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(gl_register_file file
, int index
, const glsl_type
*type
)
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(gl_register_file 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 gl_register_file 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
= reg
.file
;
134 this->index
= reg
.index
;
135 this->swizzle
= SWIZZLE_XYZW
;
137 this->reladdr
= reg
.reladdr
;
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 class ir_to_mesa_instruction
: public exec_node
{
151 /* Callers of this ralloc-based new need not call delete. It's
152 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
153 static void* operator new(size_t size
, void *ctx
)
157 node
= rzalloc_size(ctx
, size
);
158 assert(node
!= NULL
);
166 /** Pointer to the ir source this tree came from for debugging */
168 GLboolean cond_update
;
170 int sampler
; /**< sampler index */
171 int tex_target
; /**< One of TEXTURE_*_INDEX */
172 GLboolean tex_shadow
;
174 class function_entry
*function
; /* Set on OPCODE_CAL or OPCODE_BGNSUB */
177 class variable_storage
: public exec_node
{
179 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
180 : file(file
), index(index
), var(var
)
185 gl_register_file file
;
187 ir_variable
*var
; /* variable that maps to this, if any */
190 class function_entry
: public exec_node
{
192 ir_function_signature
*sig
;
195 * identifier of this function signature used by the program.
197 * At the point that Mesa instructions for function calls are
198 * generated, we don't know the address of the first instruction of
199 * the function body. So we make the BranchTarget that is called a
200 * small integer and rewrite them during set_branchtargets().
205 * Pointer to first instruction of the function body.
207 * Set during function body emits after main() is processed.
209 ir_to_mesa_instruction
*bgn_inst
;
212 * Index of the first instruction of the function body in actual
215 * Set after convertion from ir_to_mesa_instruction to prog_instruction.
219 /** Storage for the return value. */
223 class ir_to_mesa_visitor
: public ir_visitor
{
225 ir_to_mesa_visitor();
226 ~ir_to_mesa_visitor();
228 function_entry
*current_function
;
230 struct gl_context
*ctx
;
231 struct gl_program
*prog
;
232 struct gl_shader_program
*shader_program
;
233 struct gl_shader_compiler_options
*options
;
237 variable_storage
*find_variable_storage(ir_variable
*var
);
239 function_entry
*get_function_signature(ir_function_signature
*sig
);
241 src_reg
get_temp(const glsl_type
*type
);
242 void reladdr_to_temp(ir_instruction
*ir
, src_reg
*reg
, int *num_reladdr
);
244 src_reg
src_reg_for_float(float val
);
247 * \name Visit methods
249 * As typical for the visitor pattern, there must be one \c visit method for
250 * each concrete subclass of \c ir_instruction. Virtual base classes within
251 * the hierarchy should not have \c visit methods.
254 virtual void visit(ir_variable
*);
255 virtual void visit(ir_loop
*);
256 virtual void visit(ir_loop_jump
*);
257 virtual void visit(ir_function_signature
*);
258 virtual void visit(ir_function
*);
259 virtual void visit(ir_expression
*);
260 virtual void visit(ir_swizzle
*);
261 virtual void visit(ir_dereference_variable
*);
262 virtual void visit(ir_dereference_array
*);
263 virtual void visit(ir_dereference_record
*);
264 virtual void visit(ir_assignment
*);
265 virtual void visit(ir_constant
*);
266 virtual void visit(ir_call
*);
267 virtual void visit(ir_return
*);
268 virtual void visit(ir_discard
*);
269 virtual void visit(ir_texture
*);
270 virtual void visit(ir_if
*);
275 /** List of variable_storage */
278 /** List of function_entry */
279 exec_list function_signatures
;
280 int next_signature_id
;
282 /** List of ir_to_mesa_instruction */
283 exec_list instructions
;
285 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
);
287 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
288 dst_reg dst
, src_reg src0
);
290 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
291 dst_reg dst
, src_reg src0
, src_reg src1
);
293 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
295 src_reg src0
, src_reg src1
, src_reg src2
);
298 * Emit the correct dot-product instruction for the type of arguments
300 ir_to_mesa_instruction
* emit_dp(ir_instruction
*ir
,
306 void emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
307 dst_reg dst
, src_reg src0
);
309 void emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
310 dst_reg dst
, src_reg src0
, src_reg src1
);
312 void emit_scs(ir_instruction
*ir
, enum prog_opcode op
,
313 dst_reg dst
, const src_reg
&src
);
315 GLboolean
try_emit_mad(ir_expression
*ir
,
317 GLboolean
try_emit_sat(ir_expression
*ir
);
319 void emit_swz(ir_expression
*ir
);
321 bool process_move_condition(ir_rvalue
*ir
);
323 void copy_propagate(void);
328 src_reg undef_src
= src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
330 dst_reg undef_dst
= dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
);
332 dst_reg address_reg
= dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
);
335 swizzle_for_size(int size
)
337 int size_swizzles
[4] = {
338 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
339 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
340 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
341 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
344 assert((size
>= 1) && (size
<= 4));
345 return size_swizzles
[size
- 1];
348 ir_to_mesa_instruction
*
349 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
351 src_reg src0
, src_reg src1
, src_reg src2
)
353 ir_to_mesa_instruction
*inst
= new(mem_ctx
) ir_to_mesa_instruction();
356 /* If we have to do relative addressing, we want to load the ARL
357 * reg directly for one of the regs, and preload the other reladdr
358 * sources into temps.
360 num_reladdr
+= dst
.reladdr
!= NULL
;
361 num_reladdr
+= src0
.reladdr
!= NULL
;
362 num_reladdr
+= src1
.reladdr
!= NULL
;
363 num_reladdr
+= src2
.reladdr
!= NULL
;
365 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
366 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
367 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
370 emit(ir
, OPCODE_ARL
, address_reg
, *dst
.reladdr
);
373 assert(num_reladdr
== 0);
382 inst
->function
= NULL
;
384 this->instructions
.push_tail(inst
);
390 ir_to_mesa_instruction
*
391 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
392 dst_reg dst
, src_reg src0
, src_reg src1
)
394 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
397 ir_to_mesa_instruction
*
398 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
399 dst_reg dst
, src_reg src0
)
401 assert(dst
.writemask
!= 0);
402 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
405 ir_to_mesa_instruction
*
406 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
)
408 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
411 ir_to_mesa_instruction
*
412 ir_to_mesa_visitor::emit_dp(ir_instruction
*ir
,
413 dst_reg dst
, src_reg src0
, src_reg src1
,
416 static const gl_inst_opcode dot_opcodes
[] = {
417 OPCODE_DP2
, OPCODE_DP3
, OPCODE_DP4
420 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
424 * Emits Mesa scalar opcodes to produce unique answers across channels.
426 * Some Mesa opcodes are scalar-only, like ARB_fp/vp. The src X
427 * channel determines the result across all channels. So to do a vec4
428 * of this operation, we want to emit a scalar per source channel used
429 * to produce dest channels.
432 ir_to_mesa_visitor::emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
434 src_reg orig_src0
, src_reg orig_src1
)
437 int done_mask
= ~dst
.writemask
;
439 /* Mesa RCP is a scalar operation splatting results to all channels,
440 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
443 for (i
= 0; i
< 4; i
++) {
444 GLuint this_mask
= (1 << i
);
445 ir_to_mesa_instruction
*inst
;
446 src_reg src0
= orig_src0
;
447 src_reg src1
= orig_src1
;
449 if (done_mask
& this_mask
)
452 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
453 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
454 for (j
= i
+ 1; j
< 4; j
++) {
455 /* If there is another enabled component in the destination that is
456 * derived from the same inputs, generate its value on this pass as
459 if (!(done_mask
& (1 << j
)) &&
460 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
461 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
462 this_mask
|= (1 << j
);
465 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
466 src0_swiz
, src0_swiz
);
467 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
468 src1_swiz
, src1_swiz
);
470 inst
= emit(ir
, op
, dst
, src0
, src1
);
471 inst
->dst
.writemask
= this_mask
;
472 done_mask
|= this_mask
;
477 ir_to_mesa_visitor::emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
478 dst_reg dst
, src_reg src0
)
480 src_reg undef
= undef_src
;
482 undef
.swizzle
= SWIZZLE_XXXX
;
484 emit_scalar(ir
, op
, dst
, src0
, undef
);
488 * Emit an OPCODE_SCS instruction
490 * The \c SCS opcode functions a bit differently than the other Mesa (or
491 * ARB_fragment_program) opcodes. Instead of splatting its result across all
492 * four components of the destination, it writes one value to the \c x
493 * component and another value to the \c y component.
495 * \param ir IR instruction being processed
496 * \param op Either \c OPCODE_SIN or \c OPCODE_COS depending on which
498 * \param dst Destination register
499 * \param src Source register
502 ir_to_mesa_visitor::emit_scs(ir_instruction
*ir
, enum prog_opcode op
,
506 /* Vertex programs cannot use the SCS opcode.
508 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
509 emit_scalar(ir
, op
, dst
, src
);
513 const unsigned component
= (op
== OPCODE_SIN
) ? 0 : 1;
514 const unsigned scs_mask
= (1U << component
);
515 int done_mask
= ~dst
.writemask
;
518 assert(op
== OPCODE_SIN
|| op
== OPCODE_COS
);
520 /* If there are compnents in the destination that differ from the component
521 * that will be written by the SCS instrution, we'll need a temporary.
523 if (scs_mask
!= unsigned(dst
.writemask
)) {
524 tmp
= get_temp(glsl_type::vec4_type
);
527 for (unsigned i
= 0; i
< 4; i
++) {
528 unsigned this_mask
= (1U << i
);
531 if ((done_mask
& this_mask
) != 0)
534 /* The source swizzle specified which component of the source generates
535 * sine / cosine for the current component in the destination. The SCS
536 * instruction requires that this value be swizzle to the X component.
537 * Replace the current swizzle with a swizzle that puts the source in
540 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
542 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
543 src0_swiz
, src0_swiz
);
544 for (unsigned j
= i
+ 1; j
< 4; j
++) {
545 /* If there is another enabled component in the destination that is
546 * derived from the same inputs, generate its value on this pass as
549 if (!(done_mask
& (1 << j
)) &&
550 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
551 this_mask
|= (1 << j
);
555 if (this_mask
!= scs_mask
) {
556 ir_to_mesa_instruction
*inst
;
557 dst_reg tmp_dst
= dst_reg(tmp
);
559 /* Emit the SCS instruction.
561 inst
= emit(ir
, OPCODE_SCS
, tmp_dst
, src0
);
562 inst
->dst
.writemask
= scs_mask
;
564 /* Move the result of the SCS instruction to the desired location in
567 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
568 component
, component
);
569 inst
= emit(ir
, OPCODE_SCS
, dst
, tmp
);
570 inst
->dst
.writemask
= this_mask
;
572 /* Emit the SCS instruction to write directly to the destination.
574 ir_to_mesa_instruction
*inst
= emit(ir
, OPCODE_SCS
, dst
, src0
);
575 inst
->dst
.writemask
= scs_mask
;
578 done_mask
|= this_mask
;
583 ir_to_mesa_visitor::src_reg_for_float(float val
)
585 src_reg
src(PROGRAM_CONSTANT
, -1, NULL
);
587 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
588 (const gl_constant_value
*)&val
, 1, &src
.swizzle
);
594 type_size(const struct glsl_type
*type
)
599 switch (type
->base_type
) {
602 case GLSL_TYPE_FLOAT
:
604 if (type
->is_matrix()) {
605 return type
->matrix_columns
;
607 /* Regardless of size of vector, it gets a vec4. This is bad
608 * packing for things like floats, but otherwise arrays become a
609 * mess. Hopefully a later pass over the code can pack scalars
610 * down if appropriate.
614 case GLSL_TYPE_ARRAY
:
615 assert(type
->length
> 0);
616 return type_size(type
->fields
.array
) * type
->length
;
617 case GLSL_TYPE_STRUCT
:
619 for (i
= 0; i
< type
->length
; i
++) {
620 size
+= type_size(type
->fields
.structure
[i
].type
);
623 case GLSL_TYPE_SAMPLER
:
624 /* Samplers take up one slot in UNIFORMS[], but they're baked in
635 * In the initial pass of codegen, we assign temporary numbers to
636 * intermediate results. (not SSA -- variable assignments will reuse
637 * storage). Actual register allocation for the Mesa VM occurs in a
638 * pass over the Mesa IR later.
641 ir_to_mesa_visitor::get_temp(const glsl_type
*type
)
645 src
.file
= PROGRAM_TEMPORARY
;
646 src
.index
= next_temp
;
648 next_temp
+= type_size(type
);
650 if (type
->is_array() || type
->is_record()) {
651 src
.swizzle
= SWIZZLE_NOOP
;
653 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
661 ir_to_mesa_visitor::find_variable_storage(ir_variable
*var
)
664 variable_storage
*entry
;
666 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
667 entry
= (variable_storage
*)iter
.get();
669 if (entry
->var
== var
)
677 ir_to_mesa_visitor::visit(ir_variable
*ir
)
679 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
680 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
682 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
683 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
685 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
686 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
687 switch (ir
->depth_layout
) {
688 case ir_depth_layout_none
:
689 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
691 case ir_depth_layout_any
:
692 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
694 case ir_depth_layout_greater
:
695 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
697 case ir_depth_layout_less
:
698 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
700 case ir_depth_layout_unchanged
:
701 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
709 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
711 const ir_state_slot
*const slots
= ir
->state_slots
;
712 assert(ir
->state_slots
!= NULL
);
714 /* Check if this statevar's setup in the STATE file exactly
715 * matches how we'll want to reference it as a
716 * struct/array/whatever. If not, then we need to move it into
717 * temporary storage and hope that it'll get copy-propagated
720 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
721 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
726 struct variable_storage
*storage
;
728 if (i
== ir
->num_state_slots
) {
729 /* We'll set the index later. */
730 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
731 this->variables
.push_tail(storage
);
735 /* The variable_storage constructor allocates slots based on the size
736 * of the type. However, this had better match the number of state
737 * elements that we're going to copy into the new temporary.
739 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
741 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
743 this->variables
.push_tail(storage
);
744 this->next_temp
+= type_size(ir
->type
);
746 dst
= dst_reg(src_reg(PROGRAM_TEMPORARY
, storage
->index
, NULL
));
750 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
751 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
752 (gl_state_index
*)slots
[i
].tokens
);
754 if (storage
->file
== PROGRAM_STATE_VAR
) {
755 if (storage
->index
== -1) {
756 storage
->index
= index
;
758 assert(index
== storage
->index
+ (int)i
);
761 src_reg
src(PROGRAM_STATE_VAR
, index
, NULL
);
762 src
.swizzle
= slots
[i
].swizzle
;
763 emit(ir
, OPCODE_MOV
, dst
, src
);
764 /* even a float takes up a whole vec4 reg in a struct/array. */
769 if (storage
->file
== PROGRAM_TEMPORARY
&&
770 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
771 linker_error(this->shader_program
,
772 "failed to load builtin uniform `%s' "
773 "(%d/%d regs loaded)\n",
774 ir
->name
, dst
.index
- storage
->index
,
775 type_size(ir
->type
));
781 ir_to_mesa_visitor::visit(ir_loop
*ir
)
783 ir_dereference_variable
*counter
= NULL
;
785 if (ir
->counter
!= NULL
)
786 counter
= new(mem_ctx
) ir_dereference_variable(ir
->counter
);
788 if (ir
->from
!= NULL
) {
789 assert(ir
->counter
!= NULL
);
792 new(mem_ctx
) ir_assignment(counter
, ir
->from
, NULL
);
797 emit(NULL
, OPCODE_BGNLOOP
);
801 new(mem_ctx
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
803 ir_if
*if_stmt
= new(mem_ctx
) ir_if(e
);
806 new(mem_ctx
) ir_loop_jump(ir_loop_jump::jump_break
);
808 if_stmt
->then_instructions
.push_tail(brk
);
810 if_stmt
->accept(this);
813 visit_exec_list(&ir
->body_instructions
, this);
817 new(mem_ctx
) ir_expression(ir_binop_add
, counter
->type
,
818 counter
, ir
->increment
);
821 new(mem_ctx
) ir_assignment(counter
, e
, NULL
);
826 emit(NULL
, OPCODE_ENDLOOP
);
830 ir_to_mesa_visitor::visit(ir_loop_jump
*ir
)
833 case ir_loop_jump::jump_break
:
834 emit(NULL
, OPCODE_BRK
);
836 case ir_loop_jump::jump_continue
:
837 emit(NULL
, OPCODE_CONT
);
844 ir_to_mesa_visitor::visit(ir_function_signature
*ir
)
851 ir_to_mesa_visitor::visit(ir_function
*ir
)
853 /* Ignore function bodies other than main() -- we shouldn't see calls to
854 * them since they should all be inlined before we get to ir_to_mesa.
856 if (strcmp(ir
->name
, "main") == 0) {
857 const ir_function_signature
*sig
;
860 sig
= ir
->matching_signature(&empty
);
864 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
865 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
873 ir_to_mesa_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
875 int nonmul_operand
= 1 - mul_operand
;
878 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
879 if (!expr
|| expr
->operation
!= ir_binop_mul
)
882 expr
->operands
[0]->accept(this);
884 expr
->operands
[1]->accept(this);
886 ir
->operands
[nonmul_operand
]->accept(this);
889 this->result
= get_temp(ir
->type
);
890 emit(ir
, OPCODE_MAD
, dst_reg(this->result
), a
, b
, c
);
896 ir_to_mesa_visitor::try_emit_sat(ir_expression
*ir
)
898 /* Saturates were only introduced to vertex programs in
899 * NV_vertex_program3, so don't give them to drivers in the VP.
901 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
904 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
908 sat_src
->accept(this);
909 src_reg src
= this->result
;
911 /* If we generated an expression instruction into a temporary in
912 * processing the saturate's operand, apply the saturate to that
913 * instruction. Otherwise, generate a MOV to do the saturate.
915 * Note that we have to be careful to only do this optimization if
916 * the instruction in question was what generated src->result. For
917 * example, ir_dereference_array might generate a MUL instruction
918 * to create the reladdr, and return us a src reg using that
919 * reladdr. That MUL result is not the value we're trying to
922 ir_expression
*sat_src_expr
= sat_src
->as_expression();
923 ir_to_mesa_instruction
*new_inst
;
924 new_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
925 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
926 sat_src_expr
->operation
== ir_binop_add
||
927 sat_src_expr
->operation
== ir_binop_dot
)) {
928 new_inst
->saturate
= true;
930 this->result
= get_temp(ir
->type
);
931 ir_to_mesa_instruction
*inst
;
932 inst
= emit(ir
, OPCODE_MOV
, dst_reg(this->result
), src
);
933 inst
->saturate
= true;
940 ir_to_mesa_visitor::reladdr_to_temp(ir_instruction
*ir
,
941 src_reg
*reg
, int *num_reladdr
)
946 emit(ir
, OPCODE_ARL
, address_reg
, *reg
->reladdr
);
948 if (*num_reladdr
!= 1) {
949 src_reg temp
= get_temp(glsl_type::vec4_type
);
951 emit(ir
, OPCODE_MOV
, dst_reg(temp
), *reg
);
959 ir_to_mesa_visitor::emit_swz(ir_expression
*ir
)
961 /* Assume that the vector operator is in a form compatible with OPCODE_SWZ.
962 * This means that each of the operands is either an immediate value of -1,
963 * 0, or 1, or is a component from one source register (possibly with
966 uint8_t components
[4] = { 0 };
967 bool negate
[4] = { false };
968 ir_variable
*var
= NULL
;
970 for (unsigned i
= 0; i
< ir
->type
->vector_elements
; i
++) {
971 ir_rvalue
*op
= ir
->operands
[i
];
973 assert(op
->type
->is_scalar());
976 switch (op
->ir_type
) {
977 case ir_type_constant
: {
979 assert(op
->type
->is_scalar());
981 const ir_constant
*const c
= op
->as_constant();
983 components
[i
] = SWIZZLE_ONE
;
984 } else if (c
->is_zero()) {
985 components
[i
] = SWIZZLE_ZERO
;
986 } else if (c
->is_negative_one()) {
987 components
[i
] = SWIZZLE_ONE
;
990 assert(!"SWZ constant must be 0.0 or 1.0.");
997 case ir_type_dereference_variable
: {
998 ir_dereference_variable
*const deref
=
999 (ir_dereference_variable
*) op
;
1001 assert((var
== NULL
) || (deref
->var
== var
));
1002 components
[i
] = SWIZZLE_X
;
1008 case ir_type_expression
: {
1009 ir_expression
*const expr
= (ir_expression
*) op
;
1011 assert(expr
->operation
== ir_unop_neg
);
1014 op
= expr
->operands
[0];
1018 case ir_type_swizzle
: {
1019 ir_swizzle
*const swiz
= (ir_swizzle
*) op
;
1021 components
[i
] = swiz
->mask
.x
;
1027 assert(!"Should not get here.");
1033 assert(var
!= NULL
);
1035 ir_dereference_variable
*const deref
=
1036 new(mem_ctx
) ir_dereference_variable(var
);
1038 this->result
.file
= PROGRAM_UNDEFINED
;
1039 deref
->accept(this);
1040 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1042 printf("Failed to get tree for expression operand:\n");
1050 src
.swizzle
= MAKE_SWIZZLE4(components
[0],
1054 src
.negate
= ((unsigned(negate
[0]) << 0)
1055 | (unsigned(negate
[1]) << 1)
1056 | (unsigned(negate
[2]) << 2)
1057 | (unsigned(negate
[3]) << 3));
1059 /* Storage for our result. Ideally for an assignment we'd be using the
1060 * actual storage for the result here, instead.
1062 const src_reg result_src
= get_temp(ir
->type
);
1063 dst_reg result_dst
= dst_reg(result_src
);
1065 /* Limit writes to the channels that will be used by result_src later.
1066 * This does limit this temp's use as a temporary for multi-instruction
1069 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1071 emit(ir
, OPCODE_SWZ
, result_dst
, src
);
1072 this->result
= result_src
;
1076 ir_to_mesa_visitor::visit(ir_expression
*ir
)
1078 unsigned int operand
;
1079 src_reg op
[Elements(ir
->operands
)];
1083 /* Quick peephole: Emit OPCODE_MAD(a, b, c) instead of ADD(MUL(a, b), c)
1085 if (ir
->operation
== ir_binop_add
) {
1086 if (try_emit_mad(ir
, 1))
1088 if (try_emit_mad(ir
, 0))
1091 if (try_emit_sat(ir
))
1094 if (ir
->operation
== ir_quadop_vector
) {
1099 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1100 this->result
.file
= PROGRAM_UNDEFINED
;
1101 ir
->operands
[operand
]->accept(this);
1102 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1104 printf("Failed to get tree for expression operand:\n");
1105 ir
->operands
[operand
]->accept(&v
);
1108 op
[operand
] = this->result
;
1110 /* Matrix expression operands should have been broken down to vector
1111 * operations already.
1113 assert(!ir
->operands
[operand
]->type
->is_matrix());
1116 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1117 if (ir
->operands
[1]) {
1118 vector_elements
= MAX2(vector_elements
,
1119 ir
->operands
[1]->type
->vector_elements
);
1122 this->result
.file
= PROGRAM_UNDEFINED
;
1124 /* Storage for our result. Ideally for an assignment we'd be using
1125 * the actual storage for the result here, instead.
1127 result_src
= get_temp(ir
->type
);
1128 /* convenience for the emit functions below. */
1129 result_dst
= dst_reg(result_src
);
1130 /* Limit writes to the channels that will be used by result_src later.
1131 * This does limit this temp's use as a temporary for multi-instruction
1134 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1136 switch (ir
->operation
) {
1137 case ir_unop_logic_not
:
1138 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1139 * older GPUs implement SEQ using multiple instructions (i915 uses two
1140 * SGE instructions and a MUL instruction). Since our logic values are
1141 * 0.0 and 1.0, 1-x also implements !x.
1143 op
[0].negate
= ~op
[0].negate
;
1144 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], src_reg_for_float(1.0));
1147 op
[0].negate
= ~op
[0].negate
;
1151 emit(ir
, OPCODE_ABS
, result_dst
, op
[0]);
1154 emit(ir
, OPCODE_SSG
, result_dst
, op
[0]);
1157 emit_scalar(ir
, OPCODE_RCP
, result_dst
, op
[0]);
1161 emit_scalar(ir
, OPCODE_EX2
, result_dst
, op
[0]);
1165 assert(!"not reached: should be handled by ir_explog_to_explog2");
1168 emit_scalar(ir
, OPCODE_LG2
, result_dst
, op
[0]);
1171 emit_scalar(ir
, OPCODE_SIN
, result_dst
, op
[0]);
1174 emit_scalar(ir
, OPCODE_COS
, result_dst
, op
[0]);
1176 case ir_unop_sin_reduced
:
1177 emit_scs(ir
, OPCODE_SIN
, result_dst
, op
[0]);
1179 case ir_unop_cos_reduced
:
1180 emit_scs(ir
, OPCODE_COS
, result_dst
, op
[0]);
1184 emit(ir
, OPCODE_DDX
, result_dst
, op
[0]);
1187 emit(ir
, OPCODE_DDY
, result_dst
, op
[0]);
1190 case ir_unop_noise
: {
1191 const enum prog_opcode opcode
=
1192 prog_opcode(OPCODE_NOISE1
1193 + (ir
->operands
[0]->type
->vector_elements
) - 1);
1194 assert((opcode
>= OPCODE_NOISE1
) && (opcode
<= OPCODE_NOISE4
));
1196 emit(ir
, opcode
, result_dst
, op
[0]);
1201 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1204 emit(ir
, OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1208 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1211 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1213 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1217 emit(ir
, OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1219 case ir_binop_greater
:
1220 emit(ir
, OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1222 case ir_binop_lequal
:
1223 emit(ir
, OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1225 case ir_binop_gequal
:
1226 emit(ir
, OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1228 case ir_binop_equal
:
1229 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1231 case ir_binop_nequal
:
1232 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1234 case ir_binop_all_equal
:
1235 /* "==" operator producing a scalar boolean. */
1236 if (ir
->operands
[0]->type
->is_vector() ||
1237 ir
->operands
[1]->type
->is_vector()) {
1238 src_reg temp
= get_temp(glsl_type::vec4_type
);
1239 emit(ir
, OPCODE_SNE
, dst_reg(temp
), op
[0], op
[1]);
1240 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1241 emit(ir
, OPCODE_SEQ
, result_dst
, result_src
, src_reg_for_float(0.0));
1243 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1246 case ir_binop_any_nequal
:
1247 /* "!=" operator producing a scalar boolean. */
1248 if (ir
->operands
[0]->type
->is_vector() ||
1249 ir
->operands
[1]->type
->is_vector()) {
1250 src_reg temp
= get_temp(glsl_type::vec4_type
);
1251 emit(ir
, OPCODE_SNE
, dst_reg(temp
), op
[0], op
[1]);
1253 /* After the dot-product, the value will be an integer on the
1254 * range [0,4]. Zero stays zero, and positive values become 1.0.
1256 ir_to_mesa_instruction
*const dp
=
1257 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1258 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1259 /* The clamping to [0,1] can be done for free in the fragment
1260 * shader with a saturate.
1262 dp
->saturate
= true;
1264 /* Negating the result of the dot-product gives values on the range
1265 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1266 * achieved using SLT.
1268 src_reg slt_src
= result_src
;
1269 slt_src
.negate
= ~slt_src
.negate
;
1270 emit(ir
, OPCODE_SLT
, result_dst
, slt_src
, src_reg_for_float(0.0));
1273 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1278 assert(ir
->operands
[0]->type
->is_vector());
1280 /* After the dot-product, the value will be an integer on the
1281 * range [0,4]. Zero stays zero, and positive values become 1.0.
1283 ir_to_mesa_instruction
*const dp
=
1284 emit_dp(ir
, result_dst
, op
[0], op
[0],
1285 ir
->operands
[0]->type
->vector_elements
);
1286 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1287 /* The clamping to [0,1] can be done for free in the fragment
1288 * shader with a saturate.
1290 dp
->saturate
= true;
1292 /* Negating the result of the dot-product gives values on the range
1293 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1294 * is achieved using SLT.
1296 src_reg slt_src
= result_src
;
1297 slt_src
.negate
= ~slt_src
.negate
;
1298 emit(ir
, OPCODE_SLT
, result_dst
, slt_src
, src_reg_for_float(0.0));
1303 case ir_binop_logic_xor
:
1304 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1307 case ir_binop_logic_or
: {
1308 /* After the addition, the value will be an integer on the
1309 * range [0,2]. Zero stays zero, and positive values become 1.0.
1311 ir_to_mesa_instruction
*add
=
1312 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1313 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1314 /* The clamping to [0,1] can be done for free in the fragment
1315 * shader with a saturate.
1317 add
->saturate
= true;
1319 /* Negating the result of the addition gives values on the range
1320 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1321 * is achieved using SLT.
1323 src_reg slt_src
= result_src
;
1324 slt_src
.negate
= ~slt_src
.negate
;
1325 emit(ir
, OPCODE_SLT
, result_dst
, slt_src
, src_reg_for_float(0.0));
1330 case ir_binop_logic_and
:
1331 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1332 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1336 assert(ir
->operands
[0]->type
->is_vector());
1337 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1338 emit_dp(ir
, result_dst
, op
[0], op
[1],
1339 ir
->operands
[0]->type
->vector_elements
);
1343 /* sqrt(x) = x * rsq(x). */
1344 emit_scalar(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1345 emit(ir
, OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1346 /* For incoming channels <= 0, set the result to 0. */
1347 op
[0].negate
= ~op
[0].negate
;
1348 emit(ir
, OPCODE_CMP
, result_dst
,
1349 op
[0], result_src
, src_reg_for_float(0.0));
1352 emit_scalar(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1360 /* Mesa IR lacks types, ints are stored as truncated floats. */
1364 emit(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1368 emit(ir
, OPCODE_SNE
, result_dst
,
1369 op
[0], src_reg_for_float(0.0));
1372 emit(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1375 op
[0].negate
= ~op
[0].negate
;
1376 emit(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1377 result_src
.negate
= ~result_src
.negate
;
1380 emit(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1383 emit(ir
, OPCODE_FRC
, result_dst
, op
[0]);
1387 emit(ir
, OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1390 emit(ir
, OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1393 emit_scalar(ir
, OPCODE_POW
, result_dst
, op
[0], op
[1]);
1396 case ir_unop_bit_not
:
1397 case ir_binop_lshift
:
1398 case ir_binop_rshift
:
1399 case ir_binop_bit_and
:
1400 case ir_binop_bit_xor
:
1401 case ir_binop_bit_or
:
1402 case ir_unop_round_even
:
1403 assert(!"GLSL 1.30 features unsupported");
1406 case ir_quadop_vector
:
1407 /* This operation should have already been handled.
1409 assert(!"Should not get here.");
1413 this->result
= result_src
;
1418 ir_to_mesa_visitor::visit(ir_swizzle
*ir
)
1424 /* Note that this is only swizzles in expressions, not those on the left
1425 * hand side of an assignment, which do write masking. See ir_assignment
1429 ir
->val
->accept(this);
1431 assert(src
.file
!= PROGRAM_UNDEFINED
);
1433 for (i
= 0; i
< 4; i
++) {
1434 if (i
< ir
->type
->vector_elements
) {
1437 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1440 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1443 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1446 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1450 /* If the type is smaller than a vec4, replicate the last
1453 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1457 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1463 ir_to_mesa_visitor::visit(ir_dereference_variable
*ir
)
1465 variable_storage
*entry
= find_variable_storage(ir
->var
);
1466 ir_variable
*var
= ir
->var
;
1469 switch (var
->mode
) {
1470 case ir_var_uniform
:
1471 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1473 this->variables
.push_tail(entry
);
1477 /* The linker assigns locations for varyings and attributes,
1478 * including deprecated builtins (like gl_Color),
1479 * user-assigned generic attributes (glBindVertexLocation),
1480 * and user-defined varyings.
1482 * FINISHME: We would hit this path for function arguments. Fix!
1484 assert(var
->location
!= -1);
1485 entry
= new(mem_ctx
) variable_storage(var
,
1488 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1489 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1490 _mesa_add_attribute(this->prog
->Attributes
,
1492 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1494 var
->location
- VERT_ATTRIB_GENERIC0
);
1498 assert(var
->location
!= -1);
1499 entry
= new(mem_ctx
) variable_storage(var
,
1503 case ir_var_system_value
:
1504 entry
= new(mem_ctx
) variable_storage(var
,
1505 PROGRAM_SYSTEM_VALUE
,
1509 case ir_var_temporary
:
1510 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1512 this->variables
.push_tail(entry
);
1514 next_temp
+= type_size(var
->type
);
1519 printf("Failed to make storage for %s\n", var
->name
);
1524 this->result
= src_reg(entry
->file
, entry
->index
, var
->type
);
1528 ir_to_mesa_visitor::visit(ir_dereference_array
*ir
)
1532 int element_size
= type_size(ir
->type
);
1534 index
= ir
->array_index
->constant_expression_value();
1536 ir
->array
->accept(this);
1540 src
.index
+= index
->value
.i
[0] * element_size
;
1542 /* Variable index array dereference. It eats the "vec4" of the
1543 * base of the array and an index that offsets the Mesa register
1546 ir
->array_index
->accept(this);
1550 if (element_size
== 1) {
1551 index_reg
= this->result
;
1553 index_reg
= get_temp(glsl_type::float_type
);
1555 emit(ir
, OPCODE_MUL
, dst_reg(index_reg
),
1556 this->result
, src_reg_for_float(element_size
));
1559 /* If there was already a relative address register involved, add the
1560 * new and the old together to get the new offset.
1562 if (src
.reladdr
!= NULL
) {
1563 src_reg accum_reg
= get_temp(glsl_type::float_type
);
1565 emit(ir
, OPCODE_ADD
, dst_reg(accum_reg
),
1566 index_reg
, *src
.reladdr
);
1568 index_reg
= accum_reg
;
1571 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
1572 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1575 /* If the type is smaller than a vec4, replicate the last channel out. */
1576 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1577 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1579 src
.swizzle
= SWIZZLE_NOOP
;
1585 ir_to_mesa_visitor::visit(ir_dereference_record
*ir
)
1588 const glsl_type
*struct_type
= ir
->record
->type
;
1591 ir
->record
->accept(this);
1593 for (i
= 0; i
< struct_type
->length
; i
++) {
1594 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1596 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1599 /* If the type is smaller than a vec4, replicate the last channel out. */
1600 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1601 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1603 this->result
.swizzle
= SWIZZLE_NOOP
;
1605 this->result
.index
+= offset
;
1609 * We want to be careful in assignment setup to hit the actual storage
1610 * instead of potentially using a temporary like we might with the
1611 * ir_dereference handler.
1614 get_assignment_lhs(ir_dereference
*ir
, ir_to_mesa_visitor
*v
)
1616 /* The LHS must be a dereference. If the LHS is a variable indexed array
1617 * access of a vector, it must be separated into a series conditional moves
1618 * before reaching this point (see ir_vec_index_to_cond_assign).
1620 assert(ir
->as_dereference());
1621 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1623 assert(!deref_array
->array
->type
->is_vector());
1626 /* Use the rvalue deref handler for the most part. We'll ignore
1627 * swizzles in it and write swizzles using writemask, though.
1630 return dst_reg(v
->result
);
1634 * Process the condition of a conditional assignment
1636 * Examines the condition of a conditional assignment to generate the optimal
1637 * first operand of a \c CMP instruction. If the condition is a relational
1638 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1639 * used as the source for the \c CMP instruction. Otherwise the comparison
1640 * is processed to a boolean result, and the boolean result is used as the
1641 * operand to the CMP instruction.
1644 ir_to_mesa_visitor::process_move_condition(ir_rvalue
*ir
)
1646 ir_rvalue
*src_ir
= ir
;
1648 bool switch_order
= false;
1650 ir_expression
*const expr
= ir
->as_expression();
1651 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1652 bool zero_on_left
= false;
1654 if (expr
->operands
[0]->is_zero()) {
1655 src_ir
= expr
->operands
[1];
1656 zero_on_left
= true;
1657 } else if (expr
->operands
[1]->is_zero()) {
1658 src_ir
= expr
->operands
[0];
1659 zero_on_left
= false;
1663 * (a < 0) T F F ( a < 0) T F F
1664 * (0 < a) F F T (-a < 0) F F T
1665 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1666 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1667 * (a > 0) F F T (-a < 0) F F T
1668 * (0 > a) T F F ( a < 0) T F F
1669 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1670 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1672 * Note that exchanging the order of 0 and 'a' in the comparison simply
1673 * means that the value of 'a' should be negated.
1676 switch (expr
->operation
) {
1678 switch_order
= false;
1679 negate
= zero_on_left
;
1682 case ir_binop_greater
:
1683 switch_order
= false;
1684 negate
= !zero_on_left
;
1687 case ir_binop_lequal
:
1688 switch_order
= true;
1689 negate
= !zero_on_left
;
1692 case ir_binop_gequal
:
1693 switch_order
= true;
1694 negate
= zero_on_left
;
1698 /* This isn't the right kind of comparison afterall, so make sure
1699 * the whole condition is visited.
1707 src_ir
->accept(this);
1709 /* We use the OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1710 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1711 * choose which value OPCODE_CMP produces without an extra instruction
1712 * computing the condition.
1715 this->result
.negate
= ~this->result
.negate
;
1717 return switch_order
;
1721 ir_to_mesa_visitor::visit(ir_assignment
*ir
)
1727 ir
->rhs
->accept(this);
1730 l
= get_assignment_lhs(ir
->lhs
, this);
1732 /* FINISHME: This should really set to the correct maximal writemask for each
1733 * FINISHME: component written (in the loops below). This case can only
1734 * FINISHME: occur for matrices, arrays, and structures.
1736 if (ir
->write_mask
== 0) {
1737 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1738 l
.writemask
= WRITEMASK_XYZW
;
1739 } else if (ir
->lhs
->type
->is_scalar()) {
1740 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1741 * FINISHME: W component of fragment shader output zero, work correctly.
1743 l
.writemask
= WRITEMASK_XYZW
;
1746 int first_enabled_chan
= 0;
1749 assert(ir
->lhs
->type
->is_vector());
1750 l
.writemask
= ir
->write_mask
;
1752 for (int i
= 0; i
< 4; i
++) {
1753 if (l
.writemask
& (1 << i
)) {
1754 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1759 /* Swizzle a small RHS vector into the channels being written.
1761 * glsl ir treats write_mask as dictating how many channels are
1762 * present on the RHS while Mesa IR treats write_mask as just
1763 * showing which channels of the vec4 RHS get written.
1765 for (int i
= 0; i
< 4; i
++) {
1766 if (l
.writemask
& (1 << i
))
1767 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1769 swizzles
[i
] = first_enabled_chan
;
1771 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1772 swizzles
[2], swizzles
[3]);
1775 assert(l
.file
!= PROGRAM_UNDEFINED
);
1776 assert(r
.file
!= PROGRAM_UNDEFINED
);
1778 if (ir
->condition
) {
1779 const bool switch_order
= this->process_move_condition(ir
->condition
);
1780 src_reg condition
= this->result
;
1782 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1784 emit(ir
, OPCODE_CMP
, l
, condition
, src_reg(l
), r
);
1786 emit(ir
, OPCODE_CMP
, l
, condition
, r
, src_reg(l
));
1793 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1794 emit(ir
, OPCODE_MOV
, l
, r
);
1803 ir_to_mesa_visitor::visit(ir_constant
*ir
)
1806 GLfloat stack_vals
[4] = { 0 };
1807 GLfloat
*values
= stack_vals
;
1810 /* Unfortunately, 4 floats is all we can get into
1811 * _mesa_add_unnamed_constant. So, make a temp to store an
1812 * aggregate constant and move each constant value into it. If we
1813 * get lucky, copy propagation will eliminate the extra moves.
1816 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1817 src_reg temp_base
= get_temp(ir
->type
);
1818 dst_reg temp
= dst_reg(temp_base
);
1820 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1821 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1822 int size
= type_size(field_value
->type
);
1826 field_value
->accept(this);
1829 for (i
= 0; i
< (unsigned int)size
; i
++) {
1830 emit(ir
, OPCODE_MOV
, temp
, src
);
1836 this->result
= temp_base
;
1840 if (ir
->type
->is_array()) {
1841 src_reg temp_base
= get_temp(ir
->type
);
1842 dst_reg temp
= dst_reg(temp_base
);
1843 int size
= type_size(ir
->type
->fields
.array
);
1847 for (i
= 0; i
< ir
->type
->length
; i
++) {
1848 ir
->array_elements
[i
]->accept(this);
1850 for (int j
= 0; j
< size
; j
++) {
1851 emit(ir
, OPCODE_MOV
, temp
, src
);
1857 this->result
= temp_base
;
1861 if (ir
->type
->is_matrix()) {
1862 src_reg mat
= get_temp(ir
->type
);
1863 dst_reg mat_column
= dst_reg(mat
);
1865 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1866 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1867 values
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1869 src
= src_reg(PROGRAM_CONSTANT
, -1, NULL
);
1870 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1871 (gl_constant_value
*) values
,
1872 ir
->type
->vector_elements
,
1874 emit(ir
, OPCODE_MOV
, mat_column
, src
);
1883 src
.file
= PROGRAM_CONSTANT
;
1884 switch (ir
->type
->base_type
) {
1885 case GLSL_TYPE_FLOAT
:
1886 values
= &ir
->value
.f
[0];
1888 case GLSL_TYPE_UINT
:
1889 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1890 values
[i
] = ir
->value
.u
[i
];
1894 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1895 values
[i
] = ir
->value
.i
[i
];
1898 case GLSL_TYPE_BOOL
:
1899 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1900 values
[i
] = ir
->value
.b
[i
];
1904 assert(!"Non-float/uint/int/bool constant");
1907 this->result
= src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
1908 this->result
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1909 (gl_constant_value
*) values
,
1910 ir
->type
->vector_elements
,
1911 &this->result
.swizzle
);
1915 ir_to_mesa_visitor::get_function_signature(ir_function_signature
*sig
)
1917 function_entry
*entry
;
1919 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
1920 entry
= (function_entry
*)iter
.get();
1922 if (entry
->sig
== sig
)
1926 entry
= ralloc(mem_ctx
, function_entry
);
1928 entry
->sig_id
= this->next_signature_id
++;
1929 entry
->bgn_inst
= NULL
;
1931 /* Allocate storage for all the parameters. */
1932 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
1933 ir_variable
*param
= (ir_variable
*)iter
.get();
1934 variable_storage
*storage
;
1936 storage
= find_variable_storage(param
);
1939 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
1941 this->variables
.push_tail(storage
);
1943 this->next_temp
+= type_size(param
->type
);
1946 if (!sig
->return_type
->is_void()) {
1947 entry
->return_reg
= get_temp(sig
->return_type
);
1949 entry
->return_reg
= undef_src
;
1952 this->function_signatures
.push_tail(entry
);
1957 ir_to_mesa_visitor::visit(ir_call
*ir
)
1959 ir_to_mesa_instruction
*call_inst
;
1960 ir_function_signature
*sig
= ir
->get_callee();
1961 function_entry
*entry
= get_function_signature(sig
);
1964 /* Process in parameters. */
1965 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
1966 foreach_iter(exec_list_iterator
, iter
, *ir
) {
1967 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
1968 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
1970 if (param
->mode
== ir_var_in
||
1971 param
->mode
== ir_var_inout
) {
1972 variable_storage
*storage
= find_variable_storage(param
);
1975 param_rval
->accept(this);
1976 src_reg r
= this->result
;
1979 l
.file
= storage
->file
;
1980 l
.index
= storage
->index
;
1982 l
.writemask
= WRITEMASK_XYZW
;
1983 l
.cond_mask
= COND_TR
;
1985 for (i
= 0; i
< type_size(param
->type
); i
++) {
1986 emit(ir
, OPCODE_MOV
, l
, r
);
1994 assert(!sig_iter
.has_next());
1996 /* Emit call instruction */
1997 call_inst
= emit(ir
, OPCODE_CAL
);
1998 call_inst
->function
= entry
;
2000 /* Process out parameters. */
2001 sig_iter
= sig
->parameters
.iterator();
2002 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2003 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2004 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2006 if (param
->mode
== ir_var_out
||
2007 param
->mode
== ir_var_inout
) {
2008 variable_storage
*storage
= find_variable_storage(param
);
2012 r
.file
= storage
->file
;
2013 r
.index
= storage
->index
;
2015 r
.swizzle
= SWIZZLE_NOOP
;
2018 param_rval
->accept(this);
2019 dst_reg l
= dst_reg(this->result
);
2021 for (i
= 0; i
< type_size(param
->type
); i
++) {
2022 emit(ir
, OPCODE_MOV
, l
, r
);
2030 assert(!sig_iter
.has_next());
2032 /* Process return value. */
2033 this->result
= entry
->return_reg
;
2037 ir_to_mesa_visitor::visit(ir_texture
*ir
)
2039 src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
2040 dst_reg result_dst
, coord_dst
;
2041 ir_to_mesa_instruction
*inst
= NULL
;
2042 prog_opcode opcode
= OPCODE_NOP
;
2044 ir
->coordinate
->accept(this);
2046 /* Put our coords in a temp. We'll need to modify them for shadow,
2047 * projection, or LOD, so the only case we'd use it as is is if
2048 * we're doing plain old texturing. Mesa IR optimization should
2049 * handle cleaning up our mess in that case.
2051 coord
= get_temp(glsl_type::vec4_type
);
2052 coord_dst
= dst_reg(coord
);
2053 emit(ir
, OPCODE_MOV
, coord_dst
, this->result
);
2055 if (ir
->projector
) {
2056 ir
->projector
->accept(this);
2057 projector
= this->result
;
2060 /* Storage for our result. Ideally for an assignment we'd be using
2061 * the actual storage for the result here, instead.
2063 result_src
= get_temp(glsl_type::vec4_type
);
2064 result_dst
= dst_reg(result_src
);
2068 opcode
= OPCODE_TEX
;
2071 opcode
= OPCODE_TXB
;
2072 ir
->lod_info
.bias
->accept(this);
2073 lod_info
= this->result
;
2076 opcode
= OPCODE_TXL
;
2077 ir
->lod_info
.lod
->accept(this);
2078 lod_info
= this->result
;
2081 opcode
= OPCODE_TXD
;
2082 ir
->lod_info
.grad
.dPdx
->accept(this);
2084 ir
->lod_info
.grad
.dPdy
->accept(this);
2088 assert(!"GLSL 1.30 features unsupported");
2092 if (ir
->projector
) {
2093 if (opcode
== OPCODE_TEX
) {
2094 /* Slot the projector in as the last component of the coord. */
2095 coord_dst
.writemask
= WRITEMASK_W
;
2096 emit(ir
, OPCODE_MOV
, coord_dst
, projector
);
2097 coord_dst
.writemask
= WRITEMASK_XYZW
;
2098 opcode
= OPCODE_TXP
;
2100 src_reg coord_w
= coord
;
2101 coord_w
.swizzle
= SWIZZLE_WWWW
;
2103 /* For the other TEX opcodes there's no projective version
2104 * since the last slot is taken up by lod info. Do the
2105 * projective divide now.
2107 coord_dst
.writemask
= WRITEMASK_W
;
2108 emit(ir
, OPCODE_RCP
, coord_dst
, projector
);
2110 /* In the case where we have to project the coordinates "by hand,"
2111 * the shadow comparitor value must also be projected.
2113 src_reg tmp_src
= coord
;
2114 if (ir
->shadow_comparitor
) {
2115 /* Slot the shadow value in as the second to last component of the
2118 ir
->shadow_comparitor
->accept(this);
2120 tmp_src
= get_temp(glsl_type::vec4_type
);
2121 dst_reg tmp_dst
= dst_reg(tmp_src
);
2123 tmp_dst
.writemask
= WRITEMASK_Z
;
2124 emit(ir
, OPCODE_MOV
, tmp_dst
, this->result
);
2126 tmp_dst
.writemask
= WRITEMASK_XY
;
2127 emit(ir
, OPCODE_MOV
, tmp_dst
, coord
);
2130 coord_dst
.writemask
= WRITEMASK_XYZ
;
2131 emit(ir
, OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2133 coord_dst
.writemask
= WRITEMASK_XYZW
;
2134 coord
.swizzle
= SWIZZLE_XYZW
;
2138 /* If projection is done and the opcode is not OPCODE_TXP, then the shadow
2139 * comparitor was put in the correct place (and projected) by the code,
2140 * above, that handles by-hand projection.
2142 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== OPCODE_TXP
)) {
2143 /* Slot the shadow value in as the second to last component of the
2146 ir
->shadow_comparitor
->accept(this);
2147 coord_dst
.writemask
= WRITEMASK_Z
;
2148 emit(ir
, OPCODE_MOV
, coord_dst
, this->result
);
2149 coord_dst
.writemask
= WRITEMASK_XYZW
;
2152 if (opcode
== OPCODE_TXL
|| opcode
== OPCODE_TXB
) {
2153 /* Mesa IR stores lod or lod bias in the last channel of the coords. */
2154 coord_dst
.writemask
= WRITEMASK_W
;
2155 emit(ir
, OPCODE_MOV
, coord_dst
, lod_info
);
2156 coord_dst
.writemask
= WRITEMASK_XYZW
;
2159 if (opcode
== OPCODE_TXD
)
2160 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2162 inst
= emit(ir
, opcode
, result_dst
, coord
);
2164 if (ir
->shadow_comparitor
)
2165 inst
->tex_shadow
= GL_TRUE
;
2167 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2168 this->shader_program
,
2171 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2173 switch (sampler_type
->sampler_dimensionality
) {
2174 case GLSL_SAMPLER_DIM_1D
:
2175 inst
->tex_target
= (sampler_type
->sampler_array
)
2176 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2178 case GLSL_SAMPLER_DIM_2D
:
2179 inst
->tex_target
= (sampler_type
->sampler_array
)
2180 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2182 case GLSL_SAMPLER_DIM_3D
:
2183 inst
->tex_target
= TEXTURE_3D_INDEX
;
2185 case GLSL_SAMPLER_DIM_CUBE
:
2186 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2188 case GLSL_SAMPLER_DIM_RECT
:
2189 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2191 case GLSL_SAMPLER_DIM_BUF
:
2192 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2195 assert(!"Should not get here.");
2198 this->result
= result_src
;
2202 ir_to_mesa_visitor::visit(ir_return
*ir
)
2204 if (ir
->get_value()) {
2208 assert(current_function
);
2210 ir
->get_value()->accept(this);
2211 src_reg r
= this->result
;
2213 l
= dst_reg(current_function
->return_reg
);
2215 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2216 emit(ir
, OPCODE_MOV
, l
, r
);
2222 emit(ir
, OPCODE_RET
);
2226 ir_to_mesa_visitor::visit(ir_discard
*ir
)
2228 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2230 if (ir
->condition
) {
2231 ir
->condition
->accept(this);
2232 this->result
.negate
= ~this->result
.negate
;
2233 emit(ir
, OPCODE_KIL
, undef_dst
, this->result
);
2235 emit(ir
, OPCODE_KIL_NV
);
2238 fp
->UsesKill
= GL_TRUE
;
2242 ir_to_mesa_visitor::visit(ir_if
*ir
)
2244 ir_to_mesa_instruction
*cond_inst
, *if_inst
;
2245 ir_to_mesa_instruction
*prev_inst
;
2247 prev_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2249 ir
->condition
->accept(this);
2250 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2252 if (this->options
->EmitCondCodes
) {
2253 cond_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2255 /* See if we actually generated any instruction for generating
2256 * the condition. If not, then cook up a move to a temp so we
2257 * have something to set cond_update on.
2259 if (cond_inst
== prev_inst
) {
2260 src_reg temp
= get_temp(glsl_type::bool_type
);
2261 cond_inst
= emit(ir
->condition
, OPCODE_MOV
, dst_reg(temp
), result
);
2263 cond_inst
->cond_update
= GL_TRUE
;
2265 if_inst
= emit(ir
->condition
, OPCODE_IF
);
2266 if_inst
->dst
.cond_mask
= COND_NE
;
2268 if_inst
= emit(ir
->condition
, OPCODE_IF
, undef_dst
, this->result
);
2271 this->instructions
.push_tail(if_inst
);
2273 visit_exec_list(&ir
->then_instructions
, this);
2275 if (!ir
->else_instructions
.is_empty()) {
2276 emit(ir
->condition
, OPCODE_ELSE
);
2277 visit_exec_list(&ir
->else_instructions
, this);
2280 if_inst
= emit(ir
->condition
, OPCODE_ENDIF
);
2283 ir_to_mesa_visitor::ir_to_mesa_visitor()
2285 result
.file
= PROGRAM_UNDEFINED
;
2287 next_signature_id
= 1;
2288 current_function
= NULL
;
2289 mem_ctx
= ralloc_context(NULL
);
2292 ir_to_mesa_visitor::~ir_to_mesa_visitor()
2294 ralloc_free(mem_ctx
);
2297 static struct prog_src_register
2298 mesa_src_reg_from_ir_src_reg(src_reg reg
)
2300 struct prog_src_register mesa_reg
;
2302 mesa_reg
.File
= reg
.file
;
2303 assert(reg
.index
< (1 << INST_INDEX_BITS
));
2304 mesa_reg
.Index
= reg
.index
;
2305 mesa_reg
.Swizzle
= reg
.swizzle
;
2306 mesa_reg
.RelAddr
= reg
.reladdr
!= NULL
;
2307 mesa_reg
.Negate
= reg
.negate
;
2309 mesa_reg
.HasIndex2
= GL_FALSE
;
2310 mesa_reg
.RelAddr2
= 0;
2311 mesa_reg
.Index2
= 0;
2317 set_branchtargets(ir_to_mesa_visitor
*v
,
2318 struct prog_instruction
*mesa_instructions
,
2319 int num_instructions
)
2321 int if_count
= 0, loop_count
= 0;
2322 int *if_stack
, *loop_stack
;
2323 int if_stack_pos
= 0, loop_stack_pos
= 0;
2326 for (i
= 0; i
< num_instructions
; i
++) {
2327 switch (mesa_instructions
[i
].Opcode
) {
2331 case OPCODE_BGNLOOP
:
2336 mesa_instructions
[i
].BranchTarget
= -1;
2343 if_stack
= rzalloc_array(v
->mem_ctx
, int, if_count
);
2344 loop_stack
= rzalloc_array(v
->mem_ctx
, int, loop_count
);
2346 for (i
= 0; i
< num_instructions
; i
++) {
2347 switch (mesa_instructions
[i
].Opcode
) {
2349 if_stack
[if_stack_pos
] = i
;
2353 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2354 if_stack
[if_stack_pos
- 1] = i
;
2357 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2360 case OPCODE_BGNLOOP
:
2361 loop_stack
[loop_stack_pos
] = i
;
2364 case OPCODE_ENDLOOP
:
2366 /* Rewrite any breaks/conts at this nesting level (haven't
2367 * already had a BranchTarget assigned) to point to the end
2370 for (j
= loop_stack
[loop_stack_pos
]; j
< i
; j
++) {
2371 if (mesa_instructions
[j
].Opcode
== OPCODE_BRK
||
2372 mesa_instructions
[j
].Opcode
== OPCODE_CONT
) {
2373 if (mesa_instructions
[j
].BranchTarget
== -1) {
2374 mesa_instructions
[j
].BranchTarget
= i
;
2378 /* The loop ends point at each other. */
2379 mesa_instructions
[i
].BranchTarget
= loop_stack
[loop_stack_pos
];
2380 mesa_instructions
[loop_stack
[loop_stack_pos
]].BranchTarget
= i
;
2383 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
2384 function_entry
*entry
= (function_entry
*)iter
.get();
2386 if (entry
->sig_id
== mesa_instructions
[i
].BranchTarget
) {
2387 mesa_instructions
[i
].BranchTarget
= entry
->inst
;
2399 print_program(struct prog_instruction
*mesa_instructions
,
2400 ir_instruction
**mesa_instruction_annotation
,
2401 int num_instructions
)
2403 ir_instruction
*last_ir
= NULL
;
2407 for (i
= 0; i
< num_instructions
; i
++) {
2408 struct prog_instruction
*mesa_inst
= mesa_instructions
+ i
;
2409 ir_instruction
*ir
= mesa_instruction_annotation
[i
];
2411 fprintf(stdout
, "%3d: ", i
);
2413 if (last_ir
!= ir
&& ir
) {
2416 for (j
= 0; j
< indent
; j
++) {
2417 fprintf(stdout
, " ");
2423 fprintf(stdout
, " "); /* line number spacing. */
2426 indent
= _mesa_fprint_instruction_opt(stdout
, mesa_inst
, indent
,
2427 PROG_PRINT_DEBUG
, NULL
);
2433 * Count resources used by the given gpu program (number of texture
2437 count_resources(struct gl_program
*prog
)
2441 prog
->SamplersUsed
= 0;
2443 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
2444 struct prog_instruction
*inst
= &prog
->Instructions
[i
];
2446 if (_mesa_is_tex_instruction(inst
->Opcode
)) {
2447 prog
->SamplerTargets
[inst
->TexSrcUnit
] =
2448 (gl_texture_index
)inst
->TexSrcTarget
;
2449 prog
->SamplersUsed
|= 1 << inst
->TexSrcUnit
;
2450 if (inst
->TexShadow
) {
2451 prog
->ShadowSamplers
|= 1 << inst
->TexSrcUnit
;
2456 _mesa_update_shader_textures_used(prog
);
2461 * Check if the given vertex/fragment/shader program is within the
2462 * resource limits of the context (number of texture units, etc).
2463 * If any of those checks fail, record a linker error.
2465 * XXX more checks are needed...
2468 check_resources(const struct gl_context
*ctx
,
2469 struct gl_shader_program
*shader_program
,
2470 struct gl_program
*prog
)
2472 switch (prog
->Target
) {
2473 case GL_VERTEX_PROGRAM_ARB
:
2474 if (_mesa_bitcount(prog
->SamplersUsed
) >
2475 ctx
->Const
.MaxVertexTextureImageUnits
) {
2476 linker_error(shader_program
,
2477 "Too many vertex shader texture samplers");
2479 if (prog
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2480 linker_error(shader_program
, "Too many vertex shader constants");
2483 case MESA_GEOMETRY_PROGRAM
:
2484 if (_mesa_bitcount(prog
->SamplersUsed
) >
2485 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2486 linker_error(shader_program
,
2487 "Too many geometry shader texture samplers");
2489 if (prog
->Parameters
->NumParameters
>
2490 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2491 linker_error(shader_program
, "Too many geometry shader constants");
2494 case GL_FRAGMENT_PROGRAM_ARB
:
2495 if (_mesa_bitcount(prog
->SamplersUsed
) >
2496 ctx
->Const
.MaxTextureImageUnits
) {
2497 linker_error(shader_program
,
2498 "Too many fragment shader texture samplers");
2500 if (prog
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2501 linker_error(shader_program
, "Too many fragment shader constants");
2505 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2511 struct uniform_sort
{
2512 struct gl_uniform
*u
;
2516 /* The shader_program->Uniforms list is almost sorted in increasing
2517 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2518 * uniforms shared between targets. We need to add parameters in
2519 * increasing order for the targets.
2522 sort_uniforms(const void *a
, const void *b
)
2524 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2525 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2527 return u1
->pos
- u2
->pos
;
2530 /* Add the uniforms to the parameters. The linker chose locations
2531 * in our parameters lists (which weren't created yet), which the
2532 * uniforms code will use to poke values into our parameters list
2533 * when uniforms are updated.
2536 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2537 struct gl_shader
*shader
,
2538 struct gl_program
*prog
)
2541 unsigned int next_sampler
= 0, num_uniforms
= 0;
2542 struct uniform_sort
*sorted_uniforms
;
2544 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2545 shader_program
->Uniforms
->NumUniforms
);
2547 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2548 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2549 int parameter_index
= -1;
2551 switch (shader
->Type
) {
2552 case GL_VERTEX_SHADER
:
2553 parameter_index
= uniform
->VertPos
;
2555 case GL_FRAGMENT_SHADER
:
2556 parameter_index
= uniform
->FragPos
;
2558 case GL_GEOMETRY_SHADER
:
2559 parameter_index
= uniform
->GeomPos
;
2563 /* Only add uniforms used in our target. */
2564 if (parameter_index
!= -1) {
2565 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2566 sorted_uniforms
[num_uniforms
].u
= uniform
;
2571 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2574 for (i
= 0; i
< num_uniforms
; i
++) {
2575 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2576 int parameter_index
= sorted_uniforms
[i
].pos
;
2577 const glsl_type
*type
= uniform
->Type
;
2580 if (type
->is_vector() ||
2581 type
->is_scalar()) {
2582 size
= type
->vector_elements
;
2584 size
= type_size(type
) * 4;
2587 gl_register_file file
;
2588 if (type
->is_sampler() ||
2589 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2590 file
= PROGRAM_SAMPLER
;
2592 file
= PROGRAM_UNIFORM
;
2595 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2599 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2600 uniform
->Name
, size
, type
->gl_type
,
2603 /* Sampler uniform values are stored in prog->SamplerUnits,
2604 * and the entry in that array is selected by this index we
2605 * store in ParameterValues[].
2607 if (file
== PROGRAM_SAMPLER
) {
2608 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2609 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2612 /* The location chosen in the Parameters list here (returned
2613 * from _mesa_add_uniform) has to match what the linker chose.
2615 if (index
!= parameter_index
) {
2616 linker_error(shader_program
,
2617 "Allocation of uniform `%s' to target failed "
2619 uniform
->Name
, index
, parameter_index
);
2624 ralloc_free(sorted_uniforms
);
2628 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2629 struct gl_shader_program
*shader_program
,
2630 const char *name
, const glsl_type
*type
,
2633 if (type
->is_record()) {
2634 ir_constant
*field_constant
;
2636 field_constant
= (ir_constant
*)val
->components
.get_head();
2638 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2639 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2640 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2641 type
->fields
.structure
[i
].name
);
2642 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2643 field_type
, field_constant
);
2644 field_constant
= (ir_constant
*)field_constant
->next
;
2649 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2652 linker_error(shader_program
,
2653 "Couldn't find uniform for initializer %s\n", name
);
2657 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2658 ir_constant
*element
;
2659 const glsl_type
*element_type
;
2660 if (type
->is_array()) {
2661 element
= val
->array_elements
[i
];
2662 element_type
= type
->fields
.array
;
2665 element_type
= type
;
2670 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2671 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2672 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2673 conv
[j
] = element
->value
.b
[j
];
2675 values
= (void *)conv
;
2676 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2677 element_type
->vector_elements
,
2680 values
= &element
->value
;
2683 if (element_type
->is_matrix()) {
2684 _mesa_uniform_matrix(ctx
, shader_program
,
2685 element_type
->matrix_columns
,
2686 element_type
->vector_elements
,
2687 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2688 loc
+= element_type
->matrix_columns
;
2690 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2691 values
, element_type
->gl_type
);
2692 loc
+= type_size(element_type
);
2698 set_uniform_initializers(struct gl_context
*ctx
,
2699 struct gl_shader_program
*shader_program
)
2701 void *mem_ctx
= NULL
;
2703 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2704 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2709 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2710 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2711 ir_variable
*var
= ir
->as_variable();
2713 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2717 mem_ctx
= ralloc_context(NULL
);
2719 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2720 var
->type
, var
->constant_value
);
2724 ralloc_free(mem_ctx
);
2728 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
2729 * channels for copy propagation and updates following instructions to
2730 * use the original versions.
2732 * The ir_to_mesa_visitor lazily produces code assuming that this pass
2733 * will occur. As an example, a TXP production before this pass:
2735 * 0: MOV TEMP[1], INPUT[4].xyyy;
2736 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2737 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
2741 * 0: MOV TEMP[1], INPUT[4].xyyy;
2742 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2743 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
2745 * which allows for dead code elimination on TEMP[1]'s writes.
2748 ir_to_mesa_visitor::copy_propagate(void)
2750 ir_to_mesa_instruction
**acp
= rzalloc_array(mem_ctx
,
2751 ir_to_mesa_instruction
*,
2752 this->next_temp
* 4);
2753 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
2756 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2757 ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
2759 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
2760 || inst
->dst
.index
< this->next_temp
);
2762 /* First, do any copy propagation possible into the src regs. */
2763 for (int r
= 0; r
< 3; r
++) {
2764 ir_to_mesa_instruction
*first
= NULL
;
2766 int acp_base
= inst
->src
[r
].index
* 4;
2768 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
2769 inst
->src
[r
].reladdr
)
2772 /* See if we can find entries in the ACP consisting of MOVs
2773 * from the same src register for all the swizzled channels
2774 * of this src register reference.
2776 for (int i
= 0; i
< 4; i
++) {
2777 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2778 ir_to_mesa_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
2785 assert(acp_level
[acp_base
+ src_chan
] <= level
);
2790 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
2791 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
2799 /* We've now validated that we can copy-propagate to
2800 * replace this src register reference. Do it.
2802 inst
->src
[r
].file
= first
->src
[0].file
;
2803 inst
->src
[r
].index
= first
->src
[0].index
;
2806 for (int i
= 0; i
< 4; i
++) {
2807 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2808 ir_to_mesa_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
2809 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
2812 inst
->src
[r
].swizzle
= swizzle
;
2817 case OPCODE_BGNLOOP
:
2818 case OPCODE_ENDLOOP
:
2819 /* End of a basic block, clear the ACP entirely. */
2820 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
2829 /* Clear all channels written inside the block from the ACP, but
2830 * leaving those that were not touched.
2832 for (int r
= 0; r
< this->next_temp
; r
++) {
2833 for (int c
= 0; c
< 4; c
++) {
2834 if (!acp
[4 * r
+ c
])
2837 if (acp_level
[4 * r
+ c
] >= level
)
2838 acp
[4 * r
+ c
] = NULL
;
2841 if (inst
->op
== OPCODE_ENDIF
)
2846 /* Continuing the block, clear any written channels from
2849 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
2850 /* Any temporary might be written, so no copy propagation
2851 * across this instruction.
2853 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
2854 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
2855 inst
->dst
.reladdr
) {
2856 /* Any output might be written, so no copy propagation
2857 * from outputs across this instruction.
2859 for (int r
= 0; r
< this->next_temp
; r
++) {
2860 for (int c
= 0; c
< 4; c
++) {
2861 if (!acp
[4 * r
+ c
])
2864 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
2865 acp
[4 * r
+ c
] = NULL
;
2868 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
2869 inst
->dst
.file
== PROGRAM_OUTPUT
) {
2870 /* Clear where it's used as dst. */
2871 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
2872 for (int c
= 0; c
< 4; c
++) {
2873 if (inst
->dst
.writemask
& (1 << c
)) {
2874 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
2879 /* Clear where it's used as src. */
2880 for (int r
= 0; r
< this->next_temp
; r
++) {
2881 for (int c
= 0; c
< 4; c
++) {
2882 if (!acp
[4 * r
+ c
])
2885 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
2887 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
2888 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
2889 inst
->dst
.writemask
& (1 << src_chan
))
2891 acp
[4 * r
+ c
] = NULL
;
2899 /* If this is a copy, add it to the ACP. */
2900 if (inst
->op
== OPCODE_MOV
&&
2901 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
2902 !inst
->dst
.reladdr
&&
2904 !inst
->src
[0].reladdr
&&
2905 !inst
->src
[0].negate
) {
2906 for (int i
= 0; i
< 4; i
++) {
2907 if (inst
->dst
.writemask
& (1 << i
)) {
2908 acp
[4 * inst
->dst
.index
+ i
] = inst
;
2909 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
2915 ralloc_free(acp_level
);
2921 * Convert a shader's GLSL IR into a Mesa gl_program.
2923 static struct gl_program
*
2924 get_mesa_program(struct gl_context
*ctx
,
2925 struct gl_shader_program
*shader_program
,
2926 struct gl_shader
*shader
)
2928 ir_to_mesa_visitor v
;
2929 struct prog_instruction
*mesa_instructions
, *mesa_inst
;
2930 ir_instruction
**mesa_instruction_annotation
;
2932 struct gl_program
*prog
;
2934 const char *target_string
;
2936 struct gl_shader_compiler_options
*options
=
2937 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
2939 switch (shader
->Type
) {
2940 case GL_VERTEX_SHADER
:
2941 target
= GL_VERTEX_PROGRAM_ARB
;
2942 target_string
= "vertex";
2944 case GL_FRAGMENT_SHADER
:
2945 target
= GL_FRAGMENT_PROGRAM_ARB
;
2946 target_string
= "fragment";
2948 case GL_GEOMETRY_SHADER
:
2949 target
= GL_GEOMETRY_PROGRAM_NV
;
2950 target_string
= "geometry";
2953 assert(!"should not be reached");
2957 validate_ir_tree(shader
->ir
);
2959 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
2962 prog
->Parameters
= _mesa_new_parameter_list();
2963 prog
->Varying
= _mesa_new_parameter_list();
2964 prog
->Attributes
= _mesa_new_parameter_list();
2967 v
.shader_program
= shader_program
;
2968 v
.options
= options
;
2970 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
2972 /* Emit Mesa IR for main(). */
2973 visit_exec_list(shader
->ir
, &v
);
2974 v
.emit(NULL
, OPCODE_END
);
2976 /* Now emit bodies for any functions that were used. */
2978 progress
= GL_FALSE
;
2980 foreach_iter(exec_list_iterator
, iter
, v
.function_signatures
) {
2981 function_entry
*entry
= (function_entry
*)iter
.get();
2983 if (!entry
->bgn_inst
) {
2984 v
.current_function
= entry
;
2986 entry
->bgn_inst
= v
.emit(NULL
, OPCODE_BGNSUB
);
2987 entry
->bgn_inst
->function
= entry
;
2989 visit_exec_list(&entry
->sig
->body
, &v
);
2991 ir_to_mesa_instruction
*last
;
2992 last
= (ir_to_mesa_instruction
*)v
.instructions
.get_tail();
2993 if (last
->op
!= OPCODE_RET
)
2994 v
.emit(NULL
, OPCODE_RET
);
2996 ir_to_mesa_instruction
*end
;
2997 end
= v
.emit(NULL
, OPCODE_ENDSUB
);
2998 end
->function
= entry
;
3005 prog
->NumTemporaries
= v
.next_temp
;
3007 int num_instructions
= 0;
3008 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
3013 (struct prog_instruction
*)calloc(num_instructions
,
3014 sizeof(*mesa_instructions
));
3015 mesa_instruction_annotation
= ralloc_array(v
.mem_ctx
, ir_instruction
*,
3020 /* Convert ir_mesa_instructions into prog_instructions.
3022 mesa_inst
= mesa_instructions
;
3024 foreach_iter(exec_list_iterator
, iter
, v
.instructions
) {
3025 const ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*)iter
.get();
3027 mesa_inst
->Opcode
= inst
->op
;
3028 mesa_inst
->CondUpdate
= inst
->cond_update
;
3030 mesa_inst
->SaturateMode
= SATURATE_ZERO_ONE
;
3031 mesa_inst
->DstReg
.File
= inst
->dst
.file
;
3032 mesa_inst
->DstReg
.Index
= inst
->dst
.index
;
3033 mesa_inst
->DstReg
.CondMask
= inst
->dst
.cond_mask
;
3034 mesa_inst
->DstReg
.WriteMask
= inst
->dst
.writemask
;
3035 mesa_inst
->DstReg
.RelAddr
= inst
->dst
.reladdr
!= NULL
;
3036 mesa_inst
->SrcReg
[0] = mesa_src_reg_from_ir_src_reg(inst
->src
[0]);
3037 mesa_inst
->SrcReg
[1] = mesa_src_reg_from_ir_src_reg(inst
->src
[1]);
3038 mesa_inst
->SrcReg
[2] = mesa_src_reg_from_ir_src_reg(inst
->src
[2]);
3039 mesa_inst
->TexSrcUnit
= inst
->sampler
;
3040 mesa_inst
->TexSrcTarget
= inst
->tex_target
;
3041 mesa_inst
->TexShadow
= inst
->tex_shadow
;
3042 mesa_instruction_annotation
[i
] = inst
->ir
;
3044 /* Set IndirectRegisterFiles. */
3045 if (mesa_inst
->DstReg
.RelAddr
)
3046 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->DstReg
.File
;
3048 /* Update program's bitmask of indirectly accessed register files */
3049 for (unsigned src
= 0; src
< 3; src
++)
3050 if (mesa_inst
->SrcReg
[src
].RelAddr
)
3051 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->SrcReg
[src
].File
;
3053 switch (mesa_inst
->Opcode
) {
3055 if (options
->EmitNoIfs
) {
3056 linker_warning(shader_program
,
3057 "Couldn't flatten if-statement. "
3058 "This will likely result in software "
3059 "rasterization.\n");
3062 case OPCODE_BGNLOOP
:
3063 if (options
->EmitNoLoops
) {
3064 linker_warning(shader_program
,
3065 "Couldn't unroll loop. "
3066 "This will likely result in software "
3067 "rasterization.\n");
3071 if (options
->EmitNoCont
) {
3072 linker_warning(shader_program
,
3073 "Couldn't lower continue-statement. "
3074 "This will likely result in software "
3075 "rasterization.\n");
3079 inst
->function
->inst
= i
;
3080 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
3083 mesa_inst
->Comment
= strdup(inst
->function
->sig
->function_name());
3086 mesa_inst
->BranchTarget
= inst
->function
->sig_id
; /* rewritten later */
3089 prog
->NumAddressRegs
= 1;
3098 if (!shader_program
->LinkStatus
)
3102 if (!shader_program
->LinkStatus
) {
3103 free(mesa_instructions
);
3104 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
3108 set_branchtargets(&v
, mesa_instructions
, num_instructions
);
3110 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3112 printf("GLSL IR for linked %s program %d:\n", target_string
,
3113 shader_program
->Name
);
3114 _mesa_print_ir(shader
->ir
, NULL
);
3117 printf("Mesa IR for linked %s program %d:\n", target_string
,
3118 shader_program
->Name
);
3119 print_program(mesa_instructions
, mesa_instruction_annotation
,
3123 prog
->Instructions
= mesa_instructions
;
3124 prog
->NumInstructions
= num_instructions
;
3126 do_set_program_inouts(shader
->ir
, prog
);
3127 count_resources(prog
);
3129 check_resources(ctx
, shader_program
, prog
);
3131 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
3133 if ((ctx
->Shader
.Flags
& GLSL_NO_OPT
) == 0) {
3134 _mesa_optimize_program(ctx
, prog
);
3144 * Called via ctx->Driver.LinkShader()
3145 * This actually involves converting GLSL IR into Mesa gl_programs with
3146 * code lowering and other optimizations.
3149 _mesa_ir_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
3151 assert(prog
->LinkStatus
);
3153 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
3154 if (prog
->_LinkedShaders
[i
] == NULL
)
3158 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
3159 const struct gl_shader_compiler_options
*options
=
3160 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
3166 do_mat_op_to_vec(ir
);
3167 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
3169 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
3171 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
3173 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
3175 progress
= lower_quadop_vector(ir
, true) || progress
;
3177 if (options
->EmitNoIfs
) {
3178 progress
= lower_discard(ir
) || progress
;
3179 progress
= lower_if_to_cond_assign(ir
) || progress
;
3182 if (options
->EmitNoNoise
)
3183 progress
= lower_noise(ir
) || progress
;
3185 /* If there are forms of indirect addressing that the driver
3186 * cannot handle, perform the lowering pass.
3188 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
3189 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
3191 lower_variable_index_to_cond_assign(ir
,
3192 options
->EmitNoIndirectInput
,
3193 options
->EmitNoIndirectOutput
,
3194 options
->EmitNoIndirectTemp
,
3195 options
->EmitNoIndirectUniform
)
3198 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
3201 validate_ir_tree(ir
);
3204 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
3205 struct gl_program
*linked_prog
;
3207 if (prog
->_LinkedShaders
[i
] == NULL
)
3210 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
3215 switch (prog
->_LinkedShaders
[i
]->Type
) {
3216 case GL_VERTEX_SHADER
:
3217 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
3218 (struct gl_vertex_program
*)linked_prog
);
3219 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
3222 case GL_FRAGMENT_SHADER
:
3223 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
3224 (struct gl_fragment_program
*)linked_prog
);
3225 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
3228 case GL_GEOMETRY_SHADER
:
3229 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
3230 (struct gl_geometry_program
*)linked_prog
);
3231 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
3240 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
3248 * Compile a GLSL shader. Called via glCompileShader().
3251 _mesa_glsl_compile_shader(struct gl_context
*ctx
, struct gl_shader
*shader
)
3253 struct _mesa_glsl_parse_state
*state
=
3254 new(shader
) _mesa_glsl_parse_state(ctx
, shader
->Type
, shader
);
3256 const char *source
= shader
->Source
;
3257 /* Check if the user called glCompileShader without first calling
3258 * glShaderSource. This should fail to compile, but not raise a GL_ERROR.
3260 if (source
== NULL
) {
3261 shader
->CompileStatus
= GL_FALSE
;
3265 state
->error
= preprocess(state
, &source
, &state
->info_log
,
3266 &ctx
->Extensions
, ctx
->API
);
3268 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3269 printf("GLSL source for %s shader %d:\n",
3270 _mesa_glsl_shader_target_name(state
->target
), shader
->Name
);
3271 printf("%s\n", shader
->Source
);
3274 if (!state
->error
) {
3275 _mesa_glsl_lexer_ctor(state
, source
);
3276 _mesa_glsl_parse(state
);
3277 _mesa_glsl_lexer_dtor(state
);
3280 ralloc_free(shader
->ir
);
3281 shader
->ir
= new(shader
) exec_list
;
3282 if (!state
->error
&& !state
->translation_unit
.is_empty())
3283 _mesa_ast_to_hir(shader
->ir
, state
);
3285 if (!state
->error
&& !shader
->ir
->is_empty()) {
3286 validate_ir_tree(shader
->ir
);
3288 /* Do some optimization at compile time to reduce shader IR size
3289 * and reduce later work if the same shader is linked multiple times
3291 while (do_common_optimization(shader
->ir
, false, 32))
3294 validate_ir_tree(shader
->ir
);
3297 shader
->symbols
= state
->symbols
;
3299 shader
->CompileStatus
= !state
->error
;
3300 shader
->InfoLog
= state
->info_log
;
3301 shader
->Version
= state
->language_version
;
3302 memcpy(shader
->builtins_to_link
, state
->builtins_to_link
,
3303 sizeof(shader
->builtins_to_link
[0]) * state
->num_builtins_to_link
);
3304 shader
->num_builtins_to_link
= state
->num_builtins_to_link
;
3306 if (ctx
->Shader
.Flags
& GLSL_LOG
) {
3307 _mesa_write_shader_to_file(shader
);
3310 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3311 if (shader
->CompileStatus
) {
3312 printf("GLSL IR for shader %d:\n", shader
->Name
);
3313 _mesa_print_ir(shader
->ir
, NULL
);
3316 printf("GLSL shader %d failed to compile.\n", shader
->Name
);
3318 if (shader
->InfoLog
&& shader
->InfoLog
[0] != 0) {
3319 printf("GLSL shader %d info log:\n", shader
->Name
);
3320 printf("%s\n", shader
->InfoLog
);
3324 /* Retain any live IR, but trash the rest. */
3325 reparent_ir(shader
->ir
, shader
->ir
);
3332 * Link a GLSL shader program. Called via glLinkProgram().
3335 _mesa_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
3339 _mesa_clear_shader_program_data(ctx
, prog
);
3341 prog
->LinkStatus
= GL_TRUE
;
3343 for (i
= 0; i
< prog
->NumShaders
; i
++) {
3344 if (!prog
->Shaders
[i
]->CompileStatus
) {
3345 linker_error(prog
, "linking with uncompiled shader");
3346 prog
->LinkStatus
= GL_FALSE
;
3350 prog
->Varying
= _mesa_new_parameter_list();
3351 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
3352 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
3353 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
3355 if (prog
->LinkStatus
) {
3356 link_shaders(ctx
, prog
);
3359 if (prog
->LinkStatus
) {
3360 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
3361 prog
->LinkStatus
= GL_FALSE
;
3365 set_uniform_initializers(ctx
, prog
);
3367 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
3368 if (!prog
->LinkStatus
) {
3369 printf("GLSL shader program %d failed to link\n", prog
->Name
);
3372 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
3373 printf("GLSL shader program %d info log:\n", prog
->Name
);
3374 printf("%s\n", prog
->InfoLog
);