src/compiler/nir/nir_builder.h

   1 /*
   2  * Copyright © 2014-2015 Broadcom
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #ifndef NIR_BUILDER_H
  25 #define NIR_BUILDER_H
  26
  27 #include "nir_control_flow.h"
  28
  29 struct exec_list;
  30
  31 typedef struct nir_builder {
  32    nir_cursor cursor;
  33
  34    /* Whether new ALU instructions will be marked "exact" */
  35    bool exact;
  36
  37    nir_shader *shader;
  38    nir_function_impl *impl;
  39 } nir_builder;
  40
  41 static inline void
  42 nir_builder_init(nir_builder *build, nir_function_impl *impl)
  43 {
  44    memset(build, 0, sizeof(*build));
  45    build->exact = false;
  46    build->impl = impl;
  47    build->shader = impl->function->shader;
  48 }
  49
  50 static inline void
  51 nir_builder_init_simple_shader(nir_builder *build, void *mem_ctx,
  52                                gl_shader_stage stage,
  53                                const nir_shader_compiler_options *options)
  54 {
  55    build->shader = nir_shader_create(mem_ctx, stage, options);
  56    nir_function *func = nir_function_create(build->shader, "main");
  57    build->exact = false;
  58    build->impl = nir_function_impl_create(func);
  59    build->cursor = nir_after_cf_list(&build->impl->body);
  60 }
  61
  62 static inline void
  63 nir_builder_instr_insert(nir_builder *build, nir_instr *instr)
  64 {
  65    nir_instr_insert(build->cursor, instr);
  66
  67    /* Move the cursor forward. */
  68    build->cursor = nir_after_instr(instr);
  69 }
  70
  71 static inline void
  72 nir_builder_cf_insert(nir_builder *build, nir_cf_node *cf)
  73 {
  74    nir_cf_node_insert(build->cursor, cf);
  75 }
  76
  77 static inline nir_ssa_def *
  78 nir_ssa_undef(nir_builder *build, unsigned num_components, unsigned bit_size)
  79 {
  80    nir_ssa_undef_instr *undef =
  81       nir_ssa_undef_instr_create(build->shader, num_components, bit_size);
  82    if (!undef)
  83       return NULL;
  84
  85    nir_instr_insert(nir_before_cf_list(&build->impl->body), &undef->instr);
  86
  87    return &undef->def;
  88 }
  89
  90 static inline nir_ssa_def *
  91 nir_build_imm(nir_builder *build, unsigned num_components, nir_const_value value)
  92 {
  93    nir_load_const_instr *load_const =
  94       nir_load_const_instr_create(build->shader, num_components, 32);
  95    if (!load_const)
  96       return NULL;
  97
  98    load_const->value = value;
  99
 100    nir_builder_instr_insert(build, &load_const->instr);
 101
 102    return &load_const->def;
 103 }
 104
 105 static inline nir_ssa_def *
 106 nir_imm_float(nir_builder *build, float x)
 107 {
 108    nir_const_value v;
 109
 110    memset(&v, 0, sizeof(v));
 111    v.f32[0] = x;
 112
 113    return nir_build_imm(build, 1, v);
 114 }
 115
 116 static inline nir_ssa_def *
 117 nir_imm_vec4(nir_builder *build, float x, float y, float z, float w)
 118 {
 119    nir_const_value v;
 120
 121    memset(&v, 0, sizeof(v));
 122    v.f32[0] = x;
 123    v.f32[1] = y;
 124    v.f32[2] = z;
 125    v.f32[3] = w;
 126
 127    return nir_build_imm(build, 4, v);
 128 }
 129
 130 static inline nir_ssa_def *
 131 nir_imm_int(nir_builder *build, int x)
 132 {
 133    nir_const_value v;
 134
 135    memset(&v, 0, sizeof(v));
 136    v.i32[0] = x;
 137
 138    return nir_build_imm(build, 1, v);
 139 }
 140
 141 static inline nir_ssa_def *
 142 nir_imm_ivec4(nir_builder *build, int x, int y, int z, int w)
 143 {
 144    nir_const_value v;
 145
 146    memset(&v, 0, sizeof(v));
 147    v.i32[0] = x;
 148    v.i32[1] = y;
 149    v.i32[2] = z;
 150    v.i32[3] = w;
 151
 152    return nir_build_imm(build, 4, v);
 153 }
 154
 155 static inline nir_ssa_def *
 156 nir_build_alu(nir_builder *build, nir_op op, nir_ssa_def *src0,
 157               nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)
 158 {
 159    const nir_op_info *op_info = &nir_op_infos[op];
 160    nir_alu_instr *instr = nir_alu_instr_create(build->shader, op);
 161    if (!instr)
 162       return NULL;
 163
 164    instr->exact = build->exact;
 165
 166    instr->src[0].src = nir_src_for_ssa(src0);
 167    if (src1)
 168       instr->src[1].src = nir_src_for_ssa(src1);
 169    if (src2)
 170       instr->src[2].src = nir_src_for_ssa(src2);
 171    if (src3)
 172       instr->src[3].src = nir_src_for_ssa(src3);
 173
 174    /* Guess the number of components the destination temporary should have
 175     * based on our input sizes, if it's not fixed for the op.
 176     */
 177    unsigned num_components = op_info->output_size;
 178    if (num_components == 0) {
 179       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 180          if (op_info->input_sizes[i] == 0)
 181             num_components = MAX2(num_components,
 182                                   instr->src[i].src.ssa->num_components);
 183       }
 184    }
 185    assert(num_components != 0);
 186
 187    /* Figure out the bitwidth based on the source bitwidth if the instruction
 188     * is variable-width.
 189     */
 190    unsigned bit_size = nir_alu_type_get_type_size(op_info->output_type);
 191    if (bit_size == 0) {
 192       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 193          unsigned src_bit_size = instr->src[i].src.ssa->bit_size;
 194          if (nir_alu_type_get_type_size(op_info->input_types[i]) == 0) {
 195             if (bit_size)
 196                assert(src_bit_size == bit_size);
 197             else
 198                bit_size = src_bit_size;
 199          } else {
 200             assert(src_bit_size ==
 201                nir_alu_type_get_type_size(op_info->input_types[i]));
 202          }
 203       }
 204    }
 205
 206    /* Make sure we don't swizzle from outside of our source vector (like if a
 207     * scalar value was passed into a multiply with a vector).
 208     */
 209    for (unsigned i = 0; i < op_info->num_inputs; i++) {
 210       for (unsigned j = instr->src[i].src.ssa->num_components; j < 4; j++) {
 211          instr->src[i].swizzle[j] = instr->src[i].src.ssa->num_components - 1;
 212       }
 213    }
 214
 215    nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
 216                      bit_size, NULL);
 217    instr->dest.write_mask = (1 << num_components) - 1;
 218
 219    nir_builder_instr_insert(build, &instr->instr);
 220
 221    return &instr->dest.dest.ssa;
 222 }
 223
 224 #define ALU1(op)                                                          \
 225 static inline nir_ssa_def *                                               \
 226 nir_##op(nir_builder *build, nir_ssa_def *src0)                           \
 227 {                                                                         \
 228    return nir_build_alu(build, nir_op_##op, src0, NULL, NULL, NULL);      \
 229 }
 230
 231 #define ALU2(op)                                                          \
 232 static inline nir_ssa_def *                                               \
 233 nir_##op(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)        \
 234 {                                                                         \
 235    return nir_build_alu(build, nir_op_##op, src0, src1, NULL, NULL);      \
 236 }
 237
 238 #define ALU3(op)                                                          \
 239 static inline nir_ssa_def *                                               \
 240 nir_##op(nir_builder *build, nir_ssa_def *src0,                           \
 241          nir_ssa_def *src1, nir_ssa_def *src2)                            \
 242 {                                                                         \
 243    return nir_build_alu(build, nir_op_##op, src0, src1, src2, NULL);      \
 244 }
 245
 246 #define ALU4(op)                                                          \
 247 static inline nir_ssa_def *                                               \
 248 nir_##op(nir_builder *build, nir_ssa_def *src0,                           \
 249          nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)         \
 250 {                                                                         \
 251    return nir_build_alu(build, nir_op_##op, src0, src1, src2, src3);      \
 252 }
 253
 254 #include "nir_builder_opcodes.h"
 255
 256 static inline nir_ssa_def *
 257 nir_vec(nir_builder *build, nir_ssa_def **comp, unsigned num_components)
 258 {
 259    switch (num_components) {
 260    case 4:
 261       return nir_vec4(build, comp[0], comp[1], comp[2], comp[3]);
 262    case 3:
 263       return nir_vec3(build, comp[0], comp[1], comp[2]);
 264    case 2:
 265       return nir_vec2(build, comp[0], comp[1]);
 266    case 1:
 267       return comp[0];
 268    default:
 269       unreachable("bad component count");
 270       return NULL;
 271    }
 272 }
 273
 274 /**
 275  * Similar to nir_fmov, but takes a nir_alu_src instead of a nir_ssa_def.
 276  */
 277 static inline nir_ssa_def *
 278 nir_fmov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 279 {
 280    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_fmov);
 281    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 282                      nir_src_bit_size(src.src), NULL);
 283    mov->exact = build->exact;
 284    mov->dest.write_mask = (1 << num_components) - 1;
 285    mov->src[0] = src;
 286    nir_builder_instr_insert(build, &mov->instr);
 287
 288    return &mov->dest.dest.ssa;
 289 }
 290
 291 static inline nir_ssa_def *
 292 nir_imov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 293 {
 294    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_imov);
 295    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 296                      nir_src_bit_size(src.src), NULL);
 297    mov->exact = build->exact;
 298    mov->dest.write_mask = (1 << num_components) - 1;
 299    mov->src[0] = src;
 300    nir_builder_instr_insert(build, &mov->instr);
 301
 302    return &mov->dest.dest.ssa;
 303 }
 304
 305 /**
 306  * Construct an fmov or imov that reswizzles the source's components.
 307  */
 308 static inline nir_ssa_def *
 309 nir_swizzle(nir_builder *build, nir_ssa_def *src, unsigned swiz[4],
 310             unsigned num_components, bool use_fmov)
 311 {
 312    nir_alu_src alu_src = { NIR_SRC_INIT };
 313    alu_src.src = nir_src_for_ssa(src);
 314    for (unsigned i = 0; i < num_components; i++)
 315       alu_src.swizzle[i] = swiz[i];
 316
 317    return use_fmov ? nir_fmov_alu(build, alu_src, num_components) :
 318                      nir_imov_alu(build, alu_src, num_components);
 319 }
 320
 321 /* Selects the right fdot given the number of components in each source. */
 322 static inline nir_ssa_def *
 323 nir_fdot(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)
 324 {
 325    assert(src0->num_components == src1->num_components);
 326    switch (src0->num_components) {
 327    case 1: return nir_fmul(build, src0, src1);
 328    case 2: return nir_fdot2(build, src0, src1);
 329    case 3: return nir_fdot3(build, src0, src1);
 330    case 4: return nir_fdot4(build, src0, src1);
 331    default:
 332       unreachable("bad component size");
 333    }
 334
 335    return NULL;
 336 }
 337
 338 static inline nir_ssa_def *
 339 nir_channel(nir_builder *b, nir_ssa_def *def, unsigned c)
 340 {
 341    unsigned swizzle[4] = {c, c, c, c};
 342    return nir_swizzle(b, def, swizzle, 1, false);
 343 }
 344
 345 /**
 346  * Turns a nir_src into a nir_ssa_def * so it can be passed to
 347  * nir_build_alu()-based builder calls.
 348  *
 349  * See nir_ssa_for_alu_src() for alu instructions.
 350  */
 351 static inline nir_ssa_def *
 352 nir_ssa_for_src(nir_builder *build, nir_src src, int num_components)
 353 {
 354    if (src.is_ssa && src.ssa->num_components == num_components)
 355       return src.ssa;
 356
 357    nir_alu_src alu = { NIR_SRC_INIT };
 358    alu.src = src;
 359    for (int j = 0; j < 4; j++)
 360       alu.swizzle[j] = j;
 361
 362    return nir_imov_alu(build, alu, num_components);
 363 }
 364
 365 /**
 366  * Similar to nir_ssa_for_src(), but for alu src's, respecting the
 367  * nir_alu_src's swizzle.
 368  */
 369 static inline nir_ssa_def *
 370 nir_ssa_for_alu_src(nir_builder *build, nir_alu_instr *instr, unsigned srcn)
 371 {
 372    static uint8_t trivial_swizzle[4] = { 0, 1, 2, 3 };
 373    nir_alu_src *src = &instr->src[srcn];
 374    unsigned num_components = nir_ssa_alu_instr_src_components(instr, srcn);
 375
 376    if (src->src.is_ssa && (src->src.ssa->num_components == num_components) &&
 377        !src->abs && !src->negate &&
 378        (memcmp(src->swizzle, trivial_swizzle, num_components) == 0))
 379       return src->src.ssa;
 380
 381    return nir_imov_alu(build, *src, num_components);
 382 }
 383
 384 static inline nir_ssa_def *
 385 nir_load_var(nir_builder *build, nir_variable *var)
 386 {
 387    const unsigned num_components = glsl_get_vector_elements(var->type);
 388
 389    nir_intrinsic_instr *load =
 390       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_var);
 391    load->num_components = num_components;
 392    load->variables[0] = nir_deref_var_create(load, var);
 393    nir_ssa_dest_init(&load->instr, &load->dest, num_components,
 394                      glsl_get_bit_size(glsl_get_base_type(var->type)), NULL);
 395    nir_builder_instr_insert(build, &load->instr);
 396    return &load->dest.ssa;
 397 }
 398
 399 static inline void
 400 nir_store_var(nir_builder *build, nir_variable *var, nir_ssa_def *value,
 401               unsigned writemask)
 402 {
 403    const unsigned num_components = glsl_get_vector_elements(var->type);
 404
 405    nir_intrinsic_instr *store =
 406       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var);
 407    store->num_components = num_components;
 408    nir_intrinsic_set_write_mask(store, writemask);
 409    store->variables[0] = nir_deref_var_create(store, var);
 410    store->src[0] = nir_src_for_ssa(value);
 411    nir_builder_instr_insert(build, &store->instr);
 412 }
 413
 414 static inline void
 415 nir_store_deref_var(nir_builder *build, nir_deref_var *deref,
 416                     nir_ssa_def *value, unsigned writemask)
 417 {
 418    const unsigned num_components =
 419       glsl_get_vector_elements(nir_deref_tail(&deref->deref)->type);
 420
 421    nir_intrinsic_instr *store =
 422       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var);
 423    store->num_components = num_components;
 424    store->const_index[0] = writemask & ((1 << num_components) - 1);
 425    store->variables[0] = nir_deref_as_var(nir_copy_deref(store, &deref->deref));
 426    store->src[0] = nir_src_for_ssa(value);
 427    nir_builder_instr_insert(build, &store->instr);
 428 }
 429
 430 static inline void
 431 nir_copy_deref_var(nir_builder *build, nir_deref_var *dest, nir_deref_var *src)
 432 {
 433    assert(nir_deref_tail(&dest->deref)->type ==
 434           nir_deref_tail(&src->deref)->type);
 435
 436    nir_intrinsic_instr *copy =
 437       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_var);
 438    copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
 439    copy->variables[1] = nir_deref_as_var(nir_copy_deref(copy, &src->deref));
 440    nir_builder_instr_insert(build, &copy->instr);
 441 }
 442
 443 static inline void
 444 nir_copy_var(nir_builder *build, nir_variable *dest, nir_variable *src)
 445 {
 446    nir_intrinsic_instr *copy =
 447       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_var);
 448    copy->variables[0] = nir_deref_var_create(copy, dest);
 449    copy->variables[1] = nir_deref_var_create(copy, src);
 450    nir_builder_instr_insert(build, &copy->instr);
 451 }
 452
 453 static inline nir_ssa_def *
 454 nir_load_system_value(nir_builder *build, nir_intrinsic_op op, int index)
 455 {
 456    nir_intrinsic_instr *load = nir_intrinsic_instr_create(build->shader, op);
 457    load->num_components = nir_intrinsic_infos[op].dest_components;
 458    load->const_index[0] = index;
 459    nir_ssa_dest_init(&load->instr, &load->dest,
 460                      nir_intrinsic_infos[op].dest_components, 32, NULL);
 461    nir_builder_instr_insert(build, &load->instr);
 462    return &load->dest.ssa;
 463 }
 464
 465 static inline void
 466 nir_jump(nir_builder *build, nir_jump_type jump_type)
 467 {
 468    nir_jump_instr *jump = nir_jump_instr_create(build->shader, jump_type);
 469    nir_builder_instr_insert(build, &jump->instr);
 470 }
 471
 472 #endif /* NIR_BUILDER_H */