src/gallium/drivers/vc4/vc4_qir.h

   1 /*
   2  * Copyright © 2014 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 VC4_QIR_H
  25 #define VC4_QIR_H
  26
  27 #include <assert.h>
  28 #include <stdio.h>
  29 #include <stdlib.h>
  30 #include <stdbool.h>
  31 #include <stdint.h>
  32 #include <string.h>
  33
  34 #include "util/macros.h"
  35 #include "compiler/nir/nir.h"
  36 #include "util/list.h"
  37 #include "util/u_math.h"
  38
  39 #include "vc4_screen.h"
  40 #include "vc4_qpu_defines.h"
  41 #include "vc4_qpu.h"
  42 #include "kernel/vc4_packet.h"
  43 #include "pipe/p_state.h"
  44
  45 struct nir_builder;
  46
  47 enum qfile {
  48         QFILE_NULL,
  49         QFILE_TEMP,
  50         QFILE_VARY,
  51         QFILE_UNIF,
  52         QFILE_VPM,
  53         QFILE_TLB_COLOR_WRITE,
  54         QFILE_TLB_COLOR_WRITE_MS,
  55         QFILE_TLB_Z_WRITE,
  56         QFILE_TLB_STENCIL_SETUP,
  57
  58         /* Payload registers that aren't in the physical register file, so we
  59          * can just use the corresponding qpu_reg at qpu_emit time.
  60          */
  61         QFILE_FRAG_X,
  62         QFILE_FRAG_Y,
  63         QFILE_FRAG_REV_FLAG,
  64
  65         /**
  66          * Stores an immediate value in the index field that will be used
  67          * directly by qpu_load_imm().
  68          */
  69         QFILE_LOAD_IMM,
  70
  71         /**
  72          * Stores an immediate value in the index field that can be turned
  73          * into a small immediate field by qpu_encode_small_immediate().
  74          */
  75         QFILE_SMALL_IMM,
  76 };
  77
  78 struct qreg {
  79         enum qfile file;
  80         uint32_t index;
  81         int pack;
  82 };
  83
  84 static inline struct qreg qir_reg(enum qfile file, uint32_t index)
  85 {
  86         return (struct qreg){file, index};
  87 }
  88
  89 enum qop {
  90         QOP_UNDEF,
  91         QOP_MOV,
  92         QOP_FMOV,
  93         QOP_MMOV,
  94         QOP_FADD,
  95         QOP_FSUB,
  96         QOP_FMUL,
  97         QOP_V8MULD,
  98         QOP_V8MIN,
  99         QOP_V8MAX,
 100         QOP_V8ADDS,
 101         QOP_V8SUBS,
 102         QOP_MUL24,
 103         QOP_FMIN,
 104         QOP_FMAX,
 105         QOP_FMINABS,
 106         QOP_FMAXABS,
 107         QOP_ADD,
 108         QOP_SUB,
 109         QOP_SHL,
 110         QOP_SHR,
 111         QOP_ASR,
 112         QOP_MIN,
 113         QOP_MAX,
 114         QOP_AND,
 115         QOP_OR,
 116         QOP_XOR,
 117         QOP_NOT,
 118
 119         QOP_FTOI,
 120         QOP_ITOF,
 121         QOP_RCP,
 122         QOP_RSQ,
 123         QOP_EXP2,
 124         QOP_LOG2,
 125         QOP_VW_SETUP,
 126         QOP_VR_SETUP,
 127         QOP_TLB_COLOR_READ,
 128         QOP_MS_MASK,
 129         QOP_VARY_ADD_C,
 130
 131         QOP_FRAG_Z,
 132         QOP_FRAG_W,
 133
 134         /** Texture x coordinate parameter write */
 135         QOP_TEX_S,
 136         /** Texture y coordinate parameter write */
 137         QOP_TEX_T,
 138         /** Texture border color parameter or cube map z coordinate write */
 139         QOP_TEX_R,
 140         /** Texture LOD bias parameter write */
 141         QOP_TEX_B,
 142
 143         /**
 144          * Texture-unit 4-byte read with address provided direct in S
 145          * cooordinate.
 146          *
 147          * The first operand is the offset from the start of the UBO, and the
 148          * second is the uniform that has the UBO's base pointer.
 149          */
 150         QOP_TEX_DIRECT,
 151
 152         /**
 153          * Signal of texture read being necessary and then reading r4 into
 154          * the destination
 155          */
 156         QOP_TEX_RESULT,
 157
 158         QOP_LOAD_IMM,
 159
 160         /* Jumps to block->successor[0] if the qinst->cond (as a
 161          * QPU_COND_BRANCH_*) passes, or block->successor[1] if not.  Note
 162          * that block->successor[1] may be unset if the condition is ALWAYS.
 163          */
 164         QOP_BRANCH,
 165
 166         /* Emits an ADD from src[0] to src[1], where src[0] must be a
 167          * QOP_LOAD_IMM result and src[1] is a QUNIFORM_UNIFORMS_ADDRESS,
 168          * required by the kernel as part of its branch validation.
 169          */
 170         QOP_UNIFORMS_RESET,
 171 };
 172
 173 struct queued_qpu_inst {
 174         struct list_head link;
 175         uint64_t inst;
 176 };
 177
 178 struct qinst {
 179         struct list_head link;
 180
 181         enum qop op;
 182         struct qreg dst;
 183         struct qreg *src;
 184         bool sf;
 185         uint8_t cond;
 186 };
 187
 188 enum qstage {
 189         /**
 190          * Coordinate shader, runs during binning, before the VS, and just
 191          * outputs position.
 192          */
 193         QSTAGE_COORD,
 194         QSTAGE_VERT,
 195         QSTAGE_FRAG,
 196 };
 197
 198 enum quniform_contents {
 199         /**
 200          * Indicates that a constant 32-bit value is copied from the program's
 201          * uniform contents.
 202          */
 203         QUNIFORM_CONSTANT,
 204         /**
 205          * Indicates that the program's uniform contents are used as an index
 206          * into the GL uniform storage.
 207          */
 208         QUNIFORM_UNIFORM,
 209
 210         /** @{
 211          * Scaling factors from clip coordinates to relative to the viewport
 212          * center.
 213          *
 214          * This is used by the coordinate and vertex shaders to produce the
 215          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 216          * point offsets from the viewport ccenter.
 217          */
 218         QUNIFORM_VIEWPORT_X_SCALE,
 219         QUNIFORM_VIEWPORT_Y_SCALE,
 220         /** @} */
 221
 222         QUNIFORM_VIEWPORT_Z_OFFSET,
 223         QUNIFORM_VIEWPORT_Z_SCALE,
 224
 225         QUNIFORM_USER_CLIP_PLANE,
 226
 227         /**
 228          * A reference to a texture config parameter 0 uniform.
 229          *
 230          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 231          * defines texture type, miplevels, and such.  It will be found as a
 232          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 233          */
 234         QUNIFORM_TEXTURE_CONFIG_P0,
 235
 236         /**
 237          * A reference to a texture config parameter 1 uniform.
 238          *
 239          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 240          * defines texture width, height, filters, and wrap modes.  It will be
 241          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 242          * sequence.
 243          */
 244         QUNIFORM_TEXTURE_CONFIG_P1,
 245
 246         /** A reference to a texture config parameter 2 cubemap stride uniform */
 247         QUNIFORM_TEXTURE_CONFIG_P2,
 248
 249         QUNIFORM_TEXTURE_MSAA_ADDR,
 250
 251         QUNIFORM_UBO_ADDR,
 252
 253         QUNIFORM_TEXRECT_SCALE_X,
 254         QUNIFORM_TEXRECT_SCALE_Y,
 255
 256         QUNIFORM_TEXTURE_BORDER_COLOR,
 257
 258         QUNIFORM_BLEND_CONST_COLOR_X,
 259         QUNIFORM_BLEND_CONST_COLOR_Y,
 260         QUNIFORM_BLEND_CONST_COLOR_Z,
 261         QUNIFORM_BLEND_CONST_COLOR_W,
 262         QUNIFORM_BLEND_CONST_COLOR_RGBA,
 263         QUNIFORM_BLEND_CONST_COLOR_AAAA,
 264
 265         QUNIFORM_STENCIL,
 266
 267         QUNIFORM_ALPHA_REF,
 268         QUNIFORM_SAMPLE_MASK,
 269
 270         /* Placeholder uniform that will be updated by the kernel when used by
 271          * an instruction writing to QPU_W_UNIFORMS_ADDRESS.
 272          */
 273         QUNIFORM_UNIFORMS_ADDRESS,
 274 };
 275
 276 struct vc4_varying_slot {
 277         uint8_t slot;
 278         uint8_t swizzle;
 279 };
 280
 281 struct vc4_compiler_ubo_range {
 282         /**
 283          * offset in bytes from the start of the ubo where this range is
 284          * uploaded.
 285          *
 286          * Only set once used is set.
 287          */
 288         uint32_t dst_offset;
 289
 290         /**
 291          * offset in bytes from the start of the gallium uniforms where the
 292          * data comes from.
 293          */
 294         uint32_t src_offset;
 295
 296         /** size in bytes of this ubo range */
 297         uint32_t size;
 298
 299         /**
 300          * Set if this range is used by the shader for indirect uniforms
 301          * access.
 302          */
 303         bool used;
 304 };
 305
 306 struct vc4_key {
 307         struct vc4_uncompiled_shader *shader_state;
 308         struct {
 309                 enum pipe_format format;
 310                 uint8_t swizzle[4];
 311                 union {
 312                         struct {
 313                                 unsigned compare_mode:1;
 314                                 unsigned compare_func:3;
 315                                 unsigned wrap_s:3;
 316                                 unsigned wrap_t:3;
 317                                 unsigned forced_first_level:8;
 318                         };
 319                         struct {
 320                                 uint16_t msaa_width, msaa_height;
 321                         };
 322                 };
 323         } tex[VC4_MAX_TEXTURE_SAMPLERS];
 324         uint8_t ucp_enables;
 325 };
 326
 327 struct vc4_fs_key {
 328         struct vc4_key base;
 329         enum pipe_format color_format;
 330         bool depth_enabled;
 331         bool stencil_enabled;
 332         bool stencil_twoside;
 333         bool stencil_full_writemasks;
 334         bool is_points;
 335         bool is_lines;
 336         bool alpha_test;
 337         bool point_coord_upper_left;
 338         bool light_twoside;
 339         bool msaa;
 340         bool sample_coverage;
 341         bool sample_alpha_to_coverage;
 342         bool sample_alpha_to_one;
 343         uint8_t alpha_test_func;
 344         uint8_t logicop_func;
 345         uint32_t point_sprite_mask;
 346
 347         struct pipe_rt_blend_state blend;
 348 };
 349
 350 struct vc4_vs_key {
 351         struct vc4_key base;
 352
 353         /**
 354          * This is a proxy for the array of FS input semantics, which is
 355          * larger than we would want to put in the key.
 356          */
 357         uint64_t compiled_fs_id;
 358
 359         enum pipe_format attr_formats[8];
 360         bool is_coord;
 361         bool per_vertex_point_size;
 362         bool clamp_color;
 363 };
 364
 365 /** A basic block of QIR intructions. */
 366 struct qblock {
 367         struct list_head link;
 368
 369         struct list_head instructions;
 370         struct list_head qpu_inst_list;
 371
 372         struct set *predecessors;
 373         struct qblock *successors[2];
 374
 375         int index;
 376
 377         /* Instruction IPs for the first and last instruction of the block.
 378          * Set by vc4_qpu_schedule.c.
 379          */
 380         uint32_t start_qpu_ip;
 381         uint32_t end_qpu_ip;
 382
 383         /* Instruction IP for the branch instruction of the block.  Set by
 384          * vc4_qpu_schedule.c.
 385          */
 386         uint32_t branch_qpu_ip;
 387
 388         /** @{ used by vc4_qir_live_variables.c */
 389         BITSET_WORD *def;
 390         BITSET_WORD *use;
 391         BITSET_WORD *live_in;
 392         BITSET_WORD *live_out;
 393         int start_ip, end_ip;
 394         /** @} */
 395 };
 396
 397 struct vc4_compile {
 398         struct vc4_context *vc4;
 399         nir_shader *s;
 400         nir_function_impl *impl;
 401         struct exec_list *cf_node_list;
 402
 403         /**
 404          * Mapping from nir_register * or nir_ssa_def * to array of struct
 405          * qreg for the values.
 406          */
 407         struct hash_table *def_ht;
 408
 409         /* For each temp, the instruction generating its value. */
 410         struct qinst **defs;
 411         uint32_t defs_array_size;
 412
 413         /**
 414          * Inputs to the shader, arranged by TGSI declaration order.
 415          *
 416          * Not all fragment shader QFILE_VARY reads are present in this array.
 417          */
 418         struct qreg *inputs;
 419         struct qreg *outputs;
 420         bool msaa_per_sample_output;
 421         struct qreg color_reads[VC4_MAX_SAMPLES];
 422         struct qreg sample_colors[VC4_MAX_SAMPLES];
 423         uint32_t inputs_array_size;
 424         uint32_t outputs_array_size;
 425         uint32_t uniforms_array_size;
 426
 427         struct vc4_compiler_ubo_range *ubo_ranges;
 428         uint32_t ubo_ranges_array_size;
 429         /** Number of uniform areas declared in ubo_ranges. */
 430         uint32_t num_uniform_ranges;
 431         /** Number of uniform areas used for indirect addressed loads. */
 432         uint32_t num_ubo_ranges;
 433         uint32_t next_ubo_dst_offset;
 434
 435         /* State for whether we're executing on each channel currently.  0 if
 436          * yes, otherwise a block number + 1 that the channel jumped to.
 437          */
 438         struct qreg execute;
 439
 440         struct qreg line_x, point_x, point_y;
 441         struct qreg discard;
 442         struct qreg payload_FRAG_Z;
 443         struct qreg payload_FRAG_W;
 444
 445         uint8_t vattr_sizes[8];
 446
 447         /**
 448          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 449          *
 450          * This includes those that aren't part of the VPM varyings, like
 451          * point/line coordinates.
 452          */
 453         struct vc4_varying_slot *input_slots;
 454         uint32_t num_input_slots;
 455         uint32_t input_slots_array_size;
 456
 457         /**
 458          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 459          * of the output is.  Used to emit from the VS in the order that the
 460          * FS needs.
 461          */
 462         struct vc4_varying_slot *output_slots;
 463
 464         struct pipe_shader_state *shader_state;
 465         struct vc4_key *key;
 466         struct vc4_fs_key *fs_key;
 467         struct vc4_vs_key *vs_key;
 468
 469         /* Live ranges of temps. */
 470         int *temp_start, *temp_end;
 471
 472         uint32_t *uniform_data;
 473         enum quniform_contents *uniform_contents;
 474         uint32_t uniform_array_size;
 475         uint32_t num_uniforms;
 476         uint32_t num_outputs;
 477         uint32_t num_texture_samples;
 478         uint32_t output_position_index;
 479         uint32_t output_color_index;
 480         uint32_t output_point_size_index;
 481         uint32_t output_sample_mask_index;
 482
 483         struct qreg undef;
 484         enum qstage stage;
 485         uint32_t num_temps;
 486
 487         struct list_head blocks;
 488         int next_block_index;
 489         struct qblock *cur_block;
 490         struct qblock *loop_cont_block;
 491         struct qblock *loop_break_block;
 492
 493         struct list_head qpu_inst_list;
 494
 495         uint64_t *qpu_insts;
 496         uint32_t qpu_inst_count;
 497         uint32_t qpu_inst_size;
 498         uint32_t num_inputs;
 499
 500         uint32_t program_id;
 501         uint32_t variant_id;
 502 };
 503
 504 /* Special nir_load_input intrinsic index for loading the current TLB
 505  * destination color.
 506  */
 507 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 508
 509 #define VC4_NIR_MS_MASK_OUTPUT                  2000000000
 510
 511 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 512  * state-dependent value.
 513  */
 514 #define VC4_NIR_STATE_UNIFORM_OFFSET            1000000000
 515
 516 struct vc4_compile *qir_compile_init(void);
 517 void qir_compile_destroy(struct vc4_compile *c);
 518 struct qblock *qir_new_block(struct vc4_compile *c);
 519 void qir_set_emit_block(struct vc4_compile *c, struct qblock *block);
 520 void qir_link_blocks(struct qblock *predecessor, struct qblock *successor);
 521 struct qblock *qir_entry_block(struct vc4_compile *c);
 522 struct qblock *qir_exit_block(struct vc4_compile *c);
 523 struct qinst *qir_inst(enum qop op, struct qreg dst,
 524                        struct qreg src0, struct qreg src1);
 525 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 526                         struct qreg a,
 527                         struct qreg b,
 528                         struct qreg c,
 529                         struct qreg d);
 530 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 531 struct qreg qir_uniform(struct vc4_compile *c,
 532                         enum quniform_contents contents,
 533                         uint32_t data);
 534 void qir_schedule_instructions(struct vc4_compile *c);
 535 void qir_reorder_uniforms(struct vc4_compile *c);
 536 void qir_emit_uniform_stream_resets(struct vc4_compile *c);
 537
 538 struct qreg qir_emit_def(struct vc4_compile *c, struct qinst *inst);
 539 struct qinst *qir_emit_nondef(struct vc4_compile *c, struct qinst *inst);
 540
 541 struct qreg qir_get_temp(struct vc4_compile *c);
 542 void qir_calculate_live_intervals(struct vc4_compile *c);
 543 int qir_get_op_nsrc(enum qop qop);
 544 bool qir_reg_equals(struct qreg a, struct qreg b);
 545 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 546 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 547 bool qir_is_mul(struct qinst *inst);
 548 bool qir_is_raw_mov(struct qinst *inst);
 549 bool qir_is_tex(struct qinst *inst);
 550 bool qir_is_float_input(struct qinst *inst);
 551 bool qir_depends_on_flags(struct qinst *inst);
 552 bool qir_writes_r4(struct qinst *inst);
 553 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 554 uint8_t qir_channels_written(struct qinst *inst);
 555
 556 void qir_dump(struct vc4_compile *c);
 557 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 558 const char *qir_get_stage_name(enum qstage stage);
 559
 560 void qir_validate(struct vc4_compile *c);
 561
 562 void qir_optimize(struct vc4_compile *c);
 563 bool qir_opt_algebraic(struct vc4_compile *c);
 564 bool qir_opt_constant_folding(struct vc4_compile *c);
 565 bool qir_opt_copy_propagation(struct vc4_compile *c);
 566 bool qir_opt_dead_code(struct vc4_compile *c);
 567 bool qir_opt_peephole_sf(struct vc4_compile *c);
 568 bool qir_opt_small_immediates(struct vc4_compile *c);
 569 bool qir_opt_vpm(struct vc4_compile *c);
 570 void vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c);
 571 void vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c);
 572 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 573                                        enum quniform_contents contents);
 574 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 575                                           nir_ssa_def **srcs, int swiz);
 576 void vc4_nir_lower_txf_ms(nir_shader *s, struct vc4_compile *c);
 577 void qir_lower_uniforms(struct vc4_compile *c);
 578
 579 uint32_t qpu_schedule_instructions(struct vc4_compile *c);
 580
 581 void qir_SF(struct vc4_compile *c, struct qreg src);
 582
 583 static inline struct qreg
 584 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 585 {
 586         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 587 }
 588
 589 static inline struct qreg
 590 qir_uniform_f(struct vc4_compile *c, float f)
 591 {
 592         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 593 }
 594
 595 #define QIR_ALU0(name)                                                   \
 596 static inline struct qreg                                                \
 597 qir_##name(struct vc4_compile *c)                                        \
 598 {                                                                        \
 599         return qir_emit_def(c, qir_inst(QOP_##name, c->undef,            \
 600                                         c->undef, c->undef));            \
 601 }                                                                        \
 602 static inline struct qinst *                                             \
 603 qir_##name##_dest(struct vc4_compile *c, struct qreg dest)               \
 604 {                                                                        \
 605         return qir_emit_nondef(c, qir_inst(QOP_##name, dest,             \
 606                                            c->undef, c->undef));         \
 607 }
 608
 609 #define QIR_ALU1(name)                                                   \
 610 static inline struct qreg                                                \
 611 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 612 {                                                                        \
 613         return qir_emit_def(c, qir_inst(QOP_##name, c->undef,            \
 614                                         a, c->undef));                   \
 615 }                                                                        \
 616 static inline struct qinst *                                             \
 617 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 618                   struct qreg a)                                         \
 619 {                                                                        \
 620         return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a,          \
 621                                            c->undef));                   \
 622 }
 623
 624 #define QIR_ALU2(name)                                                   \
 625 static inline struct qreg                                                \
 626 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 627 {                                                                        \
 628         return qir_emit_def(c, qir_inst(QOP_##name, c->undef, a, b));    \
 629 }                                                                        \
 630 static inline struct qinst *                                             \
 631 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 632                   struct qreg a, struct qreg b)                          \
 633 {                                                                        \
 634         return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a, b));     \
 635 }
 636
 637 #define QIR_NODST_1(name)                                               \
 638 static inline struct qinst *                                            \
 639 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 640 {                                                                       \
 641         return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef,        \
 642                                            a, c->undef));               \
 643 }
 644
 645 #define QIR_NODST_2(name)                                               \
 646 static inline struct qinst *                                            \
 647 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 648 {                                                                       \
 649         return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef,        \
 650                                            a, b));                      \
 651 }
 652
 653 #define QIR_PAYLOAD(name)                                                \
 654 static inline struct qreg                                                \
 655 qir_##name(struct vc4_compile *c)                                        \
 656 {                                                                        \
 657         struct qreg *payload = &c->payload_##name;                       \
 658         if (payload->file != QFILE_NULL)                                 \
 659                 return *payload;                                         \
 660         *payload = qir_get_temp(c);                                      \
 661         struct qinst *inst = qir_inst(QOP_##name, *payload,              \
 662                                       c->undef, c->undef);               \
 663         struct qblock *entry = qir_entry_block(c);                       \
 664         list_add(&inst->link, &entry->instructions);                     \
 665         c->defs[payload->index] = inst;                                  \
 666         return *payload;                                                 \
 667 }
 668
 669 QIR_ALU1(MOV)
 670 QIR_ALU1(FMOV)
 671 QIR_ALU1(MMOV)
 672 QIR_ALU2(FADD)
 673 QIR_ALU2(FSUB)
 674 QIR_ALU2(FMUL)
 675 QIR_ALU2(V8MULD)
 676 QIR_ALU2(V8MIN)
 677 QIR_ALU2(V8MAX)
 678 QIR_ALU2(V8ADDS)
 679 QIR_ALU2(V8SUBS)
 680 QIR_ALU2(MUL24)
 681 QIR_ALU2(FMIN)
 682 QIR_ALU2(FMAX)
 683 QIR_ALU2(FMINABS)
 684 QIR_ALU2(FMAXABS)
 685 QIR_ALU1(FTOI)
 686 QIR_ALU1(ITOF)
 687
 688 QIR_ALU2(ADD)
 689 QIR_ALU2(SUB)
 690 QIR_ALU2(SHL)
 691 QIR_ALU2(SHR)
 692 QIR_ALU2(ASR)
 693 QIR_ALU2(MIN)
 694 QIR_ALU2(MAX)
 695 QIR_ALU2(AND)
 696 QIR_ALU2(OR)
 697 QIR_ALU2(XOR)
 698 QIR_ALU1(NOT)
 699
 700 QIR_ALU1(RCP)
 701 QIR_ALU1(RSQ)
 702 QIR_ALU1(EXP2)
 703 QIR_ALU1(LOG2)
 704 QIR_ALU1(VARY_ADD_C)
 705 QIR_NODST_2(TEX_S)
 706 QIR_NODST_2(TEX_T)
 707 QIR_NODST_2(TEX_R)
 708 QIR_NODST_2(TEX_B)
 709 QIR_NODST_2(TEX_DIRECT)
 710 QIR_PAYLOAD(FRAG_Z)
 711 QIR_PAYLOAD(FRAG_W)
 712 QIR_ALU0(TEX_RESULT)
 713 QIR_ALU0(TLB_COLOR_READ)
 714 QIR_NODST_1(MS_MASK)
 715
 716 static inline struct qreg
 717 qir_SEL(struct vc4_compile *c, uint8_t cond, struct qreg src0, struct qreg src1)
 718 {
 719         struct qreg t = qir_get_temp(c);
 720         struct qinst *a = qir_MOV_dest(c, t, src0);
 721         struct qinst *b = qir_MOV_dest(c, t, src1);
 722         a->cond = cond;
 723         b->cond = qpu_cond_complement(cond);
 724         return t;
 725 }
 726
 727 static inline struct qreg
 728 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 729 {
 730         struct qreg t = qir_FMOV(c, src);
 731         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 732         return t;
 733 }
 734
 735 static inline struct qreg
 736 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 737 {
 738         struct qreg t = qir_MOV(c, src);
 739         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 740         return t;
 741 }
 742
 743 static inline struct qreg
 744 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 745 {
 746         struct qreg t = qir_FMOV(c, src);
 747         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 748         return t;
 749 }
 750
 751 static inline struct qreg
 752 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 753 {
 754         struct qreg t = qir_MOV(c, src);
 755         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 756         return t;
 757 }
 758
 759 static inline void
 760 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 761 {
 762         assert(!dest.pack);
 763         dest.pack = QPU_PACK_MUL_8A + chan;
 764         qir_emit_nondef(c, qir_inst(QOP_MMOV, dest, val, c->undef));
 765 }
 766
 767 static inline struct qreg
 768 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
 769 {
 770         struct qreg dest = qir_MMOV(c, val);
 771         c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
 772         return dest;
 773 }
 774
 775 static inline struct qreg
 776 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 777 {
 778         return qir_EXP2(c, qir_FMUL(c,
 779                                     y,
 780                                     qir_LOG2(c, x)));
 781 }
 782
 783 static inline void
 784 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 785 {
 786         qir_MOV_dest(c, qir_reg(QFILE_VPM, 0), val);
 787 }
 788
 789 static inline struct qreg
 790 qir_LOAD_IMM(struct vc4_compile *c, uint32_t val)
 791 {
 792         return qir_emit_def(c, qir_inst(QOP_LOAD_IMM, c->undef,
 793                                         qir_reg(QFILE_LOAD_IMM, val), c->undef));
 794 }
 795
 796 static inline void
 797 qir_MOV_cond(struct vc4_compile *c, uint8_t cond,
 798              struct qreg dest, struct qreg src)
 799 {
 800         qir_MOV_dest(c, dest, src)->cond = cond;
 801 }
 802
 803 static inline struct qinst *
 804 qir_BRANCH(struct vc4_compile *c, uint8_t cond)
 805 {
 806         struct qinst *inst = qir_inst(QOP_BRANCH, c->undef, c->undef, c->undef);
 807         inst->cond = cond;
 808         qir_emit_nondef(c, inst);
 809         return inst;
 810 }
 811
 812 #define qir_for_each_block(block, c)                                    \
 813         list_for_each_entry(struct qblock, block, &c->blocks, link)
 814
 815 #define qir_for_each_block_rev(block, c)                                \
 816         list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
 817
 818 /* Loop over the non-NULL members of the successors array. */
 819 #define qir_for_each_successor(succ, block)                             \
 820         for (struct qblock *succ = block->successors[0];                \
 821              succ != NULL;                                              \
 822              succ = (succ == block->successors[1] ? NULL :              \
 823                      block->successors[1]))
 824
 825 #define qir_for_each_inst(inst, block)                                  \
 826         list_for_each_entry(struct qinst, inst, &block->instructions, link)
 827
 828 #define qir_for_each_inst_rev(inst, block)                                  \
 829         list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
 830
 831 #define qir_for_each_inst_safe(inst, block)                             \
 832         list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
 833
 834 #define qir_for_each_inst_inorder(inst, c)                              \
 835         qir_for_each_block(_block, c)                                   \
 836                 qir_for_each_inst(inst, _block)
 837
 838 #endif /* VC4_QIR_H */