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 * Translate GLSL IR to Mesa's gl_program representation.
33 #include "main/compiler.h"
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/shaderapi.h"
47 #include "main/shaderobj.h"
48 #include "main/uniforms.h"
49 #include "program/hash_table.h"
50 #include "program/prog_instruction.h"
51 #include "program/prog_optimize.h"
52 #include "program/prog_print.h"
53 #include "program/program.h"
54 #include "program/prog_uniform.h"
55 #include "program/prog_parameter.h"
56 #include "program/sampler.h"
59 static int swizzle_for_size(int size
);
62 * This struct is a corresponding struct to Mesa prog_src_register, with
65 typedef struct ir_to_mesa_src_reg
{
66 ir_to_mesa_src_reg(int file
, int index
, const glsl_type
*type
)
70 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
71 this->swizzle
= swizzle_for_size(type
->vector_elements
);
73 this->swizzle
= SWIZZLE_XYZW
;
80 this->file
= PROGRAM_UNDEFINED
;
87 int file
; /**< PROGRAM_* from Mesa */
88 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
89 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
90 int negate
; /**< NEGATE_XYZW mask from mesa */
91 /** Register index should be offset by the integer in this reg. */
92 ir_to_mesa_src_reg
*reladdr
;
95 typedef struct ir_to_mesa_dst_reg
{
96 int file
; /**< PROGRAM_* from Mesa */
97 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
98 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
100 /** Register index should be offset by the integer in this reg. */
101 ir_to_mesa_src_reg
*reladdr
;
102 } ir_to_mesa_dst_reg
;
104 extern ir_to_mesa_src_reg ir_to_mesa_undef
;
106 class ir_to_mesa_instruction
: public exec_node
{
108 /* Callers of this talloc-based new need not call delete. It's
109 * easier to just talloc_free 'ctx' (or any of its ancestors). */
110 static void* operator new(size_t size
, void *ctx
)
114 node
= talloc_zero_size(ctx
, size
);
115 assert(node
!= NULL
);
121 ir_to_mesa_dst_reg dst_reg
;
122 ir_to_mesa_src_reg src_reg
[3];
123 /** Pointer to the ir source this tree came from for debugging */
125 GLboolean cond_update
;
126 int sampler
; /**< sampler index */
127 int tex_target
; /**< One of TEXTURE_*_INDEX */
128 GLboolean tex_shadow
;
130 class function_entry
*function
; /* Set on OPCODE_CAL or OPCODE_BGNSUB */
133 class variable_storage
: public exec_node
{
135 variable_storage(ir_variable
*var
, int file
, int index
)
136 : file(file
), index(index
), var(var
)
143 ir_variable
*var
; /* variable that maps to this, if any */
146 class function_entry
: public exec_node
{
148 ir_function_signature
*sig
;
151 * identifier of this function signature used by the program.
153 * At the point that Mesa instructions for function calls are
154 * generated, we don't know the address of the first instruction of
155 * the function body. So we make the BranchTarget that is called a
156 * small integer and rewrite them during set_branchtargets().
161 * Pointer to first instruction of the function body.
163 * Set during function body emits after main() is processed.
165 ir_to_mesa_instruction
*bgn_inst
;
168 * Index of the first instruction of the function body in actual
171 * Set after convertion from ir_to_mesa_instruction to prog_instruction.
175 /** Storage for the return value. */
176 ir_to_mesa_src_reg return_reg
;
179 class ir_to_mesa_visitor
: public ir_visitor
{
181 ir_to_mesa_visitor();
182 ~ir_to_mesa_visitor();
184 function_entry
*current_function
;
186 struct gl_context
*ctx
;
187 struct gl_program
*prog
;
188 struct gl_shader_program
*shader_program
;
189 struct gl_shader_compiler_options
*options
;
193 variable_storage
*find_variable_storage(ir_variable
*var
);
195 function_entry
*get_function_signature(ir_function_signature
*sig
);
197 ir_to_mesa_src_reg
get_temp(const glsl_type
*type
);
198 void reladdr_to_temp(ir_instruction
*ir
,
199 ir_to_mesa_src_reg
*reg
, int *num_reladdr
);
201 struct ir_to_mesa_src_reg
src_reg_for_float(float val
);
204 * \name Visit methods
206 * As typical for the visitor pattern, there must be one \c visit method for
207 * each concrete subclass of \c ir_instruction. Virtual base classes within
208 * the hierarchy should not have \c visit methods.
211 virtual void visit(ir_variable
*);
212 virtual void visit(ir_loop
*);
213 virtual void visit(ir_loop_jump
*);
214 virtual void visit(ir_function_signature
*);
215 virtual void visit(ir_function
*);
216 virtual void visit(ir_expression
*);
217 virtual void visit(ir_swizzle
*);
218 virtual void visit(ir_dereference_variable
*);
219 virtual void visit(ir_dereference_array
*);
220 virtual void visit(ir_dereference_record
*);
221 virtual void visit(ir_assignment
*);
222 virtual void visit(ir_constant
*);
223 virtual void visit(ir_call
*);
224 virtual void visit(ir_return
*);
225 virtual void visit(ir_discard
*);
226 virtual void visit(ir_texture
*);
227 virtual void visit(ir_if
*);
230 struct ir_to_mesa_src_reg result
;
232 /** List of variable_storage */
235 /** List of function_entry */
236 exec_list function_signatures
;
237 int next_signature_id
;
239 /** List of ir_to_mesa_instruction */
240 exec_list instructions
;
242 ir_to_mesa_instruction
*ir_to_mesa_emit_op0(ir_instruction
*ir
,
243 enum prog_opcode op
);
245 ir_to_mesa_instruction
*ir_to_mesa_emit_op1(ir_instruction
*ir
,
247 ir_to_mesa_dst_reg dst
,
248 ir_to_mesa_src_reg src0
);
250 ir_to_mesa_instruction
*ir_to_mesa_emit_op2(ir_instruction
*ir
,
252 ir_to_mesa_dst_reg dst
,
253 ir_to_mesa_src_reg src0
,
254 ir_to_mesa_src_reg src1
);
256 ir_to_mesa_instruction
*ir_to_mesa_emit_op3(ir_instruction
*ir
,
258 ir_to_mesa_dst_reg dst
,
259 ir_to_mesa_src_reg src0
,
260 ir_to_mesa_src_reg src1
,
261 ir_to_mesa_src_reg src2
);
264 * Emit the correct dot-product instruction for the type of arguments
266 * \sa ir_to_mesa_emit_op2
268 void ir_to_mesa_emit_dp(ir_instruction
*ir
,
269 ir_to_mesa_dst_reg dst
,
270 ir_to_mesa_src_reg src0
,
271 ir_to_mesa_src_reg src1
,
274 void ir_to_mesa_emit_scalar_op1(ir_instruction
*ir
,
276 ir_to_mesa_dst_reg dst
,
277 ir_to_mesa_src_reg src0
);
279 void ir_to_mesa_emit_scalar_op2(ir_instruction
*ir
,
281 ir_to_mesa_dst_reg dst
,
282 ir_to_mesa_src_reg src0
,
283 ir_to_mesa_src_reg src1
);
285 GLboolean
try_emit_mad(ir_expression
*ir
,
291 ir_to_mesa_src_reg ir_to_mesa_undef
= ir_to_mesa_src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
293 ir_to_mesa_dst_reg ir_to_mesa_undef_dst
= {
294 PROGRAM_UNDEFINED
, 0, SWIZZLE_NOOP
, COND_TR
, NULL
,
297 ir_to_mesa_dst_reg ir_to_mesa_address_reg
= {
298 PROGRAM_ADDRESS
, 0, WRITEMASK_X
, COND_TR
, NULL
302 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
305 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
309 prog
->InfoLog
= talloc_vasprintf_append(prog
->InfoLog
, fmt
, args
);
312 prog
->LinkStatus
= GL_FALSE
;
316 swizzle_for_size(int size
)
318 int size_swizzles
[4] = {
319 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
320 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
321 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
322 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
325 assert((size
>= 1) && (size
<= 4));
326 return size_swizzles
[size
- 1];
329 ir_to_mesa_instruction
*
330 ir_to_mesa_visitor::ir_to_mesa_emit_op3(ir_instruction
*ir
,
332 ir_to_mesa_dst_reg dst
,
333 ir_to_mesa_src_reg src0
,
334 ir_to_mesa_src_reg src1
,
335 ir_to_mesa_src_reg src2
)
337 ir_to_mesa_instruction
*inst
= new(mem_ctx
) ir_to_mesa_instruction();
340 /* If we have to do relative addressing, we want to load the ARL
341 * reg directly for one of the regs, and preload the other reladdr
342 * sources into temps.
344 num_reladdr
+= dst
.reladdr
!= NULL
;
345 num_reladdr
+= src0
.reladdr
!= NULL
;
346 num_reladdr
+= src1
.reladdr
!= NULL
;
347 num_reladdr
+= src2
.reladdr
!= NULL
;
349 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
350 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
351 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
354 ir_to_mesa_emit_op1(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
,
359 assert(num_reladdr
== 0);
363 inst
->src_reg
[0] = src0
;
364 inst
->src_reg
[1] = src1
;
365 inst
->src_reg
[2] = src2
;
368 inst
->function
= NULL
;
370 this->instructions
.push_tail(inst
);
376 ir_to_mesa_instruction
*
377 ir_to_mesa_visitor::ir_to_mesa_emit_op2(ir_instruction
*ir
,
379 ir_to_mesa_dst_reg dst
,
380 ir_to_mesa_src_reg src0
,
381 ir_to_mesa_src_reg src1
)
383 return ir_to_mesa_emit_op3(ir
, op
, dst
, src0
, src1
, ir_to_mesa_undef
);
386 ir_to_mesa_instruction
*
387 ir_to_mesa_visitor::ir_to_mesa_emit_op1(ir_instruction
*ir
,
389 ir_to_mesa_dst_reg dst
,
390 ir_to_mesa_src_reg src0
)
392 assert(dst
.writemask
!= 0);
393 return ir_to_mesa_emit_op3(ir
, op
, dst
,
394 src0
, ir_to_mesa_undef
, ir_to_mesa_undef
);
397 ir_to_mesa_instruction
*
398 ir_to_mesa_visitor::ir_to_mesa_emit_op0(ir_instruction
*ir
,
401 return ir_to_mesa_emit_op3(ir
, op
, ir_to_mesa_undef_dst
,
408 ir_to_mesa_visitor::ir_to_mesa_emit_dp(ir_instruction
*ir
,
409 ir_to_mesa_dst_reg dst
,
410 ir_to_mesa_src_reg src0
,
411 ir_to_mesa_src_reg src1
,
414 static const gl_inst_opcode dot_opcodes
[] = {
415 OPCODE_DP2
, OPCODE_DP3
, OPCODE_DP4
418 ir_to_mesa_emit_op3(ir
, dot_opcodes
[elements
- 2],
419 dst
, src0
, src1
, ir_to_mesa_undef
);
422 inline ir_to_mesa_dst_reg
423 ir_to_mesa_dst_reg_from_src(ir_to_mesa_src_reg reg
)
425 ir_to_mesa_dst_reg dst_reg
;
427 dst_reg
.file
= reg
.file
;
428 dst_reg
.index
= reg
.index
;
429 dst_reg
.writemask
= WRITEMASK_XYZW
;
430 dst_reg
.cond_mask
= COND_TR
;
431 dst_reg
.reladdr
= reg
.reladdr
;
436 inline ir_to_mesa_src_reg
437 ir_to_mesa_src_reg_from_dst(ir_to_mesa_dst_reg reg
)
439 return ir_to_mesa_src_reg(reg
.file
, reg
.index
, NULL
);
443 * Emits Mesa scalar opcodes to produce unique answers across channels.
445 * Some Mesa opcodes are scalar-only, like ARB_fp/vp. The src X
446 * channel determines the result across all channels. So to do a vec4
447 * of this operation, we want to emit a scalar per source channel used
448 * to produce dest channels.
451 ir_to_mesa_visitor::ir_to_mesa_emit_scalar_op2(ir_instruction
*ir
,
453 ir_to_mesa_dst_reg dst
,
454 ir_to_mesa_src_reg orig_src0
,
455 ir_to_mesa_src_reg orig_src1
)
458 int done_mask
= ~dst
.writemask
;
460 /* Mesa RCP is a scalar operation splatting results to all channels,
461 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
464 for (i
= 0; i
< 4; i
++) {
465 GLuint this_mask
= (1 << i
);
466 ir_to_mesa_instruction
*inst
;
467 ir_to_mesa_src_reg src0
= orig_src0
;
468 ir_to_mesa_src_reg src1
= orig_src1
;
470 if (done_mask
& this_mask
)
473 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
474 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
475 for (j
= i
+ 1; j
< 4; j
++) {
476 if (!(done_mask
& (1 << j
)) &&
477 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
478 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
479 this_mask
|= (1 << j
);
482 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
483 src0_swiz
, src0_swiz
);
484 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
485 src1_swiz
, src1_swiz
);
487 inst
= ir_to_mesa_emit_op2(ir
, op
,
491 inst
->dst_reg
.writemask
= this_mask
;
492 done_mask
|= this_mask
;
497 ir_to_mesa_visitor::ir_to_mesa_emit_scalar_op1(ir_instruction
*ir
,
499 ir_to_mesa_dst_reg dst
,
500 ir_to_mesa_src_reg src0
)
502 ir_to_mesa_src_reg undef
= ir_to_mesa_undef
;
504 undef
.swizzle
= SWIZZLE_XXXX
;
506 ir_to_mesa_emit_scalar_op2(ir
, op
, dst
, src0
, undef
);
509 struct ir_to_mesa_src_reg
510 ir_to_mesa_visitor::src_reg_for_float(float val
)
512 ir_to_mesa_src_reg
src_reg(PROGRAM_CONSTANT
, -1, NULL
);
514 src_reg
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
515 &val
, 1, &src_reg
.swizzle
);
521 type_size(const struct glsl_type
*type
)
526 switch (type
->base_type
) {
529 case GLSL_TYPE_FLOAT
:
531 if (type
->is_matrix()) {
532 return type
->matrix_columns
;
534 /* Regardless of size of vector, it gets a vec4. This is bad
535 * packing for things like floats, but otherwise arrays become a
536 * mess. Hopefully a later pass over the code can pack scalars
537 * down if appropriate.
541 case GLSL_TYPE_ARRAY
:
542 return type_size(type
->fields
.array
) * type
->length
;
543 case GLSL_TYPE_STRUCT
:
545 for (i
= 0; i
< type
->length
; i
++) {
546 size
+= type_size(type
->fields
.structure
[i
].type
);
549 case GLSL_TYPE_SAMPLER
:
550 /* Samplers take up one slot in UNIFORMS[], but they're baked in
561 * In the initial pass of codegen, we assign temporary numbers to
562 * intermediate results. (not SSA -- variable assignments will reuse
563 * storage). Actual register allocation for the Mesa VM occurs in a
564 * pass over the Mesa IR later.
567 ir_to_mesa_visitor::get_temp(const glsl_type
*type
)
569 ir_to_mesa_src_reg src_reg
;
573 src_reg
.file
= PROGRAM_TEMPORARY
;
574 src_reg
.index
= next_temp
;
575 src_reg
.reladdr
= NULL
;
576 next_temp
+= type_size(type
);
578 if (type
->is_array() || type
->is_record()) {
579 src_reg
.swizzle
= SWIZZLE_NOOP
;
581 for (i
= 0; i
< type
->vector_elements
; i
++)
584 swizzle
[i
] = type
->vector_elements
- 1;
585 src_reg
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1],
586 swizzle
[2], swizzle
[3]);
594 ir_to_mesa_visitor::find_variable_storage(ir_variable
*var
)
597 variable_storage
*entry
;
599 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
600 entry
= (variable_storage
*)iter
.get();
602 if (entry
->var
== var
)
610 ir_to_mesa_visitor::visit(ir_variable
*ir
)
612 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
613 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
615 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
616 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
619 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
621 const struct gl_builtin_uniform_desc
*statevar
;
623 for (i
= 0; _mesa_builtin_uniform_desc
[i
].name
; i
++) {
624 if (strcmp(ir
->name
, _mesa_builtin_uniform_desc
[i
].name
) == 0)
628 if (!_mesa_builtin_uniform_desc
[i
].name
) {
629 fail_link(this->shader_program
,
630 "Failed to find builtin uniform `%s'\n", ir
->name
);
634 statevar
= &_mesa_builtin_uniform_desc
[i
];
637 if (ir
->type
->is_array()) {
638 array_count
= ir
->type
->length
;
643 /* Check if this statevar's setup in the STATE file exactly
644 * matches how we'll want to reference it as a
645 * struct/array/whatever. If not, then we need to move it into
646 * temporary storage and hope that it'll get copy-propagated
649 for (i
= 0; i
< statevar
->num_elements
; i
++) {
650 if (statevar
->elements
[i
].swizzle
!= SWIZZLE_XYZW
) {
655 struct variable_storage
*storage
;
656 ir_to_mesa_dst_reg dst
;
657 if (i
== statevar
->num_elements
) {
658 /* We'll set the index later. */
659 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
660 this->variables
.push_tail(storage
);
662 dst
= ir_to_mesa_undef_dst
;
664 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
666 this->variables
.push_tail(storage
);
667 this->next_temp
+= type_size(ir
->type
);
669 dst
= ir_to_mesa_dst_reg_from_src(ir_to_mesa_src_reg(PROGRAM_TEMPORARY
,
675 for (int a
= 0; a
< array_count
; a
++) {
676 for (unsigned int i
= 0; i
< statevar
->num_elements
; i
++) {
677 struct gl_builtin_uniform_element
*element
= &statevar
->elements
[i
];
678 int tokens
[STATE_LENGTH
];
680 memcpy(tokens
, element
->tokens
, sizeof(element
->tokens
));
681 if (ir
->type
->is_array()) {
685 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
686 (gl_state_index
*)tokens
);
688 if (storage
->file
== PROGRAM_STATE_VAR
) {
689 if (storage
->index
== -1) {
690 storage
->index
= index
;
693 (int)(storage
->index
+ a
* statevar
->num_elements
+ i
));
696 ir_to_mesa_src_reg
src(PROGRAM_STATE_VAR
, index
, NULL
);
697 src
.swizzle
= element
->swizzle
;
698 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, dst
, src
);
699 /* even a float takes up a whole vec4 reg in a struct/array. */
704 if (storage
->file
== PROGRAM_TEMPORARY
&&
705 dst
.index
!= storage
->index
+ type_size(ir
->type
)) {
706 fail_link(this->shader_program
,
707 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
708 ir
->name
, dst
.index
- storage
->index
,
709 type_size(ir
->type
));
715 ir_to_mesa_visitor::visit(ir_loop
*ir
)
717 ir_dereference_variable
*counter
= NULL
;
719 if (ir
->counter
!= NULL
)
720 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
722 if (ir
->from
!= NULL
) {
723 assert(ir
->counter
!= NULL
);
725 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
731 ir_to_mesa_emit_op0(NULL
, OPCODE_BGNLOOP
);
735 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
737 ir_if
*if_stmt
= new(ir
) ir_if(e
);
739 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
741 if_stmt
->then_instructions
.push_tail(brk
);
743 if_stmt
->accept(this);
750 visit_exec_list(&ir
->body_instructions
, this);
754 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
755 counter
, ir
->increment
);
757 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
764 ir_to_mesa_emit_op0(NULL
, OPCODE_ENDLOOP
);
768 ir_to_mesa_visitor::visit(ir_loop_jump
*ir
)
771 case ir_loop_jump::jump_break
:
772 ir_to_mesa_emit_op0(NULL
, OPCODE_BRK
);
774 case ir_loop_jump::jump_continue
:
775 ir_to_mesa_emit_op0(NULL
, OPCODE_CONT
);
782 ir_to_mesa_visitor::visit(ir_function_signature
*ir
)
789 ir_to_mesa_visitor::visit(ir_function
*ir
)
791 /* Ignore function bodies other than main() -- we shouldn't see calls to
792 * them since they should all be inlined before we get to ir_to_mesa.
794 if (strcmp(ir
->name
, "main") == 0) {
795 const ir_function_signature
*sig
;
798 sig
= ir
->matching_signature(&empty
);
802 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
803 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
811 ir_to_mesa_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
813 int nonmul_operand
= 1 - mul_operand
;
814 ir_to_mesa_src_reg a
, b
, c
;
816 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
817 if (!expr
|| expr
->operation
!= ir_binop_mul
)
820 expr
->operands
[0]->accept(this);
822 expr
->operands
[1]->accept(this);
824 ir
->operands
[nonmul_operand
]->accept(this);
827 this->result
= get_temp(ir
->type
);
828 ir_to_mesa_emit_op3(ir
, OPCODE_MAD
,
829 ir_to_mesa_dst_reg_from_src(this->result
), a
, b
, c
);
835 ir_to_mesa_visitor::reladdr_to_temp(ir_instruction
*ir
,
836 ir_to_mesa_src_reg
*reg
, int *num_reladdr
)
841 ir_to_mesa_emit_op1(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
, *reg
->reladdr
);
843 if (*num_reladdr
!= 1) {
844 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
846 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
,
847 ir_to_mesa_dst_reg_from_src(temp
), *reg
);
855 ir_to_mesa_visitor::visit(ir_expression
*ir
)
857 unsigned int operand
;
858 struct ir_to_mesa_src_reg op
[2];
859 struct ir_to_mesa_src_reg result_src
;
860 struct ir_to_mesa_dst_reg result_dst
;
862 /* Quick peephole: Emit OPCODE_MAD(a, b, c) instead of ADD(MUL(a, b), c)
864 if (ir
->operation
== ir_binop_add
) {
865 if (try_emit_mad(ir
, 1))
867 if (try_emit_mad(ir
, 0))
871 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
872 this->result
.file
= PROGRAM_UNDEFINED
;
873 ir
->operands
[operand
]->accept(this);
874 if (this->result
.file
== PROGRAM_UNDEFINED
) {
876 printf("Failed to get tree for expression operand:\n");
877 ir
->operands
[operand
]->accept(&v
);
880 op
[operand
] = this->result
;
882 /* Matrix expression operands should have been broken down to vector
883 * operations already.
885 assert(!ir
->operands
[operand
]->type
->is_matrix());
888 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
889 if (ir
->operands
[1]) {
890 vector_elements
= MAX2(vector_elements
,
891 ir
->operands
[1]->type
->vector_elements
);
894 this->result
.file
= PROGRAM_UNDEFINED
;
896 /* Storage for our result. Ideally for an assignment we'd be using
897 * the actual storage for the result here, instead.
899 result_src
= get_temp(ir
->type
);
900 /* convenience for the emit functions below. */
901 result_dst
= ir_to_mesa_dst_reg_from_src(result_src
);
902 /* Limit writes to the channels that will be used by result_src later.
903 * This does limit this temp's use as a temporary for multi-instruction
906 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
908 switch (ir
->operation
) {
909 case ir_unop_logic_not
:
910 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
,
911 op
[0], src_reg_for_float(0.0));
914 op
[0].negate
= ~op
[0].negate
;
918 ir_to_mesa_emit_op1(ir
, OPCODE_ABS
, result_dst
, op
[0]);
921 ir_to_mesa_emit_op1(ir
, OPCODE_SSG
, result_dst
, op
[0]);
924 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RCP
, result_dst
, op
[0]);
928 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_EX2
, result_dst
, op
[0]);
932 assert(!"not reached: should be handled by ir_explog_to_explog2");
935 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_LG2
, result_dst
, op
[0]);
938 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_SIN
, result_dst
, op
[0]);
941 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_COS
, result_dst
, op
[0]);
945 ir_to_mesa_emit_op1(ir
, OPCODE_DDX
, result_dst
, op
[0]);
948 ir_to_mesa_emit_op1(ir
, OPCODE_DDY
, result_dst
, op
[0]);
951 case ir_unop_noise
: {
952 const enum prog_opcode opcode
=
953 prog_opcode(OPCODE_NOISE1
954 + (ir
->operands
[0]->type
->vector_elements
) - 1);
955 assert((opcode
>= OPCODE_NOISE1
) && (opcode
<= OPCODE_NOISE4
));
957 ir_to_mesa_emit_op1(ir
, opcode
, result_dst
, op
[0]);
962 ir_to_mesa_emit_op2(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
965 ir_to_mesa_emit_op2(ir
, OPCODE_SUB
, result_dst
, op
[0], op
[1]);
969 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
972 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
974 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
978 ir_to_mesa_emit_op2(ir
, OPCODE_SLT
, result_dst
, op
[0], op
[1]);
980 case ir_binop_greater
:
981 ir_to_mesa_emit_op2(ir
, OPCODE_SGT
, result_dst
, op
[0], op
[1]);
983 case ir_binop_lequal
:
984 ir_to_mesa_emit_op2(ir
, OPCODE_SLE
, result_dst
, op
[0], op
[1]);
986 case ir_binop_gequal
:
987 ir_to_mesa_emit_op2(ir
, OPCODE_SGE
, result_dst
, op
[0], op
[1]);
990 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
992 case ir_binop_nequal
:
993 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
995 case ir_binop_all_equal
:
996 /* "==" operator producing a scalar boolean. */
997 if (ir
->operands
[0]->type
->is_vector() ||
998 ir
->operands
[1]->type
->is_vector()) {
999 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
1000 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1001 ir_to_mesa_dst_reg_from_src(temp
), op
[0], op
[1]);
1002 ir_to_mesa_emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1003 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
,
1004 result_dst
, result_src
, src_reg_for_float(0.0));
1006 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1009 case ir_binop_any_nequal
:
1010 /* "!=" operator producing a scalar boolean. */
1011 if (ir
->operands
[0]->type
->is_vector() ||
1012 ir
->operands
[1]->type
->is_vector()) {
1013 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
1014 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1015 ir_to_mesa_dst_reg_from_src(temp
), op
[0], op
[1]);
1016 ir_to_mesa_emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1017 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1018 result_dst
, result_src
, src_reg_for_float(0.0));
1020 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1025 assert(ir
->operands
[0]->type
->is_vector());
1026 ir_to_mesa_emit_dp(ir
, result_dst
, op
[0], op
[0],
1027 ir
->operands
[0]->type
->vector_elements
);
1028 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1029 result_dst
, result_src
, src_reg_for_float(0.0));
1032 case ir_binop_logic_xor
:
1033 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1036 case ir_binop_logic_or
:
1037 /* This could be a saturated add and skip the SNE. */
1038 ir_to_mesa_emit_op2(ir
, OPCODE_ADD
,
1042 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1044 result_src
, src_reg_for_float(0.0));
1047 case ir_binop_logic_and
:
1048 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1049 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
1055 assert(ir
->operands
[0]->type
->is_vector());
1056 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1057 ir_to_mesa_emit_dp(ir
, result_dst
, op
[0], op
[1],
1058 ir
->operands
[0]->type
->vector_elements
);
1062 /* sqrt(x) = x * rsq(x). */
1063 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1064 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1065 /* For incoming channels <= 0, set the result to 0. */
1066 op
[0].negate
= ~op
[0].negate
;
1067 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, result_dst
,
1068 op
[0], result_src
, src_reg_for_float(0.0));
1071 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1076 /* Mesa IR lacks types, ints are stored as truncated floats. */
1080 ir_to_mesa_emit_op1(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1084 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
,
1085 op
[0], src_reg_for_float(0.0));
1088 ir_to_mesa_emit_op1(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1091 op
[0].negate
= ~op
[0].negate
;
1092 ir_to_mesa_emit_op1(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1093 result_src
.negate
= ~result_src
.negate
;
1096 ir_to_mesa_emit_op1(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1099 ir_to_mesa_emit_op1(ir
, OPCODE_FRC
, result_dst
, op
[0]);
1103 ir_to_mesa_emit_op2(ir
, OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1106 ir_to_mesa_emit_op2(ir
, OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1109 ir_to_mesa_emit_scalar_op2(ir
, OPCODE_POW
, result_dst
, op
[0], op
[1]);
1112 case ir_unop_bit_not
:
1114 case ir_binop_lshift
:
1115 case ir_binop_rshift
:
1116 case ir_binop_bit_and
:
1117 case ir_binop_bit_xor
:
1118 case ir_binop_bit_or
:
1119 case ir_unop_round_even
:
1120 assert(!"GLSL 1.30 features unsupported");
1124 this->result
= result_src
;
1129 ir_to_mesa_visitor::visit(ir_swizzle
*ir
)
1131 ir_to_mesa_src_reg src_reg
;
1135 /* Note that this is only swizzles in expressions, not those on the left
1136 * hand side of an assignment, which do write masking. See ir_assignment
1140 ir
->val
->accept(this);
1141 src_reg
= this->result
;
1142 assert(src_reg
.file
!= PROGRAM_UNDEFINED
);
1144 for (i
= 0; i
< 4; i
++) {
1145 if (i
< ir
->type
->vector_elements
) {
1148 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.x
);
1151 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.y
);
1154 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.z
);
1157 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.w
);
1161 /* If the type is smaller than a vec4, replicate the last
1164 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1168 src_reg
.swizzle
= MAKE_SWIZZLE4(swizzle
[0],
1173 this->result
= src_reg
;
1177 ir_to_mesa_visitor::visit(ir_dereference_variable
*ir
)
1179 variable_storage
*entry
= find_variable_storage(ir
->var
);
1182 switch (ir
->var
->mode
) {
1183 case ir_var_uniform
:
1184 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_UNIFORM
,
1186 this->variables
.push_tail(entry
);
1191 /* The linker assigns locations for varyings and attributes,
1192 * including deprecated builtins (like gl_Color), user-assign
1193 * generic attributes (glBindVertexLocation), and
1194 * user-defined varyings.
1196 * FINISHME: We would hit this path for function arguments. Fix!
1198 assert(ir
->var
->location
!= -1);
1199 if (ir
->var
->mode
== ir_var_in
||
1200 ir
->var
->mode
== ir_var_inout
) {
1201 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1205 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1206 ir
->var
->location
>= VERT_ATTRIB_GENERIC0
) {
1207 _mesa_add_attribute(prog
->Attributes
,
1209 _mesa_sizeof_glsl_type(ir
->var
->type
->gl_type
),
1210 ir
->var
->type
->gl_type
,
1211 ir
->var
->location
- VERT_ATTRIB_GENERIC0
);
1214 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1221 case ir_var_temporary
:
1222 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_TEMPORARY
,
1224 this->variables
.push_tail(entry
);
1226 next_temp
+= type_size(ir
->var
->type
);
1231 printf("Failed to make storage for %s\n", ir
->var
->name
);
1236 this->result
= ir_to_mesa_src_reg(entry
->file
, entry
->index
, ir
->var
->type
);
1240 ir_to_mesa_visitor::visit(ir_dereference_array
*ir
)
1243 ir_to_mesa_src_reg src_reg
;
1244 int element_size
= type_size(ir
->type
);
1246 index
= ir
->array_index
->constant_expression_value();
1248 ir
->array
->accept(this);
1249 src_reg
= this->result
;
1252 src_reg
.index
+= index
->value
.i
[0] * element_size
;
1254 ir_to_mesa_src_reg array_base
= this->result
;
1255 /* Variable index array dereference. It eats the "vec4" of the
1256 * base of the array and an index that offsets the Mesa register
1259 ir
->array_index
->accept(this);
1261 ir_to_mesa_src_reg index_reg
;
1263 if (element_size
== 1) {
1264 index_reg
= this->result
;
1266 index_reg
= get_temp(glsl_type::float_type
);
1268 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
1269 ir_to_mesa_dst_reg_from_src(index_reg
),
1270 this->result
, src_reg_for_float(element_size
));
1273 src_reg
.reladdr
= talloc(mem_ctx
, ir_to_mesa_src_reg
);
1274 memcpy(src_reg
.reladdr
, &index_reg
, sizeof(index_reg
));
1277 /* If the type is smaller than a vec4, replicate the last channel out. */
1278 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1279 src_reg
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1281 src_reg
.swizzle
= SWIZZLE_NOOP
;
1283 this->result
= src_reg
;
1287 ir_to_mesa_visitor::visit(ir_dereference_record
*ir
)
1290 const glsl_type
*struct_type
= ir
->record
->type
;
1293 ir
->record
->accept(this);
1295 for (i
= 0; i
< struct_type
->length
; i
++) {
1296 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1298 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1300 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1301 this->result
.index
+= offset
;
1305 * We want to be careful in assignment setup to hit the actual storage
1306 * instead of potentially using a temporary like we might with the
1307 * ir_dereference handler.
1309 static struct ir_to_mesa_dst_reg
1310 get_assignment_lhs(ir_dereference
*ir
, ir_to_mesa_visitor
*v
)
1312 /* The LHS must be a dereference. If the LHS is a variable indexed array
1313 * access of a vector, it must be separated into a series conditional moves
1314 * before reaching this point (see ir_vec_index_to_cond_assign).
1316 assert(ir
->as_dereference());
1317 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1319 assert(!deref_array
->array
->type
->is_vector());
1322 /* Use the rvalue deref handler for the most part. We'll ignore
1323 * swizzles in it and write swizzles using writemask, though.
1326 return ir_to_mesa_dst_reg_from_src(v
->result
);
1330 ir_to_mesa_visitor::visit(ir_assignment
*ir
)
1332 struct ir_to_mesa_dst_reg l
;
1333 struct ir_to_mesa_src_reg r
;
1336 ir
->rhs
->accept(this);
1339 l
= get_assignment_lhs(ir
->lhs
, this);
1341 /* FINISHME: This should really set to the correct maximal writemask for each
1342 * FINISHME: component written (in the loops below). This case can only
1343 * FINISHME: occur for matrices, arrays, and structures.
1345 if (ir
->write_mask
== 0) {
1346 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1347 l
.writemask
= WRITEMASK_XYZW
;
1348 } else if (ir
->lhs
->type
->is_scalar()) {
1349 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1350 * FINISHME: W component of fragment shader output zero, work correctly.
1352 l
.writemask
= WRITEMASK_XYZW
;
1355 int first_enabled_chan
= 0;
1358 assert(ir
->lhs
->type
->is_vector());
1359 l
.writemask
= ir
->write_mask
;
1361 for (int i
= 0; i
< 4; i
++) {
1362 if (l
.writemask
& (1 << i
)) {
1363 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1368 /* Swizzle a small RHS vector into the channels being written.
1370 * glsl ir treats write_mask as dictating how many channels are
1371 * present on the RHS while Mesa IR treats write_mask as just
1372 * showing which channels of the vec4 RHS get written.
1374 for (int i
= 0; i
< 4; i
++) {
1375 if (l
.writemask
& (1 << i
))
1376 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1378 swizzles
[i
] = first_enabled_chan
;
1380 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1381 swizzles
[2], swizzles
[3]);
1384 assert(l
.file
!= PROGRAM_UNDEFINED
);
1385 assert(r
.file
!= PROGRAM_UNDEFINED
);
1387 if (ir
->condition
) {
1388 ir_to_mesa_src_reg condition
;
1390 ir
->condition
->accept(this);
1391 condition
= this->result
;
1393 /* We use the OPCODE_CMP (a < 0 ? b : c) for conditional moves,
1394 * and the condition we produced is 0.0 or 1.0. By flipping the
1395 * sign, we can choose which value OPCODE_CMP produces without
1396 * an extra computing the condition.
1398 condition
.negate
= ~condition
.negate
;
1399 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1400 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, l
,
1401 condition
, r
, ir_to_mesa_src_reg_from_dst(l
));
1406 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1407 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1416 ir_to_mesa_visitor::visit(ir_constant
*ir
)
1418 ir_to_mesa_src_reg src_reg
;
1419 GLfloat stack_vals
[4] = { 0 };
1420 GLfloat
*values
= stack_vals
;
1423 /* Unfortunately, 4 floats is all we can get into
1424 * _mesa_add_unnamed_constant. So, make a temp to store an
1425 * aggregate constant and move each constant value into it. If we
1426 * get lucky, copy propagation will eliminate the extra moves.
1429 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1430 ir_to_mesa_src_reg temp_base
= get_temp(ir
->type
);
1431 ir_to_mesa_dst_reg temp
= ir_to_mesa_dst_reg_from_src(temp_base
);
1433 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1434 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1435 int size
= type_size(field_value
->type
);
1439 field_value
->accept(this);
1440 src_reg
= this->result
;
1442 for (i
= 0; i
< (unsigned int)size
; i
++) {
1443 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, src_reg
);
1449 this->result
= temp_base
;
1453 if (ir
->type
->is_array()) {
1454 ir_to_mesa_src_reg temp_base
= get_temp(ir
->type
);
1455 ir_to_mesa_dst_reg temp
= ir_to_mesa_dst_reg_from_src(temp_base
);
1456 int size
= type_size(ir
->type
->fields
.array
);
1460 for (i
= 0; i
< ir
->type
->length
; i
++) {
1461 ir
->array_elements
[i
]->accept(this);
1462 src_reg
= this->result
;
1463 for (int j
= 0; j
< size
; j
++) {
1464 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, src_reg
);
1470 this->result
= temp_base
;
1474 if (ir
->type
->is_matrix()) {
1475 ir_to_mesa_src_reg mat
= get_temp(ir
->type
);
1476 ir_to_mesa_dst_reg mat_column
= ir_to_mesa_dst_reg_from_src(mat
);
1478 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1479 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1480 values
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1482 src_reg
= ir_to_mesa_src_reg(PROGRAM_CONSTANT
, -1, NULL
);
1483 src_reg
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1485 ir
->type
->vector_elements
,
1487 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, mat_column
, src_reg
);
1496 src_reg
.file
= PROGRAM_CONSTANT
;
1497 switch (ir
->type
->base_type
) {
1498 case GLSL_TYPE_FLOAT
:
1499 values
= &ir
->value
.f
[0];
1501 case GLSL_TYPE_UINT
:
1502 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1503 values
[i
] = ir
->value
.u
[i
];
1507 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1508 values
[i
] = ir
->value
.i
[i
];
1511 case GLSL_TYPE_BOOL
:
1512 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1513 values
[i
] = ir
->value
.b
[i
];
1517 assert(!"Non-float/uint/int/bool constant");
1520 this->result
= ir_to_mesa_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
1521 this->result
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1523 ir
->type
->vector_elements
,
1524 &this->result
.swizzle
);
1528 ir_to_mesa_visitor::get_function_signature(ir_function_signature
*sig
)
1530 function_entry
*entry
;
1532 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
1533 entry
= (function_entry
*)iter
.get();
1535 if (entry
->sig
== sig
)
1539 entry
= talloc(mem_ctx
, function_entry
);
1541 entry
->sig_id
= this->next_signature_id
++;
1542 entry
->bgn_inst
= NULL
;
1544 /* Allocate storage for all the parameters. */
1545 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
1546 ir_variable
*param
= (ir_variable
*)iter
.get();
1547 variable_storage
*storage
;
1549 storage
= find_variable_storage(param
);
1552 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
1554 this->variables
.push_tail(storage
);
1556 this->next_temp
+= type_size(param
->type
);
1559 if (!sig
->return_type
->is_void()) {
1560 entry
->return_reg
= get_temp(sig
->return_type
);
1562 entry
->return_reg
= ir_to_mesa_undef
;
1565 this->function_signatures
.push_tail(entry
);
1570 ir_to_mesa_visitor::visit(ir_call
*ir
)
1572 ir_to_mesa_instruction
*call_inst
;
1573 ir_function_signature
*sig
= ir
->get_callee();
1574 function_entry
*entry
= get_function_signature(sig
);
1577 /* Process in parameters. */
1578 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
1579 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1580 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1581 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1583 if (param
->mode
== ir_var_in
||
1584 param
->mode
== ir_var_inout
) {
1585 variable_storage
*storage
= find_variable_storage(param
);
1588 param_rval
->accept(this);
1589 ir_to_mesa_src_reg r
= this->result
;
1591 ir_to_mesa_dst_reg l
;
1592 l
.file
= storage
->file
;
1593 l
.index
= storage
->index
;
1595 l
.writemask
= WRITEMASK_XYZW
;
1596 l
.cond_mask
= COND_TR
;
1598 for (i
= 0; i
< type_size(param
->type
); i
++) {
1599 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1607 assert(!sig_iter
.has_next());
1609 /* Emit call instruction */
1610 call_inst
= ir_to_mesa_emit_op1(ir
, OPCODE_CAL
,
1611 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
1612 call_inst
->function
= entry
;
1614 /* Process out parameters. */
1615 sig_iter
= sig
->parameters
.iterator();
1616 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1617 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1618 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1620 if (param
->mode
== ir_var_out
||
1621 param
->mode
== ir_var_inout
) {
1622 variable_storage
*storage
= find_variable_storage(param
);
1625 ir_to_mesa_src_reg r
;
1626 r
.file
= storage
->file
;
1627 r
.index
= storage
->index
;
1629 r
.swizzle
= SWIZZLE_NOOP
;
1632 param_rval
->accept(this);
1633 ir_to_mesa_dst_reg l
= ir_to_mesa_dst_reg_from_src(this->result
);
1635 for (i
= 0; i
< type_size(param
->type
); i
++) {
1636 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1644 assert(!sig_iter
.has_next());
1646 /* Process return value. */
1647 this->result
= entry
->return_reg
;
1651 ir_to_mesa_visitor::visit(ir_texture
*ir
)
1653 ir_to_mesa_src_reg result_src
, coord
, lod_info
, projector
;
1654 ir_to_mesa_dst_reg result_dst
, coord_dst
;
1655 ir_to_mesa_instruction
*inst
= NULL
;
1656 prog_opcode opcode
= OPCODE_NOP
;
1658 ir
->coordinate
->accept(this);
1660 /* Put our coords in a temp. We'll need to modify them for shadow,
1661 * projection, or LOD, so the only case we'd use it as is is if
1662 * we're doing plain old texturing. Mesa IR optimization should
1663 * handle cleaning up our mess in that case.
1665 coord
= get_temp(glsl_type::vec4_type
);
1666 coord_dst
= ir_to_mesa_dst_reg_from_src(coord
);
1667 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
,
1670 if (ir
->projector
) {
1671 ir
->projector
->accept(this);
1672 projector
= this->result
;
1675 /* Storage for our result. Ideally for an assignment we'd be using
1676 * the actual storage for the result here, instead.
1678 result_src
= get_temp(glsl_type::vec4_type
);
1679 result_dst
= ir_to_mesa_dst_reg_from_src(result_src
);
1683 opcode
= OPCODE_TEX
;
1686 opcode
= OPCODE_TXB
;
1687 ir
->lod_info
.bias
->accept(this);
1688 lod_info
= this->result
;
1691 opcode
= OPCODE_TXL
;
1692 ir
->lod_info
.lod
->accept(this);
1693 lod_info
= this->result
;
1697 assert(!"GLSL 1.30 features unsupported");
1701 if (ir
->projector
) {
1702 if (opcode
== OPCODE_TEX
) {
1703 /* Slot the projector in as the last component of the coord. */
1704 coord_dst
.writemask
= WRITEMASK_W
;
1705 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, projector
);
1706 coord_dst
.writemask
= WRITEMASK_XYZW
;
1707 opcode
= OPCODE_TXP
;
1709 ir_to_mesa_src_reg coord_w
= coord
;
1710 coord_w
.swizzle
= SWIZZLE_WWWW
;
1712 /* For the other TEX opcodes there's no projective version
1713 * since the last slot is taken up by lod info. Do the
1714 * projective divide now.
1716 coord_dst
.writemask
= WRITEMASK_W
;
1717 ir_to_mesa_emit_op1(ir
, OPCODE_RCP
, coord_dst
, projector
);
1719 coord_dst
.writemask
= WRITEMASK_XYZ
;
1720 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, coord_dst
, coord
, coord_w
);
1722 coord_dst
.writemask
= WRITEMASK_XYZW
;
1723 coord
.swizzle
= SWIZZLE_XYZW
;
1727 if (ir
->shadow_comparitor
) {
1728 /* Slot the shadow value in as the second to last component of the
1731 ir
->shadow_comparitor
->accept(this);
1732 coord_dst
.writemask
= WRITEMASK_Z
;
1733 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, this->result
);
1734 coord_dst
.writemask
= WRITEMASK_XYZW
;
1737 if (opcode
== OPCODE_TXL
|| opcode
== OPCODE_TXB
) {
1738 /* Mesa IR stores lod or lod bias in the last channel of the coords. */
1739 coord_dst
.writemask
= WRITEMASK_W
;
1740 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, lod_info
);
1741 coord_dst
.writemask
= WRITEMASK_XYZW
;
1744 inst
= ir_to_mesa_emit_op1(ir
, opcode
, result_dst
, coord
);
1746 if (ir
->shadow_comparitor
)
1747 inst
->tex_shadow
= GL_TRUE
;
1749 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
1750 this->shader_program
,
1753 const glsl_type
*sampler_type
= ir
->sampler
->type
;
1755 switch (sampler_type
->sampler_dimensionality
) {
1756 case GLSL_SAMPLER_DIM_1D
:
1757 inst
->tex_target
= (sampler_type
->sampler_array
)
1758 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
1760 case GLSL_SAMPLER_DIM_2D
:
1761 inst
->tex_target
= (sampler_type
->sampler_array
)
1762 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
1764 case GLSL_SAMPLER_DIM_3D
:
1765 inst
->tex_target
= TEXTURE_3D_INDEX
;
1767 case GLSL_SAMPLER_DIM_CUBE
:
1768 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
1770 case GLSL_SAMPLER_DIM_RECT
:
1771 inst
->tex_target
= TEXTURE_RECT_INDEX
;
1773 case GLSL_SAMPLER_DIM_BUF
:
1774 assert(!"FINISHME: Implement ARB_texture_buffer_object");
1777 assert(!"Should not get here.");
1780 this->result
= result_src
;
1784 ir_to_mesa_visitor::visit(ir_return
*ir
)
1786 if (ir
->get_value()) {
1787 ir_to_mesa_dst_reg l
;
1790 assert(current_function
);
1792 ir
->get_value()->accept(this);
1793 ir_to_mesa_src_reg r
= this->result
;
1795 l
= ir_to_mesa_dst_reg_from_src(current_function
->return_reg
);
1797 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
1798 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1804 ir_to_mesa_emit_op0(ir
, OPCODE_RET
);
1808 ir_to_mesa_visitor::visit(ir_discard
*ir
)
1810 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1812 assert(ir
->condition
== NULL
); /* FINISHME */
1814 ir_to_mesa_emit_op0(ir
, OPCODE_KIL_NV
);
1815 fp
->UsesKill
= GL_TRUE
;
1819 ir_to_mesa_visitor::visit(ir_if
*ir
)
1821 ir_to_mesa_instruction
*cond_inst
, *if_inst
, *else_inst
= NULL
;
1822 ir_to_mesa_instruction
*prev_inst
;
1824 prev_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
1826 ir
->condition
->accept(this);
1827 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
1829 if (this->options
->EmitCondCodes
) {
1830 cond_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
1832 /* See if we actually generated any instruction for generating
1833 * the condition. If not, then cook up a move to a temp so we
1834 * have something to set cond_update on.
1836 if (cond_inst
== prev_inst
) {
1837 ir_to_mesa_src_reg temp
= get_temp(glsl_type::bool_type
);
1838 cond_inst
= ir_to_mesa_emit_op1(ir
->condition
, OPCODE_MOV
,
1839 ir_to_mesa_dst_reg_from_src(temp
),
1842 cond_inst
->cond_update
= GL_TRUE
;
1844 if_inst
= ir_to_mesa_emit_op0(ir
->condition
, OPCODE_IF
);
1845 if_inst
->dst_reg
.cond_mask
= COND_NE
;
1847 if_inst
= ir_to_mesa_emit_op1(ir
->condition
,
1848 OPCODE_IF
, ir_to_mesa_undef_dst
,
1852 this->instructions
.push_tail(if_inst
);
1854 visit_exec_list(&ir
->then_instructions
, this);
1856 if (!ir
->else_instructions
.is_empty()) {
1857 else_inst
= ir_to_mesa_emit_op0(ir
->condition
, OPCODE_ELSE
);
1858 visit_exec_list(&ir
->else_instructions
, this);
1861 if_inst
= ir_to_mesa_emit_op1(ir
->condition
, OPCODE_ENDIF
,
1862 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
1865 ir_to_mesa_visitor::ir_to_mesa_visitor()
1867 result
.file
= PROGRAM_UNDEFINED
;
1869 next_signature_id
= 1;
1870 current_function
= NULL
;
1871 mem_ctx
= talloc_new(NULL
);
1874 ir_to_mesa_visitor::~ir_to_mesa_visitor()
1876 talloc_free(mem_ctx
);
1879 static struct prog_src_register
1880 mesa_src_reg_from_ir_src_reg(ir_to_mesa_src_reg reg
)
1882 struct prog_src_register mesa_reg
;
1884 mesa_reg
.File
= reg
.file
;
1885 assert(reg
.index
< (1 << INST_INDEX_BITS
) - 1);
1886 mesa_reg
.Index
= reg
.index
;
1887 mesa_reg
.Swizzle
= reg
.swizzle
;
1888 mesa_reg
.RelAddr
= reg
.reladdr
!= NULL
;
1889 mesa_reg
.Negate
= reg
.negate
;
1891 mesa_reg
.HasIndex2
= GL_FALSE
;
1892 mesa_reg
.RelAddr2
= 0;
1893 mesa_reg
.Index2
= 0;
1899 set_branchtargets(ir_to_mesa_visitor
*v
,
1900 struct prog_instruction
*mesa_instructions
,
1901 int num_instructions
)
1903 int if_count
= 0, loop_count
= 0;
1904 int *if_stack
, *loop_stack
;
1905 int if_stack_pos
= 0, loop_stack_pos
= 0;
1908 for (i
= 0; i
< num_instructions
; i
++) {
1909 switch (mesa_instructions
[i
].Opcode
) {
1913 case OPCODE_BGNLOOP
:
1918 mesa_instructions
[i
].BranchTarget
= -1;
1925 if_stack
= talloc_zero_array(v
->mem_ctx
, int, if_count
);
1926 loop_stack
= talloc_zero_array(v
->mem_ctx
, int, loop_count
);
1928 for (i
= 0; i
< num_instructions
; i
++) {
1929 switch (mesa_instructions
[i
].Opcode
) {
1931 if_stack
[if_stack_pos
] = i
;
1935 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
1936 if_stack
[if_stack_pos
- 1] = i
;
1939 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
1942 case OPCODE_BGNLOOP
:
1943 loop_stack
[loop_stack_pos
] = i
;
1946 case OPCODE_ENDLOOP
:
1948 /* Rewrite any breaks/conts at this nesting level (haven't
1949 * already had a BranchTarget assigned) to point to the end
1952 for (j
= loop_stack
[loop_stack_pos
]; j
< i
; j
++) {
1953 if (mesa_instructions
[j
].Opcode
== OPCODE_BRK
||
1954 mesa_instructions
[j
].Opcode
== OPCODE_CONT
) {
1955 if (mesa_instructions
[j
].BranchTarget
== -1) {
1956 mesa_instructions
[j
].BranchTarget
= i
;
1960 /* The loop ends point at each other. */
1961 mesa_instructions
[i
].BranchTarget
= loop_stack
[loop_stack_pos
];
1962 mesa_instructions
[loop_stack
[loop_stack_pos
]].BranchTarget
= i
;
1965 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
1966 function_entry
*entry
= (function_entry
*)iter
.get();
1968 if (entry
->sig_id
== mesa_instructions
[i
].BranchTarget
) {
1969 mesa_instructions
[i
].BranchTarget
= entry
->inst
;
1981 print_program(struct prog_instruction
*mesa_instructions
,
1982 ir_instruction
**mesa_instruction_annotation
,
1983 int num_instructions
)
1985 ir_instruction
*last_ir
= NULL
;
1989 for (i
= 0; i
< num_instructions
; i
++) {
1990 struct prog_instruction
*mesa_inst
= mesa_instructions
+ i
;
1991 ir_instruction
*ir
= mesa_instruction_annotation
[i
];
1993 fprintf(stdout
, "%3d: ", i
);
1995 if (last_ir
!= ir
&& ir
) {
1998 for (j
= 0; j
< indent
; j
++) {
1999 fprintf(stdout
, " ");
2005 fprintf(stdout
, " "); /* line number spacing. */
2008 indent
= _mesa_fprint_instruction_opt(stdout
, mesa_inst
, indent
,
2009 PROG_PRINT_DEBUG
, NULL
);
2014 count_resources(struct gl_program
*prog
)
2018 prog
->SamplersUsed
= 0;
2020 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
2021 struct prog_instruction
*inst
= &prog
->Instructions
[i
];
2023 if (_mesa_is_tex_instruction(inst
->Opcode
)) {
2024 prog
->SamplerTargets
[inst
->TexSrcUnit
] =
2025 (gl_texture_index
)inst
->TexSrcTarget
;
2026 prog
->SamplersUsed
|= 1 << inst
->TexSrcUnit
;
2027 if (inst
->TexShadow
) {
2028 prog
->ShadowSamplers
|= 1 << inst
->TexSrcUnit
;
2033 _mesa_update_shader_textures_used(prog
);
2036 struct uniform_sort
{
2037 struct gl_uniform
*u
;
2041 /* The shader_program->Uniforms list is almost sorted in increasing
2042 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2043 * uniforms shared between targets. We need to add parameters in
2044 * increasing order for the targets.
2047 sort_uniforms(const void *a
, const void *b
)
2049 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2050 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2052 return u1
->pos
- u2
->pos
;
2055 /* Add the uniforms to the parameters. The linker chose locations
2056 * in our parameters lists (which weren't created yet), which the
2057 * uniforms code will use to poke values into our parameters list
2058 * when uniforms are updated.
2061 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2062 struct gl_shader
*shader
,
2063 struct gl_program
*prog
)
2066 unsigned int next_sampler
= 0, num_uniforms
= 0;
2067 struct uniform_sort
*sorted_uniforms
;
2069 sorted_uniforms
= talloc_array(NULL
, struct uniform_sort
,
2070 shader_program
->Uniforms
->NumUniforms
);
2072 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2073 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2074 int parameter_index
= -1;
2076 switch (shader
->Type
) {
2077 case GL_VERTEX_SHADER
:
2078 parameter_index
= uniform
->VertPos
;
2080 case GL_FRAGMENT_SHADER
:
2081 parameter_index
= uniform
->FragPos
;
2083 case GL_GEOMETRY_SHADER
:
2084 parameter_index
= uniform
->GeomPos
;
2088 /* Only add uniforms used in our target. */
2089 if (parameter_index
!= -1) {
2090 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2091 sorted_uniforms
[num_uniforms
].u
= uniform
;
2096 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2099 for (i
= 0; i
< num_uniforms
; i
++) {
2100 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2101 int parameter_index
= sorted_uniforms
[i
].pos
;
2102 const glsl_type
*type
= uniform
->Type
;
2105 if (type
->is_vector() ||
2106 type
->is_scalar()) {
2107 size
= type
->vector_elements
;
2109 size
= type_size(type
) * 4;
2112 gl_register_file file
;
2113 if (type
->is_sampler() ||
2114 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2115 file
= PROGRAM_SAMPLER
;
2117 file
= PROGRAM_UNIFORM
;
2120 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2124 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2125 uniform
->Name
, size
, type
->gl_type
,
2128 /* Sampler uniform values are stored in prog->SamplerUnits,
2129 * and the entry in that array is selected by this index we
2130 * store in ParameterValues[].
2132 if (file
== PROGRAM_SAMPLER
) {
2133 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2134 prog
->Parameters
->ParameterValues
[index
+ j
][0] = next_sampler
++;
2137 /* The location chosen in the Parameters list here (returned
2138 * from _mesa_add_uniform) has to match what the linker chose.
2140 if (index
!= parameter_index
) {
2141 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2142 "failed (%d vs %d)\n",
2143 uniform
->Name
, index
, parameter_index
);
2148 talloc_free(sorted_uniforms
);
2152 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2153 struct gl_shader_program
*shader_program
,
2154 const char *name
, const glsl_type
*type
,
2157 if (type
->is_record()) {
2158 ir_constant
*field_constant
;
2160 field_constant
= (ir_constant
*)val
->components
.get_head();
2162 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2163 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2164 const char *field_name
= talloc_asprintf(mem_ctx
, "%s.%s", name
,
2165 type
->fields
.structure
[i
].name
);
2166 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2167 field_type
, field_constant
);
2168 field_constant
= (ir_constant
*)field_constant
->next
;
2173 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2176 fail_link(shader_program
,
2177 "Couldn't find uniform for initializer %s\n", name
);
2181 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2182 ir_constant
*element
;
2183 const glsl_type
*element_type
;
2184 if (type
->is_array()) {
2185 element
= val
->array_elements
[i
];
2186 element_type
= type
->fields
.array
;
2189 element_type
= type
;
2194 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2195 int *conv
= talloc_array(mem_ctx
, int, element_type
->components());
2196 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2197 conv
[j
] = element
->value
.b
[j
];
2199 values
= (void *)conv
;
2200 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2201 element_type
->vector_elements
,
2204 values
= &element
->value
;
2207 if (element_type
->is_matrix()) {
2208 _mesa_uniform_matrix(ctx
, shader_program
,
2209 element_type
->matrix_columns
,
2210 element_type
->vector_elements
,
2211 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2212 loc
+= element_type
->matrix_columns
;
2214 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2215 values
, element_type
->gl_type
);
2216 loc
+= type_size(element_type
);
2222 set_uniform_initializers(struct gl_context
*ctx
,
2223 struct gl_shader_program
*shader_program
)
2225 void *mem_ctx
= NULL
;
2227 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2228 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2233 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2234 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2235 ir_variable
*var
= ir
->as_variable();
2237 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2241 mem_ctx
= talloc_new(NULL
);
2243 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2244 var
->type
, var
->constant_value
);
2248 talloc_free(mem_ctx
);
2253 * Convert a shader's GLSL IR into a Mesa gl_program.
2256 get_mesa_program(struct gl_context
*ctx
, struct gl_shader_program
*shader_program
,
2257 struct gl_shader
*shader
)
2259 ir_to_mesa_visitor v
;
2260 struct prog_instruction
*mesa_instructions
, *mesa_inst
;
2261 ir_instruction
**mesa_instruction_annotation
;
2263 struct gl_program
*prog
;
2265 const char *target_string
;
2267 struct gl_shader_compiler_options
*options
=
2268 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
2270 switch (shader
->Type
) {
2271 case GL_VERTEX_SHADER
:
2272 target
= GL_VERTEX_PROGRAM_ARB
;
2273 target_string
= "vertex";
2275 case GL_FRAGMENT_SHADER
:
2276 target
= GL_FRAGMENT_PROGRAM_ARB
;
2277 target_string
= "fragment";
2280 assert(!"should not be reached");
2284 validate_ir_tree(shader
->ir
);
2286 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
2289 prog
->Parameters
= _mesa_new_parameter_list();
2290 prog
->Varying
= _mesa_new_parameter_list();
2291 prog
->Attributes
= _mesa_new_parameter_list();
2294 v
.shader_program
= shader_program
;
2295 v
.options
= options
;
2297 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
2299 /* Emit Mesa IR for main(). */
2300 visit_exec_list(shader
->ir
, &v
);
2301 v
.ir_to_mesa_emit_op0(NULL
, OPCODE_END
);
2303 /* Now emit bodies for any functions that were used. */
2305 progress
= GL_FALSE
;
2307 foreach_iter(exec_list_iterator
, iter
, v
.function_signatures
) {
2308 function_entry
*entry
= (function_entry
*)iter
.get();
2310 if (!entry
->bgn_inst
) {
2311 v
.current_function
= entry
;
2313 entry
->bgn_inst
= v
.ir_to_mesa_emit_op0(NULL
, OPCODE_BGNSUB
);
2314 entry
->bgn_inst
->function
= entry
;
2316 visit_exec_list(&entry
->sig
->body
, &v
);
2318 ir_to_mesa_instruction
*last
;
2319 last
= (ir_to_mesa_instruction
*)v
.instructions
.get_tail();
2320 if (last
->op
!= OPCODE_RET
)
2321 v
.ir_to_mesa_emit_op0(NULL
, OPCODE_RET
);
2323 ir_to_mesa_instruction
*end
;
2324 end
= v
.ir_to_mesa_emit_op0(NULL
, OPCODE_ENDSUB
);
2325 end
->function
= entry
;
2332 prog
->NumTemporaries
= v
.next_temp
;
2334 int num_instructions
= 0;
2335 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2340 (struct prog_instruction
*)calloc(num_instructions
,
2341 sizeof(*mesa_instructions
));
2342 mesa_instruction_annotation
= talloc_array(v
.mem_ctx
, ir_instruction
*,
2345 /* Convert ir_mesa_instructions into prog_instructions.
2347 mesa_inst
= mesa_instructions
;
2349 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2350 const ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
2352 mesa_inst
->Opcode
= inst
->op
;
2353 mesa_inst
->CondUpdate
= inst
->cond_update
;
2354 mesa_inst
->DstReg
.File
= inst
->dst_reg
.file
;
2355 mesa_inst
->DstReg
.Index
= inst
->dst_reg
.index
;
2356 mesa_inst
->DstReg
.CondMask
= inst
->dst_reg
.cond_mask
;
2357 mesa_inst
->DstReg
.WriteMask
= inst
->dst_reg
.writemask
;
2358 mesa_inst
->DstReg
.RelAddr
= inst
->dst_reg
.reladdr
!= NULL
;
2359 mesa_inst
->SrcReg
[0] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[0]);
2360 mesa_inst
->SrcReg
[1] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[1]);
2361 mesa_inst
->SrcReg
[2] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[2]);
2362 mesa_inst
->TexSrcUnit
= inst
->sampler
;
2363 mesa_inst
->TexSrcTarget
= inst
->tex_target
;
2364 mesa_inst
->TexShadow
= inst
->tex_shadow
;
2365 mesa_instruction_annotation
[i
] = inst
->ir
;
2367 /* Set IndirectRegisterFiles. */
2368 if (mesa_inst
->DstReg
.RelAddr
)
2369 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->DstReg
.File
;
2371 /* Update program's bitmask of indirectly accessed register files */
2372 for (unsigned src
= 0; src
< 3; src
++)
2373 if (mesa_inst
->SrcReg
[src
].RelAddr
)
2374 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->SrcReg
[src
].File
;
2376 if (options
->EmitNoIfs
&& mesa_inst
->Opcode
== OPCODE_IF
) {
2377 fail_link(shader_program
, "Couldn't flatten if statement\n");
2380 switch (mesa_inst
->Opcode
) {
2382 inst
->function
->inst
= i
;
2383 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2386 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2389 mesa_inst
->BranchTarget
= inst
->function
->sig_id
; /* rewritten later */
2392 prog
->NumAddressRegs
= 1;
2402 set_branchtargets(&v
, mesa_instructions
, num_instructions
);
2404 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2406 printf("GLSL IR for linked %s program %d:\n", target_string
,
2407 shader_program
->Name
);
2408 _mesa_print_ir(shader
->ir
, NULL
);
2411 printf("Mesa IR for linked %s program %d:\n", target_string
,
2412 shader_program
->Name
);
2413 print_program(mesa_instructions
, mesa_instruction_annotation
,
2417 prog
->Instructions
= mesa_instructions
;
2418 prog
->NumInstructions
= num_instructions
;
2420 do_set_program_inouts(shader
->ir
, prog
);
2421 count_resources(prog
);
2423 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
2425 if ((ctx
->Shader
.Flags
& GLSL_NO_OPT
) == 0) {
2426 _mesa_optimize_program(ctx
, prog
);
2435 * Called via ctx->Driver.CompilerShader().
2437 * XXX can we remove the ctx->Driver.CompileShader() hook?
2440 _mesa_ir_compile_shader(struct gl_context
*ctx
, struct gl_shader
*shader
)
2442 assert(shader
->CompileStatus
);
2451 * Called via ctx->Driver.LinkShader()
2452 * This actually involves converting GLSL IR into Mesa gl_programs with
2453 * code lowering and other optimizations.
2456 _mesa_ir_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
2458 assert(prog
->LinkStatus
);
2460 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2461 if (prog
->_LinkedShaders
[i
] == NULL
)
2465 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
2466 const struct gl_shader_compiler_options
*options
=
2467 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
2473 do_mat_op_to_vec(ir
);
2474 do_mod_to_fract(ir
);
2475 do_div_to_mul_rcp(ir
);
2476 do_explog_to_explog2(ir
);
2478 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
2480 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
2482 if (options
->EmitNoIfs
)
2483 progress
= do_if_to_cond_assign(ir
) || progress
;
2485 if (options
->EmitNoNoise
)
2486 progress
= lower_noise(ir
) || progress
;
2488 /* If there are forms of indirect addressing that the driver
2489 * cannot handle, perform the lowering pass.
2491 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
2492 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
2494 lower_variable_index_to_cond_assign(ir
,
2495 options
->EmitNoIndirectInput
,
2496 options
->EmitNoIndirectOutput
,
2497 options
->EmitNoIndirectTemp
,
2498 options
->EmitNoIndirectUniform
)
2501 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
2504 validate_ir_tree(ir
);
2507 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2508 struct gl_program
*linked_prog
;
2511 if (prog
->_LinkedShaders
[i
] == NULL
)
2514 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
2516 switch (prog
->_LinkedShaders
[i
]->Type
) {
2517 case GL_VERTEX_SHADER
:
2518 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
2519 (struct gl_vertex_program
*)linked_prog
);
2520 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
2523 case GL_FRAGMENT_SHADER
:
2524 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
2525 (struct gl_fragment_program
*)linked_prog
);
2526 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
2533 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
2541 * Compile a GLSL shader. Called via glCompileShader().
2544 _mesa_glsl_compile_shader(struct gl_context
*ctx
, struct gl_shader
*shader
)
2546 struct _mesa_glsl_parse_state
*state
=
2547 new(shader
) _mesa_glsl_parse_state(ctx
, shader
->Type
, shader
);
2549 const char *source
= shader
->Source
;
2550 /* Check if the user called glCompileShader without first calling
2551 * glShaderSource. This should fail to compile, but not raise a GL_ERROR.
2553 if (source
== NULL
) {
2554 shader
->CompileStatus
= GL_FALSE
;
2558 state
->error
= preprocess(state
, &source
, &state
->info_log
,
2559 &ctx
->Extensions
, ctx
->API
);
2561 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2562 printf("GLSL source for shader %d:\n", shader
->Name
);
2563 printf("%s\n", shader
->Source
);
2566 if (!state
->error
) {
2567 _mesa_glsl_lexer_ctor(state
, source
);
2568 _mesa_glsl_parse(state
);
2569 _mesa_glsl_lexer_dtor(state
);
2572 talloc_free(shader
->ir
);
2573 shader
->ir
= new(shader
) exec_list
;
2574 if (!state
->error
&& !state
->translation_unit
.is_empty())
2575 _mesa_ast_to_hir(shader
->ir
, state
);
2577 if (!state
->error
&& !shader
->ir
->is_empty()) {
2578 validate_ir_tree(shader
->ir
);
2580 /* Do some optimization at compile time to reduce shader IR size
2581 * and reduce later work if the same shader is linked multiple times
2583 while (do_common_optimization(shader
->ir
, false, 32))
2586 validate_ir_tree(shader
->ir
);
2589 shader
->symbols
= state
->symbols
;
2591 shader
->CompileStatus
= !state
->error
;
2592 shader
->InfoLog
= state
->info_log
;
2593 shader
->Version
= state
->language_version
;
2594 memcpy(shader
->builtins_to_link
, state
->builtins_to_link
,
2595 sizeof(shader
->builtins_to_link
[0]) * state
->num_builtins_to_link
);
2596 shader
->num_builtins_to_link
= state
->num_builtins_to_link
;
2598 if (ctx
->Shader
.Flags
& GLSL_LOG
) {
2599 _mesa_write_shader_to_file(shader
);
2602 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2603 if (shader
->CompileStatus
) {
2604 printf("GLSL IR for shader %d:\n", shader
->Name
);
2605 _mesa_print_ir(shader
->ir
, NULL
);
2608 printf("GLSL shader %d failed to compile.\n", shader
->Name
);
2610 if (shader
->InfoLog
&& shader
->InfoLog
[0] != 0) {
2611 printf("GLSL shader %d info log:\n", shader
->Name
);
2612 printf("%s\n", shader
->InfoLog
);
2616 /* Retain any live IR, but trash the rest. */
2617 reparent_ir(shader
->ir
, shader
->ir
);
2621 if (shader
->CompileStatus
) {
2622 if (!ctx
->Driver
.CompileShader(ctx
, shader
))
2623 shader
->CompileStatus
= GL_FALSE
;
2629 * Link a GLSL shader program. Called via glLinkProgram().
2632 _mesa_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
2636 _mesa_clear_shader_program_data(ctx
, prog
);
2638 prog
->LinkStatus
= GL_TRUE
;
2640 for (i
= 0; i
< prog
->NumShaders
; i
++) {
2641 if (!prog
->Shaders
[i
]->CompileStatus
) {
2642 fail_link(prog
, "linking with uncompiled shader");
2643 prog
->LinkStatus
= GL_FALSE
;
2647 prog
->Varying
= _mesa_new_parameter_list();
2648 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
2649 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
2651 if (prog
->LinkStatus
) {
2652 link_shaders(ctx
, prog
);
2655 if (prog
->LinkStatus
) {
2656 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
2657 prog
->LinkStatus
= GL_FALSE
;
2661 set_uniform_initializers(ctx
, prog
);
2663 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2664 if (!prog
->LinkStatus
) {
2665 printf("GLSL shader program %d failed to link\n", prog
->Name
);
2668 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
2669 printf("GLSL shader program %d info log:\n", prog
->Name
);
2670 printf("%s\n", prog
->InfoLog
);