2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
33 #include "st_glsl_to_tgsi.h"
35 #include "glsl_parser_extras.h"
36 #include "ir_optimization.h"
38 #include "main/errors.h"
39 #include "main/shaderobj.h"
40 #include "main/uniforms.h"
41 #include "main/shaderapi.h"
42 #include "program/prog_instruction.h"
43 #include "program/sampler.h"
45 #include "pipe/p_context.h"
46 #include "pipe/p_screen.h"
47 #include "tgsi/tgsi_ureg.h"
48 #include "tgsi/tgsi_info.h"
49 #include "util/u_math.h"
50 #include "util/u_memory.h"
51 #include "st_program.h"
52 #include "st_mesa_to_tgsi.h"
55 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
56 #define PROGRAM_ANY_CONST ((1 << PROGRAM_STATE_VAR) | \
57 (1 << PROGRAM_CONSTANT) | \
58 (1 << PROGRAM_UNIFORM))
60 #define MAX_GLSL_TEXTURE_OFFSET 4
65 static int swizzle_for_size(int size
);
68 * This struct is a corresponding struct to TGSI ureg_src.
72 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
76 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
77 this->swizzle
= swizzle_for_size(type
->vector_elements
);
79 this->swizzle
= SWIZZLE_XYZW
;
82 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
84 this->reladdr2
= NULL
;
85 this->has_index2
= false;
86 this->double_reg2
= false;
90 st_src_reg(gl_register_file file
, int index
, int type
)
96 this->swizzle
= SWIZZLE_XYZW
;
99 this->reladdr2
= NULL
;
100 this->has_index2
= false;
101 this->double_reg2
= false;
105 st_src_reg(gl_register_file file
, int index
, int type
, int index2D
)
110 this->index2D
= index2D
;
111 this->swizzle
= SWIZZLE_XYZW
;
113 this->reladdr
= NULL
;
114 this->reladdr2
= NULL
;
115 this->has_index2
= false;
116 this->double_reg2
= false;
122 this->type
= GLSL_TYPE_ERROR
;
123 this->file
= PROGRAM_UNDEFINED
;
128 this->reladdr
= NULL
;
129 this->reladdr2
= NULL
;
130 this->has_index2
= false;
131 this->double_reg2
= false;
135 explicit st_src_reg(st_dst_reg reg
);
137 gl_register_file file
; /**< PROGRAM_* from Mesa */
138 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
140 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
141 int negate
; /**< NEGATE_XYZW mask from mesa */
142 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
143 /** Register index should be offset by the integer in this reg. */
145 st_src_reg
*reladdr2
;
148 * Is this the second half of a double register pair?
149 * currently used for input mapping only.
157 st_dst_reg(gl_register_file file
, int writemask
, int type
, int index
)
161 this->writemask
= writemask
;
162 this->cond_mask
= COND_TR
;
163 this->reladdr
= NULL
;
168 st_dst_reg(gl_register_file file
, int writemask
, int type
)
172 this->writemask
= writemask
;
173 this->cond_mask
= COND_TR
;
174 this->reladdr
= NULL
;
181 this->type
= GLSL_TYPE_ERROR
;
182 this->file
= PROGRAM_UNDEFINED
;
185 this->cond_mask
= COND_TR
;
186 this->reladdr
= NULL
;
190 explicit st_dst_reg(st_src_reg reg
);
192 gl_register_file file
; /**< PROGRAM_* from Mesa */
193 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
194 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
196 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
197 /** Register index should be offset by the integer in this reg. */
202 st_src_reg::st_src_reg(st_dst_reg reg
)
204 this->type
= reg
.type
;
205 this->file
= reg
.file
;
206 this->index
= reg
.index
;
207 this->swizzle
= SWIZZLE_XYZW
;
209 this->reladdr
= reg
.reladdr
;
211 this->reladdr2
= NULL
;
212 this->has_index2
= false;
213 this->double_reg2
= false;
214 this->array_id
= reg
.array_id
;
217 st_dst_reg::st_dst_reg(st_src_reg reg
)
219 this->type
= reg
.type
;
220 this->file
= reg
.file
;
221 this->index
= reg
.index
;
222 this->writemask
= WRITEMASK_XYZW
;
223 this->cond_mask
= COND_TR
;
224 this->reladdr
= reg
.reladdr
;
225 this->array_id
= reg
.array_id
;
228 class glsl_to_tgsi_instruction
: public exec_node
{
230 DECLARE_RALLOC_CXX_OPERATORS(glsl_to_tgsi_instruction
)
235 /** Pointer to the ir source this tree came from for debugging */
237 GLboolean cond_update
;
239 st_src_reg sampler
; /**< sampler register */
240 int sampler_array_size
; /**< 1-based size of sampler array, 1 if not array */
241 int tex_target
; /**< One of TEXTURE_*_INDEX */
242 glsl_base_type tex_type
;
243 GLboolean tex_shadow
;
245 st_src_reg tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
246 unsigned tex_offset_num_offset
;
247 int dead_mask
; /**< Used in dead code elimination */
249 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
252 class variable_storage
: public exec_node
{
254 variable_storage(ir_variable
*var
, gl_register_file file
, int index
,
255 unsigned array_id
= 0)
256 : file(file
), index(index
), var(var
), array_id(array_id
)
261 gl_register_file file
;
263 ir_variable
*var
; /* variable that maps to this, if any */
267 class immediate_storage
: public exec_node
{
269 immediate_storage(gl_constant_value
*values
, int size32
, int type
)
271 memcpy(this->values
, values
, size32
* sizeof(gl_constant_value
));
272 this->size32
= size32
;
276 /* doubles are stored across 2 gl_constant_values */
277 gl_constant_value values
[4];
278 int size32
; /**< Number of 32-bit components (1-4) */
279 int type
; /**< GL_DOUBLE, GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
282 class function_entry
: public exec_node
{
284 ir_function_signature
*sig
;
287 * identifier of this function signature used by the program.
289 * At the point that TGSI instructions for function calls are
290 * generated, we don't know the address of the first instruction of
291 * the function body. So we make the BranchTarget that is called a
292 * small integer and rewrite them during set_branchtargets().
297 * Pointer to first instruction of the function body.
299 * Set during function body emits after main() is processed.
301 glsl_to_tgsi_instruction
*bgn_inst
;
304 * Index of the first instruction of the function body in actual TGSI.
306 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
310 /** Storage for the return value. */
311 st_src_reg return_reg
;
314 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
315 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
323 struct glsl_to_tgsi_visitor
: public ir_visitor
{
325 glsl_to_tgsi_visitor();
326 ~glsl_to_tgsi_visitor();
328 function_entry
*current_function
;
330 struct gl_context
*ctx
;
331 struct gl_program
*prog
;
332 struct gl_shader_program
*shader_program
;
333 struct gl_shader
*shader
;
334 struct gl_shader_compiler_options
*options
;
338 unsigned *array_sizes
;
339 unsigned max_num_arrays
;
342 struct array_decl input_arrays
[PIPE_MAX_SHADER_INPUTS
];
343 unsigned num_input_arrays
;
344 struct array_decl output_arrays
[PIPE_MAX_SHADER_OUTPUTS
];
345 unsigned num_output_arrays
;
347 int num_address_regs
;
349 glsl_base_type sampler_types
[PIPE_MAX_SAMPLERS
];
350 int sampler_targets
[PIPE_MAX_SAMPLERS
]; /**< One of TGSI_TEXTURE_* */
351 bool indirect_addr_consts
;
352 int wpos_transform_const
;
355 bool native_integers
;
359 variable_storage
*find_variable_storage(ir_variable
*var
);
361 int add_constant(gl_register_file file
, gl_constant_value values
[8],
362 int size
, int datatype
, GLuint
*swizzle_out
);
364 function_entry
*get_function_signature(ir_function_signature
*sig
);
366 st_src_reg
get_temp(const glsl_type
*type
);
367 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
369 st_src_reg
st_src_reg_for_double(double val
);
370 st_src_reg
st_src_reg_for_float(float val
);
371 st_src_reg
st_src_reg_for_int(int val
);
372 st_src_reg
st_src_reg_for_type(int type
, int val
);
375 * \name Visit methods
377 * As typical for the visitor pattern, there must be one \c visit method for
378 * each concrete subclass of \c ir_instruction. Virtual base classes within
379 * the hierarchy should not have \c visit methods.
382 virtual void visit(ir_variable
*);
383 virtual void visit(ir_loop
*);
384 virtual void visit(ir_loop_jump
*);
385 virtual void visit(ir_function_signature
*);
386 virtual void visit(ir_function
*);
387 virtual void visit(ir_expression
*);
388 virtual void visit(ir_swizzle
*);
389 virtual void visit(ir_dereference_variable
*);
390 virtual void visit(ir_dereference_array
*);
391 virtual void visit(ir_dereference_record
*);
392 virtual void visit(ir_assignment
*);
393 virtual void visit(ir_constant
*);
394 virtual void visit(ir_call
*);
395 virtual void visit(ir_return
*);
396 virtual void visit(ir_discard
*);
397 virtual void visit(ir_texture
*);
398 virtual void visit(ir_if
*);
399 virtual void visit(ir_emit_vertex
*);
400 virtual void visit(ir_end_primitive
*);
401 virtual void visit(ir_barrier
*);
406 /** List of variable_storage */
409 /** List of immediate_storage */
410 exec_list immediates
;
411 unsigned num_immediates
;
413 /** List of function_entry */
414 exec_list function_signatures
;
415 int next_signature_id
;
417 /** List of glsl_to_tgsi_instruction */
418 exec_list instructions
;
420 glsl_to_tgsi_instruction
*emit_asm(ir_instruction
*ir
, unsigned op
,
421 st_dst_reg dst
= undef_dst
,
422 st_src_reg src0
= undef_src
,
423 st_src_reg src1
= undef_src
,
424 st_src_reg src2
= undef_src
,
425 st_src_reg src3
= undef_src
);
427 glsl_to_tgsi_instruction
*emit_asm(ir_instruction
*ir
, unsigned op
,
428 st_dst_reg dst
, st_dst_reg dst1
,
429 st_src_reg src0
= undef_src
,
430 st_src_reg src1
= undef_src
,
431 st_src_reg src2
= undef_src
,
432 st_src_reg src3
= undef_src
);
434 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
436 st_src_reg src0
, st_src_reg src1
);
439 * Emit the correct dot-product instruction for the type of arguments
441 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
447 void emit_scalar(ir_instruction
*ir
, unsigned op
,
448 st_dst_reg dst
, st_src_reg src0
);
450 void emit_scalar(ir_instruction
*ir
, unsigned op
,
451 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
453 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
455 bool try_emit_mad(ir_expression
*ir
,
457 bool try_emit_mad_for_and_not(ir_expression
*ir
,
460 void emit_swz(ir_expression
*ir
);
462 bool process_move_condition(ir_rvalue
*ir
);
464 void simplify_cmp(void);
466 void rename_temp_register(int index
, int new_index
);
467 int get_first_temp_read(int index
);
468 int get_first_temp_write(int index
);
469 int get_last_temp_read(int index
);
470 int get_last_temp_write(int index
);
472 void copy_propagate(void);
473 int eliminate_dead_code(void);
475 void merge_two_dsts(void);
476 void merge_registers(void);
477 void renumber_registers(void);
479 void emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
480 st_dst_reg
*l
, st_src_reg
*r
,
481 st_src_reg
*cond
, bool cond_swap
);
486 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 0);
487 static st_dst_reg address_reg2
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 1);
488 static st_dst_reg sampler_reladdr
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 2);
491 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
494 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
498 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
501 prog
->LinkStatus
= GL_FALSE
;
505 swizzle_for_size(int size
)
507 static const int size_swizzles
[4] = {
508 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
509 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
510 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
511 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
514 assert((size
>= 1) && (size
<= 4));
515 return size_swizzles
[size
- 1];
519 is_tex_instruction(unsigned opcode
)
521 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
526 num_inst_dst_regs(unsigned opcode
)
528 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
529 return info
->num_dst
;
533 num_inst_src_regs(unsigned opcode
)
535 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
536 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
539 glsl_to_tgsi_instruction
*
540 glsl_to_tgsi_visitor::emit_asm(ir_instruction
*ir
, unsigned op
,
541 st_dst_reg dst
, st_dst_reg dst1
,
542 st_src_reg src0
, st_src_reg src1
,
543 st_src_reg src2
, st_src_reg src3
)
545 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
546 int num_reladdr
= 0, i
, j
;
548 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
550 /* If we have to do relative addressing, we want to load the ARL
551 * reg directly for one of the regs, and preload the other reladdr
552 * sources into temps.
554 num_reladdr
+= dst
.reladdr
!= NULL
;
555 num_reladdr
+= dst1
.reladdr
!= NULL
;
556 num_reladdr
+= src0
.reladdr
!= NULL
|| src0
.reladdr2
!= NULL
;
557 num_reladdr
+= src1
.reladdr
!= NULL
|| src1
.reladdr2
!= NULL
;
558 num_reladdr
+= src2
.reladdr
!= NULL
|| src2
.reladdr2
!= NULL
;
559 num_reladdr
+= src3
.reladdr
!= NULL
|| src3
.reladdr2
!= NULL
;
561 reladdr_to_temp(ir
, &src3
, &num_reladdr
);
562 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
563 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
564 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
567 emit_arl(ir
, address_reg
, *dst
.reladdr
);
571 emit_arl(ir
, address_reg
, *dst1
.reladdr
);
574 assert(num_reladdr
== 0);
585 /* default to float, for paths where this is not initialized
586 * (since 0==UINT which is likely wrong):
588 inst
->tex_type
= GLSL_TYPE_FLOAT
;
590 inst
->function
= NULL
;
592 /* Update indirect addressing status used by TGSI */
595 case PROGRAM_STATE_VAR
:
596 case PROGRAM_CONSTANT
:
597 case PROGRAM_UNIFORM
:
598 this->indirect_addr_consts
= true;
600 case PROGRAM_IMMEDIATE
:
601 assert(!"immediates should not have indirect addressing");
608 for (i
= 0; i
< 4; i
++) {
609 if(inst
->src
[i
].reladdr
) {
610 switch(inst
->src
[i
].file
) {
611 case PROGRAM_STATE_VAR
:
612 case PROGRAM_CONSTANT
:
613 case PROGRAM_UNIFORM
:
614 this->indirect_addr_consts
= true;
616 case PROGRAM_IMMEDIATE
:
617 assert(!"immediates should not have indirect addressing");
626 this->instructions
.push_tail(inst
);
629 * This section contains the double processing.
630 * GLSL just represents doubles as single channel values,
631 * however most HW and TGSI represent doubles as pairs of register channels.
633 * so we have to fixup destination writemask/index and src swizzle/indexes.
634 * dest writemasks need to translate from single channel write mask
635 * to a dual-channel writemask, but also need to modify the index,
636 * if we are touching the Z,W fields in the pre-translated writemask.
638 * src channels have similiar index modifications along with swizzle
639 * changes to we pick the XY, ZW pairs from the correct index.
641 * GLSL [0].x -> TGSI [0].xy
642 * GLSL [0].y -> TGSI [0].zw
643 * GLSL [0].z -> TGSI [1].xy
644 * GLSL [0].w -> TGSI [1].zw
646 if (inst
->dst
[0].type
== GLSL_TYPE_DOUBLE
|| inst
->dst
[1].type
== GLSL_TYPE_DOUBLE
||
647 inst
->src
[0].type
== GLSL_TYPE_DOUBLE
) {
648 glsl_to_tgsi_instruction
*dinst
= NULL
;
649 int initial_src_swz
[4], initial_src_idx
[4];
650 int initial_dst_idx
[2], initial_dst_writemask
[2];
651 /* select the writemask for dst0 or dst1 */
652 unsigned writemask
= inst
->dst
[0].file
== PROGRAM_UNDEFINED
? inst
->dst
[1].writemask
: inst
->dst
[0].writemask
;
654 /* copy out the writemask, index and swizzles for all src/dsts. */
655 for (j
= 0; j
< 2; j
++) {
656 initial_dst_writemask
[j
] = inst
->dst
[j
].writemask
;
657 initial_dst_idx
[j
] = inst
->dst
[j
].index
;
660 for (j
= 0; j
< 4; j
++) {
661 initial_src_swz
[j
] = inst
->src
[j
].swizzle
;
662 initial_src_idx
[j
] = inst
->src
[j
].index
;
666 * scan all the components in the dst writemask
667 * generate an instruction for each of them if required.
671 int i
= u_bit_scan(&writemask
);
673 /* first time use previous instruction */
677 /* create a new instructions for subsequent attempts */
678 dinst
= new(mem_ctx
) glsl_to_tgsi_instruction();
682 this->instructions
.push_tail(dinst
);
685 /* modify the destination if we are splitting */
686 for (j
= 0; j
< 2; j
++) {
687 if (dinst
->dst
[j
].type
== GLSL_TYPE_DOUBLE
) {
688 dinst
->dst
[j
].writemask
= (i
& 1) ? WRITEMASK_ZW
: WRITEMASK_XY
;
689 dinst
->dst
[j
].index
= initial_dst_idx
[j
];
691 dinst
->dst
[j
].index
++;
693 /* if we aren't writing to a double, just get the bit of the initial writemask
695 dinst
->dst
[j
].writemask
= initial_dst_writemask
[j
] & (1 << i
);
699 /* modify the src registers */
700 for (j
= 0; j
< 4; j
++) {
701 int swz
= GET_SWZ(initial_src_swz
[j
], i
);
703 if (dinst
->src
[j
].type
== GLSL_TYPE_DOUBLE
) {
704 dinst
->src
[j
].index
= initial_src_idx
[j
];
706 dinst
->src
[j
].double_reg2
= true;
707 dinst
->src
[j
].index
++;
711 dinst
->src
[j
].swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
713 dinst
->src
[j
].swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_X
, SWIZZLE_Y
);
716 /* some opcodes are special case in what they use as sources
717 - F2D is a float src0, DLDEXP is integer src1 */
718 if (op
== TGSI_OPCODE_F2D
||
719 op
== TGSI_OPCODE_DLDEXP
||
720 (op
== TGSI_OPCODE_UCMP
&& dinst
->dst
[0].type
== GLSL_TYPE_DOUBLE
)) {
721 dinst
->src
[j
].swizzle
= MAKE_SWIZZLE4(swz
, swz
, swz
, swz
);
733 glsl_to_tgsi_instruction
*
734 glsl_to_tgsi_visitor::emit_asm(ir_instruction
*ir
, unsigned op
,
736 st_src_reg src0
, st_src_reg src1
,
737 st_src_reg src2
, st_src_reg src3
)
739 return emit_asm(ir
, op
, dst
, undef_dst
, src0
, src1
, src2
, src3
);
743 * Determines whether to use an integer, unsigned integer, or float opcode
744 * based on the operands and input opcode, then emits the result.
747 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
749 st_src_reg src0
, st_src_reg src1
)
751 int type
= GLSL_TYPE_FLOAT
;
753 if (op
== TGSI_OPCODE_MOV
)
756 assert(src0
.type
!= GLSL_TYPE_ARRAY
);
757 assert(src0
.type
!= GLSL_TYPE_STRUCT
);
758 assert(src1
.type
!= GLSL_TYPE_ARRAY
);
759 assert(src1
.type
!= GLSL_TYPE_STRUCT
);
761 if (src0
.type
== GLSL_TYPE_DOUBLE
|| src1
.type
== GLSL_TYPE_DOUBLE
)
762 type
= GLSL_TYPE_DOUBLE
;
763 else if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
764 type
= GLSL_TYPE_FLOAT
;
765 else if (native_integers
)
766 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
768 #define case5(c, f, i, u, d) \
769 case TGSI_OPCODE_##c: \
770 if (type == GLSL_TYPE_DOUBLE) \
771 op = TGSI_OPCODE_##d; \
772 else if (type == GLSL_TYPE_INT) \
773 op = TGSI_OPCODE_##i; \
774 else if (type == GLSL_TYPE_UINT) \
775 op = TGSI_OPCODE_##u; \
777 op = TGSI_OPCODE_##f; \
780 #define case4(c, f, i, u) \
781 case TGSI_OPCODE_##c: \
782 if (type == GLSL_TYPE_INT) \
783 op = TGSI_OPCODE_##i; \
784 else if (type == GLSL_TYPE_UINT) \
785 op = TGSI_OPCODE_##u; \
787 op = TGSI_OPCODE_##f; \
790 #define case3(f, i, u) case4(f, f, i, u)
791 #define case4d(f, i, u, d) case5(f, f, i, u, d)
792 #define case3fid(f, i, d) case5(f, f, i, i, d)
793 #define case2fi(f, i) case4(f, f, i, i)
794 #define case2iu(i, u) case4(i, LAST, i, u)
796 #define casecomp(c, f, i, u, d) \
797 case TGSI_OPCODE_##c: \
798 if (type == GLSL_TYPE_DOUBLE) \
799 op = TGSI_OPCODE_##d; \
800 else if (type == GLSL_TYPE_INT) \
801 op = TGSI_OPCODE_##i; \
802 else if (type == GLSL_TYPE_UINT) \
803 op = TGSI_OPCODE_##u; \
804 else if (native_integers) \
805 op = TGSI_OPCODE_##f; \
807 op = TGSI_OPCODE_##c; \
811 case3fid(ADD
, UADD
, DADD
);
812 case3fid(MUL
, UMUL
, DMUL
);
813 case3fid(MAD
, UMAD
, DMAD
);
814 case3fid(FMA
, UMAD
, DFMA
);
815 case3(DIV
, IDIV
, UDIV
);
816 case4d(MAX
, IMAX
, UMAX
, DMAX
);
817 case4d(MIN
, IMIN
, UMIN
, DMIN
);
820 casecomp(SEQ
, FSEQ
, USEQ
, USEQ
, DSEQ
);
821 casecomp(SNE
, FSNE
, USNE
, USNE
, DSNE
);
822 casecomp(SGE
, FSGE
, ISGE
, USGE
, DSGE
);
823 casecomp(SLT
, FSLT
, ISLT
, USLT
, DSLT
);
827 case3fid(SSG
, ISSG
, DSSG
);
828 case3fid(ABS
, IABS
, DABS
);
832 case2iu(IMUL_HI
, UMUL_HI
);
834 case3fid(SQRT
, SQRT
, DSQRT
);
836 case3fid(RCP
, RCP
, DRCP
);
837 case3fid(RSQ
, RSQ
, DRSQ
);
839 case3fid(FRC
, FRC
, DFRAC
);
840 case3fid(TRUNC
, TRUNC
, DTRUNC
);
841 case3fid(CEIL
, CEIL
, DCEIL
);
842 case3fid(FLR
, FLR
, DFLR
);
843 case3fid(ROUND
, ROUND
, DROUND
);
848 assert(op
!= TGSI_OPCODE_LAST
);
852 glsl_to_tgsi_instruction
*
853 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
854 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
857 static const unsigned dot_opcodes
[] = {
858 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
861 return emit_asm(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
865 * Emits TGSI scalar opcodes to produce unique answers across channels.
867 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
868 * channel determines the result across all channels. So to do a vec4
869 * of this operation, we want to emit a scalar per source channel used
870 * to produce dest channels.
873 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
875 st_src_reg orig_src0
, st_src_reg orig_src1
)
878 int done_mask
= ~dst
.writemask
;
880 /* TGSI RCP is a scalar operation splatting results to all channels,
881 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
884 for (i
= 0; i
< 4; i
++) {
885 GLuint this_mask
= (1 << i
);
886 st_src_reg src0
= orig_src0
;
887 st_src_reg src1
= orig_src1
;
889 if (done_mask
& this_mask
)
892 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
893 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
894 for (j
= i
+ 1; j
< 4; j
++) {
895 /* If there is another enabled component in the destination that is
896 * derived from the same inputs, generate its value on this pass as
899 if (!(done_mask
& (1 << j
)) &&
900 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
901 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
902 this_mask
|= (1 << j
);
905 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
906 src0_swiz
, src0_swiz
);
907 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
908 src1_swiz
, src1_swiz
);
910 dst
.writemask
= this_mask
;
911 emit_asm(ir
, op
, dst
, src0
, src1
);
912 done_mask
|= this_mask
;
917 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
918 st_dst_reg dst
, st_src_reg src0
)
920 st_src_reg undef
= undef_src
;
922 undef
.swizzle
= SWIZZLE_XXXX
;
924 emit_scalar(ir
, op
, dst
, src0
, undef
);
928 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
929 st_dst_reg dst
, st_src_reg src0
)
931 int op
= TGSI_OPCODE_ARL
;
933 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
934 op
= TGSI_OPCODE_UARL
;
936 assert(dst
.file
== PROGRAM_ADDRESS
);
937 if (dst
.index
>= this->num_address_regs
)
938 this->num_address_regs
= dst
.index
+ 1;
940 emit_asm(NULL
, op
, dst
, src0
);
944 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
945 gl_constant_value values
[8], int size
, int datatype
,
948 if (file
== PROGRAM_CONSTANT
) {
949 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
950 size
, datatype
, swizzle_out
);
953 assert(file
== PROGRAM_IMMEDIATE
);
956 immediate_storage
*entry
;
957 int size32
= size
* (datatype
== GL_DOUBLE
? 2 : 1);
960 /* Search immediate storage to see if we already have an identical
961 * immediate that we can use instead of adding a duplicate entry.
963 foreach_in_list(immediate_storage
, entry
, &this->immediates
) {
964 immediate_storage
*tmp
= entry
;
966 for (i
= 0; i
* 4 < size32
; i
++) {
967 int slot_size
= MIN2(size32
- (i
* 4), 4);
968 if (tmp
->type
!= datatype
|| tmp
->size32
!= slot_size
)
970 if (memcmp(tmp
->values
, &values
[i
* 4],
971 slot_size
* sizeof(gl_constant_value
)))
974 /* Everything matches, keep going until the full size is matched */
975 tmp
= (immediate_storage
*)tmp
->next
;
978 /* The full value matched */
985 for (i
= 0; i
* 4 < size32
; i
++) {
986 int slot_size
= MIN2(size32
- (i
* 4), 4);
987 /* Add this immediate to the list. */
988 entry
= new(mem_ctx
) immediate_storage(&values
[i
* 4], slot_size
, datatype
);
989 this->immediates
.push_tail(entry
);
990 this->num_immediates
++;
996 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
998 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
999 union gl_constant_value uval
;
1002 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
1008 glsl_to_tgsi_visitor::st_src_reg_for_double(double val
)
1010 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_DOUBLE
);
1011 union gl_constant_value uval
[2];
1013 uval
[0].u
= *(uint32_t *)&val
;
1014 uval
[1].u
= *(((uint32_t *)&val
) + 1);
1015 src
.index
= add_constant(src
.file
, uval
, 1, GL_DOUBLE
, &src
.swizzle
);
1021 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
1023 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
1024 union gl_constant_value uval
;
1026 assert(native_integers
);
1029 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
1035 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
1037 if (native_integers
)
1038 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
1039 st_src_reg_for_int(val
);
1041 return st_src_reg_for_float(val
);
1045 type_size(const struct glsl_type
*type
)
1050 switch (type
->base_type
) {
1051 case GLSL_TYPE_UINT
:
1053 case GLSL_TYPE_FLOAT
:
1054 case GLSL_TYPE_BOOL
:
1055 if (type
->is_matrix()) {
1056 return type
->matrix_columns
;
1058 /* Regardless of size of vector, it gets a vec4. This is bad
1059 * packing for things like floats, but otherwise arrays become a
1060 * mess. Hopefully a later pass over the code can pack scalars
1061 * down if appropriate.
1066 case GLSL_TYPE_DOUBLE
:
1067 if (type
->is_matrix()) {
1068 if (type
->vector_elements
<= 2)
1069 return type
->matrix_columns
;
1071 return type
->matrix_columns
* 2;
1073 /* For doubles if we have a double or dvec2 they fit in one
1074 * vec4, else they need 2 vec4s.
1076 if (type
->vector_elements
<= 2)
1082 case GLSL_TYPE_ARRAY
:
1083 assert(type
->length
> 0);
1084 return type_size(type
->fields
.array
) * type
->length
;
1085 case GLSL_TYPE_STRUCT
:
1087 for (i
= 0; i
< type
->length
; i
++) {
1088 size
+= type_size(type
->fields
.structure
[i
].type
);
1091 case GLSL_TYPE_SAMPLER
:
1092 case GLSL_TYPE_IMAGE
:
1093 /* Samplers take up one slot in UNIFORMS[], but they're baked in
1097 case GLSL_TYPE_ATOMIC_UINT
:
1098 case GLSL_TYPE_INTERFACE
:
1099 case GLSL_TYPE_VOID
:
1100 case GLSL_TYPE_ERROR
:
1101 assert(!"Invalid type in type_size");
1108 * In the initial pass of codegen, we assign temporary numbers to
1109 * intermediate results. (not SSA -- variable assignments will reuse
1113 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
1117 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
1121 if (!options
->EmitNoIndirectTemp
&&
1122 (type
->is_array() || type
->is_matrix())) {
1124 if (next_array
>= max_num_arrays
) {
1125 max_num_arrays
+= 32;
1126 array_sizes
= (unsigned*)
1127 realloc(array_sizes
, sizeof(array_sizes
[0]) * max_num_arrays
);
1130 src
.file
= PROGRAM_ARRAY
;
1131 src
.index
= next_array
<< 16 | 0x8000;
1132 array_sizes
[next_array
] = type_size(type
);
1136 src
.file
= PROGRAM_TEMPORARY
;
1137 src
.index
= next_temp
;
1138 next_temp
+= type_size(type
);
1141 if (type
->is_array() || type
->is_record()) {
1142 src
.swizzle
= SWIZZLE_NOOP
;
1144 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
1151 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1154 foreach_in_list(variable_storage
, entry
, &this->variables
) {
1155 if (entry
->var
== var
)
1163 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1165 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1166 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1168 fp
->OriginUpperLeft
= ir
->data
.origin_upper_left
;
1169 fp
->PixelCenterInteger
= ir
->data
.pixel_center_integer
;
1172 if (ir
->data
.mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1174 const ir_state_slot
*const slots
= ir
->get_state_slots();
1175 assert(slots
!= NULL
);
1177 /* Check if this statevar's setup in the STATE file exactly
1178 * matches how we'll want to reference it as a
1179 * struct/array/whatever. If not, then we need to move it into
1180 * temporary storage and hope that it'll get copy-propagated
1183 for (i
= 0; i
< ir
->get_num_state_slots(); i
++) {
1184 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1189 variable_storage
*storage
;
1191 if (i
== ir
->get_num_state_slots()) {
1192 /* We'll set the index later. */
1193 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1194 this->variables
.push_tail(storage
);
1198 /* The variable_storage constructor allocates slots based on the size
1199 * of the type. However, this had better match the number of state
1200 * elements that we're going to copy into the new temporary.
1202 assert((int) ir
->get_num_state_slots() == type_size(ir
->type
));
1204 dst
= st_dst_reg(get_temp(ir
->type
));
1206 storage
= new(mem_ctx
) variable_storage(ir
, dst
.file
, dst
.index
);
1208 this->variables
.push_tail(storage
);
1212 for (unsigned int i
= 0; i
< ir
->get_num_state_slots(); i
++) {
1213 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1214 (gl_state_index
*)slots
[i
].tokens
);
1216 if (storage
->file
== PROGRAM_STATE_VAR
) {
1217 if (storage
->index
== -1) {
1218 storage
->index
= index
;
1220 assert(index
== storage
->index
+ (int)i
);
1223 /* We use GLSL_TYPE_FLOAT here regardless of the actual type of
1224 * the data being moved since MOV does not care about the type of
1225 * data it is moving, and we don't want to declare registers with
1226 * array or struct types.
1228 st_src_reg
src(PROGRAM_STATE_VAR
, index
, GLSL_TYPE_FLOAT
);
1229 src
.swizzle
= slots
[i
].swizzle
;
1230 emit_asm(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1231 /* even a float takes up a whole vec4 reg in a struct/array. */
1236 if (storage
->file
== PROGRAM_TEMPORARY
&&
1237 dst
.index
!= storage
->index
+ (int) ir
->get_num_state_slots()) {
1238 fail_link(this->shader_program
,
1239 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1240 ir
->name
, dst
.index
- storage
->index
,
1241 type_size(ir
->type
));
1247 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1249 emit_asm(NULL
, TGSI_OPCODE_BGNLOOP
);
1251 visit_exec_list(&ir
->body_instructions
, this);
1253 emit_asm(NULL
, TGSI_OPCODE_ENDLOOP
);
1257 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1260 case ir_loop_jump::jump_break
:
1261 emit_asm(NULL
, TGSI_OPCODE_BRK
);
1263 case ir_loop_jump::jump_continue
:
1264 emit_asm(NULL
, TGSI_OPCODE_CONT
);
1271 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1278 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1280 /* Ignore function bodies other than main() -- we shouldn't see calls to
1281 * them since they should all be inlined before we get to glsl_to_tgsi.
1283 if (strcmp(ir
->name
, "main") == 0) {
1284 const ir_function_signature
*sig
;
1287 sig
= ir
->matching_signature(NULL
, &empty
, false);
1291 foreach_in_list(ir_instruction
, ir
, &sig
->body
) {
1298 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1300 int nonmul_operand
= 1 - mul_operand
;
1302 st_dst_reg result_dst
;
1304 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1305 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1308 expr
->operands
[0]->accept(this);
1310 expr
->operands
[1]->accept(this);
1312 ir
->operands
[nonmul_operand
]->accept(this);
1315 this->result
= get_temp(ir
->type
);
1316 result_dst
= st_dst_reg(this->result
);
1317 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1318 emit_asm(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1324 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1326 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1327 * implemented using multiplication, and logical-or is implemented using
1328 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1329 * As result, the logical expression (a & !b) can be rewritten as:
1333 * - (a * 1) - (a * b)
1337 * This final expression can be implemented as a single MAD(a, -b, a)
1341 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1343 const int other_operand
= 1 - try_operand
;
1346 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1347 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1350 ir
->operands
[other_operand
]->accept(this);
1352 expr
->operands
[0]->accept(this);
1355 b
.negate
= ~b
.negate
;
1357 this->result
= get_temp(ir
->type
);
1358 emit_asm(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1364 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1365 st_src_reg
*reg
, int *num_reladdr
)
1367 if (!reg
->reladdr
&& !reg
->reladdr2
)
1370 if (reg
->reladdr
) emit_arl(ir
, address_reg
, *reg
->reladdr
);
1371 if (reg
->reladdr2
) emit_arl(ir
, address_reg2
, *reg
->reladdr2
);
1373 if (*num_reladdr
!= 1) {
1374 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1376 emit_asm(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1384 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1386 unsigned int operand
;
1387 st_src_reg op
[ARRAY_SIZE(ir
->operands
)];
1388 st_src_reg result_src
;
1389 st_dst_reg result_dst
;
1391 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1393 if (ir
->operation
== ir_binop_add
) {
1394 if (try_emit_mad(ir
, 1))
1396 if (try_emit_mad(ir
, 0))
1400 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1402 if (!native_integers
&& ir
->operation
== ir_binop_logic_and
) {
1403 if (try_emit_mad_for_and_not(ir
, 1))
1405 if (try_emit_mad_for_and_not(ir
, 0))
1409 if (ir
->operation
== ir_quadop_vector
)
1410 assert(!"ir_quadop_vector should have been lowered");
1412 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1413 this->result
.file
= PROGRAM_UNDEFINED
;
1414 ir
->operands
[operand
]->accept(this);
1415 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1416 printf("Failed to get tree for expression operand:\n");
1417 ir
->operands
[operand
]->print();
1421 op
[operand
] = this->result
;
1423 /* Matrix expression operands should have been broken down to vector
1424 * operations already.
1426 assert(!ir
->operands
[operand
]->type
->is_matrix());
1429 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1430 if (ir
->operands
[1]) {
1431 vector_elements
= MAX2(vector_elements
,
1432 ir
->operands
[1]->type
->vector_elements
);
1435 this->result
.file
= PROGRAM_UNDEFINED
;
1437 /* Storage for our result. Ideally for an assignment we'd be using
1438 * the actual storage for the result here, instead.
1440 result_src
= get_temp(ir
->type
);
1441 /* convenience for the emit functions below. */
1442 result_dst
= st_dst_reg(result_src
);
1443 /* Limit writes to the channels that will be used by result_src later.
1444 * This does limit this temp's use as a temporary for multi-instruction
1447 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1449 switch (ir
->operation
) {
1450 case ir_unop_logic_not
:
1451 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1452 emit_asm(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1454 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1455 * older GPUs implement SEQ using multiple instructions (i915 uses two
1456 * SGE instructions and a MUL instruction). Since our logic values are
1457 * 0.0 and 1.0, 1-x also implements !x.
1459 op
[0].negate
= ~op
[0].negate
;
1460 emit_asm(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1464 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1465 emit_asm(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1466 else if (result_dst
.type
== GLSL_TYPE_DOUBLE
)
1467 emit_asm(ir
, TGSI_OPCODE_DNEG
, result_dst
, op
[0]);
1469 op
[0].negate
= ~op
[0].negate
;
1474 emit_asm(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1477 emit_asm(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1480 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1484 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1488 assert(!"not reached: should be handled by ir_explog_to_explog2");
1491 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1494 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1497 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1499 case ir_unop_saturate
: {
1500 glsl_to_tgsi_instruction
*inst
;
1501 inst
= emit_asm(ir
, TGSI_OPCODE_MOV
, result_dst
, op
[0]);
1502 inst
->saturate
= true;
1507 case ir_unop_dFdx_coarse
:
1508 emit_asm(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1510 case ir_unop_dFdx_fine
:
1511 emit_asm(ir
, TGSI_OPCODE_DDX_FINE
, result_dst
, op
[0]);
1514 case ir_unop_dFdy_coarse
:
1515 case ir_unop_dFdy_fine
:
1517 /* The X component contains 1 or -1 depending on whether the framebuffer
1518 * is a FBO or the window system buffer, respectively.
1519 * It is then multiplied with the source operand of DDY.
1521 static const gl_state_index transform_y_state
[STATE_LENGTH
]
1522 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
};
1524 unsigned transform_y_index
=
1525 _mesa_add_state_reference(this->prog
->Parameters
,
1528 st_src_reg transform_y
= st_src_reg(PROGRAM_STATE_VAR
,
1530 glsl_type::vec4_type
);
1531 transform_y
.swizzle
= SWIZZLE_XXXX
;
1533 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1535 emit_asm(ir
, TGSI_OPCODE_MUL
, st_dst_reg(temp
), transform_y
, op
[0]);
1536 emit_asm(ir
, ir
->operation
== ir_unop_dFdy_fine
?
1537 TGSI_OPCODE_DDY_FINE
: TGSI_OPCODE_DDY
, result_dst
, temp
);
1541 case ir_unop_frexp_sig
:
1542 emit_asm(ir
, TGSI_OPCODE_DFRACEXP
, result_dst
, undef_dst
, op
[0]);
1545 case ir_unop_frexp_exp
:
1546 emit_asm(ir
, TGSI_OPCODE_DFRACEXP
, undef_dst
, result_dst
, op
[0]);
1549 case ir_unop_noise
: {
1550 /* At some point, a motivated person could add a better
1551 * implementation of noise. Currently not even the nvidia
1552 * binary drivers do anything more than this. In any case, the
1553 * place to do this is in the GL state tracker, not the poor
1556 emit_asm(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1561 emit_asm(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1564 emit_asm(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1568 emit_asm(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1571 if (result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_DOUBLE
)
1572 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1574 emit_asm(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1577 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1578 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1580 emit_asm(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1584 emit_asm(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1586 case ir_binop_greater
:
1587 emit_asm(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1589 case ir_binop_lequal
:
1590 emit_asm(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1592 case ir_binop_gequal
:
1593 emit_asm(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1595 case ir_binop_equal
:
1596 emit_asm(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1598 case ir_binop_nequal
:
1599 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1601 case ir_binop_all_equal
:
1602 /* "==" operator producing a scalar boolean. */
1603 if (ir
->operands
[0]->type
->is_vector() ||
1604 ir
->operands
[1]->type
->is_vector()) {
1605 st_src_reg temp
= get_temp(native_integers
?
1606 glsl_type::uvec4_type
:
1607 glsl_type::vec4_type
);
1609 if (native_integers
) {
1610 st_dst_reg temp_dst
= st_dst_reg(temp
);
1611 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1613 emit_asm(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1615 /* Emit 1-3 AND operations to combine the SEQ results. */
1616 switch (ir
->operands
[0]->type
->vector_elements
) {
1620 temp_dst
.writemask
= WRITEMASK_Y
;
1621 temp1
.swizzle
= SWIZZLE_YYYY
;
1622 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1623 emit_asm(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1626 temp_dst
.writemask
= WRITEMASK_X
;
1627 temp1
.swizzle
= SWIZZLE_XXXX
;
1628 temp2
.swizzle
= SWIZZLE_YYYY
;
1629 emit_asm(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1630 temp_dst
.writemask
= WRITEMASK_Y
;
1631 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1632 temp2
.swizzle
= SWIZZLE_WWWW
;
1633 emit_asm(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1636 temp1
.swizzle
= SWIZZLE_XXXX
;
1637 temp2
.swizzle
= SWIZZLE_YYYY
;
1638 emit_asm(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1640 emit_asm(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1642 /* After the dot-product, the value will be an integer on the
1643 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1645 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1647 /* Negating the result of the dot-product gives values on the range
1648 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1649 * This is achieved using SGE.
1651 st_src_reg sge_src
= result_src
;
1652 sge_src
.negate
= ~sge_src
.negate
;
1653 emit_asm(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1656 emit_asm(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1659 case ir_binop_any_nequal
:
1660 /* "!=" operator producing a scalar boolean. */
1661 if (ir
->operands
[0]->type
->is_vector() ||
1662 ir
->operands
[1]->type
->is_vector()) {
1663 st_src_reg temp
= get_temp(native_integers
?
1664 glsl_type::uvec4_type
:
1665 glsl_type::vec4_type
);
1666 emit_asm(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1668 if (native_integers
) {
1669 st_dst_reg temp_dst
= st_dst_reg(temp
);
1670 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1672 /* Emit 1-3 OR operations to combine the SNE results. */
1673 switch (ir
->operands
[0]->type
->vector_elements
) {
1677 temp_dst
.writemask
= WRITEMASK_Y
;
1678 temp1
.swizzle
= SWIZZLE_YYYY
;
1679 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1680 emit_asm(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1683 temp_dst
.writemask
= WRITEMASK_X
;
1684 temp1
.swizzle
= SWIZZLE_XXXX
;
1685 temp2
.swizzle
= SWIZZLE_YYYY
;
1686 emit_asm(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1687 temp_dst
.writemask
= WRITEMASK_Y
;
1688 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1689 temp2
.swizzle
= SWIZZLE_WWWW
;
1690 emit_asm(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1693 temp1
.swizzle
= SWIZZLE_XXXX
;
1694 temp2
.swizzle
= SWIZZLE_YYYY
;
1695 emit_asm(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1697 /* After the dot-product, the value will be an integer on the
1698 * range [0,4]. Zero stays zero, and positive values become 1.0.
1700 glsl_to_tgsi_instruction
*const dp
=
1701 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1702 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1703 /* The clamping to [0,1] can be done for free in the fragment
1704 * shader with a saturate.
1706 dp
->saturate
= true;
1708 /* Negating the result of the dot-product gives values on the range
1709 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1710 * achieved using SLT.
1712 st_src_reg slt_src
= result_src
;
1713 slt_src
.negate
= ~slt_src
.negate
;
1714 emit_asm(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1718 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1723 assert(ir
->operands
[0]->type
->is_vector());
1725 if (native_integers
) {
1726 int dst_swizzle
= 0, op0_swizzle
, i
;
1727 st_src_reg accum
= op
[0];
1729 op0_swizzle
= op
[0].swizzle
;
1730 accum
.swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 0),
1731 GET_SWZ(op0_swizzle
, 0),
1732 GET_SWZ(op0_swizzle
, 0),
1733 GET_SWZ(op0_swizzle
, 0));
1734 for (i
= 0; i
< 4; i
++) {
1735 if (result_dst
.writemask
& (1 << i
)) {
1736 dst_swizzle
= MAKE_SWIZZLE4(i
, i
, i
, i
);
1741 assert(ir
->operands
[0]->type
->is_boolean());
1743 /* OR all the components together, since they should be either 0 or ~0
1745 switch (ir
->operands
[0]->type
->vector_elements
) {
1747 op
[0].swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 3),
1748 GET_SWZ(op0_swizzle
, 3),
1749 GET_SWZ(op0_swizzle
, 3),
1750 GET_SWZ(op0_swizzle
, 3));
1751 emit_asm(ir
, TGSI_OPCODE_OR
, result_dst
, accum
, op
[0]);
1752 accum
= st_src_reg(result_dst
);
1753 accum
.swizzle
= dst_swizzle
;
1756 op
[0].swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 2),
1757 GET_SWZ(op0_swizzle
, 2),
1758 GET_SWZ(op0_swizzle
, 2),
1759 GET_SWZ(op0_swizzle
, 2));
1760 emit_asm(ir
, TGSI_OPCODE_OR
, result_dst
, accum
, op
[0]);
1761 accum
= st_src_reg(result_dst
);
1762 accum
.swizzle
= dst_swizzle
;
1765 op
[0].swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 1),
1766 GET_SWZ(op0_swizzle
, 1),
1767 GET_SWZ(op0_swizzle
, 1),
1768 GET_SWZ(op0_swizzle
, 1));
1769 emit_asm(ir
, TGSI_OPCODE_OR
, result_dst
, accum
, op
[0]);
1772 assert(!"Unexpected vector size");
1776 /* After the dot-product, the value will be an integer on the
1777 * range [0,4]. Zero stays zero, and positive values become 1.0.
1779 glsl_to_tgsi_instruction
*const dp
=
1780 emit_dp(ir
, result_dst
, op
[0], op
[0],
1781 ir
->operands
[0]->type
->vector_elements
);
1782 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1783 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1784 /* The clamping to [0,1] can be done for free in the fragment
1785 * shader with a saturate.
1787 dp
->saturate
= true;
1788 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1789 /* Negating the result of the dot-product gives values on the range
1790 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1791 * is achieved using SLT.
1793 st_src_reg slt_src
= result_src
;
1794 slt_src
.negate
= ~slt_src
.negate
;
1795 emit_asm(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1798 /* Use SNE 0 if integers are being used as boolean values. */
1799 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1805 case ir_binop_logic_xor
:
1806 if (native_integers
)
1807 emit_asm(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1809 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1812 case ir_binop_logic_or
: {
1813 if (native_integers
) {
1814 /* If integers are used as booleans, we can use an actual "or"
1817 assert(native_integers
);
1818 emit_asm(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1820 /* After the addition, the value will be an integer on the
1821 * range [0,2]. Zero stays zero, and positive values become 1.0.
1823 glsl_to_tgsi_instruction
*add
=
1824 emit_asm(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1825 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1826 /* The clamping to [0,1] can be done for free in the fragment
1827 * shader with a saturate if floats are being used as boolean values.
1829 add
->saturate
= true;
1831 /* Negating the result of the addition gives values on the range
1832 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1833 * is achieved using SLT.
1835 st_src_reg slt_src
= result_src
;
1836 slt_src
.negate
= ~slt_src
.negate
;
1837 emit_asm(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1843 case ir_binop_logic_and
:
1844 /* If native integers are disabled, the bool args are stored as float 0.0
1845 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1846 * actual AND opcode.
1848 if (native_integers
)
1849 emit_asm(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1851 emit_asm(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1855 assert(ir
->operands
[0]->type
->is_vector());
1856 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1857 emit_dp(ir
, result_dst
, op
[0], op
[1],
1858 ir
->operands
[0]->type
->vector_elements
);
1863 emit_scalar(ir
, TGSI_OPCODE_SQRT
, result_dst
, op
[0]);
1865 /* sqrt(x) = x * rsq(x). */
1866 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1867 emit_asm(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1868 /* For incoming channels <= 0, set the result to 0. */
1869 op
[0].negate
= ~op
[0].negate
;
1870 emit_asm(ir
, TGSI_OPCODE_CMP
, result_dst
,
1871 op
[0], result_src
, st_src_reg_for_float(0.0));
1875 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1878 if (native_integers
) {
1879 emit_asm(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1882 /* fallthrough to next case otherwise */
1884 if (native_integers
) {
1885 emit_asm(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1888 /* fallthrough to next case otherwise */
1891 /* Converting between signed and unsigned integers is a no-op. */
1895 if (native_integers
) {
1896 /* Booleans are stored as integers using ~0 for true and 0 for false.
1897 * GLSL requires that int(bool) return 1 for true and 0 for false.
1898 * This conversion is done with AND, but it could be done with NEG.
1900 emit_asm(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1902 /* Booleans and integers are both stored as floats when native
1903 * integers are disabled.
1909 if (native_integers
)
1910 emit_asm(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1912 emit_asm(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1915 if (native_integers
)
1916 emit_asm(ir
, TGSI_OPCODE_F2U
, result_dst
, op
[0]);
1918 emit_asm(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1920 case ir_unop_bitcast_f2i
:
1922 result_src
.type
= GLSL_TYPE_INT
;
1924 case ir_unop_bitcast_f2u
:
1926 result_src
.type
= GLSL_TYPE_UINT
;
1928 case ir_unop_bitcast_i2f
:
1929 case ir_unop_bitcast_u2f
:
1931 result_src
.type
= GLSL_TYPE_FLOAT
;
1934 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1937 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_double(0.0));
1940 if (native_integers
)
1941 emit_asm(ir
, TGSI_OPCODE_USNE
, result_dst
, op
[0], st_src_reg_for_int(0));
1943 emit_asm(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1946 emit_asm(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1949 emit_asm(ir
, TGSI_OPCODE_CEIL
, result_dst
, op
[0]);
1952 emit_asm(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1954 case ir_unop_round_even
:
1955 emit_asm(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1958 emit_asm(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1962 emit_asm(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1965 emit_asm(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1968 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1971 case ir_unop_bit_not
:
1972 if (native_integers
) {
1973 emit_asm(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1977 if (native_integers
) {
1978 emit_asm(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1981 case ir_binop_lshift
:
1982 if (native_integers
) {
1983 emit_asm(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1986 case ir_binop_rshift
:
1987 if (native_integers
) {
1988 emit_asm(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1991 case ir_binop_bit_and
:
1992 if (native_integers
) {
1993 emit_asm(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1996 case ir_binop_bit_xor
:
1997 if (native_integers
) {
1998 emit_asm(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
2001 case ir_binop_bit_or
:
2002 if (native_integers
) {
2003 emit_asm(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
2007 assert(!"GLSL 1.30 features unsupported");
2010 case ir_binop_ubo_load
: {
2011 ir_constant
*const_uniform_block
= ir
->operands
[0]->as_constant();
2012 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
2013 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
2014 unsigned const_block
= const_uniform_block
? const_uniform_block
->value
.u
[0] + 1 : 0;
2015 st_src_reg index_reg
= get_temp(glsl_type::uint_type
);
2018 cbuf
.type
= ir
->type
->base_type
;
2019 cbuf
.file
= PROGRAM_CONSTANT
;
2021 cbuf
.reladdr
= NULL
;
2024 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
2026 if (const_offset_ir
) {
2027 /* Constant index into constant buffer */
2028 cbuf
.reladdr
= NULL
;
2029 cbuf
.index
= const_offset
/ 16;
2032 /* Relative/variable index into constant buffer */
2033 emit_asm(ir
, TGSI_OPCODE_USHR
, st_dst_reg(index_reg
), op
[1],
2034 st_src_reg_for_int(4));
2035 cbuf
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2036 memcpy(cbuf
.reladdr
, &index_reg
, sizeof(index_reg
));
2039 if (const_uniform_block
) {
2040 /* Constant constant buffer */
2041 cbuf
.reladdr2
= NULL
;
2042 cbuf
.index2D
= const_block
;
2043 cbuf
.has_index2
= true;
2046 /* Relative/variable constant buffer */
2047 cbuf
.reladdr2
= ralloc(mem_ctx
, st_src_reg
);
2049 memcpy(cbuf
.reladdr2
, &op
[0], sizeof(st_src_reg
));
2050 cbuf
.has_index2
= true;
2053 cbuf
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2054 if (cbuf
.type
== GLSL_TYPE_DOUBLE
)
2055 cbuf
.swizzle
+= MAKE_SWIZZLE4(const_offset
% 16 / 8,
2056 const_offset
% 16 / 8,
2057 const_offset
% 16 / 8,
2058 const_offset
% 16 / 8);
2060 cbuf
.swizzle
+= MAKE_SWIZZLE4(const_offset
% 16 / 4,
2061 const_offset
% 16 / 4,
2062 const_offset
% 16 / 4,
2063 const_offset
% 16 / 4);
2065 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
2066 emit_asm(ir
, TGSI_OPCODE_USNE
, result_dst
, cbuf
, st_src_reg_for_int(0));
2068 emit_asm(ir
, TGSI_OPCODE_MOV
, result_dst
, cbuf
);
2073 /* note: we have to reorder the three args here */
2074 emit_asm(ir
, TGSI_OPCODE_LRP
, result_dst
, op
[2], op
[1], op
[0]);
2077 if (this->ctx
->Const
.NativeIntegers
)
2078 emit_asm(ir
, TGSI_OPCODE_UCMP
, result_dst
, op
[0], op
[1], op
[2]);
2080 op
[0].negate
= ~op
[0].negate
;
2081 emit_asm(ir
, TGSI_OPCODE_CMP
, result_dst
, op
[0], op
[1], op
[2]);
2084 case ir_triop_bitfield_extract
:
2085 emit_asm(ir
, TGSI_OPCODE_IBFE
, result_dst
, op
[0], op
[1], op
[2]);
2087 case ir_quadop_bitfield_insert
:
2088 emit_asm(ir
, TGSI_OPCODE_BFI
, result_dst
, op
[0], op
[1], op
[2], op
[3]);
2090 case ir_unop_bitfield_reverse
:
2091 emit_asm(ir
, TGSI_OPCODE_BREV
, result_dst
, op
[0]);
2093 case ir_unop_bit_count
:
2094 emit_asm(ir
, TGSI_OPCODE_POPC
, result_dst
, op
[0]);
2096 case ir_unop_find_msb
:
2097 emit_asm(ir
, TGSI_OPCODE_IMSB
, result_dst
, op
[0]);
2099 case ir_unop_find_lsb
:
2100 emit_asm(ir
, TGSI_OPCODE_LSB
, result_dst
, op
[0]);
2102 case ir_binop_imul_high
:
2103 emit_asm(ir
, TGSI_OPCODE_IMUL_HI
, result_dst
, op
[0], op
[1]);
2106 /* In theory, MAD is incorrect here. */
2108 emit_asm(ir
, TGSI_OPCODE_FMA
, result_dst
, op
[0], op
[1], op
[2]);
2110 emit_asm(ir
, TGSI_OPCODE_MAD
, result_dst
, op
[0], op
[1], op
[2]);
2112 case ir_unop_interpolate_at_centroid
:
2113 emit_asm(ir
, TGSI_OPCODE_INTERP_CENTROID
, result_dst
, op
[0]);
2115 case ir_binop_interpolate_at_offset
:
2116 emit_asm(ir
, TGSI_OPCODE_INTERP_OFFSET
, result_dst
, op
[0], op
[1]);
2118 case ir_binop_interpolate_at_sample
:
2119 emit_asm(ir
, TGSI_OPCODE_INTERP_SAMPLE
, result_dst
, op
[0], op
[1]);
2123 emit_asm(ir
, TGSI_OPCODE_D2F
, result_dst
, op
[0]);
2126 emit_asm(ir
, TGSI_OPCODE_F2D
, result_dst
, op
[0]);
2129 emit_asm(ir
, TGSI_OPCODE_D2I
, result_dst
, op
[0]);
2132 emit_asm(ir
, TGSI_OPCODE_I2D
, result_dst
, op
[0]);
2135 emit_asm(ir
, TGSI_OPCODE_D2U
, result_dst
, op
[0]);
2138 emit_asm(ir
, TGSI_OPCODE_U2D
, result_dst
, op
[0]);
2140 case ir_unop_unpack_double_2x32
:
2141 case ir_unop_pack_double_2x32
:
2142 emit_asm(ir
, TGSI_OPCODE_MOV
, result_dst
, op
[0]);
2145 case ir_binop_ldexp
:
2146 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_DOUBLE
) {
2147 emit_asm(ir
, TGSI_OPCODE_DLDEXP
, result_dst
, op
[0], op
[1]);
2149 assert(!"Invalid ldexp for non-double opcode in glsl_to_tgsi_visitor::visit()");
2153 case ir_unop_pack_snorm_2x16
:
2154 case ir_unop_pack_unorm_2x16
:
2155 case ir_unop_pack_half_2x16
:
2156 case ir_unop_pack_snorm_4x8
:
2157 case ir_unop_pack_unorm_4x8
:
2159 case ir_unop_unpack_snorm_2x16
:
2160 case ir_unop_unpack_unorm_2x16
:
2161 case ir_unop_unpack_half_2x16
:
2162 case ir_unop_unpack_half_2x16_split_x
:
2163 case ir_unop_unpack_half_2x16_split_y
:
2164 case ir_unop_unpack_snorm_4x8
:
2165 case ir_unop_unpack_unorm_4x8
:
2167 case ir_binop_pack_half_2x16_split
:
2170 case ir_quadop_vector
:
2171 case ir_binop_vector_extract
:
2172 case ir_triop_vector_insert
:
2173 case ir_binop_carry
:
2174 case ir_binop_borrow
:
2175 /* This operation is not supported, or should have already been handled.
2177 assert(!"Invalid ir opcode in glsl_to_tgsi_visitor::visit()");
2181 this->result
= result_src
;
2186 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
2192 /* Note that this is only swizzles in expressions, not those on the left
2193 * hand side of an assignment, which do write masking. See ir_assignment
2197 ir
->val
->accept(this);
2199 assert(src
.file
!= PROGRAM_UNDEFINED
);
2200 assert(ir
->type
->vector_elements
> 0);
2202 for (i
= 0; i
< 4; i
++) {
2203 if (i
< ir
->type
->vector_elements
) {
2206 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
2209 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
2212 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
2215 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
2219 /* If the type is smaller than a vec4, replicate the last
2222 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
2226 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2231 /* Test if the variable is an array. Note that geometry and
2232 * tessellation shader inputs are outputs are always arrays (except
2233 * for patch inputs), so only the array element type is considered.
2236 is_inout_array(unsigned stage
, ir_variable
*var
, bool *is_2d
)
2238 const glsl_type
*type
= var
->type
;
2240 if ((stage
== MESA_SHADER_VERTEX
&& var
->data
.mode
== ir_var_shader_in
) ||
2241 (stage
== MESA_SHADER_FRAGMENT
&& var
->data
.mode
== ir_var_shader_out
))
2246 if (stage
== MESA_SHADER_GEOMETRY
&& var
->data
.mode
== ir_var_shader_in
) {
2247 if (!var
->type
->is_array())
2248 return false; /* a system value probably */
2250 type
= var
->type
->fields
.array
;
2254 return type
->is_array() || type
->is_matrix();
2258 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
2260 variable_storage
*entry
= find_variable_storage(ir
->var
);
2261 ir_variable
*var
= ir
->var
;
2265 switch (var
->data
.mode
) {
2266 case ir_var_uniform
:
2267 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
2268 var
->data
.location
);
2269 this->variables
.push_tail(entry
);
2271 case ir_var_shader_in
:
2272 /* The linker assigns locations for varyings and attributes,
2273 * including deprecated builtins (like gl_Color), user-assign
2274 * generic attributes (glBindVertexLocation), and
2275 * user-defined varyings.
2277 assert(var
->data
.location
!= -1);
2279 if (is_inout_array(shader
->Stage
, var
, &is_2d
)) {
2280 struct array_decl
*decl
= &input_arrays
[num_input_arrays
];
2282 decl
->mesa_index
= var
->data
.location
;
2283 decl
->array_id
= num_input_arrays
+ 1;
2285 decl
->array_size
= type_size(var
->type
->fields
.array
);
2287 decl
->array_size
= type_size(var
->type
);
2290 entry
= new(mem_ctx
) variable_storage(var
,
2296 entry
= new(mem_ctx
) variable_storage(var
,
2298 var
->data
.location
);
2300 this->variables
.push_tail(entry
);
2302 case ir_var_shader_out
:
2303 assert(var
->data
.location
!= -1);
2305 if (is_inout_array(shader
->Stage
, var
, &is_2d
)) {
2306 struct array_decl
*decl
= &output_arrays
[num_output_arrays
];
2308 decl
->mesa_index
= var
->data
.location
;
2309 decl
->array_id
= num_output_arrays
+ 1;
2311 decl
->array_size
= type_size(var
->type
->fields
.array
);
2313 decl
->array_size
= type_size(var
->type
);
2314 num_output_arrays
++;
2316 entry
= new(mem_ctx
) variable_storage(var
,
2322 entry
= new(mem_ctx
) variable_storage(var
,
2327 this->variables
.push_tail(entry
);
2329 case ir_var_system_value
:
2330 entry
= new(mem_ctx
) variable_storage(var
,
2331 PROGRAM_SYSTEM_VALUE
,
2332 var
->data
.location
);
2335 case ir_var_temporary
:
2336 st_src_reg src
= get_temp(var
->type
);
2338 entry
= new(mem_ctx
) variable_storage(var
, src
.file
, src
.index
);
2339 this->variables
.push_tail(entry
);
2345 printf("Failed to make storage for %s\n", var
->name
);
2350 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
2351 this->result
.array_id
= entry
->array_id
;
2352 if (!native_integers
)
2353 this->result
.type
= GLSL_TYPE_FLOAT
;
2357 shrink_array_declarations(struct array_decl
*arrays
, unsigned count
,
2358 GLbitfield64 usage_mask
)
2362 /* Fix array declarations by removing unused array elements at both ends
2363 * of the arrays. For example, mat4[3] where only mat[1] is used.
2365 for (i
= 0; i
< count
; i
++) {
2366 struct array_decl
*decl
= &arrays
[i
];
2368 /* Shrink the beginning. */
2369 for (j
= 0; j
< decl
->array_size
; j
++) {
2370 if (usage_mask
& BITFIELD64_BIT(decl
->mesa_index
+j
))
2378 /* Shrink the end. */
2379 for (j
= decl
->array_size
-1; j
>= 0; j
--) {
2380 if (usage_mask
& BITFIELD64_BIT(decl
->mesa_index
+j
))
2389 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
2393 int element_size
= type_size(ir
->type
);
2396 index
= ir
->array_index
->constant_expression_value();
2398 ir
->array
->accept(this);
2401 is_2D_input
= this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
&&
2402 src
.file
== PROGRAM_INPUT
&&
2403 ir
->array
->ir_type
!= ir_type_dereference_array
;
2410 src
.index2D
= index
->value
.i
[0];
2411 src
.has_index2
= true;
2413 src
.index
+= index
->value
.i
[0] * element_size
;
2415 /* Variable index array dereference. It eats the "vec4" of the
2416 * base of the array and an index that offsets the TGSI register
2419 ir
->array_index
->accept(this);
2421 st_src_reg index_reg
;
2423 if (element_size
== 1) {
2424 index_reg
= this->result
;
2426 index_reg
= get_temp(native_integers
?
2427 glsl_type::int_type
: glsl_type::float_type
);
2429 emit_asm(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2430 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2433 /* If there was already a relative address register involved, add the
2434 * new and the old together to get the new offset.
2436 if (!is_2D_input
&& src
.reladdr
!= NULL
) {
2437 st_src_reg accum_reg
= get_temp(native_integers
?
2438 glsl_type::int_type
: glsl_type::float_type
);
2440 emit_asm(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2441 index_reg
, *src
.reladdr
);
2443 index_reg
= accum_reg
;
2447 src
.reladdr2
= ralloc(mem_ctx
, st_src_reg
);
2448 memcpy(src
.reladdr2
, &index_reg
, sizeof(index_reg
));
2450 src
.has_index2
= true;
2452 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2453 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2457 /* If the type is smaller than a vec4, replicate the last channel out. */
2458 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2459 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2461 src
.swizzle
= SWIZZLE_NOOP
;
2463 /* Change the register type to the element type of the array. */
2464 src
.type
= ir
->type
->base_type
;
2470 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2473 const glsl_type
*struct_type
= ir
->record
->type
;
2476 ir
->record
->accept(this);
2478 for (i
= 0; i
< struct_type
->length
; i
++) {
2479 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2481 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2484 /* If the type is smaller than a vec4, replicate the last channel out. */
2485 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2486 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2488 this->result
.swizzle
= SWIZZLE_NOOP
;
2490 this->result
.index
+= offset
;
2491 this->result
.type
= ir
->type
->base_type
;
2495 * We want to be careful in assignment setup to hit the actual storage
2496 * instead of potentially using a temporary like we might with the
2497 * ir_dereference handler.
2500 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2502 /* The LHS must be a dereference. If the LHS is a variable indexed array
2503 * access of a vector, it must be separated into a series conditional moves
2504 * before reaching this point (see ir_vec_index_to_cond_assign).
2506 assert(ir
->as_dereference());
2507 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2509 assert(!deref_array
->array
->type
->is_vector());
2512 /* Use the rvalue deref handler for the most part. We'll ignore
2513 * swizzles in it and write swizzles using writemask, though.
2516 return st_dst_reg(v
->result
);
2520 * Process the condition of a conditional assignment
2522 * Examines the condition of a conditional assignment to generate the optimal
2523 * first operand of a \c CMP instruction. If the condition is a relational
2524 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2525 * used as the source for the \c CMP instruction. Otherwise the comparison
2526 * is processed to a boolean result, and the boolean result is used as the
2527 * operand to the CMP instruction.
2530 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2532 ir_rvalue
*src_ir
= ir
;
2534 bool switch_order
= false;
2536 ir_expression
*const expr
= ir
->as_expression();
2538 if (native_integers
) {
2539 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2540 enum glsl_base_type type
= expr
->operands
[0]->type
->base_type
;
2541 if (type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_UINT
||
2542 type
== GLSL_TYPE_BOOL
) {
2543 if (expr
->operation
== ir_binop_equal
) {
2544 if (expr
->operands
[0]->is_zero()) {
2545 src_ir
= expr
->operands
[1];
2546 switch_order
= true;
2548 else if (expr
->operands
[1]->is_zero()) {
2549 src_ir
= expr
->operands
[0];
2550 switch_order
= true;
2553 else if (expr
->operation
== ir_binop_nequal
) {
2554 if (expr
->operands
[0]->is_zero()) {
2555 src_ir
= expr
->operands
[1];
2557 else if (expr
->operands
[1]->is_zero()) {
2558 src_ir
= expr
->operands
[0];
2564 src_ir
->accept(this);
2565 return switch_order
;
2568 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2569 bool zero_on_left
= false;
2571 if (expr
->operands
[0]->is_zero()) {
2572 src_ir
= expr
->operands
[1];
2573 zero_on_left
= true;
2574 } else if (expr
->operands
[1]->is_zero()) {
2575 src_ir
= expr
->operands
[0];
2576 zero_on_left
= false;
2580 * (a < 0) T F F ( a < 0) T F F
2581 * (0 < a) F F T (-a < 0) F F T
2582 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2583 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2584 * (a > 0) F F T (-a < 0) F F T
2585 * (0 > a) T F F ( a < 0) T F F
2586 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2587 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2589 * Note that exchanging the order of 0 and 'a' in the comparison simply
2590 * means that the value of 'a' should be negated.
2593 switch (expr
->operation
) {
2595 switch_order
= false;
2596 negate
= zero_on_left
;
2599 case ir_binop_greater
:
2600 switch_order
= false;
2601 negate
= !zero_on_left
;
2604 case ir_binop_lequal
:
2605 switch_order
= true;
2606 negate
= !zero_on_left
;
2609 case ir_binop_gequal
:
2610 switch_order
= true;
2611 negate
= zero_on_left
;
2615 /* This isn't the right kind of comparison afterall, so make sure
2616 * the whole condition is visited.
2624 src_ir
->accept(this);
2626 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2627 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2628 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2629 * computing the condition.
2632 this->result
.negate
= ~this->result
.negate
;
2634 return switch_order
;
2638 glsl_to_tgsi_visitor::emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
2639 st_dst_reg
*l
, st_src_reg
*r
,
2640 st_src_reg
*cond
, bool cond_swap
)
2642 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2643 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2644 emit_block_mov(ir
, type
->fields
.structure
[i
].type
, l
, r
,
2650 if (type
->is_array()) {
2651 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2652 emit_block_mov(ir
, type
->fields
.array
, l
, r
, cond
, cond_swap
);
2657 if (type
->is_matrix()) {
2658 const struct glsl_type
*vec_type
;
2660 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2661 type
->vector_elements
, 1);
2663 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2664 emit_block_mov(ir
, vec_type
, l
, r
, cond
, cond_swap
);
2669 assert(type
->is_scalar() || type
->is_vector());
2671 r
->type
= type
->base_type
;
2673 st_src_reg l_src
= st_src_reg(*l
);
2674 l_src
.swizzle
= swizzle_for_size(type
->vector_elements
);
2676 if (native_integers
) {
2677 emit_asm(ir
, TGSI_OPCODE_UCMP
, *l
, *cond
,
2678 cond_swap
? l_src
: *r
,
2679 cond_swap
? *r
: l_src
);
2681 emit_asm(ir
, TGSI_OPCODE_CMP
, *l
, *cond
,
2682 cond_swap
? l_src
: *r
,
2683 cond_swap
? *r
: l_src
);
2686 emit_asm(ir
, TGSI_OPCODE_MOV
, *l
, *r
);
2693 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2698 ir
->rhs
->accept(this);
2701 l
= get_assignment_lhs(ir
->lhs
, this);
2703 /* FINISHME: This should really set to the correct maximal writemask for each
2704 * FINISHME: component written (in the loops below). This case can only
2705 * FINISHME: occur for matrices, arrays, and structures.
2707 if (ir
->write_mask
== 0) {
2708 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2709 l
.writemask
= WRITEMASK_XYZW
;
2710 } else if (ir
->lhs
->type
->is_scalar() &&
2711 !ir
->lhs
->type
->is_double() &&
2712 ir
->lhs
->variable_referenced()->data
.mode
== ir_var_shader_out
) {
2713 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2714 * FINISHME: W component of fragment shader output zero, work correctly.
2716 l
.writemask
= WRITEMASK_XYZW
;
2719 int first_enabled_chan
= 0;
2722 l
.writemask
= ir
->write_mask
;
2724 for (int i
= 0; i
< 4; i
++) {
2725 if (l
.writemask
& (1 << i
)) {
2726 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2731 /* Swizzle a small RHS vector into the channels being written.
2733 * glsl ir treats write_mask as dictating how many channels are
2734 * present on the RHS while TGSI treats write_mask as just
2735 * showing which channels of the vec4 RHS get written.
2737 for (int i
= 0; i
< 4; i
++) {
2738 if (l
.writemask
& (1 << i
))
2739 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2741 swizzles
[i
] = first_enabled_chan
;
2743 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2744 swizzles
[2], swizzles
[3]);
2747 assert(l
.file
!= PROGRAM_UNDEFINED
);
2748 assert(r
.file
!= PROGRAM_UNDEFINED
);
2750 if (ir
->condition
) {
2751 const bool switch_order
= this->process_move_condition(ir
->condition
);
2752 st_src_reg condition
= this->result
;
2754 emit_block_mov(ir
, ir
->lhs
->type
, &l
, &r
, &condition
, switch_order
);
2755 } else if (ir
->rhs
->as_expression() &&
2756 this->instructions
.get_tail() &&
2757 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2758 type_size(ir
->lhs
->type
) == 1 &&
2759 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
[0].writemask
) {
2760 /* To avoid emitting an extra MOV when assigning an expression to a
2761 * variable, emit the last instruction of the expression again, but
2762 * replace the destination register with the target of the assignment.
2763 * Dead code elimination will remove the original instruction.
2765 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2766 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2767 new_inst
= emit_asm(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2768 new_inst
->saturate
= inst
->saturate
;
2769 inst
->dead_mask
= inst
->dst
[0].writemask
;
2771 emit_block_mov(ir
, ir
->rhs
->type
, &l
, &r
, NULL
, false);
2777 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2780 GLdouble stack_vals
[4] = { 0 };
2781 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2782 GLenum gl_type
= GL_NONE
;
2784 static int in_array
= 0;
2785 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2787 /* Unfortunately, 4 floats is all we can get into
2788 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2789 * aggregate constant and move each constant value into it. If we
2790 * get lucky, copy propagation will eliminate the extra moves.
2792 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2793 st_src_reg temp_base
= get_temp(ir
->type
);
2794 st_dst_reg temp
= st_dst_reg(temp_base
);
2796 foreach_in_list(ir_constant
, field_value
, &ir
->components
) {
2797 int size
= type_size(field_value
->type
);
2801 field_value
->accept(this);
2804 for (i
= 0; i
< (unsigned int)size
; i
++) {
2805 emit_asm(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2811 this->result
= temp_base
;
2815 if (ir
->type
->is_array()) {
2816 st_src_reg temp_base
= get_temp(ir
->type
);
2817 st_dst_reg temp
= st_dst_reg(temp_base
);
2818 int size
= type_size(ir
->type
->fields
.array
);
2823 for (i
= 0; i
< ir
->type
->length
; i
++) {
2824 ir
->array_elements
[i
]->accept(this);
2826 for (int j
= 0; j
< size
; j
++) {
2827 emit_asm(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2833 this->result
= temp_base
;
2838 if (ir
->type
->is_matrix()) {
2839 st_src_reg mat
= get_temp(ir
->type
);
2840 st_dst_reg mat_column
= st_dst_reg(mat
);
2842 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2843 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2844 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2846 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2847 src
.index
= add_constant(file
,
2849 ir
->type
->vector_elements
,
2852 emit_asm(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2861 switch (ir
->type
->base_type
) {
2862 case GLSL_TYPE_FLOAT
:
2864 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2865 values
[i
].f
= ir
->value
.f
[i
];
2868 case GLSL_TYPE_DOUBLE
:
2869 gl_type
= GL_DOUBLE
;
2870 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2871 values
[i
* 2].i
= *(uint32_t *)&ir
->value
.d
[i
];
2872 values
[i
* 2 + 1].i
= *(((uint32_t *)&ir
->value
.d
[i
]) + 1);
2875 case GLSL_TYPE_UINT
:
2876 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2877 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2878 if (native_integers
)
2879 values
[i
].u
= ir
->value
.u
[i
];
2881 values
[i
].f
= ir
->value
.u
[i
];
2885 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2886 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2887 if (native_integers
)
2888 values
[i
].i
= ir
->value
.i
[i
];
2890 values
[i
].f
= ir
->value
.i
[i
];
2893 case GLSL_TYPE_BOOL
:
2894 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2895 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2896 values
[i
].u
= ir
->value
.b
[i
] ? ctx
->Const
.UniformBooleanTrue
: 0;
2900 assert(!"Non-float/uint/int/bool constant");
2903 this->result
= st_src_reg(file
, -1, ir
->type
);
2904 this->result
.index
= add_constant(file
,
2906 ir
->type
->vector_elements
,
2908 &this->result
.swizzle
);
2912 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2914 foreach_in_list_use_after(function_entry
, entry
, &this->function_signatures
) {
2915 if (entry
->sig
== sig
)
2919 entry
= ralloc(mem_ctx
, function_entry
);
2921 entry
->sig_id
= this->next_signature_id
++;
2922 entry
->bgn_inst
= NULL
;
2924 /* Allocate storage for all the parameters. */
2925 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2926 variable_storage
*storage
;
2928 storage
= find_variable_storage(param
);
2931 st_src_reg src
= get_temp(param
->type
);
2933 storage
= new(mem_ctx
) variable_storage(param
, src
.file
, src
.index
);
2934 this->variables
.push_tail(storage
);
2937 if (!sig
->return_type
->is_void()) {
2938 entry
->return_reg
= get_temp(sig
->return_type
);
2940 entry
->return_reg
= undef_src
;
2943 this->function_signatures
.push_tail(entry
);
2948 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2950 glsl_to_tgsi_instruction
*call_inst
;
2951 ir_function_signature
*sig
= ir
->callee
;
2952 function_entry
*entry
= get_function_signature(sig
);
2955 /* Process in parameters. */
2956 foreach_two_lists(formal_node
, &sig
->parameters
,
2957 actual_node
, &ir
->actual_parameters
) {
2958 ir_rvalue
*param_rval
= (ir_rvalue
*) actual_node
;
2959 ir_variable
*param
= (ir_variable
*) formal_node
;
2961 if (param
->data
.mode
== ir_var_function_in
||
2962 param
->data
.mode
== ir_var_function_inout
) {
2963 variable_storage
*storage
= find_variable_storage(param
);
2966 param_rval
->accept(this);
2967 st_src_reg r
= this->result
;
2970 l
.file
= storage
->file
;
2971 l
.index
= storage
->index
;
2973 l
.writemask
= WRITEMASK_XYZW
;
2974 l
.cond_mask
= COND_TR
;
2976 for (i
= 0; i
< type_size(param
->type
); i
++) {
2977 emit_asm(ir
, TGSI_OPCODE_MOV
, l
, r
);
2984 /* Emit call instruction */
2985 call_inst
= emit_asm(ir
, TGSI_OPCODE_CAL
);
2986 call_inst
->function
= entry
;
2988 /* Process out parameters. */
2989 foreach_two_lists(formal_node
, &sig
->parameters
,
2990 actual_node
, &ir
->actual_parameters
) {
2991 ir_rvalue
*param_rval
= (ir_rvalue
*) actual_node
;
2992 ir_variable
*param
= (ir_variable
*) formal_node
;
2994 if (param
->data
.mode
== ir_var_function_out
||
2995 param
->data
.mode
== ir_var_function_inout
) {
2996 variable_storage
*storage
= find_variable_storage(param
);
3000 r
.file
= storage
->file
;
3001 r
.index
= storage
->index
;
3003 r
.swizzle
= SWIZZLE_NOOP
;
3006 param_rval
->accept(this);
3007 st_dst_reg l
= st_dst_reg(this->result
);
3009 for (i
= 0; i
< type_size(param
->type
); i
++) {
3010 emit_asm(ir
, TGSI_OPCODE_MOV
, l
, r
);
3017 /* Process return value. */
3018 this->result
= entry
->return_reg
;
3022 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
3024 st_src_reg result_src
, coord
, cube_sc
, lod_info
, projector
, dx
, dy
;
3025 st_src_reg offset
[MAX_GLSL_TEXTURE_OFFSET
], sample_index
, component
;
3026 st_src_reg levels_src
;
3027 st_dst_reg result_dst
, coord_dst
, cube_sc_dst
;
3028 glsl_to_tgsi_instruction
*inst
= NULL
;
3029 unsigned opcode
= TGSI_OPCODE_NOP
;
3030 const glsl_type
*sampler_type
= ir
->sampler
->type
;
3031 ir_rvalue
*sampler_index
=
3032 _mesa_get_sampler_array_nonconst_index(ir
->sampler
);
3033 bool is_cube_array
= false;
3036 /* if we are a cube array sampler */
3037 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
3038 sampler_type
->sampler_array
)) {
3039 is_cube_array
= true;
3042 if (ir
->coordinate
) {
3043 ir
->coordinate
->accept(this);
3045 /* Put our coords in a temp. We'll need to modify them for shadow,
3046 * projection, or LOD, so the only case we'd use it as is is if
3047 * we're doing plain old texturing. The optimization passes on
3048 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
3050 coord
= get_temp(glsl_type::vec4_type
);
3051 coord_dst
= st_dst_reg(coord
);
3052 coord_dst
.writemask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
3053 emit_asm(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
3056 if (ir
->projector
) {
3057 ir
->projector
->accept(this);
3058 projector
= this->result
;
3061 /* Storage for our result. Ideally for an assignment we'd be using
3062 * the actual storage for the result here, instead.
3064 result_src
= get_temp(ir
->type
);
3065 result_dst
= st_dst_reg(result_src
);
3069 opcode
= (is_cube_array
&& ir
->shadow_comparitor
) ? TGSI_OPCODE_TEX2
: TGSI_OPCODE_TEX
;
3071 ir
->offset
->accept(this);
3072 offset
[0] = this->result
;
3076 if (is_cube_array
||
3077 sampler_type
== glsl_type::samplerCubeShadow_type
) {
3078 opcode
= TGSI_OPCODE_TXB2
;
3081 opcode
= TGSI_OPCODE_TXB
;
3083 ir
->lod_info
.bias
->accept(this);
3084 lod_info
= this->result
;
3086 ir
->offset
->accept(this);
3087 offset
[0] = this->result
;
3091 opcode
= is_cube_array
? TGSI_OPCODE_TXL2
: TGSI_OPCODE_TXL
;
3092 ir
->lod_info
.lod
->accept(this);
3093 lod_info
= this->result
;
3095 ir
->offset
->accept(this);
3096 offset
[0] = this->result
;
3100 opcode
= TGSI_OPCODE_TXD
;
3101 ir
->lod_info
.grad
.dPdx
->accept(this);
3103 ir
->lod_info
.grad
.dPdy
->accept(this);
3106 ir
->offset
->accept(this);
3107 offset
[0] = this->result
;
3111 opcode
= TGSI_OPCODE_TXQ
;
3112 ir
->lod_info
.lod
->accept(this);
3113 lod_info
= this->result
;
3115 case ir_query_levels
:
3116 opcode
= TGSI_OPCODE_TXQ
;
3117 lod_info
= undef_src
;
3118 levels_src
= get_temp(ir
->type
);
3121 opcode
= TGSI_OPCODE_TXF
;
3122 ir
->lod_info
.lod
->accept(this);
3123 lod_info
= this->result
;
3125 ir
->offset
->accept(this);
3126 offset
[0] = this->result
;
3130 opcode
= TGSI_OPCODE_TXF
;
3131 ir
->lod_info
.sample_index
->accept(this);
3132 sample_index
= this->result
;
3135 opcode
= TGSI_OPCODE_TG4
;
3136 ir
->lod_info
.component
->accept(this);
3137 component
= this->result
;
3139 ir
->offset
->accept(this);
3140 if (ir
->offset
->type
->base_type
== GLSL_TYPE_ARRAY
) {
3141 const glsl_type
*elt_type
= ir
->offset
->type
->fields
.array
;
3142 for (i
= 0; i
< ir
->offset
->type
->length
; i
++) {
3143 offset
[i
] = this->result
;
3144 offset
[i
].index
+= i
* type_size(elt_type
);
3145 offset
[i
].type
= elt_type
->base_type
;
3146 offset
[i
].swizzle
= swizzle_for_size(elt_type
->vector_elements
);
3149 offset
[0] = this->result
;
3154 opcode
= TGSI_OPCODE_LODQ
;
3158 if (ir
->projector
) {
3159 if (opcode
== TGSI_OPCODE_TEX
) {
3160 /* Slot the projector in as the last component of the coord. */
3161 coord_dst
.writemask
= WRITEMASK_W
;
3162 emit_asm(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
3163 coord_dst
.writemask
= WRITEMASK_XYZW
;
3164 opcode
= TGSI_OPCODE_TXP
;
3166 st_src_reg coord_w
= coord
;
3167 coord_w
.swizzle
= SWIZZLE_WWWW
;
3169 /* For the other TEX opcodes there's no projective version
3170 * since the last slot is taken up by LOD info. Do the
3171 * projective divide now.
3173 coord_dst
.writemask
= WRITEMASK_W
;
3174 emit_asm(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
3176 /* In the case where we have to project the coordinates "by hand,"
3177 * the shadow comparator value must also be projected.
3179 st_src_reg tmp_src
= coord
;
3180 if (ir
->shadow_comparitor
) {
3181 /* Slot the shadow value in as the second to last component of the
3184 ir
->shadow_comparitor
->accept(this);
3186 tmp_src
= get_temp(glsl_type::vec4_type
);
3187 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
3189 /* Projective division not allowed for array samplers. */
3190 assert(!sampler_type
->sampler_array
);
3192 tmp_dst
.writemask
= WRITEMASK_Z
;
3193 emit_asm(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
3195 tmp_dst
.writemask
= WRITEMASK_XY
;
3196 emit_asm(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
3199 coord_dst
.writemask
= WRITEMASK_XYZ
;
3200 emit_asm(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
3202 coord_dst
.writemask
= WRITEMASK_XYZW
;
3203 coord
.swizzle
= SWIZZLE_XYZW
;
3207 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
3208 * comparator was put in the correct place (and projected) by the code,
3209 * above, that handles by-hand projection.
3211 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
3212 /* Slot the shadow value in as the second to last component of the
3215 ir
->shadow_comparitor
->accept(this);
3217 if (is_cube_array
) {
3218 cube_sc
= get_temp(glsl_type::float_type
);
3219 cube_sc_dst
= st_dst_reg(cube_sc
);
3220 cube_sc_dst
.writemask
= WRITEMASK_X
;
3221 emit_asm(ir
, TGSI_OPCODE_MOV
, cube_sc_dst
, this->result
);
3222 cube_sc_dst
.writemask
= WRITEMASK_X
;
3225 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
3226 sampler_type
->sampler_array
) ||
3227 sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
) {
3228 coord_dst
.writemask
= WRITEMASK_W
;
3230 coord_dst
.writemask
= WRITEMASK_Z
;
3232 emit_asm(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
3233 coord_dst
.writemask
= WRITEMASK_XYZW
;
3237 if (ir
->op
== ir_txf_ms
) {
3238 coord_dst
.writemask
= WRITEMASK_W
;
3239 emit_asm(ir
, TGSI_OPCODE_MOV
, coord_dst
, sample_index
);
3240 coord_dst
.writemask
= WRITEMASK_XYZW
;
3241 } else if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
3242 opcode
== TGSI_OPCODE_TXF
) {
3243 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
3244 coord_dst
.writemask
= WRITEMASK_W
;
3245 emit_asm(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
3246 coord_dst
.writemask
= WRITEMASK_XYZW
;
3249 if (sampler_index
) {
3250 sampler_index
->accept(this);
3251 emit_arl(ir
, sampler_reladdr
, this->result
);
3254 if (opcode
== TGSI_OPCODE_TXD
)
3255 inst
= emit_asm(ir
, opcode
, result_dst
, coord
, dx
, dy
);
3256 else if (opcode
== TGSI_OPCODE_TXQ
) {
3257 if (ir
->op
== ir_query_levels
) {
3258 /* the level is stored in W */
3259 inst
= emit_asm(ir
, opcode
, st_dst_reg(levels_src
), lod_info
);
3260 result_dst
.writemask
= WRITEMASK_X
;
3261 levels_src
.swizzle
= SWIZZLE_WWWW
;
3262 emit_asm(ir
, TGSI_OPCODE_MOV
, result_dst
, levels_src
);
3264 inst
= emit_asm(ir
, opcode
, result_dst
, lod_info
);
3265 } else if (opcode
== TGSI_OPCODE_TXF
) {
3266 inst
= emit_asm(ir
, opcode
, result_dst
, coord
);
3267 } else if (opcode
== TGSI_OPCODE_TXL2
|| opcode
== TGSI_OPCODE_TXB2
) {
3268 inst
= emit_asm(ir
, opcode
, result_dst
, coord
, lod_info
);
3269 } else if (opcode
== TGSI_OPCODE_TEX2
) {
3270 inst
= emit_asm(ir
, opcode
, result_dst
, coord
, cube_sc
);
3271 } else if (opcode
== TGSI_OPCODE_TG4
) {
3272 if (is_cube_array
&& ir
->shadow_comparitor
) {
3273 inst
= emit_asm(ir
, opcode
, result_dst
, coord
, cube_sc
);
3275 inst
= emit_asm(ir
, opcode
, result_dst
, coord
, component
);
3278 inst
= emit_asm(ir
, opcode
, result_dst
, coord
);
3280 if (ir
->shadow_comparitor
)
3281 inst
->tex_shadow
= GL_TRUE
;
3283 inst
->sampler
.index
= _mesa_get_sampler_uniform_value(ir
->sampler
,
3284 this->shader_program
,
3286 if (sampler_index
) {
3287 inst
->sampler
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
3288 memcpy(inst
->sampler
.reladdr
, &sampler_reladdr
, sizeof(sampler_reladdr
));
3289 inst
->sampler_array_size
=
3290 ir
->sampler
->as_dereference_array()->array
->type
->array_size();
3292 inst
->sampler_array_size
= 1;
3296 for (i
= 0; i
< MAX_GLSL_TEXTURE_OFFSET
&& offset
[i
].file
!= PROGRAM_UNDEFINED
; i
++)
3297 inst
->tex_offsets
[i
] = offset
[i
];
3298 inst
->tex_offset_num_offset
= i
;
3301 switch (sampler_type
->sampler_dimensionality
) {
3302 case GLSL_SAMPLER_DIM_1D
:
3303 inst
->tex_target
= (sampler_type
->sampler_array
)
3304 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
3306 case GLSL_SAMPLER_DIM_2D
:
3307 inst
->tex_target
= (sampler_type
->sampler_array
)
3308 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
3310 case GLSL_SAMPLER_DIM_3D
:
3311 inst
->tex_target
= TEXTURE_3D_INDEX
;
3313 case GLSL_SAMPLER_DIM_CUBE
:
3314 inst
->tex_target
= (sampler_type
->sampler_array
)
3315 ? TEXTURE_CUBE_ARRAY_INDEX
: TEXTURE_CUBE_INDEX
;
3317 case GLSL_SAMPLER_DIM_RECT
:
3318 inst
->tex_target
= TEXTURE_RECT_INDEX
;
3320 case GLSL_SAMPLER_DIM_BUF
:
3321 inst
->tex_target
= TEXTURE_BUFFER_INDEX
;
3323 case GLSL_SAMPLER_DIM_EXTERNAL
:
3324 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
3326 case GLSL_SAMPLER_DIM_MS
:
3327 inst
->tex_target
= (sampler_type
->sampler_array
)
3328 ? TEXTURE_2D_MULTISAMPLE_ARRAY_INDEX
: TEXTURE_2D_MULTISAMPLE_INDEX
;
3331 assert(!"Should not get here.");
3334 inst
->tex_type
= ir
->type
->base_type
;
3336 this->result
= result_src
;
3340 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
3342 if (ir
->get_value()) {
3346 assert(current_function
);
3348 ir
->get_value()->accept(this);
3349 st_src_reg r
= this->result
;
3351 l
= st_dst_reg(current_function
->return_reg
);
3353 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
3354 emit_asm(ir
, TGSI_OPCODE_MOV
, l
, r
);
3360 emit_asm(ir
, TGSI_OPCODE_RET
);
3364 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
3366 if (ir
->condition
) {
3367 ir
->condition
->accept(this);
3368 st_src_reg condition
= this->result
;
3370 /* Convert the bool condition to a float so we can negate. */
3371 if (native_integers
) {
3372 st_src_reg temp
= get_temp(ir
->condition
->type
);
3373 emit_asm(ir
, TGSI_OPCODE_AND
, st_dst_reg(temp
),
3374 condition
, st_src_reg_for_float(1.0));
3378 condition
.negate
= ~condition
.negate
;
3379 emit_asm(ir
, TGSI_OPCODE_KILL_IF
, undef_dst
, condition
);
3381 /* unconditional kil */
3382 emit_asm(ir
, TGSI_OPCODE_KILL
);
3387 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
3390 glsl_to_tgsi_instruction
*if_inst
;
3392 ir
->condition
->accept(this);
3393 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
3395 if_opcode
= native_integers
? TGSI_OPCODE_UIF
: TGSI_OPCODE_IF
;
3397 if_inst
= emit_asm(ir
->condition
, if_opcode
, undef_dst
, this->result
);
3399 this->instructions
.push_tail(if_inst
);
3401 visit_exec_list(&ir
->then_instructions
, this);
3403 if (!ir
->else_instructions
.is_empty()) {
3404 emit_asm(ir
->condition
, TGSI_OPCODE_ELSE
);
3405 visit_exec_list(&ir
->else_instructions
, this);
3408 if_inst
= emit_asm(ir
->condition
, TGSI_OPCODE_ENDIF
);
3413 glsl_to_tgsi_visitor::visit(ir_emit_vertex
*ir
)
3415 assert(this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
);
3417 ir
->stream
->accept(this);
3418 emit_asm(ir
, TGSI_OPCODE_EMIT
, undef_dst
, this->result
);
3422 glsl_to_tgsi_visitor::visit(ir_end_primitive
*ir
)
3424 assert(this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
);
3426 ir
->stream
->accept(this);
3427 emit_asm(ir
, TGSI_OPCODE_ENDPRIM
, undef_dst
, this->result
);
3431 glsl_to_tgsi_visitor::visit(ir_barrier
*ir
)
3433 unreachable("Not implemented!");
3436 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
3438 result
.file
= PROGRAM_UNDEFINED
;
3443 num_input_arrays
= 0;
3444 num_output_arrays
= 0;
3445 next_signature_id
= 1;
3447 current_function
= NULL
;
3448 num_address_regs
= 0;
3450 indirect_addr_consts
= false;
3451 wpos_transform_const
= -1;
3453 native_integers
= false;
3454 mem_ctx
= ralloc_context(NULL
);
3457 shader_program
= NULL
;
3464 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
3467 ralloc_free(mem_ctx
);
3470 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
3477 * Count resources used by the given gpu program (number of texture
3481 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
3483 v
->samplers_used
= 0;
3485 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &v
->instructions
) {
3486 if (is_tex_instruction(inst
->op
)) {
3487 for (int i
= 0; i
< inst
->sampler_array_size
; i
++) {
3488 unsigned idx
= inst
->sampler
.index
+ i
;
3489 v
->samplers_used
|= 1 << idx
;
3491 debug_assert(idx
< (int)ARRAY_SIZE(v
->sampler_types
));
3492 v
->sampler_types
[idx
] = inst
->tex_type
;
3493 v
->sampler_targets
[idx
] =
3494 st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
);
3496 if (inst
->tex_shadow
) {
3497 prog
->ShadowSamplers
|= 1 << (inst
->sampler
.index
+ i
);
3502 prog
->SamplersUsed
= v
->samplers_used
;
3504 if (v
->shader_program
!= NULL
)
3505 _mesa_update_shader_textures_used(v
->shader_program
, prog
);
3509 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3510 * are read from the given src in this instruction
3513 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3515 int read_mask
= 0, comp
;
3517 /* Now, given the src swizzle and the written channels, find which
3518 * components are actually read
3520 for (comp
= 0; comp
< 4; ++comp
) {
3521 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3523 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3524 read_mask
|= 1 << coord
;
3531 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3532 * instruction is the first instruction to write to register T0. There are
3533 * several lowering passes done in GLSL IR (e.g. branches and
3534 * relative addressing) that create a large number of conditional assignments
3535 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3537 * Here is why this conversion is safe:
3538 * CMP T0, T1 T2 T0 can be expanded to:
3544 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3545 * as the original program. If (T1 < 0.0) evaluates to false, executing
3546 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3547 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3548 * because any instruction that was going to read from T0 after this was going
3549 * to read a garbage value anyway.
3552 glsl_to_tgsi_visitor::simplify_cmp(void)
3554 int tempWritesSize
= 0;
3555 unsigned *tempWrites
= NULL
;
3556 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3558 memset(outputWrites
, 0, sizeof(outputWrites
));
3560 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3561 unsigned prevWriteMask
= 0;
3563 /* Give up if we encounter relative addressing or flow control. */
3564 if (inst
->dst
[0].reladdr
||
3565 inst
->dst
[1].reladdr
||
3566 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3567 inst
->op
== TGSI_OPCODE_BGNSUB
||
3568 inst
->op
== TGSI_OPCODE_CONT
||
3569 inst
->op
== TGSI_OPCODE_END
||
3570 inst
->op
== TGSI_OPCODE_ENDSUB
||
3571 inst
->op
== TGSI_OPCODE_RET
) {
3575 if (inst
->dst
[0].file
== PROGRAM_OUTPUT
) {
3576 assert(inst
->dst
[0].index
< MAX_PROGRAM_OUTPUTS
);
3577 prevWriteMask
= outputWrites
[inst
->dst
[0].index
];
3578 outputWrites
[inst
->dst
[0].index
] |= inst
->dst
[0].writemask
;
3579 } else if (inst
->dst
[0].file
== PROGRAM_TEMPORARY
) {
3580 if (inst
->dst
[0].index
>= tempWritesSize
) {
3581 const int inc
= 4096;
3583 tempWrites
= (unsigned*)
3585 (tempWritesSize
+ inc
) * sizeof(unsigned));
3589 memset(tempWrites
+ tempWritesSize
, 0, inc
* sizeof(unsigned));
3590 tempWritesSize
+= inc
;
3593 prevWriteMask
= tempWrites
[inst
->dst
[0].index
];
3594 tempWrites
[inst
->dst
[0].index
] |= inst
->dst
[0].writemask
;
3598 /* For a CMP to be considered a conditional write, the destination
3599 * register and source register two must be the same. */
3600 if (inst
->op
== TGSI_OPCODE_CMP
3601 && !(inst
->dst
[0].writemask
& prevWriteMask
)
3602 && inst
->src
[2].file
== inst
->dst
[0].file
3603 && inst
->src
[2].index
== inst
->dst
[0].index
3604 && inst
->dst
[0].writemask
== get_src_arg_mask(inst
->dst
[0], inst
->src
[2])) {
3606 inst
->op
= TGSI_OPCODE_MOV
;
3607 inst
->src
[0] = inst
->src
[1];
3614 /* Replaces all references to a temporary register index with another index. */
3616 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3618 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3621 for (j
= 0; j
< num_inst_src_regs(inst
->op
); j
++) {
3622 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3623 inst
->src
[j
].index
== index
) {
3624 inst
->src
[j
].index
= new_index
;
3628 for (j
= 0; j
< inst
->tex_offset_num_offset
; j
++) {
3629 if (inst
->tex_offsets
[j
].file
== PROGRAM_TEMPORARY
&&
3630 inst
->tex_offsets
[j
].index
== index
) {
3631 inst
->tex_offsets
[j
].index
= new_index
;
3635 for (j
= 0; j
< num_inst_dst_regs(inst
->op
); j
++) {
3636 if (inst
->dst
[j
].file
== PROGRAM_TEMPORARY
&& inst
->dst
[j
].index
== index
) {
3637 inst
->dst
[j
].index
= new_index
;
3644 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3646 int depth
= 0; /* loop depth */
3647 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3650 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3651 for (j
= 0; j
< num_inst_src_regs(inst
->op
); j
++) {
3652 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3653 inst
->src
[j
].index
== index
) {
3654 return (depth
== 0) ? i
: loop_start
;
3657 for (j
= 0; j
< inst
->tex_offset_num_offset
; j
++) {
3658 if (inst
->tex_offsets
[j
].file
== PROGRAM_TEMPORARY
&&
3659 inst
->tex_offsets
[j
].index
== index
) {
3660 return (depth
== 0) ? i
: loop_start
;
3663 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3666 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3677 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3679 int depth
= 0; /* loop depth */
3680 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3684 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3685 for (j
= 0; j
< num_inst_dst_regs(inst
->op
); j
++) {
3686 if (inst
->dst
[j
].file
== PROGRAM_TEMPORARY
&& inst
->dst
[j
].index
== index
) {
3687 return (depth
== 0) ? i
: loop_start
;
3690 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3693 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3704 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3706 int depth
= 0; /* loop depth */
3707 int last
= -1; /* index of last instruction that reads the temporary */
3710 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3711 for (j
= 0; j
< num_inst_src_regs(inst
->op
); j
++) {
3712 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3713 inst
->src
[j
].index
== index
) {
3714 last
= (depth
== 0) ? i
: -2;
3717 for (j
= 0; j
< inst
->tex_offset_num_offset
; j
++) {
3718 if (inst
->tex_offsets
[j
].file
== PROGRAM_TEMPORARY
&&
3719 inst
->tex_offsets
[j
].index
== index
)
3720 last
= (depth
== 0) ? i
: -2;
3722 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3724 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3725 if (--depth
== 0 && last
== -2)
3735 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3737 int depth
= 0; /* loop depth */
3738 int last
= -1; /* index of last instruction that writes to the temporary */
3742 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3743 for (j
= 0; j
< num_inst_dst_regs(inst
->op
); j
++) {
3744 if (inst
->dst
[j
].file
== PROGRAM_TEMPORARY
&& inst
->dst
[j
].index
== index
)
3745 last
= (depth
== 0) ? i
: -2;
3748 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3750 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3751 if (--depth
== 0 && last
== -2)
3761 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3762 * channels for copy propagation and updates following instructions to
3763 * use the original versions.
3765 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3766 * will occur. As an example, a TXP production before this pass:
3768 * 0: MOV TEMP[1], INPUT[4].xyyy;
3769 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3770 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3774 * 0: MOV TEMP[1], INPUT[4].xyyy;
3775 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3776 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3778 * which allows for dead code elimination on TEMP[1]'s writes.
3781 glsl_to_tgsi_visitor::copy_propagate(void)
3783 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3784 glsl_to_tgsi_instruction
*,
3785 this->next_temp
* 4);
3786 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3789 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3790 assert(inst
->dst
[0].file
!= PROGRAM_TEMPORARY
3791 || inst
->dst
[0].index
< this->next_temp
);
3793 /* First, do any copy propagation possible into the src regs. */
3794 for (int r
= 0; r
< 3; r
++) {
3795 glsl_to_tgsi_instruction
*first
= NULL
;
3797 int acp_base
= inst
->src
[r
].index
* 4;
3799 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3800 inst
->src
[r
].reladdr
||
3801 inst
->src
[r
].reladdr2
)
3804 /* See if we can find entries in the ACP consisting of MOVs
3805 * from the same src register for all the swizzled channels
3806 * of this src register reference.
3808 for (int i
= 0; i
< 4; i
++) {
3809 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3810 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3817 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3822 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3823 first
->src
[0].index
!= copy_chan
->src
[0].index
||
3824 first
->src
[0].double_reg2
!= copy_chan
->src
[0].double_reg2
||
3825 first
->src
[0].index2D
!= copy_chan
->src
[0].index2D
) {
3833 /* We've now validated that we can copy-propagate to
3834 * replace this src register reference. Do it.
3836 inst
->src
[r
].file
= first
->src
[0].file
;
3837 inst
->src
[r
].index
= first
->src
[0].index
;
3838 inst
->src
[r
].index2D
= first
->src
[0].index2D
;
3839 inst
->src
[r
].has_index2
= first
->src
[0].has_index2
;
3840 inst
->src
[r
].double_reg2
= first
->src
[0].double_reg2
;
3841 inst
->src
[r
].array_id
= first
->src
[0].array_id
;
3844 for (int i
= 0; i
< 4; i
++) {
3845 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3846 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3847 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) << (3 * i
));
3849 inst
->src
[r
].swizzle
= swizzle
;
3854 case TGSI_OPCODE_BGNLOOP
:
3855 case TGSI_OPCODE_ENDLOOP
:
3856 /* End of a basic block, clear the ACP entirely. */
3857 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3860 case TGSI_OPCODE_IF
:
3861 case TGSI_OPCODE_UIF
:
3865 case TGSI_OPCODE_ENDIF
:
3866 case TGSI_OPCODE_ELSE
:
3867 /* Clear all channels written inside the block from the ACP, but
3868 * leaving those that were not touched.
3870 for (int r
= 0; r
< this->next_temp
; r
++) {
3871 for (int c
= 0; c
< 4; c
++) {
3872 if (!acp
[4 * r
+ c
])
3875 if (acp_level
[4 * r
+ c
] >= level
)
3876 acp
[4 * r
+ c
] = NULL
;
3879 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3884 /* Continuing the block, clear any written channels from
3887 for (int d
= 0; d
< 2; d
++) {
3888 if (inst
->dst
[d
].file
== PROGRAM_TEMPORARY
&& inst
->dst
[d
].reladdr
) {
3889 /* Any temporary might be written, so no copy propagation
3890 * across this instruction.
3892 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3893 } else if (inst
->dst
[d
].file
== PROGRAM_OUTPUT
&&
3894 inst
->dst
[d
].reladdr
) {
3895 /* Any output might be written, so no copy propagation
3896 * from outputs across this instruction.
3898 for (int r
= 0; r
< this->next_temp
; r
++) {
3899 for (int c
= 0; c
< 4; c
++) {
3900 if (!acp
[4 * r
+ c
])
3903 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3904 acp
[4 * r
+ c
] = NULL
;
3907 } else if (inst
->dst
[d
].file
== PROGRAM_TEMPORARY
||
3908 inst
->dst
[d
].file
== PROGRAM_OUTPUT
) {
3909 /* Clear where it's used as dst. */
3910 if (inst
->dst
[d
].file
== PROGRAM_TEMPORARY
) {
3911 for (int c
= 0; c
< 4; c
++) {
3912 if (inst
->dst
[d
].writemask
& (1 << c
))
3913 acp
[4 * inst
->dst
[d
].index
+ c
] = NULL
;
3917 /* Clear where it's used as src. */
3918 for (int r
= 0; r
< this->next_temp
; r
++) {
3919 for (int c
= 0; c
< 4; c
++) {
3920 if (!acp
[4 * r
+ c
])
3923 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3925 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
[d
].file
&&
3926 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
[d
].index
&&
3927 inst
->dst
[d
].writemask
& (1 << src_chan
)) {
3928 acp
[4 * r
+ c
] = NULL
;
3937 /* If this is a copy, add it to the ACP. */
3938 if (inst
->op
== TGSI_OPCODE_MOV
&&
3939 inst
->dst
[0].file
== PROGRAM_TEMPORARY
&&
3940 !(inst
->dst
[0].file
== inst
->src
[0].file
&&
3941 inst
->dst
[0].index
== inst
->src
[0].index
) &&
3942 !inst
->dst
[0].reladdr
&&
3944 inst
->src
[0].file
!= PROGRAM_ARRAY
&&
3945 !inst
->src
[0].reladdr
&&
3946 !inst
->src
[0].reladdr2
&&
3947 !inst
->src
[0].negate
) {
3948 for (int i
= 0; i
< 4; i
++) {
3949 if (inst
->dst
[0].writemask
& (1 << i
)) {
3950 acp
[4 * inst
->dst
[0].index
+ i
] = inst
;
3951 acp_level
[4 * inst
->dst
[0].index
+ i
] = level
;
3957 ralloc_free(acp_level
);
3962 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3965 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3966 * will occur. As an example, a TXP production after copy propagation but
3969 * 0: MOV TEMP[1], INPUT[4].xyyy;
3970 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3971 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3973 * and after this pass:
3975 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3978 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3980 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3981 glsl_to_tgsi_instruction
*,
3982 this->next_temp
* 4);
3983 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3987 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3988 assert(inst
->dst
[0].file
!= PROGRAM_TEMPORARY
3989 || inst
->dst
[0].index
< this->next_temp
);
3992 case TGSI_OPCODE_BGNLOOP
:
3993 case TGSI_OPCODE_ENDLOOP
:
3994 case TGSI_OPCODE_CONT
:
3995 case TGSI_OPCODE_BRK
:
3996 /* End of a basic block, clear the write array entirely.
3998 * This keeps us from killing dead code when the writes are
3999 * on either side of a loop, even when the register isn't touched
4000 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
4001 * dead code of this type, so it shouldn't make a difference as long as
4002 * the dead code elimination pass in the GLSL compiler does its job.
4004 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
4007 case TGSI_OPCODE_ENDIF
:
4008 case TGSI_OPCODE_ELSE
:
4009 /* Promote the recorded level of all channels written inside the
4010 * preceding if or else block to the level above the if/else block.
4012 for (int r
= 0; r
< this->next_temp
; r
++) {
4013 for (int c
= 0; c
< 4; c
++) {
4014 if (!writes
[4 * r
+ c
])
4017 if (write_level
[4 * r
+ c
] == level
)
4018 write_level
[4 * r
+ c
] = level
-1;
4021 if(inst
->op
== TGSI_OPCODE_ENDIF
)
4025 case TGSI_OPCODE_IF
:
4026 case TGSI_OPCODE_UIF
:
4028 /* fallthrough to default case to mark the condition as read */
4030 /* Continuing the block, clear any channels from the write array that
4031 * are read by this instruction.
4033 for (unsigned i
= 0; i
< ARRAY_SIZE(inst
->src
); i
++) {
4034 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
4035 /* Any temporary might be read, so no dead code elimination
4036 * across this instruction.
4038 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
4039 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
4040 /* Clear where it's used as src. */
4041 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
4042 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
4043 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
4044 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
4046 for (int c
= 0; c
< 4; c
++) {
4047 if (src_chans
& (1 << c
))
4048 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
4052 for (unsigned i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4053 if (inst
->tex_offsets
[i
].file
== PROGRAM_TEMPORARY
&& inst
->tex_offsets
[i
].reladdr
){
4054 /* Any temporary might be read, so no dead code elimination
4055 * across this instruction.
4057 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
4058 } else if (inst
->tex_offsets
[i
].file
== PROGRAM_TEMPORARY
) {
4059 /* Clear where it's used as src. */
4060 int src_chans
= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 0);
4061 src_chans
|= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 1);
4062 src_chans
|= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 2);
4063 src_chans
|= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 3);
4065 for (int c
= 0; c
< 4; c
++) {
4066 if (src_chans
& (1 << c
))
4067 writes
[4 * inst
->tex_offsets
[i
].index
+ c
] = NULL
;
4074 /* If this instruction writes to a temporary, add it to the write array.
4075 * If there is already an instruction in the write array for one or more
4076 * of the channels, flag that channel write as dead.
4078 for (unsigned i
= 0; i
< ARRAY_SIZE(inst
->dst
); i
++) {
4079 if (inst
->dst
[i
].file
== PROGRAM_TEMPORARY
&&
4080 !inst
->dst
[i
].reladdr
&&
4082 for (int c
= 0; c
< 4; c
++) {
4083 if (inst
->dst
[i
].writemask
& (1 << c
)) {
4084 if (writes
[4 * inst
->dst
[i
].index
+ c
]) {
4085 if (write_level
[4 * inst
->dst
[i
].index
+ c
] < level
)
4088 writes
[4 * inst
->dst
[i
].index
+ c
]->dead_mask
|= (1 << c
);
4090 writes
[4 * inst
->dst
[i
].index
+ c
] = inst
;
4091 write_level
[4 * inst
->dst
[i
].index
+ c
] = level
;
4098 /* Anything still in the write array at this point is dead code. */
4099 for (int r
= 0; r
< this->next_temp
; r
++) {
4100 for (int c
= 0; c
< 4; c
++) {
4101 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
4103 inst
->dead_mask
|= (1 << c
);
4107 /* Now actually remove the instructions that are completely dead and update
4108 * the writemask of other instructions with dead channels.
4110 foreach_in_list_safe(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
4111 if (!inst
->dead_mask
|| !inst
->dst
[0].writemask
)
4113 else if ((inst
->dst
[0].writemask
& ~inst
->dead_mask
) == 0) {
4118 if (inst
->dst
[0].type
== GLSL_TYPE_DOUBLE
) {
4119 if (inst
->dead_mask
== WRITEMASK_XY
||
4120 inst
->dead_mask
== WRITEMASK_ZW
)
4121 inst
->dst
[0].writemask
&= ~(inst
->dead_mask
);
4123 inst
->dst
[0].writemask
&= ~(inst
->dead_mask
);
4127 ralloc_free(write_level
);
4128 ralloc_free(writes
);
4133 /* merge DFRACEXP instructions into one. */
4135 glsl_to_tgsi_visitor::merge_two_dsts(void)
4137 foreach_in_list_safe(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
4138 glsl_to_tgsi_instruction
*inst2
;
4140 if (num_inst_dst_regs(inst
->op
) != 2)
4143 if (inst
->dst
[0].file
!= PROGRAM_UNDEFINED
&&
4144 inst
->dst
[1].file
!= PROGRAM_UNDEFINED
)
4147 inst2
= (glsl_to_tgsi_instruction
*) inst
->next
;
4150 if (inst
->src
[0].file
== inst2
->src
[0].file
&&
4151 inst
->src
[0].index
== inst2
->src
[0].index
&&
4152 inst
->src
[0].type
== inst2
->src
[0].type
&&
4153 inst
->src
[0].swizzle
== inst2
->src
[0].swizzle
)
4155 inst2
= (glsl_to_tgsi_instruction
*) inst2
->next
;
4161 if (inst
->dst
[0].file
== PROGRAM_UNDEFINED
) {
4163 inst
->dst
[0] = inst2
->dst
[0];
4164 } else if (inst
->dst
[1].file
== PROGRAM_UNDEFINED
) {
4165 inst
->dst
[1] = inst2
->dst
[1];
4176 /* Merges temporary registers together where possible to reduce the number of
4177 * registers needed to run a program.
4179 * Produces optimal code only after copy propagation and dead code elimination
4182 glsl_to_tgsi_visitor::merge_registers(void)
4184 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
4185 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
4188 /* Read the indices of the last read and first write to each temp register
4189 * into an array so that we don't have to traverse the instruction list as
4191 for (i
= 0; i
< this->next_temp
; i
++) {
4192 last_reads
[i
] = get_last_temp_read(i
);
4193 first_writes
[i
] = get_first_temp_write(i
);
4196 /* Start looking for registers with non-overlapping usages that can be
4197 * merged together. */
4198 for (i
= 0; i
< this->next_temp
; i
++) {
4199 /* Don't touch unused registers. */
4200 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
4202 for (j
= 0; j
< this->next_temp
; j
++) {
4203 /* Don't touch unused registers. */
4204 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
4206 /* We can merge the two registers if the first write to j is after or
4207 * in the same instruction as the last read from i. Note that the
4208 * register at index i will always be used earlier or at the same time
4209 * as the register at index j. */
4210 if (first_writes
[i
] <= first_writes
[j
] &&
4211 last_reads
[i
] <= first_writes
[j
]) {
4212 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
4214 /* Update the first_writes and last_reads arrays with the new
4215 * values for the merged register index, and mark the newly unused
4216 * register index as such. */
4217 last_reads
[i
] = last_reads
[j
];
4218 first_writes
[j
] = -1;
4224 ralloc_free(last_reads
);
4225 ralloc_free(first_writes
);
4228 /* Reassign indices to temporary registers by reusing unused indices created
4229 * by optimization passes. */
4231 glsl_to_tgsi_visitor::renumber_registers(void)
4236 for (i
= 0; i
< this->next_temp
; i
++) {
4237 if (get_first_temp_read(i
) < 0) continue;
4239 rename_temp_register(i
, new_index
);
4243 this->next_temp
= new_index
;
4247 * Returns a fragment program which implements the current pixel transfer ops.
4248 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
4251 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
4252 glsl_to_tgsi_visitor
*original
,
4253 int scale_and_bias
, int pixel_maps
)
4255 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
4256 struct st_context
*st
= st_context(original
->ctx
);
4257 struct gl_program
*prog
= &fp
->Base
.Base
;
4258 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
4259 st_src_reg coord
, src0
;
4261 glsl_to_tgsi_instruction
*inst
;
4263 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
4264 v
->ctx
= original
->ctx
;
4266 v
->shader_program
= NULL
;
4268 v
->glsl_version
= original
->glsl_version
;
4269 v
->native_integers
= original
->native_integers
;
4270 v
->options
= original
->options
;
4271 v
->next_temp
= original
->next_temp
;
4272 v
->num_address_regs
= original
->num_address_regs
;
4273 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
4274 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
4275 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
4276 v
->num_immediates
= original
->num_immediates
;
4279 * Get initial pixel color from the texture.
4280 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
4282 coord
= st_src_reg(PROGRAM_INPUT
, VARYING_SLOT_TEX0
, glsl_type::vec2_type
);
4283 src0
= v
->get_temp(glsl_type::vec4_type
);
4284 dst0
= st_dst_reg(src0
);
4285 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
4286 inst
->sampler_array_size
= 1;
4287 inst
->tex_target
= TEXTURE_2D_INDEX
;
4289 prog
->InputsRead
|= VARYING_BIT_TEX0
;
4290 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
4291 v
->samplers_used
|= (1 << 0);
4293 if (scale_and_bias
) {
4294 static const gl_state_index scale_state
[STATE_LENGTH
] =
4295 { STATE_INTERNAL
, STATE_PT_SCALE
,
4296 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
4297 static const gl_state_index bias_state
[STATE_LENGTH
] =
4298 { STATE_INTERNAL
, STATE_PT_BIAS
,
4299 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
4300 GLint scale_p
, bias_p
;
4301 st_src_reg scale
, bias
;
4303 scale_p
= _mesa_add_state_reference(params
, scale_state
);
4304 bias_p
= _mesa_add_state_reference(params
, bias_state
);
4306 /* MAD colorTemp, colorTemp, scale, bias; */
4307 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
4308 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
4309 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
4313 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
4314 st_dst_reg temp_dst
= st_dst_reg(temp
);
4316 assert(st
->pixel_xfer
.pixelmap_texture
);
4319 /* With a little effort, we can do four pixel map look-ups with
4320 * two TEX instructions:
4323 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
4324 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
4325 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
4326 inst
->sampler
.index
= 1;
4327 inst
->sampler_array_size
= 1;
4328 inst
->tex_target
= TEXTURE_2D_INDEX
;
4330 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
4331 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
4332 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
4333 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
4334 inst
->sampler
.index
= 1;
4335 inst
->sampler_array_size
= 1;
4336 inst
->tex_target
= TEXTURE_2D_INDEX
;
4338 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
4339 v
->samplers_used
|= (1 << 1);
4341 /* MOV colorTemp, temp; */
4342 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
4345 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
4347 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &original
->instructions
) {
4348 glsl_to_tgsi_instruction
*newinst
;
4349 st_src_reg src_regs
[3];
4351 if (inst
->dst
[0].file
== PROGRAM_OUTPUT
)
4352 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
[0].index
);
4354 for (int i
= 0; i
< 3; i
++) {
4355 src_regs
[i
] = inst
->src
[i
];
4356 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
4357 src_regs
[i
].index
== VARYING_SLOT_COL0
) {
4358 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
4359 src_regs
[i
].index
= src0
.index
;
4361 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
4362 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
4365 newinst
= v
->emit_asm(NULL
, inst
->op
, inst
->dst
[0], src_regs
[0], src_regs
[1], src_regs
[2]);
4366 newinst
->tex_target
= inst
->tex_target
;
4367 newinst
->sampler_array_size
= inst
->sampler_array_size
;
4370 /* Make modifications to fragment program info. */
4371 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
4372 original
->prog
->Parameters
);
4373 _mesa_free_parameter_list(params
);
4374 count_resources(v
, prog
);
4375 fp
->glsl_to_tgsi
= v
;
4379 * Make fragment program for glBitmap:
4380 * Sample the texture and kill the fragment if the bit is 0.
4381 * This program will be combined with the user's fragment program.
4383 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
4386 get_bitmap_visitor(struct st_fragment_program
*fp
,
4387 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
4389 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
4390 struct st_context
*st
= st_context(original
->ctx
);
4391 struct gl_program
*prog
= &fp
->Base
.Base
;
4392 st_src_reg coord
, src0
;
4394 glsl_to_tgsi_instruction
*inst
;
4396 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
4397 v
->ctx
= original
->ctx
;
4399 v
->shader_program
= NULL
;
4401 v
->glsl_version
= original
->glsl_version
;
4402 v
->native_integers
= original
->native_integers
;
4403 v
->options
= original
->options
;
4404 v
->next_temp
= original
->next_temp
;
4405 v
->num_address_regs
= original
->num_address_regs
;
4406 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
4407 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
4408 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
4409 v
->num_immediates
= original
->num_immediates
;
4411 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
4412 coord
= st_src_reg(PROGRAM_INPUT
, VARYING_SLOT_TEX0
, glsl_type::vec2_type
);
4413 src0
= v
->get_temp(glsl_type::vec4_type
);
4414 dst0
= st_dst_reg(src0
);
4415 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
4416 inst
->sampler
.index
= samplerIndex
;
4417 inst
->sampler_array_size
= 1;
4418 inst
->tex_target
= TEXTURE_2D_INDEX
;
4420 prog
->InputsRead
|= VARYING_BIT_TEX0
;
4421 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
4422 v
->samplers_used
|= (1 << samplerIndex
);
4424 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
4425 src0
.negate
= NEGATE_XYZW
;
4426 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
4427 src0
.swizzle
= SWIZZLE_XXXX
;
4428 inst
= v
->emit_asm(NULL
, TGSI_OPCODE_KILL_IF
, undef_dst
, src0
);
4430 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
4432 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &original
->instructions
) {
4433 glsl_to_tgsi_instruction
*newinst
;
4434 st_src_reg src_regs
[3];
4436 if (inst
->dst
[0].file
== PROGRAM_OUTPUT
)
4437 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
[0].index
);
4439 for (int i
= 0; i
< 3; i
++) {
4440 src_regs
[i
] = inst
->src
[i
];
4441 if (src_regs
[i
].file
== PROGRAM_INPUT
)
4442 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
4445 newinst
= v
->emit_asm(NULL
, inst
->op
, inst
->dst
[0], src_regs
[0], src_regs
[1], src_regs
[2]);
4446 newinst
->tex_target
= inst
->tex_target
;
4447 newinst
->sampler_array_size
= inst
->sampler_array_size
;
4450 /* Make modifications to fragment program info. */
4451 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
4452 count_resources(v
, prog
);
4453 fp
->glsl_to_tgsi
= v
;
4456 /* ------------------------- TGSI conversion stuff -------------------------- */
4458 unsigned branch_target
;
4463 * Intermediate state used during shader translation.
4465 struct st_translate
{
4466 struct ureg_program
*ureg
;
4468 unsigned temps_size
;
4469 struct ureg_dst
*temps
;
4471 struct ureg_dst
*arrays
;
4472 unsigned num_temp_arrays
;
4473 struct ureg_src
*constants
;
4475 struct ureg_src
*immediates
;
4477 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
4478 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
4479 struct ureg_dst address
[3];
4480 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
4481 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
4482 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
4483 unsigned *array_sizes
;
4484 struct array_decl
*input_arrays
;
4485 struct array_decl
*output_arrays
;
4487 const GLuint
*inputMapping
;
4488 const GLuint
*outputMapping
;
4490 /* For every instruction that contains a label (eg CALL), keep
4491 * details so that we can go back afterwards and emit the correct
4492 * tgsi instruction number for each label.
4494 struct label
*labels
;
4495 unsigned labels_size
;
4496 unsigned labels_count
;
4498 /* Keep a record of the tgsi instruction number that each mesa
4499 * instruction starts at, will be used to fix up labels after
4504 unsigned insn_count
;
4506 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4511 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4512 const unsigned _mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4515 TGSI_SEMANTIC_VERTEXID
,
4516 TGSI_SEMANTIC_INSTANCEID
,
4517 TGSI_SEMANTIC_VERTEXID_NOBASE
,
4518 TGSI_SEMANTIC_BASEVERTEX
,
4522 TGSI_SEMANTIC_INVOCATIONID
,
4527 TGSI_SEMANTIC_SAMPLEID
,
4528 TGSI_SEMANTIC_SAMPLEPOS
,
4529 TGSI_SEMANTIC_SAMPLEMASK
,
4533 * Make note of a branch to a label in the TGSI code.
4534 * After we've emitted all instructions, we'll go over the list
4535 * of labels built here and patch the TGSI code with the actual
4536 * location of each label.
4538 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4542 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4543 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4544 t
->labels
= (struct label
*)realloc(t
->labels
,
4545 t
->labels_size
* sizeof(struct label
));
4546 if (t
->labels
== NULL
) {
4547 static unsigned dummy
;
4553 i
= t
->labels_count
++;
4554 t
->labels
[i
].branch_target
= branch_target
;
4555 return &t
->labels
[i
].token
;
4559 * Called prior to emitting the TGSI code for each instruction.
4560 * Allocate additional space for instructions if needed.
4561 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4562 * the next TGSI instruction.
4564 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4566 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4567 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4568 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4569 if (t
->insn
== NULL
) {
4575 t
->insn
[t
->insn_count
++] = start
;
4579 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4581 static struct ureg_src
4582 emit_immediate(struct st_translate
*t
,
4583 gl_constant_value values
[4],
4586 struct ureg_program
*ureg
= t
->ureg
;
4591 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4593 return ureg_DECL_immediate_f64(ureg
, (double *)&values
[0].f
, size
);
4595 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4596 case GL_UNSIGNED_INT
:
4598 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4600 assert(!"should not get here - type must be float, int, uint, or bool");
4601 return ureg_src_undef();
4606 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4608 static struct ureg_dst
4609 dst_register(struct st_translate
*t
, gl_register_file file
, unsigned index
,
4615 case PROGRAM_UNDEFINED
:
4616 return ureg_dst_undef();
4618 case PROGRAM_TEMPORARY
:
4619 /* Allocate space for temporaries on demand. */
4620 if (index
>= t
->temps_size
) {
4621 const int inc
= 4096;
4623 t
->temps
= (struct ureg_dst
*)
4625 (t
->temps_size
+ inc
) * sizeof(struct ureg_dst
));
4627 return ureg_dst_undef();
4629 memset(t
->temps
+ t
->temps_size
, 0, inc
* sizeof(struct ureg_dst
));
4630 t
->temps_size
+= inc
;
4633 if (ureg_dst_is_undef(t
->temps
[index
]))
4634 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4636 return t
->temps
[index
];
4639 array
= index
>> 16;
4641 assert(array
< t
->num_temp_arrays
);
4643 if (ureg_dst_is_undef(t
->arrays
[array
]))
4644 t
->arrays
[array
] = ureg_DECL_array_temporary(
4645 t
->ureg
, t
->array_sizes
[array
], TRUE
);
4647 return ureg_dst_array_offset(t
->arrays
[array
],
4648 (int)(index
& 0xFFFF) - 0x8000);
4650 case PROGRAM_OUTPUT
:
4652 if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4653 assert(index
< FRAG_RESULT_MAX
);
4655 assert(index
< VARYING_SLOT_MAX
);
4657 assert(t
->outputMapping
[index
] < ARRAY_SIZE(t
->outputs
));
4658 assert(t
->outputs
[t
->outputMapping
[index
]].File
!= TGSI_FILE_NULL
);
4659 return t
->outputs
[t
->outputMapping
[index
]];
4662 struct array_decl
*decl
= &t
->output_arrays
[array_id
-1];
4663 unsigned mesa_index
= decl
->mesa_index
;
4664 int slot
= t
->outputMapping
[mesa_index
];
4666 assert(slot
!= -1 && t
->outputs
[slot
].File
== TGSI_FILE_OUTPUT
);
4667 assert(t
->outputs
[slot
].ArrayID
== array_id
);
4668 return ureg_dst_array_offset(t
->outputs
[slot
], index
- mesa_index
);
4671 case PROGRAM_ADDRESS
:
4672 return t
->address
[index
];
4675 assert(!"unknown dst register file");
4676 return ureg_dst_undef();
4681 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4683 static struct ureg_src
4684 src_register(struct st_translate
*t
, const st_src_reg
*reg
)
4686 int index
= reg
->index
;
4687 int double_reg2
= reg
->double_reg2
? 1 : 0;
4690 case PROGRAM_UNDEFINED
:
4691 return ureg_imm4f(t
->ureg
, 0, 0, 0, 0);
4693 case PROGRAM_TEMPORARY
:
4695 case PROGRAM_OUTPUT
:
4696 return ureg_src(dst_register(t
, reg
->file
, reg
->index
, reg
->array_id
));
4698 case PROGRAM_UNIFORM
:
4699 assert(reg
->index
>= 0);
4700 return reg
->index
< t
->num_constants
?
4701 t
->constants
[reg
->index
] : ureg_imm4f(t
->ureg
, 0, 0, 0, 0);
4702 case PROGRAM_STATE_VAR
:
4703 case PROGRAM_CONSTANT
: /* ie, immediate */
4704 if (reg
->has_index2
)
4705 return ureg_src_register(TGSI_FILE_CONSTANT
, reg
->index
);
4707 return reg
->index
>= 0 && reg
->index
< t
->num_constants
?
4708 t
->constants
[reg
->index
] : ureg_imm4f(t
->ureg
, 0, 0, 0, 0);
4710 case PROGRAM_IMMEDIATE
:
4711 assert(reg
->index
>= 0 && reg
->index
< t
->num_immediates
);
4712 return t
->immediates
[reg
->index
];
4715 /* GLSL inputs are 64-bit containers, so we have to
4716 * map back to the original index and add the offset after
4718 index
-= double_reg2
;
4719 if (!reg
->array_id
) {
4720 assert(t
->inputMapping
[index
] < ARRAY_SIZE(t
->inputs
));
4721 assert(t
->inputs
[t
->inputMapping
[index
]].File
!= TGSI_FILE_NULL
);
4722 return t
->inputs
[t
->inputMapping
[index
]];
4725 struct array_decl
*decl
= &t
->input_arrays
[reg
->array_id
-1];
4726 unsigned mesa_index
= decl
->mesa_index
;
4727 int slot
= t
->inputMapping
[mesa_index
];
4729 assert(slot
!= -1 && t
->inputs
[slot
].File
== TGSI_FILE_INPUT
);
4730 assert(t
->inputs
[slot
].ArrayID
== reg
->array_id
);
4731 return ureg_src_array_offset(t
->inputs
[slot
], index
- mesa_index
);
4734 case PROGRAM_ADDRESS
:
4735 return ureg_src(t
->address
[reg
->index
]);
4737 case PROGRAM_SYSTEM_VALUE
:
4738 assert(reg
->index
< (int) ARRAY_SIZE(t
->systemValues
));
4739 return t
->systemValues
[reg
->index
];
4742 assert(!"unknown src register file");
4743 return ureg_src_undef();
4748 * Create a TGSI ureg_dst register from an st_dst_reg.
4750 static struct ureg_dst
4751 translate_dst(struct st_translate
*t
,
4752 const st_dst_reg
*dst_reg
,
4753 bool saturate
, bool clamp_color
)
4755 struct ureg_dst dst
= dst_register(t
, dst_reg
->file
, dst_reg
->index
,
4758 if (dst
.File
== TGSI_FILE_NULL
)
4761 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4764 dst
= ureg_saturate(dst
);
4765 else if (clamp_color
&& dst_reg
->file
== PROGRAM_OUTPUT
) {
4766 /* Clamp colors for ARB_color_buffer_float. */
4767 switch (t
->procType
) {
4768 case TGSI_PROCESSOR_VERTEX
:
4769 /* This can only occur with a compatibility profile, which doesn't
4770 * support geometry shaders. */
4771 if (dst_reg
->index
== VARYING_SLOT_COL0
||
4772 dst_reg
->index
== VARYING_SLOT_COL1
||
4773 dst_reg
->index
== VARYING_SLOT_BFC0
||
4774 dst_reg
->index
== VARYING_SLOT_BFC1
) {
4775 dst
= ureg_saturate(dst
);
4779 case TGSI_PROCESSOR_FRAGMENT
:
4780 if (dst_reg
->index
== FRAG_RESULT_COLOR
||
4781 dst_reg
->index
>= FRAG_RESULT_DATA0
) {
4782 dst
= ureg_saturate(dst
);
4788 if (dst_reg
->reladdr
!= NULL
) {
4789 assert(dst_reg
->file
!= PROGRAM_TEMPORARY
);
4790 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4797 * Create a TGSI ureg_src register from an st_src_reg.
4799 static struct ureg_src
4800 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4802 struct ureg_src src
= src_register(t
, src_reg
);
4804 if (src_reg
->has_index2
) {
4805 /* 2D indexes occur with geometry shader inputs (attrib, vertex)
4806 * and UBO constant buffers (buffer, position).
4808 if (src_reg
->reladdr2
)
4809 src
= ureg_src_dimension_indirect(src
, ureg_src(t
->address
[1]),
4812 src
= ureg_src_dimension(src
, src_reg
->index2D
);
4815 src
= ureg_swizzle(src
,
4816 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4817 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4818 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4819 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4821 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4822 src
= ureg_negate(src
);
4824 if (src_reg
->reladdr
!= NULL
) {
4825 assert(src_reg
->file
!= PROGRAM_TEMPORARY
);
4826 src
= ureg_src_indirect(src
, ureg_src(t
->address
[0]));
4832 static struct tgsi_texture_offset
4833 translate_tex_offset(struct st_translate
*t
,
4834 const st_src_reg
*in_offset
, int idx
)
4836 struct tgsi_texture_offset offset
;
4837 struct ureg_src imm_src
;
4838 struct ureg_dst dst
;
4841 switch (in_offset
->file
) {
4842 case PROGRAM_IMMEDIATE
:
4843 assert(in_offset
->index
>= 0 && in_offset
->index
< t
->num_immediates
);
4844 imm_src
= t
->immediates
[in_offset
->index
];
4846 offset
.File
= imm_src
.File
;
4847 offset
.Index
= imm_src
.Index
;
4848 offset
.SwizzleX
= imm_src
.SwizzleX
;
4849 offset
.SwizzleY
= imm_src
.SwizzleY
;
4850 offset
.SwizzleZ
= imm_src
.SwizzleZ
;
4853 case PROGRAM_TEMPORARY
:
4854 imm_src
= ureg_src(t
->temps
[in_offset
->index
]);
4855 offset
.File
= imm_src
.File
;
4856 offset
.Index
= imm_src
.Index
;
4857 offset
.SwizzleX
= GET_SWZ(in_offset
->swizzle
, 0);
4858 offset
.SwizzleY
= GET_SWZ(in_offset
->swizzle
, 1);
4859 offset
.SwizzleZ
= GET_SWZ(in_offset
->swizzle
, 2);
4863 array
= in_offset
->index
>> 16;
4866 assert(array
< (int)t
->num_temp_arrays
);
4868 dst
= t
->arrays
[array
];
4869 offset
.File
= dst
.File
;
4870 offset
.Index
= dst
.Index
+ (in_offset
->index
& 0xFFFF) - 0x8000;
4871 offset
.SwizzleX
= GET_SWZ(in_offset
->swizzle
, 0);
4872 offset
.SwizzleY
= GET_SWZ(in_offset
->swizzle
, 1);
4873 offset
.SwizzleZ
= GET_SWZ(in_offset
->swizzle
, 2);
4883 compile_tgsi_instruction(struct st_translate
*t
,
4884 const glsl_to_tgsi_instruction
*inst
,
4885 bool clamp_dst_color_output
)
4887 struct ureg_program
*ureg
= t
->ureg
;
4889 struct ureg_dst dst
[2];
4890 struct ureg_src src
[4];
4891 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4895 unsigned tex_target
;
4897 num_dst
= num_inst_dst_regs(inst
->op
);
4898 num_src
= num_inst_src_regs(inst
->op
);
4900 for (i
= 0; i
< num_dst
; i
++)
4901 dst
[i
] = translate_dst(t
,
4904 clamp_dst_color_output
);
4906 for (i
= 0; i
< num_src
; i
++)
4907 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4910 case TGSI_OPCODE_BGNLOOP
:
4911 case TGSI_OPCODE_CAL
:
4912 case TGSI_OPCODE_ELSE
:
4913 case TGSI_OPCODE_ENDLOOP
:
4914 case TGSI_OPCODE_IF
:
4915 case TGSI_OPCODE_UIF
:
4916 assert(num_dst
== 0);
4917 ureg_label_insn(ureg
,
4921 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4924 case TGSI_OPCODE_TEX
:
4925 case TGSI_OPCODE_TXB
:
4926 case TGSI_OPCODE_TXD
:
4927 case TGSI_OPCODE_TXL
:
4928 case TGSI_OPCODE_TXP
:
4929 case TGSI_OPCODE_TXQ
:
4930 case TGSI_OPCODE_TXF
:
4931 case TGSI_OPCODE_TEX2
:
4932 case TGSI_OPCODE_TXB2
:
4933 case TGSI_OPCODE_TXL2
:
4934 case TGSI_OPCODE_TG4
:
4935 case TGSI_OPCODE_LODQ
:
4936 src
[num_src
] = t
->samplers
[inst
->sampler
.index
];
4937 assert(src
[num_src
].File
!= TGSI_FILE_NULL
);
4938 if (inst
->sampler
.reladdr
)
4940 ureg_src_indirect(src
[num_src
], ureg_src(t
->address
[2]));
4942 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4943 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
], i
);
4945 tex_target
= st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
);
4951 texoffsets
, inst
->tex_offset_num_offset
,
4955 case TGSI_OPCODE_SCS
:
4956 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4957 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4970 * Emit the TGSI instructions for inverting and adjusting WPOS.
4971 * This code is unavoidable because it also depends on whether
4972 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4975 emit_wpos_adjustment( struct st_translate
*t
,
4976 int wpos_transform_const
,
4978 GLfloat adjX
, GLfloat adjY
[2])
4980 struct ureg_program
*ureg
= t
->ureg
;
4982 assert(wpos_transform_const
>= 0);
4984 /* Fragment program uses fragment position input.
4985 * Need to replace instances of INPUT[WPOS] with temp T
4986 * where T = INPUT[WPOS] is inverted by Y.
4988 struct ureg_src wpostrans
= ureg_DECL_constant(ureg
, wpos_transform_const
);
4989 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4990 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[VARYING_SLOT_POS
]];
4992 /* First, apply the coordinate shift: */
4993 if (adjX
|| adjY
[0] || adjY
[1]) {
4994 if (adjY
[0] != adjY
[1]) {
4995 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4996 * depending on whether inversion is actually going to be applied
4997 * or not, which is determined by testing against the inversion
4998 * state variable used below, which will be either +1 or -1.
5000 struct ureg_dst adj_temp
= ureg_DECL_local_temporary(ureg
);
5002 ureg_CMP(ureg
, adj_temp
,
5003 ureg_scalar(wpostrans
, invert
? 2 : 0),
5004 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
5005 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
5006 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
5008 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
5009 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
5011 wpos_input
= ureg_src(wpos_temp
);
5013 /* MOV wpos_temp, input[wpos]
5015 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
5018 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
5019 * inversion/identity, or the other way around if we're drawing to an FBO.
5022 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
5025 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
5027 ureg_scalar(wpostrans
, 0),
5028 ureg_scalar(wpostrans
, 1));
5030 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
5033 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
5035 ureg_scalar(wpostrans
, 2),
5036 ureg_scalar(wpostrans
, 3));
5039 /* Use wpos_temp as position input from here on:
5041 t
->inputs
[t
->inputMapping
[VARYING_SLOT_POS
]] = ureg_src(wpos_temp
);
5046 * Emit fragment position/ooordinate code.
5049 emit_wpos(struct st_context
*st
,
5050 struct st_translate
*t
,
5051 const struct gl_program
*program
,
5052 struct ureg_program
*ureg
,
5053 int wpos_transform_const
)
5055 const struct gl_fragment_program
*fp
=
5056 (const struct gl_fragment_program
*) program
;
5057 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
5058 GLfloat adjX
= 0.0f
;
5059 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
5060 boolean invert
= FALSE
;
5062 /* Query the pixel center conventions supported by the pipe driver and set
5063 * adjX, adjY to help out if it cannot handle the requested one internally.
5065 * The bias of the y-coordinate depends on whether y-inversion takes place
5066 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
5067 * drawing to an FBO (causes additional inversion), and whether the the pipe
5068 * driver origin and the requested origin differ (the latter condition is
5069 * stored in the 'invert' variable).
5071 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
5073 * center shift only:
5078 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
5079 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
5080 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
5081 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
5083 * inversion and center shift:
5084 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
5085 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
5086 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
5087 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
5089 if (fp
->OriginUpperLeft
) {
5090 /* Fragment shader wants origin in upper-left */
5091 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
5092 /* the driver supports upper-left origin */
5094 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
5095 /* the driver supports lower-left origin, need to invert Y */
5096 ureg_property(ureg
, TGSI_PROPERTY_FS_COORD_ORIGIN
,
5097 TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
5104 /* Fragment shader wants origin in lower-left */
5105 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
5106 /* the driver supports lower-left origin */
5107 ureg_property(ureg
, TGSI_PROPERTY_FS_COORD_ORIGIN
,
5108 TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
5109 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
5110 /* the driver supports upper-left origin, need to invert Y */
5116 if (fp
->PixelCenterInteger
) {
5117 /* Fragment shader wants pixel center integer */
5118 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
5119 /* the driver supports pixel center integer */
5121 ureg_property(ureg
, TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
,
5122 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
5124 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
5125 /* the driver supports pixel center half integer, need to bias X,Y */
5134 /* Fragment shader wants pixel center half integer */
5135 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
5136 /* the driver supports pixel center half integer */
5138 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
5139 /* the driver supports pixel center integer, need to bias X,Y */
5140 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
5141 ureg_property(ureg
, TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
,
5142 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
5148 /* we invert after adjustment so that we avoid the MOV to temporary,
5149 * and reuse the adjustment ADD instead */
5150 emit_wpos_adjustment(t
, wpos_transform_const
, invert
, adjX
, adjY
);
5154 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
5155 * TGSI uses +1 for front, -1 for back.
5156 * This function converts the TGSI value to the GL value. Simply clamping/
5157 * saturating the value to [0,1] does the job.
5160 emit_face_var(struct gl_context
*ctx
, struct st_translate
*t
)
5162 struct ureg_program
*ureg
= t
->ureg
;
5163 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
5164 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[VARYING_SLOT_FACE
]];
5166 if (ctx
->Const
.NativeIntegers
) {
5167 ureg_FSGE(ureg
, face_temp
, face_input
, ureg_imm1f(ureg
, 0));
5170 /* MOV_SAT face_temp, input[face] */
5171 ureg_MOV(ureg
, ureg_saturate(face_temp
), face_input
);
5174 /* Use face_temp as face input from here on: */
5175 t
->inputs
[t
->inputMapping
[VARYING_SLOT_FACE
]] = ureg_src(face_temp
);
5179 emit_edgeflags(struct st_translate
*t
)
5181 struct ureg_program
*ureg
= t
->ureg
;
5182 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VARYING_SLOT_EDGE
]];
5183 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
5185 ureg_MOV(ureg
, edge_dst
, edge_src
);
5189 find_array(unsigned attr
, struct array_decl
*arrays
, unsigned count
,
5190 unsigned *array_id
, unsigned *array_size
)
5194 for (i
= 0; i
< count
; i
++) {
5195 struct array_decl
*decl
= &arrays
[i
];
5197 if (attr
== decl
->mesa_index
) {
5198 *array_id
= decl
->array_id
;
5199 *array_size
= decl
->array_size
;
5200 assert(*array_size
);
5208 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
5209 * \param program the program to translate
5210 * \param numInputs number of input registers used
5211 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
5213 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
5214 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
5216 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
5217 * \param interpLocation the TGSI_INTERPOLATE_LOC_* location for each input
5218 * \param numOutputs number of output registers used
5219 * \param outputMapping maps Mesa fragment program outputs to TGSI
5221 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
5222 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
5225 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
5227 extern "C" enum pipe_error
5228 st_translate_program(
5229 struct gl_context
*ctx
,
5231 struct ureg_program
*ureg
,
5232 glsl_to_tgsi_visitor
*program
,
5233 const struct gl_program
*proginfo
,
5235 const GLuint inputMapping
[],
5236 const GLuint inputSlotToAttr
[],
5237 const ubyte inputSemanticName
[],
5238 const ubyte inputSemanticIndex
[],
5239 const GLuint interpMode
[],
5240 const GLuint interpLocation
[],
5242 const GLuint outputMapping
[],
5243 const GLuint outputSlotToAttr
[],
5244 const ubyte outputSemanticName
[],
5245 const ubyte outputSemanticIndex
[],
5246 boolean passthrough_edgeflags
,
5247 boolean clamp_color
)
5249 struct st_translate
*t
;
5251 enum pipe_error ret
= PIPE_OK
;
5253 assert(numInputs
<= ARRAY_SIZE(t
->inputs
));
5254 assert(numOutputs
<= ARRAY_SIZE(t
->outputs
));
5256 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_FRONT_FACE
] ==
5257 TGSI_SEMANTIC_FACE
);
5258 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_VERTEX_ID
] ==
5259 TGSI_SEMANTIC_VERTEXID
);
5260 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_INSTANCE_ID
] ==
5261 TGSI_SEMANTIC_INSTANCEID
);
5262 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_SAMPLE_ID
] ==
5263 TGSI_SEMANTIC_SAMPLEID
);
5264 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_SAMPLE_POS
] ==
5265 TGSI_SEMANTIC_SAMPLEPOS
);
5266 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_SAMPLE_MASK_IN
] ==
5267 TGSI_SEMANTIC_SAMPLEMASK
);
5268 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_INVOCATION_ID
] ==
5269 TGSI_SEMANTIC_INVOCATIONID
);
5270 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
] ==
5271 TGSI_SEMANTIC_VERTEXID_NOBASE
);
5272 assert(_mesa_sysval_to_semantic
[SYSTEM_VALUE_BASE_VERTEX
] ==
5273 TGSI_SEMANTIC_BASEVERTEX
);
5275 t
= CALLOC_STRUCT(st_translate
);
5277 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
5281 t
->procType
= procType
;
5282 t
->inputMapping
= inputMapping
;
5283 t
->outputMapping
= outputMapping
;
5285 t
->num_temp_arrays
= program
->next_array
;
5286 if (t
->num_temp_arrays
)
5287 t
->arrays
= (struct ureg_dst
*)
5288 calloc(1, sizeof(t
->arrays
[0]) * t
->num_temp_arrays
);
5291 * Declare input attributes.
5294 case TGSI_PROCESSOR_FRAGMENT
:
5295 for (i
= 0; i
< numInputs
; i
++) {
5296 unsigned array_id
= 0;
5297 unsigned array_size
;
5299 if (find_array(inputSlotToAttr
[i
], program
->input_arrays
,
5300 program
->num_input_arrays
, &array_id
, &array_size
)) {
5301 /* We've found an array. Declare it so. */
5302 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
5303 inputSemanticName
[i
], inputSemanticIndex
[i
],
5304 interpMode
[i
], 0, interpLocation
[i
],
5305 array_id
, array_size
);
5306 i
+= array_size
- 1;
5309 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
5310 inputSemanticName
[i
], inputSemanticIndex
[i
],
5311 interpMode
[i
], 0, interpLocation
[i
], 0, 1);
5315 case TGSI_PROCESSOR_GEOMETRY
:
5316 for (i
= 0; i
< numInputs
; i
++) {
5317 unsigned array_id
= 0;
5318 unsigned array_size
;
5320 if (find_array(inputSlotToAttr
[i
], program
->input_arrays
,
5321 program
->num_input_arrays
, &array_id
, &array_size
)) {
5322 /* We've found an array. Declare it so. */
5323 t
->inputs
[i
] = ureg_DECL_input(ureg
, inputSemanticName
[i
],
5324 inputSemanticIndex
[i
],
5325 array_id
, array_size
);
5326 i
+= array_size
- 1;
5329 t
->inputs
[i
] = ureg_DECL_input(ureg
, inputSemanticName
[i
],
5330 inputSemanticIndex
[i
], 0, 1);
5334 case TGSI_PROCESSOR_VERTEX
:
5335 for (i
= 0; i
< numInputs
; i
++) {
5336 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
5344 * Declare output attributes.
5347 case TGSI_PROCESSOR_FRAGMENT
:
5349 case TGSI_PROCESSOR_GEOMETRY
:
5350 case TGSI_PROCESSOR_VERTEX
:
5351 for (i
= 0; i
< numOutputs
; i
++) {
5352 unsigned array_id
= 0;
5353 unsigned array_size
;
5355 if (find_array(outputSlotToAttr
[i
], program
->output_arrays
,
5356 program
->num_output_arrays
, &array_id
, &array_size
)) {
5357 /* We've found an array. Declare it so. */
5358 t
->outputs
[i
] = ureg_DECL_output_array(ureg
,
5359 outputSemanticName
[i
],
5360 outputSemanticIndex
[i
],
5361 array_id
, array_size
);
5362 i
+= array_size
- 1;
5365 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5366 outputSemanticName
[i
],
5367 outputSemanticIndex
[i
]);
5375 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
5376 if (proginfo
->InputsRead
& VARYING_BIT_POS
) {
5377 /* Must do this after setting up t->inputs. */
5378 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
,
5379 program
->wpos_transform_const
);
5382 if (proginfo
->InputsRead
& VARYING_BIT_FACE
)
5383 emit_face_var(ctx
, t
);
5385 for (i
= 0; i
< numOutputs
; i
++) {
5386 switch (outputSemanticName
[i
]) {
5387 case TGSI_SEMANTIC_POSITION
:
5388 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5389 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
5390 outputSemanticIndex
[i
]);
5391 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
5393 case TGSI_SEMANTIC_STENCIL
:
5394 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5395 TGSI_SEMANTIC_STENCIL
, /* Stencil */
5396 outputSemanticIndex
[i
]);
5397 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
5399 case TGSI_SEMANTIC_COLOR
:
5400 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5401 TGSI_SEMANTIC_COLOR
,
5402 outputSemanticIndex
[i
]);
5404 case TGSI_SEMANTIC_SAMPLEMASK
:
5405 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5406 TGSI_SEMANTIC_SAMPLEMASK
,
5407 outputSemanticIndex
[i
]);
5408 /* TODO: If we ever support more than 32 samples, this will have
5409 * to become an array.
5411 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_X
);
5414 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
5415 ret
= PIPE_ERROR_BAD_INPUT
;
5420 else if (procType
== TGSI_PROCESSOR_VERTEX
) {
5421 for (i
= 0; i
< numOutputs
; i
++) {
5422 if (outputSemanticName
[i
] == TGSI_SEMANTIC_FOG
) {
5423 /* force register to contain a fog coordinate in the form (F, 0, 0, 1). */
5425 ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_YZW
),
5426 ureg_imm4f(ureg
, 0.0f
, 0.0f
, 0.0f
, 1.0f
));
5427 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_X
);
5430 if (passthrough_edgeflags
)
5434 /* Declare address register.
5436 if (program
->num_address_regs
> 0) {
5437 assert(program
->num_address_regs
<= 3);
5438 for (int i
= 0; i
< program
->num_address_regs
; i
++)
5439 t
->address
[i
] = ureg_DECL_address(ureg
);
5442 /* Declare misc input registers
5445 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
5446 unsigned numSys
= 0;
5447 for (i
= 0; sysInputs
; i
++) {
5448 if (sysInputs
& (1 << i
)) {
5449 unsigned semName
= _mesa_sysval_to_semantic
[i
];
5450 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
5451 if (semName
== TGSI_SEMANTIC_INSTANCEID
||
5452 semName
== TGSI_SEMANTIC_VERTEXID
) {
5453 /* From Gallium perspective, these system values are always
5454 * integer, and require native integer support. However, if
5455 * native integer is supported on the vertex stage but not the
5456 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
5457 * assumes these system values are floats. To resolve the
5458 * inconsistency, we insert a U2F.
5460 struct st_context
*st
= st_context(ctx
);
5461 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
5462 assert(procType
== TGSI_PROCESSOR_VERTEX
);
5463 assert(pscreen
->get_shader_param(pscreen
, PIPE_SHADER_VERTEX
, PIPE_SHADER_CAP_INTEGERS
));
5465 if (!ctx
->Const
.NativeIntegers
) {
5466 struct ureg_dst temp
= ureg_DECL_local_temporary(t
->ureg
);
5467 ureg_U2F( t
->ureg
, ureg_writemask(temp
, TGSI_WRITEMASK_X
), t
->systemValues
[i
]);
5468 t
->systemValues
[i
] = ureg_scalar(ureg_src(temp
), 0);
5472 sysInputs
&= ~(1 << i
);
5477 t
->array_sizes
= program
->array_sizes
;
5478 t
->input_arrays
= program
->input_arrays
;
5479 t
->output_arrays
= program
->output_arrays
;
5481 /* Emit constants and uniforms. TGSI uses a single index space for these,
5482 * so we put all the translated regs in t->constants.
5484 if (proginfo
->Parameters
) {
5485 t
->constants
= (struct ureg_src
*)
5486 calloc(proginfo
->Parameters
->NumParameters
, sizeof(t
->constants
[0]));
5487 if (t
->constants
== NULL
) {
5488 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
5491 t
->num_constants
= proginfo
->Parameters
->NumParameters
;
5493 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
5494 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
5495 case PROGRAM_STATE_VAR
:
5496 case PROGRAM_UNIFORM
:
5497 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
5500 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
5501 * addressing of the const buffer.
5502 * FIXME: Be smarter and recognize param arrays:
5503 * indirect addressing is only valid within the referenced
5506 case PROGRAM_CONSTANT
:
5507 if (program
->indirect_addr_consts
)
5508 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
5510 t
->constants
[i
] = emit_immediate(t
,
5511 proginfo
->Parameters
->ParameterValues
[i
],
5512 proginfo
->Parameters
->Parameters
[i
].DataType
,
5521 if (program
->shader
) {
5522 unsigned num_ubos
= program
->shader
->NumUniformBlocks
;
5524 for (i
= 0; i
< num_ubos
; i
++) {
5525 unsigned size
= program
->shader
->UniformBlocks
[i
].UniformBufferSize
;
5526 unsigned num_const_vecs
= (size
+ 15) / 16;
5527 unsigned first
, last
;
5528 assert(num_const_vecs
> 0);
5530 last
= num_const_vecs
> 0 ? num_const_vecs
- 1 : 0;
5531 ureg_DECL_constant2D(t
->ureg
, first
, last
, i
+ 1);
5535 /* Emit immediate values.
5537 t
->immediates
= (struct ureg_src
*)
5538 calloc(program
->num_immediates
, sizeof(struct ureg_src
));
5539 if (t
->immediates
== NULL
) {
5540 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
5543 t
->num_immediates
= program
->num_immediates
;
5546 foreach_in_list(immediate_storage
, imm
, &program
->immediates
) {
5547 assert(i
< program
->num_immediates
);
5548 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size32
);
5550 assert(i
== program
->num_immediates
);
5552 /* texture samplers */
5553 for (i
= 0; i
< ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxTextureImageUnits
; i
++) {
5554 if (program
->samplers_used
& (1 << i
)) {
5557 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
5559 switch (program
->sampler_types
[i
]) {
5561 type
= TGSI_RETURN_TYPE_SINT
;
5563 case GLSL_TYPE_UINT
:
5564 type
= TGSI_RETURN_TYPE_UINT
;
5566 case GLSL_TYPE_FLOAT
:
5567 type
= TGSI_RETURN_TYPE_FLOAT
;
5570 unreachable("not reached");
5573 ureg_DECL_sampler_view( ureg
, i
, program
->sampler_targets
[i
],
5574 type
, type
, type
, type
);
5578 /* Emit each instruction in turn:
5580 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &program
->instructions
) {
5581 set_insn_start(t
, ureg_get_instruction_number(ureg
));
5582 compile_tgsi_instruction(t
, inst
, clamp_color
);
5585 /* Fix up all emitted labels:
5587 for (i
= 0; i
< t
->labels_count
; i
++) {
5588 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
5589 t
->insn
[t
->labels
[i
].branch_target
]);
5599 t
->num_constants
= 0;
5600 free(t
->immediates
);
5601 t
->num_immediates
= 0;
5604 debug_printf("%s: translate error flag set\n", __func__
);
5612 /* ----------------------------- End TGSI code ------------------------------ */
5616 shader_stage_to_ptarget(gl_shader_stage stage
)
5619 case MESA_SHADER_VERTEX
:
5620 return PIPE_SHADER_VERTEX
;
5621 case MESA_SHADER_FRAGMENT
:
5622 return PIPE_SHADER_FRAGMENT
;
5623 case MESA_SHADER_GEOMETRY
:
5624 return PIPE_SHADER_GEOMETRY
;
5625 case MESA_SHADER_COMPUTE
:
5626 return PIPE_SHADER_COMPUTE
;
5629 assert(!"should not be reached");
5630 return PIPE_SHADER_VERTEX
;
5635 * Convert a shader's GLSL IR into a Mesa gl_program, although without
5636 * generating Mesa IR.
5638 static struct gl_program
*
5639 get_mesa_program(struct gl_context
*ctx
,
5640 struct gl_shader_program
*shader_program
,
5641 struct gl_shader
*shader
)
5643 glsl_to_tgsi_visitor
* v
;
5644 struct gl_program
*prog
;
5645 GLenum target
= _mesa_shader_stage_to_program(shader
->Stage
);
5647 struct gl_shader_compiler_options
*options
=
5648 &ctx
->Const
.ShaderCompilerOptions
[_mesa_shader_enum_to_shader_stage(shader
->Type
)];
5649 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
5650 unsigned ptarget
= shader_stage_to_ptarget(shader
->Stage
);
5652 validate_ir_tree(shader
->ir
);
5654 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
5657 prog
->Parameters
= _mesa_new_parameter_list();
5658 v
= new glsl_to_tgsi_visitor();
5661 v
->shader_program
= shader_program
;
5663 v
->options
= options
;
5664 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
5665 v
->native_integers
= ctx
->Const
.NativeIntegers
;
5667 v
->have_sqrt
= pscreen
->get_shader_param(pscreen
, ptarget
,
5668 PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED
);
5669 v
->have_fma
= pscreen
->get_shader_param(pscreen
, ptarget
,
5670 PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED
);
5672 _mesa_copy_linked_program_data(shader
->Stage
, shader_program
, prog
);
5673 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
5676 /* Remove reads from output registers. */
5677 lower_output_reads(shader
->ir
);
5679 /* Emit intermediate IR for main(). */
5680 visit_exec_list(shader
->ir
, v
);
5682 /* Now emit bodies for any functions that were used. */
5684 progress
= GL_FALSE
;
5686 foreach_in_list(function_entry
, entry
, &v
->function_signatures
) {
5687 if (!entry
->bgn_inst
) {
5688 v
->current_function
= entry
;
5690 entry
->bgn_inst
= v
->emit_asm(NULL
, TGSI_OPCODE_BGNSUB
);
5691 entry
->bgn_inst
->function
= entry
;
5693 visit_exec_list(&entry
->sig
->body
, v
);
5695 glsl_to_tgsi_instruction
*last
;
5696 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
5697 if (last
->op
!= TGSI_OPCODE_RET
)
5698 v
->emit_asm(NULL
, TGSI_OPCODE_RET
);
5700 glsl_to_tgsi_instruction
*end
;
5701 end
= v
->emit_asm(NULL
, TGSI_OPCODE_ENDSUB
);
5702 end
->function
= entry
;
5710 /* Print out some information (for debugging purposes) used by the
5711 * optimization passes. */
5712 for (i
= 0; i
< v
->next_temp
; i
++) {
5713 int fr
= v
->get_first_temp_read(i
);
5714 int fw
= v
->get_first_temp_write(i
);
5715 int lr
= v
->get_last_temp_read(i
);
5716 int lw
= v
->get_last_temp_write(i
);
5718 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
5723 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
5725 v
->copy_propagate();
5726 while (v
->eliminate_dead_code());
5728 v
->merge_two_dsts();
5729 v
->merge_registers();
5730 v
->renumber_registers();
5732 /* Write the END instruction. */
5733 v
->emit_asm(NULL
, TGSI_OPCODE_END
);
5735 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
5737 _mesa_log("GLSL IR for linked %s program %d:\n",
5738 _mesa_shader_stage_to_string(shader
->Stage
),
5739 shader_program
->Name
);
5740 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
5744 prog
->Instructions
= NULL
;
5745 prog
->NumInstructions
= 0;
5747 do_set_program_inouts(shader
->ir
, prog
, shader
->Stage
);
5748 shrink_array_declarations(v
->input_arrays
, v
->num_input_arrays
,
5750 shrink_array_declarations(v
->output_arrays
, v
->num_output_arrays
,
5751 prog
->OutputsWritten
);
5752 count_resources(v
, prog
);
5754 /* This must be done before the uniform storage is associated. */
5755 if (shader
->Type
== GL_FRAGMENT_SHADER
&&
5756 prog
->InputsRead
& VARYING_BIT_POS
){
5757 static const gl_state_index wposTransformState
[STATE_LENGTH
] = {
5758 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
5761 v
->wpos_transform_const
= _mesa_add_state_reference(prog
->Parameters
,
5762 wposTransformState
);
5765 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5767 /* This has to be done last. Any operation the can cause
5768 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5769 * program constant) has to happen before creating this linkage.
5771 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5772 if (!shader_program
->LinkStatus
) {
5773 free_glsl_to_tgsi_visitor(v
);
5777 struct st_vertex_program
*stvp
;
5778 struct st_fragment_program
*stfp
;
5779 struct st_geometry_program
*stgp
;
5781 switch (shader
->Type
) {
5782 case GL_VERTEX_SHADER
:
5783 stvp
= (struct st_vertex_program
*)prog
;
5784 stvp
->glsl_to_tgsi
= v
;
5786 case GL_FRAGMENT_SHADER
:
5787 stfp
= (struct st_fragment_program
*)prog
;
5788 stfp
->glsl_to_tgsi
= v
;
5790 case GL_GEOMETRY_SHADER
:
5791 stgp
= (struct st_geometry_program
*)prog
;
5792 stgp
->glsl_to_tgsi
= v
;
5795 assert(!"should not be reached");
5806 * Called via ctx->Driver.LinkShader()
5807 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5808 * with code lowering and other optimizations.
5811 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5813 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
5814 assert(prog
->LinkStatus
);
5816 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
5817 if (prog
->_LinkedShaders
[i
] == NULL
)
5821 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5822 gl_shader_stage stage
= _mesa_shader_enum_to_shader_stage(prog
->_LinkedShaders
[i
]->Type
);
5823 const struct gl_shader_compiler_options
*options
=
5824 &ctx
->Const
.ShaderCompilerOptions
[stage
];
5825 unsigned ptarget
= shader_stage_to_ptarget(stage
);
5826 bool have_dround
= pscreen
->get_shader_param(pscreen
, ptarget
,
5827 PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED
);
5828 bool have_dfrexp
= pscreen
->get_shader_param(pscreen
, ptarget
,
5829 PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED
);
5831 /* If there are forms of indirect addressing that the driver
5832 * cannot handle, perform the lowering pass.
5834 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
||
5835 options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
) {
5836 lower_variable_index_to_cond_assign(ir
,
5837 options
->EmitNoIndirectInput
,
5838 options
->EmitNoIndirectOutput
,
5839 options
->EmitNoIndirectTemp
,
5840 options
->EmitNoIndirectUniform
);
5843 if (ctx
->Extensions
.ARB_shading_language_packing
) {
5844 unsigned lower_inst
= LOWER_PACK_SNORM_2x16
|
5845 LOWER_UNPACK_SNORM_2x16
|
5846 LOWER_PACK_UNORM_2x16
|
5847 LOWER_UNPACK_UNORM_2x16
|
5848 LOWER_PACK_SNORM_4x8
|
5849 LOWER_UNPACK_SNORM_4x8
|
5850 LOWER_UNPACK_UNORM_4x8
|
5851 LOWER_PACK_UNORM_4x8
|
5852 LOWER_PACK_HALF_2x16
|
5853 LOWER_UNPACK_HALF_2x16
;
5855 lower_packing_builtins(ir
, lower_inst
);
5858 if (!pscreen
->get_param(pscreen
, PIPE_CAP_TEXTURE_GATHER_OFFSETS
))
5859 lower_offset_arrays(ir
);
5860 do_mat_op_to_vec(ir
);
5861 lower_instructions(ir
,
5867 (have_dfrexp
? 0 : DFREXP_DLDEXP_TO_ARITH
) |
5870 (have_dround
? 0 : DOPS_TO_DFRAC
) |
5871 (options
->EmitNoPow
? POW_TO_EXP2
: 0) |
5872 (!ctx
->Const
.NativeIntegers
? INT_DIV_TO_MUL_RCP
: 0) |
5873 (options
->EmitNoSat
? SAT_TO_CLAMP
: 0));
5875 lower_ubo_reference(prog
->_LinkedShaders
[i
], ir
);
5876 do_vec_index_to_cond_assign(ir
);
5877 lower_vector_insert(ir
, true);
5878 lower_quadop_vector(ir
, false);
5880 if (options
->MaxIfDepth
== 0) {
5887 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5889 progress
= do_common_optimization(ir
, true, true, options
,
5890 ctx
->Const
.NativeIntegers
)
5893 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5897 validate_ir_tree(ir
);
5900 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
5901 struct gl_program
*linked_prog
;
5903 if (prog
->_LinkedShaders
[i
] == NULL
)
5906 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5909 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5911 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
5912 _mesa_shader_stage_to_program(i
),
5914 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5916 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5921 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5928 st_translate_stream_output_info(glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5929 const GLuint outputMapping
[],
5930 struct pipe_stream_output_info
*so
)
5933 struct gl_transform_feedback_info
*info
=
5934 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5936 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5937 so
->output
[i
].register_index
=
5938 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5939 so
->output
[i
].start_component
= info
->Outputs
[i
].ComponentOffset
;
5940 so
->output
[i
].num_components
= info
->Outputs
[i
].NumComponents
;
5941 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5942 so
->output
[i
].dst_offset
= info
->Outputs
[i
].DstOffset
;
5943 so
->output
[i
].stream
= info
->Outputs
[i
].StreamId
;
5946 for (i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
5947 so
->stride
[i
] = info
->BufferStride
[i
];
5949 so
->num_outputs
= info
->NumOutputs
;