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"
62 static int swizzle_for_size(int size
);
65 * This struct is a corresponding struct to Mesa prog_src_register, with
70 src_reg(int file
, int index
, const glsl_type
*type
)
72 this->file
= (gl_register_file
) file
;
74 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
75 this->swizzle
= swizzle_for_size(type
->vector_elements
);
77 this->swizzle
= SWIZZLE_XYZW
;
84 this->file
= PROGRAM_UNDEFINED
;
91 explicit src_reg(dst_reg reg
);
93 gl_register_file file
; /**< PROGRAM_* from Mesa */
94 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
95 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
96 int negate
; /**< NEGATE_XYZW mask from mesa */
97 /** Register index should be offset by the integer in this reg. */
103 dst_reg(int file
, int writemask
)
107 this->writemask
= writemask
;
108 this->cond_mask
= COND_TR
;
109 this->reladdr
= NULL
;
114 this->file
= PROGRAM_UNDEFINED
;
117 this->cond_mask
= COND_TR
;
118 this->reladdr
= NULL
;
121 explicit dst_reg(src_reg reg
);
123 int file
; /**< PROGRAM_* from Mesa */
124 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
125 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
127 /** Register index should be offset by the integer in this reg. */
131 src_reg::src_reg(dst_reg reg
)
133 this->file
= (gl_register_file
) reg
.file
;
134 this->index
= reg
.index
;
135 this->swizzle
= SWIZZLE_XYZW
;
137 this->reladdr
= NULL
;
140 dst_reg::dst_reg(src_reg reg
)
142 this->file
= reg
.file
;
143 this->index
= reg
.index
;
144 this->writemask
= WRITEMASK_XYZW
;
145 this->cond_mask
= COND_TR
;
146 this->reladdr
= reg
.reladdr
;
149 extern src_reg ir_to_mesa_undef
;
151 class ir_to_mesa_instruction
: public exec_node
{
153 /* Callers of this ralloc-based new need not call delete. It's
154 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
155 static void* operator new(size_t size
, void *ctx
)
159 node
= rzalloc_size(ctx
, size
);
160 assert(node
!= NULL
);
168 /** Pointer to the ir source this tree came from for debugging */
170 GLboolean cond_update
;
172 int sampler
; /**< sampler index */
173 int tex_target
; /**< One of TEXTURE_*_INDEX */
174 GLboolean tex_shadow
;
176 class function_entry
*function
; /* Set on OPCODE_CAL or OPCODE_BGNSUB */
179 class variable_storage
: public exec_node
{
181 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
182 : file(file
), index(index
), var(var
)
187 gl_register_file file
;
189 ir_variable
*var
; /* variable that maps to this, if any */
192 class function_entry
: public exec_node
{
194 ir_function_signature
*sig
;
197 * identifier of this function signature used by the program.
199 * At the point that Mesa instructions for function calls are
200 * generated, we don't know the address of the first instruction of
201 * the function body. So we make the BranchTarget that is called a
202 * small integer and rewrite them during set_branchtargets().
207 * Pointer to first instruction of the function body.
209 * Set during function body emits after main() is processed.
211 ir_to_mesa_instruction
*bgn_inst
;
214 * Index of the first instruction of the function body in actual
217 * Set after convertion from ir_to_mesa_instruction to prog_instruction.
221 /** Storage for the return value. */
225 class ir_to_mesa_visitor
: public ir_visitor
{
227 ir_to_mesa_visitor();
228 ~ir_to_mesa_visitor();
230 function_entry
*current_function
;
232 struct gl_context
*ctx
;
233 struct gl_program
*prog
;
234 struct gl_shader_program
*shader_program
;
235 struct gl_shader_compiler_options
*options
;
239 variable_storage
*find_variable_storage(ir_variable
*var
);
241 function_entry
*get_function_signature(ir_function_signature
*sig
);
243 src_reg
get_temp(const glsl_type
*type
);
244 void reladdr_to_temp(ir_instruction
*ir
, src_reg
*reg
, int *num_reladdr
);
246 src_reg
src_reg_for_float(float val
);
249 * \name Visit methods
251 * As typical for the visitor pattern, there must be one \c visit method for
252 * each concrete subclass of \c ir_instruction. Virtual base classes within
253 * the hierarchy should not have \c visit methods.
256 virtual void visit(ir_variable
*);
257 virtual void visit(ir_loop
*);
258 virtual void visit(ir_loop_jump
*);
259 virtual void visit(ir_function_signature
*);
260 virtual void visit(ir_function
*);
261 virtual void visit(ir_expression
*);
262 virtual void visit(ir_swizzle
*);
263 virtual void visit(ir_dereference_variable
*);
264 virtual void visit(ir_dereference_array
*);
265 virtual void visit(ir_dereference_record
*);
266 virtual void visit(ir_assignment
*);
267 virtual void visit(ir_constant
*);
268 virtual void visit(ir_call
*);
269 virtual void visit(ir_return
*);
270 virtual void visit(ir_discard
*);
271 virtual void visit(ir_texture
*);
272 virtual void visit(ir_if
*);
277 /** List of variable_storage */
280 /** List of function_entry */
281 exec_list function_signatures
;
282 int next_signature_id
;
284 /** List of ir_to_mesa_instruction */
285 exec_list instructions
;
287 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
);
289 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
290 dst_reg dst
, src_reg src0
);
292 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
293 dst_reg dst
, src_reg src0
, src_reg src1
);
295 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
297 src_reg src0
, src_reg src1
, src_reg src2
);
300 * Emit the correct dot-product instruction for the type of arguments
302 void emit_dp(ir_instruction
*ir
,
308 void emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
309 dst_reg dst
, src_reg src0
);
311 void emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
312 dst_reg dst
, src_reg src0
, src_reg src1
);
314 void emit_scs(ir_instruction
*ir
, enum prog_opcode op
,
315 dst_reg dst
, const src_reg
&src
);
317 GLboolean
try_emit_mad(ir_expression
*ir
,
319 GLboolean
try_emit_sat(ir_expression
*ir
);
321 void emit_swz(ir_expression
*ir
);
323 bool process_move_condition(ir_rvalue
*ir
);
325 void copy_propagate(void);
330 src_reg ir_to_mesa_undef
= src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
332 dst_reg ir_to_mesa_undef_dst
= dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
);
334 dst_reg ir_to_mesa_address_reg
= dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
);
337 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
340 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
344 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
347 prog
->LinkStatus
= GL_FALSE
;
351 swizzle_for_size(int size
)
353 int size_swizzles
[4] = {
354 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
355 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
356 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
357 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
360 assert((size
>= 1) && (size
<= 4));
361 return size_swizzles
[size
- 1];
364 ir_to_mesa_instruction
*
365 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
367 src_reg src0
, src_reg src1
, src_reg src2
)
369 ir_to_mesa_instruction
*inst
= new(mem_ctx
) ir_to_mesa_instruction();
372 /* If we have to do relative addressing, we want to load the ARL
373 * reg directly for one of the regs, and preload the other reladdr
374 * sources into temps.
376 num_reladdr
+= dst
.reladdr
!= NULL
;
377 num_reladdr
+= src0
.reladdr
!= NULL
;
378 num_reladdr
+= src1
.reladdr
!= NULL
;
379 num_reladdr
+= src2
.reladdr
!= NULL
;
381 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
382 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
383 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
386 emit(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
, *dst
.reladdr
);
389 assert(num_reladdr
== 0);
398 inst
->function
= NULL
;
400 this->instructions
.push_tail(inst
);
406 ir_to_mesa_instruction
*
407 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
408 dst_reg dst
, src_reg src0
, src_reg src1
)
410 return emit(ir
, op
, dst
, src0
, src1
, ir_to_mesa_undef
);
413 ir_to_mesa_instruction
*
414 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
415 dst_reg dst
, src_reg src0
)
417 assert(dst
.writemask
!= 0);
418 return emit(ir
, op
, dst
, src0
, ir_to_mesa_undef
, ir_to_mesa_undef
);
421 ir_to_mesa_instruction
*
422 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
)
424 return emit(ir
, op
, ir_to_mesa_undef_dst
,
425 ir_to_mesa_undef
, ir_to_mesa_undef
, ir_to_mesa_undef
);
429 ir_to_mesa_visitor::emit_dp(ir_instruction
*ir
,
430 dst_reg dst
, src_reg src0
, src_reg src1
,
433 static const gl_inst_opcode dot_opcodes
[] = {
434 OPCODE_DP2
, OPCODE_DP3
, OPCODE_DP4
437 emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
, ir_to_mesa_undef
);
441 * Emits Mesa scalar opcodes to produce unique answers across channels.
443 * Some Mesa opcodes are scalar-only, like ARB_fp/vp. The src X
444 * channel determines the result across all channels. So to do a vec4
445 * of this operation, we want to emit a scalar per source channel used
446 * to produce dest channels.
449 ir_to_mesa_visitor::emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
451 src_reg orig_src0
, src_reg orig_src1
)
454 int done_mask
= ~dst
.writemask
;
456 /* Mesa RCP is a scalar operation splatting results to all channels,
457 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
460 for (i
= 0; i
< 4; i
++) {
461 GLuint this_mask
= (1 << i
);
462 ir_to_mesa_instruction
*inst
;
463 src_reg src0
= orig_src0
;
464 src_reg src1
= orig_src1
;
466 if (done_mask
& this_mask
)
469 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
470 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
471 for (j
= i
+ 1; j
< 4; j
++) {
472 /* If there is another enabled component in the destination that is
473 * derived from the same inputs, generate its value on this pass as
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
= emit(ir
, op
, dst
, src0
, src1
);
488 inst
->dst
.writemask
= this_mask
;
489 done_mask
|= this_mask
;
494 ir_to_mesa_visitor::emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
495 dst_reg dst
, src_reg src0
)
497 src_reg undef
= ir_to_mesa_undef
;
499 undef
.swizzle
= SWIZZLE_XXXX
;
501 emit_scalar(ir
, op
, dst
, src0
, undef
);
505 * Emit an OPCODE_SCS instruction
507 * The \c SCS opcode functions a bit differently than the other Mesa (or
508 * ARB_fragment_program) opcodes. Instead of splatting its result across all
509 * four components of the destination, it writes one value to the \c x
510 * component and another value to the \c y component.
512 * \param ir IR instruction being processed
513 * \param op Either \c OPCODE_SIN or \c OPCODE_COS depending on which
515 * \param dst Destination register
516 * \param src Source register
519 ir_to_mesa_visitor::emit_scs(ir_instruction
*ir
, enum prog_opcode op
,
523 /* Vertex programs cannot use the SCS opcode.
525 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
526 emit_scalar(ir
, op
, dst
, src
);
530 const unsigned component
= (op
== OPCODE_SIN
) ? 0 : 1;
531 const unsigned scs_mask
= (1U << component
);
532 int done_mask
= ~dst
.writemask
;
535 assert(op
== OPCODE_SIN
|| op
== OPCODE_COS
);
537 /* If there are compnents in the destination that differ from the component
538 * that will be written by the SCS instrution, we'll need a temporary.
540 if (scs_mask
!= unsigned(dst
.writemask
)) {
541 tmp
= get_temp(glsl_type::vec4_type
);
544 for (unsigned i
= 0; i
< 4; i
++) {
545 unsigned this_mask
= (1U << i
);
548 if ((done_mask
& this_mask
) != 0)
551 /* The source swizzle specified which component of the source generates
552 * sine / cosine for the current component in the destination. The SCS
553 * instruction requires that this value be swizzle to the X component.
554 * Replace the current swizzle with a swizzle that puts the source in
557 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
559 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
560 src0_swiz
, src0_swiz
);
561 for (unsigned j
= i
+ 1; j
< 4; j
++) {
562 /* If there is another enabled component in the destination that is
563 * derived from the same inputs, generate its value on this pass as
566 if (!(done_mask
& (1 << j
)) &&
567 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
568 this_mask
|= (1 << j
);
572 if (this_mask
!= scs_mask
) {
573 ir_to_mesa_instruction
*inst
;
574 dst_reg tmp_dst
= dst_reg(tmp
);
576 /* Emit the SCS instruction.
578 inst
= emit(ir
, OPCODE_SCS
, tmp_dst
, src0
);
579 inst
->dst
.writemask
= scs_mask
;
581 /* Move the result of the SCS instruction to the desired location in
584 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
585 component
, component
);
586 inst
= emit(ir
, OPCODE_SCS
, dst
, tmp
);
587 inst
->dst
.writemask
= this_mask
;
589 /* Emit the SCS instruction to write directly to the destination.
591 ir_to_mesa_instruction
*inst
= emit(ir
, OPCODE_SCS
, dst
, src0
);
592 inst
->dst
.writemask
= scs_mask
;
595 done_mask
|= this_mask
;
600 ir_to_mesa_visitor::src_reg_for_float(float val
)
602 src_reg
src(PROGRAM_CONSTANT
, -1, NULL
);
604 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
605 &val
, 1, &src
.swizzle
);
611 type_size(const struct glsl_type
*type
)
616 switch (type
->base_type
) {
619 case GLSL_TYPE_FLOAT
:
621 if (type
->is_matrix()) {
622 return type
->matrix_columns
;
624 /* Regardless of size of vector, it gets a vec4. This is bad
625 * packing for things like floats, but otherwise arrays become a
626 * mess. Hopefully a later pass over the code can pack scalars
627 * down if appropriate.
631 case GLSL_TYPE_ARRAY
:
632 assert(type
->length
> 0);
633 return type_size(type
->fields
.array
) * type
->length
;
634 case GLSL_TYPE_STRUCT
:
636 for (i
= 0; i
< type
->length
; i
++) {
637 size
+= type_size(type
->fields
.structure
[i
].type
);
640 case GLSL_TYPE_SAMPLER
:
641 /* Samplers take up one slot in UNIFORMS[], but they're baked in
652 * In the initial pass of codegen, we assign temporary numbers to
653 * intermediate results. (not SSA -- variable assignments will reuse
654 * storage). Actual register allocation for the Mesa VM occurs in a
655 * pass over the Mesa IR later.
658 ir_to_mesa_visitor::get_temp(const glsl_type
*type
)
664 src
.file
= PROGRAM_TEMPORARY
;
665 src
.index
= next_temp
;
667 next_temp
+= type_size(type
);
669 if (type
->is_array() || type
->is_record()) {
670 src
.swizzle
= SWIZZLE_NOOP
;
672 for (i
= 0; i
< type
->vector_elements
; i
++)
675 swizzle
[i
] = type
->vector_elements
- 1;
676 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1],
677 swizzle
[2], swizzle
[3]);
685 ir_to_mesa_visitor::find_variable_storage(ir_variable
*var
)
688 variable_storage
*entry
;
690 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
691 entry
= (variable_storage
*)iter
.get();
693 if (entry
->var
== var
)
701 ir_to_mesa_visitor::visit(ir_variable
*ir
)
703 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
704 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
706 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
707 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
709 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
710 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
711 switch (ir
->depth_layout
) {
712 case ir_depth_layout_none
:
713 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
715 case ir_depth_layout_any
:
716 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
718 case ir_depth_layout_greater
:
719 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
721 case ir_depth_layout_less
:
722 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
724 case ir_depth_layout_unchanged
:
725 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
733 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
735 const ir_state_slot
*const slots
= ir
->state_slots
;
736 assert(ir
->state_slots
!= NULL
);
738 /* Check if this statevar's setup in the STATE file exactly
739 * matches how we'll want to reference it as a
740 * struct/array/whatever. If not, then we need to move it into
741 * temporary storage and hope that it'll get copy-propagated
744 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
745 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
750 struct variable_storage
*storage
;
752 if (i
== ir
->num_state_slots
) {
753 /* We'll set the index later. */
754 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
755 this->variables
.push_tail(storage
);
757 dst
= ir_to_mesa_undef_dst
;
759 /* The variable_storage constructor allocates slots based on the size
760 * of the type. However, this had better match the number of state
761 * elements that we're going to copy into the new temporary.
763 assert(ir
->num_state_slots
== type_size(ir
->type
));
765 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
767 this->variables
.push_tail(storage
);
768 this->next_temp
+= type_size(ir
->type
);
770 dst
= dst_reg(src_reg(PROGRAM_TEMPORARY
, storage
->index
, NULL
));
774 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
775 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
776 (gl_state_index
*)slots
[i
].tokens
);
778 if (storage
->file
== PROGRAM_STATE_VAR
) {
779 if (storage
->index
== -1) {
780 storage
->index
= index
;
782 assert(index
== storage
->index
+ (int)i
);
785 src_reg
src(PROGRAM_STATE_VAR
, index
, NULL
);
786 src
.swizzle
= slots
[i
].swizzle
;
787 emit(ir
, OPCODE_MOV
, dst
, src
);
788 /* even a float takes up a whole vec4 reg in a struct/array. */
793 if (storage
->file
== PROGRAM_TEMPORARY
&&
794 dst
.index
!= storage
->index
+ ir
->num_state_slots
) {
795 fail_link(this->shader_program
,
796 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
797 ir
->name
, dst
.index
- storage
->index
,
798 type_size(ir
->type
));
804 ir_to_mesa_visitor::visit(ir_loop
*ir
)
806 ir_dereference_variable
*counter
= NULL
;
808 if (ir
->counter
!= NULL
)
809 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
811 if (ir
->from
!= NULL
) {
812 assert(ir
->counter
!= NULL
);
814 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
820 emit(NULL
, OPCODE_BGNLOOP
);
824 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
826 ir_if
*if_stmt
= new(ir
) ir_if(e
);
828 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
830 if_stmt
->then_instructions
.push_tail(brk
);
832 if_stmt
->accept(this);
839 visit_exec_list(&ir
->body_instructions
, this);
843 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
844 counter
, ir
->increment
);
846 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
853 emit(NULL
, OPCODE_ENDLOOP
);
857 ir_to_mesa_visitor::visit(ir_loop_jump
*ir
)
860 case ir_loop_jump::jump_break
:
861 emit(NULL
, OPCODE_BRK
);
863 case ir_loop_jump::jump_continue
:
864 emit(NULL
, OPCODE_CONT
);
871 ir_to_mesa_visitor::visit(ir_function_signature
*ir
)
878 ir_to_mesa_visitor::visit(ir_function
*ir
)
880 /* Ignore function bodies other than main() -- we shouldn't see calls to
881 * them since they should all be inlined before we get to ir_to_mesa.
883 if (strcmp(ir
->name
, "main") == 0) {
884 const ir_function_signature
*sig
;
887 sig
= ir
->matching_signature(&empty
);
891 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
892 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
900 ir_to_mesa_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
902 int nonmul_operand
= 1 - mul_operand
;
905 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
906 if (!expr
|| expr
->operation
!= ir_binop_mul
)
909 expr
->operands
[0]->accept(this);
911 expr
->operands
[1]->accept(this);
913 ir
->operands
[nonmul_operand
]->accept(this);
916 this->result
= get_temp(ir
->type
);
917 emit(ir
, OPCODE_MAD
, dst_reg(this->result
), a
, b
, c
);
923 ir_to_mesa_visitor::try_emit_sat(ir_expression
*ir
)
925 /* Saturates were only introduced to vertex programs in
926 * NV_vertex_program3, so don't give them to drivers in the VP.
928 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
931 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
935 sat_src
->accept(this);
936 src_reg src
= this->result
;
938 this->result
= get_temp(ir
->type
);
939 ir_to_mesa_instruction
*inst
;
940 inst
= emit(ir
, OPCODE_MOV
, dst_reg(this->result
), src
);
941 inst
->saturate
= true;
947 ir_to_mesa_visitor::reladdr_to_temp(ir_instruction
*ir
,
948 src_reg
*reg
, int *num_reladdr
)
953 emit(ir
, OPCODE_ARL
, ir_to_mesa_address_reg
, *reg
->reladdr
);
955 if (*num_reladdr
!= 1) {
956 src_reg temp
= get_temp(glsl_type::vec4_type
);
958 emit(ir
, OPCODE_MOV
, dst_reg(temp
), *reg
);
966 ir_to_mesa_visitor::emit_swz(ir_expression
*ir
)
968 /* Assume that the vector operator is in a form compatible with OPCODE_SWZ.
969 * This means that each of the operands is either an immediate value of -1,
970 * 0, or 1, or is a component from one source register (possibly with
973 uint8_t components
[4] = { 0 };
974 bool negate
[4] = { false };
975 ir_variable
*var
= NULL
;
977 for (unsigned i
= 0; i
< ir
->type
->vector_elements
; i
++) {
978 ir_rvalue
*op
= ir
->operands
[i
];
980 assert(op
->type
->is_scalar());
983 switch (op
->ir_type
) {
984 case ir_type_constant
: {
986 assert(op
->type
->is_scalar());
988 const ir_constant
*const c
= op
->as_constant();
990 components
[i
] = SWIZZLE_ONE
;
991 } else if (c
->is_zero()) {
992 components
[i
] = SWIZZLE_ZERO
;
993 } else if (c
->is_negative_one()) {
994 components
[i
] = SWIZZLE_ONE
;
997 assert(!"SWZ constant must be 0.0 or 1.0.");
1004 case ir_type_dereference_variable
: {
1005 ir_dereference_variable
*const deref
=
1006 (ir_dereference_variable
*) op
;
1008 assert((var
== NULL
) || (deref
->var
== var
));
1009 components
[i
] = SWIZZLE_X
;
1015 case ir_type_expression
: {
1016 ir_expression
*const expr
= (ir_expression
*) op
;
1018 assert(expr
->operation
== ir_unop_neg
);
1021 op
= expr
->operands
[0];
1025 case ir_type_swizzle
: {
1026 ir_swizzle
*const swiz
= (ir_swizzle
*) op
;
1028 components
[i
] = swiz
->mask
.x
;
1034 assert(!"Should not get here.");
1040 assert(var
!= NULL
);
1042 ir_dereference_variable
*const deref
=
1043 new(mem_ctx
) ir_dereference_variable(var
);
1045 this->result
.file
= PROGRAM_UNDEFINED
;
1046 deref
->accept(this);
1047 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1049 printf("Failed to get tree for expression operand:\n");
1057 src
.swizzle
= MAKE_SWIZZLE4(components
[0],
1061 src
.negate
= ((unsigned(negate
[0]) << 0)
1062 | (unsigned(negate
[1]) << 1)
1063 | (unsigned(negate
[2]) << 2)
1064 | (unsigned(negate
[3]) << 3));
1066 /* Storage for our result. Ideally for an assignment we'd be using the
1067 * actual storage for the result here, instead.
1069 const src_reg result_src
= get_temp(ir
->type
);
1070 dst_reg result_dst
= dst_reg(result_src
);
1072 /* Limit writes to the channels that will be used by result_src later.
1073 * This does limit this temp's use as a temporary for multi-instruction
1076 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1078 emit(ir
, OPCODE_SWZ
, result_dst
, src
);
1079 this->result
= result_src
;
1083 ir_to_mesa_visitor::visit(ir_expression
*ir
)
1085 unsigned int operand
;
1086 src_reg op
[Elements(ir
->operands
)];
1090 /* Quick peephole: Emit OPCODE_MAD(a, b, c) instead of ADD(MUL(a, b), c)
1092 if (ir
->operation
== ir_binop_add
) {
1093 if (try_emit_mad(ir
, 1))
1095 if (try_emit_mad(ir
, 0))
1098 if (try_emit_sat(ir
))
1101 if (ir
->operation
== ir_quadop_vector
) {
1106 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1107 this->result
.file
= PROGRAM_UNDEFINED
;
1108 ir
->operands
[operand
]->accept(this);
1109 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1111 printf("Failed to get tree for expression operand:\n");
1112 ir
->operands
[operand
]->accept(&v
);
1115 op
[operand
] = this->result
;
1117 /* Matrix expression operands should have been broken down to vector
1118 * operations already.
1120 assert(!ir
->operands
[operand
]->type
->is_matrix());
1123 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1124 if (ir
->operands
[1]) {
1125 vector_elements
= MAX2(vector_elements
,
1126 ir
->operands
[1]->type
->vector_elements
);
1129 this->result
.file
= PROGRAM_UNDEFINED
;
1131 /* Storage for our result. Ideally for an assignment we'd be using
1132 * the actual storage for the result here, instead.
1134 result_src
= get_temp(ir
->type
);
1135 /* convenience for the emit functions below. */
1136 result_dst
= dst_reg(result_src
);
1137 /* Limit writes to the channels that will be used by result_src later.
1138 * This does limit this temp's use as a temporary for multi-instruction
1141 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1143 switch (ir
->operation
) {
1144 case ir_unop_logic_not
:
1145 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], src_reg_for_float(0.0));
1148 op
[0].negate
= ~op
[0].negate
;
1152 emit(ir
, OPCODE_ABS
, result_dst
, op
[0]);
1155 emit(ir
, OPCODE_SSG
, result_dst
, op
[0]);
1158 emit_scalar(ir
, OPCODE_RCP
, result_dst
, op
[0]);
1162 emit_scalar(ir
, OPCODE_EX2
, result_dst
, op
[0]);
1166 assert(!"not reached: should be handled by ir_explog_to_explog2");
1169 emit_scalar(ir
, OPCODE_LG2
, result_dst
, op
[0]);
1172 emit_scalar(ir
, OPCODE_SIN
, result_dst
, op
[0]);
1175 emit_scalar(ir
, OPCODE_COS
, result_dst
, op
[0]);
1177 case ir_unop_sin_reduced
:
1178 emit_scs(ir
, OPCODE_SIN
, result_dst
, op
[0]);
1180 case ir_unop_cos_reduced
:
1181 emit_scs(ir
, OPCODE_COS
, result_dst
, op
[0]);
1185 emit(ir
, OPCODE_DDX
, result_dst
, op
[0]);
1188 emit(ir
, OPCODE_DDY
, result_dst
, op
[0]);
1191 case ir_unop_noise
: {
1192 const enum prog_opcode opcode
=
1193 prog_opcode(OPCODE_NOISE1
1194 + (ir
->operands
[0]->type
->vector_elements
) - 1);
1195 assert((opcode
>= OPCODE_NOISE1
) && (opcode
<= OPCODE_NOISE4
));
1197 emit(ir
, opcode
, result_dst
, op
[0]);
1202 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1205 emit(ir
, OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1209 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1212 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1214 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1218 emit(ir
, OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1220 case ir_binop_greater
:
1221 emit(ir
, OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1223 case ir_binop_lequal
:
1224 emit(ir
, OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1226 case ir_binop_gequal
:
1227 emit(ir
, OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1229 case ir_binop_equal
:
1230 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1232 case ir_binop_nequal
:
1233 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1235 case ir_binop_all_equal
:
1236 /* "==" operator producing a scalar boolean. */
1237 if (ir
->operands
[0]->type
->is_vector() ||
1238 ir
->operands
[1]->type
->is_vector()) {
1239 src_reg temp
= get_temp(glsl_type::vec4_type
);
1240 emit(ir
, OPCODE_SNE
, dst_reg(temp
), op
[0], op
[1]);
1241 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1242 emit(ir
, OPCODE_SEQ
, result_dst
, result_src
, src_reg_for_float(0.0));
1244 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1247 case ir_binop_any_nequal
:
1248 /* "!=" operator producing a scalar boolean. */
1249 if (ir
->operands
[0]->type
->is_vector() ||
1250 ir
->operands
[1]->type
->is_vector()) {
1251 src_reg temp
= get_temp(glsl_type::vec4_type
);
1252 emit(ir
, OPCODE_SNE
, dst_reg(temp
), op
[0], op
[1]);
1253 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1254 emit(ir
, OPCODE_SNE
, result_dst
, result_src
, src_reg_for_float(0.0));
1256 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1261 assert(ir
->operands
[0]->type
->is_vector());
1262 emit_dp(ir
, result_dst
, op
[0], op
[0],
1263 ir
->operands
[0]->type
->vector_elements
);
1264 emit(ir
, OPCODE_SNE
, result_dst
, result_src
, src_reg_for_float(0.0));
1267 case ir_binop_logic_xor
:
1268 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1271 case ir_binop_logic_or
:
1272 /* This could be a saturated add and skip the SNE. */
1273 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1274 emit(ir
, OPCODE_SNE
, result_dst
, result_src
, src_reg_for_float(0.0));
1277 case ir_binop_logic_and
:
1278 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1279 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1283 assert(ir
->operands
[0]->type
->is_vector());
1284 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1285 emit_dp(ir
, result_dst
, op
[0], op
[1],
1286 ir
->operands
[0]->type
->vector_elements
);
1290 /* sqrt(x) = x * rsq(x). */
1291 emit_scalar(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1292 emit(ir
, OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1293 /* For incoming channels <= 0, set the result to 0. */
1294 op
[0].negate
= ~op
[0].negate
;
1295 emit(ir
, OPCODE_CMP
, result_dst
,
1296 op
[0], result_src
, src_reg_for_float(0.0));
1299 emit_scalar(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1304 /* Mesa IR lacks types, ints are stored as truncated floats. */
1308 emit(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1312 emit(ir
, OPCODE_SNE
, result_dst
,
1313 op
[0], src_reg_for_float(0.0));
1316 emit(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1319 op
[0].negate
= ~op
[0].negate
;
1320 emit(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1321 result_src
.negate
= ~result_src
.negate
;
1324 emit(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1327 emit(ir
, OPCODE_FRC
, result_dst
, op
[0]);
1331 emit(ir
, OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1334 emit(ir
, OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1337 emit_scalar(ir
, OPCODE_POW
, result_dst
, op
[0], op
[1]);
1340 case ir_unop_bit_not
:
1342 case ir_binop_lshift
:
1343 case ir_binop_rshift
:
1344 case ir_binop_bit_and
:
1345 case ir_binop_bit_xor
:
1346 case ir_binop_bit_or
:
1347 case ir_unop_round_even
:
1348 assert(!"GLSL 1.30 features unsupported");
1351 case ir_quadop_vector
:
1352 /* This operation should have already been handled.
1354 assert(!"Should not get here.");
1358 this->result
= result_src
;
1363 ir_to_mesa_visitor::visit(ir_swizzle
*ir
)
1369 /* Note that this is only swizzles in expressions, not those on the left
1370 * hand side of an assignment, which do write masking. See ir_assignment
1374 ir
->val
->accept(this);
1376 assert(src
.file
!= PROGRAM_UNDEFINED
);
1378 for (i
= 0; i
< 4; i
++) {
1379 if (i
< ir
->type
->vector_elements
) {
1382 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1385 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1388 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1391 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1395 /* If the type is smaller than a vec4, replicate the last
1398 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1402 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1408 ir_to_mesa_visitor::visit(ir_dereference_variable
*ir
)
1410 variable_storage
*entry
= find_variable_storage(ir
->var
);
1411 ir_variable
*var
= ir
->var
;
1414 switch (var
->mode
) {
1415 case ir_var_uniform
:
1416 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1418 this->variables
.push_tail(entry
);
1422 /* The linker assigns locations for varyings and attributes,
1423 * including deprecated builtins (like gl_Color), user-assign
1424 * generic attributes (glBindVertexLocation), and
1425 * user-defined varyings.
1427 * FINISHME: We would hit this path for function arguments. Fix!
1429 assert(var
->location
!= -1);
1430 entry
= new(mem_ctx
) variable_storage(var
,
1433 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1434 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1435 _mesa_add_attribute(this->prog
->Attributes
,
1437 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1439 var
->location
- VERT_ATTRIB_GENERIC0
);
1443 assert(var
->location
!= -1);
1444 entry
= new(mem_ctx
) variable_storage(var
,
1448 case ir_var_system_value
:
1449 entry
= new(mem_ctx
) variable_storage(var
,
1450 PROGRAM_SYSTEM_VALUE
,
1454 case ir_var_temporary
:
1455 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1457 this->variables
.push_tail(entry
);
1459 next_temp
+= type_size(var
->type
);
1464 printf("Failed to make storage for %s\n", var
->name
);
1469 this->result
= src_reg(entry
->file
, entry
->index
, var
->type
);
1473 ir_to_mesa_visitor::visit(ir_dereference_array
*ir
)
1477 int element_size
= type_size(ir
->type
);
1479 index
= ir
->array_index
->constant_expression_value();
1481 ir
->array
->accept(this);
1485 src
.index
+= index
->value
.i
[0] * element_size
;
1487 src_reg array_base
= this->result
;
1488 /* Variable index array dereference. It eats the "vec4" of the
1489 * base of the array and an index that offsets the Mesa register
1492 ir
->array_index
->accept(this);
1496 if (element_size
== 1) {
1497 index_reg
= this->result
;
1499 index_reg
= get_temp(glsl_type::float_type
);
1501 emit(ir
, OPCODE_MUL
, dst_reg(index_reg
),
1502 this->result
, src_reg_for_float(element_size
));
1505 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
1506 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1509 /* If the type is smaller than a vec4, replicate the last channel out. */
1510 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1511 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1513 src
.swizzle
= SWIZZLE_NOOP
;
1519 ir_to_mesa_visitor::visit(ir_dereference_record
*ir
)
1522 const glsl_type
*struct_type
= ir
->record
->type
;
1525 ir
->record
->accept(this);
1527 for (i
= 0; i
< struct_type
->length
; i
++) {
1528 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1530 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1533 /* If the type is smaller than a vec4, replicate the last channel out. */
1534 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1535 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1537 this->result
.swizzle
= SWIZZLE_NOOP
;
1539 this->result
.index
+= offset
;
1543 * We want to be careful in assignment setup to hit the actual storage
1544 * instead of potentially using a temporary like we might with the
1545 * ir_dereference handler.
1548 get_assignment_lhs(ir_dereference
*ir
, ir_to_mesa_visitor
*v
)
1550 /* The LHS must be a dereference. If the LHS is a variable indexed array
1551 * access of a vector, it must be separated into a series conditional moves
1552 * before reaching this point (see ir_vec_index_to_cond_assign).
1554 assert(ir
->as_dereference());
1555 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1557 assert(!deref_array
->array
->type
->is_vector());
1560 /* Use the rvalue deref handler for the most part. We'll ignore
1561 * swizzles in it and write swizzles using writemask, though.
1564 return dst_reg(v
->result
);
1568 * Process the condition of a conditional assignment
1570 * Examines the condition of a conditional assignment to generate the optimal
1571 * first operand of a \c CMP instruction. If the condition is a relational
1572 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1573 * used as the source for the \c CMP instruction. Otherwise the comparison
1574 * is processed to a boolean result, and the boolean result is used as the
1575 * operand to the CMP instruction.
1578 ir_to_mesa_visitor::process_move_condition(ir_rvalue
*ir
)
1580 ir_rvalue
*src_ir
= ir
;
1582 bool switch_order
= false;
1584 ir_expression
*const expr
= ir
->as_expression();
1585 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1586 bool zero_on_left
= false;
1588 if (expr
->operands
[0]->is_zero()) {
1589 src_ir
= expr
->operands
[1];
1590 zero_on_left
= true;
1591 } else if (expr
->operands
[1]->is_zero()) {
1592 src_ir
= expr
->operands
[0];
1593 zero_on_left
= false;
1597 * (a < 0) T F F ( a < 0) T F F
1598 * (0 < a) F F T (-a < 0) F F T
1599 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1600 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1601 * (a > 0) F F T (-a < 0) F F T
1602 * (0 > a) T F F ( a < 0) T F F
1603 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1604 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1606 * Note that exchanging the order of 0 and 'a' in the comparison simply
1607 * means that the value of 'a' should be negated.
1610 switch (expr
->operation
) {
1612 switch_order
= false;
1613 negate
= zero_on_left
;
1616 case ir_binop_greater
:
1617 switch_order
= false;
1618 negate
= !zero_on_left
;
1621 case ir_binop_lequal
:
1622 switch_order
= true;
1623 negate
= !zero_on_left
;
1626 case ir_binop_gequal
:
1627 switch_order
= true;
1628 negate
= zero_on_left
;
1632 /* This isn't the right kind of comparison afterall, so make sure
1633 * the whole condition is visited.
1641 src_ir
->accept(this);
1643 /* We use the OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1644 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1645 * choose which value OPCODE_CMP produces without an extra instruction
1646 * computing the condition.
1649 this->result
.negate
= ~this->result
.negate
;
1651 return switch_order
;
1655 ir_to_mesa_visitor::visit(ir_assignment
*ir
)
1661 ir
->rhs
->accept(this);
1664 l
= get_assignment_lhs(ir
->lhs
, this);
1666 /* FINISHME: This should really set to the correct maximal writemask for each
1667 * FINISHME: component written (in the loops below). This case can only
1668 * FINISHME: occur for matrices, arrays, and structures.
1670 if (ir
->write_mask
== 0) {
1671 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1672 l
.writemask
= WRITEMASK_XYZW
;
1673 } else if (ir
->lhs
->type
->is_scalar()) {
1674 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1675 * FINISHME: W component of fragment shader output zero, work correctly.
1677 l
.writemask
= WRITEMASK_XYZW
;
1680 int first_enabled_chan
= 0;
1683 assert(ir
->lhs
->type
->is_vector());
1684 l
.writemask
= ir
->write_mask
;
1686 for (int i
= 0; i
< 4; i
++) {
1687 if (l
.writemask
& (1 << i
)) {
1688 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1693 /* Swizzle a small RHS vector into the channels being written.
1695 * glsl ir treats write_mask as dictating how many channels are
1696 * present on the RHS while Mesa IR treats write_mask as just
1697 * showing which channels of the vec4 RHS get written.
1699 for (int i
= 0; i
< 4; i
++) {
1700 if (l
.writemask
& (1 << i
))
1701 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1703 swizzles
[i
] = first_enabled_chan
;
1705 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1706 swizzles
[2], swizzles
[3]);
1709 assert(l
.file
!= PROGRAM_UNDEFINED
);
1710 assert(r
.file
!= PROGRAM_UNDEFINED
);
1712 if (ir
->condition
) {
1713 const bool switch_order
= this->process_move_condition(ir
->condition
);
1714 src_reg condition
= this->result
;
1716 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1718 emit(ir
, OPCODE_CMP
, l
, condition
, src_reg(l
), r
);
1720 emit(ir
, OPCODE_CMP
, l
, condition
, r
, src_reg(l
));
1727 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1728 emit(ir
, OPCODE_MOV
, l
, r
);
1737 ir_to_mesa_visitor::visit(ir_constant
*ir
)
1740 GLfloat stack_vals
[4] = { 0 };
1741 GLfloat
*values
= stack_vals
;
1744 /* Unfortunately, 4 floats is all we can get into
1745 * _mesa_add_unnamed_constant. So, make a temp to store an
1746 * aggregate constant and move each constant value into it. If we
1747 * get lucky, copy propagation will eliminate the extra moves.
1750 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1751 src_reg temp_base
= get_temp(ir
->type
);
1752 dst_reg temp
= dst_reg(temp_base
);
1754 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1755 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1756 int size
= type_size(field_value
->type
);
1760 field_value
->accept(this);
1763 for (i
= 0; i
< (unsigned int)size
; i
++) {
1764 emit(ir
, OPCODE_MOV
, temp
, src
);
1770 this->result
= temp_base
;
1774 if (ir
->type
->is_array()) {
1775 src_reg temp_base
= get_temp(ir
->type
);
1776 dst_reg temp
= dst_reg(temp_base
);
1777 int size
= type_size(ir
->type
->fields
.array
);
1781 for (i
= 0; i
< ir
->type
->length
; i
++) {
1782 ir
->array_elements
[i
]->accept(this);
1784 for (int j
= 0; j
< size
; j
++) {
1785 emit(ir
, OPCODE_MOV
, temp
, src
);
1791 this->result
= temp_base
;
1795 if (ir
->type
->is_matrix()) {
1796 src_reg mat
= get_temp(ir
->type
);
1797 dst_reg mat_column
= dst_reg(mat
);
1799 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1800 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1801 values
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1803 src
= src_reg(PROGRAM_CONSTANT
, -1, NULL
);
1804 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1806 ir
->type
->vector_elements
,
1808 emit(ir
, OPCODE_MOV
, mat_column
, src
);
1817 src
.file
= PROGRAM_CONSTANT
;
1818 switch (ir
->type
->base_type
) {
1819 case GLSL_TYPE_FLOAT
:
1820 values
= &ir
->value
.f
[0];
1822 case GLSL_TYPE_UINT
:
1823 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1824 values
[i
] = ir
->value
.u
[i
];
1828 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1829 values
[i
] = ir
->value
.i
[i
];
1832 case GLSL_TYPE_BOOL
:
1833 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1834 values
[i
] = ir
->value
.b
[i
];
1838 assert(!"Non-float/uint/int/bool constant");
1841 this->result
= src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
1842 this->result
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1844 ir
->type
->vector_elements
,
1845 &this->result
.swizzle
);
1849 ir_to_mesa_visitor::get_function_signature(ir_function_signature
*sig
)
1851 function_entry
*entry
;
1853 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
1854 entry
= (function_entry
*)iter
.get();
1856 if (entry
->sig
== sig
)
1860 entry
= ralloc(mem_ctx
, function_entry
);
1862 entry
->sig_id
= this->next_signature_id
++;
1863 entry
->bgn_inst
= NULL
;
1865 /* Allocate storage for all the parameters. */
1866 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
1867 ir_variable
*param
= (ir_variable
*)iter
.get();
1868 variable_storage
*storage
;
1870 storage
= find_variable_storage(param
);
1873 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
1875 this->variables
.push_tail(storage
);
1877 this->next_temp
+= type_size(param
->type
);
1880 if (!sig
->return_type
->is_void()) {
1881 entry
->return_reg
= get_temp(sig
->return_type
);
1883 entry
->return_reg
= ir_to_mesa_undef
;
1886 this->function_signatures
.push_tail(entry
);
1891 ir_to_mesa_visitor::visit(ir_call
*ir
)
1893 ir_to_mesa_instruction
*call_inst
;
1894 ir_function_signature
*sig
= ir
->get_callee();
1895 function_entry
*entry
= get_function_signature(sig
);
1898 /* Process in parameters. */
1899 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
1900 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1901 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1902 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1904 if (param
->mode
== ir_var_in
||
1905 param
->mode
== ir_var_inout
) {
1906 variable_storage
*storage
= find_variable_storage(param
);
1909 param_rval
->accept(this);
1910 src_reg r
= this->result
;
1913 l
.file
= storage
->file
;
1914 l
.index
= storage
->index
;
1916 l
.writemask
= WRITEMASK_XYZW
;
1917 l
.cond_mask
= COND_TR
;
1919 for (i
= 0; i
< type_size(param
->type
); i
++) {
1920 emit(ir
, OPCODE_MOV
, l
, r
);
1928 assert(!sig_iter
.has_next());
1930 /* Emit call instruction */
1931 call_inst
= emit(ir
, OPCODE_CAL
, ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
1932 call_inst
->function
= entry
;
1934 /* Process out parameters. */
1935 sig_iter
= sig
->parameters
.iterator();
1936 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1937 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1938 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1940 if (param
->mode
== ir_var_out
||
1941 param
->mode
== ir_var_inout
) {
1942 variable_storage
*storage
= find_variable_storage(param
);
1946 r
.file
= storage
->file
;
1947 r
.index
= storage
->index
;
1949 r
.swizzle
= SWIZZLE_NOOP
;
1952 param_rval
->accept(this);
1953 dst_reg l
= dst_reg(this->result
);
1955 for (i
= 0; i
< type_size(param
->type
); i
++) {
1956 emit(ir
, OPCODE_MOV
, l
, r
);
1964 assert(!sig_iter
.has_next());
1966 /* Process return value. */
1967 this->result
= entry
->return_reg
;
1971 ir_to_mesa_visitor::visit(ir_texture
*ir
)
1973 src_reg result_src
, coord
, lod_info
, projector
;
1974 dst_reg result_dst
, coord_dst
;
1975 ir_to_mesa_instruction
*inst
= NULL
;
1976 prog_opcode opcode
= OPCODE_NOP
;
1978 ir
->coordinate
->accept(this);
1980 /* Put our coords in a temp. We'll need to modify them for shadow,
1981 * projection, or LOD, so the only case we'd use it as is is if
1982 * we're doing plain old texturing. Mesa IR optimization should
1983 * handle cleaning up our mess in that case.
1985 coord
= get_temp(glsl_type::vec4_type
);
1986 coord_dst
= dst_reg(coord
);
1987 emit(ir
, OPCODE_MOV
, coord_dst
, this->result
);
1989 if (ir
->projector
) {
1990 ir
->projector
->accept(this);
1991 projector
= this->result
;
1994 /* Storage for our result. Ideally for an assignment we'd be using
1995 * the actual storage for the result here, instead.
1997 result_src
= get_temp(glsl_type::vec4_type
);
1998 result_dst
= dst_reg(result_src
);
2002 opcode
= OPCODE_TEX
;
2005 opcode
= OPCODE_TXB
;
2006 ir
->lod_info
.bias
->accept(this);
2007 lod_info
= this->result
;
2010 opcode
= OPCODE_TXL
;
2011 ir
->lod_info
.lod
->accept(this);
2012 lod_info
= this->result
;
2016 assert(!"GLSL 1.30 features unsupported");
2020 if (ir
->projector
) {
2021 if (opcode
== OPCODE_TEX
) {
2022 /* Slot the projector in as the last component of the coord. */
2023 coord_dst
.writemask
= WRITEMASK_W
;
2024 emit(ir
, OPCODE_MOV
, coord_dst
, projector
);
2025 coord_dst
.writemask
= WRITEMASK_XYZW
;
2026 opcode
= OPCODE_TXP
;
2028 src_reg coord_w
= coord
;
2029 coord_w
.swizzle
= SWIZZLE_WWWW
;
2031 /* For the other TEX opcodes there's no projective version
2032 * since the last slot is taken up by lod info. Do the
2033 * projective divide now.
2035 coord_dst
.writemask
= WRITEMASK_W
;
2036 emit(ir
, OPCODE_RCP
, coord_dst
, projector
);
2038 /* In the case where we have to project the coordinates "by hand,"
2039 * the shadow comparitor value must also be projected.
2041 src_reg tmp_src
= coord
;
2042 if (ir
->shadow_comparitor
) {
2043 /* Slot the shadow value in as the second to last component of the
2046 ir
->shadow_comparitor
->accept(this);
2048 tmp_src
= get_temp(glsl_type::vec4_type
);
2049 dst_reg tmp_dst
= dst_reg(tmp_src
);
2051 tmp_dst
.writemask
= WRITEMASK_Z
;
2052 emit(ir
, OPCODE_MOV
, tmp_dst
, this->result
);
2054 tmp_dst
.writemask
= WRITEMASK_XY
;
2055 emit(ir
, OPCODE_MOV
, tmp_dst
, coord
);
2058 coord_dst
.writemask
= WRITEMASK_XYZ
;
2059 emit(ir
, OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2061 coord_dst
.writemask
= WRITEMASK_XYZW
;
2062 coord
.swizzle
= SWIZZLE_XYZW
;
2066 /* If projection is done and the opcode is not OPCODE_TXP, then the shadow
2067 * comparitor was put in the correct place (and projected) by the code,
2068 * above, that handles by-hand projection.
2070 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== OPCODE_TXP
)) {
2071 /* Slot the shadow value in as the second to last component of the
2074 ir
->shadow_comparitor
->accept(this);
2075 coord_dst
.writemask
= WRITEMASK_Z
;
2076 emit(ir
, OPCODE_MOV
, coord_dst
, this->result
);
2077 coord_dst
.writemask
= WRITEMASK_XYZW
;
2080 if (opcode
== OPCODE_TXL
|| opcode
== OPCODE_TXB
) {
2081 /* Mesa IR stores lod or lod bias in the last channel of the coords. */
2082 coord_dst
.writemask
= WRITEMASK_W
;
2083 emit(ir
, OPCODE_MOV
, coord_dst
, lod_info
);
2084 coord_dst
.writemask
= WRITEMASK_XYZW
;
2087 inst
= emit(ir
, opcode
, result_dst
, coord
);
2089 if (ir
->shadow_comparitor
)
2090 inst
->tex_shadow
= GL_TRUE
;
2092 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2093 this->shader_program
,
2096 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2098 switch (sampler_type
->sampler_dimensionality
) {
2099 case GLSL_SAMPLER_DIM_1D
:
2100 inst
->tex_target
= (sampler_type
->sampler_array
)
2101 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2103 case GLSL_SAMPLER_DIM_2D
:
2104 inst
->tex_target
= (sampler_type
->sampler_array
)
2105 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2107 case GLSL_SAMPLER_DIM_3D
:
2108 inst
->tex_target
= TEXTURE_3D_INDEX
;
2110 case GLSL_SAMPLER_DIM_CUBE
:
2111 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2113 case GLSL_SAMPLER_DIM_RECT
:
2114 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2116 case GLSL_SAMPLER_DIM_BUF
:
2117 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2120 assert(!"Should not get here.");
2123 this->result
= result_src
;
2127 ir_to_mesa_visitor::visit(ir_return
*ir
)
2129 if (ir
->get_value()) {
2133 assert(current_function
);
2135 ir
->get_value()->accept(this);
2136 src_reg r
= this->result
;
2138 l
= dst_reg(current_function
->return_reg
);
2140 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2141 emit(ir
, OPCODE_MOV
, l
, r
);
2147 emit(ir
, OPCODE_RET
);
2151 ir_to_mesa_visitor::visit(ir_discard
*ir
)
2153 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2155 if (ir
->condition
) {
2156 ir
->condition
->accept(this);
2157 this->result
.negate
= ~this->result
.negate
;
2158 emit(ir
, OPCODE_KIL
, ir_to_mesa_undef_dst
, this->result
);
2160 emit(ir
, OPCODE_KIL_NV
);
2163 fp
->UsesKill
= GL_TRUE
;
2167 ir_to_mesa_visitor::visit(ir_if
*ir
)
2169 ir_to_mesa_instruction
*cond_inst
, *if_inst
, *else_inst
= NULL
;
2170 ir_to_mesa_instruction
*prev_inst
;
2172 prev_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2174 ir
->condition
->accept(this);
2175 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2177 if (this->options
->EmitCondCodes
) {
2178 cond_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2180 /* See if we actually generated any instruction for generating
2181 * the condition. If not, then cook up a move to a temp so we
2182 * have something to set cond_update on.
2184 if (cond_inst
== prev_inst
) {
2185 src_reg temp
= get_temp(glsl_type::bool_type
);
2186 cond_inst
= emit(ir
->condition
, OPCODE_MOV
, dst_reg(temp
), result
);
2188 cond_inst
->cond_update
= GL_TRUE
;
2190 if_inst
= emit(ir
->condition
, OPCODE_IF
);
2191 if_inst
->dst
.cond_mask
= COND_NE
;
2193 if_inst
= emit(ir
->condition
, OPCODE_IF
, ir_to_mesa_undef_dst
,
2197 this->instructions
.push_tail(if_inst
);
2199 visit_exec_list(&ir
->then_instructions
, this);
2201 if (!ir
->else_instructions
.is_empty()) {
2202 else_inst
= emit(ir
->condition
, OPCODE_ELSE
);
2203 visit_exec_list(&ir
->else_instructions
, this);
2206 if_inst
= emit(ir
->condition
, OPCODE_ENDIF
,
2207 ir_to_mesa_undef_dst
, ir_to_mesa_undef
);
2210 ir_to_mesa_visitor::ir_to_mesa_visitor()
2212 result
.file
= PROGRAM_UNDEFINED
;
2214 next_signature_id
= 1;
2215 current_function
= NULL
;
2216 mem_ctx
= ralloc_context(NULL
);
2219 ir_to_mesa_visitor::~ir_to_mesa_visitor()
2221 ralloc_free(mem_ctx
);
2224 static struct prog_src_register
2225 mesa_src_reg_from_ir_src_reg(src_reg reg
)
2227 struct prog_src_register mesa_reg
;
2229 mesa_reg
.File
= reg
.file
;
2230 assert(reg
.index
< (1 << INST_INDEX_BITS
));
2231 mesa_reg
.Index
= reg
.index
;
2232 mesa_reg
.Swizzle
= reg
.swizzle
;
2233 mesa_reg
.RelAddr
= reg
.reladdr
!= NULL
;
2234 mesa_reg
.Negate
= reg
.negate
;
2236 mesa_reg
.HasIndex2
= GL_FALSE
;
2237 mesa_reg
.RelAddr2
= 0;
2238 mesa_reg
.Index2
= 0;
2244 set_branchtargets(ir_to_mesa_visitor
*v
,
2245 struct prog_instruction
*mesa_instructions
,
2246 int num_instructions
)
2248 int if_count
= 0, loop_count
= 0;
2249 int *if_stack
, *loop_stack
;
2250 int if_stack_pos
= 0, loop_stack_pos
= 0;
2253 for (i
= 0; i
< num_instructions
; i
++) {
2254 switch (mesa_instructions
[i
].Opcode
) {
2258 case OPCODE_BGNLOOP
:
2263 mesa_instructions
[i
].BranchTarget
= -1;
2270 if_stack
= rzalloc_array(v
->mem_ctx
, int, if_count
);
2271 loop_stack
= rzalloc_array(v
->mem_ctx
, int, loop_count
);
2273 for (i
= 0; i
< num_instructions
; i
++) {
2274 switch (mesa_instructions
[i
].Opcode
) {
2276 if_stack
[if_stack_pos
] = i
;
2280 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2281 if_stack
[if_stack_pos
- 1] = i
;
2284 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2287 case OPCODE_BGNLOOP
:
2288 loop_stack
[loop_stack_pos
] = i
;
2291 case OPCODE_ENDLOOP
:
2293 /* Rewrite any breaks/conts at this nesting level (haven't
2294 * already had a BranchTarget assigned) to point to the end
2297 for (j
= loop_stack
[loop_stack_pos
]; j
< i
; j
++) {
2298 if (mesa_instructions
[j
].Opcode
== OPCODE_BRK
||
2299 mesa_instructions
[j
].Opcode
== OPCODE_CONT
) {
2300 if (mesa_instructions
[j
].BranchTarget
== -1) {
2301 mesa_instructions
[j
].BranchTarget
= i
;
2305 /* The loop ends point at each other. */
2306 mesa_instructions
[i
].BranchTarget
= loop_stack
[loop_stack_pos
];
2307 mesa_instructions
[loop_stack
[loop_stack_pos
]].BranchTarget
= i
;
2310 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
2311 function_entry
*entry
= (function_entry
*)iter
.get();
2313 if (entry
->sig_id
== mesa_instructions
[i
].BranchTarget
) {
2314 mesa_instructions
[i
].BranchTarget
= entry
->inst
;
2326 print_program(struct prog_instruction
*mesa_instructions
,
2327 ir_instruction
**mesa_instruction_annotation
,
2328 int num_instructions
)
2330 ir_instruction
*last_ir
= NULL
;
2334 for (i
= 0; i
< num_instructions
; i
++) {
2335 struct prog_instruction
*mesa_inst
= mesa_instructions
+ i
;
2336 ir_instruction
*ir
= mesa_instruction_annotation
[i
];
2338 fprintf(stdout
, "%3d: ", i
);
2340 if (last_ir
!= ir
&& ir
) {
2343 for (j
= 0; j
< indent
; j
++) {
2344 fprintf(stdout
, " ");
2350 fprintf(stdout
, " "); /* line number spacing. */
2353 indent
= _mesa_fprint_instruction_opt(stdout
, mesa_inst
, indent
,
2354 PROG_PRINT_DEBUG
, NULL
);
2360 * Count resources used by the given gpu program (number of texture
2364 count_resources(struct gl_program
*prog
)
2368 prog
->SamplersUsed
= 0;
2370 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
2371 struct prog_instruction
*inst
= &prog
->Instructions
[i
];
2373 if (_mesa_is_tex_instruction(inst
->Opcode
)) {
2374 prog
->SamplerTargets
[inst
->TexSrcUnit
] =
2375 (gl_texture_index
)inst
->TexSrcTarget
;
2376 prog
->SamplersUsed
|= 1 << inst
->TexSrcUnit
;
2377 if (inst
->TexShadow
) {
2378 prog
->ShadowSamplers
|= 1 << inst
->TexSrcUnit
;
2383 _mesa_update_shader_textures_used(prog
);
2388 * Check if the given vertex/fragment/shader program is within the
2389 * resource limits of the context (number of texture units, etc).
2390 * If any of those checks fail, record a linker error.
2392 * XXX more checks are needed...
2395 check_resources(const struct gl_context
*ctx
,
2396 struct gl_shader_program
*shader_program
,
2397 struct gl_program
*prog
)
2399 switch (prog
->Target
) {
2400 case GL_VERTEX_PROGRAM_ARB
:
2401 if (_mesa_bitcount(prog
->SamplersUsed
) >
2402 ctx
->Const
.MaxVertexTextureImageUnits
) {
2403 fail_link(shader_program
, "Too many vertex shader texture samplers");
2405 if (prog
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2406 fail_link(shader_program
, "Too many vertex shader constants");
2409 case MESA_GEOMETRY_PROGRAM
:
2410 if (_mesa_bitcount(prog
->SamplersUsed
) >
2411 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2412 fail_link(shader_program
, "Too many geometry shader texture samplers");
2414 if (prog
->Parameters
->NumParameters
>
2415 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2416 fail_link(shader_program
, "Too many geometry shader constants");
2419 case GL_FRAGMENT_PROGRAM_ARB
:
2420 if (_mesa_bitcount(prog
->SamplersUsed
) >
2421 ctx
->Const
.MaxTextureImageUnits
) {
2422 fail_link(shader_program
, "Too many fragment shader texture samplers");
2424 if (prog
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2425 fail_link(shader_program
, "Too many fragment shader constants");
2429 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2435 struct uniform_sort
{
2436 struct gl_uniform
*u
;
2440 /* The shader_program->Uniforms list is almost sorted in increasing
2441 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2442 * uniforms shared between targets. We need to add parameters in
2443 * increasing order for the targets.
2446 sort_uniforms(const void *a
, const void *b
)
2448 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2449 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2451 return u1
->pos
- u2
->pos
;
2454 /* Add the uniforms to the parameters. The linker chose locations
2455 * in our parameters lists (which weren't created yet), which the
2456 * uniforms code will use to poke values into our parameters list
2457 * when uniforms are updated.
2460 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2461 struct gl_shader
*shader
,
2462 struct gl_program
*prog
)
2465 unsigned int next_sampler
= 0, num_uniforms
= 0;
2466 struct uniform_sort
*sorted_uniforms
;
2468 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2469 shader_program
->Uniforms
->NumUniforms
);
2471 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2472 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2473 int parameter_index
= -1;
2475 switch (shader
->Type
) {
2476 case GL_VERTEX_SHADER
:
2477 parameter_index
= uniform
->VertPos
;
2479 case GL_FRAGMENT_SHADER
:
2480 parameter_index
= uniform
->FragPos
;
2482 case GL_GEOMETRY_SHADER
:
2483 parameter_index
= uniform
->GeomPos
;
2487 /* Only add uniforms used in our target. */
2488 if (parameter_index
!= -1) {
2489 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2490 sorted_uniforms
[num_uniforms
].u
= uniform
;
2495 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2498 for (i
= 0; i
< num_uniforms
; i
++) {
2499 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2500 int parameter_index
= sorted_uniforms
[i
].pos
;
2501 const glsl_type
*type
= uniform
->Type
;
2504 if (type
->is_vector() ||
2505 type
->is_scalar()) {
2506 size
= type
->vector_elements
;
2508 size
= type_size(type
) * 4;
2511 gl_register_file file
;
2512 if (type
->is_sampler() ||
2513 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2514 file
= PROGRAM_SAMPLER
;
2516 file
= PROGRAM_UNIFORM
;
2519 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2523 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2524 uniform
->Name
, size
, type
->gl_type
,
2527 /* Sampler uniform values are stored in prog->SamplerUnits,
2528 * and the entry in that array is selected by this index we
2529 * store in ParameterValues[].
2531 if (file
== PROGRAM_SAMPLER
) {
2532 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2533 prog
->Parameters
->ParameterValues
[index
+ j
][0] = next_sampler
++;
2536 /* The location chosen in the Parameters list here (returned
2537 * from _mesa_add_uniform) has to match what the linker chose.
2539 if (index
!= parameter_index
) {
2540 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2541 "failed (%d vs %d)\n",
2542 uniform
->Name
, index
, parameter_index
);
2547 ralloc_free(sorted_uniforms
);
2551 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2552 struct gl_shader_program
*shader_program
,
2553 const char *name
, const glsl_type
*type
,
2556 if (type
->is_record()) {
2557 ir_constant
*field_constant
;
2559 field_constant
= (ir_constant
*)val
->components
.get_head();
2561 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2562 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2563 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2564 type
->fields
.structure
[i
].name
);
2565 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2566 field_type
, field_constant
);
2567 field_constant
= (ir_constant
*)field_constant
->next
;
2572 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2575 fail_link(shader_program
,
2576 "Couldn't find uniform for initializer %s\n", name
);
2580 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2581 ir_constant
*element
;
2582 const glsl_type
*element_type
;
2583 if (type
->is_array()) {
2584 element
= val
->array_elements
[i
];
2585 element_type
= type
->fields
.array
;
2588 element_type
= type
;
2593 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2594 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2595 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2596 conv
[j
] = element
->value
.b
[j
];
2598 values
= (void *)conv
;
2599 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2600 element_type
->vector_elements
,
2603 values
= &element
->value
;
2606 if (element_type
->is_matrix()) {
2607 _mesa_uniform_matrix(ctx
, shader_program
,
2608 element_type
->matrix_columns
,
2609 element_type
->vector_elements
,
2610 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2611 loc
+= element_type
->matrix_columns
;
2613 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2614 values
, element_type
->gl_type
);
2615 loc
+= type_size(element_type
);
2621 set_uniform_initializers(struct gl_context
*ctx
,
2622 struct gl_shader_program
*shader_program
)
2624 void *mem_ctx
= NULL
;
2626 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2627 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2632 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2633 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2634 ir_variable
*var
= ir
->as_variable();
2636 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2640 mem_ctx
= ralloc_context(NULL
);
2642 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2643 var
->type
, var
->constant_value
);
2647 ralloc_free(mem_ctx
);
2651 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
2652 * channels for copy propagation and updates following instructions to
2653 * use the original versions.
2655 * The ir_to_mesa_visitor lazily produces code assuming that this pass
2656 * will occur. As an example, a TXP production before this pass:
2658 * 0: MOV TEMP[1], INPUT[4].xyyy;
2659 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2660 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
2664 * 0: MOV TEMP[1], INPUT[4].xyyy;
2665 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2666 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
2668 * which allows for dead code elimination on TEMP[1]'s writes.
2671 ir_to_mesa_visitor::copy_propagate(void)
2673 ir_to_mesa_instruction
**acp
= rzalloc_array(mem_ctx
,
2674 ir_to_mesa_instruction
*,
2675 this->next_temp
* 4);
2676 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
2679 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2680 ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
2682 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
2683 || inst
->dst
.index
< this->next_temp
);
2685 /* First, do any copy propagation possible into the src regs. */
2686 for (int r
= 0; r
< 3; r
++) {
2687 ir_to_mesa_instruction
*first
= NULL
;
2689 int acp_base
= inst
->src
[r
].index
* 4;
2691 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
2692 inst
->src
[r
].reladdr
)
2695 /* See if we can find entries in the ACP consisting of MOVs
2696 * from the same src register for all the swizzled channels
2697 * of this src register reference.
2699 for (int i
= 0; i
< 4; i
++) {
2700 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2701 ir_to_mesa_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
2708 assert(acp_level
[acp_base
+ src_chan
] <= level
);
2713 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
2714 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
2722 /* We've now validated that we can copy-propagate to
2723 * replace this src register reference. Do it.
2725 inst
->src
[r
].file
= first
->src
[0].file
;
2726 inst
->src
[r
].index
= first
->src
[0].index
;
2729 for (int i
= 0; i
< 4; i
++) {
2730 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2731 ir_to_mesa_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
2732 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
2735 inst
->src
[r
].swizzle
= swizzle
;
2740 case OPCODE_BGNLOOP
:
2741 case OPCODE_ENDLOOP
:
2742 /* End of a basic block, clear the ACP entirely. */
2743 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
2752 /* Clear all channels written inside the block from the ACP, but
2753 * leaving those that were not touched.
2755 for (int r
= 0; r
< this->next_temp
; r
++) {
2756 for (int c
= 0; c
< 4; c
++) {
2757 if (!acp
[4 * r
+ c
])
2760 if (acp_level
[4 * r
+ c
] >= level
)
2761 acp
[4 * r
+ c
] = NULL
;
2764 if (inst
->op
== OPCODE_ENDIF
)
2769 /* Continuing the block, clear any written channels from
2772 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
2773 /* Any temporary might be written, so no copy propagation
2774 * across this instruction.
2776 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
2777 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
2778 inst
->dst
.reladdr
) {
2779 /* Any output might be written, so no copy propagation
2780 * from outputs across this instruction.
2782 for (int r
= 0; r
< this->next_temp
; r
++) {
2783 for (int c
= 0; c
< 4; c
++) {
2784 if (!acp
[4 * r
+ c
])
2787 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
2788 acp
[4 * r
+ c
] = NULL
;
2791 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
2792 inst
->dst
.file
== PROGRAM_OUTPUT
) {
2793 /* Clear where it's used as dst. */
2794 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
2795 for (int c
= 0; c
< 4; c
++) {
2796 if (inst
->dst
.writemask
& (1 << c
)) {
2797 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
2802 /* Clear where it's used as src. */
2803 for (int r
= 0; r
< this->next_temp
; r
++) {
2804 for (int c
= 0; c
< 4; c
++) {
2805 if (!acp
[4 * r
+ c
])
2808 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
2810 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
2811 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
2812 inst
->dst
.writemask
& (1 << src_chan
))
2814 acp
[4 * r
+ c
] = NULL
;
2822 /* If this is a copy, add it to the ACP. */
2823 if (inst
->op
== OPCODE_MOV
&&
2824 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
2825 !inst
->dst
.reladdr
&&
2827 !inst
->src
[0].reladdr
&&
2828 !inst
->src
[0].negate
) {
2829 for (int i
= 0; i
< 4; i
++) {
2830 if (inst
->dst
.writemask
& (1 << i
)) {
2831 acp
[4 * inst
->dst
.index
+ i
] = inst
;
2832 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
2838 ralloc_free(acp_level
);
2844 * Convert a shader's GLSL IR into a Mesa gl_program.
2846 static struct gl_program
*
2847 get_mesa_program(struct gl_context
*ctx
,
2848 struct gl_shader_program
*shader_program
,
2849 struct gl_shader
*shader
)
2851 ir_to_mesa_visitor v
;
2852 struct prog_instruction
*mesa_instructions
, *mesa_inst
;
2853 ir_instruction
**mesa_instruction_annotation
;
2855 struct gl_program
*prog
;
2857 const char *target_string
;
2859 struct gl_shader_compiler_options
*options
=
2860 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
2862 switch (shader
->Type
) {
2863 case GL_VERTEX_SHADER
:
2864 target
= GL_VERTEX_PROGRAM_ARB
;
2865 target_string
= "vertex";
2867 case GL_FRAGMENT_SHADER
:
2868 target
= GL_FRAGMENT_PROGRAM_ARB
;
2869 target_string
= "fragment";
2871 case GL_GEOMETRY_SHADER
:
2872 target
= GL_GEOMETRY_PROGRAM_NV
;
2873 target_string
= "geometry";
2876 assert(!"should not be reached");
2880 validate_ir_tree(shader
->ir
);
2882 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
2885 prog
->Parameters
= _mesa_new_parameter_list();
2886 prog
->Varying
= _mesa_new_parameter_list();
2887 prog
->Attributes
= _mesa_new_parameter_list();
2890 v
.shader_program
= shader_program
;
2891 v
.options
= options
;
2893 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
2895 /* Emit Mesa IR for main(). */
2896 visit_exec_list(shader
->ir
, &v
);
2897 v
.emit(NULL
, OPCODE_END
);
2899 /* Now emit bodies for any functions that were used. */
2901 progress
= GL_FALSE
;
2903 foreach_iter(exec_list_iterator
, iter
, v
.function_signatures
) {
2904 function_entry
*entry
= (function_entry
*)iter
.get();
2906 if (!entry
->bgn_inst
) {
2907 v
.current_function
= entry
;
2909 entry
->bgn_inst
= v
.emit(NULL
, OPCODE_BGNSUB
);
2910 entry
->bgn_inst
->function
= entry
;
2912 visit_exec_list(&entry
->sig
->body
, &v
);
2914 ir_to_mesa_instruction
*last
;
2915 last
= (ir_to_mesa_instruction
*)v
.instructions
.get_tail();
2916 if (last
->op
!= OPCODE_RET
)
2917 v
.emit(NULL
, OPCODE_RET
);
2919 ir_to_mesa_instruction
*end
;
2920 end
= v
.emit(NULL
, OPCODE_ENDSUB
);
2921 end
->function
= entry
;
2928 prog
->NumTemporaries
= v
.next_temp
;
2930 int num_instructions
= 0;
2931 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2936 (struct prog_instruction
*)calloc(num_instructions
,
2937 sizeof(*mesa_instructions
));
2938 mesa_instruction_annotation
= ralloc_array(v
.mem_ctx
, ir_instruction
*,
2943 /* Convert ir_mesa_instructions into prog_instructions.
2945 mesa_inst
= mesa_instructions
;
2947 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
2948 const ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
2950 mesa_inst
->Opcode
= inst
->op
;
2951 mesa_inst
->CondUpdate
= inst
->cond_update
;
2953 mesa_inst
->SaturateMode
= SATURATE_ZERO_ONE
;
2954 mesa_inst
->DstReg
.File
= inst
->dst
.file
;
2955 mesa_inst
->DstReg
.Index
= inst
->dst
.index
;
2956 mesa_inst
->DstReg
.CondMask
= inst
->dst
.cond_mask
;
2957 mesa_inst
->DstReg
.WriteMask
= inst
->dst
.writemask
;
2958 mesa_inst
->DstReg
.RelAddr
= inst
->dst
.reladdr
!= NULL
;
2959 mesa_inst
->SrcReg
[0] = mesa_src_reg_from_ir_src_reg(inst
->src
[0]);
2960 mesa_inst
->SrcReg
[1] = mesa_src_reg_from_ir_src_reg(inst
->src
[1]);
2961 mesa_inst
->SrcReg
[2] = mesa_src_reg_from_ir_src_reg(inst
->src
[2]);
2962 mesa_inst
->TexSrcUnit
= inst
->sampler
;
2963 mesa_inst
->TexSrcTarget
= inst
->tex_target
;
2964 mesa_inst
->TexShadow
= inst
->tex_shadow
;
2965 mesa_instruction_annotation
[i
] = inst
->ir
;
2967 /* Set IndirectRegisterFiles. */
2968 if (mesa_inst
->DstReg
.RelAddr
)
2969 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->DstReg
.File
;
2971 /* Update program's bitmask of indirectly accessed register files */
2972 for (unsigned src
= 0; src
< 3; src
++)
2973 if (mesa_inst
->SrcReg
[src
].RelAddr
)
2974 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->SrcReg
[src
].File
;
2976 if (options
->EmitNoIfs
&& mesa_inst
->Opcode
== OPCODE_IF
) {
2977 fail_link(shader_program
, "Couldn't flatten if statement\n");
2980 switch (mesa_inst
->Opcode
) {
2982 inst
->function
->inst
= i
;
2983 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2986 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
2989 mesa_inst
->BranchTarget
= inst
->function
->sig_id
; /* rewritten later */
2992 prog
->NumAddressRegs
= 1;
3001 if (!shader_program
->LinkStatus
)
3005 if (!shader_program
->LinkStatus
) {
3006 free(mesa_instructions
);
3007 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
3011 set_branchtargets(&v
, mesa_instructions
, num_instructions
);
3013 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3015 printf("GLSL IR for linked %s program %d:\n", target_string
,
3016 shader_program
->Name
);
3017 _mesa_print_ir(shader
->ir
, NULL
);
3020 printf("Mesa IR for linked %s program %d:\n", target_string
,
3021 shader_program
->Name
);
3022 print_program(mesa_instructions
, mesa_instruction_annotation
,
3026 prog
->Instructions
= mesa_instructions
;
3027 prog
->NumInstructions
= num_instructions
;
3029 do_set_program_inouts(shader
->ir
, prog
);
3030 count_resources(prog
);
3032 check_resources(ctx
, shader_program
, prog
);
3034 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
3036 if ((ctx
->Shader
.Flags
& GLSL_NO_OPT
) == 0) {
3037 _mesa_optimize_program(ctx
, prog
);
3047 * Called via ctx->Driver.LinkShader()
3048 * This actually involves converting GLSL IR into Mesa gl_programs with
3049 * code lowering and other optimizations.
3052 _mesa_ir_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
3054 assert(prog
->LinkStatus
);
3056 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
3057 if (prog
->_LinkedShaders
[i
] == NULL
)
3061 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
3062 const struct gl_shader_compiler_options
*options
=
3063 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
3069 do_mat_op_to_vec(ir
);
3070 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
3072 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
3074 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
3076 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
3078 progress
= lower_quadop_vector(ir
, true) || progress
;
3080 if (options
->EmitNoIfs
) {
3081 progress
= lower_discard(ir
) || progress
;
3082 progress
= lower_if_to_cond_assign(ir
) || progress
;
3085 if (options
->EmitNoNoise
)
3086 progress
= lower_noise(ir
) || progress
;
3088 /* If there are forms of indirect addressing that the driver
3089 * cannot handle, perform the lowering pass.
3091 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
3092 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
3094 lower_variable_index_to_cond_assign(ir
,
3095 options
->EmitNoIndirectInput
,
3096 options
->EmitNoIndirectOutput
,
3097 options
->EmitNoIndirectTemp
,
3098 options
->EmitNoIndirectUniform
)
3101 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
3104 validate_ir_tree(ir
);
3107 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
3108 struct gl_program
*linked_prog
;
3110 if (prog
->_LinkedShaders
[i
] == NULL
)
3113 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
3118 switch (prog
->_LinkedShaders
[i
]->Type
) {
3119 case GL_VERTEX_SHADER
:
3120 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
3121 (struct gl_vertex_program
*)linked_prog
);
3122 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
3125 case GL_FRAGMENT_SHADER
:
3126 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
3127 (struct gl_fragment_program
*)linked_prog
);
3128 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
3131 case GL_GEOMETRY_SHADER
:
3132 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
3133 (struct gl_geometry_program
*)linked_prog
);
3134 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
3143 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
3151 * Compile a GLSL shader. Called via glCompileShader().
3154 _mesa_glsl_compile_shader(struct gl_context
*ctx
, struct gl_shader
*shader
)
3156 struct _mesa_glsl_parse_state
*state
=
3157 new(shader
) _mesa_glsl_parse_state(ctx
, shader
->Type
, shader
);
3159 const char *source
= shader
->Source
;
3160 /* Check if the user called glCompileShader without first calling
3161 * glShaderSource. This should fail to compile, but not raise a GL_ERROR.
3163 if (source
== NULL
) {
3164 shader
->CompileStatus
= GL_FALSE
;
3168 state
->error
= preprocess(state
, &source
, &state
->info_log
,
3169 &ctx
->Extensions
, ctx
->API
);
3171 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3172 printf("GLSL source for shader %d:\n", shader
->Name
);
3173 printf("%s\n", shader
->Source
);
3176 if (!state
->error
) {
3177 _mesa_glsl_lexer_ctor(state
, source
);
3178 _mesa_glsl_parse(state
);
3179 _mesa_glsl_lexer_dtor(state
);
3182 ralloc_free(shader
->ir
);
3183 shader
->ir
= new(shader
) exec_list
;
3184 if (!state
->error
&& !state
->translation_unit
.is_empty())
3185 _mesa_ast_to_hir(shader
->ir
, state
);
3187 if (!state
->error
&& !shader
->ir
->is_empty()) {
3188 validate_ir_tree(shader
->ir
);
3190 /* Do some optimization at compile time to reduce shader IR size
3191 * and reduce later work if the same shader is linked multiple times
3193 while (do_common_optimization(shader
->ir
, false, 32))
3196 validate_ir_tree(shader
->ir
);
3199 shader
->symbols
= state
->symbols
;
3201 shader
->CompileStatus
= !state
->error
;
3202 shader
->InfoLog
= state
->info_log
;
3203 shader
->Version
= state
->language_version
;
3204 memcpy(shader
->builtins_to_link
, state
->builtins_to_link
,
3205 sizeof(shader
->builtins_to_link
[0]) * state
->num_builtins_to_link
);
3206 shader
->num_builtins_to_link
= state
->num_builtins_to_link
;
3208 if (ctx
->Shader
.Flags
& GLSL_LOG
) {
3209 _mesa_write_shader_to_file(shader
);
3212 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3213 if (shader
->CompileStatus
) {
3214 printf("GLSL IR for shader %d:\n", shader
->Name
);
3215 _mesa_print_ir(shader
->ir
, NULL
);
3218 printf("GLSL shader %d failed to compile.\n", shader
->Name
);
3220 if (shader
->InfoLog
&& shader
->InfoLog
[0] != 0) {
3221 printf("GLSL shader %d info log:\n", shader
->Name
);
3222 printf("%s\n", shader
->InfoLog
);
3226 /* Retain any live IR, but trash the rest. */
3227 reparent_ir(shader
->ir
, shader
->ir
);
3234 * Link a GLSL shader program. Called via glLinkProgram().
3237 _mesa_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
3241 _mesa_clear_shader_program_data(ctx
, prog
);
3243 prog
->LinkStatus
= GL_TRUE
;
3245 for (i
= 0; i
< prog
->NumShaders
; i
++) {
3246 if (!prog
->Shaders
[i
]->CompileStatus
) {
3247 fail_link(prog
, "linking with uncompiled shader");
3248 prog
->LinkStatus
= GL_FALSE
;
3252 prog
->Varying
= _mesa_new_parameter_list();
3253 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
3254 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
3255 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
3257 if (prog
->LinkStatus
) {
3258 link_shaders(ctx
, prog
);
3261 if (prog
->LinkStatus
) {
3262 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
3263 prog
->LinkStatus
= GL_FALSE
;
3267 set_uniform_initializers(ctx
, prog
);
3269 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3270 if (!prog
->LinkStatus
) {
3271 printf("GLSL shader program %d failed to link\n", prog
->Name
);
3274 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
3275 printf("GLSL shader program %d info log:\n", prog
->Name
);
3276 printf("%s\n", prog
->InfoLog
);