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