src/mesa/drivers/dri/i965/brw_fs.h

   1 /*
   2  * Copyright © 2010 Intel Corporation
   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  * Authors:
  24  *    Eric Anholt <eric@anholt.net>
  25  *
  26  */
  27
  28 #pragma once
  29
  30 #include "brw_shader.h"
  31 #include "brw_ir_fs.h"
  32
  33 extern "C" {
  34
  35 #include <sys/types.h>
  36
  37 #include "main/macros.h"
  38 #include "main/shaderobj.h"
  39 #include "main/uniforms.h"
  40 #include "program/prog_parameter.h"
  41 #include "program/prog_print.h"
  42 #include "program/prog_optimize.h"
  43 #include "util/register_allocate.h"
  44 #include "program/hash_table.h"
  45 #include "brw_context.h"
  46 #include "brw_eu.h"
  47 #include "brw_wm.h"
  48 #include "intel_asm_annotation.h"
  49 }
  50 #include "glsl/glsl_types.h"
  51 #include "glsl/ir.h"
  52 #include "glsl/nir/nir.h"
  53 #include "program/sampler.h"
  54
  55 struct bblock_t;
  56 namespace {
  57    struct acp_entry;
  58 }
  59
  60 namespace brw {
  61    class fs_live_variables;
  62 }
  63
  64 /**
  65  * The fragment shader front-end.
  66  *
  67  * Translates either GLSL IR or Mesa IR (for ARB_fragment_program) into FS IR.
  68  */
  69 class fs_visitor : public backend_visitor
  70 {
  71 public:
  72    const fs_reg reg_null_f;
  73    const fs_reg reg_null_d;
  74    const fs_reg reg_null_ud;
  75
  76    fs_visitor(struct brw_context *brw,
  77               void *mem_ctx,
  78               const struct brw_wm_prog_key *key,
  79               struct brw_wm_prog_data *prog_data,
  80               struct gl_shader_program *shader_prog,
  81               struct gl_fragment_program *fp,
  82               unsigned dispatch_width);
  83
  84    fs_visitor(struct brw_context *brw,
  85               void *mem_ctx,
  86               const struct brw_vs_prog_key *key,
  87               struct brw_vs_prog_data *prog_data,
  88               struct gl_shader_program *shader_prog,
  89               struct gl_vertex_program *cp,
  90               unsigned dispatch_width);
  91
  92    ~fs_visitor();
  93    void init();
  94
  95    fs_reg *variable_storage(ir_variable *var);
  96    fs_reg vgrf(const glsl_type *const type);
  97    fs_reg vgrf(int num_components);
  98    void import_uniforms(fs_visitor *v);
  99    void setup_uniform_clipplane_values();
 100    void compute_clip_distance();
 101
 102    void visit(ir_variable *ir);
 103    void visit(ir_assignment *ir);
 104    void visit(ir_dereference_variable *ir);
 105    void visit(ir_dereference_record *ir);
 106    void visit(ir_dereference_array *ir);
 107    void visit(ir_expression *ir);
 108    void visit(ir_texture *ir);
 109    void visit(ir_if *ir);
 110    void visit(ir_constant *ir);
 111    void visit(ir_swizzle *ir);
 112    void visit(ir_return *ir);
 113    void visit(ir_loop *ir);
 114    void visit(ir_loop_jump *ir);
 115    void visit(ir_discard *ir);
 116    void visit(ir_call *ir);
 117    void visit(ir_function *ir);
 118    void visit(ir_function_signature *ir);
 119    void visit(ir_emit_vertex *);
 120    void visit(ir_end_primitive *);
 121
 122    uint32_t gather_channel(int orig_chan, uint32_t sampler);
 123    void swizzle_result(ir_texture_opcode op, int dest_components,
 124                        fs_reg orig_val, uint32_t sampler);
 125
 126    fs_inst *emit(fs_inst *inst);
 127    void emit(exec_list list);
 128
 129    fs_inst *emit(enum opcode opcode);
 130    fs_inst *emit(enum opcode opcode, const fs_reg &dst);
 131    fs_inst *emit(enum opcode opcode, const fs_reg &dst, const fs_reg &src0);
 132    fs_inst *emit(enum opcode opcode, const fs_reg &dst, const fs_reg &src0,
 133                  const fs_reg &src1);
 134    fs_inst *emit(enum opcode opcode, const fs_reg &dst,
 135                  const fs_reg &src0, const fs_reg &src1, const fs_reg &src2);
 136    fs_inst *emit(enum opcode opcode, const fs_reg &dst,
 137                  fs_reg src[], int sources);
 138
 139    fs_inst *MOV(const fs_reg &dst, const fs_reg &src);
 140    fs_inst *NOT(const fs_reg &dst, const fs_reg &src);
 141    fs_inst *RNDD(const fs_reg &dst, const fs_reg &src);
 142    fs_inst *RNDE(const fs_reg &dst, const fs_reg &src);
 143    fs_inst *RNDZ(const fs_reg &dst, const fs_reg &src);
 144    fs_inst *FRC(const fs_reg &dst, const fs_reg &src);
 145    fs_inst *ADD(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 146    fs_inst *MUL(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 147    fs_inst *MACH(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 148    fs_inst *MAC(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 149    fs_inst *SHL(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 150    fs_inst *SHR(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 151    fs_inst *ASR(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 152    fs_inst *AND(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 153    fs_inst *OR(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 154    fs_inst *XOR(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 155    fs_inst *IF(enum brw_predicate predicate);
 156    fs_inst *IF(const fs_reg &src0, const fs_reg &src1,
 157                enum brw_conditional_mod condition);
 158    fs_inst *CMP(fs_reg dst, fs_reg src0, fs_reg src1,
 159                 enum brw_conditional_mod condition);
 160    fs_inst *LRP(const fs_reg &dst, const fs_reg &a, const fs_reg &y,
 161                 const fs_reg &x);
 162    fs_inst *DEP_RESOLVE_MOV(int grf);
 163    fs_inst *BFREV(const fs_reg &dst, const fs_reg &value);
 164    fs_inst *BFE(const fs_reg &dst, const fs_reg &bits, const fs_reg &offset,
 165                 const fs_reg &value);
 166    fs_inst *BFI1(const fs_reg &dst, const fs_reg &bits, const fs_reg &offset);
 167    fs_inst *BFI2(const fs_reg &dst, const fs_reg &bfi1_dst,
 168                  const fs_reg &insert, const fs_reg &base);
 169    fs_inst *FBH(const fs_reg &dst, const fs_reg &value);
 170    fs_inst *FBL(const fs_reg &dst, const fs_reg &value);
 171    fs_inst *CBIT(const fs_reg &dst, const fs_reg &value);
 172    fs_inst *MAD(const fs_reg &dst, const fs_reg &c, const fs_reg &b,
 173                 const fs_reg &a);
 174    fs_inst *ADDC(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 175    fs_inst *SUBB(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 176    fs_inst *SEL(const fs_reg &dst, const fs_reg &src0, const fs_reg &src1);
 177
 178    int type_size(const struct glsl_type *type);
 179    fs_inst *get_instruction_generating_reg(fs_inst *start,
 180                                            fs_inst *end,
 181                                            const fs_reg &reg);
 182
 183    fs_inst *LOAD_PAYLOAD(const fs_reg &dst, fs_reg *src, int sources);
 184
 185    exec_list VARYING_PULL_CONSTANT_LOAD(const fs_reg &dst,
 186                                         const fs_reg &surf_index,
 187                                         const fs_reg &varying_offset,
 188                                         uint32_t const_offset);
 189
 190    bool run_fs();
 191    bool run_vs();
 192    void optimize();
 193    void allocate_registers();
 194    void assign_binding_table_offsets();
 195    void setup_payload_gen4();
 196    void setup_payload_gen6();
 197    void setup_vs_payload();
 198    void fixup_3src_null_dest();
 199    void assign_curb_setup();
 200    void calculate_urb_setup();
 201    void assign_urb_setup();
 202    void assign_vs_urb_setup();
 203    bool assign_regs(bool allow_spilling);
 204    void assign_regs_trivial();
 205    void get_used_mrfs(bool *mrf_used);
 206    void setup_payload_interference(struct ra_graph *g, int payload_reg_count,
 207                                    int first_payload_node);
 208    void setup_mrf_hack_interference(struct ra_graph *g,
 209                                     int first_mrf_hack_node);
 210    int choose_spill_reg(struct ra_graph *g);
 211    void spill_reg(int spill_reg);
 212    void split_virtual_grfs();
 213    bool compact_virtual_grfs();
 214    void move_uniform_array_access_to_pull_constants();
 215    void assign_constant_locations();
 216    void demote_pull_constants();
 217    void invalidate_live_intervals();
 218    void calculate_live_intervals();
 219    void calculate_register_pressure();
 220    bool opt_algebraic();
 221    bool opt_redundant_discard_jumps();
 222    bool opt_cse();
 223    bool opt_cse_local(bblock_t *block);
 224    bool opt_copy_propagate();
 225    bool try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry);
 226    bool try_constant_propagate(fs_inst *inst, acp_entry *entry);
 227    bool opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
 228                                  exec_list *acp);
 229    bool opt_register_renaming();
 230    bool register_coalesce();
 231    bool compute_to_mrf();
 232    bool dead_code_eliminate();
 233    bool remove_duplicate_mrf_writes();
 234    bool virtual_grf_interferes(int a, int b);
 235    void schedule_instructions(instruction_scheduler_mode mode);
 236    void insert_gen4_send_dependency_workarounds();
 237    void insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
 238                                                     fs_inst *inst);
 239    void insert_gen4_post_send_dependency_workarounds(bblock_t *block,
 240                                                      fs_inst *inst);
 241    void vfail(const char *msg, va_list args);
 242    void fail(const char *msg, ...);
 243    void no16(const char *msg, ...);
 244    void lower_uniform_pull_constant_loads();
 245    bool lower_load_payload();
 246    bool opt_combine_constants();
 247
 248    void emit_dummy_fs();
 249    void emit_repclear_shader();
 250    fs_reg *emit_fragcoord_interpolation(bool pixel_center_integer,
 251                                         bool origin_upper_left);
 252    fs_inst *emit_linterp(const fs_reg &attr, const fs_reg &interp,
 253                          glsl_interp_qualifier interpolation_mode,
 254                          bool is_centroid, bool is_sample);
 255    fs_reg *emit_frontfacing_interpolation();
 256    fs_reg *emit_samplepos_setup();
 257    fs_reg *emit_sampleid_setup();
 258    void emit_general_interpolation(fs_reg attr, const char *name,
 259                                    const glsl_type *type,
 260                                    glsl_interp_qualifier interpolation_mode,
 261                                    int location, bool mod_centroid,
 262                                    bool mod_sample);
 263    fs_reg *emit_vs_system_value(int location);
 264    void emit_interpolation_setup_gen4();
 265    void emit_interpolation_setup_gen6();
 266    void compute_sample_position(fs_reg dst, fs_reg int_sample_pos);
 267    fs_reg rescale_texcoord(fs_reg coordinate, int coord_components,
 268                            bool is_rect, uint32_t sampler, int texunit);
 269    fs_inst *emit_texture_gen4(ir_texture_opcode op, fs_reg dst,
 270                               fs_reg coordinate, int coord_components,
 271                               fs_reg shadow_comp,
 272                               fs_reg lod, fs_reg lod2, int grad_components,
 273                               uint32_t sampler);
 274    fs_inst *emit_texture_gen5(ir_texture_opcode op, fs_reg dst,
 275                               fs_reg coordinate, int coord_components,
 276                               fs_reg shadow_comp,
 277                               fs_reg lod, fs_reg lod2, int grad_components,
 278                               fs_reg sample_index, uint32_t sampler,
 279                               bool has_offset);
 280    fs_inst *emit_texture_gen7(ir_texture_opcode op, fs_reg dst,
 281                               fs_reg coordinate, int coord_components,
 282                               fs_reg shadow_comp,
 283                               fs_reg lod, fs_reg lod2, int grad_components,
 284                               fs_reg sample_index, fs_reg mcs, fs_reg sampler,
 285                               fs_reg offset_value);
 286    void emit_texture(ir_texture_opcode op,
 287                      const glsl_type *dest_type,
 288                      fs_reg coordinate, int components,
 289                      fs_reg shadow_c,
 290                      fs_reg lod, fs_reg dpdy, int grad_components,
 291                      fs_reg sample_index,
 292                      fs_reg offset,
 293                      fs_reg mcs,
 294                      int gather_component,
 295                      bool is_cube_array,
 296                      bool is_rect,
 297                      uint32_t sampler,
 298                      fs_reg sampler_reg,
 299                      int texunit);
 300    fs_reg emit_mcs_fetch(fs_reg coordinate, int components, fs_reg sampler);
 301    void emit_gen6_gather_wa(uint8_t wa, fs_reg dst);
 302    void resolve_source_modifiers(fs_reg *src);
 303    fs_reg fix_math_operand(fs_reg src);
 304    fs_inst *emit_math(enum opcode op, fs_reg dst, fs_reg src0);
 305    fs_inst *emit_math(enum opcode op, fs_reg dst, fs_reg src0, fs_reg src1);
 306    void emit_lrp(const fs_reg &dst, const fs_reg &x, const fs_reg &y,
 307                  const fs_reg &a);
 308    void emit_minmax(enum brw_conditional_mod conditionalmod, const fs_reg &dst,
 309                     const fs_reg &src0, const fs_reg &src1);
 310    bool try_emit_saturate(ir_expression *ir);
 311    bool try_emit_line(ir_expression *ir);
 312    bool try_emit_mad(ir_expression *ir);
 313    bool try_replace_with_sel();
 314    bool try_opt_frontfacing_ternary(ir_if *ir);
 315    bool opt_peephole_sel();
 316    bool opt_peephole_predicated_break();
 317    bool opt_saturate_propagation();
 318    bool opt_cmod_propagation();
 319    void emit_bool_to_cond_code(ir_rvalue *condition);
 320    void emit_if_gen6(ir_if *ir);
 321    void emit_unspill(bblock_t *block, fs_inst *inst, fs_reg reg,
 322                      uint32_t spill_offset, int count);
 323    void emit_spill(bblock_t *block, fs_inst *inst, fs_reg reg,
 324                    uint32_t spill_offset, int count);
 325
 326    void emit_fragment_program_code();
 327    void setup_fp_regs();
 328    fs_reg get_fp_src_reg(const prog_src_register *src);
 329    fs_reg get_fp_dst_reg(const prog_dst_register *dst);
 330    void emit_fp_alu1(enum opcode opcode,
 331                      const struct prog_instruction *fpi,
 332                      fs_reg dst, fs_reg src);
 333    void emit_fp_alu2(enum opcode opcode,
 334                      const struct prog_instruction *fpi,
 335                      fs_reg dst, fs_reg src0, fs_reg src1);
 336    void emit_fp_scalar_write(const struct prog_instruction *fpi,
 337                              fs_reg dst, fs_reg src);
 338    void emit_fp_scalar_math(enum opcode opcode,
 339                             const struct prog_instruction *fpi,
 340                             fs_reg dst, fs_reg src);
 341
 342    void emit_fp_minmax(const struct prog_instruction *fpi,
 343                        fs_reg dst, fs_reg src0, fs_reg src1);
 344
 345    void emit_fp_sop(enum brw_conditional_mod conditional_mod,
 346                     const struct prog_instruction *fpi,
 347                     fs_reg dst, fs_reg src0, fs_reg src1, fs_reg one);
 348
 349    void emit_nir_code();
 350    void nir_setup_inputs(nir_shader *shader);
 351    void nir_setup_outputs(nir_shader *shader);
 352    void nir_setup_uniforms(nir_shader *shader);
 353    void nir_setup_uniform(nir_variable *var);
 354    void nir_setup_builtin_uniform(nir_variable *var);
 355    void nir_emit_system_values(nir_shader *shader);
 356    void nir_emit_impl(nir_function_impl *impl);
 357    void nir_emit_cf_list(exec_list *list);
 358    void nir_emit_if(nir_if *if_stmt);
 359    void nir_emit_loop(nir_loop *loop);
 360    void nir_emit_block(nir_block *block);
 361    void nir_emit_instr(nir_instr *instr);
 362    void nir_emit_alu(nir_alu_instr *instr);
 363    void nir_emit_intrinsic(nir_intrinsic_instr *instr);
 364    void nir_emit_texture(nir_tex_instr *instr);
 365    void nir_emit_jump(nir_jump_instr *instr);
 366    fs_reg get_nir_src(nir_src src);
 367    fs_reg get_nir_dest(nir_dest dest);
 368    void emit_percomp(fs_inst *inst, unsigned wr_mask);
 369
 370    bool optimize_frontfacing_ternary(nir_alu_instr *instr,
 371                                      const fs_reg &result);
 372
 373    int setup_color_payload(fs_reg *dst, fs_reg color, unsigned components,
 374                            bool use_2nd_half);
 375    void emit_alpha_test();
 376    fs_inst *emit_single_fb_write(fs_reg color1, fs_reg color2,
 377                                  fs_reg src0_alpha, unsigned components,
 378                                  bool use_2nd_half = false);
 379    void emit_fb_writes();
 380    void emit_urb_writes();
 381
 382    void emit_shader_time_begin();
 383    void emit_shader_time_end();
 384    fs_inst *SHADER_TIME_ADD(enum shader_time_shader_type type, fs_reg value);
 385
 386    void emit_untyped_atomic(unsigned atomic_op, unsigned surf_index,
 387                             fs_reg dst, fs_reg offset, fs_reg src0,
 388                             fs_reg src1);
 389
 390    void emit_untyped_surface_read(unsigned surf_index, fs_reg dst,
 391                                   fs_reg offset);
 392
 393    void emit_interpolate_expression(ir_expression *ir);
 394
 395    bool try_rewrite_rhs_to_dst(ir_assignment *ir,
 396                                fs_reg dst,
 397                                fs_reg src,
 398                                fs_inst *pre_rhs_inst,
 399                                fs_inst *last_rhs_inst);
 400    void emit_assignment_writes(fs_reg &l, fs_reg &r,
 401                                const glsl_type *type, bool predicated);
 402    void resolve_ud_negate(fs_reg *reg);
 403    void resolve_bool_comparison(ir_rvalue *rvalue, fs_reg *reg);
 404
 405    fs_reg get_timestamp();
 406
 407    struct brw_reg interp_reg(int location, int channel);
 408    void setup_uniform_values(ir_variable *ir);
 409    void setup_builtin_uniform_values(ir_variable *ir);
 410    int implied_mrf_writes(fs_inst *inst);
 411
 412    virtual void dump_instructions();
 413    virtual void dump_instructions(const char *name);
 414    void dump_instruction(backend_instruction *inst);
 415    void dump_instruction(backend_instruction *inst, FILE *file);
 416
 417    void visit_atomic_counter_intrinsic(ir_call *ir);
 418
 419    const void *const key;
 420    struct brw_stage_prog_data *prog_data;
 421    unsigned int sanity_param_count;
 422
 423    int *param_size;
 424
 425    int *virtual_grf_start;
 426    int *virtual_grf_end;
 427    brw::fs_live_variables *live_intervals;
 428
 429    int *regs_live_at_ip;
 430
 431    /** Number of uniform variable components visited. */
 432    unsigned uniforms;
 433
 434    /** Byte-offset for the next available spot in the scratch space buffer. */
 435    unsigned last_scratch;
 436
 437    /**
 438     * Array mapping UNIFORM register numbers to the pull parameter index,
 439     * or -1 if this uniform register isn't being uploaded as a pull constant.
 440     */
 441    int *pull_constant_loc;
 442
 443    /**
 444     * Array mapping UNIFORM register numbers to the push parameter index,
 445     * or -1 if this uniform register isn't being uploaded as a push constant.
 446     */
 447    int *push_constant_loc;
 448
 449    struct hash_table *variable_ht;
 450    fs_reg frag_depth;
 451    fs_reg sample_mask;
 452    fs_reg outputs[VARYING_SLOT_MAX];
 453    unsigned output_components[VARYING_SLOT_MAX];
 454    fs_reg dual_src_output;
 455    bool do_dual_src;
 456    int first_non_payload_grf;
 457    /** Either BRW_MAX_GRF or GEN7_MRF_HACK_START */
 458    unsigned max_grf;
 459
 460    fs_reg *fp_temp_regs;
 461    fs_reg *fp_input_regs;
 462
 463    fs_reg *nir_locals;
 464    fs_reg *nir_globals;
 465    fs_reg nir_inputs;
 466    fs_reg nir_outputs;
 467    fs_reg nir_uniforms;
 468    fs_reg *nir_system_values;
 469
 470    /** @{ debug annotation info */
 471    const char *current_annotation;
 472    const void *base_ir;
 473    /** @} */
 474
 475    bool failed;
 476    char *fail_msg;
 477    bool simd16_unsupported;
 478    char *no16_msg;
 479
 480    /* Result of last visit() method. */
 481    fs_reg result;
 482
 483    /** Register numbers for thread payload fields. */
 484    struct {
 485       uint8_t source_depth_reg;
 486       uint8_t source_w_reg;
 487       uint8_t aa_dest_stencil_reg;
 488       uint8_t dest_depth_reg;
 489       uint8_t sample_pos_reg;
 490       uint8_t sample_mask_in_reg;
 491       uint8_t barycentric_coord_reg[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 492
 493       /** The number of thread payload registers the hardware will supply. */
 494       uint8_t num_regs;
 495    } payload;
 496
 497    bool source_depth_to_render_target;
 498    bool runtime_check_aads_emit;
 499
 500    fs_reg pixel_x;
 501    fs_reg pixel_y;
 502    fs_reg wpos_w;
 503    fs_reg pixel_w;
 504    fs_reg delta_x[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 505    fs_reg delta_y[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT];
 506    fs_reg shader_start_time;
 507    fs_reg userplane[MAX_CLIP_PLANES];
 508
 509    unsigned grf_used;
 510    bool spilled_any_registers;
 511
 512    const unsigned dispatch_width; /**< 8 or 16 */
 513 };
 514
 515 /**
 516  * The fragment shader code generator.
 517  *
 518  * Translates FS IR to actual i965 assembly code.
 519  */
 520 class fs_generator
 521 {
 522 public:
 523    fs_generator(struct brw_context *brw,
 524                 void *mem_ctx,
 525                 const void *key,
 526                 struct brw_stage_prog_data *prog_data,
 527                 struct gl_program *fp,
 528                 bool runtime_check_aads_emit,
 529                 const char *stage_abbrev);
 530    ~fs_generator();
 531
 532    void enable_debug(const char *shader_name);
 533    int generate_code(const cfg_t *cfg, int dispatch_width);
 534    const unsigned *get_assembly(unsigned int *assembly_size);
 535
 536 private:
 537    void fire_fb_write(fs_inst *inst,
 538                       struct brw_reg payload,
 539                       struct brw_reg implied_header,
 540                       GLuint nr);
 541    void generate_fb_write(fs_inst *inst, struct brw_reg payload);
 542    void generate_urb_write(fs_inst *inst, struct brw_reg payload);
 543    void generate_blorp_fb_write(fs_inst *inst);
 544    void generate_pixel_xy(struct brw_reg dst, bool is_x);
 545    void generate_linterp(fs_inst *inst, struct brw_reg dst,
 546                          struct brw_reg *src);
 547    void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
 548                      struct brw_reg sampler_index);
 549    void generate_math_gen6(fs_inst *inst,
 550                            struct brw_reg dst,
 551                            struct brw_reg src0,
 552                            struct brw_reg src1);
 553    void generate_math_gen4(fs_inst *inst,
 554                            struct brw_reg dst,
 555                            struct brw_reg src);
 556    void generate_math_g45(fs_inst *inst,
 557                           struct brw_reg dst,
 558                           struct brw_reg src);
 559    void generate_ddx(enum opcode op, struct brw_reg dst, struct brw_reg src);
 560    void generate_ddy(enum opcode op, struct brw_reg dst, struct brw_reg src,
 561                      bool negate_value);
 562    void generate_scratch_write(fs_inst *inst, struct brw_reg src);
 563    void generate_scratch_read(fs_inst *inst, struct brw_reg dst);
 564    void generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst);
 565    void generate_uniform_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 566                                             struct brw_reg index,
 567                                             struct brw_reg offset);
 568    void generate_uniform_pull_constant_load_gen7(fs_inst *inst,
 569                                                  struct brw_reg dst,
 570                                                  struct brw_reg surf_index,
 571                                                  struct brw_reg offset);
 572    void generate_varying_pull_constant_load(fs_inst *inst, struct brw_reg dst,
 573                                             struct brw_reg index,
 574                                             struct brw_reg offset);
 575    void generate_varying_pull_constant_load_gen7(fs_inst *inst,
 576                                                  struct brw_reg dst,
 577                                                  struct brw_reg index,
 578                                                  struct brw_reg offset);
 579    void generate_mov_dispatch_to_flags(fs_inst *inst);
 580
 581    void generate_pixel_interpolator_query(fs_inst *inst,
 582                                           struct brw_reg dst,
 583                                           struct brw_reg src,
 584                                           struct brw_reg msg_data,
 585                                           unsigned msg_type);
 586
 587    void generate_set_omask(fs_inst *inst,
 588                            struct brw_reg dst,
 589                            struct brw_reg sample_mask);
 590
 591    void generate_set_sample_id(fs_inst *inst,
 592                                struct brw_reg dst,
 593                                struct brw_reg src0,
 594                                struct brw_reg src1);
 595
 596    void generate_set_simd4x2_offset(fs_inst *inst,
 597                                     struct brw_reg dst,
 598                                     struct brw_reg offset);
 599    void generate_discard_jump(fs_inst *inst);
 600
 601    void generate_pack_half_2x16_split(fs_inst *inst,
 602                                       struct brw_reg dst,
 603                                       struct brw_reg x,
 604                                       struct brw_reg y);
 605    void generate_unpack_half_2x16_split(fs_inst *inst,
 606                                         struct brw_reg dst,
 607                                         struct brw_reg src);
 608
 609    void generate_shader_time_add(fs_inst *inst,
 610                                  struct brw_reg payload,
 611                                  struct brw_reg offset,
 612                                  struct brw_reg value);
 613
 614    void generate_untyped_atomic(fs_inst *inst,
 615                                 struct brw_reg dst,
 616                                 struct brw_reg payload,
 617                                 struct brw_reg atomic_op,
 618                                 struct brw_reg surf_index);
 619
 620    void generate_untyped_surface_read(fs_inst *inst,
 621                                       struct brw_reg dst,
 622                                       struct brw_reg payload,
 623                                       struct brw_reg surf_index);
 624
 625    bool patch_discard_jumps_to_fb_writes();
 626
 627    struct brw_context *brw;
 628    struct gl_context *ctx;
 629
 630    struct brw_compile *p;
 631    const void * const key;
 632    struct brw_stage_prog_data * const prog_data;
 633
 634    const struct gl_program *prog;
 635
 636    unsigned dispatch_width; /**< 8 or 16 */
 637
 638    exec_list discard_halt_patches;
 639    bool runtime_check_aads_emit;
 640    bool debug_flag;
 641    const char *shader_name;
 642    const char *stage_abbrev;
 643    void *mem_ctx;
 644 };
 645
 646 bool brw_do_channel_expressions(struct exec_list *instructions);
 647 bool brw_do_vector_splitting(struct exec_list *instructions);