {
 348         ac_atomic_swap,
 349         ac_atomic_add,
 350         ac_atomic_sub,
 351         ac_atomic_smin,
 352         ac_atomic_umin,
 353         ac_atomic_smax,
 354         ac_atomic_umax,
 355         ac_atomic_and,
 356         ac_atomic_or,
 357         ac_atomic_xor,
 358 };
 359
 360 enum ac_image_dim {
 361         ac_image_1d,
 362         ac_image_2d,
 363         ac_image_3d,
 364         ac_image_cube, // includes cube arrays
 365         ac_image_1darray,
 366         ac_image_2darray,
 367         ac_image_2dmsaa,
 368         ac_image_2darraymsaa,
 369 };
 370
 371 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 372 enum ac_image_cache_policy {
 373         ac_glc = 1 << 0,
 374         ac_slc = 1 << 1,
 375 };
 376
 377 struct ac_image_args {
 378         enum ac_image_opcode opcode : 4;
 379         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 380         enum ac_image_dim dim : 3;
 381         unsigned dmask : 4;
 382         unsigned cache_policy : 2;
 383         bool unorm : 1;
 384         bool level_zero : 1;
 385         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 386
 387         LLVMValueRef resource;
 388         LLVMValueRef sampler;
 389         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 390         LLVMValueRef offset;
 391         LLVMValueRef bias;
 392         LLVMValueRef compare;
 393         LLVMValueRef derivs[6];
 394         LLVMValueRef coords[4];
 395         LLVMValueRef lod; // also used by ac_image_get_resinfo
 396 };
 397
 398 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 399                                    struct ac_image_args *a);
 400 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 401                                     LLVMValueRef args[2]);
 402 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 403                                      LLVMValueRef args[2]);
 404 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 405                                      LLVMValueRef args[2]);
 406 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 407                                  LLVMValueRef args[2], unsigned bits, bool hi);
 408 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 409                                  LLVMValueRef args[2], unsigned bits, bool hi);
 410 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 411 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 412 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 413                           LLVMValueRef offset, LLVMValueRef width,
 414                           bool is_signed);
 415 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 416                            LLVMValueRef s1, LLVMValueRef s2);
 417 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 418                            LLVMValueRef s1, LLVMValueRef s2);
 419
 420 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
 421
 422 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 423                            unsigned bitsize);
 424
 425 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 426                             unsigned bitsize);
 427
 428 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 429                             unsigned bitsize);
 430
 431 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 432
 433 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 434                                        LLVMValueRef src0);
 435
 436 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 437                             LLVMValueRef main_fn,
 438                             uint8_t *vs_output_param_offset,
 439                             uint32_t num_outputs,
 440                             uint8_t *num_param_exports);
 441 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 442
 443 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 444 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 445                          LLVMValueRef dw_addr);
 446 void ac_lds_store(struct ac_llvm_context *ctx,
 447                   LLVMValueRef dw_addr, LLVMValueRef value);
 448
 449 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 450                          LLVMTypeRef dst_type,
 451                          LLVMValueRef src0);
 452
 453 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 454 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 455
 456 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 457 void ac_build_break(struct ac_llvm_context *ctx);
 458 void ac_build_continue(struct ac_llvm_context *ctx);
 459 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 460 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 461 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 462 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 463                  int lable_id);
 464 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 465                   int lable_id);
 466
 467 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 468                              const char *name);
 469 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 470                                    const char *name);
 471
 472 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 473                          LLVMTypeRef type);
 474
 475 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 476                             unsigned count);
 477
 478 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 479                              unsigned rshift, unsigned bitwidth);
 480
 481 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 482                               LLVMValueRef *addr, bool is_array_tex);
 483
 484 LLVMValueRef
 485 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 486
 487 LLVMValueRef
 488 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 489
 490 LLVMValueRef
 491 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 492
 493 LLVMValueRef
 494 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 495
 496 LLVMValueRef
 497 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 498
 499 LLVMValueRef
 500 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 501
 502 LLVMValueRef
 503 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 504
 505 LLVMValueRef
 506 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 507                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 508
 509 LLVMValueRef
 510 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 511
 512 #ifdef __cplusplus
 513 }
 514 #endif
 515
 516 #endif