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
,
288 bool process_move_condition(ir_rvalue
*ir
);
293 ir_to_mesa_src_reg ir_to_mesa_undef
= ir_to_mesa_src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
295 ir_to_mesa_dst_reg ir_to_mesa_undef_dst
= {
296 PROGRAM_UNDEFINED
, 0, SWIZZLE_NOOP
, COND_TR
, NULL
,
299 ir_to_mesa_dst_reg ir_to_mesa_address_reg
= {
300 PROGRAM_ADDRESS
, 0, WRITEMASK_X
, COND_TR
, NULL
304 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
307 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
311 prog
->InfoLog
= talloc_vasprintf_append(prog
->InfoLog
, fmt
, args
);
314 prog
->LinkStatus
= GL_FALSE
;
318 swizzle_for_size(int size
)
320 int size_swizzles
[4] = {
321 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
322 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
323 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
324 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
327 assert((size
>= 1) && (size
<= 4));
328 return size_swizzles
[size
- 1];
331 ir_to_mesa_instruction
*
332 ir_to_mesa_visitor::ir_to_mesa_emit_op3(ir_instruction
*ir
,
334 ir_to_mesa_dst_reg dst
,
335 ir_to_mesa_src_reg src0
,
336 ir_to_mesa_src_reg src1
,
337 ir_to_mesa_src_reg src2
)
339 ir_to_mesa_instruction
*inst
= new(mem_ctx
) ir_to_mesa_instruction();
342 /* If we have to do relative addressing, we want to load the ARL
343 * reg directly for one of the regs, and preload the other reladdr
344 * sources into temps.
346 num_reladdr
+= dst
.reladdr
!= NULL
;
347 num_reladdr
+= src0
.reladdr
!= NULL
;
348 num_reladdr
+= src1
.reladdr
!= NULL
;
349 num_reladdr
+= src2
.reladdr
!= NULL
;
351 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
352 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
353 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
356 ir_to_mesa_emit_op1(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
,
361 assert(num_reladdr
== 0);
365 inst
->src_reg
[0] = src0
;
366 inst
->src_reg
[1] = src1
;
367 inst
->src_reg
[2] = src2
;
370 inst
->function
= NULL
;
372 this->instructions
.push_tail(inst
);
378 ir_to_mesa_instruction
*
379 ir_to_mesa_visitor::ir_to_mesa_emit_op2(ir_instruction
*ir
,
381 ir_to_mesa_dst_reg dst
,
382 ir_to_mesa_src_reg src0
,
383 ir_to_mesa_src_reg src1
)
385 return ir_to_mesa_emit_op3(ir
, op
, dst
, src0
, src1
, ir_to_mesa_undef
);
388 ir_to_mesa_instruction
*
389 ir_to_mesa_visitor::ir_to_mesa_emit_op1(ir_instruction
*ir
,
391 ir_to_mesa_dst_reg dst
,
392 ir_to_mesa_src_reg src0
)
394 assert(dst
.writemask
!= 0);
395 return ir_to_mesa_emit_op3(ir
, op
, dst
,
396 src0
, ir_to_mesa_undef
, ir_to_mesa_undef
);
399 ir_to_mesa_instruction
*
400 ir_to_mesa_visitor::ir_to_mesa_emit_op0(ir_instruction
*ir
,
403 return ir_to_mesa_emit_op3(ir
, op
, ir_to_mesa_undef_dst
,
410 ir_to_mesa_visitor::ir_to_mesa_emit_dp(ir_instruction
*ir
,
411 ir_to_mesa_dst_reg dst
,
412 ir_to_mesa_src_reg src0
,
413 ir_to_mesa_src_reg src1
,
416 static const gl_inst_opcode dot_opcodes
[] = {
417 OPCODE_DP2
, OPCODE_DP3
, OPCODE_DP4
420 ir_to_mesa_emit_op3(ir
, dot_opcodes
[elements
- 2],
421 dst
, src0
, src1
, ir_to_mesa_undef
);
424 inline ir_to_mesa_dst_reg
425 ir_to_mesa_dst_reg_from_src(ir_to_mesa_src_reg reg
)
427 ir_to_mesa_dst_reg dst_reg
;
429 dst_reg
.file
= reg
.file
;
430 dst_reg
.index
= reg
.index
;
431 dst_reg
.writemask
= WRITEMASK_XYZW
;
432 dst_reg
.cond_mask
= COND_TR
;
433 dst_reg
.reladdr
= reg
.reladdr
;
438 inline ir_to_mesa_src_reg
439 ir_to_mesa_src_reg_from_dst(ir_to_mesa_dst_reg reg
)
441 return ir_to_mesa_src_reg(reg
.file
, reg
.index
, NULL
);
445 * Emits Mesa scalar opcodes to produce unique answers across channels.
447 * Some Mesa opcodes are scalar-only, like ARB_fp/vp. The src X
448 * channel determines the result across all channels. So to do a vec4
449 * of this operation, we want to emit a scalar per source channel used
450 * to produce dest channels.
453 ir_to_mesa_visitor::ir_to_mesa_emit_scalar_op2(ir_instruction
*ir
,
455 ir_to_mesa_dst_reg dst
,
456 ir_to_mesa_src_reg orig_src0
,
457 ir_to_mesa_src_reg orig_src1
)
460 int done_mask
= ~dst
.writemask
;
462 /* Mesa RCP is a scalar operation splatting results to all channels,
463 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
466 for (i
= 0; i
< 4; i
++) {
467 GLuint this_mask
= (1 << i
);
468 ir_to_mesa_instruction
*inst
;
469 ir_to_mesa_src_reg src0
= orig_src0
;
470 ir_to_mesa_src_reg src1
= orig_src1
;
472 if (done_mask
& this_mask
)
475 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
476 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
477 for (j
= i
+ 1; j
< 4; j
++) {
478 if (!(done_mask
& (1 << j
)) &&
479 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
480 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
481 this_mask
|= (1 << j
);
484 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
485 src0_swiz
, src0_swiz
);
486 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
487 src1_swiz
, src1_swiz
);
489 inst
= ir_to_mesa_emit_op2(ir
, op
,
493 inst
->dst_reg
.writemask
= this_mask
;
494 done_mask
|= this_mask
;
499 ir_to_mesa_visitor::ir_to_mesa_emit_scalar_op1(ir_instruction
*ir
,
501 ir_to_mesa_dst_reg dst
,
502 ir_to_mesa_src_reg src0
)
504 ir_to_mesa_src_reg undef
= ir_to_mesa_undef
;
506 undef
.swizzle
= SWIZZLE_XXXX
;
508 ir_to_mesa_emit_scalar_op2(ir
, op
, dst
, src0
, undef
);
511 struct ir_to_mesa_src_reg
512 ir_to_mesa_visitor::src_reg_for_float(float val
)
514 ir_to_mesa_src_reg
src_reg(PROGRAM_CONSTANT
, -1, NULL
);
516 src_reg
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
517 &val
, 1, &src_reg
.swizzle
);
523 type_size(const struct glsl_type
*type
)
528 switch (type
->base_type
) {
531 case GLSL_TYPE_FLOAT
:
533 if (type
->is_matrix()) {
534 return type
->matrix_columns
;
536 /* Regardless of size of vector, it gets a vec4. This is bad
537 * packing for things like floats, but otherwise arrays become a
538 * mess. Hopefully a later pass over the code can pack scalars
539 * down if appropriate.
543 case GLSL_TYPE_ARRAY
:
544 return type_size(type
->fields
.array
) * type
->length
;
545 case GLSL_TYPE_STRUCT
:
547 for (i
= 0; i
< type
->length
; i
++) {
548 size
+= type_size(type
->fields
.structure
[i
].type
);
551 case GLSL_TYPE_SAMPLER
:
552 /* Samplers take up one slot in UNIFORMS[], but they're baked in
563 * In the initial pass of codegen, we assign temporary numbers to
564 * intermediate results. (not SSA -- variable assignments will reuse
565 * storage). Actual register allocation for the Mesa VM occurs in a
566 * pass over the Mesa IR later.
569 ir_to_mesa_visitor::get_temp(const glsl_type
*type
)
571 ir_to_mesa_src_reg src_reg
;
575 src_reg
.file
= PROGRAM_TEMPORARY
;
576 src_reg
.index
= next_temp
;
577 src_reg
.reladdr
= NULL
;
578 next_temp
+= type_size(type
);
580 if (type
->is_array() || type
->is_record()) {
581 src_reg
.swizzle
= SWIZZLE_NOOP
;
583 for (i
= 0; i
< type
->vector_elements
; i
++)
586 swizzle
[i
] = type
->vector_elements
- 1;
587 src_reg
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1],
588 swizzle
[2], swizzle
[3]);
596 ir_to_mesa_visitor::find_variable_storage(ir_variable
*var
)
599 variable_storage
*entry
;
601 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
602 entry
= (variable_storage
*)iter
.get();
604 if (entry
->var
== var
)
612 ir_to_mesa_visitor::visit(ir_variable
*ir
)
614 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
615 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
617 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
618 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
621 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
623 const struct gl_builtin_uniform_desc
*statevar
;
625 for (i
= 0; _mesa_builtin_uniform_desc
[i
].name
; i
++) {
626 if (strcmp(ir
->name
, _mesa_builtin_uniform_desc
[i
].name
) == 0)
630 if (!_mesa_builtin_uniform_desc
[i
].name
) {
631 fail_link(this->shader_program
,
632 "Failed to find builtin uniform `%s'\n", ir
->name
);
636 statevar
= &_mesa_builtin_uniform_desc
[i
];
639 if (ir
->type
->is_array()) {
640 array_count
= ir
->type
->length
;
645 /* Check if this statevar's setup in the STATE file exactly
646 * matches how we'll want to reference it as a
647 * struct/array/whatever. If not, then we need to move it into
648 * temporary storage and hope that it'll get copy-propagated
651 for (i
= 0; i
< statevar
->num_elements
; i
++) {
652 if (statevar
->elements
[i
].swizzle
!= SWIZZLE_XYZW
) {
657 struct variable_storage
*storage
;
658 ir_to_mesa_dst_reg dst
;
659 if (i
== statevar
->num_elements
) {
660 /* We'll set the index later. */
661 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
662 this->variables
.push_tail(storage
);
664 dst
= ir_to_mesa_undef_dst
;
666 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
668 this->variables
.push_tail(storage
);
669 this->next_temp
+= type_size(ir
->type
);
671 dst
= ir_to_mesa_dst_reg_from_src(ir_to_mesa_src_reg(PROGRAM_TEMPORARY
,
677 for (int a
= 0; a
< array_count
; a
++) {
678 for (unsigned int i
= 0; i
< statevar
->num_elements
; i
++) {
679 struct gl_builtin_uniform_element
*element
= &statevar
->elements
[i
];
680 int tokens
[STATE_LENGTH
];
682 memcpy(tokens
, element
->tokens
, sizeof(element
->tokens
));
683 if (ir
->type
->is_array()) {
687 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
688 (gl_state_index
*)tokens
);
690 if (storage
->file
== PROGRAM_STATE_VAR
) {
691 if (storage
->index
== -1) {
692 storage
->index
= index
;
695 (int)(storage
->index
+ a
* statevar
->num_elements
+ i
));
698 ir_to_mesa_src_reg
src(PROGRAM_STATE_VAR
, index
, NULL
);
699 src
.swizzle
= element
->swizzle
;
700 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, dst
, src
);
701 /* even a float takes up a whole vec4 reg in a struct/array. */
706 if (storage
->file
== PROGRAM_TEMPORARY
&&
707 dst
.index
!= storage
->index
+ type_size(ir
->type
)) {
708 fail_link(this->shader_program
,
709 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
710 ir
->name
, dst
.index
- storage
->index
,
711 type_size(ir
->type
));
717 ir_to_mesa_visitor::visit(ir_loop
*ir
)
719 ir_dereference_variable
*counter
= NULL
;
721 if (ir
->counter
!= NULL
)
722 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
724 if (ir
->from
!= NULL
) {
725 assert(ir
->counter
!= NULL
);
727 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
733 ir_to_mesa_emit_op0(NULL
, OPCODE_BGNLOOP
);
737 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
739 ir_if
*if_stmt
= new(ir
) ir_if(e
);
741 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
743 if_stmt
->then_instructions
.push_tail(brk
);
745 if_stmt
->accept(this);
752 visit_exec_list(&ir
->body_instructions
, this);
756 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
757 counter
, ir
->increment
);
759 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
766 ir_to_mesa_emit_op0(NULL
, OPCODE_ENDLOOP
);
770 ir_to_mesa_visitor::visit(ir_loop_jump
*ir
)
773 case ir_loop_jump::jump_break
:
774 ir_to_mesa_emit_op0(NULL
, OPCODE_BRK
);
776 case ir_loop_jump::jump_continue
:
777 ir_to_mesa_emit_op0(NULL
, OPCODE_CONT
);
784 ir_to_mesa_visitor::visit(ir_function_signature
*ir
)
791 ir_to_mesa_visitor::visit(ir_function
*ir
)
793 /* Ignore function bodies other than main() -- we shouldn't see calls to
794 * them since they should all be inlined before we get to ir_to_mesa.
796 if (strcmp(ir
->name
, "main") == 0) {
797 const ir_function_signature
*sig
;
800 sig
= ir
->matching_signature(&empty
);
804 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
805 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
813 ir_to_mesa_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
815 int nonmul_operand
= 1 - mul_operand
;
816 ir_to_mesa_src_reg a
, b
, c
;
818 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
819 if (!expr
|| expr
->operation
!= ir_binop_mul
)
822 expr
->operands
[0]->accept(this);
824 expr
->operands
[1]->accept(this);
826 ir
->operands
[nonmul_operand
]->accept(this);
829 this->result
= get_temp(ir
->type
);
830 ir_to_mesa_emit_op3(ir
, OPCODE_MAD
,
831 ir_to_mesa_dst_reg_from_src(this->result
), a
, b
, c
);
837 ir_to_mesa_visitor::reladdr_to_temp(ir_instruction
*ir
,
838 ir_to_mesa_src_reg
*reg
, int *num_reladdr
)
843 ir_to_mesa_emit_op1(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
, *reg
->reladdr
);
845 if (*num_reladdr
!= 1) {
846 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
848 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
,
849 ir_to_mesa_dst_reg_from_src(temp
), *reg
);
857 ir_to_mesa_visitor::visit(ir_expression
*ir
)
859 unsigned int operand
;
860 struct ir_to_mesa_src_reg op
[2];
861 struct ir_to_mesa_src_reg result_src
;
862 struct ir_to_mesa_dst_reg result_dst
;
864 /* Quick peephole: Emit OPCODE_MAD(a, b, c) instead of ADD(MUL(a, b), c)
866 if (ir
->operation
== ir_binop_add
) {
867 if (try_emit_mad(ir
, 1))
869 if (try_emit_mad(ir
, 0))
873 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
874 this->result
.file
= PROGRAM_UNDEFINED
;
875 ir
->operands
[operand
]->accept(this);
876 if (this->result
.file
== PROGRAM_UNDEFINED
) {
878 printf("Failed to get tree for expression operand:\n");
879 ir
->operands
[operand
]->accept(&v
);
882 op
[operand
] = this->result
;
884 /* Matrix expression operands should have been broken down to vector
885 * operations already.
887 assert(!ir
->operands
[operand
]->type
->is_matrix());
890 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
891 if (ir
->operands
[1]) {
892 vector_elements
= MAX2(vector_elements
,
893 ir
->operands
[1]->type
->vector_elements
);
896 this->result
.file
= PROGRAM_UNDEFINED
;
898 /* Storage for our result. Ideally for an assignment we'd be using
899 * the actual storage for the result here, instead.
901 result_src
= get_temp(ir
->type
);
902 /* convenience for the emit functions below. */
903 result_dst
= ir_to_mesa_dst_reg_from_src(result_src
);
904 /* Limit writes to the channels that will be used by result_src later.
905 * This does limit this temp's use as a temporary for multi-instruction
908 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
910 switch (ir
->operation
) {
911 case ir_unop_logic_not
:
912 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
,
913 op
[0], src_reg_for_float(0.0));
916 op
[0].negate
= ~op
[0].negate
;
920 ir_to_mesa_emit_op1(ir
, OPCODE_ABS
, result_dst
, op
[0]);
923 ir_to_mesa_emit_op1(ir
, OPCODE_SSG
, result_dst
, op
[0]);
926 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RCP
, result_dst
, op
[0]);
930 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_EX2
, result_dst
, op
[0]);
934 assert(!"not reached: should be handled by ir_explog_to_explog2");
937 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_LG2
, result_dst
, op
[0]);
940 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_SIN
, result_dst
, op
[0]);
943 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_COS
, result_dst
, op
[0]);
947 ir_to_mesa_emit_op1(ir
, OPCODE_DDX
, result_dst
, op
[0]);
950 ir_to_mesa_emit_op1(ir
, OPCODE_DDY
, result_dst
, op
[0]);
953 case ir_unop_noise
: {
954 const enum prog_opcode opcode
=
955 prog_opcode(OPCODE_NOISE1
956 + (ir
->operands
[0]->type
->vector_elements
) - 1);
957 assert((opcode
>= OPCODE_NOISE1
) && (opcode
<= OPCODE_NOISE4
));
959 ir_to_mesa_emit_op1(ir
, opcode
, result_dst
, op
[0]);
964 ir_to_mesa_emit_op2(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
967 ir_to_mesa_emit_op2(ir
, OPCODE_SUB
, result_dst
, op
[0], op
[1]);
971 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
974 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
976 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
980 ir_to_mesa_emit_op2(ir
, OPCODE_SLT
, result_dst
, op
[0], op
[1]);
982 case ir_binop_greater
:
983 ir_to_mesa_emit_op2(ir
, OPCODE_SGT
, result_dst
, op
[0], op
[1]);
985 case ir_binop_lequal
:
986 ir_to_mesa_emit_op2(ir
, OPCODE_SLE
, result_dst
, op
[0], op
[1]);
988 case ir_binop_gequal
:
989 ir_to_mesa_emit_op2(ir
, OPCODE_SGE
, result_dst
, op
[0], op
[1]);
992 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
994 case ir_binop_nequal
:
995 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
997 case ir_binop_all_equal
:
998 /* "==" operator producing a scalar boolean. */
999 if (ir
->operands
[0]->type
->is_vector() ||
1000 ir
->operands
[1]->type
->is_vector()) {
1001 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
1002 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1003 ir_to_mesa_dst_reg_from_src(temp
), op
[0], op
[1]);
1004 ir_to_mesa_emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1005 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
,
1006 result_dst
, result_src
, src_reg_for_float(0.0));
1008 ir_to_mesa_emit_op2(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1011 case ir_binop_any_nequal
:
1012 /* "!=" operator producing a scalar boolean. */
1013 if (ir
->operands
[0]->type
->is_vector() ||
1014 ir
->operands
[1]->type
->is_vector()) {
1015 ir_to_mesa_src_reg temp
= get_temp(glsl_type::vec4_type
);
1016 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1017 ir_to_mesa_dst_reg_from_src(temp
), op
[0], op
[1]);
1018 ir_to_mesa_emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1019 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1020 result_dst
, result_src
, src_reg_for_float(0.0));
1022 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1027 assert(ir
->operands
[0]->type
->is_vector());
1028 ir_to_mesa_emit_dp(ir
, result_dst
, op
[0], op
[0],
1029 ir
->operands
[0]->type
->vector_elements
);
1030 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1031 result_dst
, result_src
, src_reg_for_float(0.0));
1034 case ir_binop_logic_xor
:
1035 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1038 case ir_binop_logic_or
:
1039 /* This could be a saturated add and skip the SNE. */
1040 ir_to_mesa_emit_op2(ir
, OPCODE_ADD
,
1044 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
,
1046 result_src
, src_reg_for_float(0.0));
1049 case ir_binop_logic_and
:
1050 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1051 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
1057 assert(ir
->operands
[0]->type
->is_vector());
1058 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1059 ir_to_mesa_emit_dp(ir
, result_dst
, op
[0], op
[1],
1060 ir
->operands
[0]->type
->vector_elements
);
1064 /* sqrt(x) = x * rsq(x). */
1065 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1066 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1067 /* For incoming channels <= 0, set the result to 0. */
1068 op
[0].negate
= ~op
[0].negate
;
1069 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, result_dst
,
1070 op
[0], result_src
, src_reg_for_float(0.0));
1073 ir_to_mesa_emit_scalar_op1(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1078 /* Mesa IR lacks types, ints are stored as truncated floats. */
1082 ir_to_mesa_emit_op1(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1086 ir_to_mesa_emit_op2(ir
, OPCODE_SNE
, result_dst
,
1087 op
[0], src_reg_for_float(0.0));
1090 ir_to_mesa_emit_op1(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1093 op
[0].negate
= ~op
[0].negate
;
1094 ir_to_mesa_emit_op1(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1095 result_src
.negate
= ~result_src
.negate
;
1098 ir_to_mesa_emit_op1(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1101 ir_to_mesa_emit_op1(ir
, OPCODE_FRC
, result_dst
, op
[0]);
1105 ir_to_mesa_emit_op2(ir
, OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1108 ir_to_mesa_emit_op2(ir
, OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1111 ir_to_mesa_emit_scalar_op2(ir
, OPCODE_POW
, result_dst
, op
[0], op
[1]);
1114 case ir_unop_bit_not
:
1116 case ir_binop_lshift
:
1117 case ir_binop_rshift
:
1118 case ir_binop_bit_and
:
1119 case ir_binop_bit_xor
:
1120 case ir_binop_bit_or
:
1121 case ir_unop_round_even
:
1122 assert(!"GLSL 1.30 features unsupported");
1126 this->result
= result_src
;
1131 ir_to_mesa_visitor::visit(ir_swizzle
*ir
)
1133 ir_to_mesa_src_reg src_reg
;
1137 /* Note that this is only swizzles in expressions, not those on the left
1138 * hand side of an assignment, which do write masking. See ir_assignment
1142 ir
->val
->accept(this);
1143 src_reg
= this->result
;
1144 assert(src_reg
.file
!= PROGRAM_UNDEFINED
);
1146 for (i
= 0; i
< 4; i
++) {
1147 if (i
< ir
->type
->vector_elements
) {
1150 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.x
);
1153 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.y
);
1156 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.z
);
1159 swizzle
[i
] = GET_SWZ(src_reg
.swizzle
, ir
->mask
.w
);
1163 /* If the type is smaller than a vec4, replicate the last
1166 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1170 src_reg
.swizzle
= MAKE_SWIZZLE4(swizzle
[0],
1175 this->result
= src_reg
;
1179 ir_to_mesa_visitor::visit(ir_dereference_variable
*ir
)
1181 variable_storage
*entry
= find_variable_storage(ir
->var
);
1184 switch (ir
->var
->mode
) {
1185 case ir_var_uniform
:
1186 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_UNIFORM
,
1188 this->variables
.push_tail(entry
);
1193 /* The linker assigns locations for varyings and attributes,
1194 * including deprecated builtins (like gl_Color), user-assign
1195 * generic attributes (glBindVertexLocation), and
1196 * user-defined varyings.
1198 * FINISHME: We would hit this path for function arguments. Fix!
1200 assert(ir
->var
->location
!= -1);
1201 if (ir
->var
->mode
== ir_var_in
||
1202 ir
->var
->mode
== ir_var_inout
) {
1203 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1207 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1208 ir
->var
->location
>= VERT_ATTRIB_GENERIC0
) {
1209 _mesa_add_attribute(prog
->Attributes
,
1211 _mesa_sizeof_glsl_type(ir
->var
->type
->gl_type
),
1212 ir
->var
->type
->gl_type
,
1213 ir
->var
->location
- VERT_ATTRIB_GENERIC0
);
1216 entry
= new(mem_ctx
) variable_storage(ir
->var
,
1223 case ir_var_temporary
:
1224 entry
= new(mem_ctx
) variable_storage(ir
->var
, PROGRAM_TEMPORARY
,
1226 this->variables
.push_tail(entry
);
1228 next_temp
+= type_size(ir
->var
->type
);
1233 printf("Failed to make storage for %s\n", ir
->var
->name
);
1238 this->result
= ir_to_mesa_src_reg(entry
->file
, entry
->index
, ir
->var
->type
);
1242 ir_to_mesa_visitor::visit(ir_dereference_array
*ir
)
1245 ir_to_mesa_src_reg src_reg
;
1246 int element_size
= type_size(ir
->type
);
1248 index
= ir
->array_index
->constant_expression_value();
1250 ir
->array
->accept(this);
1251 src_reg
= this->result
;
1254 src_reg
.index
+= index
->value
.i
[0] * element_size
;
1256 ir_to_mesa_src_reg array_base
= this->result
;
1257 /* Variable index array dereference. It eats the "vec4" of the
1258 * base of the array and an index that offsets the Mesa register
1261 ir
->array_index
->accept(this);
1263 ir_to_mesa_src_reg index_reg
;
1265 if (element_size
== 1) {
1266 index_reg
= this->result
;
1268 index_reg
= get_temp(glsl_type::float_type
);
1270 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
,
1271 ir_to_mesa_dst_reg_from_src(index_reg
),
1272 this->result
, src_reg_for_float(element_size
));
1275 src_reg
.reladdr
= talloc(mem_ctx
, ir_to_mesa_src_reg
);
1276 memcpy(src_reg
.reladdr
, &index_reg
, sizeof(index_reg
));
1279 /* If the type is smaller than a vec4, replicate the last channel out. */
1280 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1281 src_reg
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1283 src_reg
.swizzle
= SWIZZLE_NOOP
;
1285 this->result
= src_reg
;
1289 ir_to_mesa_visitor::visit(ir_dereference_record
*ir
)
1292 const glsl_type
*struct_type
= ir
->record
->type
;
1295 ir
->record
->accept(this);
1297 for (i
= 0; i
< struct_type
->length
; i
++) {
1298 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1300 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1302 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1303 this->result
.index
+= offset
;
1307 * We want to be careful in assignment setup to hit the actual storage
1308 * instead of potentially using a temporary like we might with the
1309 * ir_dereference handler.
1311 static struct ir_to_mesa_dst_reg
1312 get_assignment_lhs(ir_dereference
*ir
, ir_to_mesa_visitor
*v
)
1314 /* The LHS must be a dereference. If the LHS is a variable indexed array
1315 * access of a vector, it must be separated into a series conditional moves
1316 * before reaching this point (see ir_vec_index_to_cond_assign).
1318 assert(ir
->as_dereference());
1319 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1321 assert(!deref_array
->array
->type
->is_vector());
1324 /* Use the rvalue deref handler for the most part. We'll ignore
1325 * swizzles in it and write swizzles using writemask, though.
1328 return ir_to_mesa_dst_reg_from_src(v
->result
);
1332 * Process the condition of a conditional assignment
1334 * Examines the condition of a conditional assignment to generate the optimal
1335 * first operand of a \c CMP instruction. If the condition is a relational
1336 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1337 * used as the source for the \c CMP instruction. Otherwise the comparison
1338 * is processed to a boolean result, and the boolean result is used as the
1339 * operand to the CMP instruction.
1342 ir_to_mesa_visitor::process_move_condition(ir_rvalue
*ir
)
1344 ir_rvalue
*src_ir
= ir
;
1346 bool switch_order
= false;
1348 ir_expression
*const expr
= ir
->as_expression();
1349 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1350 bool zero_on_left
= false;
1352 if (expr
->operands
[0]->is_zero()) {
1353 src_ir
= expr
->operands
[1];
1354 zero_on_left
= true;
1355 } else if (expr
->operands
[1]->is_zero()) {
1356 src_ir
= expr
->operands
[0];
1357 zero_on_left
= false;
1361 * (a < 0) T F F ( a < 0) T F F
1362 * (0 < a) F F T (-a < 0) F F T
1363 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1364 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1365 * (a > 0) F F T (-a < 0) F F T
1366 * (0 > a) T F F ( a < 0) T F F
1367 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1368 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1370 * Note that exchanging the order of 0 and 'a' in the comparison simply
1371 * means that the value of 'a' should be negated.
1374 switch (expr
->operation
) {
1376 switch_order
= false;
1377 negate
= zero_on_left
;
1380 case ir_binop_greater
:
1381 switch_order
= false;
1382 negate
= !zero_on_left
;
1385 case ir_binop_lequal
:
1386 switch_order
= true;
1387 negate
= !zero_on_left
;
1390 case ir_binop_gequal
:
1391 switch_order
= true;
1392 negate
= zero_on_left
;
1396 /* This isn't the right kind of comparison afterall, so make sure
1397 * the whole condition is visited.
1405 src_ir
->accept(this);
1407 /* We use the OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1408 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1409 * choose which value OPCODE_CMP produces without an extra instruction
1410 * computing the condition.
1413 this->result
.negate
= ~this->result
.negate
;
1415 return switch_order
;
1419 ir_to_mesa_visitor::visit(ir_assignment
*ir
)
1421 struct ir_to_mesa_dst_reg l
;
1422 struct ir_to_mesa_src_reg r
;
1425 ir
->rhs
->accept(this);
1428 l
= get_assignment_lhs(ir
->lhs
, this);
1430 /* FINISHME: This should really set to the correct maximal writemask for each
1431 * FINISHME: component written (in the loops below). This case can only
1432 * FINISHME: occur for matrices, arrays, and structures.
1434 if (ir
->write_mask
== 0) {
1435 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1436 l
.writemask
= WRITEMASK_XYZW
;
1437 } else if (ir
->lhs
->type
->is_scalar()) {
1438 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1439 * FINISHME: W component of fragment shader output zero, work correctly.
1441 l
.writemask
= WRITEMASK_XYZW
;
1444 int first_enabled_chan
= 0;
1447 assert(ir
->lhs
->type
->is_vector());
1448 l
.writemask
= ir
->write_mask
;
1450 for (int i
= 0; i
< 4; i
++) {
1451 if (l
.writemask
& (1 << i
)) {
1452 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1457 /* Swizzle a small RHS vector into the channels being written.
1459 * glsl ir treats write_mask as dictating how many channels are
1460 * present on the RHS while Mesa IR treats write_mask as just
1461 * showing which channels of the vec4 RHS get written.
1463 for (int i
= 0; i
< 4; i
++) {
1464 if (l
.writemask
& (1 << i
))
1465 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1467 swizzles
[i
] = first_enabled_chan
;
1469 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1470 swizzles
[2], swizzles
[3]);
1473 assert(l
.file
!= PROGRAM_UNDEFINED
);
1474 assert(r
.file
!= PROGRAM_UNDEFINED
);
1476 if (ir
->condition
) {
1477 const bool switch_order
= this->process_move_condition(ir
->condition
);
1478 ir_to_mesa_src_reg condition
= this->result
;
1480 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1482 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, l
,
1483 condition
, ir_to_mesa_src_reg_from_dst(l
), r
);
1485 ir_to_mesa_emit_op3(ir
, OPCODE_CMP
, l
,
1486 condition
, r
, ir_to_mesa_src_reg_from_dst(l
));
1493 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1494 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1503 ir_to_mesa_visitor::visit(ir_constant
*ir
)
1505 ir_to_mesa_src_reg src_reg
;
1506 GLfloat stack_vals
[4] = { 0 };
1507 GLfloat
*values
= stack_vals
;
1510 /* Unfortunately, 4 floats is all we can get into
1511 * _mesa_add_unnamed_constant. So, make a temp to store an
1512 * aggregate constant and move each constant value into it. If we
1513 * get lucky, copy propagation will eliminate the extra moves.
1516 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1517 ir_to_mesa_src_reg temp_base
= get_temp(ir
->type
);
1518 ir_to_mesa_dst_reg temp
= ir_to_mesa_dst_reg_from_src(temp_base
);
1520 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1521 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1522 int size
= type_size(field_value
->type
);
1526 field_value
->accept(this);
1527 src_reg
= this->result
;
1529 for (i
= 0; i
< (unsigned int)size
; i
++) {
1530 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, src_reg
);
1536 this->result
= temp_base
;
1540 if (ir
->type
->is_array()) {
1541 ir_to_mesa_src_reg temp_base
= get_temp(ir
->type
);
1542 ir_to_mesa_dst_reg temp
= ir_to_mesa_dst_reg_from_src(temp_base
);
1543 int size
= type_size(ir
->type
->fields
.array
);
1547 for (i
= 0; i
< ir
->type
->length
; i
++) {
1548 ir
->array_elements
[i
]->accept(this);
1549 src_reg
= this->result
;
1550 for (int j
= 0; j
< size
; j
++) {
1551 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, temp
, src_reg
);
1557 this->result
= temp_base
;
1561 if (ir
->type
->is_matrix()) {
1562 ir_to_mesa_src_reg mat
= get_temp(ir
->type
);
1563 ir_to_mesa_dst_reg mat_column
= ir_to_mesa_dst_reg_from_src(mat
);
1565 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1566 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1567 values
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1569 src_reg
= ir_to_mesa_src_reg(PROGRAM_CONSTANT
, -1, NULL
);
1570 src_reg
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1572 ir
->type
->vector_elements
,
1574 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, mat_column
, src_reg
);
1583 src_reg
.file
= PROGRAM_CONSTANT
;
1584 switch (ir
->type
->base_type
) {
1585 case GLSL_TYPE_FLOAT
:
1586 values
= &ir
->value
.f
[0];
1588 case GLSL_TYPE_UINT
:
1589 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1590 values
[i
] = ir
->value
.u
[i
];
1594 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1595 values
[i
] = ir
->value
.i
[i
];
1598 case GLSL_TYPE_BOOL
:
1599 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1600 values
[i
] = ir
->value
.b
[i
];
1604 assert(!"Non-float/uint/int/bool constant");
1607 this->result
= ir_to_mesa_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
1608 this->result
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1610 ir
->type
->vector_elements
,
1611 &this->result
.swizzle
);
1615 ir_to_mesa_visitor::get_function_signature(ir_function_signature
*sig
)
1617 function_entry
*entry
;
1619 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
1620 entry
= (function_entry
*)iter
.get();
1622 if (entry
->sig
== sig
)
1626 entry
= talloc(mem_ctx
, function_entry
);
1628 entry
->sig_id
= this->next_signature_id
++;
1629 entry
->bgn_inst
= NULL
;
1631 /* Allocate storage for all the parameters. */
1632 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
1633 ir_variable
*param
= (ir_variable
*)iter
.get();
1634 variable_storage
*storage
;
1636 storage
= find_variable_storage(param
);
1639 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
1641 this->variables
.push_tail(storage
);
1643 this->next_temp
+= type_size(param
->type
);
1646 if (!sig
->return_type
->is_void()) {
1647 entry
->return_reg
= get_temp(sig
->return_type
);
1649 entry
->return_reg
= ir_to_mesa_undef
;
1652 this->function_signatures
.push_tail(entry
);
1657 ir_to_mesa_visitor::visit(ir_call
*ir
)
1659 ir_to_mesa_instruction
*call_inst
;
1660 ir_function_signature
*sig
= ir
->get_callee();
1661 function_entry
*entry
= get_function_signature(sig
);
1664 /* Process in parameters. */
1665 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
1666 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1667 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1668 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1670 if (param
->mode
== ir_var_in
||
1671 param
->mode
== ir_var_inout
) {
1672 variable_storage
*storage
= find_variable_storage(param
);
1675 param_rval
->accept(this);
1676 ir_to_mesa_src_reg r
= this->result
;
1678 ir_to_mesa_dst_reg l
;
1679 l
.file
= storage
->file
;
1680 l
.index
= storage
->index
;
1682 l
.writemask
= WRITEMASK_XYZW
;
1683 l
.cond_mask
= COND_TR
;
1685 for (i
= 0; i
< type_size(param
->type
); i
++) {
1686 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1694 assert(!sig_iter
.has_next());
1696 /* Emit call instruction */
1697 call_inst
= ir_to_mesa_emit_op1(ir
, OPCODE_CAL
,
1698 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
1699 call_inst
->function
= entry
;
1701 /* Process out parameters. */
1702 sig_iter
= sig
->parameters
.iterator();
1703 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1704 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1705 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1707 if (param
->mode
== ir_var_out
||
1708 param
->mode
== ir_var_inout
) {
1709 variable_storage
*storage
= find_variable_storage(param
);
1712 ir_to_mesa_src_reg r
;
1713 r
.file
= storage
->file
;
1714 r
.index
= storage
->index
;
1716 r
.swizzle
= SWIZZLE_NOOP
;
1719 param_rval
->accept(this);
1720 ir_to_mesa_dst_reg l
= ir_to_mesa_dst_reg_from_src(this->result
);
1722 for (i
= 0; i
< type_size(param
->type
); i
++) {
1723 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1731 assert(!sig_iter
.has_next());
1733 /* Process return value. */
1734 this->result
= entry
->return_reg
;
1738 ir_to_mesa_visitor::visit(ir_texture
*ir
)
1740 ir_to_mesa_src_reg result_src
, coord
, lod_info
, projector
;
1741 ir_to_mesa_dst_reg result_dst
, coord_dst
;
1742 ir_to_mesa_instruction
*inst
= NULL
;
1743 prog_opcode opcode
= OPCODE_NOP
;
1745 ir
->coordinate
->accept(this);
1747 /* Put our coords in a temp. We'll need to modify them for shadow,
1748 * projection, or LOD, so the only case we'd use it as is is if
1749 * we're doing plain old texturing. Mesa IR optimization should
1750 * handle cleaning up our mess in that case.
1752 coord
= get_temp(glsl_type::vec4_type
);
1753 coord_dst
= ir_to_mesa_dst_reg_from_src(coord
);
1754 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
,
1757 if (ir
->projector
) {
1758 ir
->projector
->accept(this);
1759 projector
= this->result
;
1762 /* Storage for our result. Ideally for an assignment we'd be using
1763 * the actual storage for the result here, instead.
1765 result_src
= get_temp(glsl_type::vec4_type
);
1766 result_dst
= ir_to_mesa_dst_reg_from_src(result_src
);
1770 opcode
= OPCODE_TEX
;
1773 opcode
= OPCODE_TXB
;
1774 ir
->lod_info
.bias
->accept(this);
1775 lod_info
= this->result
;
1778 opcode
= OPCODE_TXL
;
1779 ir
->lod_info
.lod
->accept(this);
1780 lod_info
= this->result
;
1784 assert(!"GLSL 1.30 features unsupported");
1788 if (ir
->projector
) {
1789 if (opcode
== OPCODE_TEX
) {
1790 /* Slot the projector in as the last component of the coord. */
1791 coord_dst
.writemask
= WRITEMASK_W
;
1792 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, projector
);
1793 coord_dst
.writemask
= WRITEMASK_XYZW
;
1794 opcode
= OPCODE_TXP
;
1796 ir_to_mesa_src_reg coord_w
= coord
;
1797 coord_w
.swizzle
= SWIZZLE_WWWW
;
1799 /* For the other TEX opcodes there's no projective version
1800 * since the last slot is taken up by lod info. Do the
1801 * projective divide now.
1803 coord_dst
.writemask
= WRITEMASK_W
;
1804 ir_to_mesa_emit_op1(ir
, OPCODE_RCP
, coord_dst
, projector
);
1806 coord_dst
.writemask
= WRITEMASK_XYZ
;
1807 ir_to_mesa_emit_op2(ir
, OPCODE_MUL
, coord_dst
, coord
, coord_w
);
1809 coord_dst
.writemask
= WRITEMASK_XYZW
;
1810 coord
.swizzle
= SWIZZLE_XYZW
;
1814 if (ir
->shadow_comparitor
) {
1815 /* Slot the shadow value in as the second to last component of the
1818 ir
->shadow_comparitor
->accept(this);
1819 coord_dst
.writemask
= WRITEMASK_Z
;
1820 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, this->result
);
1821 coord_dst
.writemask
= WRITEMASK_XYZW
;
1824 if (opcode
== OPCODE_TXL
|| opcode
== OPCODE_TXB
) {
1825 /* Mesa IR stores lod or lod bias in the last channel of the coords. */
1826 coord_dst
.writemask
= WRITEMASK_W
;
1827 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, coord_dst
, lod_info
);
1828 coord_dst
.writemask
= WRITEMASK_XYZW
;
1831 inst
= ir_to_mesa_emit_op1(ir
, opcode
, result_dst
, coord
);
1833 if (ir
->shadow_comparitor
)
1834 inst
->tex_shadow
= GL_TRUE
;
1836 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
1837 this->shader_program
,
1840 const glsl_type
*sampler_type
= ir
->sampler
->type
;
1842 switch (sampler_type
->sampler_dimensionality
) {
1843 case GLSL_SAMPLER_DIM_1D
:
1844 inst
->tex_target
= (sampler_type
->sampler_array
)
1845 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
1847 case GLSL_SAMPLER_DIM_2D
:
1848 inst
->tex_target
= (sampler_type
->sampler_array
)
1849 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
1851 case GLSL_SAMPLER_DIM_3D
:
1852 inst
->tex_target
= TEXTURE_3D_INDEX
;
1854 case GLSL_SAMPLER_DIM_CUBE
:
1855 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
1857 case GLSL_SAMPLER_DIM_RECT
:
1858 inst
->tex_target
= TEXTURE_RECT_INDEX
;
1860 case GLSL_SAMPLER_DIM_BUF
:
1861 assert(!"FINISHME: Implement ARB_texture_buffer_object");
1864 assert(!"Should not get here.");
1867 this->result
= result_src
;
1871 ir_to_mesa_visitor::visit(ir_return
*ir
)
1873 if (ir
->get_value()) {
1874 ir_to_mesa_dst_reg l
;
1877 assert(current_function
);
1879 ir
->get_value()->accept(this);
1880 ir_to_mesa_src_reg r
= this->result
;
1882 l
= ir_to_mesa_dst_reg_from_src(current_function
->return_reg
);
1884 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
1885 ir_to_mesa_emit_op1(ir
, OPCODE_MOV
, l
, r
);
1891 ir_to_mesa_emit_op0(ir
, OPCODE_RET
);
1895 ir_to_mesa_visitor::visit(ir_discard
*ir
)
1897 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1899 assert(ir
->condition
== NULL
); /* FINISHME */
1901 ir_to_mesa_emit_op0(ir
, OPCODE_KIL_NV
);
1902 fp
->UsesKill
= GL_TRUE
;
1906 ir_to_mesa_visitor::visit(ir_if
*ir
)
1908 ir_to_mesa_instruction
*cond_inst
, *if_inst
, *else_inst
= NULL
;
1909 ir_to_mesa_instruction
*prev_inst
;
1911 prev_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
1913 ir
->condition
->accept(this);
1914 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
1916 if (this->options
->EmitCondCodes
) {
1917 cond_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
1919 /* See if we actually generated any instruction for generating
1920 * the condition. If not, then cook up a move to a temp so we
1921 * have something to set cond_update on.
1923 if (cond_inst
== prev_inst
) {
1924 ir_to_mesa_src_reg temp
= get_temp(glsl_type::bool_type
);
1925 cond_inst
= ir_to_mesa_emit_op1(ir
->condition
, OPCODE_MOV
,
1926 ir_to_mesa_dst_reg_from_src(temp
),
1929 cond_inst
->cond_update
= GL_TRUE
;
1931 if_inst
= ir_to_mesa_emit_op0(ir
->condition
, OPCODE_IF
);
1932 if_inst
->dst_reg
.cond_mask
= COND_NE
;
1934 if_inst
= ir_to_mesa_emit_op1(ir
->condition
,
1935 OPCODE_IF
, ir_to_mesa_undef_dst
,
1939 this->instructions
.push_tail(if_inst
);
1941 visit_exec_list(&ir
->then_instructions
, this);
1943 if (!ir
->else_instructions
.is_empty()) {
1944 else_inst
= ir_to_mesa_emit_op0(ir
->condition
, OPCODE_ELSE
);
1945 visit_exec_list(&ir
->else_instructions
, this);
1948 if_inst
= ir_to_mesa_emit_op1(ir
->condition
, OPCODE_ENDIF
,
1949 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
1952 ir_to_mesa_visitor::ir_to_mesa_visitor()
1954 result
.file
= PROGRAM_UNDEFINED
;
1956 next_signature_id
= 1;
1957 current_function
= NULL
;
1958 mem_ctx
= talloc_new(NULL
);
1961 ir_to_mesa_visitor::~ir_to_mesa_visitor()
1963 talloc_free(mem_ctx
);
1966 static struct prog_src_register
1967 mesa_src_reg_from_ir_src_reg(ir_to_mesa_src_reg reg
)
1969 struct prog_src_register mesa_reg
;
1971 mesa_reg
.File
= reg
.file
;
1972 assert(reg
.index
< (1 << INST_INDEX_BITS
) - 1);
1973 mesa_reg
.Index
= reg
.index
;
1974 mesa_reg
.Swizzle
= reg
.swizzle
;
1975 mesa_reg
.RelAddr
= reg
.reladdr
!= NULL
;
1976 mesa_reg
.Negate
= reg
.negate
;
1978 mesa_reg
.HasIndex2
= GL_FALSE
;
1979 mesa_reg
.RelAddr2
= 0;
1980 mesa_reg
.Index2
= 0;
1986 set_branchtargets(ir_to_mesa_visitor
*v
,
1987 struct prog_instruction
*mesa_instructions
,
1988 int num_instructions
)
1990 int if_count
= 0, loop_count
= 0;
1991 int *if_stack
, *loop_stack
;
1992 int if_stack_pos
= 0, loop_stack_pos
= 0;
1995 for (i
= 0; i
< num_instructions
; i
++) {
1996 switch (mesa_instructions
[i
].Opcode
) {
2000 case OPCODE_BGNLOOP
:
2005 mesa_instructions
[i
].BranchTarget
= -1;
2012 if_stack
= talloc_zero_array(v
->mem_ctx
, int, if_count
);
2013 loop_stack
= talloc_zero_array(v
->mem_ctx
, int, loop_count
);
2015 for (i
= 0; i
< num_instructions
; i
++) {
2016 switch (mesa_instructions
[i
].Opcode
) {
2018 if_stack
[if_stack_pos
] = i
;
2022 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2023 if_stack
[if_stack_pos
- 1] = i
;
2026 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2029 case OPCODE_BGNLOOP
:
2030 loop_stack
[loop_stack_pos
] = i
;
2033 case OPCODE_ENDLOOP
:
2035 /* Rewrite any breaks/conts at this nesting level (haven't
2036 * already had a BranchTarget assigned) to point to the end
2039 for (j
= loop_stack
[loop_stack_pos
]; j
< i
; j
++) {
2040 if (mesa_instructions
[j
].Opcode
== OPCODE_BRK
||
2041 mesa_instructions
[j
].Opcode
== OPCODE_CONT
) {
2042 if (mesa_instructions
[j
].BranchTarget
== -1) {
2043 mesa_instructions
[j
].BranchTarget
= i
;
2047 /* The loop ends point at each other. */
2048 mesa_instructions
[i
].BranchTarget
= loop_stack
[loop_stack_pos
];
2049 mesa_instructions
[loop_stack
[loop_stack_pos
]].BranchTarget
= i
;
2052 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
2053 function_entry
*entry
= (function_entry
*)iter
.get();
2055 if (entry
->sig_id
== mesa_instructions
[i
].BranchTarget
) {
2056 mesa_instructions
[i
].BranchTarget
= entry
->inst
;
2068 print_program(struct prog_instruction
*mesa_instructions
,
2069 ir_instruction
**mesa_instruction_annotation
,
2070 int num_instructions
)
2072 ir_instruction
*last_ir
= NULL
;
2076 for (i
= 0; i
< num_instructions
; i
++) {
2077 struct prog_instruction
*mesa_inst
= mesa_instructions
+ i
;
2078 ir_instruction
*ir
= mesa_instruction_annotation
[i
];
2080 fprintf(stdout
, "%3d: ", i
);
2082 if (last_ir
!= ir
&& ir
) {
2085 for (j
= 0; j
< indent
; j
++) {
2086 fprintf(stdout
, " ");
2092 fprintf(stdout
, " "); /* line number spacing. */
2095 indent
= _mesa_fprint_instruction_opt(stdout
, mesa_inst
, indent
,
2096 PROG_PRINT_DEBUG
, NULL
);
2101 count_resources(struct gl_program
*prog
)
2105 prog
->SamplersUsed
= 0;
2107 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
2108 struct prog_instruction
*inst
= &prog
->Instructions
[i
];
2110 if (_mesa_is_tex_instruction(inst
->Opcode
)) {
2111 prog
->SamplerTargets
[inst
->TexSrcUnit
] =
2112 (gl_texture_index
)inst
->TexSrcTarget
;
2113 prog
->SamplersUsed
|= 1 << inst
->TexSrcUnit
;
2114 if (inst
->TexShadow
) {
2115 prog
->ShadowSamplers
|= 1 << inst
->TexSrcUnit
;
2120 _mesa_update_shader_textures_used(prog
);
2123 struct uniform_sort
{
2124 struct gl_uniform
*u
;
2128 /* The shader_program->Uniforms list is almost sorted in increasing
2129 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2130 * uniforms shared between targets. We need to add parameters in
2131 * increasing order for the targets.
2134 sort_uniforms(const void *a
, const void *b
)
2136 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2137 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2139 return u1
->pos
- u2
->pos
;
2142 /* Add the uniforms to the parameters. The linker chose locations
2143 * in our parameters lists (which weren't created yet), which the
2144 * uniforms code will use to poke values into our parameters list
2145 * when uniforms are updated.
2148 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2149 struct gl_shader
*shader
,
2150 struct gl_program
*prog
)
2153 unsigned int next_sampler
= 0, num_uniforms
= 0;
2154 struct uniform_sort
*sorted_uniforms
;
2156 sorted_uniforms
= talloc_array(NULL
, struct uniform_sort
,
2157 shader_program
->Uniforms
->NumUniforms
);
2159 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2160 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2161 int parameter_index
= -1;
2163 switch (shader
->Type
) {
2164 case GL_VERTEX_SHADER
:
2165 parameter_index
= uniform
->VertPos
;
2167 case GL_FRAGMENT_SHADER
:
2168 parameter_index
= uniform
->FragPos
;
2170 case GL_GEOMETRY_SHADER
:
2171 parameter_index
= uniform
->GeomPos
;
2175 /* Only add uniforms used in our target. */
2176 if (parameter_index
!= -1) {
2177 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2178 sorted_uniforms
[num_uniforms
].u
= uniform
;
2183 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2186 for (i
= 0; i
< num_uniforms
; i
++) {
2187 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2188 int parameter_index
= sorted_uniforms
[i
].pos
;
2189 const glsl_type
*type
= uniform
->Type
;
2192 if (type
->is_vector() ||
2193 type
->is_scalar()) {
2194 size
= type
->vector_elements
;
2196 size
= type_size(type
) * 4;
2199 gl_register_file file
;
2200 if (type
->is_sampler() ||
2201 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2202 file
= PROGRAM_SAMPLER
;
2204 file
= PROGRAM_UNIFORM
;
2207 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2211 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2212 uniform
->Name
, size
, type
->gl_type
,
2215 /* Sampler uniform values are stored in prog->SamplerUnits,
2216 * and the entry in that array is selected by this index we
2217 * store in ParameterValues[].
2219 if (file
== PROGRAM_SAMPLER
) {
2220 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2221 prog
->Parameters
->ParameterValues
[index
+ j
][0] = next_sampler
++;
2224 /* The location chosen in the Parameters list here (returned
2225 * from _mesa_add_uniform) has to match what the linker chose.
2227 if (index
!= parameter_index
) {
2228 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2229 "failed (%d vs %d)\n",
2230 uniform
->Name
, index
, parameter_index
);
2235 talloc_free(sorted_uniforms
);
2239 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2240 struct gl_shader_program
*shader_program
,
2241 const char *name
, const glsl_type
*type
,
2244 if (type
->is_record()) {
2245 ir_constant
*field_constant
;
2247 field_constant
= (ir_constant
*)val
->components
.get_head();
2249 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2250 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2251 const char *field_name
= talloc_asprintf(mem_ctx
, "%s.%s", name
,
2252 type
->fields
.structure
[i
].name
);
2253 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2254 field_type
, field_constant
);
2255 field_constant
= (ir_constant
*)field_constant
->next
;
2260 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2263 fail_link(shader_program
,
2264 "Couldn't find uniform for initializer %s\n", name
);
2268 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2269 ir_constant
*element
;
2270 const glsl_type
*element_type
;
2271 if (type
->is_array()) {
2272 element
= val
->array_elements
[i
];
2273 element_type
= type
->fields
.array
;
2276 element_type
= type
;
2281 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2282 int *conv
= talloc_array(mem_ctx
, int, element_type
->components());
2283 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2284 conv
[j
] = element
->value
.b
[j
];
2286 values
= (void *)conv
;
2287 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2288 element_type
->vector_elements
,
2291 values
= &element
->value
;
2294 if (element_type
->is_matrix()) {
2295 _mesa_uniform_matrix(ctx
, shader_program
,
2296 element_type
->matrix_columns
,
2297 element_type
->vector_elements
,
2298 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2299 loc
+= element_type
->matrix_columns
;
2301 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2302 values
, element_type
->gl_type
);
2303 loc
+= type_size(element_type
);
2309 set_uniform_initializers(struct gl_context
*ctx
,
2310 struct gl_shader_program
*shader_program
)
2312 void *mem_ctx
= NULL
;
2314 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2315 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2320 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2321 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2322 ir_variable
*var
= ir
->as_variable();
2324 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2328 mem_ctx
= talloc_new(NULL
);
2330 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2331 var
->type
, var
->constant_value
);
2335 talloc_free(mem_ctx
);
2340 * Convert a shader's GLSL IR into a Mesa gl_program.
2343 get_mesa_program(struct gl_context
*ctx
, struct gl_shader_program
*shader_program
,
2344 struct gl_shader
*shader
)
2346 ir_to_mesa_visitor v
;
2347 struct prog_instruction
*mesa_instructions
, *mesa_inst
;
2348 ir_instruction
**mesa_instruction_annotation
;
2350 struct gl_program
*prog
;
2352 const char *target_string
;
2354 struct gl_shader_compiler_options
*options
=
2355 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
2357 switch (shader
->Type
) {
2358 case GL_VERTEX_SHADER
:
2359 target
= GL_VERTEX_PROGRAM_ARB
;
2360 target_string
= "vertex";
2362 case GL_FRAGMENT_SHADER
:
2363 target
= GL_FRAGMENT_PROGRAM_ARB
;
2364 target_string
= "fragment";
2367 assert(!"should not be reached");
2371 validate_ir_tree(shader
->ir
);
2373 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
2376 prog
->Parameters
= _mesa_new_parameter_list();
2377 prog
->Varying
= _mesa_new_parameter_list();
2378 prog
->Attributes
= _mesa_new_parameter_list();
2381 v
.shader_program
= shader_program
;
2382 v
.options
= options
;
2384 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
2386 /* Emit Mesa IR for main(). */
2387 visit_exec_list(shader
->ir
, &v
);
2388 v
.ir_to_mesa_emit_op0(NULL
, OPCODE_END
);
2390 /* Now emit bodies for any functions that were used. */
2392 progress
= GL_FALSE
;
2394 foreach_iter(exec_list_iterator
, iter
, v
.function_signatures
) {
2395 function_entry
*entry
= (function_entry
*)iter
.get();
2397 if (!entry
->bgn_inst
) {
2398 v
.current_function
= entry
;
2400 entry
->bgn_inst
= v
.ir_to_mesa_emit_op0(NULL
, OPCODE_BGNSUB
);
2401 entry
->bgn_inst
->function
= entry
;
2403 visit_exec_list(&entry
->sig
->body
, &v
);
2405 ir_to_mesa_instruction
*last
;
2406 last
= (ir_to_mesa_instruction
*)v
.instructions
.get_tail();
2407 if (last
->op
!= OPCODE_RET
)
2408 v
.ir_to_mesa_emit_op0(NULL
, OPCODE_RET
);
2410 ir_to_mesa_instruction
*end
;
2411 end
= v
.ir_to_mesa_emit_op0(NULL
, OPCODE_ENDSUB
);
2412 end
->function
= entry
;
2419 prog
->NumTemporaries
= v
.next_temp
;
2421 int num_instructions
= 0;
2422 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2427 (struct prog_instruction
*)calloc(num_instructions
,
2428 sizeof(*mesa_instructions
));
2429 mesa_instruction_annotation
= talloc_array(v
.mem_ctx
, ir_instruction
*,
2432 /* Convert ir_mesa_instructions into prog_instructions.
2434 mesa_inst
= mesa_instructions
;
2436 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2437 const ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
2439 mesa_inst
->Opcode
= inst
->op
;
2440 mesa_inst
->CondUpdate
= inst
->cond_update
;
2441 mesa_inst
->DstReg
.File
= inst
->dst_reg
.file
;
2442 mesa_inst
->DstReg
.Index
= inst
->dst_reg
.index
;
2443 mesa_inst
->DstReg
.CondMask
= inst
->dst_reg
.cond_mask
;
2444 mesa_inst
->DstReg
.WriteMask
= inst
->dst_reg
.writemask
;
2445 mesa_inst
->DstReg
.RelAddr
= inst
->dst_reg
.reladdr
!= NULL
;
2446 mesa_inst
->SrcReg
[0] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[0]);
2447 mesa_inst
->SrcReg
[1] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[1]);
2448 mesa_inst
->SrcReg
[2] = mesa_src_reg_from_ir_src_reg(inst
->src_reg
[2]);
2449 mesa_inst
->TexSrcUnit
= inst
->sampler
;
2450 mesa_inst
->TexSrcTarget
= inst
->tex_target
;
2451 mesa_inst
->TexShadow
= inst
->tex_shadow
;
2452 mesa_instruction_annotation
[i
] = inst
->ir
;
2454 /* Set IndirectRegisterFiles. */
2455 if (mesa_inst
->DstReg
.RelAddr
)
2456 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->DstReg
.File
;
2458 /* Update program's bitmask of indirectly accessed register files */
2459 for (unsigned src
= 0; src
< 3; src
++)
2460 if (mesa_inst
->SrcReg
[src
].RelAddr
)
2461 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->SrcReg
[src
].File
;
2463 if (options
->EmitNoIfs
&& mesa_inst
->Opcode
== OPCODE_IF
) {
2464 fail_link(shader_program
, "Couldn't flatten if statement\n");
2467 switch (mesa_inst
->Opcode
) {
2469 inst
->function
->inst
= i
;
2470 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2473 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2476 mesa_inst
->BranchTarget
= inst
->function
->sig_id
; /* rewritten later */
2479 prog
->NumAddressRegs
= 1;
2489 set_branchtargets(&v
, mesa_instructions
, num_instructions
);
2491 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2493 printf("GLSL IR for linked %s program %d:\n", target_string
,
2494 shader_program
->Name
);
2495 _mesa_print_ir(shader
->ir
, NULL
);
2498 printf("Mesa IR for linked %s program %d:\n", target_string
,
2499 shader_program
->Name
);
2500 print_program(mesa_instructions
, mesa_instruction_annotation
,
2504 prog
->Instructions
= mesa_instructions
;
2505 prog
->NumInstructions
= num_instructions
;
2507 do_set_program_inouts(shader
->ir
, prog
);
2508 count_resources(prog
);
2510 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
2512 if ((ctx
->Shader
.Flags
& GLSL_NO_OPT
) == 0) {
2513 _mesa_optimize_program(ctx
, prog
);
2522 * Called via ctx->Driver.CompilerShader().
2524 * XXX can we remove the ctx->Driver.CompileShader() hook?
2527 _mesa_ir_compile_shader(struct gl_context
*ctx
, struct gl_shader
*shader
)
2529 assert(shader
->CompileStatus
);
2538 * Called via ctx->Driver.LinkShader()
2539 * This actually involves converting GLSL IR into Mesa gl_programs with
2540 * code lowering and other optimizations.
2543 _mesa_ir_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
2545 assert(prog
->LinkStatus
);
2547 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2548 if (prog
->_LinkedShaders
[i
] == NULL
)
2552 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
2553 const struct gl_shader_compiler_options
*options
=
2554 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
2560 do_mat_op_to_vec(ir
);
2561 do_mod_to_fract(ir
);
2562 do_div_to_mul_rcp(ir
);
2563 do_explog_to_explog2(ir
);
2565 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
2567 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
2569 if (options
->EmitNoIfs
)
2570 progress
= do_if_to_cond_assign(ir
) || progress
;
2572 if (options
->EmitNoNoise
)
2573 progress
= lower_noise(ir
) || progress
;
2575 /* If there are forms of indirect addressing that the driver
2576 * cannot handle, perform the lowering pass.
2578 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
2579 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
2581 lower_variable_index_to_cond_assign(ir
,
2582 options
->EmitNoIndirectInput
,
2583 options
->EmitNoIndirectOutput
,
2584 options
->EmitNoIndirectTemp
,
2585 options
->EmitNoIndirectUniform
)
2588 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
2591 validate_ir_tree(ir
);
2594 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2595 struct gl_program
*linked_prog
;
2598 if (prog
->_LinkedShaders
[i
] == NULL
)
2601 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
2603 switch (prog
->_LinkedShaders
[i
]->Type
) {
2604 case GL_VERTEX_SHADER
:
2605 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
2606 (struct gl_vertex_program
*)linked_prog
);
2607 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
2610 case GL_FRAGMENT_SHADER
:
2611 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
2612 (struct gl_fragment_program
*)linked_prog
);
2613 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
2620 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
2628 * Compile a GLSL shader. Called via glCompileShader().
2631 _mesa_glsl_compile_shader(struct gl_context
*ctx
, struct gl_shader
*shader
)
2633 struct _mesa_glsl_parse_state
*state
=
2634 new(shader
) _mesa_glsl_parse_state(ctx
, shader
->Type
, shader
);
2636 const char *source
= shader
->Source
;
2637 /* Check if the user called glCompileShader without first calling
2638 * glShaderSource. This should fail to compile, but not raise a GL_ERROR.
2640 if (source
== NULL
) {
2641 shader
->CompileStatus
= GL_FALSE
;
2645 state
->error
= preprocess(state
, &source
, &state
->info_log
,
2646 &ctx
->Extensions
, ctx
->API
);
2648 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2649 printf("GLSL source for shader %d:\n", shader
->Name
);
2650 printf("%s\n", shader
->Source
);
2653 if (!state
->error
) {
2654 _mesa_glsl_lexer_ctor(state
, source
);
2655 _mesa_glsl_parse(state
);
2656 _mesa_glsl_lexer_dtor(state
);
2659 talloc_free(shader
->ir
);
2660 shader
->ir
= new(shader
) exec_list
;
2661 if (!state
->error
&& !state
->translation_unit
.is_empty())
2662 _mesa_ast_to_hir(shader
->ir
, state
);
2664 if (!state
->error
&& !shader
->ir
->is_empty()) {
2665 validate_ir_tree(shader
->ir
);
2667 /* Do some optimization at compile time to reduce shader IR size
2668 * and reduce later work if the same shader is linked multiple times
2670 while (do_common_optimization(shader
->ir
, false, 32))
2673 validate_ir_tree(shader
->ir
);
2676 shader
->symbols
= state
->symbols
;
2678 shader
->CompileStatus
= !state
->error
;
2679 shader
->InfoLog
= state
->info_log
;
2680 shader
->Version
= state
->language_version
;
2681 memcpy(shader
->builtins_to_link
, state
->builtins_to_link
,
2682 sizeof(shader
->builtins_to_link
[0]) * state
->num_builtins_to_link
);
2683 shader
->num_builtins_to_link
= state
->num_builtins_to_link
;
2685 if (ctx
->Shader
.Flags
& GLSL_LOG
) {
2686 _mesa_write_shader_to_file(shader
);
2689 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2690 if (shader
->CompileStatus
) {
2691 printf("GLSL IR for shader %d:\n", shader
->Name
);
2692 _mesa_print_ir(shader
->ir
, NULL
);
2695 printf("GLSL shader %d failed to compile.\n", shader
->Name
);
2697 if (shader
->InfoLog
&& shader
->InfoLog
[0] != 0) {
2698 printf("GLSL shader %d info log:\n", shader
->Name
);
2699 printf("%s\n", shader
->InfoLog
);
2703 /* Retain any live IR, but trash the rest. */
2704 reparent_ir(shader
->ir
, shader
->ir
);
2708 if (shader
->CompileStatus
) {
2709 if (!ctx
->Driver
.CompileShader(ctx
, shader
))
2710 shader
->CompileStatus
= GL_FALSE
;
2716 * Link a GLSL shader program. Called via glLinkProgram().
2719 _mesa_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
2723 _mesa_clear_shader_program_data(ctx
, prog
);
2725 prog
->LinkStatus
= GL_TRUE
;
2727 for (i
= 0; i
< prog
->NumShaders
; i
++) {
2728 if (!prog
->Shaders
[i
]->CompileStatus
) {
2729 fail_link(prog
, "linking with uncompiled shader");
2730 prog
->LinkStatus
= GL_FALSE
;
2734 prog
->Varying
= _mesa_new_parameter_list();
2735 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
2736 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
2738 if (prog
->LinkStatus
) {
2739 link_shaders(ctx
, prog
);
2742 if (prog
->LinkStatus
) {
2743 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
2744 prog
->LinkStatus
= GL_FALSE
;
2748 set_uniform_initializers(ctx
, prog
);
2750 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
2751 if (!prog
->LinkStatus
) {
2752 printf("GLSL shader program %d failed to link\n", prog
->Name
);
2755 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
2756 printf("GLSL shader program %d info log:\n", prog
->Name
);
2757 printf("%s\n", prog
->InfoLog
);