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    undef->def.bit_size = bit_size;
  83    if (!undef)
  84       return NULL;
  85
  86    nir_instr_insert(nir_before_cf_list(&build->impl->body), &undef->instr);
  87
  88    return &undef->def;
  89 }
  90
  91 static inline nir_ssa_def *
  92 nir_build_imm(nir_builder *build, unsigned num_components, nir_const_value value)
  93 {
  94    nir_load_const_instr *load_const =
  95       nir_load_const_instr_create(build->shader, num_components, 32);
  96    if (!load_const)
  97       return NULL;
  98
  99    load_const->value = value;
 100
 101    nir_builder_instr_insert(build, &load_const->instr);
 102
 103    return &load_const->def;
 104 }
 105
 106 static inline nir_ssa_def *
 107 nir_imm_float(nir_builder *build, float x)
 108 {
 109    nir_const_value v;
 110
 111    memset(&v, 0, sizeof(v));
 112    v.f32[0] = x;
 113
 114    return nir_build_imm(build, 1, v);
 115 }
 116
 117 static inline nir_ssa_def *
 118 nir_imm_vec4(nir_builder *build, float x, float y, float z, float w)
 119 {
 120    nir_const_value v;
 121
 122    memset(&v, 0, sizeof(v));
 123    v.f32[0] = x;
 124    v.f32[1] = y;
 125    v.f32[2] = z;
 126    v.f32[3] = w;
 127
 128    return nir_build_imm(build, 4, v);
 129 }
 130
 131 static inline nir_ssa_def *
 132 nir_imm_int(nir_builder *build, int x)
 133 {
 134    nir_const_value v;
 135
 136    memset(&v, 0, sizeof(v));
 137    v.i32[0] = x;
 138
 139    return nir_build_imm(build, 1, v);
 140 }
 141
 142 static inline nir_ssa_def *
 143 nir_imm_ivec4(nir_builder *build, int x, int y, int z, int w)
 144 {
 145    nir_const_value v;
 146
 147    memset(&v, 0, sizeof(v));
 148    v.i32[0] = x;
 149    v.i32[1] = y;
 150    v.i32[2] = z;
 151    v.i32[3] = w;
 152
 153    return nir_build_imm(build, 4, v);
 154 }
 155
 156 static inline nir_ssa_def *
 157 nir_build_alu(nir_builder *build, nir_op op, nir_ssa_def *src0,
 158               nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)
 159 {
 160    const nir_op_info *op_info = &nir_op_infos[op];
 161    nir_alu_instr *instr = nir_alu_instr_create(build->shader, op);
 162    if (!instr)
 163       return NULL;
 164
 165    instr->exact = build->exact;
 166
 167    instr->src[0].src = nir_src_for_ssa(src0);
 168    if (src1)
 169       instr->src[1].src = nir_src_for_ssa(src1);
 170    if (src2)
 171       instr->src[2].src = nir_src_for_ssa(src2);
 172    if (src3)
 173       instr->src[3].src = nir_src_for_ssa(src3);
 174
 175    /* Guess the number of components the destination temporary should have
 176     * based on our input sizes, if it's not fixed for the op.
 177     */
 178    unsigned num_components = op_info->output_size;
 179    if (num_components == 0) {
 180       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 181          if (op_info->input_sizes[i] == 0)
 182             num_components = MAX2(num_components,
 183                                   instr->src[i].src.ssa->num_components);
 184       }
 185    }
 186    assert(num_components != 0);
 187
 188    /* Figure out the bitwidth based on the source bitwidth if the instruction
 189     * is variable-width.
 190     */
 191    unsigned bit_size = nir_alu_type_get_type_size(op_info->output_type);
 192    if (bit_size == 0) {
 193       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 194          unsigned src_bit_size = instr->src[i].src.ssa->bit_size;
 195          if (nir_alu_type_get_type_size(op_info->input_types[i]) == 0) {
 196             if (bit_size)
 197                assert(src_bit_size == bit_size);
 198             else
 199                bit_size = src_bit_size;
 200          } else {
 201             assert(src_bit_size ==
 202                nir_alu_type_get_type_size(op_info->input_types[i]));
 203          }
 204       }
 205    }
 206
 207    /* Make sure we don't swizzle from outside of our source vector (like if a
 208     * scalar value was passed into a multiply with a vector).
 209     */
 210    for (unsigned i = 0; i < op_info->num_inputs; i++) {
 211       for (unsigned j = instr->src[i].src.ssa->num_components; j < 4; j++) {
 212          instr->src[i].swizzle[j] = instr->src[i].src.ssa->num_components - 1;
 213       }
 214    }
 215
 216    nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
 217                      bit_size, NULL);
 218    instr->dest.write_mask = (1 << num_components) - 1;
 219
 220    nir_builder_instr_insert(build, &instr->instr);
 221
 222    return &instr->dest.dest.ssa;
 223 }
 224
 225 #define ALU1(op)                                                          \
 226 static inline nir_ssa_def *                                               \
 227 nir_##op(nir_builder *build, nir_ssa_def *src0)                           \
 228 {                                                                         \
 229    return nir_build_alu(build, nir_op_##op, src0, NULL, NULL, NULL);      \
 230 }
 231
 232 #define ALU2(op)                                                          \
 233 static inline nir_ssa_def *                                               \
 234 nir_##op(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)        \
 235 {                                                                         \
 236    return nir_build_alu(build, nir_op_##op, src0, src1, NULL, NULL);      \
 237 }
 238
 239 #define ALU3(op)                                                          \
 240 static inline nir_ssa_def *                                               \
 241 nir_##op(nir_builder *build, nir_ssa_def *src0,                           \
 242          nir_ssa_def *src1, nir_ssa_def *src2)                            \
 243 {                                                                         \
 244    return nir_build_alu(build, nir_op_##op, src0, src1, src2, NULL);      \
 245 }
 246
 247 #define ALU4(op)                                                          \
 248 static inline nir_ssa_def *                                               \
 249 nir_##op(nir_builder *build, nir_ssa_def *src0,                           \
 250          nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)         \
 251 {                                                                         \
 252    return nir_build_alu(build, nir_op_##op, src0, src1, src2, src3);      \
 253 }
 254
 255 #include "nir_builder_opcodes.h"
 256
 257 static inline nir_ssa_def *
 258 nir_vec(nir_builder *build, nir_ssa_def **comp, unsigned num_components)
 259 {
 260    switch (num_components) {
 261    case 4:
 262       return nir_vec4(build, comp[0], comp[1], comp[2], comp[3]);
 263    case 3:
 264       return nir_vec3(build, comp[0], comp[1], comp[2]);
 265    case 2:
 266       return nir_vec2(build, comp[0], comp[1]);
 267    case 1:
 268       return comp[0];
 269    default:
 270       unreachable("bad component count");
 271       return NULL;
 272    }
 273 }
 274
 275 /**
 276  * Similar to nir_fmov, but takes a nir_alu_src instead of a nir_ssa_def.
 277  */
 278 static inline nir_ssa_def *
 279 nir_fmov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 280 {
 281    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_fmov);
 282    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 283                      nir_src_bit_size(src.src), NULL);
 284    mov->exact = build->exact;
 285    mov->dest.write_mask = (1 << num_components) - 1;
 286    mov->src[0] = src;
 287    nir_builder_instr_insert(build, &mov->instr);
 288
 289    return &mov->dest.dest.ssa;
 290 }
 291
 292 static inline nir_ssa_def *
 293 nir_imov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 294 {
 295    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_imov);
 296    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 297                      nir_src_bit_size(src.src), NULL);
 298    mov->exact = build->exact;
 299    mov->dest.write_mask = (1 << num_components) - 1;
 300    mov->src[0] = src;
 301    nir_builder_instr_insert(build, &mov->instr);
 302
 303    return &mov->dest.dest.ssa;
 304 }
 305
 306 /**
 307  * Construct an fmov or imov that reswizzles the source's components.
 308  */
 309 static inline nir_ssa_def *
 310 nir_swizzle(nir_builder *build, nir_ssa_def *src, unsigned swiz[4],
 311             unsigned num_components, bool use_fmov)
 312 {
 313    nir_alu_src alu_src = { NIR_SRC_INIT };
 314    alu_src.src = nir_src_for_ssa(src);
 315    for (unsigned i = 0; i < num_components; i++)
 316       alu_src.swizzle[i] = swiz[i];
 317
 318    return use_fmov ? nir_fmov_alu(build, alu_src, num_components) :
 319                      nir_imov_alu(build, alu_src, num_components);
 320 }
 321
 322 /* Selects the right fdot given the number of components in each source. */
 323 static inline nir_ssa_def *
 324 nir_fdot(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)
 325 {
 326    assert(src0->num_components == src1->num_components);
 327    switch (src0->num_components) {
 328    case 1: return nir_fmul(build, src0, src1);
 329    case 2: return nir_fdot2(build, src0, src1);
 330    case 3: return nir_fdot3(build, src0, src1);
 331    case 4: return nir_fdot4(build, src0, src1);
 332    default:
 333       unreachable("bad component size");
 334    }
 335
 336    return NULL;
 337 }
 338
 339 static inline nir_ssa_def *
 340 nir_channel(nir_builder *b, nir_ssa_def *def, unsigned c)
 341 {
 342    unsigned swizzle[4] = {c, c, c, c};
 343    return nir_swizzle(b, def, swizzle, 1, false);
 344 }
 345
 346 /**
 347  * Turns a nir_src into a nir_ssa_def * so it can be passed to
 348  * nir_build_alu()-based builder calls.
 349  *
 350  * See nir_ssa_for_alu_src() for alu instructions.
 351  */
 352 static inline nir_ssa_def *
 353 nir_ssa_for_src(nir_builder *build, nir_src src, int num_components)
 354 {
 355    if (src.is_ssa && src.ssa->num_components == num_components)
 356       return src.ssa;
 357
 358    nir_alu_src alu = { NIR_SRC_INIT };
 359    alu.src = src;
 360    for (int j = 0; j < 4; j++)
 361       alu.swizzle[j] = j;
 362
 363    return nir_imov_alu(build, alu, num_components);
 364 }
 365
 366 /**
 367  * Similar to nir_ssa_for_src(), but for alu src's, respecting the
 368  * nir_alu_src's swizzle.
 369  */
 370 static inline nir_ssa_def *
 371 nir_ssa_for_alu_src(nir_builder *build, nir_alu_instr *instr, unsigned srcn)
 372 {
 373    static uint8_t trivial_swizzle[4] = { 0, 1, 2, 3 };
 374    nir_alu_src *src = &instr->src[srcn];
 375    unsigned num_components = nir_ssa_alu_instr_src_components(instr, srcn);
 376
 377    if (src->src.is_ssa && (src->src.ssa->num_components == num_components) &&
 378        !src->abs && !src->negate &&
 379        (memcmp(src->swizzle, trivial_swizzle, num_components) == 0))
 380       return src->src.ssa;
 381
 382    return nir_imov_alu(build, *src, num_components);
 383 }
 384
 385 static inline nir_ssa_def *
 386 nir_load_var(nir_builder *build, nir_variable *var)
 387 {
 388    const unsigned num_components = glsl_get_vector_elements(var->type);
 389
 390    nir_intrinsic_instr *load =
 391       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_var);
 392    load->num_components = num_components;
 393    load->variables[0] = nir_deref_var_create(load, var);
 394    nir_ssa_dest_init(&load->instr, &load->dest, num_components,
 395                      glsl_get_bit_size(glsl_get_base_type(var->type)), NULL);
 396    nir_builder_instr_insert(build, &load->instr);
 397    return &load->dest.ssa;
 398 }
 399
 400 static inline void
 401 nir_store_var(nir_builder *build, nir_variable *var, nir_ssa_def *value,
 402               unsigned writemask)
 403 {
 404    const unsigned num_components = glsl_get_vector_elements(var->type);
 405
 406    nir_intrinsic_instr *store =
 407       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var);
 408    store->num_components = num_components;
 409    nir_intrinsic_set_write_mask(store, writemask);
 410    store->variables[0] = nir_deref_var_create(store, var);
 411    store->src[0] = nir_src_for_ssa(value);
 412    nir_builder_instr_insert(build, &store->instr);
 413 }
 414
 415 static inline void
 416 nir_store_deref_var(nir_builder *build, nir_deref_var *deref,
 417                     nir_ssa_def *value, unsigned writemask)
 418 {
 419    const unsigned num_components =
 420       glsl_get_vector_elements(nir_deref_tail(&deref->deref)->type);
 421
 422    nir_intrinsic_instr *store =
 423       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var);
 424    store->num_components = num_components;
 425    store->const_index[0] = writemask & ((1 << num_components) - 1);
 426    store->variables[0] = nir_deref_as_var(nir_copy_deref(store, &deref->deref));
 427    store->src[0] = nir_src_for_ssa(value);
 428    nir_builder_instr_insert(build, &store->instr);
 429 }
 430
 431 static inline void
 432 nir_copy_deref_var(nir_builder *build, nir_deref_var *dest, nir_deref_var *src)
 433 {
 434    assert(nir_deref_tail(&dest->deref)->type ==
 435           nir_deref_tail(&src->deref)->type);
 436
 437    nir_intrinsic_instr *copy =
 438       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_var);
 439    copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
 440    copy->variables[1] = nir_deref_as_var(nir_copy_deref(copy, &src->deref));
 441    nir_builder_instr_insert(build, &copy->instr);
 442 }
 443
 444 static inline void
 445 nir_copy_var(nir_builder *build, nir_variable *dest, nir_variable *src)
 446 {
 447    nir_intrinsic_instr *copy =
 448       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_var);
 449    copy->variables[0] = nir_deref_var_create(copy, dest);
 450    copy->variables[1] = nir_deref_var_create(copy, src);
 451    nir_builder_instr_insert(build, &copy->instr);
 452 }
 453
 454 static inline nir_ssa_def *
 455 nir_load_system_value(nir_builder *build, nir_intrinsic_op op, int index)
 456 {
 457    nir_intrinsic_instr *load = nir_intrinsic_instr_create(build->shader, op);
 458    load->num_components = nir_intrinsic_infos[op].dest_components;
 459    load->const_index[0] = index;
 460    nir_ssa_dest_init(&load->instr, &load->dest,
 461                      nir_intrinsic_infos[op].dest_components, 32, NULL);
 462    nir_builder_instr_insert(build, &load->instr);
 463    return &load->dest.ssa;
 464 }
 465
 466 static inline void
 467 nir_jump(nir_builder *build, nir_jump_type jump_type)
 468 {
 469    nir_jump_instr *jump = nir_jump_instr_create(build->shader, jump_type);
 470    nir_builder_instr_insert(build, &jump->instr);
 471 }
 472
 473 #endif /* NIR_BUILDER_H */