2 * Copyright © 2010 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
30 #include "brw_shader.h"
34 #include <sys/types.h>
36 #include "main/macros.h"
37 #include "main/shaderobj.h"
38 #include "main/uniforms.h"
39 #include "program/prog_parameter.h"
40 #include "program/prog_print.h"
41 #include "program/prog_optimize.h"
42 #include "program/register_allocate.h"
43 #include "program/sampler.h"
44 #include "program/hash_table.h"
45 #include "brw_context.h"
49 #include "glsl/glsl_types.h"
58 FIXED_HW_REG
, /* a struct brw_reg */
59 UNIFORM
, /* prog_data->params[reg] */
64 /* Callers of this ralloc-based new need not call delete. It's
65 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
66 static void* operator new(size_t size
, void *ctx
)
70 node
= ralloc_size(ctx
, size
);
78 memset(this, 0, sizeof(*this));
82 /** Generic unset register constructor. */
86 this->file
= BAD_FILE
;
89 /** Immediate value constructor. */
94 this->type
= BRW_REGISTER_TYPE_F
;
98 /** Immediate value constructor. */
103 this->type
= BRW_REGISTER_TYPE_D
;
107 /** Immediate value constructor. */
112 this->type
= BRW_REGISTER_TYPE_UD
;
116 /** Fixed brw_reg Immediate value constructor. */
117 fs_reg(struct brw_reg fixed_hw_reg
)
120 this->file
= FIXED_HW_REG
;
121 this->fixed_hw_reg
= fixed_hw_reg
;
122 this->type
= fixed_hw_reg
.type
;
125 fs_reg(enum register_file file
, int reg
);
126 fs_reg(enum register_file file
, int reg
, uint32_t type
);
127 fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
);
129 bool equals(const fs_reg
&r
) const
131 return (file
== r
.file
&&
133 reg_offset
== r
.reg_offset
&&
135 negate
== r
.negate
&&
137 memcmp(&fixed_hw_reg
, &r
.fixed_hw_reg
,
138 sizeof(fixed_hw_reg
)) == 0 &&
143 /** Register file: ARF, GRF, MRF, IMM. */
144 enum register_file file
;
146 * Register number. For ARF/MRF, it's the hardware register. For
147 * GRF, it's a virtual register number until register allocation
151 * For virtual registers, this is a hardware register offset from
152 * the start of the register block (for example, a constant index
153 * in an array access).
156 /** Register type. BRW_REGISTER_TYPE_* */
161 struct brw_reg fixed_hw_reg
;
162 int smear
; /* -1, or a channel of the reg to smear to all channels. */
164 /** Value for file == IMM */
172 static const fs_reg reg_undef
;
173 static const fs_reg
reg_null_f(ARF
, BRW_ARF_NULL
, BRW_REGISTER_TYPE_F
);
174 static const fs_reg
reg_null_d(ARF
, BRW_ARF_NULL
, BRW_REGISTER_TYPE_D
);
176 class ip_record
: public exec_node
{
178 static void* operator new(size_t size
, void *ctx
)
182 node
= rzalloc_size(ctx
, size
);
183 assert(node
!= NULL
);
196 class fs_inst
: public exec_node
{
198 /* Callers of this ralloc-based new need not call delete. It's
199 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
200 static void* operator new(size_t size
, void *ctx
)
204 node
= rzalloc_size(ctx
, size
);
205 assert(node
!= NULL
);
212 memset(this, 0, sizeof(*this));
213 this->opcode
= BRW_OPCODE_NOP
;
214 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
216 this->dst
= reg_undef
;
217 this->src
[0] = reg_undef
;
218 this->src
[1] = reg_undef
;
219 this->src
[2] = reg_undef
;
227 fs_inst(enum opcode opcode
)
230 this->opcode
= opcode
;
233 fs_inst(enum opcode opcode
, fs_reg dst
)
236 this->opcode
= opcode
;
240 assert(dst
.reg_offset
>= 0);
243 fs_inst(enum opcode opcode
, fs_reg dst
, fs_reg src0
)
246 this->opcode
= opcode
;
251 assert(dst
.reg_offset
>= 0);
252 if (src
[0].file
== GRF
)
253 assert(src
[0].reg_offset
>= 0);
256 fs_inst(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
259 this->opcode
= opcode
;
265 assert(dst
.reg_offset
>= 0);
266 if (src
[0].file
== GRF
)
267 assert(src
[0].reg_offset
>= 0);
268 if (src
[1].file
== GRF
)
269 assert(src
[1].reg_offset
>= 0);
272 fs_inst(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
, fs_reg src2
)
275 this->opcode
= opcode
;
282 assert(dst
.reg_offset
>= 0);
283 if (src
[0].file
== GRF
)
284 assert(src
[0].reg_offset
>= 0);
285 if (src
[1].file
== GRF
)
286 assert(src
[1].reg_offset
>= 0);
287 if (src
[2].file
== GRF
)
288 assert(src
[2].reg_offset
>= 0);
291 bool equals(fs_inst
*inst
)
293 return (opcode
== inst
->opcode
&&
294 dst
.equals(inst
->dst
) &&
295 src
[0].equals(inst
->src
[0]) &&
296 src
[1].equals(inst
->src
[1]) &&
297 src
[2].equals(inst
->src
[2]) &&
298 saturate
== inst
->saturate
&&
299 predicated
== inst
->predicated
&&
300 conditional_mod
== inst
->conditional_mod
&&
301 mlen
== inst
->mlen
&&
302 base_mrf
== inst
->base_mrf
&&
303 sampler
== inst
->sampler
&&
304 target
== inst
->target
&&
306 header_present
== inst
->header_present
&&
307 shadow_compare
== inst
->shadow_compare
&&
308 offset
== inst
->offset
);
316 /* The SINCOS and INT_DIV_QUOTIENT_AND_REMAINDER math functions return 2,
317 * but we don't currently use them...nor do we have an opcode for them.
325 return (opcode
== SHADER_OPCODE_TEX
||
326 opcode
== FS_OPCODE_TXB
||
327 opcode
== SHADER_OPCODE_TXD
||
328 opcode
== SHADER_OPCODE_TXF
||
329 opcode
== SHADER_OPCODE_TXL
||
330 opcode
== SHADER_OPCODE_TXS
);
335 return (opcode
== SHADER_OPCODE_RCP
||
336 opcode
== SHADER_OPCODE_RSQ
||
337 opcode
== SHADER_OPCODE_SQRT
||
338 opcode
== SHADER_OPCODE_EXP2
||
339 opcode
== SHADER_OPCODE_LOG2
||
340 opcode
== SHADER_OPCODE_SIN
||
341 opcode
== SHADER_OPCODE_COS
||
342 opcode
== SHADER_OPCODE_INT_QUOTIENT
||
343 opcode
== SHADER_OPCODE_INT_REMAINDER
||
344 opcode
== SHADER_OPCODE_POW
);
347 enum opcode opcode
; /* BRW_OPCODE_* or FS_OPCODE_* */
352 bool predicate_inverse
;
353 int conditional_mod
; /**< BRW_CONDITIONAL_* */
355 int mlen
; /**< SEND message length */
356 int base_mrf
; /**< First MRF in the SEND message, if mlen is nonzero. */
358 int target
; /**< MRT target. */
362 bool force_uncompressed
;
364 uint32_t offset
; /* spill/unspill offset */
367 * Annotation for the generated IR. One of the two can be set.
370 const char *annotation
;
374 class fs_visitor
: public ir_visitor
378 fs_visitor(struct brw_wm_compile
*c
, struct gl_shader_program
*prog
,
379 struct brw_shader
*shader
)
384 this->fp
= (struct gl_fragment_program
*)
385 prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
]->Program
;
387 this->intel
= &brw
->intel
;
388 this->ctx
= &intel
->ctx
;
389 this->mem_ctx
= ralloc_context(NULL
);
390 this->shader
= shader
;
391 this->failed
= false;
392 this->variable_ht
= hash_table_ctor(0,
393 hash_table_pointer_hash
,
394 hash_table_pointer_compare
);
396 /* There's a question that appears to be left open in the spec:
397 * How do implicit dst conversions interact with the CMP
398 * instruction or conditional mods? On gen6, the instruction:
400 * CMP null<d> src0<f> src1<f>
402 * will do src1 - src0 and compare that result as if it was an
403 * integer. On gen4, it will do src1 - src0 as float, convert
404 * the result to int, and compare as int. In between, it
405 * appears that it does src1 - src0 and does the compare in the
406 * execution type so dst type doesn't matter.
408 if (this->intel
->gen
> 4)
409 this->reg_null_cmp
= reg_null_d
;
411 this->reg_null_cmp
= reg_null_f
;
413 this->frag_depth
= NULL
;
414 memset(this->outputs
, 0, sizeof(this->outputs
));
415 this->first_non_payload_grf
= 0;
416 this->max_grf
= intel
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
418 this->current_annotation
= NULL
;
419 this->base_ir
= NULL
;
421 this->virtual_grf_sizes
= NULL
;
422 this->virtual_grf_next
= 0;
423 this->virtual_grf_array_size
= 0;
424 this->virtual_grf_def
= NULL
;
425 this->virtual_grf_use
= NULL
;
426 this->live_intervals_valid
= false;
428 this->kill_emitted
= false;
429 this->force_uncompressed_stack
= 0;
430 this->force_sechalf_stack
= 0;
435 ralloc_free(this->mem_ctx
);
436 hash_table_dtor(this->variable_ht
);
439 fs_reg
*variable_storage(ir_variable
*var
);
440 int virtual_grf_alloc(int size
);
441 void import_uniforms(fs_visitor
*v
);
443 void visit(ir_variable
*ir
);
444 void visit(ir_assignment
*ir
);
445 void visit(ir_dereference_variable
*ir
);
446 void visit(ir_dereference_record
*ir
);
447 void visit(ir_dereference_array
*ir
);
448 void visit(ir_expression
*ir
);
449 void visit(ir_texture
*ir
);
450 void visit(ir_if
*ir
);
451 void visit(ir_constant
*ir
);
452 void visit(ir_swizzle
*ir
);
453 void visit(ir_return
*ir
);
454 void visit(ir_loop
*ir
);
455 void visit(ir_loop_jump
*ir
);
456 void visit(ir_discard
*ir
);
457 void visit(ir_call
*ir
);
458 void visit(ir_function
*ir
);
459 void visit(ir_function_signature
*ir
);
461 void swizzle_result(ir_texture
*ir
, fs_reg orig_val
, int sampler
);
463 fs_inst
*emit(fs_inst inst
);
465 fs_inst
*emit(enum opcode opcode
)
467 return emit(fs_inst(opcode
));
470 fs_inst
*emit(enum opcode opcode
, fs_reg dst
)
472 return emit(fs_inst(opcode
, dst
));
475 fs_inst
*emit(enum opcode opcode
, fs_reg dst
, fs_reg src0
)
477 return emit(fs_inst(opcode
, dst
, src0
));
480 fs_inst
*emit(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
482 return emit(fs_inst(opcode
, dst
, src0
, src1
));
485 fs_inst
*emit(enum opcode opcode
, fs_reg dst
,
486 fs_reg src0
, fs_reg src1
, fs_reg src2
)
488 return emit(fs_inst(opcode
, dst
, src0
, src1
, src2
));
491 int type_size(const struct glsl_type
*type
);
492 fs_inst
*get_instruction_generating_reg(fs_inst
*start
,
497 void setup_paramvalues_refs();
498 void assign_curb_setup();
499 void calculate_urb_setup();
500 void assign_urb_setup();
502 void assign_regs_trivial();
503 int choose_spill_reg(struct ra_graph
*g
);
504 void spill_reg(int spill_reg
);
505 void split_virtual_grfs();
506 void setup_pull_constants();
507 void calculate_live_intervals();
508 bool propagate_constants();
509 bool opt_algebraic();
510 bool register_coalesce();
511 bool compute_to_mrf();
512 bool dead_code_eliminate();
513 bool remove_dead_constants();
514 bool remove_duplicate_mrf_writes();
515 bool virtual_grf_interferes(int a
, int b
);
516 void schedule_instructions();
517 void patch_discard_jumps_to_fb_writes();
518 void fail(const char *msg
, ...);
520 void push_force_uncompressed();
521 void pop_force_uncompressed();
522 void push_force_sechalf();
523 void pop_force_sechalf();
525 void generate_code();
526 void generate_fb_write(fs_inst
*inst
);
527 void generate_pixel_xy(struct brw_reg dst
, bool is_x
);
528 void generate_linterp(fs_inst
*inst
, struct brw_reg dst
,
529 struct brw_reg
*src
);
530 void generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
531 void generate_math1_gen7(fs_inst
*inst
,
534 void generate_math2_gen7(fs_inst
*inst
,
537 struct brw_reg src1
);
538 void generate_math1_gen6(fs_inst
*inst
,
541 void generate_math2_gen6(fs_inst
*inst
,
544 struct brw_reg src1
);
545 void generate_math_gen4(fs_inst
*inst
,
548 void generate_discard(fs_inst
*inst
);
549 void generate_ddx(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
550 void generate_ddy(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
551 void generate_spill(fs_inst
*inst
, struct brw_reg src
);
552 void generate_unspill(fs_inst
*inst
, struct brw_reg dst
);
553 void generate_pull_constant_load(fs_inst
*inst
, struct brw_reg dst
);
555 void emit_dummy_fs();
556 fs_reg
*emit_fragcoord_interpolation(ir_variable
*ir
);
557 fs_reg
*emit_frontfacing_interpolation(ir_variable
*ir
);
558 fs_reg
*emit_general_interpolation(ir_variable
*ir
);
559 void emit_interpolation_setup_gen4();
560 void emit_interpolation_setup_gen6();
561 fs_inst
*emit_texture_gen4(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
,
563 fs_inst
*emit_texture_gen5(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
,
565 fs_inst
*emit_texture_gen7(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
,
567 fs_inst
*emit_math(enum opcode op
, fs_reg dst
, fs_reg src0
);
568 fs_inst
*emit_math(enum opcode op
, fs_reg dst
, fs_reg src0
, fs_reg src1
);
569 bool try_emit_saturate(ir_expression
*ir
);
570 bool try_emit_mad(ir_expression
*ir
, int mul_arg
);
571 void emit_bool_to_cond_code(ir_rvalue
*condition
);
572 void emit_if_gen6(ir_if
*ir
);
573 void emit_unspill(fs_inst
*inst
, fs_reg reg
, uint32_t spill_offset
);
575 void emit_color_write(int target
, int index
, int first_color_mrf
);
576 void emit_fb_writes();
577 bool try_rewrite_rhs_to_dst(ir_assignment
*ir
,
580 fs_inst
*pre_rhs_inst
,
581 fs_inst
*last_rhs_inst
);
582 void emit_assignment_writes(fs_reg
&l
, fs_reg
&r
,
583 const glsl_type
*type
, bool predicated
);
584 void resolve_ud_negate(fs_reg
*reg
);
585 void resolve_bool_comparison(ir_rvalue
*rvalue
, fs_reg
*reg
);
587 struct brw_reg
interp_reg(int location
, int channel
);
588 int setup_uniform_values(int loc
, const glsl_type
*type
);
589 void setup_builtin_uniform_values(ir_variable
*ir
);
590 int implied_mrf_writes(fs_inst
*inst
);
592 struct brw_context
*brw
;
593 const struct gl_fragment_program
*fp
;
594 struct intel_context
*intel
;
595 struct gl_context
*ctx
;
596 struct brw_wm_compile
*c
;
597 struct brw_compile
*p
;
598 struct brw_shader
*shader
;
599 struct gl_shader_program
*prog
;
601 exec_list instructions
;
602 exec_list discard_halt_patches
;
604 /* Delayed setup of c->prog_data.params[] due to realloc of
605 * ParamValues[] during compile.
607 int param_index
[MAX_UNIFORMS
* 4];
608 int param_offset
[MAX_UNIFORMS
* 4];
610 int *virtual_grf_sizes
;
611 int virtual_grf_next
;
612 int virtual_grf_array_size
;
613 int *virtual_grf_def
;
614 int *virtual_grf_use
;
615 bool live_intervals_valid
;
617 /* This is the map from UNIFORM hw_reg + reg_offset as generated by
618 * the visitor to the packed uniform number after
619 * remove_dead_constants() that represents the actual uploaded
624 struct hash_table
*variable_ht
;
625 ir_variable
*frag_depth
;
626 fs_reg outputs
[BRW_MAX_DRAW_BUFFERS
];
627 int first_non_payload_grf
;
629 int urb_setup
[FRAG_ATTRIB_MAX
];
632 /** @{ debug annotation info */
633 const char *current_annotation
;
634 ir_instruction
*base_ir
;
640 /* Result of last visit() method. */
647 fs_reg delta_x
[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
];
648 fs_reg delta_y
[BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
];
653 int force_uncompressed_stack
;
654 int force_sechalf_stack
;
656 class fs_bblock
*bblock
;
659 bool brw_do_channel_expressions(struct exec_list
*instructions
);
660 bool brw_do_vector_splitting(struct exec_list
*instructions
);
661 bool brw_fs_precompile(struct gl_context
*ctx
, struct gl_shader_program
*prog
);