2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
23 * DEALINGS IN THE SOFTWARE.
27 * \file ir_to_mesa.cpp
29 * Translates the IR to ARB_fragment_program text if possible,
35 #include "ir_visitor.h"
36 #include "ir_print_visitor.h"
37 #include "ir_expression_flattening.h"
38 #include "glsl_types.h"
39 #include "glsl_parser_extras.h"
40 #include "../glsl/program.h"
41 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "main/uniforms.h"
48 #include "program/prog_instruction.h"
49 #include "program/prog_optimize.h"
50 #include "program/prog_print.h"
51 #include "program/program.h"
52 #include "program/prog_uniform.h"
53 #include "program/prog_parameter.h"
56 static int swizzle_for_size(int size
);
59 * This struct is a corresponding struct to Mesa prog_src_register, with
62 typedef struct ir_to_mesa_src_reg
{
63 ir_to_mesa_src_reg(int file
, int index
, const glsl_type
*type
)
67 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
68 this->swizzle
= swizzle_for_size(type
->vector_elements
);
70 this->swizzle
= SWIZZLE_XYZW
;
77 this->file
= PROGRAM_UNDEFINED
;
80 int file
; /**< PROGRAM_* from Mesa */
81 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
82 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
83 int negate
; /**< NEGATE_XYZW mask from mesa */
84 /** Register index should be offset by the integer in this reg. */
85 ir_to_mesa_src_reg
*reladdr
;
88 typedef struct ir_to_mesa_dst_reg
{
89 int file
; /**< PROGRAM_* from Mesa */
90 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
91 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
93 /** Register index should be offset by the integer in this reg. */
94 ir_to_mesa_src_reg
*reladdr
;
97 extern ir_to_mesa_src_reg ir_to_mesa_undef
;
99 class ir_to_mesa_instruction
: public exec_node
{
102 ir_to_mesa_dst_reg dst_reg
;
103 ir_to_mesa_src_reg src_reg
[3];
104 /** Pointer to the ir source this tree came from for debugging */
106 GLboolean cond_update
;
107 int sampler
; /**< sampler index */
108 int tex_target
; /**< One of TEXTURE_*_INDEX */
109 GLboolean tex_shadow
;
111 class function_entry
*function
; /* Set on OPCODE_CAL or OPCODE_BGNSUB */
114 class variable_storage
: public exec_node
{
116 variable_storage(ir_variable
*var
, int file
, int index
)
117 : file(file
), index(index
), var(var
)
124 ir_variable
*var
; /* variable that maps to this, if any */
127 class function_entry
: public exec_node
{
129 ir_function_signature
*sig
;
132 * identifier of this function signature used by the program.
134 * At the point that Mesa instructions for function calls are
135 * generated, we don't know the address of the first instruction of
136 * the function body. So we make the BranchTarget that is called a
137 * small integer and rewrite them during set_branchtargets().
142 * Pointer to first instruction of the function body.
144 * Set during function body emits after main() is processed.
146 ir_to_mesa_instruction
*bgn_inst
;
149 * Index of the first instruction of the function body in actual
152 * Set after convertion from ir_to_mesa_instruction to prog_instruction.
156 /** Storage for the return value. */
157 ir_to_mesa_src_reg return_reg
;
160 class ir_to_mesa_visitor
: public ir_visitor
{
162 ir_to_mesa_visitor();
164 function_entry
*current_function
;
167 struct gl_program
*prog
;
171 variable_storage
*find_variable_storage(ir_variable
*var
);
173 function_entry
*get_function_signature(ir_function_signature
*sig
);
175 ir_to_mesa_src_reg
get_temp(const glsl_type
*type
);
176 void reladdr_to_temp(ir_instruction
*ir
,
177 ir_to_mesa_src_reg
*reg
, int *num_reladdr
);
179 struct ir_to_mesa_src_reg
src_reg_for_float(float val
);
182 * \name Visit methods
184 * As typical for the visitor pattern, there must be one \c visit method for
185 * each concrete subclass of \c ir_instruction. Virtual base classes within
186 * the hierarchy should not have \c visit methods.
189 virtual void visit(ir_variable
*);
190 virtual void visit(ir_loop
*);
191 virtual void visit(ir_loop_jump
*);
192 virtual void visit(ir_function_signature
*);
193 virtual void visit(ir_function
*);
194 virtual void visit(ir_expression
*);
195 virtual void visit(ir_swizzle
*);
196 virtual void visit(ir_dereference_variable
*);
197 virtual void visit(ir_dereference_array
*);
198 virtual void visit(ir_dereference_record
*);
199 virtual void visit(ir_assignment
*);
200 virtual void visit(ir_constant
*);
201 virtual void visit(ir_call
*);
202 virtual void visit(ir_return
*);
203 virtual void visit(ir_discard
*);
204 virtual void visit(ir_texture
*);
205 virtual void visit(ir_if
*);
208 struct ir_to_mesa_src_reg result
;
210 /** List of variable_storage */
213 /** List of function_entry */
214 exec_list function_signatures
;
215 int next_signature_id
;
217 /** List of ir_to_mesa_instruction */
218 exec_list instructions
;
220 ir_to_mesa_instruction
*ir_to_mesa_emit_op0(ir_instruction
*ir
,
221 enum prog_opcode op
);
223 ir_to_mesa_instruction
*ir_to_mesa_emit_op1(ir_instruction
*ir
,
225 ir_to_mesa_dst_reg dst
,
226 ir_to_mesa_src_reg src0
);
228 ir_to_mesa_instruction
*ir_to_mesa_emit_op2(ir_instruction
*ir
,
230 ir_to_mesa_dst_reg dst
,
231 ir_to_mesa_src_reg src0
,
232 ir_to_mesa_src_reg src1
);
234 ir_to_mesa_instruction
*ir_to_mesa_emit_op3(ir_instruction
*ir
,
236 ir_to_mesa_dst_reg dst
,
237 ir_to_mesa_src_reg src0
,
238 ir_to_mesa_src_reg src1
,
239 ir_to_mesa_src_reg src2
);
241 void ir_to_mesa_emit_scalar_op1(ir_instruction
*ir
,
243 ir_to_mesa_dst_reg dst
,
244 ir_to_mesa_src_reg src0
);
246 void ir_to_mesa_emit_scalar_op2(ir_instruction
*ir
,
248 ir_to_mesa_dst_reg dst
,
249 ir_to_mesa_src_reg src0
,
250 ir_to_mesa_src_reg src1
);
252 GLboolean
try_emit_mad(ir_expression
*ir
,
255 void add_aggregate_uniform(ir_instruction
*ir
,
257 const struct glsl_type
*type
,
258 struct ir_to_mesa_dst_reg temp
);
261 int sampler_map_size
;
263 void map_sampler(int location
, int sampler
);
264 int get_sampler_number(int location
);
269 ir_to_mesa_src_reg ir_to_mesa_undef
= ir_to_mesa_src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
271 ir_to_mesa_dst_reg ir_to_mesa_undef_dst
= {
272 PROGRAM_UNDEFINED
, 0, SWIZZLE_NOOP
, COND_TR
, NULL
,
275 ir_to_mesa_dst_reg ir_to_mesa_address_reg
= {
276 PROGRAM_ADDRESS
, 0, WRITEMASK_X
, COND_TR
, NULL
279 static int swizzle_for_size(int size
)
281 int size_swizzles
[4] = {
282 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
283 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
284 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
285 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
288 return size_swizzles
[size
- 1];
291 ir_to_mesa_instruction
*
292 ir_to_mesa_visitor::ir_to_mesa_emit_op3(ir_instruction
*ir
,
294 ir_to_mesa_dst_reg dst
,
295 ir_to_mesa_src_reg src0
,
296 ir_to_mesa_src_reg src1
,
297 ir_to_mesa_src_reg src2
)
299 ir_to_mesa_instruction
*inst
= new(mem_ctx
) ir_to_mesa_instruction();
302 /* If we have to do relative addressing, we want to load the ARL
303 * reg directly for one of the regs, and preload the other reladdr
304 * sources into temps.
306 num_reladdr
+= dst
.reladdr
!= NULL
;
307 num_reladdr
+= src0
.reladdr
!= NULL
;
308 num_reladdr
+= src1
.reladdr
!= NULL
;
309 num_reladdr
+= src2
.reladdr
!= NULL
;
311 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
312 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
313 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
316 ir_to_mesa_emit_op1(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
,
321 assert(num_reladdr
== 0);
325 inst
->src_reg
[0] = src0
;
326 inst
->src_reg
[1] = src1
;
327 inst
->src_reg
[2] = src2
;
330 inst
->function
= NULL
;
332 this->instructions
.push_tail(inst
);
338 ir_to_mesa_instruction
*
339 ir_to_mesa_visitor::ir_to_mesa_emit_op2(ir_instruction
*ir
,
341 ir_to_mesa_dst_reg dst
,
342 ir_to_mesa_src_reg src0
,
343 ir_to_mesa_src_reg src1
)
345 return ir_to_mesa_emit_op3(ir
, op
, dst
, src0
, src1
, ir_to_mesa_undef
);
348 ir_to_mesa_instruction
*
349 ir_to_mesa_visitor::ir_to_mesa_emit_op1(ir_instruction
*ir
,
351 ir_to_mesa_dst_reg dst
,
352 ir_to_mesa_src_reg src0
)
354 return ir_to_mesa_emit_op3(ir
, op
, dst
,
355 src0
, ir_to_mesa_undef
, ir_to_mesa_undef
);
358 ir_to_mesa_instruction
*
359 ir_to_mesa_visitor::ir_to_mesa_emit_op0(ir_instruction
*ir
,
362 return ir_to_mesa_emit_op3(ir
, op
, ir_to_mesa_undef_dst
,
369 ir_to_mesa_visitor::map_sampler(int location
, int sampler
)
371 if (this->sampler_map_size
<= location
) {
372 this->sampler_map
= talloc_realloc(this->mem_ctx
, this->sampler_map
,
374 this->sampler_map_size
= location
+ 1;
377 this->sampler_map
[location
] = sampler
;
381 ir_to_mesa_visitor::get_sampler_number(int location
)
383 assert(location
< this->sampler_map_size
);
384 return this->sampler_map
[location
];
387 inline ir_to_mesa_dst_reg
388 ir_to_mesa_dst_reg_from_src(ir_to_mesa_src_reg reg
)
390 ir_to_mesa_dst_reg dst_reg
;
392 dst_reg
.file
= reg
.file
;
393 dst_reg
.index
= reg
.index
;
394 dst_reg
.writemask
= WRITEMASK_XYZW
;
395 dst_reg
.cond_mask
= COND_TR
;
396 dst_reg
.reladdr
= reg
.reladdr
;
401 inline ir_to_mesa_src_reg
402 ir_to_mesa_src_reg_from_dst(ir_to_mesa_dst_reg reg
)
404 return ir_to_mesa_src_reg(reg
.file
, reg
.index
, NULL
);
408 * Emits Mesa scalar opcodes to produce unique answers across channels.
410 * Some Mesa opcodes are scalar-only, like ARB_fp/vp. The src X
411 * channel determines the result across all channels. So to do a vec4
412 * of this operation, we want to emit a scalar per source channel used
413 * to produce dest channels.
416 ir_to_mesa_visitor::ir_to_mesa_emit_scalar_op2(ir_instruction
*ir
,
418 ir_to_mesa_dst_reg dst
,
419 ir_to_mesa_src_reg orig_src0
,
420 ir_to_mesa_src_reg orig_src1
)
423 int done_mask
= ~dst
.writemask
;
425 /* Mesa RCP is a scalar operation splatting results to all channels,
426 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
429 for (i
= 0; i
< 4; i
++) {
430 GLuint this_mask
= (1 << i
);
431 ir_to_mesa_instruction
*inst
;
432 ir_to_mesa_src_reg src0
= orig_src0
;
433 ir_to_mesa_src_reg src1
= orig_src1
;
435 if (done_mask
& this_mask
)
438 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
439 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
440 for (j
= i
+ 1; j
< 4; j
++) {
441 if (!(done_mask
& (1 << j
)) &&
442 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
443 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
444 this_mask
|= (1 << j
);
447 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
448 src0_swiz
, src0_swiz
);
449 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
450 src1_swiz
, src1_swiz
);
452 inst
= ir_to_mesa_emit_op2(ir
, op
,
456 inst
->dst_reg
.writemask
= this_mask
;
457 done_mask
|= this_mask
;
462 ir_to_mesa_visitor::ir_to_mesa_emit_scalar_op1(ir_instruction
*ir
,
464 ir_to_mesa_dst_reg dst
,
465 ir_to_mesa_src_reg src0
)
467 ir_to_mesa_src_reg undef
= ir_to_mesa_undef
;
469 undef
.swizzle
= SWIZZLE_XXXX
;
471 ir_to_mesa_emit_scalar_op2(ir
, op
, dst
, src0
, undef
);
474 struct ir_to_mesa_src_reg
475 ir_to_mesa_visitor::src_reg_for_float(float val
)
477 ir_to_mesa_src_reg
src_reg(PROGRAM_CONSTANT
, -1, NULL
);
479 src_reg
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
480 &val
, 1, &src_reg
.swizzle
);
486 type_size(const struct glsl_type
*type
)
491 switch (type
->base_type
) {
494 case GLSL_TYPE_FLOAT
:
496 if (type
->is_matrix()) {
497 return type
->matrix_columns
;
499 /* Regardless of size of vector, it gets a vec4. This is bad
500 * packing for things like floats, but otherwise arrays become a
501 * mess. Hopefully a later pass over the code can pack scalars
502 * down if appropriate.
506 case GLSL_TYPE_ARRAY
:
507 return type_size(type
->fields
.array
) * type
->length
;
508 case GLSL_TYPE_STRUCT
:
510 for (i
= 0; i
< type
->length
; i
++) {
511 size
+= type_size(type
->fields
.structure
[i
].type
);
520 * In the initial pass of codegen, we assign temporary numbers to
521 * intermediate results. (not SSA -- variable assignments will reuse
522 * storage). Actual register allocation for the Mesa VM occurs in a
523 * pass over the Mesa IR later.
526 ir_to_mesa_visitor::get_temp(const glsl_type
*type
)
528 ir_to_mesa_src_reg src_reg
;
532 src_reg
.file
= PROGRAM_TEMPORARY
;
533 src_reg
.index
= next_temp
;
534 src_reg
.reladdr
= NULL
;
535 next_temp
+= type_size(type
);
537 if (type
->is_array() || type
->is_record()) {
538 src_reg
.swizzle
= SWIZZLE_NOOP
;
540 for (i
= 0; i
< type
->vector_elements
; i
++)
543 swizzle
[i
] = type
->vector_elements
- 1;
544 src_reg
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1],
545 swizzle
[2], swizzle
[3]);
553 ir_to_mesa_visitor::find_variable_storage(ir_variable
*var
)
556 variable_storage
*entry
;
558 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
559 entry
= (variable_storage
*)iter
.get();
561 if (entry
->var
== var
)
569 ir_to_mesa_visitor::visit(ir_variable
*ir
)
571 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
572 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
574 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
575 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
580 ir_to_mesa_visitor::visit(ir_loop
*ir
)
584 assert(!ir
->increment
);
585 assert(!ir
->counter
);
587 ir_to_mesa_emit_op0(NULL
, OPCODE_BGNLOOP
);
588 visit_exec_list(&ir
->body_instructions
, this);
589 ir_to_mesa_emit_op0(NULL
, OPCODE_ENDLOOP
);
593 ir_to_mesa_visitor::visit(ir_loop_jump
*ir
)
596 case ir_loop_jump::jump_break
:
597 ir_to_mesa_emit_op0(NULL
, OPCODE_BRK
);
599 case ir_loop_jump::jump_continue
:
600 ir_to_mesa_emit_op0(NULL
, OPCODE_CONT
);
607 ir_to_mesa_visitor::visit(ir_function_signature
*ir
)
614 ir_to_mesa_visitor::visit(ir_function
*ir
)
616 /* Ignore function bodies other than main() -- we shouldn't see calls to
617 * them since they should all be inlined before we get to ir_to_mesa.
619 if (strcmp(ir
->name
, "main") == 0) {
620 const ir_function_signature
*sig
;
623 sig
= ir
->matching_signature(&empty
);
627 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
628 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
636 ir_to_mesa_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
638 int nonmul_operand
= 1 - mul_operand
;
639 ir_to_mesa_src_reg a
, b
, c
;
641 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
642 if (!expr
|| expr
->operation
!= ir_binop_mul
)
645 expr
->operands
[0]->accept(this);
647 expr
->operands
[1]->accept(this);
649 ir
->operands
[nonmul_operand
]->accept(this);
652 this->result
= get_temp(ir
->type
);
653 ir_to_mesa_emit_op3(ir
, OPCODE_MAD
,
654 ir_to_mesa_dst_reg_from_src(this->result
), a
, b
, c
);
660 ir_to_mesa_visitor::reladdr_to_temp(ir_instruction
*ir
,
661 ir_to_mesa_src_reg
*reg
, int *num_reladdr
)
666 ir_to_mesa_emit_op1(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
, *reg
->reladdr
);
668 if (*num_reladdr
!= 1) {
669 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
671 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
,
672 ir_to_mesa_dst_reg_from_src(temp
), *reg
);
680 ir_to_mesa_visitor::visit(ir_expression
*ir
)
682 unsigned int operand
;
683 struct ir_to_mesa_src_reg op
[2];
684 struct ir_to_mesa_src_reg result_src
;
685 struct ir_to_mesa_dst_reg result_dst
;
686 const glsl_type
*vec4_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
, 4, 1);
687 const glsl_type
*vec3_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
, 3, 1);
688 const glsl_type
*vec2_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
, 2, 1);
690 /* Quick peephole: Emit OPCODE_MAD(a, b, c) instead of ADD(MUL(a, b), c)
692 if (ir
->operation
== ir_binop_add
) {
693 if (try_emit_mad(ir
, 1))
695 if (try_emit_mad(ir
, 0))
699 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
700 this->result
.file
= PROGRAM_UNDEFINED
;
701 ir
->operands
[operand
]->accept(this);
702 if (this->result
.file
== PROGRAM_UNDEFINED
) {
704 printf("Failed to get tree for expression operand:\n");
705 ir
->operands
[operand
]->accept(&v
);
708 op
[operand
] = this->result
;
710 /* Matrix expression operands should have been broken down to vector
711 * operations already.
713 assert(!ir
->operands
[operand
]->type
->is_matrix());
716 this->result
.file
= PROGRAM_UNDEFINED
;
718 /* Storage for our result. Ideally for an assignment we'd be using
719 * the actual storage for the result here, instead.
721 result_src
= get_temp(ir
->type
);
722 /* convenience for the emit functions below. */
723 result_dst
= ir_to_mesa_dst_reg_from_src(result_src
);
724 /* Limit writes to the channels that will be used by result_src later.
725 * This does limit this temp's use as a temporary for multi-instruction
728 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
730 switch (ir
->operation
) {
731 case ir_unop_logic_not
:
732 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
,
733 op
[0], src_reg_for_float(0.0));
736 op
[0].negate
= ~op
[0].negate
;
740 ir_to_mesa_emit_op1(ir
, OPCODE_ABS
, result_dst
, op
[0]);
743 ir_to_mesa_emit_op1(ir
, OPCODE_SSG
, result_dst
, op
[0]);
746 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RCP
, result_dst
, op
[0]);
750 ir_to_mesa_emit_scalar_op2(ir
, OPCODE_POW
, result_dst
,
751 src_reg_for_float(M_E
), op
[0]);
754 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_EX2
, result_dst
, op
[0]);
757 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_LOG
, result_dst
, op
[0]);
760 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_LG2
, result_dst
, op
[0]);
763 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_SIN
, result_dst
, op
[0]);
766 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_COS
, result_dst
, op
[0]);
770 ir_to_mesa_emit_op1(ir
, OPCODE_DDX
, result_dst
, op
[0]);
773 ir_to_mesa_emit_op1(ir
, OPCODE_DDY
, result_dst
, op
[0]);
777 ir_to_mesa_emit_op2(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
780 ir_to_mesa_emit_op2(ir
, OPCODE_SUB
, result_dst
, op
[0], op
[1]);
784 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
787 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
789 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
793 ir_to_mesa_emit_op2(ir
, OPCODE_SLT
, result_dst
, op
[0], op
[1]);
795 case ir_binop_greater
:
796 ir_to_mesa_emit_op2(ir
, OPCODE_SGT
, result_dst
, op
[0], op
[1]);
798 case ir_binop_lequal
:
799 ir_to_mesa_emit_op2(ir
, OPCODE_SLE
, result_dst
, op
[0], op
[1]);
801 case ir_binop_gequal
:
802 ir_to_mesa_emit_op2(ir
, OPCODE_SGE
, result_dst
, op
[0], op
[1]);
805 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
807 case ir_binop_logic_xor
:
808 case ir_binop_nequal
:
809 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
812 case ir_binop_logic_or
:
813 /* This could be a saturated add and skip the SNE. */
814 ir_to_mesa_emit_op2(ir
, OPCODE_ADD
,
818 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
820 result_src
, src_reg_for_float(0.0));
823 case ir_binop_logic_and
:
824 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
825 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
831 if (ir
->operands
[0]->type
== vec4_type
) {
832 assert(ir
->operands
[1]->type
== vec4_type
);
833 ir_to_mesa_emit_op2(ir
, OPCODE_DP4
,
836 } else if (ir
->operands
[0]->type
== vec3_type
) {
837 assert(ir
->operands
[1]->type
== vec3_type
);
838 ir_to_mesa_emit_op2(ir
, OPCODE_DP3
,
841 } else if (ir
->operands
[0]->type
== vec2_type
) {
842 assert(ir
->operands
[1]->type
== vec2_type
);
843 ir_to_mesa_emit_op2(ir
, OPCODE_DP2
,
850 ir_to_mesa_emit_op2(ir
, OPCODE_XPD
, result_dst
, op
[0], op
[1]);
854 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
855 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RCP
, result_dst
, result_src
);
856 /* For incoming channels < 0, set the result to 0. */
857 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, result_dst
,
858 op
[0], src_reg_for_float(0.0), result_src
);
861 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
866 /* Mesa IR lacks types, ints are stored as truncated floats. */
870 ir_to_mesa_emit_op1(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
874 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
,
875 result_src
, src_reg_for_float(0.0));
878 ir_to_mesa_emit_op1(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
881 op
[0].negate
= ~op
[0].negate
;
882 ir_to_mesa_emit_op1(ir
, OPCODE_FLR
, result_dst
, op
[0]);
883 result_src
.negate
= ~result_src
.negate
;
886 ir_to_mesa_emit_op1(ir
, OPCODE_FLR
, result_dst
, op
[0]);
889 ir_to_mesa_emit_op1(ir
, OPCODE_FRC
, result_dst
, op
[0]);
893 ir_to_mesa_emit_op2(ir
, OPCODE_MIN
, result_dst
, op
[0], op
[1]);
896 ir_to_mesa_emit_op2(ir
, OPCODE_MAX
, result_dst
, op
[0], op
[1]);
899 ir_to_mesa_emit_scalar_op2(ir
, OPCODE_POW
, result_dst
, op
[0], op
[1]);
902 case ir_unop_bit_not
:
904 case ir_binop_lshift
:
905 case ir_binop_rshift
:
906 case ir_binop_bit_and
:
907 case ir_binop_bit_xor
:
908 case ir_binop_bit_or
:
909 assert(!"GLSL 1.30 features unsupported");
913 this->result
= result_src
;
918 ir_to_mesa_visitor::visit(ir_swizzle
*ir
)
920 ir_to_mesa_src_reg src_reg
;
924 /* Note that this is only swizzles in expressions, not those on the left
925 * hand side of an assignment, which do write masking. See ir_assignment
929 ir
->val
->accept(this);
930 src_reg
= this->result
;
931 assert(src_reg
.file
!= PROGRAM_UNDEFINED
);
933 for (i
= 0; i
< 4; i
++) {
934 if (i
< ir
->type
->vector_elements
) {
937 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.x
);
940 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.y
);
943 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.z
);
946 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.w
);
950 /* If the type is smaller than a vec4, replicate the last
953 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
957 src_reg
.swizzle
= MAKE_SWIZZLE4(swizzle
[0],
962 this->result
= src_reg
;
965 static const struct {
968 int tokens
[STATE_LENGTH
];
972 {"gl_DepthRange", "near",
973 {STATE_DEPTH_RANGE
, 0, 0}, SWIZZLE_XXXX
},
974 {"gl_DepthRange", "far",
975 {STATE_DEPTH_RANGE
, 0, 0}, SWIZZLE_YYYY
},
976 {"gl_DepthRange", "diff",
977 {STATE_DEPTH_RANGE
, 0, 0}, SWIZZLE_ZZZZ
},
979 {"gl_ClipPlane", NULL
,
980 {STATE_CLIPPLANE
, 0, 0}, SWIZZLE_XYZW
, true}
983 {STATE_POINT_SIZE
}, SWIZZLE_XXXX
},
984 {"gl_Point", "sizeMin",
985 {STATE_POINT_SIZE
}, SWIZZLE_YYYY
},
986 {"gl_Point", "sizeMax",
987 {STATE_POINT_SIZE
}, SWIZZLE_ZZZZ
},
988 {"gl_Point", "fadeThresholdSize",
989 {STATE_POINT_SIZE
}, SWIZZLE_WWWW
},
990 {"gl_Point", "distanceConstantAttenuation",
991 {STATE_POINT_ATTENUATION
}, SWIZZLE_XXXX
},
992 {"gl_Point", "distanceLinearAttenuation",
993 {STATE_POINT_ATTENUATION
}, SWIZZLE_YYYY
},
994 {"gl_Point", "distanceQuadraticAttenuation",
995 {STATE_POINT_ATTENUATION
}, SWIZZLE_ZZZZ
},
997 {"gl_FrontMaterial", "emission",
998 {STATE_MATERIAL
, 0, STATE_EMISSION
}, SWIZZLE_XYZW
},
999 {"gl_FrontMaterial", "ambient",
1000 {STATE_MATERIAL
, 0, STATE_AMBIENT
}, SWIZZLE_XYZW
},
1001 {"gl_FrontMaterial", "diffuse",
1002 {STATE_MATERIAL
, 0, STATE_DIFFUSE
}, SWIZZLE_XYZW
},
1003 {"gl_FrontMaterial", "specular",
1004 {STATE_MATERIAL
, 0, STATE_SPECULAR
}, SWIZZLE_XYZW
},
1005 {"gl_FrontMaterial", "shininess",
1006 {STATE_MATERIAL
, 0, STATE_SHININESS
}, SWIZZLE_XXXX
},
1008 {"gl_BackMaterial", "emission",
1009 {STATE_MATERIAL
, 1, STATE_EMISSION
}, SWIZZLE_XYZW
},
1010 {"gl_BackMaterial", "ambient",
1011 {STATE_MATERIAL
, 1, STATE_AMBIENT
}, SWIZZLE_XYZW
},
1012 {"gl_BackMaterial", "diffuse",
1013 {STATE_MATERIAL
, 1, STATE_DIFFUSE
}, SWIZZLE_XYZW
},
1014 {"gl_BackMaterial", "specular",
1015 {STATE_MATERIAL
, 1, STATE_SPECULAR
}, SWIZZLE_XYZW
},
1016 {"gl_BackMaterial", "shininess",
1017 {STATE_MATERIAL
, 1, STATE_SHININESS
}, SWIZZLE_XXXX
},
1019 {"gl_LightSource", "ambient",
1020 {STATE_LIGHT
, 0, STATE_AMBIENT
}, SWIZZLE_XYZW
, true},
1021 {"gl_LightSource", "diffuse",
1022 {STATE_LIGHT
, 0, STATE_DIFFUSE
}, SWIZZLE_XYZW
, true},
1023 {"gl_LightSource", "specular",
1024 {STATE_LIGHT
, 0, STATE_SPECULAR
}, SWIZZLE_XYZW
, true},
1025 {"gl_LightSource", "position",
1026 {STATE_LIGHT
, 0, STATE_POSITION
}, SWIZZLE_XYZW
, true},
1027 {"gl_LightSource", "halfVector",
1028 {STATE_LIGHT
, 0, STATE_HALF_VECTOR
}, SWIZZLE_XYZW
, true},
1029 {"gl_LightSource", "spotDirection",
1030 {STATE_LIGHT
, 0, STATE_SPOT_DIRECTION
}, SWIZZLE_XYZW
, true},
1031 {"gl_LightSource", "spotCosCutoff",
1032 {STATE_LIGHT
, 0, STATE_SPOT_DIRECTION
}, SWIZZLE_WWWW
, true},
1033 {"gl_LightSource", "spotCutoff",
1034 {STATE_LIGHT
, 0, STATE_SPOT_CUTOFF
}, SWIZZLE_XXXX
, true},
1035 {"gl_LightSource", "spotExponent",
1036 {STATE_LIGHT
, 0, STATE_ATTENUATION
}, SWIZZLE_WWWW
, true},
1037 {"gl_LightSource", "constantAttenuation",
1038 {STATE_LIGHT
, 0, STATE_ATTENUATION
}, SWIZZLE_XXXX
, true},
1039 {"gl_LightSource", "linearAttenuation",
1040 {STATE_LIGHT
, 0, STATE_ATTENUATION
}, SWIZZLE_YYYY
, true},
1041 {"gl_LightSource", "quadraticAttenuation",
1042 {STATE_LIGHT
, 0, STATE_ATTENUATION
}, SWIZZLE_ZZZZ
, true},
1044 {"gl_LightModel", NULL
,
1045 {STATE_LIGHTMODEL_AMBIENT
, 0}, SWIZZLE_XYZW
},
1047 {"gl_FrontLightModelProduct", NULL
,
1048 {STATE_LIGHTMODEL_SCENECOLOR
, 0}, SWIZZLE_XYZW
},
1049 {"gl_BackLightModelProduct", NULL
,
1050 {STATE_LIGHTMODEL_SCENECOLOR
, 1}, SWIZZLE_XYZW
},
1052 {"gl_FrontLightProduct", "ambient",
1053 {STATE_LIGHTPROD
, 0, 0, STATE_AMBIENT
}, SWIZZLE_XYZW
, true},
1054 {"gl_FrontLightProduct", "diffuse",
1055 {STATE_LIGHTPROD
, 0, 0, STATE_DIFFUSE
}, SWIZZLE_XYZW
, true},
1056 {"gl_FrontLightProduct", "specular",
1057 {STATE_LIGHTPROD
, 0, 0, STATE_SPECULAR
}, SWIZZLE_XYZW
, true},
1059 {"gl_BackLightProduct", "ambient",
1060 {STATE_LIGHTPROD
, 0, 1, STATE_AMBIENT
}, SWIZZLE_XYZW
, true},
1061 {"gl_BackLightProduct", "diffuse",
1062 {STATE_LIGHTPROD
, 0, 1, STATE_DIFFUSE
}, SWIZZLE_XYZW
, true},
1063 {"gl_BackLightProduct", "specular",
1064 {STATE_LIGHTPROD
, 0, 1, STATE_SPECULAR
}, SWIZZLE_XYZW
, true},
1066 {"gl_TextureEnvColor", "ambient",
1067 {STATE_TEXENV_COLOR
, 0}, SWIZZLE_XYZW
, true},
1069 {"gl_EyePlaneS", NULL
,
1070 {STATE_TEXGEN
, 0, STATE_TEXGEN_EYE_S
}, SWIZZLE_XYZW
, true},
1071 {"gl_EyePlaneT", NULL
,
1072 {STATE_TEXGEN
, 0, STATE_TEXGEN_EYE_T
}, SWIZZLE_XYZW
, true},
1073 {"gl_EyePlaneR", NULL
,
1074 {STATE_TEXGEN
, 0, STATE_TEXGEN_EYE_R
}, SWIZZLE_XYZW
, true},
1075 {"gl_EyePlaneQ", NULL
,
1076 {STATE_TEXGEN
, 0, STATE_TEXGEN_EYE_Q
}, SWIZZLE_XYZW
, true},
1078 {"gl_ObjectPlaneS", NULL
,
1079 {STATE_TEXGEN
, 0, STATE_TEXGEN_OBJECT_S
}, SWIZZLE_XYZW
, true},
1080 {"gl_ObjectPlaneT", NULL
,
1081 {STATE_TEXGEN
, 0, STATE_TEXGEN_OBJECT_T
}, SWIZZLE_XYZW
, true},
1082 {"gl_ObjectPlaneR", NULL
,
1083 {STATE_TEXGEN
, 0, STATE_TEXGEN_OBJECT_R
}, SWIZZLE_XYZW
, true},
1084 {"gl_ObjectPlaneQ", NULL
,
1085 {STATE_TEXGEN
, 0, STATE_TEXGEN_OBJECT_Q
}, SWIZZLE_XYZW
, true},
1088 {STATE_FOG_COLOR
}, SWIZZLE_XYZW
},
1089 {"gl_Fog", "density",
1090 {STATE_FOG_PARAMS
}, SWIZZLE_XXXX
},
1092 {STATE_FOG_PARAMS
}, SWIZZLE_YYYY
},
1094 {STATE_FOG_PARAMS
}, SWIZZLE_ZZZZ
},
1096 {STATE_FOG_PARAMS
}, SWIZZLE_WWWW
},
1099 static ir_to_mesa_src_reg
1100 get_builtin_uniform_reg(struct gl_program
*prog
,
1101 const char *name
, int array_index
, const char *field
)
1104 ir_to_mesa_src_reg src_reg
;
1105 int tokens
[STATE_LENGTH
];
1107 for (i
= 0; i
< Elements(statevars
); i
++) {
1108 if (strcmp(statevars
[i
].name
, name
) != 0)
1110 if (!field
&& statevars
[i
].field
) {
1111 assert(!"FINISHME: whole-structure state var dereference");
1113 if (field
&& strcmp(statevars
[i
].field
, field
) != 0)
1118 if (i
== Elements(statevars
)) {
1119 printf("builtin uniform %s%s%s not found\n",
1122 field
? field
: "");
1126 memcpy(&tokens
, statevars
[i
].tokens
, sizeof(tokens
));
1127 if (statevars
[i
].array_indexed
)
1128 tokens
[1] = array_index
;
1130 src_reg
.file
= PROGRAM_STATE_VAR
;
1131 src_reg
.index
= _mesa_add_state_reference(prog
->Parameters
,
1132 (gl_state_index
*)tokens
);
1133 src_reg
.swizzle
= statevars
[i
].swizzle
;
1135 src_reg
.reladdr
= false;
1141 add_matrix_ref(struct gl_program
*prog
, int *tokens
)
1146 /* Add a ref for each column. It looks like the reason we do
1147 * it this way is that _mesa_add_state_reference doesn't work
1148 * for things that aren't vec4s, so the tokens[2]/tokens[3]
1149 * range has to be equal.
1151 for (i
= 0; i
< 4; i
++) {
1154 int pos
= _mesa_add_state_reference(prog
->Parameters
,
1155 (gl_state_index
*)tokens
);
1159 assert(base_pos
+ i
== pos
);
1165 static variable_storage
*
1166 get_builtin_matrix_ref(void *mem_ctx
, struct gl_program
*prog
, ir_variable
*var
,
1167 ir_rvalue
*array_index
)
1170 * NOTE: The ARB_vertex_program extension specified that matrices get
1171 * loaded in registers in row-major order. With GLSL, we want column-
1172 * major order. So, we need to transpose all matrices here...
1174 static const struct {
1179 { "gl_ModelViewMatrix", STATE_MODELVIEW_MATRIX
, STATE_MATRIX_TRANSPOSE
},
1180 { "gl_ModelViewMatrixInverse", STATE_MODELVIEW_MATRIX
, STATE_MATRIX_INVTRANS
},
1181 { "gl_ModelViewMatrixTranspose", STATE_MODELVIEW_MATRIX
, 0 },
1182 { "gl_ModelViewMatrixInverseTranspose", STATE_MODELVIEW_MATRIX
, STATE_MATRIX_INVERSE
},
1184 { "gl_ProjectionMatrix", STATE_PROJECTION_MATRIX
, STATE_MATRIX_TRANSPOSE
},
1185 { "gl_ProjectionMatrixInverse", STATE_PROJECTION_MATRIX
, STATE_MATRIX_INVTRANS
},
1186 { "gl_ProjectionMatrixTranspose", STATE_PROJECTION_MATRIX
, 0 },
1187 { "gl_ProjectionMatrixInverseTranspose", STATE_PROJECTION_MATRIX
, STATE_MATRIX_INVERSE
},
1189 { "gl_ModelViewProjectionMatrix", STATE_MVP_MATRIX
, STATE_MATRIX_TRANSPOSE
},
1190 { "gl_ModelViewProjectionMatrixInverse", STATE_MVP_MATRIX
, STATE_MATRIX_INVTRANS
},
1191 { "gl_ModelViewProjectionMatrixTranspose", STATE_MVP_MATRIX
, 0 },
1192 { "gl_ModelViewProjectionMatrixInverseTranspose", STATE_MVP_MATRIX
, STATE_MATRIX_INVERSE
},
1194 { "gl_TextureMatrix", STATE_TEXTURE_MATRIX
, STATE_MATRIX_TRANSPOSE
},
1195 { "gl_TextureMatrixInverse", STATE_TEXTURE_MATRIX
, STATE_MATRIX_INVTRANS
},
1196 { "gl_TextureMatrixTranspose", STATE_TEXTURE_MATRIX
, 0 },
1197 { "gl_TextureMatrixInverseTranspose", STATE_TEXTURE_MATRIX
, STATE_MATRIX_INVERSE
},
1199 { "gl_NormalMatrix", STATE_MODELVIEW_MATRIX
, STATE_MATRIX_INVERSE
},
1203 variable_storage
*entry
;
1205 /* C++ gets angry when we try to use an int as a gl_state_index, so we use
1206 * ints for gl_state_index. Make sure they're compatible.
1208 assert(sizeof(gl_state_index
) == sizeof(int));
1210 for (i
= 0; i
< Elements(matrices
); i
++) {
1211 if (strcmp(var
->name
, matrices
[i
].name
) == 0) {
1212 int tokens
[STATE_LENGTH
];
1215 tokens
[0] = matrices
[i
].matrix
;
1216 tokens
[4] = matrices
[i
].modifier
;
1217 if (matrices
[i
].matrix
== STATE_TEXTURE_MATRIX
) {
1218 ir_constant
*index
= array_index
->constant_expression_value();
1220 tokens
[1] = index
->value
.i
[0];
1221 base_pos
= add_matrix_ref(prog
, tokens
);
1223 for (i
= 0; i
< var
->type
->length
; i
++) {
1225 int pos
= add_matrix_ref(prog
, tokens
);
1229 assert(base_pos
+ (int)i
* 4 == pos
);
1233 tokens
[1] = 0; /* unused array index */
1234 base_pos
= add_matrix_ref(prog
, tokens
);
1236 tokens
[4] = matrices
[i
].modifier
;
1238 entry
= new(mem_ctx
) variable_storage(var
,
1249 /* Recursively add all the members of the aggregate uniform as uniform names
1250 * to Mesa, moving those uniforms to our structured temporary.
1253 ir_to_mesa_visitor::add_aggregate_uniform(ir_instruction
*ir
,
1255 const struct glsl_type
*type
,
1256 struct ir_to_mesa_dst_reg temp
)
1260 if (type
->is_record()) {
1261 void *mem_ctx
= talloc_new(NULL
);
1263 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1264 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
1265 add_aggregate_uniform(ir
,
1266 talloc_asprintf(mem_ctx
, "%s.%s", name
,
1267 type
->fields
.structure
[i
].name
),
1269 temp
.index
+= type_size(field_type
);
1272 talloc_free(mem_ctx
);
1277 assert(type
->is_vector() || type
->is_scalar() || !"FINISHME: other types");
1281 if (type
->is_vector() ||
1282 type
->is_scalar()) {
1283 len
= type
->vector_elements
;
1285 len
= type_size(type
) * 4;
1288 loc
= _mesa_add_uniform(this->prog
->Parameters
,
1295 ir_to_mesa_src_reg
uniform(PROGRAM_UNIFORM
, loc
, type
);
1297 for (int i
= 0; i
< type_size(type
); i
++) {
1298 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, uniform
);
1306 ir_to_mesa_visitor::visit(ir_dereference_variable
*ir
)
1308 variable_storage
*entry
= find_variable_storage(ir
->var
);
1313 switch (ir
->var
->mode
) {
1314 case ir_var_uniform
:
1315 entry
= get_builtin_matrix_ref(this->mem_ctx
, this->prog
, ir
->var
,
1320 /* FINISHME: Fix up uniform name for arrays and things */
1321 if (ir
->var
->type
->base_type
== GLSL_TYPE_SAMPLER
) {
1322 /* FINISHME: we whack the location of the var here, which
1323 * is probably not expected. But we need to communicate
1324 * mesa's sampler number to the tex instruction.
1326 int sampler
= _mesa_add_sampler(this->prog
->Parameters
,
1328 ir
->var
->type
->gl_type
);
1329 map_sampler(ir
->var
->location
, sampler
);
1331 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_SAMPLER
,
1333 this->variables
.push_tail(entry
);
1337 assert(ir
->var
->type
->gl_type
!= 0 &&
1338 ir
->var
->type
->gl_type
!= GL_INVALID_ENUM
);
1340 /* Oh, the joy of aggregate types in Mesa. Like constants,
1341 * we can only really do vec4s. So, make a temp, chop the
1342 * aggregate up into vec4s, and move those vec4s to the temp.
1344 if (ir
->var
->type
->is_record()) {
1345 ir_to_mesa_src_reg temp
= get_temp(ir
->var
->type
);
1347 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1350 this->variables
.push_tail(entry
);
1352 add_aggregate_uniform(ir
->var
, ir
->var
->name
, ir
->var
->type
,
1353 ir_to_mesa_dst_reg_from_src(temp
));
1357 if (ir
->var
->type
->is_vector() ||
1358 ir
->var
->type
->is_scalar()) {
1359 len
= ir
->var
->type
->vector_elements
;
1361 len
= type_size(ir
->var
->type
) * 4;
1364 loc
= _mesa_add_uniform(this->prog
->Parameters
,
1367 ir
->var
->type
->gl_type
,
1370 /* Always mark the uniform used at this point. If it isn't
1371 * used, dead code elimination should have nuked the decl already.
1373 this->prog
->Parameters
->Parameters
[loc
].Used
= GL_TRUE
;
1375 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_UNIFORM
, loc
);
1376 this->variables
.push_tail(entry
);
1381 /* The linker assigns locations for varyings and attributes,
1382 * including deprecated builtins (like gl_Color), user-assign
1383 * generic attributes (glBindVertexLocation), and
1384 * user-defined varyings.
1386 * FINISHME: We would hit this path for function arguments. Fix!
1388 assert(ir
->var
->location
!= -1);
1389 if (ir
->var
->mode
== ir_var_in
||
1390 ir
->var
->mode
== ir_var_inout
) {
1391 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1395 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1396 ir
->var
->location
>= VERT_ATTRIB_GENERIC0
) {
1397 _mesa_add_attribute(prog
->Attributes
,
1399 type_size(ir
->var
->type
) * 4,
1400 ir
->var
->type
->gl_type
,
1401 ir
->var
->location
- VERT_ATTRIB_GENERIC0
);
1404 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1411 case ir_var_temporary
:
1412 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_TEMPORARY
,
1414 this->variables
.push_tail(entry
);
1416 next_temp
+= type_size(ir
->var
->type
);
1421 printf("Failed to make storage for %s\n", ir
->var
->name
);
1426 this->result
= ir_to_mesa_src_reg(entry
->file
, entry
->index
, ir
->var
->type
);
1430 ir_to_mesa_visitor::visit(ir_dereference_array
*ir
)
1432 ir_variable
*var
= ir
->variable_referenced();
1434 ir_to_mesa_src_reg src_reg
;
1435 ir_dereference_variable
*deref_var
= ir
->array
->as_dereference_variable();
1436 int element_size
= type_size(ir
->type
);
1438 index
= ir
->array_index
->constant_expression_value();
1440 if (deref_var
&& strncmp(deref_var
->var
->name
,
1442 strlen("gl_TextureMatrix")) == 0) {
1443 struct variable_storage
*entry
;
1445 entry
= get_builtin_matrix_ref(this->mem_ctx
, this->prog
, deref_var
->var
,
1449 ir_to_mesa_src_reg
src_reg(entry
->file
, entry
->index
, ir
->type
);
1452 src_reg
.reladdr
= NULL
;
1454 ir_to_mesa_src_reg index_reg
= get_temp(glsl_type::float_type
);
1456 ir
->array_index
->accept(this);
1457 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
1458 ir_to_mesa_dst_reg_from_src(index_reg
),
1459 this->result
, src_reg_for_float(element_size
));
1461 src_reg
.reladdr
= talloc(mem_ctx
, ir_to_mesa_src_reg
);
1462 memcpy(src_reg
.reladdr
, &index_reg
, sizeof(index_reg
));
1465 this->result
= src_reg
;
1469 if (strncmp(var
->name
, "gl_", 3) == 0 && var
->mode
== ir_var_uniform
&&
1470 !var
->type
->is_matrix()) {
1471 ir_dereference_record
*record
= NULL
;
1472 if (ir
->array
->ir_type
== ir_type_dereference_record
)
1473 record
= (ir_dereference_record
*)ir
->array
;
1475 assert(index
|| !"FINISHME: variable-indexed builtin uniform access");
1477 this->result
= get_builtin_uniform_reg(prog
,
1480 record
? record
->field
: NULL
);
1483 ir
->array
->accept(this);
1484 src_reg
= this->result
;
1487 src_reg
.index
+= index
->value
.i
[0] * element_size
;
1489 ir_to_mesa_src_reg array_base
= this->result
;
1490 /* Variable index array dereference. It eats the "vec4" of the
1491 * base of the array and an index that offsets the Mesa register
1494 ir
->array_index
->accept(this);
1496 ir_to_mesa_src_reg index_reg
;
1498 if (element_size
== 1) {
1499 index_reg
= this->result
;
1501 index_reg
= get_temp(glsl_type::float_type
);
1503 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
1504 ir_to_mesa_dst_reg_from_src(index_reg
),
1505 this->result
, src_reg_for_float(element_size
));
1508 src_reg
.reladdr
= talloc(mem_ctx
, ir_to_mesa_src_reg
);
1509 memcpy(src_reg
.reladdr
, &index_reg
, sizeof(index_reg
));
1512 /* If the type is smaller than a vec4, replicate the last channel out. */
1513 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1514 src_reg
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1516 src_reg
.swizzle
= SWIZZLE_NOOP
;
1518 this->result
= src_reg
;
1522 ir_to_mesa_visitor::visit(ir_dereference_record
*ir
)
1525 const glsl_type
*struct_type
= ir
->record
->type
;
1527 ir_variable
*var
= ir
->record
->variable_referenced();
1529 if (strncmp(var
->name
, "gl_", 3) == 0 && var
->mode
== ir_var_uniform
) {
1532 this->result
= get_builtin_uniform_reg(prog
,
1539 ir
->record
->accept(this);
1541 for (i
= 0; i
< struct_type
->length
; i
++) {
1542 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1544 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1546 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1547 this->result
.index
+= offset
;
1551 * We want to be careful in assignment setup to hit the actual storage
1552 * instead of potentially using a temporary like we might with the
1553 * ir_dereference handler.
1555 * Thanks to ir_swizzle_swizzle, and ir_vec_index_to_swizzle, we
1556 * should only see potentially one variable array index of a vector,
1557 * and one swizzle, before getting to actual vec4 storage. So handle
1558 * those, then go use ir_dereference to handle the rest.
1560 static struct ir_to_mesa_dst_reg
1561 get_assignment_lhs(ir_instruction
*ir
, ir_to_mesa_visitor
*v
,
1562 ir_to_mesa_src_reg
*r
)
1564 struct ir_to_mesa_dst_reg dst_reg
;
1567 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1568 /* This should have been handled by ir_vec_index_to_cond_assign */
1570 assert(!deref_array
->array
->type
->is_vector());
1573 /* Use the rvalue deref handler for the most part. We'll ignore
1574 * swizzles in it and write swizzles using writemask, though.
1577 dst_reg
= ir_to_mesa_dst_reg_from_src(v
->result
);
1579 if ((swiz
= ir
->as_swizzle())) {
1586 int new_r_swizzle
[4];
1587 int orig_r_swizzle
= r
->swizzle
;
1590 for (i
= 0; i
< 4; i
++) {
1591 new_r_swizzle
[i
] = GET_SWZ(orig_r_swizzle
, 0);
1594 dst_reg
.writemask
= 0;
1595 for (i
= 0; i
< 4; i
++) {
1596 if (i
< swiz
->mask
.num_components
) {
1597 dst_reg
.writemask
|= 1 << swizzles
[i
];
1598 new_r_swizzle
[swizzles
[i
]] = GET_SWZ(orig_r_swizzle
, i
);
1602 r
->swizzle
= MAKE_SWIZZLE4(new_r_swizzle
[0],
1612 ir_to_mesa_visitor::visit(ir_assignment
*ir
)
1614 struct ir_to_mesa_dst_reg l
;
1615 struct ir_to_mesa_src_reg r
;
1618 ir
->rhs
->accept(this);
1621 l
= get_assignment_lhs(ir
->lhs
, this, &r
);
1623 assert(l
.file
!= PROGRAM_UNDEFINED
);
1624 assert(r
.file
!= PROGRAM_UNDEFINED
);
1626 if (ir
->condition
) {
1627 ir_to_mesa_src_reg condition
;
1629 ir
->condition
->accept(this);
1630 condition
= this->result
;
1632 /* We use the OPCODE_CMP (a < 0 ? b : c) for conditional moves,
1633 * and the condition we produced is 0.0 or 1.0. By flipping the
1634 * sign, we can choose which value OPCODE_CMP produces without
1635 * an extra computing the condition.
1637 condition
.negate
= ~condition
.negate
;
1638 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1639 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, l
,
1640 condition
, r
, ir_to_mesa_src_reg_from_dst(l
));
1645 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1646 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1655 ir_to_mesa_visitor::visit(ir_constant
*ir
)
1657 ir_to_mesa_src_reg src_reg
;
1658 GLfloat stack_vals
[4];
1659 GLfloat
*values
= stack_vals
;
1662 /* Unfortunately, 4 floats is all we can get into
1663 * _mesa_add_unnamed_constant. So, make a temp to store an
1664 * aggregate constant and move each constant value into it. If we
1665 * get lucky, copy propagation will eliminate the extra moves.
1668 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1669 ir_to_mesa_src_reg temp_base
= get_temp(ir
->type
);
1670 ir_to_mesa_dst_reg temp
= ir_to_mesa_dst_reg_from_src(temp_base
);
1672 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1673 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1674 int size
= type_size(field_value
->type
);
1678 field_value
->accept(this);
1679 src_reg
= this->result
;
1681 for (i
= 0; i
< (unsigned int)size
; i
++) {
1682 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, src_reg
);
1688 this->result
= temp_base
;
1692 if (ir
->type
->is_array()) {
1693 ir_to_mesa_src_reg temp_base
= get_temp(ir
->type
);
1694 ir_to_mesa_dst_reg temp
= ir_to_mesa_dst_reg_from_src(temp_base
);
1695 int size
= type_size(ir
->type
->fields
.array
);
1699 for (i
= 0; i
< ir
->type
->length
; i
++) {
1700 ir
->array_elements
[i
]->accept(this);
1701 src_reg
= this->result
;
1702 for (int j
= 0; j
< size
; j
++) {
1703 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, src_reg
);
1709 this->result
= temp_base
;
1713 if (ir
->type
->is_matrix()) {
1714 ir_to_mesa_src_reg mat
= get_temp(ir
->type
);
1715 ir_to_mesa_dst_reg mat_column
= ir_to_mesa_dst_reg_from_src(mat
);
1717 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1718 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1719 values
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1721 src_reg
= ir_to_mesa_src_reg(PROGRAM_CONSTANT
, -1, NULL
);
1722 src_reg
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1724 ir
->type
->vector_elements
,
1726 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, mat_column
, src_reg
);
1734 src_reg
.file
= PROGRAM_CONSTANT
;
1735 switch (ir
->type
->base_type
) {
1736 case GLSL_TYPE_FLOAT
:
1737 values
= &ir
->value
.f
[0];
1739 case GLSL_TYPE_UINT
:
1740 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1741 values
[i
] = ir
->value
.u
[i
];
1745 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1746 values
[i
] = ir
->value
.i
[i
];
1749 case GLSL_TYPE_BOOL
:
1750 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1751 values
[i
] = ir
->value
.b
[i
];
1755 assert(!"Non-float/uint/int/bool constant");
1758 this->result
= ir_to_mesa_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
1759 this->result
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1761 ir
->type
->vector_elements
,
1762 &this->result
.swizzle
);
1766 ir_to_mesa_visitor::get_function_signature(ir_function_signature
*sig
)
1768 function_entry
*entry
;
1770 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
1771 entry
= (function_entry
*)iter
.get();
1773 if (entry
->sig
== sig
)
1777 entry
= talloc(mem_ctx
, function_entry
);
1779 entry
->sig_id
= this->next_signature_id
++;
1780 entry
->bgn_inst
= NULL
;
1782 /* Allocate storage for all the parameters. */
1783 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
1784 ir_variable
*param
= (ir_variable
*)iter
.get();
1785 variable_storage
*storage
;
1787 storage
= find_variable_storage(param
);
1790 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
1792 this->variables
.push_tail(storage
);
1794 this->next_temp
+= type_size(param
->type
);
1797 if (!sig
->return_type
->is_void()) {
1798 entry
->return_reg
= get_temp(sig
->return_type
);
1800 entry
->return_reg
= ir_to_mesa_undef
;
1803 this->function_signatures
.push_tail(entry
);
1808 ir_to_mesa_visitor::visit(ir_call
*ir
)
1810 ir_to_mesa_instruction
*call_inst
;
1811 ir_function_signature
*sig
= ir
->get_callee();
1812 function_entry
*entry
= get_function_signature(sig
);
1815 /* Process in parameters. */
1816 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
1817 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1818 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1819 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1821 if (param
->mode
== ir_var_in
||
1822 param
->mode
== ir_var_inout
) {
1823 variable_storage
*storage
= find_variable_storage(param
);
1826 param_rval
->accept(this);
1827 ir_to_mesa_src_reg r
= this->result
;
1829 ir_to_mesa_dst_reg l
;
1830 l
.file
= storage
->file
;
1831 l
.index
= storage
->index
;
1833 l
.writemask
= WRITEMASK_XYZW
;
1834 l
.cond_mask
= COND_TR
;
1836 for (i
= 0; i
< type_size(param
->type
); i
++) {
1837 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1845 assert(!sig_iter
.has_next());
1847 /* Emit call instruction */
1848 call_inst
= ir_to_mesa_emit_op1(ir
, OPCODE_CAL
,
1849 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
1850 call_inst
->function
= entry
;
1852 /* Process out parameters. */
1853 sig_iter
= sig
->parameters
.iterator();
1854 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1855 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1856 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1858 if (param
->mode
== ir_var_out
||
1859 param
->mode
== ir_var_inout
) {
1860 variable_storage
*storage
= find_variable_storage(param
);
1863 ir_to_mesa_src_reg r
;
1864 r
.file
= storage
->file
;
1865 r
.index
= storage
->index
;
1867 r
.swizzle
= SWIZZLE_NOOP
;
1870 param_rval
->accept(this);
1871 ir_to_mesa_dst_reg l
= ir_to_mesa_dst_reg_from_src(this->result
);
1873 for (i
= 0; i
< type_size(param
->type
); i
++) {
1874 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1882 assert(!sig_iter
.has_next());
1884 /* Process return value. */
1885 this->result
= entry
->return_reg
;
1890 ir_to_mesa_visitor::visit(ir_texture
*ir
)
1892 ir_to_mesa_src_reg result_src
, coord
, lod_info
, projector
;
1893 ir_to_mesa_dst_reg result_dst
, coord_dst
;
1894 ir_to_mesa_instruction
*inst
= NULL
;
1895 prog_opcode opcode
= OPCODE_NOP
;
1897 ir
->coordinate
->accept(this);
1899 /* Put our coords in a temp. We'll need to modify them for shadow,
1900 * projection, or LOD, so the only case we'd use it as is is if
1901 * we're doing plain old texturing. Mesa IR optimization should
1902 * handle cleaning up our mess in that case.
1904 coord
= get_temp(glsl_type::vec4_type
);
1905 coord_dst
= ir_to_mesa_dst_reg_from_src(coord
);
1906 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
,
1909 if (ir
->projector
) {
1910 ir
->projector
->accept(this);
1911 projector
= this->result
;
1914 /* Storage for our result. Ideally for an assignment we'd be using
1915 * the actual storage for the result here, instead.
1917 result_src
= get_temp(glsl_type::vec4_type
);
1918 result_dst
= ir_to_mesa_dst_reg_from_src(result_src
);
1922 opcode
= OPCODE_TEX
;
1925 opcode
= OPCODE_TXB
;
1926 ir
->lod_info
.bias
->accept(this);
1927 lod_info
= this->result
;
1930 opcode
= OPCODE_TXL
;
1931 ir
->lod_info
.lod
->accept(this);
1932 lod_info
= this->result
;
1936 assert(!"GLSL 1.30 features unsupported");
1940 if (ir
->projector
) {
1941 if (opcode
== OPCODE_TEX
) {
1942 /* Slot the projector in as the last component of the coord. */
1943 coord_dst
.writemask
= WRITEMASK_W
;
1944 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, projector
);
1945 coord_dst
.writemask
= WRITEMASK_XYZW
;
1946 opcode
= OPCODE_TXP
;
1948 ir_to_mesa_src_reg coord_w
= coord
;
1949 coord_w
.swizzle
= SWIZZLE_WWWW
;
1951 /* For the other TEX opcodes there's no projective version
1952 * since the last slot is taken up by lod info. Do the
1953 * projective divide now.
1955 coord_dst
.writemask
= WRITEMASK_W
;
1956 ir_to_mesa_emit_op1(ir
, OPCODE_RCP
, coord_dst
, projector
);
1958 coord_dst
.writemask
= WRITEMASK_XYZ
;
1959 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, coord_dst
, coord
, coord_w
);
1961 coord_dst
.writemask
= WRITEMASK_XYZW
;
1962 coord
.swizzle
= SWIZZLE_XYZW
;
1966 if (ir
->shadow_comparitor
) {
1967 /* Slot the shadow value in as the second to last component of the
1970 ir
->shadow_comparitor
->accept(this);
1971 coord_dst
.writemask
= WRITEMASK_Z
;
1972 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, this->result
);
1973 coord_dst
.writemask
= WRITEMASK_XYZW
;
1976 if (opcode
== OPCODE_TXL
|| opcode
== OPCODE_TXB
) {
1977 /* Mesa IR stores lod or lod bias in the last channel of the coords. */
1978 coord_dst
.writemask
= WRITEMASK_W
;
1979 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, lod_info
);
1980 coord_dst
.writemask
= WRITEMASK_XYZW
;
1983 inst
= ir_to_mesa_emit_op1(ir
, opcode
, result_dst
, coord
);
1985 if (ir
->shadow_comparitor
)
1986 inst
->tex_shadow
= GL_TRUE
;
1988 ir_dereference_variable
*sampler
= ir
->sampler
->as_dereference_variable();
1989 assert(sampler
); /* FINISHME: sampler arrays */
1990 /* generate the mapping, remove when we generate storage at
1993 sampler
->accept(this);
1995 inst
->sampler
= get_sampler_number(sampler
->var
->location
);
1997 switch (sampler
->type
->sampler_dimensionality
) {
1998 case GLSL_SAMPLER_DIM_1D
:
1999 inst
->tex_target
= TEXTURE_1D_INDEX
;
2001 case GLSL_SAMPLER_DIM_2D
:
2002 inst
->tex_target
= TEXTURE_2D_INDEX
;
2004 case GLSL_SAMPLER_DIM_3D
:
2005 inst
->tex_target
= TEXTURE_3D_INDEX
;
2007 case GLSL_SAMPLER_DIM_CUBE
:
2008 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2011 assert(!"FINISHME: other texture targets");
2014 this->result
= result_src
;
2018 ir_to_mesa_visitor::visit(ir_return
*ir
)
2020 assert(current_function
);
2022 if (ir
->get_value()) {
2023 ir_to_mesa_dst_reg l
;
2026 ir
->get_value()->accept(this);
2027 ir_to_mesa_src_reg r
= this->result
;
2029 l
= ir_to_mesa_dst_reg_from_src(current_function
->return_reg
);
2031 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2032 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
2038 ir_to_mesa_emit_op0(ir
, OPCODE_RET
);
2042 ir_to_mesa_visitor::visit(ir_discard
*ir
)
2044 assert(ir
->condition
== NULL
); /* FINISHME */
2046 ir_to_mesa_emit_op0(ir
, OPCODE_KIL_NV
);
2050 ir_to_mesa_visitor::visit(ir_if
*ir
)
2052 ir_to_mesa_instruction
*cond_inst
, *if_inst
, *else_inst
= NULL
;
2053 ir_to_mesa_instruction
*prev_inst
;
2055 prev_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2057 ir
->condition
->accept(this);
2058 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2060 if (ctx
->Shader
.EmitCondCodes
) {
2061 cond_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2063 /* See if we actually generated any instruction for generating
2064 * the condition. If not, then cook up a move to a temp so we
2065 * have something to set cond_update on.
2067 if (cond_inst
== prev_inst
) {
2068 ir_to_mesa_src_reg temp
= get_temp(glsl_type::bool_type
);
2069 cond_inst
= ir_to_mesa_emit_op1(ir
->condition
, OPCODE_MOV
,
2070 ir_to_mesa_dst_reg_from_src(temp
),
2073 cond_inst
->cond_update
= GL_TRUE
;
2075 if_inst
= ir_to_mesa_emit_op0(ir
->condition
, OPCODE_IF
);
2076 if_inst
->dst_reg
.cond_mask
= COND_NE
;
2078 if_inst
= ir_to_mesa_emit_op1(ir
->condition
,
2079 OPCODE_IF
, ir_to_mesa_undef_dst
,
2083 this->instructions
.push_tail(if_inst
);
2085 visit_exec_list(&ir
->then_instructions
, this);
2087 if (!ir
->else_instructions
.is_empty()) {
2088 else_inst
= ir_to_mesa_emit_op0(ir
->condition
, OPCODE_ELSE
);
2089 visit_exec_list(&ir
->else_instructions
, this);
2092 if_inst
= ir_to_mesa_emit_op1(ir
->condition
, OPCODE_ENDIF
,
2093 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
2096 ir_to_mesa_visitor::ir_to_mesa_visitor()
2098 result
.file
= PROGRAM_UNDEFINED
;
2100 next_signature_id
= 1;
2102 sampler_map_size
= 0;
2103 current_function
= NULL
;
2106 static struct prog_src_register
2107 mesa_src_reg_from_ir_src_reg(ir_to_mesa_src_reg reg
)
2109 struct prog_src_register mesa_reg
;
2111 mesa_reg
.File
= reg
.file
;
2112 assert(reg
.index
< (1 << INST_INDEX_BITS
) - 1);
2113 mesa_reg
.Index
= reg
.index
;
2114 mesa_reg
.Swizzle
= reg
.swizzle
;
2115 mesa_reg
.RelAddr
= reg
.reladdr
!= NULL
;
2116 mesa_reg
.Negate
= reg
.negate
;
2123 set_branchtargets(ir_to_mesa_visitor
*v
,
2124 struct prog_instruction
*mesa_instructions
,
2125 int num_instructions
)
2127 int if_count
= 0, loop_count
= 0;
2128 int *if_stack
, *loop_stack
;
2129 int if_stack_pos
= 0, loop_stack_pos
= 0;
2132 for (i
= 0; i
< num_instructions
; i
++) {
2133 switch (mesa_instructions
[i
].Opcode
) {
2137 case OPCODE_BGNLOOP
:
2142 mesa_instructions
[i
].BranchTarget
= -1;
2149 if_stack
= (int *)calloc(if_count
, sizeof(*if_stack
));
2150 loop_stack
= (int *)calloc(loop_count
, sizeof(*loop_stack
));
2152 for (i
= 0; i
< num_instructions
; i
++) {
2153 switch (mesa_instructions
[i
].Opcode
) {
2155 if_stack
[if_stack_pos
] = i
;
2159 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2160 if_stack
[if_stack_pos
- 1] = i
;
2163 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2166 case OPCODE_BGNLOOP
:
2167 loop_stack
[loop_stack_pos
] = i
;
2170 case OPCODE_ENDLOOP
:
2172 /* Rewrite any breaks/conts at this nesting level (haven't
2173 * already had a BranchTarget assigned) to point to the end
2176 for (j
= loop_stack
[loop_stack_pos
]; j
< i
; j
++) {
2177 if (mesa_instructions
[j
].Opcode
== OPCODE_BRK
||
2178 mesa_instructions
[j
].Opcode
== OPCODE_CONT
) {
2179 if (mesa_instructions
[j
].BranchTarget
== -1) {
2180 mesa_instructions
[j
].BranchTarget
= i
;
2184 /* The loop ends point at each other. */
2185 mesa_instructions
[i
].BranchTarget
= loop_stack
[loop_stack_pos
];
2186 mesa_instructions
[loop_stack
[loop_stack_pos
]].BranchTarget
= i
;
2189 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
2190 function_entry
*entry
= (function_entry
*)iter
.get();
2192 if (entry
->sig_id
== mesa_instructions
[i
].BranchTarget
) {
2193 mesa_instructions
[i
].BranchTarget
= entry
->inst
;
2207 print_program(struct prog_instruction
*mesa_instructions
,
2208 ir_instruction
**mesa_instruction_annotation
,
2209 int num_instructions
)
2211 ir_instruction
*last_ir
= NULL
;
2215 for (i
= 0; i
< num_instructions
; i
++) {
2216 struct prog_instruction
*mesa_inst
= mesa_instructions
+ i
;
2217 ir_instruction
*ir
= mesa_instruction_annotation
[i
];
2219 fprintf(stdout
, "%3d: ", i
);
2221 if (last_ir
!= ir
&& ir
) {
2224 for (j
= 0; j
< indent
; j
++) {
2225 fprintf(stdout
, " ");
2231 fprintf(stdout
, " "); /* line number spacing. */
2234 indent
= _mesa_fprint_instruction_opt(stdout
, mesa_inst
, indent
,
2235 PROG_PRINT_DEBUG
, NULL
);
2240 mark_input(struct gl_program
*prog
,
2244 prog
->InputsRead
|= BITFIELD64_BIT(index
);
2248 if (index
>= FRAG_ATTRIB_TEX0
&& index
<= FRAG_ATTRIB_TEX7
) {
2249 for (i
= 0; i
< 8; i
++) {
2250 prog
->InputsRead
|= BITFIELD64_BIT(FRAG_ATTRIB_TEX0
+ i
);
2253 assert(!"FINISHME: Mark InputsRead for varying arrays");
2259 mark_output(struct gl_program
*prog
,
2263 prog
->OutputsWritten
|= BITFIELD64_BIT(index
);
2267 if (index
>= VERT_RESULT_TEX0
&& index
<= VERT_RESULT_TEX7
) {
2268 for (i
= 0; i
< 8; i
++) {
2269 prog
->OutputsWritten
|= BITFIELD64_BIT(FRAG_ATTRIB_TEX0
+ i
);
2272 assert(!"FINISHME: Mark OutputsWritten for varying arrays");
2278 count_resources(struct gl_program
*prog
)
2282 prog
->InputsRead
= 0;
2283 prog
->OutputsWritten
= 0;
2284 prog
->SamplersUsed
= 0;
2286 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
2287 struct prog_instruction
*inst
= &prog
->Instructions
[i
];
2290 switch (inst
->DstReg
.File
) {
2291 case PROGRAM_OUTPUT
:
2292 mark_output(prog
, inst
->DstReg
.Index
, inst
->DstReg
.RelAddr
);
2295 mark_input(prog
, inst
->DstReg
.Index
, inst
->DstReg
.RelAddr
);
2301 for (reg
= 0; reg
< _mesa_num_inst_src_regs(inst
->Opcode
); reg
++) {
2302 switch (inst
->SrcReg
[reg
].File
) {
2303 case PROGRAM_OUTPUT
:
2304 mark_output(prog
, inst
->SrcReg
[reg
].Index
,
2305 inst
->SrcReg
[reg
].RelAddr
);
2308 mark_input(prog
, inst
->SrcReg
[reg
].Index
, inst
->SrcReg
[reg
].RelAddr
);
2315 /* Instead of just using the uniform's value to map to a
2316 * sampler, Mesa first allocates a separate number for the
2317 * sampler (_mesa_add_sampler), then we reindex it down to a
2318 * small integer (sampler_map[], SamplersUsed), then that gets
2319 * mapped to the uniform's value, and we get an actual sampler.
2321 if (_mesa_is_tex_instruction(inst
->Opcode
)) {
2322 prog
->SamplerTargets
[inst
->TexSrcUnit
] =
2323 (gl_texture_index
)inst
->TexSrcTarget
;
2324 prog
->SamplersUsed
|= 1 << inst
->TexSrcUnit
;
2325 if (inst
->TexShadow
) {
2326 prog
->ShadowSamplers
|= 1 << inst
->TexSrcUnit
;
2331 _mesa_update_shader_textures_used(prog
);
2334 /* Each stage has some uniforms in its Parameters list. The Uniforms
2335 * list for the linked shader program has a pointer to these uniforms
2336 * in each of the stage's Parameters list, so that their values can be
2337 * updated when a uniform is set.
2340 link_uniforms_to_shared_uniform_list(struct gl_uniform_list
*uniforms
,
2341 struct gl_program
*prog
)
2345 for (i
= 0; i
< prog
->Parameters
->NumParameters
; i
++) {
2346 const struct gl_program_parameter
*p
= prog
->Parameters
->Parameters
+ i
;
2348 if (p
->Type
== PROGRAM_UNIFORM
|| p
->Type
== PROGRAM_SAMPLER
) {
2349 struct gl_uniform
*uniform
=
2350 _mesa_append_uniform(uniforms
, p
->Name
, prog
->Target
, i
);
2352 uniform
->Initialized
= p
->Initialized
;
2358 get_mesa_program(GLcontext
*ctx
, struct gl_shader_program
*shader_program
,
2359 struct gl_shader
*shader
)
2361 void *mem_ctx
= shader_program
;
2362 ir_to_mesa_visitor v
;
2363 struct prog_instruction
*mesa_instructions
, *mesa_inst
;
2364 ir_instruction
**mesa_instruction_annotation
;
2366 struct gl_program
*prog
;
2368 const char *target_string
;
2371 switch (shader
->Type
) {
2372 case GL_VERTEX_SHADER
:
2373 target
= GL_VERTEX_PROGRAM_ARB
;
2374 target_string
= "vertex";
2376 case GL_FRAGMENT_SHADER
:
2377 target
= GL_FRAGMENT_PROGRAM_ARB
;
2378 target_string
= "fragment";
2381 assert(!"should not be reached");
2385 validate_ir_tree(shader
->ir
);
2387 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
2390 prog
->Parameters
= _mesa_new_parameter_list();
2391 prog
->Varying
= _mesa_new_parameter_list();
2392 prog
->Attributes
= _mesa_new_parameter_list();
2396 v
.mem_ctx
= talloc_new(NULL
);
2398 /* Emit Mesa IR for main(). */
2399 visit_exec_list(shader
->ir
, &v
);
2400 v
.ir_to_mesa_emit_op0(NULL
, OPCODE_END
);
2402 /* Now emit bodies for any functions that were used. */
2404 progress
= GL_FALSE
;
2406 foreach_iter(exec_list_iterator
, iter
, v
.function_signatures
) {
2407 function_entry
*entry
= (function_entry
*)iter
.get();
2409 if (!entry
->bgn_inst
) {
2410 v
.current_function
= entry
;
2412 entry
->bgn_inst
= v
.ir_to_mesa_emit_op0(NULL
, OPCODE_BGNSUB
);
2413 entry
->bgn_inst
->function
= entry
;
2415 visit_exec_list(&entry
->sig
->body
, &v
);
2417 ir_to_mesa_instruction
*last
;
2418 last
= (ir_to_mesa_instruction
*)v
.instructions
.get_tail();
2419 if (last
->op
!= OPCODE_RET
)
2420 v
.ir_to_mesa_emit_op0(NULL
, OPCODE_RET
);
2422 ir_to_mesa_instruction
*end
;
2423 end
= v
.ir_to_mesa_emit_op0(NULL
, OPCODE_ENDSUB
);
2424 end
->function
= entry
;
2431 prog
->NumTemporaries
= v
.next_temp
;
2433 int num_instructions
= 0;
2434 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2439 (struct prog_instruction
*)calloc(num_instructions
,
2440 sizeof(*mesa_instructions
));
2441 mesa_instruction_annotation
= talloc_array(mem_ctx
, ir_instruction
*,
2444 mesa_inst
= mesa_instructions
;
2446 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2447 ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
2449 mesa_inst
->Opcode
= inst
->op
;
2450 mesa_inst
->CondUpdate
= inst
->cond_update
;
2451 mesa_inst
->DstReg
.File
= inst
->dst_reg
.file
;
2452 mesa_inst
->DstReg
.Index
= inst
->dst_reg
.index
;
2453 mesa_inst
->DstReg
.CondMask
= inst
->dst_reg
.cond_mask
;
2454 mesa_inst
->DstReg
.WriteMask
= inst
->dst_reg
.writemask
;
2455 mesa_inst
->DstReg
.RelAddr
= inst
->dst_reg
.reladdr
!= NULL
;
2456 mesa_inst
->SrcReg
[0] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[0]);
2457 mesa_inst
->SrcReg
[1] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[1]);
2458 mesa_inst
->SrcReg
[2] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[2]);
2459 mesa_inst
->TexSrcUnit
= inst
->sampler
;
2460 mesa_inst
->TexSrcTarget
= inst
->tex_target
;
2461 mesa_inst
->TexShadow
= inst
->tex_shadow
;
2462 mesa_instruction_annotation
[i
] = inst
->ir
;
2464 if (ctx
->Shader
.EmitNoIfs
&& mesa_inst
->Opcode
== OPCODE_IF
) {
2465 shader_program
->InfoLog
=
2466 talloc_asprintf_append(shader_program
->InfoLog
,
2467 "Couldn't flatten if statement\n");
2468 shader_program
->LinkStatus
= false;
2471 switch (mesa_inst
->Opcode
) {
2473 inst
->function
->inst
= i
;
2474 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2477 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2480 mesa_inst
->BranchTarget
= inst
->function
->sig_id
; /* rewritten later */
2483 prog
->NumAddressRegs
= 1;
2493 set_branchtargets(&v
, mesa_instructions
, num_instructions
);
2494 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2495 printf("Mesa %s program:\n", target_string
);
2496 print_program(mesa_instructions
, mesa_instruction_annotation
,
2500 prog
->Instructions
= mesa_instructions
;
2501 prog
->NumInstructions
= num_instructions
;
2503 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
2505 if ((ctx
->Shader
.Flags
& GLSL_NO_OPT
) == 0) {
2506 _mesa_optimize_program(ctx
, prog
);
2515 _mesa_glsl_compile_shader(GLcontext
*ctx
, struct gl_shader
*shader
)
2517 struct _mesa_glsl_parse_state
*state
=
2518 new(shader
) _mesa_glsl_parse_state(ctx
, shader
->Type
, shader
);
2520 const char *source
= shader
->Source
;
2521 state
->error
= preprocess(state
, &source
, &state
->info_log
,
2524 if (!state
->error
) {
2525 _mesa_glsl_lexer_ctor(state
, source
);
2526 _mesa_glsl_parse(state
);
2527 _mesa_glsl_lexer_dtor(state
);
2530 shader
->ir
= new(shader
) exec_list
;
2531 if (!state
->error
&& !state
->translation_unit
.is_empty())
2532 _mesa_ast_to_hir(shader
->ir
, state
);
2534 if (!state
->error
&& !shader
->ir
->is_empty()) {
2535 validate_ir_tree(shader
->ir
);
2538 do_mat_op_to_vec(shader
->ir
);
2539 do_mod_to_fract(shader
->ir
);
2540 do_div_to_mul_rcp(shader
->ir
);
2542 /* Optimization passes */
2547 progress
= do_function_inlining(shader
->ir
) || progress
;
2548 progress
= do_if_simplification(shader
->ir
) || progress
;
2549 progress
= do_copy_propagation(shader
->ir
) || progress
;
2550 progress
= do_dead_code_local(shader
->ir
) || progress
;
2551 progress
= do_dead_code_unlinked(shader
->ir
) || progress
;
2552 progress
= do_tree_grafting(shader
->ir
) || progress
;
2553 progress
= do_constant_variable_unlinked(shader
->ir
) || progress
;
2554 progress
= do_constant_folding(shader
->ir
) || progress
;
2555 progress
= do_algebraic(shader
->ir
) || progress
;
2556 progress
= do_if_return(shader
->ir
) || progress
;
2557 if (ctx
->Shader
.EmitNoIfs
)
2558 progress
= do_if_to_cond_assign(shader
->ir
) || progress
;
2560 progress
= do_vec_index_to_swizzle(shader
->ir
) || progress
;
2561 /* Do this one after the previous to let the easier pass handle
2562 * constant vector indexing.
2564 progress
= do_vec_index_to_cond_assign(shader
->ir
) || progress
;
2566 progress
= do_swizzle_swizzle(shader
->ir
) || progress
;
2569 validate_ir_tree(shader
->ir
);
2572 shader
->symbols
= state
->symbols
;
2574 shader
->CompileStatus
= !state
->error
;
2575 shader
->InfoLog
= state
->info_log
;
2576 shader
->Version
= state
->language_version
;
2577 memcpy(shader
->builtins_to_link
, state
->builtins_to_link
,
2578 sizeof(shader
->builtins_to_link
[0]) * state
->num_builtins_to_link
);
2579 shader
->num_builtins_to_link
= state
->num_builtins_to_link
;
2581 if (ctx
->Shader
.Flags
& GLSL_LOG
) {
2582 _mesa_write_shader_to_file(shader
);
2585 /* Retain any live IR, but trash the rest. */
2586 reparent_ir(shader
->ir
, shader
);
2592 _mesa_glsl_link_shader(GLcontext
*ctx
, struct gl_shader_program
*prog
)
2596 _mesa_clear_shader_program_data(ctx
, prog
);
2598 prog
->LinkStatus
= GL_TRUE
;
2600 for (i
= 0; i
< prog
->NumShaders
; i
++) {
2601 if (!prog
->Shaders
[i
]->CompileStatus
) {
2603 talloc_asprintf_append(prog
->InfoLog
,
2604 "linking with uncompiled shader");
2605 prog
->LinkStatus
= GL_FALSE
;
2609 prog
->Varying
= _mesa_new_parameter_list();
2610 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
2611 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
2613 if (prog
->LinkStatus
) {
2616 /* We don't use the linker's uniforms list, and cook up our own at
2619 free(prog
->Uniforms
);
2620 prog
->Uniforms
= _mesa_new_uniform_list();
2623 if (prog
->LinkStatus
) {
2624 for (i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
2625 struct gl_program
*linked_prog
;
2628 linked_prog
= get_mesa_program(ctx
, prog
,
2629 prog
->_LinkedShaders
[i
]);
2630 count_resources(linked_prog
);
2632 link_uniforms_to_shared_uniform_list(prog
->Uniforms
, linked_prog
);
2634 switch (prog
->_LinkedShaders
[i
]->Type
) {
2635 case GL_VERTEX_SHADER
:
2636 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
2637 (struct gl_vertex_program
*)linked_prog
);
2638 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
2641 case GL_FRAGMENT_SHADER
:
2642 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
2643 (struct gl_fragment_program
*)linked_prog
);
2644 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
2649 prog
->LinkStatus
= GL_FALSE
;