2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
34 #include "main/compiler.h"
36 #include "ir_visitor.h"
37 #include "ir_print_visitor.h"
38 #include "ir_expression_flattening.h"
39 #include "glsl_types.h"
40 #include "glsl_parser_extras.h"
41 #include "../glsl/program.h"
42 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "main/uniforms.h"
52 #include "program/prog_instruction.h"
53 #include "program/prog_optimize.h"
54 #include "program/prog_print.h"
55 #include "program/program.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
59 #include "pipe/p_compiler.h"
60 #include "pipe/p_context.h"
61 #include "pipe/p_screen.h"
62 #include "pipe/p_shader_tokens.h"
63 #include "pipe/p_state.h"
64 #include "util/u_math.h"
65 #include "tgsi/tgsi_ureg.h"
66 #include "tgsi/tgsi_info.h"
67 #include "st_context.h"
68 #include "st_program.h"
69 #include "st_glsl_to_tgsi.h"
70 #include "st_mesa_to_tgsi.h"
73 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
74 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
75 (1 << PROGRAM_ENV_PARAM) | \
76 (1 << PROGRAM_STATE_VAR) | \
77 (1 << PROGRAM_NAMED_PARAM) | \
78 (1 << PROGRAM_CONSTANT) | \
79 (1 << PROGRAM_UNIFORM))
82 * Maximum number of temporary registers.
84 * It is too big for stack allocated arrays -- it will cause stack overflow on
85 * Windows and likely Mac OS X.
87 #define MAX_TEMPS 4096
89 /* will be 4 for GLSL 4.00 */
90 #define MAX_GLSL_TEXTURE_OFFSET 1
95 static int swizzle_for_size(int size
);
98 * This struct is a corresponding struct to TGSI ureg_src.
102 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
106 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
107 this->swizzle
= swizzle_for_size(type
->vector_elements
);
109 this->swizzle
= SWIZZLE_XYZW
;
111 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
112 this->reladdr
= NULL
;
115 st_src_reg(gl_register_file file
, int index
, int type
)
120 this->swizzle
= SWIZZLE_XYZW
;
122 this->reladdr
= NULL
;
127 this->type
= GLSL_TYPE_ERROR
;
128 this->file
= PROGRAM_UNDEFINED
;
132 this->reladdr
= NULL
;
135 explicit st_src_reg(st_dst_reg reg
);
137 gl_register_file file
; /**< PROGRAM_* from Mesa */
138 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
139 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
140 int negate
; /**< NEGATE_XYZW mask from mesa */
141 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
142 /** Register index should be offset by the integer in this reg. */
148 st_dst_reg(gl_register_file file
, int writemask
, int type
)
152 this->writemask
= writemask
;
153 this->cond_mask
= COND_TR
;
154 this->reladdr
= NULL
;
160 this->type
= GLSL_TYPE_ERROR
;
161 this->file
= PROGRAM_UNDEFINED
;
164 this->cond_mask
= COND_TR
;
165 this->reladdr
= NULL
;
168 explicit st_dst_reg(st_src_reg reg
);
170 gl_register_file file
; /**< PROGRAM_* from Mesa */
171 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
172 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
174 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
175 /** Register index should be offset by the integer in this reg. */
179 st_src_reg::st_src_reg(st_dst_reg reg
)
181 this->type
= reg
.type
;
182 this->file
= reg
.file
;
183 this->index
= reg
.index
;
184 this->swizzle
= SWIZZLE_XYZW
;
186 this->reladdr
= reg
.reladdr
;
189 st_dst_reg::st_dst_reg(st_src_reg reg
)
191 this->type
= reg
.type
;
192 this->file
= reg
.file
;
193 this->index
= reg
.index
;
194 this->writemask
= WRITEMASK_XYZW
;
195 this->cond_mask
= COND_TR
;
196 this->reladdr
= reg
.reladdr
;
199 class glsl_to_tgsi_instruction
: public exec_node
{
201 /* Callers of this ralloc-based new need not call delete. It's
202 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
203 static void* operator new(size_t size
, void *ctx
)
207 node
= rzalloc_size(ctx
, size
);
208 assert(node
!= NULL
);
216 /** Pointer to the ir source this tree came from for debugging */
218 GLboolean cond_update
;
220 int sampler
; /**< sampler index */
221 int tex_target
; /**< One of TEXTURE_*_INDEX */
222 GLboolean tex_shadow
;
223 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
224 unsigned tex_offset_num_offset
;
225 int dead_mask
; /**< Used in dead code elimination */
227 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
230 class variable_storage
: public exec_node
{
232 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
233 : file(file
), index(index
), var(var
)
238 gl_register_file file
;
240 ir_variable
*var
; /* variable that maps to this, if any */
243 class immediate_storage
: public exec_node
{
245 immediate_storage(gl_constant_value
*values
, int size
, int type
)
247 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
252 gl_constant_value values
[4];
253 int size
; /**< Number of components (1-4) */
254 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
257 class function_entry
: public exec_node
{
259 ir_function_signature
*sig
;
262 * identifier of this function signature used by the program.
264 * At the point that TGSI instructions for function calls are
265 * generated, we don't know the address of the first instruction of
266 * the function body. So we make the BranchTarget that is called a
267 * small integer and rewrite them during set_branchtargets().
272 * Pointer to first instruction of the function body.
274 * Set during function body emits after main() is processed.
276 glsl_to_tgsi_instruction
*bgn_inst
;
279 * Index of the first instruction of the function body in actual TGSI.
281 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
285 /** Storage for the return value. */
286 st_src_reg return_reg
;
289 class glsl_to_tgsi_visitor
: public ir_visitor
{
291 glsl_to_tgsi_visitor();
292 ~glsl_to_tgsi_visitor();
294 function_entry
*current_function
;
296 struct gl_context
*ctx
;
297 struct gl_program
*prog
;
298 struct gl_shader_program
*shader_program
;
299 struct gl_shader_compiler_options
*options
;
303 int num_address_regs
;
305 bool indirect_addr_temps
;
306 bool indirect_addr_consts
;
307 int num_clip_distances
;
310 bool native_integers
;
312 variable_storage
*find_variable_storage(ir_variable
*var
);
314 int add_constant(gl_register_file file
, gl_constant_value values
[4],
315 int size
, int datatype
, GLuint
*swizzle_out
);
317 function_entry
*get_function_signature(ir_function_signature
*sig
);
319 st_src_reg
get_temp(const glsl_type
*type
);
320 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
322 st_src_reg
st_src_reg_for_float(float val
);
323 st_src_reg
st_src_reg_for_int(int val
);
324 st_src_reg
st_src_reg_for_type(int type
, int val
);
327 * \name Visit methods
329 * As typical for the visitor pattern, there must be one \c visit method for
330 * each concrete subclass of \c ir_instruction. Virtual base classes within
331 * the hierarchy should not have \c visit methods.
334 virtual void visit(ir_variable
*);
335 virtual void visit(ir_loop
*);
336 virtual void visit(ir_loop_jump
*);
337 virtual void visit(ir_function_signature
*);
338 virtual void visit(ir_function
*);
339 virtual void visit(ir_expression
*);
340 virtual void visit(ir_swizzle
*);
341 virtual void visit(ir_dereference_variable
*);
342 virtual void visit(ir_dereference_array
*);
343 virtual void visit(ir_dereference_record
*);
344 virtual void visit(ir_assignment
*);
345 virtual void visit(ir_constant
*);
346 virtual void visit(ir_call
*);
347 virtual void visit(ir_return
*);
348 virtual void visit(ir_discard
*);
349 virtual void visit(ir_texture
*);
350 virtual void visit(ir_if
*);
355 /** List of variable_storage */
358 /** List of immediate_storage */
359 exec_list immediates
;
362 /** List of function_entry */
363 exec_list function_signatures
;
364 int next_signature_id
;
366 /** List of glsl_to_tgsi_instruction */
367 exec_list instructions
;
369 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
371 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
372 st_dst_reg dst
, st_src_reg src0
);
374 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
375 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
377 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
379 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
381 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
383 st_src_reg src0
, st_src_reg src1
);
386 * Emit the correct dot-product instruction for the type of arguments
388 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
394 void emit_scalar(ir_instruction
*ir
, unsigned op
,
395 st_dst_reg dst
, st_src_reg src0
);
397 void emit_scalar(ir_instruction
*ir
, unsigned op
,
398 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
400 void try_emit_float_set(ir_instruction
*ir
, unsigned op
, st_dst_reg dst
);
402 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
404 void emit_scs(ir_instruction
*ir
, unsigned op
,
405 st_dst_reg dst
, const st_src_reg
&src
);
407 bool try_emit_mad(ir_expression
*ir
,
409 bool try_emit_mad_for_and_not(ir_expression
*ir
,
411 bool try_emit_sat(ir_expression
*ir
);
413 void emit_swz(ir_expression
*ir
);
415 bool process_move_condition(ir_rvalue
*ir
);
417 void simplify_cmp(void);
419 void rename_temp_register(int index
, int new_index
);
420 int get_first_temp_read(int index
);
421 int get_first_temp_write(int index
);
422 int get_last_temp_read(int index
);
423 int get_last_temp_write(int index
);
425 void copy_propagate(void);
426 void eliminate_dead_code(void);
427 int eliminate_dead_code_advanced(void);
428 void merge_registers(void);
429 void renumber_registers(void);
434 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
436 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
438 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
441 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
444 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
448 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
451 prog
->LinkStatus
= GL_FALSE
;
455 swizzle_for_size(int size
)
457 int size_swizzles
[4] = {
458 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
459 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
460 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
461 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
464 assert((size
>= 1) && (size
<= 4));
465 return size_swizzles
[size
- 1];
469 is_tex_instruction(unsigned opcode
)
471 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
476 num_inst_dst_regs(unsigned opcode
)
478 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
479 return info
->num_dst
;
483 num_inst_src_regs(unsigned opcode
)
485 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
486 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
489 glsl_to_tgsi_instruction
*
490 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
492 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
494 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
495 int num_reladdr
= 0, i
;
497 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
499 /* If we have to do relative addressing, we want to load the ARL
500 * reg directly for one of the regs, and preload the other reladdr
501 * sources into temps.
503 num_reladdr
+= dst
.reladdr
!= NULL
;
504 num_reladdr
+= src0
.reladdr
!= NULL
;
505 num_reladdr
+= src1
.reladdr
!= NULL
;
506 num_reladdr
+= src2
.reladdr
!= NULL
;
508 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
509 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
510 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
513 emit_arl(ir
, address_reg
, *dst
.reladdr
);
516 assert(num_reladdr
== 0);
526 inst
->function
= NULL
;
528 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
529 this->num_address_regs
= 1;
531 /* Update indirect addressing status used by TGSI */
534 case PROGRAM_TEMPORARY
:
535 this->indirect_addr_temps
= true;
537 case PROGRAM_LOCAL_PARAM
:
538 case PROGRAM_ENV_PARAM
:
539 case PROGRAM_STATE_VAR
:
540 case PROGRAM_NAMED_PARAM
:
541 case PROGRAM_CONSTANT
:
542 case PROGRAM_UNIFORM
:
543 this->indirect_addr_consts
= true;
545 case PROGRAM_IMMEDIATE
:
546 assert(!"immediates should not have indirect addressing");
553 for (i
=0; i
<3; i
++) {
554 if(inst
->src
[i
].reladdr
) {
555 switch(inst
->src
[i
].file
) {
556 case PROGRAM_TEMPORARY
:
557 this->indirect_addr_temps
= true;
559 case PROGRAM_LOCAL_PARAM
:
560 case PROGRAM_ENV_PARAM
:
561 case PROGRAM_STATE_VAR
:
562 case PROGRAM_NAMED_PARAM
:
563 case PROGRAM_CONSTANT
:
564 case PROGRAM_UNIFORM
:
565 this->indirect_addr_consts
= true;
567 case PROGRAM_IMMEDIATE
:
568 assert(!"immediates should not have indirect addressing");
577 this->instructions
.push_tail(inst
);
580 try_emit_float_set(ir
, op
, dst
);
586 glsl_to_tgsi_instruction
*
587 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
588 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
590 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
593 glsl_to_tgsi_instruction
*
594 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
595 st_dst_reg dst
, st_src_reg src0
)
597 assert(dst
.writemask
!= 0);
598 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
601 glsl_to_tgsi_instruction
*
602 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
604 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
608 * Emits the code to convert the result of float SET instructions to integers.
611 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
614 if ((op
== TGSI_OPCODE_SEQ
||
615 op
== TGSI_OPCODE_SNE
||
616 op
== TGSI_OPCODE_SGE
||
617 op
== TGSI_OPCODE_SLT
))
619 st_src_reg src
= st_src_reg(dst
);
620 src
.negate
= ~src
.negate
;
621 dst
.type
= GLSL_TYPE_FLOAT
;
622 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
627 * Determines whether to use an integer, unsigned integer, or float opcode
628 * based on the operands and input opcode, then emits the result.
631 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
633 st_src_reg src0
, st_src_reg src1
)
635 int type
= GLSL_TYPE_FLOAT
;
637 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
638 type
= GLSL_TYPE_FLOAT
;
639 else if (native_integers
)
640 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
642 #define case4(c, f, i, u) \
643 case TGSI_OPCODE_##c: \
644 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
645 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
646 else op = TGSI_OPCODE_##f; \
648 #define case3(f, i, u) case4(f, f, i, u)
649 #define case2fi(f, i) case4(f, f, i, i)
650 #define case2iu(i, u) case4(i, LAST, i, u)
656 case3(DIV
, IDIV
, UDIV
);
657 case3(MAX
, IMAX
, UMAX
);
658 case3(MIN
, IMIN
, UMIN
);
663 case3(SGE
, ISGE
, USGE
);
664 case3(SLT
, ISLT
, USLT
);
671 assert(op
!= TGSI_OPCODE_LAST
);
675 glsl_to_tgsi_instruction
*
676 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
677 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
680 static const unsigned dot_opcodes
[] = {
681 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
684 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
688 * Emits TGSI scalar opcodes to produce unique answers across channels.
690 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
691 * channel determines the result across all channels. So to do a vec4
692 * of this operation, we want to emit a scalar per source channel used
693 * to produce dest channels.
696 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
698 st_src_reg orig_src0
, st_src_reg orig_src1
)
701 int done_mask
= ~dst
.writemask
;
703 /* TGSI RCP is a scalar operation splatting results to all channels,
704 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
707 for (i
= 0; i
< 4; i
++) {
708 GLuint this_mask
= (1 << i
);
709 glsl_to_tgsi_instruction
*inst
;
710 st_src_reg src0
= orig_src0
;
711 st_src_reg src1
= orig_src1
;
713 if (done_mask
& this_mask
)
716 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
717 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
718 for (j
= i
+ 1; j
< 4; j
++) {
719 /* If there is another enabled component in the destination that is
720 * derived from the same inputs, generate its value on this pass as
723 if (!(done_mask
& (1 << j
)) &&
724 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
725 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
726 this_mask
|= (1 << j
);
729 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
730 src0_swiz
, src0_swiz
);
731 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
732 src1_swiz
, src1_swiz
);
734 inst
= emit(ir
, op
, dst
, src0
, src1
);
735 inst
->dst
.writemask
= this_mask
;
736 done_mask
|= this_mask
;
741 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
742 st_dst_reg dst
, st_src_reg src0
)
744 st_src_reg undef
= undef_src
;
746 undef
.swizzle
= SWIZZLE_XXXX
;
748 emit_scalar(ir
, op
, dst
, src0
, undef
);
752 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
753 st_dst_reg dst
, st_src_reg src0
)
755 int op
= TGSI_OPCODE_ARL
;
757 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
758 op
= TGSI_OPCODE_UARL
;
760 emit(NULL
, op
, dst
, src0
);
764 * Emit an TGSI_OPCODE_SCS instruction
766 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
767 * Instead of splatting its result across all four components of the
768 * destination, it writes one value to the \c x component and another value to
769 * the \c y component.
771 * \param ir IR instruction being processed
772 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
773 * on which value is desired.
774 * \param dst Destination register
775 * \param src Source register
778 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
780 const st_src_reg
&src
)
782 /* Vertex programs cannot use the SCS opcode.
784 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
785 emit_scalar(ir
, op
, dst
, src
);
789 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
790 const unsigned scs_mask
= (1U << component
);
791 int done_mask
= ~dst
.writemask
;
794 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
796 /* If there are compnents in the destination that differ from the component
797 * that will be written by the SCS instrution, we'll need a temporary.
799 if (scs_mask
!= unsigned(dst
.writemask
)) {
800 tmp
= get_temp(glsl_type::vec4_type
);
803 for (unsigned i
= 0; i
< 4; i
++) {
804 unsigned this_mask
= (1U << i
);
805 st_src_reg src0
= src
;
807 if ((done_mask
& this_mask
) != 0)
810 /* The source swizzle specified which component of the source generates
811 * sine / cosine for the current component in the destination. The SCS
812 * instruction requires that this value be swizzle to the X component.
813 * Replace the current swizzle with a swizzle that puts the source in
816 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
818 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
819 src0_swiz
, src0_swiz
);
820 for (unsigned j
= i
+ 1; j
< 4; j
++) {
821 /* If there is another enabled component in the destination that is
822 * derived from the same inputs, generate its value on this pass as
825 if (!(done_mask
& (1 << j
)) &&
826 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
827 this_mask
|= (1 << j
);
831 if (this_mask
!= scs_mask
) {
832 glsl_to_tgsi_instruction
*inst
;
833 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
835 /* Emit the SCS instruction.
837 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
838 inst
->dst
.writemask
= scs_mask
;
840 /* Move the result of the SCS instruction to the desired location in
843 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
844 component
, component
);
845 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
846 inst
->dst
.writemask
= this_mask
;
848 /* Emit the SCS instruction to write directly to the destination.
850 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
851 inst
->dst
.writemask
= scs_mask
;
854 done_mask
|= this_mask
;
859 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
860 gl_constant_value values
[4], int size
, int datatype
,
863 if (file
== PROGRAM_CONSTANT
) {
864 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
865 size
, datatype
, swizzle_out
);
868 immediate_storage
*entry
;
869 assert(file
== PROGRAM_IMMEDIATE
);
871 /* Search immediate storage to see if we already have an identical
872 * immediate that we can use instead of adding a duplicate entry.
874 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
875 entry
= (immediate_storage
*)iter
.get();
877 if (entry
->size
== size
&&
878 entry
->type
== datatype
&&
879 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
885 /* Add this immediate to the list. */
886 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
887 this->immediates
.push_tail(entry
);
888 this->num_immediates
++;
894 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
896 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
897 union gl_constant_value uval
;
900 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
906 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
908 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
909 union gl_constant_value uval
;
911 assert(native_integers
);
914 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
920 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
923 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
924 st_src_reg_for_int(val
);
926 return st_src_reg_for_float(val
);
930 type_size(const struct glsl_type
*type
)
935 switch (type
->base_type
) {
938 case GLSL_TYPE_FLOAT
:
940 if (type
->is_matrix()) {
941 return type
->matrix_columns
;
943 /* Regardless of size of vector, it gets a vec4. This is bad
944 * packing for things like floats, but otherwise arrays become a
945 * mess. Hopefully a later pass over the code can pack scalars
946 * down if appropriate.
950 case GLSL_TYPE_ARRAY
:
951 assert(type
->length
> 0);
952 return type_size(type
->fields
.array
) * type
->length
;
953 case GLSL_TYPE_STRUCT
:
955 for (i
= 0; i
< type
->length
; i
++) {
956 size
+= type_size(type
->fields
.structure
[i
].type
);
959 case GLSL_TYPE_SAMPLER
:
960 /* Samplers take up one slot in UNIFORMS[], but they're baked in
971 * In the initial pass of codegen, we assign temporary numbers to
972 * intermediate results. (not SSA -- variable assignments will reuse
976 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
980 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
981 src
.file
= PROGRAM_TEMPORARY
;
982 src
.index
= next_temp
;
984 next_temp
+= type_size(type
);
986 if (type
->is_array() || type
->is_record()) {
987 src
.swizzle
= SWIZZLE_NOOP
;
989 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
997 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1000 variable_storage
*entry
;
1002 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1003 entry
= (variable_storage
*)iter
.get();
1005 if (entry
->var
== var
)
1013 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1015 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1016 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1018 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1019 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1022 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1024 const ir_state_slot
*const slots
= ir
->state_slots
;
1025 assert(ir
->state_slots
!= NULL
);
1027 /* Check if this statevar's setup in the STATE file exactly
1028 * matches how we'll want to reference it as a
1029 * struct/array/whatever. If not, then we need to move it into
1030 * temporary storage and hope that it'll get copy-propagated
1033 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1034 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1039 variable_storage
*storage
;
1041 if (i
== ir
->num_state_slots
) {
1042 /* We'll set the index later. */
1043 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1044 this->variables
.push_tail(storage
);
1048 /* The variable_storage constructor allocates slots based on the size
1049 * of the type. However, this had better match the number of state
1050 * elements that we're going to copy into the new temporary.
1052 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1054 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1056 this->variables
.push_tail(storage
);
1057 this->next_temp
+= type_size(ir
->type
);
1059 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1060 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1064 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1065 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1066 (gl_state_index
*)slots
[i
].tokens
);
1068 if (storage
->file
== PROGRAM_STATE_VAR
) {
1069 if (storage
->index
== -1) {
1070 storage
->index
= index
;
1072 assert(index
== storage
->index
+ (int)i
);
1075 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1076 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1077 src
.swizzle
= slots
[i
].swizzle
;
1078 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1079 /* even a float takes up a whole vec4 reg in a struct/array. */
1084 if (storage
->file
== PROGRAM_TEMPORARY
&&
1085 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1086 fail_link(this->shader_program
,
1087 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1088 ir
->name
, dst
.index
- storage
->index
,
1089 type_size(ir
->type
));
1095 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1097 ir_dereference_variable
*counter
= NULL
;
1099 if (ir
->counter
!= NULL
)
1100 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1102 if (ir
->from
!= NULL
) {
1103 assert(ir
->counter
!= NULL
);
1105 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1111 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1115 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1117 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1119 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1121 if_stmt
->then_instructions
.push_tail(brk
);
1123 if_stmt
->accept(this);
1130 visit_exec_list(&ir
->body_instructions
, this);
1132 if (ir
->increment
) {
1134 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1135 counter
, ir
->increment
);
1137 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1144 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1148 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1151 case ir_loop_jump::jump_break
:
1152 emit(NULL
, TGSI_OPCODE_BRK
);
1154 case ir_loop_jump::jump_continue
:
1155 emit(NULL
, TGSI_OPCODE_CONT
);
1162 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1169 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1171 /* Ignore function bodies other than main() -- we shouldn't see calls to
1172 * them since they should all be inlined before we get to glsl_to_tgsi.
1174 if (strcmp(ir
->name
, "main") == 0) {
1175 const ir_function_signature
*sig
;
1178 sig
= ir
->matching_signature(&empty
);
1182 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1183 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1191 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1193 int nonmul_operand
= 1 - mul_operand
;
1195 st_dst_reg result_dst
;
1197 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1198 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1201 expr
->operands
[0]->accept(this);
1203 expr
->operands
[1]->accept(this);
1205 ir
->operands
[nonmul_operand
]->accept(this);
1208 this->result
= get_temp(ir
->type
);
1209 result_dst
= st_dst_reg(this->result
);
1210 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1211 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1217 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1219 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1220 * implemented using multiplication, and logical-or is implemented using
1221 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1222 * As result, the logical expression (a & !b) can be rewritten as:
1226 * - (a * 1) - (a * b)
1230 * This final expression can be implemented as a single MAD(a, -b, a)
1234 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1236 const int other_operand
= 1 - try_operand
;
1239 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1240 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1243 ir
->operands
[other_operand
]->accept(this);
1245 expr
->operands
[0]->accept(this);
1248 b
.negate
= ~b
.negate
;
1250 this->result
= get_temp(ir
->type
);
1251 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1257 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1259 /* Saturates were only introduced to vertex programs in
1260 * NV_vertex_program3, so don't give them to drivers in the VP.
1262 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1265 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1269 sat_src
->accept(this);
1270 st_src_reg src
= this->result
;
1272 /* If we generated an expression instruction into a temporary in
1273 * processing the saturate's operand, apply the saturate to that
1274 * instruction. Otherwise, generate a MOV to do the saturate.
1276 * Note that we have to be careful to only do this optimization if
1277 * the instruction in question was what generated src->result. For
1278 * example, ir_dereference_array might generate a MUL instruction
1279 * to create the reladdr, and return us a src reg using that
1280 * reladdr. That MUL result is not the value we're trying to
1283 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1284 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1285 sat_src_expr
->operation
== ir_binop_add
||
1286 sat_src_expr
->operation
== ir_binop_dot
)) {
1287 glsl_to_tgsi_instruction
*new_inst
;
1288 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1289 new_inst
->saturate
= true;
1291 this->result
= get_temp(ir
->type
);
1292 st_dst_reg result_dst
= st_dst_reg(this->result
);
1293 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1294 glsl_to_tgsi_instruction
*inst
;
1295 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1296 inst
->saturate
= true;
1303 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1304 st_src_reg
*reg
, int *num_reladdr
)
1309 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1311 if (*num_reladdr
!= 1) {
1312 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1314 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1322 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1324 unsigned int operand
;
1325 st_src_reg op
[Elements(ir
->operands
)];
1326 st_src_reg result_src
;
1327 st_dst_reg result_dst
;
1329 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1331 if (ir
->operation
== ir_binop_add
) {
1332 if (try_emit_mad(ir
, 1))
1334 if (try_emit_mad(ir
, 0))
1338 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1340 if (ir
->operation
== ir_binop_logic_and
) {
1341 if (try_emit_mad_for_and_not(ir
, 1))
1343 if (try_emit_mad_for_and_not(ir
, 0))
1347 if (try_emit_sat(ir
))
1350 if (ir
->operation
== ir_quadop_vector
)
1351 assert(!"ir_quadop_vector should have been lowered");
1353 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1354 this->result
.file
= PROGRAM_UNDEFINED
;
1355 ir
->operands
[operand
]->accept(this);
1356 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1358 printf("Failed to get tree for expression operand:\n");
1359 ir
->operands
[operand
]->accept(&v
);
1362 op
[operand
] = this->result
;
1364 /* Matrix expression operands should have been broken down to vector
1365 * operations already.
1367 assert(!ir
->operands
[operand
]->type
->is_matrix());
1370 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1371 if (ir
->operands
[1]) {
1372 vector_elements
= MAX2(vector_elements
,
1373 ir
->operands
[1]->type
->vector_elements
);
1376 this->result
.file
= PROGRAM_UNDEFINED
;
1378 /* Storage for our result. Ideally for an assignment we'd be using
1379 * the actual storage for the result here, instead.
1381 result_src
= get_temp(ir
->type
);
1382 /* convenience for the emit functions below. */
1383 result_dst
= st_dst_reg(result_src
);
1384 /* Limit writes to the channels that will be used by result_src later.
1385 * This does limit this temp's use as a temporary for multi-instruction
1388 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1390 switch (ir
->operation
) {
1391 case ir_unop_logic_not
:
1392 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1393 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1395 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1396 * older GPUs implement SEQ using multiple instructions (i915 uses two
1397 * SGE instructions and a MUL instruction). Since our logic values are
1398 * 0.0 and 1.0, 1-x also implements !x.
1400 op
[0].negate
= ~op
[0].negate
;
1401 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1405 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1406 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1408 op
[0].negate
= ~op
[0].negate
;
1413 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1414 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1417 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1420 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1424 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1428 assert(!"not reached: should be handled by ir_explog_to_explog2");
1431 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1434 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1437 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1439 case ir_unop_sin_reduced
:
1440 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1442 case ir_unop_cos_reduced
:
1443 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1447 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1450 op
[0].negate
= ~op
[0].negate
;
1451 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1454 case ir_unop_noise
: {
1455 /* At some point, a motivated person could add a better
1456 * implementation of noise. Currently not even the nvidia
1457 * binary drivers do anything more than this. In any case, the
1458 * place to do this is in the GL state tracker, not the poor
1461 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1466 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1469 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1473 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1476 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1477 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1479 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1482 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1483 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1485 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1489 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1491 case ir_binop_greater
:
1492 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1494 case ir_binop_lequal
:
1495 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1497 case ir_binop_gequal
:
1498 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1500 case ir_binop_equal
:
1501 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1503 case ir_binop_nequal
:
1504 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1506 case ir_binop_all_equal
:
1507 /* "==" operator producing a scalar boolean. */
1508 if (ir
->operands
[0]->type
->is_vector() ||
1509 ir
->operands
[1]->type
->is_vector()) {
1510 st_src_reg temp
= get_temp(native_integers
?
1511 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1512 glsl_type::vec4_type
);
1514 if (native_integers
) {
1515 st_dst_reg temp_dst
= st_dst_reg(temp
);
1516 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1518 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1520 /* Emit 1-3 AND operations to combine the SEQ results. */
1521 switch (ir
->operands
[0]->type
->vector_elements
) {
1525 temp_dst
.writemask
= WRITEMASK_Y
;
1526 temp1
.swizzle
= SWIZZLE_YYYY
;
1527 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1528 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1531 temp_dst
.writemask
= WRITEMASK_X
;
1532 temp1
.swizzle
= SWIZZLE_XXXX
;
1533 temp2
.swizzle
= SWIZZLE_YYYY
;
1534 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1535 temp_dst
.writemask
= WRITEMASK_Y
;
1536 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1537 temp2
.swizzle
= SWIZZLE_WWWW
;
1538 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1541 temp1
.swizzle
= SWIZZLE_XXXX
;
1542 temp2
.swizzle
= SWIZZLE_YYYY
;
1543 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1545 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1547 /* After the dot-product, the value will be an integer on the
1548 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1550 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1552 /* Negating the result of the dot-product gives values on the range
1553 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1554 * This is achieved using SGE.
1556 st_src_reg sge_src
= result_src
;
1557 sge_src
.negate
= ~sge_src
.negate
;
1558 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1561 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1564 case ir_binop_any_nequal
:
1565 /* "!=" operator producing a scalar boolean. */
1566 if (ir
->operands
[0]->type
->is_vector() ||
1567 ir
->operands
[1]->type
->is_vector()) {
1568 st_src_reg temp
= get_temp(native_integers
?
1569 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1570 glsl_type::vec4_type
);
1571 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1573 if (native_integers
) {
1574 st_dst_reg temp_dst
= st_dst_reg(temp
);
1575 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1577 /* Emit 1-3 OR operations to combine the SNE results. */
1578 switch (ir
->operands
[0]->type
->vector_elements
) {
1582 temp_dst
.writemask
= WRITEMASK_Y
;
1583 temp1
.swizzle
= SWIZZLE_YYYY
;
1584 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1585 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1588 temp_dst
.writemask
= WRITEMASK_X
;
1589 temp1
.swizzle
= SWIZZLE_XXXX
;
1590 temp2
.swizzle
= SWIZZLE_YYYY
;
1591 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1592 temp_dst
.writemask
= WRITEMASK_Y
;
1593 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1594 temp2
.swizzle
= SWIZZLE_WWWW
;
1595 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1598 temp1
.swizzle
= SWIZZLE_XXXX
;
1599 temp2
.swizzle
= SWIZZLE_YYYY
;
1600 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1602 /* After the dot-product, the value will be an integer on the
1603 * range [0,4]. Zero stays zero, and positive values become 1.0.
1605 glsl_to_tgsi_instruction
*const dp
=
1606 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1607 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1608 /* The clamping to [0,1] can be done for free in the fragment
1609 * shader with a saturate.
1611 dp
->saturate
= true;
1613 /* Negating the result of the dot-product gives values on the range
1614 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1615 * achieved using SLT.
1617 st_src_reg slt_src
= result_src
;
1618 slt_src
.negate
= ~slt_src
.negate
;
1619 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1623 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1628 assert(ir
->operands
[0]->type
->is_vector());
1630 /* After the dot-product, the value will be an integer on the
1631 * range [0,4]. Zero stays zero, and positive values become 1.0.
1633 glsl_to_tgsi_instruction
*const dp
=
1634 emit_dp(ir
, result_dst
, op
[0], op
[0],
1635 ir
->operands
[0]->type
->vector_elements
);
1636 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1637 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1638 /* The clamping to [0,1] can be done for free in the fragment
1639 * shader with a saturate.
1641 dp
->saturate
= true;
1642 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1643 /* Negating the result of the dot-product gives values on the range
1644 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1645 * is achieved using SLT.
1647 st_src_reg slt_src
= result_src
;
1648 slt_src
.negate
= ~slt_src
.negate
;
1649 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1652 /* Use SNE 0 if integers are being used as boolean values. */
1653 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1658 case ir_binop_logic_xor
:
1659 if (native_integers
)
1660 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1662 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1665 case ir_binop_logic_or
: {
1666 if (native_integers
) {
1667 /* If integers are used as booleans, we can use an actual "or"
1670 assert(native_integers
);
1671 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1673 /* After the addition, the value will be an integer on the
1674 * range [0,2]. Zero stays zero, and positive values become 1.0.
1676 glsl_to_tgsi_instruction
*add
=
1677 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1678 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1679 /* The clamping to [0,1] can be done for free in the fragment
1680 * shader with a saturate if floats are being used as boolean values.
1682 add
->saturate
= true;
1684 /* Negating the result of the addition gives values on the range
1685 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1686 * is achieved using SLT.
1688 st_src_reg slt_src
= result_src
;
1689 slt_src
.negate
= ~slt_src
.negate
;
1690 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1696 case ir_binop_logic_and
:
1697 /* If native integers are disabled, the bool args are stored as float 0.0
1698 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1699 * actual AND opcode.
1701 if (native_integers
)
1702 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1704 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1708 assert(ir
->operands
[0]->type
->is_vector());
1709 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1710 emit_dp(ir
, result_dst
, op
[0], op
[1],
1711 ir
->operands
[0]->type
->vector_elements
);
1715 /* sqrt(x) = x * rsq(x). */
1716 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1717 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1718 /* For incoming channels <= 0, set the result to 0. */
1719 op
[0].negate
= ~op
[0].negate
;
1720 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1721 op
[0], result_src
, st_src_reg_for_float(0.0));
1724 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1727 if (native_integers
) {
1728 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1731 /* fallthrough to next case otherwise */
1733 if (native_integers
) {
1734 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1737 /* fallthrough to next case otherwise */
1740 /* Converting between signed and unsigned integers is a no-op. */
1744 if (native_integers
) {
1745 /* Booleans are stored as integers using ~0 for true and 0 for false.
1746 * GLSL requires that int(bool) return 1 for true and 0 for false.
1747 * This conversion is done with AND, but it could be done with NEG.
1749 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1751 /* Booleans and integers are both stored as floats when native
1752 * integers are disabled.
1758 if (native_integers
)
1759 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1761 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1764 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1767 if (native_integers
)
1768 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1770 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1773 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1776 op
[0].negate
= ~op
[0].negate
;
1777 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1778 result_src
.negate
= ~result_src
.negate
;
1781 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1783 case ir_unop_round_even
:
1784 emit(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1787 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1791 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1794 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1797 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1800 case ir_unop_bit_not
:
1801 if (native_integers
) {
1802 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1806 if (native_integers
) {
1807 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1810 case ir_binop_lshift
:
1811 if (native_integers
) {
1812 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1815 case ir_binop_rshift
:
1816 if (native_integers
) {
1817 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1820 case ir_binop_bit_and
:
1821 if (native_integers
) {
1822 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1825 case ir_binop_bit_xor
:
1826 if (native_integers
) {
1827 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1830 case ir_binop_bit_or
:
1831 if (native_integers
) {
1832 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1836 assert(!"GLSL 1.30 features unsupported");
1839 case ir_quadop_vector
:
1840 /* This operation should have already been handled.
1842 assert(!"Should not get here.");
1846 this->result
= result_src
;
1851 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1857 /* Note that this is only swizzles in expressions, not those on the left
1858 * hand side of an assignment, which do write masking. See ir_assignment
1862 ir
->val
->accept(this);
1864 assert(src
.file
!= PROGRAM_UNDEFINED
);
1866 for (i
= 0; i
< 4; i
++) {
1867 if (i
< ir
->type
->vector_elements
) {
1870 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1873 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1876 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1879 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1883 /* If the type is smaller than a vec4, replicate the last
1886 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1890 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1896 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1898 variable_storage
*entry
= find_variable_storage(ir
->var
);
1899 ir_variable
*var
= ir
->var
;
1902 switch (var
->mode
) {
1903 case ir_var_uniform
:
1904 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1906 this->variables
.push_tail(entry
);
1910 /* The linker assigns locations for varyings and attributes,
1911 * including deprecated builtins (like gl_Color), user-assign
1912 * generic attributes (glBindVertexLocation), and
1913 * user-defined varyings.
1915 * FINISHME: We would hit this path for function arguments. Fix!
1917 assert(var
->location
!= -1);
1918 entry
= new(mem_ctx
) variable_storage(var
,
1923 assert(var
->location
!= -1);
1924 entry
= new(mem_ctx
) variable_storage(var
,
1928 case ir_var_system_value
:
1929 entry
= new(mem_ctx
) variable_storage(var
,
1930 PROGRAM_SYSTEM_VALUE
,
1934 case ir_var_temporary
:
1935 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1937 this->variables
.push_tail(entry
);
1939 next_temp
+= type_size(var
->type
);
1944 printf("Failed to make storage for %s\n", var
->name
);
1949 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1950 if (!native_integers
)
1951 this->result
.type
= GLSL_TYPE_FLOAT
;
1955 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1959 int element_size
= type_size(ir
->type
);
1961 index
= ir
->array_index
->constant_expression_value();
1963 ir
->array
->accept(this);
1967 src
.index
+= index
->value
.i
[0] * element_size
;
1969 /* Variable index array dereference. It eats the "vec4" of the
1970 * base of the array and an index that offsets the TGSI register
1973 ir
->array_index
->accept(this);
1975 st_src_reg index_reg
;
1977 if (element_size
== 1) {
1978 index_reg
= this->result
;
1980 index_reg
= get_temp(native_integers
?
1981 glsl_type::int_type
: glsl_type::float_type
);
1983 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1984 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
1987 /* If there was already a relative address register involved, add the
1988 * new and the old together to get the new offset.
1990 if (src
.reladdr
!= NULL
) {
1991 st_src_reg accum_reg
= get_temp(native_integers
?
1992 glsl_type::int_type
: glsl_type::float_type
);
1994 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
1995 index_reg
, *src
.reladdr
);
1997 index_reg
= accum_reg
;
2000 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2001 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2004 /* If the type is smaller than a vec4, replicate the last channel out. */
2005 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2006 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2008 src
.swizzle
= SWIZZLE_NOOP
;
2014 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2017 const glsl_type
*struct_type
= ir
->record
->type
;
2020 ir
->record
->accept(this);
2022 for (i
= 0; i
< struct_type
->length
; i
++) {
2023 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2025 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2028 /* If the type is smaller than a vec4, replicate the last channel out. */
2029 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2030 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2032 this->result
.swizzle
= SWIZZLE_NOOP
;
2034 this->result
.index
+= offset
;
2038 * We want to be careful in assignment setup to hit the actual storage
2039 * instead of potentially using a temporary like we might with the
2040 * ir_dereference handler.
2043 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2045 /* The LHS must be a dereference. If the LHS is a variable indexed array
2046 * access of a vector, it must be separated into a series conditional moves
2047 * before reaching this point (see ir_vec_index_to_cond_assign).
2049 assert(ir
->as_dereference());
2050 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2052 assert(!deref_array
->array
->type
->is_vector());
2055 /* Use the rvalue deref handler for the most part. We'll ignore
2056 * swizzles in it and write swizzles using writemask, though.
2059 return st_dst_reg(v
->result
);
2063 * Process the condition of a conditional assignment
2065 * Examines the condition of a conditional assignment to generate the optimal
2066 * first operand of a \c CMP instruction. If the condition is a relational
2067 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2068 * used as the source for the \c CMP instruction. Otherwise the comparison
2069 * is processed to a boolean result, and the boolean result is used as the
2070 * operand to the CMP instruction.
2073 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2075 ir_rvalue
*src_ir
= ir
;
2077 bool switch_order
= false;
2079 ir_expression
*const expr
= ir
->as_expression();
2080 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2081 bool zero_on_left
= false;
2083 if (expr
->operands
[0]->is_zero()) {
2084 src_ir
= expr
->operands
[1];
2085 zero_on_left
= true;
2086 } else if (expr
->operands
[1]->is_zero()) {
2087 src_ir
= expr
->operands
[0];
2088 zero_on_left
= false;
2092 * (a < 0) T F F ( a < 0) T F F
2093 * (0 < a) F F T (-a < 0) F F T
2094 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2095 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2096 * (a > 0) F F T (-a < 0) F F T
2097 * (0 > a) T F F ( a < 0) T F F
2098 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2099 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2101 * Note that exchanging the order of 0 and 'a' in the comparison simply
2102 * means that the value of 'a' should be negated.
2105 switch (expr
->operation
) {
2107 switch_order
= false;
2108 negate
= zero_on_left
;
2111 case ir_binop_greater
:
2112 switch_order
= false;
2113 negate
= !zero_on_left
;
2116 case ir_binop_lequal
:
2117 switch_order
= true;
2118 negate
= !zero_on_left
;
2121 case ir_binop_gequal
:
2122 switch_order
= true;
2123 negate
= zero_on_left
;
2127 /* This isn't the right kind of comparison afterall, so make sure
2128 * the whole condition is visited.
2136 src_ir
->accept(this);
2138 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2139 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2140 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2141 * computing the condition.
2144 this->result
.negate
= ~this->result
.negate
;
2146 return switch_order
;
2150 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2156 ir
->rhs
->accept(this);
2159 l
= get_assignment_lhs(ir
->lhs
, this);
2161 /* FINISHME: This should really set to the correct maximal writemask for each
2162 * FINISHME: component written (in the loops below). This case can only
2163 * FINISHME: occur for matrices, arrays, and structures.
2165 if (ir
->write_mask
== 0) {
2166 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2167 l
.writemask
= WRITEMASK_XYZW
;
2168 } else if (ir
->lhs
->type
->is_scalar() &&
2169 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2170 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2171 * FINISHME: W component of fragment shader output zero, work correctly.
2173 l
.writemask
= WRITEMASK_XYZW
;
2176 int first_enabled_chan
= 0;
2179 l
.writemask
= ir
->write_mask
;
2181 for (int i
= 0; i
< 4; i
++) {
2182 if (l
.writemask
& (1 << i
)) {
2183 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2188 /* Swizzle a small RHS vector into the channels being written.
2190 * glsl ir treats write_mask as dictating how many channels are
2191 * present on the RHS while TGSI treats write_mask as just
2192 * showing which channels of the vec4 RHS get written.
2194 for (int i
= 0; i
< 4; i
++) {
2195 if (l
.writemask
& (1 << i
))
2196 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2198 swizzles
[i
] = first_enabled_chan
;
2200 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2201 swizzles
[2], swizzles
[3]);
2204 assert(l
.file
!= PROGRAM_UNDEFINED
);
2205 assert(r
.file
!= PROGRAM_UNDEFINED
);
2207 if (ir
->condition
) {
2208 const bool switch_order
= this->process_move_condition(ir
->condition
);
2209 st_src_reg condition
= this->result
;
2211 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2212 st_src_reg l_src
= st_src_reg(l
);
2213 st_src_reg condition_temp
= condition
;
2214 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2216 if (native_integers
) {
2217 /* This is necessary because TGSI's CMP instruction expects the
2218 * condition to be a float, and we store booleans as integers.
2219 * If TGSI had a UCMP instruction or similar, this extra
2220 * instruction would not be necessary.
2222 condition_temp
= get_temp(glsl_type::vec4_type
);
2223 condition
.negate
= 0;
2224 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2225 condition_temp
.swizzle
= condition
.swizzle
;
2229 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2231 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2237 } else if (ir
->rhs
->as_expression() &&
2238 this->instructions
.get_tail() &&
2239 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2240 type_size(ir
->lhs
->type
) == 1 &&
2241 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2242 /* To avoid emitting an extra MOV when assigning an expression to a
2243 * variable, emit the last instruction of the expression again, but
2244 * replace the destination register with the target of the assignment.
2245 * Dead code elimination will remove the original instruction.
2247 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2248 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2249 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2250 new_inst
->saturate
= inst
->saturate
;
2251 inst
->dead_mask
= inst
->dst
.writemask
;
2253 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2254 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2263 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2266 GLfloat stack_vals
[4] = { 0 };
2267 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2268 GLenum gl_type
= GL_NONE
;
2270 static int in_array
= 0;
2271 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2273 /* Unfortunately, 4 floats is all we can get into
2274 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2275 * aggregate constant and move each constant value into it. If we
2276 * get lucky, copy propagation will eliminate the extra moves.
2278 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2279 st_src_reg temp_base
= get_temp(ir
->type
);
2280 st_dst_reg temp
= st_dst_reg(temp_base
);
2282 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2283 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2284 int size
= type_size(field_value
->type
);
2288 field_value
->accept(this);
2291 for (i
= 0; i
< (unsigned int)size
; i
++) {
2292 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2298 this->result
= temp_base
;
2302 if (ir
->type
->is_array()) {
2303 st_src_reg temp_base
= get_temp(ir
->type
);
2304 st_dst_reg temp
= st_dst_reg(temp_base
);
2305 int size
= type_size(ir
->type
->fields
.array
);
2310 for (i
= 0; i
< ir
->type
->length
; i
++) {
2311 ir
->array_elements
[i
]->accept(this);
2313 for (int j
= 0; j
< size
; j
++) {
2314 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2320 this->result
= temp_base
;
2325 if (ir
->type
->is_matrix()) {
2326 st_src_reg mat
= get_temp(ir
->type
);
2327 st_dst_reg mat_column
= st_dst_reg(mat
);
2329 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2330 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2331 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2333 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2334 src
.index
= add_constant(file
,
2336 ir
->type
->vector_elements
,
2339 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2348 switch (ir
->type
->base_type
) {
2349 case GLSL_TYPE_FLOAT
:
2351 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2352 values
[i
].f
= ir
->value
.f
[i
];
2355 case GLSL_TYPE_UINT
:
2356 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2357 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2358 if (native_integers
)
2359 values
[i
].u
= ir
->value
.u
[i
];
2361 values
[i
].f
= ir
->value
.u
[i
];
2365 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2366 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2367 if (native_integers
)
2368 values
[i
].i
= ir
->value
.i
[i
];
2370 values
[i
].f
= ir
->value
.i
[i
];
2373 case GLSL_TYPE_BOOL
:
2374 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2375 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2376 if (native_integers
)
2377 values
[i
].b
= ir
->value
.b
[i
];
2379 values
[i
].f
= ir
->value
.b
[i
];
2383 assert(!"Non-float/uint/int/bool constant");
2386 this->result
= st_src_reg(file
, -1, ir
->type
);
2387 this->result
.index
= add_constant(file
,
2389 ir
->type
->vector_elements
,
2391 &this->result
.swizzle
);
2395 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2397 function_entry
*entry
;
2399 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2400 entry
= (function_entry
*)iter
.get();
2402 if (entry
->sig
== sig
)
2406 entry
= ralloc(mem_ctx
, function_entry
);
2408 entry
->sig_id
= this->next_signature_id
++;
2409 entry
->bgn_inst
= NULL
;
2411 /* Allocate storage for all the parameters. */
2412 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2413 ir_variable
*param
= (ir_variable
*)iter
.get();
2414 variable_storage
*storage
;
2416 storage
= find_variable_storage(param
);
2419 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2421 this->variables
.push_tail(storage
);
2423 this->next_temp
+= type_size(param
->type
);
2426 if (!sig
->return_type
->is_void()) {
2427 entry
->return_reg
= get_temp(sig
->return_type
);
2429 entry
->return_reg
= undef_src
;
2432 this->function_signatures
.push_tail(entry
);
2437 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2439 glsl_to_tgsi_instruction
*call_inst
;
2440 ir_function_signature
*sig
= ir
->get_callee();
2441 function_entry
*entry
= get_function_signature(sig
);
2444 /* Process in parameters. */
2445 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2446 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2447 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2448 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2450 if (param
->mode
== ir_var_in
||
2451 param
->mode
== ir_var_inout
) {
2452 variable_storage
*storage
= find_variable_storage(param
);
2455 param_rval
->accept(this);
2456 st_src_reg r
= this->result
;
2459 l
.file
= storage
->file
;
2460 l
.index
= storage
->index
;
2462 l
.writemask
= WRITEMASK_XYZW
;
2463 l
.cond_mask
= COND_TR
;
2465 for (i
= 0; i
< type_size(param
->type
); i
++) {
2466 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2474 assert(!sig_iter
.has_next());
2476 /* Emit call instruction */
2477 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2478 call_inst
->function
= entry
;
2480 /* Process out parameters. */
2481 sig_iter
= sig
->parameters
.iterator();
2482 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2483 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2484 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2486 if (param
->mode
== ir_var_out
||
2487 param
->mode
== ir_var_inout
) {
2488 variable_storage
*storage
= find_variable_storage(param
);
2492 r
.file
= storage
->file
;
2493 r
.index
= storage
->index
;
2495 r
.swizzle
= SWIZZLE_NOOP
;
2498 param_rval
->accept(this);
2499 st_dst_reg l
= st_dst_reg(this->result
);
2501 for (i
= 0; i
< type_size(param
->type
); i
++) {
2502 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2510 assert(!sig_iter
.has_next());
2512 /* Process return value. */
2513 this->result
= entry
->return_reg
;
2517 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2519 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2520 st_dst_reg result_dst
, coord_dst
;
2521 glsl_to_tgsi_instruction
*inst
= NULL
;
2522 unsigned opcode
= TGSI_OPCODE_NOP
;
2524 if (ir
->coordinate
) {
2525 ir
->coordinate
->accept(this);
2527 /* Put our coords in a temp. We'll need to modify them for shadow,
2528 * projection, or LOD, so the only case we'd use it as is is if
2529 * we're doing plain old texturing. The optimization passes on
2530 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2532 coord
= get_temp(glsl_type::vec4_type
);
2533 coord_dst
= st_dst_reg(coord
);
2534 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2537 if (ir
->projector
) {
2538 ir
->projector
->accept(this);
2539 projector
= this->result
;
2542 /* Storage for our result. Ideally for an assignment we'd be using
2543 * the actual storage for the result here, instead.
2545 result_src
= get_temp(glsl_type::vec4_type
);
2546 result_dst
= st_dst_reg(result_src
);
2550 opcode
= TGSI_OPCODE_TEX
;
2553 opcode
= TGSI_OPCODE_TXB
;
2554 ir
->lod_info
.bias
->accept(this);
2555 lod_info
= this->result
;
2558 opcode
= TGSI_OPCODE_TXL
;
2559 ir
->lod_info
.lod
->accept(this);
2560 lod_info
= this->result
;
2563 opcode
= TGSI_OPCODE_TXD
;
2564 ir
->lod_info
.grad
.dPdx
->accept(this);
2566 ir
->lod_info
.grad
.dPdy
->accept(this);
2570 opcode
= TGSI_OPCODE_TXQ
;
2571 ir
->lod_info
.lod
->accept(this);
2572 lod_info
= this->result
;
2575 opcode
= TGSI_OPCODE_TXF
;
2576 ir
->lod_info
.lod
->accept(this);
2577 lod_info
= this->result
;
2579 ir
->offset
->accept(this);
2580 offset
= this->result
;
2585 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2587 if (ir
->projector
) {
2588 if (opcode
== TGSI_OPCODE_TEX
) {
2589 /* Slot the projector in as the last component of the coord. */
2590 coord_dst
.writemask
= WRITEMASK_W
;
2591 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2592 coord_dst
.writemask
= WRITEMASK_XYZW
;
2593 opcode
= TGSI_OPCODE_TXP
;
2595 st_src_reg coord_w
= coord
;
2596 coord_w
.swizzle
= SWIZZLE_WWWW
;
2598 /* For the other TEX opcodes there's no projective version
2599 * since the last slot is taken up by LOD info. Do the
2600 * projective divide now.
2602 coord_dst
.writemask
= WRITEMASK_W
;
2603 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2605 /* In the case where we have to project the coordinates "by hand,"
2606 * the shadow comparator value must also be projected.
2608 st_src_reg tmp_src
= coord
;
2609 if (ir
->shadow_comparitor
) {
2610 /* Slot the shadow value in as the second to last component of the
2613 ir
->shadow_comparitor
->accept(this);
2615 tmp_src
= get_temp(glsl_type::vec4_type
);
2616 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2618 /* Projective division not allowed for array samplers. */
2619 assert(!sampler_type
->sampler_array
);
2621 tmp_dst
.writemask
= WRITEMASK_Z
;
2622 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2624 tmp_dst
.writemask
= WRITEMASK_XY
;
2625 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2628 coord_dst
.writemask
= WRITEMASK_XYZ
;
2629 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2631 coord_dst
.writemask
= WRITEMASK_XYZW
;
2632 coord
.swizzle
= SWIZZLE_XYZW
;
2636 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2637 * comparator was put in the correct place (and projected) by the code,
2638 * above, that handles by-hand projection.
2640 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2641 /* Slot the shadow value in as the second to last component of the
2644 ir
->shadow_comparitor
->accept(this);
2646 /* XXX This will need to be updated for cubemap array samplers. */
2647 if (sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2648 sampler_type
->sampler_array
) {
2649 coord_dst
.writemask
= WRITEMASK_W
;
2651 coord_dst
.writemask
= WRITEMASK_Z
;
2654 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2655 coord_dst
.writemask
= WRITEMASK_XYZW
;
2658 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2659 opcode
== TGSI_OPCODE_TXF
) {
2660 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2661 coord_dst
.writemask
= WRITEMASK_W
;
2662 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2663 coord_dst
.writemask
= WRITEMASK_XYZW
;
2666 if (opcode
== TGSI_OPCODE_TXD
)
2667 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2668 else if (opcode
== TGSI_OPCODE_TXQ
)
2669 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2670 else if (opcode
== TGSI_OPCODE_TXF
) {
2671 inst
= emit(ir
, opcode
, result_dst
, coord
);
2673 inst
= emit(ir
, opcode
, result_dst
, coord
);
2675 if (ir
->shadow_comparitor
)
2676 inst
->tex_shadow
= GL_TRUE
;
2678 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2679 this->shader_program
,
2683 inst
->tex_offset_num_offset
= 1;
2684 inst
->tex_offsets
[0].Index
= offset
.index
;
2685 inst
->tex_offsets
[0].File
= offset
.file
;
2686 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2687 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2688 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2691 switch (sampler_type
->sampler_dimensionality
) {
2692 case GLSL_SAMPLER_DIM_1D
:
2693 inst
->tex_target
= (sampler_type
->sampler_array
)
2694 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2696 case GLSL_SAMPLER_DIM_2D
:
2697 inst
->tex_target
= (sampler_type
->sampler_array
)
2698 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2700 case GLSL_SAMPLER_DIM_3D
:
2701 inst
->tex_target
= TEXTURE_3D_INDEX
;
2703 case GLSL_SAMPLER_DIM_CUBE
:
2704 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2706 case GLSL_SAMPLER_DIM_RECT
:
2707 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2709 case GLSL_SAMPLER_DIM_BUF
:
2710 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2712 case GLSL_SAMPLER_DIM_EXTERNAL
:
2713 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2716 assert(!"Should not get here.");
2719 this->result
= result_src
;
2723 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2725 if (ir
->get_value()) {
2729 assert(current_function
);
2731 ir
->get_value()->accept(this);
2732 st_src_reg r
= this->result
;
2734 l
= st_dst_reg(current_function
->return_reg
);
2736 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2737 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2743 emit(ir
, TGSI_OPCODE_RET
);
2747 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2749 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2751 if (ir
->condition
) {
2752 ir
->condition
->accept(this);
2753 this->result
.negate
= ~this->result
.negate
;
2754 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2756 emit(ir
, TGSI_OPCODE_KILP
);
2759 fp
->UsesKill
= GL_TRUE
;
2763 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2765 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2766 glsl_to_tgsi_instruction
*prev_inst
;
2768 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2770 ir
->condition
->accept(this);
2771 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2773 if (this->options
->EmitCondCodes
) {
2774 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2776 /* See if we actually generated any instruction for generating
2777 * the condition. If not, then cook up a move to a temp so we
2778 * have something to set cond_update on.
2780 if (cond_inst
== prev_inst
) {
2781 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2782 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2784 cond_inst
->cond_update
= GL_TRUE
;
2786 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2787 if_inst
->dst
.cond_mask
= COND_NE
;
2789 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2792 this->instructions
.push_tail(if_inst
);
2794 visit_exec_list(&ir
->then_instructions
, this);
2796 if (!ir
->else_instructions
.is_empty()) {
2797 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2798 visit_exec_list(&ir
->else_instructions
, this);
2801 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2804 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2806 result
.file
= PROGRAM_UNDEFINED
;
2808 next_signature_id
= 1;
2810 current_function
= NULL
;
2811 num_address_regs
= 0;
2812 indirect_addr_temps
= false;
2813 indirect_addr_consts
= false;
2814 mem_ctx
= ralloc_context(NULL
);
2817 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2819 ralloc_free(mem_ctx
);
2822 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2829 * Count resources used by the given gpu program (number of texture
2833 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2835 v
->samplers_used
= 0;
2837 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2838 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2840 if (is_tex_instruction(inst
->op
)) {
2841 v
->samplers_used
|= 1 << inst
->sampler
;
2843 prog
->SamplerTargets
[inst
->sampler
] =
2844 (gl_texture_index
)inst
->tex_target
;
2845 if (inst
->tex_shadow
) {
2846 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2851 prog
->SamplersUsed
= v
->samplers_used
;
2852 _mesa_update_shader_textures_used(prog
);
2856 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2857 struct gl_shader_program
*shader_program
,
2858 const char *name
, const glsl_type
*type
,
2861 if (type
->is_record()) {
2862 ir_constant
*field_constant
;
2864 field_constant
= (ir_constant
*)val
->components
.get_head();
2866 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2867 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2868 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2869 type
->fields
.structure
[i
].name
);
2870 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2871 field_type
, field_constant
);
2872 field_constant
= (ir_constant
*)field_constant
->next
;
2877 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2880 fail_link(shader_program
,
2881 "Couldn't find uniform for initializer %s\n", name
);
2885 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2886 ir_constant
*element
;
2887 const glsl_type
*element_type
;
2888 if (type
->is_array()) {
2889 element
= val
->array_elements
[i
];
2890 element_type
= type
->fields
.array
;
2893 element_type
= type
;
2898 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2899 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2900 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2901 conv
[j
] = element
->value
.b
[j
];
2903 values
= (void *)conv
;
2904 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2905 element_type
->vector_elements
,
2908 values
= &element
->value
;
2911 if (element_type
->is_matrix()) {
2912 _mesa_uniform_matrix(ctx
, shader_program
,
2913 element_type
->matrix_columns
,
2914 element_type
->vector_elements
,
2915 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2917 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2918 values
, element_type
->gl_type
);
2926 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
2927 * are read from the given src in this instruction
2930 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
2932 int read_mask
= 0, comp
;
2934 /* Now, given the src swizzle and the written channels, find which
2935 * components are actually read
2937 for (comp
= 0; comp
< 4; ++comp
) {
2938 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
2940 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
2941 read_mask
|= 1 << coord
;
2948 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
2949 * instruction is the first instruction to write to register T0. There are
2950 * several lowering passes done in GLSL IR (e.g. branches and
2951 * relative addressing) that create a large number of conditional assignments
2952 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
2954 * Here is why this conversion is safe:
2955 * CMP T0, T1 T2 T0 can be expanded to:
2961 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
2962 * as the original program. If (T1 < 0.0) evaluates to false, executing
2963 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
2964 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
2965 * because any instruction that was going to read from T0 after this was going
2966 * to read a garbage value anyway.
2969 glsl_to_tgsi_visitor::simplify_cmp(void)
2971 unsigned *tempWrites
;
2972 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
2974 tempWrites
= new unsigned[MAX_TEMPS
];
2978 memset(tempWrites
, 0, sizeof(tempWrites
));
2979 memset(outputWrites
, 0, sizeof(outputWrites
));
2981 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2982 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2983 unsigned prevWriteMask
= 0;
2985 /* Give up if we encounter relative addressing or flow control. */
2986 if (inst
->dst
.reladdr
||
2987 tgsi_get_opcode_info(inst
->op
)->is_branch
||
2988 inst
->op
== TGSI_OPCODE_BGNSUB
||
2989 inst
->op
== TGSI_OPCODE_CONT
||
2990 inst
->op
== TGSI_OPCODE_END
||
2991 inst
->op
== TGSI_OPCODE_ENDSUB
||
2992 inst
->op
== TGSI_OPCODE_RET
) {
2996 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
2997 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
2998 prevWriteMask
= outputWrites
[inst
->dst
.index
];
2999 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3000 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3001 assert(inst
->dst
.index
< MAX_TEMPS
);
3002 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3003 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3006 /* For a CMP to be considered a conditional write, the destination
3007 * register and source register two must be the same. */
3008 if (inst
->op
== TGSI_OPCODE_CMP
3009 && !(inst
->dst
.writemask
& prevWriteMask
)
3010 && inst
->src
[2].file
== inst
->dst
.file
3011 && inst
->src
[2].index
== inst
->dst
.index
3012 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3014 inst
->op
= TGSI_OPCODE_MOV
;
3015 inst
->src
[0] = inst
->src
[1];
3019 delete [] tempWrites
;
3022 /* Replaces all references to a temporary register index with another index. */
3024 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3026 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3027 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3030 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3031 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3032 inst
->src
[j
].index
== index
) {
3033 inst
->src
[j
].index
= new_index
;
3037 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3038 inst
->dst
.index
= new_index
;
3044 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3046 int depth
= 0; /* loop depth */
3047 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3050 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3051 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3053 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3054 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3055 inst
->src
[j
].index
== index
) {
3056 return (depth
== 0) ? i
: loop_start
;
3060 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3063 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3076 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3078 int depth
= 0; /* loop depth */
3079 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3082 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3083 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3085 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3086 return (depth
== 0) ? i
: loop_start
;
3089 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3092 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3105 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3107 int depth
= 0; /* loop depth */
3108 int last
= -1; /* index of last instruction that reads the temporary */
3111 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3112 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3114 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3115 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3116 inst
->src
[j
].index
== index
) {
3117 last
= (depth
== 0) ? i
: -2;
3121 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3123 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3124 if (--depth
== 0 && last
== -2)
3136 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3138 int depth
= 0; /* loop depth */
3139 int last
= -1; /* index of last instruction that writes to the temporary */
3142 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3143 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3145 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3146 last
= (depth
== 0) ? i
: -2;
3148 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3150 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3151 if (--depth
== 0 && last
== -2)
3163 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3164 * channels for copy propagation and updates following instructions to
3165 * use the original versions.
3167 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3168 * will occur. As an example, a TXP production before this pass:
3170 * 0: MOV TEMP[1], INPUT[4].xyyy;
3171 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3172 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3176 * 0: MOV TEMP[1], INPUT[4].xyyy;
3177 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3178 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3180 * which allows for dead code elimination on TEMP[1]'s writes.
3183 glsl_to_tgsi_visitor::copy_propagate(void)
3185 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3186 glsl_to_tgsi_instruction
*,
3187 this->next_temp
* 4);
3188 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3191 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3192 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3194 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3195 || inst
->dst
.index
< this->next_temp
);
3197 /* First, do any copy propagation possible into the src regs. */
3198 for (int r
= 0; r
< 3; r
++) {
3199 glsl_to_tgsi_instruction
*first
= NULL
;
3201 int acp_base
= inst
->src
[r
].index
* 4;
3203 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3204 inst
->src
[r
].reladdr
)
3207 /* See if we can find entries in the ACP consisting of MOVs
3208 * from the same src register for all the swizzled channels
3209 * of this src register reference.
3211 for (int i
= 0; i
< 4; i
++) {
3212 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3213 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3220 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3225 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3226 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3234 /* We've now validated that we can copy-propagate to
3235 * replace this src register reference. Do it.
3237 inst
->src
[r
].file
= first
->src
[0].file
;
3238 inst
->src
[r
].index
= first
->src
[0].index
;
3241 for (int i
= 0; i
< 4; i
++) {
3242 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3243 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3244 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3247 inst
->src
[r
].swizzle
= swizzle
;
3252 case TGSI_OPCODE_BGNLOOP
:
3253 case TGSI_OPCODE_ENDLOOP
:
3254 /* End of a basic block, clear the ACP entirely. */
3255 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3258 case TGSI_OPCODE_IF
:
3262 case TGSI_OPCODE_ENDIF
:
3263 case TGSI_OPCODE_ELSE
:
3264 /* Clear all channels written inside the block from the ACP, but
3265 * leaving those that were not touched.
3267 for (int r
= 0; r
< this->next_temp
; r
++) {
3268 for (int c
= 0; c
< 4; c
++) {
3269 if (!acp
[4 * r
+ c
])
3272 if (acp_level
[4 * r
+ c
] >= level
)
3273 acp
[4 * r
+ c
] = NULL
;
3276 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3281 /* Continuing the block, clear any written channels from
3284 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3285 /* Any temporary might be written, so no copy propagation
3286 * across this instruction.
3288 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3289 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3290 inst
->dst
.reladdr
) {
3291 /* Any output might be written, so no copy propagation
3292 * from outputs across this instruction.
3294 for (int r
= 0; r
< this->next_temp
; r
++) {
3295 for (int c
= 0; c
< 4; c
++) {
3296 if (!acp
[4 * r
+ c
])
3299 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3300 acp
[4 * r
+ c
] = NULL
;
3303 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3304 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3305 /* Clear where it's used as dst. */
3306 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3307 for (int c
= 0; c
< 4; c
++) {
3308 if (inst
->dst
.writemask
& (1 << c
)) {
3309 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3314 /* Clear where it's used as src. */
3315 for (int r
= 0; r
< this->next_temp
; r
++) {
3316 for (int c
= 0; c
< 4; c
++) {
3317 if (!acp
[4 * r
+ c
])
3320 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3322 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3323 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3324 inst
->dst
.writemask
& (1 << src_chan
))
3326 acp
[4 * r
+ c
] = NULL
;
3334 /* If this is a copy, add it to the ACP. */
3335 if (inst
->op
== TGSI_OPCODE_MOV
&&
3336 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3337 !inst
->dst
.reladdr
&&
3339 !inst
->src
[0].reladdr
&&
3340 !inst
->src
[0].negate
) {
3341 for (int i
= 0; i
< 4; i
++) {
3342 if (inst
->dst
.writemask
& (1 << i
)) {
3343 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3344 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3350 ralloc_free(acp_level
);
3355 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3357 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3358 * will occur. As an example, a TXP production after copy propagation but
3361 * 0: MOV TEMP[1], INPUT[4].xyyy;
3362 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3363 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3365 * and after this pass:
3367 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3369 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3370 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3373 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3377 for (i
=0; i
< this->next_temp
; i
++) {
3378 int last_read
= get_last_temp_read(i
);
3381 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3382 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3384 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3397 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3398 * code elimination. This is less primitive than eliminate_dead_code(), as it
3399 * is per-channel and can detect consecutive writes without a read between them
3400 * as dead code. However, there is some dead code that can be eliminated by
3401 * eliminate_dead_code() but not this function - for example, this function
3402 * cannot eliminate an instruction writing to a register that is never read and
3403 * is the only instruction writing to that register.
3405 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3409 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3411 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3412 glsl_to_tgsi_instruction
*,
3413 this->next_temp
* 4);
3414 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3418 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3419 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3421 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3422 || inst
->dst
.index
< this->next_temp
);
3425 case TGSI_OPCODE_BGNLOOP
:
3426 case TGSI_OPCODE_ENDLOOP
:
3427 case TGSI_OPCODE_CONT
:
3428 case TGSI_OPCODE_BRK
:
3429 /* End of a basic block, clear the write array entirely.
3431 * This keeps us from killing dead code when the writes are
3432 * on either side of a loop, even when the register isn't touched
3433 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
3434 * dead code of this type, so it shouldn't make a difference as long as
3435 * the dead code elimination pass in the GLSL compiler does its job.
3437 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3440 case TGSI_OPCODE_ENDIF
:
3441 case TGSI_OPCODE_ELSE
:
3442 /* Promote the recorded level of all channels written inside the
3443 * preceding if or else block to the level above the if/else block.
3445 for (int r
= 0; r
< this->next_temp
; r
++) {
3446 for (int c
= 0; c
< 4; c
++) {
3447 if (!writes
[4 * r
+ c
])
3450 if (write_level
[4 * r
+ c
] == level
)
3451 write_level
[4 * r
+ c
] = level
-1;
3455 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3460 case TGSI_OPCODE_IF
:
3462 /* fallthrough to default case to mark the condition as read */
3465 /* Continuing the block, clear any channels from the write array that
3466 * are read by this instruction.
3468 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3469 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3470 /* Any temporary might be read, so no dead code elimination
3471 * across this instruction.
3473 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3474 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3475 /* Clear where it's used as src. */
3476 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3477 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3478 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3479 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3481 for (int c
= 0; c
< 4; c
++) {
3482 if (src_chans
& (1 << c
)) {
3483 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3491 /* If this instruction writes to a temporary, add it to the write array.
3492 * If there is already an instruction in the write array for one or more
3493 * of the channels, flag that channel write as dead.
3495 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3496 !inst
->dst
.reladdr
&&
3498 for (int c
= 0; c
< 4; c
++) {
3499 if (inst
->dst
.writemask
& (1 << c
)) {
3500 if (writes
[4 * inst
->dst
.index
+ c
]) {
3501 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3504 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3506 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3507 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3513 /* Anything still in the write array at this point is dead code. */
3514 for (int r
= 0; r
< this->next_temp
; r
++) {
3515 for (int c
= 0; c
< 4; c
++) {
3516 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3518 inst
->dead_mask
|= (1 << c
);
3522 /* Now actually remove the instructions that are completely dead and update
3523 * the writemask of other instructions with dead channels.
3525 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3526 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3528 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3530 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3535 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3538 ralloc_free(write_level
);
3539 ralloc_free(writes
);
3544 /* Merges temporary registers together where possible to reduce the number of
3545 * registers needed to run a program.
3547 * Produces optimal code only after copy propagation and dead code elimination
3550 glsl_to_tgsi_visitor::merge_registers(void)
3552 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3553 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3556 /* Read the indices of the last read and first write to each temp register
3557 * into an array so that we don't have to traverse the instruction list as
3559 for (i
=0; i
< this->next_temp
; i
++) {
3560 last_reads
[i
] = get_last_temp_read(i
);
3561 first_writes
[i
] = get_first_temp_write(i
);
3564 /* Start looking for registers with non-overlapping usages that can be
3565 * merged together. */
3566 for (i
=0; i
< this->next_temp
; i
++) {
3567 /* Don't touch unused registers. */
3568 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3570 for (j
=0; j
< this->next_temp
; j
++) {
3571 /* Don't touch unused registers. */
3572 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3574 /* We can merge the two registers if the first write to j is after or
3575 * in the same instruction as the last read from i. Note that the
3576 * register at index i will always be used earlier or at the same time
3577 * as the register at index j. */
3578 if (first_writes
[i
] <= first_writes
[j
] &&
3579 last_reads
[i
] <= first_writes
[j
])
3581 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3583 /* Update the first_writes and last_reads arrays with the new
3584 * values for the merged register index, and mark the newly unused
3585 * register index as such. */
3586 last_reads
[i
] = last_reads
[j
];
3587 first_writes
[j
] = -1;
3593 ralloc_free(last_reads
);
3594 ralloc_free(first_writes
);
3597 /* Reassign indices to temporary registers by reusing unused indices created
3598 * by optimization passes. */
3600 glsl_to_tgsi_visitor::renumber_registers(void)
3605 for (i
=0; i
< this->next_temp
; i
++) {
3606 if (get_first_temp_read(i
) < 0) continue;
3608 rename_temp_register(i
, new_index
);
3612 this->next_temp
= new_index
;
3616 * Returns a fragment program which implements the current pixel transfer ops.
3617 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3620 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3621 glsl_to_tgsi_visitor
*original
,
3622 int scale_and_bias
, int pixel_maps
)
3624 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3625 struct st_context
*st
= st_context(original
->ctx
);
3626 struct gl_program
*prog
= &fp
->Base
.Base
;
3627 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3628 st_src_reg coord
, src0
;
3630 glsl_to_tgsi_instruction
*inst
;
3632 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3633 v
->ctx
= original
->ctx
;
3635 v
->shader_program
= NULL
;
3636 v
->glsl_version
= original
->glsl_version
;
3637 v
->native_integers
= original
->native_integers
;
3638 v
->options
= original
->options
;
3639 v
->next_temp
= original
->next_temp
;
3640 v
->num_address_regs
= original
->num_address_regs
;
3641 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3642 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3643 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3644 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3647 * Get initial pixel color from the texture.
3648 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3650 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3651 src0
= v
->get_temp(glsl_type::vec4_type
);
3652 dst0
= st_dst_reg(src0
);
3653 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3655 inst
->tex_target
= TEXTURE_2D_INDEX
;
3657 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3658 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3659 v
->samplers_used
|= (1 << 0);
3661 if (scale_and_bias
) {
3662 static const gl_state_index scale_state
[STATE_LENGTH
] =
3663 { STATE_INTERNAL
, STATE_PT_SCALE
,
3664 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3665 static const gl_state_index bias_state
[STATE_LENGTH
] =
3666 { STATE_INTERNAL
, STATE_PT_BIAS
,
3667 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3668 GLint scale_p
, bias_p
;
3669 st_src_reg scale
, bias
;
3671 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3672 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3674 /* MAD colorTemp, colorTemp, scale, bias; */
3675 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3676 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3677 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3681 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3682 st_dst_reg temp_dst
= st_dst_reg(temp
);
3684 assert(st
->pixel_xfer
.pixelmap_texture
);
3686 /* With a little effort, we can do four pixel map look-ups with
3687 * two TEX instructions:
3690 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3691 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3692 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3694 inst
->tex_target
= TEXTURE_2D_INDEX
;
3696 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3697 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3698 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3699 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3701 inst
->tex_target
= TEXTURE_2D_INDEX
;
3703 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3704 v
->samplers_used
|= (1 << 1);
3706 /* MOV colorTemp, temp; */
3707 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3710 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3712 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3713 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3714 glsl_to_tgsi_instruction
*newinst
;
3715 st_src_reg src_regs
[3];
3717 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3718 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3720 for (int i
=0; i
<3; i
++) {
3721 src_regs
[i
] = inst
->src
[i
];
3722 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3723 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3725 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3726 src_regs
[i
].index
= src0
.index
;
3728 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3729 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3732 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3733 newinst
->tex_target
= inst
->tex_target
;
3736 /* Make modifications to fragment program info. */
3737 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3738 original
->prog
->Parameters
);
3739 _mesa_free_parameter_list(params
);
3740 count_resources(v
, prog
);
3741 fp
->glsl_to_tgsi
= v
;
3745 * Make fragment program for glBitmap:
3746 * Sample the texture and kill the fragment if the bit is 0.
3747 * This program will be combined with the user's fragment program.
3749 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3752 get_bitmap_visitor(struct st_fragment_program
*fp
,
3753 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3755 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3756 struct st_context
*st
= st_context(original
->ctx
);
3757 struct gl_program
*prog
= &fp
->Base
.Base
;
3758 st_src_reg coord
, src0
;
3760 glsl_to_tgsi_instruction
*inst
;
3762 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3763 v
->ctx
= original
->ctx
;
3765 v
->shader_program
= NULL
;
3766 v
->glsl_version
= original
->glsl_version
;
3767 v
->native_integers
= original
->native_integers
;
3768 v
->options
= original
->options
;
3769 v
->next_temp
= original
->next_temp
;
3770 v
->num_address_regs
= original
->num_address_regs
;
3771 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3772 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3773 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3774 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3776 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3777 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3778 src0
= v
->get_temp(glsl_type::vec4_type
);
3779 dst0
= st_dst_reg(src0
);
3780 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3781 inst
->sampler
= samplerIndex
;
3782 inst
->tex_target
= TEXTURE_2D_INDEX
;
3784 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3785 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3786 v
->samplers_used
|= (1 << samplerIndex
);
3788 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3789 src0
.negate
= NEGATE_XYZW
;
3790 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3791 src0
.swizzle
= SWIZZLE_XXXX
;
3792 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3794 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3796 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3797 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3798 glsl_to_tgsi_instruction
*newinst
;
3799 st_src_reg src_regs
[3];
3801 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3802 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3804 for (int i
=0; i
<3; i
++) {
3805 src_regs
[i
] = inst
->src
[i
];
3806 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3807 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3810 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3811 newinst
->tex_target
= inst
->tex_target
;
3814 /* Make modifications to fragment program info. */
3815 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3816 count_resources(v
, prog
);
3817 fp
->glsl_to_tgsi
= v
;
3820 /* ------------------------- TGSI conversion stuff -------------------------- */
3822 unsigned branch_target
;
3827 * Intermediate state used during shader translation.
3829 struct st_translate
{
3830 struct ureg_program
*ureg
;
3832 struct ureg_dst temps
[MAX_TEMPS
];
3833 struct ureg_src
*constants
;
3834 struct ureg_src
*immediates
;
3835 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3836 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3837 struct ureg_dst address
[1];
3838 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3839 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3841 /* Extra info for handling point size clamping in vertex shader */
3842 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3843 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3844 GLint pointSizeOutIndex
; /**< Temp point size output register */
3845 GLboolean prevInstWrotePointSize
;
3847 const GLuint
*inputMapping
;
3848 const GLuint
*outputMapping
;
3850 /* For every instruction that contains a label (eg CALL), keep
3851 * details so that we can go back afterwards and emit the correct
3852 * tgsi instruction number for each label.
3854 struct label
*labels
;
3855 unsigned labels_size
;
3856 unsigned labels_count
;
3858 /* Keep a record of the tgsi instruction number that each mesa
3859 * instruction starts at, will be used to fix up labels after
3864 unsigned insn_count
;
3866 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3871 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3872 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3874 TGSI_SEMANTIC_VERTEXID
,
3875 TGSI_SEMANTIC_INSTANCEID
3879 * Make note of a branch to a label in the TGSI code.
3880 * After we've emitted all instructions, we'll go over the list
3881 * of labels built here and patch the TGSI code with the actual
3882 * location of each label.
3884 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
3888 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3889 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3890 t
->labels
= (struct label
*)realloc(t
->labels
,
3891 t
->labels_size
* sizeof(struct label
));
3892 if (t
->labels
== NULL
) {
3893 static unsigned dummy
;
3899 i
= t
->labels_count
++;
3900 t
->labels
[i
].branch_target
= branch_target
;
3901 return &t
->labels
[i
].token
;
3905 * Called prior to emitting the TGSI code for each instruction.
3906 * Allocate additional space for instructions if needed.
3907 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
3908 * the next TGSI instruction.
3910 static void set_insn_start(struct st_translate
*t
, unsigned start
)
3912 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3913 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3914 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
3915 if (t
->insn
== NULL
) {
3921 t
->insn
[t
->insn_count
++] = start
;
3925 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
3927 static struct ureg_src
3928 emit_immediate(struct st_translate
*t
,
3929 gl_constant_value values
[4],
3932 struct ureg_program
*ureg
= t
->ureg
;
3937 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
3939 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
3940 case GL_UNSIGNED_INT
:
3942 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
3944 assert(!"should not get here - type must be float, int, uint, or bool");
3945 return ureg_src_undef();
3950 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
3952 static struct ureg_dst
3953 dst_register(struct st_translate
*t
,
3954 gl_register_file file
,
3958 case PROGRAM_UNDEFINED
:
3959 return ureg_dst_undef();
3961 case PROGRAM_TEMPORARY
:
3962 if (ureg_dst_is_undef(t
->temps
[index
]))
3963 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
3965 return t
->temps
[index
];
3967 case PROGRAM_OUTPUT
:
3968 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
3969 t
->prevInstWrotePointSize
= GL_TRUE
;
3971 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
3972 assert(index
< VERT_RESULT_MAX
);
3973 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
3974 assert(index
< FRAG_RESULT_MAX
);
3976 assert(index
< GEOM_RESULT_MAX
);
3978 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3980 return t
->outputs
[t
->outputMapping
[index
]];
3982 case PROGRAM_ADDRESS
:
3983 return t
->address
[index
];
3986 assert(!"unknown dst register file");
3987 return ureg_dst_undef();
3992 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
3994 static struct ureg_src
3995 src_register(struct st_translate
*t
,
3996 gl_register_file file
,
4000 case PROGRAM_UNDEFINED
:
4001 return ureg_src_undef();
4003 case PROGRAM_TEMPORARY
:
4005 assert(index
< Elements(t
->temps
));
4006 if (ureg_dst_is_undef(t
->temps
[index
]))
4007 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4008 return ureg_src(t
->temps
[index
]);
4010 case PROGRAM_NAMED_PARAM
:
4011 case PROGRAM_ENV_PARAM
:
4012 case PROGRAM_LOCAL_PARAM
:
4013 case PROGRAM_UNIFORM
:
4015 return t
->constants
[index
];
4016 case PROGRAM_STATE_VAR
:
4017 case PROGRAM_CONSTANT
: /* ie, immediate */
4019 return ureg_DECL_constant(t
->ureg
, 0);
4021 return t
->constants
[index
];
4023 case PROGRAM_IMMEDIATE
:
4024 return t
->immediates
[index
];
4027 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4028 return t
->inputs
[t
->inputMapping
[index
]];
4030 case PROGRAM_OUTPUT
:
4031 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4032 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4034 case PROGRAM_ADDRESS
:
4035 return ureg_src(t
->address
[index
]);
4037 case PROGRAM_SYSTEM_VALUE
:
4038 assert(index
< Elements(t
->systemValues
));
4039 return t
->systemValues
[index
];
4042 assert(!"unknown src register file");
4043 return ureg_src_undef();
4048 * Create a TGSI ureg_dst register from an st_dst_reg.
4050 static struct ureg_dst
4051 translate_dst(struct st_translate
*t
,
4052 const st_dst_reg
*dst_reg
,
4055 struct ureg_dst dst
= dst_register(t
,
4059 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4062 dst
= ureg_saturate(dst
);
4064 if (dst_reg
->reladdr
!= NULL
)
4065 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4071 * Create a TGSI ureg_src register from an st_src_reg.
4073 static struct ureg_src
4074 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4076 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4078 src
= ureg_swizzle(src
,
4079 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4080 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4081 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4082 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4084 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4085 src
= ureg_negate(src
);
4087 if (src_reg
->reladdr
!= NULL
) {
4088 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4089 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4090 * set the bit for src.Negate. So we have to do the operation manually
4091 * here to work around the compiler's problems. */
4092 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4093 struct ureg_src addr
= ureg_src(t
->address
[0]);
4095 src
.IndirectFile
= addr
.File
;
4096 src
.IndirectIndex
= addr
.Index
;
4097 src
.IndirectSwizzle
= addr
.SwizzleX
;
4099 if (src_reg
->file
!= PROGRAM_INPUT
&&
4100 src_reg
->file
!= PROGRAM_OUTPUT
) {
4101 /* If src_reg->index was negative, it was set to zero in
4102 * src_register(). Reassign it now. But don't do this
4103 * for input/output regs since they get remapped while
4104 * const buffers don't.
4106 src
.Index
= src_reg
->index
;
4113 static struct tgsi_texture_offset
4114 translate_tex_offset(struct st_translate
*t
,
4115 const struct tgsi_texture_offset
*in_offset
)
4117 struct tgsi_texture_offset offset
;
4119 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4121 offset
.File
= TGSI_FILE_IMMEDIATE
;
4122 offset
.Index
= in_offset
->Index
;
4123 offset
.SwizzleX
= in_offset
->SwizzleX
;
4124 offset
.SwizzleY
= in_offset
->SwizzleY
;
4125 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4131 compile_tgsi_instruction(struct st_translate
*t
,
4132 const glsl_to_tgsi_instruction
*inst
)
4134 struct ureg_program
*ureg
= t
->ureg
;
4136 struct ureg_dst dst
[1];
4137 struct ureg_src src
[4];
4138 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4143 num_dst
= num_inst_dst_regs(inst
->op
);
4144 num_src
= num_inst_src_regs(inst
->op
);
4147 dst
[0] = translate_dst(t
,
4151 for (i
= 0; i
< num_src
; i
++)
4152 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4155 case TGSI_OPCODE_BGNLOOP
:
4156 case TGSI_OPCODE_CAL
:
4157 case TGSI_OPCODE_ELSE
:
4158 case TGSI_OPCODE_ENDLOOP
:
4159 case TGSI_OPCODE_IF
:
4160 assert(num_dst
== 0);
4161 ureg_label_insn(ureg
,
4165 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4168 case TGSI_OPCODE_TEX
:
4169 case TGSI_OPCODE_TXB
:
4170 case TGSI_OPCODE_TXD
:
4171 case TGSI_OPCODE_TXL
:
4172 case TGSI_OPCODE_TXP
:
4173 case TGSI_OPCODE_TXQ
:
4174 case TGSI_OPCODE_TXF
:
4175 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4176 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4177 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4182 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4183 texoffsets
, inst
->tex_offset_num_offset
,
4187 case TGSI_OPCODE_SCS
:
4188 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4189 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4202 * Emit the TGSI instructions for inverting and adjusting WPOS.
4203 * This code is unavoidable because it also depends on whether
4204 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4207 emit_wpos_adjustment( struct st_translate
*t
,
4208 const struct gl_program
*program
,
4210 GLfloat adjX
, GLfloat adjY
[2])
4212 struct ureg_program
*ureg
= t
->ureg
;
4214 /* Fragment program uses fragment position input.
4215 * Need to replace instances of INPUT[WPOS] with temp T
4216 * where T = INPUT[WPOS] by y is inverted.
4218 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4219 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4220 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4222 /* XXX: note we are modifying the incoming shader here! Need to
4223 * do this before emitting the constant decls below, or this
4226 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4227 wposTransformState
);
4229 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4230 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4231 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4233 /* First, apply the coordinate shift: */
4234 if (adjX
|| adjY
[0] || adjY
[1]) {
4235 if (adjY
[0] != adjY
[1]) {
4236 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4237 * depending on whether inversion is actually going to be applied
4238 * or not, which is determined by testing against the inversion
4239 * state variable used below, which will be either +1 or -1.
4241 struct ureg_dst adj_temp
= ureg_DECL_temporary(ureg
);
4243 ureg_CMP(ureg
, adj_temp
,
4244 ureg_scalar(wpostrans
, invert
? 2 : 0),
4245 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4246 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4247 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4249 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4250 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4252 wpos_input
= ureg_src(wpos_temp
);
4254 /* MOV wpos_temp, input[wpos]
4256 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4259 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4260 * inversion/identity, or the other way around if we're drawing to an FBO.
4263 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4266 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4268 ureg_scalar(wpostrans
, 0),
4269 ureg_scalar(wpostrans
, 1));
4271 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4274 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4276 ureg_scalar(wpostrans
, 2),
4277 ureg_scalar(wpostrans
, 3));
4280 /* Use wpos_temp as position input from here on:
4282 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4287 * Emit fragment position/ooordinate code.
4290 emit_wpos(struct st_context
*st
,
4291 struct st_translate
*t
,
4292 const struct gl_program
*program
,
4293 struct ureg_program
*ureg
)
4295 const struct gl_fragment_program
*fp
=
4296 (const struct gl_fragment_program
*) program
;
4297 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4298 GLfloat adjX
= 0.0f
;
4299 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4300 boolean invert
= FALSE
;
4302 /* Query the pixel center conventions supported by the pipe driver and set
4303 * adjX, adjY to help out if it cannot handle the requested one internally.
4305 * The bias of the y-coordinate depends on whether y-inversion takes place
4306 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4307 * drawing to an FBO (causes additional inversion), and whether the the pipe
4308 * driver origin and the requested origin differ (the latter condition is
4309 * stored in the 'invert' variable).
4311 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4313 * center shift only:
4318 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4319 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4320 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4321 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4323 * inversion and center shift:
4324 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4325 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4326 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4327 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4329 if (fp
->OriginUpperLeft
) {
4330 /* Fragment shader wants origin in upper-left */
4331 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4332 /* the driver supports upper-left origin */
4334 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4335 /* the driver supports lower-left origin, need to invert Y */
4336 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4343 /* Fragment shader wants origin in lower-left */
4344 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4345 /* the driver supports lower-left origin */
4346 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4347 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4348 /* the driver supports upper-left origin, need to invert Y */
4354 if (fp
->PixelCenterInteger
) {
4355 /* Fragment shader wants pixel center integer */
4356 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4357 /* the driver supports pixel center integer */
4359 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4361 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4362 /* the driver supports pixel center half integer, need to bias X,Y */
4371 /* Fragment shader wants pixel center half integer */
4372 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4373 /* the driver supports pixel center half integer */
4375 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4376 /* the driver supports pixel center integer, need to bias X,Y */
4377 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4378 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4384 /* we invert after adjustment so that we avoid the MOV to temporary,
4385 * and reuse the adjustment ADD instead */
4386 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4390 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4391 * TGSI uses +1 for front, -1 for back.
4392 * This function converts the TGSI value to the GL value. Simply clamping/
4393 * saturating the value to [0,1] does the job.
4396 emit_face_var(struct st_translate
*t
)
4398 struct ureg_program
*ureg
= t
->ureg
;
4399 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4400 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4402 /* MOV_SAT face_temp, input[face] */
4403 face_temp
= ureg_saturate(face_temp
);
4404 ureg_MOV(ureg
, face_temp
, face_input
);
4406 /* Use face_temp as face input from here on: */
4407 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4411 emit_edgeflags(struct st_translate
*t
)
4413 struct ureg_program
*ureg
= t
->ureg
;
4414 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4415 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4417 ureg_MOV(ureg
, edge_dst
, edge_src
);
4421 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4422 * \param program the program to translate
4423 * \param numInputs number of input registers used
4424 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4426 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4427 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4429 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4430 * \param numOutputs number of output registers used
4431 * \param outputMapping maps Mesa fragment program outputs to TGSI
4433 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4434 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4437 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4439 extern "C" enum pipe_error
4440 st_translate_program(
4441 struct gl_context
*ctx
,
4443 struct ureg_program
*ureg
,
4444 glsl_to_tgsi_visitor
*program
,
4445 const struct gl_program
*proginfo
,
4447 const GLuint inputMapping
[],
4448 const ubyte inputSemanticName
[],
4449 const ubyte inputSemanticIndex
[],
4450 const GLuint interpMode
[],
4452 const GLuint outputMapping
[],
4453 const ubyte outputSemanticName
[],
4454 const ubyte outputSemanticIndex
[],
4455 boolean passthrough_edgeflags
)
4457 struct st_translate
*t
;
4459 enum pipe_error ret
= PIPE_OK
;
4461 assert(numInputs
<= Elements(t
->inputs
));
4462 assert(numOutputs
<= Elements(t
->outputs
));
4464 t
= CALLOC_STRUCT(st_translate
);
4466 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4470 memset(t
, 0, sizeof *t
);
4472 t
->procType
= procType
;
4473 t
->inputMapping
= inputMapping
;
4474 t
->outputMapping
= outputMapping
;
4476 t
->pointSizeOutIndex
= -1;
4477 t
->prevInstWrotePointSize
= GL_FALSE
;
4479 if (program
->shader_program
) {
4480 for (i
= 0; i
< program
->shader_program
->NumUserUniformStorage
; i
++) {
4481 struct gl_uniform_storage
*const storage
=
4482 &program
->shader_program
->UniformStorage
[i
];
4484 _mesa_uniform_detach_all_driver_storage(storage
);
4489 * Declare input attributes.
4491 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4492 for (i
= 0; i
< numInputs
; i
++) {
4493 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4494 inputSemanticName
[i
],
4495 inputSemanticIndex
[i
],
4499 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4500 /* Must do this after setting up t->inputs, and before
4501 * emitting constant references, below:
4503 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4506 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4510 * Declare output attributes.
4512 for (i
= 0; i
< numOutputs
; i
++) {
4513 switch (outputSemanticName
[i
]) {
4514 case TGSI_SEMANTIC_POSITION
:
4515 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4516 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4517 outputSemanticIndex
[i
]);
4518 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4520 case TGSI_SEMANTIC_STENCIL
:
4521 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4522 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4523 outputSemanticIndex
[i
]);
4524 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4526 case TGSI_SEMANTIC_COLOR
:
4527 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4528 TGSI_SEMANTIC_COLOR
,
4529 outputSemanticIndex
[i
]);
4532 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4533 ret
= PIPE_ERROR_BAD_INPUT
;
4538 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4539 for (i
= 0; i
< numInputs
; i
++) {
4540 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4542 inputSemanticName
[i
],
4543 inputSemanticIndex
[i
]);
4546 for (i
= 0; i
< numOutputs
; i
++) {
4547 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4548 outputSemanticName
[i
],
4549 outputSemanticIndex
[i
]);
4553 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4555 for (i
= 0; i
< numInputs
; i
++) {
4556 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4559 for (i
= 0; i
< numOutputs
; i
++) {
4560 if (outputSemanticName
[i
] == TGSI_SEMANTIC_CLIPDIST
) {
4561 int mask
= ((1 << (program
->num_clip_distances
- 4*outputSemanticIndex
[i
])) - 1) & TGSI_WRITEMASK_XYZW
;
4562 t
->outputs
[i
] = ureg_DECL_output_masked(ureg
,
4563 outputSemanticName
[i
],
4564 outputSemanticIndex
[i
],
4567 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4568 outputSemanticName
[i
],
4569 outputSemanticIndex
[i
]);
4571 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4572 /* Writing to the point size result register requires special
4573 * handling to implement clamping.
4575 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4576 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4577 /* XXX: note we are modifying the incoming shader here! Need to
4578 * do this before emitting the constant decls below, or this
4581 unsigned pointSizeClampConst
=
4582 _mesa_add_state_reference(proginfo
->Parameters
,
4583 pointSizeClampState
);
4584 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4585 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4586 t
->pointSizeResult
= t
->outputs
[i
];
4587 t
->pointSizeOutIndex
= i
;
4588 t
->outputs
[i
] = psizregtemp
;
4591 if (passthrough_edgeflags
)
4595 /* Declare address register.
4597 if (program
->num_address_regs
> 0) {
4598 assert(program
->num_address_regs
== 1);
4599 t
->address
[0] = ureg_DECL_address(ureg
);
4602 /* Declare misc input registers
4605 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4606 unsigned numSys
= 0;
4607 for (i
= 0; sysInputs
; i
++) {
4608 if (sysInputs
& (1 << i
)) {
4609 unsigned semName
= mesa_sysval_to_semantic
[i
];
4610 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4612 sysInputs
&= ~(1 << i
);
4617 if (program
->indirect_addr_temps
) {
4618 /* If temps are accessed with indirect addressing, declare temporaries
4619 * in sequential order. Else, we declare them on demand elsewhere.
4620 * (Note: the number of temporaries is equal to program->next_temp)
4622 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4623 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4624 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4628 /* Emit constants and uniforms. TGSI uses a single index space for these,
4629 * so we put all the translated regs in t->constants.
4631 if (proginfo
->Parameters
) {
4632 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4633 if (t
->constants
== NULL
) {
4634 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4638 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4639 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4640 case PROGRAM_ENV_PARAM
:
4641 case PROGRAM_LOCAL_PARAM
:
4642 case PROGRAM_STATE_VAR
:
4643 case PROGRAM_NAMED_PARAM
:
4644 case PROGRAM_UNIFORM
:
4645 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4648 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4649 * addressing of the const buffer.
4650 * FIXME: Be smarter and recognize param arrays:
4651 * indirect addressing is only valid within the referenced
4654 case PROGRAM_CONSTANT
:
4655 if (program
->indirect_addr_consts
)
4656 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4658 t
->constants
[i
] = emit_immediate(t
,
4659 proginfo
->Parameters
->ParameterValues
[i
],
4660 proginfo
->Parameters
->Parameters
[i
].DataType
,
4669 /* Emit immediate values.
4671 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4672 if (t
->immediates
== NULL
) {
4673 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4677 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4678 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4679 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4682 /* texture samplers */
4683 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4684 if (program
->samplers_used
& (1 << i
)) {
4685 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4689 /* Emit each instruction in turn:
4691 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4692 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4693 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4695 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4696 /* The previous instruction wrote to the (fake) vertex point size
4697 * result register. Now we need to clamp that value to the min/max
4698 * point size range, putting the result into the real point size
4700 * Note that we can't do this easily at the end of program due to
4701 * possible early return.
4703 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4705 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4706 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4707 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4708 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4709 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4710 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4712 t
->prevInstWrotePointSize
= GL_FALSE
;
4715 /* Fix up all emitted labels:
4717 for (i
= 0; i
< t
->labels_count
; i
++) {
4718 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4719 t
->insn
[t
->labels
[i
].branch_target
]);
4722 if (program
->shader_program
) {
4723 /* This has to be done last. Any operation the can cause
4724 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4725 * program constant) has to happen before creating this linkage.
4727 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4728 if (program
->shader_program
->_LinkedShaders
[i
] == NULL
)
4731 _mesa_associate_uniform_storage(ctx
, program
->shader_program
,
4732 program
->shader_program
->_LinkedShaders
[i
]->Program
->Parameters
);
4741 FREE(t
->immediates
);
4744 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4752 /* ----------------------------- End TGSI code ------------------------------ */
4755 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4756 * generating Mesa IR.
4758 static struct gl_program
*
4759 get_mesa_program(struct gl_context
*ctx
,
4760 struct gl_shader_program
*shader_program
,
4761 struct gl_shader
*shader
,
4762 int num_clip_distances
)
4764 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4765 struct gl_program
*prog
;
4766 struct pipe_screen
* screen
= st_context(ctx
)->pipe
->screen
;
4767 unsigned pipe_shader_type
;
4769 const char *target_string
;
4771 struct gl_shader_compiler_options
*options
=
4772 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4774 switch (shader
->Type
) {
4775 case GL_VERTEX_SHADER
:
4776 target
= GL_VERTEX_PROGRAM_ARB
;
4777 target_string
= "vertex";
4778 pipe_shader_type
= PIPE_SHADER_VERTEX
;
4780 case GL_FRAGMENT_SHADER
:
4781 target
= GL_FRAGMENT_PROGRAM_ARB
;
4782 target_string
= "fragment";
4783 pipe_shader_type
= PIPE_SHADER_FRAGMENT
;
4785 case GL_GEOMETRY_SHADER
:
4786 target
= GL_GEOMETRY_PROGRAM_NV
;
4787 target_string
= "geometry";
4788 pipe_shader_type
= PIPE_SHADER_GEOMETRY
;
4791 assert(!"should not be reached");
4795 validate_ir_tree(shader
->ir
);
4797 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4800 prog
->Parameters
= _mesa_new_parameter_list();
4803 v
->shader_program
= shader_program
;
4804 v
->options
= options
;
4805 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4806 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4807 v
->num_clip_distances
= num_clip_distances
;
4809 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4812 if (!screen
->get_shader_param(screen
, pipe_shader_type
,
4813 PIPE_SHADER_CAP_OUTPUT_READ
)) {
4814 /* Remove reads to output registers, and to varyings in vertex shaders. */
4815 lower_output_reads(shader
->ir
);
4819 /* Emit intermediate IR for main(). */
4820 visit_exec_list(shader
->ir
, v
);
4822 /* Now emit bodies for any functions that were used. */
4824 progress
= GL_FALSE
;
4826 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4827 function_entry
*entry
= (function_entry
*)iter
.get();
4829 if (!entry
->bgn_inst
) {
4830 v
->current_function
= entry
;
4832 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4833 entry
->bgn_inst
->function
= entry
;
4835 visit_exec_list(&entry
->sig
->body
, v
);
4837 glsl_to_tgsi_instruction
*last
;
4838 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4839 if (last
->op
!= TGSI_OPCODE_RET
)
4840 v
->emit(NULL
, TGSI_OPCODE_RET
);
4842 glsl_to_tgsi_instruction
*end
;
4843 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4844 end
->function
= entry
;
4852 /* Print out some information (for debugging purposes) used by the
4853 * optimization passes. */
4854 for (i
=0; i
< v
->next_temp
; i
++) {
4855 int fr
= v
->get_first_temp_read(i
);
4856 int fw
= v
->get_first_temp_write(i
);
4857 int lr
= v
->get_last_temp_read(i
);
4858 int lw
= v
->get_last_temp_write(i
);
4860 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4865 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4867 v
->copy_propagate();
4868 while (v
->eliminate_dead_code_advanced());
4870 /* FIXME: These passes to optimize temporary registers don't work when there
4871 * is indirect addressing of the temporary register space. We need proper
4872 * array support so that we don't have to give up these passes in every
4873 * shader that uses arrays.
4875 if (!v
->indirect_addr_temps
) {
4876 v
->eliminate_dead_code();
4877 v
->merge_registers();
4878 v
->renumber_registers();
4881 /* Write the END instruction. */
4882 v
->emit(NULL
, TGSI_OPCODE_END
);
4884 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4886 printf("GLSL IR for linked %s program %d:\n", target_string
,
4887 shader_program
->Name
);
4888 _mesa_print_ir(shader
->ir
, NULL
);
4894 prog
->Instructions
= NULL
;
4895 prog
->NumInstructions
= 0;
4897 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
4898 count_resources(v
, prog
);
4900 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4902 /* This has to be done last. Any operation the can cause
4903 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4904 * program constant) has to happen before creating this linkage.
4906 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
4907 if (!shader_program
->LinkStatus
) {
4911 struct st_vertex_program
*stvp
;
4912 struct st_fragment_program
*stfp
;
4913 struct st_geometry_program
*stgp
;
4915 switch (shader
->Type
) {
4916 case GL_VERTEX_SHADER
:
4917 stvp
= (struct st_vertex_program
*)prog
;
4918 stvp
->glsl_to_tgsi
= v
;
4920 case GL_FRAGMENT_SHADER
:
4921 stfp
= (struct st_fragment_program
*)prog
;
4922 stfp
->glsl_to_tgsi
= v
;
4924 case GL_GEOMETRY_SHADER
:
4925 stgp
= (struct st_geometry_program
*)prog
;
4926 stgp
->glsl_to_tgsi
= v
;
4929 assert(!"should not be reached");
4937 * Searches through the IR for a declaration of gl_ClipDistance and returns the
4938 * declared size of the gl_ClipDistance array. Returns 0 if gl_ClipDistance is
4939 * not declared in the IR.
4941 int get_clip_distance_size(exec_list
*ir
)
4943 foreach_iter (exec_list_iterator
, iter
, *ir
) {
4944 ir_instruction
*inst
= (ir_instruction
*)iter
.get();
4945 ir_variable
*var
= inst
->as_variable();
4946 if (var
== NULL
) continue;
4947 if (!strcmp(var
->name
, "gl_ClipDistance")) {
4948 return var
->type
->length
;
4958 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4960 struct gl_shader
*shader
;
4961 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4962 type
== GL_GEOMETRY_SHADER_ARB
);
4963 shader
= rzalloc(NULL
, struct gl_shader
);
4965 shader
->Type
= type
;
4966 shader
->Name
= name
;
4967 _mesa_init_shader(ctx
, shader
);
4972 struct gl_shader_program
*
4973 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4975 struct gl_shader_program
*shProg
;
4976 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4978 shProg
->Name
= name
;
4979 _mesa_init_shader_program(ctx
, shProg
);
4986 * Called via ctx->Driver.LinkShader()
4987 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4988 * with code lowering and other optimizations.
4991 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4993 int num_clip_distances
[MESA_SHADER_TYPES
];
4994 assert(prog
->LinkStatus
);
4996 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4997 if (prog
->_LinkedShaders
[i
] == NULL
)
5001 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5002 const struct gl_shader_compiler_options
*options
=
5003 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5005 /* We have to determine the length of the gl_ClipDistance array before
5006 * the array is lowered to two vec4s by lower_clip_distance().
5008 num_clip_distances
[i
] = get_clip_distance_size(ir
);
5014 do_mat_op_to_vec(ir
);
5015 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5016 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
5017 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5019 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5021 progress
= do_common_optimization(ir
, true, true,
5022 options
->MaxUnrollIterations
)
5025 progress
= lower_quadop_vector(ir
, false) || progress
;
5026 progress
= lower_clip_distance(ir
) || progress
;
5028 if (options
->MaxIfDepth
== 0)
5029 progress
= lower_discard(ir
) || progress
;
5031 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5033 if (options
->EmitNoNoise
)
5034 progress
= lower_noise(ir
) || progress
;
5036 /* If there are forms of indirect addressing that the driver
5037 * cannot handle, perform the lowering pass.
5039 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5040 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5042 lower_variable_index_to_cond_assign(ir
,
5043 options
->EmitNoIndirectInput
,
5044 options
->EmitNoIndirectOutput
,
5045 options
->EmitNoIndirectTemp
,
5046 options
->EmitNoIndirectUniform
)
5049 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5052 validate_ir_tree(ir
);
5055 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5056 struct gl_program
*linked_prog
;
5058 if (prog
->_LinkedShaders
[i
] == NULL
)
5061 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
],
5062 num_clip_distances
[i
]);
5065 static const GLenum targets
[] = {
5066 GL_VERTEX_PROGRAM_ARB
,
5067 GL_FRAGMENT_PROGRAM_ARB
,
5068 GL_GEOMETRY_PROGRAM_NV
5071 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5073 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5074 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5076 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5081 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5088 st_translate_stream_output_info(struct glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5089 const GLuint outputMapping
[],
5090 struct pipe_stream_output_info
*so
)
5092 static unsigned comps_to_mask
[] = {
5100 struct gl_transform_feedback_info
*info
=
5101 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5103 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5104 assert(info
->Outputs
[i
].NumComponents
< Elements(comps_to_mask
));
5105 so
->output
[i
].register_index
=
5106 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5107 so
->output
[i
].register_mask
=
5108 comps_to_mask
[info
->Outputs
[i
].NumComponents
]
5109 << info
->Outputs
[i
].ComponentOffset
;
5110 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5112 so
->num_outputs
= info
->NumOutputs
;