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