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         /* CONST is the only address space that selects SMEM loads */
  41         AC_CONST_ADDR_SPACE = HAVE_LLVM >= 0x700 ? 4 : 2,
  42         AC_LOCAL_ADDR_SPACE = 3,
  43         AC_CONST_32BIT_ADDR_SPACE = 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 v4f32;
  75         LLVMTypeRef v8i32;
  76
  77         LLVMValueRef i16_0;
  78         LLVMValueRef i16_1;
  79         LLVMValueRef i32_0;
  80         LLVMValueRef i32_1;
  81         LLVMValueRef i64_0;
  82         LLVMValueRef i64_1;
  83         LLVMValueRef f32_0;
  84         LLVMValueRef f32_1;
  85         LLVMValueRef f64_0;
  86         LLVMValueRef f64_1;
  87         LLVMValueRef i1true;
  88         LLVMValueRef i1false;
  89
  90         struct ac_llvm_flow *flow;
  91         unsigned flow_depth;
  92         unsigned flow_depth_max;
  93
  94         unsigned range_md_kind;
  95         unsigned invariant_load_md_kind;
  96         unsigned uniform_md_kind;
  97         unsigned fpmath_md_kind;
  98         LLVMValueRef fpmath_md_2p5_ulp;
  99         LLVMValueRef empty_md;
 100
 101         enum chip_class chip_class;
 102         enum radeon_family family;
 103
 104         LLVMValueRef lds;
 105 };
 106
 107 void
 108 ac_llvm_context_init(struct ac_llvm_context *ctx,
 109                      enum chip_class chip_class, enum radeon_family family);
 110
 111 void
 112 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
 113
 114 int
 115 ac_get_llvm_num_components(LLVMValueRef value);
 116
 117 int
 118 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 119
 120 LLVMValueRef
 121 ac_llvm_extract_elem(struct ac_llvm_context *ac,
 122                      LLVMValueRef value,
 123                      int index);
 124
 125 unsigned ac_get_type_size(LLVMTypeRef type);
 126
 127 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 128 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
 129 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
 130 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
 131
 132 LLVMValueRef
 133 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
 134                    LLVMTypeRef return_type, LLVMValueRef *params,
 135                    unsigned param_count, unsigned attrib_mask);
 136
 137 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 138
 139 LLVMValueRef
 140 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
 141              unsigned count_incoming, LLVMValueRef *values,
 142              LLVMBasicBlockRef *blocks);
 143
 144 void ac_build_s_barrier(struct ac_llvm_context *ctx);
 145 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
 146                                    LLVMValueRef *pvgpr);
 147
 148 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
 149
 150 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 151
 152 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
 153
 154 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
 155
 156 LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
 157
 158 LLVMValueRef
 159 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
 160                                unsigned value_count, unsigned component);
 161
 162 LLVMValueRef
 163 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
 164                                 LLVMValueRef *values,
 165                                 unsigned value_count,
 166                                 unsigned value_stride,
 167                                 bool load,
 168                                 bool always_vector);
 169 LLVMValueRef
 170 ac_build_gather_values(struct ac_llvm_context *ctx,
 171                        LLVMValueRef *values,
 172                        unsigned value_count);
 173 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 174                                      LLVMValueRef value,
 175                                      unsigned num_channels);
 176
 177 LLVMValueRef
 178 ac_build_fdiv(struct ac_llvm_context *ctx,
 179               LLVMValueRef num,
 180               LLVMValueRef den);
 181
 182 void
 183 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 184                        bool is_deriv, bool is_array, bool is_lod,
 185                        LLVMValueRef *coords_arg,
 186                        LLVMValueRef *derivs_arg);
 187
 188
 189 LLVMValueRef
 190 ac_build_fs_interp(struct ac_llvm_context *ctx,
 191                    LLVMValueRef llvm_chan,
 192                    LLVMValueRef attr_number,
 193                    LLVMValueRef params,
 194                    LLVMValueRef i,
 195                    LLVMValueRef j);
 196
 197 LLVMValueRef
 198 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
 199                        LLVMValueRef parameter,
 200                        LLVMValueRef llvm_chan,
 201                        LLVMValueRef attr_number,
 202                        LLVMValueRef params);
 203
 204 LLVMValueRef
 205 ac_build_gep0(struct ac_llvm_context *ctx,
 206               LLVMValueRef base_ptr,
 207               LLVMValueRef index);
 208 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 209                                   LLVMValueRef index);
 210
 211 void
 212 ac_build_indexed_store(struct ac_llvm_context *ctx,
 213                        LLVMValueRef base_ptr, LLVMValueRef index,
 214                        LLVMValueRef value);
 215
 216 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
 217                            LLVMValueRef index);
 218 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
 219                                      LLVMValueRef base_ptr, LLVMValueRef index);
 220 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
 221                                    LLVMValueRef base_ptr, LLVMValueRef index);
 222 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
 223                                    LLVMValueRef base_ptr, LLVMValueRef index);
 224
 225 void
 226 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 227                             LLVMValueRef rsrc,
 228                             LLVMValueRef vdata,
 229                             unsigned num_channels,
 230                             LLVMValueRef voffset,
 231                             LLVMValueRef soffset,
 232                             unsigned inst_offset,
 233                             bool glc,
 234                             bool slc,
 235                             bool writeonly_memory,
 236                             bool swizzle_enable_hint);
 237 LLVMValueRef
 238 ac_build_buffer_load(struct ac_llvm_context *ctx,
 239                      LLVMValueRef rsrc,
 240                      int num_channels,
 241                      LLVMValueRef vindex,
 242                      LLVMValueRef voffset,
 243                      LLVMValueRef soffset,
 244                      unsigned inst_offset,
 245                      unsigned glc,
 246                      unsigned slc,
 247                      bool can_speculate,
 248                      bool allow_smem);
 249
 250 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 251                                          LLVMValueRef rsrc,
 252                                          LLVMValueRef vindex,
 253                                          LLVMValueRef voffset,
 254                                          unsigned num_channels,
 255                                          bool glc,
 256                                          bool can_speculate);
 257
 258 /* load_format that handles the stride & element count better if idxen is
 259  * disabled by LLVM. */
 260 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 261                                                   LLVMValueRef rsrc,
 262                                                   LLVMValueRef vindex,
 263                                                   LLVMValueRef voffset,
 264                                                   unsigned num_channels,
 265                                                   bool glc,
 266                                                   bool can_speculate);
 267
 268 LLVMValueRef
 269 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
 270                             LLVMValueRef rsrc,
 271                             LLVMValueRef vindex,
 272                             LLVMValueRef voffset,
 273                                 LLVMValueRef soffset,
 274                                 LLVMValueRef immoffset);
 275
 276 LLVMValueRef
 277 ac_get_thread_id(struct ac_llvm_context *ctx);
 278
 279 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
 280 #define AC_TID_MASK_TOP      0xfffffffd
 281 #define AC_TID_MASK_LEFT     0xfffffffe
 282
 283 LLVMValueRef
 284 ac_build_ddxy(struct ac_llvm_context *ctx,
 285               uint32_t mask,
 286               int idx,
 287               LLVMValueRef val);
 288
 289 #define AC_SENDMSG_GS 2
 290 #define AC_SENDMSG_GS_DONE 3
 291
 292 #define AC_SENDMSG_GS_OP_NOP      (0 << 4)
 293 #define AC_SENDMSG_GS_OP_CUT      (1 << 4)
 294 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 295 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 296
 297 void ac_build_sendmsg(struct ac_llvm_context *ctx,
 298                       uint32_t msg,
 299                       LLVMValueRef wave_id);
 300
 301 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
 302                            LLVMValueRef arg,
 303                            LLVMTypeRef dst_type);
 304
 305 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
 306                           LLVMValueRef arg,
 307                           LLVMTypeRef dst_type);
 308 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 309                            LLVMValueRef b);
 310 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 311                            LLVMValueRef b);
 312 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
 313                            LLVMValueRef b);
 314 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
 315                            LLVMValueRef b);
 316 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 317 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 318
 319 struct ac_export_args {
 320         LLVMValueRef out[4];
 321         unsigned target;
 322         unsigned enabled_channels;
 323         bool compr;
 324         bool done;
 325         bool valid_mask;
 326 };
 327
 328 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
 329
 330 void ac_build_export_null(struct ac_llvm_context *ctx);
 331
 332 enum ac_image_opcode {
 333         ac_image_sample,
 334         ac_image_gather4,
 335         ac_image_load,
 336         ac_image_load_mip,
 337         ac_image_store,
 338         ac_image_store_mip,
 339         ac_image_get_lod,
 340         ac_image_get_resinfo,
 341         ac_image_atomic,
 342         ac_image_atomic_cmpswap,
 343 };
 344
 345 enum ac_atomic_op {
 346         ac_atomic_swap,
 347         ac_atomic_add,
 348         ac_atomic_sub,
 349         ac_atomic_smin,
 350         ac_atomic_umin,
 351         ac_atomic_smax,
 352         ac_atomic_umax,
 353         ac_atomic_and,
 354         ac_atomic_or,
 355         ac_atomic_xor,
 356 };
 357
 358 enum ac_image_dim {
 359         ac_image_1d,
 360         ac_image_2d,
 361         ac_image_3d,
 362         ac_image_cube, // includes cube arrays
 363         ac_image_1darray,
 364         ac_image_2darray,
 365         ac_image_2dmsaa,
 366         ac_image_2darraymsaa,
 367 };
 368
 369 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
 370 enum ac_image_cache_policy {
 371         ac_glc = 1 << 0,
 372         ac_slc = 1 << 1,
 373 };
 374
 375 struct ac_image_args {
 376         enum ac_image_opcode opcode : 4;
 377         enum ac_atomic_op atomic : 4; /* for the ac_image_atomic opcode */
 378         enum ac_image_dim dim : 3;
 379         unsigned dmask : 4;
 380         unsigned cache_policy : 2;
 381         bool unorm : 1;
 382         bool level_zero : 1;
 383         unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 384
 385         LLVMValueRef resource;
 386         LLVMValueRef sampler;
 387         LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
 388         LLVMValueRef offset;
 389         LLVMValueRef bias;
 390         LLVMValueRef compare;
 391         LLVMValueRef derivs[6];
 392         LLVMValueRef coords[4];
 393         LLVMValueRef lod; // also used by ac_image_get_resinfo
 394 };
 395
 396 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 397                                    struct ac_image_args *a);
 398 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 399                                     LLVMValueRef args[2]);
 400 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
 401                                      LLVMValueRef args[2]);
 402 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
 403                                      LLVMValueRef args[2]);
 404 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
 405                                  LLVMValueRef args[2], unsigned bits, bool hi);
 406 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
 407                                  LLVMValueRef args[2], unsigned bits, bool hi);
 408 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 409 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 410 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 411                           LLVMValueRef offset, LLVMValueRef width,
 412                           bool is_signed);
 413 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 414                            LLVMValueRef s1, LLVMValueRef s2);
 415 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 416                            LLVMValueRef s1, LLVMValueRef s2);
 417
 418 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
 419
 420 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
 421                            unsigned bitsize);
 422
 423 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 424                             unsigned bitsize);
 425
 426 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
 427                             unsigned bitsize);
 428
 429 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 430
 431 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
 432                                        LLVMValueRef src0);
 433
 434 void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
 435                             LLVMValueRef main_fn,
 436                             uint8_t *vs_output_param_offset,
 437                             uint32_t num_outputs,
 438                             uint8_t *num_param_exports);
 439 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 440
 441 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
 442 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
 443                          LLVMValueRef dw_addr);
 444 void ac_lds_store(struct ac_llvm_context *ctx,
 445                   LLVMValueRef dw_addr, LLVMValueRef value);
 446
 447 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
 448                          LLVMTypeRef dst_type,
 449                          LLVMValueRef src0);
 450
 451 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
 452 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 453
 454 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
 455 void ac_build_break(struct ac_llvm_context *ctx);
 456 void ac_build_continue(struct ac_llvm_context *ctx);
 457 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
 458 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
 459 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
 460 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
 461                  int lable_id);
 462 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
 463                   int lable_id);
 464
 465 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
 466                              const char *name);
 467 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
 468                                    const char *name);
 469
 470 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 471                          LLVMTypeRef type);
 472
 473 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
 474                             unsigned count);
 475
 476 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
 477                              unsigned rshift, unsigned bitwidth);
 478
 479 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 480                               LLVMValueRef *addr, bool is_array_tex);
 481
 482 LLVMValueRef
 483 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 484
 485 LLVMValueRef
 486 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 487
 488 LLVMValueRef
 489 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane);
 490
 491 LLVMValueRef
 492 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
 493
 494 LLVMValueRef
 495 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 496
 497 LLVMValueRef
 498 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
 499
 500 LLVMValueRef
 501 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
 502
 503 LLVMValueRef
 504 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
 505                 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
 506
 507 LLVMValueRef
 508 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
 509
 510 #ifdef __cplusplus
 511 }
 512 #endif
 513
 514 #endif