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
 167 struct queued_qpu_inst {
 168         struct list_head link;
 169         uint64_t inst;
 170 };
 171
 172 struct qinst {
 173         struct list_head link;
 174
 175         enum qop op;
 176         struct qreg dst;
 177         struct qreg *src;
 178         bool sf;
 179         uint8_t cond;
 180 };
 181
 182 enum qstage {
 183         /**
 184          * Coordinate shader, runs during binning, before the VS, and just
 185          * outputs position.
 186          */
 187         QSTAGE_COORD,
 188         QSTAGE_VERT,
 189         QSTAGE_FRAG,
 190 };
 191
 192 enum quniform_contents {
 193         /**
 194          * Indicates that a constant 32-bit value is copied from the program's
 195          * uniform contents.
 196          */
 197         QUNIFORM_CONSTANT,
 198         /**
 199          * Indicates that the program's uniform contents are used as an index
 200          * into the GL uniform storage.
 201          */
 202         QUNIFORM_UNIFORM,
 203
 204         /** @{
 205          * Scaling factors from clip coordinates to relative to the viewport
 206          * center.
 207          *
 208          * This is used by the coordinate and vertex shaders to produce the
 209          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 210          * point offsets from the viewport ccenter.
 211          */
 212         QUNIFORM_VIEWPORT_X_SCALE,
 213         QUNIFORM_VIEWPORT_Y_SCALE,
 214         /** @} */
 215
 216         QUNIFORM_VIEWPORT_Z_OFFSET,
 217         QUNIFORM_VIEWPORT_Z_SCALE,
 218
 219         QUNIFORM_USER_CLIP_PLANE,
 220
 221         /**
 222          * A reference to a texture config parameter 0 uniform.
 223          *
 224          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 225          * defines texture type, miplevels, and such.  It will be found as a
 226          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 227          */
 228         QUNIFORM_TEXTURE_CONFIG_P0,
 229
 230         /**
 231          * A reference to a texture config parameter 1 uniform.
 232          *
 233          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 234          * defines texture width, height, filters, and wrap modes.  It will be
 235          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 236          * sequence.
 237          */
 238         QUNIFORM_TEXTURE_CONFIG_P1,
 239
 240         /** A reference to a texture config parameter 2 cubemap stride uniform */
 241         QUNIFORM_TEXTURE_CONFIG_P2,
 242
 243         QUNIFORM_TEXTURE_MSAA_ADDR,
 244
 245         QUNIFORM_UBO_ADDR,
 246
 247         QUNIFORM_TEXRECT_SCALE_X,
 248         QUNIFORM_TEXRECT_SCALE_Y,
 249
 250         QUNIFORM_TEXTURE_BORDER_COLOR,
 251
 252         QUNIFORM_BLEND_CONST_COLOR_X,
 253         QUNIFORM_BLEND_CONST_COLOR_Y,
 254         QUNIFORM_BLEND_CONST_COLOR_Z,
 255         QUNIFORM_BLEND_CONST_COLOR_W,
 256         QUNIFORM_BLEND_CONST_COLOR_RGBA,
 257         QUNIFORM_BLEND_CONST_COLOR_AAAA,
 258
 259         QUNIFORM_STENCIL,
 260
 261         QUNIFORM_ALPHA_REF,
 262         QUNIFORM_SAMPLE_MASK,
 263 };
 264
 265 struct vc4_varying_slot {
 266         uint8_t slot;
 267         uint8_t swizzle;
 268 };
 269
 270 struct vc4_compiler_ubo_range {
 271         /**
 272          * offset in bytes from the start of the ubo where this range is
 273          * uploaded.
 274          *
 275          * Only set once used is set.
 276          */
 277         uint32_t dst_offset;
 278
 279         /**
 280          * offset in bytes from the start of the gallium uniforms where the
 281          * data comes from.
 282          */
 283         uint32_t src_offset;
 284
 285         /** size in bytes of this ubo range */
 286         uint32_t size;
 287
 288         /**
 289          * Set if this range is used by the shader for indirect uniforms
 290          * access.
 291          */
 292         bool used;
 293 };
 294
 295 struct vc4_key {
 296         struct vc4_uncompiled_shader *shader_state;
 297         struct {
 298                 enum pipe_format format;
 299                 uint8_t swizzle[4];
 300                 union {
 301                         struct {
 302                                 unsigned compare_mode:1;
 303                                 unsigned compare_func:3;
 304                                 unsigned wrap_s:3;
 305                                 unsigned wrap_t:3;
 306                         };
 307                         struct {
 308                                 uint16_t msaa_width, msaa_height;
 309                         };
 310                 };
 311         } tex[VC4_MAX_TEXTURE_SAMPLERS];
 312         uint8_t ucp_enables;
 313 };
 314
 315 struct vc4_fs_key {
 316         struct vc4_key base;
 317         enum pipe_format color_format;
 318         bool depth_enabled;
 319         bool stencil_enabled;
 320         bool stencil_twoside;
 321         bool stencil_full_writemasks;
 322         bool is_points;
 323         bool is_lines;
 324         bool alpha_test;
 325         bool point_coord_upper_left;
 326         bool light_twoside;
 327         bool msaa;
 328         bool sample_coverage;
 329         bool sample_alpha_to_coverage;
 330         bool sample_alpha_to_one;
 331         uint8_t alpha_test_func;
 332         uint8_t logicop_func;
 333         uint32_t point_sprite_mask;
 334
 335         struct pipe_rt_blend_state blend;
 336 };
 337
 338 struct vc4_vs_key {
 339         struct vc4_key base;
 340
 341         /**
 342          * This is a proxy for the array of FS input semantics, which is
 343          * larger than we would want to put in the key.
 344          */
 345         uint64_t compiled_fs_id;
 346
 347         enum pipe_format attr_formats[8];
 348         bool is_coord;
 349         bool per_vertex_point_size;
 350         bool clamp_color;
 351 };
 352
 353 /** A basic block of QIR intructions. */
 354 struct qblock {
 355         struct list_head link;
 356
 357         struct list_head instructions;
 358         struct list_head qpu_inst_list;
 359
 360         struct set *predecessors;
 361         struct qblock *successors[2];
 362
 363         int index;
 364
 365         /** @{ used by vc4_qir_live_variables.c */
 366         BITSET_WORD *def;
 367         BITSET_WORD *use;
 368         BITSET_WORD *live_in;
 369         BITSET_WORD *live_out;
 370         int start_ip, end_ip;
 371         /** @} */
 372 };
 373
 374 struct vc4_compile {
 375         struct vc4_context *vc4;
 376         nir_shader *s;
 377         nir_function_impl *impl;
 378         struct exec_list *cf_node_list;
 379
 380         /**
 381          * Mapping from nir_register * or nir_ssa_def * to array of struct
 382          * qreg for the values.
 383          */
 384         struct hash_table *def_ht;
 385
 386         /* For each temp, the instruction generating its value. */
 387         struct qinst **defs;
 388         uint32_t defs_array_size;
 389
 390         /**
 391          * Inputs to the shader, arranged by TGSI declaration order.
 392          *
 393          * Not all fragment shader QFILE_VARY reads are present in this array.
 394          */
 395         struct qreg *inputs;
 396         struct qreg *outputs;
 397         bool msaa_per_sample_output;
 398         struct qreg color_reads[VC4_MAX_SAMPLES];
 399         struct qreg sample_colors[VC4_MAX_SAMPLES];
 400         uint32_t inputs_array_size;
 401         uint32_t outputs_array_size;
 402         uint32_t uniforms_array_size;
 403
 404         struct vc4_compiler_ubo_range *ubo_ranges;
 405         uint32_t ubo_ranges_array_size;
 406         /** Number of uniform areas declared in ubo_ranges. */
 407         uint32_t num_uniform_ranges;
 408         /** Number of uniform areas used for indirect addressed loads. */
 409         uint32_t num_ubo_ranges;
 410         uint32_t next_ubo_dst_offset;
 411
 412         /* State for whether we're executing on each channel currently.  0 if
 413          * yes, otherwise a block number + 1 that the channel jumped to.
 414          */
 415         struct qreg execute;
 416
 417         struct qreg line_x, point_x, point_y;
 418         struct qreg discard;
 419         struct qreg payload_FRAG_Z;
 420         struct qreg payload_FRAG_W;
 421
 422         uint8_t vattr_sizes[8];
 423
 424         /**
 425          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 426          *
 427          * This includes those that aren't part of the VPM varyings, like
 428          * point/line coordinates.
 429          */
 430         struct vc4_varying_slot *input_slots;
 431         uint32_t num_input_slots;
 432         uint32_t input_slots_array_size;
 433
 434         /**
 435          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 436          * of the output is.  Used to emit from the VS in the order that the
 437          * FS needs.
 438          */
 439         struct vc4_varying_slot *output_slots;
 440
 441         struct pipe_shader_state *shader_state;
 442         struct vc4_key *key;
 443         struct vc4_fs_key *fs_key;
 444         struct vc4_vs_key *vs_key;
 445
 446         /* Live ranges of temps. */
 447         int *temp_start, *temp_end;
 448
 449         uint32_t *uniform_data;
 450         enum quniform_contents *uniform_contents;
 451         uint32_t uniform_array_size;
 452         uint32_t num_uniforms;
 453         uint32_t num_outputs;
 454         uint32_t num_texture_samples;
 455         uint32_t output_position_index;
 456         uint32_t output_color_index;
 457         uint32_t output_point_size_index;
 458         uint32_t output_sample_mask_index;
 459
 460         struct qreg undef;
 461         enum qstage stage;
 462         uint32_t num_temps;
 463
 464         struct list_head blocks;
 465         int next_block_index;
 466         struct qblock *cur_block;
 467         struct qblock *loop_cont_block;
 468         struct qblock *loop_break_block;
 469
 470         struct list_head qpu_inst_list;
 471
 472         uint64_t *qpu_insts;
 473         uint32_t qpu_inst_count;
 474         uint32_t qpu_inst_size;
 475         uint32_t num_inputs;
 476
 477         uint32_t program_id;
 478         uint32_t variant_id;
 479 };
 480
 481 /* Special nir_load_input intrinsic index for loading the current TLB
 482  * destination color.
 483  */
 484 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 485
 486 #define VC4_NIR_MS_MASK_OUTPUT                  2000000000
 487
 488 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 489  * state-dependent value.
 490  */
 491 #define VC4_NIR_STATE_UNIFORM_OFFSET            1000000000
 492
 493 struct vc4_compile *qir_compile_init(void);
 494 void qir_compile_destroy(struct vc4_compile *c);
 495 struct qblock *qir_new_block(struct vc4_compile *c);
 496 void qir_set_emit_block(struct vc4_compile *c, struct qblock *block);
 497 void qir_link_blocks(struct qblock *predecessor, struct qblock *successor);
 498 struct qblock *qir_entry_block(struct vc4_compile *c);
 499 struct qblock *qir_exit_block(struct vc4_compile *c);
 500 struct qinst *qir_inst(enum qop op, struct qreg dst,
 501                        struct qreg src0, struct qreg src1);
 502 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 503                         struct qreg a,
 504                         struct qreg b,
 505                         struct qreg c,
 506                         struct qreg d);
 507 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 508 struct qreg qir_uniform(struct vc4_compile *c,
 509                         enum quniform_contents contents,
 510                         uint32_t data);
 511 void qir_schedule_instructions(struct vc4_compile *c);
 512 void qir_reorder_uniforms(struct vc4_compile *c);
 513
 514 struct qreg qir_emit_def(struct vc4_compile *c, struct qinst *inst);
 515 struct qinst *qir_emit_nondef(struct vc4_compile *c, struct qinst *inst);
 516
 517 struct qreg qir_get_temp(struct vc4_compile *c);
 518 void qir_calculate_live_intervals(struct vc4_compile *c);
 519 int qir_get_op_nsrc(enum qop qop);
 520 bool qir_reg_equals(struct qreg a, struct qreg b);
 521 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 522 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 523 bool qir_is_mul(struct qinst *inst);
 524 bool qir_is_raw_mov(struct qinst *inst);
 525 bool qir_is_tex(struct qinst *inst);
 526 bool qir_is_float_input(struct qinst *inst);
 527 bool qir_depends_on_flags(struct qinst *inst);
 528 bool qir_writes_r4(struct qinst *inst);
 529 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 530 uint8_t qir_channels_written(struct qinst *inst);
 531
 532 void qir_dump(struct vc4_compile *c);
 533 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 534 const char *qir_get_stage_name(enum qstage stage);
 535
 536 void qir_validate(struct vc4_compile *c);
 537
 538 void qir_optimize(struct vc4_compile *c);
 539 bool qir_opt_algebraic(struct vc4_compile *c);
 540 bool qir_opt_constant_folding(struct vc4_compile *c);
 541 bool qir_opt_copy_propagation(struct vc4_compile *c);
 542 bool qir_opt_dead_code(struct vc4_compile *c);
 543 bool qir_opt_peephole_sf(struct vc4_compile *c);
 544 bool qir_opt_small_immediates(struct vc4_compile *c);
 545 bool qir_opt_vpm(struct vc4_compile *c);
 546 void vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c);
 547 void vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c);
 548 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 549                                        enum quniform_contents contents);
 550 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 551                                           nir_ssa_def **srcs, int swiz);
 552 void vc4_nir_lower_txf_ms(nir_shader *s, struct vc4_compile *c);
 553 void qir_lower_uniforms(struct vc4_compile *c);
 554
 555 uint32_t qpu_schedule_instructions(struct vc4_compile *c);
 556
 557 void qir_SF(struct vc4_compile *c, struct qreg src);
 558
 559 static inline struct qreg
 560 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 561 {
 562         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 563 }
 564
 565 static inline struct qreg
 566 qir_uniform_f(struct vc4_compile *c, float f)
 567 {
 568         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 569 }
 570
 571 #define QIR_ALU0(name)                                                   \
 572 static inline struct qreg                                                \
 573 qir_##name(struct vc4_compile *c)                                        \
 574 {                                                                        \
 575         return qir_emit_def(c, qir_inst(QOP_##name, c->undef,            \
 576                                         c->undef, c->undef));            \
 577 }                                                                        \
 578 static inline struct qinst *                                             \
 579 qir_##name##_dest(struct vc4_compile *c, struct qreg dest)               \
 580 {                                                                        \
 581         return qir_emit_nondef(c, qir_inst(QOP_##name, dest,             \
 582                                            c->undef, c->undef));         \
 583 }
 584
 585 #define QIR_ALU1(name)                                                   \
 586 static inline struct qreg                                                \
 587 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 588 {                                                                        \
 589         return qir_emit_def(c, qir_inst(QOP_##name, c->undef,            \
 590                                         a, c->undef));                   \
 591 }                                                                        \
 592 static inline struct qinst *                                             \
 593 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 594                   struct qreg a)                                         \
 595 {                                                                        \
 596         return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a,          \
 597                                            c->undef));                   \
 598 }
 599
 600 #define QIR_ALU2(name)                                                   \
 601 static inline struct qreg                                                \
 602 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 603 {                                                                        \
 604         return qir_emit_def(c, qir_inst(QOP_##name, c->undef, a, b));    \
 605 }                                                                        \
 606 static inline struct qinst *                                             \
 607 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 608                   struct qreg a, struct qreg b)                          \
 609 {                                                                        \
 610         return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a, b));     \
 611 }
 612
 613 #define QIR_NODST_1(name)                                               \
 614 static inline struct qinst *                                            \
 615 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 616 {                                                                       \
 617         return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef,        \
 618                                            a, c->undef));               \
 619 }
 620
 621 #define QIR_NODST_2(name)                                               \
 622 static inline struct qinst *                                            \
 623 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 624 {                                                                       \
 625         return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef,        \
 626                                            a, b));                      \
 627 }
 628
 629 #define QIR_PAYLOAD(name)                                                \
 630 static inline struct qreg                                                \
 631 qir_##name(struct vc4_compile *c)                                        \
 632 {                                                                        \
 633         struct qreg *payload = &c->payload_##name;                       \
 634         if (payload->file != QFILE_NULL)                                 \
 635                 return *payload;                                         \
 636         *payload = qir_get_temp(c);                                      \
 637         struct qinst *inst = qir_inst(QOP_##name, *payload,              \
 638                                       c->undef, c->undef);               \
 639         struct qblock *entry = qir_entry_block(c);                       \
 640         list_add(&inst->link, &entry->instructions);                     \
 641         c->defs[payload->index] = inst;                                  \
 642         return *payload;                                                 \
 643 }
 644
 645 QIR_ALU1(MOV)
 646 QIR_ALU1(FMOV)
 647 QIR_ALU1(MMOV)
 648 QIR_ALU2(FADD)
 649 QIR_ALU2(FSUB)
 650 QIR_ALU2(FMUL)
 651 QIR_ALU2(V8MULD)
 652 QIR_ALU2(V8MIN)
 653 QIR_ALU2(V8MAX)
 654 QIR_ALU2(V8ADDS)
 655 QIR_ALU2(V8SUBS)
 656 QIR_ALU2(MUL24)
 657 QIR_ALU2(FMIN)
 658 QIR_ALU2(FMAX)
 659 QIR_ALU2(FMINABS)
 660 QIR_ALU2(FMAXABS)
 661 QIR_ALU1(FTOI)
 662 QIR_ALU1(ITOF)
 663
 664 QIR_ALU2(ADD)
 665 QIR_ALU2(SUB)
 666 QIR_ALU2(SHL)
 667 QIR_ALU2(SHR)
 668 QIR_ALU2(ASR)
 669 QIR_ALU2(MIN)
 670 QIR_ALU2(MAX)
 671 QIR_ALU2(AND)
 672 QIR_ALU2(OR)
 673 QIR_ALU2(XOR)
 674 QIR_ALU1(NOT)
 675
 676 QIR_ALU1(RCP)
 677 QIR_ALU1(RSQ)
 678 QIR_ALU1(EXP2)
 679 QIR_ALU1(LOG2)
 680 QIR_ALU1(VARY_ADD_C)
 681 QIR_NODST_2(TEX_S)
 682 QIR_NODST_2(TEX_T)
 683 QIR_NODST_2(TEX_R)
 684 QIR_NODST_2(TEX_B)
 685 QIR_NODST_2(TEX_DIRECT)
 686 QIR_PAYLOAD(FRAG_Z)
 687 QIR_PAYLOAD(FRAG_W)
 688 QIR_ALU0(TEX_RESULT)
 689 QIR_ALU0(TLB_COLOR_READ)
 690 QIR_NODST_1(MS_MASK)
 691
 692 static inline struct qreg
 693 qir_SEL(struct vc4_compile *c, uint8_t cond, struct qreg src0, struct qreg src1)
 694 {
 695         struct qreg t = qir_get_temp(c);
 696         struct qinst *a = qir_MOV_dest(c, t, src0);
 697         struct qinst *b = qir_MOV_dest(c, t, src1);
 698         a->cond = cond;
 699         b->cond = qpu_cond_complement(cond);
 700         return t;
 701 }
 702
 703 static inline struct qreg
 704 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 705 {
 706         struct qreg t = qir_FMOV(c, src);
 707         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 708         return t;
 709 }
 710
 711 static inline struct qreg
 712 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 713 {
 714         struct qreg t = qir_MOV(c, src);
 715         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 716         return t;
 717 }
 718
 719 static inline struct qreg
 720 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 721 {
 722         struct qreg t = qir_FMOV(c, src);
 723         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 724         return t;
 725 }
 726
 727 static inline struct qreg
 728 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 729 {
 730         struct qreg t = qir_MOV(c, src);
 731         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 732         return t;
 733 }
 734
 735 static inline void
 736 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 737 {
 738         assert(!dest.pack);
 739         dest.pack = QPU_PACK_MUL_8A + chan;
 740         qir_emit_nondef(c, qir_inst(QOP_MMOV, dest, val, c->undef));
 741 }
 742
 743 static inline struct qreg
 744 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
 745 {
 746         struct qreg dest = qir_MMOV(c, val);
 747         c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
 748         return dest;
 749 }
 750
 751 static inline struct qreg
 752 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 753 {
 754         return qir_EXP2(c, qir_FMUL(c,
 755                                     y,
 756                                     qir_LOG2(c, x)));
 757 }
 758
 759 static inline void
 760 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 761 {
 762         qir_MOV_dest(c, qir_reg(QFILE_VPM, 0), val);
 763 }
 764
 765 static inline struct qreg
 766 qir_LOAD_IMM(struct vc4_compile *c, uint32_t val)
 767 {
 768         return qir_emit_def(c, qir_inst(QOP_LOAD_IMM, c->undef,
 769                                         qir_reg(QFILE_LOAD_IMM, val), c->undef));
 770 }
 771
 772 static inline void
 773 qir_MOV_cond(struct vc4_compile *c, uint8_t cond,
 774              struct qreg dest, struct qreg src)
 775 {
 776         qir_MOV_dest(c, dest, src)->cond = cond;
 777 }
 778
 779 static inline struct qinst *
 780 qir_BRANCH(struct vc4_compile *c, uint8_t cond)
 781 {
 782         struct qinst *inst = qir_inst(QOP_BRANCH, c->undef, c->undef, c->undef);
 783         inst->cond = cond;
 784         qir_emit_nondef(c, inst);
 785         return inst;
 786 }
 787
 788 #define qir_for_each_block(block, c)                                    \
 789         list_for_each_entry(struct qblock, block, &c->blocks, link)
 790
 791 #define qir_for_each_block_rev(block, c)                                \
 792         list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
 793
 794 /* Loop over the non-NULL members of the successors array. */
 795 #define qir_for_each_successor(succ, block)                             \
 796         for (struct qblock *succ = block->successors[0];                \
 797              succ != NULL;                                              \
 798              succ = (succ == block->successors[1] ? NULL :              \
 799                      block->successors[1]))
 800
 801 #define qir_for_each_inst(inst, block)                                  \
 802         list_for_each_entry(struct qinst, inst, &block->instructions, link)
 803
 804 #define qir_for_each_inst_rev(inst, block)                                  \
 805         list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
 806
 807 #define qir_for_each_inst_safe(inst, block)                             \
 808         list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
 809
 810 #define qir_for_each_inst_inorder(inst, c)                              \
 811         qir_for_each_block(_block, c)                                   \
 812                 qir_for_each_inst(inst, _block)
 813
 814 #endif /* VC4_QIR_H */